Can large-ensembles downscaled using generative AI be used for climate attribution?

Attributing rainfall extremes in the context of climate change, requires climate model data run at sufficiently high resolution both in the present and pre-industrial climates. Nevertheless, climate runs for the latter do not exist at sufficiently high resolution to adequately represent rainfall extremes involving convection or over regions with complex topography due to the coarseness of the model’s spatial resolution. This limits our ability to make robust attribution statements on many types of rainfall events and their consequent flood impacts, as high resolution spatial-patterns of rainfall are also required to produce realistic flood inundation mapping for flood modelling. Furthermore, large-ensemble climate model simulations that are valuable for attribution are very expensive to run at convection permitting resolution. In this work we downscale our large ensemble of attribution runs (HadGEM3-A) providing over 5000 years of data in both present and pre-industrial climates to convection-permitting resolution for England and Wales using a generative AI approach. We use the diffusion model CPMGEM from Addison et al. (2024), trained on UK Climate model projections to map from coarse to convection-permitting resolution over England and Wales, to downscale the attribution runs. We test the ability of the diffusion model to transfer to generating high-resolution precipitation for the attribution system, in both present and pre-industrial climates. Exploratory testing and validation of the results may be able to provide answers on whether data from such ML techniques are suitable for attribution studies. If they are, the capacity to attribute rainfall extremes and flooding will increase greatly. We then intend to redo a couple of past attribution studies that used the coarser data, with the newly downscaled data from the ML model to compare results.