Author(s): Roberto Bentivoglio; Elvin Isufi; Sebastiaan Nicolas Jonkman; Riccardo Taormina
Linked Author(s):
Keywords: No Keywords
Abstract: Numerical methods for flood modelling are accurate but computationally expensive. Surrogate models based on deep learning emerged as alternative solutions to speed up flood mapping while keeping accurate results. In particular, hydraulic-based graph neural networks seem a promising way ahead as they showed transferability to domains where they were not trained on. However, they have been so far corroborated only for regular meshes and they also require initial conditions from the numerical solver. In this paper, we aim to overcome these two issues and propose an extension of hydraulic-based graph neural networks to time-varying boundary conditions. This is achieved by using ghost cells, which enforce the solution at the domain’s boundary. We validate the proposed method on a dataset of irregular meshes, over varying topographies and input hydrograph discharges. The results show that the model can replicate the overall dynamics of the flood evolution over unseen meshes, topographies, and boundary conditions, without any input from the numerical model, opening up possibilities for simulating also realistic case studies.
Year: 2024