PMBRA surrogate-assisted optimization

Reproducibility package for the manuscript “Surrogate-Assisted Multi-objective Structural Optimization of a Permanent-Magnet Bridged Reluctance Actuator.”

Repository contents

• data/: the 4,096-case full-factorial FEA table, the 168-point intermediate grid, the seven newly simulated air-gap points within the 15-point sweep, and the available Pareto-region FEA validation workbooks.
• dnn/: DNN definition, training, cross-validation, intermediate-grid validation, air-gap validation, configurations, archived results, and preserved split information.
• surrogate_baselines/: linear, ridge, quadratic, decision-tree, and k-nearest-neighbor surrogate comparison material.
• benchmark/: NSGA-II, DN-NSGA-II, SPD-DN-NSGA-II, CEC2019 functions/reference data, and the deterministic rerun schedule.
• ablation/: SPD-off, constant-JR, and q = 64, 512, and 1024 ablation settings with the archived run-level and summary results.
• statistics/: run-level metrics and Mann–Whitney, Holm, Vargha–Delaney, Friedman, and Wilcoxon outputs.
• PMBRA_optimization/: optimization inputs, raw Pareto results, constraint results, and direct FEA validation files.
• figures/: source data grouped for Figs. 6–12.
• maxwell/: the ANSYS Maxwell project and run notes.
• environment/: recorded Python, MATLAB, and ANSYS versions.

See MANIFEST.md for the file-level mapping and checksums.txt for SHA-256 integrity values.

Python environment

python -m pip install -r requirements.txt

The recorded core environment is Python 3.12.13, TensorFlow 2.21.0, NumPy 2.5.2, and pandas 3.0.5. The DNN validation protocol uses an 80/20 hold-out split with seed 42, shuffled five-fold cross-validation with split seed 20260826, and network seeds 7, 21, 42, 84, and 126.

The archived DNN implementation combines hold-out, five-fold, and 168-point intermediate-grid validation in one pipeline. Therefore, dnn/cross_validation.py and dnn/intermediate_validation.py preserve the same complete validation script under task-oriented filenames.

Benchmark protocol

The CEC2019 comparison uses 22 problems, 31 runs, population size 100*n_var, and a total evaluation budget of 10000*n_var, including initial-population evaluations. The supplied rerun seed schedule starts at 20260826. The original RNG states of the archived historical runs were not preserved; the schedule applies to new reruns.

Run the principal reproducibility workflows from the repository root:

python dnn/train_dnn.py
python surrogate_baselines/run_surrogate_baselines.py
python statistics/reproduce_statistics.py
python tools/audit_repository.py

run('benchmark/CEC2019/run_cec2019_comparison.m')
run_ablation('smoke')   % after addpath('ablation')
run_ablation('paper')   % full 31-run study

The DNN scripts train on the signed Maxwell Force_y output; reported q_F is its positive magnitude. Since the model depth is 1 m, the numerical magnitude in kN equals q_F in kN/m. Actuator thrust is then F [N] = q_F [kN/m] * (0.2*N_s) [mm].

For release review, python tools/audit_repository.py --strict intentionally returns exit code 2 while the scientific limitations below remain. Regenerate SHA-256 values after any release edit with python tools/generate_checksums.py.

Known archive limitations

• maxwell/Project.aedt stores default I = 5 A.
• No Project.aedtresults cache or solver-profile/convergence export is included, so the archived mesh/pass/runtime claims cannot be independently checked from this package.
• The unrounded raw 16-row direct-FEA export is unavailable. published_validation_points_16.csv contains only the displayed values from manuscript Tables 13 and 14.
• The manuscript ablation uses a fixed MaxFE = 10000 for every problem. The archived rows contain 10,000 evaluations for two-variable cases but 15,000 for three-variable cases, so the archived ablation is not equal-budget. The corrected script passes an exact MaxFE override and stops every paper-mode run at 10,000. The pre-audit script also configured the constant-JR variant as 3/8, whereas the manuscript states JR = 0.1.
• Recalculation from statistics/raw_run_metrics.csv gives the reported rPSP, IGDX, and IGDF suite tests, but gives rHV Friedman chi^2 = 15.6364, p = 0.0004024, not the manuscript's chi^2 = 28.78, p = 5.64e-7.
• Source scripts for the PMBRA-specific optimization, Figs. 8--12, and the 100-repeat timing experiment are not included; only their available data/products and protocol metadata are archived.

Data boundary

The repository excludes solved *.aedtresults caches, temporary files, manuscript sources, and local virtual environments. Third-party CEC2019 materials retain their original provenance and terms. See AUDIT_REPORT.md for the release audit and unresolved blockers.