Downscaling with a machine learning-based emulator of a local-scale UK climate model

Research output: Contribution to conferenceOther Conference Contribution

Abstract

High resolution rainfall projections are useful for planning for climate change [1] but are expensive to produce using physical simulations. We make novel use of a state-of-the-art generative machine learning (ML) method, diffusion models [2], to more cheaply generate high resolution (8.8km) daily mean rainfall samples over England and Wales conditioned on low resolution (60km) climate model variables. The downscaling model is trained on output from the Met Office UK convection-permitting model (CPM) [3]. We then apply it to predict high-resolution rainfall based on either coarsened CPM output or output from the Met Office HadGEM3 general circulation model (GCM). The downscaling model is stochastic and able to produce samples of high-resolution rainfall that have realistic spatial structure, which previous methods struggle to achieve. It is also easy to train and should better estimate the probability of extreme events compared to previous generative ML approaches.

The downscaling model samples match well the rainfall distribution of CPM simulation output. We use as our conditioning variables We obtained further improvements by also including high-resolution, location-specific parameters that are learnt during the ML training phase. We will discuss the challenges of applying the model trained on coarsened CPM variables to GCM variables and present results about the method’s ability to reproduce the spatial and temporal behaviour of rainfall and extreme events that are better represented in the CPM than the GCM due to the CPM’s ability to model atmospheric convection.

References

[1] Kendon, E. J. et al. (2021). Update to the UKCP Local (2.2km) projections. Science report, Met Office Hadley Centre, Exeter, UK. [Online]. Available: https://www.metoffice.gov.uk/pub/data/weather/uk/ukcp18/science-reports/ukcp18_local_update_report_2021.pdf

[2] Song, Y. et al. (2021). Score-Based Generative Modeling through Stochastic Differential Equations. ICLR.

[3] Met Office Hadley Centre. (2019). UKCP18 Local Projections at 2.2km Resolution for 1980-2080, Centre for Environmental Data Analysis. [Online]. Available: https://catalogue.ceda.ac.uk/uuid/d5822183143c4011a2bb304ee7c0baf7
Original languageEnglish
DOIs
Publication statusUnpublished - 28 Apr 2023

Structured keywords

  • Interactive Artificial Intelligence CDT

Fingerprint

Dive into the research topics of 'Downscaling with a machine learning-based emulator of a local-scale UK climate model'. Together they form a unique fingerprint.

Cite this