A mechanism-data hybrid pretraining-finetuning framework for rapid and reliable urban pluvial flooding prediction

Urban pluvial flooding (UPF) evolves rapidly in space and time, making physically realistic and near-real-time prediction difficult, particularly in urban settings where full-field observations are scarce. Mechanistic models can represent inundation processes with strong physical interpretability but remain computationally expensive, whereas purely data-driven models are efficient but depend heavily on large amounts of high-quality training data. To address this trade-off, this study develops a mechanism-data hybrid pretraining-finetuning framework for rapid UPF prediction. In the pretraining stage, hydrodynamic priors are learned from randomized environmental scenarios, mechanistic teacher signals, and conservation-based constraints. In the finetuning stage, limited real-world observations are used to adapt the pretrained model to local conditions. Applied to a real UPF event, the hybrid model achieved the highest site-scale simulation accuracy with NSE values above 0.9. It also produced physically coherent inundation patterns, completed a 24 h simulation 32 times faster than the mechanistic model, and remained robust under environmental and rainfall perturbations. These results demonstrate that mechanism-data hybrid learning can balance mechanistic fidelity, local observational adaptation, and computational efficiency, providing a practical route for rapid UPF prediction under limited observations.