Explainable Ensemble Learning for Web‐Crippling Resistance of Sigma Cold‐Formed Steel Sections under Interior–Two–Flange Loading

Cold-formed steel (CFS) sigma sections are common in floors and purlins but are prone to web crippling under interior-two-flange (ITF) bearing. Equations in AISI S100, EN 1993-1-3 and related standards were mainly calibrated for channel sections and give poor, highly scattered predictions for sigma profiles. This study develops explainable machine-learning (ML) surrogates for ITF web-crippling capacity. A validated finite-element (FE) database of 648 cases was created, spanning practical ranges of thickness, bearing length, yield strength and geometry. Code equations and recent proposals showed large bias (often negative R2), whereas boosted trees (CatBoost, XGBoost, LightGBM, GBRT) and kernel methods (SVR, GPR) achieved near-perfect accuracy (R2 ≥ 0.99; RMSE ≈ 0.20–0.25 kN). SHAP analysis identifies thickness and yield strength as the dominant drivers, followed by section depth and bearing length, consistent with mechanics. Reliability calibration using the AISI S100 framework provides design-ready factors for top models (φ ≈ 0.90, Ω ≈ 1.78). Interpretable ML surrogates thus enable accurate, reliability-aligned design of sigma sections under ITF loading.