Hybrid Physics-AI Approaches for Urban Flood Prediction: GPU Hydrodynamics, Data Assimilation, and AI Surrogates

Urban pluvial flood prediction demands rapid operational response while maintaining street-scale realism amid strong urban heterogeneity and uncertain forcings. We present a suite of hybrid physics–AI developments that address this trade-off through complementary components designed for flexible coupling and discussion. First, multi-GPU-accelerated hydrodynamic modeling reduces latency, enabling city-scale, high-resolution scenario exploration. Second, to exploit sparse and heterogeneous observations (e.g., gauges and camera-derived depths), we introduce real-time data assimilation methodologies, such as particle filtering, and multivariate geostatistical data fusion via co-kriging. The latter translates limited measurements and auxiliary covariates into spatially distributed, uncertainty-aware inundation updates. In parallel, we introduce AI surrogates to complement physical modeling: a rapid emulator trained on high-fidelity physics simulations and deep-learning super-resolution methods that bridge the scale gap between coarse forcings and street-level impacts. We conclude by discussing alternative deployment pathways for these components, evaluating key trade-offs among speed, physical consistency, observational influence, and robustness under extreme events.