Probabilistic flood modelling with multi-scale hydraulic-based graph neural networks

Traditional flood modelling approaches, which rely on deterministic methods, often fail to account for the inherent uncertainties of flood events, such as discharge estimates or flood defence parameters. Probabilistic flood modelling addresses this gap by quantifying the likelihood of various scenarios, based on the probability of occurrence of its inputs. However, the large number of required numerical simulations makes this framework computationally expensive. In this study, we explore the use of a multi-scale graph neural networks inspired by finite volume methods to accelerate probabilistic flood simulations. This approach is applied to several dike rings in the Netherlands - regions enclosed by levees - considering uncertainties in dike breach locations and inflow discharges. To improve the reliability of the model, we select among the output simulations only the ones that approximately preserve the total flood volumes over time, as calculated from the inflow boundary conditions. Our model generates thousands of flood scenarios with orders-of-magnitude speedups compared to traditional methods. The resulting output maps provide the expected frequencies of inundation extents for specific water depths, offering a robust tool for efficient and comprehensive flood risk assessment.