Physics‐Guided Multi‐Stage Deep Learning for Progressive Collapse Assessment of Prestressed Concrete Frames With Infill Walls

This study proposes a physics‐guided multi‐stage deep learning (PG‐MSDL) model for rapid prediction of incremental dynamic analysis (IDA) curves of prestressed concrete (PC) frames with infill walls, considering interior and exterior column removal conditions. Eighteen high‐dimensional parameters, related to physical and geometric properties, including sectional dimensions and material properties of concrete, reinforcement, steel strands, and infill walls, are used as input variables. Purely data‐driven models, such as XGBoost and Transformer, are first investigated, but their prediction accuracy is limited. By integrating physics‐based performances and finite element results, the PG‐MSDL model markedly improves the prediction of IDA curve evolution and key performance points, enhancing robustness and accuracy under highly nonlinear responses. Furthermore, a Sobol global sensitivity analysis based on the surrogate model reveals that infill wall properties dominate load‐related responses, whereas displacement responses are more sensitive to structural parameter interactions. The proposed approach provides an efficient and reliable method for predicting progressive collapse behavior of PC frames under high‐dimensional uncertainty.