Machine Learning Models for Predicting Seismic Response of a Novel Two-Stage Friction Pendulum Isolated Bridge Structure

Bridges are critical infrastructure, and their vulnerability to seismic events necessitates efficient methods for predicting structural responses. Traditional methods, such as Nonlinear Time History Analysis (NLTHA), are computationally intensive, time-consuming, and require an expert analyst. Additionally, innovative structural elements like Two-Stage Friction Pendulum Bearings (TSFPBs) are not readily available in commercial software. This study proposes a Machine Learning (ML) approach as a faster and simplified alternative to NLTHA for predicting the structural response of TSFPB-isolated bridge structures. A comprehensive dataset of 500,000 data points was meticulously generated through NLTHA. Force-displacement laboratory tests under various loading and displacement conditions were conducted to validate the TSFPB nonlinear element developed in OpenSees. Utilizing this extensive dataset, seven ML models were trained and evaluated using statistical key performance indicators (KPIs). Results revealed that Extreme Gradient Boosting (XGB) achieved the highest accuracy ( $\mathrm{R}^{2} =0.959$ ) on the testing dataset, while Adaptive Boosting (AB) showed the lowest accuracy ( $\mathrm{R}^{2} =0.761$ ). Furthermore, a comparison of different surrogate models was conducted for global feature interpretation, while SHAP analysis was used for local interpretation to identify the influence of input parameters on individual structural responses. A graphical user interface (GUI) was developed using the best-performing model (i.e., XGB), offering engineers an accessible tool for seismic response prediction. This study demonstrates the potential of ML as an efficient alternative to NLTHA while providing valuable insights into key parameters influencing the seismic performance of TSFPB-isolated bridges, enhancing decision-making and design optimization.