Atomizer/studies/bracket_stiffness_optimization_atomizerfield/README.md

Bracket Stiffness Optimization - AtomizerField

Multi-objective bracket geometry optimization with neural network acceleration.

Created: 2025-11-26 Protocol: Protocol 11 (Multi-Objective NSGA-II) Status: Ready to Run

---

1. Engineering Problem

1.1 Objective

Optimize an L-bracket mounting structure to maximize structural stiffness while minimizing mass, subject to manufacturing constraints.

1.2 Physical System

• Component: L-shaped mounting bracket
• Material: Steel (density ρ defined in NX model)
• Loading: Static force applied in Z-direction
• Boundary Conditions: Fixed support at mounting face
• Analysis Type: Linear static (Nastran SOL 101)

---

2. Mathematical Formulation

2.1 Objectives

Objective	Goal	Weight	Formula	Units
Stiffness	maximize	1.0	k = \frac{F}{\delta_{max}}	N/mm
Mass	minimize	0.1	m = \sum_{e} \rho_e V_e	kg

Where:

• k = structural stiffness
• F = applied force magnitude (N)
• \delta_{max} = maximum absolute displacement (mm)
• \rho_e = element material density (kg/mm³)
• V_e = element volume (mm³)

2.2 Design Variables

Parameter	Symbol	Bounds	Units	Description
Support Angle	\theta	[20, 70]	degrees	Angle of the support arm relative to base
Tip Thickness	t	[30, 60]	mm	Thickness at the bracket tip

Design Space:

\mathbf{x} = [\theta, t]^T \in \mathbb{R}^2 20 \leq \theta \leq 70 30 \leq t \leq 60

2.3 Constraints

Constraint	Type	Formula	Threshold	Handling
Mass Limit	Inequality	g_1(\mathbf{x}) = m - m_{max}	m_{max} = 0.2 kg	Infeasible if violated

Feasible Region:

\mathcal{F} = \{\mathbf{x} : g_1(\mathbf{x}) \leq 0\}

2.4 Multi-Objective Formulation

Pareto Optimization Problem:

\max_{\mathbf{x} \in \mathcal{F}} \quad k(\mathbf{x}) \min_{\mathbf{x} \in \mathcal{F}} \quad m(\mathbf{x})

Pareto Dominance: Solution \mathbf{x}_1 dominates \mathbf{x}_2 if:

• k(\mathbf{x}_1) \geq k(\mathbf{x}_2) and m(\mathbf{x}_1) \leq m(\mathbf{x}_2)
• With at least one strict inequality

---

3. Optimization Algorithm

3.1 NSGA-II Configuration

Parameter	Value	Description
Algorithm	NSGA-II	Non-dominated Sorting Genetic Algorithm II
Population	auto	Managed by Optuna
Directions	['maximize', 'minimize']	(stiffness, mass)
Sampler	NSGAIISampler	Multi-objective sampler
Trials	100	50 FEA + 50 neural

NSGA-II Properties:

• Fast non-dominated sorting: O(MN^2) where M = objectives, N = population
• Crowding distance for diversity preservation
• Binary tournament selection with crowding comparison

3.2 Return Format

def objective(trial) -> Tuple[float, float]:
    # ... simulation and extraction ...
    return (stiffness, mass)  # Tuple, NOT negated

---

4. Simulation Pipeline

4.1 Trial Execution Flow

┌─────────────────────────────────────────────────────────────────────┐
│                         TRIAL n EXECUTION                            │
├─────────────────────────────────────────────────────────────────────┤
│                                                                      │
│  1. OPTUNA SAMPLES (NSGA-II)                                        │
│     θ = trial.suggest_float("support_angle", 20, 70)                │
│     t = trial.suggest_float("tip_thickness", 30, 60)                │
│                                                                      │
│  2. NX PARAMETER UPDATE                                             │
│     Module: optimization_engine/nx_updater.py                       │
│     Action: Bracket.prt expressions ← {θ, t}                        │
│                                                                      │
│  3. HOOK: PRE_SOLVE                                                 │
│     → Log trial start, validate design bounds                       │
│                                                                      │
│  4. NX SIMULATION (Nastran SOL 101)                                 │
│     Module: optimization_engine/solve_simulation.py                 │
│     Input: Bracket_sim1.sim                                         │
│     Output: .dat, .op2, .f06                                        │
│                                                                      │
│  5. HOOK: POST_SOLVE                                                │
│     → Run export_displacement_field.py                              │
│     → Generate export_field_dz.fld                                  │
│                                                                      │
│  6. RESULT EXTRACTION                                               │
│     Mass ← bdf_mass_extractor(.dat)                                 │
│     Stiffness ← stiffness_calculator(.fld, .op2)                   │
│                                                                      │
│  7. HOOK: POST_EXTRACTION                                           │
│     → Export BDF/OP2 to training data directory                     │
│                                                                      │
│  8. CONSTRAINT EVALUATION                                           │
│     mass ≤ 0.2 kg → feasible/infeasible                            │
│                                                                      │
│  9. RETURN TO OPTUNA                                                │
│     return (stiffness, mass)                                        │
│                                                                      │
└─────────────────────────────────────────────────────────────────────┘

4.2 Hooks Configuration

Hook Point	Function	Purpose
PRE_SOLVE	log_trial_start()	Log design variables, trial number
POST_SOLVE	export_field_data()	Run NX journal for .fld export
POST_EXTRACTION	export_training_data()	Save BDF/OP2 for neural training

---

5. Result Extraction Methods

5.1 Mass Extraction

Attribute	Value
Extractor	bdf_mass_extractor
Module	optimization_engine.extractors.bdf_mass_extractor
Function	extract_mass_from_bdf()
Source	bracket_sim1-solution_1.dat
Output	kg

Algorithm:

m = \sum_{e=1}^{N_{elem}} m_e = \sum_{e=1}^{N_{elem}} \rho_e \cdot V_e

Where element volume V_e is computed from BDF geometry (CTETRA, CHEXA, etc.).

Code:

from optimization_engine.extractors.bdf_mass_extractor import extract_mass_from_bdf

mass_kg = extract_mass_from_bdf("1_setup/model/bracket_sim1-solution_1.dat")

5.2 Stiffness Extraction

Attribute	Value
Extractor	stiffness_calculator
Module	optimization_engine.extractors.stiffness_calculator
Displacement Source	export_field_dz.fld
Force Source	bracket_sim1-solution_1.op2
Components	Force: F_z, Displacement: \delta_z
Output	N/mm

Algorithm:

k = \frac{F_z}{\delta_{max,z}}

Where:

• F_z = applied force in Z-direction (extracted from OP2 OLOAD resultant)
• \delta_{max,z} = \max_{i \in nodes} |u_{z,i}| (from field export)

Code:

from optimization_engine.extractors.stiffness_calculator import StiffnessCalculator

calculator = StiffnessCalculator(
    field_file="1_setup/model/export_field_dz.fld",
    op2_file="1_setup/model/bracket_sim1-solution_1.op2",
    force_component="fz",
    displacement_component="z"
)
result = calculator.calculate()
stiffness = result['stiffness']  # N/mm

5.3 Field Data Format

NX Field Export (.fld):

FIELD: [ResultProbe] : [TABLE]
RESULT TYPE: Displacement
COMPONENT: Z
START DATA
step, node_id, value
0, 396, -0.086716040968895
0, 397, -0.091234567890123
...
END DATA

---

6. Neural Acceleration (AtomizerField)

6.1 Configuration

Setting	Value	Description
enabled	true	Neural surrogate active
min_training_points	50	FEA trials before auto-training
auto_train	true	Trigger training automatically
epochs	100	Training epochs
validation_split	0.2	20% holdout for validation
retrain_threshold	25	Retrain after N new FEA points
model_type	parametric	Input: design params only

6.2 Surrogate Model

Input: \mathbf{x} = [\theta, t]^T \in \mathbb{R}^2

Output: \hat{\mathbf{y}} = [\hat{k}, \hat{m}]^T \in \mathbb{R}^2

Architecture: Parametric neural network (MLP)

Training Objective:

\mathcal{L} = \frac{1}{N} \sum_{i=1}^{N} \left[ (k_i - \hat{k}_i)^2 + (m_i - \hat{m}_i)^2 \right]

6.3 Training Data Location

atomizer_field_training_data/bracket_stiffness_optimization_atomizerfield/
├── trial_0001/
│   ├── input/model.bdf       # Mesh + design parameters
│   ├── output/model.op2      # FEA displacement/stress results
│   └── metadata.json         # {support_angle, tip_thickness, stiffness, mass}
├── trial_0002/
└── ...

6.4 Expected Performance

Metric	Value
FEA time per trial	10-30 min
Neural time per trial	~4.5 ms
Speedup	~2,200x
Expected R²	> 0.95 (after 50 samples)

---

7. Study File Structure

bracket_stiffness_optimization_atomizerfield/
│
├── 1_setup/                              # INPUT CONFIGURATION
│   ├── model/                            # NX Model Files
│   │   ├── Bracket.prt                   # Parametric part
│   │   │   └── Expressions: support_angle, tip_thickness
│   │   ├── Bracket_sim1.sim              # Simulation (SOL 101)
│   │   ├── Bracket_fem1.fem              # FEM mesh (auto-updated)
│   │   ├── bracket_sim1-solution_1.dat   # Nastran BDF input
│   │   ├── bracket_sim1-solution_1.op2   # Binary results
│   │   ├── bracket_sim1-solution_1.f06   # Text summary
│   │   ├── export_displacement_field.py  # Field export journal
│   │   └── export_field_dz.fld           # Z-displacement field
│   │
│   ├── optimization_config.json          # Study configuration
│   └── workflow_config.json              # Workflow metadata
│
├── 2_results/                            # OUTPUT (auto-generated)
│   ├── study.db                          # Optuna SQLite database
│   ├── optimization_history.json         # Trial history
│   ├── pareto_front.json                 # Pareto-optimal solutions
│   ├── optimization.log                  # Structured log
│   └── reports/                          # Generated reports
│       └── optimization_report.md        # Full results report
│
├── run_optimization.py                   # Entry point
├── reset_study.py                        # Database reset
└── README.md                             # This blueprint

---

8. Results Location

After optimization completes, results will be generated in 2_results/:

File	Description	Format
study.db	Optuna database with all trials	SQLite
optimization_history.json	Full trial history	JSON
pareto_front.json	Pareto-optimal solutions	JSON
optimization.log	Execution log	Text
reports/optimization_report.md	Full Results Report	Markdown

8.1 Results Report Contents

The generated optimization_report.md will contain:

1. Optimization Summary - Best solutions, convergence status
2. Pareto Front Analysis - All non-dominated solutions with trade-off visualization
3. Parameter Correlations - Design variable vs objective relationships
4. Convergence History - Objective values over trials
5. Constraint Satisfaction - Feasibility statistics
6. Neural Surrogate Performance - Training loss, validation R², prediction accuracy
7. Algorithm Statistics - NSGA-II population diversity, hypervolume indicator
8. Recommendations - Suggested optimal configurations

---

9. Quick Start

Staged Workflow (Recommended)

# STAGE 1: DISCOVER - Clean old files, run ONE solve, discover available outputs
python run_optimization.py --discover

# STAGE 2: VALIDATE - Run single trial to validate extraction works
python run_optimization.py --validate

# STAGE 3: TEST - Run 3-trial integration test
python run_optimization.py --test

# STAGE 4: TRAIN - Collect FEA training data for neural surrogate
python run_optimization.py --train --trials 50

# STAGE 5: RUN - Official optimization
python run_optimization.py --run --trials 100

# With neural acceleration (after training)
python run_optimization.py --run --trials 100 --enable-nn --resume

Stage Descriptions

Stage	Command	Purpose	When to Use
DISCOVER	--discover	Scan model, clean files, run 1 solve, report outputs	First time setup
VALIDATE	--validate	Run 1 trial with full extraction pipeline	After discover
TEST	--test	Run 3 trials, check consistency	Before long runs
TRAIN	--train	Collect FEA data for neural network	Building surrogate
RUN	--run	Official optimization	Production runs

Additional Options

# Clean old Nastran files before any stage
python run_optimization.py --discover --clean

# Resume from existing study
python run_optimization.py --run --trials 50 --resume

# Reset study (delete database)
python reset_study.py
python reset_study.py --clean  # Also clean Nastran files

---

10. Dashboard Access

Live Monitoring

Dashboard	URL	Purpose
Atomizer Dashboard	http://localhost:3003	Live optimization monitoring, Pareto plots
Optuna Dashboard	http://localhost:8081	Trial history, hyperparameter importance

Starting Dashboards

# Start Atomizer Dashboard (from project root)
cd atomizer-dashboard/frontend && npm run dev
cd atomizer-dashboard/backend && python -m uvicorn api.main:app --port 8000

# Start Optuna Dashboard (for this study)
optuna-dashboard sqlite:///2_results/study.db --port 8081

What You'll See

Atomizer Dashboard (localhost:3003):

• Real-time Pareto front visualization
• Parallel coordinates plot for design variables
• Trial progress and success/failure rates
• Study comparison across multiple optimizations

Optuna Dashboard (localhost:8081):

• Trial history with all parameters and objectives
• Hyperparameter importance analysis
• Optimization history plots
• Slice plots for parameter sensitivity

---

11. Configuration Reference

File: 1_setup/optimization_config.json

Section	Key	Description
optimization_settings.protocol	protocol_11_multi_objective	Algorithm selection
optimization_settings.sampler	NSGAIISampler	Optuna sampler
optimization_settings.n_trials	100	Total trials
design_variables[]	[support_angle, tip_thickness]	Params to optimize
objectives[]	[stiffness, mass]	Objectives with goals
constraints[]	[mass_limit]	Constraints with thresholds
result_extraction.*	Extractor configs	How to get results
neural_acceleration.*	Neural settings	AtomizerField config
training_data_export.*	Export settings	Training data location

---

12. References

• Deb, K. et al. (2002). A Fast and Elitist Multiobjective Genetic Algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation.
• pyNastran Documentation: BDF/OP2 parsing
• Optuna Documentation: Multi-objective optimization with NSGA-II