Generating extreme floods using multi-millennial high-resolution simulations: A proof-of-concept over the UK

A critical challenge in catastrophe modeling is the requirement for high-resolution stochastic event sets that span tens of thousands of years to accurately sample extreme tail risks. Traditional weather generators often rely on simplified statistical assumptions that struggle to maintain complex multi-variable physical consistencies and to realistically capture climate change. Conversely, dynamical downscaling of General Circulation Models (GCMs) is computationally prohibitive to generate the multi-millennial simulations covering large domains required for robust risk assessment.

We present a computationally efficient modeling chain that leverages generative diffusion models to overcome these limitations. Our framework is rooted in GCM runs, allowing it to account for climate change and capture its impacts across variables, space and varying global warming levels. Specifically, we leverage the CESM2 Single Model Initial-condition Large Ensemble (SMILE) to sample internal natural variability and generate events more extreme than in the historical record. The methodology employs an emulator based on autoregressive video diffusion to produce synthetic GCM-resolution atmospheric states (see presentation EGU26-19946), enabling us to go beyond the length of the SMILE time series. The emulated fields are processed through a diffusion model trained on reanalysis data downscaling them to ~10km resolution (see presentations EGU26-19822 and EGU26-20546). This modeling chain preserves seasonal dependencies and atmospheric patterns while providing the stable, multi-decadal sequences necessary to generate river flow time series using the process-based Wflow hydrological model (see presentation EGU26-4924).

We prove the validity of this framework over the United Kingdom, where we show it successfully reproduces event frequencies and severity, and generates convincing reconstructions of historical events. We discuss the most extreme floods of the simulations and critically assess their realism. Our modelling chain illustrates that the use of machine learning (diffusion models) enables in-house hydroclimatic modelling from the GCM to catchment scale over periods and domain sizes much larger than previously possible.