d5ffba099e
Permanently integrates the Atomizer-Field GNN surrogate system: - neural_models/: Graph Neural Network for FEA field prediction - batch_parser.py: Parse training data from FEA exports - train.py: Neural network training pipeline - predict.py: Inference engine for fast predictions This enables 600x-2200x speedup over traditional FEA by replacing expensive simulations with millisecond neural network predictions. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
127 lines
5.1 KiB
Markdown
127 lines
5.1 KiB
Markdown
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 |