Explainable machine learning surrogates for structural simulation: benchmark datasets and code

Parameterized benchmark datasets and the complete analysis code for a study of nine regression models used as surrogates for finite-element structural simulation (cantilever beam, fixed-base portal frame, six-panel Pratt truss). Includes solver verification against closed-form solutions, stratified splits, 10-fold cross-validation, held-out and extrapolation evaluation, SHAP interpretability analysis compared with solver-based sensitivity indices, Friedman and Holm-corrected statistical comparisons, learning curves, a solver-versus-surrogate timing benchmark, and a 100,000-candidate design-screening study with conformal uncertainty intervals.