A Graph Network-Based Real-Time Precise Predicting Model for Progressive Collapse of Steel Frame Structures

Progressive collapse is initiated by local component failure and evolves into global structural instability through continuous damage propagation. Rapid prediction of the structural response before an actual collapse-inducing event occurs is important for collapse-resistant design, pre-disaster consequence assessment, and safety evaluation. However, challenges remain in predicting three-dimensional and multi-component collapse scenarios, including geometry representation, topology changes, and generalization across different failure conditions. To address these issues, this study proposes a graph neural network model for fast prediction of progressive collapse in steel-framed buildings after component failure. The model adopts a graph encoder-decoder architecture to represent component geometry and structural connectivity, employs multi-scale message passing to infer global damage patterns from local failures, and incorporates a multi-source feature joint training strategy for single-, double-, and multi-column failure scenarios. Validation results show that the model achieves 94% accuracy compared with simulation data and a 6900-fold acceleration, enabling real-time collapse prediction for pre-disaster consequence forecasting. This approach improves the balance between simulation accuracy and efficiency and provides a new pathway for large-scale collapse analysis.