Atomizer/studies/uav_arm_atomizerfield_test

UAV Arm AtomizerField Test Study

Overview

This document summarizes the setup of the UAV arm AtomizerField neural surrogate test study, demonstrating the integration of Graph Neural Networks (GNN) for 600x-500,000x speedup in FEA-based optimization.

Study Location

studies/uav_arm_atomizerfield_test/
├── 1_setup/
│   ├── model/           # NX model files (copied from uav_arm_optimization)
│   └── workflow_config.json  # Neural surrogate configuration
├── 2_results/           # Will contain optimization results
├── run_optimization.py  # Neural-enhanced runner script
└── reset_study.py      # Clean reset utility

Neural Surrogate Configuration

The study is configured with a phased optimization strategy:

Phase 1: Initial FEA Exploration (Trials 0-30)

• Purpose: Collect diverse training data using pure FEA
• Neural Surrogate: DISABLED
• Training Data Export: ENABLED to atomizer_field_training_data/uav_arm_test/

Phase 2: Neural Training (Trials 31-40)

• Purpose: Continue FEA while training neural network
• Neural Surrogate: DISABLED (training in background)
• Action Required: Run AtomizerField training scripts

Phase 3: Neural Exploitation (Trials 41-180)

• Purpose: Rapid optimization using neural surrogate
• Neural Surrogate: ENABLED (600x speedup)
• Confidence Threshold: 85% (fallback to FEA if lower)

Phase 4: Final Validation (Trials 181-200)

• Purpose: Validate best designs with FEA
• Neural Surrogate: DISABLED
• Ensures accuracy of final results

Key Features

1. Training Data Export

• Automatically exports NX Nastran .dat (input) and .op2 (results) files
• Creates structured directory with metadata.json for each trial
• Compatible with AtomizerField batch_parser.py

2. Confidence-Based Fallback

• Neural predictions include confidence estimate
• Automatically falls back to FEA when confidence < 85%
• Ensures reliability while maximizing speedup

3. Hybrid Optimization

• Smart switching between FEA and NN based on:
  • Current optimization phase
  • Prediction confidence
  • Validation frequency
  • Drift detection

4. Performance Tracking

• Tracks speedup metrics for each neural prediction
• Exports performance report after optimization
• Shows time saved and accuracy achieved

Running the Test

Step 1: Initial FEA Trials (Collect Training Data)

cd studies/uav_arm_atomizerfield_test
python run_optimization.py --trials 30

This will:

• Run 30 FEA trials to explore design space
• Export training data to atomizer_field_training_data/uav_arm_test/
• Create optimization database in 2_results/study.db

Step 2: Train Neural Network (AtomizerField)

cd atomizer-field
python batch_parser.py --data-dir ../atomizer_field_training_data/uav_arm_test
python train.py --epochs 200 --model GraphUNet

Step 3: Enable Neural Surrogate

Update workflow_config.json:

{
  "neural_surrogate": {
    "enabled": true,  // Change from false to true
    "model_checkpoint": "atomizer-field/checkpoints/uav_arm_model/best.pt"
  }
}

Step 4: Continue Optimization with Neural Acceleration

python run_optimization.py --trials 170 --resume --enable-nn

This will:

• Use neural network for trials 41-180 (140 trials)
• Achieve 600x+ speedup (50ms vs 30 minutes per evaluation)
• Fall back to FEA when confidence is low
• Validate final 20 designs with FEA

Expected Results

Without Neural Surrogate (Pure FEA)

• 200 trials × 30 minutes = 100 hours
• Limited design space exploration
• High computational cost

With Neural Surrogate

• 50 FEA trials × 30 minutes = 25 hours
• 150 NN trials × 50ms = 7.5 seconds
• Total: ~25 hours (75% reduction)
• 600x more designs evaluated in exploitation phase

Monitoring Progress

The script provides real-time feedback:

Trial 42: Used neural network (confidence: 94.2%, time: 0.048s)
Trial 43: Neural confidence too low (72.1%), using FEA
Trial 44: Used neural network (confidence: 91.8%, time: 0.051s)

Final summary:

============================================================
NEURAL NETWORK SPEEDUP SUMMARY
============================================================
Trials using neural network: 140/200 (70.0%)
Average NN inference time: 0.052 seconds
Average NN confidence: 92.3%
Estimated speedup: 34,615x
Time saved: ~70.0 hours
============================================================

Design Variables (4)

1. beam_half_core_thickness: 20-30 mm
2. beam_face_thickness: 1-3 mm
3. holes_diameter: 180-280 mm
4. hole_count: 8-14 (integer)

Objectives (2)

1. Minimize mass (target < 120g)
2. Maximize fundamental frequency (target > 150 Hz)

Constraints (3)

1. Max displacement < 1.5mm (850g camera load)
2. Max stress < 120 MPa (Al 6061-T6, SF=2.3)
3. Min frequency > 150 Hz (avoid rotor resonance)

Files Created

1. run_optimization.py: Neural-enhanced optimization runner

   • Uses NeuralOptimizationRunner from runner_with_neural.py
   • Integrates with existing NX solver and extractors
   • Command-line flags for training and enabling NN

2. workflow_config.json: Complete neural surrogate configuration

   • Neural model settings (checkpoint, confidence, device)
   • Hybrid optimization phases
   • Training data export configuration
   • Performance tracking settings

3. reset_study.py: Clean reset utility

   • Removes results and training data
   • Preserves setup and model files

Next Steps

1. Run initial FEA trials to generate training data
2. Train AtomizerField model on collected data
3. Enable neural surrogate and continue optimization
4. Analyze speedup metrics and validate accuracy
5. Deploy to production if successful

Integration Status

✅ Neural surrogate module created (optimization_engine/neural_surrogate.py) ✅ Neural runner created (optimization_engine/runner_with_neural.py) ✅ Training data exporter integrated (optimization_engine/training_data_exporter.py) ✅ UAV arm test study configured ⏳ Waiting to run initial trials and train model

Technical Details

• Neural Architecture: Graph U-Net with 718k parameters
• Input: FEA mesh topology + design variables
• Output: Stress, displacement, frequency predictions
• Physics Loss: Enforces equilibrium and boundary conditions
• Ensemble: 3 models for uncertainty quantification
• Device: CUDA GPU for 10x faster inference

This test study demonstrates the seamless integration of AtomizerField neural surrogates with Atomizer, enabling dramatic speedup in engineering optimization while maintaining accuracy through confidence-based fallback and validation.