A physics-guided deep learning method for construction-induced settlement prediction during pipe-roof underpassing of an operating high-speed railway

Under ultra-shallow-buried, water-rich soft ground conditions, deformation control is highly sensitive when the pipe-roof method is used to underpass an operating high-speed railway. Excavation disturbance may induce coupled settlement of the tunnel crown, ground surface, and existing track, directly threatening the operational safety of the high-speed railway. To address the limited availability of field monitoring data, the significant spatiotemporal coupling of settlement responses, and the discrepancy between numerical simulations and actual engineering conditions, a physics-guided deep learning method for construction-induced settlement prediction, termed PG-STGAT, is proposed. In this method, crown, ground, and rail monitoring points are represented as graph nodes, and the graph structure is constructed based on spatial proximity and deformation-transfer mechanisms. A temporal convolutional network and a graph attention network are adopted to extract the temporal evolution and spatial correlation features of settlement, while data fitting, temporal continuity, spatial coordination, and deformation-transfer consistency constraints are introduced into the loss function. Nine sets of FLAC3D simulation data under multiple working conditions are used for pretraining, followed by fine-tuning with field monitoring data to achieve transfer from simulation data to real engineering data. The engineering case results show that PG-STGAT achieves RMSE, MAE, MSE, WMAPE, and R 2 values of 0.1048 mm, 0.0614 mm, 0.0110 mm 2 , 7.97%, and 0.9869 on the test set, respectively. Ablation analysis verifies the effectiveness of physics guidance, graph attention, and temporal feature extraction. With only 5% of the field monitoring data used for fine-tuning, the R 2 remains 0.9129, and the model maintains good predictive performance under 20% noise. Model interpretability analysis further reveals the information-transfer relationships among crown, ground, and rail settlement and verifies the physical consistency of the prediction results. The proposed data-physics fusion approach provides a potentially effective tool for accurate prediction of construction-induced settlement during pipe-roof underpassing of an operating high-speed railway.