Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Generative Diffusion-based Downscaling for Climate

Robbie A. Watt, Laura A. Mansfield

TL;DR

This study addresses the need for high-resolution regional climate information derived from coarse-model outputs. It compares a diffusion-based generative downscaling model against a baseline U‑Net, using ERA5 data to recover $0.25\degree$ resolution from $2\degree$ inputs over the continental USA. The diffusion model yields higher fidelity, especially at fine scales, and naturally produces ensembles for probabilistic risk assessment. The work demonstrates the practical potential of diffusion-based downscaling for climate studies and outlines next steps toward applying the approach to CMIP-scale outputs, precipitation, and non-stationarity challenges.

Abstract

Downscaling, or super-resolution, provides decision-makers with detailed, high-resolution information about the potential risks and impacts of climate change, based on climate model output. Machine learning algorithms are proving themselves to be efficient and accurate approaches to downscaling. Here, we show how a generative, diffusion-based approach to downscaling gives accurate downscaled results. We focus on an idealised setting where we recover ERA5 at $0.25\degree$~resolution from coarse grained version at $2\degree$~resolution. The diffusion-based method provides superior accuracy compared to a standard U-Net, particularly at the fine scales, as highlighted by a spectral decomposition. Additionally, the generative approach provides users with a probability distribution which can be used for risk assessment. This research highlights the potential of diffusion-based downscaling techniques in providing reliable and detailed climate predictions.

Generative Diffusion-based Downscaling for Climate

TL;DR

This study addresses the need for high-resolution regional climate information derived from coarse-model outputs. It compares a diffusion-based generative downscaling model against a baseline U‑Net, using ERA5 data to recover resolution from inputs over the continental USA. The diffusion model yields higher fidelity, especially at fine scales, and naturally produces ensembles for probabilistic risk assessment. The work demonstrates the practical potential of diffusion-based downscaling for climate studies and outlines next steps toward applying the approach to CMIP-scale outputs, precipitation, and non-stationarity challenges.

Abstract

Downscaling, or super-resolution, provides decision-makers with detailed, high-resolution information about the potential risks and impacts of climate change, based on climate model output. Machine learning algorithms are proving themselves to be efficient and accurate approaches to downscaling. Here, we show how a generative, diffusion-based approach to downscaling gives accurate downscaled results. We focus on an idealised setting where we recover ERA5 at ~resolution from coarse grained version at ~resolution. The diffusion-based method provides superior accuracy compared to a standard U-Net, particularly at the fine scales, as highlighted by a spectral decomposition. Additionally, the generative approach provides users with a probability distribution which can be used for risk assessment. This research highlights the potential of diffusion-based downscaling techniques in providing reliable and detailed climate predictions.
Paper Structure (14 sections, 3 equations, 3 figures, 2 tables)

This paper contains 14 sections, 3 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Downscaling
  4. Machine learning
  5. Methods
  6. Data
  7. Baseline U-Net for downscaling
  8. Diffusion-based generative model
  9. Results
  10. Example
  11. Probabilistic Predictions
  12. Metrics
  13. Spectra
  14. Conclusions

Figures (3)

  • Figure 1: Maps comparing (a-c) coarse resolution (the input) to the fine resolution for (d-f) the truth, (g-i) the U-Net downscaled prediction and (j-l) one generated diffusion downscaled prediction, for all three variables at a single timestep. (m-n) show 1 standard deviation across a 30-member ensemble generated by diffusion.
  • Figure 2: Maps comparing the error metrics for the U-Net downscaled prediction and the diffusion downscaled prediction, for all three variables.
  • Figure 3: (a-c) Power spectrum on a log-scale for all three variables for the truth, diffusion, U-Net and linear interpolation, and (d-f) the difference between the truth and the predicted power spectra for diffusion, U-Net and linear interpolation.