A new three-dimensional model of train-track-bridge coupled system based on meshless method and its graph neural network-based surrogate model

Abstract

A model of train–track–bridge coupled system is proposed to study the interactions between structures in greater detail. The new model employs a meshless method to numerically simulate the box girder bridge and track slab. In the dynamic analysis, the system at each time step is abstracted into a graph structure and trained using a graph neural network to develop a surrogate prediction model. The graph neural network node connections in the bridge top plate are determined by the meshless method. Multiple numerical examples demonstrate the differences in structural response between the proposed model and the conventional model and evaluate the performance and self-evolutionary capabilities of the surrogate model. The results indicate that, compared to the proposed model, the conventional model underestimates vertical responses by approximately 17 %–69 % and lateral responses by one to two orders of magnitude. The surrogate model demonstrates good displacement prediction capabilities for the bridge on the training dataset, achieving an $R^2$ value as high as 0.99. Furthermore, it exhibits robust prediction and self-evolutionary capabilities on the test dataset under topological changes, with prediction accuracy decreasing by only about 2 %. However, the prediction performance for rail responses is relatively poor, with an $R^2$ value as low as 0.29.