Interpretable machine learning for predicting reinforced concrete beam column joint shear strength

Abstract Accurate prediction of RC beam–column joint shear strength is essential for seismic design and performance-based assessment; however, existing design codes are primarily based on simplified empirical formulations that inadequately capture the complex interactions among joint geometry, material properties, reinforcement detailing, and axial load effects. This study presents a comprehensive ML framework for predicting joint shear strength using a curated database of 194 experimentally tested RC beam–column joints compiled from the literature. Ten key input variables representing joint geometry, concrete strength, reinforcement indices, and axial load were adopted to model the experimental joint shear capacity (V Exp ). Five ML algorithms LR, ANN, KNN, RF, and SVM were developed and evaluated using an 80/20 training–testing split and tenfold CV. Among the investigated models, RF demonstrated superior predictive performance and robustness, achieving the highest accuracy (R 2 = 0.978, RMSE = 62.27 kN) with stable generalization across validation schemes. Model transparency was ensured through SHAP and PDPs, which identified joint dimensions and reinforcement-related parameters as the dominant contributors to shear resistance. In addition, a practical RF-based graphical user interface (GUI) was developed to facilitate real-time prediction and engineering implementation. Comparative evaluation against major international design codes confirmed the improved accuracy and reduced bias of the proposed ML approach.