Downscaling UK rainfall using machine-learning emulation of a convection-permitting model

Climate change is causing the intensification of rainfall extremes in the UK. Physics-based numerical simulations for creating precipitation projections are computationally expensive and must be run many times to quantify the natural variability of precipitation. Local-scale projections such as those from the Met Office's 2.2km convection-permitting model are possible but the computational expense of these simulations requires trade-offs in the duration, domain size, ensemble size and emission scenarios for which to produce projections.

Here, we apply state-of-the-art machine learning methods to predict precipitation from the 2.2km model given large-scale predictors that are represented in GCMs. By conditioning on outputs from a physical model, rainfall can be downscaled in both past and future climates. We test the extent these methods can reproduce the complex spatial and temporal structure of rainfall, with which past statistical approaches struggle. We are interested in the methods’ ability to capture the distribution of extreme rainfall and to reproduce extreme events. Our methods are neural-network-based and explore generative approaches for representing the stochastic component of high-resolution precipitation. Compared to physical models, these approaches are computationally much cheaper and have a simple interface allowing them to be used to downscale other large GCM datasets.