Physics-informed Graph Attention-Transformer networks for multi-scale residual displacement prediction in progressive collapse of RC frames

Nonlinear dynamic finite element analysis (FEA) is the mainstream method for assessing progressive collapse of structures, but its prohibitive computational cost limits rapid evaluation of damaged building systems. This paper proposes a physics-informed Graph Attention-Transformer (GAT-Transformer) framework as an engineering informatics surrogate for predicting residual displacements of reinforced concrete (RC) frames under progressive collapse. The framework transforms structural layouts, initial column-removal scenarios, and physical attributes into graph-based representations, in which beam-column joints are represented as graph nodes, physical beams and columns as graph edges, and slab-aware diagonal connections as virtual edges encoding in-plane slab interactions. A local Graph Attention Network captures neighborhood response redistribution, while a global Graph Transformer with feature bypass learns long-range structural interactions and preserves macro-level descriptors. To address the scale imbalance between horizontal and vertical responses, a decoupled multi-head regression scheme is introduced for multi-scale displacement prediction. Using 9968 parametric cases generated from high-fidelity simulations, the model achieves R 2 = 0.951 on the validation set with an overall MAE of 18.2 mm, reducing MAE by 24.6% relative to GraphGPS. Topology-disjoint testing and extrapolation results further demonstrate spatial generalization across unseen structural configurations and extended structural scales. Validation against a full-scale two-story RC-frame blast test shows good agreement with a relative error of 11.5% in residual displacement. The developed model completes a single-case inference within 5.2 ms and substantially accelerates progressive-collapse response assessment.