Prediction of Beam-Column Joint Shear Strength Using Machine Learning and a Proposed Design Equation

Authors: Tufail Mabood
Journal:  Emirati Journal of Civil Engineering and Applications
Volume: Vol 4 Issue 1
Keywords: Beam–Column Joint; Joint Shear Strength; Machine Learning Models; Gradient Boosting; Random Forest; Stacked Ensemble Learning; Reinforced Concrete (RC) Structures.

Abstract:

Accurate prediction of shear strength in reinforced concrete beam–column joints is essential for seismic design, yet traditional code-based equations, such as those in ACI 318M-14, often oversimplify complex interactions among geometric, material, and reinforcement parameters, leading to conservatism and high scatter. This study develops three machine learning models—Gradient Boosting, Random Forest, and a Stacked Ensemble—using a dataset of 98 experimental specimens with inputs including concrete compressive strength, joint dimensions, reinforcement ratios, axial load, and eccentricity. Gradient Boosting achieved the highest accuracy (R² = 0.994, COV = 1.95%), followed by the Stacked Ensemble (R² = 0.986), while Random Forest showed greater variability (R² = 0.827). Correlation analysis, SHAP values, and feature importance rankings confirmed the dominant roles of concrete strength, beam reinforcement ratio, and column dimensions, aligning with mechanical principles of strut capacity and confinement. A novel empirical equation was calibrated from these insights, offering improved accuracy over balanced ACI provisions. Target permutation validation verified meaningful physical learning. The findings demonstrate the superiority of machine learning for capturing nonlinear joint behavior and provide a more reliable basis for design and retrofitting.