Adaptive simulation and data-driven hybrid modeling for predicting shear strength and failure modes of interior reinforced concrete beam-column joints

This study investigates the key factors influencing failure modes and shear strength in interior reinforced concrete beam-column joints (RCBCJs) using an integrated machine learning framework. A dataset of 200 experimental records was analyzed, with Principal Component Analysis (PCA) refining essential parameters and the Synthetic Minority Over-sampling Technique (SMOTE) addressing class imbalance through the generation of 136 synthetic instances, equilibrating the minority class to 140 samples. Five classification and regression algorithms were evaluated, with the Random Forest (RF) model demonstrating superior predictive performance. For shear strength prediction, the RF model achieved a training relative absolute error (RAE) of 0.11 and a coefficient of determination (R² = 0.99), outperforming conventional design codes (ACI 318–14, EN1998-I:2004). Testing yielded an RAE of 0.27 and R² = 0.94, demonstrating robust generalizability. In failure mode classification, the model attained 98 % training accuracy and 84 % testing accuracy, surpassing the performance of empirical code-based methods.

SHapley Additive exPlanations (SHAP) analysis revealed beam width (b_b) and column height (h_c) as the most influential factors for failure modes (mean absolute SHAP = 0.09 and 0.05). For shear strength, column height (h_c) had the highest impact (mean absolute SHAP = 76.52), followed by top (A_sb,top; 64.02) and bottom (A_sb,bot; 54.74) beam reinforcement areas. The RF model consistently surpassed existing design standards, validating its capacity to capture complex parameter interactions. To bridge research and practice, a user-friendly graphical user interface (GUI) was developed, enabling streamlined RCBCJ design optimization by integrating data-driven insights with structural engineering principles.