Atomizer/.claude/ATOMIZER_CONTEXT.md

Atomizer Session Context

What is Atomizer?

Atomizer is an LLM-first FEA (Finite Element Analysis) optimization framework. Users describe optimization problems in natural language, and Claude orchestrates the entire workflow: model introspection, config generation, optimization execution, and results analysis.

Philosophy: LLM-driven optimization. Engineers describe what they want; AI handles the rest.

---

Session Initialization Checklist

On EVERY new session, perform these steps:

Step 1: Identify Working Directory

If in: c:\Users\Antoine\Atomizer\           → Project root (full capabilities)
If in: c:\Users\Antoine\Atomizer\studies\*  → Inside a study (load study context)
If elsewhere:                               → Limited context (warn user)

Step 2: Detect Study Context

If working directory contains optimization_config.json:

1. Read the config to understand the study
2. Check 2_results/study.db for optimization status
3. Summarize study state to user

Python utility for study detection:

# Get study state for current directory
python -m optimization_engine.study_state .

# Get all studies in Atomizer
python -c "from optimization_engine.study_state import get_all_studies; from pathlib import Path; [print(f'{s[\"study_name\"]}: {s[\"status\"]}') for s in get_all_studies(Path('.'))]"

Step 3: Route to Task Protocol

Use keyword matching to load appropriate context:

User Intent	Keywords	Load Protocol	Action
Create study	"create", "new", "set up", "optimize"	OP_01 + SYS_12	Launch study builder
Run optimization	"run", "start", "execute", "trials"	OP_02 + SYS_15	Execute optimization
Check progress	"status", "progress", "how many"	OP_03	Query study.db
Analyze results	"results", "best", "Pareto", "analyze"	OP_04	Generate analysis
Neural acceleration	"neural", "surrogate", "turbo", "NN", "SAT"	SYS_14 + SYS_16	Method selection
NX/CAD help	"NX", "model", "mesh", "expression"	MCP + nx-docs	Use Siemens MCP
Physics insights	"zernike", "stress view", "insight"	SYS_16	Generate insights
Troubleshoot	"error", "failed", "fix", "debug"	OP_06	Diagnose issues

---

Quick Reference

Core Commands

# Optimization workflow
python run_optimization.py --start --trials 50  # Run optimization
python run_optimization.py --start --resume     # Continue interrupted run
python run_optimization.py --test               # Single trial test

# Neural acceleration
python run_nn_optimization.py --turbo --nn-trials 5000  # Fast NN exploration
python -m optimization_engine.method_selector config.json study.db  # Get recommendation

# Dashboard
cd atomizer-dashboard && npm run dev     # Start at http://localhost:3003

When to Use --resume

Scenario	Use --resume?
First run of new study	NO
First run with seeding (e.g., V15 from V14)	NO - seeding is automatic
Continue interrupted run	YES
Add more trials to completed study	YES

Key: --resume continues existing study.db. Seeding from source_studies in config happens automatically on first run - don't confuse seeding with resuming!

Study Structure (100% standardized)

Studies are organized by geometry type:

studies/
├── M1_Mirror/                    # Mirror optimization studies
│   ├── m1_mirror_adaptive_V14/
│   ├── m1_mirror_cost_reduction_V3/
│   └── m1_mirror_cost_reduction_V4/
├── Simple_Bracket/               # Bracket studies
├── UAV_Arm/                      # UAV arm studies
├── Drone_Gimbal/                 # Gimbal studies
├── Simple_Beam/                  # Beam studies
└── _Other/                       # Test/experimental studies

Individual study structure:

studies/{geometry_type}/{study_name}/
├── optimization_config.json    # Problem definition
├── run_optimization.py         # FEA optimization script
├── run_turbo_optimization.py   # GNN-Turbo acceleration (optional)
├── README.md                   # MANDATORY documentation
├── STUDY_REPORT.md             # Results template
├── 1_setup/
│   ├── optimization_config.json # Config copy for reference
│   └── model/
│       ├── Model.prt           # NX part file
│       ├── Model_sim1.sim      # NX simulation
│       └── Model_fem1.fem      # FEM definition
├── 2_iterations/               # FEA trial folders (trial_NNNN/)
│   ├── trial_0001/             # Zero-padded, NEVER reset
│   ├── trial_0002/
│   └── ...
├── 3_results/
│   ├── study.db                # Optuna-compatible database
│   ├── optimization.log        # Logs
│   └── turbo_report.json       # NN results (if run)
└── 3_insights/                 # Study Insights (SYS_16)
    ├── zernike_*.html          # Zernike WFE visualizations
    ├── stress_*.html           # Stress field visualizations
    └── design_space_*.html     # Parameter exploration

IMPORTANT: When creating a new study, always place it under the appropriate geometry type folder!

Available Extractors (SYS_12)

ID	Physics	Function	Notes
E1	Displacement	extract_displacement()	mm
E2	Frequency	extract_frequency()	Hz
E3	Von Mises Stress	extract_solid_stress()	Specify element_type!
E4	BDF Mass	extract_mass_from_bdf()	kg
E5	CAD Mass	extract_mass_from_expression()	kg
E8-10	Zernike WFE (standard)	extract_zernike_*()	nm (mirrors)
E12-14	Phase 2	Principal stress, strain energy, SPC forces
E15-18	Phase 3	Temperature, heat flux, modal mass
E20	Zernike Analytic	extract_zernike_analytic()	nm (parabola-based)
E22	Zernike OPD	extract_zernike_opd()	nm (RECOMMENDED)

Critical: For stress extraction, specify element type:

• Shell (CQUAD4): element_type='cquad4'
• Solid (CTETRA): element_type='ctetra'

---

Protocol System Overview

┌─────────────────────────────────────────────────────────────────┐
│ Layer 0: BOOTSTRAP (.claude/skills/00_BOOTSTRAP.md)            │
│ Purpose: Task routing, quick reference                          │
└─────────────────────────────────────────────────────────────────┘
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│ Layer 1: OPERATIONS (docs/protocols/operations/OP_*.md)        │
│ OP_01: Create Study    OP_02: Run Optimization                  │
│ OP_03: Monitor         OP_04: Analyze Results                   │
│ OP_05: Export Data     OP_06: Troubleshoot                      │
└─────────────────────────────────────────────────────────────────┘
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│ Layer 2: SYSTEM (docs/protocols/system/SYS_*.md)               │
│ SYS_10: IMSO (single-obj)    SYS_11: Multi-objective           │
│ SYS_12: Extractors           SYS_13: Dashboard                  │
│ SYS_14: Neural Accel         SYS_15: Method Selector            │
│ SYS_16: SAT (Self-Aware Turbo) - VALIDATED v3, WS=205.58        │
│ SYS_17: Context Engineering                                      │
└─────────────────────────────────────────────────────────────────┘
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│ Layer 3: EXTENSIONS (docs/protocols/extensions/EXT_*.md)       │
│ EXT_01: Create Extractor  EXT_02: Create Hook                   │
│ EXT_03: Create Protocol   EXT_04: Create Skill                  │
└─────────────────────────────────────────────────────────────────┘

---

Subagent Routing

For complex tasks, Claude should spawn specialized subagents:

Task	Subagent Type	Context to Load
Create study from description	general-purpose	core/study-creation-core.md, SYS_12
Explore codebase	Explore	(built-in)
Plan architecture	Plan	(built-in)
NX API lookup	general-purpose	Use MCP siemens-docs tools

---

Environment Setup

CRITICAL: Always use the atomizer conda environment:

conda activate atomizer
python run_optimization.py

DO NOT:

• Install packages with pip/conda (everything is installed)
• Create new virtual environments
• Use system Python

NX Open Requirements:

• NX 2506 installed at C:\Program Files\Siemens\NX2506\
• Use run_journal.exe for NX automation

---

Template Registry

Available study templates for quick creation:

Template	Objectives	Extractors	Example Study
multi_objective_structural	mass, stress, stiffness	E1, E3, E4	bracket_pareto_3obj
frequency_optimization	frequency, mass	E2, E4	uav_arm_optimization
mirror_wavefront	Zernike RMS	E8-E10	m1_mirror_zernike
shell_structural	mass, stress	E1, E3, E4	beam_pareto_4var
thermal_structural	temperature, stress	E3, E15	(template only)

Python utility for templates:

# List all templates
python -m optimization_engine.templates

# Get template details in code
from optimization_engine.templates import get_template, suggest_template
template = suggest_template(n_objectives=2, physics_type="structural")

---

Auto-Documentation Protocol

When Claude creates/modifies extractors or protocols:

1. Code change → Update optimization_engine/extractors/__init__.py
2. Doc update → Update SYS_12_EXTRACTOR_LIBRARY.md
3. Quick ref → Update .claude/skills/01_CHEATSHEET.md
4. Commit → Use structured message: feat: Add E{N} {name} extractor

---

Key Principles

1. Conversation first - Don't ask user to edit JSON manually
2. Validate everything - Catch errors before FEA runs
3. Explain decisions - Say why you chose a sampler/protocol
4. NEVER modify master files - Copy NX files to study directory
5. ALWAYS reuse code - Check extractors before writing new code
6. Proactive documentation - Update docs after code changes

---

Base Classes (Phase 2 - Code Deduplication)

New studies should use these base classes instead of duplicating code:

ConfigDrivenRunner (FEA Optimization)

# run_optimization.py - Now just ~30 lines instead of ~300
from optimization_engine.base_runner import ConfigDrivenRunner

runner = ConfigDrivenRunner(__file__)
runner.run()  # Handles --discover, --validate, --test, --run

ConfigDrivenSurrogate (Neural Acceleration)

# run_nn_optimization.py - Now just ~30 lines instead of ~600
from optimization_engine.generic_surrogate import ConfigDrivenSurrogate

surrogate = ConfigDrivenSurrogate(__file__)
surrogate.run()  # Handles --train, --turbo, --all

Templates: optimization_engine/templates/run_*_template.py

---

CRITICAL: NXSolver Initialization Pattern

NEVER pass full config dict to NXSolver. This causes TypeError: expected str, bytes or os.PathLike object, not dict.

WRONG

self.nx_solver = NXSolver(self.config)  # ❌ NEVER DO THIS

CORRECT - FEARunner Pattern

Always wrap NXSolver in a FEARunner class with explicit parameters:

class FEARunner:
    def __init__(self, config: Dict):
        self.config = config
        self.nx_solver = None
        self.master_model_dir = SETUP_DIR / "model"

    def setup(self):
        import re
        nx_settings = self.config.get('nx_settings', {})
        nx_install_dir = nx_settings.get('nx_install_path', 'C:\\Program Files\\Siemens\\NX2506')

        version_match = re.search(r'NX(\d+)', nx_install_dir)
        nastran_version = version_match.group(1) if version_match else "2506"

        self.nx_solver = NXSolver(
            master_model_dir=str(self.master_model_dir),
            nx_install_dir=nx_install_dir,
            nastran_version=nastran_version,
            timeout=nx_settings.get('simulation_timeout_s', 600),
            use_iteration_folders=True,
            study_name=self.config.get('study_name', 'my_study')
        )

    def run_fea(self, params, iter_num):
        if self.nx_solver is None:
            self.setup()
        # ... run simulation

Reference implementations:

• studies/m1_mirror_adaptive_V14/run_optimization.py
• studies/m1_mirror_adaptive_V15/run_optimization.py

---

Skill Registry (Phase 3 - Consolidated Skills)

All skills now have YAML frontmatter with metadata for versioning and dependency tracking.

Skill ID	Name	Type	Version	Location
SKILL_000	Bootstrap	bootstrap	2.0	.claude/skills/00_BOOTSTRAP.md
SKILL_001	Cheatsheet	reference	2.0	.claude/skills/01_CHEATSHEET.md
SKILL_002	Context Loader	loader	2.0	.claude/skills/02_CONTEXT_LOADER.md
SKILL_CORE_001	Study Creation Core	core	2.4	.claude/skills/core/study-creation-core.md

Deprecated Skills

Old File	Reason	Replacement
create-study.md	Duplicate of core skill	core/study-creation-core.md

Skill Metadata Format

All skills use YAML frontmatter:

---
skill_id: SKILL_XXX
version: X.X
last_updated: YYYY-MM-DD
type: bootstrap|reference|loader|core|module
code_dependencies:
  - path/to/code.py
requires_skills:
  - SKILL_YYY
replaces: old-skill.md  # if applicable
---

---

Subagent Commands (Phase 5 - Specialized Agents)

Atomizer provides specialized subagent commands for complex tasks:

Command	Purpose	When to Use
/study-builder	Create new optimization studies	"create study", "set up optimization"
/nx-expert	NX Open API help, model automation	"how to in NX", "update mesh"
/protocol-auditor	Validate configs and code quality	"validate config", "check study"
/results-analyzer	Analyze optimization results	"analyze results", "best solution"

Command Files

.claude/commands/
├── study-builder.md      # Create studies from descriptions
├── nx-expert.md          # NX Open / Simcenter expertise
├── protocol-auditor.md   # Config and code validation
├── results-analyzer.md   # Results analysis and reporting
└── dashboard.md          # Dashboard control

Subagent Invocation Pattern

# Master agent delegates to specialized subagent
Task(
    subagent_type='general-purpose',
    prompt='''
    Load context from .claude/commands/study-builder.md

    User request: "{user's request}"

    Follow the workflow in the command file.
    ''',
    description='Study builder task'
)

---

Auto-Documentation (Phase 4 - Self-Expanding Knowledge)

Atomizer can auto-generate documentation from code:

# Generate all documentation
python -m optimization_engine.auto_doc all

# Generate only extractor docs
python -m optimization_engine.auto_doc extractors

# Generate only template docs
python -m optimization_engine.auto_doc templates

Generated Files:

• docs/generated/EXTRACTORS.md - Full extractor reference (auto-generated)
• docs/generated/EXTRACTOR_CHEATSHEET.md - Quick reference table
• docs/generated/TEMPLATES.md - Study templates reference

When to Run Auto-Doc:

1. After adding a new extractor
2. After modifying template registry
3. Before major releases

---

Trial Management System (v2.3)

New unified trial management ensures consistency across all optimization methods:

Key Components

Component	Path	Purpose
TrialManager	optimization_engine/utils/trial_manager.py	Unified trial folder + DB management
DashboardDB	optimization_engine/utils/dashboard_db.py	Optuna-compatible database wrapper

Trial Naming Convention

2_iterations/
├── trial_0001/    # Zero-padded, monotonically increasing
├── trial_0002/    # NEVER reset, NEVER overwritten
├── trial_0003/
└── ...

Key principles:

• Trial numbers NEVER reset (monotonically increasing)
• Folders NEVER get overwritten
• Database is always in sync with filesystem
• Surrogate predictions (5K) are NOT trials - only FEA results

Usage

from optimization_engine.utils.trial_manager import TrialManager

tm = TrialManager(study_dir)

# Start new trial
trial = tm.new_trial(params={'rib_thickness': 10.5})

# After FEA completes
tm.complete_trial(
    trial_number=trial['trial_number'],
    objectives={'wfe_40_20': 5.63, 'mass_kg': 118.67},
    weighted_sum=42.5,
    is_feasible=True
)

Database Schema (Optuna-Compatible)

The DashboardDB class creates Optuna-compatible schema for dashboard integration:

• trials - Main trial records with state, datetime, value
• trial_values - Objective values (supports multiple objectives)
• trial_params - Design parameter values
• trial_user_attributes - Metadata (source, solve_time, etc.)
• studies - Study metadata (directions, name)

---

Version Info

Component	Version	Last Updated
ATOMIZER_CONTEXT	2.0	2026-01-20
Documentation Structure	2.0	2026-01-20
BaseOptimizationRunner	1.0	2025-12-07
GenericSurrogate	1.0	2025-12-07
Study State Detector	1.0	2025-12-07
Template Registry	1.0	2025-12-07
Extractor Library	1.4	2025-12-12
Method Selector	2.1	2025-12-07
Protocol System	2.1	2025-12-12
Skill System	2.1	2025-12-12
Auto-Doc Generator	1.0	2025-12-07
Subagent Commands	1.0	2025-12-07
FEARunner Pattern	1.0	2025-12-12
Study Insights	1.0	2025-12-20
TrialManager	1.0	2025-12-28
DashboardDB	1.0	2025-12-28
GNN-Turbo System	2.3	2025-12-28

---

Atomizer: LLM-driven structural optimization for engineering.