System-level seismic fragility prediction of high-speed railway track-bridge system based on deep learning

Rapid post-earthquake assessment of high-speed railway track–bridge systems (HSRTBS) is essential for restoring service and managing seismic risk. Conventional nonlinear time-history analysis (NTHA) is the standard tool for response estimation, but its high computational cost in multi-component systems makes it impractical for rapid use. To address this gap, this paper proposes a deep learning (DL) framework that predicts the seismic response and fragility of HSRTBS directly from ground-motion (GM) records. The core of the framework is the CNN–ResNet–LSTM (C-R-L) model, which takes GM time histories as input and outputs both the response time histories and the peak responses for multiple components. A Bayesian optimization (BO) algorithm tunes the key hyperparameters to improve generalization, and ablation studies quantify the contribution of each module. The trained model predicts component-level seismic responses with high accuracy: the coefficient of determination ( R 2 ) for time-history prediction reaches 0.966, and peak-response errors are mostly between 2% and 12%. Embedded in the fragility analysis pipeline, the model further reproduces the system-level fragility curves of HSRTBS. The predicted curves agree closely with those from conventional analysis, with R 2 values above 0.95 in most cases and an average error of within 5%. The proposed framework therefore preserves predictive accuracy while substantially reducing the time burden of repeated NTHA, providing a practical route to rapid post-earthquake damage and system-risk assessment of HSRTBS.