Highway bridges supported by the unbonded laminated rubber bearings (ULRBs) are critical components of the transportation infrastructure systems. Accurate and efficient assessment of damage states and seismic response prediction for individual bridges at regional scale is essential for evaluating seismic vulnerability and enhancing infrastructure resilience. However, such assessments are typically resource-intensive, requiring significant labor, time, and computational power. This study introduces an approach based on machine learning (ML) algorithms to develop predictive models for rapid damage states assessment and seismic response prediction, incorporating various types of input data. The ML includes five ensemble learning algorithms and three deep neural networks (DNNs), notably TabNet, a novel architecture optimized for tabular data processing. Using a comprehensive dataset generated through nonlinear time-history analyses (NTHA) of 120 ULRB-supported highway bridges subjected to 320 ground motions, a total of 38,400 data samples were obtained for model training and evaluation. Results demonstrate that TabNet outperformed other models in predicting damage states, achieving accuracies of 93.6 % and 90.5 % across two test sets. For predicting the displacement of ULRBs, CatBoost demonstrated superior performance, achieving R2 values of 0.949 and 0.905 for the two test sets. Furthermore, model interpretability was enhanced using SHAP analysis, which identified the ULRB friction coefficient, elastic stiffness, and spectral accelerations at 0.8 s and 1.0 s as key predictors for bridge damage states and seismic response. This study represents the first attempt to apply TabNet for damage assessment and response prediction in highway bridges. The results provide a valuable reference for applying ML techniques to evaluation of highway bridge performance under earthquake events.
