Flood risk from AI-based, seasonal weather forecasts for the River Elbe

Inland flooding is one of the costliest natural hazards, inflicting substantial economic and societal damage annually, with floods causing almost USD 5 billion of insured losses across Europe in 2023 alone. Due to its proximity to major cities such as Berlin and Hamburg, a significant proportion of European exposures are vulnerable to extreme events over the Elbe River catchment. These risks need to be robustly quantified both to ensure adequate societal preparedness and so that (re)insurers are sufficiently well capitalised, which highlights the need to estimate the tails of the flood risk distribution.  As fluvial flooding is driven by the frequency, duration, and intensity of weather events, standard approaches to assess flood use extreme value theory to extrapolate from observations and simulate new and unprecedented weather events and thus river response. These methods often fall short in generating spatially coherent and physically plausible weather events, particularly those that differ substantially from the historical record, limiting flood risk estimation. 

Ensembles of weather forecasts over extended lead times could offer a promising alternative to statistical extrapolation by generating a diverse set of realistic weather outcomes. While this is not computationally feasible with numerical forecast models, artificial intelligence (AI) weather models, particularly FourCastNet based on Spectral Fourier Neural Operators (FCN SFNO), can rapidly produce large ensembles of weather forecasts while maintaining stability over long lead periods. Crucially, FCN SFNO enables forecasts to decouple from their initial conditions, facilitating the generation of numerous plausible, unseen weather events.  

Leveraging NVIDIA Earth-2, a platform for developing AI augmented weather forecasting pipelines, we demonstrate the use of the FCN SFNO-based huge ensemble (HENS) pipeline to generate a counterfactual analysis of winter seasons for the Elbe basin. Our AI-driven weather simulations are integrated with hydrological models to connect the weather events and the subsequent river response. The resulting ensemble improves our estimate of present-day flood risk, driven by a wide array of physically plausible flood events that are grouped into seasonally coherent blocks. Our approach not only surpasses the limitations of standard statistical methods but also offers an efficient, scalable, and reliable framework for flood risk estimation and management globally.