A Cascaded Classification-Regression Framework for Shear Strength Prediction of Cold-Formed Steel Screw Connections

Existing AISI S100 provisions for cold-formed steel (CFS) screw connections lack codified strength equations for screw shear and net section fracture, and traditional machine learning (ML) models struggle to predict these minority failure modes due to imbalanced experimental datasets. This study proposes a cascaded ML framework that first classifies the failure mode and then predicts strength using mode-specific regressors. Two cascade strategies are evaluated: a Hard Classification Cascade (HC-C) and a novel Probability-Weighted Cascade (PW-C) that weights predictions by class probabilities to mitigate error propagation from misclassification. The predictive performance of the two cascaded models is benchmarked against a single regressor without classification. The superior PW-C model is then compared with AISI S100, and its resistance factor ϕ is subsequently calibrated in accordance with LRFD. Results show that the proposed cascaded models outperform the direct regression model, with PW-C improving the R² for minority-class screw shear from 0.765 to 0.933 and for net section fracture from 0.784 to 0.912. Compared with AISI S100 provisions, PW-C extends coverage to the currently unaddressed failure modes and effectively captures screw group effects on shear strength based on a database of 564 tests. Reliability analysis yields an overall ϕ_c of 0.64 for the PW-C model, with a recommended divisor of 1.15 for direct application within the AISI design framework. This work provides a practical, data-driven pathway for updating design codes to cover failure modes beyond current specification limits.