Stacked generalisation for improved prediction of joint shear in beam-column joints

Understanding shear and bond mechanisms becomes crucial in addressing the complexities and uncertainties inherent in designing beam-column joints, especially under seismic conditions. This study proposes employing machine learning as a viable approach for predicting joint shear strength. The study analyses a dataset containing 670 beam-column joints (312 interior and 358 exterior beam-column joints) statistical summaries featuring eight input parameters encompassing the joint's cross-sectional dimensions, reinforcement details, and material properties. Machine learning algorithms, including K-Nearest Neighbours, Support Vector Regressor, Decision Tree, Multilayer Perceptron, Random Forest, Extreme Gradient Boosting, Adaptive Neuro Fuzzy Inference System, Elastic Net, and Ridge, are fine-tuned using Optuna and evaluated based on performance metrics such as Root Mean Squared Error and R-squared. The study evaluates the performance of machine learning models for predicting joint shear strength in interior (IBCJ) and exterior (EBCJ) beam-column joints, emphasizing the effectiveness of ensemble techniques like stacking regressors. The Stacking Regressor consistently outperformed traditional models and design codes (CSA, ACI, AIJ, IS, GB, EN, and NZS), achieving RMSE values of 1.1407–1.2170 (R²: 0.8180–0.84) for IBCJ and RMSE of 1.02 (R²: 0.84) for EBCJ, compared to code errors exceeding 1.8. SHAP analysis revealed that concrete compressive strength (importance: 0.85 for IBCJ, 0.89 for EBCJ), column reinforcement percentage, and top reinforcement percentage were the most influential features. These findings highlight the potential of ML models to capture complex, non-linear structural behaviour more accurately than conventional methods. Keywords: Beam-Column Joint, Joint Shear Stress, Ensemble techniques, Feature Importance, Stacking Regressor, Algorithms, Machine learning