Atomizer/projects/hydrotech-beam/studies/01_doe_landscape

Study: 01_doe_landscape — Hydrotech Beam

See ../../README.md for project overview.

Purpose

Map the design space of the Hydrotech sandwich I-beam to identify feasible regions, characterize variable sensitivities, and prepare data for Phase 2 (TPE optimization). This is Phase 1 of the two-phase strategy (DEC-HB-002).

Quick Facts

Item	Value
Objective	Minimize mass (kg)
Constraints	Tip displacement ≤ 10 mm, Von Mises stress ≤ 130 MPa
Design variables	4 (3 continuous + 1 integer)
Algorithm	Phase 1: LHS DoE (50 trials + 1 baseline)
Total budget	51 evaluations
Constraint handling	Deb's feasibility rules (Optuna constraint interface)
Status	✅ Pipeline operational — first results obtained 2026-02-11
Solver	DesigncenterNX 2512 (NX Nastran SOL 101)
Solve time	~12 s/trial (~10 min for full DOE)
First results	Displacement=17.93mm, Stress=111.9MPa

Design Variables (confirmed via NX binary introspection)

ID	NX Expression	Range	Type	Baseline	Unit
DV1	beam_half_core_thickness	10–40	Continuous	25.162	mm
DV2	beam_face_thickness	10–40	Continuous	21.504	mm
DV3	holes_diameter	150–450	Continuous	300	mm
DV4	hole_count	5–15	Integer	10	—

Usage

Prerequisites

pip install -r requirements.txt

Requires: Python 3.10+, optuna, scipy, numpy, pandas.

Development Run (stub solver)

# From the study directory:
cd projects/hydrotech-beam/studies/01_doe_landscape/

# Run with synthetic results (for pipeline testing):
python run_doe.py --backend stub

# With verbose logging:
python run_doe.py --backend stub -v

# Custom study name:
python run_doe.py --backend stub --study-name my_test_run

Production Run (NXOpen on Windows)

# On dalidou (Windows node with DesigncenterNX 2512):
cd C:\Users\antoi\Atomizer\projects\hydrotech-beam\studies\01_doe_landscape
conda activate atomizer
python run_doe.py --backend nxopen --model-dir "../../models"

⚠️ --model-dir can be relative — the code resolves it with Path.resolve(). NX requires fully resolved paths (no .. components).

Resume Interrupted Study

python run_doe.py --backend stub --resume --study-name hydrotech_beam_doe_phase1

CLI Options

Flag	Default	Description
--backend	stub	stub (testing) or nxopen (real NX)
--model-dir	—	Path to NX model files (required for nxopen)
--study-name	hydrotech_beam_doe_phase1	Optuna study name
--n-samples	50	Number of LHS sample points
--seed	42	Random seed for reproducibility
--results-dir	results/	Output directory
--resume	false	Resume existing study
--clean	false	Delete Optuna DB and start fresh (history DB preserved)
-v	false	Verbose (DEBUG) logging

Files

File	Description
run_doe.py	Main entry point — orchestrates the DoE study
sampling.py	LHS generation with stratified integer sampling
geometric_checks.py	Pre-flight geometric feasibility filter
nx_interface.py	AtomizerNXSolver wrapping optimization_engine
iteration_manager.py	Smart retention — last 10 + best 3 with full data
requirements.txt	Python dependencies
OPTIMIZATION_STRATEGY.md	Full strategy document
results/	Output directory (CSV, JSON, Optuna DB)
iterations/	Per-trial archived outputs

Output Files

After a study run, results/ will contain:

File	Description
doe_results.csv	All trials — DVs, objectives, constraints, status
doe_summary.json	Study metadata, statistics, best feasible design
optuna_study.db	SQLite database (Optuna persistence, resume support)
history.db	Persistent SQLite — all DVs, results, feasibility (survives --clean)
history.csv	CSV mirror of history DB

Iteration Folders

Each trial produces iterations/iterNNN/ containing:

File	Description
params.json	Design variable values for this trial
params.exp	NX expression file used for this trial
results.json	Extracted results (mass, displacement, stress, feasibility)
*.op2	Nastran binary results (for post-processing)
*.f06	Nastran text output (for debugging)

Smart retention policy (runs every 5 iterations):

• Keep last 10 iterations with full data
• Keep best 3 feasible iterations with full data
• Strip model files from all others, keep solver outputs

Architecture

run_doe.py
  ├── sampling.py              Generate 50 LHS + 1 baseline
  │     └── scipy.stats.qmc.LatinHypercube
  ├── geometric_checks.py      Pre-flight feasibility filter
  │     ├── Hole overlap:  span/(n-1) - d ≥ 30mm
  │     └── Web clearance: 500 - 2·face - d > 0
  ├── nx_interface.py          AtomizerNXSolver (wraps optimization_engine)
  │     ├── NXStubSolver       → synthetic results (development)
  │     └── AtomizerNXSolver   → real DesigncenterNX 2512 SOL 101 (production)
  ├── iteration_manager.py     Smart retention (last 10 + best 3)
  ├── Optuna study
  │     ├── SQLite storage (resume support)
  │     ├── Enqueued trials (deterministic LHS)
  │     └── Deb's feasibility rules (constraint interface)
  └── history.db               Persistent history (append-only, survives --clean)

Pipeline per Trial

Trial N
  │
  ├── 1. Suggest DVs (from enqueued LHS point)
  │
  ├── 2. Geometric pre-check
  │     ├── FAIL → record as infeasible, skip NX
  │     └── PASS ↓
  │
  ├── 3. Restore master model from backup
  │
  ├── 4. NX evaluation (IN-PLACE in models/ directory)
  │     ├── Write .exp file with trial DVs
  │     ├── Open .sim (loads model chain)
  │     ├── Import expressions, rebuild geometry
  │     ├── Update FEM (remesh)
  │     ├── Solve SOL 101 (~12s)
  │     └── Extract: mass (p173 → _temp_mass.txt), displacement, stress (via OP2)
  │
  ├── 5. Archive outputs to iterations/iterNNN/
  │     └── params.json, params.exp, results.json, OP2, F06
  │
  ├── 6. Record results + constraint violations
  │
  ├── 7. Log to Optuna DB + history.db + CSV
  │
  └── 8. Smart retention check (every 5 iterations)

Phase 1 → Phase 2 Gate Criteria

Before proceeding to Phase 2 (TPE optimization), these checks must pass:

Check	Threshold	Action if FAIL
Feasible points found	≥ 5	Expand bounds or relax constraints (escalate to CEO)
NX solve success rate	≥ 80%	Investigate failures, fix model, re-run
No systematic crashes at bounds	Visual check	Tighten bounds away from failure region

Key Design Decisions

• Baseline as Trial 0 — LAC lesson: always validate baseline first
• Pre-flight geometric filter — catches infeasible geometry before NX (saves compute, avoids crashes)
• Stratified integer sampling — ensures all 11 hole_count levels (5-15) are covered
• Deb's feasibility rules — no penalty weight tuning; feasible always beats infeasible
• SQLite persistence — study can be interrupted and resumed
• No surrogates — LAC lesson: direct FEA via TPE beats surrogate + L-BFGS

NX Integration Notes

Solver: DesigncenterNX 2512 (Siemens rebrand of NX). Install: C:\Program Files\Siemens\DesigncenterNX2512.

The nx_interface.py module provides:

• NXStubSolver — synthetic results from simplified beam mechanics (for pipeline testing)
• AtomizerNXSolver — wraps optimization_engine NXSolver for real DesigncenterNX 2512 solves

Expression names are exact from binary introspection. Critical: beam_lenght has a typo in NX (no 'h') — use exact spelling.

Outputs extracted from NX:

Output	Source	Unit	Notes
Mass	Expression p173 → _temp_mass.txt	kg	Journal extracts after solve
Tip displacement	OP2 parse via pyNastran	mm	Max Tz at free end
Max VM stress	OP2 parse via pyNastran	MPa	⚠️ pyNastran returns kPa — divide by 1000

Critical Lessons Learned (2026-02-11)

1. Path resolution: Use Path.resolve() NOT Path.absolute() on Windows. .absolute() doesn't resolve .. components — NX can't follow them.
2. File references: NX .sim files have absolute internal refs to .fem/.prt. Never copy model files to iteration folders — solve in-place with backup/restore.
3. NX version: Must match exactly. Model files from 2512 won't load in 2412 ("Part file is from a newer version").
4. pyNastran: Warns about unsupported NX version 2512 but works fine. Safe to ignore.

References

• OPTIMIZATION_STRATEGY.md — Full strategy document
• ../../BREAKDOWN.md — Tech Lead's technical analysis
• ../../DECISIONS.md — Decision log
• ../../CONTEXT.md — Project context & expression map

---

Code: 🏗️ Study Builder (Technical Lead) | Strategy: ⚡ Optimizer Agent | Updated: 2026-02-11