AI based seasonal large ensembles for fluvial flood risk: Evaluation over the Elbe basin

Abstract. A key challenge in risk analysis is to identify hazard events that are plausible and yet extrapolate beyond historical observations with appropriate frequency. For flood risk management, this can be done with large ensembles of synthetic, physically plausible weather scenarios that extend beyond the historical record to sample low-likelihood, high-impact events. Traditional statistical approaches for synthetic weather generation are often limited in variability and physical realism. Here, we show for the first time that a machine learning weather prognostic model, combined with a diagnostic precipitation model, can generate seasonal-scale ensembles suitable for flood risk assessment. Specifically, we adapt the huge ensembles (HENS) approach using a Spherical Fourier Neural Operator (SFNO)-based model combined with an Adaptive Fourier Neural Operator (AFNO)-based diagnostic precipitation model, using Nvidia Earth-2 stack, in a framework which we call “PrecipHENS”, to produce > 1000 synthetic European winter seasons of precipitation and temperature at 0.25° resolution in 112 GPU hours on NVIDIA L40s GPUs. In an Elbe River case study, PrecipHENS reproduces key features of the precipitation and temperature climatology, preserves spatial and temporal dependence – including decay of extremal co‑occurrence with distance – and generates a wider diversity of extreme events than an industry-standard conditional multivariate extreme value model benchmark. Principal component analysis of extreme precipitation fields shows that PrecipHENS spans a much broader space of storm structures (≈ 81 % of 1 × 1 grid cells) than the benchmark (≈ 50 %) or the historical record (≈ 19 %), indicating plausible novelty rather than repetition of past patterns. Coupled with a hydrological model, the AI-generated weather sequences produce river flow simulations consistent with historical climatology and extreme discharge patterns. These results demonstrate the potential of AI-based weather models to support event set generation for flood hazard and risk applications. Beyond flood risk, such AI-based large-ensemble weather generation offers a general framework for applications that benefit from expanding the physically plausible sample space, including risk assessment, climate-impact analysis, storyline development and statistical characterisation of extremes.