Anto01
857c01e7ca
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generated_hooks/ (legacy generator outputs) - Remove optimization_validation/ (empty skeleton) - Remove reports/ (superseded by optimization_engine/reporting/) - Remove root-level stale files: DEVELOPMENT.md, INSTALL_INSTRUCTIONS.md, config.py, atomizer_paths.py,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>
1.9 KiB
1.9 KiB
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)