A Transformer–GNN Hybrid Framework for Predicting Structural Nonlinear Responses in Urban-Scale Seismic Performance Assessment

The rapid and accurate computation of seismic responses for large-scale urban building clusters is crucial for post-earthquake damage assessment, emergency response, disaster scenario simulation, and pre-disaster planning. However, traditional numerical approaches encounter the long-standing accuracy-scalability trade-off, which limits their applicability in urban-scale analyses. To overcome this limitation, a Transformer-Graph Neural Network (GNN) hybrid framework is developed for efficient prediction of nonlinear seismic response time histories of shear wall structures. The framework consists of three tightly integrated modules: a GNN-based structural property encoder, a Transformer-based ground motion encoder, and an autoregressive response decoder with cross-attention fusion. This architecture effectively integrates static structural attributes and dynamic ground motion features, enabling the model to capture their nonlinear coupling and temporal evolution during seismic excitation. The model is trained and evaluated on a large-scale dataset comprising approximately 2.3 million samples. On the validation set and the test set, the model achieves overall R² values of 0.93 and 0.90, respectively, and mean relative errors of 15.85% and 18.83% for peak inter-story drift prediction. Moreover, the model achieves an effective per-sample latency of 0.35 ms and a speedup of (Formula presented.) over conventional nonlinear time-history analysis (NLTHA). The proposed framework is scalable to urban-scale building clusters, enabling near-real-time seismic response computation and supporting rapid post-earthquake loss assessment.