Rapid 2D hydrodynamic flood modeling using deep learning surrogates

Hydrodynamic flood models improve the hydrologic and hydraulic prediction of storm events. However, the computationally intensive numerical solutions required for 2D hydrodynamics have historically prevented their implementation in rapid modeling (such as flood forecasting or probabilistic modeling). This study examines whether several Deep Neural Network (DNN) architectures are suitable for optimizing flood models by solving the spatial discretizations of hydrodynamic models. Pluvial flooding events were simulated in a low-relief, high-resolution urban environment using HEC-RAS 2D. These simulations were assembled into a training set for four different DNN architectures (Feedforward DNN, Bayesian DNN, Physics-informed DNN, and LSTM). The architectures were then used to model the spatially-discretized hydrodynamics of two study areas. The DNNs were compared to the hydrodynamic flood models, and showed good capacity to simulate hydrodynamics, with a median RMSE of up to 2 mm for cell flooding depths after fine tuning. The DNNs also improved computation time significantly, being between 15.9 and 52.2 times faster than conventional hydrodynamic models. Notably, the DNNs were also up to an order of magnitude faster than a comparable GPU-optimized hydrodynamic model. Negligible differences in fitting capabilities were observed between HEC-RAS’ Full Momentum Equations and Diffusion Wave Equations once the networks were fine-tuned. Important numerical stability considerations were discovered that impact the selection of hydrodynamic formulation, DNN architecture, and forecast target. Overall, the results from this study show that DNNs can greatly optimize flood modeling, and enable rapid hydrodynamics. •Deep Neural Networks were used as surrogate models for 2D hydrodynamics.•Four architectures were trained on 2D HEC-RAS models in a low-relief urban area.•Architecture-dependent differences were found across surrogate models.•Surrogate models are faster than both conventional and GPU-optimized hydrodynamics.•Important computational stability considerations for hydrodynamics were discovered.