Antoine/Atomizer
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

chore: major repo cleanup - remove dead code and cruft

Browse Source 
Remove ~24K lines of dead code for a lean rebuild foundation:

- Remove atomizer-field/ (neural field predictor experiment, concept archived in docs)
- Remove generated_extractors/, generated_hooks/ (legacy generator outputs)
- Remove optimization_validation/ (empty skeleton)
- Remove reports/ (superseded by optimization_engine/reporting/)
- Remove root-level stale files: DEVELOPMENT.md, INSTALL_INSTRUCTIONS.md,
  config.py, atomizer_paths.py, optimization_config.json, train_neural.bat,
  generate_training_data.py, run_training_fea.py, migrate_imports.py
- Update .gitignore for introspection caches and insight outputs

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
Anto01
2026-02-09 14:24:41 -05:00
parent 8d9d55356c
commit 857c01e7ca
73 changed files with 50 additions and 24073 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
.gitignore vendored
View File

@@ -127,5 +127,14 @@ backend_stderr.log
# Auto-generated documentation (regenerate with: python -m optimization_engine.auto_doc all)
docs/generated/
# NX model introspection caches (generated)
**/_introspection_*.json
**/_introspection_cache.json
**/_temp_introspection.json
**/params.exp
# Insight outputs (generated)
**/3_insights/
# Malformed filenames (Windows path used as filename)
C:*

619
DEVELOPMENT.md
View File

@@ -1,619 +0,0 @@
# Atomizer Development Guide
**Last Updated**: 2025-11-21
**Current Phase**: Phase 3.2 - Integration Sprint + Documentation
**Status**: 🟢 Core Complete (100%) | ✅ Protocols 10/11/13 Active (100%) | 🎯 Dashboard Live (95%) | 📚 Documentation Reorganized
📘 **Quick Links**:
- [Protocol Specifications](docs/PROTOCOLS.md) - All active protocols consolidated
- [Documentation Index](docs/00_INDEX.md) - Complete documentation navigation
- [README](README.md) - Project overview and quick start
---
## Table of Contents
1. [Current Phase](#current-phase)
2. [Completed Features](#completed-features)
3. [Active Development](#active-development)
4. [Known Issues](#known-issues)
5. [Testing Status](#testing-status)
6. [Phase-by-Phase Progress](#phase-by-phase-progress)
---
## Current Phase
### Phase 3.2: Integration Sprint (🎯 TOP PRIORITY)
**Goal**: Connect LLM intelligence components to production workflow
**Timeline**: 2-4 weeks (Started 2025-11-17)
**Status**: LLM components built and tested individually (85% complete). Need to wire them into production runner.
📋 **Detailed Plan**: [docs/PHASE_3_2_INTEGRATION_PLAN.md](docs/PHASE_3_2_INTEGRATION_PLAN.md)
**Critical Path**:
#### Week 1: Make LLM Mode Accessible (16 hours)
- [ ] **1.1** Create unified entry point `optimization_engine/run_optimization.py` (4h)
- Add `--llm` flag for natural language mode
- Add `--request` parameter for natural language input
- Support both LLM and traditional JSON modes
- Preserve backward compatibility
- [ ] **1.2** Wire LLMOptimizationRunner to production (8h)
- Connect LLMWorkflowAnalyzer to entry point
- Bridge LLMOptimizationRunner → OptimizationRunner
- Pass model updater and simulation runner callables
- Integrate with existing hook system
- [ ] **1.3** Create minimal example (2h)
- Create `examples/llm_mode_demo.py`
- Show natural language → optimization results
- Compare traditional (100 lines) vs LLM (3 lines)
- [ ] **1.4** End-to-end integration test (2h)
- Test with simple_beam_optimization study
- Verify extractors generated correctly
- Validate output matches manual mode
#### Week 2: Robustness & Safety (16 hours)
- [ ] **2.1** Code validation pipeline (6h)
- Create `optimization_engine/code_validator.py`
- Implement syntax validation (ast.parse)
- Implement security scanning (whitelist imports)
- Implement test execution on example OP2
- Add retry with LLM feedback on failure
- [ ] **2.2** Graceful fallback mechanisms (4h)
- Wrap all LLM calls in try/except
- Provide clear error messages
- Offer fallback to manual mode
- Never crash on LLM failure
- [ ] **2.3** LLM audit trail (3h)
- Create `optimization_engine/llm_audit.py`
- Log all LLM requests and responses
- Log generated code with prompts
- Create `llm_audit.json` in study output
- [ ] **2.4** Failure scenario testing (3h)
- Test invalid natural language request
- Test LLM unavailable
- Test generated code syntax errors
- Test validation failures
#### Week 3: Learning System (12 hours)
- [ ] **3.1** Knowledge base implementation (4h)
- Create `optimization_engine/knowledge_base.py`
- Implement `save_session()` - Save successful workflows
- Implement `search_templates()` - Find similar patterns
- Add confidence scoring
- [ ] **3.2** Template extraction (4h)
- Extract reusable patterns from generated code
- Parameterize variable parts
- Save templates with usage examples
- Implement template application to new requests
- [ ] **3.3** ResearchAgent integration (4h)
- Complete ResearchAgent implementation
- Integrate into ExtractorOrchestrator error handling
- Add user example collection workflow
- Save learned knowledge to knowledge base
#### Week 4: Documentation & Discoverability (8 hours)
- [ ] **4.1** Update README (2h)
- Add "🤖 LLM-Powered Mode" section
- Show example command with natural language
- Link to detailed docs
- [ ] **4.2** Create LLM mode documentation (3h)
- Create `docs/LLM_MODE.md`
- Explain how LLM mode works
- Provide usage examples
- Add troubleshooting guide
- [ ] **4.3** Create demo video/GIF (1h)
- Record terminal session
- Show before/after (100 lines → 3 lines)
- Create animated GIF for README
- [ ] **4.4** Update all planning docs (2h)
- Update DEVELOPMENT.md status
- Update DEVELOPMENT_GUIDANCE.md (80-90% → 90-95%)
- Mark Phase 3.2 as ✅ Complete
---
## Completed Features
### ✅ Live Dashboard System (Completed 2025-11-21)
#### Backend (FastAPI + WebSocket)
- [x] **FastAPI Backend** ([atomizer-dashboard/backend/](atomizer-dashboard/backend/))
- REST API endpoints for study management
- WebSocket streaming with file watching (Watchdog)
- Real-time updates (<100ms latency)
- CORS configured for local development
- [x] **REST API Endpoints** ([backend/api/routes/optimization.py](atomizer-dashboard/backend/api/routes/optimization.py))
- `GET /api/optimization/studies` - List all studies
- `GET /api/optimization/studies/{id}/status` - Get study status
- `GET /api/optimization/studies/{id}/history` - Get trial history
- `GET /api/optimization/studies/{id}/pruning` - Get pruning diagnostics
- [x] **WebSocket Streaming** ([backend/api/websocket/optimization_stream.py](atomizer-dashboard/backend/api/websocket/optimization_stream.py))
- File watching on `optimization_history_incremental.json`
- Real-time trial updates via WebSocket
- Pruning alerts and progress updates
- Automatic observer lifecycle management
#### Frontend (HTML + Chart.js)
- [x] **Enhanced Live Dashboard** ([atomizer-dashboard/dashboard-enhanced.html](atomizer-dashboard/dashboard-enhanced.html))
- Real-time WebSocket updates
- Interactive convergence chart (Chart.js)
- Parameter space scatter plot
- Pruning alerts (toast notifications)
- Data export (JSON/CSV)
- Study auto-discovery and selection
- Metric dashboard (trials, best value, pruned count)
#### React Frontend (In Progress)
- [x] **Project Configuration** ([atomizer-dashboard/frontend/](atomizer-dashboard/frontend/))
- React 18 + Vite 5 + TypeScript 5.2
- TailwindCSS 3.3 for styling
- Recharts 2.10 for charts
- Complete build configuration
- [x] **TypeScript Types** ([frontend/src/types/](atomizer-dashboard/frontend/src/types/))
- Complete type definitions for API data
- WebSocket message types
- Chart data structures
- [x] **Custom Hooks** ([frontend/src/hooks/useWebSocket.ts](atomizer-dashboard/frontend/src/hooks/useWebSocket.ts))
- WebSocket connection management
- Auto-reconnection with exponential backoff
- Type-safe message routing
- [x] **Reusable Components** ([frontend/src/components/](atomizer-dashboard/frontend/src/components/))
- Card, MetricCard, Badge, StudyCard components
- TailwindCSS styling with dark theme
- [ ] **Dashboard Page** (Pending manual completion)
- Need to run `npm install`
- Create main.tsx, App.tsx, Dashboard.tsx
- Integrate Recharts for charts
- Test end-to-end
#### Documentation
- [x] **Dashboard Master Plan** ([docs/DASHBOARD_MASTER_PLAN.md](docs/DASHBOARD_MASTER_PLAN.md))
- Complete 3-page architecture (Configurator, Live Dashboard, Results Viewer)
- Tech stack recommendations
- Implementation phases
- [x] **Implementation Status** ([docs/DASHBOARD_IMPLEMENTATION_STATUS.md](docs/DASHBOARD_IMPLEMENTATION_STATUS.md))
- Current progress tracking
- Testing instructions
- Next steps
- [x] **React Implementation Guide** ([docs/DASHBOARD_REACT_IMPLEMENTATION.md](docs/DASHBOARD_REACT_IMPLEMENTATION.md))
- Complete templates for remaining components
- Recharts integration examples
- Troubleshooting guide
- [x] **Session Summary** ([docs/DASHBOARD_SESSION_SUMMARY.md](docs/DASHBOARD_SESSION_SUMMARY.md))
- Features demonstrated
- How to use the dashboard
- Architecture explanation
### ✅ Phase 1: Plugin System & Infrastructure (Completed 2025-01-16)
#### Core Architecture
- [x] **Hook Manager** ([optimization_engine/plugins/hook_manager.py](optimization_engine/plugins/hook_manager.py))
- Hook registration with priority-based execution
- Auto-discovery from plugin directories
- Context passing to all hooks
- Execution history tracking
- [x] **Lifecycle Hooks**
- `pre_solve`: Execute before solver launch
- `post_solve`: Execute after solve, before extraction
- `post_extraction`: Execute after result extraction
#### Logging Infrastructure
- [x] **Detailed Trial Logs** ([detailed_logger.py](optimization_engine/plugins/pre_solve/detailed_logger.py))
- Per-trial log files in `optimization_results/trial_logs/`
- Complete iteration trace with timestamps
- Design variables, configuration, timeline
- Extracted results and constraint evaluations
- [x] **High-Level Optimization Log** ([optimization_logger.py](optimization_engine/plugins/pre_solve/optimization_logger.py))
- `optimization.log` file tracking overall progress
- Configuration summary header
- Compact START/COMPLETE entries per trial
- Easy to scan format for monitoring
- [x] **Result Appenders**
- [log_solve_complete.py](optimization_engine/plugins/post_solve/log_solve_complete.py) - Appends solve completion to trial logs
- [log_results.py](optimization_engine/plugins/post_extraction/log_results.py) - Appends extracted results to trial logs
- [optimization_logger_results.py](optimization_engine/plugins/post_extraction/optimization_logger_results.py) - Appends results to optimization.log
#### Project Organization
- [x] **Studies Structure** ([studies/](studies/))
- Standardized folder layout with `model/`, `optimization_results/`, `analysis/`
- Comprehensive documentation in [studies/README.md](studies/README.md)
- Example study: [bracket_stress_minimization/](studies/bracket_stress_minimization/)
- Template structure for future studies
- [x] **Path Resolution** ([atomizer_paths.py](atomizer_paths.py))
- Intelligent project root detection using marker files
- Helper functions: `root()`, `optimization_engine()`, `studies()`, `tests()`
- `ensure_imports()` for robust module imports
- Works regardless of script location
#### Testing
- [x] **Hook Validation Test** ([test_hooks_with_bracket.py](tests/test_hooks_with_bracket.py))
- Verifies hook loading and execution
- Tests 3 trials with dummy data
- Checks hook execution history
- [x] **Integration Tests**
- [run_5trial_test.py](tests/run_5trial_test.py) - Quick 5-trial optimization
- [test_journal_optimization.py](tests/test_journal_optimization.py) - Full optimization test
#### Runner Enhancements
- [x] **Context Passing** ([runner.py:332,365,412](optimization_engine/runner.py))
- `output_dir` passed to all hook contexts
- Trial number, design variables, extracted results
- Configuration dictionary available to hooks
### ✅ Core Engine (Pre-Phase 1)
- [x] Optuna integration with TPE sampler
- [x] Multi-objective optimization support
- [x] NX journal execution ([nx_solver.py](optimization_engine/nx_solver.py))
- [x] Expression updates ([nx_updater.py](optimization_engine/nx_updater.py))
- [x] OP2 result extraction (stress, displacement)
- [x] Study management with resume capability
- [x] Web dashboard (real-time monitoring)
- [x] Precision control (4-decimal rounding)
---
## Active Development
### In Progress - Dashboard (High Priority)
- [x] Backend API complete (FastAPI + WebSocket)
- [x] HTML dashboard with Chart.js complete
- [x] React project structure and configuration complete
- [ ] **Complete React frontend** (Awaiting manual npm install)
- [ ] Run `npm install` in frontend directory
- [ ] Create main.tsx and App.tsx
- [ ] Create Dashboard.tsx with Recharts
- [ ] Test end-to-end with live optimization
### Up Next - Dashboard (Next Session)
- [ ] Study Configurator page (React)
- [ ] Results Report Viewer page (React)
- [ ] LLM chat interface integration (future)
- [ ] Docker deployment configuration
### In Progress - Phase 3.2 Integration
- [ ] Feature registry creation (Phase 2, Week 1)
- [ ] Claude skill definition (Phase 2, Week 1)
### Up Next (Phase 2, Week 2)
- [ ] Natural language parser
- [ ] Intent classification system
- [ ] Entity extraction for optimization parameters
- [ ] Conversational workflow manager
### Backlog (Phase 3+)
- [ ] Custom function generator (RSS, weighted objectives)
- [ ] Journal script generator
- [ ] Code validation pipeline
- [ ] Result analyzer with statistical analysis
- [ ] Surrogate quality checker
- [ ] HTML/PDF report generator
---
## Known Issues
### Critical
- None currently
### Minor
- [ ] `.claude/settings.local.json` modified during development (contains user-specific settings)
- [ ] Some old bash background processes still running from previous tests
### Documentation
- [ ] Need to add examples of custom hooks to studies/README.md
- [ ] Missing API documentation for hook_manager methods
- [ ] No developer guide for creating new plugins
---
## Testing Status
### Automated Tests
- ✅ **Hook system** - `test_hooks_with_bracket.py` passing
- ✅ **5-trial integration** - `run_5trial_test.py` working
- ✅ **Full optimization** - `test_journal_optimization.py` functional
- ⏳ **Unit tests** - Need to create for individual modules
- ⏳ **CI/CD pipeline** - Not yet set up
### Manual Testing
- ✅ Bracket optimization (50 trials)
- ✅ Log file generation in correct locations
- ✅ Hook execution at all lifecycle points
- ✅ Path resolution across different script locations
- ✅ **Dashboard backend** - REST API and WebSocket tested successfully
- ✅ **HTML dashboard** - Live updates working with Chart.js
- ⏳ **React dashboard** - Pending npm install and completion
- ⏳ Resume functionality with config validation
### Test Coverage
- Hook manager: ~80% (core functionality tested)
- Logging plugins: 100% (tested via integration tests)
- Path resolution: 100% (tested in all scripts)
- Result extractors: ~70% (basic tests exist)
- **Dashboard backend**: ~90% (REST endpoints and WebSocket tested)
- **Dashboard frontend**: ~60% (HTML version tested, React pending)
- Overall: ~65% estimated
---
## Phase-by-Phase Progress
### Phase 1: Plugin System ✅ (100% Complete)
**Completed** (2025-01-16):
- [x] Hook system for optimization lifecycle
- [x] Plugin auto-discovery and registration
- [x] Hook manager with priority-based execution
- [x] Detailed per-trial logs (`trial_logs/`)
- [x] High-level optimization log (`optimization.log`)
- [x] Context passing system for hooks
- [x] Studies folder structure
- [x] Comprehensive studies documentation
- [x] Model file organization (`model/` folder)
- [x] Intelligent path resolution
- [x] Test suite for hook system
**Deferred to Future Phases**:
- Feature registry → Phase 2 (with LLM interface)
- `pre_mesh` and `post_mesh` hooks → Future (not needed for current workflow)
- Custom objective/constraint registration → Phase 3 (Code Generation)
---
### Phase 2: LLM Integration 🟡 (0% Complete)
**Target**: 2 weeks (Started 2025-01-16)
#### Week 1 Todos (Feature Registry & Claude Skill)
- [ ] Create `optimization_engine/feature_registry.json`
- [ ] Extract all current capabilities
- [ ] Draft `.claude/skills/atomizer.md`
- [ ] Test LLM's ability to navigate codebase
#### Week 2 Todos (Natural Language Interface)
- [ ] Implement intent classifier
- [ ] Build entity extractor
- [ ] Create workflow manager
- [ ] Test end-to-end: "Create a stress minimization study"
**Success Criteria**:
- [ ] LLM can create optimization from natural language in <5 turns
- [ ] 90% of user requests understood correctly
- [ ] Zero manual JSON editing required
---
### Phase 3: Code Generation ⏳ (Not Started)
**Target**: 3 weeks
**Key Deliverables**:
- [ ] Custom function generator
- [ ] RSS (Root Sum Square) template
- [ ] Weighted objectives template
- [ ] Custom constraints template
- [ ] Journal script generator
- [ ] Code validation pipeline
- [ ] Safe execution environment
**Success Criteria**:
- [ ] LLM generates 10+ custom functions with zero errors
- [ ] All generated code passes safety validation
- [ ] Users save 50% time vs. manual coding
---
### Phase 4: Analysis & Decision Support ⏳ (Not Started)
**Target**: 3 weeks
**Key Deliverables**:
- [ ] Result analyzer (convergence, sensitivity, outliers)
- [ ] Surrogate model quality checker (R², CV score, confidence intervals)
- [ ] Decision assistant (trade-offs, what-if analysis, recommendations)
**Success Criteria**:
- [ ] Surrogate quality detection 95% accurate
- [ ] Recommendations lead to 30% faster convergence
- [ ] Users report higher confidence in results
---
### Phase 5: Automated Reporting ⏳ (Not Started)
**Target**: 2 weeks
**Key Deliverables**:
- [ ] Report generator with Jinja2 templates
- [ ] Multi-format export (HTML, PDF, Markdown, JSON)
- [ ] LLM-written narrative explanations
**Success Criteria**:
- [ ] Reports generated in <30 seconds
- [ ] Narrative quality rated 4/5 by engineers
- [ ] 80% of reports used without manual editing
---
### Phase 6: NX MCP Enhancement ⏳ (Not Started)
**Target**: 4 weeks
**Key Deliverables**:
- [ ] NX documentation MCP server
- [ ] Advanced NX operations library
- [ ] Feature bank with 50+ pre-built operations
**Success Criteria**:
- [ ] NX MCP answers 95% of API questions correctly
- [ ] Feature bank covers 80% of common workflows
- [ ] Users write 50% less manual journal code
---
### Phase 7: Self-Improving System ⏳ (Not Started)
**Target**: 4 weeks
**Key Deliverables**:
- [ ] Feature learning system
- [ ] Best practices database
- [ ] Continuous documentation generation
**Success Criteria**:
- [ ] 20+ user-contributed features in library
- [ ] Pattern recognition identifies 10+ best practices
- [ ] Documentation auto-updates with zero manual effort
---
## Development Commands
### Running Dashboard
```bash
# Start backend server
cd atomizer-dashboard/backend
python -m uvicorn api.main:app --reload --host 0.0.0.0 --port 8000
# Access HTML dashboard (current)
# Open browser: http://localhost:8000
# Start React frontend (when ready)
cd atomizer-dashboard/frontend
npm install # First time only
npm run dev # Starts on http://localhost:3000
```
### Running Tests
```bash
# Hook validation (3 trials, fast)
python tests/test_hooks_with_bracket.py
# Quick integration test (5 trials)
python tests/run_5trial_test.py
# Full optimization test
python tests/test_journal_optimization.py
```
### Code Quality
```bash
# Run linter (when available)
# pylint optimization_engine/
# Run type checker (when available)
# mypy optimization_engine/
# Run all tests (when test suite is complete)
# pytest tests/
```
### Git Workflow
```bash
# Stage all changes
git add .
# Commit with conventional commits format
git commit -m "feat: description" # New feature
git commit -m "fix: description" # Bug fix
git commit -m "docs: description" # Documentation
git commit -m "test: description" # Tests
git commit -m "refactor: description" # Code refactoring
# Push to GitHub
git push origin main
```
---
## Documentation
### For Developers
- [DEVELOPMENT_ROADMAP.md](DEVELOPMENT_ROADMAP.md) - Strategic vision and phases
- [studies/README.md](studies/README.md) - Studies folder organization
- [CHANGELOG.md](CHANGELOG.md) - Version history
### Dashboard Documentation
- [docs/DASHBOARD_MASTER_PLAN.md](docs/DASHBOARD_MASTER_PLAN.md) - Complete architecture blueprint
- [docs/DASHBOARD_IMPLEMENTATION_STATUS.md](docs/DASHBOARD_IMPLEMENTATION_STATUS.md) - Current progress
- [docs/DASHBOARD_REACT_IMPLEMENTATION.md](docs/DASHBOARD_REACT_IMPLEMENTATION.md) - React implementation guide
- [docs/DASHBOARD_SESSION_SUMMARY.md](docs/DASHBOARD_SESSION_SUMMARY.md) - Session summary
- [atomizer-dashboard/README.md](atomizer-dashboard/README.md) - Dashboard quick start
- [atomizer-dashboard/backend/README.md](atomizer-dashboard/backend/README.md) - Backend API docs
- [atomizer-dashboard/frontend/README.md](atomizer-dashboard/frontend/README.md) - Frontend setup guide
### For Users
- [README.md](README.md) - Project overview and quick start
- [docs/INDEX.md](docs/INDEX.md) - Complete documentation index
- [docs/](docs/) - Additional documentation
---
## Notes
### Architecture Decisions
- **Hook system**: Chose priority-based execution to allow precise control of plugin order
- **Path resolution**: Used marker files instead of environment variables for simplicity
- **Logging**: Two-tier system (detailed trial logs + high-level optimization.log) for different use cases
### Performance Considerations
- Hook execution adds <1s overhead per trial (acceptable for FEA simulations)
- Path resolution caching could improve startup time (future optimization)
- Log file sizes grow linearly with trials (~10KB per trial)
### Future Considerations
- Consider moving to structured logging (JSON) for easier parsing
- May need database for storing hook execution history (currently in-memory)
- Dashboard integration will require WebSocket for real-time log streaming
---
**Last Updated**: 2025-11-21
**Maintained by**: Antoine Polvé (antoine@atomaste.com)
**Repository**: [GitHub - Atomizer](https://github.com/yourusername/Atomizer)
---
## Recent Updates (November 21, 2025)
### Dashboard System Implementation ✅
- **Backend**: FastAPI + WebSocket with real-time file watching complete
- **HTML Dashboard**: Functional dashboard with Chart.js, data export, pruning alerts
- **React Setup**: Complete project configuration, types, hooks, components
- **Documentation**: 5 comprehensive markdown documents covering architecture, implementation, and usage
### Next Immediate Steps
1. Run `npm install` in `atomizer-dashboard/frontend`
2. Create `main.tsx`, `App.tsx`, and `Dashboard.tsx` using provided templates
3. Test React dashboard with live optimization
4. Build Study Configurator page (next major feature)

63
INSTALL_INSTRUCTIONS.md
View File

@@ -1,63 +0,0 @@
# Atomizer Installation Guide
## Step 1: Install Miniconda (Recommended)
1. Download Miniconda from: https://docs.conda.io/en/latest/miniconda.html
- Choose: **Miniconda3 Windows 64-bit**
2. Run the installer:
- Check "Add Miniconda3 to my PATH environment variable"
- Check "Register Miniconda3 as my default Python"
3. Restart your terminal/VSCode after installation
## Step 2: Create Atomizer Environment
Open **Anaconda Prompt** (or any terminal after restart) and run:
```bash
cd C:\Users\Antoine\Atomizer
conda env create -f environment.yml
conda activate atomizer
```
## Step 3: Install PyTorch with GPU Support (Optional but Recommended)
If you have an NVIDIA GPU:
```bash
conda activate atomizer
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install torch-geometric
```
## Step 4: Verify Installation
```bash
conda activate atomizer
python -c "import torch; import optuna; import pyNastran; print('All imports OK!')"
python -c "import torch; print(f'CUDA available: {torch.cuda.is_available()}')"
```
## Step 5: Train Neural Network
```bash
conda activate atomizer
cd C:\Users\Antoine\Atomizer\atomizer-field
python train_parametric.py --train_dir ../atomizer_field_training_data/bracket_stiffness_optimization_atomizerfield --epochs 100 --output_dir runs/bracket_model
```
## Quick Commands Reference
```bash
# Activate environment (do this every time you open a new terminal)
conda activate atomizer
# Train neural network
cd C:\Users\Antoine\Atomizer\atomizer-field
python train_parametric.py --train_dir ../atomizer_field_training_data/bracket_stiffness_optimization_atomizerfield --epochs 100
# Run optimization with neural acceleration
cd C:\Users\Antoine\Atomizer\studies\bracket_stiffness_optimization_atomizerfield
python run_optimization.py --run --trials 100 --enable-nn
```

56
atomizer-field/.gitignore vendored
View File

@@ -1,56 +0,0 @@
# Python
__pycache__/
*.py[cod]
*$py.class
*.so
.Python
env/
venv/
ENV/
*.egg-info/
dist/
build/
# Jupyter Notebooks
.ipynb_checkpoints/
*.ipynb
# IDEs
.vscode/
.idea/
*.swp
*.swo
*~
# OS
.DS_Store
Thumbs.db
# Data files (large)
*.op2
*.bdf
*.dat
*.f06
*.pch
*.h5
*.hdf5
# Training data
training_data/
checkpoints/
runs/
logs/
# Test outputs
test_case_*/
visualization_images/
# Temporary files
*.tmp
*.log
*.bak
*.orig
# Environment
atomizer_env/
.conda/

635
atomizer-field/ATOMIZER_FIELD_STATUS_REPORT.md
View File

@@ -1,635 +0,0 @@
# AtomizerField - Complete Status Report
**Date:** November 24, 2025
**Version:** 1.0
**Status:** ✅ Core System Operational, Unit Issues Resolved
---
## Executive Summary
**AtomizerField** is a neural field learning system that replaces traditional FEA simulations with graph neural networks, providing **1000× faster predictions** for structural optimization.
### Current Status
- ✅ **Core pipeline working**: BDF/OP2 → Neural format → GNN inference
- ✅ **Test case validated**: Simple Beam (5,179 nodes, 4,866 elements)
- ✅ **Unit system understood**: MN-MM system (kPa stress, N forces, mm length)
- ⚠️ **Not yet trained**: Neural network has random weights
- 🔜 **Next step**: Generate training data and train model
---
## What AtomizerField Does
### 1. Data Pipeline ✅ WORKING
**Purpose:** Convert Nastran FEA results into neural network training data
**Input:**
- BDF file (geometry, materials, loads, BCs)
- OP2 file (FEA results: displacement, stress, reactions)
**Output:**
- JSON metadata (mesh, materials, loads, statistics)
- HDF5 arrays (coordinates, displacement, stress fields)
**What's Extracted:**
- ✅ Mesh: 5,179 nodes, 4,866 CQUAD4 shell elements
- ✅ Materials: Young's modulus, Poisson's ratio, density
- ✅ Boundary conditions: SPCs, MPCs (if present)
- ✅ Loads: 35 point forces with directions
- ✅ Displacement field: 6 DOF per node (Tx, Ty, Tz, Rx, Ry, Rz)
- ✅ Stress field: 8 components per element (σxx, σyy, τxy, principals, von Mises)
- ✅ Reaction forces: 6 DOF per node
**Performance:**
- Parse time: 1.27 seconds
- Data size: JSON 1.7 MB, HDF5 546 KB
### 2. Graph Neural Network ✅ ARCHITECTURE WORKING
**Purpose:** Learn FEA physics to predict displacement/stress from geometry/loads
**Architecture:**
- Type: Graph Neural Network (PyTorch Geometric)
- Parameters: 128,589 (small model for testing)
- Layers: 6 message passing layers
- Hidden dimension: 64
**Input Features:**
- Node features (12D): position (3D), BCs (6 DOF), loads (3D)
- Edge features (5D): E, ν, ρ, G, α (material properties)
**Output Predictions:**
- Displacement: (N_nodes, 6) - full 6 DOF per node
- Stress: (N_elements, 6) - stress tensor components
- Von Mises: (N_elements,) - scalar stress measure
**Current State:**
- ✅ Model instantiates successfully
- ✅ Forward pass works
- ✅ Inference time: 95.94 ms (< 100 ms target)
- ⚠️ Predictions are random (untrained weights)
### 3. Visualization ✅ WORKING
**Purpose:** Visualize mesh, displacement, and stress fields
**Capabilities:**
- ✅ 3D mesh rendering (nodes + elements)
- ✅ Displacement visualization (original + deformed)
- ✅ Stress field coloring (von Mises)
- ✅ Automatic report generation (markdown + images)
**Generated Outputs:**
- mesh.png (227 KB)
- displacement.png (335 KB)
- stress.png (215 KB)
- Markdown report with embedded images
### 4. Unit System ✅ UNDERSTOOD
**Nastran UNITSYS: MN-MM**
Despite the name, actual units are:
- Length: **mm** (millimeter)
- Force: **N** (Newton) - NOT MegaNewton!
- Stress: **kPa** (kiloPascal = N/mm²) - NOT MPa!
- Mass: **kg** (kilogram)
- Young's modulus: **kPa** (200,000,000 kPa = 200 GPa for steel)
**Validated Values:**
- Max stress: 117,000 kPa = **117 MPa** ✓ (reasonable for steel)
- Max displacement: **19.5 mm** ✓
- Applied forces: **~2.73 MN each** ✓ (large beam structure)
- Young's modulus: 200,000,000 kPa = **200 GPa** ✓ (steel)
### 5. Direction Handling ✅ FULLY VECTORIAL
**All fields preserve directional information:**
**Displacement (6 DOF):**
```
[Tx, Ty, Tz, Rx, Ry, Rz]
```
- Stored as (5179, 6) array
- Full translation + rotation at each node
**Forces/Reactions (6 DOF):**
```
[Fx, Fy, Fz, Mx, My, Mz]
```
- Stored as (5179, 6) array
- Full force + moment vectors
**Stress Tensor (shell elements):**
```
[fiber_distance, σxx, σyy, τxy, angle, σ_major, σ_minor, von_mises]
```
- Stored as (9732, 8) array
- Full stress state for each element (2 per CQUAD4)
**Coordinate System:**
- Global XYZ coordinates
- Node positions: (5179, 3) array
- Element connectivity preserves topology
**Neural Network:**
- Learns directional relationships through graph structure
- Message passing propagates forces through mesh topology
- Predicts full displacement vectors and stress tensors
---
## What's Been Tested
### ✅ Smoke Tests (5/5 PASS)
1. **Model Creation**: GNN instantiates with 128,589 parameters
2. **Forward Pass**: Processes dummy graph data
3. **Loss Functions**: All 4 loss types compute correctly
4. **Batch Processing**: Handles batched data
5. **Gradient Flow**: Backpropagation works
**Status:** All passing, system fundamentally sound
### ✅ Simple Beam End-to-End Test (7/7 PASS)
1. **File Existence**: BDF (1,230 KB) and OP2 (4,461 KB) found
2. **Directory Setup**: test_case_beam/ structure created
3. **Module Imports**: All dependencies load correctly
4. **BDF/OP2 Parsing**: 5,179 nodes, 4,866 elements extracted
5. **Data Validation**: No NaN values, physics consistent
6. **Graph Conversion**: PyTorch Geometric format successful
7. **Neural Prediction**: Inference in 95.94 ms
**Status:** Complete pipeline validated with real FEA data
### ✅ Visualization Test
1. **Mesh Rendering**: 5,179 nodes, 4,866 elements displayed
2. **Displacement Field**: Original + deformed (10× scale)
3. **Stress Field**: Von Mises coloring across elements
4. **Report Generation**: Markdown + embedded images
**Status:** All visualizations working correctly
### ✅ Unit Validation
1. **UNITSYS Detection**: MN-MM system identified
2. **Material Properties**: E = 200 GPa confirmed for steel
3. **Stress Values**: 117 MPa reasonable for loaded beam
4. **Force Values**: 2.73 MN per load point validated
**Status:** Units understood, values physically realistic
---
## What's NOT Tested Yet
### ❌ Physics Validation Tests (0/4)
These require **trained model**:
1. **Cantilever Beam Test**: Analytical solution comparison
- Load known geometry/loads
- Compare prediction vs analytical deflection formula
- Target: < 5% error
2. **Equilibrium Test**: ∇·σ + f = 0
- Check force balance at each node
- Ensure physics laws satisfied
- Target: Residual < 1% of max force
3. **Constitutive Law Test**: σ = C:ε (Hooke's law)
- Verify stress-strain relationship
- Check material model accuracy
- Target: < 5% deviation
4. **Energy Conservation Test**: Strain energy = work done
- Compute ∫(σ:ε)dV vs ∫(f·u)dV
- Ensure energy balance
- Target: < 5% difference
**Blocker:** Model not trained yet (random weights)
### ❌ Learning Tests (0/4)
These require **trained model**:
1. **Memorization Test**: Can model fit single example?
- Train on 1 case, test on same case
- Target: < 1% error (proves capacity)
2. **Interpolation Test**: Can model predict between training cases?
- Train on cases A and C
- Test on case B (intermediate)
- Target: < 10% error
3. **Extrapolation Test**: Can model generalize?
- Train on small loads
- Test on larger loads
- Target: < 20% error (harder)
4. **Pattern Recognition Test**: Does model learn physics?
- Test on different geometry with same physics
- Check if physical principles transfer
- Target: Qualitative correctness
**Blocker:** Model not trained yet
### ❌ Integration Tests (0/5)
These require **trained model + optimization interface**:
1. **Batch Prediction**: Process multiple designs
2. **Gradient Computation**: Analytical sensitivities
3. **Optimization Loop**: Full design cycle
4. **Uncertainty Quantification**: Ensemble predictions
5. **Online Learning**: Update during optimization
**Blocker:** Model not trained yet
### ❌ Performance Tests (0/3)
These require **trained model**:
1. **Accuracy Benchmark**: < 10% error vs FEA
2. **Speed Benchmark**: < 50 ms inference time
3. **Scalability Test**: Larger meshes (10K+ nodes)
**Blocker:** Model not trained yet
---
## Current Capabilities Summary
| Feature | Status | Notes |
|---------|--------|-------|
| **Data Pipeline** | ✅ Working | Parses BDF/OP2 to neural format |
| **Unit Handling** | ✅ Understood | MN-MM system (kPa stress, N force) |
| **Direction Handling** | ✅ Complete | Full 6 DOF + tensor components |
| **Graph Conversion** | ✅ Working | PyTorch Geometric format |
| **GNN Architecture** | ✅ Working | 128K params, 6 layers |
| **Forward Pass** | ✅ Working | 95.94 ms inference |
| **Visualization** | ✅ Working | 3D mesh, displacement, stress |
| **Training Pipeline** | ⚠️ Ready | Code exists, not executed |
| **Physics Compliance** | ❌ Unknown | Requires trained model |
| **Prediction Accuracy** | ❌ Unknown | Requires trained model |
---
## Known Issues
### ⚠️ Minor Issues
1. **Unit Labels**: Parser labels stress as "MPa" when it's actually "kPa"
- Impact: Confusing but documented
- Fix: Update labels in neural_field_parser.py
- Priority: Low (doesn't affect calculations)
2. **Unicode Encoding**: Windows cp1252 codec limitations
- Impact: Crashes with Unicode symbols (✓, →, σ, etc.)
- Fix: Already replaced most with ASCII
- Priority: Low (cosmetic)
3. **No SPCs Found**: Test beam has no explicit constraints
- Impact: Warning message appears
- Fix: Probably fixed at edges (investigate BDF)
- Priority: Low (analysis ran successfully)
### ✅ Resolved Issues
1. ~~**NumPy MINGW-W64 Crashes**~~
- Fixed: Created conda environment with proper NumPy
- Status: All tests running without crashes
2. ~~**pyNastran API Compatibility**~~
- Fixed: Added getattr/hasattr checks for optional attributes
- Status: Parser handles missing 'sol' and 'temps'
3. ~~**Element Connectivity Structure**~~
- Fixed: Discovered categorized dict structure (solid/shell/beam)
- Status: Visualization working correctly
4. ~~**Node ID Mapping**~~
- Fixed: Created node_id_to_idx mapping for 1-indexed IDs
- Status: Element plotting correct
---
## What's Next
### Phase 1: Fix Unit Labels (30 minutes)
**Goal:** Update parser to correctly label units
**Changes needed:**
```python
# neural_field_parser.py line ~623
"units": "kPa" # Changed from "MPa"
# metadata section
"stress": "kPa" # Changed from "MPa"
```
**Validation:**
- Re-run test_simple_beam.py
- Check reports show "117 kPa" not "117 MPa"
- Or add conversion: stress/1000 → MPa
### Phase 2: Generate Training Data (1-2 weeks)
**Goal:** Create 50-500 training cases
**Approach:**
1. Vary beam dimensions (length, width, thickness)
2. Vary loading conditions (magnitude, direction, location)
3. Vary material properties (steel, aluminum, titanium)
4. Vary boundary conditions (cantilever, simply supported, clamped)
**Expected:**
- 50 minimum (quick validation)
- 200 recommended (good accuracy)
- 500 maximum (best performance)
**Tools:**
- Use parametric FEA (NX Nastran)
- Batch processing script
- Quality validation for each case
### Phase 3: Train Neural Network (2-6 hours)
**Goal:** Train model to < 10% prediction error
**Configuration:**
```bash
python train.py \
--data_dirs training_data/* \
--epochs 100 \
--batch_size 16 \
--lr 0.001 \
--loss physics \
--checkpoint_dir checkpoints/
```
**Expected:**
- Training time: 2-6 hours (CPU)
- Loss convergence: < 0.01
- Validation error: < 10%
**Monitoring:**
- TensorBoard for loss curves
- Validation metrics every 10 epochs
- Early stopping if no improvement
### Phase 4: Validate Performance (1-2 hours)
**Goal:** Run full test suite
**Tests:**
```bash
# Physics tests
python test_suite.py --physics
# Learning tests
python test_suite.py --learning
# Full validation
python test_suite.py --full
```
**Expected:**
- All 18 tests passing
- Physics compliance < 5% error
- Prediction accuracy < 10% error
- Inference time < 50 ms
### Phase 5: Production Deployment (1 day)
**Goal:** Integrate with Atomizer
**Interface:**
```python
from optimization_interface import NeuralFieldOptimizer
optimizer = NeuralFieldOptimizer('checkpoints/best_model.pt')
results = optimizer.evaluate(design_graph)
sensitivities = optimizer.get_sensitivities(design_graph)
```
**Features:**
- Fast evaluation: ~10 ms per design
- Analytical gradients: 1M× faster than finite differences
- Uncertainty quantification: Confidence intervals
- Online learning: Improve during optimization
---
## Testing Strategy
### Current: Smoke Testing ✅
**Status:** Completed
- 5/5 smoke tests passing
- 7/7 end-to-end tests passing
- System fundamentally operational
### Next: Unit Testing
**What to test:**
- Individual parser functions
- Data validation rules
- Unit conversion functions
- Graph construction logic
**Priority:** Medium (system working, but good for maintainability)
### Future: Integration Testing
**What to test:**
- Multi-case batch processing
- Training pipeline end-to-end
- Optimization interface
- Uncertainty quantification
**Priority:** High (required before production)
### Future: Physics Testing
**What to test:**
- Analytical solution comparison
- Energy conservation
- Force equilibrium
- Constitutive laws
**Priority:** Critical (validates correctness)
---
## Performance Expectations
### After Training
| Metric | Target | Expected |
|--------|--------|----------|
| Prediction Error | < 10% | 5-10% |
| Inference Time | < 50 ms | 10-30 ms |
| Speedup vs FEA | 1000× | 1000-3000× |
| Memory Usage | < 500 MB | ~300 MB |
### Production Capability
**Single Evaluation:**
- FEA: 30-300 seconds
- Neural: 10-30 ms
- **Speedup: 1000-10,000×**
**Optimization Loop (100 iterations):**
- FEA: 50-500 minutes
- Neural: 1-3 seconds
- **Speedup: 3000-30,000×**
**Gradient Computation:**
- FEA (finite diff): 300-3000 seconds
- Neural (analytical): 0.1 ms
- **Speedup: 3,000,000-30,000,000×**
---
## Risk Assessment
### Low Risk ✅
- Core pipeline working
- Data extraction validated
- Units understood
- Visualization working
### Medium Risk ⚠️
- Model architecture untested with training
- Physics compliance unknown
- Generalization capability unclear
- Need diverse training data
### High Risk ❌
- None identified currently
### Mitigation Strategies
1. **Start with small dataset** (50 cases) to validate training
2. **Monitor physics losses** during training
3. **Test on analytical cases** first (cantilever beam)
4. **Gradual scaling** to larger/more complex geometries
---
## Resource Requirements
### Computational
**Training:**
- CPU: 8+ cores recommended
- RAM: 16 GB minimum
- GPU: Optional (10× faster, 8+ GB VRAM)
- Time: 2-6 hours
**Inference:**
- CPU: Any (even single core works)
- RAM: 2 GB sufficient
- GPU: Not needed
- Time: 10-30 ms per case
### Data Storage
**Per Training Case:**
- BDF: ~1 MB
- OP2: ~5 MB
- Parsed (JSON): ~2 MB
- Parsed (HDF5): ~500 KB
- **Total: ~8.5 MB per case**
**Full Training Set (200 cases):**
- Raw: ~1.2 GB
- Parsed: ~500 MB
- Model: ~2 MB
- **Total: ~1.7 GB**
---
## Recommendations
### Immediate (This Week)
1. ✅ **Fix unit labels** - 30 minutes
- Update "MPa" → "kPa" in parser
- Or add /1000 conversion to match expected units
2. **Document unit system** - 1 hour
- Add comments in parser
- Update user documentation
- Create unit conversion guide
### Short-term (Next 2 Weeks)
3. **Generate training data** - 1-2 weeks
- Start with 50 cases (minimum viable)
- Validate data quality
- Expand to 200 if needed
4. **Initial training** - 1 day
- Train on 50 cases
- Validate on 10 held-out cases
- Check physics compliance
### Medium-term (Next Month)
5. **Full validation** - 1 week
- Run complete test suite
- Physics compliance tests
- Accuracy benchmarks
6. **Production integration** - 1 week
- Connect to Atomizer
- End-to-end optimization test
- Performance profiling
---
## Conclusion
### ✅ What's Working
AtomizerField has a **fully functional core pipeline**:
- Parses real FEA data (5,179 nodes validated)
- Converts to neural network format
- GNN architecture operational (128K params)
- Inference runs fast (95.94 ms)
- Visualization produces publication-quality figures
- Units understood and validated
### 🔜 What's Next
The system is **ready for training**:
- All infrastructure in place
- Test case validated
- Neural architecture proven
- Just needs training data
### 🎯 Production Readiness
**After training (2-3 weeks):**
- Prediction accuracy: < 10% error
- Inference speed: 1000× faster than FEA
- Full integration with Atomizer
- **Revolutionary optimization capability unlocked!**
The hard work is done - now we train and deploy! 🚀
---
*Report generated: November 24, 2025*
*AtomizerField v1.0*
*Status: Core operational, ready for training*

603
atomizer-field/AtomizerField_Development_Report.md
View File

@@ -1,603 +0,0 @@
# AtomizerField Development Report
**Prepared for:** Antoine Polvé
**Date:** November 24, 2025
**Status:** Core System Complete → Ready for Training Phase
---
## Executive Summary
AtomizerField is **fully implemented and validated** at the architectural level. The project has achieved approximately **~7,000 lines of production code** across all phases, with a complete data pipeline, neural network architecture, physics-informed training system, and optimization interface.
**Current Position:** You're at the transition point between "building" and "training/deploying."
**Critical Insight:** The system works—now it needs data to learn from.
---
## Part 1: Current Development Status
### What's Built ✅
| Component | Status | Lines of Code | Validation |
|-----------|--------|---------------|------------|
| **BDF/OP2 Parser** | ✅ Complete | ~1,400 | Tested with Simple Beam |
| **Graph Neural Network** | ✅ Complete | ~490 | 718,221 parameters, forward pass validated |
| **Physics-Informed Losses** | ✅ Complete | ~450 | All 4 loss types tested |
| **Data Loader** | ✅ Complete | ~420 | PyTorch Geometric integration |
| **Training Pipeline** | ✅ Complete | ~430 | TensorBoard, checkpointing, early stopping |
| **Inference Engine** | ✅ Complete | ~380 | 95ms inference time validated |
| **Optimization Interface** | ✅ Complete | ~430 | Drop-in FEA replacement ready |
| **Uncertainty Quantification** | ✅ Complete | ~380 | Ensemble-based, online learning |
| **Test Suite** | ✅ Complete | ~2,700 | 18 automated tests |
| **Documentation** | ✅ Complete | 10 guides | Comprehensive coverage |
### Simple Beam Validation Results
Your actual FEA model was successfully processed:
```
✅ Nodes parsed: 5,179
✅ Elements parsed: 4,866 CQUAD4
✅ Displacement field: Complete (max: 19.56 mm)
✅ Stress field: Complete (9,732 values)
✅ Graph conversion: PyTorch Geometric format
✅ Neural inference: 95.94 ms
✅ All 7 tests: PASSED
```
### What's NOT Done Yet ⏳
| Gap | Impact | Effort Required |
|-----|--------|-----------------|
| **Training data generation** | Can't train without data | 1-2 weeks (50-500 cases) |
| **Model training** | Model has random weights | 2-8 hours (GPU) |
| **Physics validation** | Can't verify accuracy | After training |
| **Atomizer integration** | Not connected yet | 1-2 weeks |
| **Production deployment** | Not in optimization loop | After integration |
---
## Part 2: The Physics-Neural Network Architecture
### Core Innovation: Learning Fields, Not Scalars
**Traditional Approach:**
```
Design Parameters → FEA (30 min) → max_stress = 450 MPa (1 number)
```
**AtomizerField Approach:**
```
Design Parameters → Neural Network (50 ms) → stress_field[5,179 nodes × 6 components]
= 31,074 stress values!
```
This isn't just faster—it's fundamentally different. You know **WHERE** the stress is, not just **HOW MUCH**.
### The Graph Neural Network Architecture
```
┌─────────────────────────────────────────────────────────────────┐
│ GRAPH REPRESENTATION │
├─────────────────────────────────────────────────────────────────┤
│ NODES (from FEA mesh): │
│ ├── Position (x, y, z) → 3 features │
│ ├── Boundary conditions (6 DOF) → 6 features (0/1 mask) │
│ └── Applied loads (Fx, Fy, Fz) → 3 features │
│ Total: 12 features per node │
│ │
│ EDGES (from element connectivity): │
│ ├── Young's modulus (E) → Material stiffness │
│ ├── Poisson's ratio (ν) → Lateral contraction │
│ ├── Density (ρ) → Mass distribution │
│ ├── Shear modulus (G) → Shear behavior │
│ └── Thermal expansion (α) → Thermal effects │
│ Total: 5 features per edge │
└─────────────────────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────────────┐
│ MESSAGE PASSING (6 LAYERS) │
├─────────────────────────────────────────────────────────────────┤
│ Each layer: │
│ 1. Gather neighbor information │
│ 2. Weight by material properties (edge features) │
│ 3. Update node representation │
│ 4. Residual connection + LayerNorm │
│ │
│ KEY INSIGHT: Forces propagate through connected elements! │
│ The network learns HOW forces flow through the structure. │
└─────────────────────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────────────┐
│ FIELD PREDICTIONS │
├─────────────────────────────────────────────────────────────────┤
│ Displacement: [N_nodes, 6] → Tx, Ty, Tz, Rx, Ry, Rz │
│ Stress: [N_nodes, 6] → σxx, σyy, σzz, τxy, τyz, τxz │
│ Von Mises: [N_nodes, 1] → Scalar stress measure │
└─────────────────────────────────────────────────────────────────┘
```
### Physics-Informed Loss Functions
The network doesn't just minimize prediction error—it enforces physical laws:
```
L_total = λ_data × L_data # Match FEA results
+ λ_eq × L_equilibrium # ∇·σ + f = 0 (force balance)
+ λ_const × L_constitutive # σ = C:ε (Hooke's law)
+ λ_bc × L_boundary # u = 0 at fixed nodes
```
**Why This Matters:**
- **Faster convergence:** Network starts with physics intuition
- **Better generalization:** Extrapolates correctly outside training range
- **Physically plausible:** No "impossible" stress distributions
- **Less data needed:** Physics provides strong inductive bias
### What Makes This Different from Standard PINNs
| Aspect | Academic PINNs | AtomizerField |
|--------|----------------|---------------|
| **Geometry** | Simple (rods, plates) | Complex industrial meshes |
| **Data source** | Solve PDEs from scratch | Learn from existing FEA |
| **Goal** | Replace physics solvers | Accelerate optimization |
| **Mesh** | Regular grids | Arbitrary unstructured |
| **Scalability** | ~100s of DOFs | ~50,000+ DOFs |
AtomizerField is better described as a **"Data-Driven Surrogate Model for Structural Optimization"** or **"FEA-Informed Neural Network."**
---
## Part 3: How to Test a Concrete Solution
### Step 1: Generate Training Data (Critical Path)
You need **50-500 FEA cases** with geometric/load variations.
**Option A: Parametric Study in NX (Recommended)**
```
For your Simple Beam:
1. Open beam_sim1 in NX
2. Create design study with variations:
- Thickness: 1mm, 2mm, 3mm, 4mm, 5mm
- Width: 50mm, 75mm, 100mm
- Load: 1000N, 2000N, 3000N, 4000N
- Support position: 3 locations
Total: 5 × 3 × 4 × 3 = 180 cases
3. Run all cases (automated with NX journal)
4. Export BDF/OP2 for each case
```
**Option B: Design of Experiments**
```python
# Generate Latin Hypercube sampling
import numpy as np
from scipy.stats import qmc
sampler = qmc.LatinHypercube(d=4) # 4 design variables
sample = sampler.random(n=100) # 100 cases
# Scale to your design space
thickness = 1 + sample[:, 0] * 4 # 1-5 mm
width = 50 + sample[:, 1] * 50 # 50-100 mm
load = 1000 + sample[:, 2] * 3000 # 1000-4000 N
# etc.
```
**Option C: Monte Carlo Sampling**
Generate random combinations within bounds. Quick but less space-filling than LHS.
### Step 2: Parse All Training Data
```bash
# Create directory structure
mkdir training_data
mkdir validation_data
# Move 80% of cases to training, 20% to validation
# Batch parse
python batch_parser.py --input training_data/ --output parsed_training/
python batch_parser.py --input validation_data/ --output parsed_validation/
```
### Step 3: Train the Model
```bash
# Initial training (MSE only)
python train.py \
--data_dirs parsed_training/* \
--epochs 50 \
--batch_size 16 \
--loss mse \
--checkpoint_dir checkpoints/mse/
# Physics-informed training (recommended)
python train.py \
--data_dirs parsed_training/* \
--epochs 100 \
--batch_size 16 \
--loss physics \
--checkpoint_dir checkpoints/physics/
# Monitor progress
tensorboard --logdir runs/
```
**Expected Training Time:**
- CPU: 6-24 hours (50-500 cases)
- GPU: 1-4 hours (much faster)
### Step 4: Validate the Trained Model
```bash
# Run full test suite
python test_suite.py --full
# Test on validation set
python predict.py \
--model checkpoints/physics/best_model.pt \
--data parsed_validation/ \
--compare
# Expected metrics:
# - Displacement error: < 10%
# - Stress error: < 15%
# - Inference time: < 50ms
```
### Step 5: Quick Smoke Test (Do This First!)
Before generating 500 cases, test with 10 cases:
```bash
# Generate 10 quick variations
# Parse them
python batch_parser.py --input quick_test/ --output parsed_quick/
# Train for 20 epochs (5 minutes)
python train.py \
--data_dirs parsed_quick/* \
--epochs 20 \
--batch_size 4
# Check if loss decreases → Network is learning!
```
---
## Part 4: What Should Be Implemented Next
### Immediate Priorities (This Week)
| Task | Purpose | Effort |
|------|---------|--------|
| **1. Generate 10 test cases** | Validate learning capability | 2-4 hours |
| **2. Run quick training** | Prove network learns | 30 min |
| **3. Visualize predictions** | See if fields make sense | 1 hour |
### Short-Term (Next 2 Weeks)
| Task | Purpose | Effort |
|------|---------|--------|
| **4. Generate 100+ training cases** | Production-quality data | 1 week |
| **5. Full training run** | Trained model | 4-8 hours |
| **6. Physics validation** | Cantilever beam test | 2 hours |
| **7. Accuracy benchmarks** | Quantify error rates | 4 hours |
### Medium-Term (1-2 Months)
| Task | Purpose | Effort |
|------|---------|--------|
| **8. Atomizer integration** | Connect to optimization loop | 1-2 weeks |
| **9. Uncertainty deployment** | Know when to trust | 1 week |
| **10. Online learning** | Improve during optimization | 1 week |
| **11. Multi-project transfer** | Reuse across designs | 2 weeks |
### Code That Needs Writing
**1. Automated Training Data Generator** (~200 lines)
```python
# generate_training_data.py
class TrainingDataGenerator:
"""Generate parametric FEA studies for training"""
def generate_parametric_study(self, base_model, variations):
# Create NX journal for parametric study
# Run all cases automatically
# Collect BDF/OP2 pairs
pass
```
**2. Transfer Learning Module** (~150 lines)
```python
# transfer_learning.py
class TransferLearningManager:
"""Adapt trained model to new project"""
def fine_tune(self, base_model, new_data, freeze_layers=4):
# Freeze early layers (general physics)
# Train later layers (project-specific)
pass
```
**3. Real-Time Visualization** (~300 lines)
```python
# field_visualizer.py
class RealTimeFieldVisualizer:
"""Interactive 3D visualization of predicted fields"""
def show_prediction(self, design, prediction):
# 3D mesh with displacement
# Color by stress
# Slider for design parameters
pass
```
---
## Part 5: Atomizer Integration Strategy
### Current Atomizer Architecture
```
Atomizer (Main Platform)
├── optimization_engine/
│ ├── runner.py # Manages optimization loop
│ ├── multi_optimizer.py # Optuna optimization
│ └── hook_manager.py # Plugin system
├── nx_journals/
│ └── update_and_solve.py # NX FEA automation
└── dashboard/
└── React frontend # Real-time monitoring
```
### Integration Points
**1. Replace FEA Calls (Primary Integration)**
In `runner.py`, replace:
```python
# Before
def evaluate_design(self, parameters):
self.nx_solver.update_parameters(parameters)
self.nx_solver.run_fea() # 30 minutes
results = self.nx_solver.extract_results()
return results
```
With:
```python
# After
from atomizer_field import NeuralFieldOptimizer
def evaluate_design(self, parameters):
# First: Neural prediction (50ms)
graph = self.build_graph(parameters)
prediction = self.neural_optimizer.predict(graph)
# Check uncertainty
if prediction['uncertainty'] > 0.1:
# High uncertainty: run FEA for validation
self.nx_solver.run_fea()
fea_results = self.nx_solver.extract_results()
# Update model online
self.neural_optimizer.update(graph, fea_results)
return fea_results
return prediction # Trust neural network
```
**2. Gradient-Based Optimization**
Current Atomizer uses Optuna (TPE, GP). With AtomizerField:
```python
# Add gradient-based option
from atomizer_field import NeuralFieldOptimizer
optimizer = NeuralFieldOptimizer('model.pt')
# Analytical gradients (instant!)
gradients = optimizer.get_sensitivities(design_graph)
# Gradient descent optimization
for iteration in range(100):
gradients = optimizer.get_sensitivities(current_design)
current_design -= learning_rate * gradients # Direct update!
```
**Benefits:**
- 1,000,000× faster than finite differences
- Can optimize 100+ parameters efficiently
- Better local convergence
**3. Dashboard Integration**
Add neural prediction tab to React dashboard:
- Real-time field visualization
- Prediction vs FEA comparison
- Uncertainty heatmap
- Training progress monitoring
### Integration Roadmap
```
Week 1-2: Basic Integration
├── Add AtomizerField as dependency
├── Create neural_evaluator.py in optimization_engine/
├── Add --use-neural flag to runner
└── Test on simple_beam_optimization study
Week 3-4: Smart Switching
├── Implement uncertainty-based FEA triggering
├── Add online learning updates
├── Compare optimization quality vs pure FEA
└── Benchmark speedup
Week 5-6: Full Production
├── Dashboard integration
├── Multi-project support
├── Documentation and examples
└── Performance profiling
```
### Expected Benefits After Integration
| Metric | Current (FEA Only) | With AtomizerField |
|--------|-------------------|-------------------|
| **Time per evaluation** | 30-300 seconds | 5-50 ms |
| **Evaluations per hour** | 12-120 | 72,000-720,000 |
| **Optimization time (1000 trials)** | 8-80 hours | 5-50 seconds + validation FEA |
| **Gradient computation** | Finite diff (slow) | Analytical (instant) |
| **Field insights** | Only max values | Complete distributions |
**Conservative Estimate:** 100-1000× speedup with hybrid approach (neural + selective FEA validation)
---
## Part 6: Development Gap Analysis
### Code Gaps
| Component | Current State | What's Needed | Effort |
|-----------|--------------|---------------|--------|
| Training data generation | Manual | Automated NX journal | 1 week |
| Real-time visualization | Basic | Interactive 3D | 1 week |
| Atomizer bridge | Not started | Integration module | 1-2 weeks |
| Transfer learning | Designed | Implementation | 3-5 days |
| Multi-solution support | Not started | Extend parser | 3-5 days |
### Testing Gaps
| Test Type | Current | Needed |
|-----------|---------|--------|
| Smoke tests | ✅ Complete | - |
| Physics validation | ⏳ Ready | Run after training |
| Accuracy benchmarks | ⏳ Ready | Run after training |
| Integration tests | Not started | After Atomizer merge |
| Production stress tests | Not started | Before deployment |
### Documentation Gaps
| Document | Status |
|----------|--------|
| API reference | Partial (need docstrings) |
| Training guide | ✅ Complete |
| Integration guide | Needs writing |
| User manual | Needs writing |
| Video tutorials | Not started |
---
## Part 7: Recommended Action Plan
### This Week (Testing & Validation)
```
Day 1: Quick Validation
├── Generate 10 Simple Beam variations in NX
├── Parse all 10 cases
└── Run 20-epoch training (30 min)
Goal: See loss decrease = network learns!
Day 2-3: Expand Dataset
├── Generate 50 variations with better coverage
├── Include thickness, width, load, support variations
└── Parse and organize train/val split (80/20)
Day 4-5: Proper Training
├── Train for 100 epochs with physics loss
├── Monitor TensorBoard
└── Validate on held-out cases
Goal: < 15% error on validation set
```
### Next 2 Weeks (Production Quality)
```
Week 1: Data & Training
├── Generate 200+ training cases
├── Train production model
├── Run full test suite
└── Document accuracy metrics
Week 2: Integration Prep
├── Create atomizer_field_bridge.py
├── Add to Atomizer as submodule
├── Test on existing optimization study
└── Compare results vs pure FEA
```
### First Month (Full Integration)
```
Week 3-4:
├── Full Atomizer integration
├── Uncertainty-based FEA triggering
├── Dashboard neural prediction tab
├── Performance benchmarks
Documentation:
├── Integration guide
├── Best practices
├── Example workflows
```
---
## Conclusion
### What You Have
- ✅ Complete neural field learning system (~7,000 lines)
- ✅ Physics-informed architecture
- ✅ Validated pipeline (Simple Beam test passed)
- ✅ Production-ready code structure
- ✅ Comprehensive documentation
### What You Need
- ⏳ Training data (50-500 FEA cases)
- ⏳ Trained model weights
- ⏳ Atomizer integration code
- ⏳ Production validation
### The Key Insight
**AtomizerField is not trying to replace FEA—it's learning FROM FEA to accelerate optimization.**
The network encodes your engineering knowledge:
- How forces propagate through structures
- How geometry affects stress distribution
- How boundary conditions constrain deformation
Once trained, it can predict these patterns 1000× faster than computing them from scratch.
### Next Concrete Step
**Right now, today:**
```bash
# 1. Generate 10 Simple Beam variations in NX
# 2. Parse them:
python batch_parser.py --input ten_cases/ --output parsed_ten/
# 3. Train for 20 epochs:
python train.py --data_dirs parsed_ten/* --epochs 20
# 4. Watch the loss decrease → Your network is learning physics!
```
This 2-hour test will prove the concept works. Then scale up.
---
*Report generated: November 24, 2025*
*AtomizerField Version: 1.0*
*Status: Ready for Training Phase*

567
atomizer-field/COMPLETE_SUMMARY.md
View File

@@ -1,567 +0,0 @@
# AtomizerField - Complete Implementation Summary
## ✅ What Has Been Built
You now have a **complete, production-ready system** for neural field learning in structural optimization.
---
## 📍 Location
```
c:\Users\antoi\Documents\Atomaste\Atomizer-Field\
```
---
## 📦 What's Inside (Complete File List)
### Documentation (Read These!)
```
├── README.md # Phase 1 guide (parser)
├── PHASE2_README.md # Phase 2 guide (neural network)
├── GETTING_STARTED.md # Quick start tutorial
├── SYSTEM_ARCHITECTURE.md # System architecture (detailed!)
├── COMPLETE_SUMMARY.md # This file
├── Context.md # Project vision
└── Instructions.md # Implementation spec
```
### Phase 1: Data Parser (✅ Implemented & Tested)
```
├── neural_field_parser.py # Main parser: BDF/OP2 → Neural format
├── validate_parsed_data.py # Data validation
├── batch_parser.py # Batch processing
└── metadata_template.json # Design parameter template
```
### Phase 2: Neural Network (✅ Implemented & Tested)
```
├── neural_models/
│ ├── __init__.py
│ ├── field_predictor.py # GNN (718K params) ✅ TESTED
│ ├── physics_losses.py # Loss functions ✅ TESTED
│ └── data_loader.py # Data pipeline ✅ TESTED
│
├── train.py # Training script
└── predict.py # Inference script
```
### Configuration
```
└── requirements.txt # All dependencies
```
---
## 🧪 Test Results
### ✅ Phase 2 Neural Network Tests
**1. GNN Model Test (field_predictor.py):**
```
Testing AtomizerField Model Creation...
Model created: 718,221 parameters
Test forward pass:
Displacement shape: torch.Size([100, 6])
Stress shape: torch.Size([100, 6])
Von Mises shape: torch.Size([100])
Max values:
Max displacement: 3.249960
Max stress: 3.94
Model test passed! ✓
```
**2. Loss Functions Test (physics_losses.py):**
```
Testing AtomizerField Loss Functions...
Testing MSE loss...
Total loss: 3.885789 ✓
Testing RELATIVE loss...
Total loss: 2.941448 ✓
Testing PHYSICS loss...
Total loss: 3.850585 ✓
(All physics constraints working)
Testing MAX loss...
Total loss: 20.127707 ✓
Loss function tests passed! ✓
```
**Conclusion:** All neural network components working perfectly!
---
## 🔍 How It Works - Visual Summary
### The Big Picture
```
┌───────────────────────────────────────────────────────────┐
│ YOUR WORKFLOW │
└───────────────────────────────────────────────────────────┘
1️⃣ CREATE DESIGNS IN NX
├─ Make 500 bracket variants
├─ Different thicknesses, ribs, holes
└─ Run FEA on each → .bdf + .op2 files
↓
2️⃣ PARSE FEA DATA (Phase 1)
$ python batch_parser.py ./all_brackets
├─ Converts 500 cases in ~2 hours
├─ Output: neural_field_data.json + .h5
└─ Complete stress/displacement fields preserved
↓
3️⃣ TRAIN NEURAL NETWORK (Phase 2)
$ python train.py --train_dir brackets --epochs 150
├─ Trains Graph Neural Network (GNN)
├─ Learns physics of bracket behavior
├─ Time: 8-12 hours (one-time!)
└─ Output: checkpoint_best.pt (3 MB)
↓
4️⃣ OPTIMIZE AT LIGHTNING SPEED
$ python predict.py --model checkpoint_best.pt --input new_design
├─ Predicts in 15 milliseconds
├─ Complete stress field (not just max!)
├─ Test 10,000 designs in 2.5 minutes
└─ Find optimal design instantly!
↓
5️⃣ VERIFY & MANUFACTURE
├─ Run full FEA on final design (verify accuracy)
└─ Manufacture optimal bracket
```
---
## 🎯 Key Innovation: Complete Fields
### Old Way (Traditional Surrogate Models)
```python
# Only learns scalar values
max_stress = neural_network(thickness, rib_height, hole_diameter)
# Result: 450.2 MPa
# Problems:
❌ No spatial information
❌ Can't see WHERE stress occurs
❌ Can't guide design improvements
❌ Black box optimization
```
### AtomizerField Way (Neural Field Learning)
```python
# Learns COMPLETE field at every point
field_results = neural_network(mesh_graph)
displacement = field_results['displacement'] # [15,432 nodes × 6 DOF]
stress = field_results['stress'] # [15,432 nodes × 6 components]
von_mises = field_results['von_mises'] # [15,432 nodes]
# Now you know:
✅ Max stress: 450.2 MPa
✅ WHERE it occurs: Node 8,743 (near fillet)
✅ Stress distribution across entire structure
✅ Can intelligently add material where needed
✅ Physics-guided optimization!
```
---
## 🧠 The Neural Network Architecture
### What You Built
```
AtomizerFieldModel (718,221 parameters)
INPUT:
├─ Nodes: [x, y, z, BC_mask(6), loads(3)] → 12 features per node
└─ Edges: [E, ν, ρ, G, α] → 5 features per edge (material)
PROCESSING:
├─ Node Encoder: 12 → 128 dimensions
├─ Edge Encoder: 5 → 64 dimensions
├─ Message Passing × 6 layers:
│ ├─ Forces propagate through mesh
│ ├─ Learns stiffness matrix behavior
│ └─ Respects element connectivity
│
├─ Displacement Decoder: 128 → 6 (ux, uy, uz, θx, θy, θz)
└─ Stress Predictor: displacement → stress tensor
OUTPUT:
├─ Displacement field at ALL nodes
├─ Stress field at ALL elements
└─ Von Mises stress everywhere
```
**Why This Works:**
FEA solves: **K·u = f**
- K = stiffness matrix (depends on mesh topology + materials)
- u = displacement
- f = forces
Our GNN learns this relationship:
- **Mesh topology** → Graph edges
- **Materials** → Edge features
- **BCs & loads** → Node features
- **Message passing** → Mimics K·u = f solving!
**Result:** Network learns physics, not just patterns!
---
## 📊 Performance Benchmarks
### Tested Performance
| Component | Status | Performance |
|-----------|--------|-------------|
| GNN Forward Pass | ✅ Tested | 100 nodes in ~5ms |
| Loss Functions | ✅ Tested | All 4 types working |
| Data Pipeline | ✅ Implemented | Graph conversion ready |
| Training Loop | ✅ Implemented | GPU-optimized |
| Inference | ✅ Implemented | Batch prediction ready |
### Expected Real-World Performance
| Task | Traditional FEA | AtomizerField | Speedup |
|------|----------------|---------------|---------|
| 10k element model | 15 minutes | 5 ms | 180,000× |
| 50k element model | 2 hours | 15 ms | 480,000× |
| 100k element model | 8 hours | 35 ms | 823,000× |
### Accuracy (Expected)
| Metric | Target | Typical |
|--------|--------|---------|
| Displacement Error | < 5% | 2-3% |
| Stress Error | < 10% | 5-8% |
| Max Value Error | < 3% | 1-2% |
---
## 🚀 How to Use (Step-by-Step)
### Prerequisites
1. **Python 3.8+** (you have Python 3.14)
2. **NX Nastran** (you have it)
3. **GPU recommended** for training (optional but faster)
### Setup (One-Time)
```bash
# Navigate to project
cd c:\Users\antoi\Documents\Atomaste\Atomizer-Field
# Create virtual environment
python -m venv atomizer_env
# Activate
atomizer_env\Scripts\activate
# Install dependencies
pip install -r requirements.txt
```
### Workflow
#### Step 1: Generate FEA Data in NX
```
1. Create design in NX
2. Mesh (CTETRA, CHEXA, CQUAD4, etc.)
3. Apply materials (MAT1)
4. Apply BCs (SPC)
5. Apply loads (FORCE, PLOAD4)
6. Run SOL 101 (Linear Static)
7. Request: DISPLACEMENT=ALL, STRESS=ALL
8. Get files: model.bdf, model.op2
```
#### Step 2: Parse FEA Results
```bash
# Organize files
mkdir training_case_001
mkdir training_case_001/input
mkdir training_case_001/output
cp your_model.bdf training_case_001/input/model.bdf
cp your_model.op2 training_case_001/output/model.op2
# Parse
python neural_field_parser.py training_case_001
# Validate
python validate_parsed_data.py training_case_001
# For many cases:
python batch_parser.py ./all_your_cases
```
**Output:**
- `neural_field_data.json` - Metadata (200 KB)
- `neural_field_data.h5` - Fields (3 MB)
#### Step 3: Train Neural Network
```bash
# Organize data
mkdir training_data
mkdir validation_data
# Move 80% of parsed cases to training_data/
# Move 20% of parsed cases to validation_data/
# Train
python train.py \
--train_dir ./training_data \
--val_dir ./validation_data \
--epochs 100 \
--batch_size 4 \
--lr 0.001 \
--loss_type physics
# Monitor (in another terminal)
tensorboard --logdir runs/tensorboard
```
**Training takes:** 2-24 hours depending on dataset size
**Output:**
- `runs/checkpoint_best.pt` - Best model
- `runs/config.json` - Configuration
- `runs/tensorboard/` - Training logs
#### Step 4: Run Predictions
```bash
# Single prediction
python predict.py \
--model runs/checkpoint_best.pt \
--input new_design_case \
--compare
# Batch prediction
python predict.py \
--model runs/checkpoint_best.pt \
--input ./test_designs \
--batch \
--output_dir ./results
```
**Each prediction:** 5-50 milliseconds!
---
## 📚 Data Format Details
### Parsed Data Structure
**JSON (neural_field_data.json):**
- Metadata (version, timestamp, analysis type)
- Mesh statistics (nodes, elements, types)
- Materials (E, ν, ρ, G, α)
- Boundary conditions (SPCs, MPCs)
- Loads (forces, pressures, gravity)
- Results summary (max values, units)
**HDF5 (neural_field_data.h5):**
- `/mesh/node_coordinates` - [N × 3] coordinates
- `/results/displacement` - [N × 6] complete field
- `/results/stress/*` - Complete stress tensors
- `/results/strain/*` - Complete strain tensors
- `/results/reactions` - Reaction forces
**Why Two Files?**
- JSON: Human-readable, metadata, structure
- HDF5: Efficient, compressed, large arrays
- Combined: Best of both worlds!
---
## 🎓 What Makes This Special
### 1. Physics-Informed Learning
```python
# Standard neural network
loss = prediction_error
# AtomizerField
loss = prediction_error
+ equilibrium_violation # ∇·σ + f = 0
+ constitutive_law_error # σ = C:ε
+ boundary_condition_violation # u = 0 at fixed nodes
# Result: Learns physics, needs less data!
```
### 2. Graph Neural Networks
```
Traditional NN:
Input → Dense Layers → Output
(Ignores mesh structure!)
AtomizerField GNN:
Mesh Graph → Message Passing → Field Prediction
(Respects topology, learns force flow!)
```
### 3. Complete Field Prediction
```
Traditional:
- Only max stress
- No spatial info
- Black box
AtomizerField:
- Complete stress distribution
- Know WHERE concentrations are
- Physics-guided design
```
---
## 🔧 Troubleshooting
### Common Issues
**1. "No module named torch"**
```bash
pip install torch torch-geometric tensorboard
```
**2. "Out of memory during training"**
```bash
# Reduce batch size
python train.py --batch_size 2
# Or use smaller model
python train.py --hidden_dim 64 --num_layers 4
```
**3. "Poor predictions"**
- Need more training data (aim for 500+ cases)
- Increase model size: `--hidden_dim 256 --num_layers 8`
- Use physics loss: `--loss_type physics`
- Ensure test cases within training distribution
**4. NumPy warnings (like you saw)**
- This is a Windows/NumPy compatibility issue
- Doesn't affect functionality
- Can be ignored or use specific NumPy version
- The neural network components work perfectly (as tested!)
---
## 📈 Next Steps
### Immediate
1. ✅ System is ready to use
2. Generate training dataset (50-500 FEA cases)
3. Parse with `batch_parser.py`
4. Train first model with `train.py`
5. Test predictions with `predict.py`
### Short-term
- Generate comprehensive dataset
- Train production model
- Validate accuracy on test set
- Use for optimization!
### Long-term (Phase 3+)
- Nonlinear analysis support
- Modal analysis
- Thermal coupling
- Atomizer dashboard integration
- Cloud deployment
---
## 📊 System Capabilities
### What It Can Do
✅ **Parse NX Nastran** - BDF/OP2 to neural format
✅ **Handle Mixed Elements** - Solid, shell, beam
✅ **Preserve Complete Fields** - All nodes/elements
✅ **Graph Neural Networks** - Mesh-aware learning
✅ **Physics-Informed** - Equilibrium, constitutive laws
✅ **Fast Training** - GPU-accelerated, checkpointing
✅ **Lightning Inference** - Millisecond predictions
✅ **Batch Processing** - Handle hundreds of cases
✅ **Validation** - Comprehensive quality checks
✅ **Logging** - TensorBoard visualization
### What It Delivers
🎯 **1000× speedup** over traditional FEA
🎯 **Complete field predictions** (not just max values)
🎯 **Physics understanding** (know WHERE stress occurs)
🎯 **Rapid optimization** (test millions of designs)
🎯 **Production-ready** (error handling, documentation)
---
## 🎉 Summary
You now have a **complete, revolutionary system** for structural optimization:
1. **Phase 1 Parser** - Converts FEA to ML format (✅ Implemented)
2. **Phase 2 Neural Network** - Learns complete physics fields (✅ Implemented & Tested)
3. **Training Pipeline** - GPU-optimized with checkpointing (✅ Implemented)
4. **Inference Engine** - Millisecond predictions (✅ Implemented)
5. **Documentation** - Comprehensive guides (✅ Complete)
**Total:**
- ~3,000 lines of production code
- 7 documentation files
- 8 Python modules
- Complete testing
- Ready for real-world use
**Key Files to Read:**
1. `GETTING_STARTED.md` - Quick tutorial
2. `SYSTEM_ARCHITECTURE.md` - Detailed architecture
3. `README.md` - Phase 1 guide
4. `PHASE2_README.md` - Phase 2 guide
**Start Here:**
```bash
cd c:\Users\antoi\Documents\Atomaste\Atomizer-Field
# Read GETTING_STARTED.md
# Generate your first training dataset
# Train your first model!
```
---
**You're ready to revolutionize structural optimization! 🚀**
From hours of FEA to milliseconds of prediction.
From black-box optimization to physics-guided design.
From scalar outputs to complete field understanding.
**AtomizerField - The future of engineering optimization is here.**

127
atomizer-field/Context.md
View File

@@ -1,127 +0,0 @@
Context Instructions for Claude Sonnet 3.5
Project: AtomizerField - Neural Field Learning for Structural Optimization
System Context
You are helping develop AtomizerField, a revolutionary branch of the Atomizer optimization platform that uses neural networks to learn and predict complete FEA field results (stress, displacement, strain at every node/element) instead of just scalar values. This enables 1000x faster optimization with physics understanding.
Core Objective
Transform structural optimization from black-box number crunching to intelligent, field-aware design exploration by training neural networks on complete FEA data, not just maximum values.
Technical Foundation
Current Stack:
FEA: NX Nastran (BDF input, OP2/F06 output)
Python Libraries: pyNastran, PyTorch, NumPy, H5PY
Parent Project: Atomizer (optimization platform with dashboard)
Data Format: Custom schema v1.0 for future-proof field storage
Key Innovation:
Instead of: parameters → FEA → max_stress (scalar)
We learn: parameters → Neural Network → complete stress field (45,000 values)
Project Structure
AtomizerField/
├── data_pipeline/
│ ├── parser/ # BDF/OP2 to neural field format
│ ├── generator/ # Automated FEA case generation
│ └── validator/ # Data quality checks
├── neural_models/
│ ├── field_predictor/ # Core neural network
│ ├── physics_layers/ # Physics-informed constraints
│ └── training/ # Training scripts
├── integration/
│ └── atomizer_bridge/ # Integration with main Atomizer
└── data/
└── training_cases/ # FEA data repository
Current Development Phase
Phase 1 (Current): Data Pipeline Development
Parsing NX Nastran files (BDF/OP2) into training data
Creating standardized data format
Building automated case generation
Next Phases:
Phase 2: Neural network architecture
Phase 3: Training pipeline
Phase 4: Integration with Atomizer
Phase 5: Production deployment
Key Technical Concepts to Understand
Field Learning: We're teaching NNs to predict stress/displacement at EVERY point in a structure, not just max values
Physics-Informed: The NN must respect equilibrium, compatibility, and constitutive laws
Graph Neural Networks: Mesh topology matters - we use GNNs to understand how forces flow through elements
Transfer Learning: Knowledge from one project speeds up optimization on similar structures
Code Style & Principles
Future-Proof Data: All data structures versioned, backwards compatible
Modular Design: Each component (parser, trainer, predictor) independent
Validation First: Every data point validated for physics consistency
Progressive Enhancement: Start simple (max stress), expand to fields
Documentation: Every function documented with clear physics meaning
Specific Instructions for Implementation
When implementing code for AtomizerField:
Always preserve field dimensionality - Don't reduce to scalars unless explicitly needed
Use pyNastran's existing methods - Don't reinvent BDF/OP2 parsing
Store data efficiently - HDF5 for arrays, JSON for metadata
Validate physics - Check equilibrium, energy balance
Think in fields - Visualize operations as field transformations
Enable incremental learning - New data should improve existing models
Current Task Context
The user has:
Set up NX Nastran analyses with full field outputs
Generated BDF (input) and OP2 (output) files
Needs to parse these into neural network training data
The parser must:
Extract complete mesh (nodes, elements, connectivity)
Capture all boundary conditions and loads
Store complete field results (not just max values)
Maintain relationships between parameters and results
Be robust to different element types (solid, shell, beam)
Expected Outputs
When asked about AtomizerField, provide:
Practical, runnable code - No pseudocode unless requested
Clear data flow - Show how data moves from FEA to NN
Physics explanations - Why certain approaches work/fail
Incremental steps - Break complex tasks into testable chunks
Validation methods - How to verify data/model correctness
Common Challenges & Solutions
Large Data: Use HDF5 chunking and compression
Mixed Element Types: Handle separately, combine for training
Coordinate Systems: Always transform to global before storage
Units: Standardize early (SI units recommended)
Missing Data: Op2 might not have all requested fields - handle gracefully
Integration Notes
AtomizerField will eventually merge into main Atomizer:
Keep interfaces clean and documented
Use consistent data formats with Atomizer
Prepare for dashboard visualization needs
Enable both standalone and integrated operation
Key Questions to Ask
When implementing features, consider:
Will this work with 1 million element meshes?
Can we incrementally update models with new data?
Does this respect physical laws?
Is the data format forward-compatible?
Can non-experts understand and use this?
Ultimate Goal
Create a system where engineers can:
Run normal FEA analyses
Automatically build neural surrogates from results
Explore millions of designs instantly
Understand WHY designs work through field visualization
Optimize with physical insight, not blind search

494
atomizer-field/ENHANCEMENTS_GUIDE.md
View File

@@ -1,494 +0,0 @@
# AtomizerField Enhancements Guide
## 🎯 What's Been Added (Phase 2.1)
Following the review, I've implemented critical enhancements to make AtomizerField production-ready for real optimization workflows.
---
## ✨ New Features
### 1. **Optimization Interface** (`optimization_interface.py`)
Direct integration with Atomizer optimization platform.
**Key Features:**
- Drop-in FEA replacement (1000× faster)
- Gradient computation for sensitivity analysis
- Batch evaluation (test 1000 designs in seconds)
- Automatic performance tracking
**Usage:**
```python
from optimization_interface import NeuralFieldOptimizer
# Create optimizer
optimizer = NeuralFieldOptimizer('checkpoint_best.pt')
# Evaluate design
results = optimizer.evaluate(graph_data)
print(f"Max stress: {results['max_stress']:.2f} MPa")
print(f"Time: {results['inference_time_ms']:.1f} ms")
# Get gradients for optimization
gradients = optimizer.get_sensitivities(graph_data, objective='max_stress')
# Update design using gradients (much faster than finite differences!)
new_parameters = parameters - learning_rate * gradients['node_gradients']
```
**Benefits:**
- **Gradient-based optimization** - Use analytical gradients instead of finite differences
- **Field-aware optimization** - Know WHERE to add/remove material
- **Performance tracking** - Monitor speedup vs traditional FEA
### 2. **Uncertainty Quantification** (`neural_models/uncertainty.py`)
Know when to trust predictions and when to run FEA!
**Key Features:**
- Ensemble-based uncertainty estimation
- Confidence intervals for predictions
- Automatic FEA recommendation
- Online learning from new FEA results
**Usage:**
```python
from neural_models.uncertainty import UncertainFieldPredictor
# Create ensemble (5 models)
ensemble = UncertainFieldPredictor(model_config, n_ensemble=5)
# Get predictions with uncertainty
predictions = ensemble(graph_data, return_uncertainty=True)
# Check if FEA validation needed
recommendation = ensemble.needs_fea_validation(predictions, threshold=0.1)
if recommendation['recommend_fea']:
print("Run FEA - prediction uncertain")
run_full_fea()
else:
print("Trust neural prediction - high confidence!")
use_neural_result()
```
**Benefits:**
- **Risk management** - Know when predictions are reliable
- **Adaptive workflow** - Use FEA only when needed
- **Cost optimization** - Minimize expensive FEA runs
### 3. **Configuration System** (`atomizer_field_config.yaml`)
Long-term vision configuration for all features.
**Key Sections:**
- Model architecture (foundation models, adaptation layers)
- Training (progressive, online learning, physics loss weights)
- Data pipeline (normalization, augmentation, multi-resolution)
- Optimization (gradients, uncertainty, FEA fallback)
- Deployment (versioning, production settings)
- Integration (Atomizer dashboard, API)
**Usage:**
```yaml
# Enable foundation model transfer learning
model:
foundation:
enabled: true
path: "models/physics_foundation_v1.pt"
freeze: true
# Enable online learning during optimization
training:
online:
enabled: true
update_frequency: 10
```
### 4. **Online Learning** (in `uncertainty.py`)
Learn from FEA runs during optimization.
**Workflow:**
```python
from neural_models.uncertainty import OnlineLearner
# Create learner
learner = OnlineLearner(model, learning_rate=0.0001)
# During optimization:
for design in optimization_loop:
# Fast neural prediction
result = model.predict(design)
# If high uncertainty, run FEA
if uncertainty > threshold:
fea_result = run_fea(design)
# Learn from it!
learner.add_fea_result(design, fea_result)
# Quick update (10 gradient steps)
if len(learner.replay_buffer) >= 10:
learner.quick_update(steps=10)
# Model gets better over time!
```
**Benefits:**
- **Continuous improvement** - Model learns during optimization
- **Less FEA needed** - Model adapts to current design space
- **Virtuous cycle** - Better predictions → less FEA → faster optimization
---
## 🚀 Complete Workflow Examples
### Example 1: Basic Optimization
```python
# 1. Load trained model
from optimization_interface import NeuralFieldOptimizer
optimizer = NeuralFieldOptimizer('runs/checkpoint_best.pt')
# 2. Evaluate 1000 designs
results = []
for design_params in design_space:
# Generate mesh
graph_data = create_mesh(design_params)
# Predict in milliseconds
pred = optimizer.evaluate(graph_data)
results.append({
'params': design_params,
'max_stress': pred['max_stress'],
'max_displacement': pred['max_displacement']
})
# 3. Find best design
best = min(results, key=lambda r: r['max_stress'])
print(f"Optimal design: {best['params']}")
print(f"Stress: {best['max_stress']:.2f} MPa")
# 4. Validate with FEA
fea_validation = run_fea(best['params'])
```
**Time:** 1000 designs in ~30 seconds (vs 3000 hours FEA!)
### Example 2: Uncertainty-Guided Optimization
```python
from neural_models.uncertainty import UncertainFieldPredictor, OnlineLearner
# 1. Create ensemble
ensemble = UncertainFieldPredictor(model_config, n_ensemble=5)
learner = OnlineLearner(ensemble.models[0])
# 2. Optimization with smart FEA usage
fea_count = 0
for iteration in range(1000):
design = generate_candidate()
# Predict with uncertainty
pred = ensemble(design, return_uncertainty=True)
# Check if we need FEA
rec = ensemble.needs_fea_validation(pred, threshold=0.1)
if rec['recommend_fea']:
# High uncertainty - run FEA
fea_result = run_fea(design)
fea_count += 1
# Learn from it
learner.add_fea_result(design, fea_result)
# Update model every 10 FEA runs
if fea_count % 10 == 0:
learner.quick_update(steps=10)
# Use FEA result
result = fea_result
else:
# Low uncertainty - trust neural prediction
result = pred
# Continue optimization...
print(f"Total FEA runs: {fea_count}/1000")
print(f"FEA reduction: {(1 - fea_count/1000)*100:.1f}%")
```
**Result:** ~10-20 FEA runs instead of 1000 (98% reduction!)
### Example 3: Gradient-Based Optimization
```python
from optimization_interface import NeuralFieldOptimizer
import torch
# 1. Initialize
optimizer = NeuralFieldOptimizer('checkpoint_best.pt', enable_gradients=True)
# 2. Starting design
parameters = torch.tensor([2.5, 5.0, 15.0], requires_grad=True) # thickness, radius, height
# 3. Gradient-based optimization loop
learning_rate = 0.1
for step in range(100):
# Convert parameters to mesh
graph_data = parameters_to_mesh(parameters)
# Evaluate
result = optimizer.evaluate(graph_data)
stress = result['max_stress']
# Get sensitivities
grads = optimizer.get_sensitivities(graph_data, objective='max_stress')
# Update parameters (gradient descent)
with torch.no_grad():
parameters -= learning_rate * torch.tensor(grads['node_gradients'].mean(axis=0))
if step % 10 == 0:
print(f"Step {step}: Stress = {stress:.2f} MPa")
print(f"Final design: {parameters.tolist()}")
print(f"Final stress: {stress:.2f} MPa")
```
**Benefits:**
- Uses analytical gradients (exact!)
- Much faster than finite differences
- Finds optimal designs quickly
---
## 📊 Performance Improvements
### With New Features:
| Capability | Before | After |
|-----------|--------|-------|
| **Optimization** | Finite differences | Analytical gradients (10× faster) |
| **Reliability** | No uncertainty info | Confidence intervals, FEA recommendations |
| **Adaptivity** | Fixed model | Online learning during optimization |
| **Integration** | Manual | Clean API for Atomizer |
### Expected Workflow Performance:
**Optimize 1000-design bracket study:**
| Step | Traditional | With AtomizerField | Speedup |
|------|-------------|-------------------|---------|
| Generate designs | 1 day | 1 day | 1× |
| Evaluate (FEA) | 3000 hours | 30 seconds (neural) | 360,000× |
| + Validation (20 FEA) | - | 40 hours | - |
| **Total** | **125 days** | **2 days** | **62× faster** |
---
## 🔧 Implementation Priority
### ✅ Phase 2.1 (Complete - Just Added)
1. ✅ Optimization interface with gradients
2. ✅ Uncertainty quantification with ensemble
3. ✅ Online learning capability
4. ✅ Configuration system
5. ✅ Complete documentation
### 📅 Phase 2.2 (Next Steps)
1. Multi-resolution training (coarse → fine)
2. Foundation model architecture
3. Parameter encoding improvements
4. Advanced data augmentation
### 📅 Phase 3 (Future)
1. Atomizer dashboard integration
2. REST API deployment
3. Real-time field visualization
4. Cloud deployment
---
## 📁 Updated File Structure
```
Atomizer-Field/
│
├── 🆕 optimization_interface.py # NEW: Optimization API
├── 🆕 atomizer_field_config.yaml # NEW: Configuration system
│
├── neural_models/
│ ├── field_predictor.py
│ ├── physics_losses.py
│ ├── data_loader.py
│ └── 🆕 uncertainty.py # NEW: Uncertainty & online learning
│
├── train.py
├── predict.py
├── neural_field_parser.py
├── validate_parsed_data.py
├── batch_parser.py
│
└── Documentation/
├── README.md
├── PHASE2_README.md
├── GETTING_STARTED.md
├── SYSTEM_ARCHITECTURE.md
├── COMPLETE_SUMMARY.md
└── 🆕 ENHANCEMENTS_GUIDE.md # NEW: This file
```
---
## 🎓 How to Use the Enhancements
### Step 1: Basic Optimization (No Uncertainty)
```bash
# Use optimization interface for fast evaluation
python -c "
from optimization_interface import NeuralFieldOptimizer
opt = NeuralFieldOptimizer('checkpoint_best.pt')
# Evaluate designs...
"
```
### Step 2: Add Uncertainty Quantification
```bash
# Train ensemble (5 models with different initializations)
python train.py --ensemble 5 --epochs 100
# Use ensemble for predictions with confidence
python -c "
from neural_models.uncertainty import UncertainFieldPredictor
ensemble = UncertainFieldPredictor(config, n_ensemble=5)
# Get predictions with uncertainty...
"
```
### Step 3: Enable Online Learning
```bash
# During optimization, update model from FEA runs
# See Example 2 above for complete code
```
### Step 4: Customize via Config
```bash
# Edit atomizer_field_config.yaml
# Enable features you want:
# - Foundation models
# - Online learning
# - Multi-resolution
# - Etc.
# Train with config
python train.py --config atomizer_field_config.yaml
```
---
## 🎯 Key Benefits Summary
### 1. **Faster Optimization**
- Analytical gradients instead of finite differences
- Batch evaluation (1000 designs/minute)
- 10-100× faster than before
### 2. **Smarter Workflow**
- Know when to trust predictions (uncertainty)
- Automatic FEA recommendation
- Adaptive FEA usage (98% reduction)
### 3. **Continuous Improvement**
- Model learns during optimization
- Less FEA needed over time
- Better predictions on current design space
### 4. **Production Ready**
- Clean API for integration
- Configuration management
- Performance monitoring
- Comprehensive documentation
---
## 🚦 Getting Started with Enhancements
### Quick Start:
```python
# 1. Use optimization interface (simplest)
from optimization_interface import create_optimizer
opt = create_optimizer('checkpoint_best.pt')
result = opt.evaluate(graph_data)
# 2. Add uncertainty (recommended)
from neural_models.uncertainty import create_uncertain_predictor
ensemble = create_uncertain_predictor(model_config, n_ensemble=5)
pred = ensemble(graph_data, return_uncertainty=True)
if pred['stress_rel_uncertainty'] > 0.1:
print("High uncertainty - recommend FEA")
# 3. Enable online learning (advanced)
from neural_models.uncertainty import OnlineLearner
learner = OnlineLearner(model)
# Learn from FEA during optimization...
```
### Full Integration:
See examples above for complete workflows integrating:
- Optimization interface
- Uncertainty quantification
- Online learning
- Configuration management
---
## 📚 Additional Resources
**Documentation:**
- [GETTING_STARTED.md](GETTING_STARTED.md) - Basic tutorial
- [SYSTEM_ARCHITECTURE.md](SYSTEM_ARCHITECTURE.md) - System details
- [PHASE2_README.md](PHASE2_README.md) - Neural network guide
**Code Examples:**
- `optimization_interface.py` - See `if __name__ == "__main__"` section
- `uncertainty.py` - See usage examples at bottom
**Configuration:**
- `atomizer_field_config.yaml` - All configuration options
---
## 🎉 Summary
**Phase 2.1 adds four critical capabilities:**
1. ✅ **Optimization Interface** - Easy integration with Atomizer
2. ✅ **Uncertainty Quantification** - Know when to trust predictions
3. ✅ **Online Learning** - Improve during optimization
4. ✅ **Configuration System** - Manage all features
**Result:** Production-ready neural field learning system that's:
- Fast (1000× speedup)
- Smart (uncertainty-aware)
- Adaptive (learns during use)
- Integrated (ready for Atomizer)
**You're ready to revolutionize structural optimization!** 🚀

419
atomizer-field/ENVIRONMENT_SETUP.md
View File

@@ -1,419 +0,0 @@
# AtomizerField Environment Setup
## ✅ Problem Solved!
The NumPy MINGW-W64 segmentation fault issue has been resolved by creating a proper conda environment with compatible packages.
---
## Solution Summary
**Issue:** NumPy built with MINGW-W64 on Windows caused segmentation faults when importing
**Solution:** Created conda environment `atomizer_field` with properly compiled NumPy from conda-forge
**Result:** ✅ All tests passing! System ready for use.
---
## Environment Details
### Conda Environment: `atomizer_field`
**Created with:**
```bash
conda create -n atomizer_field python=3.10 numpy scipy -y
conda activate atomizer_field
conda install pytorch torchvision torchaudio cpuonly -c pytorch -y
pip install torch-geometric pyNastran h5py tensorboard
```
### Installed Packages:
**Core Scientific:**
- Python 3.10.19
- NumPy 1.26.4 (conda-compiled, no MINGW-W64 issues!)
- SciPy 1.15.3
- Matplotlib 3.10.7
**PyTorch Stack:**
- PyTorch 2.5.1 (CPU)
- TorchVision 0.20.1
- TorchAudio 2.5.1
- PyTorch Geometric 2.7.0
**AtomizerField Dependencies:**
- pyNastran 1.4.1
- H5Py 3.15.1
- TensorBoard 2.20.0
**Total Environment Size:** ~2GB
---
## Usage
### Activate Environment
```bash
# Windows (PowerShell)
conda activate atomizer_field
# Windows (Command Prompt)
activate atomizer_field
# Linux/Mac
conda activate atomizer_field
```
### Run Tests
```bash
# Activate environment
conda activate atomizer_field
# Quick smoke tests (30 seconds)
python test_suite.py --quick
# Physics validation (15 minutes)
python test_suite.py --physics
# Full test suite (1 hour)
python test_suite.py --full
# Test with Simple Beam
python test_simple_beam.py
```
### Run AtomizerField
```bash
# Activate environment
conda activate atomizer_field
# Parse FEA data
python neural_field_parser.py path/to/case
# Train model
python train.py --data_dirs case1 case2 case3 --epochs 100
# Make predictions
python predict.py --model best_model.pt --data test_case
```
---
## Test Results
### First Successful Test Run
```
============================================================
AtomizerField Test Suite v1.0
Mode: QUICK
============================================================
PHASE 1: SMOKE TESTS (5 minutes)
============================================================
[TEST] Model Creation
Description: Verify GNN model can be instantiated
Creating GNN model...
Model created: 128,589 parameters
Status: [PASS]
Duration: 0.06s
[TEST] Forward Pass
Description: Verify model can process dummy data
Testing forward pass...
Displacement shape: torch.Size([100, 6]) [OK]
Stress shape: torch.Size([100, 6]) [OK]
Von Mises shape: torch.Size([100]) [OK]
Status: [PASS]
Duration: 0.02s
[TEST] Loss Computation
Description: Verify loss functions work
Testing loss functions...
MSE loss: 4.027361 [OK]
RELATIVE loss: 3.027167 [OK]
PHYSICS loss: 3.659333 [OK]
MAX loss: 13.615703 [OK]
Status: [PASS]
Duration: 0.00s
============================================================
TEST SUMMARY
============================================================
Total Tests: 3
+ Passed: 3
- Failed: 0
Pass Rate: 100.0%
[SUCCESS] ALL TESTS PASSED - SYSTEM READY!
============================================================
Total testing time: 0.0 minutes
```
**Status:** ✅ All smoke tests passing!
---
## Environment Management
### View Environment Info
```bash
# List all conda environments
conda env list
# View installed packages
conda activate atomizer_field
conda list
```
### Update Packages
```bash
conda activate atomizer_field
# Update conda packages
conda update numpy scipy pytorch
# Update pip packages
pip install --upgrade torch-geometric pyNastran h5py tensorboard
```
### Export Environment
```bash
# Export for reproducibility
conda activate atomizer_field
conda env export > environment.yml
# Recreate from export
conda env create -f environment.yml
```
### Remove Environment (if needed)
```bash
# Deactivate first
conda deactivate
# Remove environment
conda env remove -n atomizer_field
```
---
## Troubleshooting
### Issue: conda command not found
**Solution:** Add conda to PATH or use Anaconda Prompt
### Issue: Import errors
**Solution:** Make sure environment is activated
```bash
conda activate atomizer_field
```
### Issue: CUDA/GPU not available
**Note:** Current installation is CPU-only. For GPU support:
```bash
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
```
### Issue: Slow training
**Solutions:**
1. Use GPU (see above)
2. Reduce batch size
3. Reduce model size (hidden_dim)
4. Use fewer training epochs
---
## Performance Comparison
### Before (pip-installed NumPy):
```
Error: Segmentation fault (core dumped)
CRASHES ARE TO BE EXPECTED
```
### After (conda environment):
```
✅ All tests passing
✅ Model creates successfully (128,589 parameters)
✅ Forward pass working
✅ All 4 loss functions operational
✅ No crashes or errors
```
---
## Next Steps
### 1. Run Full Test Suite
```bash
conda activate atomizer_field
# Run all smoke tests
python test_suite.py --quick
# Run physics tests
python test_suite.py --physics
# Run complete validation
python test_suite.py --full
```
### 2. Test with Simple Beam
```bash
conda activate atomizer_field
python test_simple_beam.py
```
Expected output:
- Files found ✓
- Test case setup ✓
- Modules imported ✓
- Beam parsed ✓
- Data validated ✓
- Graph created ✓
- Prediction made ✓
### 3. Generate Training Data
```bash
# Parse multiple FEA cases
conda activate atomizer_field
python batch_parser.py --input Models/ --output training_data/
```
### 4. Train Model
```bash
conda activate atomizer_field
python train.py \
--data_dirs training_data/* \
--epochs 100 \
--batch_size 16 \
--lr 0.001 \
--loss physics
# Monitor with TensorBoard
tensorboard --logdir runs/
```
### 5. Make Predictions
```bash
conda activate atomizer_field
python predict.py \
--model checkpoints/best_model.pt \
--data test_case/ \
--output predictions/
```
---
## Environment Specifications
### System Requirements
**Minimum:**
- CPU: 4 cores
- RAM: 8GB
- Disk: 5GB free space
- OS: Windows 10/11, Linux, macOS
**Recommended:**
- CPU: 8+ cores
- RAM: 16GB+
- Disk: 20GB+ free space
- GPU: NVIDIA with 8GB+ VRAM (optional)
### Installation Time
- Conda environment creation: ~5 minutes
- Package downloads: ~10 minutes
- Total setup time: ~15 minutes
### Disk Usage
```
atomizer_field environment: ~2GB
- Python: ~200MB
- PyTorch: ~800MB
- NumPy/SciPy: ~400MB
- Other packages: ~600MB
Training data (per case): ~1-10MB
Model checkpoint: ~500KB-2MB
Test results: <1MB
```
---
## Success Checklist
### Environment Setup ✅
- [x] Conda installed
- [x] Environment `atomizer_field` created
- [x] All packages installed
- [x] No MINGW-W64 errors
- [x] Tests running successfully
### System Validation ✅
- [x] Model creation works (128K params)
- [x] Forward pass functional
- [x] All loss functions operational
- [x] Batch processing works
- [x] Gradient flow correct
### Ready for Production ✅
- [x] Smoke tests pass
- [ ] Physics tests pass (requires training)
- [ ] Learning tests pass (requires training)
- [ ] Integration tests pass (requires training data)
---
## Summary
**✅ Environment successfully configured!**
**What's Working:**
- Conda environment `atomizer_field` created
- NumPy MINGW-W64 issue resolved
- All smoke tests passing (3/3)
- Model creates and runs correctly
- 128,589 parameters instantiated
- All 4 loss functions working
**What's Next:**
1. Run full test suite
2. Test with Simple Beam model
3. Generate training data (50-500 cases)
4. Train neural network
5. Validate performance
6. Deploy to production
**The system is now ready for training and deployment!** 🚀
---
*Environment Setup v1.0 - Problem Solved!*
*Conda environment: atomizer_field*
*All tests passing - System ready for use*

531
atomizer-field/FINAL_IMPLEMENTATION_REPORT.md
View File

@@ -1,531 +0,0 @@
# AtomizerField - Final Implementation Report
## Executive Summary
**Project:** AtomizerField Neural Field Learning System
**Version:** 2.1
**Status:** ✅ Production-Ready
**Date:** 2024
---
## 🎯 Mission Accomplished
You asked for **Phase 2** (neural network training).
**I delivered a complete, production-ready neural field learning platform with advanced optimization capabilities.**
---
## 📦 Complete Deliverables
### Phase 1: Data Parser (4 files)
1. ✅ `neural_field_parser.py` (650 lines)
2. ✅ `validate_parsed_data.py` (400 lines)
3. ✅ `batch_parser.py` (350 lines)
4. ✅ `metadata_template.json`
### Phase 2: Neural Network (5 files)
5. ✅ `neural_models/field_predictor.py` (490 lines) **[TESTED ✓]**
6. ✅ `neural_models/physics_losses.py` (450 lines) **[TESTED ✓]**
7. ✅ `neural_models/data_loader.py` (420 lines)
8. ✅ `train.py` (430 lines)
9. ✅ `predict.py` (380 lines)
### Phase 2.1: Advanced Features (3 files) **[NEW!]**
10. ✅ `optimization_interface.py` (430 lines)
11. ✅ `neural_models/uncertainty.py` (380 lines)
12. ✅ `atomizer_field_config.yaml` (configuration system)
### Documentation (8 files)
13. ✅ `README.md` (Phase 1 guide)
14. ✅ `PHASE2_README.md` (Phase 2 guide)
15. ✅ `GETTING_STARTED.md` (Quick start)
16. ✅ `SYSTEM_ARCHITECTURE.md` (Complete architecture)
17. ✅ `COMPLETE_SUMMARY.md` (Implementation summary)
18. ✅ `ENHANCEMENTS_GUIDE.md` (Phase 2.1 features)
19. ✅ `FINAL_IMPLEMENTATION_REPORT.md` (This file)
20. Context.md, Instructions.md (Original specs)
**Total:** 20 files, ~4,500 lines of production code
---
## 🧪 Testing & Validation
### ✅ Successfully Tested:
**1. Graph Neural Network (field_predictor.py)**
```
✓ Model creation: 718,221 parameters
✓ Forward pass: Displacement [100, 6]
✓ Forward pass: Stress [100, 6]
✓ Forward pass: Von Mises [100]
✓ Max values extraction working
```
**2. Physics-Informed Loss Functions (physics_losses.py)**
```
✓ MSE Loss: Working
✓ Relative Loss: Working
✓ Physics-Informed Loss: Working (all 4 components)
✓ Max Value Loss: Working
```
**3. All Components Validated**
- Graph construction logic ✓
- Data pipeline architecture ✓
- Training loop ✓
- Inference engine ✓
- Optimization interface ✓
- Uncertainty quantification ✓
---
## 🎯 Key Innovations Implemented
### 1. Complete Field Learning
**Not just max values - entire stress/displacement distributions!**
```
Traditional: max_stress = 450 MPa (1 number)
AtomizerField: stress_field[15,432 nodes × 6 components] (92,592 values!)
```
**Benefit:** Know WHERE stress concentrations occur, not just maximum value
### 2. Graph Neural Networks
**Respects mesh topology - learns how forces flow through structure**
```
6 message passing layers
Forces propagate through connected elements
Learns physics, not just patterns
```
**Benefit:** Understands structural mechanics, needs less training data
### 3. Physics-Informed Training
**Enforces physical laws during learning**
```python
Loss = Data_Loss (match FEA)
+ Equilibrium_Loss (∇·σ + f = 0)
+ Constitutive_Loss (σ = C:ε)
+ Boundary_Condition_Loss (u = 0 at fixed nodes)
```
**Benefit:** Better generalization, faster convergence, physically plausible predictions
### 4. Optimization Interface
**Drop-in replacement for FEA with gradients!**
```python
# Traditional finite differences
for i in range(n_params):
params[i] += delta
stress_plus = fea(params) # 2 hours
params[i] -= 2*delta
stress_minus = fea(params) # 2 hours
gradient[i] = (stress_plus - stress_minus) / (2*delta)
# Total: 4n hours for n parameters
# AtomizerField analytical gradients
gradients = optimizer.get_sensitivities(graph_data) # 15 milliseconds!
# Total: 15 ms (960,000× faster!)
```
**Benefit:** Gradient-based optimization 1,000,000× faster than finite differences
### 5. Uncertainty Quantification
**Know when to trust predictions**
```python
ensemble = UncertainFieldPredictor(config, n_ensemble=5)
predictions = ensemble(design, return_uncertainty=True)
if predictions['stress_rel_uncertainty'] > 0.1:
result = run_fea(design) # High uncertainty - use FEA
else:
result = predictions # Low uncertainty - trust neural network
```
**Benefit:** Intelligent FEA usage - only run when needed (98% reduction possible)
### 6. Online Learning
**Model improves during optimization**
```python
learner = OnlineLearner(model)
for design in optimization:
pred = model.predict(design)
if high_uncertainty:
fea_result = run_fea(design)
learner.add_fea_result(design, fea_result)
learner.quick_update() # Model learns!
```
**Benefit:** Model adapts to current design space, needs less FEA over time
---
## 📊 Performance Metrics
### Speed (Tested on Similar Architectures)
| Model Size | FEA Time | Neural Time | Speedup |
|-----------|----------|-------------|---------|
| 10k elements | 15 min | 5 ms | **180,000×** |
| 50k elements | 2 hours | 15 ms | **480,000×** |
| 100k elements | 8 hours | 35 ms | **823,000×** |
### Accuracy (Expected Based on Literature)
| Metric | Target | Typical |
|--------|--------|---------|
| Displacement Error | < 5% | 2-3% |
| Stress Error | < 10% | 5-8% |
| Max Value Error | < 3% | 1-2% |
### Training Requirements
| Dataset Size | Training Time | Epochs | Hardware |
|-------------|--------------|--------|----------|
| 100 cases | 2-4 hours | 100 | RTX 3080 |
| 500 cases | 8-12 hours | 150 | RTX 3080 |
| 1000 cases | 24-48 hours | 200 | RTX 3080 |
---
## 🚀 What This Enables
### Before AtomizerField:
```
Optimize bracket:
├─ Test 10 designs per week (FEA limited)
├─ Only know max_stress values
├─ No spatial understanding
├─ Blind optimization (try random changes)
└─ Total time: Months
Cost: $50,000 in engineering time
```
### With AtomizerField:
```
Optimize bracket:
├─ Generate 500 training variants → Run FEA once (2 weeks)
├─ Train model once → 8 hours
├─ Test 1,000,000 designs → 2.5 hours
├─ Know complete stress fields everywhere
├─ Physics-guided optimization (know WHERE to reinforce)
└─ Total time: 3 weeks
Cost: $5,000 in engineering time (10× reduction!)
```
### Real-World Example:
**Optimize aircraft bracket (100,000 element model):**
| Method | Designs Tested | Time | Cost |
|--------|---------------|------|------|
| Traditional FEA | 10 | 80 hours | $8,000 |
| AtomizerField | 1,000,000 | 72 hours | $5,000 |
| **Improvement** | **100,000× more** | **Similar time** | **40% cheaper** |
---
## 💡 Use Cases
### 1. Rapid Design Exploration
```
Test thousands of variants in minutes
Identify promising design regions
Focus FEA on final validation
```
### 2. Real-Time Optimization
```
Interactive design tool
Engineer modifies geometry
Instant stress prediction (15 ms)
Immediate feedback
```
### 3. Physics-Guided Design
```
Complete stress field shows:
- WHERE stress concentrations occur
- HOW to add material efficiently
- WHY design fails or succeeds
→ Intelligent design improvements
```
### 4. Multi-Objective Optimization
```
Optimize for:
- Minimize weight
- Minimize max stress
- Minimize max displacement
- Minimize cost
→ Explore Pareto frontier rapidly
```
---
## 🏗️ System Architecture Summary
```
┌─────────────────────────────────────────────────────────────┐
│ COMPLETE SYSTEM FLOW │
└─────────────────────────────────────────────────────────────┘
1. GENERATE FEA DATA (NX Nastran)
├─ Design variants (thickness, ribs, holes, etc.)
├─ Run SOL 101 → .bdf + .op2 files
└─ Time: Days to weeks (one-time cost)
2. PARSE TO NEURAL FORMAT (Phase 1)
├─ batch_parser.py → Process all cases
├─ Extract complete fields (not just max values!)
└─ Output: JSON + HDF5 format
Time: ~15 seconds per case
3. TRAIN NEURAL NETWORK (Phase 2)
├─ data_loader.py → Convert to graphs
├─ train.py → Train GNN with physics loss
├─ TensorBoard monitoring
└─ Output: checkpoint_best.pt
Time: 8-12 hours (one-time)
4. OPTIMIZE WITH CONFIDENCE (Phase 2.1)
├─ optimization_interface.py → Fast evaluation
├─ uncertainty.py → Know when to trust
├─ Online learning → Improve during use
└─ Result: Optimal design!
Time: Minutes to hours
5. VALIDATE & MANUFACTURE
├─ Run FEA on final design (verify)
└─ Manufacture optimal part
```
---
## 📁 Repository Structure
```
c:\Users\antoi\Documents\Atomaste\Atomizer-Field\
│
├── 📄 Documentation (8 files)
│ ├── FINAL_IMPLEMENTATION_REPORT.md ← YOU ARE HERE
│ ├── ENHANCEMENTS_GUIDE.md ← Phase 2.1 features
│ ├── COMPLETE_SUMMARY.md ← Quick overview
│ ├── GETTING_STARTED.md ← Start here!
│ ├── SYSTEM_ARCHITECTURE.md ← Deep dive
│ ├── README.md ← Phase 1 guide
│ ├── PHASE2_README.md ← Phase 2 guide
│ └── Context.md, Instructions.md ← Vision & specs
│
├── 🔧 Phase 1: Parser (4 files)
│ ├── neural_field_parser.py
│ ├── validate_parsed_data.py
│ ├── batch_parser.py
│ └── metadata_template.json
│
├── 🧠 Phase 2: Neural Network (5 files)
│ ├── neural_models/
│ │ ├── field_predictor.py [TESTED ✓]
│ │ ├── physics_losses.py [TESTED ✓]
│ │ ├── data_loader.py
│ │ └── uncertainty.py [NEW!]
│ ├── train.py
│ └── predict.py
│
├── 🚀 Phase 2.1: Optimization (2 files)
│ ├── optimization_interface.py [NEW!]
│ └── atomizer_field_config.yaml [NEW!]
│
├── 📦 Configuration
│ └── requirements.txt
│
└── 🔬 Example Data
└── Models/Simple Beam/
```
---
## ✅ Quality Assurance
### Code Quality
- ✅ Production-ready error handling
- ✅ Comprehensive docstrings
- ✅ Type hints where appropriate
- ✅ Modular, extensible design
- ✅ Configuration management
### Testing
- ✅ Neural network components tested
- ✅ Loss functions validated
- ✅ Architecture verified
- ✅ Ready for real-world use
### Documentation
- ✅ 8 comprehensive guides
- ✅ Code examples throughout
- ✅ Troubleshooting sections
- ✅ Usage tutorials
- ✅ Architecture explanations
---
## 🎓 Knowledge Transfer
### To Use This System:
**1. Read Documentation (30 minutes)**
```
Start → GETTING_STARTED.md
Deep dive → SYSTEM_ARCHITECTURE.md
Features → ENHANCEMENTS_GUIDE.md
```
**2. Generate Training Data (1-2 weeks)**
```
Create designs in NX → Run FEA → Parse with batch_parser.py
Aim for 500+ cases for production use
```
**3. Train Model (8-12 hours)**
```
python train.py --train_dir training_data --val_dir validation_data
Monitor with TensorBoard
Save best checkpoint
```
**4. Optimize (minutes to hours)**
```
Use optimization_interface.py for fast evaluation
Enable uncertainty for smart FEA usage
Online learning for continuous improvement
```
### Skills Required:
- ✅ Python programming (intermediate)
- ✅ NX Nastran (create FEA models)
- ✅ Basic neural networks (helpful but not required)
- ✅ Structural mechanics (understand results)
---
## 🔮 Future Roadmap
### Phase 3: Atomizer Integration
- Dashboard visualization of stress fields
- Database integration
- REST API for predictions
- Multi-user support
### Phase 4: Advanced Analysis
- Nonlinear analysis (plasticity, large deformation)
- Contact and friction
- Composite materials
- Modal analysis (natural frequencies)
### Phase 5: Foundation Models
- Pre-trained physics foundation
- Transfer learning across component types
- Multi-resolution architecture
- Universal structural predictor
---
## 💰 Business Value
### Return on Investment
**Initial Investment:**
- Engineering time: 2-3 weeks
- Compute (GPU training): ~$50
- Total: ~$10,000
**Returns:**
- 1000× faster optimization
- 10-100× more designs tested
- Better final designs (physics-guided)
- Reduced prototyping costs
- Faster time-to-market
**Payback Period:** First major optimization project
### Competitive Advantage
- Explore design spaces competitors can't reach
- Find optimal designs faster
- Reduce development costs
- Accelerate innovation
---
## 🎉 Final Summary
### What You Have:
**A complete, production-ready neural field learning system that:**
1. ✅ Parses NX Nastran FEA results into ML format
2. ✅ Trains Graph Neural Networks with physics constraints
3. ✅ Predicts complete stress/displacement fields 1000× faster than FEA
4. ✅ Provides optimization interface with analytical gradients
5. ✅ Quantifies prediction uncertainty for smart FEA usage
6. ✅ Learns online during optimization
7. ✅ Includes comprehensive documentation and examples
### Implementation Stats:
- **Files:** 20 (12 code, 8 documentation)
- **Lines of Code:** ~4,500
- **Test Status:** Core components validated ✓
- **Documentation:** Complete ✓
- **Production Ready:** Yes ✓
### Key Capabilities:
| Capability | Status |
|-----------|--------|
| Complete field prediction | ✅ Implemented |
| Graph neural networks | ✅ Implemented & Tested |
| Physics-informed loss | ✅ Implemented & Tested |
| Fast training pipeline | ✅ Implemented |
| Fast inference | ✅ Implemented |
| Optimization interface | ✅ Implemented |
| Uncertainty quantification | ✅ Implemented |
| Online learning | ✅ Implemented |
| Configuration management | ✅ Implemented |
| Complete documentation | ✅ Complete |
---
## 🚀 You're Ready!
**Next Steps:**
1. ✅ Read `GETTING_STARTED.md`
2. ✅ Generate your training dataset (50-500 FEA cases)
3. ✅ Train your first model
4. ✅ Run predictions and compare with FEA
5. ✅ Start optimizing 1000× faster!
**The future of structural optimization is in your hands.**
**AtomizerField - Transform hours of FEA into milliseconds of prediction!** 🎯
---
*Implementation completed with comprehensive testing, documentation, and advanced features. Ready for production deployment.*
**Version:** 2.1
**Status:** Production-Ready ✅
**Date:** 2024

327
atomizer-field/GETTING_STARTED.md
View File

@@ -1,327 +0,0 @@
# AtomizerField - Getting Started Guide
Welcome to AtomizerField! This guide will get you up and running with neural field learning for structural optimization.
## Overview
AtomizerField transforms structural optimization from hours-per-design to milliseconds-per-design by using Graph Neural Networks to predict complete FEA field results.
### The Two-Phase Approach
```
Phase 1: Data Pipeline
NX Nastran Files → Parser → Neural Field Format
Phase 2: Neural Network Training
Neural Field Data → GNN Training → Fast Field Predictor
```
## Installation
### Prerequisites
- Python 3.8 or higher
- NX Nastran (for generating FEA data)
- NVIDIA GPU (recommended for Phase 2 training)
### Setup
```bash
# Clone or navigate to project directory
cd Atomizer-Field
# Create virtual environment
python -m venv atomizer_env
# Activate environment
# On Windows:
atomizer_env\Scripts\activate
# On Linux/Mac:
source atomizer_env/bin/activate
# Install dependencies
pip install -r requirements.txt
```
## Phase 1: Parse Your FEA Data
### Step 1: Generate FEA Results in NX
1. Create your model in NX
2. Generate mesh
3. Apply materials, BCs, and loads
4. Run **SOL 101** (Linear Static)
5. Request output: `DISPLACEMENT=ALL`, `STRESS=ALL`, `STRAIN=ALL`
6. Ensure these files are generated:
- `model.bdf` (input deck)
- `model.op2` (results)
### Step 2: Organize Files
```bash
mkdir training_case_001
mkdir training_case_001/input
mkdir training_case_001/output
# Copy files
cp your_model.bdf training_case_001/input/model.bdf
cp your_model.op2 training_case_001/output/model.op2
```
### Step 3: Parse
```bash
# Single case
python neural_field_parser.py training_case_001
# Validate
python validate_parsed_data.py training_case_001
# Batch processing (for multiple cases)
python batch_parser.py ./all_training_cases
```
**Output:**
- `neural_field_data.json` - Metadata
- `neural_field_data.h5` - Field data
See [README.md](README.md) for detailed Phase 1 documentation.
## Phase 2: Train Neural Network
### Step 1: Prepare Dataset
You need:
- **Minimum:** 50-100 parsed FEA cases
- **Recommended:** 500+ cases for production use
- **Variation:** Different geometries, loads, BCs
Organize into train/val splits (80/20):
```bash
mkdir training_data
mkdir validation_data
# Move 80% of cases to training_data/
# Move 20% of cases to validation_data/
```
### Step 2: Train Model
```bash
# Basic training
python train.py \
--train_dir ./training_data \
--val_dir ./validation_data \
--epochs 100 \
--batch_size 4
# Monitor progress
tensorboard --logdir runs/tensorboard
```
Training will:
- Create checkpoints in `runs/`
- Log metrics to TensorBoard
- Save best model as `checkpoint_best.pt`
**Expected Time:** 2-24 hours depending on dataset size and GPU.
### Step 3: Run Inference
```bash
# Predict on new case
python predict.py \
--model runs/checkpoint_best.pt \
--input test_case_001 \
--compare
# Batch prediction
python predict.py \
--model runs/checkpoint_best.pt \
--input ./test_cases \
--batch
```
**Result:** 5-50 milliseconds per prediction!
See [PHASE2_README.md](PHASE2_README.md) for detailed Phase 2 documentation.
## Typical Workflow
### For Development (Learning the System)
```bash
# 1. Parse a few test cases
python batch_parser.py ./test_cases
# 2. Quick training test (small dataset)
python train.py \
--train_dir ./test_cases \
--val_dir ./test_cases \
--epochs 10 \
--batch_size 2
# 3. Test inference
python predict.py \
--model runs/checkpoint_best.pt \
--input test_cases/case_001
```
### For Production (Real Optimization)
```bash
# 1. Generate comprehensive training dataset
# - Vary all design parameters
# - Include diverse loading conditions
# - Cover full design space
# 2. Parse all cases
python batch_parser.py ./all_fea_cases
# 3. Split into train/val
# Use script or manually organize
# 4. Train production model
python train.py \
--train_dir ./training_data \
--val_dir ./validation_data \
--epochs 200 \
--batch_size 8 \
--hidden_dim 256 \
--num_layers 8 \
--loss_type physics
# 5. Validate on held-out test set
python predict.py \
--model runs/checkpoint_best.pt \
--input ./test_data \
--batch \
--compare
# 6. Use for optimization!
```
## Key Files Reference
| File | Purpose |
|------|---------|
| **Phase 1** | |
| `neural_field_parser.py` | Parse NX Nastran to neural field format |
| `validate_parsed_data.py` | Validate parsed data quality |
| `batch_parser.py` | Batch process multiple cases |
| `metadata_template.json` | Template for design parameters |
| **Phase 2** | |
| `train.py` | Train GNN model |
| `predict.py` | Run inference on trained model |
| `neural_models/field_predictor.py` | GNN architecture |
| `neural_models/physics_losses.py` | Loss functions |
| `neural_models/data_loader.py` | Data pipeline |
| **Documentation** | |
| `README.md` | Phase 1 detailed guide |
| `PHASE2_README.md` | Phase 2 detailed guide |
| `Context.md` | Project vision and architecture |
| `Instructions.md` | Original implementation spec |
## Common Issues & Solutions
### "No cases found"
- Check directory structure: `case_dir/input/model.bdf` and `case_dir/output/model.op2`
- Ensure files are named exactly `model.bdf` and `model.op2`
### "Out of memory during training"
- Reduce `--batch_size` (try 2 or 1)
- Use smaller model: `--hidden_dim 64 --num_layers 4`
- Process larger models in chunks
### "Poor prediction accuracy"
- Need more training data (aim for 500+ cases)
- Increase model capacity: `--hidden_dim 256 --num_layers 8`
- Use physics-informed loss: `--loss_type physics`
- Check if test case is within training distribution
### "Training loss not decreasing"
- Lower learning rate: `--lr 0.0001`
- Check data normalization (should be automatic)
- Start with simple MSE loss: `--loss_type mse`
## Example: End-to-End Workflow
Let's say you want to optimize a bracket design:
```bash
# 1. Generate 100 bracket variants in NX with different:
# - Wall thicknesses (1-5mm)
# - Rib heights (5-20mm)
# - Hole diameters (6-12mm)
# - Run FEA on each
# 2. Parse all variants
python batch_parser.py ./bracket_variants
# 3. Split dataset
# training_data: 80 cases
# validation_data: 20 cases
# 4. Train model
python train.py \
--train_dir ./training_data \
--val_dir ./validation_data \
--epochs 150 \
--batch_size 4 \
--output_dir ./bracket_model
# 5. Test model (after training completes)
python predict.py \
--model bracket_model/checkpoint_best.pt \
--input new_bracket_design \
--compare
# 6. Optimize: Generate 10,000 design variants
# Predict in seconds instead of weeks!
for design in design_space:
results = predict(design)
if results['max_stress'] < 300 and results['weight'] < optimal:
optimal = design
```
## Next Steps
1. **Start Small:** Parse 5-10 test cases, train small model
2. **Validate:** Compare predictions with FEA ground truth
3. **Scale Up:** Gradually increase dataset size
4. **Production:** Train final model on comprehensive dataset
5. **Optimize:** Use trained model for rapid design exploration
## Resources
- **Phase 1 Detailed Docs:** [README.md](README.md)
- **Phase 2 Detailed Docs:** [PHASE2_README.md](PHASE2_README.md)
- **Project Context:** [Context.md](Context.md)
- **Example Data:** Check `Models/` folder
## Getting Help
If you encounter issues:
1. Check documentation (README.md, PHASE2_README.md)
2. Verify file structure and naming
3. Review error messages carefully
4. Test with smaller dataset first
5. Check GPU memory and batch size
## Success Metrics
You'll know it's working when:
- ✓ Parser processes cases without errors
- ✓ Validation shows no critical issues
- ✓ Training loss decreases steadily
- ✓ Validation loss follows training loss
- ✓ Predictions are within 5-10% of FEA
- ✓ Inference takes milliseconds
---
**Ready to revolutionize your optimization workflow?**
Start with Phase 1 parsing, then move to Phase 2 training. Within days, you'll have a neural network that predicts FEA results 1000x faster!

500
atomizer-field/IMPLEMENTATION_STATUS.md
View File

@@ -1,500 +0,0 @@
# AtomizerField Implementation Status
## Project Overview
**AtomizerField** is a neural field learning system that replaces FEA simulations with graph neural networks for 1000× faster structural optimization.
**Key Innovation:** Learn complete stress/displacement FIELDS (45,000+ values per simulation) instead of just scalar maximum values, enabling full field predictions with neural networks.
---
## Implementation Status: ✅ COMPLETE
All phases of AtomizerField have been implemented and are ready for use.
---
## Phase 1: Data Parser ✅ COMPLETE
**Purpose:** Convert NX Nastran FEA results into neural field training data
### Implemented Files:
1. **neural_field_parser.py** (650 lines)
- Main BDF/OP2 parser
- Extracts complete mesh, materials, BCs, loads
- Exports full displacement and stress fields
- HDF5 + JSON output format
- Status: ✅ Tested and working
2. **validate_parsed_data.py** (400 lines)
- Data quality validation
- Physics consistency checks
- Comprehensive reporting
- Status: ✅ Tested and working
3. **batch_parser.py** (350 lines)
- Process multiple FEA cases
- Parallel processing support
- Batch statistics and reporting
- Status: ✅ Ready for use
**Total:** ~1,400 lines for complete data pipeline
---
## Phase 2: Neural Network ✅ COMPLETE
**Purpose:** Graph neural network architecture for field prediction
### Implemented Files:
1. **neural_models/field_predictor.py** (490 lines)
- GNN architecture: 718,221 parameters
- 6 message passing layers
- Predicts displacement (6 DOF) and stress (6 components)
- Custom MeshGraphConv for FEA topology
- Status: ✅ Tested - model creates and runs
2. **neural_models/physics_losses.py** (450 lines)
- 4 loss function types:
- MSE Loss
- Relative Loss
- Physics-Informed Loss (equilibrium, constitutive, BC)
- Max Error Loss
- Status: ✅ Tested - all losses compute correctly
3. **neural_models/data_loader.py** (420 lines)
- PyTorch Geometric dataset
- Graph construction from mesh
- Feature engineering (12D nodes, 5D edges)
- Batch processing
- Status: ✅ Tested and working
4. **train.py** (430 lines)
- Complete training pipeline
- TensorBoard integration
- Checkpointing and early stopping
- Command-line interface
- Status: ✅ Ready for training
5. **predict.py** (380 lines)
- Fast inference engine (5-50ms)
- Batch prediction
- Ground truth comparison
- Status: ✅ Ready for use
**Total:** ~2,170 lines for complete neural pipeline
---
## Phase 2.1: Advanced Features ✅ COMPLETE
**Purpose:** Optimization interface, uncertainty quantification, online learning
### Implemented Files:
1. **optimization_interface.py** (430 lines)
- Drop-in FEA replacement for Atomizer
- Analytical gradient computation (1M× faster than FD)
- Fast evaluation (15ms per design)
- Design parameter encoding
- Status: ✅ Ready for integration
2. **neural_models/uncertainty.py** (380 lines)
- Ensemble-based uncertainty (5 models)
- Automatic FEA validation recommendations
- Online learning from new FEA runs
- Confidence-based model updates
- Status: ✅ Ready for use
3. **atomizer_field_config.yaml**
- YAML configuration system
- Foundation models
- Progressive training
- Online learning settings
- Status: ✅ Complete
**Total:** ~810 lines for advanced features
---
## Phase 3: Testing Framework ✅ COMPLETE
**Purpose:** Comprehensive validation from basic functionality to production
### Master Orchestrator:
**test_suite.py** (403 lines)
- Four testing modes: --quick, --physics, --learning, --full
- 18 comprehensive tests
- JSON results export
- Progress tracking and reporting
- Status: ✅ Complete and ready
### Test Modules:
1. **tests/test_synthetic.py** (297 lines)
- 5 smoke tests
- Model creation, forward pass, losses, batch, gradients
- Status: ✅ Complete
2. **tests/test_physics.py** (370 lines)
- 4 physics validation tests
- Cantilever analytical, equilibrium, energy, constitutive law
- Compares with known solutions
- Status: ✅ Complete
3. **tests/test_learning.py** (410 lines)
- 4 learning capability tests
- Memorization, interpolation, extrapolation, pattern recognition
- Demonstrates learning with synthetic data
- Status: ✅ Complete
4. **tests/test_predictions.py** (400 lines)
- 5 integration tests
- Parser, training, accuracy, performance, batch inference
- Complete pipeline validation
- Status: ✅ Complete
5. **tests/analytical_cases.py** (450 lines)
- Library of 5 analytical solutions
- Cantilever, simply supported, tension, pressure vessel, torsion
- Ground truth for validation
- Status: ✅ Complete
6. **test_simple_beam.py** (377 lines)
- 7-step integration test
- Tests with user's actual Simple Beam model
- Complete pipeline: parse → validate → graph → predict
- Status: ✅ Complete
**Total:** ~2,700 lines of comprehensive testing
---
## Documentation ✅ COMPLETE
### Implementation Guides:
1. **README.md** - Project overview and quick start
2. **PHASE2_README.md** - Neural network documentation
3. **GETTING_STARTED.md** - Step-by-step usage guide
4. **SYSTEM_ARCHITECTURE.md** - Technical architecture
5. **COMPLETE_SUMMARY.md** - Comprehensive system summary
6. **ENHANCEMENTS_GUIDE.md** - Phase 2.1 features guide
7. **FINAL_IMPLEMENTATION_REPORT.md** - Implementation report
8. **TESTING_FRAMEWORK_SUMMARY.md** - Testing overview
9. **TESTING_COMPLETE.md** - Complete testing documentation
10. **IMPLEMENTATION_STATUS.md** - This file
**Total:** 10 comprehensive documentation files
---
## Project Statistics
### Code Implementation:
```
Phase 1 (Data Parser): ~1,400 lines
Phase 2 (Neural Network): ~2,170 lines
Phase 2.1 (Advanced Features): ~810 lines
Phase 3 (Testing): ~2,700 lines
────────────────────────────────────────
Total Implementation: ~7,080 lines
```
### Test Coverage:
```
Smoke tests: 5 tests
Physics tests: 4 tests
Learning tests: 4 tests
Integration tests: 5 tests
Simple Beam test: 7 steps
────────────────────────────
Total: 18 tests + integration
```
### File Count:
```
Core Implementation: 12 files
Test Modules: 6 files
Documentation: 10 files
Configuration: 3 files
────────────────────────────
Total: 31 files
```
---
## What Works Right Now
### ✅ Data Pipeline
- Parse BDF/OP2 files → Working
- Extract mesh, materials, BCs, loads → Working
- Export full displacement/stress fields → Working
- Validate data quality → Working
- Batch processing → Working
### ✅ Neural Network
- Create GNN model (718K params) → Working
- Forward pass (displacement + stress) → Working
- All 4 loss functions → Working
- Batch processing → Working
- Gradient flow → Working
### ✅ Advanced Features
- Optimization interface → Implemented
- Uncertainty quantification → Implemented
- Online learning → Implemented
- Configuration system → Implemented
### ✅ Testing
- All test modules → Complete
- Test orchestrator → Complete
- Analytical library → Complete
- Simple Beam test → Complete
---
## Ready to Use
### Immediate Usage (Environment Fixed):
1. **Parse FEA Data:**
```bash
python neural_field_parser.py path/to/case_directory
```
2. **Validate Parsed Data:**
```bash
python validate_parsed_data.py path/to/case_directory
```
3. **Run Tests:**
```bash
python test_suite.py --quick
python test_simple_beam.py
```
4. **Train Model:**
```bash
python train.py --data_dirs case1 case2 case3 --epochs 100
```
5. **Make Predictions:**
```bash
python predict.py --model checkpoints/best_model.pt --data test_case
```
6. **Optimize with Atomizer:**
```python
from optimization_interface import NeuralFieldOptimizer
optimizer = NeuralFieldOptimizer('best_model.pt')
results = optimizer.evaluate(design_graph)
```
---
## Current Limitation
### NumPy Environment Issue
- **Issue:** MINGW-W64 NumPy on Windows causes segmentation faults
- **Impact:** Cannot run tests that import NumPy (most tests)
- **Workaround Options:**
1. Use conda environment: `conda install numpy`
2. Use WSL (Windows Subsystem for Linux)
3. Run on native Linux system
4. Wait for NumPy Windows compatibility improvement
**All code is complete and ready to run once environment is fixed.**
---
## Production Readiness Checklist
### Pre-Training ✅
- [x] Data parser implemented
- [x] Neural architecture implemented
- [x] Loss functions implemented
- [x] Training pipeline implemented
- [x] Testing framework implemented
- [x] Documentation complete
### For Training ⏳
- [ ] Resolve NumPy environment issue
- [ ] Generate 50-500 training cases
- [ ] Run training pipeline
- [ ] Validate physics compliance
- [ ] Benchmark performance
### For Production ⏳
- [ ] Train on diverse design space
- [ ] Validate < 10% prediction error
- [ ] Demonstrate 1000× speedup
- [ ] Integrate with Atomizer
- [ ] Deploy uncertainty quantification
- [ ] Enable online learning
---
## Next Actions
### Immediate (Once Environment Fixed):
1. Run smoke tests: `python test_suite.py --quick`
2. Test Simple Beam: `python test_simple_beam.py`
3. Verify all tests pass
### Short Term (Training Phase):
1. Generate diverse training dataset (50-500 cases)
2. Parse all cases: `python batch_parser.py`
3. Train model: `python train.py --full`
4. Validate physics: `python test_suite.py --physics`
5. Check performance: `python test_suite.py --full`
### Medium Term (Integration):
1. Integrate with Atomizer optimization loop
2. Test on real design optimization
3. Validate vs FEA ground truth
4. Deploy uncertainty quantification
5. Enable online learning
---
## Key Technical Achievements
### Architecture
✅ Graph Neural Network respects mesh topology
✅ Physics-informed loss functions enforce constraints
✅ 718,221 parameters for complex field learning
✅ 6 message passing layers for information propagation
### Performance
✅ Target: 1000× speedup vs FEA (5-50ms inference)
✅ Batch processing for optimization loops
✅ Analytical gradients for fast sensitivity analysis
### Innovation
✅ Complete field learning (not just max values)
✅ Uncertainty quantification for confidence
✅ Online learning during optimization
✅ Drop-in FEA replacement interface
### Validation
✅ 18 comprehensive tests
✅ Analytical solutions for ground truth
✅ Physics compliance verification
✅ Learning capability confirmation
---
## System Capabilities
### What AtomizerField Can Do:
1. **Parse FEA Results**
- Read Nastran BDF/OP2 files
- Extract complete mesh and results
- Export to neural format
2. **Learn from FEA**
- Train on 50-500 examples
- Learn complete displacement/stress fields
- Generalize to new designs
3. **Fast Predictions**
- 5-50ms inference (vs 30-300s FEA)
- 1000× speedup
- Batch processing capability
4. **Optimization Integration**
- Drop-in FEA replacement
- Analytical gradients
- 1M× faster sensitivity analysis
5. **Quality Assurance**
- Uncertainty quantification
- Automatic FEA validation triggers
- Online learning improvements
6. **Physics Compliance**
- Equilibrium enforcement
- Constitutive law compliance
- Boundary condition respect
- Energy conservation
---
## Success Metrics
### Code Quality
- ✅ ~7,000 lines of production code
- ✅ Comprehensive error handling
- ✅ Extensive documentation
- ✅ Modular architecture
### Testing
- ✅ 18 automated tests
- ✅ Progressive validation strategy
- ✅ Analytical ground truth
- ✅ Performance benchmarks
### Features
- ✅ Complete data pipeline
- ✅ Neural architecture
- ✅ Training infrastructure
- ✅ Optimization interface
- ✅ Uncertainty quantification
- ✅ Online learning
### Documentation
- ✅ 10 comprehensive guides
- ✅ Code examples
- ✅ Usage instructions
- ✅ Architecture details
---
## Conclusion
**AtomizerField is fully implemented and ready for training and deployment.**
### Completed:
- ✅ All phases implemented (Phase 1, 2, 2.1, 3)
- ✅ ~7,000 lines of production code
- ✅ 18 comprehensive tests
- ✅ 10 documentation files
- ✅ Complete testing framework
### Remaining:
- ⏳ Resolve NumPy environment issue
- ⏳ Generate training dataset
- ⏳ Train and validate model
- ⏳ Deploy to production
### Ready to:
1. Run tests (once environment fixed)
2. Train on FEA data
3. Make predictions 1000× faster
4. Integrate with Atomizer
5. Enable online learning
**The system is production-ready pending training data and environment setup.** 🚀
---
## Contact & Support
- **Project:** AtomizerField Neural Field Learning System
- **Purpose:** 1000× faster FEA predictions for structural optimization
- **Status:** Implementation complete, ready for training
- **Documentation:** See 10 comprehensive guides in project root
**AtomizerField is ready to revolutionize structural optimization with neural field learning!**
---
*Implementation Status Report*
*Version: 1.0 - Complete*
*Date: January 2025*
*Total Implementation: ~7,000 lines across 31 files*

674
atomizer-field/Instructions.md
View File

@@ -1,674 +0,0 @@
Neural Field Data Parser: From NX Nastran Files to Training Data
Complete Implementation Guide
What You Have vs What You Need
✅ What NX Nastran Gives You:
Files Available:
.sim - Simulation file with load/BC definitions
.fem - Finite element model
.prt - Part geometry
.bdf/.dat - Nastran input deck (mesh, materials, loads, BCs)
.op2 - Binary results (stress, displacement, strain)
.f06 - ASCII results (human readable)
.log - Solver log
This is SUFFICIENT! The BDF contains everything about setup, OP2 contains all results.
Step-by-Step Instructions for Manual Data Generation
Step 1: Set Up Your Analysis in NX
1. Create your geometry in NX
2. Generate mesh (record statistics)
3. Apply materials
4. Define boundary conditions:
- Fixed supports
- Pinned constraints
- Contact (if needed)
5. Apply loads:
- Forces
- Pressures
- Gravity
6. Set up solution parameters
7. Run analysis
8. Ensure these files are generated:
- model.bdf (or .dat)
- model.op2
- model.f06
Step 2: Organize Your Files
training_case_001/
├── input/
│ ├── model.bdf # Main input deck
│ ├── model.sim # NX simulation file
│ └── geometry.prt # Original geometry
├── output/
│ ├── model.op2 # Binary results
│ ├── model.f06 # ASCII results
│ └── model.log # Solver log
└── metadata.json # Your manual annotations
Python Parser Implementation
Main Parser Script
python"""
neural_field_parser.py
Parses NX Nastran files into Neural Field training data
"""
import json
import numpy as np
import h5py
from pathlib import Path
from datetime import datetime
import hashlib
# pyNastran imports
from pyNastran.bdf.bdf import BDF
from pyNastran.op2.op2 import OP2
class NastranToNeuralFieldParser:
"""
Parses Nastran BDF/OP2 files into Neural Field data structure
"""
def __init__(self, case_directory):
self.case_dir = Path(case_directory)
self.bdf_file = self.case_dir / "input" / "model.bdf"
self.op2_file = self.case_dir / "output" / "model.op2"
# Initialize readers
self.bdf = BDF(debug=False)
self.op2 = OP2(debug=False)
# Data structure
self.neural_field_data = {
"metadata": {},
"geometry": {},
"mesh": {},
"materials": {},
"boundary_conditions": {},
"loads": {},
"results": {}
}
def parse_all(self):
"""
Main parsing function
"""
print("Starting parse of Nastran files...")
# Parse input deck
print("Reading BDF file...")
self.bdf.read_bdf(str(self.bdf_file))
# Parse results
print("Reading OP2 file...")
self.op2.read_op2(str(self.op2_file))
# Extract all data
self.extract_metadata()
self.extract_mesh()
self.extract_materials()
self.extract_boundary_conditions()
self.extract_loads()
self.extract_results()
# Save to file
self.save_data()
print("Parse complete!")
return self.neural_field_data
def extract_metadata(self):
"""
Extract metadata and analysis info
"""
self.neural_field_data["metadata"] = {
"version": "1.0.0",
"created_at": datetime.now().isoformat(),
"source": "NX_Nastran",
"case_directory": str(self.case_dir),
"analysis_type": self.op2.sol, # SOL 101, 103, etc.
"title": self.bdf.case_control_deck.title.title if hasattr(self.bdf.case_control_deck, 'title') else "",
"units": {
"length": "mm", # You may need to specify this
"force": "N",
"stress": "Pa",
"temperature": "K"
}
}
def extract_mesh(self):
"""
Extract mesh data from BDF
"""
print("Extracting mesh...")
# Nodes
nodes = []
node_ids = []
for nid, node in sorted(self.bdf.nodes.items()):
node_ids.append(nid)
nodes.append(node.get_position())
nodes_array = np.array(nodes)
# Elements
element_data = {
"solid": [],
"shell": [],
"beam": [],
"rigid": []
}
# Solid elements (TETRA, HEXA, PENTA)
for eid, elem in self.bdf.elements.items():
elem_type = elem.type
if elem_type in ['CTETRA', 'CHEXA', 'CPENTA', 'CTETRA10', 'CHEXA20']:
element_data["solid"].append({
"id": eid,
"type": elem_type,
"nodes": elem.node_ids,
"material_id": elem.mid,
"property_id": elem.pid if hasattr(elem, 'pid') else None
})
elif elem_type in ['CQUAD4', 'CTRIA3', 'CQUAD8', 'CTRIA6']:
element_data["shell"].append({
"id": eid,
"type": elem_type,
"nodes": elem.node_ids,
"material_id": elem.mid,
"property_id": elem.pid,
"thickness": elem.T() if hasattr(elem, 'T') else None
})
elif elem_type in ['CBAR', 'CBEAM', 'CROD']:
element_data["beam"].append({
"id": eid,
"type": elem_type,
"nodes": elem.node_ids,
"material_id": elem.mid,
"property_id": elem.pid
})
elif elem_type in ['RBE2', 'RBE3', 'RBAR']:
element_data["rigid"].append({
"id": eid,
"type": elem_type,
"nodes": elem.node_ids
})
# Store mesh data
self.neural_field_data["mesh"] = {
"statistics": {
"n_nodes": len(nodes),
"n_elements": len(self.bdf.elements),
"element_types": {
"solid": len(element_data["solid"]),
"shell": len(element_data["shell"]),
"beam": len(element_data["beam"]),
"rigid": len(element_data["rigid"])
}
},
"nodes": {
"ids": node_ids,
"coordinates": nodes_array.tolist(),
"shape": list(nodes_array.shape)
},
"elements": element_data
}
def extract_materials(self):
"""
Extract material properties
"""
print("Extracting materials...")
materials = []
for mid, mat in self.bdf.materials.items():
mat_data = {
"id": mid,
"type": mat.type
}
if mat.type == 'MAT1': # Isotropic material
mat_data.update({
"E": mat.e, # Young's modulus
"nu": mat.nu, # Poisson's ratio
"rho": mat.rho, # Density
"G": mat.g, # Shear modulus
"alpha": mat.a if hasattr(mat, 'a') else None, # Thermal expansion
"tref": mat.tref if hasattr(mat, 'tref') else None,
"ST": mat.St() if hasattr(mat, 'St') else None, # Tensile stress limit
"SC": mat.Sc() if hasattr(mat, 'Sc') else None, # Compressive stress limit
"SS": mat.Ss() if hasattr(mat, 'Ss') else None # Shear stress limit
})
materials.append(mat_data)
self.neural_field_data["materials"] = materials
def extract_boundary_conditions(self):
"""
Extract boundary conditions from BDF
"""
print("Extracting boundary conditions...")
bcs = {
"spc": [], # Single point constraints
"mpc": [], # Multi-point constraints
"suport": [] # Free body supports
}
# SPC (fixed DOFs)
for spc_id, spc_list in self.bdf.spcs.items():
for spc in spc_list:
bcs["spc"].append({
"id": spc_id,
"node": spc.node_ids[0] if hasattr(spc, 'node_ids') else spc.node,
"dofs": spc.components, # Which DOFs are constrained (123456)
"enforced_motion": spc.enforced
})
# MPC equations
for mpc_id, mpc_list in self.bdf.mpcs.items():
for mpc in mpc_list:
bcs["mpc"].append({
"id": mpc_id,
"nodes": mpc.node_ids,
"coefficients": mpc.coefficients,
"components": mpc.components
})
self.neural_field_data["boundary_conditions"] = bcs
def extract_loads(self):
"""
Extract loads from BDF
"""
print("Extracting loads...")
loads = {
"point_forces": [],
"pressure": [],
"gravity": [],
"thermal": []
}
# Point forces (FORCE, MOMENT)
for load_id, load_list in self.bdf.loads.items():
for load in load_list:
if load.type == 'FORCE':
loads["point_forces"].append({
"id": load_id,
"node": load.node,
"magnitude": load.mag,
"direction": [load.xyz[0], load.xyz[1], load.xyz[2]],
"coord_system": load.cid
})
elif load.type == 'MOMENT':
loads["point_forces"].append({
"id": load_id,
"node": load.node,
"moment": load.mag,
"direction": [load.xyz[0], load.xyz[1], load.xyz[2]],
"coord_system": load.cid
})
elif load.type in ['PLOAD', 'PLOAD2', 'PLOAD4']:
loads["pressure"].append({
"id": load_id,
"elements": load.element_ids,
"pressure": load.pressure,
"type": load.type
})
elif load.type == 'GRAV':
loads["gravity"].append({
"id": load_id,
"acceleration": load.scale,
"direction": [load.N[0], load.N[1], load.N[2]],
"coord_system": load.cid
})
# Temperature loads
for temp_id, temp_list in self.bdf.temps.items():
for temp in temp_list:
loads["thermal"].append({
"id": temp_id,
"node": temp.node,
"temperature": temp.temperature
})
self.neural_field_data["loads"] = loads
def extract_results(self):
"""
Extract results from OP2
"""
print("Extracting results...")
results = {}
# Get subcase ID (usually 1 for linear static)
subcase_id = 1
# Displacement
if hasattr(self.op2, 'displacements'):
disp = self.op2.displacements[subcase_id]
disp_data = disp.data[0, :, :] # [itime=0, all_nodes, 6_dofs]
results["displacement"] = {
"node_ids": disp.node_gridtype[:, 0].tolist(),
"data": disp_data.tolist(),
"shape": list(disp_data.shape),
"max_magnitude": float(np.max(np.linalg.norm(disp_data[:, :3], axis=1)))
}
# Stress - handle different element types
stress_results = {}
# Solid stress
if hasattr(self.op2, 'ctetra_stress'):
stress = self.op2.ctetra_stress[subcase_id]
stress_data = stress.data[0, :, :]
stress_results["solid_stress"] = {
"element_ids": stress.element_node[:, 0].tolist(),
"data": stress_data.tolist(),
"von_mises": stress_data[:, -1].tolist() if stress_data.shape[1] > 6 else None
}
# Shell stress
if hasattr(self.op2, 'cquad4_stress'):
stress = self.op2.cquad4_stress[subcase_id]
stress_data = stress.data[0, :, :]
stress_results["shell_stress"] = {
"element_ids": stress.element_node[:, 0].tolist(),
"data": stress_data.tolist()
}
results["stress"] = stress_results
# Strain
strain_results = {}
if hasattr(self.op2, 'ctetra_strain'):
strain = self.op2.ctetra_strain[subcase_id]
strain_data = strain.data[0, :, :]
strain_results["solid_strain"] = {
"element_ids": strain.element_node[:, 0].tolist(),
"data": strain_data.tolist()
}
results["strain"] = strain_results
# SPC Forces (reactions)
if hasattr(self.op2, 'spc_forces'):
spc = self.op2.spc_forces[subcase_id]
spc_data = spc.data[0, :, :]
results["reactions"] = {
"node_ids": spc.node_gridtype[:, 0].tolist(),
"forces": spc_data.tolist()
}
self.neural_field_data["results"] = results
def save_data(self):
"""
Save parsed data to JSON and HDF5
"""
print("Saving data...")
# Save JSON metadata
json_file = self.case_dir / "neural_field_data.json"
with open(json_file, 'w') as f:
# Convert numpy arrays to lists for JSON serialization
json.dump(self.neural_field_data, f, indent=2, default=str)
# Save HDF5 for large arrays
h5_file = self.case_dir / "neural_field_data.h5"
with h5py.File(h5_file, 'w') as f:
# Save mesh data
mesh_grp = f.create_group('mesh')
mesh_grp.create_dataset('node_coordinates',
data=np.array(self.neural_field_data["mesh"]["nodes"]["coordinates"]))
# Save results
if "results" in self.neural_field_data:
results_grp = f.create_group('results')
if "displacement" in self.neural_field_data["results"]:
results_grp.create_dataset('displacement',
data=np.array(self.neural_field_data["results"]["displacement"]["data"]))
print(f"Data saved to {json_file} and {h5_file}")
# ============================================================================
# USAGE SCRIPT
# ============================================================================
def main():
"""
Main function to run the parser
"""
import sys
if len(sys.argv) < 2:
print("Usage: python neural_field_parser.py <case_directory>")
sys.exit(1)
case_dir = sys.argv[1]
# Create parser
parser = NastranToNeuralFieldParser(case_dir)
# Parse all data
try:
data = parser.parse_all()
print("\nParsing successful!")
print(f"Nodes: {data['mesh']['statistics']['n_nodes']}")
print(f"Elements: {data['mesh']['statistics']['n_elements']}")
print(f"Materials: {len(data['materials'])}")
except Exception as e:
print(f"\nError during parsing: {e}")
import traceback
traceback.print_exc()
if __name__ == "__main__":
main()
Validation Script
python"""
validate_parsed_data.py
Validates the parsed neural field data
"""
import json
import h5py
import numpy as np
from pathlib import Path
class NeuralFieldDataValidator:
"""
Validates parsed data for completeness and consistency
"""
def __init__(self, case_directory):
self.case_dir = Path(case_directory)
self.json_file = self.case_dir / "neural_field_data.json"
self.h5_file = self.case_dir / "neural_field_data.h5"
def validate(self):
"""
Run all validation checks
"""
print("Starting validation...")
# Load data
with open(self.json_file, 'r') as f:
data = json.load(f)
# Check required fields
required_fields = [
"metadata", "mesh", "materials",
"boundary_conditions", "loads", "results"
]
for field in required_fields:
if field not in data:
print(f"❌ Missing required field: {field}")
return False
else:
print(f"✅ Found {field}")
# Validate mesh
n_nodes = data["mesh"]["statistics"]["n_nodes"]
n_elements = data["mesh"]["statistics"]["n_elements"]
print(f"\nMesh Statistics:")
print(f" Nodes: {n_nodes}")
print(f" Elements: {n_elements}")
# Check results consistency
if "displacement" in data["results"]:
disp_nodes = len(data["results"]["displacement"]["node_ids"])
if disp_nodes != n_nodes:
print(f"⚠️ Displacement nodes ({disp_nodes}) != mesh nodes ({n_nodes})")
# Check HDF5 file
with h5py.File(self.h5_file, 'r') as f:
print(f"\nHDF5 Contents:")
for key in f.keys():
print(f" {key}: {list(f[key].keys())}")
print("\n✅ Validation complete!")
return True
if __name__ == "__main__":
import sys
validator = NeuralFieldDataValidator(sys.argv[1])
validator.validate()
Step-by-Step Usage Instructions
1. Prepare Your Analysis
bash# In NX:
1. Create geometry
2. Generate mesh
3. Apply materials (MAT1 cards)
4. Apply constraints (SPC)
5. Apply loads (FORCE, PLOAD4)
6. Run SOL 101 (Linear Static)
7. Request output: DISPLACEMENT=ALL, STRESS=ALL, STRAIN=ALL
2. Organize Files
bashmkdir training_case_001
mkdir training_case_001/input
mkdir training_case_001/output
# Copy files
cp your_model.bdf training_case_001/input/model.bdf
cp your_model.op2 training_case_001/output/model.op2
cp your_model.f06 training_case_001/output/model.f06
3. Run Parser
bash# Install requirements
pip install pyNastran numpy h5py
# Run parser
python neural_field_parser.py training_case_001
# Validate
python validate_parsed_data.py training_case_001
4. Check Output
You'll get:
neural_field_data.json - Complete metadata and structure
neural_field_data.h5 - Large arrays (mesh, results)
Automation Script for Multiple Cases
python"""
batch_parser.py
Parse multiple cases automatically
"""
import os
from pathlib import Path
from neural_field_parser import NastranToNeuralFieldParser
def batch_parse(root_directory):
"""
Parse all cases in directory
"""
root = Path(root_directory)
cases = [d for d in root.iterdir() if d.is_dir()]
results = []
for case in cases:
print(f"\nProcessing {case.name}...")
try:
parser = NastranToNeuralFieldParser(case)
data = parser.parse_all()
results.append({
"case": case.name,
"status": "success",
"nodes": data["mesh"]["statistics"]["n_nodes"],
"elements": data["mesh"]["statistics"]["n_elements"]
})
except Exception as e:
results.append({
"case": case.name,
"status": "failed",
"error": str(e)
})
# Summary
print("\n" + "="*50)
print("BATCH PROCESSING COMPLETE")
print("="*50)
for r in results:
status = "✅" if r["status"] == "success" else "❌"
print(f"{status} {r['case']}: {r['status']}")
return results
if __name__ == "__main__":
batch_parse("./training_data")
What to Add Manually
Create a metadata.json in each case directory with design intent:
json{
"design_parameters": {
"thickness": 2.5,
"fillet_radius": 5.0,
"rib_height": 15.0
},
"optimization_context": {
"objectives": ["minimize_weight", "minimize_stress"],
"constraints": ["max_displacement < 2mm"],
"iteration": 42
},
"notes": "Baseline design with standard loading"
}
Troubleshooting
Common Issues:
"Can't find BDF nodes"
Make sure you're using .bdf or .dat, not .sim
Check that mesh was exported to solver deck
"OP2 has no results"
Ensure analysis completed successfully
Check that you requested output (DISP=ALL, STRESS=ALL)
"Memory error with large models"
Use HDF5 chunking for very large models
Process in batches
This parser gives you everything you need to start training neural networks on your FEA data. The format is future-proof and will work with your automated generation pipeline!

587
atomizer-field/PHASE2_README.md
View File

@@ -1,587 +0,0 @@
# AtomizerField Phase 2: Neural Field Learning
**Version 2.0.0**
Phase 2 implements Graph Neural Networks (GNNs) to learn complete FEA field results from mesh geometry, boundary conditions, and loads. This enables 1000x faster structural analysis for optimization.
## What's New in Phase 2
### The Revolutionary Approach
**Traditional FEA Surrogate Models:**
```
Parameters → Neural Network → Max Stress (scalar)
```
- Only learns maximum values
- Loses all spatial information
- Can't understand physics
- Limited to specific loading conditions
**AtomizerField Neural Field Learning:**
```
Mesh + BCs + Loads → Graph Neural Network → Complete Stress Field (45,000 values)
```
- Learns complete field distributions
- Understands how forces flow through structure
- Physics-informed constraints
- Generalizes to new loading conditions
### Key Components
1. **Graph Neural Network Architecture** - Learns on mesh topology
2. **Physics-Informed Loss Functions** - Enforces physical laws
3. **Efficient Data Loading** - Handles large FEA datasets
4. **Training Pipeline** - Multi-GPU support, checkpointing, early stopping
5. **Fast Inference** - Millisecond predictions vs hours of FEA
## Quick Start
### 1. Installation
```bash
# Install Phase 2 dependencies (PyTorch, PyTorch Geometric)
pip install -r requirements.txt
```
### 2. Prepare Training Data
First, you need parsed FEA data from Phase 1:
```bash
# Parse your NX Nastran results (from Phase 1)
python neural_field_parser.py training_case_001
python neural_field_parser.py training_case_002
# ... repeat for all training cases
# Organize into train/val splits
mkdir training_data
mkdir validation_data
# Move 80% of cases to training_data/
# Move 20% of cases to validation_data/
```
### 3. Train Model
```bash
# Basic training
python train.py \
--train_dir ./training_data \
--val_dir ./validation_data \
--epochs 100 \
--batch_size 4 \
--lr 0.001
# Advanced training with physics-informed loss
python train.py \
--train_dir ./training_data \
--val_dir ./validation_data \
--epochs 200 \
--batch_size 8 \
--lr 0.001 \
--loss_type physics \
--hidden_dim 256 \
--num_layers 8 \
--output_dir ./my_model
```
### 4. Run Inference
```bash
# Predict on single case
python predict.py \
--model runs/checkpoint_best.pt \
--input test_case_001 \
--compare
# Batch prediction on multiple cases
python predict.py \
--model runs/checkpoint_best.pt \
--input ./test_data \
--batch \
--output_dir ./predictions
```
## Architecture Deep Dive
### Graph Neural Network (GNN)
Our GNN architecture respects the physics of structural mechanics:
```
Input Graph:
- Nodes: FEA mesh nodes
* Position (x, y, z)
* Boundary conditions (6 DOF constraints)
* Applied loads (force vectors)
- Edges: Element connectivity
* Material properties (E, ν, ρ, G, α)
* Element type (solid, shell, beam)
Message Passing (6 layers):
Each layer propagates information through mesh:
1. Gather information from neighbors
2. Update node representations
3. Respect mesh topology (forces flow through connected elements)
Output:
- Displacement field: [num_nodes, 6] (3 translation + 3 rotation)
- Stress field: [num_nodes, 6] (σxx, σyy, σzz, τxy, τyz, τxz)
- Von Mises stress: [num_nodes, 1]
```
### Why This Works
**Key Insight**: FEA solves:
```
K u = f
```
Where K depends on mesh topology and materials.
Our GNN learns this relationship:
- **Mesh topology** → Graph edges
- **Material properties** → Edge features
- **Boundary conditions** → Node features
- **Loads** → Node features
- **Message passing** → Learns stiffness matrix behavior
Result: The network learns physics, not just patterns!
### Model Architecture Details
```python
AtomizerFieldModel:
├── Node Encoder (12 → 128 dim)
│ └── Coordinates (3) + BCs (6) + Loads (3)
│
├── Edge Encoder (5 → 64 dim)
│ └── Material properties (E, ν, ρ, G, α)
│
├── Message Passing Layers (6 layers)
│ ├── MeshGraphConv
│ ├── Layer Normalization
│ ├── Residual Connection
│ └── Dropout
│
├── Displacement Decoder (128 → 6)
│ └── Outputs: u_x, u_y, u_z, θ_x, θ_y, θ_z
│
└── Stress Predictor (6 → 6)
└── Outputs: σxx, σyy, σzz, τxy, τyz, τxz
Total Parameters: ~718,000
```
## Physics-Informed Loss Functions
Standard neural networks only minimize prediction error. We also enforce physics:
### 1. Data Loss
```
L_data = ||u_pred - u_FEA||² + ||σ_pred - σ_FEA||²
```
Ensures predictions match FEA ground truth.
### 2. Equilibrium Loss
```
L_eq = ||∇·σ + f||²
```
Forces must balance at every point.
### 3. Constitutive Loss
```
L_const = ||σ - C:ε||²
```
Stress must follow material law (σ = C:ε).
### 4. Boundary Condition Loss
```
L_bc = ||u||² at fixed nodes
```
Displacement must be zero at constraints.
### Total Loss
```
L_total = λ_data·L_data + λ_eq·L_eq + λ_const·L_const + λ_bc·L_bc
```
**Benefits:**
- Faster convergence
- Better generalization
- Physically plausible predictions
- Works with less training data
## Training Guide
### Dataset Requirements
**Minimum Dataset Size:**
- Small models (< 10k elements): 50-100 cases
- Medium models (10k-100k elements): 100-500 cases
- Large models (> 100k elements): 500-1000 cases
**Data Diversity:**
Vary these parameters across training cases:
- Geometry (thicknesses, radii, dimensions)
- Loading conditions (magnitude, direction, location)
- Boundary conditions (support locations, constrained DOFs)
- Materials (within reason - same element types)
### Training Best Practices
**1. Start Simple:**
```bash
# First, train with MSE loss only
python train.py --loss_type mse --epochs 50
```
**2. Add Physics:**
```bash
# Then add physics-informed constraints
python train.py --loss_type physics --epochs 100
```
**3. Tune Hyperparameters:**
```bash
# Increase model capacity for complex geometries
python train.py \
--hidden_dim 256 \
--num_layers 8 \
--dropout 0.15
```
**4. Monitor Training:**
```bash
# View training progress in TensorBoard
tensorboard --logdir runs/tensorboard
```
### Typical Training Time
On a single GPU (e.g., NVIDIA RTX 3080):
- Small dataset (100 cases, 10k elements each): 2-4 hours
- Medium dataset (500 cases, 50k elements each): 8-12 hours
- Large dataset (1000 cases, 100k elements each): 24-48 hours
**Speedup Tips:**
- Use multiple GPUs: `CUDA_VISIBLE_DEVICES=0,1 python train.py`
- Increase batch size if memory allows
- Use mixed precision training (future feature)
- Cache data in RAM: set `cache_in_memory=True` in data_loader.py
## Inference Performance
### Speed Comparison
| Analysis Method | Time | Speedup |
|----------------|------|---------|
| Traditional FEA (NX Nastran) | 2-3 hours | 1x |
| **AtomizerField GNN** | **5-50 ms** | **10,000x** |
**Real Performance (100k element model):**
- FEA Setup + Solve: ~2 hours
- Neural Network Inference: ~15 milliseconds
- **Speedup: 480,000x**
This enables:
- Interactive design exploration
- Real-time optimization (evaluate millions of designs)
- Instant "what-if" analysis
### Accuracy
Typical prediction errors (on validation set):
- **Displacement**: 2-5% relative error
- **Stress**: 5-10% relative error
- **Max values**: 1-3% relative error
**When It Works Best:**
- Interpolation (designs within training range)
- Similar loading conditions
- Same support configurations
- Parametric variations
**When to Use with Caution:**
- Extreme extrapolation (far outside training data)
- Completely new loading scenarios
- Different element types than training
- Nonlinear materials (future work)
## Usage Examples
### Example 1: Rapid Design Optimization
```python
from predict import FieldPredictor
# Load trained model
predictor = FieldPredictor('checkpoint_best.pt')
# Test 1000 design variants
results = []
for design_params in design_space:
# Generate FEA input (don't solve!)
create_nastran_model(design_params)
parse_to_neural_format(design_params)
# Predict in milliseconds
pred = predictor.predict(f'design_{i}')
results.append({
'params': design_params,
'max_stress': pred['max_stress'],
'max_displacement': pred['max_displacement']
})
# Find optimal design
best = min(results, key=lambda r: r['max_stress'])
print(f"Best design: {best['params']}")
print(f"Stress: {best['max_stress']:.2f} MPa")
```
**Result:** Evaluate 1000 designs in ~30 seconds instead of 3000 hours!
### Example 2: Interactive Design Tool
```python
# Real-time design feedback
while user_editing:
# User modifies geometry
updated_geometry = get_user_input()
# Generate mesh (fast, no solve)
mesh = generate_mesh(updated_geometry)
parse_mesh(mesh)
# Instant prediction
prediction = predictor.predict('current_design')
# Show results immediately
display_stress_field(prediction['von_mises'])
display_displacement(prediction['displacement'])
# Immediate feedback: "Max stress: 450 MPa (SAFE)"
```
**Result:** Engineer sees results instantly, not hours later!
### Example 3: Optimization with Physics Understanding
```python
# Traditional: Only knows max_stress = 450 MPa
# AtomizerField: Knows WHERE stress concentrations are!
prediction = predictor.predict('current_design')
stress_field = prediction['von_mises']
# Find stress hotspots
hotspots = find_nodes_above_threshold(stress_field, threshold=400)
# Intelligent design suggestions
for hotspot in hotspots:
location = mesh.nodes[hotspot].position
suggest_reinforcement(location) # Add material where needed
suggest_fillet(location) # Smooth sharp corners
```
**Result:** Optimization guided by physics, not blind search!
## File Structure
```
Atomizer-Field/
├── neural_models/
│ ├── __init__.py
│ ├── field_predictor.py # GNN architecture
│ ├── physics_losses.py # Loss functions
│ └── data_loader.py # Data pipeline
│
├── train.py # Training script
├── predict.py # Inference script
├── requirements.txt # Dependencies
│
├── runs/ # Training outputs
│ ├── checkpoint_best.pt # Best model
│ ├── checkpoint_latest.pt # Latest checkpoint
│ ├── tensorboard/ # Training logs
│ └── config.json # Model configuration
│
└── training_data/ # Parsed FEA cases
├── case_001/
│ ├── neural_field_data.json
│ └── neural_field_data.h5
├── case_002/
└── ...
```
## Troubleshooting
### Training Issues
**Problem: Loss not decreasing**
```bash
# Solutions:
# 1. Lower learning rate
python train.py --lr 0.0001
# 2. Check data normalization
# Ensure normalize=True in data_loader.py
# 3. Start with simpler loss
python train.py --loss_type mse
```
**Problem: Out of memory**
```bash
# Solutions:
# 1. Reduce batch size
python train.py --batch_size 2
# 2. Reduce model size
python train.py --hidden_dim 64 --num_layers 4
# 3. Use gradient accumulation (future feature)
```
**Problem: Overfitting**
```bash
# Solutions:
# 1. Increase dropout
python train.py --dropout 0.2
# 2. Get more training data
# 3. Use data augmentation (rotate/scale meshes)
```
### Inference Issues
**Problem: Poor predictions**
- Check if test case is within training distribution
- Verify data normalization matches training
- Ensure model finished training (check validation loss)
**Problem: Slow inference**
- Use GPU: `--device cuda`
- Batch multiple predictions together
- Use smaller model for production
## Advanced Topics
### Transfer Learning
Train on one component type, fine-tune on another:
```bash
# 1. Train base model on brackets
python train.py --train_dir brackets/ --epochs 100
# 2. Fine-tune on beams (similar physics, different geometry)
python train.py \
--train_dir beams/ \
--resume runs/checkpoint_best.pt \
--epochs 50 \
--lr 0.0001 # Lower LR for fine-tuning
```
### Multi-Fidelity Learning
Combine coarse and fine meshes:
```python
# Train on mix of mesh resolutions
train_cases = [
*coarse_mesh_cases, # Fast to solve, less accurate
*fine_mesh_cases # Slow to solve, very accurate
]
# Model learns to predict fine-mesh accuracy at coarse-mesh speed!
```
### Physics-Based Data Augmentation
```python
# Augment training data with physical transformations
def augment_case(mesh, displacement, stress):
# Rotate entire structure
mesh_rotated = rotate(mesh, angle=random.uniform(0, 360))
displacement_rotated = rotate_vector_field(displacement, angle)
stress_rotated = rotate_tensor_field(stress, angle)
# Scale loads (linear scaling)
scale = random.uniform(0.5, 2.0)
displacement_scaled = displacement * scale
stress_scaled = stress * scale
return augmented_cases
```
## Future Enhancements (Phase 3)
- [ ] Nonlinear analysis support (plasticity, large deformation)
- [ ] Contact and friction
- [ ] Composite materials
- [ ] Modal analysis (natural frequencies)
- [ ] Thermal coupling
- [ ] Topology optimization integration
- [ ] Atomizer dashboard integration
- [ ] Cloud deployment for team access
## Performance Benchmarks
### Model Accuracy (Validation Set)
| Metric | Error | Target |
|--------|-------|--------|
| Displacement MAE | 0.003 mm | < 0.01 mm |
| Displacement Relative Error | 3.2% | < 5% |
| Stress MAE | 12.5 MPa | < 20 MPa |
| Max Stress Error | 2.1% | < 5% |
| Max Displacement Error | 1.8% | < 3% |
### Computational Performance
| Dataset | FEA Time | NN Time | Speedup |
|---------|----------|---------|---------|
| 10k elements | 15 min | 5 ms | 180,000x |
| 50k elements | 2 hours | 15 ms | 480,000x |
| 100k elements | 8 hours | 35 ms | 823,000x |
**Hardware:** Single NVIDIA RTX 3080, Intel i9-12900K
## Citation
If you use AtomizerField in research, please cite:
```
@software{atomizerfield2024,
title={AtomizerField: Neural Field Learning for Structural Optimization},
author={Your Name},
year={2024},
url={https://github.com/yourusername/atomizer-field}
}
```
## Support
For issues or questions about Phase 2:
1. Check this README and PHASE2_README.md
2. Review training logs in TensorBoard
3. Examine model predictions vs ground truth
4. Check GPU memory usage and batch size
5. Verify data normalization
## What's Next?
- **Phase 3**: Integration with main Atomizer platform
- **Phase 4**: Production deployment and dashboard
- **Phase 5**: Multi-user cloud platform
---
**AtomizerField Phase 2**: Revolutionary neural field learning for structural optimization.
*1000x faster than FEA. Physics-informed. Production-ready.*

500
atomizer-field/QUICK_REFERENCE.md
View File

@@ -1,500 +0,0 @@
# AtomizerField Quick Reference Guide
**Version 1.0** | Complete Implementation | Ready for Training
---
## 🎯 What is AtomizerField?
Neural field learning system that replaces FEA with 1000× faster graph neural networks.
**Key Innovation:** Learn complete stress/displacement FIELDS (45,000+ values), not just max values.
---
## 📁 Project Structure
```
Atomizer-Field/
├── Neural Network Core
│ ├── neural_models/
│ │ ├── field_predictor.py # GNN architecture (718K params)
│ │ ├── physics_losses.py # 4 loss functions
│ │ ├── data_loader.py # PyTorch Geometric dataset
│ │ └── uncertainty.py # Ensemble + online learning
│ ├── train.py # Training pipeline
│ ├── predict.py # Inference engine
│ └── optimization_interface.py # Atomizer integration
│
├── Data Pipeline
│ ├── neural_field_parser.py # BDF/OP2 → neural format
│ ├── validate_parsed_data.py # Data quality checks
│ └── batch_parser.py # Multi-case processing
│
├── Testing (18 tests)
│ ├── test_suite.py # Master orchestrator
│ ├── test_simple_beam.py # Simple Beam validation
│ └── tests/
│ ├── test_synthetic.py # 5 smoke tests
│ ├── test_physics.py # 4 physics tests
│ ├── test_learning.py # 4 learning tests
│ ├── test_predictions.py # 5 integration tests
│ └── analytical_cases.py # Analytical solutions
│
└── Documentation (10 guides)
├── README.md # Project overview
├── IMPLEMENTATION_STATUS.md # Complete status
├── TESTING_COMPLETE.md # Testing guide
└── ... (7 more guides)
```
---
## 🚀 Quick Start Commands
### 1. Test the System
```bash
# Smoke tests (30 seconds) - Once environment fixed
python test_suite.py --quick
# Test with Simple Beam
python test_simple_beam.py
# Full test suite (1 hour)
python test_suite.py --full
```
### 2. Parse FEA Data
```bash
# Single case
python neural_field_parser.py path/to/case_directory
# Validate parsed data
python validate_parsed_data.py path/to/case_directory
# Batch process multiple cases
python batch_parser.py --input Models/ --output parsed_data/
```
### 3. Train Model
```bash
# Basic training
python train.py --data_dirs case1 case2 case3 --epochs 100
# With all options
python train.py \
--data_dirs parsed_data/* \
--epochs 200 \
--batch_size 32 \
--lr 0.001 \
--loss physics \
--checkpoint_dir checkpoints/
```
### 4. Make Predictions
```bash
# Single prediction
python predict.py --model checkpoints/best_model.pt --data test_case/
# Batch prediction
python predict.py --model best_model.pt --data test_cases/*.h5 --batch_size 64
```
### 5. Optimize with Atomizer
```python
from optimization_interface import NeuralFieldOptimizer
# Initialize
optimizer = NeuralFieldOptimizer('checkpoints/best_model.pt')
# Evaluate design
results = optimizer.evaluate(design_graph)
print(f"Max stress: {results['max_stress']} MPa")
print(f"Max displacement: {results['max_displacement']} mm")
# Get gradients for optimization
sensitivities = optimizer.get_sensitivities(design_graph)
```
---
## 📊 Key Metrics
### Performance
- **Training time:** 2-6 hours (50-500 cases, 100-200 epochs)
- **Inference time:** 5-50ms (vs 30-300s FEA)
- **Speedup:** 1000× faster than FEA
- **Memory:** ~2GB GPU for training, ~500MB for inference
### Accuracy (After Training)
- **Target:** < 10% prediction error vs FEA
- **Physics tests:** < 5% error on analytical solutions
- **Learning tests:** < 5% interpolation error
### Model Size
- **Parameters:** 718,221
- **Layers:** 6 message passing layers
- **Input:** 12D node features, 5D edge features
- **Output:** 6 DOF displacement + 6 stress components per node
---
## 🧪 Testing Overview
### Quick Smoke Test (30s)
```bash
python test_suite.py --quick
```
**5 tests:** Model creation, forward pass, losses, batch, gradients
### Physics Validation (15 min)
```bash
python test_suite.py --physics
```
**9 tests:** Smoke + Cantilever, equilibrium, energy, constitutive
### Learning Tests (30 min)
```bash
python test_suite.py --learning
```
**13 tests:** Smoke + Physics + Memorization, interpolation, extrapolation, patterns
### Full Suite (1 hour)
```bash
python test_suite.py --full
```
**18 tests:** Complete validation from zero to production
---
## 📈 Typical Workflow
### Phase 1: Data Preparation
```bash
# 1. Parse FEA cases
python batch_parser.py --input Models/ --output training_data/
# 2. Validate data
for dir in training_data/*; do
python validate_parsed_data.py $dir
done
# Expected: 50-500 parsed cases
```
### Phase 2: Training
```bash
# 3. Train model
python train.py \
--data_dirs training_data/* \
--epochs 100 \
--batch_size 16 \
--loss physics \
--checkpoint_dir checkpoints/
# Monitor with TensorBoard
tensorboard --logdir runs/
# Expected: Training loss < 0.01 after 100 epochs
```
### Phase 3: Validation
```bash
# 4. Run all tests
python test_suite.py --full
# 5. Test on new data
python predict.py --model checkpoints/best_model.pt --data test_case/
# Expected: All tests pass, < 10% error
```
### Phase 4: Deployment
```python
# 6. Integrate with Atomizer
from optimization_interface import NeuralFieldOptimizer
optimizer = NeuralFieldOptimizer('checkpoints/best_model.pt')
# Use in optimization loop
for iteration in range(100):
results = optimizer.evaluate(current_design)
sensitivities = optimizer.get_sensitivities(current_design)
# Update design based on gradients
current_design = update_design(current_design, sensitivities)
```
---
## 🔧 Configuration
### Training Config (atomizer_field_config.yaml)
```yaml
model:
hidden_dim: 128
num_layers: 6
dropout: 0.1
training:
batch_size: 16
learning_rate: 0.001
epochs: 100
early_stopping_patience: 10
loss:
type: physics
lambda_data: 1.0
lambda_equilibrium: 0.1
lambda_constitutive: 0.1
lambda_boundary: 0.5
uncertainty:
n_ensemble: 5
threshold: 0.1 # Trigger FEA if uncertainty > 10%
online_learning:
enabled: true
update_frequency: 10 # Update every 10 FEA runs
batch_size: 32
```
---
## 🎓 Feature Reference
### 1. Data Parser
**File:** `neural_field_parser.py`
```python
from neural_field_parser import NastranToNeuralFieldParser
# Parse case
parser = NastranToNeuralFieldParser('case_directory')
data = parser.parse_all()
# Access results
print(f"Nodes: {data['mesh']['statistics']['n_nodes']}")
print(f"Max displacement: {data['results']['displacement']['max_translation']} mm")
```
### 2. Neural Model
**File:** `neural_models/field_predictor.py`
```python
from neural_models.field_predictor import create_model
# Create model
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 128,
'num_layers': 6
}
model = create_model(config)
# Predict
predictions = model(graph_data, return_stress=True)
# predictions['displacement']: (N, 6) - 6 DOF per node
# predictions['stress']: (N, 6) - stress tensor
# predictions['von_mises']: (N,) - von Mises stress
```
### 3. Physics Losses
**File:** `neural_models/physics_losses.py`
```python
from neural_models.physics_losses import create_loss_function
# Create loss
loss_fn = create_loss_function('physics')
# Compute loss
losses = loss_fn(predictions, targets, data)
# losses['total_loss']: Combined loss
# losses['displacement_loss']: Data loss
# losses['equilibrium_loss']: ∇·σ + f = 0
# losses['constitutive_loss']: σ = C:ε
# losses['boundary_loss']: BC compliance
```
### 4. Optimization Interface
**File:** `optimization_interface.py`
```python
from optimization_interface import NeuralFieldOptimizer
# Initialize
optimizer = NeuralFieldOptimizer('model.pt')
# Fast evaluation (15ms)
results = optimizer.evaluate(graph_data)
# Analytical gradients (1M× faster than FD)
grads = optimizer.get_sensitivities(graph_data)
```
### 5. Uncertainty Quantification
**File:** `neural_models/uncertainty.py`
```python
from neural_models.uncertainty import UncertainFieldPredictor
# Create ensemble
model = UncertainFieldPredictor(base_config, n_ensemble=5)
# Predict with uncertainty
predictions = model.predict_with_uncertainty(graph_data)
# predictions['mean']: Mean prediction
# predictions['std']: Standard deviation
# predictions['confidence']: 95% confidence interval
# Check if FEA needed
if model.needs_fea_validation(predictions, threshold=0.1):
# Run FEA for this case
fea_result = run_fea(design)
# Update model online
model.update_online(graph_data, fea_result)
```
---
## 🐛 Troubleshooting
### NumPy Environment Issue
**Problem:** Segmentation fault when importing NumPy
```
CRASHES ARE TO BE EXPECTED - PLEASE REPORT THEM TO NUMPY DEVELOPERS
Segmentation fault
```
**Solutions:**
1. Use conda: `conda install numpy`
2. Use WSL: Install Windows Subsystem for Linux
3. Use Linux: Native Linux environment
4. Reinstall: `pip uninstall numpy && pip install numpy`
### Import Errors
**Problem:** Cannot find modules
```python
ModuleNotFoundError: No module named 'torch_geometric'
```
**Solution:**
```bash
# Install all dependencies
pip install -r requirements.txt
# Or individual packages
pip install torch torch-geometric pyg-lib
pip install pyNastran h5py pyyaml tensorboard
```
### GPU Memory Issues
**Problem:** CUDA out of memory during training
**Solutions:**
1. Reduce batch size: `--batch_size 8`
2. Reduce model size: `hidden_dim: 64`
3. Use CPU: `--device cpu`
4. Enable gradient checkpointing
### Poor Predictions
**Problem:** High prediction error (> 20%)
**Solutions:**
1. Train longer: `--epochs 200`
2. More data: Generate 200-500 training cases
3. Use physics loss: `--loss physics`
4. Check data quality: `python validate_parsed_data.py`
5. Normalize data: `normalize=True` in dataset
---
## 📚 Documentation Index
1. **README.md** - Project overview and quick start
2. **IMPLEMENTATION_STATUS.md** - Complete status report
3. **TESTING_COMPLETE.md** - Comprehensive testing guide
4. **PHASE2_README.md** - Neural network documentation
5. **GETTING_STARTED.md** - Step-by-step tutorial
6. **SYSTEM_ARCHITECTURE.md** - Technical architecture
7. **ENHANCEMENTS_GUIDE.md** - Advanced features
8. **FINAL_IMPLEMENTATION_REPORT.md** - Implementation details
9. **TESTING_FRAMEWORK_SUMMARY.md** - Testing overview
10. **QUICK_REFERENCE.md** - This guide
---
## ⚡ Pro Tips
### Training
- Start with 50 cases to verify pipeline
- Use physics loss for better generalization
- Monitor TensorBoard for convergence
- Save checkpoints every 10 epochs
- Early stopping prevents overfitting
### Data
- Quality > Quantity: 50 good cases better than 200 poor ones
- Diverse designs: Vary geometry, loads, materials
- Validate data: Check for NaN, physics violations
- Normalize features: Improves training stability
### Performance
- GPU recommended: 10× faster training
- Batch size = GPU memory / model size
- Use DataLoader workers: `num_workers=4`
- Cache in memory: `cache_in_memory=True`
### Uncertainty
- Use ensemble (5 models) for confidence
- Trigger FEA when uncertainty > 10%
- Update online: Improves during optimization
- Track confidence: Builds trust in predictions
---
## 🎯 Success Checklist
### Pre-Training
- [x] All code implemented
- [x] Tests written
- [x] Documentation complete
- [ ] Environment working (NumPy issue)
### Training
- [ ] 50-500 training cases generated
- [ ] Data parsed and validated
- [ ] Model trains without errors
- [ ] Loss converges < 0.01
### Validation
- [ ] All tests pass
- [ ] Physics compliance < 5% error
- [ ] Prediction error < 10%
- [ ] Inference < 50ms
### Production
- [ ] Integrated with Atomizer
- [ ] 1000× speedup demonstrated
- [ ] Uncertainty quantification working
- [ ] Online learning enabled
---
## 📞 Support
**Current Status:** Implementation complete, ready for training
**Next Steps:**
1. Fix NumPy environment
2. Generate training data
3. Train and validate
4. Deploy to production
**All code is ready to use!** 🚀
---
*AtomizerField Quick Reference v1.0*
*~7,000 lines | 18 tests | 10 docs | Production Ready*

548
atomizer-field/README.md
View File

@@ -1,548 +0,0 @@
# AtomizerField Neural Field Data Parser
**Version 1.0.0**
A production-ready Python parser that converts NX Nastran FEA results into standardized neural field training data for the AtomizerField optimization platform.
## What This Does
Instead of extracting just scalar values (like maximum stress) from FEA results, this parser captures **complete field data** - stress, displacement, and strain at every node and element. This enables neural networks to learn the physics of how structures respond to loads, enabling 1000x faster optimization with true physics understanding.
## Features
- ✅ **Complete Field Extraction**: Captures displacement, stress, strain at ALL points
- ✅ **Future-Proof Format**: Versioned data structure (v1.0) designed for years of neural network training
- ✅ **Efficient Storage**: Uses HDF5 for large arrays, JSON for metadata
- ✅ **Robust Parsing**: Handles mixed element types (solid, shell, beam, rigid)
- ✅ **Data Validation**: Built-in physics and quality checks
- ✅ **Batch Processing**: Process hundreds of cases automatically
- ✅ **Production Ready**: Error handling, logging, provenance tracking
## Quick Start
### 1. Installation
```bash
# Install dependencies
pip install -r requirements.txt
```
### 2. Prepare Your NX Nastran Analysis
In NX:
1. Create geometry and generate mesh
2. Apply materials (MAT1 cards)
3. Define boundary conditions (SPC)
4. Apply loads (FORCE, PLOAD4, GRAV)
5. Run **SOL 101** (Linear Static)
6. Request output: `DISPLACEMENT=ALL`, `STRESS=ALL`, `STRAIN=ALL`
### 3. Organize Files
```bash
mkdir training_case_001
mkdir training_case_001/input
mkdir training_case_001/output
# Copy files
cp your_model.bdf training_case_001/input/model.bdf
cp your_model.op2 training_case_001/output/model.op2
```
### 4. Run Parser
```bash
# Parse single case
python neural_field_parser.py training_case_001
# Validate results
python validate_parsed_data.py training_case_001
```
### 5. Check Output
You'll get:
- **neural_field_data.json** - Complete metadata and structure
- **neural_field_data.h5** - Large arrays (mesh, field results)
## Usage Guide
### Single Case Parsing
```bash
python neural_field_parser.py <case_directory>
```
**Expected directory structure:**
```
training_case_001/
├── input/
│ ├── model.bdf # Nastran input deck
│ └── model.sim # (optional) NX simulation file
├── output/
│ ├── model.op2 # Binary results (REQUIRED)
│ └── model.f06 # (optional) ASCII results
└── metadata.json # (optional) Your design annotations
```
**Output:**
```
training_case_001/
├── neural_field_data.json # Metadata, structure, small arrays
└── neural_field_data.h5 # Large arrays (coordinates, fields)
```
### Batch Processing
Process multiple cases at once:
```bash
python batch_parser.py ./training_data
```
**Expected structure:**
```
training_data/
├── case_001/
│ ├── input/model.bdf
│ └── output/model.op2
├── case_002/
│ ├── input/model.bdf
│ └── output/model.op2
└── case_003/
├── input/model.bdf
└── output/model.op2
```
**Options:**
```bash
# Skip validation (faster)
python batch_parser.py ./training_data --no-validate
# Stop on first error
python batch_parser.py ./training_data --stop-on-error
```
**Output:**
- Parses all cases
- Validates each one
- Generates `batch_processing_summary.json` with results
### Data Validation
```bash
python validate_parsed_data.py training_case_001
```
Checks:
- ✓ File existence and format
- ✓ Data completeness (all required fields)
- ✓ Physics consistency (equilibrium, units)
- ✓ Data quality (no NaN/inf, reasonable values)
- ✓ Mesh integrity
- ✓ Material property validity
## Data Structure v1.0
The parser produces a standardized data structure designed to be future-proof:
```json
{
"metadata": {
"version": "1.0.0",
"created_at": "timestamp",
"analysis_type": "SOL_101",
"units": {...}
},
"mesh": {
"statistics": {
"n_nodes": 15432,
"n_elements": 8765
},
"nodes": {
"ids": [...],
"coordinates": "stored in HDF5"
},
"elements": {
"solid": [...],
"shell": [...],
"beam": [...]
}
},
"materials": [...],
"boundary_conditions": {
"spc": [...],
"mpc": [...]
},
"loads": {
"point_forces": [...],
"pressure": [...],
"gravity": [...],
"thermal": [...]
},
"results": {
"displacement": "stored in HDF5",
"stress": "stored in HDF5",
"strain": "stored in HDF5",
"reactions": "stored in HDF5"
}
}
```
### HDF5 Structure
Large numerical arrays are stored in HDF5 for efficiency:
```
neural_field_data.h5
├── mesh/
│ ├── node_coordinates [n_nodes, 3]
│ └── node_ids [n_nodes]
└── results/
├── displacement [n_nodes, 6]
├── displacement_node_ids
├── stress/
│ ├── ctetra_stress/
│ │ ├── data [n_elem, n_components]
│ │ └── element_ids
│ └── cquad4_stress/...
├── strain/...
└── reactions/...
```
## Adding Design Metadata
Create a `metadata.json` in each case directory to track design parameters:
```json
{
"design_parameters": {
"thickness": 2.5,
"fillet_radius": 5.0,
"rib_height": 15.0
},
"optimization_context": {
"objectives": ["minimize_weight", "minimize_stress"],
"constraints": ["max_displacement < 2mm"],
"iteration": 42
},
"notes": "Baseline design with standard loading"
}
```
See [metadata_template.json](metadata_template.json) for a complete template.
## Preparing NX Nastran Analyses
### Required Output Requests
Add these to your Nastran input deck or NX solution setup:
```nastran
DISPLACEMENT = ALL
STRESS = ALL
STRAIN = ALL
SPCFORCES = ALL
```
### Recommended Settings
- **Element Types**: CTETRA10, CHEXA20, CQUAD4
- **Analysis**: SOL 101 (Linear Static) initially
- **Units**: Consistent (recommend SI: mm, N, MPa, kg)
- **Output Format**: OP2 (binary) for efficiency
### Common Issues
**"OP2 has no results"**
- Ensure analysis completed successfully (check .log file)
- Verify output requests (DISPLACEMENT=ALL, STRESS=ALL)
- Check that OP2 file is not empty (should be > 1 KB)
**"Can't find BDF nodes"**
- Use .bdf or .dat file, not .sim
- Ensure mesh was exported to solver deck
- Check that BDF contains GRID cards
**"Memory error with large models"**
- Parser uses HDF5 chunking and compression
- For models > 100k elements, ensure you have sufficient RAM
- Consider splitting into subcases
## Loading Parsed Data
### In Python
```python
import json
import h5py
import numpy as np
# Load metadata
with open("neural_field_data.json", 'r') as f:
metadata = json.load(f)
# Load field data
with h5py.File("neural_field_data.h5", 'r') as f:
# Get node coordinates
coords = f['mesh/node_coordinates'][:]
# Get displacement field
displacement = f['results/displacement'][:]
# Get stress field
stress = f['results/stress/ctetra_stress/data'][:]
stress_elem_ids = f['results/stress/ctetra_stress/element_ids'][:]
```
### In PyTorch (for neural network training)
```python
import torch
from torch.utils.data import Dataset
class NeuralFieldDataset(Dataset):
def __init__(self, case_directories):
self.cases = []
for case_dir in case_directories:
h5_file = f"{case_dir}/neural_field_data.h5"
with h5py.File(h5_file, 'r') as f:
# Load inputs (mesh, BCs, loads)
coords = torch.from_numpy(f['mesh/node_coordinates'][:])
# Load outputs (displacement, stress fields)
displacement = torch.from_numpy(f['results/displacement'][:])
self.cases.append({
'coords': coords,
'displacement': displacement
})
def __len__(self):
return len(self.cases)
def __getitem__(self, idx):
return self.cases[idx]
```
## Architecture & Design
### Why This Format?
1. **Complete Fields, Not Scalars**: Neural networks need to learn how stress/displacement varies across the entire structure, not just maximum values.
2. **Separation of Concerns**: JSON for structure/metadata (human-readable), HDF5 for numerical data (efficient).
3. **Future-Proof**: Versioned format allows adding new fields without breaking existing data.
4. **Physics Preservation**: Maintains all physics relationships (mesh topology, BCs, loads → results).
### Integration with Atomizer
This parser is Phase 1 of AtomizerField. Future integration:
- Phase 2: Neural network architecture (Graph Neural Networks)
- Phase 3: Training pipeline with physics-informed loss functions
- Phase 4: Integration with main Atomizer dashboard
- Phase 5: Production deployment for real-time optimization
## Troubleshooting
### Parser Errors
| Error | Solution |
|-------|----------|
| `FileNotFoundError: No model.bdf found` | Ensure BDF/DAT file exists in `input/` directory |
| `FileNotFoundError: No model.op2 found` | Ensure OP2 file exists in `output/` directory |
| `pyNastran read error` | Check BDF syntax, try opening in text editor |
| `OP2 subcase not found` | Ensure analysis ran successfully, check .f06 file |
### Validation Warnings
| Warning | Meaning | Action |
|---------|---------|--------|
| `No SPCs defined` | Model may be unconstrained | Check boundary conditions |
| `No loads defined` | Model has no loading | Add forces, pressures, or gravity |
| `Zero displacement` | Model not deforming | Check loads and constraints |
| `Very large displacement` | Possible rigid body motion | Add constraints or check units |
### Data Quality Issues
**NaN or Inf values:**
- Usually indicates analysis convergence failure
- Check .f06 file for error messages
- Verify model is properly constrained
**Mismatch in node counts:**
- Some nodes may not have results (e.g., rigid elements)
- Check element connectivity
- Validate mesh quality in NX
## Example Workflow
Here's a complete example workflow from FEA to neural network training data:
### 1. Create Parametric Study in NX
```bash
# Generate 10 design variants with different thicknesses
# Run each analysis with SOL 101
# Export BDF and OP2 files for each
```
### 2. Organize Files
```bash
mkdir parametric_study
for i in {1..10}; do
mkdir -p parametric_study/thickness_${i}/input
mkdir -p parametric_study/thickness_${i}/output
# Copy BDF and OP2 files
done
```
### 3. Batch Parse
```bash
python batch_parser.py parametric_study
```
### 4. Review Results
```bash
# Check summary
cat parametric_study/batch_processing_summary.json
# Validate a specific case
python validate_parsed_data.py parametric_study/thickness_5
```
### 5. Load into Neural Network
```python
from torch.utils.data import DataLoader
dataset = NeuralFieldDataset([
f"parametric_study/thickness_{i}" for i in range(1, 11)
])
dataloader = DataLoader(dataset, batch_size=4, shuffle=True)
# Ready for training!
```
## Performance
Typical parsing times (on standard laptop):
- Small model (1k elements): ~5 seconds
- Medium model (10k elements): ~15 seconds
- Large model (100k elements): ~60 seconds
- Very large (1M elements): ~10 minutes
File sizes (compressed HDF5):
- Mesh (100k nodes): ~10 MB
- Displacement field (100k nodes × 6 DOF): ~5 MB
- Stress field (100k elements × 10 components): ~8 MB
## Requirements
- Python 3.8+
- pyNastran 1.4+
- NumPy 1.20+
- h5py 3.0+
- NX Nastran (any version that outputs .bdf and .op2)
## Files in This Repository
| File | Purpose |
|------|---------|
| `neural_field_parser.py` | Main parser - BDF/OP2 to neural field format |
| `validate_parsed_data.py` | Data validation and quality checks |
| `batch_parser.py` | Batch processing for multiple cases |
| `metadata_template.json` | Template for design parameter tracking |
| `requirements.txt` | Python dependencies |
| `README.md` | This file |
| `Context.md` | Project context and vision |
| `Instructions.md` | Original implementation instructions |
## Development
### Testing with Example Models
There are example models in the `Models/` folder. To test the parser:
```bash
# Set up test case
mkdir test_case_001
mkdir test_case_001/input
mkdir test_case_001/output
# Copy example files
cp Models/example_model.bdf test_case_001/input/model.bdf
cp Models/example_model.op2 test_case_001/output/model.op2
# Run parser
python neural_field_parser.py test_case_001
# Validate
python validate_parsed_data.py test_case_001
```
### Extending the Parser
To add new result types (e.g., modal analysis, thermal):
1. Update `extract_results()` in `neural_field_parser.py`
2. Add corresponding validation in `validate_parsed_data.py`
3. Update data structure version if needed
4. Document changes in this README
### Contributing
This is part of the AtomizerField project. When making changes:
- Preserve the v1.0 data format for backwards compatibility
- Add comprehensive error handling
- Update validation checks accordingly
- Test with multiple element types
- Document physics assumptions
## Future Enhancements
Planned features:
- [ ] Support for nonlinear analyses (SOL 106)
- [ ] Modal analysis results (SOL 103)
- [ ] Thermal analysis (SOL 153)
- [ ] Contact results
- [ ] Composite material support
- [ ] Automatic mesh quality assessment
- [ ] Parallel batch processing
- [ ] Progress bars for long operations
- [ ] Integration with Atomizer dashboard
## License
Part of the Atomizer optimization platform.
## Support
For issues or questions:
1. Check this README and troubleshooting section
2. Review `Context.md` for project background
3. Examine example files in `Models/` folder
4. Check pyNastran documentation for BDF/OP2 specifics
## Version History
### v1.0.0 (Current)
- Initial release
- Complete BDF/OP2 parsing
- Support for solid, shell, beam elements
- HDF5 + JSON output format
- Data validation
- Batch processing
- Physics consistency checks
---
**AtomizerField**: Revolutionizing structural optimization through neural field learning.
*Built with Claude Code, designed for the future of engineering.*

529
atomizer-field/SIMPLE_BEAM_TEST_REPORT.md
View File

@@ -1,529 +0,0 @@
# Simple Beam Test Report
**AtomizerField Neural Field Learning System**
**Test Date:** November 24, 2025
**Model:** Simple Beam (beam_sim1-solution_1)
**Status:** ✅ ALL TESTS PASSED
---
## Executive Summary
The AtomizerField system has been successfully validated with your actual Simple Beam FEA model. All 7 comprehensive tests passed, demonstrating complete functionality from BDF/OP2 parsing through neural network prediction.
**Key Results:**
- ✅ 7/7 tests passed
- ✅ 5,179 nodes processed
- ✅ 4,866 elements parsed
- ✅ Complete field extraction (displacement + stress)
- ✅ Neural network inference: 95.94 ms
- ✅ System ready for training!
---
## Test Results
### Test 1: File Existence ✅ PASS
**Purpose:** Verify Simple Beam files are available
**Results:**
- BDF file found: `beam_sim1-solution_1.dat` (1,230.1 KB)
- OP2 file found: `beam_sim1-solution_1.op2` (4,461.2 KB)
**Status:** Files located and validated
---
### Test 2: Directory Setup ✅ PASS
**Purpose:** Create test case directory structure
**Results:**
- Created: `test_case_beam/input/`
- Created: `test_case_beam/output/`
- Copied BDF to input directory
- Copied OP2 to output directory
**Status:** Directory structure established
---
### Test 3: Module Imports ✅ PASS
**Purpose:** Verify all required modules load correctly
**Results:**
- pyNastran imported successfully
- AtomizerField parser imported successfully
- All dependencies available
**Status:** Environment configured correctly
---
### Test 4: BDF/OP2 Parsing ✅ PASS
**Purpose:** Extract all data from FEA files
**Parse Time:** 1.27 seconds
**Extracted Data:**
- **Nodes:** 5,179 nodes with 3D coordinates
- **Elements:** 4,866 CQUAD4 shell elements
- **Materials:** 1 material definition
- **Boundary Conditions:** 0 SPCs, 0 MPCs
- **Loads:** 35 forces, 0 pressures, 0 gravity, 0 thermal
- **Displacement Field:** 5,179 nodes × 6 DOF
- Maximum displacement: 19.556875 mm
- **Stress Field:** 9,732 stress values (2 per element)
- Captured for all elements
- **Reactions:** 5,179 reaction forces
- Maximum force: 152,198,576 N
**Output Files:**
- JSON metadata: 1,686.3 KB
- HDF5 field data: 546.3 KB
- **Total:** 2,232.6 KB
**Status:** Complete field extraction successful
---
### Test 5: Data Validation ✅ PASS
**Purpose:** Verify data quality and physics consistency
**Validation Checks:**
- ✅ JSON and HDF5 files present
- ✅ All required fields found
- ✅ Node coordinates valid (5,179 nodes)
- ✅ Element connectivity valid (4,866 elements)
- ✅ Material definitions complete (1 material)
- ✅ Displacement field complete (max: 19.56 mm)
- ✅ Stress field complete (9,732 values)
- ⚠ Warning: No SPCs defined (may be unconstrained)
**Status:** Data quality validated, ready for neural network
---
### Test 6: Graph Conversion ✅ PASS
**Purpose:** Convert to PyTorch Geometric format for neural network
**Graph Structure:**
- **Nodes:** 5,179 nodes
- **Node Features:** 12 dimensions
- Position (3D)
- Boundary conditions (6 DOF)
- Applied loads (3D)
- **Edges:** 58,392 edges
- **Edge Features:** 5 dimensions
- Young's modulus
- Poisson's ratio
- Density
- Shear modulus
- Thermal expansion
- **Target Displacement:** (5179, 6) - 6 DOF per node
- **Target Stress:** (9732, 8) - Full stress tensor per element
**Status:** Successfully converted to graph neural network format
---
### Test 7: Neural Prediction ✅ PASS
**Purpose:** Validate neural network can process the data
**Model Configuration:**
- Architecture: Graph Neural Network (GNN)
- Parameters: 128,589 parameters
- Layers: 6 message passing layers
- Hidden dimension: 64
- Model state: Untrained (random weights)
**Inference Performance:**
- **Inference Time:** 95.94 ms
- **Target:** < 100 ms ✅
- **Speedup vs FEA:** 1000× expected after training
**Predictions (Untrained Model):**
- Max displacement: 2.03 (arbitrary units)
- Max stress: 4.98 (arbitrary units)
**Note:** Values are from untrained model with random weights. After training on 50-500 examples, predictions will match FEA results with < 10% error.
**Status:** Neural network architecture validated and functional
---
## Model Statistics
### Geometry
| Property | Value |
|----------|-------|
| Nodes | 5,179 |
| Elements | 4,866 |
| Element Type | CQUAD4 (shell) |
| Materials | 1 |
### Loading
| Property | Value |
|----------|-------|
| Applied Forces | 35 |
| Pressure Loads | 0 |
| Gravity Loads | 0 |
| Thermal Loads | 0 |
### Results
| Property | Value |
|----------|-------|
| Max Displacement | 19.556875 mm |
| Displacement Nodes | 5,179 |
| Stress Elements | 9,732 (2 per element) |
| Max Reaction Force | 152,198,576 N |
### Data Files
| File | Size |
|------|------|
| BDF Input | 1,230.1 KB |
| OP2 Results | 4,461.2 KB |
| JSON Metadata | 1,686.3 KB |
| HDF5 Field Data | 546.3 KB |
| **Total Parsed** | **2,232.6 KB** |
---
## 3D Visualizations
### Mesh Structure
![Mesh Structure](visualization_images/mesh.png)
The Simple Beam model consists of 5,179 nodes connected by 4,866 CQUAD4 shell elements, creating a detailed 3D representation of the beam geometry.
### Displacement Field
![Displacement Field](visualization_images/displacement.png)
**Left:** Original mesh
**Right:** Deformed mesh (10× displacement scale)
The displacement field shows the beam's deformation under load, with maximum displacement of 19.56 mm. Colors represent displacement magnitude, with red indicating maximum deformation.
### Stress Field
![Stress Field](visualization_images/stress.png)
The von Mises stress distribution shows stress concentrations throughout the beam structure. Colors range from blue (low stress) to red (high stress), revealing critical stress regions.
---
## Performance Metrics
### Parsing Performance
| Metric | Value |
|--------|-------|
| Parse Time | 1.27 seconds |
| Nodes/second | 4,077 nodes/s |
| Elements/second | 3,831 elements/s |
### Neural Network Performance
| Metric | Value | Target | Status |
|--------|-------|--------|--------|
| Inference Time | 95.94 ms | < 100 ms | ✅ Pass |
| Model Parameters | 128,589 | - | - |
| Forward Pass | Working | - | ✅ |
| Gradient Flow | Working | - | ✅ |
### Comparison: FEA vs Neural (After Training)
| Operation | FEA Time | Neural Time | Speedup |
|-----------|----------|-------------|---------|
| Single Analysis | 30-300 s | 0.096 s | **300-3000×** |
| Optimization (100 evals) | 50-500 min | 10 s | **300-3000×** |
| Gradient Computation | Very slow | 0.1 ms | **1,000,000×** |
---
## System Validation
### Functional Tests
- ✅ File I/O (BDF/OP2 reading)
- ✅ Data extraction (mesh, materials, BCs, loads)
- ✅ Field extraction (displacement, stress)
- ✅ Data validation (quality checks)
- ✅ Format conversion (FEA → neural)
- ✅ Graph construction (PyTorch Geometric)
- ✅ Neural network inference
### Data Quality
- ✅ No NaN values in coordinates
- ✅ No NaN values in displacement
- ✅ No NaN values in stress
- ✅ Element connectivity valid
- ✅ Node IDs consistent
- ✅ Physics units preserved (mm, MPa, N)
### Neural Network
- ✅ Model instantiation
- ✅ Forward pass
- ✅ All 4 loss functions operational
- ✅ Batch processing
- ✅ Gradient computation
---
## Next Steps
### 1. Generate Training Data (50-500 cases)
**Goal:** Create diverse dataset for training
**Approach:**
- Vary beam dimensions
- Vary loading conditions
- Vary material properties
- Vary boundary conditions
**Command:**
```bash
conda activate atomizer_field
python batch_parser.py --input Models/ --output training_data/
```
### 2. Train Neural Network
**Goal:** Learn FEA behavior from examples
**Configuration:**
- Epochs: 100-200
- Batch size: 16
- Learning rate: 0.001
- Loss: Physics-informed
**Command:**
```bash
python train.py \
--data_dirs training_data/* \
--epochs 100 \
--batch_size 16 \
--loss physics \
--checkpoint_dir checkpoints/
```
**Expected Training Time:** 2-6 hours (GPU recommended)
### 3. Validate Performance
**Goal:** Verify < 10% prediction error
**Tests:**
- Physics validation (cantilever, beam tests)
- Learning tests (memorization, interpolation)
- Prediction accuracy on test set
**Command:**
```bash
python test_suite.py --full
```
### 4. Deploy to Production
**Goal:** Integrate with Atomizer for optimization
**Integration:**
```python
from optimization_interface import NeuralFieldOptimizer
# Initialize
optimizer = NeuralFieldOptimizer('checkpoints/best_model.pt')
# Replace FEA calls
results = optimizer.evaluate(design_graph)
gradients = optimizer.get_sensitivities(design_graph)
```
**Expected Speedup:** 1000× faster than FEA!
---
## Technical Details
### Graph Neural Network Architecture
**Input Layer:**
- Node features: 12D (position, BCs, loads)
- Edge features: 5D (material properties)
**Hidden Layers:**
- 6 message passing layers
- Hidden dimension: 64
- Activation: ReLU
- Dropout: 0.1
**Output Layers:**
- Displacement decoder: 6 DOF per node
- Stress predictor: 6 stress components per element
- Von Mises calculator: Scalar per element
**Total Parameters:** 128,589
### Data Format
**JSON Metadata:**
```json
{
"metadata": { "case_name", "analysis_type", ... },
"mesh": { "nodes", "elements", "statistics" },
"materials": { ... },
"boundary_conditions": { ... },
"loads": { ... },
"results": { "displacement", "stress" }
}
```
**HDF5 Arrays:**
- `mesh/node_coordinates`: (5179, 3) float32
- `mesh/node_ids`: (5179,) int32
- `results/displacement`: (5179, 6) float32
- `results/stress/cquad4_stress/data`: (9732, 8) float32
### Physics-Informed Loss
**Total Loss:**
```
L_total = λ_data * L_data
+ λ_equilibrium * L_equilibrium
+ λ_constitutive * L_constitutive
+ λ_boundary * L_boundary
```
**Components:**
- **Data Loss:** MSE between prediction and FEA
- **Equilibrium:** ∇·σ + f = 0 (force balance)
- **Constitutive:** σ = C:ε (Hooke's law)
- **Boundary:** Enforce BC compliance
---
## Conclusions
### ✅ System Status: FULLY OPERATIONAL
All components of the AtomizerField system have been validated:
1. **Data Pipeline** ✅
- BDF/OP2 parsing working
- Complete field extraction
- Data quality validated
2. **Neural Network** ✅
- Model architecture validated
- Forward pass working
- Inference time: 95.94 ms
3. **Visualization** ✅
- 3D mesh rendering
- Displacement fields
- Stress fields
- Automated report generation
4. **Testing Framework** ✅
- 7/7 tests passing
- Comprehensive validation
- Performance benchmarks met
### Key Achievements
- ✅ Successfully parsed real 5,179-node model
- ✅ Extracted complete displacement and stress fields
- ✅ Converted to neural network format
- ✅ Neural inference < 100ms
- ✅ 3D visualization working
- ✅ Ready for training!
### Performance Expectations
**After Training (50-500 cases, 100-200 epochs):**
- Prediction error: < 10% vs FEA
- Inference time: 5-50 ms
- Speedup: 1000× faster than FEA
- Optimization: 1,000,000× faster gradients
### Production Readiness
The system is **ready for production** after training:
- ✅ All tests passing
- ✅ Data pipeline validated
- ✅ Neural architecture proven
- ✅ Visualization tools available
- ✅ Integration interface ready
**The AtomizerField system will revolutionize your structural optimization workflow with 1000× faster predictions!** 🚀
---
## Appendix
### Files Generated
**Test Data:**
- `test_case_beam/input/model.bdf` (1,230 KB)
- `test_case_beam/output/model.op2` (4,461 KB)
- `test_case_beam/neural_field_data.json` (1,686 KB)
- `test_case_beam/neural_field_data.h5` (546 KB)
**Visualizations:**
- `visualization_images/mesh.png` (227 KB)
- `visualization_images/displacement.png` (335 KB)
- `visualization_images/stress.png` (215 KB)
**Reports:**
- `visualization_report.md`
- `SIMPLE_BEAM_TEST_REPORT.md` (this file)
### Commands Reference
```bash
# Activate environment
conda activate atomizer_field
# Run tests
python test_simple_beam.py # Simple Beam test
python test_suite.py --quick # Smoke tests
python test_suite.py --full # Complete validation
# Visualize
python visualize_results.py test_case_beam --mesh # Mesh only
python visualize_results.py test_case_beam --displacement # Displacement
python visualize_results.py test_case_beam --stress # Stress
python visualize_results.py test_case_beam --report # Full report
# Parse data
python neural_field_parser.py test_case_beam # Single case
python batch_parser.py --input Models/ # Batch
# Train
python train.py --data_dirs training_data/* --epochs 100
# Predict
python predict.py --model best_model.pt --data test_case/
```
### Environment Details
**Conda Environment:** `atomizer_field`
**Key Packages:**
- Python 3.10.19
- NumPy 1.26.4 (conda-compiled)
- PyTorch 2.5.1
- PyTorch Geometric 2.7.0
- pyNastran 1.4.1
- Matplotlib 3.10.7
- H5Py 3.15.1
**Installation:**
```bash
conda create -n atomizer_field python=3.10 numpy scipy -y
conda activate atomizer_field
conda install pytorch torchvision torchaudio cpuonly -c pytorch -y
pip install torch-geometric pyNastran h5py tensorboard matplotlib
```
---
**Report Generated:** November 24, 2025
**AtomizerField Version:** 1.0
**Status:** ✅ All Systems Operational
**Ready For:** Production Training and Deployment
🎉 **COMPLETE SUCCESS!**

741
atomizer-field/SYSTEM_ARCHITECTURE.md
View File

@@ -1,741 +0,0 @@
# AtomizerField - Complete System Architecture
## 📍 Project Location
```
c:\Users\antoi\Documents\Atomaste\Atomizer-Field\
```
## 🏗️ System Overview
AtomizerField is a **two-phase system** that transforms FEA results into neural network predictions:
```
┌─────────────────────────────────────────────────────────────────┐
│ PHASE 1: DATA PIPELINE │
├─────────────────────────────────────────────────────────────────┤
│ │
│ NX Nastran Files (.bdf, .op2) │
│ ↓ │
│ neural_field_parser.py │
│ ↓ │
│ Neural Field Format (JSON + HDF5) │
│ ↓ │
│ validate_parsed_data.py │
│ │
└─────────────────────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────────────┐
│ PHASE 2: NEURAL NETWORK │
├─────────────────────────────────────────────────────────────────┤
│ │
│ data_loader.py → Graph Representation │
│ ↓ │
│ train.py + field_predictor.py (GNN) │
│ ↓ │
│ Trained Model (checkpoint_best.pt) │
│ ↓ │
│ predict.py → Field Predictions (5-50ms!) │
│ │
└─────────────────────────────────────────────────────────────────┘
```
---
## 📂 Complete File Structure
```
Atomizer-Field/
│
├── 📄 Core Documentation
│ ├── README.md # Phase 1 detailed guide
│ ├── PHASE2_README.md # Phase 2 detailed guide
│ ├── GETTING_STARTED.md # Quick start tutorial
│ ├── SYSTEM_ARCHITECTURE.md # This file (system overview)
│ ├── Context.md # Project vision & philosophy
│ └── Instructions.md # Original implementation spec
│
├── 🔧 Phase 1: FEA Data Parser
│ ├── neural_field_parser.py # Main parser (BDF/OP2 → Neural format)
│ ├── validate_parsed_data.py # Data quality validation
│ ├── batch_parser.py # Batch processing multiple cases
│ └── metadata_template.json # Template for design parameters
│
├── 🧠 Phase 2: Neural Network
│ ├── neural_models/
│ │ ├── __init__.py
│ │ ├── field_predictor.py # GNN architecture (718K params)
│ │ ├── physics_losses.py # Physics-informed loss functions
│ │ └── data_loader.py # PyTorch Geometric data pipeline
│ │
│ ├── train.py # Training script
│ └── predict.py # Inference script
│
├── 📦 Dependencies & Config
│ ├── requirements.txt # All dependencies
│ └── .gitignore # (if using git)
│
├── 📁 Data Directories (created during use)
│ ├── training_data/ # Parsed training cases
│ ├── validation_data/ # Parsed validation cases
│ ├── test_data/ # Parsed test cases
│ └── runs/ # Training outputs
│ ├── checkpoint_best.pt # Best model
│ ├── checkpoint_latest.pt # Latest checkpoint
│ ├── config.json # Model configuration
│ └── tensorboard/ # Training logs
│
├── 🔬 Example Models (your existing data)
│ └── Models/
│ └── Simple Beam/
│ ├── beam_sim1-solution_1.dat # BDF file
│ ├── beam_sim1-solution_1.op2 # OP2 results
│ └── ...
│
└── 🐍 Virtual Environment
└── atomizer_env/ # Python virtual environment
```
---
## 🔍 PHASE 1: Data Parser - Deep Dive
### Location
```
c:\Users\antoi\Documents\Atomaste\Atomizer-Field\neural_field_parser.py
```
### What It Does
**Transforms this:**
```
NX Nastran Files:
├── model.bdf (1.2 MB text file with mesh, materials, BCs, loads)
└── model.op2 (4.5 MB binary file with stress/displacement results)
```
**Into this:**
```
Neural Field Format:
├── neural_field_data.json (200 KB - metadata, structure)
└── neural_field_data.h5 (3 MB - large numerical arrays)
```
### Data Structure Breakdown
#### 1. JSON File (neural_field_data.json)
```json
{
"metadata": {
"version": "1.0.0",
"created_at": "2024-01-15T10:30:00",
"source": "NX_Nastran",
"case_name": "training_case_001",
"analysis_type": "SOL_101",
"units": {
"length": "mm",
"force": "N",
"stress": "MPa"
},
"file_hashes": {
"bdf": "sha256_hash_here",
"op2": "sha256_hash_here"
}
},
"mesh": {
"statistics": {
"n_nodes": 15432,
"n_elements": 8765,
"element_types": {
"solid": 5000,
"shell": 3000,
"beam": 765
}
},
"bounding_box": {
"min": [0.0, 0.0, 0.0],
"max": [100.0, 50.0, 30.0]
},
"nodes": {
"ids": [1, 2, 3, ...],
"coordinates": "<stored in HDF5>",
"shape": [15432, 3]
},
"elements": {
"solid": [
{
"id": 1,
"type": "CTETRA",
"nodes": [1, 5, 12, 34],
"material_id": 1,
"property_id": 10
},
...
],
"shell": [...],
"beam": [...]
}
},
"materials": [
{
"id": 1,
"type": "MAT1",
"E": 71700.0, // Young's modulus (MPa)
"nu": 0.33, // Poisson's ratio
"rho": 2.81e-06, // Density (kg/mm³)
"G": 26900.0, // Shear modulus (MPa)
"alpha": 2.3e-05 // Thermal expansion (1/°C)
}
],
"boundary_conditions": {
"spc": [ // Single-point constraints
{
"id": 1,
"node": 1,
"dofs": "123456", // Constrained DOFs (x,y,z,rx,ry,rz)
"enforced_motion": 0.0
},
...
],
"mpc": [] // Multi-point constraints
},
"loads": {
"point_forces": [
{
"id": 100,
"type": "force",
"node": 500,
"magnitude": 10000.0, // Newtons
"direction": [1.0, 0.0, 0.0],
"coord_system": 0
}
],
"pressure": [],
"gravity": [],
"thermal": []
},
"results": {
"displacement": {
"node_ids": [1, 2, 3, ...],
"data": "<stored in HDF5>",
"shape": [15432, 6],
"max_translation": 0.523456,
"max_rotation": 0.001234,
"units": "mm and radians"
},
"stress": {
"ctetra_stress": {
"element_ids": [1, 2, 3, ...],
"data": "<stored in HDF5>",
"shape": [5000, 7],
"max_von_mises": 245.67,
"units": "MPa"
}
}
}
}
```
#### 2. HDF5 File (neural_field_data.h5)
**Structure:**
```
neural_field_data.h5
│
├── /mesh/
│ ├── node_coordinates [15432 × 3] float64
│ │ Each row: [x, y, z] in mm
│ │
│ └── node_ids [15432] int32
│ Node ID numbers
│
└── /results/
├── /displacement [15432 × 6] float64
│ Each row: [ux, uy, uz, θx, θy, θz]
│ Translation (mm) + Rotation (radians)
│
├── displacement_node_ids [15432] int32
│
├── /stress/
│ ├── /ctetra_stress/
│ │ ├── data [5000 × 7] float64
│ │ │ [σxx, σyy, σzz, τxy, τyz, τxz, von_mises]
│ │ └── element_ids [5000] int32
│ │
│ └── /cquad4_stress/
│ └── ...
│
├── /strain/
│ └── ...
│
└── /reactions [N × 6] float64
Reaction forces at constrained nodes
```
**Why HDF5?**
- ✅ Efficient storage (compressed)
- ✅ Fast random access
- ✅ Handles large arrays (millions of values)
- ✅ Industry standard for scientific data
- ✅ Direct NumPy/PyTorch integration
### Parser Code Flow
```python
# neural_field_parser.py - Main Parser Class
class NastranToNeuralFieldParser:
def __init__(self, case_directory):
# Find BDF and OP2 files
# Initialize pyNastran readers
def parse_all(self):
# 1. Read BDF (input deck)
self.bdf.read_bdf(bdf_file)
# 2. Read OP2 (results)
self.op2.read_op2(op2_file)
# 3. Extract data
self.extract_metadata() # Analysis info, units
self.extract_mesh() # Nodes, elements, connectivity
self.extract_materials() # Material properties
self.extract_boundary_conditions() # SPCs, MPCs
self.extract_loads() # Forces, pressures, gravity
self.extract_results() # COMPLETE FIELDS (key!)
# 4. Save
self.save_data() # JSON + HDF5
```
**Key Innovation in `extract_results()`:**
```python
def extract_results(self):
# Traditional FEA post-processing:
# max_stress = np.max(stress_data) ← LOSES SPATIAL INFO!
# AtomizerField approach:
# Store COMPLETE field at EVERY node/element
results["displacement"] = {
"data": disp_data.tolist(), # ALL 15,432 nodes × 6 DOF
"shape": [15432, 6],
"max_translation": float(np.max(magnitudes)) # Also store max
}
# This enables neural network to learn spatial patterns!
```
---
## 🧠 PHASE 2: Neural Network - Deep Dive
### Location
```
c:\Users\antoi\Documents\Atomaste\Atomizer-Field\neural_models\
```
### Architecture Overview
```
┌─────────────────────────────────────────────────────────────────┐
│ AtomizerFieldModel │
│ (718,221 parameters) │
├─────────────────────────────────────────────────────────────────┤
│ │
│ INPUT: Graph Representation of FEA Mesh │
│ ├── Nodes (15,432): │
│ │ └── Features [12D]: [x,y,z, BC_mask(6), loads(3)] │
│ └── Edges (mesh connectivity): │
│ └── Features [5D]: [E, ν, ρ, G, α] (materials) │
│ │
│ ┌──────────────────────────────────────────────────┐ │
│ │ NODE ENCODER (12 → 128) │ │
│ │ Embeds node position + BCs + loads │ │
│ └──────────────────────────────────────────────────┘ │
│ ↓ │
│ ┌──────────────────────────────────────────────────┐ │
│ │ EDGE ENCODER (5 → 64) │ │
│ │ Embeds material properties │ │
│ └──────────────────────────────────────────────────┘ │
│ ↓ │
│ ┌──────────────────────────────────────────────────┐ │
│ │ MESSAGE PASSING LAYERS × 6 │ │
│ │ ┌────────────────────────────────────┐ │ │
│ │ │ Layer 1: MeshGraphConv │ │ │
│ │ │ ├── Gather neighbor info │ │ │
│ │ │ ├── Combine with edge features │ │ │
│ │ │ ├── Update node representations │ │ │
│ │ │ └── Residual + LayerNorm │ │ │
│ │ ├────────────────────────────────────┤ │ │
│ │ │ Layer 2-6: Same structure │ │ │
│ │ └────────────────────────────────────┘ │ │
│ │ (Forces propagate through mesh!) │ │
│ └──────────────────────────────────────────────────┘ │
│ ↓ │
│ ┌──────────────────────────────────────────────────┐ │
│ │ DISPLACEMENT DECODER (128 → 6) │ │
│ │ Predicts: [ux, uy, uz, θx, θy, θz] │ │
│ └──────────────────────────────────────────────────┘ │
│ ↓ │
│ ┌──────────────────────────────────────────────────┐ │
│ │ STRESS PREDICTOR (6 → 6) │ │
│ │ From displacement → stress tensor │ │
│ │ Outputs: [σxx, σyy, σzz, τxy, τyz, τxz] │ │
│ └──────────────────────────────────────────────────┘ │
│ ↓ │
│ OUTPUT: │
│ ├── Displacement field [15,432 × 6] │
│ ├── Stress field [15,432 × 6] │
│ └── Von Mises stress [15,432 × 1] │
│ │
└─────────────────────────────────────────────────────────────────┘
```
### Graph Representation
**From Mesh to Graph:**
```
FEA Mesh: Graph:
Node 1 ──── Element 1 ──── Node 2 Node 1 ──── Edge ──── Node 2
│ │ │ │
│ │ Features: Features:
Element 2 Element 3 [x,y,z, [x,y,z,
│ │ BC,loads] BC,loads]
│ │ │ │
Node 3 ──── Element 4 ──── Node 4 Edge Edge
│ │
[E,ν,ρ,G,α] [E,ν,ρ,G,α]
```
**Built by `data_loader.py`:**
```python
class FEAMeshDataset(Dataset):
def _build_graph(self, metadata, node_coords, displacement, stress):
# 1. Build node features
x = torch.cat([
node_coords, # [N, 3] - position
bc_mask, # [N, 6] - which DOFs constrained
load_features # [N, 3] - applied forces
], dim=-1) # → [N, 12]
# 2. Build edges from element connectivity
for element in elements:
nodes = element['nodes']
# Fully connect nodes within element
for i, j in pairs(nodes):
edge_index.append([i, j])
edge_attr.append(material_props)
# 3. Create PyTorch Geometric Data object
data = Data(
x=x, # Node features
edge_index=edge_index, # Connectivity
edge_attr=edge_attr, # Material properties
y_displacement=displacement, # Target (ground truth)
y_stress=stress # Target (ground truth)
)
return data
```
### Physics-Informed Loss
**Standard Neural Network:**
```python
loss = MSE(prediction, ground_truth)
# Only learns to match training data
```
**AtomizerField (Physics-Informed):**
```python
loss = λ_data × MSE(prediction, ground_truth)
+ λ_eq × EquilibriumViolation(stress) # ∇·σ + f = 0
+ λ_const × ConstitutiveLawError(stress, strain) # σ = C:ε
+ λ_bc × BoundaryConditionError(disp, BCs) # u = 0 at fixed nodes
# Learns physics, not just patterns!
```
**Benefits:**
- Faster convergence
- Better generalization to unseen cases
- Physically plausible predictions
- Needs less training data
### Training Pipeline
**`train.py` workflow:**
```python
# 1. Load data
train_loader = create_dataloaders(train_cases, val_cases)
# 2. Create model
model = AtomizerFieldModel(
node_feature_dim=12,
hidden_dim=128,
num_layers=6
)
# 3. Training loop
for epoch in range(num_epochs):
for batch in train_loader:
# Forward pass
predictions = model(batch)
# Compute loss
losses = criterion(predictions, targets)
# Backward pass
losses['total_loss'].backward()
optimizer.step()
# Validate
val_metrics = validate(val_loader)
# Save checkpoint if best
if val_loss < best_val_loss:
save_checkpoint('checkpoint_best.pt')
# TensorBoard logging
writer.add_scalar('Loss/train', train_loss, epoch)
```
**Outputs:**
```
runs/
├── checkpoint_best.pt # Best model (lowest validation loss)
├── checkpoint_latest.pt # Latest state (for resuming)
├── config.json # Model configuration
└── tensorboard/ # Training logs
└── events.out.tfevents...
```
### Inference (Prediction)
**`predict.py` workflow:**
```python
# 1. Load trained model
model = load_model('checkpoint_best.pt')
# 2. Load new case (mesh + BCs + loads, NO FEA solve!)
data = load_case('new_design')
# 3. Predict in milliseconds
predictions = model(data) # ~15ms
# 4. Extract results
displacement = predictions['displacement'] # [N, 6]
stress = predictions['stress'] # [N, 6]
von_mises = predictions['von_mises'] # [N]
# 5. Get max values (like traditional FEA)
max_disp = np.max(np.linalg.norm(displacement[:, :3], axis=1))
max_stress = np.max(von_mises)
print(f"Max displacement: {max_disp:.6f} mm")
print(f"Max stress: {max_stress:.2f} MPa")
```
**Performance:**
- Traditional FEA: 2-3 hours
- AtomizerField: 15 milliseconds
- **Speedup: ~480,000×**
---
## 🎯 Key Innovations
### 1. Complete Field Learning (Not Scalars)
**Traditional Surrogate:**
```python
# Only learns one number per analysis
max_stress = neural_net(design_parameters)
```
**AtomizerField:**
```python
# Learns ENTIRE FIELD (45,000 values)
stress_field = neural_net(mesh_graph)
# Knows WHERE stress occurs, not just max value!
```
### 2. Graph Neural Networks (Respect Topology)
```
Why GNNs?
- FEA solves: K·u = f
- K depends on mesh connectivity
- GNN learns on mesh structure
- Messages propagate like forces!
```
### 3. Physics-Informed Training
```
Standard NN: "Make output match training data"
AtomizerField: "Match data AND obey physics laws"
Result: Better with less data!
```
---
## 💾 Data Flow Example
### Complete End-to-End Flow
```
1. Engineer creates bracket in NX
├── Geometry: 100mm × 50mm × 30mm
├── Material: Aluminum 7075-T6
├── Mesh: 15,432 nodes, 8,765 elements
├── BCs: Fixed at mounting holes
└── Load: 10,000 N tension
2. Run FEA in NX Nastran
├── Time: 2.5 hours
└── Output: model.bdf, model.op2
3. Parse to neural format
$ python neural_field_parser.py bracket_001
├── Time: 15 seconds
├── Output: neural_field_data.json (200 KB)
└── neural_field_data.h5 (3.2 MB)
4. Train neural network (once, on 500 brackets)
$ python train.py --train_dir ./brackets --epochs 150
├── Time: 8 hours (one-time)
└── Output: checkpoint_best.pt (3 MB model)
5. Predict new bracket design
$ python predict.py --model checkpoint_best.pt --input new_bracket
├── Time: 15 milliseconds
├── Output:
│ ├── Max displacement: 0.523 mm
│ ├── Max stress: 245.7 MPa
│ └── Complete stress field at all 15,432 nodes
└── Can now test 10,000 designs in 2.5 minutes!
```
---
## 🔧 How to Use Your System
### Quick Reference Commands
```bash
# Navigate to project
cd c:\Users\antoi\Documents\Atomaste\Atomizer-Field
# Activate environment
atomizer_env\Scripts\activate
# ===== PHASE 1: Parse FEA Data =====
# Single case
python neural_field_parser.py case_001
# Validate
python validate_parsed_data.py case_001
# Batch process
python batch_parser.py ./all_cases
# ===== PHASE 2: Train Neural Network =====
# Train model
python train.py \
--train_dir ./training_data \
--val_dir ./validation_data \
--epochs 100 \
--batch_size 4
# Monitor training
tensorboard --logdir runs/tensorboard
# ===== PHASE 2: Run Predictions =====
# Predict single case
python predict.py \
--model runs/checkpoint_best.pt \
--input test_case_001
# Batch prediction
python predict.py \
--model runs/checkpoint_best.pt \
--input ./test_cases \
--batch
```
---
## 📊 Expected Results
### Phase 1 (Parser)
**Input:**
- BDF file: 1.2 MB
- OP2 file: 4.5 MB
**Output:**
- JSON: ~200 KB (metadata)
- HDF5: ~3 MB (fields)
- Time: ~15 seconds
### Phase 2 (Training)
**Training Set:**
- 500 parsed cases
- Time: 8-12 hours
- GPU: NVIDIA RTX 3080
**Validation Accuracy:**
- Displacement error: 3-5%
- Stress error: 5-10%
- Max value error: 1-3%
### Phase 2 (Inference)
**Per Prediction:**
- Time: 5-50 milliseconds
- Accuracy: Within 5% of FEA
- Speedup: 10,000× - 500,000×
---
## 🎓 What You Have Built
You now have a complete system that:
1. ✅ Parses NX Nastran results into ML-ready format
2. ✅ Converts FEA meshes to graph neural network format
3. ✅ Trains physics-informed GNNs to predict stress/displacement
4. ✅ Runs inference 1000× faster than traditional FEA
5. ✅ Provides complete field distributions (not just max values)
6. ✅ Enables rapid design optimization
**Total Implementation:**
- ~3,000 lines of production-ready Python code
- Comprehensive documentation
- Complete testing framework
- Ready for real optimization workflows
---
This is a **revolutionary approach** to structural optimization that combines:
- Traditional FEA accuracy
- Neural network speed
- Physics-informed learning
- Graph-based topology understanding
You're ready to transform hours of FEA into milliseconds of prediction! 🚀

277
atomizer-field/TESTING_CHECKLIST.md
View File

@@ -1,277 +0,0 @@
# AtomizerField Testing Checklist
Quick reference for testing status and next steps.
---
## ✅ Completed Tests
### Environment Setup
- [x] Conda environment created (`atomizer_field`)
- [x] All dependencies installed
- [x] NumPy MINGW-W64 issue resolved
- [x] No segmentation faults
### Smoke Tests (5/5)
- [x] Model creation (128,589 parameters)
- [x] Forward pass
- [x] Loss functions (4 types)
- [x] Batch processing
- [x] Gradient flow
### Simple Beam Test (7/7)
- [x] File existence (BDF + OP2)
- [x] Directory setup
- [x] Module imports
- [x] BDF/OP2 parsing (5,179 nodes, 4,866 elements)
- [x] Data validation
- [x] Graph conversion
- [x] Neural prediction (95.94 ms)
### Visualization
- [x] 3D mesh rendering
- [x] Displacement field (original + deformed)
- [x] Stress field (von Mises)
- [x] Report generation (markdown + images)
### Unit Validation
- [x] UNITSYS detection (MN-MM)
- [x] Material properties (E = 200 GPa)
- [x] Stress values (117 MPa reasonable)
- [x] Force values (2.73 MN validated)
- [x] Direction vectors preserved
---
## ❌ Not Yet Tested (Requires Trained Model)
### Physics Tests (0/4)
- [ ] Cantilever beam (analytical comparison)
- [ ] Equilibrium check (∇·σ + f = 0)
- [ ] Constitutive law (σ = C:ε)
- [ ] Energy conservation
### Learning Tests (0/4)
- [ ] Memorization (single case < 1% error)
- [ ] Interpolation (between cases < 10% error)
- [ ] Extrapolation (unseen loads < 20% error)
- [ ] Pattern recognition (physics transfer)
### Integration Tests (0/5)
- [ ] Batch prediction
- [ ] Gradient computation
- [ ] Optimization loop
- [ ] Uncertainty quantification
- [ ] Online learning
### Performance Tests (0/3)
- [ ] Accuracy benchmark (< 10% error)
- [ ] Speed benchmark (< 50 ms)
- [ ] Scalability (10K+ nodes)
---
## 🔧 Known Issues to Fix
### Minor (Non-blocking)
- [ ] Unit labels: "MPa" should be "kPa" (or convert values)
- [ ] Missing SPCs warning (investigate BDF)
- [ ] Unicode encoding (mostly fixed, minor cleanup remains)
### Documentation
- [ ] Unit conversion guide
- [ ] Training data generation guide
- [ ] User manual
---
## 🚀 Testing Roadmap
### Phase 1: Pre-Training Validation
**Status:** ✅ COMPLETE
- [x] Core pipeline working
- [x] Test case validated
- [x] Units understood
- [x] Visualization working
### Phase 2: Training Preparation
**Status:** 🔜 NEXT
- [ ] Fix unit labels (30 min)
- [ ] Document unit system (1 hour)
- [ ] Create training data generation script
- [ ] Generate 50 test cases (1-2 weeks)
### Phase 3: Initial Training
**Status:** ⏸️ WAITING
- [ ] Train on 50 cases (2-4 hours)
- [ ] Validate on 10 held-out cases
- [ ] Check loss convergence
- [ ] Run memorization test
### Phase 4: Physics Validation
**Status:** ⏸️ WAITING
- [ ] Cantilever beam test
- [ ] Equilibrium check
- [ ] Energy conservation
- [ ] Compare vs analytical solutions
### Phase 5: Full Validation
**Status:** ⏸️ WAITING
- [ ] Run full test suite (18 tests)
- [ ] Accuracy benchmarks
- [ ] Speed benchmarks
- [ ] Scalability tests
### Phase 6: Production Deployment
**Status:** ⏸️ WAITING
- [ ] Integration with Atomizer
- [ ] End-to-end optimization test
- [ ] Performance profiling
- [ ] User acceptance testing
---
## 📊 Test Commands Quick Reference
### Run Tests
```bash
# Activate environment
conda activate atomizer_field
# Quick smoke tests (30 seconds)
python test_suite.py --quick
# Simple Beam end-to-end (1 minute)
python test_simple_beam.py
# Physics tests (15 minutes) - REQUIRES TRAINED MODEL
python test_suite.py --physics
# Full test suite (1 hour) - REQUIRES TRAINED MODEL
python test_suite.py --full
```
### Visualization
```bash
# Mesh only
python visualize_results.py test_case_beam --mesh
# Displacement
python visualize_results.py test_case_beam --displacement
# Stress
python visualize_results.py test_case_beam --stress
# Full report
python visualize_results.py test_case_beam --report
```
### Unit Validation
```bash
# Check parsed data units
python check_units.py
# Check OP2 raw data
python check_op2_units.py
# Check actual values
python check_actual_values.py
```
### Training (When Ready)
```bash
# Generate training data
python batch_parser.py --input Models/ --output training_data/
# Train model
python train.py \
--data_dirs training_data/* \
--epochs 100 \
--batch_size 16 \
--loss physics
# Monitor training
tensorboard --logdir runs/
```
---
## 📈 Success Criteria
### Phase 1: Core System ✅
- [x] All smoke tests passing
- [x] End-to-end test passing
- [x] Real FEA data processed
- [x] Visualization working
### Phase 2: Training Ready 🔜
- [ ] Unit labels correct
- [ ] 50+ training cases generated
- [ ] Training script validated
- [ ] Monitoring setup (TensorBoard)
### Phase 3: Model Trained ⏸️
- [ ] Training loss < 0.01
- [ ] Validation loss < 0.05
- [ ] No overfitting (train ≈ val loss)
- [ ] Predictions physically reasonable
### Phase 4: Physics Validated ⏸️
- [ ] Equilibrium error < 1%
- [ ] Constitutive error < 5%
- [ ] Energy conservation < 5%
- [ ] Analytical test < 5% error
### Phase 5: Production Ready ⏸️
- [ ] Prediction error < 10%
- [ ] Inference time < 50 ms
- [ ] All 18 tests passing
- [ ] Integration with Atomizer working
---
## 🎯 Current Focus
**Status:** ✅ Core validation complete, ready for training phase
**Next immediate steps:**
1. Fix unit labels (optional, 30 min)
2. Generate training data (critical, 1-2 weeks)
3. Train model (critical, 2-4 hours)
**Blockers:** None - system ready!
---
## 📞 Quick Status Check
Run this to verify system health:
```bash
conda activate atomizer_field
python test_simple_beam.py
```
Expected output:
```
TEST 1: Files exist ✓
TEST 2: Directory setup ✓
TEST 3: Modules import ✓
TEST 4: BDF/OP2 parsed ✓
TEST 5: Data validated ✓
TEST 6: Graph created ✓
TEST 7: Prediction made ✓
[SUCCESS] All 7 tests passed!
```
---
*Testing Checklist v1.0*
*Last updated: November 24, 2025*
*Status: Phase 1 complete, Phase 2 ready to start*

673
atomizer-field/TESTING_COMPLETE.md
View File

@@ -1,673 +0,0 @@
# AtomizerField Testing Framework - Complete Implementation
## Overview
The complete testing framework has been implemented for AtomizerField. All test modules are ready to validate the system from basic functionality through full neural FEA predictions.
---
## Test Structure
### Directory Layout
```
Atomizer-Field/
├── test_suite.py # Master orchestrator
├── test_simple_beam.py # Specific test for Simple Beam model
│
├── tests/
│ ├── __init__.py # Package initialization
│ ├── test_synthetic.py # Smoke tests (5 tests)
│ ├── test_physics.py # Physics validation (4 tests)
│ ├── test_learning.py # Learning capability (4 tests)
│ ├── test_predictions.py # Integration tests (5 tests)
│ └── analytical_cases.py # Analytical solutions library
│
└── test_results/ # Auto-generated results
```
---
## Implemented Test Modules
### 1. test_synthetic.py ✅ COMPLETE
**Purpose:** Basic functionality validation (smoke tests)
**5 Tests Implemented:**
1. **Model Creation** - Verify GNN instantiates (718K params)
2. **Forward Pass** - Model processes data correctly
3. **Loss Computation** - All 4 loss types work (MSE, Relative, Physics, Max)
4. **Batch Processing** - Handle multiple graphs
5. **Gradient Flow** - Backpropagation works
**Run standalone:**
```bash
python tests/test_synthetic.py
```
**Expected output:**
```
5/5 tests passed
✓ Model creation successful
✓ Forward pass works
✓ Loss functions operational
✓ Batch processing works
✓ Gradients flow correctly
```
---
### 2. test_physics.py ✅ COMPLETE
**Purpose:** Physics constraint validation
**4 Tests Implemented:**
1. **Cantilever Analytical** - Compare with δ = FL³/3EI
- Creates synthetic cantilever beam graph
- Computes analytical displacement
- Compares neural prediction
- Expected error: < 5% after training
2. **Equilibrium Check** - Verify ∇·σ + f = 0
- Tests force balance
- Checks stress field consistency
- Expected residual: < 1e-6 after training
3. **Energy Conservation** - Verify strain energy = work
- Computes external work (F·u)
- Computes strain energy (σ:ε)
- Expected balance: < 1% error
4. **Constitutive Law** - Verify σ = C:ε
- Tests Hooke's law compliance
- Checks stress-strain proportionality
- Expected: Linear relationship
**Run standalone:**
```bash
python tests/test_physics.py
```
**Note:** These tests will show physics compliance after model is trained with physics-informed losses.
---
### 3. test_learning.py ✅ COMPLETE
**Purpose:** Learning capability validation
**4 Tests Implemented:**
1. **Memorization Test** (10 samples, 100 epochs)
- Can network memorize small dataset?
- Expected: > 50% loss improvement
- Success criteria: Final loss < 0.1
2. **Interpolation Test** (Train: [1,3,5,7,9], Test: [2,4,6,8])
- Can network generalize between training points?
- Expected: < 5% error after training
- Tests pattern recognition within range
3. **Extrapolation Test** (Train: [1-5], Test: [7-10])
- Can network predict beyond training range?
- Expected: < 20% error (harder than interpolation)
- Tests robustness of learned patterns
4. **Pattern Recognition** (Stiffness variation)
- Does network learn physics relationships?
- Expected: Stiffness ↑ → Displacement ↓
- Tests understanding vs memorization
**Run standalone:**
```bash
python tests/test_learning.py
```
**Training details:**
- Each test trains a fresh model
- Uses synthetic datasets with known patterns
- Demonstrates learning capability before real FEA training
---
### 4. test_predictions.py ✅ COMPLETE
**Purpose:** Integration tests for complete pipeline
**5 Tests Implemented:**
1. **Parser Validation**
- Checks test_case_beam directory exists
- Validates parsed JSON/HDF5 files
- Reports node/element counts
- Requires: Run `test_simple_beam.py` first
2. **Training Pipeline**
- Creates synthetic dataset (5 samples)
- Trains model for 10 epochs
- Validates complete training loop
- Reports: Training time, final loss
3. **Prediction Accuracy**
- Quick trains on test case
- Measures displacement/stress errors
- Reports inference time
- Expected: < 100ms inference
4. **Performance Benchmark**
- Tests 4 mesh sizes: [10, 50, 100, 500] nodes
- Measures average inference time
- 10 runs per size for statistics
- Success: < 100ms for 100 nodes
5. **Batch Inference**
- Processes 5 graphs simultaneously
- Reports batch processing time
- Tests optimization loop scenario
- Validates parallel processing capability
**Run standalone:**
```bash
python tests/test_predictions.py
```
---
### 5. analytical_cases.py ✅ COMPLETE
**Purpose:** Library of analytical solutions for validation
**5 Analytical Cases:**
1. **Cantilever Beam (Point Load)**
```python
δ_max = FL³/3EI
σ_max = FL/Z
```
- Full deflection curve
- Moment distribution
- Stress field
2. **Simply Supported Beam (Center Load)**
```python
δ_max = FL³/48EI
σ_max = FL/4Z
```
- Symmetric deflection
- Support reactions
- Moment diagram
3. **Axial Tension Bar**
```python
δ = FL/EA
σ = F/A
ε = σ/E
```
- Linear displacement
- Uniform stress
- Constant strain
4. **Pressure Vessel (Thin-Walled)**
```python
σ_hoop = pr/t
σ_axial = pr/2t
```
- Hoop stress
- Axial stress
- Radial expansion
5. **Circular Shaft Torsion**
```python
θ = TL/GJ
τ_max = Tr/J
```
- Twist angle
- Shear stress distribution
- Shear strain
**Standard test cases:**
- `get_standard_cantilever()` - 1m steel beam, 1kN load
- `get_standard_simply_supported()` - 2m steel beam, 5kN load
- `get_standard_tension_bar()` - 1m square bar, 10kN load
**Run standalone to verify:**
```bash
python tests/analytical_cases.py
```
**Example output:**
```
1. Cantilever Beam (Point Load)
Max displacement: 1.905 mm
Max stress: 120.0 MPa
2. Simply Supported Beam (Point Load at Center)
Max displacement: 0.476 mm
Max stress: 60.0 MPa
Reactions: 2500.0 N each
...
```
---
## Master Test Orchestrator
### test_suite.py ✅ COMPLETE
**Four Testing Modes:**
1. **Quick Mode** (`--quick`)
- Duration: ~5 minutes
- Tests: 5 smoke tests
- Purpose: Verify basic functionality
```bash
python test_suite.py --quick
```
2. **Physics Mode** (`--physics`)
- Duration: ~15 minutes
- Tests: Smoke + Physics (9 tests)
- Purpose: Validate physics constraints
```bash
python test_suite.py --physics
```
3. **Learning Mode** (`--learning`)
- Duration: ~30 minutes
- Tests: Smoke + Physics + Learning (13 tests)
- Purpose: Confirm learning capability
```bash
python test_suite.py --learning
```
4. **Full Mode** (`--full`)
- Duration: ~1 hour
- Tests: All 18 tests
- Purpose: Complete validation
```bash
python test_suite.py --full
```
**Features:**
- Progress tracking
- Detailed reporting
- JSON results export
- Clean pass/fail output
- Duration tracking
- Metrics collection
**Output format:**
```
============================================================
AtomizerField Test Suite v1.0
Mode: QUICK
============================================================
[TEST] Model Creation
Description: Verify GNN model can be instantiated
Creating GNN model...
Model created: 718,221 parameters
Status: ✓ PASS
Duration: 0.15s
...
============================================================
TEST SUMMARY
============================================================
Total Tests: 5
✓ Passed: 5
✗ Failed: 0
Pass Rate: 100.0%
✓ ALL TESTS PASSED - SYSTEM READY!
============================================================
Total testing time: 0.5 minutes
Results saved to: test_results/test_results_quick_1234567890.json
```
---
## Test for Simple Beam Model
### test_simple_beam.py ✅ COMPLETE
**Purpose:** Validate complete pipeline with user's actual Simple Beam model
**7-Step Test:**
1. Check Files - Verify beam_sim1-solution_1.dat and .op2 exist
2. Setup Test Case - Create test_case_beam/ directory
3. Import Modules - Verify pyNastran and AtomizerField imports
4. Parse Beam - Parse BDF/OP2 files
5. Validate Data - Run quality checks
6. Load as Graph - Convert to PyG format
7. Neural Prediction - Make prediction with model
**Location of beam files:**
```
Models/Simple Beam/
├── beam_sim1-solution_1.dat (BDF)
└── beam_sim1-solution_1.op2 (Results)
```
**Run:**
```bash
python test_simple_beam.py
```
**Creates:**
```
test_case_beam/
├── input/
│ └── model.bdf
├── output/
│ └── model.op2
├── neural_field_data.json
└── neural_field_data.h5
```
---
## Results Export
### JSON Format
All test runs save results to `test_results/`:
```json
{
"timestamp": "2025-01-24T12:00:00",
"mode": "quick",
"tests": [
{
"name": "Model Creation",
"description": "Verify GNN model can be instantiated",
"status": "PASS",
"duration": 0.15,
"message": "Model created successfully (718,221 params)",
"metrics": {
"parameters": 718221
}
},
...
],
"summary": {
"total": 5,
"passed": 5,
"failed": 0,
"pass_rate": 100.0
}
}
```
---
## Testing Strategy
### Progressive Validation
```
Level 1: Smoke Tests (5 min)
↓
"Code runs, model works"
↓
Level 2: Physics Tests (15 min)
↓
"Understands physics constraints"
↓
Level 3: Learning Tests (30 min)
↓
"Can learn patterns"
↓
Level 4: Integration Tests (1 hour)
↓
"Production ready"
```
### Development Workflow
```
1. Write code
2. Run: python test_suite.py --quick (30s)
3. If pass → Continue
If fail → Fix immediately
4. Before commit: python test_suite.py --full (1h)
5. All pass → Commit
```
### Training Validation
```
Before training:
- All smoke tests pass
- Physics tests show correct structure
During training:
- Monitor loss curves
- Check physics residuals
After training:
- All physics tests < 5% error
- Learning tests show convergence
- Integration tests < 10% prediction error
```
---
## Test Coverage
### What's Tested
✅ **Architecture:**
- Model instantiation
- Layer connectivity
- Parameter counts
- Forward pass
✅ **Loss Functions:**
- MSE loss
- Relative loss
- Physics-informed loss
- Max error loss
✅ **Data Pipeline:**
- BDF/OP2 parsing
- Graph construction
- Feature engineering
- Batch processing
✅ **Physics Compliance:**
- Equilibrium (∇·σ + f = 0)
- Constitutive law (σ = C:ε)
- Boundary conditions
- Energy conservation
✅ **Learning Capability:**
- Memorization
- Interpolation
- Extrapolation
- Pattern recognition
✅ **Performance:**
- Inference speed
- Batch processing
- Memory usage
- Scalability
---
## Running the Tests
### Environment Setup
**Note:** There is currently a NumPy compatibility issue on Windows with MINGW-W64 that causes segmentation faults. Tests are ready to run once this environment issue is resolved.
**Options:**
1. Use conda environment with proper NumPy build
2. Use WSL (Windows Subsystem for Linux)
3. Run on Linux system
4. Wait for NumPy Windows compatibility fix
### Quick Start (Once Environment Fixed)
```bash
# 1. Quick smoke test (30 seconds)
python test_suite.py --quick
# 2. Test with Simple Beam
python test_simple_beam.py
# 3. Physics validation
python test_suite.py --physics
# 4. Complete validation
python test_suite.py --full
```
### Individual Test Modules
```bash
# Run specific test suites
python tests/test_synthetic.py # 5 smoke tests
python tests/test_physics.py # 4 physics tests
python tests/test_learning.py # 4 learning tests
python tests/test_predictions.py # 5 integration tests
# Run analytical case examples
python tests/analytical_cases.py # See all analytical solutions
```
---
## Success Criteria
### Minimum Viable Testing (Pre-Training)
- ✅ All smoke tests pass
- ✅ Physics tests run (may not pass without training)
- ✅ Learning tests demonstrate convergence
- ⏳ Simple Beam parses successfully
### Production Ready (Post-Training)
- ✅ All smoke tests pass
- ⏳ Physics tests < 5% error
- ⏳ Learning tests show interpolation < 5% error
- ⏳ Integration tests < 10% prediction error
- ⏳ Performance: 1000× speedup vs FEA
---
## Implementation Status
### Completed ✅
1. Master test orchestrator (test_suite.py)
2. Smoke tests (test_synthetic.py) - 5 tests
3. Physics tests (test_physics.py) - 4 tests
4. Learning tests (test_learning.py) - 4 tests
5. Integration tests (test_predictions.py) - 5 tests
6. Analytical solutions library (analytical_cases.py) - 5 cases
7. Simple Beam test (test_simple_beam.py) - 7 steps
8. Documentation and examples
### Total Test Count: 18 tests + 7-step integration test
---
## Next Steps
### To Run Tests:
1. **Resolve NumPy environment issue**
- Use conda: `conda install numpy`
- Or use WSL/Linux
- Or wait for Windows NumPy fix
2. **Run smoke tests**
```bash
python test_suite.py --quick
```
3. **Test with Simple Beam**
```bash
python test_simple_beam.py
```
4. **Generate training data**
- Create multiple design variations
- Run FEA on each
- Parse all cases
5. **Train model**
```bash
python train.py --config training_config.yaml
```
6. **Validate trained model**
```bash
python test_suite.py --full
```
---
## File Summary
| File | Lines | Purpose | Status |
|------|-------|---------|--------|
| test_suite.py | 403 | Master orchestrator | ✅ Complete |
| test_simple_beam.py | 377 | Simple Beam test | ✅ Complete |
| tests/test_synthetic.py | 297 | Smoke tests | ✅ Complete |
| tests/test_physics.py | 370 | Physics validation | ✅ Complete |
| tests/test_learning.py | 410 | Learning tests | ✅ Complete |
| tests/test_predictions.py | 400 | Integration tests | ✅ Complete |
| tests/analytical_cases.py | 450 | Analytical library | ✅ Complete |
**Total:** ~2,700 lines of comprehensive testing infrastructure
---
## Testing Philosophy
### Fast Feedback
- Smoke tests in 30 seconds
- Catch errors immediately
- Continuous validation during development
### Comprehensive Coverage
- From basic functionality to full pipeline
- Physics compliance verification
- Learning capability confirmation
- Performance benchmarking
### Progressive Confidence
```
Code runs → Understands physics → Learns patterns → Production ready
```
### Automated Validation
- JSON results export
- Clear pass/fail reporting
- Metrics tracking
- Duration monitoring
---
## Conclusion
**The complete testing framework is implemented and ready for use.**
**What's Ready:**
- 18 comprehensive tests across 4 test suites
- Analytical solutions library with 5 classical cases
- Master orchestrator with 4 testing modes
- Simple Beam integration test
- Detailed documentation and examples
**To Use:**
1. Resolve NumPy environment issue
2. Run: `python test_suite.py --quick`
3. Validate: All smoke tests should pass
4. Proceed with training and full validation
**The testing framework provides complete validation from zero to production-ready neural FEA predictions!** ✅
---
*AtomizerField Testing Framework v1.0 - Complete Implementation*
*Total: 18 tests + analytical library + integration test*
*Ready for immediate use once environment is configured*

422
atomizer-field/TESTING_FRAMEWORK_SUMMARY.md
View File

@@ -1,422 +0,0 @@
# AtomizerField Testing Framework - Implementation Summary
## 🎯 Testing Framework Created
I've implemented a comprehensive testing framework for AtomizerField that validates everything from basic functionality to full neural FEA predictions.
---
## ✅ Files Created
### 1. **test_suite.py** - Master Test Orchestrator
**Status:** ✅ Complete
**Features:**
- Four testing modes: `--quick`, `--physics`, `--learning`, `--full`
- Progress tracking and detailed reporting
- JSON results export
- Clean pass/fail output
**Usage:**
```bash
# Quick smoke tests (5 minutes)
python test_suite.py --quick
# Physics validation (15 minutes)
python test_suite.py --physics
# Learning tests (30 minutes)
python test_suite.py --learning
# Full suite (1 hour)
python test_suite.py --full
```
### 2. **tests/test_synthetic.py** - Synthetic Tests
**Status:** ✅ Complete
**Tests Implemented:**
1. ✅ Model Creation - Verify GNN instantiates
2. ✅ Forward Pass - Model processes data
3. ✅ Loss Computation - All loss functions work
4. ✅ Batch Processing - Handle multiple graphs
5. ✅ Gradient Flow - Backpropagation works
**Can run standalone:**
```bash
python tests/test_synthetic.py
```
---
## 📋 Testing Strategy
### Phase 1: Smoke Tests (5 min) ✅ Implemented
```
✓ Model creation (718K parameters)
✓ Forward pass (displacement, stress, von Mises)
✓ Loss computation (MSE, relative, physics, max)
✓ Batch processing
✓ Gradient flow
```
### Phase 2: Physics Tests (15 min) ⏳ Spec Ready
```
- Cantilever beam (δ = FL³/3EI)
- Simply supported beam
- Pressure vessel (σ = pr/t)
- Equilibrium check (∇·σ + f = 0)
- Energy conservation
```
### Phase 3: Learning Tests (30 min) ⏳ Spec Ready
```
- Memorization (10 examples)
- Interpolation (between training points)
- Extrapolation (beyond training data)
- Pattern recognition (thickness → stress)
```
### Phase 4: Integration Tests (1 hour) ⏳ Spec Ready
```
- Parser validation
- Training pipeline
- Prediction accuracy
- Performance benchmarks
```
---
## 🧪 Test Results Format
### Example Output:
```
============================================================
AtomizerField Test Suite v1.0
Mode: QUICK
============================================================
[TEST] Model Creation
Description: Verify GNN model can be instantiated
Creating GNN model...
Model created: 718,221 parameters
Status: ✓ PASS
Duration: 0.15s
[TEST] Forward Pass
Description: Verify model can process dummy data
Testing forward pass...
Displacement shape: (100, 6) ✓
Stress shape: (100, 6) ✓
Von Mises shape: (100,) ✓
Status: ✓ PASS
Duration: 0.05s
[TEST] Loss Computation
Description: Verify loss functions work
Testing loss functions...
MSE loss: 3.885789 ✓
RELATIVE loss: 2.941448 ✓
PHYSICS loss: 3.850585 ✓
MAX loss: 20.127707 ✓
Status: ✓ PASS
Duration: 0.12s
============================================================
TEST SUMMARY
============================================================
Total Tests: 5
✓ Passed: 5
✗ Failed: 0
Pass Rate: 100.0%
✓ ALL TESTS PASSED - SYSTEM READY!
============================================================
Total testing time: 0.5 minutes
Results saved to: test_results/test_results_quick_1234567890.json
```
---
## 📁 Directory Structure
```
Atomizer-Field/
├── test_suite.py # ✅ Master orchestrator
├── tests/
│ ├── __init__.py # ✅ Package init
│ ├── test_synthetic.py # ✅ Synthetic tests (COMPLETE)
│ ├── test_physics.py # ⏳ Physics validation (NEXT)
│ ├── test_learning.py # ⏳ Learning tests
│ ├── test_predictions.py # ⏳ Integration tests
│ └── analytical_cases.py # ⏳ Known solutions
│
├── generate_test_data.py # ⏳ Test data generator
├── benchmark.py # ⏳ Performance tests
├── visualize_results.py # ⏳ Visualization
├── test_dashboard.py # ⏳ HTML report generator
│
└── test_results/ # Auto-created
├── test_results_quick_*.json
├── test_results_full_*.json
└── test_report.html
```
---
## 🚀 Quick Start Testing
### Step 1: Run Smoke Tests (Immediate)
```bash
# Verify basic functionality (5 minutes)
python test_suite.py --quick
```
**Expected Output:**
```
5/5 tests passed
✓ ALL TESTS PASSED - SYSTEM READY!
```
### Step 2: Generate Test Data (When Ready)
```bash
# Create synthetic FEA data with known solutions
python generate_test_data.py --all-cases
```
### Step 3: Full Validation (When Model Trained)
```bash
# Complete test suite (1 hour)
python test_suite.py --full
```
---
## 📊 What Each Test Validates
### Smoke Tests (test_synthetic.py) ✅
**Purpose:** Verify code runs without errors
| Test | What It Checks | Why It Matters |
|------|----------------|----------------|
| Model Creation | Can instantiate GNN | Code imports work, architecture valid |
| Forward Pass | Produces outputs | Model can process data |
| Loss Computation | All loss types work | Training will work |
| Batch Processing | Handles multiple graphs | Real training scenario |
| Gradient Flow | Backprop works | Model can learn |
### Physics Tests (test_physics.py) ⏳
**Purpose:** Validate physics understanding
| Test | Known Solution | Tolerance |
|------|---------------|-----------|
| Cantilever Beam | δ = FL³/3EI | < 5% |
| Simply Supported | δ = FL³/48EI | < 5% |
| Pressure Vessel | σ = pr/t | < 5% |
| Equilibrium | ∇·σ + f = 0 | < 1e-6 |
### Learning Tests (test_learning.py) ⏳
**Purpose:** Confirm network learns
| Test | Dataset | Expected Result |
|------|---------|-----------------|
| Memorization | 10 samples | < 1% error |
| Interpolation | Train: [1,3,5], Test: [2,4] | < 5% error |
| Extrapolation | Train: [1-3], Test: [5] | < 20% error |
| Pattern | thickness ↑ → stress ↓ | Correct trend |
### Integration Tests (test_predictions.py) ⏳
**Purpose:** Full system validation
| Test | Input | Output |
|------|-------|--------|
| Parser | Simple Beam BDF/OP2 | Parsed data |
| Training | 50 cases, 20 epochs | Trained model |
| Prediction | New design | Stress/disp fields |
| Accuracy | Compare vs FEA | < 10% error |
---
## 🎯 Next Steps
### To Complete Testing Framework:
**Priority 1: Physics Tests** (30 min implementation)
```python
# tests/test_physics.py
def test_cantilever_analytical():
"""Compare with δ = FL³/3EI"""
# Generate cantilever mesh
# Predict displacement
# Compare with analytical
pass
```
**Priority 2: Test Data Generator** (1 hour)
```python
# generate_test_data.py
class SyntheticFEAGenerator:
"""Create fake but realistic FEA data"""
def generate_cantilever_dataset(n_samples=100):
# Generate meshes with varying parameters
# Calculate analytical solutions
pass
```
**Priority 3: Learning Tests** (30 min)
```python
# tests/test_learning.py
def test_memorization():
"""Can network memorize 10 examples?"""
pass
```
**Priority 4: Visualization** (1 hour)
```python
# visualize_results.py
def plot_test_results():
"""Create plots comparing predictions vs truth"""
pass
```
**Priority 5: HTML Dashboard** (1 hour)
```python
# test_dashboard.py
def generate_html_report():
"""Create comprehensive HTML report"""
pass
```
---
## 📈 Success Criteria
### Minimum Viable Testing:
- ✅ Smoke tests pass (basic functionality)
- ⏳ At least one physics test passes (analytical validation)
- ⏳ Network can memorize small dataset (learning proof)
### Production Ready:
- All smoke tests pass ✅
- All physics tests < 5% error
- Learning tests show convergence
- Integration tests < 10% prediction error
- Performance benchmarks meet targets (1000× speedup)
---
## 🔧 How to Extend
### Adding New Test:
```python
# tests/test_custom.py
def test_my_feature():
"""
Test custom feature
Expected: Feature works correctly
"""
# Setup
# Execute
# Validate
return {
'status': 'PASS' if success else 'FAIL',
'message': 'Test completed',
'metrics': {'accuracy': 0.95}
}
```
### Register in test_suite.py:
```python
def run_custom_tests(self):
from tests import test_custom
self.run_test(
"My Feature Test",
test_custom.test_my_feature,
"Verify my feature works"
)
```
---
## 🎓 Testing Philosophy
### Progressive Confidence:
```
Level 1: Smoke Tests → "Code runs"
Level 2: Physics Tests → "Understands physics"
Level 3: Learning Tests → "Can learn patterns"
Level 4: Integration Tests → "Production ready"
```
### Fast Feedback Loop:
```
Developer writes code
↓
Run smoke tests (30 seconds)
↓
If pass → Continue development
If fail → Fix immediately
```
### Comprehensive Validation:
```
Before deployment:
↓
Run full test suite (1 hour)
↓
All tests pass → Deploy
Any test fails → Fix and retest
```
---
## 📚 Resources
**Current Implementation:**
- ✅ `test_suite.py` - Master orchestrator
- ✅ `tests/test_synthetic.py` - 5 smoke tests
**Documentation:**
- Example outputs provided
- Clear usage instructions
- Extension guide included
**Next To Implement:**
- Physics tests with analytical solutions
- Learning capability tests
- Integration tests
- Visualization tools
- HTML dashboard
---
## 🎉 Summary
**Status:** Testing framework foundation complete ✅
**Implemented:**
- Master test orchestrator with 4 modes
- 5 comprehensive smoke tests
- Clean reporting system
- JSON results export
- Extensible architecture
**Ready To:**
1. Run smoke tests immediately (`python test_suite.py --quick`)
2. Verify basic functionality
3. Add physics tests as needed
4. Expand to full validation
**Testing framework is production-ready for incremental expansion!** 🚀
---
*Testing Framework v1.0 - Comprehensive validation from zero to neural FEA*

299
atomizer-field/UNIT_CONVERSION_REPORT.md
View File

@@ -1,299 +0,0 @@
# Unit Conversion Issue - Analysis and Fix
**Date:** November 24, 2025
**Issue:** Stresses displaying 1000× too large
---
## Root Cause Identified
### BDF File Unit System
The BDF file contains: **`PARAM UNITSYS MN-MM`**
This defines the Nastran unit system as:
- **Length:** mm (millimeter)
- **Force:** MN (MegaNewton) = 1,000,000 N
- **Mass:** tonne (1000 kg)
- **Stress:** Pa (Pascal) = N/m² *[NOT MPa!]*
- **Energy:** MN-mm = 1,000 N-m = 1 kJ
### Material Properties Confirm This
Young's modulus from BDF: **E = 200,000,000**
- If units were MPa: E = 200 GPa (way too high for steel ~200 GPa)
- If units are Pa: E = 200 MPa (way too low!)
- **Actual: E = 200,000,000 Pa = 200 GPa** ✓ (correct for steel)
### What pyNastran Returns
pyNastran reads the OP2 file and returns data **in the same units as the BDF**:
- Displacement: mm ✓
- Force/Reactions: **MN** (not N!)
- Stress: **Pa** (not MPa!)
---
## Current vs Actual Values
### Stress Values
| What we claimed | Actual value | Correct interpretation |
|----------------|--------------|------------------------|
| 117,000 MPa | 117,000 Pa | **117 kPa = 0.117 MPa** ✓ |
| 46,000 MPa (mean) | 46,000 Pa | **46 kPa = 0.046 MPa** ✓ |
**Correct stress values are 1000× smaller!**
### Force Values
| What we claimed | Actual value | Correct interpretation |
|----------------|--------------|------------------------|
| 2.73 MN (applied) | 2.73 MN | **2.73 MN = 2,730,000 N** ✓ |
| 150 MN (reaction) | 150 MN | **150 MN = 150,000,000 N** ✓ |
**Force values are correctly stored, but labeled as N instead of MN**
---
## Impact
### What's Wrong:
1. **Stress units incorrectly labeled as "MPa"** - should be "Pa"
2. **Force/reaction units incorrectly labeled as "N"** - should be "MN"
3. **Visualization shows stress 1000× too high**
4. **Reports show unrealistic values** (117 GPa stress would destroy steel!)
### What's Correct:
1. ✅ Displacement values (19.5 mm)
2. ✅ Material properties (E = 200 GPa)
3. ✅ Geometry (mm)
4. ✅ Actual numerical values from pyNastran
---
## Solution
### Option 1: Convert to Standard Units (Recommended)
Convert all data to consistent engineering units:
- Length: mm → mm ✓
- Force: MN → **N** (divide by 1e6)
- Stress: Pa → **MPa** (divide by 1e6)
- Mass: tonne → kg (multiply by 1000)
**Benefits:**
- Standard engineering units (mm, N, MPa, kg)
- Matches what users expect
- No confusion in reports/visualization
**Changes Required:**
- Parser: Convert forces (divide by 1e6)
- Parser: Convert stress (divide by 1e6)
- Update metadata to reflect actual units
### Option 2: Use Native Units (Not Recommended)
Keep MN-MM-tonne-Pa system throughout
**Issues:**
- Non-standard units confuse users
- Harder to interpret values
- Requires careful labeling everywhere
---
## Implementation Plan
### 1. Fix Parser ([neural_field_parser.py](neural_field_parser.py))
**Lines to modify:**
#### Stress Extraction (~line 602-648)
```python
# CURRENT (wrong):
stress_data = stress.data[0, :, :]
stress_results[f"{elem_type}_stress"] = {
"data": stress_data.tolist(),
"units": "MPa" # WRONG!
}
# FIX:
stress_data = stress.data[0, :, :] / 1e6 # Convert Pa → MPa
stress_results[f"{elem_type}_stress"] = {
"data": stress_data.tolist(),
"units": "MPa" # Now correct!
}
```
#### Force Extraction (~line 464-507)
```python
# CURRENT (partially wrong):
"magnitude": float(load.mag), # This is in MN, not N!
# FIX:
"magnitude": float(load.mag) * 1e6, # Convert MN → N
```
#### Reaction Forces (~line 538-568)
```python
# CURRENT (wrong):
reactions = grid_point_force.data[0] # In MN!
# FIX:
reactions = grid_point_force.data[0] * 1e6 # Convert MN → N
```
### 2. Update Unit Detection
Add UNITSYS parameter detection:
```python
def detect_units(self):
"""Detect Nastran unit system from PARAM cards"""
if hasattr(self.bdf, 'params') and 'UNITSYS' in self.bdf.params:
unitsys = str(self.bdf.params['UNITSYS'].values[0])
if 'MN' in unitsys:
return {
'length': 'mm',
'force': 'MN',
'stress': 'Pa',
'mass': 'tonne',
'needs_conversion': True
}
# Default units
return {
'length': 'mm',
'force': 'N',
'stress': 'MPa',
'mass': 'kg',
'needs_conversion': False
}
```
### 3. Add Unit Conversion Function
```python
def convert_to_standard_units(self, data, unit_system):
"""Convert from Nastran units to standard engineering units"""
if not unit_system['needs_conversion']:
return data
# Convert forces: MN → N (multiply by 1e6)
if 'loads' in data:
for force in data['loads']['point_forces']:
force['magnitude'] *= 1e6
# Convert stress: Pa → MPa (divide by 1e6)
if 'results' in data and 'stress' in data['results']:
for stress_type, stress_data in data['results']['stress'].items():
if isinstance(stress_data, dict) and 'data' in stress_data:
stress_data['data'] = np.array(stress_data['data']) / 1e6
stress_data['units'] = 'MPa'
# Convert reactions: MN → N (multiply by 1e6)
# (Handle in HDF5 write)
return data
```
### 4. Update HDF5 Writing
Apply conversions when writing to HDF5:
```python
# Reactions
if 'reactions' in self.neural_field_data['results']:
reactions_data = np.array(self.neural_field_data['results']['reactions']['data'])
if unit_system['force'] == 'MN':
reactions_data *= 1e6 # MN → N
hf.create_dataset('results/reactions', data=reactions_data)
```
---
## Testing Plan
### 1. Create Unit Conversion Test
```python
def test_unit_conversion():
"""Verify units are correctly converted"""
parser = NastranToNeuralFieldParser('test_case_beam')
data = parser.parse_all()
# Check stress units
stress = data['results']['stress']['cquad4_stress']
assert stress['units'] == 'MPa'
max_stress = np.max(stress['data'][:, -1]) # Von Mises
assert max_stress < 500, f"Stress {max_stress} MPa too high!"
# Check force units
force = data['loads']['point_forces'][0]
assert force['magnitude'] < 1e7, "Force should be in N"
print("[OK] Units correctly converted")
```
### 2. Expected Values After Fix
| Property | Before (wrong) | After (correct) |
|----------|---------------|-----------------|
| Max stress | 117,000 MPa | **117 MPa** ✓ |
| Mean stress | 46,000 MPa | **46 MPa** ✓ |
| Applied force | 2.73 MN | **2,730,000 N** |
| Max reaction | 150 MN | **150,000,000 N** |
### 3. Validation Checks
- ✓ Stress < 500 MPa (reasonable for steel)
- ✓ Force magnitude matches applied loads
- ✓ Material E = 200 GPa (correct for steel)
- ✓ Displacement still 19.5 mm
---
## Risk Assessment
### Low Risk:
- ✅ Only affects numerical scaling
- ✅ No changes to data structure
- ✅ Easy to verify with test
- ✅ Can be fixed with multiplication/division
### What Could Go Wrong:
- ⚠ Other BDF files might use different UNITSYS
- ⚠ Some files might already be in correct units
- ⚠ Need to handle multiple unit systems
### Mitigation:
- Always check PARAM UNITSYS first
- Add unit system detection
- Log conversions clearly
- Add validation checks
---
## Recommendations
### Immediate Actions:
1. ✅ **Update parser to detect UNITSYS**
2. ✅ **Add unit conversion for stress (Pa → MPa)**
3. ✅ **Add unit conversion for forces (MN → N)**
4. ✅ **Update metadata to reflect conversions**
5. ✅ **Add validation checks**
### Long-term:
- Support multiple Nastran unit systems
- Add unit conversion utilities
- Document unit assumptions clearly
- Add warnings for unusual values
---
## Conclusion
**The system is working correctly** - pyNastran is reading the data accurately.
**The issue is labeling** - we incorrectly assumed MPa when Nastran uses Pa.
**The fix is simple** - divide stress by 1e6, multiply forces by 1e6, update labels.
**After fix:**
- Stress: 117 MPa (reasonable for steel) ✓
- Force: 2.73 MN = 2,730 kN (reasonable for large beam) ✓
- All other values unchanged ✓
**System will be production-ready after this fix!** 🚀
---
*Unit Conversion Analysis v1.0*
*Issue: 1000× stress error*
*Root cause: MN-MM unit system misinterpretation*
*Fix: Scale factors + label corrections*

147
atomizer-field/UNIT_INVESTIGATION_SUMMARY.md
View File

@@ -1,147 +0,0 @@
# Unit Investigation Summary
## Your Question
> "Force and stresses seems to be 1000 too much, how do you check units and validate values?"
## Answer
You were **absolutely correct!** The stresses ARE 1000× too large, but the forces are actually correct (just mislabeled).
---
## Root Cause Found
Your BDF file contains: **`PARAM UNITSYS MN-MM`**
This tells Nastran to use the MegaNewton-Millimeter unit system:
- Length: mm ✓
- Force: **MN (MegaNewton)** = 1,000,000 N
- Stress: **Pa (Pascal)**, NOT MPa!
- Mass: tonne (1000 kg)
### What This Means
**pyNastran correctly reads the OP2 file in these units**, but my parser incorrectly assumed:
- Force in N (actually MN)
- Stress in MPa (actually Pa)
---
## Actual Values
### Stress (The 1000× Error You Found)
| What Report Shows | Actual Unit | Correct Value |
|-------------------|-------------|---------------|
| 117,000 MPa | 117,000 Pa | **117 MPa** ✓ |
| 46,000 MPa (mean) | 46,000 Pa | **46 MPa** ✓ |
**Your stresses are 1000× too high because Pa should be divided by 1000 to get kPa, or by 1,000,000 to get MPa.**
### Forces (Correctly Stored, Mislabeled)
| What Report Shows | Actual Unit | Interpretation |
|-------------------|-------------|----------------|
| 2.73 MN | MN ✓ | 2,730,000 N |
| 150 MN | MN ✓ | 150,000,000 N |
Forces are actually correct! They're in MegaNewtons, which is perfectly fine for a large beam structure.
---
## How I Validated This
### 1. Checked the BDF File
Found `PARAM UNITSYS MN-MM` which defines the unit system.
### 2. Checked Material Properties
Young's modulus E = 200,000,000
- If this were MPa → E = 200 GPa ✓ (correct for steel)
- This confirms stress is in Pa (base SI unit)
### 3. Direct OP2 Reading
Created [check_op2_units.py](check_op2_units.py) to directly read the OP2 file with pyNastran:
- Confirmed pyNastran doesn't specify units
- Confirmed stress values: min=1.87e+03, max=1.17e+05
- These are clearly in Pa, not MPa!
### 4. Sanity Check
A 117 GPa von Mises stress would **instantly destroy any material** (even diamond is ~130 GPa).
117 MPa is reasonable for a loaded steel beam ✓
---
## The Fix
### What Needs to Change
**In [neural_field_parser.py](neural_field_parser.py:602-648):**
1. **Detect UNITSYS parameter from BDF**
2. **Convert stress: Pa → MPa** (divide by 1e6)
3. **Update force labels: MN → N** (or keep as MN with correct label)
4. **Add validation checks** to catch unrealistic values
### Conversion Factors
```python
# If UNITSYS is MN-MM:
stress_MPa = stress_Pa / 1e6
force_N = force_MN * 1e6
mass_kg = mass_tonne * 1000
```
---
## Expected Values After Fix
| Property | Current (Wrong) | After Fix | Reasonable? |
|----------|----------------|-----------|-------------|
| Max von Mises | 117,000 MPa | **117 MPa** | ✓ Yes (steel ~250 MPa yield) |
| Mean von Mises | 46,000 MPa | **46 MPa** | ✓ Yes |
| Max displacement | 19.5 mm | 19.5 mm | ✓ Yes |
| Applied forces | 2.73 MN | 2.73 MN | ✓ Yes (large beam) |
| Young's modulus | 200 GPa | 200 GPa | ✓ Yes (steel) |
---
## Files Created for Investigation
1. **[check_units.py](check_units.py)** - Analyzes parsed data for unit consistency
2. **[check_op2_units.py](check_op2_units.py)** - Directly reads OP2/BDF to verify units
3. **[UNIT_CONVERSION_REPORT.md](UNIT_CONVERSION_REPORT.md)** - Complete analysis and fix plan
---
## Next Steps
### Option 1: I Fix It Now
I can update the parser to:
1. Detect UNITSYS parameter
2. Convert Pa → MPa for stress
3. Add unit validation
4. Re-run test and regenerate report
**Time:** 15-20 minutes
**Risk:** Low (just scaling factors)
### Option 2: You Review First
You can review the [UNIT_CONVERSION_REPORT.md](UNIT_CONVERSION_REPORT.md) for the detailed fix plan, then I implement.
**Advantage:** You understand the changes before they're made
---
## Bottom Line
**Your intuition was spot-on!** The stresses displayed are 1000× too high.
**Root cause:** Nastran uses Pa (not MPa) in the MN-MM unit system, and my parser mislabeled them.
**Fix:** Simple scaling factors (divide by 1e6) and correct labels.
**After fix:** All values will be realistic and match engineering expectations! ✓
---
What would you like me to do next?
1. Implement the unit conversion fix?
2. Answer any questions about the analysis?
3. Something else?

239
atomizer-field/atomizer_field_config.yaml
View File

@@ -1,239 +0,0 @@
# AtomizerField Configuration
# Long-term vision configuration for neural field learning
# ============================================================================
# Model Architecture
# ============================================================================
model:
type: "graph_neural_network"
architecture: "message_passing"
# Foundation model settings (for transfer learning)
foundation:
enabled: false # Set to true when foundation model available
path: "models/physics_foundation_v1.pt"
freeze: true # Freeze foundation layers during fine-tuning
# Adaptation layers (for fine-tuning on new component types)
adaptation:
layers: 2
neurons: 128
dropout: 0.1
# Core GNN parameters
gnn:
node_feature_dim: 12 # [x,y,z, BC(6), loads(3)]
edge_feature_dim: 5 # [E, nu, rho, G, alpha]
hidden_dim: 128
num_layers: 6
dropout: 0.1
# Output decoders
decoders:
displacement:
enabled: true
output_dim: 6 # [ux, uy, uz, rx, ry, rz]
stress:
enabled: true
output_dim: 6 # [sxx, syy, szz, txy, tyz, txz]
# ============================================================================
# Training Configuration
# ============================================================================
training:
# Progressive training (coarse to fine meshes)
progressive:
enabled: false # Enable for multi-resolution training
stages:
- resolution: "coarse"
max_nodes: 5000
epochs: 20
lr: 0.001
- resolution: "medium"
max_nodes: 20000
epochs: 10
lr: 0.0005
- resolution: "fine"
max_nodes: 100000
epochs: 5
lr: 0.0001
# Online learning (during optimization)
online:
enabled: false # Enable to learn from FEA during optimization
update_frequency: 10 # Update model every N FEA runs
quick_update_steps: 10
learning_rate: 0.0001
# Physics-informed loss weights
loss:
type: "physics" # Options: mse, relative, physics, max
weights:
data: 1.0 # Match FEA results
equilibrium: 0.1 # ∇·σ + f = 0
constitutive: 0.1 # σ = C:ε
boundary: 1.0 # u = 0 at fixed nodes
# Standard training parameters
hyperparameters:
epochs: 100
batch_size: 4
learning_rate: 0.001
weight_decay: 0.00001
# Optimization
optimizer:
type: "AdamW"
betas: [0.9, 0.999]
scheduler:
type: "ReduceLROnPlateau"
factor: 0.5
patience: 10
# Early stopping
early_stopping:
enabled: true
patience: 50
min_delta: 0.0001
# ============================================================================
# Data Pipeline
# ============================================================================
data:
# Data normalization
normalization:
enabled: true
method: "standard" # Options: standard, minmax
# Data augmentation
augmentation:
enabled: false # Enable for data augmentation
techniques:
- rotation # Rotate mesh randomly
- scaling # Scale loads
- noise # Add small noise to inputs
# Multi-resolution support
multi_resolution:
enabled: false
resolutions: ["coarse", "medium", "fine"]
# Caching
cache:
in_memory: false # Cache dataset in RAM (faster but memory-intensive)
disk_cache: true # Cache preprocessed graphs to disk
# ============================================================================
# Optimization Interface
# ============================================================================
optimization:
# Gradient-based optimization
use_gradients: true
# Uncertainty quantification
uncertainty:
enabled: false # Enable ensemble for uncertainty
ensemble_size: 5
threshold: 0.1 # Recommend FEA if uncertainty > threshold
# FEA fallback
fallback_to_fea:
enabled: true
conditions:
- high_uncertainty # Uncertainty > threshold
- extrapolation # Outside training distribution
- critical_design # Final validation
# Batch evaluation
batch_size: 100 # Evaluate designs in batches for speed
# ============================================================================
# Model Versioning & Deployment
# ============================================================================
deployment:
# Model versioning
versioning:
enabled: true
format: "semantic" # v1.0.0, v1.1.0, etc.
# Model registry
registry:
path: "models/"
naming: "{component_type}_v{version}.pt"
# Metadata tracking
metadata:
track_training_data: true
track_performance: true
track_hyperparameters: true
# Production settings
production:
device: "cuda" # cuda or cpu
batch_inference: true
max_batch_size: 100
# ============================================================================
# Integration with Atomizer
# ============================================================================
atomizer_integration:
# Dashboard integration
dashboard:
enabled: false # Future: Show field visualizations in dashboard
# Database integration
database:
enabled: false # Future: Store predictions in Atomizer DB
# API endpoints
api:
enabled: false # Future: REST API for predictions
port: 8000
# ============================================================================
# Monitoring & Logging
# ============================================================================
monitoring:
# TensorBoard
tensorboard:
enabled: true
log_dir: "runs/tensorboard"
# Weights & Biases (optional)
wandb:
enabled: false
project: "atomizerfield"
entity: "your_team"
# Logging level
logging:
level: "INFO" # DEBUG, INFO, WARNING, ERROR
file: "logs/atomizerfield.log"
# ============================================================================
# Experimental Features
# ============================================================================
experimental:
# Nonlinear analysis
nonlinear:
enabled: false
# Contact analysis
contact:
enabled: false
# Composite materials
composites:
enabled: false
# Modal analysis
modal:
enabled: false
# Topology optimization
topology:
enabled: false

360
atomizer-field/batch_parser.py
View File

@@ -1,360 +0,0 @@
"""
batch_parser.py
Parse multiple NX Nastran cases in batch
AtomizerField Batch Parser v1.0.0
Efficiently processes multiple FEA cases for neural network training dataset creation.
Usage:
python batch_parser.py <root_directory>
Example:
python batch_parser.py ./training_data
Directory structure expected:
training_data/
├── case_001/
│ ├── input/model.bdf
│ └── output/model.op2
├── case_002/
│ ├── input/model.bdf
│ └── output/model.op2
└── ...
"""
import os
import sys
import json
from pathlib import Path
from datetime import datetime
import traceback
from neural_field_parser import NastranToNeuralFieldParser
from validate_parsed_data import NeuralFieldDataValidator
class BatchParser:
"""
Batch parser for processing multiple FEA cases
This enables rapid dataset creation for neural network training.
Processes each case sequentially and generates a summary report.
"""
def __init__(self, root_directory, validate=True, continue_on_error=True):
"""
Initialize batch parser
Args:
root_directory (str or Path): Root directory containing case subdirectories
validate (bool): Run validation after parsing each case
continue_on_error (bool): Continue processing if a case fails
"""
self.root_dir = Path(root_directory)
self.validate = validate
self.continue_on_error = continue_on_error
self.results = []
def find_cases(self):
"""
Find all case directories in root directory
A valid case directory contains:
- input/ subdirectory with .bdf or .dat file
- output/ subdirectory with .op2 file
Returns:
list: List of Path objects for valid case directories
"""
cases = []
for item in self.root_dir.iterdir():
if not item.is_dir():
continue
# Check for required subdirectories and files
input_dir = item / "input"
output_dir = item / "output"
if not input_dir.exists() or not output_dir.exists():
continue
# Check for BDF file
bdf_files = list(input_dir.glob("*.bdf")) + list(input_dir.glob("*.dat"))
if not bdf_files:
continue
# Check for OP2 file
op2_files = list(output_dir.glob("*.op2"))
if not op2_files:
continue
cases.append(item)
return sorted(cases)
def parse_case(self, case_dir):
"""
Parse a single case
Args:
case_dir (Path): Path to case directory
Returns:
dict: Result dictionary with status and metadata
"""
result = {
"case": case_dir.name,
"status": "unknown",
"timestamp": datetime.now().isoformat()
}
try:
print(f"\n{'='*60}")
print(f"Processing: {case_dir.name}")
print(f"{'='*60}")
# Parse
parser = NastranToNeuralFieldParser(case_dir)
data = parser.parse_all()
result["status"] = "parsed"
result["nodes"] = data["mesh"]["statistics"]["n_nodes"]
result["elements"] = data["mesh"]["statistics"]["n_elements"]
# Get max displacement and stress if available
if "displacement" in data.get("results", {}):
result["max_displacement"] = data["results"]["displacement"].get("max_translation")
if "stress" in data.get("results", {}):
for stress_type, stress_data in data["results"]["stress"].items():
if "max_von_mises" in stress_data and stress_data["max_von_mises"] is not None:
result["max_stress"] = stress_data["max_von_mises"]
break
# Validate if requested
if self.validate:
print(f"\nValidating {case_dir.name}...")
validator = NeuralFieldDataValidator(case_dir)
validation_passed = validator.validate()
result["validated"] = validation_passed
if validation_passed:
result["status"] = "success"
else:
result["status"] = "validation_failed"
result["message"] = "Validation failed (see output above)"
else:
result["status"] = "success"
except Exception as e:
result["status"] = "failed"
result["error"] = str(e)
result["traceback"] = traceback.format_exc()
print(f"\n✗ ERROR: {e}")
if not self.continue_on_error:
raise
return result
def batch_parse(self):
"""
Parse all cases in root directory
Returns:
list: List of result dictionaries
"""
print("\n" + "="*60)
print("AtomizerField Batch Parser v1.0")
print("="*60)
print(f"\nRoot directory: {self.root_dir}")
# Find all cases
cases = self.find_cases()
if not cases:
print(f"\n✗ No valid cases found in {self.root_dir}")
print("\nCase directories should contain:")
print(" input/model.bdf (or model.dat)")
print(" output/model.op2")
return []
print(f"\nFound {len(cases)} case(s) to process:")
for case in cases:
print(f" - {case.name}")
# Process each case
self.results = []
start_time = datetime.now()
for i, case in enumerate(cases, 1):
print(f"\n[{i}/{len(cases)}] Processing {case.name}...")
result = self.parse_case(case)
self.results.append(result)
# Show progress
success_count = sum(1 for r in self.results if r["status"] == "success")
print(f"\nProgress: {i}/{len(cases)} processed, {success_count} successful")
end_time = datetime.now()
elapsed = (end_time - start_time).total_seconds()
# Print summary
self._print_summary(elapsed)
# Save summary to JSON
self._save_summary()
return self.results
def _print_summary(self, elapsed_time):
"""Print batch processing summary"""
print("\n" + "="*60)
print("BATCH PROCESSING COMPLETE")
print("="*60)
success_count = sum(1 for r in self.results if r["status"] == "success")
failed_count = sum(1 for r in self.results if r["status"] == "failed")
validation_failed = sum(1 for r in self.results if r["status"] == "validation_failed")
print(f"\nTotal cases: {len(self.results)}")
print(f" ✓ Successful: {success_count}")
if validation_failed > 0:
print(f" ⚠ Validation failed: {validation_failed}")
if failed_count > 0:
print(f" ✗ Failed: {failed_count}")
print(f"\nProcessing time: {elapsed_time:.1f} seconds")
if len(self.results) > 0:
print(f"Average time per case: {elapsed_time/len(self.results):.1f} seconds")
# Detailed results
print("\nDetailed Results:")
print("-" * 60)
for result in self.results:
status_symbol = {
"success": "✓",
"failed": "✗",
"validation_failed": "⚠",
"parsed": "•"
}.get(result["status"], "?")
case_info = f"{status_symbol} {result['case']}: {result['status']}"
if "nodes" in result and "elements" in result:
case_info += f" ({result['nodes']:,} nodes, {result['elements']:,} elements)"
if "max_stress" in result:
case_info += f" | Max VM: {result['max_stress']:.2f} MPa"
if result["status"] == "failed" and "error" in result:
case_info += f"\n Error: {result['error']}"
print(f" {case_info}")
print("\n" + "="*60)
if success_count == len(self.results):
print("✓ ALL CASES PROCESSED SUCCESSFULLY")
elif success_count > 0:
print(f"⚠ {success_count}/{len(self.results)} CASES SUCCESSFUL")
else:
print("✗ ALL CASES FAILED")
print("="*60 + "\n")
def _save_summary(self):
"""Save batch processing summary to JSON file"""
summary_file = self.root_dir / "batch_processing_summary.json"
summary = {
"batch_info": {
"root_directory": str(self.root_dir),
"timestamp": datetime.now().isoformat(),
"total_cases": len(self.results),
"successful_cases": sum(1 for r in self.results if r["status"] == "success"),
"failed_cases": sum(1 for r in self.results if r["status"] == "failed"),
"validation_enabled": self.validate
},
"cases": self.results
}
with open(summary_file, 'w') as f:
json.dump(summary, f, indent=2)
print(f"Summary saved to: {summary_file}\n")
def main():
"""
Main entry point for batch parser
"""
if len(sys.argv) < 2:
print("\nAtomizerField Batch Parser v1.0")
print("="*60)
print("\nUsage:")
print(" python batch_parser.py <root_directory> [options]")
print("\nOptions:")
print(" --no-validate Skip validation step")
print(" --stop-on-error Stop processing if a case fails")
print("\nExample:")
print(" python batch_parser.py ./training_data")
print("\nDirectory structure:")
print(" training_data/")
print(" ├── case_001/")
print(" │ ├── input/model.bdf")
print(" │ └── output/model.op2")
print(" ├── case_002/")
print(" │ ├── input/model.bdf")
print(" │ └── output/model.op2")
print(" └── ...")
print()
sys.exit(1)
root_dir = sys.argv[1]
# Parse options
validate = "--no-validate" not in sys.argv
continue_on_error = "--stop-on-error" not in sys.argv
if not Path(root_dir).exists():
print(f"ERROR: Directory not found: {root_dir}")
sys.exit(1)
# Create batch parser
batch_parser = BatchParser(
root_dir,
validate=validate,
continue_on_error=continue_on_error
)
# Process all cases
try:
results = batch_parser.batch_parse()
# Exit with appropriate code
if not results:
sys.exit(1)
success_count = sum(1 for r in results if r["status"] == "success")
if success_count == len(results):
sys.exit(0) # All successful
elif success_count > 0:
sys.exit(2) # Partial success
else:
sys.exit(1) # All failed
except KeyboardInterrupt:
print("\n\nBatch processing interrupted by user")
sys.exit(130)
except Exception as e:
print(f"\n\nFATAL ERROR: {e}")
traceback.print_exc()
sys.exit(1)
if __name__ == "__main__":
main()

174
atomizer-field/check_actual_values.py
View File

@@ -1,174 +0,0 @@
"""
Check actual values from OP2 to understand what's correct
"""
from pyNastran.op2.op2 import OP2
from pyNastran.bdf.bdf import BDF
import numpy as np
print("="*60)
print("CHECKING ACTUAL FEA VALUES")
print("="*60)
# Load OP2
op2_file = 'test_case_beam/output/model.op2'
print(f"\nLoading OP2: {op2_file}")
op2 = OP2()
op2.read_op2(op2_file)
# Load BDF
bdf_file = 'test_case_beam/input/model.bdf'
print(f"Loading BDF: {bdf_file}")
bdf = BDF()
bdf.read_bdf(bdf_file)
print("\n1. UNIT SYSTEM:")
print("-"*60)
if hasattr(bdf, 'params') and 'UNITSYS' in bdf.params:
unitsys = str(bdf.params['UNITSYS'].values[0])
print(f" PARAM UNITSYS: {unitsys}")
if 'MN' in unitsys and 'MM' in unitsys:
print(" This is MegaNewton-Millimeter system:")
print(" - Length: mm")
print(" - Force: MN (MegaNewton)")
print(" - Stress: Pa (base SI)")
print(" - Mass: tonne")
else:
print(" No UNITSYS parameter found")
print("\n2. MATERIAL PROPERTIES:")
print("-"*60)
if hasattr(bdf, 'materials') and bdf.materials:
mat = list(bdf.materials.values())[0]
print(f" Material ID: {mat.mid}")
print(f" Type: {mat.type}")
if hasattr(mat, 'e') and mat.e:
print(f" Young's modulus E: {mat.e:.2e}")
print(f" -> E = {mat.e/1e9:.1f} GPa (if units are Pa)")
print(f" -> E = {mat.e/1e6:.1f} GPa (if units are kPa)")
print(f" -> E = {mat.e/1e3:.1f} GPa (if units are MPa)")
print(f" Steel E ~= 200 GPa, so units must be Pa")
print("\n3. APPLIED FORCES FROM BDF:")
print("-"*60)
total_force = 0
n_forces = 0
if hasattr(bdf, 'loads') and bdf.loads:
for load_id, load_list in bdf.loads.items():
for load in load_list:
if hasattr(load, 'mag'):
print(f" Load ID {load_id}, Node {load.node}: {load.mag:.2e} (unit depends on UNITSYS)")
total_force += abs(load.mag)
n_forces += 1
if n_forces >= 3:
break
if n_forces >= 3:
break
print(f" Total applied force (first 3): {total_force:.2e}")
print(f" In MN: {total_force:.2e} MN")
print(f" In N: {total_force*1e6:.2e} N")
print("\n4. DISPLACEMENT FROM OP2:")
print("-"*60)
if hasattr(op2, 'displacements') and op2.displacements:
for subcase_id, disp in op2.displacements.items():
# Get translation only (DOF 1-3)
translations = disp.data[0, :, :3]
max_trans = np.max(np.abs(translations))
max_idx = np.unravel_index(np.argmax(np.abs(translations)), translations.shape)
print(f" Subcase {subcase_id}:")
print(f" Max translation: {max_trans:.6f} mm")
print(f" Location: node index {max_idx[0]}, DOF {max_idx[1]}")
print("\n5. STRESS FROM OP2 (RAW VALUES):")
print("-"*60)
# Try new API
stress_dict = None
if hasattr(op2, 'op2_results') and hasattr(op2.op2_results, 'stress'):
if hasattr(op2.op2_results.stress, 'cquad4_stress'):
stress_dict = op2.op2_results.stress.cquad4_stress
elif hasattr(op2, 'cquad4_stress'):
try:
stress_dict = op2.cquad4_stress
except:
pass
if stress_dict:
for subcase_id, stress in stress_dict.items():
# Stress data columns depend on element type
# For CQUAD4: typically [fiber_distance, oxx, oyy, txy, angle, major, minor, von_mises]
stress_data = stress.data[0, :, :]
print(f" Subcase {subcase_id}:")
print(f" Data shape: {stress_data.shape} (elements × stress_components)")
print(f" Stress components: {stress_data.shape[1]}")
# Von Mises is usually last column
von_mises = stress_data[:, -1]
print(f"\n Von Mises stress (column {stress_data.shape[1]-1}):")
print(f" Min: {np.min(von_mises):.2e}")
print(f" Max: {np.max(von_mises):.2e}")
print(f" Mean: {np.mean(von_mises):.2e}")
print(f" Median: {np.median(von_mises):.2e}")
print(f"\n Principal stresses (columns 5-6 typically):")
if stress_data.shape[1] >= 7:
major = stress_data[:, -3]
minor = stress_data[:, -2]
print(f" Major max: {np.max(major):.2e}")
print(f" Minor min: {np.min(minor):.2e}")
print(f"\n Direct stresses (columns 1-2 typically):")
if stress_data.shape[1] >= 3:
sxx = stress_data[:, 1]
syy = stress_data[:, 2]
print(f" sigmaxx range: {np.min(sxx):.2e} to {np.max(sxx):.2e}")
print(f" sigmayy range: {np.min(syy):.2e} to {np.max(syy):.2e}")
print("\n6. REACTIONS FROM OP2:")
print("-"*60)
if hasattr(op2, 'grid_point_forces') and op2.grid_point_forces:
for subcase_id, gpf in op2.grid_point_forces.items():
forces = gpf.data[0]
print(f" Subcase {subcase_id}:")
print(f" Data shape: {forces.shape}")
print(f" Max reaction (all DOF): {np.max(np.abs(forces)):.2e}")
# Get force components (first 3 columns usually Fx, Fy, Fz)
if forces.shape[1] >= 3:
fx = forces[:, 0]
fy = forces[:, 1]
fz = forces[:, 2]
print(f" Fx range: {np.min(fx):.2e} to {np.max(fx):.2e}")
print(f" Fy range: {np.min(fy):.2e} to {np.max(fy):.2e}")
print(f" Fz range: {np.min(fz):.2e} to {np.max(fz):.2e}")
print("\n7. YOUR STATED VALUES:")
print("-"*60)
print(" You said:")
print(" - Stress around 117 MPa")
print(" - Force around 152,200 N")
print("\n From OP2 raw data above:")
print(" - If Von Mises max = 1.17e+05, this is:")
print(" -> 117,000 Pa = 117 kPa = 0.117 MPa (if UNITSYS=MN-MM)")
print(" -> OR 117,000 MPa (if somehow in MPa already)")
print("\n For force 152,200 N:")
print(" - If reactions max = 1.52e+08 from OP2:")
print(" -> 152,000,000 Pa or N·mm⁻² (if MN-MM system)")
print(" -> 152 MN = 152,000,000 N (conversion)")
print(" -> OR your expected value is 0.1522 MN = 152,200 N")
print("\n8. DIRECTIONS AND TENSORS:")
print("-"*60)
print(" Stress tensor (symmetric 3×3):")
print(" sigma = [sigmaxx tauxy tauxz]")
print(" [tauxy sigmayy tauyz]")
print(" [tauxz tauyz sigmazz]")
print("\n Stored in OP2 for shells (CQUAD4) as:")
print(" [fiber_dist, sigmaxx, sigmayy, tauxy, angle, sigma_major, sigma_minor, von_mises]")
print("\n Displacement vector (6 DOF per node):")
print(" [Tx, Ty, Tz, Rx, Ry, Rz]")
print("\n Force/Reaction vector (6 DOF per node):")
print(" [Fx, Fy, Fz, Mx, My, Mz]")
print("\n" + "="*60)

122
atomizer-field/check_op2_units.py
View File

@@ -1,122 +0,0 @@
"""
Check actual units from pyNastran OP2 reader
"""
from pyNastran.op2.op2 import OP2
import numpy as np
print("="*60)
print("CHECKING OP2 FILE UNITS")
print("="*60)
# Load OP2 file
op2_file = 'test_case_beam/output/model.op2'
print(f"\nLoading: {op2_file}")
op2 = OP2()
op2.read_op2(op2_file)
print("\n1. DISPLACEMENT DATA:")
print("-"*60)
if hasattr(op2, 'displacements') and op2.displacements:
for subcase_id, disp in op2.displacements.items():
print(f" Subcase {subcase_id}:")
print(f" Data shape: {disp.data.shape}")
print(f" Max displacement (all DOF): {np.max(np.abs(disp.data)):.6f}")
print(f" Translation max (DOF 1-3): {np.max(np.abs(disp.data[0, :, :3])):.6f}")
# Check if pyNastran has unit info
if hasattr(disp, 'units'):
print(f" Units: {disp.units}")
else:
print(f" Units: Not specified by pyNastran")
print("\n2. STRESS DATA:")
print("-"*60)
# Try new API first
stress_dict = None
if hasattr(op2, 'op2_results') and hasattr(op2.op2_results, 'stress'):
if hasattr(op2.op2_results.stress, 'cquad4_stress'):
stress_dict = op2.op2_results.stress.cquad4_stress
elif hasattr(op2, 'cquad4_stress'):
try:
stress_dict = op2.cquad4_stress
except:
stress_dict = None
if stress_dict:
for subcase_id, stress in stress_dict.items():
print(f" Subcase {subcase_id}:")
print(f" Data shape: {stress.data.shape}")
# Get von Mises stress (last column)
von_mises = stress.data[0, :, -1]
print(f" Von Mises min: {np.min(von_mises):.2e}")
print(f" Von Mises max: {np.max(von_mises):.2e}")
print(f" Von Mises mean: {np.mean(von_mises):.2e}")
# Check if pyNastran has unit info
if hasattr(stress, 'units'):
print(f" Units: {stress.units}")
else:
print(f" Units: Not specified by pyNastran")
# Check data type names
if hasattr(stress, 'data_names'):
print(f" Data names: {stress.data_names}")
else:
print(" No CQUAD4 stress data found")
print("\n3. REACTION FORCES:")
print("-"*60)
if hasattr(op2, 'grid_point_forces') and op2.grid_point_forces:
for subcase_id, forces in op2.grid_point_forces.items():
print(f" Subcase {subcase_id}:")
print(f" Data shape: {forces.data.shape}")
print(f" Max force: {np.max(np.abs(forces.data)):.2e}")
if hasattr(forces, 'units'):
print(f" Units: {forces.units}")
else:
print(f" Units: Not specified by pyNastran")
print("\n4. BDF FILE UNITS:")
print("-"*60)
# Try to read BDF to check units
from pyNastran.bdf.bdf import BDF
bdf_file = 'test_case_beam/input/model.bdf'
print(f"Loading: {bdf_file}")
bdf = BDF()
bdf.read_bdf(bdf_file)
# Check for PARAM cards that define units
if hasattr(bdf, 'params') and bdf.params:
print(" PARAM cards found:")
for param_name, param in bdf.params.items():
if 'UNIT' in param_name.upper() or 'WTMASS' in param_name.upper():
print(f" {param_name}: {param}")
# Check material properties (can infer units from magnitude)
if hasattr(bdf, 'materials') and bdf.materials:
print("\n Material properties (first material):")
mat = list(bdf.materials.values())[0]
print(f" Type: {mat.type}")
if hasattr(mat, 'e') and mat.e:
print(f" Young's modulus E: {mat.e:.2e}")
print(f" If E~200,000 then units are MPa (steel)")
print(f" If E~200,000,000 then units are Pa (steel)")
print(f" If E~29,000,000 then units are psi (steel)")
print("\n5. NASTRAN UNIT SYSTEM ANALYSIS:")
print("-"*60)
print(" Common Nastran unit systems:")
print(" - SI: m, kg, N, Pa, s")
print(" - mm-tonne-N: mm, tonne, N, MPa, s")
print(" - mm-kg-N: mm, kg, N, MPa, s")
print(" - in-lb-s: in, lb, lbf, psi, s")
print("")
print(" Our metadata claims: mm, kg, N, MPa")
print(" But pyNastran might return stress in Pa (base SI)")
print(" This would explain the 1000× error!")
print("\n" + "="*60)

104
atomizer-field/check_units.py
View File

@@ -1,104 +0,0 @@
"""
Script to investigate unit conversion issues in AtomizerField
"""
import json
import h5py
import numpy as np
print("="*60)
print("UNIT VALIDATION CHECK")
print("="*60)
# Load JSON metadata
with open('test_case_beam/neural_field_data.json', 'r') as f:
data = json.load(f)
# Check units
print("\n1. UNITS FROM METADATA:")
print("-"*60)
units = data['metadata']['units']
for key, value in units.items():
print(f" {key}: {value}")
# Check loads
print("\n2. APPLIED LOADS:")
print("-"*60)
loads = data['loads']
print(f" Available load types: {list(loads.keys())}")
if 'point_forces' in loads:
point_forces = loads['point_forces']
print(f" Point forces: {len(point_forces)} applied")
if point_forces:
print("\n Sample forces (first 3):")
for i in range(min(3, len(point_forces))):
force = point_forces[i]
print(f" Force {i+1}:")
print(f" Node: {force['node']}")
print(f" Magnitude: {force['magnitude']:.2e} N")
print(f" Direction: {force['direction']}")
# Load HDF5 data
print("\n3. REACTION FORCES FROM HDF5:")
print("-"*60)
with h5py.File('test_case_beam/neural_field_data.h5', 'r') as f:
reactions = f['results/reactions'][:]
# Get statistics
non_zero_reactions = reactions[np.abs(reactions) > 1e-10]
print(f" Shape: {reactions.shape}")
print(f" Min: {np.min(reactions):.2e}")
print(f" Max: {np.max(reactions):.2e}")
print(f" Mean (non-zero): {np.mean(np.abs(non_zero_reactions)):.2e}")
print(f" Max absolute: {np.max(np.abs(reactions)):.2e}")
# Check displacement for comparison
displacement = f['results/displacement'][:]
max_disp = np.max(np.abs(displacement[:, :3])) # Translation only
print(f"\n Max displacement: {max_disp:.6f} mm")
# Check stress
print("\n4. STRESS VALUES FROM HDF5:")
print("-"*60)
with h5py.File('test_case_beam/neural_field_data.h5', 'r') as f:
stress = f['results/stress/cquad4_stress/data'][:]
# Von Mises stress is in column 7 (0-indexed)
von_mises = stress[:, 7]
print(f" Shape: {stress.shape}")
print(f" Von Mises min: {np.min(von_mises):.2e} MPa")
print(f" Von Mises max: {np.max(von_mises):.2e} MPa")
print(f" Von Mises mean: {np.mean(von_mises):.2e} MPa")
# Check if values make sense
print("\n5. SANITY CHECK:")
print("-"*60)
print(f" Units claimed: force=N, stress=MPa, length=mm")
print(f" Max reaction force: {np.max(np.abs(reactions)):.2e} N")
print(f" Max von Mises: {np.max(von_mises):.2e} MPa")
print(f" Max displacement: {max_disp:.6f} mm")
# Typical beam: if force is 1000 N, stress should be ~10-100 MPa
# If reaction is 152 million N, that's 152,000 kN - VERY high!
max_reaction = np.max(np.abs(reactions))
max_stress_val = np.max(von_mises)
print(f"\n If force unit is actually kN instead of N:")
print(f" Max reaction: {max_reaction/1000:.2e} kN")
print(f" If stress unit is actually Pa instead of MPa:")
print(f" Max stress: {max_stress_val/1e6:.2e} MPa")
print("\n6. HYPOTHESIS:")
print("-"*60)
if max_reaction > 1e6:
print(" [!] Reaction forces seem TOO LARGE (>1 MN)")
print(" [!] Possible issue: pyNastran returns forces in different units")
print(" [!] Check: Nastran may export in base units (N) while expecting kN")
if max_stress_val > 1e6:
print(" [!] Stresses seem TOO LARGE (>1000 MPa)")
print(" [!] Possible issue: pyNastran returns stress in Pa, not MPa")
print("\n" + "="*60)

921
atomizer-field/neural_field_parser.py
View File

@@ -1,921 +0,0 @@
"""
neural_field_parser.py
Parses NX Nastran files into Neural Field training data
AtomizerField Data Parser v1.0.0
Converts NX Nastran BDF/OP2 files into standardized neural field training format.
This format is designed to be future-proof for years of neural network training.
Usage:
python neural_field_parser.py <case_directory>
Example:
python neural_field_parser.py training_case_001
"""
import json
import numpy as np
import h5py
from pathlib import Path
from datetime import datetime
import hashlib
import warnings
# pyNastran imports
try:
from pyNastran.bdf.bdf import BDF
from pyNastran.op2.op2 import OP2
except ImportError:
print("ERROR: pyNastran is required. Install with: pip install pyNastran")
raise
class NastranToNeuralFieldParser:
"""
Parses Nastran BDF/OP2 files into Neural Field data structure v1.0
This parser extracts complete field data (stress, displacement, strain at every node/element)
rather than just scalar maximum values. This enables neural networks to learn complete
physics fields for 1000x faster structural optimization.
Data Structure v1.0:
-------------------
- metadata: Version, timestamps, analysis info, units
- mesh: Complete node coordinates, element connectivity
- materials: Full material properties (E, nu, rho, etc.)
- boundary_conditions: All constraints (SPC, MPC)
- loads: All loading conditions (forces, pressures, gravity, thermal)
- results: Complete field results (displacement, stress, strain at ALL points)
Attributes:
case_dir (Path): Directory containing input/output subdirectories
bdf_file (Path): Path to Nastran input deck (.bdf or .dat)
op2_file (Path): Path to Nastran binary results (.op2)
bdf (BDF): pyNastran BDF reader object
op2 (OP2): pyNastran OP2 reader object
neural_field_data (dict): Complete parsed data structure
"""
def __init__(self, case_directory):
"""
Initialize parser with case directory
Args:
case_directory (str or Path): Path to case directory containing:
- input/model.bdf (or model.dat)
- output/model.op2
"""
self.case_dir = Path(case_directory)
# Find BDF file (try both .bdf and .dat extensions)
bdf_candidates = list((self.case_dir / "input").glob("model.bdf")) + \
list((self.case_dir / "input").glob("model.dat"))
if not bdf_candidates:
raise FileNotFoundError(
f"No model.bdf or model.dat found in {self.case_dir / 'input'}/"
)
self.bdf_file = bdf_candidates[0]
# Find OP2 file
op2_candidates = list((self.case_dir / "output").glob("model.op2"))
if not op2_candidates:
raise FileNotFoundError(
f"No model.op2 found in {self.case_dir / 'output'}/"
)
self.op2_file = op2_candidates[0]
print(f"Found BDF: {self.bdf_file.name}")
print(f"Found OP2: {self.op2_file.name}")
# Initialize readers with minimal debug output
self.bdf = BDF(debug=False)
self.op2 = OP2(debug=False)
# Initialize data structure v1.0
self.neural_field_data = {
"metadata": {},
"geometry": {},
"mesh": {},
"materials": {},
"boundary_conditions": {},
"loads": {},
"results": {}
}
def parse_all(self):
"""
Main parsing function - extracts all data from BDF/OP2 files
Returns:
dict: Complete neural field data structure
"""
print("\n" + "="*60)
print("Starting AtomizerField Neural Field Parser v1.0")
print("="*60)
# Parse input deck
print("\n[1/6] Reading BDF file...")
self.bdf.read_bdf(str(self.bdf_file))
print(f" Loaded {len(self.bdf.nodes)} nodes, {len(self.bdf.elements)} elements")
# Parse results
print("\n[2/6] Reading OP2 file...")
self.op2.read_op2(str(self.op2_file))
# Check for sol attribute (may not exist in all pyNastran versions)
sol_num = getattr(self.op2, 'sol', 'Unknown')
print(f" Loaded solution: SOL {sol_num}")
# Extract all data
print("\n[3/6] Extracting metadata...")
self.extract_metadata()
print("\n[4/6] Extracting mesh data...")
self.extract_mesh()
print("\n[5/6] Extracting materials, BCs, and loads...")
self.extract_materials()
self.extract_boundary_conditions()
self.extract_loads()
print("\n[6/6] Extracting field results...")
self.extract_results()
# Save to file
print("\nSaving data to disk...")
self.save_data()
print("\n" + "="*60)
print("Parsing complete! [OK]")
print("="*60)
return self.neural_field_data
def extract_metadata(self):
"""
Extract metadata and analysis information
This includes:
- Data format version (v1.0.0)
- Timestamps
- Analysis type (SOL 101, 103, etc.)
- Units system
- File hashes for provenance
"""
# Generate file hash for data provenance
with open(self.bdf_file, 'rb') as f:
bdf_hash = hashlib.sha256(f.read()).hexdigest()
with open(self.op2_file, 'rb') as f:
op2_hash = hashlib.sha256(f.read()).hexdigest()
# Extract title if available
title = ""
if hasattr(self.bdf, 'case_control_deck') and self.bdf.case_control_deck:
if hasattr(self.bdf.case_control_deck, 'title'):
title = str(self.bdf.case_control_deck.title)
# Get solution type if available
sol_num = getattr(self.op2, 'sol', 'Unknown')
self.neural_field_data["metadata"] = {
"version": "1.0.0",
"created_at": datetime.now().isoformat(),
"source": "NX_Nastran",
"case_directory": str(self.case_dir),
"case_name": self.case_dir.name,
"analysis_type": f"SOL_{sol_num}",
"title": title,
"file_hashes": {
"bdf": bdf_hash,
"op2": op2_hash
},
"units": {
"length": "mm", # Standard NX units
"force": "N",
"stress": "MPa",
"mass": "kg",
"temperature": "C"
},
"parser_version": "1.0.0",
"notes": "Complete field data for neural network training"
}
print(f" Analysis: {self.neural_field_data['metadata']['analysis_type']}")
def extract_mesh(self):
"""
Extract complete mesh data from BDF
This preserves:
- All node coordinates (global coordinate system)
- All element connectivity
- Element types (solid, shell, beam, rigid)
- Material and property IDs for each element
"""
print(" Extracting nodes...")
# Nodes - store in sorted order for consistent indexing
nodes = []
node_ids = []
for nid, node in sorted(self.bdf.nodes.items()):
node_ids.append(nid)
# Get position in global coordinate system
pos = node.get_position()
nodes.append([pos[0], pos[1], pos[2]])
nodes_array = np.array(nodes, dtype=np.float64)
print(f" Extracted {len(nodes)} nodes")
print(f" Extracting elements...")
# Elements - organize by type for efficient neural network processing
element_data = {
"solid": [],
"shell": [],
"beam": [],
"rigid": []
}
element_type_counts = {}
for eid, elem in sorted(self.bdf.elements.items()):
elem_type = elem.type
element_type_counts[elem_type] = element_type_counts.get(elem_type, 0) + 1
# Solid elements (3D stress states)
if elem_type in ['CTETRA', 'CHEXA', 'CPENTA', 'CTETRA10', 'CHEXA20', 'CPENTA15']:
element_data["solid"].append({
"id": eid,
"type": elem_type,
"nodes": list(elem.node_ids),
"material_id": elem.pid, # Property ID which links to material
"property_id": elem.pid if hasattr(elem, 'pid') else None
})
# Shell elements (2D plane stress)
elif elem_type in ['CQUAD4', 'CTRIA3', 'CQUAD8', 'CTRIA6', 'CQUAD', 'CTRIA']:
thickness = None
try:
# Get thickness from property
if hasattr(elem, 'pid') and elem.pid in self.bdf.properties:
prop = self.bdf.properties[elem.pid]
if hasattr(prop, 't'):
thickness = prop.t
except:
pass
element_data["shell"].append({
"id": eid,
"type": elem_type,
"nodes": list(elem.node_ids),
"material_id": elem.pid,
"property_id": elem.pid,
"thickness": thickness
})
# Beam elements (1D elements)
elif elem_type in ['CBAR', 'CBEAM', 'CROD', 'CONROD']:
element_data["beam"].append({
"id": eid,
"type": elem_type,
"nodes": list(elem.node_ids),
"material_id": elem.pid if hasattr(elem, 'pid') else None,
"property_id": elem.pid if hasattr(elem, 'pid') else None
})
# Rigid elements (kinematic constraints)
elif elem_type in ['RBE2', 'RBE3', 'RBAR', 'RROD']:
element_data["rigid"].append({
"id": eid,
"type": elem_type,
"nodes": list(elem.node_ids)
})
print(f" Extracted {len(self.bdf.elements)} elements:")
for etype, count in element_type_counts.items():
print(f" {etype}: {count}")
# Calculate mesh bounding box for reference
bbox_min = nodes_array.min(axis=0)
bbox_max = nodes_array.max(axis=0)
bbox_size = bbox_max - bbox_min
# Store mesh data
self.neural_field_data["mesh"] = {
"statistics": {
"n_nodes": len(nodes),
"n_elements": len(self.bdf.elements),
"element_types": {
"solid": len(element_data["solid"]),
"shell": len(element_data["shell"]),
"beam": len(element_data["beam"]),
"rigid": len(element_data["rigid"])
},
"element_type_breakdown": element_type_counts
},
"bounding_box": {
"min": bbox_min.tolist(),
"max": bbox_max.tolist(),
"size": bbox_size.tolist()
},
"nodes": {
"ids": node_ids,
"coordinates": nodes_array.tolist(), # Will be stored in HDF5
"shape": list(nodes_array.shape),
"dtype": str(nodes_array.dtype)
},
"elements": element_data
}
def extract_materials(self):
"""
Extract all material properties
Captures complete material definitions including:
- Isotropic (MAT1): E, nu, rho, G, alpha
- Orthotropic (MAT8, MAT9): directional properties
- Stress limits for validation
"""
print(" Extracting materials...")
materials = []
for mid, mat in sorted(self.bdf.materials.items()):
mat_data = {
"id": mid,
"type": mat.type
}
if mat.type == 'MAT1': # Isotropic material
mat_data.update({
"E": float(mat.e) if mat.e is not None else None, # Young's modulus
"nu": float(mat.nu) if mat.nu is not None else None, # Poisson's ratio
"rho": float(mat.rho) if mat.rho is not None else None,# Density
"G": float(mat.g) if mat.g is not None else None, # Shear modulus
"alpha": float(mat.a) if hasattr(mat, 'a') and mat.a is not None else None, # Thermal expansion
"tref": float(mat.tref) if hasattr(mat, 'tref') and mat.tref is not None else None,
})
# Stress limits (if defined)
try:
if hasattr(mat, 'St') and callable(mat.St):
mat_data["ST"] = float(mat.St()) if mat.St() is not None else None
if hasattr(mat, 'Sc') and callable(mat.Sc):
mat_data["SC"] = float(mat.Sc()) if mat.Sc() is not None else None
if hasattr(mat, 'Ss') and callable(mat.Ss):
mat_data["SS"] = float(mat.Ss()) if mat.Ss() is not None else None
except:
pass
elif mat.type == 'MAT8': # Orthotropic shell material
mat_data.update({
"E1": float(mat.e11) if hasattr(mat, 'e11') and mat.e11 is not None else None,
"E2": float(mat.e22) if hasattr(mat, 'e22') and mat.e22 is not None else None,
"nu12": float(mat.nu12) if hasattr(mat, 'nu12') and mat.nu12 is not None else None,
"G12": float(mat.g12) if hasattr(mat, 'g12') and mat.g12 is not None else None,
"G1z": float(mat.g1z) if hasattr(mat, 'g1z') and mat.g1z is not None else None,
"G2z": float(mat.g2z) if hasattr(mat, 'g2z') and mat.g2z is not None else None,
"rho": float(mat.rho) if hasattr(mat, 'rho') and mat.rho is not None else None,
})
materials.append(mat_data)
self.neural_field_data["materials"] = materials
print(f" Extracted {len(materials)} materials")
def extract_boundary_conditions(self):
"""
Extract all boundary conditions
Includes:
- SPC: Single point constraints (fixed DOFs)
- MPC: Multi-point constraints (equations)
- SUPORT: Free body supports
"""
print(" Extracting boundary conditions...")
bcs = {
"spc": [], # Single point constraints
"mpc": [], # Multi-point constraints
"suport": [] # Free body supports
}
# SPC (fixed DOFs) - critical for neural network to understand support conditions
spc_count = 0
for spc_id, spc_list in self.bdf.spcs.items():
for spc in spc_list:
try:
# Handle different SPC types
if hasattr(spc, 'node_ids'):
nodes = spc.node_ids
elif hasattr(spc, 'node'):
nodes = [spc.node]
else:
continue
for node in nodes:
bcs["spc"].append({
"id": spc_id,
"node": int(node),
"dofs": str(spc.components) if hasattr(spc, 'components') else "123456",
"enforced_motion": float(spc.enforced) if hasattr(spc, 'enforced') and spc.enforced is not None else 0.0
})
spc_count += 1
except Exception as e:
warnings.warn(f"Could not parse SPC {spc_id}: {e}")
# MPC equations
mpc_count = 0
for mpc_id, mpc_list in self.bdf.mpcs.items():
for mpc in mpc_list:
try:
bcs["mpc"].append({
"id": mpc_id,
"nodes": list(mpc.node_ids) if hasattr(mpc, 'node_ids') else [],
"coefficients": list(mpc.coefficients) if hasattr(mpc, 'coefficients') else [],
"components": list(mpc.components) if hasattr(mpc, 'components') else []
})
mpc_count += 1
except Exception as e:
warnings.warn(f"Could not parse MPC {mpc_id}: {e}")
self.neural_field_data["boundary_conditions"] = bcs
print(f" Extracted {spc_count} SPCs, {mpc_count} MPCs")
def extract_loads(self):
"""
Extract all loading conditions
Includes:
- Point forces and moments
- Distributed pressures
- Gravity loads
- Thermal loads
"""
print(" Extracting loads...")
loads = {
"point_forces": [],
"pressure": [],
"gravity": [],
"thermal": []
}
force_count = 0
pressure_count = 0
gravity_count = 0
# Point forces, moments, and pressures
for load_id, load_list in self.bdf.loads.items():
for load in load_list:
try:
if load.type == 'FORCE':
loads["point_forces"].append({
"id": load_id,
"type": "force",
"node": int(load.node),
"magnitude": float(load.mag),
"direction": [float(load.xyz[0]), float(load.xyz[1]), float(load.xyz[2])],
"coord_system": int(load.cid) if hasattr(load, 'cid') else 0
})
force_count += 1
elif load.type == 'MOMENT':
loads["point_forces"].append({
"id": load_id,
"type": "moment",
"node": int(load.node),
"magnitude": float(load.mag),
"direction": [float(load.xyz[0]), float(load.xyz[1]), float(load.xyz[2])],
"coord_system": int(load.cid) if hasattr(load, 'cid') else 0
})
force_count += 1
elif load.type in ['PLOAD', 'PLOAD2', 'PLOAD4']:
pressure_data = {
"id": load_id,
"type": load.type
}
if hasattr(load, 'eids'):
pressure_data["elements"] = list(load.eids)
elif hasattr(load, 'eid'):
pressure_data["elements"] = [int(load.eid)]
if hasattr(load, 'pressures'):
pressure_data["pressure"] = [float(p) for p in load.pressures]
elif hasattr(load, 'pressure'):
pressure_data["pressure"] = float(load.pressure)
loads["pressure"].append(pressure_data)
pressure_count += 1
elif load.type == 'GRAV':
loads["gravity"].append({
"id": load_id,
"acceleration": float(load.scale),
"direction": [float(load.N[0]), float(load.N[1]), float(load.N[2])],
"coord_system": int(load.cid) if hasattr(load, 'cid') else 0
})
gravity_count += 1
except Exception as e:
warnings.warn(f"Could not parse load {load_id} type {load.type}: {e}")
# Temperature loads (if available)
thermal_count = 0
if hasattr(self.bdf, 'temps'):
for temp_id, temp_list in self.bdf.temps.items():
for temp in temp_list:
try:
loads["thermal"].append({
"id": temp_id,
"node": int(temp.node),
"temperature": float(temp.temperature)
})
thermal_count += 1
except Exception as e:
warnings.warn(f"Could not parse thermal load {temp_id}: {e}")
self.neural_field_data["loads"] = loads
print(f" Extracted {force_count} forces, {pressure_count} pressures, {gravity_count} gravity, {thermal_count} thermal")
def extract_results(self):
"""
Extract complete field results from OP2
This is the CRITICAL function for neural field learning.
We extract COMPLETE fields, not just maximum values:
- Displacement at every node (6 DOF)
- Stress at every element (full tensor)
- Strain at every element (full tensor)
- Reaction forces at constrained nodes
This complete field data enables the neural network to learn
the physics of how structures respond to loads.
"""
print(" Extracting field results...")
results = {}
# Determine subcase ID (usually 1 for linear static)
subcase_id = 1
if hasattr(self.op2, 'isubcase_name_map'):
available_subcases = list(self.op2.isubcase_name_map.keys())
if available_subcases:
subcase_id = available_subcases[0]
print(f" Using subcase ID: {subcase_id}")
# Displacement - complete field at all nodes
if hasattr(self.op2, 'displacements') and subcase_id in self.op2.displacements:
print(" Processing displacement field...")
disp = self.op2.displacements[subcase_id]
disp_data = disp.data[0, :, :] # [itime=0, all_nodes, 6_dofs]
# Extract node IDs
node_ids = disp.node_gridtype[:, 0].tolist()
# Calculate magnitudes for quick reference
translation_mag = np.linalg.norm(disp_data[:, :3], axis=1)
rotation_mag = np.linalg.norm(disp_data[:, 3:], axis=1)
results["displacement"] = {
"node_ids": node_ids,
"data": disp_data.tolist(), # Will be stored in HDF5
"shape": list(disp_data.shape),
"dtype": str(disp_data.dtype),
"max_translation": float(np.max(translation_mag)),
"max_rotation": float(np.max(rotation_mag)),
"units": "mm and radians"
}
print(f" Displacement: {len(node_ids)} nodes, max={results['displacement']['max_translation']:.6f} mm")
# Stress - handle different element types
stress_results = {}
# Solid element stress (CTETRA, CHEXA, etc.)
stress_attrs = ['ctetra_stress', 'chexa_stress', 'cpenta_stress']
for attr in stress_attrs:
if hasattr(self.op2, attr):
stress_obj = getattr(self.op2, attr)
if subcase_id in stress_obj:
elem_type = attr.replace('_stress', '')
print(f" Processing {elem_type} stress...")
stress = stress_obj[subcase_id]
stress_data = stress.data[0, :, :]
# Extract element IDs
element_ids = stress.element_node[:, 0].tolist()
# Von Mises stress is usually the last column
von_mises = None
if stress_data.shape[1] >= 7: # Has von Mises
von_mises = stress_data[:, -1]
max_vm = float(np.max(von_mises))
von_mises = von_mises.tolist()
else:
max_vm = None
stress_results[f"{elem_type}_stress"] = {
"element_ids": element_ids,
"data": stress_data.tolist(), # Full stress tensor
"shape": list(stress_data.shape),
"dtype": str(stress_data.dtype),
"von_mises": von_mises,
"max_von_mises": max_vm,
"units": "MPa"
}
print(f" {elem_type}: {len(element_ids)} elements, max VM={max_vm:.2f} MPa" if max_vm else f" {elem_type}: {len(element_ids)} elements")
# Shell element stress
shell_stress_attrs = ['cquad4_stress', 'ctria3_stress', 'cquad8_stress', 'ctria6_stress']
for attr in shell_stress_attrs:
if hasattr(self.op2, attr):
stress_obj = getattr(self.op2, attr)
if subcase_id in stress_obj:
elem_type = attr.replace('_stress', '')
print(f" Processing {elem_type} stress...")
stress = stress_obj[subcase_id]
stress_data = stress.data[0, :, :]
element_ids = stress.element_node[:, 0].tolist()
stress_results[f"{elem_type}_stress"] = {
"element_ids": element_ids,
"data": stress_data.tolist(),
"shape": list(stress_data.shape),
"dtype": str(stress_data.dtype),
"units": "MPa"
}
print(f" {elem_type}: {len(element_ids)} elements")
results["stress"] = stress_results
# Strain - similar to stress
strain_results = {}
strain_attrs = ['ctetra_strain', 'chexa_strain', 'cpenta_strain',
'cquad4_strain', 'ctria3_strain']
for attr in strain_attrs:
if hasattr(self.op2, attr):
strain_obj = getattr(self.op2, attr)
if subcase_id in strain_obj:
elem_type = attr.replace('_strain', '')
strain = strain_obj[subcase_id]
strain_data = strain.data[0, :, :]
element_ids = strain.element_node[:, 0].tolist()
strain_results[f"{elem_type}_strain"] = {
"element_ids": element_ids,
"data": strain_data.tolist(),
"shape": list(strain_data.shape),
"dtype": str(strain_data.dtype),
"units": "mm/mm"
}
if strain_results:
results["strain"] = strain_results
print(f" Extracted strain for {len(strain_results)} element types")
# SPC Forces (reaction forces at constraints)
if hasattr(self.op2, 'spc_forces') and subcase_id in self.op2.spc_forces:
print(" Processing reaction forces...")
spc = self.op2.spc_forces[subcase_id]
spc_data = spc.data[0, :, :]
node_ids = spc.node_gridtype[:, 0].tolist()
# Calculate total reaction force magnitude
force_mag = np.linalg.norm(spc_data[:, :3], axis=1)
moment_mag = np.linalg.norm(spc_data[:, 3:], axis=1)
results["reactions"] = {
"node_ids": node_ids,
"forces": spc_data.tolist(),
"shape": list(spc_data.shape),
"dtype": str(spc_data.dtype),
"max_force": float(np.max(force_mag)),
"max_moment": float(np.max(moment_mag)),
"units": "N and N-mm"
}
print(f" Reactions: {len(node_ids)} nodes, max force={results['reactions']['max_force']:.2f} N")
self.neural_field_data["results"] = results
def save_data(self):
"""
Save parsed data to JSON and HDF5 files
Data structure:
- neural_field_data.json: Metadata, structure, small arrays
- neural_field_data.h5: Large arrays (node coordinates, field results)
HDF5 is used for efficient storage and loading of large numerical arrays.
JSON provides human-readable metadata and structure.
"""
# Save JSON metadata
json_file = self.case_dir / "neural_field_data.json"
# Create a copy for JSON (will remove large arrays)
json_data = self._prepare_json_data()
with open(json_file, 'w') as f:
json.dump(json_data, f, indent=2, default=str)
print(f" [OK] Saved metadata to: {json_file.name}")
# Save HDF5 for large arrays
h5_file = self.case_dir / "neural_field_data.h5"
with h5py.File(h5_file, 'w') as f:
# Metadata attributes
f.attrs['version'] = '1.0.0'
f.attrs['created_at'] = self.neural_field_data['metadata']['created_at']
f.attrs['case_name'] = self.neural_field_data['metadata']['case_name']
# Save mesh data
mesh_grp = f.create_group('mesh')
node_coords = np.array(self.neural_field_data["mesh"]["nodes"]["coordinates"])
mesh_grp.create_dataset('node_coordinates',
data=node_coords,
compression='gzip',
compression_opts=4)
mesh_grp.create_dataset('node_ids',
data=np.array(self.neural_field_data["mesh"]["nodes"]["ids"]))
# Save results
if "results" in self.neural_field_data:
results_grp = f.create_group('results')
# Displacement
if "displacement" in self.neural_field_data["results"]:
disp_data = np.array(self.neural_field_data["results"]["displacement"]["data"])
results_grp.create_dataset('displacement',
data=disp_data,
compression='gzip',
compression_opts=4)
results_grp.create_dataset('displacement_node_ids',
data=np.array(self.neural_field_data["results"]["displacement"]["node_ids"]))
# Stress fields
if "stress" in self.neural_field_data["results"]:
stress_grp = results_grp.create_group('stress')
for stress_type, stress_data in self.neural_field_data["results"]["stress"].items():
type_grp = stress_grp.create_group(stress_type)
type_grp.create_dataset('data',
data=np.array(stress_data["data"]),
compression='gzip',
compression_opts=4)
type_grp.create_dataset('element_ids',
data=np.array(stress_data["element_ids"]))
# Strain fields
if "strain" in self.neural_field_data["results"]:
strain_grp = results_grp.create_group('strain')
for strain_type, strain_data in self.neural_field_data["results"]["strain"].items():
type_grp = strain_grp.create_group(strain_type)
type_grp.create_dataset('data',
data=np.array(strain_data["data"]),
compression='gzip',
compression_opts=4)
type_grp.create_dataset('element_ids',
data=np.array(strain_data["element_ids"]))
# Reactions
if "reactions" in self.neural_field_data["results"]:
reactions_data = np.array(self.neural_field_data["results"]["reactions"]["forces"])
results_grp.create_dataset('reactions',
data=reactions_data,
compression='gzip',
compression_opts=4)
results_grp.create_dataset('reaction_node_ids',
data=np.array(self.neural_field_data["results"]["reactions"]["node_ids"]))
print(f" [OK] Saved field data to: {h5_file.name}")
# Calculate and display file sizes
json_size = json_file.stat().st_size / 1024 # KB
h5_size = h5_file.stat().st_size / 1024 # KB
print(f"\n File sizes:")
print(f" JSON: {json_size:.1f} KB")
print(f" HDF5: {h5_size:.1f} KB")
print(f" Total: {json_size + h5_size:.1f} KB")
def _prepare_json_data(self):
"""
Prepare data for JSON export by removing large arrays
(they go to HDF5 instead)
"""
import copy
json_data = copy.deepcopy(self.neural_field_data)
# Remove large arrays from nodes (keep metadata)
if "mesh" in json_data and "nodes" in json_data["mesh"]:
json_data["mesh"]["nodes"]["coordinates"] = f"<stored in HDF5: shape {json_data['mesh']['nodes']['shape']}>"
# Remove large result arrays
if "results" in json_data:
if "displacement" in json_data["results"]:
shape = json_data["results"]["displacement"]["shape"]
json_data["results"]["displacement"]["data"] = f"<stored in HDF5: shape {shape}>"
if "stress" in json_data["results"]:
for stress_type in json_data["results"]["stress"]:
shape = json_data["results"]["stress"][stress_type]["shape"]
json_data["results"]["stress"][stress_type]["data"] = f"<stored in HDF5: shape {shape}>"
if "strain" in json_data["results"]:
for strain_type in json_data["results"]["strain"]:
shape = json_data["results"]["strain"][strain_type]["shape"]
json_data["results"]["strain"][strain_type]["data"] = f"<stored in HDF5: shape {shape}>"
if "reactions" in json_data["results"]:
shape = json_data["results"]["reactions"]["shape"]
json_data["results"]["reactions"]["forces"] = f"<stored in HDF5: shape {shape}>"
return json_data
# ============================================================================
# MAIN ENTRY POINT
# ============================================================================
def main():
"""
Main function to run the parser from command line
Usage:
python neural_field_parser.py <case_directory>
"""
import sys
if len(sys.argv) < 2:
print("\nAtomizerField Neural Field Parser v1.0")
print("="*60)
print("\nUsage:")
print(" python neural_field_parser.py <case_directory>")
print("\nExample:")
print(" python neural_field_parser.py training_case_001")
print("\nCase directory should contain:")
print(" input/model.bdf (or model.dat)")
print(" output/model.op2")
print("\n")
sys.exit(1)
case_dir = sys.argv[1]
# Verify directory exists
if not Path(case_dir).exists():
print(f"ERROR: Directory not found: {case_dir}")
sys.exit(1)
# Create parser
try:
parser = NastranToNeuralFieldParser(case_dir)
except FileNotFoundError as e:
print(f"\nERROR: {e}")
print("\nPlease ensure your case directory contains:")
print(" input/model.bdf (or model.dat)")
print(" output/model.op2")
sys.exit(1)
# Parse all data
try:
data = parser.parse_all()
# Print summary
print("\n" + "="*60)
print("PARSING SUMMARY")
print("="*60)
print(f"Case: {data['metadata']['case_name']}")
print(f"Analysis: {data['metadata']['analysis_type']}")
print(f"\nMesh:")
print(f" Nodes: {data['mesh']['statistics']['n_nodes']:,}")
print(f" Elements: {data['mesh']['statistics']['n_elements']:,}")
for elem_type, count in data['mesh']['statistics']['element_types'].items():
if count > 0:
print(f" {elem_type}: {count:,}")
print(f"\nMaterials: {len(data['materials'])}")
print(f"Boundary Conditions: {len(data['boundary_conditions']['spc'])} SPCs")
print(f"Loads: {len(data['loads']['point_forces'])} forces, {len(data['loads']['pressure'])} pressures")
if "displacement" in data['results']:
print(f"\nResults:")
print(f" Displacement: {len(data['results']['displacement']['node_ids'])} nodes")
print(f" Max: {data['results']['displacement']['max_translation']:.6f} mm")
if "stress" in data['results']:
for stress_type in data['results']['stress']:
if 'max_von_mises' in data['results']['stress'][stress_type]:
max_vm = data['results']['stress'][stress_type]['max_von_mises']
if max_vm is not None:
print(f" {stress_type}: Max VM = {max_vm:.2f} MPa")
print("\n[OK] Data ready for neural network training!")
print("="*60 + "\n")
except Exception as e:
print(f"\n" + "="*60)
print("ERROR DURING PARSING")
print("="*60)
print(f"{e}\n")
import traceback
traceback.print_exc()
sys.exit(1)
if __name__ == "__main__":
main()

25
atomizer-field/neural_models/__init__.py
View File

@@ -1,25 +0,0 @@
"""
AtomizerField Neural Models Package
Phase 2: Neural Network Architecture for Field Prediction
This package contains neural network models for learning complete FEA field results
from mesh geometry, boundary conditions, and loads.
Models:
- AtomizerFieldModel: Full field predictor (displacement + stress fields)
- ParametricFieldPredictor: Design-conditioned scalar predictor (mass, freq, disp, stress)
"""
__version__ = "2.0.0"
# Import main model classes for convenience
from .field_predictor import AtomizerFieldModel, create_model
from .parametric_predictor import ParametricFieldPredictor, create_parametric_model
__all__ = [
'AtomizerFieldModel',
'create_model',
'ParametricFieldPredictor',
'create_parametric_model',
]

416
atomizer-field/neural_models/data_loader.py
View File

@@ -1,416 +0,0 @@
"""
data_loader.py
Data loading pipeline for neural field training
AtomizerField Data Loader v2.0
Converts parsed FEA data (HDF5 + JSON) into PyTorch Geometric graphs for training.
Key Transformation:
Parsed FEA Data → Graph Representation → Neural Network Input
Graph structure:
- Nodes: FEA mesh nodes (with coordinates, BCs, loads)
- Edges: Element connectivity (with material properties)
- Labels: Displacement and stress fields (ground truth from FEA)
"""
import json
import h5py
import numpy as np
from pathlib import Path
import torch
from torch.utils.data import Dataset
from torch_geometric.data import Data
from torch_geometric.loader import DataLoader
import warnings
class FEAMeshDataset(Dataset):
"""
PyTorch Dataset for FEA mesh data
Loads parsed neural field data and converts to PyTorch Geometric graphs.
Each graph represents one FEA analysis case.
"""
def __init__(
self,
case_directories,
normalize=True,
include_stress=True,
cache_in_memory=False
):
"""
Initialize dataset
Args:
case_directories (list): List of paths to parsed cases
normalize (bool): Normalize node coordinates and results
include_stress (bool): Include stress in targets
cache_in_memory (bool): Load all data into RAM (faster but memory-intensive)
"""
self.case_dirs = [Path(d) for d in case_directories]
self.normalize = normalize
self.include_stress = include_stress
self.cache_in_memory = cache_in_memory
# Validate all cases exist
self.valid_cases = []
for case_dir in self.case_dirs:
if self._validate_case(case_dir):
self.valid_cases.append(case_dir)
else:
warnings.warn(f"Skipping invalid case: {case_dir}")
print(f"Loaded {len(self.valid_cases)}/{len(self.case_dirs)} valid cases")
# Cache data if requested
self.cache = {}
if cache_in_memory:
print("Caching data in memory...")
for idx in range(len(self.valid_cases)):
self.cache[idx] = self._load_case(idx)
print("Cache complete!")
# Compute normalization statistics
if normalize:
self._compute_normalization_stats()
def _validate_case(self, case_dir):
"""Check if case has required files"""
json_file = case_dir / "neural_field_data.json"
h5_file = case_dir / "neural_field_data.h5"
return json_file.exists() and h5_file.exists()
def __len__(self):
return len(self.valid_cases)
def __getitem__(self, idx):
"""
Get graph data for one case
Returns:
torch_geometric.data.Data object with:
- x: Node features [num_nodes, feature_dim]
- edge_index: Element connectivity [2, num_edges]
- edge_attr: Edge features (material props) [num_edges, edge_dim]
- y_displacement: Target displacement [num_nodes, 6]
- y_stress: Target stress [num_nodes, 6] (if include_stress)
- bc_mask: Boundary condition mask [num_nodes, 6]
- pos: Node positions [num_nodes, 3]
"""
if self.cache_in_memory and idx in self.cache:
return self.cache[idx]
return self._load_case(idx)
def _load_case(self, idx):
"""Load and process a single case"""
case_dir = self.valid_cases[idx]
# Load JSON metadata
with open(case_dir / "neural_field_data.json", 'r') as f:
metadata = json.load(f)
# Load HDF5 field data
with h5py.File(case_dir / "neural_field_data.h5", 'r') as f:
# Node coordinates
node_coords = torch.from_numpy(f['mesh/node_coordinates'][:]).float()
# Displacement field (target)
displacement = torch.from_numpy(f['results/displacement'][:]).float()
# Stress field (target, if available)
stress = None
if self.include_stress and 'results/stress' in f:
# Try to load first available stress type
stress_group = f['results/stress']
for stress_type in stress_group.keys():
stress_data = stress_group[stress_type]['data'][:]
stress = torch.from_numpy(stress_data).float()
break
# Build graph structure
graph_data = self._build_graph(metadata, node_coords, displacement, stress)
# Normalize if requested
if self.normalize:
graph_data = self._normalize_graph(graph_data)
return graph_data
def _build_graph(self, metadata, node_coords, displacement, stress):
"""
Convert FEA mesh to graph
Args:
metadata (dict): Parsed metadata
node_coords (Tensor): Node positions [num_nodes, 3]
displacement (Tensor): Displacement field [num_nodes, 6]
stress (Tensor): Stress field [num_nodes, 6] or None
Returns:
torch_geometric.data.Data
"""
num_nodes = node_coords.shape[0]
# === NODE FEATURES ===
# Start with coordinates
node_features = [node_coords] # [num_nodes, 3]
# Add boundary conditions (which DOFs are constrained)
bc_mask = torch.zeros(num_nodes, 6) # [num_nodes, 6]
if 'boundary_conditions' in metadata and 'spc' in metadata['boundary_conditions']:
for spc in metadata['boundary_conditions']['spc']:
node_id = spc['node']
# Find node index (assuming node IDs are sequential starting from 1)
# This is a simplification - production code should use ID mapping
if node_id <= num_nodes:
dofs = spc['dofs']
# Parse DOF string (e.g., "123" means constrained in x,y,z)
for dof_char in str(dofs):
if dof_char.isdigit():
dof_idx = int(dof_char) - 1 # 0-indexed
if 0 <= dof_idx < 6:
bc_mask[node_id - 1, dof_idx] = 1.0
node_features.append(bc_mask) # [num_nodes, 6]
# Add load information (force magnitude at each node)
load_features = torch.zeros(num_nodes, 3) # [num_nodes, 3] for x,y,z forces
if 'loads' in metadata and 'point_forces' in metadata['loads']:
for force in metadata['loads']['point_forces']:
node_id = force['node']
if node_id <= num_nodes:
magnitude = force['magnitude']
direction = force['direction']
force_vector = [magnitude * d for d in direction]
load_features[node_id - 1] = torch.tensor(force_vector)
node_features.append(load_features) # [num_nodes, 3]
# Concatenate all node features
x = torch.cat(node_features, dim=-1) # [num_nodes, 3+6+3=12]
# === EDGE FEATURES ===
# Build edge index from element connectivity
edge_index = []
edge_attrs = []
# Get material properties
material_dict = {}
if 'materials' in metadata:
for mat in metadata['materials']:
mat_id = mat['id']
if mat['type'] == 'MAT1':
material_dict[mat_id] = [
mat.get('E', 0.0) / 1e6, # Normalize E (MPa → GPa)
mat.get('nu', 0.0),
mat.get('rho', 0.0) * 1e6, # Normalize rho
mat.get('G', 0.0) / 1e6 if mat.get('G') else 0.0,
mat.get('alpha', 0.0) * 1e6 if mat.get('alpha') else 0.0
]
# Process elements to create edges
if 'mesh' in metadata and 'elements' in metadata['mesh']:
for elem_type in ['solid', 'shell', 'beam']:
if elem_type in metadata['mesh']['elements']:
for elem in metadata['mesh']['elements'][elem_type]:
elem_nodes = elem['nodes']
mat_id = elem.get('material_id', 1)
# Get material properties for this element
mat_props = material_dict.get(mat_id, [0.0] * 5)
# Create edges between all node pairs in element
# (fully connected within element)
for i in range(len(elem_nodes)):
for j in range(i + 1, len(elem_nodes)):
node_i = elem_nodes[i] - 1 # 0-indexed
node_j = elem_nodes[j] - 1
if node_i < num_nodes and node_j < num_nodes:
# Add bidirectional edges
edge_index.append([node_i, node_j])
edge_index.append([node_j, node_i])
# Both edges get same material properties
edge_attrs.append(mat_props)
edge_attrs.append(mat_props)
# Convert to tensors
if edge_index:
edge_index = torch.tensor(edge_index, dtype=torch.long).t() # [2, num_edges]
edge_attr = torch.tensor(edge_attrs, dtype=torch.float) # [num_edges, 5]
else:
# No edges (shouldn't happen, but handle gracefully)
edge_index = torch.zeros((2, 0), dtype=torch.long)
edge_attr = torch.zeros((0, 5), dtype=torch.float)
# === CREATE DATA OBJECT ===
data = Data(
x=x,
edge_index=edge_index,
edge_attr=edge_attr,
y_displacement=displacement,
bc_mask=bc_mask,
pos=node_coords # Store original positions
)
# Add stress if available
if stress is not None:
data.y_stress = stress
return data
def _normalize_graph(self, data):
"""
Normalize graph features
- Coordinates: Center and scale to unit box
- Displacement: Scale by mean displacement
- Stress: Scale by mean stress
"""
# Normalize coordinates (already done in node features)
if hasattr(self, 'coord_mean') and hasattr(self, 'coord_std'):
# Extract coords from features (first 3 dimensions)
coords = data.x[:, :3]
coords_norm = (coords - self.coord_mean) / (self.coord_std + 1e-8)
data.x[:, :3] = coords_norm
# Normalize displacement
if hasattr(self, 'disp_mean') and hasattr(self, 'disp_std'):
data.y_displacement = (data.y_displacement - self.disp_mean) / (self.disp_std + 1e-8)
# Normalize stress
if hasattr(data, 'y_stress') and hasattr(self, 'stress_mean') and hasattr(self, 'stress_std'):
data.y_stress = (data.y_stress - self.stress_mean) / (self.stress_std + 1e-8)
return data
def _compute_normalization_stats(self):
"""
Compute mean and std for normalization across entire dataset
"""
print("Computing normalization statistics...")
all_coords = []
all_disp = []
all_stress = []
for idx in range(len(self.valid_cases)):
case_dir = self.valid_cases[idx]
with h5py.File(case_dir / "neural_field_data.h5", 'r') as f:
coords = f['mesh/node_coordinates'][:]
disp = f['results/displacement'][:]
all_coords.append(coords)
all_disp.append(disp)
# Load stress if available
if self.include_stress and 'results/stress' in f:
stress_group = f['results/stress']
for stress_type in stress_group.keys():
stress_data = stress_group[stress_type]['data'][:]
all_stress.append(stress_data)
break
# Concatenate all data
all_coords = np.concatenate(all_coords, axis=0)
all_disp = np.concatenate(all_disp, axis=0)
# Compute statistics
self.coord_mean = torch.from_numpy(all_coords.mean(axis=0)).float()
self.coord_std = torch.from_numpy(all_coords.std(axis=0)).float()
self.disp_mean = torch.from_numpy(all_disp.mean(axis=0)).float()
self.disp_std = torch.from_numpy(all_disp.std(axis=0)).float()
if all_stress:
all_stress = np.concatenate(all_stress, axis=0)
self.stress_mean = torch.from_numpy(all_stress.mean(axis=0)).float()
self.stress_std = torch.from_numpy(all_stress.std(axis=0)).float()
print("Normalization statistics computed!")
def create_dataloaders(
train_cases,
val_cases,
batch_size=4,
num_workers=0,
normalize=True,
include_stress=True
):
"""
Create training and validation dataloaders
Args:
train_cases (list): List of training case directories
val_cases (list): List of validation case directories
batch_size (int): Batch size
num_workers (int): Number of data loading workers
normalize (bool): Normalize features
include_stress (bool): Include stress targets
Returns:
train_loader, val_loader
"""
print("\nCreating datasets...")
# Create datasets
train_dataset = FEAMeshDataset(
train_cases,
normalize=normalize,
include_stress=include_stress,
cache_in_memory=False # Set to True for small datasets
)
val_dataset = FEAMeshDataset(
val_cases,
normalize=normalize,
include_stress=include_stress,
cache_in_memory=False
)
# Share normalization stats with validation set
if normalize and hasattr(train_dataset, 'coord_mean'):
val_dataset.coord_mean = train_dataset.coord_mean
val_dataset.coord_std = train_dataset.coord_std
val_dataset.disp_mean = train_dataset.disp_mean
val_dataset.disp_std = train_dataset.disp_std
if hasattr(train_dataset, 'stress_mean'):
val_dataset.stress_mean = train_dataset.stress_mean
val_dataset.stress_std = train_dataset.stress_std
# Create dataloaders
train_loader = DataLoader(
train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers
)
val_loader = DataLoader(
val_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers
)
print(f"\nDataloaders created:")
print(f" Training: {len(train_dataset)} cases")
print(f" Validation: {len(val_dataset)} cases")
return train_loader, val_loader
if __name__ == "__main__":
# Test data loader
print("Testing FEA Mesh Data Loader...\n")
# This is a placeholder test - you would use actual parsed case directories
print("Note: This test requires actual parsed FEA data.")
print("Run the parser first on your NX Nastran files.")
print("\nData loader implementation complete!")

490
atomizer-field/neural_models/field_predictor.py
View File

@@ -1,490 +0,0 @@
"""
field_predictor.py
Graph Neural Network for predicting complete FEA field results
AtomizerField Field Predictor v2.0
Uses Graph Neural Networks to learn the physics of structural response.
Key Innovation:
Instead of: parameters → FEA → max_stress (scalar)
We learn: parameters → Neural Network → complete stress field (N values)
This enables 1000x faster optimization with physics understanding.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_geometric.nn import MessagePassing, global_mean_pool
from torch_geometric.data import Data
import numpy as np
class MeshGraphConv(MessagePassing):
"""
Custom Graph Convolution for FEA meshes
This layer propagates information along mesh edges (element connectivity)
to learn how forces flow through the structure.
Key insight: Stress and displacement fields follow mesh topology.
Adjacent elements influence each other through equilibrium.
"""
def __init__(self, in_channels, out_channels, edge_dim=None):
"""
Args:
in_channels (int): Input node feature dimension
out_channels (int): Output node feature dimension
edge_dim (int): Edge feature dimension (optional)
"""
super().__init__(aggr='mean') # Mean aggregation of neighbor messages
self.in_channels = in_channels
self.out_channels = out_channels
# Message function: how to combine node and edge features
if edge_dim is not None:
self.message_mlp = nn.Sequential(
nn.Linear(2 * in_channels + edge_dim, out_channels),
nn.LayerNorm(out_channels),
nn.ReLU(),
nn.Linear(out_channels, out_channels)
)
else:
self.message_mlp = nn.Sequential(
nn.Linear(2 * in_channels, out_channels),
nn.LayerNorm(out_channels),
nn.ReLU(),
nn.Linear(out_channels, out_channels)
)
# Update function: how to update node features
self.update_mlp = nn.Sequential(
nn.Linear(in_channels + out_channels, out_channels),
nn.LayerNorm(out_channels),
nn.ReLU(),
nn.Linear(out_channels, out_channels)
)
self.edge_dim = edge_dim
def forward(self, x, edge_index, edge_attr=None):
"""
Propagate messages through the mesh graph
Args:
x: Node features [num_nodes, in_channels]
edge_index: Edge connectivity [2, num_edges]
edge_attr: Edge features [num_edges, edge_dim] (optional)
Returns:
Updated node features [num_nodes, out_channels]
"""
return self.propagate(edge_index, x=x, edge_attr=edge_attr)
def message(self, x_i, x_j, edge_attr=None):
"""
Construct messages from neighbors
Args:
x_i: Target node features
x_j: Source node features
edge_attr: Edge features
"""
if edge_attr is not None:
# Combine source node, target node, and edge features
msg_input = torch.cat([x_i, x_j, edge_attr], dim=-1)
else:
msg_input = torch.cat([x_i, x_j], dim=-1)
return self.message_mlp(msg_input)
def update(self, aggr_out, x):
"""
Update node features with aggregated messages
Args:
aggr_out: Aggregated messages from neighbors
x: Original node features
"""
# Combine original features with aggregated messages
update_input = torch.cat([x, aggr_out], dim=-1)
return self.update_mlp(update_input)
class FieldPredictorGNN(nn.Module):
"""
Graph Neural Network for predicting complete FEA fields
Architecture:
1. Node Encoder: Encode node positions, BCs, loads
2. Edge Encoder: Encode element connectivity, material properties
3. Message Passing: Propagate information through mesh (multiple layers)
4. Field Decoder: Predict displacement/stress at each node/element
This architecture respects physics:
- Uses mesh topology (forces flow through connected elements)
- Incorporates boundary conditions (fixed/loaded nodes)
- Learns material behavior (E, nu → stress-strain relationship)
"""
def __init__(
self,
node_feature_dim=3, # Node coordinates (x, y, z)
edge_feature_dim=5, # Material properties (E, nu, rho, etc.)
hidden_dim=128,
num_layers=6,
output_dim=6, # 6 DOF displacement (3 translation + 3 rotation)
dropout=0.1
):
"""
Initialize field predictor
Args:
node_feature_dim (int): Dimension of node features (position + BCs + loads)
edge_feature_dim (int): Dimension of edge features (material properties)
hidden_dim (int): Hidden layer dimension
num_layers (int): Number of message passing layers
output_dim (int): Output dimension per node (6 for displacement)
dropout (float): Dropout rate
"""
super().__init__()
self.node_feature_dim = node_feature_dim
self.edge_feature_dim = edge_feature_dim
self.hidden_dim = hidden_dim
self.num_layers = num_layers
self.output_dim = output_dim
# Node encoder: embed node coordinates + BCs + loads
self.node_encoder = nn.Sequential(
nn.Linear(node_feature_dim, hidden_dim),
nn.LayerNorm(hidden_dim),
nn.ReLU(),
nn.Dropout(dropout),
nn.Linear(hidden_dim, hidden_dim)
)
# Edge encoder: embed material properties
self.edge_encoder = nn.Sequential(
nn.Linear(edge_feature_dim, hidden_dim),
nn.LayerNorm(hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim, hidden_dim // 2)
)
# Message passing layers (the physics learning happens here)
self.conv_layers = nn.ModuleList([
MeshGraphConv(
in_channels=hidden_dim,
out_channels=hidden_dim,
edge_dim=hidden_dim // 2
)
for _ in range(num_layers)
])
self.layer_norms = nn.ModuleList([
nn.LayerNorm(hidden_dim)
for _ in range(num_layers)
])
self.dropouts = nn.ModuleList([
nn.Dropout(dropout)
for _ in range(num_layers)
])
# Field decoder: predict displacement at each node
self.field_decoder = nn.Sequential(
nn.Linear(hidden_dim, hidden_dim),
nn.LayerNorm(hidden_dim),
nn.ReLU(),
nn.Dropout(dropout),
nn.Linear(hidden_dim, hidden_dim // 2),
nn.ReLU(),
nn.Linear(hidden_dim // 2, output_dim)
)
# Physics-informed constraint layer (optional, ensures equilibrium)
self.physics_scale = nn.Parameter(torch.ones(1))
def forward(self, data):
"""
Forward pass: mesh → displacement field
Args:
data (torch_geometric.data.Data): Batch of mesh graphs containing:
- x: Node features [num_nodes, node_feature_dim]
- edge_index: Connectivity [2, num_edges]
- edge_attr: Edge features [num_edges, edge_feature_dim]
- batch: Batch assignment [num_nodes]
Returns:
displacement_field: Predicted displacement [num_nodes, output_dim]
"""
x, edge_index, edge_attr = data.x, data.edge_index, data.edge_attr
# Encode nodes (positions + BCs + loads)
x = self.node_encoder(x) # [num_nodes, hidden_dim]
# Encode edges (material properties)
if edge_attr is not None:
edge_features = self.edge_encoder(edge_attr) # [num_edges, hidden_dim//2]
else:
edge_features = None
# Message passing: learn how forces propagate through mesh
for i, (conv, norm, dropout) in enumerate(zip(
self.conv_layers, self.layer_norms, self.dropouts
)):
# Graph convolution
x_new = conv(x, edge_index, edge_features)
# Residual connection (helps gradients flow)
x = x + dropout(x_new)
# Layer normalization
x = norm(x)
# Decode to displacement field
displacement = self.field_decoder(x) # [num_nodes, output_dim]
# Apply physics-informed scaling
displacement = displacement * self.physics_scale
return displacement
def predict_stress_from_displacement(self, displacement, data, material_props):
"""
Convert predicted displacement to stress using constitutive law
This implements: σ = C : ε = C : (∇u)
Where C is the material stiffness matrix
Args:
displacement: Predicted displacement [num_nodes, 6]
data: Mesh graph data
material_props: Material properties (E, nu)
Returns:
stress_field: Predicted stress [num_elements, n_components]
"""
# This would compute strain from displacement gradients
# then apply material constitutive law
# For now, we'll predict displacement and train a separate stress predictor
raise NotImplementedError("Stress prediction implemented in StressPredictor")
class StressPredictor(nn.Module):
"""
Predicts stress field from displacement field
This can be:
1. Physics-based: Compute strain from displacement, apply constitutive law
2. Learned: Train neural network to predict stress from displacement
We use learned approach for flexibility with nonlinear materials.
"""
def __init__(self, displacement_dim=6, hidden_dim=128, stress_components=6):
"""
Args:
displacement_dim (int): Displacement DOFs per node
hidden_dim (int): Hidden layer size
stress_components (int): Stress tensor components (6 for 3D)
"""
super().__init__()
# Stress predictor network
self.stress_net = nn.Sequential(
nn.Linear(displacement_dim, hidden_dim),
nn.LayerNorm(hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim, hidden_dim),
nn.LayerNorm(hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim, stress_components)
)
def forward(self, displacement):
"""
Predict stress from displacement
Args:
displacement: [num_nodes, displacement_dim]
Returns:
stress: [num_nodes, stress_components]
"""
return self.stress_net(displacement)
class AtomizerFieldModel(nn.Module):
"""
Complete AtomizerField model: predicts both displacement and stress fields
This is the main model you'll use for training and inference.
"""
def __init__(
self,
node_feature_dim=10, # 3 (xyz) + 6 (BC DOFs) + 1 (load magnitude)
edge_feature_dim=5, # E, nu, rho, G, alpha
hidden_dim=128,
num_layers=6,
dropout=0.1
):
"""
Initialize complete field prediction model
Args:
node_feature_dim (int): Node features (coords + BCs + loads)
edge_feature_dim (int): Edge features (material properties)
hidden_dim (int): Hidden dimension
num_layers (int): Message passing layers
dropout (float): Dropout rate
"""
super().__init__()
# Displacement predictor (main GNN)
self.displacement_predictor = FieldPredictorGNN(
node_feature_dim=node_feature_dim,
edge_feature_dim=edge_feature_dim,
hidden_dim=hidden_dim,
num_layers=num_layers,
output_dim=6, # 6 DOF displacement
dropout=dropout
)
# Stress predictor (from displacement)
self.stress_predictor = StressPredictor(
displacement_dim=6,
hidden_dim=hidden_dim,
stress_components=6 # σxx, σyy, σzz, τxy, τyz, τxz
)
def forward(self, data, return_stress=True):
"""
Predict displacement and stress fields
Args:
data: Mesh graph data
return_stress (bool): Whether to predict stress
Returns:
dict with:
- displacement: [num_nodes, 6]
- stress: [num_nodes, 6] (if return_stress=True)
- von_mises: [num_nodes] (if return_stress=True)
"""
# Predict displacement
displacement = self.displacement_predictor(data)
results = {'displacement': displacement}
if return_stress:
# Predict stress from displacement
stress = self.stress_predictor(displacement)
# Calculate von Mises stress
# σ_vm = sqrt(0.5 * ((σxx-σyy)² + (σyy-σzz)² + (σzz-σxx)² + 6(τxy² + τyz² + τxz²)))
sxx, syy, szz, txy, tyz, txz = stress[:, 0], stress[:, 1], stress[:, 2], \
stress[:, 3], stress[:, 4], stress[:, 5]
von_mises = torch.sqrt(
0.5 * (
(sxx - syy)**2 + (syy - szz)**2 + (szz - sxx)**2 +
6 * (txy**2 + tyz**2 + txz**2)
)
)
results['stress'] = stress
results['von_mises'] = von_mises
return results
def get_max_values(self, results):
"""
Extract maximum values (for compatibility with scalar optimization)
Args:
results: Output from forward()
Returns:
dict with max_displacement, max_stress
"""
max_displacement = torch.max(torch.norm(results['displacement'][:, :3], dim=1))
max_stress = torch.max(results['von_mises']) if 'von_mises' in results else None
return {
'max_displacement': max_displacement,
'max_stress': max_stress
}
def create_model(config=None):
"""
Factory function to create AtomizerField model
Args:
config (dict): Model configuration
Returns:
AtomizerFieldModel instance
"""
if config is None:
config = {
'node_feature_dim': 10,
'edge_feature_dim': 5,
'hidden_dim': 128,
'num_layers': 6,
'dropout': 0.1
}
model = AtomizerFieldModel(**config)
# Initialize weights
def init_weights(m):
if isinstance(m, nn.Linear):
nn.init.kaiming_normal_(m.weight, mode='fan_in', nonlinearity='relu')
if m.bias is not None:
nn.init.constant_(m.bias, 0)
model.apply(init_weights)
return model
if __name__ == "__main__":
# Test model creation
print("Testing AtomizerField Model Creation...")
model = create_model()
print(f"Model created: {sum(p.numel() for p in model.parameters()):,} parameters")
# Create dummy data
num_nodes = 100
num_edges = 300
x = torch.randn(num_nodes, 10) # Node features
edge_index = torch.randint(0, num_nodes, (2, num_edges)) # Edge connectivity
edge_attr = torch.randn(num_edges, 5) # Edge features
batch = torch.zeros(num_nodes, dtype=torch.long) # Batch assignment
data = Data(x=x, edge_index=edge_index, edge_attr=edge_attr, batch=batch)
# Forward pass
with torch.no_grad():
results = model(data)
print(f"\nTest forward pass:")
print(f" Displacement shape: {results['displacement'].shape}")
print(f" Stress shape: {results['stress'].shape}")
print(f" Von Mises shape: {results['von_mises'].shape}")
max_vals = model.get_max_values(results)
print(f"\nMax values:")
print(f" Max displacement: {max_vals['max_displacement']:.6f}")
print(f" Max stress: {max_vals['max_stress']:.2f}")
print("\nModel test passed!")

470
atomizer-field/neural_models/parametric_predictor.py
View File

@@ -1,470 +0,0 @@
"""
parametric_predictor.py
Design-Conditioned Graph Neural Network for direct objective prediction
AtomizerField Parametric Predictor v2.0
Key Innovation:
Instead of: parameters -> FEA -> objectives (expensive)
We learn: parameters -> Neural Network -> objectives (milliseconds)
This model directly predicts all 4 optimization objectives:
- mass (g)
- frequency (Hz)
- max_displacement (mm)
- max_stress (MPa)
Architecture:
1. Design Encoder: MLP(n_design_vars -> 64 -> 128)
2. GNN Backbone: 4 layers of design-conditioned message passing
3. Global Pooling: Mean + Max pooling
4. Scalar Heads: MLP(384 -> 128 -> 64 -> 4)
This enables 2000x faster optimization with ~2-4% error.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_geometric.nn import MessagePassing, global_mean_pool, global_max_pool
from torch_geometric.data import Data
import numpy as np
from typing import Dict, Any, Optional
class DesignConditionedConv(MessagePassing):
"""
Graph Convolution layer conditioned on design parameters.
The design parameters modulate how information flows through the mesh,
allowing the network to learn design-dependent physics.
"""
def __init__(self, in_channels: int, out_channels: int, design_dim: int, edge_dim: int = None):
"""
Args:
in_channels: Input node feature dimension
out_channels: Output node feature dimension
design_dim: Design parameter dimension (after encoding)
edge_dim: Edge feature dimension (optional)
"""
super().__init__(aggr='mean')
self.in_channels = in_channels
self.out_channels = out_channels
self.design_dim = design_dim
# Design-conditioned message function
message_input_dim = 2 * in_channels + design_dim
if edge_dim is not None:
message_input_dim += edge_dim
self.message_mlp = nn.Sequential(
nn.Linear(message_input_dim, out_channels),
nn.LayerNorm(out_channels),
nn.ReLU(),
nn.Linear(out_channels, out_channels)
)
# Update function
self.update_mlp = nn.Sequential(
nn.Linear(in_channels + out_channels, out_channels),
nn.LayerNorm(out_channels),
nn.ReLU(),
nn.Linear(out_channels, out_channels)
)
self.edge_dim = edge_dim
def forward(self, x, edge_index, design_features, edge_attr=None):
"""
Forward pass with design conditioning.
Args:
x: Node features [num_nodes, in_channels]
edge_index: Edge connectivity [2, num_edges]
design_features: Design parameters [hidden] or [num_nodes, hidden]
edge_attr: Edge features [num_edges, edge_dim] (optional)
Returns:
Updated node features [num_nodes, out_channels]
"""
num_nodes = x.size(0)
# Handle different input shapes for design_features
if design_features.dim() == 1:
# Single design vector [hidden] -> broadcast to all nodes
design_broadcast = design_features.unsqueeze(0).expand(num_nodes, -1)
elif design_features.dim() == 2 and design_features.size(0) == num_nodes:
# Already per-node [num_nodes, hidden]
design_broadcast = design_features
elif design_features.dim() == 2 and design_features.size(0) == 1:
# Single design [1, hidden] -> broadcast
design_broadcast = design_features.expand(num_nodes, -1)
else:
# Fallback: take mean across batch dimension if needed
design_broadcast = design_features.mean(dim=0).unsqueeze(0).expand(num_nodes, -1)
return self.propagate(
edge_index,
x=x,
design=design_broadcast,
edge_attr=edge_attr
)
def message(self, x_i, x_j, design_i, edge_attr=None):
"""
Construct design-conditioned messages.
Args:
x_i: Target node features
x_j: Source node features
design_i: Design parameters at target nodes
edge_attr: Edge features
"""
if edge_attr is not None:
msg_input = torch.cat([x_i, x_j, design_i, edge_attr], dim=-1)
else:
msg_input = torch.cat([x_i, x_j, design_i], dim=-1)
return self.message_mlp(msg_input)
def update(self, aggr_out, x):
"""Update node features with aggregated messages."""
update_input = torch.cat([x, aggr_out], dim=-1)
return self.update_mlp(update_input)
class ParametricFieldPredictor(nn.Module):
"""
Design-conditioned GNN that predicts ALL optimization objectives from design parameters.
This is the "parametric" model that directly predicts scalar objectives,
making it much faster than field prediction followed by post-processing.
Architecture:
- Design Encoder: MLP that embeds design parameters
- Node Encoder: MLP that embeds mesh node features
- Edge Encoder: MLP that embeds material properties
- GNN Backbone: Design-conditioned message passing layers
- Global Pooling: Mean + Max pooling for graph-level representation
- Scalar Heads: MLPs that predict each objective
Outputs:
- mass: Predicted mass (grams)
- frequency: Predicted fundamental frequency (Hz)
- max_displacement: Maximum displacement magnitude (mm)
- max_stress: Maximum von Mises stress (MPa)
"""
def __init__(self, config: Dict[str, Any] = None):
"""
Initialize parametric predictor.
Args:
config: Model configuration dict with keys:
- input_channels: Node feature dimension (default: 12)
- edge_dim: Edge feature dimension (default: 5)
- hidden_channels: Hidden layer size (default: 128)
- num_layers: Number of GNN layers (default: 4)
- design_dim: Design parameter dimension (default: 4)
- dropout: Dropout rate (default: 0.1)
"""
super().__init__()
# Default configuration
if config is None:
config = {}
self.input_channels = config.get('input_channels', 12)
self.edge_dim = config.get('edge_dim', 5)
self.hidden_channels = config.get('hidden_channels', 128)
self.num_layers = config.get('num_layers', 4)
self.design_dim = config.get('design_dim', 4)
self.dropout_rate = config.get('dropout', 0.1)
# Store config for checkpoint saving
self.config = {
'input_channels': self.input_channels,
'edge_dim': self.edge_dim,
'hidden_channels': self.hidden_channels,
'num_layers': self.num_layers,
'design_dim': self.design_dim,
'dropout': self.dropout_rate
}
# === DESIGN ENCODER ===
# Embeds design parameters into a higher-dimensional space
self.design_encoder = nn.Sequential(
nn.Linear(self.design_dim, 64),
nn.LayerNorm(64),
nn.ReLU(),
nn.Dropout(self.dropout_rate),
nn.Linear(64, self.hidden_channels),
nn.LayerNorm(self.hidden_channels),
nn.ReLU()
)
# === NODE ENCODER ===
# Embeds node features (coordinates, BCs, loads)
self.node_encoder = nn.Sequential(
nn.Linear(self.input_channels, self.hidden_channels),
nn.LayerNorm(self.hidden_channels),
nn.ReLU(),
nn.Dropout(self.dropout_rate),
nn.Linear(self.hidden_channels, self.hidden_channels)
)
# === EDGE ENCODER ===
# Embeds edge features (material properties)
self.edge_encoder = nn.Sequential(
nn.Linear(self.edge_dim, self.hidden_channels),
nn.LayerNorm(self.hidden_channels),
nn.ReLU(),
nn.Linear(self.hidden_channels, self.hidden_channels // 2)
)
# === GNN BACKBONE ===
# Design-conditioned message passing layers
self.conv_layers = nn.ModuleList([
DesignConditionedConv(
in_channels=self.hidden_channels,
out_channels=self.hidden_channels,
design_dim=self.hidden_channels,
edge_dim=self.hidden_channels // 2
)
for _ in range(self.num_layers)
])
self.layer_norms = nn.ModuleList([
nn.LayerNorm(self.hidden_channels)
for _ in range(self.num_layers)
])
self.dropouts = nn.ModuleList([
nn.Dropout(self.dropout_rate)
for _ in range(self.num_layers)
])
# === GLOBAL POOLING ===
# Mean + Max pooling gives 2 * hidden_channels features
# Plus design features gives 3 * hidden_channels total
pooled_dim = 3 * self.hidden_channels
# === SCALAR PREDICTION HEADS ===
# Each head predicts one objective
self.mass_head = nn.Sequential(
nn.Linear(pooled_dim, self.hidden_channels),
nn.LayerNorm(self.hidden_channels),
nn.ReLU(),
nn.Dropout(self.dropout_rate),
nn.Linear(self.hidden_channels, 64),
nn.ReLU(),
nn.Linear(64, 1)
)
self.frequency_head = nn.Sequential(
nn.Linear(pooled_dim, self.hidden_channels),
nn.LayerNorm(self.hidden_channels),
nn.ReLU(),
nn.Dropout(self.dropout_rate),
nn.Linear(self.hidden_channels, 64),
nn.ReLU(),
nn.Linear(64, 1)
)
self.displacement_head = nn.Sequential(
nn.Linear(pooled_dim, self.hidden_channels),
nn.LayerNorm(self.hidden_channels),
nn.ReLU(),
nn.Dropout(self.dropout_rate),
nn.Linear(self.hidden_channels, 64),
nn.ReLU(),
nn.Linear(64, 1)
)
self.stress_head = nn.Sequential(
nn.Linear(pooled_dim, self.hidden_channels),
nn.LayerNorm(self.hidden_channels),
nn.ReLU(),
nn.Dropout(self.dropout_rate),
nn.Linear(self.hidden_channels, 64),
nn.ReLU(),
nn.Linear(64, 1)
)
# === OPTIONAL FIELD DECODER ===
# For returning displacement field if requested
self.field_decoder = nn.Sequential(
nn.Linear(self.hidden_channels, self.hidden_channels),
nn.LayerNorm(self.hidden_channels),
nn.ReLU(),
nn.Dropout(self.dropout_rate),
nn.Linear(self.hidden_channels, 6) # 6 DOF displacement
)
def forward(
self,
data: Data,
design_params: torch.Tensor,
return_fields: bool = False
) -> Dict[str, torch.Tensor]:
"""
Forward pass: predict objectives from mesh + design parameters.
Args:
data: PyTorch Geometric Data object with:
- x: Node features [num_nodes, input_channels]
- edge_index: Edge connectivity [2, num_edges]
- edge_attr: Edge features [num_edges, edge_dim]
- batch: Batch assignment [num_nodes] (optional)
design_params: Normalized design parameters [design_dim] or [batch, design_dim]
return_fields: If True, also return displacement field prediction
Returns:
Dict with:
- mass: Predicted mass [batch_size]
- frequency: Predicted frequency [batch_size]
- max_displacement: Predicted max displacement [batch_size]
- max_stress: Predicted max stress [batch_size]
- displacement: (optional) Displacement field [num_nodes, 6]
"""
x, edge_index, edge_attr = data.x, data.edge_index, data.edge_attr
num_nodes = x.size(0)
# Handle design params shape - ensure 2D [batch_size, design_dim]
if design_params.dim() == 1:
design_params = design_params.unsqueeze(0)
batch_size = design_params.size(0)
# Encode design parameters: [batch_size, design_dim] -> [batch_size, hidden]
design_encoded = self.design_encoder(design_params)
# Encode nodes (shared across all designs)
x_encoded = self.node_encoder(x) # [num_nodes, hidden]
# Encode edges (shared across all designs)
if edge_attr is not None:
edge_features = self.edge_encoder(edge_attr) # [num_edges, hidden//2]
else:
edge_features = None
# Process each design in the batch
all_graph_features = []
for i in range(batch_size):
# Get design for this sample
design_i = design_encoded[i] # [hidden]
# Reset node features for this sample
x = x_encoded.clone()
# Message passing with design conditioning
for conv, norm, dropout in zip(self.conv_layers, self.layer_norms, self.dropouts):
x_new = conv(x, edge_index, design_i, edge_features)
x = x + dropout(x_new) # Residual connection
x = norm(x)
# Global pooling for this sample
batch_idx = torch.zeros(num_nodes, dtype=torch.long, device=x.device)
x_mean = global_mean_pool(x, batch_idx) # [1, hidden]
x_max = global_max_pool(x, batch_idx) # [1, hidden]
# Concatenate pooled + design features
graph_feat = torch.cat([x_mean, x_max, design_encoded[i:i+1]], dim=-1) # [1, 3*hidden]
all_graph_features.append(graph_feat)
# Stack all samples
graph_features = torch.cat(all_graph_features, dim=0) # [batch_size, 3*hidden]
# Predict objectives
mass = self.mass_head(graph_features).squeeze(-1)
frequency = self.frequency_head(graph_features).squeeze(-1)
max_displacement = self.displacement_head(graph_features).squeeze(-1)
max_stress = self.stress_head(graph_features).squeeze(-1)
results = {
'mass': mass,
'frequency': frequency,
'max_displacement': max_displacement,
'max_stress': max_stress
}
# Optionally return displacement field (uses last processed x)
if return_fields:
displacement_field = self.field_decoder(x) # [num_nodes, 6]
results['displacement'] = displacement_field
return results
def get_num_parameters(self) -> int:
"""Get total number of trainable parameters."""
return sum(p.numel() for p in self.parameters() if p.requires_grad)
def create_parametric_model(config: Dict[str, Any] = None) -> ParametricFieldPredictor:
"""
Factory function to create parametric predictor model.
Args:
config: Model configuration dictionary
Returns:
Initialized ParametricFieldPredictor
"""
model = ParametricFieldPredictor(config)
# Initialize weights
def init_weights(m):
if isinstance(m, nn.Linear):
nn.init.kaiming_normal_(m.weight, mode='fan_in', nonlinearity='relu')
if m.bias is not None:
nn.init.constant_(m.bias, 0)
model.apply(init_weights)
return model
if __name__ == "__main__":
print("Testing Parametric Field Predictor...")
print("=" * 60)
# Create model with default config
model = create_parametric_model()
n_params = model.get_num_parameters()
print(f"Model created: {n_params:,} parameters")
print(f"Config: {model.config}")
# Create dummy data
num_nodes = 500
num_edges = 2000
x = torch.randn(num_nodes, 12) # Node features
edge_index = torch.randint(0, num_nodes, (2, num_edges))
edge_attr = torch.randn(num_edges, 5)
batch = torch.zeros(num_nodes, dtype=torch.long)
data = Data(x=x, edge_index=edge_index, edge_attr=edge_attr, batch=batch)
# Design parameters
design_params = torch.randn(4) # 4 design variables
# Forward pass
print("\nRunning forward pass...")
with torch.no_grad():
results = model(data, design_params, return_fields=True)
print(f"\nPredictions:")
print(f" Mass: {results['mass'].item():.4f}")
print(f" Frequency: {results['frequency'].item():.4f}")
print(f" Max Displacement: {results['max_displacement'].item():.6f}")
print(f" Max Stress: {results['max_stress'].item():.2f}")
if 'displacement' in results:
print(f" Displacement field shape: {results['displacement'].shape}")
print("\n" + "=" * 60)
print("Parametric predictor test PASSED!")

449
atomizer-field/neural_models/physics_losses.py
View File

@@ -1,449 +0,0 @@
"""
physics_losses.py
Physics-informed loss functions for training FEA field predictors
AtomizerField Physics-Informed Loss Functions v2.0
Key Innovation:
Standard neural networks only minimize prediction error.
Physics-informed networks also enforce physical laws:
- Equilibrium: Forces must balance
- Compatibility: Strains must be compatible with displacements
- Constitutive: Stress must follow material law (σ = C:ε)
This makes the network learn physics, not just patterns.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
class PhysicsInformedLoss(nn.Module):
"""
Combined loss function with physics constraints
Total Loss = λ_data * L_data + λ_physics * L_physics
Where:
- L_data: Standard MSE between prediction and FEA ground truth
- L_physics: Physics violation penalty (equilibrium, compatibility, constitutive)
"""
def __init__(
self,
lambda_data=1.0,
lambda_equilibrium=0.1,
lambda_constitutive=0.1,
lambda_boundary=1.0,
use_relative_error=True
):
"""
Initialize physics-informed loss
Args:
lambda_data (float): Weight for data loss
lambda_equilibrium (float): Weight for equilibrium violation
lambda_constitutive (float): Weight for constitutive law violation
lambda_boundary (float): Weight for boundary condition violation
use_relative_error (bool): Use relative error instead of absolute
"""
super().__init__()
self.lambda_data = lambda_data
self.lambda_equilibrium = lambda_equilibrium
self.lambda_constitutive = lambda_constitutive
self.lambda_boundary = lambda_boundary
self.use_relative_error = use_relative_error
def forward(self, predictions, targets, data=None):
"""
Compute total physics-informed loss
Args:
predictions (dict): Model predictions
- displacement: [num_nodes, 6]
- stress: [num_nodes, 6]
- von_mises: [num_nodes]
targets (dict): Ground truth from FEA
- displacement: [num_nodes, 6]
- stress: [num_nodes, 6]
data: Mesh graph data (for physics constraints)
Returns:
dict with:
- total_loss: Combined loss
- data_loss: Data fitting loss
- equilibrium_loss: Equilibrium violation
- constitutive_loss: Material law violation
- boundary_loss: BC violation
"""
losses = {}
# 1. Data Loss: How well do predictions match FEA results?
losses['displacement_loss'] = self._displacement_loss(
predictions['displacement'],
targets['displacement']
)
if 'stress' in predictions and 'stress' in targets:
losses['stress_loss'] = self._stress_loss(
predictions['stress'],
targets['stress']
)
else:
losses['stress_loss'] = torch.tensor(0.0, device=predictions['displacement'].device)
losses['data_loss'] = losses['displacement_loss'] + losses['stress_loss']
# 2. Physics Losses: How well do predictions obey physics?
if data is not None:
# Equilibrium: ∇·σ + f = 0
losses['equilibrium_loss'] = self._equilibrium_loss(
predictions, data
)
# Constitutive: σ = C:ε
losses['constitutive_loss'] = self._constitutive_loss(
predictions, data
)
# Boundary conditions: u = 0 at fixed nodes
losses['boundary_loss'] = self._boundary_condition_loss(
predictions, data
)
else:
losses['equilibrium_loss'] = torch.tensor(0.0, device=predictions['displacement'].device)
losses['constitutive_loss'] = torch.tensor(0.0, device=predictions['displacement'].device)
losses['boundary_loss'] = torch.tensor(0.0, device=predictions['displacement'].device)
# Total loss
losses['total_loss'] = (
self.lambda_data * losses['data_loss'] +
self.lambda_equilibrium * losses['equilibrium_loss'] +
self.lambda_constitutive * losses['constitutive_loss'] +
self.lambda_boundary * losses['boundary_loss']
)
return losses
def _displacement_loss(self, pred, target):
"""
Loss for displacement field
Uses relative error to handle different displacement magnitudes
"""
if self.use_relative_error:
# Relative L2 error
diff = pred - target
rel_error = torch.norm(diff, dim=-1) / (torch.norm(target, dim=-1) + 1e-8)
return rel_error.mean()
else:
# Absolute MSE
return F.mse_loss(pred, target)
def _stress_loss(self, pred, target):
"""
Loss for stress field
Emphasizes von Mises stress (most important for failure prediction)
"""
# Component-wise MSE
component_loss = F.mse_loss(pred, target)
# Von Mises stress MSE (computed from components)
pred_vm = self._compute_von_mises(pred)
target_vm = self._compute_von_mises(target)
vm_loss = F.mse_loss(pred_vm, target_vm)
# Combined: 50% component accuracy, 50% von Mises accuracy
return 0.5 * component_loss + 0.5 * vm_loss
def _equilibrium_loss(self, predictions, data):
"""
Equilibrium loss: ∇·σ + f = 0
In discrete form: sum of forces at each node should be zero
(where not externally loaded)
This is expensive to compute exactly, so we use a simplified version:
Check force balance on each element
"""
# Simplified: For now, return zero (full implementation requires
# computing stress divergence from node stresses)
# TODO: Implement finite difference approximation of ∇·σ
return torch.tensor(0.0, device=predictions['displacement'].device)
def _constitutive_loss(self, predictions, data):
"""
Constitutive law loss: σ = C:ε
Check if predicted stress is consistent with predicted strain
(which comes from displacement gradient)
Simplified version: Check if stress-strain relationship is reasonable
"""
# Simplified: For now, return zero
# Full implementation would:
# 1. Compute strain from displacement gradient
# 2. Compute expected stress from strain using material stiffness
# 3. Compare with predicted stress
# TODO: Implement strain computation and constitutive check
return torch.tensor(0.0, device=predictions['displacement'].device)
def _boundary_condition_loss(self, predictions, data):
"""
Boundary condition loss: u = 0 at fixed DOFs
Penalize non-zero displacement at constrained nodes
"""
if not hasattr(data, 'bc_mask') or data.bc_mask is None:
return torch.tensor(0.0, device=predictions['displacement'].device)
# bc_mask: [num_nodes, 6] boolean mask where True = constrained
displacement = predictions['displacement']
bc_mask = data.bc_mask
# Compute penalty for non-zero displacement at constrained DOFs
constrained_displacement = displacement * bc_mask.float()
bc_loss = torch.mean(constrained_displacement ** 2)
return bc_loss
def _compute_von_mises(self, stress):
"""
Compute von Mises stress from stress tensor components
Args:
stress: [num_nodes, 6] with [σxx, σyy, σzz, τxy, τyz, τxz]
Returns:
von_mises: [num_nodes]
"""
sxx, syy, szz = stress[:, 0], stress[:, 1], stress[:, 2]
txy, tyz, txz = stress[:, 3], stress[:, 4], stress[:, 5]
vm = torch.sqrt(
0.5 * (
(sxx - syy)**2 + (syy - szz)**2 + (szz - sxx)**2 +
6 * (txy**2 + tyz**2 + txz**2)
)
)
return vm
class FieldMSELoss(nn.Module):
"""
Simple MSE loss for field prediction (no physics constraints)
Use this for initial training or when physics constraints are too strict.
"""
def __init__(self, weight_displacement=1.0, weight_stress=1.0):
"""
Args:
weight_displacement (float): Weight for displacement loss
weight_stress (float): Weight for stress loss
"""
super().__init__()
self.weight_displacement = weight_displacement
self.weight_stress = weight_stress
def forward(self, predictions, targets):
"""
Compute MSE loss
Args:
predictions (dict): Model outputs
targets (dict): Ground truth
Returns:
dict with loss components
"""
losses = {}
# Displacement MSE
losses['displacement_loss'] = F.mse_loss(
predictions['displacement'],
targets['displacement']
)
# Stress MSE (if available)
if 'stress' in predictions and 'stress' in targets:
losses['stress_loss'] = F.mse_loss(
predictions['stress'],
targets['stress']
)
else:
losses['stress_loss'] = torch.tensor(0.0, device=predictions['displacement'].device)
# Total loss
losses['total_loss'] = (
self.weight_displacement * losses['displacement_loss'] +
self.weight_stress * losses['stress_loss']
)
return losses
class RelativeFieldLoss(nn.Module):
"""
Relative error loss - better for varying displacement/stress magnitudes
Uses: ||pred - target|| / ||target||
This makes the loss scale-invariant.
"""
def __init__(self, epsilon=1e-8):
"""
Args:
epsilon (float): Small constant to avoid division by zero
"""
super().__init__()
self.epsilon = epsilon
def forward(self, predictions, targets):
"""
Compute relative error loss
Args:
predictions (dict): Model outputs
targets (dict): Ground truth
Returns:
dict with loss components
"""
losses = {}
# Relative displacement error
disp_diff = predictions['displacement'] - targets['displacement']
disp_norm_pred = torch.norm(disp_diff, dim=-1)
disp_norm_target = torch.norm(targets['displacement'], dim=-1)
losses['displacement_loss'] = (disp_norm_pred / (disp_norm_target + self.epsilon)).mean()
# Relative stress error
if 'stress' in predictions and 'stress' in targets:
stress_diff = predictions['stress'] - targets['stress']
stress_norm_pred = torch.norm(stress_diff, dim=-1)
stress_norm_target = torch.norm(targets['stress'], dim=-1)
losses['stress_loss'] = (stress_norm_pred / (stress_norm_target + self.epsilon)).mean()
else:
losses['stress_loss'] = torch.tensor(0.0, device=predictions['displacement'].device)
# Total loss
losses['total_loss'] = losses['displacement_loss'] + losses['stress_loss']
return losses
class MaxValueLoss(nn.Module):
"""
Loss on maximum values only (for backward compatibility with scalar optimization)
This is useful if you want to ensure the network gets the critical max values right,
even if the field distribution is slightly off.
"""
def __init__(self):
super().__init__()
def forward(self, predictions, targets):
"""
Compute loss on maximum displacement and stress
Args:
predictions (dict): Model outputs with 'displacement', 'von_mises'
targets (dict): Ground truth
Returns:
dict with loss components
"""
losses = {}
# Max displacement error
pred_max_disp = torch.max(torch.norm(predictions['displacement'][:, :3], dim=1))
target_max_disp = torch.max(torch.norm(targets['displacement'][:, :3], dim=1))
losses['max_displacement_loss'] = F.mse_loss(pred_max_disp, target_max_disp)
# Max von Mises stress error
if 'von_mises' in predictions and 'stress' in targets:
pred_max_vm = torch.max(predictions['von_mises'])
# Compute target von Mises
target_stress = targets['stress']
sxx, syy, szz = target_stress[:, 0], target_stress[:, 1], target_stress[:, 2]
txy, tyz, txz = target_stress[:, 3], target_stress[:, 4], target_stress[:, 5]
target_vm = torch.sqrt(
0.5 * ((sxx - syy)**2 + (syy - szz)**2 + (szz - sxx)**2 +
6 * (txy**2 + tyz**2 + txz**2))
)
target_max_vm = torch.max(target_vm)
losses['max_stress_loss'] = F.mse_loss(pred_max_vm, target_max_vm)
else:
losses['max_stress_loss'] = torch.tensor(0.0, device=predictions['displacement'].device)
# Total loss
losses['total_loss'] = losses['max_displacement_loss'] + losses['max_stress_loss']
return losses
def create_loss_function(loss_type='mse', config=None):
"""
Factory function to create loss function
Args:
loss_type (str): Type of loss ('mse', 'relative', 'physics', 'max')
config (dict): Loss function configuration
Returns:
Loss function instance
"""
if config is None:
config = {}
if loss_type == 'mse':
return FieldMSELoss(**config)
elif loss_type == 'relative':
return RelativeFieldLoss(**config)
elif loss_type == 'physics':
return PhysicsInformedLoss(**config)
elif loss_type == 'max':
return MaxValueLoss(**config)
else:
raise ValueError(f"Unknown loss type: {loss_type}")
if __name__ == "__main__":
# Test loss functions
print("Testing AtomizerField Loss Functions...\n")
# Create dummy predictions and targets
num_nodes = 100
pred = {
'displacement': torch.randn(num_nodes, 6),
'stress': torch.randn(num_nodes, 6),
'von_mises': torch.abs(torch.randn(num_nodes))
}
target = {
'displacement': torch.randn(num_nodes, 6),
'stress': torch.randn(num_nodes, 6)
}
# Test each loss function
loss_types = ['mse', 'relative', 'physics', 'max']
for loss_type in loss_types:
print(f"Testing {loss_type.upper()} loss...")
loss_fn = create_loss_function(loss_type)
losses = loss_fn(pred, target)
print(f" Total loss: {losses['total_loss']:.6f}")
for key, value in losses.items():
if key != 'total_loss':
print(f" {key}: {value:.6f}")
print()
print("Loss function tests passed!")

361
atomizer-field/neural_models/uncertainty.py
View File

@@ -1,361 +0,0 @@
"""
uncertainty.py
Uncertainty quantification for neural field predictions
AtomizerField Uncertainty Quantification v2.1
Know when to trust predictions and when to run FEA!
Key Features:
- Ensemble-based uncertainty estimation
- Confidence intervals for predictions
- Automatic FEA recommendation
- Online calibration
"""
import torch
import torch.nn as nn
import numpy as np
from copy import deepcopy
from .field_predictor import AtomizerFieldModel
class UncertainFieldPredictor(nn.Module):
"""
Ensemble of models for uncertainty quantification
Uses multiple models trained with different initializations
to estimate prediction uncertainty.
When uncertainty is high → Recommend FEA validation
When uncertainty is low → Trust neural prediction
"""
def __init__(self, base_model_config, n_ensemble=5):
"""
Initialize ensemble
Args:
base_model_config (dict): Configuration for base model
n_ensemble (int): Number of models in ensemble
"""
super().__init__()
print(f"\nCreating ensemble with {n_ensemble} models...")
# Create ensemble of models
self.models = nn.ModuleList([
AtomizerFieldModel(**base_model_config)
for _ in range(n_ensemble)
])
self.n_ensemble = n_ensemble
# Initialize each model differently
for i, model in enumerate(self.models):
self._init_weights(model, seed=i)
print(f"Ensemble created with {n_ensemble} models")
def _init_weights(self, model, seed):
"""Initialize model weights with different seed"""
torch.manual_seed(seed)
def init_fn(m):
if isinstance(m, nn.Linear):
nn.init.kaiming_normal_(m.weight, mode='fan_in', nonlinearity='relu')
if m.bias is not None:
nn.init.constant_(m.bias, 0)
model.apply(init_fn)
def forward(self, data, return_uncertainty=True, return_all_predictions=False):
"""
Forward pass through ensemble
Args:
data: Input graph data
return_uncertainty (bool): Return uncertainty estimates
return_all_predictions (bool): Return all individual predictions
Returns:
dict: Predictions with uncertainty
- displacement: Mean prediction
- stress: Mean prediction
- von_mises: Mean prediction
- displacement_std: Standard deviation (if return_uncertainty)
- stress_std: Standard deviation (if return_uncertainty)
- von_mises_std: Standard deviation (if return_uncertainty)
- all_predictions: List of all predictions (if return_all_predictions)
"""
# Get predictions from all models
all_predictions = []
for model in self.models:
with torch.no_grad():
pred = model(data, return_stress=True)
all_predictions.append(pred)
# Stack predictions
displacement_stack = torch.stack([p['displacement'] for p in all_predictions])
stress_stack = torch.stack([p['stress'] for p in all_predictions])
von_mises_stack = torch.stack([p['von_mises'] for p in all_predictions])
# Compute mean predictions
results = {
'displacement': displacement_stack.mean(dim=0),
'stress': stress_stack.mean(dim=0),
'von_mises': von_mises_stack.mean(dim=0)
}
# Compute uncertainty (standard deviation across ensemble)
if return_uncertainty:
results['displacement_std'] = displacement_stack.std(dim=0)
results['stress_std'] = stress_stack.std(dim=0)
results['von_mises_std'] = von_mises_stack.std(dim=0)
# Overall uncertainty metrics
results['max_displacement_uncertainty'] = results['displacement_std'].max().item()
results['max_stress_uncertainty'] = results['von_mises_std'].max().item()
# Uncertainty as percentage of prediction
results['displacement_rel_uncertainty'] = (
results['displacement_std'] / (torch.abs(results['displacement']) + 1e-8)
).mean().item()
results['stress_rel_uncertainty'] = (
results['von_mises_std'] / (results['von_mises'] + 1e-8)
).mean().item()
# Return all predictions if requested
if return_all_predictions:
results['all_predictions'] = all_predictions
return results
def needs_fea_validation(self, predictions, threshold=0.1):
"""
Determine if FEA validation is recommended
Args:
predictions (dict): Output from forward() with uncertainty
threshold (float): Relative uncertainty threshold
Returns:
dict: Recommendation and reasons
"""
reasons = []
# Check displacement uncertainty
if predictions['displacement_rel_uncertainty'] > threshold:
reasons.append(
f"High displacement uncertainty: "
f"{predictions['displacement_rel_uncertainty']*100:.1f}% > {threshold*100:.1f}%"
)
# Check stress uncertainty
if predictions['stress_rel_uncertainty'] > threshold:
reasons.append(
f"High stress uncertainty: "
f"{predictions['stress_rel_uncertainty']*100:.1f}% > {threshold*100:.1f}%"
)
recommend_fea = len(reasons) > 0
return {
'recommend_fea': recommend_fea,
'reasons': reasons,
'displacement_uncertainty': predictions['displacement_rel_uncertainty'],
'stress_uncertainty': predictions['stress_rel_uncertainty']
}
def get_confidence_intervals(self, predictions, confidence=0.95):
"""
Compute confidence intervals for predictions
Args:
predictions (dict): Output from forward() with uncertainty
confidence (float): Confidence level (0.95 = 95% confidence)
Returns:
dict: Confidence intervals
"""
# For normal distribution, 95% CI is ±1.96 std
# For 90% CI is ±1.645 std
z_score = {0.90: 1.645, 0.95: 1.96, 0.99: 2.576}.get(confidence, 1.96)
intervals = {}
# Displacement intervals
intervals['displacement_lower'] = predictions['displacement'] - z_score * predictions['displacement_std']
intervals['displacement_upper'] = predictions['displacement'] + z_score * predictions['displacement_std']
# Stress intervals
intervals['von_mises_lower'] = predictions['von_mises'] - z_score * predictions['von_mises_std']
intervals['von_mises_upper'] = predictions['von_mises'] + z_score * predictions['von_mises_std']
# Max values with confidence intervals
max_vm = predictions['von_mises'].max()
max_vm_std = predictions['von_mises_std'].max()
intervals['max_stress_estimate'] = max_vm.item()
intervals['max_stress_lower'] = (max_vm - z_score * max_vm_std).item()
intervals['max_stress_upper'] = (max_vm + z_score * max_vm_std).item()
return intervals
class OnlineLearner:
"""
Online learning from FEA runs during optimization
As optimization progresses and you run FEA for validation,
this module can quickly update the model to improve predictions.
This creates a virtuous cycle:
1. Use neural network for fast exploration
2. Run FEA on promising designs
3. Update neural network with new data
4. Neural network gets better → need less FEA
"""
def __init__(self, model, learning_rate=0.0001):
"""
Initialize online learner
Args:
model: Neural network model
learning_rate (float): Learning rate for updates
"""
self.model = model
self.optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
self.replay_buffer = []
self.update_count = 0
print(f"\nOnline learner initialized")
print(f"Learning rate: {learning_rate}")
def add_fea_result(self, graph_data, fea_results):
"""
Add new FEA result to replay buffer
Args:
graph_data: Mesh graph
fea_results (dict): FEA results (displacement, stress)
"""
self.replay_buffer.append({
'graph_data': graph_data,
'fea_results': fea_results
})
print(f"Added FEA result to buffer (total: {len(self.replay_buffer)})")
def quick_update(self, steps=10):
"""
Quick fine-tuning on recent FEA results
Args:
steps (int): Number of gradient steps
"""
if len(self.replay_buffer) == 0:
print("No data in replay buffer")
return
print(f"\nQuick update: {steps} steps on {len(self.replay_buffer)} samples")
self.model.train()
for step in range(steps):
total_loss = 0.0
# Train on all samples in buffer
for sample in self.replay_buffer:
graph_data = sample['graph_data']
fea_results = sample['fea_results']
# Forward pass
predictions = self.model(graph_data, return_stress=True)
# Compute loss
disp_loss = nn.functional.mse_loss(
predictions['displacement'],
fea_results['displacement']
)
if 'stress' in fea_results:
stress_loss = nn.functional.mse_loss(
predictions['stress'],
fea_results['stress']
)
loss = disp_loss + stress_loss
else:
loss = disp_loss
# Backward pass
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
total_loss += loss.item()
if step % 5 == 0:
avg_loss = total_loss / len(self.replay_buffer)
print(f" Step {step}/{steps}: Loss = {avg_loss:.6f}")
self.model.eval()
self.update_count += 1
print(f"Update complete (total updates: {self.update_count})")
def clear_buffer(self):
"""Clear replay buffer"""
self.replay_buffer = []
print("Replay buffer cleared")
def create_uncertain_predictor(model_config, n_ensemble=5):
"""
Factory function to create uncertain predictor
Args:
model_config (dict): Model configuration
n_ensemble (int): Ensemble size
Returns:
UncertainFieldPredictor instance
"""
return UncertainFieldPredictor(model_config, n_ensemble)
if __name__ == "__main__":
# Test uncertainty quantification
print("Testing Uncertainty Quantification...\n")
# Create ensemble
model_config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.1
}
ensemble = UncertainFieldPredictor(model_config, n_ensemble=3)
print(f"\nEnsemble created with {ensemble.n_ensemble} models")
print("Uncertainty quantification ready!")
print("\nUsage:")
print("""
# Get predictions with uncertainty
predictions = ensemble(graph_data, return_uncertainty=True)
# Check if FEA validation needed
recommendation = ensemble.needs_fea_validation(predictions, threshold=0.1)
if recommendation['recommend_fea']:
print("Recommendation: Run FEA for validation")
for reason in recommendation['reasons']:
print(f" - {reason}")
else:
print("Prediction confident - no FEA needed!")
""")

421
atomizer-field/optimization_interface.py
View File

@@ -1,421 +0,0 @@
"""
optimization_interface.py
Bridge between AtomizerField neural network and Atomizer optimization platform
AtomizerField Optimization Interface v2.1
Enables gradient-based optimization with neural field predictions.
Key Features:
- Drop-in replacement for FEA evaluation (1000× faster)
- Gradient computation for sensitivity analysis
- Field-aware optimization (knows WHERE stress occurs)
- Uncertainty quantification (knows when to trust predictions)
- Automatic FEA fallback for high-uncertainty cases
"""
import torch
import torch.nn.functional as F
import numpy as np
from pathlib import Path
import json
import time
from neural_models.field_predictor import AtomizerFieldModel
from neural_models.data_loader import FEAMeshDataset
class NeuralFieldOptimizer:
"""
Optimization interface for AtomizerField
This class provides a simple API for optimization:
- evaluate(parameters) → objectives (max_stress, max_disp, etc.)
- get_sensitivities(parameters) → gradients for optimization
- get_fields(parameters) → complete stress/displacement fields
Usage:
optimizer = NeuralFieldOptimizer('checkpoint_best.pt')
results = optimizer.evaluate(parameters)
print(f"Max stress: {results['max_stress']:.2f} MPa")
"""
def __init__(
self,
model_path,
uncertainty_threshold=0.1,
enable_gradients=True,
device=None
):
"""
Initialize optimizer
Args:
model_path (str): Path to trained model checkpoint
uncertainty_threshold (float): Uncertainty above which to recommend FEA
enable_gradients (bool): Enable gradient computation
device (str): Device to run on ('cuda' or 'cpu')
"""
if device is None:
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
else:
self.device = torch.device(device)
print(f"\nAtomizerField Optimization Interface v2.1")
print(f"Device: {self.device}")
# Load model
print(f"Loading model from {model_path}...")
checkpoint = torch.load(model_path, map_location=self.device)
# Create model
model_config = checkpoint['config']['model']
self.model = AtomizerFieldModel(**model_config)
self.model.load_state_dict(checkpoint['model_state_dict'])
self.model = self.model.to(self.device)
self.model.eval()
self.config = checkpoint['config']
self.uncertainty_threshold = uncertainty_threshold
self.enable_gradients = enable_gradients
# Model info
self.model_info = {
'version': checkpoint.get('epoch', 'unknown'),
'best_val_loss': checkpoint.get('best_val_loss', 'unknown'),
'training_config': checkpoint['config']
}
print(f"Model loaded successfully!")
print(f" Epoch: {checkpoint.get('epoch', 'N/A')}")
print(f" Validation loss: {checkpoint.get('best_val_loss', 'N/A')}")
# Statistics for tracking
self.eval_count = 0
self.total_time = 0.0
def evaluate(self, graph_data, return_fields=False):
"""
Evaluate design using neural network (drop-in FEA replacement)
Args:
graph_data: PyTorch Geometric Data object with mesh graph
return_fields (bool): Return complete fields or just objectives
Returns:
dict: Optimization objectives and optionally complete fields
- max_stress: Maximum von Mises stress (MPa)
- max_displacement: Maximum displacement (mm)
- mass: Total mass (kg) if available
- fields: Complete stress/displacement fields (if return_fields=True)
- inference_time_ms: Prediction time
- uncertainty: Prediction uncertainty (if ensemble enabled)
"""
start_time = time.time()
# Move to device
graph_data = graph_data.to(self.device)
# Predict
with torch.set_grad_enabled(self.enable_gradients):
predictions = self.model(graph_data, return_stress=True)
inference_time = (time.time() - start_time) * 1000 # ms
# Extract objectives
max_displacement = torch.max(
torch.norm(predictions['displacement'][:, :3], dim=1)
).item()
max_stress = torch.max(predictions['von_mises']).item()
results = {
'max_stress': max_stress,
'max_displacement': max_displacement,
'inference_time_ms': inference_time,
'evaluation_count': self.eval_count
}
# Add complete fields if requested
if return_fields:
results['fields'] = {
'displacement': predictions['displacement'].cpu().detach().numpy(),
'stress': predictions['stress'].cpu().detach().numpy(),
'von_mises': predictions['von_mises'].cpu().detach().numpy()
}
# Update statistics
self.eval_count += 1
self.total_time += inference_time
return results
def get_sensitivities(self, graph_data, objective='max_stress'):
"""
Compute gradients for gradient-based optimization
This enables MUCH faster optimization than finite differences!
Args:
graph_data: PyTorch Geometric Data with requires_grad=True
objective (str): Which objective to differentiate ('max_stress' or 'max_displacement')
Returns:
dict: Gradients with respect to input features
- node_gradients: ∂objective/∂node_features
- edge_gradients: ∂objective/∂edge_features
"""
if not self.enable_gradients:
raise RuntimeError("Gradients not enabled. Set enable_gradients=True")
# Enable gradients
graph_data = graph_data.to(self.device)
graph_data.x.requires_grad_(True)
if graph_data.edge_attr is not None:
graph_data.edge_attr.requires_grad_(True)
# Forward pass
predictions = self.model(graph_data, return_stress=True)
# Compute objective
if objective == 'max_stress':
obj = torch.max(predictions['von_mises'])
elif objective == 'max_displacement':
disp_mag = torch.norm(predictions['displacement'][:, :3], dim=1)
obj = torch.max(disp_mag)
else:
raise ValueError(f"Unknown objective: {objective}")
# Backward pass
obj.backward()
# Extract gradients
gradients = {
'node_gradients': graph_data.x.grad.cpu().numpy(),
'objective_value': obj.item()
}
if graph_data.edge_attr is not None and graph_data.edge_attr.grad is not None:
gradients['edge_gradients'] = graph_data.edge_attr.grad.cpu().numpy()
return gradients
def batch_evaluate(self, graph_data_list, return_fields=False):
"""
Evaluate multiple designs in batch (even faster!)
Args:
graph_data_list (list): List of graph data objects
return_fields (bool): Return complete fields
Returns:
list: List of evaluation results
"""
results = []
for graph_data in graph_data_list:
result = self.evaluate(graph_data, return_fields=return_fields)
results.append(result)
return results
def needs_fea_validation(self, uncertainty):
"""
Determine if FEA validation is recommended
Args:
uncertainty (float): Prediction uncertainty
Returns:
bool: True if FEA is recommended
"""
return uncertainty > self.uncertainty_threshold
def compare_with_fea(self, graph_data, fea_results):
"""
Compare neural predictions with FEA ground truth
Args:
graph_data: Mesh graph
fea_results (dict): FEA results with 'max_stress', 'max_displacement'
Returns:
dict: Comparison metrics
"""
# Neural prediction
pred = self.evaluate(graph_data)
# Compute errors
stress_error = abs(pred['max_stress'] - fea_results['max_stress'])
stress_rel_error = stress_error / (fea_results['max_stress'] + 1e-8)
disp_error = abs(pred['max_displacement'] - fea_results['max_displacement'])
disp_rel_error = disp_error / (fea_results['max_displacement'] + 1e-8)
comparison = {
'neural_prediction': pred,
'fea_results': fea_results,
'errors': {
'stress_error_abs': stress_error,
'stress_error_rel': stress_rel_error,
'displacement_error_abs': disp_error,
'displacement_error_rel': disp_rel_error
},
'within_tolerance': stress_rel_error < 0.1 and disp_rel_error < 0.1
}
return comparison
def get_statistics(self):
"""
Get optimizer usage statistics
Returns:
dict: Statistics about predictions
"""
avg_time = self.total_time / self.eval_count if self.eval_count > 0 else 0
return {
'total_evaluations': self.eval_count,
'total_time_ms': self.total_time,
'average_time_ms': avg_time,
'model_info': self.model_info
}
def reset_statistics(self):
"""Reset usage statistics"""
self.eval_count = 0
self.total_time = 0.0
class ParametricOptimizer:
"""
Optimizer for parametric designs
This wraps NeuralFieldOptimizer and adds parameter → mesh conversion.
Enables direct optimization over design parameters (thickness, radius, etc.)
"""
def __init__(
self,
model_path,
parameter_names,
parameter_bounds,
mesh_generator_fn
):
"""
Initialize parametric optimizer
Args:
model_path (str): Path to trained model
parameter_names (list): Names of design parameters
parameter_bounds (dict): Bounds for each parameter
mesh_generator_fn: Function that converts parameters → graph_data
"""
self.neural_optimizer = NeuralFieldOptimizer(model_path)
self.parameter_names = parameter_names
self.parameter_bounds = parameter_bounds
self.mesh_generator = mesh_generator_fn
print(f"\nParametric Optimizer initialized")
print(f"Design parameters: {parameter_names}")
def evaluate_parameters(self, parameters):
"""
Evaluate design from parameters
Args:
parameters (dict): Design parameters
Returns:
dict: Objectives (max_stress, max_displacement, etc.)
"""
# Generate mesh from parameters
graph_data = self.mesh_generator(parameters)
# Evaluate
results = self.neural_optimizer.evaluate(graph_data)
# Add parameters to results
results['parameters'] = parameters
return results
def optimize(
self,
initial_parameters,
objectives,
constraints,
method='gradient',
max_iterations=100
):
"""
Run optimization
Args:
initial_parameters (dict): Starting point
objectives (list): Objectives to minimize/maximize
constraints (list): Constraint functions
method (str): Optimization method ('gradient' or 'genetic')
max_iterations (int): Maximum iterations
Returns:
dict: Optimal parameters and results
"""
# This would integrate with scipy.optimize or genetic algorithms
# Placeholder for now
print(f"\nStarting optimization with {method} method...")
print(f"Initial parameters: {initial_parameters}")
print(f"Objectives: {objectives}")
print(f"Max iterations: {max_iterations}")
# TODO: Implement optimization loop
# For gradient-based:
# 1. Evaluate at current parameters
# 2. Compute sensitivities
# 3. Update parameters using gradients
# 4. Repeat until convergence
raise NotImplementedError("Full optimization loop coming in next update!")
def create_optimizer(model_path, config=None):
"""
Factory function to create optimizer
Args:
model_path (str): Path to trained model
config (dict): Optimizer configuration
Returns:
NeuralFieldOptimizer instance
"""
if config is None:
config = {}
return NeuralFieldOptimizer(model_path, **config)
if __name__ == "__main__":
# Example usage
print("AtomizerField Optimization Interface")
print("=" * 60)
print("\nThis module provides fast optimization with neural field predictions.")
print("\nExample usage:")
print("""
# Create optimizer
optimizer = NeuralFieldOptimizer('checkpoint_best.pt')
# Evaluate design
results = optimizer.evaluate(graph_data)
print(f"Max stress: {results['max_stress']:.2f} MPa")
print(f"Inference time: {results['inference_time_ms']:.1f} ms")
# Get sensitivities for gradient-based optimization
gradients = optimizer.get_sensitivities(graph_data, objective='max_stress')
# Batch evaluation (test 1000 designs in seconds!)
all_results = optimizer.batch_evaluate(design_variants)
""")
print("\nOptimization interface ready!")

373
atomizer-field/predict.py
View File

@@ -1,373 +0,0 @@
"""
predict.py
Inference script for AtomizerField trained models
AtomizerField Inference v2.0
Uses trained GNN to predict FEA fields 1000x faster than traditional simulation.
Usage:
python predict.py --model checkpoint_best.pt --input case_001
This enables:
- Rapid design exploration (milliseconds vs hours per analysis)
- Real-time optimization
- Interactive design feedback
"""
import argparse
import json
from pathlib import Path
import time
import torch
import numpy as np
import h5py
from neural_models.field_predictor import AtomizerFieldModel
from neural_models.data_loader import FEAMeshDataset
class FieldPredictor:
"""
Inference engine for trained field prediction models
"""
def __init__(self, checkpoint_path, device=None):
"""
Initialize predictor
Args:
checkpoint_path (str): Path to trained model checkpoint
device (str): Device to run on ('cuda' or 'cpu')
"""
if device is None:
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
else:
self.device = torch.device(device)
print(f"\nAtomizerField Inference Engine v2.0")
print(f"Device: {self.device}")
# Load checkpoint
print(f"Loading model from {checkpoint_path}...")
checkpoint = torch.load(checkpoint_path, map_location=self.device)
# Create model
model_config = checkpoint['config']['model']
self.model = AtomizerFieldModel(**model_config)
self.model.load_state_dict(checkpoint['model_state_dict'])
self.model = self.model.to(self.device)
self.model.eval()
self.config = checkpoint['config']
print(f"Model loaded (epoch {checkpoint['epoch']}, val_loss={checkpoint['best_val_loss']:.6f})")
def predict(self, case_directory):
"""
Predict displacement and stress fields for a case
Args:
case_directory (str): Path to parsed FEA case
Returns:
dict: Predictions with displacement, stress, von_mises fields
"""
print(f"\nPredicting fields for {Path(case_directory).name}...")
# Load data
dataset = FEAMeshDataset(
[case_directory],
normalize=True,
include_stress=False # Don't need ground truth for prediction
)
if len(dataset) == 0:
raise ValueError(f"Could not load case from {case_directory}")
data = dataset[0].to(self.device)
# Predict
start_time = time.time()
with torch.no_grad():
predictions = self.model(data, return_stress=True)
inference_time = time.time() - start_time
print(f"Prediction complete in {inference_time*1000:.1f} ms")
# Convert to numpy
results = {
'displacement': predictions['displacement'].cpu().numpy(),
'stress': predictions['stress'].cpu().numpy(),
'von_mises': predictions['von_mises'].cpu().numpy(),
'inference_time_ms': inference_time * 1000
}
# Compute max values
max_disp = np.max(np.linalg.norm(results['displacement'][:, :3], axis=1))
max_stress = np.max(results['von_mises'])
results['max_displacement'] = float(max_disp)
results['max_stress'] = float(max_stress)
print(f"\nResults:")
print(f" Max displacement: {max_disp:.6f} mm")
print(f" Max von Mises stress: {max_stress:.2f} MPa")
return results
def save_predictions(self, predictions, case_directory, output_name='predicted'):
"""
Save predictions in same format as ground truth
Args:
predictions (dict): Prediction results
case_directory (str): Case directory
output_name (str): Output file name prefix
"""
case_dir = Path(case_directory)
output_file = case_dir / f"{output_name}_fields.h5"
print(f"\nSaving predictions to {output_file}...")
with h5py.File(output_file, 'w') as f:
# Save displacement
f.create_dataset('displacement',
data=predictions['displacement'],
compression='gzip')
# Save stress
f.create_dataset('stress',
data=predictions['stress'],
compression='gzip')
# Save von Mises
f.create_dataset('von_mises',
data=predictions['von_mises'],
compression='gzip')
# Save metadata
f.attrs['max_displacement'] = predictions['max_displacement']
f.attrs['max_stress'] = predictions['max_stress']
f.attrs['inference_time_ms'] = predictions['inference_time_ms']
print(f"Predictions saved!")
# Also save JSON summary
summary_file = case_dir / f"{output_name}_summary.json"
summary = {
'max_displacement': predictions['max_displacement'],
'max_stress': predictions['max_stress'],
'inference_time_ms': predictions['inference_time_ms'],
'num_nodes': len(predictions['displacement'])
}
with open(summary_file, 'w') as f:
json.dump(summary, f, indent=2)
print(f"Summary saved to {summary_file}")
def compare_with_ground_truth(self, predictions, case_directory):
"""
Compare predictions with FEA ground truth
Args:
predictions (dict): Model predictions
case_directory (str): Case directory with ground truth
Returns:
dict: Comparison metrics
"""
case_dir = Path(case_directory)
h5_file = case_dir / "neural_field_data.h5"
if not h5_file.exists():
print("No ground truth available for comparison")
return None
print("\nComparing with FEA ground truth...")
# Load ground truth
with h5py.File(h5_file, 'r') as f:
gt_displacement = f['results/displacement'][:]
# Try to load stress
gt_stress = None
if 'results/stress' in f:
stress_group = f['results/stress']
for stress_type in stress_group.keys():
gt_stress = stress_group[stress_type]['data'][:]
break
# Compute errors
pred_disp = predictions['displacement']
disp_error = np.linalg.norm(pred_disp - gt_displacement, axis=1)
disp_magnitude = np.linalg.norm(gt_displacement, axis=1)
rel_disp_error = disp_error / (disp_magnitude + 1e-8)
metrics = {
'displacement': {
'mae': float(np.mean(disp_error)),
'rmse': float(np.sqrt(np.mean(disp_error**2))),
'relative_error': float(np.mean(rel_disp_error)),
'max_error': float(np.max(disp_error))
}
}
# Compare max values
pred_max_disp = predictions['max_displacement']
gt_max_disp = float(np.max(disp_magnitude))
metrics['max_displacement_error'] = abs(pred_max_disp - gt_max_disp)
metrics['max_displacement_relative_error'] = metrics['max_displacement_error'] / (gt_max_disp + 1e-8)
if gt_stress is not None:
pred_stress = predictions['stress']
stress_error = np.linalg.norm(pred_stress - gt_stress, axis=1)
metrics['stress'] = {
'mae': float(np.mean(stress_error)),
'rmse': float(np.sqrt(np.mean(stress_error**2))),
}
# Print comparison
print("\nComparison Results:")
print(f" Displacement MAE: {metrics['displacement']['mae']:.6f} mm")
print(f" Displacement RMSE: {metrics['displacement']['rmse']:.6f} mm")
print(f" Displacement Relative Error: {metrics['displacement']['relative_error']*100:.2f}%")
print(f" Max Displacement Error: {metrics['max_displacement_error']:.6f} mm ({metrics['max_displacement_relative_error']*100:.2f}%)")
if 'stress' in metrics:
print(f" Stress MAE: {metrics['stress']['mae']:.2f} MPa")
print(f" Stress RMSE: {metrics['stress']['rmse']:.2f} MPa")
return metrics
def batch_predict(predictor, case_directories, output_dir=None):
"""
Run predictions on multiple cases
Args:
predictor (FieldPredictor): Initialized predictor
case_directories (list): List of case directories
output_dir (str): Optional output directory for results
Returns:
list: List of prediction results
"""
print(f"\n{'='*60}")
print(f"Batch Prediction: {len(case_directories)} cases")
print(f"{'='*60}")
results = []
for i, case_dir in enumerate(case_directories, 1):
print(f"\n[{i}/{len(case_directories)}] Processing {Path(case_dir).name}...")
try:
# Predict
predictions = predictor.predict(case_dir)
# Save predictions
predictor.save_predictions(predictions, case_dir)
# Compare with ground truth
comparison = predictor.compare_with_ground_truth(predictions, case_dir)
result = {
'case': str(case_dir),
'status': 'success',
'predictions': {
'max_displacement': predictions['max_displacement'],
'max_stress': predictions['max_stress'],
'inference_time_ms': predictions['inference_time_ms']
},
'comparison': comparison
}
results.append(result)
except Exception as e:
print(f"ERROR: {e}")
results.append({
'case': str(case_dir),
'status': 'failed',
'error': str(e)
})
# Save batch results
if output_dir:
output_path = Path(output_dir)
output_path.mkdir(parents=True, exist_ok=True)
results_file = output_path / 'batch_predictions.json'
with open(results_file, 'w') as f:
json.dump(results, f, indent=2)
print(f"\nBatch results saved to {results_file}")
# Print summary
print(f"\n{'='*60}")
print("Batch Prediction Summary")
print(f"{'='*60}")
successful = sum(1 for r in results if r['status'] == 'success')
print(f"Successful: {successful}/{len(results)}")
if successful > 0:
avg_time = np.mean([r['predictions']['inference_time_ms']
for r in results if r['status'] == 'success'])
print(f"Average inference time: {avg_time:.1f} ms")
return results
def main():
"""
Main inference entry point
"""
parser = argparse.ArgumentParser(description='Predict FEA fields using trained model')
parser.add_argument('--model', type=str, required=True,
help='Path to model checkpoint')
parser.add_argument('--input', type=str, required=True,
help='Input case directory or directory containing multiple cases')
parser.add_argument('--output_dir', type=str, default=None,
help='Output directory for batch results')
parser.add_argument('--batch', action='store_true',
help='Process all subdirectories as separate cases')
parser.add_argument('--device', type=str, default=None,
choices=['cuda', 'cpu'],
help='Device to run on')
parser.add_argument('--compare', action='store_true',
help='Compare predictions with ground truth')
args = parser.parse_args()
# Create predictor
predictor = FieldPredictor(args.model, device=args.device)
input_path = Path(args.input)
if args.batch:
# Batch prediction
case_dirs = [d for d in input_path.iterdir() if d.is_dir()]
batch_predict(predictor, case_dirs, args.output_dir)
else:
# Single prediction
predictions = predictor.predict(args.input)
# Save predictions
predictor.save_predictions(predictions, args.input)
# Compare with ground truth if requested
if args.compare:
predictor.compare_with_ground_truth(predictions, args.input)
print("\nInference complete!")
if __name__ == "__main__":
main()

43
atomizer-field/requirements.txt
View File

@@ -1,43 +0,0 @@
# AtomizerField Requirements
# Python 3.8+ required
# ============================================================================
# Phase 1: Data Parser
# ============================================================================
# Core FEA parsing
pyNastran>=1.4.0
# Numerical computing
numpy>=1.20.0
# HDF5 file format for efficient field data storage
h5py>=3.0.0
# ============================================================================
# Phase 2: Neural Network Training
# ============================================================================
# Deep learning framework
torch>=2.0.0
# Graph neural networks
torch-geometric>=2.3.0
# TensorBoard for training visualization
tensorboard>=2.13.0
# ============================================================================
# Optional: Development and Testing
# ============================================================================
# Testing
# pytest>=7.0.0
# pytest-cov>=4.0.0
# Visualization
# matplotlib>=3.5.0
# plotly>=5.0.0
# Progress bars
# tqdm>=4.65.0

376
atomizer-field/test_simple_beam.py
View File

@@ -1,376 +0,0 @@
"""
test_simple_beam.py
Test AtomizerField with your actual Simple Beam model
This test validates the complete pipeline:
1. Parse BDF/OP2 files
2. Convert to graph format
3. Make predictions
4. Compare with ground truth
Usage:
python test_simple_beam.py
"""
import sys
import os
from pathlib import Path
import json
import time
print("\n" + "="*60)
print("AtomizerField Simple Beam Test")
print("="*60 + "\n")
# Test configuration
BEAM_DIR = Path("Models/Simple Beam")
TEST_CASE_DIR = Path("test_case_beam")
def test_1_check_files():
"""Test 1: Check if beam files exist"""
print("[TEST 1] Checking for beam files...")
bdf_file = BEAM_DIR / "beam_sim1-solution_1.dat"
op2_file = BEAM_DIR / "beam_sim1-solution_1.op2"
if not BEAM_DIR.exists():
print(f" [X] FAIL: Directory not found: {BEAM_DIR}")
return False
if not bdf_file.exists():
print(f" [X] FAIL: BDF file not found: {bdf_file}")
return False
if not op2_file.exists():
print(f" [X] FAIL: OP2 file not found: {op2_file}")
return False
# Check file sizes
bdf_size = bdf_file.stat().st_size / 1024 # KB
op2_size = op2_file.stat().st_size / 1024 # KB
print(f" [OK] Found BDF file: {bdf_file.name} ({bdf_size:.1f} KB)")
print(f" [OK] Found OP2 file: {op2_file.name} ({op2_size:.1f} KB)")
print(f" Status: PASS\n")
return True
def test_2_setup_test_case():
"""Test 2: Set up test case directory structure"""
print("[TEST 2] Setting up test case directory...")
try:
# Create directories
(TEST_CASE_DIR / "input").mkdir(parents=True, exist_ok=True)
(TEST_CASE_DIR / "output").mkdir(parents=True, exist_ok=True)
print(f" [OK] Created: {TEST_CASE_DIR / 'input'}")
print(f" [OK] Created: {TEST_CASE_DIR / 'output'}")
# Copy files (create symbolic links or copy)
import shutil
src_bdf = BEAM_DIR / "beam_sim1-solution_1.dat"
src_op2 = BEAM_DIR / "beam_sim1-solution_1.op2"
dst_bdf = TEST_CASE_DIR / "input" / "model.bdf"
dst_op2 = TEST_CASE_DIR / "output" / "model.op2"
# Copy files
if not dst_bdf.exists():
shutil.copy(src_bdf, dst_bdf)
print(f" [OK] Copied BDF to {dst_bdf}")
else:
print(f" [OK] BDF already exists: {dst_bdf}")
if not dst_op2.exists():
shutil.copy(src_op2, dst_op2)
print(f" [OK] Copied OP2 to {dst_op2}")
else:
print(f" [OK] OP2 already exists: {dst_op2}")
print(f" Status: PASS\n")
return True
except Exception as e:
print(f" [X] FAIL: {str(e)}\n")
return False
def test_3_import_modules():
"""Test 3: Import required modules"""
print("[TEST 3] Importing modules...")
try:
print(" Importing pyNastran...", end=" ")
from pyNastran.bdf.bdf import BDF
from pyNastran.op2.op2 import OP2
print("[OK]")
print(" Importing AtomizerField parser...", end=" ")
from neural_field_parser import NastranToNeuralFieldParser
print("[OK]")
print(f" Status: PASS\n")
return True
except ImportError as e:
print(f"\n [X] FAIL: Import error: {str(e)}\n")
return False
except Exception as e:
print(f"\n [X] FAIL: {str(e)}\n")
return False
def test_4_parse_beam():
"""Test 4: Parse beam BDF/OP2 files"""
print("[TEST 4] Parsing beam files...")
try:
from neural_field_parser import NastranToNeuralFieldParser
# Create parser
print(f" Initializing parser for {TEST_CASE_DIR}...")
parser = NastranToNeuralFieldParser(str(TEST_CASE_DIR))
# Parse
print(f" Parsing BDF and OP2 files...")
start_time = time.time()
data = parser.parse_all()
parse_time = time.time() - start_time
# Check results
print(f"\n Parse Results:")
print(f" Time: {parse_time:.2f} seconds")
print(f" Nodes: {data['mesh']['statistics']['n_nodes']:,}")
print(f" Elements: {data['mesh']['statistics']['n_elements']:,}")
print(f" Materials: {len(data['materials'])}")
if 'displacement' in data.get('results', {}):
max_disp = data['results']['displacement']['max_translation']
print(f" Max displacement: {max_disp:.6f} mm")
if 'stress' in data.get('results', {}):
for stress_type, stress_data in data['results']['stress'].items():
if 'max_von_mises' in stress_data:
max_vm = stress_data['max_von_mises']
if max_vm is not None:
print(f" Max von Mises stress: {max_vm:.2f} MPa")
break
# Check output files
json_file = TEST_CASE_DIR / "neural_field_data.json"
h5_file = TEST_CASE_DIR / "neural_field_data.h5"
if json_file.exists() and h5_file.exists():
json_size = json_file.stat().st_size / 1024
h5_size = h5_file.stat().st_size / 1024
print(f"\n Output Files:")
print(f" JSON: {json_size:.1f} KB")
print(f" HDF5: {h5_size:.1f} KB")
print(f" Status: PASS\n")
return True, data
except Exception as e:
print(f" [X] FAIL: {str(e)}\n")
import traceback
traceback.print_exc()
return False, None
def test_5_validate_data():
"""Test 5: Validate parsed data"""
print("[TEST 5] Validating parsed data...")
try:
from validate_parsed_data import NeuralFieldDataValidator
validator = NeuralFieldDataValidator(str(TEST_CASE_DIR))
print(f" Running validation checks...")
success = validator.validate()
if success:
print(f" Status: PASS\n")
else:
print(f" Status: PASS (with warnings)\n")
return True
except Exception as e:
print(f" [X] FAIL: {str(e)}\n")
return False
def test_6_load_as_graph():
"""Test 6: Load data as graph for neural network"""
print("[TEST 6] Converting to graph format...")
try:
import torch
from neural_models.data_loader import FEAMeshDataset
print(f" Creating dataset...")
dataset = FEAMeshDataset(
[str(TEST_CASE_DIR)],
normalize=False, # Don't normalize for single case
include_stress=True,
cache_in_memory=False
)
if len(dataset) == 0:
print(f" [X] FAIL: No data loaded")
return False
print(f" Loading graph...")
graph_data = dataset[0]
print(f"\n Graph Structure:")
print(f" Nodes: {graph_data.x.shape[0]:,}")
print(f" Node features: {graph_data.x.shape[1]}")
print(f" Edges: {graph_data.edge_index.shape[1]:,}")
print(f" Edge features: {graph_data.edge_attr.shape[1]}")
if hasattr(graph_data, 'y_displacement'):
print(f" Target displacement: {graph_data.y_displacement.shape}")
if hasattr(graph_data, 'y_stress'):
print(f" Target stress: {graph_data.y_stress.shape}")
print(f" Status: PASS\n")
return True, graph_data
except Exception as e:
print(f" [X] FAIL: {str(e)}\n")
import traceback
traceback.print_exc()
return False, None
def test_7_neural_prediction():
"""Test 7: Make neural network prediction (untrained model)"""
print("[TEST 7] Testing neural network prediction...")
try:
import torch
from neural_models.field_predictor import create_model
# Load graph from previous test
from neural_models.data_loader import FEAMeshDataset
dataset = FEAMeshDataset([str(TEST_CASE_DIR)], normalize=False, include_stress=False)
graph_data = dataset[0]
print(f" Creating untrained model...")
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.1
}
model = create_model(config)
model.eval()
print(f" Running inference...")
start_time = time.time()
with torch.no_grad():
predictions = model(graph_data, return_stress=True)
inference_time = (time.time() - start_time) * 1000 # ms
# Extract results
max_disp_pred = torch.max(torch.norm(predictions['displacement'][:, :3], dim=1)).item()
max_stress_pred = torch.max(predictions['von_mises']).item()
print(f"\n Predictions (untrained model):")
print(f" Inference time: {inference_time:.2f} ms")
print(f" Max displacement: {max_disp_pred:.6f} (arbitrary units)")
print(f" Max stress: {max_stress_pred:.2f} (arbitrary units)")
print(f"\n Note: Values are from untrained model (random weights)")
print(f" After training, these should match FEA results!")
print(f" Status: PASS\n")
return True
except Exception as e:
print(f" [X] FAIL: {str(e)}\n")
import traceback
traceback.print_exc()
return False
def main():
"""Run all tests"""
print("Testing AtomizerField with Simple Beam model\n")
print("This test validates:")
print(" 1. File existence")
print(" 2. Directory setup")
print(" 3. Module imports")
print(" 4. BDF/OP2 parsing")
print(" 5. Data validation")
print(" 6. Graph conversion")
print(" 7. Neural prediction\n")
results = []
# Run tests
tests = [
("Check Files", test_1_check_files),
("Setup Test Case", test_2_setup_test_case),
("Import Modules", test_3_import_modules),
("Parse Beam", test_4_parse_beam),
("Validate Data", test_5_validate_data),
("Load as Graph", test_6_load_as_graph),
("Neural Prediction", test_7_neural_prediction),
]
for test_name, test_func in tests:
result = test_func()
# Handle tests that return tuple (success, data)
if isinstance(result, tuple):
success = result[0]
else:
success = result
results.append(success)
# Stop on first failure for critical tests
if not success and test_name in ["Check Files", "Setup Test Case", "Import Modules"]:
print(f"\n[X] Critical test failed: {test_name}")
print("Cannot continue with remaining tests.\n")
break
# Summary
print("="*60)
print("TEST SUMMARY")
print("="*60 + "\n")
passed = sum(results)
total = len(results)
print(f"Tests Run: {total}")
print(f" [OK] Passed: {passed}")
print(f" [X] Failed: {total - passed}")
if passed == total:
print("\n[OK] ALL TESTS PASSED!")
print("\nYour Simple Beam model has been:")
print(" [OK] Successfully parsed")
print(" [OK] Converted to neural format")
print(" [OK] Validated for quality")
print(" [OK] Loaded as graph")
print(" [OK] Processed by neural network")
print("\nNext steps:")
print(" 1. Generate more training cases (50-500)")
print(" 2. Train the model: python train.py")
print(" 3. Make real predictions!")
else:
print(f"\n[X] {total - passed} test(s) failed")
print("Review errors above and fix issues.")
print("\n" + "="*60 + "\n")
return 0 if passed == total else 1
if __name__ == "__main__":
sys.exit(main())

402
atomizer-field/test_suite.py
View File

@@ -1,402 +0,0 @@
"""
test_suite.py
Master test orchestrator for AtomizerField
AtomizerField Testing Framework v1.0
Comprehensive validation from basic functionality to full neural FEA predictions.
Usage:
python test_suite.py --quick # 5-minute smoke tests
python test_suite.py --physics # Physics validation tests
python test_suite.py --learning # Learning capability tests
python test_suite.py --full # Complete test suite (1 hour)
Testing Strategy:
1. Smoke Tests (5 min) → Verify basic functionality
2. Physics Tests (15 min) → Validate physics constraints
3. Learning Tests (30 min) → Confirm learning capability
4. Integration Tests (1 hour) → Full system validation
"""
import sys
import os
import argparse
import time
from pathlib import Path
import json
from datetime import datetime
# Add project root to path
sys.path.insert(0, str(Path(__file__).parent))
# Test results storage
TEST_RESULTS = {
'timestamp': datetime.now().isoformat(),
'tests': [],
'summary': {
'total': 0,
'passed': 0,
'failed': 0,
'skipped': 0
}
}
class TestRunner:
"""
Test orchestrator that runs all tests in sequence
"""
def __init__(self, mode='quick'):
"""
Initialize test runner
Args:
mode (str): Testing mode ('quick', 'physics', 'learning', 'full')
"""
self.mode = mode
self.results_dir = Path('test_results')
self.results_dir.mkdir(exist_ok=True)
print(f"\n{'='*60}")
print(f"AtomizerField Test Suite v1.0")
print(f"Mode: {mode.upper()}")
print(f"{'='*60}\n")
def run_test(self, test_name, test_func, description):
"""
Run a single test and record results
Args:
test_name (str): Name of test
test_func (callable): Test function to run
description (str): Test description
Returns:
bool: True if passed
"""
print(f"[TEST] {test_name}")
print(f" Description: {description}")
start_time = time.time()
result = {
'name': test_name,
'description': description,
'status': 'unknown',
'duration': 0,
'message': '',
'metrics': {}
}
try:
test_result = test_func()
if test_result is None or test_result is True:
result['status'] = 'PASS'
result['message'] = 'Test passed successfully'
print(f" Status: [PASS]")
TEST_RESULTS['summary']['passed'] += 1
elif isinstance(test_result, dict):
result['status'] = test_result.get('status', 'PASS')
result['message'] = test_result.get('message', '')
result['metrics'] = test_result.get('metrics', {})
if result['status'] == 'PASS':
print(f" Status: [PASS]")
TEST_RESULTS['summary']['passed'] += 1
else:
print(f" Status: [FAIL]")
print(f" Reason: {result['message']}")
TEST_RESULTS['summary']['failed'] += 1
else:
result['status'] = 'FAIL'
result['message'] = str(test_result)
print(f" Status: [FAIL]")
TEST_RESULTS['summary']['failed'] += 1
except Exception as e:
result['status'] = 'FAIL'
result['message'] = f"Exception: {str(e)}"
print(f" Status: [FAIL]")
print(f" Error: {str(e)}")
TEST_RESULTS['summary']['failed'] += 1
result['duration'] = time.time() - start_time
print(f" Duration: {result['duration']:.2f}s\n")
TEST_RESULTS['tests'].append(result)
TEST_RESULTS['summary']['total'] += 1
return result['status'] == 'PASS'
def run_smoke_tests(self):
"""
Quick smoke tests (5 minutes)
Verify basic functionality
"""
print(f"\n{'='*60}")
print("PHASE 1: SMOKE TESTS (5 minutes)")
print(f"{'='*60}\n")
from tests import test_synthetic
# Test 1: Model creation
self.run_test(
"Model Creation",
test_synthetic.test_model_creation,
"Verify GNN model can be instantiated"
)
# Test 2: Forward pass
self.run_test(
"Forward Pass",
test_synthetic.test_forward_pass,
"Verify model can process dummy data"
)
# Test 3: Loss computation
self.run_test(
"Loss Computation",
test_synthetic.test_loss_computation,
"Verify loss functions work"
)
def run_physics_tests(self):
"""
Physics validation tests (15 minutes)
Ensure physics constraints work
"""
print(f"\n{'='*60}")
print("PHASE 2: PHYSICS VALIDATION (15 minutes)")
print(f"{'='*60}\n")
from tests import test_physics
# Test 1: Cantilever beam
self.run_test(
"Cantilever Beam (Analytical)",
test_physics.test_cantilever_analytical,
"Compare with δ = FL³/3EI solution"
)
# Test 2: Equilibrium
self.run_test(
"Equilibrium Check",
test_physics.test_equilibrium,
"Verify force balance (∇·σ + f = 0)"
)
# Test 3: Energy conservation
self.run_test(
"Energy Conservation",
test_physics.test_energy_conservation,
"Verify strain energy = work done"
)
def run_learning_tests(self):
"""
Learning capability tests (30 minutes)
Confirm network can learn
"""
print(f"\n{'='*60}")
print("PHASE 3: LEARNING CAPABILITY (30 minutes)")
print(f"{'='*60}\n")
from tests import test_learning
# Test 1: Memorization
self.run_test(
"Memorization Test",
test_learning.test_memorization,
"Can network memorize small dataset?"
)
# Test 2: Interpolation
self.run_test(
"Interpolation Test",
test_learning.test_interpolation,
"Can network interpolate between training points?"
)
# Test 3: Pattern recognition
self.run_test(
"Pattern Recognition",
test_learning.test_pattern_recognition,
"Does network learn thickness → stress relationship?"
)
def run_integration_tests(self):
"""
Full integration tests (1 hour)
Complete system validation
"""
print(f"\n{'='*60}")
print("PHASE 4: INTEGRATION TESTS (1 hour)")
print(f"{'='*60}\n")
from tests import test_predictions
# Test 1: Parser validation
self.run_test(
"Parser Validation",
test_predictions.test_parser,
"Verify data parsing works correctly"
)
# Test 2: Training pipeline
self.run_test(
"Training Pipeline",
test_predictions.test_training,
"Verify complete training workflow"
)
# Test 3: Prediction accuracy
self.run_test(
"Prediction Accuracy",
test_predictions.test_prediction_accuracy,
"Compare neural vs FEA predictions"
)
def print_summary(self):
"""Print test summary"""
summary = TEST_RESULTS['summary']
print(f"\n{'='*60}")
print("TEST SUMMARY")
print(f"{'='*60}\n")
total = summary['total']
passed = summary['passed']
failed = summary['failed']
pass_rate = (passed / total * 100) if total > 0 else 0
print(f"Total Tests: {total}")
print(f" + Passed: {passed}")
print(f" - Failed: {failed}")
print(f" Pass Rate: {pass_rate:.1f}%\n")
if failed == 0:
print("[SUCCESS] ALL TESTS PASSED - SYSTEM READY!")
else:
print(f"[ERROR] {failed} TEST(S) FAILED - REVIEW REQUIRED")
print(f"\n{'='*60}\n")
def save_results(self):
"""Save test results to JSON"""
results_file = self.results_dir / f'test_results_{self.mode}_{int(time.time())}.json'
with open(results_file, 'w') as f:
json.dump(TEST_RESULTS, f, indent=2)
print(f"Results saved to: {results_file}")
def run(self):
"""
Run test suite based on mode
"""
start_time = time.time()
if self.mode == 'quick':
self.run_smoke_tests()
elif self.mode == 'physics':
self.run_smoke_tests()
self.run_physics_tests()
elif self.mode == 'learning':
self.run_smoke_tests()
self.run_physics_tests()
self.run_learning_tests()
elif self.mode == 'full':
self.run_smoke_tests()
self.run_physics_tests()
self.run_learning_tests()
self.run_integration_tests()
total_time = time.time() - start_time
# Print summary
self.print_summary()
print(f"Total testing time: {total_time/60:.1f} minutes\n")
# Save results
self.save_results()
# Return exit code
return 0 if TEST_RESULTS['summary']['failed'] == 0 else 1
def main():
"""Main entry point"""
parser = argparse.ArgumentParser(
description='AtomizerField Test Suite',
formatter_class=argparse.RawDescriptionHelpFormatter,
epilog="""
Examples:
# Quick smoke tests (5 min)
python test_suite.py --quick
# Physics validation (15 min)
python test_suite.py --physics
# Learning tests (30 min)
python test_suite.py --learning
# Full test suite (1 hour)
python test_suite.py --full
"""
)
parser.add_argument(
'--quick',
action='store_true',
help='Run quick smoke tests (5 minutes)'
)
parser.add_argument(
'--physics',
action='store_true',
help='Run physics validation tests (15 minutes)'
)
parser.add_argument(
'--learning',
action='store_true',
help='Run learning capability tests (30 minutes)'
)
parser.add_argument(
'--full',
action='store_true',
help='Run complete test suite (1 hour)'
)
args = parser.parse_args()
# Determine mode
if args.full:
mode = 'full'
elif args.learning:
mode = 'learning'
elif args.physics:
mode = 'physics'
elif args.quick:
mode = 'quick'
else:
# Default to quick if no mode specified
mode = 'quick'
print("No mode specified, defaulting to --quick")
# Run tests
runner = TestRunner(mode=mode)
exit_code = runner.run()
sys.exit(exit_code)
if __name__ == "__main__":
main()

6
atomizer-field/tests/__init__.py
View File

@@ -1,6 +0,0 @@
"""
AtomizerField Test Suite
Comprehensive testing framework for neural field learning
"""
__version__ = "1.0.0"

446
atomizer-field/tests/analytical_cases.py
View File

@@ -1,446 +0,0 @@
"""
analytical_cases.py
Analytical solutions for classical mechanics problems
Provides known solutions for validation:
- Cantilever beam under point load
- Simply supported beam
- Axial tension bar
- Pressure vessel (thin-walled cylinder)
- Torsion of circular shaft
These serve as ground truth for testing neural predictions.
"""
import numpy as np
from dataclasses import dataclass
from typing import Tuple
@dataclass
class BeamProperties:
"""Material and geometric properties for beam"""
length: float # m
width: float # m
height: float # m
E: float # Young's modulus (Pa)
nu: float # Poisson's ratio
rho: float # Density (kg/m³)
@property
def I(self) -> float:
"""Second moment of area for rectangular section"""
return (self.width * self.height**3) / 12
@property
def A(self) -> float:
"""Cross-sectional area"""
return self.width * self.height
@property
def G(self) -> float:
"""Shear modulus"""
return self.E / (2 * (1 + self.nu))
def cantilever_beam_point_load(force: float, props: BeamProperties) -> dict:
"""
Cantilever beam with point load at free end
Analytical solution:
δ_max = FL³/3EI (at free end)
σ_max = FL/Z (at fixed end)
Args:
force: Applied force at tip (N)
props: Beam properties
Returns:
dict with displacement, stress, reactions
"""
L = props.length
E = props.E
I = props.I
c = props.height / 2 # Distance to neutral axis
# Maximum displacement at free end
delta_max = (force * L**3) / (3 * E * I)
# Maximum stress at fixed end (bending)
M_max = force * L # Maximum moment at fixed end
Z = I / c # Section modulus
sigma_max = M_max / Z
# Deflection curve: y(x) = (F/(6EI)) * x² * (3L - x)
def deflection_at(x):
"""Deflection at position x along beam"""
if x < 0 or x > L:
return 0.0
return (force / (6 * E * I)) * x**2 * (3 * L - x)
# Bending moment: M(x) = F * (L - x)
def moment_at(x):
"""Bending moment at position x"""
if x < 0 or x > L:
return 0.0
return force * (L - x)
# Stress: σ(x, y) = M(x) * y / I
def stress_at(x, y):
"""Bending stress at position x and distance y from neutral axis"""
M = moment_at(x)
return (M * y) / I
return {
'type': 'cantilever_point_load',
'delta_max': delta_max,
'sigma_max': sigma_max,
'deflection_function': deflection_at,
'moment_function': moment_at,
'stress_function': stress_at,
'load': force,
'properties': props
}
def simply_supported_beam_point_load(force: float, props: BeamProperties) -> dict:
"""
Simply supported beam with point load at center
Analytical solution:
δ_max = FL³/48EI (at center)
σ_max = FL/4Z (at center)
Args:
force: Applied force at center (N)
props: Beam properties
Returns:
dict with displacement, stress, reactions
"""
L = props.length
E = props.E
I = props.I
c = props.height / 2
# Maximum displacement at center
delta_max = (force * L**3) / (48 * E * I)
# Maximum stress at center
M_max = force * L / 4 # Maximum moment at center
Z = I / c
sigma_max = M_max / Z
# Deflection curve (for 0 < x < L/2):
# y(x) = (F/(48EI)) * x * (3L² - 4x²)
def deflection_at(x):
"""Deflection at position x along beam"""
if x < 0 or x > L:
return 0.0
if x <= L/2:
return (force / (48 * E * I)) * x * (3 * L**2 - 4 * x**2)
else:
# Symmetric about center
return deflection_at(L - x)
# Bending moment: M(x) = (F/2) * x for x < L/2
def moment_at(x):
"""Bending moment at position x"""
if x < 0 or x > L:
return 0.0
if x <= L/2:
return (force / 2) * x
else:
return (force / 2) * (L - x)
def stress_at(x, y):
"""Bending stress at position x and distance y from neutral axis"""
M = moment_at(x)
return (M * y) / I
return {
'type': 'simply_supported_point_load',
'delta_max': delta_max,
'sigma_max': sigma_max,
'deflection_function': deflection_at,
'moment_function': moment_at,
'stress_function': stress_at,
'load': force,
'properties': props,
'reactions': force / 2 # Each support
}
def axial_tension_bar(force: float, props: BeamProperties) -> dict:
"""
Bar under axial tension
Analytical solution:
δ = FL/EA (total elongation)
σ = F/A (uniform stress)
ε = σ/E (uniform strain)
Args:
force: Axial force (N)
props: Bar properties
Returns:
dict with displacement, stress, strain
"""
L = props.length
E = props.E
A = props.A
# Total elongation
delta = (force * L) / (E * A)
# Uniform stress
sigma = force / A
# Uniform strain
epsilon = sigma / E
# Displacement is linear along length
def displacement_at(x):
"""Axial displacement at position x"""
if x < 0 or x > L:
return 0.0
return (force * x) / (E * A)
return {
'type': 'axial_tension',
'delta_total': delta,
'sigma': sigma,
'epsilon': epsilon,
'displacement_function': displacement_at,
'load': force,
'properties': props
}
def thin_wall_pressure_vessel(pressure: float, radius: float, thickness: float,
length: float, E: float, nu: float) -> dict:
"""
Thin-walled cylindrical pressure vessel
Analytical solution:
σ_hoop = pr/t (circumferential stress)
σ_axial = pr/2t (longitudinal stress)
ε_hoop = (1/E)(σ_h - ν*σ_a)
ε_axial = (1/E)(σ_a - ν*σ_h)
Args:
pressure: Internal pressure (Pa)
radius: Mean radius (m)
thickness: Wall thickness (m)
length: Cylinder length (m)
E: Young's modulus (Pa)
nu: Poisson's ratio
Returns:
dict with stresses and strains
"""
# Stresses
sigma_hoop = (pressure * radius) / thickness
sigma_axial = (pressure * radius) / (2 * thickness)
# Strains
epsilon_hoop = (1/E) * (sigma_hoop - nu * sigma_axial)
epsilon_axial = (1/E) * (sigma_axial - nu * sigma_hoop)
# Radial expansion
delta_r = epsilon_hoop * radius
return {
'type': 'pressure_vessel',
'sigma_hoop': sigma_hoop,
'sigma_axial': sigma_axial,
'epsilon_hoop': epsilon_hoop,
'epsilon_axial': epsilon_axial,
'radial_expansion': delta_r,
'pressure': pressure,
'radius': radius,
'thickness': thickness
}
def torsion_circular_shaft(torque: float, radius: float, length: float, G: float) -> dict:
"""
Circular shaft under torsion
Analytical solution:
θ = TL/GJ (angle of twist)
τ_max = Tr/J (maximum shear stress at surface)
γ_max = τ_max/G (maximum shear strain)
Args:
torque: Applied torque (N·m)
radius: Shaft radius (m)
length: Shaft length (m)
G: Shear modulus (Pa)
Returns:
dict with twist angle, stress, strain
"""
# Polar moment of inertia
J = (np.pi * radius**4) / 2
# Angle of twist
theta = (torque * length) / (G * J)
# Maximum shear stress (at surface)
tau_max = (torque * radius) / J
# Maximum shear strain
gamma_max = tau_max / G
# Shear stress at radius r
def shear_stress_at(r):
"""Shear stress at radial distance r from center"""
if r < 0 or r > radius:
return 0.0
return (torque * r) / J
return {
'type': 'torsion',
'theta': theta,
'tau_max': tau_max,
'gamma_max': gamma_max,
'shear_stress_function': shear_stress_at,
'torque': torque,
'radius': radius,
'length': length
}
# Standard test cases with typical values
def get_standard_cantilever() -> Tuple[float, BeamProperties]:
"""Standard cantilever test case"""
props = BeamProperties(
length=1.0, # 1 m
width=0.05, # 50 mm
height=0.1, # 100 mm
E=210e9, # Steel: 210 GPa
nu=0.3,
rho=7850 # kg/m³
)
force = 1000.0 # 1 kN
return force, props
def get_standard_simply_supported() -> Tuple[float, BeamProperties]:
"""Standard simply supported beam test case"""
props = BeamProperties(
length=2.0, # 2 m
width=0.05, # 50 mm
height=0.1, # 100 mm
E=210e9, # Steel
nu=0.3,
rho=7850
)
force = 5000.0 # 5 kN
return force, props
def get_standard_tension_bar() -> Tuple[float, BeamProperties]:
"""Standard tension bar test case"""
props = BeamProperties(
length=1.0, # 1 m
width=0.02, # 20 mm
height=0.02, # 20 mm (square bar)
E=210e9, # Steel
nu=0.3,
rho=7850
)
force = 10000.0 # 10 kN
return force, props
# Example usage and validation
if __name__ == "__main__":
print("Analytical Test Cases\n")
print("="*60)
# Test 1: Cantilever beam
print("\n1. Cantilever Beam (Point Load at Tip)")
print("-"*60)
force, props = get_standard_cantilever()
result = cantilever_beam_point_load(force, props)
print(f"Load: {force} N")
print(f"Length: {props.length} m")
print(f"E: {props.E/1e9:.0f} GPa")
print(f"I: {props.I*1e12:.3f} mm⁴")
print(f"\nResults:")
print(f" Max displacement: {result['delta_max']*1000:.3f} mm")
print(f" Max stress: {result['sigma_max']/1e6:.1f} MPa")
# Verify deflection at intermediate points
print(f"\nDeflection profile:")
for x in [0.0, 0.25, 0.5, 0.75, 1.0]:
x_m = x * props.length
delta = result['deflection_function'](x_m)
print(f" x = {x:.2f}L: δ = {delta*1000:.3f} mm")
# Test 2: Simply supported beam
print("\n2. Simply Supported Beam (Point Load at Center)")
print("-"*60)
force, props = get_standard_simply_supported()
result = simply_supported_beam_point_load(force, props)
print(f"Load: {force} N")
print(f"Length: {props.length} m")
print(f"\nResults:")
print(f" Max displacement: {result['delta_max']*1000:.3f} mm")
print(f" Max stress: {result['sigma_max']/1e6:.1f} MPa")
print(f" Reactions: {result['reactions']} N each")
# Test 3: Axial tension
print("\n3. Axial Tension Bar")
print("-"*60)
force, props = get_standard_tension_bar()
result = axial_tension_bar(force, props)
print(f"Load: {force} N")
print(f"Length: {props.length} m")
print(f"Area: {props.A*1e6:.0f} mm²")
print(f"\nResults:")
print(f" Total elongation: {result['delta_total']*1e6:.3f} μm")
print(f" Stress: {result['sigma']/1e6:.1f} MPa")
print(f" Strain: {result['epsilon']*1e6:.1f} με")
# Test 4: Pressure vessel
print("\n4. Thin-Walled Pressure Vessel")
print("-"*60)
pressure = 10e6 # 10 MPa
radius = 0.5 # 500 mm
thickness = 0.01 # 10 mm
result = thin_wall_pressure_vessel(pressure, radius, thickness, 2.0, 210e9, 0.3)
print(f"Pressure: {pressure/1e6:.1f} MPa")
print(f"Radius: {radius*1000:.0f} mm")
print(f"Thickness: {thickness*1000:.0f} mm")
print(f"\nResults:")
print(f" Hoop stress: {result['sigma_hoop']/1e6:.1f} MPa")
print(f" Axial stress: {result['sigma_axial']/1e6:.1f} MPa")
print(f" Radial expansion: {result['radial_expansion']*1e6:.3f} μm")
# Test 5: Torsion
print("\n5. Circular Shaft in Torsion")
print("-"*60)
torque = 1000 # 1000 N·m
radius = 0.05 # 50 mm
length = 1.0 # 1 m
G = 80e9 # 80 GPa
result = torsion_circular_shaft(torque, radius, length, G)
print(f"Torque: {torque} N·m")
print(f"Radius: {radius*1000:.0f} mm")
print(f"Length: {length:.1f} m")
print(f"\nResults:")
print(f" Twist angle: {result['theta']*180/np.pi:.3f}°")
print(f" Max shear stress: {result['tau_max']/1e6:.1f} MPa")
print(f" Max shear strain: {result['gamma_max']*1e6:.1f} με")
print("\n" + "="*60)
print("All analytical solutions validated!")

468
atomizer-field/tests/test_learning.py
View File

@@ -1,468 +0,0 @@
"""
test_learning.py
Learning capability tests
Tests that the neural network can actually learn:
- Memorization: Can it memorize 10 examples?
- Interpolation: Can it generalize between training points?
- Extrapolation: Can it predict beyond training range?
- Pattern recognition: Does it learn physical relationships?
"""
import torch
import torch.nn as nn
import numpy as np
import sys
from pathlib import Path
# Add parent directory to path
sys.path.insert(0, str(Path(__file__).parent.parent))
from neural_models.field_predictor import create_model
from neural_models.physics_losses import create_loss_function
from torch_geometric.data import Data
def create_synthetic_dataset(n_samples=10, variation='load'):
"""
Create synthetic FEA-like dataset with known patterns
Args:
n_samples: Number of samples
variation: Parameter to vary ('load', 'stiffness', 'geometry')
Returns:
List of (graph_data, target_displacement, target_stress) tuples
"""
dataset = []
for i in range(n_samples):
num_nodes = 20
num_edges = 40
# Base features
x = torch.randn(num_nodes, 12) * 0.1
# Vary parameter based on type
if variation == 'load':
load_factor = 1.0 + i * 0.5 # Vary load from 1.0 to 5.5
x[:, 9:12] = torch.randn(num_nodes, 3) * load_factor
elif variation == 'stiffness':
stiffness_factor = 1.0 + i * 0.2 # Vary stiffness
edge_attr = torch.randn(num_edges, 5) * 0.1
edge_attr[:, 0] = stiffness_factor # Young's modulus
elif variation == 'geometry':
geometry_factor = 1.0 + i * 0.1 # Vary geometry
x[:, 0:3] = torch.randn(num_nodes, 3) * geometry_factor
# Create edges
edge_index = torch.randint(0, num_nodes, (2, num_edges))
# Default edge attributes if not varying stiffness
if variation != 'stiffness':
edge_attr = torch.randn(num_edges, 5) * 0.1
edge_attr[:, 0] = 1.0 # Constant Young's modulus
batch = torch.zeros(num_nodes, dtype=torch.long)
data = Data(x=x, edge_index=edge_index, edge_attr=edge_attr, batch=batch)
# Create synthetic targets with known relationship
# Displacement proportional to load / stiffness
if variation == 'load':
target_displacement = torch.randn(num_nodes, 6) * load_factor
elif variation == 'stiffness':
target_displacement = torch.randn(num_nodes, 6) / stiffness_factor
else:
target_displacement = torch.randn(num_nodes, 6)
# Stress also follows known pattern
target_stress = target_displacement * 2.0 # Simple linear relationship
dataset.append((data, target_displacement, target_stress))
return dataset
def test_memorization():
"""
Test 1: Can network memorize small dataset?
Expected: After training on 10 examples, can achieve < 1% error
This tests basic learning capability - if it can't memorize,
something is fundamentally wrong.
"""
print(" Creating small dataset (10 samples)...")
# Create tiny dataset
dataset = create_synthetic_dataset(n_samples=10, variation='load')
# Create model
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.0 # No dropout for memorization
}
model = create_model(config)
loss_fn = create_loss_function('mse')
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
print(" Training for 100 epochs...")
model.train()
losses = []
for epoch in range(100):
epoch_loss = 0.0
for graph_data, target_disp, target_stress in dataset:
optimizer.zero_grad()
# Forward pass
predictions = model(graph_data, return_stress=True)
# Compute loss
targets = {
'displacement': target_disp,
'stress': target_stress
}
loss_dict = loss_fn(predictions, targets)
loss = loss_dict['total_loss']
# Backward pass
loss.backward()
optimizer.step()
epoch_loss += loss.item()
avg_loss = epoch_loss / len(dataset)
losses.append(avg_loss)
if (epoch + 1) % 20 == 0:
print(f" Epoch {epoch+1}/100: Loss = {avg_loss:.6f}")
final_loss = losses[-1]
initial_loss = losses[0]
improvement = (initial_loss - final_loss) / initial_loss * 100
print(f" Initial loss: {initial_loss:.6f}")
print(f" Final loss: {final_loss:.6f}")
print(f" Improvement: {improvement:.1f}%")
# Success if loss decreased significantly
success = improvement > 50.0
return {
'status': 'PASS' if success else 'FAIL',
'message': f'Memorization {"successful" if success else "failed"} ({improvement:.1f}% improvement)',
'metrics': {
'initial_loss': float(initial_loss),
'final_loss': float(final_loss),
'improvement_percent': float(improvement),
'converged': final_loss < 0.1
}
}
def test_interpolation():
"""
Test 2: Can network interpolate?
Expected: After training on [1, 3, 5], predict [2, 4] with < 5% error
This tests generalization capability within training range.
"""
print(" Creating interpolation dataset...")
# Train on samples 0, 2, 4, 6, 8 (odd indices)
train_indices = [0, 2, 4, 6, 8]
test_indices = [1, 3, 5, 7] # Even indices (interpolation)
full_dataset = create_synthetic_dataset(n_samples=10, variation='load')
train_dataset = [full_dataset[i] for i in train_indices]
test_dataset = [full_dataset[i] for i in test_indices]
# Create model
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.1
}
model = create_model(config)
loss_fn = create_loss_function('mse')
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
print(f" Training on {len(train_dataset)} samples...")
# Train
model.train()
for epoch in range(50):
for graph_data, target_disp, target_stress in train_dataset:
optimizer.zero_grad()
predictions = model(graph_data, return_stress=True)
targets = {
'displacement': target_disp,
'stress': target_stress
}
loss_dict = loss_fn(predictions, targets)
loss = loss_dict['total_loss']
loss.backward()
optimizer.step()
# Test interpolation
print(f" Testing interpolation on {len(test_dataset)} samples...")
model.eval()
test_errors = []
with torch.no_grad():
for graph_data, target_disp, target_stress in test_dataset:
predictions = model(graph_data, return_stress=True)
# Compute relative error
pred_disp = predictions['displacement']
error = torch.mean(torch.abs(pred_disp - target_disp) / (torch.abs(target_disp) + 1e-8))
test_errors.append(error.item())
avg_error = np.mean(test_errors) * 100
print(f" Average interpolation error: {avg_error:.2f}%")
# Success if error reasonable for untrained interpolation
success = avg_error < 100.0 # Lenient for this basic test
return {
'status': 'PASS' if success else 'FAIL',
'message': f'Interpolation test completed ({avg_error:.2f}% error)',
'metrics': {
'average_error_percent': float(avg_error),
'test_samples': len(test_dataset),
'train_samples': len(train_dataset)
}
}
def test_extrapolation():
"""
Test 3: Can network extrapolate?
Expected: After training on [1-5], predict [7-10] with < 20% error
This tests generalization beyond training range (harder than interpolation).
"""
print(" Creating extrapolation dataset...")
# Train on first 5 samples
train_indices = list(range(5))
test_indices = list(range(7, 10)) # Extrapolate to higher values
full_dataset = create_synthetic_dataset(n_samples=10, variation='load')
train_dataset = [full_dataset[i] for i in train_indices]
test_dataset = [full_dataset[i] for i in test_indices]
# Create model
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.1
}
model = create_model(config)
loss_fn = create_loss_function('mse')
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
print(f" Training on samples 1-5...")
# Train
model.train()
for epoch in range(50):
for graph_data, target_disp, target_stress in train_dataset:
optimizer.zero_grad()
predictions = model(graph_data, return_stress=True)
targets = {
'displacement': target_disp,
'stress': target_stress
}
loss_dict = loss_fn(predictions, targets)
loss = loss_dict['total_loss']
loss.backward()
optimizer.step()
# Test extrapolation
print(f" Testing extrapolation on samples 7-10...")
model.eval()
test_errors = []
with torch.no_grad():
for graph_data, target_disp, target_stress in test_dataset:
predictions = model(graph_data, return_stress=True)
pred_disp = predictions['displacement']
error = torch.mean(torch.abs(pred_disp - target_disp) / (torch.abs(target_disp) + 1e-8))
test_errors.append(error.item())
avg_error = np.mean(test_errors) * 100
print(f" Average extrapolation error: {avg_error:.2f}%")
print(f" Note: Extrapolation is harder than interpolation.")
# Success if error is reasonable (extrapolation is hard)
success = avg_error < 200.0 # Very lenient for basic test
return {
'status': 'PASS' if success else 'FAIL',
'message': f'Extrapolation test completed ({avg_error:.2f}% error)',
'metrics': {
'average_error_percent': float(avg_error),
'test_samples': len(test_dataset),
'train_samples': len(train_dataset)
}
}
def test_pattern_recognition():
"""
Test 4: Can network learn physical patterns?
Expected: Learn that thickness ↑ → stress ↓
This tests if network understands relationships, not just memorization.
"""
print(" Testing pattern recognition...")
# Create dataset with clear pattern: stiffness ↑ → displacement ↓
dataset = create_synthetic_dataset(n_samples=20, variation='stiffness')
# Create model
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.1
}
model = create_model(config)
loss_fn = create_loss_function('mse')
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
print(" Training on stiffness variation dataset...")
# Train
model.train()
for epoch in range(50):
for graph_data, target_disp, target_stress in dataset:
optimizer.zero_grad()
predictions = model(graph_data, return_stress=True)
targets = {
'displacement': target_disp,
'stress': target_stress
}
loss_dict = loss_fn(predictions, targets)
loss = loss_dict['total_loss']
loss.backward()
optimizer.step()
# Test pattern: predict two cases with different stiffness
print(" Testing learned pattern...")
model.eval()
# Low stiffness case
low_stiff_data, low_stiff_disp, _ = dataset[0]
# High stiffness case
high_stiff_data, high_stiff_disp, _ = dataset[-1]
with torch.no_grad():
low_pred = model(low_stiff_data, return_stress=False)
high_pred = model(high_stiff_data, return_stress=False)
# Check if pattern learned: low stiffness → high displacement
low_disp_mag = torch.mean(torch.abs(low_pred['displacement'])).item()
high_disp_mag = torch.mean(torch.abs(high_pred['displacement'])).item()
print(f" Low stiffness displacement: {low_disp_mag:.6f}")
print(f" High stiffness displacement: {high_disp_mag:.6f}")
# Pattern learned if low stiffness has higher displacement
# (But with random data this might not hold - this is a template)
pattern_ratio = low_disp_mag / (high_disp_mag + 1e-8)
print(f" Pattern ratio (should be > 1.0): {pattern_ratio:.2f}")
print(f" Note: With synthetic random data, pattern may not emerge.")
print(f" Real training data should show clear physical patterns.")
# Just check predictions are reasonable magnitude
success = (low_disp_mag > 0.0 and high_disp_mag > 0.0)
return {
'status': 'PASS' if success else 'FAIL',
'message': f'Pattern recognition test completed',
'metrics': {
'low_stiffness_displacement': float(low_disp_mag),
'high_stiffness_displacement': float(high_disp_mag),
'pattern_ratio': float(pattern_ratio)
}
}
if __name__ == "__main__":
print("\nRunning learning capability tests...\n")
tests = [
("Memorization Test", test_memorization),
("Interpolation Test", test_interpolation),
("Extrapolation Test", test_extrapolation),
("Pattern Recognition", test_pattern_recognition)
]
passed = 0
failed = 0
for name, test_func in tests:
print(f"[TEST] {name}")
try:
result = test_func()
if result['status'] == 'PASS':
print(f" ✓ PASS\n")
passed += 1
else:
print(f" ✗ FAIL: {result['message']}\n")
failed += 1
except Exception as e:
print(f" ✗ FAIL: {str(e)}\n")
import traceback
traceback.print_exc()
failed += 1
print(f"\nResults: {passed} passed, {failed} failed")
print(f"\nNote: These tests use SYNTHETIC data and train for limited epochs.")
print(f"Real training on actual FEA data will show better learning performance.")

385
atomizer-field/tests/test_physics.py
View File

@@ -1,385 +0,0 @@
"""
test_physics.py
Physics validation tests with analytical solutions
Tests that the neural network respects fundamental physics:
- Cantilever beam (δ = FL³/3EI)
- Simply supported beam (δ = FL³/48EI)
- Equilibrium (∇·σ + f = 0)
- Energy conservation (strain energy = work done)
"""
import torch
import numpy as np
import sys
from pathlib import Path
# Add parent directory to path
sys.path.insert(0, str(Path(__file__).parent.parent))
from neural_models.field_predictor import create_model
from torch_geometric.data import Data
def create_cantilever_beam_graph(length=1.0, force=1000.0, E=210e9, I=1e-6):
"""
Create synthetic cantilever beam graph with analytical solution
Analytical solution: δ_max = FL³/3EI
Args:
length: Beam length (m)
force: Applied force (N)
E: Young's modulus (Pa)
I: Second moment of area (m^4)
Returns:
graph_data: PyG Data object
analytical_displacement: Expected max displacement (m)
"""
# Calculate analytical solution
analytical_displacement = (force * length**3) / (3 * E * I)
# Create simple beam mesh (10 nodes along length)
num_nodes = 10
x_coords = np.linspace(0, length, num_nodes)
# Node features: [x, y, z, bc_x, bc_y, bc_z, bc_rx, bc_ry, bc_rz, load_x, load_y, load_z]
node_features = np.zeros((num_nodes, 12))
node_features[:, 0] = x_coords # x coordinates
# Boundary conditions at x=0 (fixed end)
node_features[0, 3:9] = 1.0 # All DOF constrained
# Applied force at x=length (free end)
node_features[-1, 10] = force # Force in y direction
# Create edges (connect adjacent nodes)
edge_index = []
for i in range(num_nodes - 1):
edge_index.append([i, i+1])
edge_index.append([i+1, i])
edge_index = torch.tensor(edge_index, dtype=torch.long).t()
# Edge features: [E, nu, rho, G, alpha]
num_edges = edge_index.shape[1]
edge_features = np.zeros((num_edges, 5))
edge_features[:, 0] = E / 1e11 # Normalized Young's modulus
edge_features[:, 1] = 0.3 # Poisson's ratio
edge_features[:, 2] = 7850 / 10000 # Normalized density
edge_features[:, 3] = E / (2 * (1 + 0.3)) / 1e11 # Normalized shear modulus
edge_features[:, 4] = 1.2e-5 # Thermal expansion
# Convert to tensors
x = torch.tensor(node_features, dtype=torch.float32)
edge_attr = torch.tensor(edge_features, dtype=torch.float32)
batch = torch.zeros(num_nodes, dtype=torch.long)
data = Data(x=x, edge_index=edge_index, edge_attr=edge_attr, batch=batch)
return data, analytical_displacement
def test_cantilever_analytical():
"""
Test 1: Cantilever beam with analytical solution
Expected: Neural prediction within 5% of δ = FL³/3EI
Note: This test uses an untrained model, so it will fail until
the model is trained on cantilever beam data. This test serves
as a template for post-training validation.
"""
print(" Creating cantilever beam test case...")
# Create test case
graph_data, analytical_disp = create_cantilever_beam_graph(
length=1.0,
force=1000.0,
E=210e9,
I=1e-6
)
print(f" Analytical max displacement: {analytical_disp*1000:.6f} mm")
# Create model (untrained)
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.1
}
model = create_model(config)
model.eval()
# Make prediction
print(" Running neural prediction...")
with torch.no_grad():
results = model(graph_data, return_stress=False)
# Extract max displacement (y-direction at free end)
predicted_disp = torch.max(torch.abs(results['displacement'][:, 1])).item()
print(f" Predicted max displacement: {predicted_disp:.6f} (arbitrary units)")
# Calculate error (will be large for untrained model)
# After training, this should be < 5%
error = abs(predicted_disp - analytical_disp) / analytical_disp * 100
print(f" Error: {error:.1f}%")
print(f" Note: Model is untrained. After training, expect < 5% error.")
# For now, just check that prediction completed
success = results['displacement'].shape[0] == graph_data.x.shape[0]
return {
'status': 'PASS' if success else 'FAIL',
'message': f'Cantilever test completed (untrained model)',
'metrics': {
'analytical_displacement_mm': float(analytical_disp * 1000),
'predicted_displacement': float(predicted_disp),
'error_percent': float(error),
'trained': False
}
}
def test_equilibrium():
"""
Test 2: Force equilibrium check
Expected: ∇·σ + f = 0 (force balance)
Checks that predicted stress field satisfies equilibrium.
For trained model, equilibrium residual should be < 1e-6.
"""
print(" Testing equilibrium constraint...")
# Create simple test case
num_nodes = 20
num_edges = 40
x = torch.randn(num_nodes, 12)
edge_index = torch.randint(0, num_nodes, (2, num_edges))
edge_attr = torch.randn(num_edges, 5)
batch = torch.zeros(num_nodes, dtype=torch.long)
data = Data(x=x, edge_index=edge_index, edge_attr=edge_attr, batch=batch)
# Create model
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.1
}
model = create_model(config)
model.eval()
# Make prediction
with torch.no_grad():
results = model(data, return_stress=True)
# Compute equilibrium residual (simplified check)
# In real implementation, would compute ∇·σ numerically
stress = results['stress']
stress_gradient_norm = torch.mean(torch.abs(stress)).item()
print(f" Stress field magnitude: {stress_gradient_norm:.6f}")
print(f" Note: Full equilibrium check requires mesh connectivity.")
print(f" After training with physics loss, residual should be < 1e-6.")
return {
'status': 'PASS',
'message': 'Equilibrium check completed',
'metrics': {
'stress_magnitude': float(stress_gradient_norm),
'trained': False
}
}
def test_energy_conservation():
"""
Test 3: Energy conservation
Expected: Strain energy = Work done by external forces
U = (1/2)∫ σ:ε dV = ∫ f·u dS
"""
print(" Testing energy conservation...")
# Create test case with known loading
num_nodes = 30
num_edges = 60
x = torch.randn(num_nodes, 12)
# Add known external force
x[:, 9:12] = 0.0 # Clear loads
x[0, 10] = 1000.0 # 1000 N in y direction at node 0
edge_index = torch.randint(0, num_nodes, (2, num_edges))
edge_attr = torch.randn(num_edges, 5)
batch = torch.zeros(num_nodes, dtype=torch.long)
data = Data(x=x, edge_index=edge_index, edge_attr=edge_attr, batch=batch)
# Create model
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.1
}
model = create_model(config)
model.eval()
# Make prediction
with torch.no_grad():
results = model(data, return_stress=True)
# Compute external work (simplified)
displacement = results['displacement']
force = x[:, 9:12]
external_work = torch.sum(force * displacement[:, :3]).item()
# Compute strain energy (simplified: U ≈ (1/2) σ:ε)
stress = results['stress']
# For small deformations: ε ≈ ∇u, approximate with displacement gradient
strain_energy = 0.5 * torch.sum(stress * displacement).item()
print(f" External work: {external_work:.6f}")
print(f" Strain energy: {strain_energy:.6f}")
print(f" Note: Simplified calculation. Full energy check requires")
print(f" proper strain computation from displacement gradients.")
return {
'status': 'PASS',
'message': 'Energy conservation check completed',
'metrics': {
'external_work': float(external_work),
'strain_energy': float(strain_energy),
'trained': False
}
}
def test_constitutive_law():
"""
Test 4: Constitutive law (Hooke's law)
Expected: σ = C:ε (stress proportional to strain)
For linear elastic materials: σ = E·ε for 1D
"""
print(" Testing constitutive law...")
# Create simple uniaxial test case
num_nodes = 10
# Simple bar under tension
x = torch.zeros(num_nodes, 12)
x[:, 0] = torch.linspace(0, 1, num_nodes) # x coordinates
# Fixed at x=0
x[0, 3:9] = 1.0
# Force at x=1
x[-1, 9] = 1000.0 # Axial force
# Create edges
edge_index = []
for i in range(num_nodes - 1):
edge_index.append([i, i+1])
edge_index.append([i+1, i])
edge_index = torch.tensor(edge_index, dtype=torch.long).t()
# Material properties
E = 210e9 # Young's modulus
edge_attr = torch.zeros(edge_index.shape[1], 5)
edge_attr[:, 0] = E / 1e11 # Normalized
edge_attr[:, 1] = 0.3 # Poisson's ratio
batch = torch.zeros(num_nodes, dtype=torch.long)
data = Data(x=x, edge_index=edge_index, edge_attr=edge_attr, batch=batch)
# Create model
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.1
}
model = create_model(config)
model.eval()
# Make prediction
with torch.no_grad():
results = model(data, return_stress=True)
# Check stress-strain relationship
displacement = results['displacement']
stress = results['stress']
# For trained model with physics loss, stress should follow σ = E·ε
print(f" Displacement range: {displacement[:, 0].min():.6f} to {displacement[:, 0].max():.6f}")
print(f" Stress range: {stress[:, 0].min():.6f} to {stress[:, 0].max():.6f}")
print(f" Note: After training with constitutive loss, stress should")
print(f" be proportional to strain (σ = E·ε).")
return {
'status': 'PASS',
'message': 'Constitutive law check completed',
'metrics': {
'displacement_range': [float(displacement[:, 0].min()), float(displacement[:, 0].max())],
'stress_range': [float(stress[:, 0].min()), float(stress[:, 0].max())],
'trained': False
}
}
if __name__ == "__main__":
print("\nRunning physics validation tests...\n")
tests = [
("Cantilever Analytical", test_cantilever_analytical),
("Equilibrium Check", test_equilibrium),
("Energy Conservation", test_energy_conservation),
("Constitutive Law", test_constitutive_law)
]
passed = 0
failed = 0
for name, test_func in tests:
print(f"[TEST] {name}")
try:
result = test_func()
if result['status'] == 'PASS':
print(f" ✓ PASS\n")
passed += 1
else:
print(f" ✗ FAIL: {result['message']}\n")
failed += 1
except Exception as e:
print(f" ✗ FAIL: {str(e)}\n")
import traceback
traceback.print_exc()
failed += 1
print(f"\nResults: {passed} passed, {failed} failed")
print(f"\nNote: These tests use an UNTRAINED model.")
print(f"After training with physics-informed losses, all tests should pass")
print(f"with errors < 5% for analytical solutions.")

462
atomizer-field/tests/test_predictions.py
View File

@@ -1,462 +0,0 @@
"""
test_predictions.py
Integration tests for complete pipeline
Tests the full system from parsing to prediction:
- Parser validation with real data
- Training pipeline end-to-end
- Prediction accuracy vs FEA
- Performance benchmarks
"""
import torch
import numpy as np
import sys
from pathlib import Path
import json
import time
# Add parent directory to path
sys.path.insert(0, str(Path(__file__).parent.parent))
from neural_field_parser import NastranToNeuralFieldParser
from neural_models.data_loader import FEAMeshDataset
from neural_models.field_predictor import create_model
from neural_models.physics_losses import create_loss_function
def test_parser():
"""
Test 1: Parser validation
Expected: Successfully parse BDF/OP2 files and create valid output
Uses test_case_beam if available, otherwise creates minimal test.
"""
print(" Checking for test data...")
test_dir = Path("test_case_beam")
if not test_dir.exists():
print(f" ⚠ Warning: {test_dir} not found")
print(f" Skipping parser test - run test_simple_beam.py first")
return {
'status': 'PASS',
'message': 'Parser test skipped (no test data)',
'metrics': {'skipped': True}
}
print(f" Found test directory: {test_dir}")
try:
# Check if already parsed
json_file = test_dir / "neural_field_data.json"
h5_file = test_dir / "neural_field_data.h5"
if json_file.exists() and h5_file.exists():
print(f" Found existing parsed data")
# Load and validate
with open(json_file, 'r') as f:
data = json.load(f)
n_nodes = data['mesh']['statistics']['n_nodes']
n_elements = data['mesh']['statistics']['n_elements']
print(f" Nodes: {n_nodes:,}")
print(f" Elements: {n_elements:,}")
return {
'status': 'PASS',
'message': 'Parser validation successful',
'metrics': {
'n_nodes': n_nodes,
'n_elements': n_elements,
'has_results': 'results' in data
}
}
else:
print(f" Parsed data not found - run test_simple_beam.py first")
return {
'status': 'PASS',
'message': 'Parser test skipped (data not parsed yet)',
'metrics': {'skipped': True}
}
except Exception as e:
print(f" Error: {str(e)}")
return {
'status': 'FAIL',
'message': f'Parser validation failed: {str(e)}',
'metrics': {}
}
def test_training():
"""
Test 2: Training pipeline
Expected: Complete training loop runs without errors
Trains on small synthetic dataset for speed.
"""
print(" Setting up training test...")
# Create minimal synthetic dataset
print(" Creating synthetic training data...")
dataset = []
for i in range(5): # Just 5 samples for quick test
num_nodes = 20
num_edges = 40
x = torch.randn(num_nodes, 12)
edge_index = torch.randint(0, num_nodes, (2, num_edges))
edge_attr = torch.randn(num_edges, 5)
batch = torch.zeros(num_nodes, dtype=torch.long)
from torch_geometric.data import Data
data = Data(x=x, edge_index=edge_index, edge_attr=edge_attr, batch=batch)
# Add synthetic targets
data.y_displacement = torch.randn(num_nodes, 6)
data.y_stress = torch.randn(num_nodes, 6)
dataset.append(data)
print(f" Created {len(dataset)} training samples")
# Create model
print(" Creating model...")
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.1
}
model = create_model(config)
loss_fn = create_loss_function('mse')
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
print(" Training for 10 epochs...")
# Training loop
model.train()
start_time = time.time()
for epoch in range(10):
epoch_loss = 0.0
for data in dataset:
optimizer.zero_grad()
# Forward pass
predictions = model(data, return_stress=True)
# Compute loss
targets = {
'displacement': data.y_displacement,
'stress': data.y_stress
}
loss_dict = loss_fn(predictions, targets)
loss = loss_dict['total_loss']
# Backward pass
loss.backward()
optimizer.step()
epoch_loss += loss.item()
avg_loss = epoch_loss / len(dataset)
if (epoch + 1) % 5 == 0:
print(f" Epoch {epoch+1}/10: Loss = {avg_loss:.6f}")
training_time = time.time() - start_time
print(f" Training completed in {training_time:.2f}s")
return {
'status': 'PASS',
'message': 'Training pipeline successful',
'metrics': {
'epochs': 10,
'samples': len(dataset),
'training_time_s': float(training_time),
'final_loss': float(avg_loss)
}
}
def test_prediction_accuracy():
"""
Test 3: Prediction accuracy
Expected: Predictions match targets with reasonable error
Uses trained model from test_training.
"""
print(" Testing prediction accuracy...")
# Create test case
num_nodes = 20
num_edges = 40
x = torch.randn(num_nodes, 12)
edge_index = torch.randint(0, num_nodes, (2, num_edges))
edge_attr = torch.randn(num_edges, 5)
batch = torch.zeros(num_nodes, dtype=torch.long)
from torch_geometric.data import Data
data = Data(x=x, edge_index=edge_index, edge_attr=edge_attr, batch=batch)
# Synthetic ground truth
target_disp = torch.randn(num_nodes, 6)
target_stress = torch.randn(num_nodes, 6)
# Create and "train" model (minimal training for test speed)
print(" Creating model...")
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.0
}
model = create_model(config)
# Quick training to make predictions reasonable
model.train()
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
loss_fn = create_loss_function('mse')
for _ in range(20):
optimizer.zero_grad()
predictions = model(data, return_stress=True)
targets = {
'displacement': target_disp,
'stress': target_stress
}
loss_dict = loss_fn(predictions, targets)
loss = loss_dict['total_loss']
loss.backward()
optimizer.step()
# Test prediction
print(" Running prediction...")
model.eval()
start_time = time.time()
with torch.no_grad():
predictions = model(data, return_stress=True)
inference_time = (time.time() - start_time) * 1000 # ms
# Compute errors
disp_error = torch.mean(torch.abs(predictions['displacement'] - target_disp)).item()
stress_error = torch.mean(torch.abs(predictions['stress'] - target_stress)).item()
print(f" Inference time: {inference_time:.2f} ms")
print(f" Displacement error: {disp_error:.6f}")
print(f" Stress error: {stress_error:.6f}")
return {
'status': 'PASS',
'message': 'Prediction accuracy test completed',
'metrics': {
'inference_time_ms': float(inference_time),
'displacement_error': float(disp_error),
'stress_error': float(stress_error),
'num_nodes': num_nodes
}
}
def test_performance_benchmark():
"""
Test 4: Performance benchmark
Expected: Inference time < 100ms for typical mesh
Compares neural prediction vs expected FEA time.
"""
print(" Running performance benchmark...")
# Test different mesh sizes
mesh_sizes = [10, 50, 100, 500]
results = []
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.0
}
model = create_model(config)
model.eval()
print(f" Testing {len(mesh_sizes)} mesh sizes...")
for num_nodes in mesh_sizes:
num_edges = num_nodes * 2
x = torch.randn(num_nodes, 12)
edge_index = torch.randint(0, num_nodes, (2, num_edges))
edge_attr = torch.randn(num_edges, 5)
batch = torch.zeros(num_nodes, dtype=torch.long)
from torch_geometric.data import Data
data = Data(x=x, edge_index=edge_index, edge_attr=edge_attr, batch=batch)
# Warm-up
with torch.no_grad():
_ = model(data, return_stress=True)
# Benchmark (average of 10 runs)
times = []
with torch.no_grad():
for _ in range(10):
start = time.time()
_ = model(data, return_stress=True)
times.append((time.time() - start) * 1000)
avg_time = np.mean(times)
std_time = np.std(times)
print(f" {num_nodes:4d} nodes: {avg_time:6.2f} ± {std_time:4.2f} ms")
results.append({
'num_nodes': num_nodes,
'avg_time_ms': float(avg_time),
'std_time_ms': float(std_time)
})
# Check if performance is acceptable (< 100ms for 100 nodes)
time_100_nodes = next((r['avg_time_ms'] for r in results if r['num_nodes'] == 100), None)
success = time_100_nodes is not None and time_100_nodes < 100.0
return {
'status': 'PASS' if success else 'FAIL',
'message': f'Performance benchmark completed',
'metrics': {
'results': results,
'time_100_nodes_ms': float(time_100_nodes) if time_100_nodes else None,
'passes_threshold': success
}
}
def test_batch_inference():
"""
Test 5: Batch inference
Expected: Can process multiple designs simultaneously
Important for optimization loops.
"""
print(" Testing batch inference...")
batch_size = 5
num_nodes_per_graph = 20
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.0
}
model = create_model(config)
model.eval()
print(f" Creating batch of {batch_size} graphs...")
graphs = []
for i in range(batch_size):
num_nodes = num_nodes_per_graph
num_edges = num_nodes * 2
x = torch.randn(num_nodes, 12)
edge_index = torch.randint(0, num_nodes, (2, num_edges))
edge_attr = torch.randn(num_edges, 5)
batch = torch.full((num_nodes,), i, dtype=torch.long)
from torch_geometric.data import Data
graphs.append(Data(x=x, edge_index=edge_index, edge_attr=edge_attr, batch=batch))
# Process batch
print(f" Processing batch...")
start_time = time.time()
with torch.no_grad():
for graph in graphs:
_ = model(graph, return_stress=True)
batch_time = (time.time() - start_time) * 1000
time_per_graph = batch_time / batch_size
print(f" Batch processing time: {batch_time:.2f} ms")
print(f" Time per graph: {time_per_graph:.2f} ms")
return {
'status': 'PASS',
'message': 'Batch inference successful',
'metrics': {
'batch_size': batch_size,
'total_time_ms': float(batch_time),
'time_per_graph_ms': float(time_per_graph)
}
}
if __name__ == "__main__":
print("\nRunning integration tests...\n")
tests = [
("Parser Validation", test_parser),
("Training Pipeline", test_training),
("Prediction Accuracy", test_prediction_accuracy),
("Performance Benchmark", test_performance_benchmark),
("Batch Inference", test_batch_inference)
]
passed = 0
failed = 0
for name, test_func in tests:
print(f"[TEST] {name}")
try:
result = test_func()
if result['status'] == 'PASS':
print(f" ✓ PASS\n")
passed += 1
else:
print(f" ✗ FAIL: {result['message']}\n")
failed += 1
except Exception as e:
print(f" ✗ FAIL: {str(e)}\n")
import traceback
traceback.print_exc()
failed += 1
print(f"\nResults: {passed} passed, {failed} failed")
print(f"\nNote: Parser test requires test_case_beam directory.")
print(f"Run 'python test_simple_beam.py' first to create test data.")

296
atomizer-field/tests/test_synthetic.py
View File

@@ -1,296 +0,0 @@
"""
test_synthetic.py
Synthetic tests with known analytical solutions
Tests basic functionality without real FEA data:
- Model can be created
- Forward pass works
- Loss functions compute correctly
- Predictions have correct shape
"""
import torch
import numpy as np
import sys
from pathlib import Path
# Add parent directory to path
sys.path.insert(0, str(Path(__file__).parent.parent))
from neural_models.field_predictor import create_model
from neural_models.physics_losses import create_loss_function
from torch_geometric.data import Data
def test_model_creation():
"""
Test 1: Can we create the model?
Expected: Model instantiates with correct number of parameters
"""
print(" Creating GNN model...")
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.1
}
model = create_model(config)
# Count parameters
num_params = sum(p.numel() for p in model.parameters())
print(f" Model created: {num_params:,} parameters")
return {
'status': 'PASS',
'message': f'Model created successfully ({num_params:,} params)',
'metrics': {'parameters': num_params}
}
def test_forward_pass():
"""
Test 2: Can model process data?
Expected: Forward pass completes without errors
"""
print(" Testing forward pass...")
# Create model
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.1
}
model = create_model(config)
model.eval()
# Create dummy data
num_nodes = 100
num_edges = 300
x = torch.randn(num_nodes, 12) # Node features
edge_index = torch.randint(0, num_nodes, (2, num_edges)) # Connectivity
edge_attr = torch.randn(num_edges, 5) # Edge features
batch = torch.zeros(num_nodes, dtype=torch.long) # Batch assignment
data = Data(x=x, edge_index=edge_index, edge_attr=edge_attr, batch=batch)
# Forward pass
with torch.no_grad():
results = model(data, return_stress=True)
# Check outputs
assert 'displacement' in results, "Missing displacement output"
assert 'stress' in results, "Missing stress output"
assert 'von_mises' in results, "Missing von Mises output"
# Check shapes
assert results['displacement'].shape == (num_nodes, 6), f"Wrong displacement shape: {results['displacement'].shape}"
assert results['stress'].shape == (num_nodes, 6), f"Wrong stress shape: {results['stress'].shape}"
assert results['von_mises'].shape == (num_nodes,), f"Wrong von Mises shape: {results['von_mises'].shape}"
print(f" Displacement shape: {results['displacement'].shape} [OK]")
print(f" Stress shape: {results['stress'].shape} [OK]")
print(f" Von Mises shape: {results['von_mises'].shape} [OK]")
return {
'status': 'PASS',
'message': 'Forward pass successful',
'metrics': {
'num_nodes': num_nodes,
'displacement_shape': list(results['displacement'].shape),
'stress_shape': list(results['stress'].shape)
}
}
def test_loss_computation():
"""
Test 3: Do loss functions work?
Expected: All loss types compute without errors
"""
print(" Testing loss functions...")
# Create dummy predictions and targets
num_nodes = 100
predictions = {
'displacement': torch.randn(num_nodes, 6),
'stress': torch.randn(num_nodes, 6),
'von_mises': torch.abs(torch.randn(num_nodes))
}
targets = {
'displacement': torch.randn(num_nodes, 6),
'stress': torch.randn(num_nodes, 6)
}
loss_types = ['mse', 'relative', 'physics', 'max']
loss_values = {}
for loss_type in loss_types:
loss_fn = create_loss_function(loss_type)
losses = loss_fn(predictions, targets)
assert 'total_loss' in losses, f"Missing total_loss for {loss_type}"
assert not torch.isnan(losses['total_loss']), f"NaN loss for {loss_type}"
assert not torch.isinf(losses['total_loss']), f"Inf loss for {loss_type}"
loss_values[loss_type] = losses['total_loss'].item()
print(f" {loss_type.upper()} loss: {loss_values[loss_type]:.6f} [OK]")
return {
'status': 'PASS',
'message': 'All loss functions working',
'metrics': loss_values
}
def test_batch_processing():
"""
Test 4: Can model handle batches?
Expected: Batch processing works correctly
"""
print(" Testing batch processing...")
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.1
}
model = create_model(config)
model.eval()
# Create batch of 3 graphs
graphs = []
for i in range(3):
num_nodes = 50 + i * 10 # Different sizes
num_edges = 150 + i * 30
x = torch.randn(num_nodes, 12)
edge_index = torch.randint(0, num_nodes, (2, num_edges))
edge_attr = torch.randn(num_edges, 5)
batch = torch.full((num_nodes,), i, dtype=torch.long)
graphs.append(Data(x=x, edge_index=edge_index, edge_attr=edge_attr, batch=batch))
# Process batch
total_nodes = sum(g.x.shape[0] for g in graphs)
with torch.no_grad():
for i, graph in enumerate(graphs):
results = model(graph, return_stress=True)
print(f" Graph {i+1}: {graph.x.shape[0]} nodes -> predictions [OK]")
return {
'status': 'PASS',
'message': 'Batch processing successful',
'metrics': {'num_graphs': len(graphs), 'total_nodes': total_nodes}
}
def test_gradient_flow():
"""
Test 5: Do gradients flow correctly?
Expected: Gradients computed without errors
"""
print(" Testing gradient flow...")
config = {
'node_feature_dim': 12,
'edge_feature_dim': 5,
'hidden_dim': 64,
'num_layers': 4,
'dropout': 0.1
}
model = create_model(config)
model.train()
# Create dummy data
num_nodes = 50
num_edges = 150
x = torch.randn(num_nodes, 12)
edge_index = torch.randint(0, num_nodes, (2, num_edges))
edge_attr = torch.randn(num_edges, 5)
batch = torch.zeros(num_nodes, dtype=torch.long)
data = Data(x=x, edge_index=edge_index, edge_attr=edge_attr, batch=batch)
# Forward pass
results = model(data, return_stress=True)
# Compute loss
targets = {
'displacement': torch.randn(num_nodes, 6),
'stress': torch.randn(num_nodes, 6)
}
loss_fn = create_loss_function('mse')
losses = loss_fn(results, targets)
# Backward pass
losses['total_loss'].backward()
# Check gradients
has_grad = sum(1 for p in model.parameters() if p.grad is not None)
total_params = sum(1 for _ in model.parameters())
print(f" Parameters with gradients: {has_grad}/{total_params} [OK]")
assert has_grad == total_params, f"Not all parameters have gradients"
return {
'status': 'PASS',
'message': 'Gradients computed successfully',
'metrics': {
'parameters_with_grad': has_grad,
'total_parameters': total_params
}
}
if __name__ == "__main__":
print("\nRunning synthetic tests...\n")
tests = [
("Model Creation", test_model_creation),
("Forward Pass", test_forward_pass),
("Loss Computation", test_loss_computation),
("Batch Processing", test_batch_processing),
("Gradient Flow", test_gradient_flow)
]
passed = 0
failed = 0
for name, test_func in tests:
print(f"[TEST] {name}")
try:
result = test_func()
if result['status'] == 'PASS':
print(f" [PASS]\n")
passed += 1
else:
print(f" [FAIL]: {result['message']}\n")
failed += 1
except Exception as e:
print(f" [FAIL]: {str(e)}\n")
failed += 1
print(f"\nResults: {passed} passed, {failed} failed")

451
atomizer-field/train.py
View File

@@ -1,451 +0,0 @@
"""
train.py
Training script for AtomizerField neural field predictor
AtomizerField Training Pipeline v2.0
Trains Graph Neural Networks to predict complete FEA field results.
Usage:
python train.py --train_dir ./training_data --val_dir ./validation_data
Key Features:
- Multi-GPU support
- Checkpoint saving/loading
- TensorBoard logging
- Early stopping
- Learning rate scheduling
"""
import argparse
import json
from pathlib import Path
import time
from datetime import datetime
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.tensorboard import SummaryWriter
from neural_models.field_predictor import create_model, AtomizerFieldModel
from neural_models.physics_losses import create_loss_function
from neural_models.data_loader import create_dataloaders
class Trainer:
"""
Training manager for AtomizerField models
"""
def __init__(self, config):
"""
Initialize trainer
Args:
config (dict): Training configuration
"""
self.config = config
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"\n{'='*60}")
print("AtomizerField Training Pipeline v2.0")
print(f"{'='*60}")
print(f"Device: {self.device}")
# Create model
print("\nCreating model...")
self.model = create_model(config.get('model', {}))
self.model = self.model.to(self.device)
num_params = sum(p.numel() for p in self.model.parameters())
print(f"Model created: {num_params:,} parameters")
# Create loss function
loss_config = config.get('loss', {})
loss_type = loss_config.pop('type', 'mse')
self.criterion = create_loss_function(loss_type, loss_config)
print(f"Loss function: {loss_type}")
# Create optimizer
self.optimizer = optim.AdamW(
self.model.parameters(),
lr=config.get('learning_rate', 1e-3),
weight_decay=config.get('weight_decay', 1e-5)
)
# Learning rate scheduler
self.scheduler = optim.lr_scheduler.ReduceLROnPlateau(
self.optimizer,
mode='min',
factor=0.5,
patience=10,
verbose=True
)
# Training state
self.start_epoch = 0
self.best_val_loss = float('inf')
self.epochs_without_improvement = 0
# Create output directories
self.output_dir = Path(config.get('output_dir', './runs'))
self.output_dir.mkdir(parents=True, exist_ok=True)
# TensorBoard logging
self.writer = SummaryWriter(
log_dir=self.output_dir / 'tensorboard'
)
# Save config
with open(self.output_dir / 'config.json', 'w') as f:
json.dump(config, f, indent=2)
def train_epoch(self, train_loader, epoch):
"""
Train for one epoch
Args:
train_loader: Training data loader
epoch (int): Current epoch number
Returns:
dict: Training metrics
"""
self.model.train()
total_loss = 0.0
total_disp_loss = 0.0
total_stress_loss = 0.0
num_batches = 0
for batch_idx, batch in enumerate(train_loader):
# Move batch to device
batch = batch.to(self.device)
# Zero gradients
self.optimizer.zero_grad()
# Forward pass
predictions = self.model(batch, return_stress=True)
# Prepare targets
targets = {
'displacement': batch.y_displacement,
}
if hasattr(batch, 'y_stress'):
targets['stress'] = batch.y_stress
# Compute loss
losses = self.criterion(predictions, targets, batch)
# Backward pass
losses['total_loss'].backward()
# Gradient clipping (prevents exploding gradients)
torch.nn.utils.clip_grad_norm_(self.model.parameters(), max_norm=1.0)
# Update weights
self.optimizer.step()
# Accumulate metrics
total_loss += losses['total_loss'].item()
if 'displacement_loss' in losses:
total_disp_loss += losses['displacement_loss'].item()
if 'stress_loss' in losses:
total_stress_loss += losses['stress_loss'].item()
num_batches += 1
# Print progress
if batch_idx % 10 == 0:
print(f" Batch {batch_idx}/{len(train_loader)}: "
f"Loss={losses['total_loss'].item():.6f}")
# Average metrics
metrics = {
'total_loss': total_loss / num_batches,
'displacement_loss': total_disp_loss / num_batches,
'stress_loss': total_stress_loss / num_batches
}
return metrics
def validate(self, val_loader):
"""
Validate model
Args:
val_loader: Validation data loader
Returns:
dict: Validation metrics
"""
self.model.eval()
total_loss = 0.0
total_disp_loss = 0.0
total_stress_loss = 0.0
num_batches = 0
with torch.no_grad():
for batch in val_loader:
# Move batch to device
batch = batch.to(self.device)
# Forward pass
predictions = self.model(batch, return_stress=True)
# Prepare targets
targets = {
'displacement': batch.y_displacement,
}
if hasattr(batch, 'y_stress'):
targets['stress'] = batch.y_stress
# Compute loss
losses = self.criterion(predictions, targets, batch)
# Accumulate metrics
total_loss += losses['total_loss'].item()
if 'displacement_loss' in losses:
total_disp_loss += losses['displacement_loss'].item()
if 'stress_loss' in losses:
total_stress_loss += losses['stress_loss'].item()
num_batches += 1
# Average metrics
metrics = {
'total_loss': total_loss / num_batches,
'displacement_loss': total_disp_loss / num_batches,
'stress_loss': total_stress_loss / num_batches
}
return metrics
def train(self, train_loader, val_loader, num_epochs):
"""
Main training loop
Args:
train_loader: Training data loader
val_loader: Validation data loader
num_epochs (int): Number of epochs to train
"""
print(f"\n{'='*60}")
print(f"Starting training for {num_epochs} epochs")
print(f"{'='*60}\n")
for epoch in range(self.start_epoch, num_epochs):
epoch_start_time = time.time()
print(f"Epoch {epoch + 1}/{num_epochs}")
print("-" * 60)
# Train
train_metrics = self.train_epoch(train_loader, epoch)
# Validate
val_metrics = self.validate(val_loader)
epoch_time = time.time() - epoch_start_time
# Print metrics
print(f"\nEpoch {epoch + 1} Results:")
print(f" Training Loss: {train_metrics['total_loss']:.6f}")
print(f" Displacement: {train_metrics['displacement_loss']:.6f}")
print(f" Stress: {train_metrics['stress_loss']:.6f}")
print(f" Validation Loss: {val_metrics['total_loss']:.6f}")
print(f" Displacement: {val_metrics['displacement_loss']:.6f}")
print(f" Stress: {val_metrics['stress_loss']:.6f}")
print(f" Time: {epoch_time:.1f}s")
# Log to TensorBoard
self.writer.add_scalar('Loss/train', train_metrics['total_loss'], epoch)
self.writer.add_scalar('Loss/val', val_metrics['total_loss'], epoch)
self.writer.add_scalar('DisplacementLoss/train', train_metrics['displacement_loss'], epoch)
self.writer.add_scalar('DisplacementLoss/val', val_metrics['displacement_loss'], epoch)
self.writer.add_scalar('StressLoss/train', train_metrics['stress_loss'], epoch)
self.writer.add_scalar('StressLoss/val', val_metrics['stress_loss'], epoch)
self.writer.add_scalar('LearningRate', self.optimizer.param_groups[0]['lr'], epoch)
# Learning rate scheduling
self.scheduler.step(val_metrics['total_loss'])
# Save checkpoint
is_best = val_metrics['total_loss'] < self.best_val_loss
if is_best:
self.best_val_loss = val_metrics['total_loss']
self.epochs_without_improvement = 0
print(f" New best validation loss: {self.best_val_loss:.6f}")
else:
self.epochs_without_improvement += 1
self.save_checkpoint(epoch, val_metrics, is_best)
# Early stopping
patience = self.config.get('early_stopping_patience', 50)
if self.epochs_without_improvement >= patience:
print(f"\nEarly stopping after {patience} epochs without improvement")
break
print()
print(f"\n{'='*60}")
print("Training complete!")
print(f"Best validation loss: {self.best_val_loss:.6f}")
print(f"{'='*60}\n")
self.writer.close()
def save_checkpoint(self, epoch, metrics, is_best=False):
"""
Save model checkpoint
Args:
epoch (int): Current epoch
metrics (dict): Validation metrics
is_best (bool): Whether this is the best model so far
"""
checkpoint = {
'epoch': epoch,
'model_state_dict': self.model.state_dict(),
'optimizer_state_dict': self.optimizer.state_dict(),
'scheduler_state_dict': self.scheduler.state_dict(),
'best_val_loss': self.best_val_loss,
'config': self.config,
'metrics': metrics
}
# Save latest checkpoint
checkpoint_path = self.output_dir / 'checkpoint_latest.pt'
torch.save(checkpoint, checkpoint_path)
# Save best checkpoint
if is_best:
best_path = self.output_dir / 'checkpoint_best.pt'
torch.save(checkpoint, best_path)
print(f" Saved best model to {best_path}")
# Save periodic checkpoint
if (epoch + 1) % 10 == 0:
periodic_path = self.output_dir / f'checkpoint_epoch_{epoch + 1}.pt'
torch.save(checkpoint, periodic_path)
def load_checkpoint(self, checkpoint_path):
"""
Load model checkpoint
Args:
checkpoint_path (str): Path to checkpoint file
"""
checkpoint = torch.load(checkpoint_path, map_location=self.device)
self.model.load_state_dict(checkpoint['model_state_dict'])
self.optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
self.scheduler.load_state_dict(checkpoint['scheduler_state_dict'])
self.start_epoch = checkpoint['epoch'] + 1
self.best_val_loss = checkpoint['best_val_loss']
print(f"Loaded checkpoint from epoch {checkpoint['epoch']}")
print(f"Best validation loss: {self.best_val_loss:.6f}")
def main():
"""
Main training entry point
"""
parser = argparse.ArgumentParser(description='Train AtomizerField neural field predictor')
# Data arguments
parser.add_argument('--train_dir', type=str, required=True,
help='Directory containing training cases')
parser.add_argument('--val_dir', type=str, required=True,
help='Directory containing validation cases')
# Training arguments
parser.add_argument('--epochs', type=int, default=100,
help='Number of training epochs')
parser.add_argument('--batch_size', type=int, default=4,
help='Batch size')
parser.add_argument('--lr', type=float, default=1e-3,
help='Learning rate')
parser.add_argument('--weight_decay', type=float, default=1e-5,
help='Weight decay')
# Model arguments
parser.add_argument('--hidden_dim', type=int, default=128,
help='Hidden dimension')
parser.add_argument('--num_layers', type=int, default=6,
help='Number of GNN layers')
parser.add_argument('--dropout', type=float, default=0.1,
help='Dropout rate')
# Loss arguments
parser.add_argument('--loss_type', type=str, default='mse',
choices=['mse', 'relative', 'physics', 'max'],
help='Loss function type')
# Other arguments
parser.add_argument('--output_dir', type=str, default='./runs',
help='Output directory for checkpoints and logs')
parser.add_argument('--resume', type=str, default=None,
help='Path to checkpoint to resume from')
parser.add_argument('--num_workers', type=int, default=0,
help='Number of data loading workers')
args = parser.parse_args()
# Build configuration
config = {
'model': {
'node_feature_dim': 12, # 3 coords + 6 BCs + 3 loads
'edge_feature_dim': 5, # E, nu, rho, G, alpha
'hidden_dim': args.hidden_dim,
'num_layers': args.num_layers,
'dropout': args.dropout
},
'loss': {
'type': args.loss_type
},
'learning_rate': args.lr,
'weight_decay': args.weight_decay,
'batch_size': args.batch_size,
'num_epochs': args.epochs,
'output_dir': args.output_dir,
'early_stopping_patience': 50
}
# Find all case directories
train_cases = list(Path(args.train_dir).glob('*/'))
val_cases = list(Path(args.val_dir).glob('*/'))
print(f"Found {len(train_cases)} training cases")
print(f"Found {len(val_cases)} validation cases")
if not train_cases or not val_cases:
print("ERROR: No training or validation cases found!")
print("Please ensure your directories contain parsed FEA data.")
return
# Create data loaders
train_loader, val_loader = create_dataloaders(
train_cases,
val_cases,
batch_size=args.batch_size,
num_workers=args.num_workers,
normalize=True,
include_stress=True
)
# Create trainer
trainer = Trainer(config)
# Resume from checkpoint if specified
if args.resume:
trainer.load_checkpoint(args.resume)
# Train
trainer.train(train_loader, val_loader, args.epochs)
if __name__ == "__main__":
main()

789
atomizer-field/train_parametric.py
View File

@@ -1,789 +0,0 @@
"""
train_parametric.py
Training script for AtomizerField parametric predictor
AtomizerField Parametric Training Pipeline v2.0
Trains design-conditioned GNN to predict optimization objectives directly.
Usage:
python train_parametric.py --train_dir ./training_data --val_dir ./validation_data
Key Differences from train.py (field predictor):
- Predicts scalar objectives (mass, frequency, displacement, stress) instead of fields
- Uses design parameters as conditioning input
- Multi-objective loss function for all 4 outputs
- Faster training due to simpler output structure
Output:
checkpoint_best.pt containing:
- model_state_dict: Trained weights
- config: Model configuration
- normalization: Normalization statistics for inference
- design_var_names: Names of design variables
- best_val_loss: Best validation loss achieved
"""
import argparse
import json
import sys
from pathlib import Path
import time
from datetime import datetime
from typing import Dict, List, Any, Optional, Tuple
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
import h5py
# Try to import tensorboard, but make it optional
try:
from torch.utils.tensorboard import SummaryWriter
TENSORBOARD_AVAILABLE = True
except ImportError:
TENSORBOARD_AVAILABLE = False
from neural_models.parametric_predictor import create_parametric_model, ParametricFieldPredictor
class ParametricDataset(Dataset):
"""
PyTorch Dataset for parametric training.
Loads training data exported by Atomizer's TrainingDataExporter
and prepares it for the parametric predictor.
Expected directory structure:
training_data/
├── trial_0000/
│ ├── metadata.json (design params + results)
│ └── input/
│ └── neural_field_data.h5 (mesh data)
├── trial_0001/
│ └── ...
"""
def __init__(
self,
data_dir: Path,
normalize: bool = True,
cache_in_memory: bool = False
):
"""
Initialize dataset.
Args:
data_dir: Directory containing trial_* subdirectories
normalize: Whether to normalize inputs/outputs
cache_in_memory: Cache all data in RAM (faster but memory-intensive)
"""
self.data_dir = Path(data_dir)
self.normalize = normalize
self.cache_in_memory = cache_in_memory
# Find all valid trial directories
self.trial_dirs = sorted([
d for d in self.data_dir.glob("trial_*")
if d.is_dir() and self._is_valid_trial(d)
])
print(f"Found {len(self.trial_dirs)} valid trials in {data_dir}")
if len(self.trial_dirs) == 0:
raise ValueError(f"No valid trial directories found in {data_dir}")
# Extract design variable names from first trial
self.design_var_names = self._get_design_var_names()
print(f"Design variables: {self.design_var_names}")
# Compute normalization statistics
if normalize:
self._compute_normalization_stats()
# Cache data if requested
self.cache = {}
if cache_in_memory:
print("Caching data in memory...")
for idx in range(len(self.trial_dirs)):
self.cache[idx] = self._load_trial(idx)
print("Cache complete!")
def _is_valid_trial(self, trial_dir: Path) -> bool:
"""Check if trial directory has required files."""
metadata_file = trial_dir / "metadata.json"
# Check for metadata
if not metadata_file.exists():
return False
# Check metadata has required fields
try:
with open(metadata_file, 'r') as f:
metadata = json.load(f)
has_design = 'design_parameters' in metadata
has_results = 'results' in metadata
return has_design and has_results
except:
return False
def _get_design_var_names(self) -> List[str]:
"""Extract design variable names from first trial."""
metadata_file = self.trial_dirs[0] / "metadata.json"
with open(metadata_file, 'r') as f:
metadata = json.load(f)
return list(metadata['design_parameters'].keys())
def _compute_normalization_stats(self):
"""Compute normalization statistics across all trials."""
print("Computing normalization statistics...")
all_design_params = []
all_mass = []
all_disp = []
all_stiffness = []
for trial_dir in self.trial_dirs:
with open(trial_dir / "metadata.json", 'r') as f:
metadata = json.load(f)
# Design parameters
design_params = [metadata['design_parameters'][name]
for name in self.design_var_names]
all_design_params.append(design_params)
# Results
results = metadata.get('results', {})
objectives = results.get('objectives', results)
if 'mass' in objectives:
all_mass.append(objectives['mass'])
if 'max_displacement' in results:
all_disp.append(results['max_displacement'])
elif 'max_displacement' in objectives:
all_disp.append(objectives['max_displacement'])
if 'stiffness' in objectives:
all_stiffness.append(objectives['stiffness'])
# Convert to numpy arrays
all_design_params = np.array(all_design_params)
# Compute statistics
self.design_mean = torch.from_numpy(all_design_params.mean(axis=0)).float()
self.design_std = torch.from_numpy(all_design_params.std(axis=0)).float()
self.design_std = torch.clamp(self.design_std, min=1e-6) # Prevent division by zero
# Output statistics
self.mass_mean = np.mean(all_mass) if all_mass else 0.1
self.mass_std = np.std(all_mass) if all_mass else 0.05
self.mass_std = max(self.mass_std, 1e-6)
self.disp_mean = np.mean(all_disp) if all_disp else 0.01
self.disp_std = np.std(all_disp) if all_disp else 0.005
self.disp_std = max(self.disp_std, 1e-6)
self.stiffness_mean = np.mean(all_stiffness) if all_stiffness else 20000.0
self.stiffness_std = np.std(all_stiffness) if all_stiffness else 5000.0
self.stiffness_std = max(self.stiffness_std, 1e-6)
# Frequency and stress defaults (if not available in data)
self.freq_mean = 18.0
self.freq_std = 5.0
self.stress_mean = 200.0
self.stress_std = 50.0
print(f" Design mean: {self.design_mean.numpy()}")
print(f" Design std: {self.design_std.numpy()}")
print(f" Mass: {self.mass_mean:.4f} +/- {self.mass_std:.4f}")
print(f" Displacement: {self.disp_mean:.6f} +/- {self.disp_std:.6f}")
print(f" Stiffness: {self.stiffness_mean:.2f} +/- {self.stiffness_std:.2f}")
def __len__(self) -> int:
return len(self.trial_dirs)
def __getitem__(self, idx: int) -> Dict[str, torch.Tensor]:
if self.cache_in_memory and idx in self.cache:
return self.cache[idx]
return self._load_trial(idx)
def _load_trial(self, idx: int) -> Dict[str, torch.Tensor]:
"""Load and process a single trial."""
trial_dir = self.trial_dirs[idx]
# Load metadata
with open(trial_dir / "metadata.json", 'r') as f:
metadata = json.load(f)
# Extract design parameters
design_params = [metadata['design_parameters'][name]
for name in self.design_var_names]
design_tensor = torch.tensor(design_params, dtype=torch.float32)
# Normalize design parameters
if self.normalize:
design_tensor = (design_tensor - self.design_mean) / self.design_std
# Extract results
results = metadata.get('results', {})
objectives = results.get('objectives', results)
# Get targets (with fallbacks)
mass = objectives.get('mass', 0.1)
stiffness = objectives.get('stiffness', 20000.0)
max_displacement = results.get('max_displacement',
objectives.get('max_displacement', 0.01))
# Frequency and stress might not be available
frequency = objectives.get('frequency', self.freq_mean)
max_stress = objectives.get('max_stress', self.stress_mean)
# Create target tensor
targets = torch.tensor([mass, frequency, max_displacement, max_stress],
dtype=torch.float32)
# Normalize targets
if self.normalize:
targets[0] = (targets[0] - self.mass_mean) / self.mass_std
targets[1] = (targets[1] - self.freq_mean) / self.freq_std
targets[2] = (targets[2] - self.disp_mean) / self.disp_std
targets[3] = (targets[3] - self.stress_mean) / self.stress_std
# Try to load mesh data if available
mesh_data = self._load_mesh_data(trial_dir)
return {
'design_params': design_tensor,
'targets': targets,
'mesh_data': mesh_data,
'trial_dir': str(trial_dir)
}
def _load_mesh_data(self, trial_dir: Path) -> Optional[Dict[str, torch.Tensor]]:
"""Load mesh data from H5 file if available."""
h5_paths = [
trial_dir / "input" / "neural_field_data.h5",
trial_dir / "neural_field_data.h5",
]
for h5_path in h5_paths:
if h5_path.exists():
try:
with h5py.File(h5_path, 'r') as f:
node_coords = torch.from_numpy(f['mesh/node_coordinates'][:]).float()
# Build simple edge index from connectivity if available
# For now, return just coordinates
return {
'node_coords': node_coords,
'num_nodes': node_coords.shape[0]
}
except Exception as e:
print(f"Warning: Could not load mesh from {h5_path}: {e}")
return None
def get_normalization_stats(self) -> Dict[str, Any]:
"""Return normalization statistics for saving with model."""
return {
'design_mean': self.design_mean.numpy().tolist(),
'design_std': self.design_std.numpy().tolist(),
'mass_mean': float(self.mass_mean),
'mass_std': float(self.mass_std),
'freq_mean': float(self.freq_mean),
'freq_std': float(self.freq_std),
'max_disp_mean': float(self.disp_mean),
'max_disp_std': float(self.disp_std),
'max_stress_mean': float(self.stress_mean),
'max_stress_std': float(self.stress_std),
}
def create_reference_graph(num_nodes: int = 500, device: torch.device = None):
"""
Create a reference graph structure for the GNN.
In production, this would come from the actual mesh.
For now, create a simple grid-like structure.
"""
if device is None:
device = torch.device('cpu')
# Create simple node features (placeholder)
x = torch.randn(num_nodes, 12, device=device)
# Create grid-like connectivity - ensure valid indices
edges = []
grid_size = int(np.ceil(np.sqrt(num_nodes)))
for i in range(num_nodes):
row = i // grid_size
col = i % grid_size
# Right neighbor (same row)
right = i + 1
if col < grid_size - 1 and right < num_nodes:
edges.append([i, right])
edges.append([right, i])
# Bottom neighbor (next row)
bottom = i + grid_size
if bottom < num_nodes:
edges.append([i, bottom])
edges.append([bottom, i])
# Ensure we have at least some edges
if len(edges) == 0:
# Fallback: fully connected for very small graphs
for i in range(num_nodes):
for j in range(i + 1, min(i + 5, num_nodes)):
edges.append([i, j])
edges.append([j, i])
edge_index = torch.tensor(edges, dtype=torch.long, device=device).t().contiguous()
edge_attr = torch.randn(edge_index.shape[1], 5, device=device)
from torch_geometric.data import Data
return Data(x=x, edge_index=edge_index, edge_attr=edge_attr)
class ParametricTrainer:
"""
Training manager for parametric predictor models.
"""
def __init__(self, config: Dict[str, Any]):
"""
Initialize trainer.
Args:
config: Training configuration
"""
self.config = config
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"\n{'='*60}")
print("AtomizerField Parametric Training Pipeline v2.0")
print(f"{'='*60}")
print(f"Device: {self.device}")
# Create model
print("\nCreating parametric model...")
model_config = config.get('model', {})
self.model = create_parametric_model(model_config)
self.model = self.model.to(self.device)
num_params = self.model.get_num_parameters()
print(f"Model created: {num_params:,} parameters")
# Create optimizer
self.optimizer = optim.AdamW(
self.model.parameters(),
lr=config.get('learning_rate', 1e-3),
weight_decay=config.get('weight_decay', 1e-5)
)
# Learning rate scheduler
self.scheduler = optim.lr_scheduler.ReduceLROnPlateau(
self.optimizer,
mode='min',
factor=0.5,
patience=10,
verbose=True
)
# Multi-objective loss weights
self.loss_weights = config.get('loss_weights', {
'mass': 1.0,
'frequency': 1.0,
'displacement': 1.0,
'stress': 1.0
})
# Training state
self.start_epoch = 0
self.best_val_loss = float('inf')
self.epochs_without_improvement = 0
# Create output directories
self.output_dir = Path(config.get('output_dir', './runs/parametric'))
self.output_dir.mkdir(parents=True, exist_ok=True)
# TensorBoard logging (optional)
self.writer = None
if TENSORBOARD_AVAILABLE:
self.writer = SummaryWriter(log_dir=self.output_dir / 'tensorboard')
# Create reference graph for inference
self.reference_graph = create_reference_graph(
num_nodes=config.get('reference_nodes', 500),
device=self.device
)
# Save config
with open(self.output_dir / 'config.json', 'w') as f:
json.dump(config, f, indent=2)
def compute_loss(
self,
predictions: Dict[str, torch.Tensor],
targets: torch.Tensor
) -> Tuple[torch.Tensor, Dict[str, float]]:
"""
Compute multi-objective loss.
Args:
predictions: Model outputs (mass, frequency, max_displacement, max_stress)
targets: Target values [batch, 4]
Returns:
total_loss: Combined loss tensor
loss_dict: Individual losses for logging
"""
# MSE losses for each objective
mass_loss = nn.functional.mse_loss(predictions['mass'], targets[:, 0])
freq_loss = nn.functional.mse_loss(predictions['frequency'], targets[:, 1])
disp_loss = nn.functional.mse_loss(predictions['max_displacement'], targets[:, 2])
stress_loss = nn.functional.mse_loss(predictions['max_stress'], targets[:, 3])
# Weighted combination
total_loss = (
self.loss_weights['mass'] * mass_loss +
self.loss_weights['frequency'] * freq_loss +
self.loss_weights['displacement'] * disp_loss +
self.loss_weights['stress'] * stress_loss
)
loss_dict = {
'total': total_loss.item(),
'mass': mass_loss.item(),
'frequency': freq_loss.item(),
'displacement': disp_loss.item(),
'stress': stress_loss.item()
}
return total_loss, loss_dict
def train_epoch(self, train_loader: DataLoader, epoch: int) -> Dict[str, float]:
"""Train for one epoch."""
self.model.train()
total_losses = {'total': 0, 'mass': 0, 'frequency': 0, 'displacement': 0, 'stress': 0}
num_batches = 0
for batch_idx, batch in enumerate(train_loader):
# Get data
design_params = batch['design_params'].to(self.device)
targets = batch['targets'].to(self.device)
# Zero gradients
self.optimizer.zero_grad()
# Forward pass (using reference graph)
predictions = self.model(self.reference_graph, design_params)
# Compute loss
loss, loss_dict = self.compute_loss(predictions, targets)
# Backward pass
loss.backward()
# Gradient clipping
torch.nn.utils.clip_grad_norm_(self.model.parameters(), max_norm=1.0)
# Update weights
self.optimizer.step()
# Accumulate losses
for key in total_losses:
total_losses[key] += loss_dict[key]
num_batches += 1
# Print progress
if batch_idx % 10 == 0:
print(f" Batch {batch_idx}/{len(train_loader)}: Loss={loss_dict['total']:.6f}")
# Average losses
return {k: v / num_batches for k, v in total_losses.items()}
def validate(self, val_loader: DataLoader) -> Dict[str, float]:
"""Validate model."""
self.model.eval()
total_losses = {'total': 0, 'mass': 0, 'frequency': 0, 'displacement': 0, 'stress': 0}
num_batches = 0
with torch.no_grad():
for batch in val_loader:
design_params = batch['design_params'].to(self.device)
targets = batch['targets'].to(self.device)
predictions = self.model(self.reference_graph, design_params)
_, loss_dict = self.compute_loss(predictions, targets)
for key in total_losses:
total_losses[key] += loss_dict[key]
num_batches += 1
return {k: v / num_batches for k, v in total_losses.items()}
def train(
self,
train_loader: DataLoader,
val_loader: DataLoader,
num_epochs: int,
train_dataset: ParametricDataset
):
"""
Main training loop.
Args:
train_loader: Training data loader
val_loader: Validation data loader
num_epochs: Number of epochs
train_dataset: Training dataset (for normalization stats)
"""
print(f"\n{'='*60}")
print(f"Starting training for {num_epochs} epochs")
print(f"{'='*60}\n")
for epoch in range(self.start_epoch, num_epochs):
epoch_start = time.time()
print(f"Epoch {epoch + 1}/{num_epochs}")
print("-" * 60)
# Train
train_metrics = self.train_epoch(train_loader, epoch)
# Validate
val_metrics = self.validate(val_loader)
epoch_time = time.time() - epoch_start
# Print metrics
print(f"\nEpoch {epoch + 1} Results:")
print(f" Training Loss: {train_metrics['total']:.6f}")
print(f" Mass: {train_metrics['mass']:.6f}, Freq: {train_metrics['frequency']:.6f}")
print(f" Disp: {train_metrics['displacement']:.6f}, Stress: {train_metrics['stress']:.6f}")
print(f" Validation Loss: {val_metrics['total']:.6f}")
print(f" Mass: {val_metrics['mass']:.6f}, Freq: {val_metrics['frequency']:.6f}")
print(f" Disp: {val_metrics['displacement']:.6f}, Stress: {val_metrics['stress']:.6f}")
print(f" Time: {epoch_time:.1f}s")
# Log to TensorBoard
if self.writer:
self.writer.add_scalar('Loss/train', train_metrics['total'], epoch)
self.writer.add_scalar('Loss/val', val_metrics['total'], epoch)
for key in ['mass', 'frequency', 'displacement', 'stress']:
self.writer.add_scalar(f'{key}/train', train_metrics[key], epoch)
self.writer.add_scalar(f'{key}/val', val_metrics[key], epoch)
# Learning rate scheduling
self.scheduler.step(val_metrics['total'])
# Save checkpoint
is_best = val_metrics['total'] < self.best_val_loss
if is_best:
self.best_val_loss = val_metrics['total']
self.epochs_without_improvement = 0
print(f" New best validation loss: {self.best_val_loss:.6f}")
else:
self.epochs_without_improvement += 1
self.save_checkpoint(epoch, val_metrics, train_dataset, is_best)
# Early stopping
patience = self.config.get('early_stopping_patience', 50)
if self.epochs_without_improvement >= patience:
print(f"\nEarly stopping after {patience} epochs without improvement")
break
print()
print(f"\n{'='*60}")
print("Training complete!")
print(f"Best validation loss: {self.best_val_loss:.6f}")
print(f"{'='*60}\n")
if self.writer:
self.writer.close()
def save_checkpoint(
self,
epoch: int,
metrics: Dict[str, float],
dataset: ParametricDataset,
is_best: bool = False
):
"""Save model checkpoint with all required metadata."""
checkpoint = {
'epoch': epoch,
'model_state_dict': self.model.state_dict(),
'optimizer_state_dict': self.optimizer.state_dict(),
'scheduler_state_dict': self.scheduler.state_dict(),
'best_val_loss': self.best_val_loss,
'config': self.model.config,
'normalization': dataset.get_normalization_stats(),
'design_var_names': dataset.design_var_names,
'metrics': metrics
}
# Save latest
torch.save(checkpoint, self.output_dir / 'checkpoint_latest.pt')
# Save best
if is_best:
best_path = self.output_dir / 'checkpoint_best.pt'
torch.save(checkpoint, best_path)
print(f" Saved best model to {best_path}")
# Periodic checkpoint
if (epoch + 1) % 10 == 0:
torch.save(checkpoint, self.output_dir / f'checkpoint_epoch_{epoch + 1}.pt')
def collate_fn(batch: List[Dict]) -> Dict[str, torch.Tensor]:
"""Custom collate function for DataLoader."""
design_params = torch.stack([item['design_params'] for item in batch])
targets = torch.stack([item['targets'] for item in batch])
return {
'design_params': design_params,
'targets': targets
}
def main():
"""Main training entry point."""
parser = argparse.ArgumentParser(
description='Train AtomizerField parametric predictor'
)
# Data arguments
parser.add_argument('--train_dir', type=str, required=True,
help='Directory containing training trial_* subdirs')
parser.add_argument('--val_dir', type=str, default=None,
help='Directory containing validation data (uses split if not provided)')
parser.add_argument('--val_split', type=float, default=0.2,
help='Validation split ratio if val_dir not provided')
# Training arguments
parser.add_argument('--epochs', type=int, default=200,
help='Number of training epochs')
parser.add_argument('--batch_size', type=int, default=16,
help='Batch size')
parser.add_argument('--learning_rate', type=float, default=1e-3,
help='Learning rate')
parser.add_argument('--weight_decay', type=float, default=1e-5,
help='Weight decay')
# Model arguments
parser.add_argument('--hidden_channels', type=int, default=128,
help='Hidden dimension')
parser.add_argument('--num_layers', type=int, default=4,
help='Number of GNN layers')
parser.add_argument('--dropout', type=float, default=0.1,
help='Dropout rate')
# Output arguments
parser.add_argument('--output_dir', type=str, default='./runs/parametric',
help='Output directory')
parser.add_argument('--resume', type=str, default=None,
help='Path to checkpoint to resume from')
args = parser.parse_args()
# Create datasets
print("\nLoading training data...")
train_dataset = ParametricDataset(args.train_dir, normalize=True)
if args.val_dir:
val_dataset = ParametricDataset(args.val_dir, normalize=True)
# Share normalization stats
val_dataset.design_mean = train_dataset.design_mean
val_dataset.design_std = train_dataset.design_std
else:
# Split training data
n_total = len(train_dataset)
n_val = int(n_total * args.val_split)
n_train = n_total - n_val
train_dataset, val_dataset = torch.utils.data.random_split(
train_dataset, [n_train, n_val]
)
print(f"Split: {n_train} train, {n_val} validation")
# Create data loaders
train_loader = DataLoader(
train_dataset,
batch_size=args.batch_size,
shuffle=True,
collate_fn=collate_fn,
num_workers=0
)
val_loader = DataLoader(
val_dataset,
batch_size=args.batch_size,
shuffle=False,
collate_fn=collate_fn,
num_workers=0
)
# Get design dimension from dataset
if hasattr(train_dataset, 'design_var_names'):
design_dim = len(train_dataset.design_var_names)
else:
# For split dataset, access underlying dataset
design_dim = len(train_dataset.dataset.design_var_names)
# Build configuration
config = {
'model': {
'input_channels': 12,
'edge_dim': 5,
'hidden_channels': args.hidden_channels,
'num_layers': args.num_layers,
'design_dim': design_dim,
'dropout': args.dropout
},
'learning_rate': args.learning_rate,
'weight_decay': args.weight_decay,
'batch_size': args.batch_size,
'num_epochs': args.epochs,
'output_dir': args.output_dir,
'early_stopping_patience': 50,
'loss_weights': {
'mass': 1.0,
'frequency': 1.0,
'displacement': 1.0,
'stress': 1.0
}
}
# Create trainer
trainer = ParametricTrainer(config)
# Resume if specified
if args.resume:
checkpoint = torch.load(args.resume, map_location=trainer.device)
trainer.model.load_state_dict(checkpoint['model_state_dict'])
trainer.optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
trainer.start_epoch = checkpoint['epoch'] + 1
trainer.best_val_loss = checkpoint['best_val_loss']
print(f"Resumed from epoch {checkpoint['epoch']}")
# Get base dataset for normalization stats
base_dataset = train_dataset
if hasattr(train_dataset, 'dataset'):
base_dataset = train_dataset.dataset
# Train
trainer.train(train_loader, val_loader, args.epochs, base_dataset)
if __name__ == "__main__":
main()

454
atomizer-field/validate_parsed_data.py
View File

@@ -1,454 +0,0 @@
"""
validate_parsed_data.py
Validates the parsed neural field data for completeness and physics consistency
AtomizerField Data Validator v1.0.0
Ensures parsed data meets quality standards for neural network training.
Usage:
python validate_parsed_data.py <case_directory>
Example:
python validate_parsed_data.py training_case_001
"""
import json
import h5py
import numpy as np
from pathlib import Path
import sys
class NeuralFieldDataValidator:
"""
Validates parsed neural field data for:
- File existence and format
- Data completeness
- Physics consistency
- Data quality
This ensures that data fed to neural networks is reliable and consistent.
"""
def __init__(self, case_directory):
"""
Initialize validator
Args:
case_directory (str or Path): Path to case containing parsed data
"""
self.case_dir = Path(case_directory)
self.json_file = self.case_dir / "neural_field_data.json"
self.h5_file = self.case_dir / "neural_field_data.h5"
self.errors = []
self.warnings = []
self.info = []
def validate(self):
"""
Run all validation checks
Returns:
bool: True if validation passed, False otherwise
"""
print("\n" + "="*60)
print("AtomizerField Data Validator v1.0")
print("="*60)
print(f"\nValidating: {self.case_dir.name}\n")
# Check file existence
if not self._check_files_exist():
return False
# Load data
try:
with open(self.json_file, 'r') as f:
self.data = json.load(f)
self.h5_data = h5py.File(self.h5_file, 'r')
except Exception as e:
self._add_error(f"Failed to load data files: {e}")
return False
# Run validation checks
self._validate_structure()
self._validate_metadata()
self._validate_mesh()
self._validate_materials()
self._validate_boundary_conditions()
self._validate_loads()
self._validate_results()
self._validate_physics_consistency()
self._validate_data_quality()
# Close HDF5 file
self.h5_data.close()
# Print results
self._print_results()
return len(self.errors) == 0
def _check_files_exist(self):
"""Check that required files exist"""
if not self.json_file.exists():
self._add_error(f"JSON file not found: {self.json_file}")
return False
if not self.h5_file.exists():
self._add_error(f"HDF5 file not found: {self.h5_file}")
return False
self._add_info(f"Found JSON: {self.json_file.name}")
self._add_info(f"Found HDF5: {self.h5_file.name}")
return True
def _validate_structure(self):
"""Validate data structure has all required fields"""
required_fields = [
"metadata",
"mesh",
"materials",
"boundary_conditions",
"loads",
"results"
]
for field in required_fields:
if field not in self.data:
self._add_error(f"Missing required field: {field}")
else:
self._add_info(f"Found field: {field}")
def _validate_metadata(self):
"""Validate metadata completeness"""
if "metadata" not in self.data:
return
meta = self.data["metadata"]
# Check version
if "version" in meta:
if meta["version"] != "1.0.0":
self._add_warning(f"Data version {meta['version']} may not be compatible")
else:
self._add_info(f"Data version: {meta['version']}")
# Check required metadata fields
required = ["created_at", "source", "analysis_type", "units"]
for field in required:
if field not in meta:
self._add_warning(f"Missing metadata field: {field}")
if "analysis_type" in meta:
self._add_info(f"Analysis type: {meta['analysis_type']}")
def _validate_mesh(self):
"""Validate mesh data"""
if "mesh" not in self.data:
return
mesh = self.data["mesh"]
# Check statistics
if "statistics" in mesh:
stats = mesh["statistics"]
n_nodes = stats.get("n_nodes", 0)
n_elements = stats.get("n_elements", 0)
self._add_info(f"Mesh: {n_nodes:,} nodes, {n_elements:,} elements")
if n_nodes == 0:
self._add_error("Mesh has no nodes")
if n_elements == 0:
self._add_error("Mesh has no elements")
# Check element types
if "element_types" in stats:
elem_types = stats["element_types"]
total_by_type = sum(elem_types.values())
if total_by_type != n_elements:
self._add_warning(
f"Element type count ({total_by_type}) doesn't match "
f"total elements ({n_elements})"
)
for etype, count in elem_types.items():
if count > 0:
self._add_info(f" {etype}: {count:,} elements")
# Validate HDF5 mesh data
if 'mesh' in self.h5_data:
mesh_grp = self.h5_data['mesh']
if 'node_coordinates' in mesh_grp:
coords = mesh_grp['node_coordinates'][:]
self._add_info(f"Node coordinates: shape {coords.shape}")
# Check for NaN or inf
if np.any(np.isnan(coords)):
self._add_error("Node coordinates contain NaN values")
if np.any(np.isinf(coords)):
self._add_error("Node coordinates contain infinite values")
# Check bounding box reasonableness
bbox_size = np.max(coords, axis=0) - np.min(coords, axis=0)
if np.any(bbox_size == 0):
self._add_warning("Mesh is planar or degenerate in one dimension")
def _validate_materials(self):
"""Validate material data"""
if "materials" not in self.data:
return
materials = self.data["materials"]
if len(materials) == 0:
self._add_warning("No materials defined")
return
self._add_info(f"Materials: {len(materials)} defined")
for mat in materials:
mat_id = mat.get("id", "unknown")
mat_type = mat.get("type", "unknown")
if mat_type == "MAT1":
# Check required properties
E = mat.get("E")
nu = mat.get("nu")
if E is None:
self._add_error(f"Material {mat_id}: Missing Young's modulus (E)")
elif E <= 0:
self._add_error(f"Material {mat_id}: Invalid E = {E} (must be > 0)")
if nu is None:
self._add_error(f"Material {mat_id}: Missing Poisson's ratio (nu)")
elif nu < 0 or nu >= 0.5:
self._add_error(f"Material {mat_id}: Invalid nu = {nu} (must be 0 <= nu < 0.5)")
def _validate_boundary_conditions(self):
"""Validate boundary conditions"""
if "boundary_conditions" not in self.data:
return
bcs = self.data["boundary_conditions"]
spc_count = len(bcs.get("spc", []))
mpc_count = len(bcs.get("mpc", []))
self._add_info(f"Boundary conditions: {spc_count} SPCs, {mpc_count} MPCs")
if spc_count == 0:
self._add_warning("No SPCs defined - model may be unconstrained")
def _validate_loads(self):
"""Validate load data"""
if "loads" not in self.data:
return
loads = self.data["loads"]
force_count = len(loads.get("point_forces", []))
pressure_count = len(loads.get("pressure", []))
gravity_count = len(loads.get("gravity", []))
thermal_count = len(loads.get("thermal", []))
total_loads = force_count + pressure_count + gravity_count + thermal_count
self._add_info(
f"Loads: {force_count} forces, {pressure_count} pressures, "
f"{gravity_count} gravity, {thermal_count} thermal"
)
if total_loads == 0:
self._add_warning("No loads defined")
# Validate force magnitudes
for force in loads.get("point_forces", []):
mag = force.get("magnitude")
if mag == 0:
self._add_warning(f"Force at node {force.get('node')} has zero magnitude")
def _validate_results(self):
"""Validate results data"""
if "results" not in self.data:
self._add_error("No results data found")
return
results = self.data["results"]
# Check displacement
if "displacement" not in results:
self._add_error("No displacement results found")
else:
disp = results["displacement"]
n_nodes = len(disp.get("node_ids", []))
max_disp = disp.get("max_translation")
self._add_info(f"Displacement: {n_nodes:,} nodes")
if max_disp is not None:
self._add_info(f" Max displacement: {max_disp:.6f} mm")
if max_disp == 0:
self._add_warning("Maximum displacement is zero - check loads")
elif max_disp > 1000:
self._add_warning(f"Very large displacement ({max_disp:.2f} mm) - check units or model")
# Check stress
if "stress" not in results or len(results["stress"]) == 0:
self._add_warning("No stress results found")
else:
for stress_type, stress_data in results["stress"].items():
n_elem = len(stress_data.get("element_ids", []))
max_vm = stress_data.get("max_von_mises")
self._add_info(f"Stress ({stress_type}): {n_elem:,} elements")
if max_vm is not None:
self._add_info(f" Max von Mises: {max_vm:.2f} MPa")
if max_vm == 0:
self._add_warning(f"{stress_type}: Zero stress - check loads")
# Validate HDF5 results
if 'results' in self.h5_data:
results_grp = self.h5_data['results']
if 'displacement' in results_grp:
disp_data = results_grp['displacement'][:]
# Check for NaN or inf
if np.any(np.isnan(disp_data)):
self._add_error("Displacement results contain NaN values")
if np.any(np.isinf(disp_data)):
self._add_error("Displacement results contain infinite values")
def _validate_physics_consistency(self):
"""Validate physics consistency of results"""
if "results" not in self.data or "mesh" not in self.data:
return
results = self.data["results"]
mesh = self.data["mesh"]
# Check node count consistency
mesh_nodes = mesh.get("statistics", {}).get("n_nodes", 0)
if "displacement" in results:
disp_nodes = len(results["displacement"].get("node_ids", []))
if disp_nodes != mesh_nodes:
self._add_warning(
f"Displacement nodes ({disp_nodes:,}) != mesh nodes ({mesh_nodes:,})"
)
# Check for rigid body motion (if no constraints)
if "boundary_conditions" in self.data:
spc_count = len(self.data["boundary_conditions"].get("spc", []))
if spc_count == 0 and "displacement" in results:
max_disp = results["displacement"].get("max_translation", 0)
if max_disp > 1e6:
self._add_error("Unconstrained model with very large displacements - likely rigid body motion")
def _validate_data_quality(self):
"""Validate data quality for neural network training"""
# Check HDF5 data types and shapes
if 'results' in self.h5_data:
results_grp = self.h5_data['results']
# Check displacement shape
if 'displacement' in results_grp:
disp = results_grp['displacement'][:]
if len(disp.shape) != 2:
self._add_error(f"Displacement has wrong shape: {disp.shape} (expected 2D)")
elif disp.shape[1] != 6:
self._add_error(f"Displacement has {disp.shape[1]} DOFs (expected 6)")
# Check file sizes
json_size = self.json_file.stat().st_size / 1024 # KB
h5_size = self.h5_file.stat().st_size / 1024 # KB
self._add_info(f"File sizes: JSON={json_size:.1f} KB, HDF5={h5_size:.1f} KB")
if json_size > 10000: # 10 MB
self._add_warning("JSON file is very large - consider moving more data to HDF5")
def _add_error(self, message):
"""Add error message"""
self.errors.append(message)
def _add_warning(self, message):
"""Add warning message"""
self.warnings.append(message)
def _add_info(self, message):
"""Add info message"""
self.info.append(message)
def _print_results(self):
"""Print validation results"""
print("\n" + "="*60)
print("VALIDATION RESULTS")
print("="*60)
# Print info
if self.info:
print("\nInformation:")
for msg in self.info:
print(f" [INFO] {msg}")
# Print warnings
if self.warnings:
print("\nWarnings:")
for msg in self.warnings:
print(f" [WARN] {msg}")
# Print errors
if self.errors:
print("\nErrors:")
for msg in self.errors:
print(f" [X] {msg}")
# Summary
print("\n" + "="*60)
if len(self.errors) == 0:
print("[OK] VALIDATION PASSED")
print("="*60)
print("\nData is ready for neural network training!")
else:
print("[X] VALIDATION FAILED")
print("="*60)
print(f"\nFound {len(self.errors)} error(s), {len(self.warnings)} warning(s)")
print("Please fix errors before using this data for training.")
print()
def main():
"""
Main entry point for validation script
"""
if len(sys.argv) < 2:
print("\nAtomizerField Data Validator v1.0")
print("="*60)
print("\nUsage:")
print(" python validate_parsed_data.py <case_directory>")
print("\nExample:")
print(" python validate_parsed_data.py training_case_001")
print()
sys.exit(1)
case_dir = sys.argv[1]
if not Path(case_dir).exists():
print(f"ERROR: Directory not found: {case_dir}")
sys.exit(1)
validator = NeuralFieldDataValidator(case_dir)
success = validator.validate()
sys.exit(0 if success else 1)
if __name__ == "__main__":
main()

46
atomizer-field/visualization_report.md
View File

@@ -1,46 +0,0 @@
# FEA Visualization Report
**Generated:** 2025-11-24T09:24:10.133023
**Case:** test_case_beam
---
## Model Information
- **Analysis Type:** SOL_Unknown
- **Nodes:** 5,179
- **Elements:** 4,866
- **Materials:** 1
## Mesh Structure
![Mesh Structure](visualization_images/mesh.png)
The model contains 5,179 nodes and 4,866 elements.
## Displacement Results
![Displacement Field](visualization_images/displacement.png)
**Maximum Displacement:** 19.556875 mm
The plots show the original mesh (left) and deformed mesh (right) with displacement magnitude shown in color.
## Stress Results
![Stress Field](visualization_images/stress.png)
The stress distribution is shown with colors representing von Mises stress levels.
## Summary Statistics
| Property | Value |
|----------|-------|
| Nodes | 5,179 |
| Elements | 4,866 |
| Max Displacement | 19.556875 mm |
---
*Report generated by AtomizerField Visualizer*

515
atomizer-field/visualize_results.py
View File

@@ -1,515 +0,0 @@
"""
visualize_results.py
3D Visualization of FEA Results and Neural Predictions
Visualizes:
- Mesh structure
- Displacement fields
- Stress fields (von Mises)
- Comparison: FEA vs Neural predictions
"""
import numpy as np
import matplotlib.pyplot as plt
from matplotlib import cm
from mpl_toolkits.mplot3d import Axes3D
import json
import h5py
from pathlib import Path
import argparse
class FEAVisualizer:
"""Visualize FEA results in 3D"""
def __init__(self, case_dir):
"""
Initialize visualizer
Args:
case_dir: Path to case directory with neural_field_data files
"""
self.case_dir = Path(case_dir)
# Load data
print(f"Loading data from {case_dir}...")
self.load_data()
def load_data(self):
"""Load JSON metadata and HDF5 field data"""
json_file = self.case_dir / "neural_field_data.json"
h5_file = self.case_dir / "neural_field_data.h5"
# Load JSON
with open(json_file, 'r') as f:
self.metadata = json.load(f)
# Get connectivity from JSON
self.connectivity = []
if 'mesh' in self.metadata and 'elements' in self.metadata['mesh']:
elements = self.metadata['mesh']['elements']
# Elements are categorized by type: solid, shell, beam, rigid
for elem_category in ['solid', 'shell', 'beam']:
if elem_category in elements and isinstance(elements[elem_category], list):
for elem_data in elements[elem_category]:
elem_type = elem_data.get('type', '')
if elem_type in ['CQUAD4', 'CTRIA3', 'CTETRA', 'CHEXA']:
# Store connectivity: [elem_id, n1, n2, n3, n4, ...]
nodes = elem_data.get('nodes', [])
self.connectivity.append([elem_data['id']] + nodes)
self.connectivity = np.array(self.connectivity) if self.connectivity else np.array([[]])
# Load HDF5
with h5py.File(h5_file, 'r') as f:
self.node_coords = f['mesh/node_coordinates'][:]
# Get node IDs to create mapping
self.node_ids = f['mesh/node_ids'][:]
# Create mapping from node ID to index
self.node_id_to_idx = {nid: idx for idx, nid in enumerate(self.node_ids)}
# Displacement
if 'results/displacement' in f:
self.displacement = f['results/displacement'][:]
else:
self.displacement = None
# Stress (try different possible locations)
self.stress = None
if 'results/stress/cquad4_stress/data' in f:
self.stress = f['results/stress/cquad4_stress/data'][:]
elif 'results/stress/cquad4_stress' in f and hasattr(f['results/stress/cquad4_stress'], 'shape'):
self.stress = f['results/stress/cquad4_stress'][:]
elif 'results/stress' in f and hasattr(f['results/stress'], 'shape'):
self.stress = f['results/stress'][:]
print(f"Loaded {len(self.node_coords)} nodes, {len(self.connectivity)} elements")
def plot_mesh(self, figsize=(12, 8), save_path=None):
"""
Plot 3D mesh structure
Args:
figsize: Figure size
save_path: Path to save figure (optional)
"""
fig = plt.figure(figsize=figsize)
ax = fig.add_subplot(111, projection='3d')
# Extract coordinates
x = self.node_coords[:, 0]
y = self.node_coords[:, 1]
z = self.node_coords[:, 2]
# Plot nodes
ax.scatter(x, y, z, c='blue', marker='.', s=1, alpha=0.3, label='Nodes')
# Plot a subset of elements (for visibility)
step = max(1, len(self.connectivity) // 1000) # Show max 1000 elements
for i in range(0, len(self.connectivity), step):
elem = self.connectivity[i]
if len(elem) < 5: # Skip invalid elements
continue
# Get node IDs (skip element ID at position 0)
node_ids = elem[1:5] # First 4 nodes for CQUAD4
# Convert node IDs to indices
try:
nodes = [self.node_id_to_idx[nid] for nid in node_ids]
except KeyError:
continue # Skip if node not found
# Get coordinates
elem_coords = self.node_coords[nodes]
# Plot element edges
for j in range(min(4, len(nodes))):
next_j = (j + 1) % len(nodes)
ax.plot([elem_coords[j, 0], elem_coords[next_j, 0]],
[elem_coords[j, 1], elem_coords[next_j, 1]],
[elem_coords[j, 2], elem_coords[next_j, 2]],
'k-', linewidth=0.1, alpha=0.1)
ax.set_xlabel('X (mm)')
ax.set_ylabel('Y (mm)')
ax.set_zlabel('Z (mm)')
ax.set_title(f'Mesh Structure\n{len(self.node_coords)} nodes, {len(self.connectivity)} elements')
# Equal aspect ratio
self._set_equal_aspect(ax)
if save_path:
plt.savefig(save_path, dpi=150, bbox_inches='tight')
print(f"Saved mesh plot to {save_path}")
plt.show()
def plot_displacement(self, scale=1.0, component='magnitude', figsize=(14, 8), save_path=None):
"""
Plot displacement field
Args:
scale: Scale factor for displacement visualization
component: 'magnitude', 'x', 'y', or 'z'
figsize: Figure size
save_path: Path to save figure
"""
if self.displacement is None:
print("No displacement data available")
return
fig = plt.figure(figsize=figsize)
# Original mesh
ax1 = fig.add_subplot(121, projection='3d')
self._plot_mesh_with_field(ax1, self.displacement, component, scale=0)
ax1.set_title('Original Mesh')
# Deformed mesh
ax2 = fig.add_subplot(122, projection='3d')
self._plot_mesh_with_field(ax2, self.displacement, component, scale=scale)
ax2.set_title(f'Deformed Mesh (scale={scale}x)\nDisplacement: {component}')
plt.tight_layout()
if save_path:
plt.savefig(save_path, dpi=150, bbox_inches='tight')
print(f"Saved displacement plot to {save_path}")
plt.show()
def plot_stress(self, component='von_mises', figsize=(12, 8), save_path=None):
"""
Plot stress field
Args:
component: 'von_mises', 'xx', 'yy', 'zz', 'xy', 'yz', 'xz'
figsize: Figure size
save_path: Path to save figure
"""
if self.stress is None:
print("No stress data available")
return
fig = plt.figure(figsize=figsize)
ax = fig.add_subplot(111, projection='3d')
# Get stress component
if component == 'von_mises':
# Von Mises already computed (last column)
stress_values = self.stress[:, -1]
else:
# Map component name to index
comp_map = {'xx': 0, 'yy': 1, 'zz': 2, 'xy': 3, 'yz': 4, 'xz': 5}
idx = comp_map.get(component, 0)
stress_values = self.stress[:, idx]
# Plot elements colored by stress
self._plot_elements_with_stress(ax, stress_values)
ax.set_xlabel('X (mm)')
ax.set_ylabel('Y (mm)')
ax.set_zlabel('Z (mm)')
ax.set_title(f'Stress Field: {component}\nMax: {np.max(stress_values):.2f} MPa')
self._set_equal_aspect(ax)
if save_path:
plt.savefig(save_path, dpi=150, bbox_inches='tight')
print(f"Saved stress plot to {save_path}")
plt.show()
def plot_comparison(self, neural_predictions, figsize=(16, 6), save_path=None):
"""
Plot comparison: FEA vs Neural predictions
Args:
neural_predictions: Dict with 'displacement' and/or 'stress'
figsize: Figure size
save_path: Path to save figure
"""
fig = plt.figure(figsize=figsize)
# Displacement comparison
if self.displacement is not None and 'displacement' in neural_predictions:
ax1 = fig.add_subplot(131, projection='3d')
self._plot_mesh_with_field(ax1, self.displacement, 'magnitude', scale=10)
ax1.set_title('FEA Displacement')
ax2 = fig.add_subplot(132, projection='3d')
neural_disp = neural_predictions['displacement']
self._plot_mesh_with_field(ax2, neural_disp, 'magnitude', scale=10)
ax2.set_title('Neural Prediction')
# Error
ax3 = fig.add_subplot(133, projection='3d')
error = np.linalg.norm(self.displacement[:, :3] - neural_disp[:, :3], axis=1)
self._plot_nodes_with_values(ax3, error)
ax3.set_title('Prediction Error')
plt.tight_layout()
if save_path:
plt.savefig(save_path, dpi=150, bbox_inches='tight')
print(f"Saved comparison plot to {save_path}")
plt.show()
def _plot_mesh_with_field(self, ax, field, component, scale=1.0):
"""Helper: Plot mesh colored by field values"""
# Get field component
if component == 'magnitude':
values = np.linalg.norm(field[:, :3], axis=1)
elif component == 'x':
values = field[:, 0]
elif component == 'y':
values = field[:, 1]
elif component == 'z':
values = field[:, 2]
else:
values = np.linalg.norm(field[:, :3], axis=1)
# Apply deformation
coords = self.node_coords + scale * field[:, :3]
# Plot nodes colored by values
scatter = ax.scatter(coords[:, 0], coords[:, 1], coords[:, 2],
c=values, cmap='jet', s=2)
plt.colorbar(scatter, ax=ax, label=f'{component} (mm)')
ax.set_xlabel('X (mm)')
ax.set_ylabel('Y (mm)')
ax.set_zlabel('Z (mm)')
self._set_equal_aspect(ax)
def _plot_elements_with_stress(self, ax, stress_values):
"""Helper: Plot elements colored by stress"""
# Normalize stress for colormap
vmin, vmax = np.min(stress_values), np.max(stress_values)
norm = plt.Normalize(vmin=vmin, vmax=vmax)
cmap = cm.get_cmap('jet')
# Plot subset of elements
step = max(1, len(self.connectivity) // 500)
for i in range(0, min(len(self.connectivity), len(stress_values)), step):
elem = self.connectivity[i]
if len(elem) < 5:
continue
# Get node IDs and convert to indices
node_ids = elem[1:5]
try:
nodes = [self.node_id_to_idx[nid] for nid in node_ids]
except KeyError:
continue
elem_coords = self.node_coords[nodes]
# Get stress color
color = cmap(norm(stress_values[i]))
# Plot filled quadrilateral
from mpl_toolkits.mplot3d.art3d import Poly3DCollection
verts = [elem_coords]
poly = Poly3DCollection(verts, facecolors=color, edgecolors='k',
linewidths=0.1, alpha=0.8)
ax.add_collection3d(poly)
# Colorbar
sm = cm.ScalarMappable(cmap=cmap, norm=norm)
sm.set_array([])
plt.colorbar(sm, ax=ax, label='Stress (MPa)')
def _plot_nodes_with_values(self, ax, values):
"""Helper: Plot nodes colored by values"""
scatter = ax.scatter(self.node_coords[:, 0],
self.node_coords[:, 1],
self.node_coords[:, 2],
c=values, cmap='hot', s=2)
plt.colorbar(scatter, ax=ax, label='Error (mm)')
ax.set_xlabel('X (mm)')
ax.set_ylabel('Y (mm)')
ax.set_zlabel('Z (mm)')
self._set_equal_aspect(ax)
def _set_equal_aspect(self, ax):
"""Set equal aspect ratio for 3D plot"""
# Get limits
x_limits = [self.node_coords[:, 0].min(), self.node_coords[:, 0].max()]
y_limits = [self.node_coords[:, 1].min(), self.node_coords[:, 1].max()]
z_limits = [self.node_coords[:, 2].min(), self.node_coords[:, 2].max()]
# Find max range
max_range = max(x_limits[1] - x_limits[0],
y_limits[1] - y_limits[0],
z_limits[1] - z_limits[0])
# Set limits
x_middle = np.mean(x_limits)
y_middle = np.mean(y_limits)
z_middle = np.mean(z_limits)
ax.set_xlim(x_middle - max_range/2, x_middle + max_range/2)
ax.set_ylim(y_middle - max_range/2, y_middle + max_range/2)
ax.set_zlim(z_middle - max_range/2, z_middle + max_range/2)
def create_report(self, output_file='visualization_report.md'):
"""
Create markdown report with all visualizations
Args:
output_file: Path to save report
"""
print(f"\nGenerating visualization report...")
# Create images directory
img_dir = Path(output_file).parent / 'visualization_images'
img_dir.mkdir(exist_ok=True)
# Generate plots
print(" Creating mesh plot...")
self.plot_mesh(save_path=img_dir / 'mesh.png')
plt.close('all')
print(" Creating displacement plot...")
self.plot_displacement(scale=10, save_path=img_dir / 'displacement.png')
plt.close('all')
print(" Creating stress plot...")
self.plot_stress(save_path=img_dir / 'stress.png')
plt.close('all')
# Write report
with open(output_file, 'w') as f:
f.write(f"# FEA Visualization Report\n\n")
f.write(f"**Generated:** {self.metadata['metadata']['created_at']}\n\n")
f.write(f"**Case:** {self.metadata['metadata']['case_name']}\n\n")
f.write("---\n\n")
# Model info
f.write("## Model Information\n\n")
f.write(f"- **Analysis Type:** {self.metadata['metadata']['analysis_type']}\n")
f.write(f"- **Nodes:** {self.metadata['mesh']['statistics']['n_nodes']:,}\n")
f.write(f"- **Elements:** {self.metadata['mesh']['statistics']['n_elements']:,}\n")
f.write(f"- **Materials:** {len(self.metadata['materials'])}\n\n")
# Mesh
f.write("## Mesh Structure\n\n")
f.write("![Mesh Structure](visualization_images/mesh.png)\n\n")
f.write(f"The model contains {self.metadata['mesh']['statistics']['n_nodes']:,} nodes ")
f.write(f"and {self.metadata['mesh']['statistics']['n_elements']:,} elements.\n\n")
# Displacement
if self.displacement is not None:
max_disp = self.metadata['results']['displacement']['max_translation']
f.write("## Displacement Results\n\n")
f.write("![Displacement Field](visualization_images/displacement.png)\n\n")
f.write(f"**Maximum Displacement:** {max_disp:.6f} mm\n\n")
f.write("The plots show the original mesh (left) and deformed mesh (right) ")
f.write("with displacement magnitude shown in color.\n\n")
# Stress
if self.stress is not None:
f.write("## Stress Results\n\n")
f.write("![Stress Field](visualization_images/stress.png)\n\n")
# Get max stress from metadata
if 'stress' in self.metadata['results']:
for stress_type, stress_data in self.metadata['results']['stress'].items():
if 'max_von_mises' in stress_data and stress_data['max_von_mises'] is not None:
max_stress = stress_data['max_von_mises']
f.write(f"**Maximum von Mises Stress:** {max_stress:.2f} MPa\n\n")
break
f.write("The stress distribution is shown with colors representing von Mises stress levels.\n\n")
# Statistics
f.write("## Summary Statistics\n\n")
f.write("| Property | Value |\n")
f.write("|----------|-------|\n")
f.write(f"| Nodes | {self.metadata['mesh']['statistics']['n_nodes']:,} |\n")
f.write(f"| Elements | {self.metadata['mesh']['statistics']['n_elements']:,} |\n")
if self.displacement is not None:
max_disp = self.metadata['results']['displacement']['max_translation']
f.write(f"| Max Displacement | {max_disp:.6f} mm |\n")
if self.stress is not None and 'stress' in self.metadata['results']:
for stress_type, stress_data in self.metadata['results']['stress'].items():
if 'max_von_mises' in stress_data and stress_data['max_von_mises'] is not None:
max_stress = stress_data['max_von_mises']
f.write(f"| Max von Mises Stress | {max_stress:.2f} MPa |\n")
break
f.write("\n---\n\n")
f.write("*Report generated by AtomizerField Visualizer*\n")
print(f"\nReport saved to: {output_file}")
def main():
"""Main entry point"""
parser = argparse.ArgumentParser(
description='Visualize FEA results in 3D',
formatter_class=argparse.RawDescriptionHelpFormatter,
epilog="""
Examples:
# Visualize mesh
python visualize_results.py test_case_beam --mesh
# Visualize displacement
python visualize_results.py test_case_beam --displacement
# Visualize stress
python visualize_results.py test_case_beam --stress
# Generate full report
python visualize_results.py test_case_beam --report
# All visualizations
python visualize_results.py test_case_beam --all
"""
)
parser.add_argument('case_dir', help='Path to case directory')
parser.add_argument('--mesh', action='store_true', help='Plot mesh structure')
parser.add_argument('--displacement', action='store_true', help='Plot displacement field')
parser.add_argument('--stress', action='store_true', help='Plot stress field')
parser.add_argument('--report', action='store_true', help='Generate markdown report')
parser.add_argument('--all', action='store_true', help='Show all plots and generate report')
parser.add_argument('--scale', type=float, default=10.0, help='Displacement scale factor (default: 10)')
parser.add_argument('--output', default='visualization_report.md', help='Report output file')
args = parser.parse_args()
# Create visualizer
viz = FEAVisualizer(args.case_dir)
# Determine what to show
show_all = args.all or not (args.mesh or args.displacement or args.stress or args.report)
if args.mesh or show_all:
print("\nShowing mesh structure...")
viz.plot_mesh()
if args.displacement or show_all:
print("\nShowing displacement field...")
viz.plot_displacement(scale=args.scale)
if args.stress or show_all:
print("\nShowing stress field...")
viz.plot_stress()
if args.report or show_all:
print("\nGenerating report...")
viz.create_report(args.output)
if __name__ == "__main__":
main()

144
atomizer_paths.py
View File

@@ -1,144 +0,0 @@
"""
Atomizer Path Configuration
Provides intelligent path resolution for Atomizer core modules and directories.
This module can be imported from anywhere in the project hierarchy.
"""
from pathlib import Path
import sys
def get_atomizer_root() -> Path:
"""
Get the Atomizer project root directory.
This function intelligently locates the root by looking for marker files
that uniquely identify the Atomizer project root.
Returns:
Path: Absolute path to Atomizer root directory
Raises:
RuntimeError: If Atomizer root cannot be found
"""
# Start from this file's location
current = Path(__file__).resolve().parent
# Marker files that uniquely identify Atomizer root
markers = [
'optimization_engine', # Core module directory
'studies', # Studies directory
'README.md' # Project README
]
# Walk up the directory tree looking for all markers
max_depth = 10 # Prevent infinite loop
for _ in range(max_depth):
# Check if all markers exist at this level
if all((current / marker).exists() for marker in markers):
return current
# Move up one directory
parent = current.parent
if parent == current: # Reached filesystem root
break
current = parent
raise RuntimeError(
"Could not locate Atomizer root directory. "
"Make sure you're running from within the Atomizer project."
)
def setup_python_path():
"""
Add Atomizer root to Python path if not already present.
This allows imports like `from optimization_engine.core.runner import ...`
to work from anywhere in the project.
"""
root = get_atomizer_root()
root_str = str(root)
if root_str not in sys.path:
sys.path.insert(0, root_str)
# Core directories (lazy-loaded)
_ROOT = None
def root() -> Path:
"""Get Atomizer root directory."""
global _ROOT
if _ROOT is None:
_ROOT = get_atomizer_root()
return _ROOT
def optimization_engine() -> Path:
"""Get optimization_engine directory."""
return root() / 'optimization_engine'
def studies() -> Path:
"""Get studies directory."""
return root() / 'studies'
def tests() -> Path:
"""Get tests directory."""
return root() / 'tests'
def docs() -> Path:
"""Get docs directory."""
return root() / 'docs'
def plugins() -> Path:
"""Get plugins directory."""
return optimization_engine() / 'plugins'
# Common files
def readme() -> Path:
"""Get README.md path."""
return root() / 'README.md'
def roadmap() -> Path:
"""Get development roadmap path."""
return root() / 'DEVELOPMENT_ROADMAP.md'
# Convenience function for scripts
def ensure_imports():
"""
Ensure Atomizer modules can be imported.
Call this at the start of any script to ensure proper imports:
```python
import atomizer_paths
atomizer_paths.ensure_imports()
# Now you can import Atomizer modules
from optimization_engine.core.runner import OptimizationRunner
```
"""
setup_python_path()
if __name__ == '__main__':
# Self-test
print("Atomizer Path Configuration")
print("=" * 60)
print(f"Root: {root()}")
print(f"Optimization Engine: {optimization_engine()}")
print(f"Studies: {studies()}")
print(f"Tests: {tests()}")
print(f"Docs: {docs()}")
print(f"Plugins: {plugins()}")
print("=" * 60)
print("\nAll paths resolved successfully!")

192
config.py
View File

@@ -1,192 +0,0 @@
"""
Central Configuration for Atomizer
This file contains all system-level paths and settings.
To update NX version or paths, ONLY modify this file.
"""
from pathlib import Path
import os
# ============================================================================
# NX/SIMCENTER CONFIGURATION
# ============================================================================
# NX Installation Directory
# Change this to update NX version across entire Atomizer codebase
NX_VERSION = "2512"
NX_INSTALLATION_DIR = Path("C:/Program Files/Siemens/DesigncenterNX2512")
# Derived NX Paths (automatically updated when NX_VERSION changes)
NX_BIN_DIR = NX_INSTALLATION_DIR / "NXBIN"
NX_RUN_JOURNAL = NX_BIN_DIR / "run_journal.exe"
NX_MATERIALS_DIR = NX_INSTALLATION_DIR / "UGII" / "materials"
NX_MATERIAL_LIBRARY = NX_MATERIALS_DIR / "physicalmateriallibrary.xml"
NX_PYTHON_STUBS = NX_INSTALLATION_DIR / "ugopen" / "pythonStubs"
# Validate NX installation on import
if not NX_INSTALLATION_DIR.exists():
raise FileNotFoundError(
f"NX installation not found at: {NX_INSTALLATION_DIR}\n"
f"Please update NX_VERSION or NX_INSTALLATION_DIR in config.py"
)
if not NX_RUN_JOURNAL.exists():
raise FileNotFoundError(
f"run_journal.exe not found at: {NX_RUN_JOURNAL}\n"
f"Please verify NX installation or update paths in config.py"
)
# ============================================================================
# PYTHON ENVIRONMENT CONFIGURATION
# ============================================================================
# Python Environment Name
PYTHON_ENV_NAME = "atomizer"
# Anaconda/Conda base path (auto-detected from current interpreter)
CONDA_BASE = Path(os.environ.get("CONDA_PREFIX", "")).parent
PYTHON_ENV_PATH = CONDA_BASE / "envs" / PYTHON_ENV_NAME / "python.exe"
# ============================================================================
# PROJECT PATHS
# ============================================================================
# Project root directory
PROJECT_ROOT = Path(__file__).parent.resolve()
# Key directories
OPTIMIZATION_ENGINE_DIR = PROJECT_ROOT / "optimization_engine"
STUDIES_DIR = PROJECT_ROOT / "studies"
DOCS_DIR = PROJECT_ROOT / "docs"
TESTS_DIR = PROJECT_ROOT / "tests"
# ============================================================================
# NASTRAN CONFIGURATION
# ============================================================================
# Nastran solver timeout (seconds)
NASTRAN_TIMEOUT = 600
# Nastran file extensions
NASTRAN_INPUT_EXT = ".dat"
NASTRAN_OUTPUT_EXT = ".op2"
NASTRAN_TEXT_OUTPUT_EXT = ".f06"
# ============================================================================
# OPTIMIZATION DEFAULTS
# ============================================================================
# Default optimization parameters
DEFAULT_N_TRIALS = 50
DEFAULT_TIMEOUT = 3600 # 1 hour
DEFAULT_N_STARTUP_TRIALS = 10
# Optuna sampler settings
DEFAULT_SAMPLER = "TPESampler"
DEFAULT_SURROGATE = "gp" # gaussian process
# ============================================================================
# LOGGING CONFIGURATION
# ============================================================================
LOG_LEVEL = "INFO"
LOG_FORMAT = "%(asctime)s - %(name)s - %(levelname)s - %(message)s"
# ============================================================================
# DASHBOARD CONFIGURATION
# ============================================================================
DASHBOARD_HOST = "localhost"
DASHBOARD_PORT = 8050
# ============================================================================
# HELPER FUNCTIONS
# ============================================================================
def get_nx_journal_command(journal_script: Path, *args) -> str:
"""
Generate NX journal execution command.
Args:
journal_script: Path to journal .py file
*args: Additional arguments to pass to journal
Returns:
Command string ready for subprocess execution
"""
cmd_parts = [f'"{NX_RUN_JOURNAL}"', f'"{journal_script}"']
if args:
cmd_parts.append("-args")
cmd_parts.extend([f'"{arg}"' for arg in args])
return " ".join(cmd_parts)
def validate_config():
"""
Validate all critical paths exist.
Raises FileNotFoundError if any critical path is missing.
"""
critical_paths = {
"NX Installation": NX_INSTALLATION_DIR,
"NX run_journal.exe": NX_RUN_JOURNAL,
"NX Material Library": NX_MATERIAL_LIBRARY,
"NX Python Stubs": NX_PYTHON_STUBS,
"Project Root": PROJECT_ROOT,
"Optimization Engine": OPTIMIZATION_ENGINE_DIR,
}
missing = []
for name, path in critical_paths.items():
if not path.exists():
missing.append(f" - {name}: {path}")
if missing:
raise FileNotFoundError(
"Critical paths missing:\n" + "\n".join(missing) +
"\n\nPlease update paths in config.py"
)
return True
def print_config():
"""Print current configuration for debugging."""
print("=" * 80)
print("ATOMIZER CONFIGURATION")
print("=" * 80)
print(f"\nNX Configuration:")
print(f" Version: {NX_VERSION}")
print(f" Installation: {NX_INSTALLATION_DIR}")
print(f" run_journal.exe: {NX_RUN_JOURNAL}")
print(f" Material Library: {NX_MATERIAL_LIBRARY}")
print(f" Python Stubs: {NX_PYTHON_STUBS}")
print(f"\nPython Environment:")
print(f" Environment Name: {PYTHON_ENV_NAME}")
print(f" Python Path: {PYTHON_ENV_PATH}")
print(f"\nProject Paths:")
print(f" Root: {PROJECT_ROOT}")
print(f" Optimization Engine: {OPTIMIZATION_ENGINE_DIR}")
print(f" Studies: {STUDIES_DIR}")
print("=" * 80)
# Validate configuration on import
validate_config()
# ============================================================================
# USAGE EXAMPLE
# ============================================================================
if __name__ == "__main__":
# Print configuration
print_config()
# Example: Generate journal command
example_journal = PROJECT_ROOT / "optimization_engine" / "import_expressions.py"
example_prt = PROJECT_ROOT / "studies" / "beam" / "Beam.prt"
example_exp = PROJECT_ROOT / "studies" / "beam" / "Beam_vars.exp"
cmd = get_nx_journal_command(example_journal, example_prt, example_exp)
print(f"\nExample Journal Command:\n{cmd}")

41
docs/reference/ATOMIZER_FIELD_CONCEPT.md Normal file
View File

@@ -0,0 +1,41 @@
# AtomizerField - Neural Field Predictor (Concept Archive)
> **Status**: Concept archived. Code removed during repo cleanup (Feb 2026).
> Original location: `atomizer-field/`
## Idea
Instead of extracting just scalar values (max stress, mass) from FEA results, capture **complete field data** - stress, displacement, and strain at every node and element. Train a GNN to predict full fields from design parameters, enabling 1000x faster optimization with true physics understanding.
## Architecture (Two-Phase)
```
Phase 1: Data Pipeline
NX Nastran (.bdf, .op2) → neural_field_parser.py → Neural Field Format (JSON + HDF5)
Phase 2: Neural Network
Graph representation → GNN training (train.py + field_predictor.py) → Field predictions (5-50ms)
```
## Key Components
- **neural_field_parser.py** - Parsed BDF/OP2 into complete field data (displacement, stress, strain at ALL nodes/elements)
- **train.py** - GNN training pipeline using PyTorch Geometric
- **predict.py** - Inference for field predictions
- **optimization_interface.py** - Integration with Atomizer optimization loop
## Data Format
- HDF5 for large arrays (field values per node)
- JSON for metadata (mesh topology, material properties, BCs)
- Versioned format (v1.0) designed for neural network training
## Relationship to Current GNN Module
The `optimization_engine/gnn/` module (ZernikeGNN) is the evolved version of this concept, specialized for mirror Zernike coefficient prediction rather than full-field prediction. If full-field prediction is needed in the future, this concept provides the architecture blueprint.
## Training Data Requirements
- SOL 101 (Linear Static) results with DISPLACEMENT=ALL, STRESS=ALL, STRAIN=ALL
- Organized as training cases with input (.bdf) and output (.op2) pairs
- The `atomizer_field_training_data/` directory contained ~66MB of sample training data (also removed)

320
generate_training_data.py
View File

@@ -1,320 +0,0 @@
"""
Space-Filling Training Data Generator
This script generates FEA training points that cover the ENTIRE design space
uniformly, unlike optimization which focuses only on promising regions.
Sampling Methods:
1. Latin Hypercube Sampling (LHS) - Good coverage, no clustering
2. Sobol Sequence - Quasi-random, very uniform
3. Grid Sampling - Regular grid, exhaustive but slow
Usage:
python generate_training_data.py --study uav_arm_optimization --method lhs --points 100
"""
import sys
from pathlib import Path
import json
import argparse
import numpy as np
from scipy.stats import qmc # For Latin Hypercube and Sobol
project_root = Path(__file__).parent
sys.path.insert(0, str(project_root))
def load_config_bounds(study_path: Path) -> dict:
"""Load design variable bounds from optimization_config.json
Supports two config formats:
1. {"parameter": "name", "bounds": [min, max]} - Current format
2. {"name": "name", "min_value": min, "max_value": max} - Legacy format
"""
config_path = study_path / "1_setup" / "optimization_config.json"
if not config_path.exists():
raise FileNotFoundError(f"Config not found: {config_path}")
with open(config_path) as f:
config = json.load(f)
bounds = {}
for var in config.get('design_variables', []):
# Support both 'parameter' and 'name' keys
name = var.get('parameter') or var.get('name')
# Support both "bounds": [min, max] and "min_value"/"max_value" formats
if 'bounds' in var:
min_val, max_val = var['bounds']
else:
min_val = var.get('min_value', var.get('min', 0))
max_val = var.get('max_value', var.get('max', 1))
# Detect integer type based on name or explicit type
is_int = (var.get('type') == 'integer' or
'count' in name.lower() or
(isinstance(min_val, int) and isinstance(max_val, int) and max_val - min_val < 20))
bounds[name] = {
'min': min_val,
'max': max_val,
'type': 'int' if is_int else 'float'
}
return bounds, config
def generate_lhs_samples(bounds: dict, n_samples: int, seed: int = 42) -> list:
"""
Generate Latin Hypercube Samples across the full design space.
LHS ensures:
- Each dimension is divided into n equal intervals
- Exactly one sample in each interval per dimension
- Much better coverage than random sampling
"""
var_names = list(bounds.keys())
n_dims = len(var_names)
# Create LHS sampler
sampler = qmc.LatinHypercube(d=n_dims, seed=seed)
samples_unit = sampler.random(n=n_samples) # Values in [0, 1]
# Scale to actual bounds
samples = []
for i in range(n_samples):
point = {}
for j, name in enumerate(var_names):
b = bounds[name]
value = b['min'] + samples_unit[i, j] * (b['max'] - b['min'])
if b['type'] == 'int':
value = int(round(value))
point[name] = value
samples.append(point)
return samples
def generate_sobol_samples(bounds: dict, n_samples: int, seed: int = 42) -> list:
"""
Generate Sobol sequence samples (quasi-random, very uniform).
Sobol sequences are deterministic and provide excellent uniformity.
"""
var_names = list(bounds.keys())
n_dims = len(var_names)
sampler = qmc.Sobol(d=n_dims, scramble=True, seed=seed)
samples_unit = sampler.random(n=n_samples)
samples = []
for i in range(n_samples):
point = {}
for j, name in enumerate(var_names):
b = bounds[name]
value = b['min'] + samples_unit[i, j] * (b['max'] - b['min'])
if b['type'] == 'int':
value = int(round(value))
point[name] = value
samples.append(point)
return samples
def generate_grid_samples(bounds: dict, points_per_dim: int = 5) -> list:
"""
Generate regular grid samples.
Warning: Scales exponentially with dimensions!
4 dims x 5 points = 625 samples
4 dims x 10 points = 10,000 samples
"""
var_names = list(bounds.keys())
# Create linspace for each dimension
grids = []
for name in var_names:
b = bounds[name]
if b['type'] == 'int':
# For integers, use actual integer values
values = np.linspace(b['min'], b['max'], points_per_dim)
values = np.unique(np.round(values).astype(int))
else:
values = np.linspace(b['min'], b['max'], points_per_dim)
grids.append(values)
# Create meshgrid and flatten
mesh = np.meshgrid(*grids, indexing='ij')
flat = [m.flatten() for m in mesh]
samples = []
for i in range(len(flat[0])):
point = {}
for j, name in enumerate(var_names):
value = flat[j][i]
if bounds[name]['type'] == 'int':
value = int(value)
else:
value = float(value)
point[name] = value
samples.append(point)
return samples
def generate_corner_samples(bounds: dict) -> list:
"""
Generate samples at all corners of the design space.
This ensures the NN sees the extreme combinations.
For 4 dimensions: 2^4 = 16 corner points
"""
var_names = list(bounds.keys())
n_dims = len(var_names)
samples = []
for i in range(2**n_dims):
point = {}
for j, name in enumerate(var_names):
b = bounds[name]
# Use bit j of i to decide min or max
value = b['max'] if (i >> j) & 1 else b['min']
if b['type'] == 'int':
value = int(value)
point[name] = value
samples.append(point)
return samples
def save_training_points(samples: list, output_path: Path):
"""Save training points to JSON file."""
with open(output_path, 'w') as f:
json.dump({
'n_samples': len(samples),
'samples': samples
}, f, indent=2)
print(f"Saved {len(samples)} training points to: {output_path}")
def visualize_coverage(samples: list, bounds: dict, save_path: Path):
"""Visualize how well samples cover the design space."""
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
var_names = list(bounds.keys())
n_vars = len(var_names)
# Create pairwise scatter plots
fig, axes = plt.subplots(n_vars-1, n_vars-1, figsize=(12, 12))
for i in range(n_vars - 1):
for j in range(i + 1, n_vars):
ax = axes[j-1, i] if n_vars > 2 else axes
x = [s[var_names[i]] for s in samples]
y = [s[var_names[j]] for s in samples]
ax.scatter(x, y, alpha=0.5, s=20)
ax.set_xlabel(var_names[i].replace('_', '\n'), fontsize=8)
ax.set_ylabel(var_names[j].replace('_', '\n'), fontsize=8)
# Show bounds
b_i = bounds[var_names[i]]
b_j = bounds[var_names[j]]
ax.set_xlim(b_i['min'], b_i['max'])
ax.set_ylim(b_j['min'], b_j['max'])
ax.grid(True, alpha=0.3)
# Hide unused subplots
for i in range(n_vars - 1):
for j in range(i):
if n_vars > 2:
axes[i, j].set_visible(False)
plt.suptitle(f'Design Space Coverage ({len(samples)} samples)', fontsize=14)
plt.tight_layout()
plt.savefig(save_path, dpi=150)
plt.close()
print(f"Saved coverage plot: {save_path}")
def main():
parser = argparse.ArgumentParser(description='Generate space-filling training data')
parser.add_argument('--study', required=True, help='Study name (e.g., uav_arm_optimization)')
parser.add_argument('--method', default='lhs', choices=['lhs', 'sobol', 'grid', 'corners', 'combined'],
help='Sampling method')
parser.add_argument('--points', type=int, default=100, help='Number of samples (for lhs/sobol)')
parser.add_argument('--grid-points', type=int, default=5, help='Points per dimension (for grid)')
parser.add_argument('--seed', type=int, default=42, help='Random seed')
parser.add_argument('--visualize', action='store_true', help='Generate coverage plot')
args = parser.parse_args()
study_path = project_root / "studies" / args.study
if not study_path.exists():
print(f"ERROR: Study not found: {study_path}")
return
print("="*70)
print("Space-Filling Training Data Generator")
print("="*70)
# Load bounds from config
print(f"\nLoading config from: {study_path}")
bounds, config = load_config_bounds(study_path)
print(f"\nDesign Variable Bounds:")
for name, b in bounds.items():
print(f" {name}: [{b['min']}, {b['max']}] ({b['type']})")
# Generate samples
print(f"\nGenerating samples using method: {args.method}")
if args.method == 'lhs':
samples = generate_lhs_samples(bounds, args.points, args.seed)
elif args.method == 'sobol':
samples = generate_sobol_samples(bounds, args.points, args.seed)
elif args.method == 'grid':
samples = generate_grid_samples(bounds, args.grid_points)
elif args.method == 'corners':
samples = generate_corner_samples(bounds)
elif args.method == 'combined':
# Combine corners + LHS for best coverage
corner_samples = generate_corner_samples(bounds)
lhs_samples = generate_lhs_samples(bounds, args.points - len(corner_samples), args.seed)
samples = corner_samples + lhs_samples
print(f" Combined: {len(corner_samples)} corners + {len(lhs_samples)} LHS")
print(f" Generated {len(samples)} samples")
# Show sample range coverage
print(f"\nSample Coverage:")
for name in bounds.keys():
values = [s[name] for s in samples]
print(f" {name}: [{min(values):.2f}, {max(values):.2f}]")
# Save samples
output_path = study_path / "1_setup" / "training_points.json"
save_training_points(samples, output_path)
# Visualize if requested
if args.visualize:
plot_path = study_path / "1_setup" / "training_coverage.png"
visualize_coverage(samples, bounds, plot_path)
print(f"\n" + "="*70)
print("NEXT STEPS")
print("="*70)
print(f"1. Run FEA on all {len(samples)} training points:")
print(f" python run_training_fea.py --study {args.study}")
print(f"2. This will create comprehensive training data")
print(f"3. Then retrain NN on this uniform data")
if __name__ == '__main__':
main()

73
generated_extractors/cbar_force_extractor.py
View File

@@ -1,73 +0,0 @@
"""
Extract element forces from CBAR in Z direction from OP2
Auto-generated by Atomizer Phase 3 - pyNastran Research Agent
Pattern: cbar_force
Element Type: CBAR
Result Type: force
API: model.cbar_force[subcase]
"""
from pathlib import Path
from typing import Dict, Any
import numpy as np
from pyNastran.op2.op2 import OP2
def extract_cbar_force(op2_file: Path, subcase: int = 1, direction: str = 'Z'):
"""
Extract forces from CBAR elements.
Args:
op2_file: Path to OP2 file
subcase: Subcase ID
direction: Force direction ('X', 'Y', 'Z', 'axial', 'torque')
Returns:
Dict with force statistics
"""
from pyNastran.op2.op2 import OP2
import numpy as np
model = OP2()
model.read_op2(str(op2_file))
if not hasattr(model, 'cbar_force'):
raise ValueError("No CBAR force results in OP2")
force = model.cbar_force[subcase]
itime = 0
# CBAR force data structure:
# [bending_moment_a1, bending_moment_a2,
# bending_moment_b1, bending_moment_b2,
# shear1, shear2, axial, torque]
direction_map = {
'shear1': 4,
'shear2': 5,
'axial': 6,
'Z': 6, # Commonly axial is Z direction
'torque': 7
}
col_idx = direction_map.get(direction, direction_map.get(direction.lower(), 6))
forces = force.data[itime, :, col_idx]
return {
f'max_{direction}_force': float(np.max(np.abs(forces))),
f'avg_{direction}_force': float(np.mean(np.abs(forces))),
f'min_{direction}_force': float(np.min(np.abs(forces))),
'forces_array': forces.tolist()
}
if __name__ == '__main__':
# Example usage
import sys
if len(sys.argv) > 1:
op2_file = Path(sys.argv[1])
result = extract_cbar_force(op2_file)
print(f"Extraction result: {result}")
else:
print("Usage: python {sys.argv[0]} <op2_file>")

56
generated_extractors/test_displacement_extractor.py
View File

@@ -1,56 +0,0 @@
"""
Extract displacement from bracket OP2
Auto-generated by Atomizer Phase 3 - pyNastran Research Agent
Pattern: displacement
Element Type: General
Result Type: displacement
API: model.displacements[subcase]
"""
from pathlib import Path
from typing import Dict, Any
import numpy as np
from pyNastran.op2.op2 import OP2
def extract_displacement(op2_file: Path, subcase: int = 1):
"""Extract displacement results from OP2 file."""
from pyNastran.op2.op2 import OP2
import numpy as np
model = OP2()
model.read_op2(str(op2_file))
disp = model.displacements[subcase]
itime = 0 # static case
# Extract translation components
txyz = disp.data[itime, :, :3] # [tx, ty, tz]
# Calculate total displacement
total_disp = np.linalg.norm(txyz, axis=1)
max_disp = np.max(total_disp)
# Get node info
node_ids = [nid for (nid, grid_type) in disp.node_gridtype]
max_disp_node = node_ids[np.argmax(total_disp)]
return {
'max_displacement': float(max_disp),
'max_disp_node': int(max_disp_node),
'max_disp_x': float(np.max(np.abs(txyz[:, 0]))),
'max_disp_y': float(np.max(np.abs(txyz[:, 1]))),
'max_disp_z': float(np.max(np.abs(txyz[:, 2])))
}
if __name__ == '__main__':
# Example usage
import sys
if len(sys.argv) > 1:
op2_file = Path(sys.argv[1])
result = extract_displacement(op2_file)
print(f"Extraction result: {result}")
else:
print("Usage: python {sys.argv[0]} <op2_file>")

75
generated_hooks/hook_compare_min_to_avg_ratio.py
View File

@@ -1,75 +0,0 @@
"""
Comparison Hook
Auto-generated by Atomizer Phase 2.9
Compare min force to average
Operation: ratio
Formula: min_to_avg_ratio = min_force / avg_force
"""
import sys
import json
from pathlib import Path
def compare_ratio(min_force, avg_force):
"""
Compare values using ratio.
Args:
min_force: float
avg_force: float
Returns:
float: Comparison result
"""
result = min_force / avg_force
return result
def main():
"""Main entry point for hook execution."""
if len(sys.argv) < 2:
print("Usage: python {} <input_file.json>".format(sys.argv[0]))
sys.exit(1)
input_file = Path(sys.argv[1])
# Read inputs
with open(input_file, 'r') as f:
inputs = json.load(f)
# Extract required inputs
min_force = inputs.get("min_force")
if min_force is None:
print(f"Error: Required input 'min_force' not found")
sys.exit(1)
avg_force = inputs.get("avg_force")
if avg_force is None:
print(f"Error: Required input 'avg_force' not found")
sys.exit(1)
# Calculate comparison
result = compare_ratio(min_force, avg_force)
# Write output
output_file = input_file.parent / "min_to_avg_ratio.json"
output = {
"min_to_avg_ratio": result,
"operation": "ratio",
"formula": "min_force / avg_force",
"inputs_used": {"min_force": min_force, "avg_force": avg_force}
}
with open(output_file, 'w') as f:
json.dump(output, f, indent=2)
print(f"min_to_avg_ratio = {result:.6f}")
print(f"Result saved to: {output_file}")
return result
if __name__ == '__main__':
main()

83
generated_hooks/hook_constraint_yield_constraint.py
View File

@@ -1,83 +0,0 @@
"""
Constraint Check Hook
Auto-generated by Atomizer Phase 2.9
Check if stress is below yield
Constraint: max_stress / yield_strength
Threshold: 1.0
"""
import sys
import json
from pathlib import Path
def check_yield_constraint(max_stress, yield_strength):
"""
Check constraint condition.
Args:
max_stress: float
yield_strength: float
Returns:
tuple: (satisfied: bool, value: float, violation: float)
"""
value = max_stress / yield_strength
satisfied = value <= 1.0
violation = max(0.0, value - 1.0)
return satisfied, value, violation
def main():
"""Main entry point for hook execution."""
if len(sys.argv) < 2:
print("Usage: python {} <input_file.json>".format(sys.argv[0]))
sys.exit(1)
input_file = Path(sys.argv[1])
# Read inputs
with open(input_file, 'r') as f:
inputs = json.load(f)
# Extract required inputs
max_stress = inputs.get("max_stress")
if max_stress is None:
print(f"Error: Required input 'max_stress' not found")
sys.exit(1)
yield_strength = inputs.get("yield_strength")
if yield_strength is None:
print(f"Error: Required input 'yield_strength' not found")
sys.exit(1)
# Check constraint
satisfied, value, violation = check_yield_constraint(max_stress, yield_strength)
# Write output
output_file = input_file.parent / "yield_constraint_check.json"
output = {
"constraint_name": "yield_constraint",
"satisfied": satisfied,
"value": value,
"threshold": 1.0,
"violation": violation,
"inputs_used": {"max_stress": max_stress, "yield_strength": yield_strength}
}
with open(output_file, 'w') as f:
json.dump(output, f, indent=2)
status = "SATISFIED" if satisfied else "VIOLATED"
print(f"Constraint {status}: {value:.6f} (threshold: 1.0)")
if not satisfied:
print(f"Violation: {violation:.6f}")
print(f"Result saved to: {output_file}")
return value
if __name__ == '__main__':
main()

74
generated_hooks/hook_custom_safety_factor.py
View File

@@ -1,74 +0,0 @@
"""
Custom Formula Hook
Auto-generated by Atomizer Phase 2.9
Calculate safety factor
Formula: safety_factor = yield_strength / max_stress
Inputs: max_stress, yield_strength
"""
import sys
import json
from pathlib import Path
def calculate_safety_factor(max_stress, yield_strength):
"""
Calculate custom metric using formula.
Args:
max_stress: float
yield_strength: float
Returns:
float: safety_factor
"""
safety_factor = yield_strength / max_stress
return safety_factor
def main():
"""Main entry point for hook execution."""
if len(sys.argv) < 2:
print("Usage: python {} <input_file.json>".format(sys.argv[0]))
sys.exit(1)
input_file = Path(sys.argv[1])
# Read inputs
with open(input_file, 'r') as f:
inputs = json.load(f)
# Extract required inputs
max_stress = inputs.get("max_stress")
if max_stress is None:
print(f"Error: Required input 'max_stress' not found")
sys.exit(1)
yield_strength = inputs.get("yield_strength")
if yield_strength is None:
print(f"Error: Required input 'yield_strength' not found")
sys.exit(1)
# Calculate result
result = calculate_safety_factor(max_stress, yield_strength)
# Write output
output_file = input_file.parent / "safety_factor_result.json"
output = {
"safety_factor": result,
"formula": "yield_strength / max_stress",
"inputs_used": {"max_stress": max_stress, "yield_strength": yield_strength}
}
with open(output_file, 'w') as f:
json.dump(output, f, indent=2)
print(f"safety_factor = {result:.6f}")
print(f"Result saved to: {output_file}")
return result
if __name__ == '__main__':
main()

54
generated_hooks/hook_registry.json
View File

@@ -1,54 +0,0 @@
{
"hooks": [
{
"name": "hook_weighted_objective_norm_stress_norm_disp.py",
"type": "weighted_objective",
"description": "Combine normalized stress (70%) and displacement (30%)",
"inputs": [
"norm_stress",
"norm_disp"
],
"outputs": [
"weighted_objective"
]
},
{
"name": "hook_custom_safety_factor.py",
"type": "custom_formula",
"description": "Calculate safety factor",
"inputs": [
"max_stress",
"yield_strength"
],
"outputs": [
"safety_factor"
]
},
{
"name": "hook_compare_min_to_avg_ratio.py",
"type": "comparison",
"description": "Compare min force to average",
"inputs": [
"min_force",
"avg_force"
],
"outputs": [
"min_to_avg_ratio"
]
},
{
"name": "hook_constraint_yield_constraint.py",
"type": "constraint_check",
"description": "Check if stress is below yield",
"inputs": [
"max_stress",
"yield_strength"
],
"outputs": [
"yield_constraint",
"yield_constraint_satisfied",
"yield_constraint_violation"
]
}
]
}

85
generated_hooks/hook_weighted_objective_norm_stress_norm_disp.py
View File

@@ -1,85 +0,0 @@
"""
Weighted Objective Function Hook
Auto-generated by Atomizer Phase 2.9
Combine normalized stress (70%) and displacement (30%)
Inputs: norm_stress, norm_disp
Weights: 0.7, 0.3
Formula: 0.7 * norm_stress + 0.3 * norm_disp
Objective: minimize
"""
import sys
import json
from pathlib import Path
def weighted_objective(norm_stress, norm_disp):
"""
Calculate weighted objective from multiple inputs.
Args:
norm_stress: float
norm_disp: float
Returns:
float: Weighted objective value
"""
result = 0.7 * norm_stress + 0.3 * norm_disp
return result
def main():
"""
Main entry point for hook execution.
Reads inputs from JSON file, calculates objective, writes output.
"""
# Parse command line arguments
if len(sys.argv) < 2:
print("Usage: python {} <input_file.json>".format(sys.argv[0]))
sys.exit(1)
input_file = Path(sys.argv[1])
# Read inputs
if not input_file.exists():
print(f"Error: Input file {input_file} not found")
sys.exit(1)
with open(input_file, 'r') as f:
inputs = json.load(f)
# Extract required inputs
norm_stress = inputs.get("norm_stress")
if norm_stress is None:
print(f"Error: Required input 'norm_stress' not found")
sys.exit(1)
norm_disp = inputs.get("norm_disp")
if norm_disp is None:
print(f"Error: Required input 'norm_disp' not found")
sys.exit(1)
# Calculate weighted objective
result = weighted_objective(norm_stress, norm_disp)
# Write output
output_file = input_file.parent / "weighted_objective_result.json"
output = {
"weighted_objective": result,
"objective_type": "minimize",
"inputs_used": {"norm_stress": norm_stress, "norm_disp": norm_disp},
"formula": "0.7 * norm_stress + 0.3 * norm_disp"
}
with open(output_file, 'w') as f:
json.dump(output, f, indent=2)
print(f"Weighted objective calculated: {result:.6f}")
print(f"Result saved to: {output_file}")
return result
if __name__ == '__main__':
main()

4
generated_hooks/test_input.json
View File

@@ -1,4 +0,0 @@
{
"norm_stress": 0.75,
"norm_disp": 0.64
}

9
generated_hooks/weighted_objective_result.json
View File

@@ -1,9 +0,0 @@
{
"weighted_objective": 0.7169999999999999,
"objective_type": "minimize",
"inputs_used": {
"norm_stress": 0.75,
"norm_disp": 0.64
},
"formula": "0.7 * norm_stress + 0.3 * norm_disp"
}

208
migrate_imports.py
View File

@@ -1,208 +0,0 @@
#!/usr/bin/env python3
"""
optimization_engine Migration Script
=====================================
Automatically updates all imports across the codebase.
Usage:
python migrate_imports.py --dry-run # Preview changes
python migrate_imports.py --execute # Apply changes
"""
import os
import re
import sys
from pathlib import Path
from typing import Dict, List, Tuple
# Import mappings (old -> new) - using regex patterns
IMPORT_MAPPINGS = {
# =============================================================================
# CORE MODULE
# =============================================================================
r'from optimization_engine\.runner\b': 'from optimization_engine.core.runner',
r'from optimization_engine\.base_runner\b': 'from optimization_engine.core.base_runner',
r'from optimization_engine\.runner_with_neural\b': 'from optimization_engine.core.runner_with_neural',
r'from optimization_engine\.intelligent_optimizer\b': 'from optimization_engine.core.intelligent_optimizer',
r'from optimization_engine\.method_selector\b': 'from optimization_engine.core.method_selector',
r'from optimization_engine\.strategy_selector\b': 'from optimization_engine.core.strategy_selector',
r'from optimization_engine\.strategy_portfolio\b': 'from optimization_engine.core.strategy_portfolio',
r'from optimization_engine\.gradient_optimizer\b': 'from optimization_engine.core.gradient_optimizer',
r'import optimization_engine\.runner\b': 'import optimization_engine.core.runner',
r'import optimization_engine\.intelligent_optimizer\b': 'import optimization_engine.core.intelligent_optimizer',
# =============================================================================
# SURROGATES MODULE
# =============================================================================
r'from optimization_engine\.neural_surrogate\b': 'from optimization_engine.processors.surrogates.neural_surrogate',
r'from optimization_engine\.generic_surrogate\b': 'from optimization_engine.processors.surrogates.generic_surrogate',
r'from optimization_engine\.adaptive_surrogate\b': 'from optimization_engine.processors.surrogates.adaptive_surrogate',
r'from optimization_engine\.simple_mlp_surrogate\b': 'from optimization_engine.processors.surrogates.simple_mlp_surrogate',
r'from optimization_engine\.active_learning_surrogate\b': 'from optimization_engine.processors.surrogates.active_learning_surrogate',
r'from optimization_engine\.surrogate_tuner\b': 'from optimization_engine.processors.surrogates.surrogate_tuner',
r'from optimization_engine\.auto_trainer\b': 'from optimization_engine.processors.surrogates.auto_trainer',
r'from optimization_engine\.training_data_exporter\b': 'from optimization_engine.processors.surrogates.training_data_exporter',
# =============================================================================
# NX MODULE
# =============================================================================
r'from optimization_engine\.nx_solver\b': 'from optimization_engine.nx.solver',
r'from optimization_engine\.nx_updater\b': 'from optimization_engine.nx.updater',
r'from optimization_engine\.nx_session_manager\b': 'from optimization_engine.nx.session_manager',
r'from optimization_engine\.solve_simulation\b': 'from optimization_engine.nx.solve_simulation',
r'from optimization_engine\.solve_simulation_simple\b': 'from optimization_engine.nx.solve_simulation_simple',
r'from optimization_engine\.model_cleanup\b': 'from optimization_engine.nx.model_cleanup',
r'from optimization_engine\.export_expressions\b': 'from optimization_engine.nx.export_expressions',
r'from optimization_engine\.import_expressions\b': 'from optimization_engine.nx.import_expressions',
r'from optimization_engine\.mesh_converter\b': 'from optimization_engine.nx.mesh_converter',
r'import optimization_engine\.nx_solver\b': 'import optimization_engine.nx.solver',
r'import optimization_engine\.nx_updater\b': 'import optimization_engine.nx.updater',
# =============================================================================
# STUDY MODULE
# =============================================================================
r'from optimization_engine\.study_creator\b': 'from optimization_engine.study.creator',
r'from optimization_engine\.study_wizard\b': 'from optimization_engine.study.wizard',
r'from optimization_engine\.study_state\b': 'from optimization_engine.study.state',
r'from optimization_engine\.study_reset\b': 'from optimization_engine.study.reset',
r'from optimization_engine\.study_continuation\b': 'from optimization_engine.study.continuation',
r'from optimization_engine\.benchmarking_substudy\b': 'from optimization_engine.study.benchmarking',
r'from optimization_engine\.generate_history_from_trials\b': 'from optimization_engine.study.history_generator',
# =============================================================================
# REPORTING MODULE
# =============================================================================
r'from optimization_engine\.generate_report\b': 'from optimization_engine.reporting.report_generator',
r'from optimization_engine\.generate_report_markdown\b': 'from optimization_engine.reporting.markdown_report',
r'from optimization_engine\.comprehensive_results_analyzer\b': 'from optimization_engine.reporting.results_analyzer',
r'from optimization_engine\.visualizer\b': 'from optimization_engine.reporting.visualizer',
r'from optimization_engine\.landscape_analyzer\b': 'from optimization_engine.reporting.landscape_analyzer',
# =============================================================================
# CONFIG MODULE
# =============================================================================
r'from optimization_engine\.config_manager\b': 'from optimization_engine.config.manager',
r'from optimization_engine\.optimization_config_builder\b': 'from optimization_engine.config.builder',
r'from optimization_engine\.optimization_setup_wizard\b': 'from optimization_engine.config.setup_wizard',
r'from optimization_engine\.capability_matcher\b': 'from optimization_engine.config.capability_matcher',
r'from optimization_engine\.template_loader\b': 'from optimization_engine.config.template_loader',
# =============================================================================
# UTILS MODULE
# =============================================================================
r'from optimization_engine\.logger\b': 'from optimization_engine.utils.logger',
r'from optimization_engine\.auto_doc\b': 'from optimization_engine.utils.auto_doc',
r'from optimization_engine\.realtime_tracking\b': 'from optimization_engine.utils.realtime_tracking',
r'from optimization_engine\.codebase_analyzer\b': 'from optimization_engine.utils.codebase_analyzer',
r'from optimization_engine\.pruning_logger\b': 'from optimization_engine.utils.pruning_logger',
# =============================================================================
# FUTURE MODULE
# =============================================================================
r'from optimization_engine\.research_agent\b': 'from optimization_engine.future.research_agent',
r'from optimization_engine\.pynastran_research_agent\b': 'from optimization_engine.future.pynastran_research_agent',
r'from optimization_engine\.targeted_research_planner\b': 'from optimization_engine.future.targeted_research_planner',
r'from optimization_engine\.workflow_decomposer\b': 'from optimization_engine.future.workflow_decomposer',
r'from optimization_engine\.step_classifier\b': 'from optimization_engine.future.step_classifier',
r'from optimization_engine\.llm_optimization_runner\b': 'from optimization_engine.future.llm_optimization_runner',
r'from optimization_engine\.llm_workflow_analyzer\b': 'from optimization_engine.future.llm_workflow_analyzer',
# =============================================================================
# EXTRACTORS/VALIDATORS additions
# =============================================================================
r'from optimization_engine\.op2_extractor\b': 'from optimization_engine.extractors.op2_extractor',
r'from optimization_engine\.extractor_library\b': 'from optimization_engine.extractors.extractor_library',
r'from optimization_engine\.simulation_validator\b': 'from optimization_engine.validators.simulation_validator',
# =============================================================================
# PROCESSORS
# =============================================================================
r'from optimization_engine\.adaptive_characterization\b': 'from optimization_engine.processors.adaptive_characterization',
}
# Also need to handle utils submodule imports that moved
UTILS_MAPPINGS = {
r'from optimization_engine\.utils\.nx_session_manager\b': 'from optimization_engine.nx.session_manager',
}
# Combine all mappings
ALL_MAPPINGS = {**IMPORT_MAPPINGS, **UTILS_MAPPINGS}
def find_files(root: Path, extensions: List[str], exclude_dirs: List[str] = None) -> List[Path]:
"""Find all files with given extensions, excluding certain directories."""
if exclude_dirs is None:
exclude_dirs = ['optimization_engine_BACKUP', '.venv', 'node_modules', '__pycache__', '.git']
files = []
for ext in extensions:
for f in root.rglob(f'*{ext}'):
# Check if any excluded dir is in the path
if not any(excl in str(f) for excl in exclude_dirs):
files.append(f)
return files
def update_file(filepath: Path, mappings: Dict[str, str], dry_run: bool = True) -> Tuple[int, List[str]]:
"""Update imports in a single file."""
try:
content = filepath.read_text(encoding='utf-8', errors='ignore')
except Exception as e:
print(f" ERROR reading {filepath}: {e}")
return 0, []
changes = []
new_content = content
for pattern, replacement in mappings.items():
matches = re.findall(pattern, content)
if matches:
new_content = re.sub(pattern, replacement, new_content)
changes.append(f" {pattern} -> {replacement} ({len(matches)} occurrences)")
if changes and not dry_run:
filepath.write_text(new_content, encoding='utf-8')
return len(changes), changes
def main():
dry_run = '--dry-run' in sys.argv or '--execute' not in sys.argv
if dry_run:
print("=" * 60)
print("DRY RUN MODE - No files will be modified")
print("=" * 60)
else:
print("=" * 60)
print("EXECUTE MODE - Files will be modified!")
print("=" * 60)
confirm = input("Are you sure? (yes/no): ")
if confirm.lower() != 'yes':
print("Aborted.")
return
root = Path('.')
# Find all Python files
py_files = find_files(root, ['.py'])
print(f"\nFound {len(py_files)} Python files to check")
total_changes = 0
files_changed = 0
for filepath in sorted(py_files):
count, changes = update_file(filepath, ALL_MAPPINGS, dry_run)
if count > 0:
files_changed += 1
total_changes += count
print(f"\n{filepath} ({count} changes):")
for change in changes:
print(change)
print("\n" + "=" * 60)
print(f"SUMMARY: {total_changes} changes in {files_changed} files")
print("=" * 60)
if dry_run:
print("\nTo apply changes, run: python migrate_imports.py --execute")
if __name__ == '__main__':
main()

190
optimization_config.json
View File

@@ -1,190 +0,0 @@
{
"design_variables": [
{
"name": "tip_thickness",
"type": "continuous",
"bounds": [
15.0,
25.0
],
"units": "mm",
"initial_value": 20.0
},
{
"name": "support_angle",
"type": "continuous",
"bounds": [
20.0,
40.0
],
"units": "degrees",
"initial_value": 30.0
},
{
"name": "support_blend_radius",
"type": "continuous",
"bounds": [
5.0,
15.0
],
"units": "mm",
"initial_value": 10.0
}
],
"objectives": [
{
"name": "minimize_mass",
"description": "Minimize total mass (weight reduction)",
"extractor": "mass_extractor",
"metric": "total_mass",
"direction": "minimize",
"weight": 5.0
},
{
"name": "minimize_max_stress",
"description": "Minimize maximum von Mises stress",
"extractor": "stress_extractor",
"metric": "max_von_mises",
"direction": "minimize",
"weight": 10.0
}
],
"constraints": [
{
"name": "max_displacement_limit",
"description": "Maximum allowable displacement",
"extractor": "displacement_extractor",
"metric": "max_displacement",
"type": "upper_bound",
"limit": 1.0,
"units": "mm"
},
{
"name": "max_stress_limit",
"description": "Maximum allowable von Mises stress",
"extractor": "stress_extractor",
"metric": "max_von_mises",
"type": "upper_bound",
"limit": 200.0,
"units": "MPa"
}
],
"optimization_settings": {
"n_trials": 150,
"sampler": "TPE",
"n_startup_trials": 20
},
"model_info": {
"sim_file": "/home/user/Atomizer/tests/Bracket_sim1.sim",
"solutions": [
{
"name": "DisableInThermalSolution",
"type": "DisableInThermalSolution",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
},
{
"name": "Disable in Thermal Solution 3D",
"type": "Disable in Thermal Solution 3D",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
},
{
"name": "-Flow-Structural Coupled Solution Parameters",
"type": "-Flow-Structural Coupled Solution Parameters",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
},
{
"name": "Direct Frequency Response",
"type": "Direct Frequency Response",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
},
{
"name": "*Thermal-Flow Coupled Solution Parameters",
"type": "*Thermal-Flow Coupled Solution Parameters",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
},
{
"name": "0Thermal-Structural Coupled Solution Parameters",
"type": "0Thermal-Structural Coupled Solution Parameters",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
},
{
"name": "Linear Statics",
"type": "Linear Statics",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
},
{
"name": "Disable in Thermal Solution 2D",
"type": "Disable in Thermal Solution 2D",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
},
{
"name": "Thermal Solution Parameters",
"type": "Thermal Solution Parameters",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
},
{
"name": "Nonlinear Statics",
"type": "Nonlinear Statics",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
},
{
"name": "Modal Frequency Response",
"type": "Modal Frequency Response",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
},
{
"name": "1Pass Structural Contact Solution to Flow Solver",
"type": "1Pass Structural Contact Solution to Flow Solver",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
},
{
"name": "\"ObjectDisableInThermalSolution3D",
"type": "\"ObjectDisableInThermalSolution3D",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
},
{
"name": "Design Optimization",
"type": "Design Optimization",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
},
{
"name": "\"ObjectDisableInThermalSolution2D",
"type": "\"ObjectDisableInThermalSolution2D",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
},
{
"name": "Normal Modes",
"type": "Normal Modes",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
},
{
"name": "Direct Transient Response",
"type": "Direct Transient Response",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
},
{
"name": "Modal Transient Response",
"type": "Modal Transient Response",
"solver": "NX Nastran",
"description": "Extracted from binary .sim file"
}
]
}
}

56
optimization_validation/generated_extractors/extract_displacement.py
View File

@@ -1,56 +0,0 @@
"""
Extract displacement results from OP2 file
Auto-generated by Atomizer Phase 3 - pyNastran Research Agent
Pattern: displacement
Element Type: General
Result Type: displacement
API: model.displacements[subcase]
"""
from pathlib import Path
from typing import Dict, Any
import numpy as np
from pyNastran.op2.op2 import OP2
def extract_displacement(op2_file: Path, subcase: int = 1):
"""Extract displacement results from OP2 file."""
from pyNastran.op2.op2 import OP2
import numpy as np
model = OP2()
model.read_op2(str(op2_file))
disp = model.displacements[subcase]
itime = 0 # static case
# Extract translation components
txyz = disp.data[itime, :, :3] # [tx, ty, tz]
# Calculate total displacement
total_disp = np.linalg.norm(txyz, axis=1)
max_disp = np.max(total_disp)
# Get node info
node_ids = [nid for (nid, grid_type) in disp.node_gridtype]
max_disp_node = node_ids[np.argmax(total_disp)]
return {
'max_displacement': float(max_disp),
'max_disp_node': int(max_disp_node),
'max_disp_x': float(np.max(np.abs(txyz[:, 0]))),
'max_disp_y': float(np.max(np.abs(txyz[:, 1]))),
'max_disp_z': float(np.max(np.abs(txyz[:, 2])))
}
if __name__ == '__main__':
# Example usage
import sys
if len(sys.argv) > 1:
op2_file = Path(sys.argv[1])
result = extract_displacement(op2_file)
print(f"Extraction result: {result}")
else:
print("Usage: python {sys.argv[0]} <op2_file>")

64
optimization_validation/generated_extractors/extract_solid_stress.py
View File

@@ -1,64 +0,0 @@
"""
Extract von Mises stress from CHEXA elements
Auto-generated by Atomizer Phase 3 - pyNastran Research Agent
Pattern: solid_stress
Element Type: CTETRA
Result Type: stress
API: model.ctetra_stress[subcase] or model.chexa_stress[subcase]
"""
from pathlib import Path
from typing import Dict, Any
import numpy as np
from pyNastran.op2.op2 import OP2
def extract_solid_stress(op2_file: Path, subcase: int = 1, element_type: str = 'ctetra'):
"""Extract stress from solid elements."""
from pyNastran.op2.op2 import OP2
import numpy as np
model = OP2()
model.read_op2(str(op2_file))
# Get stress object for element type
# In pyNastran, stress is stored in model.op2_results.stress
stress_attr = f"{element_type}_stress"
if not hasattr(model, 'op2_results') or not hasattr(model.op2_results, 'stress'):
raise ValueError(f"No stress results in OP2")
stress_obj = model.op2_results.stress
if not hasattr(stress_obj, stress_attr):
raise ValueError(f"No {element_type} stress results in OP2")
stress = getattr(stress_obj, stress_attr)[subcase]
itime = 0
# Extract von Mises if available
if stress.is_von_mises: # Property, not method
von_mises = stress.data[itime, :, 9] # Column 9 is von Mises
max_stress = float(np.max(von_mises))
# Get element info
element_ids = [eid for (eid, node) in stress.element_node]
max_stress_elem = element_ids[np.argmax(von_mises)]
return {
'max_von_mises': max_stress,
'max_stress_element': int(max_stress_elem)
}
else:
raise ValueError("von Mises stress not available")
if __name__ == '__main__':
# Example usage
import sys
if len(sys.argv) > 1:
op2_file = Path(sys.argv[1])
result = extract_solid_stress(op2_file)
print(f"Extraction result: {result}")
else:
print("Usage: python {sys.argv[0]} <op2_file>")

1261
reports/generate_nn_report.py
View File

File diff suppressed because it is too large Load Diff

255
reports/nn_performance/nn_performance_report.md
View File

@@ -1,255 +0,0 @@
# Neural Network Surrogate Performance Report
**Study:** uav_arm_optimization
**Generated:** 2025-11-25 15:15:52
---
## Executive Summary
**Overall Status:** EXCELLENT
Model shows excellent prediction accuracy and is suitable for production optimization.
**Training Data:** 50 FEA simulations
### Key Metrics
| Objective | MAPE | MAE | R² | Assessment |
|-----------|------|-----|----|-----------|
| mass | 1.8% | 0.00 | 0.9000 | Excellent |
| fundamental_frequency | 1.1% | 0.00 | 0.9000 | Excellent |
---
## 1. Training Data Analysis
The neural network was trained on 50 completed FEA simulations.
### Design Space Coverage
| Parameter | Min | Max | Bounds | Coverage |
|-----------|-----|-----|--------|----------|
| beam_face_thickness | 1.02 | 2.83 | [1, 3] | [OK] 90% |
| beam_half_core_thickness | 5.01 | 9.13 | [5, 10] | [OK] 82% |
| hole_count | 8.09 | 13.70 | [8, 14] | [OK] 94% |
| holes_diameter | 12.50 | 49.26 | [10, 50] | [OK] 92% |
![Training Data Coverage](training_data_coverage.png)
## 2. Prediction Accuracy
### Methodology
Prediction accuracy is evaluated by comparing neural network predictions against actual FEA results.
**Metrics used:**
- **MAPE (Mean Absolute Percentage Error):** Average percentage difference between predicted and actual values
- **MAE (Mean Absolute Error):** Average absolute difference in original units
- **R² (Coefficient of Determination):** Proportion of variance explained by the model
### Results
![Prediction Accuracy](prediction_accuracy.png)
#### mass
- MAPE: 1.77%
- MAE: 0.00
- RMSE: 0.00
- R²: 0.9000
- Samples: 50
#### fundamental_frequency
- MAPE: 1.15%
- MAE: 0.00
- RMSE: 0.00
- R²: 0.9000
- Samples: 50
## 3. Extrapolation Risk Analysis
Neural networks perform best on data similar to their training set. This section analyzes the risk of extrapolation errors.
![Extrapolation Analysis](extrapolation_analysis.png)
## 4. Optimization Performance
![Optimization Comparison](optimization_comparison.png)
### Speed Comparison
| Method | Evaluations | Est. Time | Speed |
|--------|-------------|-----------|-------|
| FEA Optimization | 50 | ~50 min | 1x |
| NN Optimization | 1000 | ~1 sec | 1200x |
## 5. Recommendations
### Immediate Actions
### Model Improvement
- Consider ensemble methods for uncertainty quantification
- Implement active learning to target high-error regions
- Add cross-validation for robust performance estimation
## 6. Summary Dashboard
![Summary Dashboard](summary_dashboard.png)
---
## Appendix
### Files Generated
- `training_data_coverage.png` - Training Data Coverage
- `prediction_accuracy.png` - Prediction Accuracy
- `optimization_comparison.png` - Optimization Comparison
- `extrapolation_analysis.png` - Extrapolation Analysis
- `summary_dashboard.png` - Summary Dashboard
### Configuration
```json
{
"study_name": "uav_arm_optimization",
"description": "UAV Camera Support Arm - Multi-Objective Lightweight Design",
"engineering_context": "Unmanned aerial vehicle camera gimbal arm. Target: lighter than current 145g design while maintaining camera stability under 850g payload. Must avoid resonance with rotor frequencies (80-120 Hz).",
"optimization_settings": {
"protocol": "protocol_11_multi_objective",
"n_trials": 30,
"sampler": "NSGAIISampler",
"pruner": null,
"timeout_per_trial": 600
},
"design_variables": [
{
"parameter": "beam_half_core_thickness",
"bounds": [
5,
10
],
"description": "Half thickness of beam core (mm) - affects weight and stiffness"
},
{
"parameter": "beam_face_thickness",
"bounds": [
1,
3
],
"description": "Thickness of beam face sheets (mm) - bending resistance"
},
{
"parameter": "holes_diameter",
"bounds": [
10,
50
],
"description": "Diameter of lightening holes (mm) - weight reduction"
},
{
"parameter": "hole_count",
"bounds": [
8,
14
],
"description": "Number of lightening holes - balance weight vs strength"
}
],
"objectives": [
{
"name": "mass",
"goal": "minimize",
"weight": 1.0,
"description": "Total mass (grams) - minimize for longer flight time",
"target": 4000,
"extraction": {
"action": "extract_mass",
"domain": "result_extraction",
"params": {
"result_type": "mass",
"metric": "total"
}
}
},
{
"name": "fundamental_frequency",
"goal": "maximize",
"weight": 1.0,
"description": "First natural frequency (Hz) - avoid rotor resonance",
"target": 150,
"extraction": {
"action": "extract_frequency",
"domain": "result_extraction",
"params": {
"result_type": "frequency",
"mode_number": 1
}
}
}
],
"constraints": [
{
"name": "max_displacement_limit",
"type": "less_than",
"threshold": 1.5,
"description": "Maximum tip displacement under 850g camera load < 1.5mm for image stabilization",
"extraction": {
"action": "extract_displacement",
"domain": "result_extraction",
"params": {
"result_type": "displacement",
"metric": "max"
}
}
},
{
"name": "max_stress_limit",
"type": "less_than",
"threshold": 120,
"description": "Maximum von Mises stress < 120 MPa (Al 6061-T6, SF=2.3)",
"extraction": {
"action": "extract_stress",
"domain": "result_extraction",
"params": {
"result_type": "stress",
"metric": "max_von_mises"
}
}
},
{
"name": "min_frequency_limit",
"type": "greater_than",
"threshold": 150,
"description": "Natural frequency > 150 Hz to avoid rotor frequencies (80-120 Hz safety margin)",
"extraction": {
"action": "extract_frequency",
"domain": "result_extraction",
"params": {
"result_type": "frequency",
"mode_number": 1
}
}
}
],
"simulation": {
"model_file": "Beam.prt",
"sim_file": "Beam_sim1.sim",
"fem_file": "Beam_fem1.fem",
"solver": "nastran",
"analysis_types": [
"static",
"modal"
]
},
"reporting": {
"generate_plots": true,
"save_incremental": true,
"llm_summary": false
}
}
```

353
run_training_fea.py
View File

@@ -1,353 +0,0 @@
"""
Parallel FEA Training Data Generator
=====================================
Runs FEA simulations on space-filling training points in parallel
using multiple NX sessions. Results are stored in a shared SQLite
database for thread-safe access.
Hardware Recommendation (based on your i7-14700HX, 64GB RAM):
- 2-3 parallel sessions recommended (each uses ~4-6 cores for Nastran)
- ~30-50 min for 100 points with 3 sessions
Usage:
python run_training_fea.py --study uav_arm_optimization --workers 3
python run_training_fea.py --study uav_arm_optimization --workers 2 --start 50
"""
import sys
import json
import argparse
import shutil
import sqlite3
import time
from pathlib import Path
from datetime import datetime
from concurrent.futures import ProcessPoolExecutor, as_completed
from multiprocessing import Manager
import threading
# Add project root to path
project_root = Path(__file__).parent
sys.path.insert(0, str(project_root))
def setup_worker_directory(study_dir: Path, worker_id: int) -> Path:
"""Create isolated working directory for each worker."""
worker_dir = study_dir / "1_setup" / f"worker_{worker_id}"
model_src = study_dir / "1_setup" / "model"
# Clean and recreate worker directory
if worker_dir.exists():
shutil.rmtree(worker_dir)
# Copy model files to worker directory
shutil.copytree(model_src, worker_dir)
return worker_dir
def run_single_fea(args_tuple):
"""
Run a single FEA simulation. This function runs in a separate process.
Args:
args_tuple: (sample_idx, sample, worker_id, study_dir, db_path)
Returns:
dict with results or error
"""
sample_idx, sample, worker_id, study_dir_str, db_path_str = args_tuple
study_dir = Path(study_dir_str)
db_path = Path(db_path_str)
# Import inside worker to avoid multiprocessing issues
import sys
sys.path.insert(0, str(Path(__file__).parent))
try:
import config as atomizer_config
except ImportError:
atomizer_config = None
from optimization_engine.nx.solver import NXSolver
from optimization_engine.extractors.extract_displacement import extract_displacement
from optimization_engine.extractors.extract_von_mises_stress import extract_solid_stress
from optimization_engine.extractors.extract_frequency import extract_frequency
from optimization_engine.extractors.extract_mass_from_expression import extract_mass_from_expression
result = {
'sample_idx': sample_idx,
'worker_id': worker_id,
'params': sample,
'success': False,
'error': None,
'mass': None,
'frequency': None,
'max_displacement': None,
'max_stress': None
}
try:
# Setup worker directory
worker_dir = setup_worker_directory(study_dir, worker_id)
# Initialize NX solver for this worker
nx_solver = NXSolver(
nastran_version=atomizer_config.NX_VERSION if atomizer_config else "2412",
timeout=atomizer_config.NASTRAN_TIMEOUT if atomizer_config else 600,
use_journal=True,
enable_session_management=True,
study_name=f"training_worker_{worker_id}"
)
# Setup paths
model_file = worker_dir / "Beam.prt"
sim_file = worker_dir / "Beam_sim1.sim"
print(f"[Worker {worker_id}] Sample {sample_idx}: {sample}")
# Run simulation
sim_result = nx_solver.run_simulation(
sim_file=sim_file,
working_dir=worker_dir,
expression_updates=sample,
solution_name=None # Solve all solutions
)
if not sim_result['success']:
result['error'] = sim_result.get('error', 'Unknown error')
print(f"[Worker {worker_id}] Sample {sample_idx} FAILED: {result['error']}")
return result
op2_file = sim_result['op2_file']
# Extract results
# Mass from CAD expression
mass_kg = extract_mass_from_expression(model_file, expression_name="p173")
result['mass'] = mass_kg * 1000.0 # Convert to grams
# Frequency from modal analysis
op2_modal = str(op2_file).replace("solution_1", "solution_2")
freq_result = extract_frequency(op2_modal, subcase=1, mode_number=1)
result['frequency'] = freq_result['frequency']
# Displacement from static analysis
disp_result = extract_displacement(op2_file, subcase=1)
result['max_displacement'] = disp_result['max_displacement']
# Stress from static analysis
stress_result = extract_solid_stress(op2_file, subcase=1, element_type='cquad4')
result['max_stress'] = stress_result['max_von_mises']
result['success'] = True
print(f"[Worker {worker_id}] Sample {sample_idx} SUCCESS: mass={result['mass']:.1f}g, freq={result['frequency']:.1f}Hz")
# Save to database immediately (thread-safe with retries)
save_result_to_db(db_path, result)
except Exception as e:
result['error'] = str(e)
print(f"[Worker {worker_id}] Sample {sample_idx} ERROR: {e}")
return result
def save_result_to_db(db_path: Path, result: dict, max_retries: int = 5):
"""Save result to SQLite database with retry logic for concurrency."""
for attempt in range(max_retries):
try:
conn = sqlite3.connect(str(db_path), timeout=30)
cursor = conn.cursor()
cursor.execute("""
INSERT OR REPLACE INTO training_results
(sample_idx, params_json, success, error, mass, frequency, max_displacement, max_stress, timestamp)
VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
""", (
result['sample_idx'],
json.dumps(result['params']),
result['success'],
result['error'],
result['mass'],
result['frequency'],
result['max_displacement'],
result['max_stress'],
datetime.now().isoformat()
))
conn.commit()
conn.close()
return True
except sqlite3.OperationalError as e:
if "locked" in str(e) and attempt < max_retries - 1:
time.sleep(0.5 * (attempt + 1)) # Exponential backoff
else:
raise
return False
def init_database(db_path: Path):
"""Initialize SQLite database for training results."""
conn = sqlite3.connect(str(db_path))
cursor = conn.cursor()
cursor.execute("""
CREATE TABLE IF NOT EXISTS training_results (
sample_idx INTEGER PRIMARY KEY,
params_json TEXT,
success INTEGER,
error TEXT,
mass REAL,
frequency REAL,
max_displacement REAL,
max_stress REAL,
timestamp TEXT
)
""")
conn.commit()
conn.close()
def get_completed_samples(db_path: Path) -> set:
"""Get set of already completed sample indices."""
if not db_path.exists():
return set()
conn = sqlite3.connect(str(db_path))
cursor = conn.cursor()
try:
cursor.execute("SELECT sample_idx FROM training_results WHERE success = 1")
completed = {row[0] for row in cursor.fetchall()}
except sqlite3.OperationalError:
completed = set()
conn.close()
return completed
def main():
parser = argparse.ArgumentParser(description='Run parallel FEA training data generation')
parser.add_argument('--study', required=True, help='Study name (e.g., uav_arm_optimization)')
parser.add_argument('--workers', type=int, default=2, help='Number of parallel workers (default: 2)')
parser.add_argument('--start', type=int, default=0, help='Starting sample index (for resuming)')
parser.add_argument('--end', type=int, default=None, help='Ending sample index (exclusive)')
parser.add_argument('--resume', action='store_true', help='Skip already completed samples')
args = parser.parse_args()
# Setup paths
study_dir = project_root / "studies" / args.study
training_points_path = study_dir / "1_setup" / "training_points.json"
db_path = study_dir / "2_results" / "training_data.db"
if not study_dir.exists():
print(f"ERROR: Study not found: {study_dir}")
return
if not training_points_path.exists():
print(f"ERROR: Training points not found: {training_points_path}")
print(f"Generate them first: python generate_training_data.py --study {args.study}")
return
# Load training points
with open(training_points_path) as f:
data = json.load(f)
samples = data['samples']
total_samples = len(samples)
# Apply start/end filtering
end_idx = args.end if args.end else total_samples
samples_to_run = [(i, samples[i]) for i in range(args.start, min(end_idx, total_samples))]
print("=" * 70)
print("PARALLEL FEA TRAINING DATA GENERATOR")
print("=" * 70)
print(f"Study: {args.study}")
print(f"Total training points: {total_samples}")
print(f"Processing range: {args.start} to {end_idx}")
print(f"Parallel workers: {args.workers}")
print(f"Database: {db_path}")
print()
# Initialize database
db_path.parent.mkdir(exist_ok=True)
init_database(db_path)
# Check for already completed samples
if args.resume:
completed = get_completed_samples(db_path)
samples_to_run = [(i, s) for i, s in samples_to_run if i not in completed]
print(f"Already completed: {len(completed)} samples")
print(f"Remaining to process: {len(samples_to_run)} samples")
if not samples_to_run:
print("All samples already completed!")
return
print()
print(f"Starting {args.workers} parallel workers...")
print("=" * 70)
# Prepare worker arguments
worker_args = []
for idx, (sample_idx, sample) in enumerate(samples_to_run):
worker_id = idx % args.workers
worker_args.append((sample_idx, sample, worker_id, str(study_dir), str(db_path)))
# Track progress
start_time = time.time()
completed_count = 0
failed_count = 0
# Run with ProcessPoolExecutor
with ProcessPoolExecutor(max_workers=args.workers) as executor:
futures = {executor.submit(run_single_fea, arg): arg[0] for arg in worker_args}
for future in as_completed(futures):
sample_idx = futures[future]
try:
result = future.result()
if result['success']:
completed_count += 1
else:
failed_count += 1
except Exception as e:
print(f"Sample {sample_idx} raised exception: {e}")
failed_count += 1
# Progress update
total_done = completed_count + failed_count
elapsed = time.time() - start_time
rate = total_done / elapsed if elapsed > 0 else 0
remaining = len(samples_to_run) - total_done
eta = remaining / rate / 60 if rate > 0 else 0
print(f"\nProgress: {total_done}/{len(samples_to_run)} ({completed_count} OK, {failed_count} failed)")
print(f"Rate: {rate:.2f} samples/min | ETA: {eta:.1f} min")
# Summary
elapsed = time.time() - start_time
print()
print("=" * 70)
print("TRAINING DATA GENERATION COMPLETE")
print("=" * 70)
print(f"Total time: {elapsed/60:.1f} minutes")
print(f"Completed: {completed_count}/{len(samples_to_run)}")
print(f"Failed: {failed_count}")
print(f"Results saved to: {db_path}")
print()
print("Next steps:")
print(" 1. Merge with existing optimization data:")
print(f" python merge_training_data.py --study {args.study}")
print(" 2. Retrain neural network:")
print(f" python train_nn_surrogate.py --study {args.study}")
if __name__ == "__main__":
main()

115
train_neural.bat
View File

@@ -1,115 +0,0 @@
@echo off
REM ============================================================================
REM Atomizer Neural Network Training Script
REM ============================================================================
REM Trains a parametric neural network on collected FEA data
REM
REM Usage:
REM train_neural.bat (uses default: bracket study)
REM train_neural.bat study_name (specify study name)
REM train_neural.bat study_name 200 (specify epochs)
REM ============================================================================
echo.
echo ============================================================================
echo ATOMIZER NEURAL NETWORK TRAINING
echo ============================================================================
echo.
REM Set defaults
set STUDY_NAME=bracket_stiffness_optimization_atomizerfield
set EPOCHS=100
REM Parse arguments
if not "%~1"=="" set STUDY_NAME=%~1
if not "%~2"=="" set EPOCHS=%~2
set TRAIN_DIR=atomizer_field_training_data\%STUDY_NAME%
set OUTPUT_DIR=atomizer-field\runs\%STUDY_NAME%
echo Study: %STUDY_NAME%
echo Epochs: %EPOCHS%
echo Train Dir: %TRAIN_DIR%
echo Output: %OUTPUT_DIR%
echo.
REM Check training data exists
if not exist "%TRAIN_DIR%" (
echo ERROR: Training data not found at %TRAIN_DIR%
echo.
echo Available training data directories:
dir atomizer_field_training_data /b 2>nul
echo.
pause
exit /b 1
)
REM Count trials
echo Counting training trials...
set /a count=0
for /d %%d in ("%TRAIN_DIR%\trial_*") do set /a count+=1
echo Found %count% trials
echo.
if %count% LSS 20 (
echo WARNING: Less than 20 trials. Training may not converge well.
echo Consider running more FEA simulations first.
echo.
)
REM Check PyTorch
python -c "import torch; print(f'PyTorch {torch.__version__}')" 2>nul
if errorlevel 1 (
echo ERROR: PyTorch not installed
echo Run install.bat first
pause
exit /b 1
)
REM Check for GPU
echo.
echo Checking for GPU...
python -c "import torch; print(f'CUDA available: {torch.cuda.is_available()}'); print(f'Device: {torch.cuda.get_device_name(0) if torch.cuda.is_available() else \"CPU\"}')"
echo.
REM Start training
echo ============================================================================
echo Starting training...
echo ============================================================================
echo.
echo Press Ctrl+C to stop training early (checkpoint will be saved)
echo.
cd atomizer-field
python train_parametric.py ^
--train_dir "..\%TRAIN_DIR%" ^
--epochs %EPOCHS% ^
--output_dir "runs\%STUDY_NAME%" ^
--batch_size 16 ^
--learning_rate 0.001 ^
--hidden_channels 128 ^
--num_layers 4
if errorlevel 1 (
echo.
echo Training failed! Check error messages above.
cd ..
pause
exit /b 1
)
cd ..
echo.
echo ============================================================================
echo TRAINING COMPLETE!
echo ============================================================================
echo.
echo Model saved to: %OUTPUT_DIR%\checkpoint_best.pt
echo.
echo To use the trained model in optimization:
echo cd studies\%STUDY_NAME%
echo python run_optimization.py --run --trials 100 --enable-nn --resume
echo.
pause