Machine Learning Emulation of Precipitation from km-scale Regional Climate Simulations using a Diffusion Model

Dynamical downscaling of climate simulations to local scales is valuable for understanding climate change impacts and planning adaptation measures, but is very computationally expensive. We present CPMGEM (Convection-Permitting Model Generative EMulator) [1]: a novel application of a generative machine learning model, a diffusion model, that skilfully emulates precipitation simulations by Met Office’s UK convection-permitting model (CPM). This achieves similar results to dynamical downscaling at a fraction of the computational cost. This emulator enables stochastic generation of high-resolution (8.8km) daily-mean precipitation samples, fine enough for use in applications such as flood modelling, conditioned on coarse (60km) weather states from a general circulation model (GCM).

We trained the emulator to produce output over England and Wales, using Met Office simulations from the United Kingdom Climate Projections (UKCP) Local product, covering 1980-2080. The output precipitation has a similarly realistic spatial structure and intensity distribution to the CPM simulations. Our generative emulator outputs a well-calibrated spread of predictions and reproduces the small-scale structure and frequency of extreme intensities better than a deterministic model. We also find evidence that the emulator captures the main features of the CPM-simulated 21st century climate change but exhibits some error in the magnitude.

Potential applications include producing high-resolution precipitation predictions for large-ensemble climate simulations and predictions using inputs from different GCMs and scenarios to better sample uncertainty. We will also discuss extending the emulator to predict sub-daily precipitation by downscaling temporally as well as spatially.

[1] Addison, H., Kendon, E. J., Ravuri, S., Aitchison, L. and Watson, P.A.G., 2024. Machine learning emulation of precipitation from km-scale regional climate simulations using a diffusion model. arXiv preprint arXiv:2407.14158