A Hybrid Machine Learning Approach for Predicting Joint Shear Capacity in Beam-Column Connections

Abstract

Accurately predicting the shear strength of beam-column connections is crucial for maintaining the structural integrity and stability of buildings, especially in seismic conditions. This study aims to address this challenge by developing and evaluating multiple machine learning regression models for estimating joint shear capacity. A dataset consisting of 445 beam-column connections with 17 key influencing variables was compiled and used to train seven distinct regression models. Among them, the four best-performing models—Quadratic Support Vector Machine (QSVM), Rational Quadratic Gaussian Process Regression (RQGPR), Kernel Ridge Regression (KRR), and Ensemble Boosting (EB)—were selected based on their predictive accuracy. To further enhance performance, these models were combined into a hybrid ensemble model, capitalizing on their complementary strengths to improve shear strength estimation. The hybrid model exhibited superior predictive performance, achieving a test RMSE of 0.0246 and an R2 value of 0.9605, significantly surpassing the accuracy of the best standalone model (RQGPR). This reinforces the advantage of ensemble learning in minimizing error and enhancing generalization. The findings of this research highlight the growing role of machine learning in structural engineering, particularly in advancing shear strength prediction methodologies. By demonstrating that a hybrid model can outperform traditional single-model approaches, this study provides valuable insights for developing safer, more resilient structures and optimizing modern engineering practices with artificial intelligence. © The Author(s), under exclusive licence to Shiraz University 2025.