Predictive framework for shear strength and failure modes of exterior reinforced concrete beam–column joints using machine learning

Abstract This study examines exterior reinforced concrete (RC) beam–column joints (BCJs) subjected to seismic forces, to develop a predictive framework for shear strength and failure modes. The analysis incorporates data from 277 experimental results, utilizing eight classification algorithms (logistic regression, artificial neural network, Bayesian networks, k‐nearest neighbors, random forest, Naïve Bayes, decision tree, and the RIPPER algorithm) and nine regression algorithms (ridge regression, support vector regression, LASSO regression, random forest, random tree, linear regression, artificial neural network, Gaussian processes, and meta‐classifiers [stochastic]). To address significant class imbalance, the synthetic minority over‐sampling technique was employed, increasing the number of minority instances by approximately 4417% and generating 265 synthetic instances, resulting in 271 minority instances. The random forest algorithm demonstrates high accuracy, achieving a 97% success rate in training and 90% in testing for failure mode identification, while also providing robust predictive capabilities for shear strength. SHapley Additive exPlanations analysis revealed that the tensile strength of column reinforcement and joint transverse reinforcement were the most influential features for failure modes, while bottom beam reinforcement area and column height were critical for shear strength prediction. Comparative analyses confirm the superiority of the random forest algorithm over traditional design formulas and strut‐and‐tie models. A user‐friendly graphical interface is developed to enhance the practical application of the framework, advancing the understanding and efficiency of exterior BCJ design in real‐world engineering projects.