Machine learning prediction of seismic response parameters of reinforced concrete buildings under different site and ground motion conditions

Abstract

Predicting seismic response parameters is vital for designing reinforced concrete structures, but traditional approaches can become computationally intensive when multiple seismic scenarios are evaluated. This study presents a unified machine learning framework to predict storey-level seismic responses of a 10-storey RC special moment-resisting frame building. It explicitly integrates seismic hazard features, site conditions, and response variability. A comprehensive dataset of around 1300 samples was generated using ETABS-based nonlinear analyses, considering seismic zones, soil types, and ground motion intensity measures such as peak ground acceleration, velocity, and displacement as inputs. The key response parameters analysed include storey displacement, inter-storey drift ratio, storey shear, and overturning moment. Three machine learning models, viz., Random Forest, Support Vector Regression, and Extreme Gradient Boosting, were developed and optimised via cross-validation. Among these methods, XGBoost performed the best, with test R² scores of 0.9936 for displacement, 0.9955 for inter-storey drift ratio, 0.9973 for storey shear, and 0.9941 for overturning moment. It reduced prediction errors by approximately 30% to 90% compared to RF and SVR, with the greatest improvement observed in storey shear prediction, resulting in a 21% increase in R² over SVR. These findings demonstrate that the proposed framework effectively models complex nonlinear seismic behaviour while significantly reducing computational costs. This method provides a practical, efficient tool for rapid seismic performance assessments and supports data-driven structural design aligned with Indian seismic code standards.