Atomizer/studies/M1_Mirror/m1_mirror_surrogate_turbo

M1 Mirror Surrogate Turbo Optimization

GNN Surrogate Training + Iterative Turbo Optimization

Train a Graph Neural Network on ~1000 FEA trials from V6-V12, then use it to accelerate optimization with iterative FEA validation.

Strategy Overview

V6-V9: ~794 trials (Z-only extraction)
   │
   ├─── Re-extract with ZernikeOPD ────┐
   │                                    │
V11-V12: ~200+ trials (ZernikeOPD)     │
   │                                    │
   └─────────────── Aggregate ──────────┘
                       │
                       ▼
               Train GNN (~1000 samples)
                       │
                       ▼
        ┌──────────────────────────────┐
        │        TURBO LOOP            │
        ├──────────────────────────────┤
        │ 1. GNN: 5000 trials (fast)   │
        │ 2. Select 5 diverse designs  │
        │ 3. FEA: Validate 5 (~25 min) │
        │ 4. Update if accuracy drops  │
        │ 5. Check convergence         │
        └──────────────────────────────┘
                       │
                       ▼
            Best design (WS < 280?)

Why This Approach?

1. V6-V9 data is valuable but used wrong extraction method (Z-only Zernike)
2. Re-extraction unifies all data under ZernikeOPD with extract_relative()
3. ~1000 samples is 3x more training data than typical single-study GNN
4. Turbo mode accelerates exploration: 5000 GNN predictions in ~2 min vs ~8 hours FEA
5. Iterative validation ensures GNN predictions match reality

Source Studies

Study	Trials	Original Method	Action
V6	196	Z-only	Re-extract
V7	199	Z-only	Re-extract
V8	199	Z-only	Re-extract
V9	200	Z-only	Re-extract
V11	199	ZernikeOPD	Use directly
V12	TBD	ZernikeOPD	Use directly

V10 excluded - abandoned study with potential quality issues.

Configuration

Design Variables (10 total)

Variable	Min	Max	Baseline	Units
Whiffle (4)
whiffle_min	30.0	72.0	61.92	mm
whiffle_outer_to_vertical	70.0	85.0	81.75	mm
whiffle_triangle_closeness	65.0	120.0	65.54	mm
blank_backface_angle	4.2	4.5	4.42	deg
Lateral (6)
lateral_inner_angle	25.0	30.0	29.90	deg
lateral_outer_angle	11.0	17.0	11.76	deg
lateral_outer_pivot	7.0	12.0	11.13	deg
lateral_inner_pivot	5.0	12.0	6.52	deg
lateral_middle_pivot	15.0	27.0	22.39	deg
lateral_closeness	7.0	12.0	8.69	deg

Turbo Settings

Setting	Value
GNN trials per iteration	5000
FEA validations per iteration	5
Max iterations	20
Convergence patience	2 iterations
GNN accuracy threshold	R² > 0.95

Weighted Sum Formula

WS = 5*(40-20 RMS) + 5*(60-20 RMS) + 3*(MFG 90) + 1*(mass)

Usage

Phase 1: Re-extract V6-V9

cd studies/M1_Mirror/m1_mirror_surrogate_turbo
python scripts/01_reextract_studies.py

# Or single study:
python scripts/01_reextract_studies.py --study V7

Time estimate: 2-3 hours for ~794 trials

Phase 2: Aggregate Training Data

python scripts/02_aggregate_gnn_data.py

# Custom train/val split:
python scripts/02_aggregate_gnn_data.py --train-ratio 0.85

Time estimate: 30 minutes

Phase 3: Train GNN

python scripts/03_train_gnn.py --epochs 200

# Quick test:
python scripts/03_train_gnn.py --epochs 50 --quick

Time estimate: 2-4 hours (GPU), longer on CPU

Phase 4: Run Turbo Optimization

python run_turbo_optimization.py

# GNN-only exploration (no FEA):
python run_turbo_optimization.py --no-fea

# Custom settings:
python run_turbo_optimization.py --gnn-trials 10000 --fea-per-iter 3

Time estimate: ~45 min per iteration (with FEA)

Directory Structure

m1_mirror_surrogate_turbo/
├── 1_setup/
│   ├── optimization_config.json    # Full config with all 10 variables
│   ├── source_studies.json         # Metadata about source studies
│   └── model/                      # NX model files (V11 best)
├── 2_iterations/
│   └── iter{N}/                    # FEA validation working directories
├── 3_results/
│   ├── reextracted/                # Re-extracted V6-V9 data
│   │   ├── V6/
│   │   ├── V7/
│   │   ├── V8/
│   │   └── V9/
│   ├── gnn_data/                   # Aggregated training data
│   │   ├── train/
│   │   ├── val/
│   │   └── dataset_meta.json
│   ├── checkpoints/                # GNN model saves
│   │   ├── best_model.pt
│   │   └── training_history.json
│   ├── turbo_logs/                 # Per-iteration results
│   │   └── iter_0000.json
│   └── study.db                    # Track validation trials
├── scripts/
│   ├── 01_reextract_studies.py     # Batch re-extraction
│   ├── 02_aggregate_gnn_data.py    # Data aggregation
│   └── 03_train_gnn.py             # Training wrapper
├── run_turbo_optimization.py       # Main turbo loop
├── README.md
└── STUDY_REPORT.md

Expected Outcomes

1. GNN with R² > 0.95 for all subcase predictions
2. Find design with WS < 280 (improvement over V11's 284.19)
3. GNN predictions within 10% of FEA validation
4. Validated best design ready for manufacturing

Current Baseline (V11 Trial 190)

Metric	Value
Weighted Sum	284.19
40-20 RMS	6.50 nm
60-20 RMS	14.08 nm
MFG 90	28.89 nm
Mass	94.66 kg

Success Criteria

• GNN validation R² > 0.95 for all subcases
• Find design with WS < 280
• GNN-predicted objectives within 10% of FEA
• Complete at least 3 turbo iterations

---

Created: 2025-12-23 Atomizer Framework