Seismic performance prediction and interpretation of RC bridges under vehicle-bridge interaction: From VBI system simulation to ensemble learning surrogate models

Abstract

To accurately predict the dynamic response of vehicle-bridge interaction (VBI) systems under seismic excitations, this study proposes a framework based on LightGBM (LGBM) for dynamic response prediction. The framework is enhanced with Bayesian optimization and interpretability analysis for model improvement and feature understanding, through thousands of nonlinear time - history analyses on a typical Chinese RC T-beam bridge, a comprehensive VBI model is established. It accounts for vehicle dynamic response on uneven bridge surfaces and the impact of random traffic loads on bridge response. For the non - stationary distribution of random traffic loads and traffic data growth, the study uses the generalized extreme value distribution to model the extreme values of stationary vehicle loads in each time interval. This discretizes the continuous stochastic process into a combination of stationary processes in the time domain, enabling non - stationary evaluation of vehicle load effects. Results indicate that the proposed LGBM model achieves superior accuracy in predicting six key dynamic response, with R² values exceeding 92 %, significantly outperforming traditional models such as RF, GBDT, and SVM. This demonstrates higher predictive accuracy and computational efficiency, providing a reliable tool for effectively predicting dynamic responses in complex bridge systems under seismic actions. Additionally, Bayesian optimization increased hyperparameter tuning efficiency by 6–8 times compared to grid and random search methods, yielding better prediction results. Feature importance analysis shows that seismic intensity-related input variables (e.g., Sa₁₀) dominate pier drift ratio and deck displacement predictions, while geometric parameters such as pier height significantly influence overall demand estimation, highlighting their importance under varying structural conditions. SHAP analysis reveals negative effects of some geometric features (e.g., pier height H_c and span length L_m) on seismic demand, while road quality and vehicle stiffness are more influential among vehicle parameters, with vehicle speed and weight contributing less. Feature interaction analysis shows notable nonlinear interactions, such as between lateral shear displacement and pier height, significantly affecting abutment displacement prediction. These findings reveal coupled effects under VBI, providing deeper insights for seismic design optimization.