Explainable machine learning models for predicting shear capacity of exterior reinforced concrete beam–column joints

The performance of reinforced concrete (RC) beam–column joints play a decisive role in how frame structures behave, especially when subjected to seismic forces. Despite this importance, the provisions found in commonly used design codes such as ACI 352R-02 and Eurocode 8 are still largely based on simplified assumptions and empirical expressions. These approaches are often unable to capture the combined influence of material characteristics, reinforcement configuration, and joint proportions, all of which interact in a highly nonlinear manner. To address these limitations, the present study employs a data driven approach by developing machine learning models aimed at estimating the shear capacity of exterior RC joints. A comprehensive open-source database containing 203 experimental tests serves as the foundation for model training. Several algorithms, including Random Forest, Extreme Gradient Boosting, and neural-network-based models, are evaluated and compared with predictions derived from existing code equations. Explainability techniques such as SHAP analysis and permutation-based importance are incorporated to clarify how the models reach their predictions and to highlight the parameters that most strongly influence joint capacity. Across all evaluations, the machine learning models show noticeably better accuracy and more reliable generalization than the traditional formulations. Variables such as column width (b c ), beam width (b b ), and column depth (h c ) consistently emerge as the dominant factors, while material properties provide secondary contributions. The results emphasize that combining ML tools with explainable AI can offer structural engineers a more transparent and robust framework for assessing RC beam–column joints and improving current design practices.