Reproduce “Machine learning emulation of a local-scale UK climate model” on OpenResearch

About this paper

Henry Addison, Elizabeth Kendon, Suman Ravuri, Laurence Aitchison, Peter AG Watson

The authors develop a probabilistic model generating rainfall samples from coarse weather conditions. A conditional score-based diffusion model is trained on coarsened convection-permitting model rainfall and can use global climate model inputs. With learned location-specific features, mean-normalized bias is within about 10% of target mean per pixel on an 8.8-kilometer grid; this indicates distributional agreement, not event prediction or cost savings.

Key insights

• Diffusion models effectively generate realistic, diverse, and sharp high-resolution rainfall samples, demonstrating their suitability for emulating convection-permitting model outputs in climate downscaling.
• The inclusion of self-learnt, location-specific parameters significantly improves model performance, notably by reducing spatial biases to generally within 10% of the target mean at each pixel.
• While trained on coarsened CPM data, the model can be applied to GCM inputs, but this transfer may introduce new regional biases, highlighting the need for future work in inter-model transferability.

Summary by alphaXiv.