- history.db: SQLite append-only, never deleted by --clean
- history.csv: Auto-exported after each trial (live updates)
- Logs: DVs, results, feasibility, status, solve time, iter path
- Cross-study queries: full lineage across all runs/phases
- --clean only resets Optuna DB, history preserved
- Each iteration gets full model files in iterations/iterNNN/ (openable in NX)
- Retention: keep last 10 + best 3 with full models, strip the rest
- Stripped iterations keep solver outputs (OP2, F06, params, results)
- All paths resolved to absolute before passing to NX (fixes reference issue)
- iteration_manager.py: reusable for future studies
- One-time backup of model files at study start (_model_backup/)
- Restore clean state before each trial (files stay in models/, NX refs intact)
- If a trial corrupts the model, next trial starts clean
- Best of both: NX reference integrity + trial isolation
- Each iteration gets: params.json, results.json, OP2, F06, mass files
- Model directory stays clean (no solver output buildup)
- Study folder is self-contained with full trial history
- solve_simulation.py: FEM finder now excludes idealized parts, falls back to loading .fem
- solve_simulation.py: hole_count written as [Constant] not [MilliMeter] in .exp
- run_doe.py: dual logging to console + results/doe_run.log
Complete the NXOpenSolver class in nx_interface.py with production-ready
evaluate() and close() methods, following proven patterns from
M1_Mirror/SAT3_Trajectory_V7.
Pipeline per trial:
1. NXSolver.create_iteration_folder() — HEEDS-style isolation with fresh
model copies + params.exp generation
2. NXSolver.run_simulation() — journal-based solve via run_journal.exe
(handles expression import, geometry rebuild, FEM update, SOL 101)
3. extract_displacement() — max displacement from OP2
4. extract_solid_stress() — max von Mises with auto-detect element type
(tries all solid types first, falls back to CQUAD4 shell)
5. extract_mass_from_expression() — reads _temp_mass.txt from journal,
with _temp_part_properties.json fallback
Key decisions:
- Auto-detect element type for stress (element_type=None) instead of
hardcoding CQUAD4 — the beam model may use solid or shell elements
- Lazy solver init on first evaluate() call for clean error handling
- OP2 fallback path: tries solver result first, then expected naming
convention (beam_sim1-solution_1.op2)
- Mass fallback: _temp_mass.txt -> _temp_part_properties.json
- LAC-compliant close(): only uses session_manager.cleanup_stale_locks(),
never kills NX processes directly
Expression mapping (confirmed from binary introspection):
- beam_half_core_thickness, beam_face_thickness, holes_diameter, hole_count
- Mass output: p173 (body_property147.mass, kg)
Refs: OP_09, OPTIMIZATION_STRATEGY.md §8.2
- Use optimization_engine.nx.updater.NXParameterUpdater for expression updates (.exp import method)
- Use optimization_engine.nx.solver.NXSolver for journal-based solving (run_journal.exe)
- Use optimization_engine.extractors for displacement, stress, and mass extraction
- Displacement: extract_displacement() from pyNastran OP2
- Stress: extract_solid_stress() with cquad4 support (shell elements), kPa→MPa conversion
- Mass: extract_mass_from_expression() reads from temp file written by solve journal
- Support for iteration folders (HEEDS-style clean state between trials)
- Proper error handling with TrialResult(success=False, error_message=...)
- 600s timeout per trial (matching existing NXSolver default)
- Keep stub solver and create_solver() factory working
- Maintain run_doe.py interface compatibility
New project structure with knowledge base integration:
- kb/ with components, materials, fea, dev generations
- models/ for reference NX files (golden copies)
- studies/ for self-contained optimization campaigns
- deliverables/ for final outputs
- DECISIONS.md decision log (6 decisions tracked)
- BREAKDOWN.md (moved from 1_breakdown/)
- Gen 001 created from intake + technical breakdown
KB extension file: atomizer/shared/skills/knowledge-base-atomizer-ext.md
Refs: DEC-HB-004, DEC-HB-005, DEC-HB-006
