Files

Antoine ec5e42d733 feat: Add M1 mirror Zernike optimization with correct RMS calculation

Major improvements to telescope mirror optimization workflow:

Assembly FEM Workflow (solve_simulation.py):
- Fixed multi-part assembly FEM update sequence
- Use ImportFromFile() for reliable expression updates
- Add DuplicateNodesCheckBuilder with MergeOccurrenceNodes=True
- Switch to Foreground solve mode for multi-subcase solutions
- Add detailed logging and diagnostics for node merge operations

Zernike RMS Calculation:
- CRITICAL FIX: Use correct surface-based RMS formula
  - Global RMS = sqrt(mean(W^2)) from actual WFE values
  - Filtered RMS = sqrt(mean(W_residual^2)) after removing low-order fit
  - This matches zernike_Post_Script_NX.py (optical standard)
- Previous WRONG formula was: sqrt(sum(coeffs^2))
- Add compute_rms_filter_j1to3() for optician workload metric

Subcase Mapping:
- Fix subcase mapping to match NX model:
  - Subcase 1 = 90 deg (polishing orientation)
  - Subcase 2 = 20 deg (reference)
  - Subcase 3 = 40 deg
  - Subcase 4 = 60 deg

New Study: M1 Mirror Zernike Optimization
- Full optimization config with 11 design variables
- 3 objectives: rel_filtered_rms_40_vs_20, rel_filtered_rms_60_vs_20, mfg_90_optician_workload
- Neural surrogate support for accelerated optimization

Documentation:
- Update ZERNIKE_INTEGRATION.md with correct RMS formula
- Update ASSEMBLY_FEM_WORKFLOW.md with expression import and node merge details
- Add reference scripts from original zernike_Post_Script_NX.py

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

2025-11-28 16:30:15 -05:00

4.0 KiB

Raw Blame History

M1 Mirror Zernike Optimization Dashboard

Study Overview

Objective: Optimize telescope primary mirror (M1) support structure to minimize wavefront error across different gravity orientations.

Method: Hybrid FEA + Neural Network acceleration using Zernike polynomial decomposition.

Quick Start

1. Prepare Model Files

Copy your NX model files to:

studies/m1_mirror_zernike_optimization/1_setup/model/

Required files:

ASSY_M1.prt (or your assembly name)
ASSY_M1_assyfem1.afm
ASSY_M1_assyfem1_sim1.sim
Associated .fem and _i.prt files

2. Run FEA Trials (Build Training Data)

cd studies/m1_mirror_zernike_optimization
python run_optimization.py --run --trials 40

This will:

Run ~40 FEA trials (10-15 min each = ~8-10 hours)
Extract 50 Zernike coefficients for each subcase (20/40/60/90 deg)
Store all data in Optuna database

3. Train Neural Surrogate

python run_optimization.py --train-surrogate

Trains MLP to predict 200 outputs (50 coefficients x 4 subcases) from design variables.

4. Run Neural-Accelerated Optimization

python run_optimization.py --run --trials 1000 --enable-nn

1000 trials in ~seconds!

5. View Results

Optuna Dashboard:

optuna-dashboard sqlite:///2_results/study.db --port 8081

Open http://localhost:8081

Design Variables

Variable	Range	Baseline	Units	Status
whiffle_min	35-55	40.55	mm	Enabled
whiffle_outer_to_vertical	68-80	75.67	deg	Enabled
inner_circular_rib_dia	480-620	534.00	mm	Enabled
whiffle_triangle_closeness	50-65	60.00	mm	Disabled
blank_backface_angle	3.5-5.0	4.23	deg	Disabled
lateral_inner_angle	25-28.5	26.79	deg	Disabled
lateral_outer_angle	13-17	14.64	deg	Disabled
lateral_outer_pivot	9-12	10.40	mm	Disabled
lateral_inner_pivot	9-12	10.07	mm	Disabled
lateral_middle_pivot	18-23	20.73	mm	Disabled
lateral_closeness	9.5-12.5	11.02	mm	Disabled

Edit 1_setup/optimization_config.json to enable/disable variables.

Objectives

Objective	Weight	Target	Description
rel_filtered_rms_40_vs_20	5	4 nm	WFE at 40° relative to 20° reference
rel_filtered_rms_60_vs_20	5	10 nm	WFE at 60° relative to 20° reference
mfg_90_optician_workload	1	20 nm	Polishing workload at 90° orientation

Strategy: Weighted sum minimization (normalized by targets)

Neural Surrogate Architecture

Input: Design variables (3-11 depending on enabled)
Output: 200 values (50 Zernike coefficients × 4 subcases)
Architecture: MLP with 4 layers, 128 hidden units
Training: ~40 FEA samples, 200 epochs

File Structure

m1_mirror_zernike_optimization/
├── 1_setup/
│   ├── optimization_config.json    # Configuration
│   └── model/                      # NX model files (add yours here)
├── 2_results/
│   ├── study.db                    # Optuna database
│   └── zernike_surrogate/          # Trained neural model
│       └── checkpoint_best.pt
├── run_optimization.py             # Main script
└── DASHBOARD.md                    # This file

Commands Reference

# Run FEA optimization
python run_optimization.py --run --trials 40

# Train neural surrogate
python run_optimization.py --train-surrogate

# Run with neural acceleration
python run_optimization.py --run --trials 1000 --enable-nn

# Check status
python run_optimization.py --status

# Launch Optuna dashboard
optuna-dashboard sqlite:///2_results/study.db --port 8081

Tips

Start small: Run 5-10 FEA trials first to verify everything works
Check Zernike extraction: Verify OP2 has correct subcases (20/40/60/90)
Enable variables gradually: Start with 3, add more after initial exploration
Neural validation: After finding good neural designs, verify top candidates with FEA

4.0 KiB Raw Blame History Unescape Escape