Transfer learning-enhanced neural networks for seismic response prediction of high-speed railway simply supported bridges

The seismic analysis of Chinese high-speed railway bridge-track system (i.e., CRTS II ballastless track structure) is crucial for assessing vehicle operational safety and facilitating the seismic design of bridge bearings and piers. Currently, the design code utilizes bridge models for seismic design but neglects the influence of the track structure situated above the bridge, thereby overlooking the vulnerability of components in the track structure. Developing full models of bridge-track systems would significantly increase computational intensity and time costs, especially for assessing the seismic performance of high-speed railway bridge lines. To tackle this issue, this paper introduces a transfer learning-enhanced neural network to rapidly predict the seismic responses of bridge-track systems with limited labeled data. Pairs of bridge models, one with and one without the presence of track structure, are developed to establish the relationship of seismic responses between bridge-only and bridge-track system models. The seismic responses derived from bridge models are utilized as input, while seismic responses from bridge-track system models serve as output for training gated recurrent unit neural networks. Transfer learning techniques, based on Maximum Mean Discrepancy (MMD), are employed to facilitate feature transfer between various high-speed railway bridge-track systems with varying spans, pier heights, and different types of bearings. The application of transfer learning significantly decreases data acquisition costs while improving the predictive accuracy of neural networks. Analysis results indicate that the proposed framework displays strong generalizability across new models and is both computationally efficient and effective in predicting the seismic responses of high-speed railway bridge-track systems. This method provides an alternative for rapidly evaluating the seismic performance of high-speed railway bridge lines. • A transfer learning-based neural network approach predicts seismic responses of bridge-track systems with limited data. • Transfer learning using Maximum Mean Discrepancy reduces data acquisition costs and enhances neural network performance. • Seismic analysis of high-speed railway bridge-track systems with varying spans, pier heights, and bearings was conducted.