Physics-informed data augmentation and explainable machine learning for parametric design of stainless steel columns

Compared with code-based design, data-driven prediction learns context-specific response patterns from tests and simulations, offering finer-grained and less conservative estimates. Reliable data-driven prediction in structural engineering is often hampered by limited and imbalanced datasets as well as the weak coupling between learned models and practical tools. This work presents an integrated framework that combines physics-informed generative data augmentation, Bayesian-optimised machine learning, model interpretability, and parametric design for stainless steel tubular columns. The generative stage is two-fold: a conditional model synthesizes feasible feature vectors under design labels, after which a physics-aware residual is generated under the joint context of the features and the Eurocode baseline. A post-generation rebalancing enforces adequate representation across cross-section classes and non-dimensional slenderness bands. Tree-structured Parzen Estimator optimisation is used to tune both the physics-informed CTGAN/CTVAE generators and the gradient-boosted learners (XGBoost, LightGBM). Across multiple train/test settings, the augmented learners consistently improve predictive fidelity and robustness relative to Eurocode formulas and to models trained without augmentation. SHAP analyses confirm that learned effects are congruent with mechanics and remain stable across datasets and algorithms. The tuned predictor is embedded in a Rhino-Grasshopper component that provides real-time what-if exploration and side-by-side comparison with the Eurocode baseline inside a familiar parametric workflow. The study establishes a reproducible pathway for combining mechanics-aware augmentation, explainable learning, and CAD-centric deployment to support reliable design under data scarcity.