Mechanics-Aware Machine Learning Classification of Reinforced Concrete Shear Wall Failure Modes Using Dimensional and Non-Dimensional Features

Abstract

Accurate classification of failure modes in reinforced concrete (RC) shear walls is essential for performance-based seismic design, yet current code-based criteria rely mainly on simple geometric indicators and cannot fully capture the nonlinear interaction among geometry, axial load, reinforcement, and boundary confinement. This study develops a mechanics-aware machine learning (ML) framework to classify four experimentally observed failure modes flexural (F), flexure shear (FS), shear (S), and sliding (SL) using both dimensional and non-dimensional feature spaces. The experimental data used in this study are analysed through two complementary representations: a dimensional dataset (CSDL1, 435 specimens) based on conventional geometric and material parameters, and a non-dimensional, mechanics-informed dataset (CSDL2, 569 specimens) expressed in terms of physically motivated ratios (A_b/A_g, l_w/t_w, P/(f_c′A_g), ρ_fy/f_c′), both constructed from the same body of published experimental studies. Three representative tree-based algorithms—Decision Tree, Random Forest, and XGBoost were trained and evaluated using 70/30 train–test splits, 10-fold cross-validation, confusion matrices, and AUC–ROC metrics. All models achieved strong multi-class performance with average AUC values above 0.85. Random Forest provided the most stable generalisation across both feature representations, while XGBoost attained comparable accuracy. More importantly, the non-dimensional feature space enhanced physical interpretability: SHAP analysis consistently identified the boundary-to-gross area ratio (A_b/A_g) and wall slenderness (l_w/t_w) as the dominant predictors for all failure modes, followed by reinforcement and axial-load ratios. The ML-derived transition thresholds (A_b/A_g ≈ 0.08–0.12, l_w/t_w ≈ 8–12, M/(Vl_w) ≈ 0.4, and P/(f_c′A_g) ≈ 0.1) align well with conceptual limits in ACI 318 and Eurocode 8. The study therefore demonstrates that mechanics-informed, non-dimensional features not only sustain high predictive accuracy but also recover the underlying physics of wall behaviour, enabling design-oriented summary tables and a quick failure-mode classification checklist for practical seismic design and assessment.