From Inundation to Systemic Collapse: A Dynamic Flood Risk Assessment Framework Coupling Hydro-Deep Learning with Cascading Failure Analysis

With the increasing complexity of urban systems, traditional flood risk assessments often fail to capture the systemic vulnerability arising from infrastructure interdependencies. This study integrates high-fidelity hydrodynamic modeling (HiPIMS), physics-guided deep learning, and complex network theory to develop a novel dynamic flood chain risk assessment framework. To enhance hazard prediction, a Physically-Guided Spatiotemporal Mixture-of-Experts (PG-ST-MoE) network is constructed. By leveraging hydrodynamic outputs as guidance, this network dynamically predicts high-precision spatiotemporal flood probabilities, effectively bridging the gap between idealized physical simulations and real-world flood occurrences. Crucially, the framework transcends static hazard mapping to analyze disaster chain effects. By simulating cascading failures within infrastructure-community networks, it quantifies how localized physical damage propagates into widespread functional paralysis and identifies functional islands where critical services are severed despite the absence of direct flooding. The framework has been deployed in the San Isabel Basin, South America, demonstrating the capability to reveal hidden systemic risks in data-scarce regions. This study offers a paradigm shift from static exposure assessment to dynamic chain-reaction analysis, providing actionable insights for preventing systemic collapse and enhancing adaptive emergency management.