A two-stage physics-driven graph neural network framework for predicting tunnelling-induced deformation of frame structures

Shield tunnelling-induced ground deformation can adversely affect adjacent buildings. Rapid and physically rational prediction of structural deformation responses remains a critical issue for risk control in underground engineering. This study proposes an integrated data-mechanics framework for predicting tunnelling-induced deformation of adjacent frame structures based on centrifuge test data.First, a hybrid neural network-radial basis function model (NN-RBF) is developed to predict buildingaffected vertical settlement and horizontal displacement profiles. An elastic foundation beam model is then introduced to evaluate the base deformation response from the reconstructed ground displacement profiles. Furthermore, a physics-driven graph neural network (TMP-GNN) based on Timoshenko beam theory and the principle of minimum potential energy is constructed to predict the nodal displacement field of frame structures. The results show that the NN-RBF model accurately characterizes complex ground displacement patterns, while the elastic foundation beam model reasonably captures the primary deformation characteristics of the foundation slab. The TMP-GNN can be trained without nodal displacement labels and reconstructs overall deformed configurations under different soil-structure interaction conditions. By combining data-driven modelling with mechanics-based constraints, the proposed method provides a topology-aware and physicsconstrained approach for rapid assessment of tunnelling-induced building deformation.