Graph neural network–based framework for predicting seismic damage in shield tunnels with contact loss defects

Contact loss defects (CLDs) at the tunnel–soil interface can significantly affect the seismic response of shield tunnels, while conventional finite element methods (FEM) are too time-consuming for rapid decision-making. This paper proposes a framework for seismic damage prediction of shield tunnels with CLD that directly maps field-detected CLD and stratum parameters to tunnel damage distributions. The framework integrates a multilayer perceptron (MLP) and a graph neural network (GNN) to encode finite element results into graph data and learn spatial damage patterns. It is trained and validated on a large dataset of simulated seismic responses covering diverse CLD scenarios and stratum conditions, and tested on real detection cases. The model achieves an R² of 0.98, RMSE of 4.2, and MAPE of 0.08, while reducing computation time by 90-fold compared with FEM. These results demonstrate the framework's effectiveness, efficiency, and scalability for rapid post-earthquake assessment of shield tunnels.