Atomizer/docs/protocols/operations/OP_05_EXPORT_TRAINING_DATA.md

OP_05: Export Training Data

Overview

This protocol covers exporting FEA simulation data for training neural network surrogates. Proper data export enables Protocol 14 (Neural Acceleration).

---

When to Use

Trigger	Action
"export training data"	Follow this protocol
"neural network data"	Follow this protocol
Planning >50 trials	Consider export for acceleration
Want to train surrogate	Follow this protocol

---

Quick Reference

Export Command:

python run_optimization.py --export-training

Output Structure:

atomizer_field_training_data/{study_name}/
├── trial_0001/
│   ├── input/model.bdf
│   ├── output/model.op2
│   └── metadata.json
├── trial_0002/
│   └── ...
└── study_summary.json

Recommended Data Volume:

Complexity	Training Samples	Validation Samples
Simple (2-3 params)	50-100	20-30
Medium (4-6 params)	100-200	30-50
Complex (7+ params)	200-500	50-100

---

Configuration

Enable Export in Config

Add to optimization_config.json:

{
  "training_data_export": {
    "enabled": true,
    "export_dir": "atomizer_field_training_data/my_study",
    "export_bdf": true,
    "export_op2": true,
    "export_fields": ["displacement", "stress"],
    "include_failed": false
  }
}

Configuration Options

Option	Type	Default	Description
enabled	bool	false	Enable export
export_dir	string	-	Output directory
export_bdf	bool	true	Save Nastran input
export_op2	bool	true	Save binary results
export_fields	list	all	Which result fields
include_failed	bool	false	Include failed trials

---

Export Workflow

Step 1: Run with Export Enabled

conda activate atomizer
cd studies/my_study
python run_optimization.py --export-training

Or run standard optimization with config export enabled.

Step 2: Verify Export

ls atomizer_field_training_data/my_study/
# Should see trial_0001/, trial_0002/, etc.

# Check a trial
ls atomizer_field_training_data/my_study/trial_0001/
# input/model.bdf
# output/model.op2
# metadata.json

Step 3: Check Metadata

cat atomizer_field_training_data/my_study/trial_0001/metadata.json

{
  "trial_number": 1,
  "design_parameters": {
    "thickness": 5.2,
    "width": 30.0
  },
  "objectives": {
    "mass": 0.234,
    "max_stress": 198.5
  },
  "constraints_satisfied": true,
  "simulation_time": 145.2
}

Step 4: Check Study Summary

cat atomizer_field_training_data/my_study/study_summary.json

{
  "study_name": "my_study",
  "total_trials": 50,
  "successful_exports": 47,
  "failed_exports": 3,
  "design_parameters": ["thickness", "width"],
  "objectives": ["mass", "max_stress"],
  "export_timestamp": "2025-12-05T15:30:00"
}

---

Data Quality Checks

Verify Sample Count

from pathlib import Path
import json

export_dir = Path("atomizer_field_training_data/my_study")
trials = list(export_dir.glob("trial_*"))
print(f"Exported trials: {len(trials)}")

# Check for missing files
for trial_dir in trials:
    bdf = trial_dir / "input" / "model.bdf"
    op2 = trial_dir / "output" / "model.op2"
    meta = trial_dir / "metadata.json"

    if not all([bdf.exists(), op2.exists(), meta.exists()]):
        print(f"Missing files in {trial_dir}")

Check Parameter Coverage

import json
import numpy as np

# Load all metadata
params = []
for trial_dir in export_dir.glob("trial_*"):
    with open(trial_dir / "metadata.json") as f:
        meta = json.load(f)
        params.append(meta["design_parameters"])

# Check coverage
import pandas as pd
df = pd.DataFrame(params)
print(df.describe())

# Look for gaps
for col in df.columns:
    print(f"{col}: min={df[col].min():.2f}, max={df[col].max():.2f}")

---

Space-Filling Sampling

For best neural network training, use space-filling designs:

Latin Hypercube Sampling

from scipy.stats import qmc

# Generate space-filling samples
n_samples = 100
n_params = 4

sampler = qmc.LatinHypercube(d=n_params)
samples = sampler.random(n=n_samples)

# Scale to parameter bounds
lower = [2.0, 20.0, 5.0, 1.0]
upper = [10.0, 50.0, 15.0, 5.0]
scaled = qmc.scale(samples, lower, upper)

Sobol Sequence

sampler = qmc.Sobol(d=n_params)
samples = sampler.random(n=n_samples)
scaled = qmc.scale(samples, lower, upper)

---

Next Steps After Export

1. Parse to Neural Format

cd atomizer-field
python batch_parser.py ../atomizer_field_training_data/my_study

2. Split Train/Validation

from sklearn.model_selection import train_test_split

# 80/20 split
train_trials, val_trials = train_test_split(
    all_trials,
    test_size=0.2,
    random_state=42
)

3. Train Model

python train_parametric.py \
  --train_dir ../training_data/parsed \
  --val_dir ../validation_data/parsed \
  --epochs 200

See SYS_14_NEURAL_ACCELERATION for full training workflow.

---

Troubleshooting

Symptom	Cause	Solution
No export directory	Export not enabled	Add training_data_export to config
Missing OP2 files	Solve failed	Check include_failed: false
Incomplete metadata	Extraction error	Check extractor logs
Low sample count	Too many failures	Relax constraints

---

Cross-References

• Related: SYS_14_NEURAL_ACCELERATION
• Preceded By: OP_02_RUN_OPTIMIZATION
• Skill: .claude/skills/modules/neural-acceleration.md

---

Version History

Version	Date	Changes
1.0	2025-12-05	Initial release