Author(s): Juan-Manuel Perez-Garcia-De-Carellan; Mario Morales-Hernandez; Pilar Garcia-Navarro
Linked Author(s): Pilar García-Navarro
Keywords: Deep learning; Flood modelling; Hydrodynamic simulation; U-Net; Super-resolution
Abstract: High-resolution flood simulations are critical for risk management, yet conventional solvers remain computationally prohibitive for real-time applications. Addressing this via deep learning-based super-resolution, we extend [3] by coupling the TRITON simulator with a U-Net architecture. The model, trained on paired constant-rainfall simulations using coarse maps and fine topography, successfully generalizes to realistic time-varying events. Results demonstrate a 4x super-resolution enhancement with ~1 cm mean errors, validating the approach as a real-time surrogate model. Furthermore, we conduct sensitivity analyses on architectural modifications, loss functions, training data volume, and domain decomposition strategies to optimize performance, highlighting the potential of this hybrid approach for operational flood management.
Year: 2026