Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TAUDiff: Highly efficient kilometer-scale downscaling using generative diffusion models

Rahul Sundar, Yucong Hu, Nishant Parashar, Antoine Blanchard, Boyko Dodov

TL;DR

TAUDiff tackles spectral bias in climate-downscaling by coupling a deterministic mean-field video-prediction model with a compact diffusion-based correction, enabling efficient and dynamically consistent wind-field downscaling at kilometer scales. The method is trained on ERA5 reanalysis and evaluated against baselines, including a full end-to-end diffusion model, showing superior recovery of spatial and temporal spectra and realistic dynamical features such as vorticity. A proof-of-concept km-scale extension demonstrates downscaling from $0.25^{\circ}$ to $0.0625^{\circ}$ by combining diffusion at a coarser grid with a deterministic upscaling, achieving reasonable inference times (as low as a few minutes per year) and maintaining physically plausible extremal-event statistics. The work suggests TAUDiff can enable rapid, multi-resolution extreme-weather datasets for risk assessment and economic-loss estimation with reduced computational burden.

Abstract

Deterministic regression-based downscaling models for climate variables often suffer from spectral bias, which can be mitigated by generative models like diffusion models. To enable efficient and reliable simulation of extreme weather events, it is crucial to achieve rapid turnaround, dynamical consistency, and accurate spatio-temporal spectral recovery. We propose an efficient correction diffusion model, TAUDiff, that combines a deterministic spatio-temporal model for mean field downscaling with a smaller generative diffusion model for recovering the fine-scale stochastic features. We demonstrate the efficacy of this approach on downscaling atmospheric wind velocity fields obtained from coarse GCM simulations. We then extend TAUDiff for computationally efficient kilometer-scale downscaling of atmospheric wind velocity fields. Owing to low inference times, our approach can ensure quicker simulation of extreme events necessary for estimating associated risks and economic losses.

TAUDiff: Highly efficient kilometer-scale downscaling using generative diffusion models

TL;DR

TAUDiff tackles spectral bias in climate-downscaling by coupling a deterministic mean-field video-prediction model with a compact diffusion-based correction, enabling efficient and dynamically consistent wind-field downscaling at kilometer scales. The method is trained on ERA5 reanalysis and evaluated against baselines, including a full end-to-end diffusion model, showing superior recovery of spatial and temporal spectra and realistic dynamical features such as vorticity. A proof-of-concept km-scale extension demonstrates downscaling from to by combining diffusion at a coarser grid with a deterministic upscaling, achieving reasonable inference times (as low as a few minutes per year) and maintaining physically plausible extremal-event statistics. The work suggests TAUDiff can enable rapid, multi-resolution extreme-weather datasets for risk assessment and economic-loss estimation with reduced computational burden.

Abstract

Deterministic regression-based downscaling models for climate variables often suffer from spectral bias, which can be mitigated by generative models like diffusion models. To enable efficient and reliable simulation of extreme weather events, it is crucial to achieve rapid turnaround, dynamical consistency, and accurate spatio-temporal spectral recovery. We propose an efficient correction diffusion model, TAUDiff, that combines a deterministic spatio-temporal model for mean field downscaling with a smaller generative diffusion model for recovering the fine-scale stochastic features. We demonstrate the efficacy of this approach on downscaling atmospheric wind velocity fields obtained from coarse GCM simulations. We then extend TAUDiff for computationally efficient kilometer-scale downscaling of atmospheric wind velocity fields. Owing to low inference times, our approach can ensure quicker simulation of extreme events necessary for estimating associated risks and economic losses.

Paper Structure

This paper contains 15 sections, 5 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Overview
  4. TAUDiff Component 1: Mean model
  5. TAUDiff Component 2: Correction diffusion model
  6. Results and discussion
  7. Model validation and GCM downscaling
  8. Experimental protocol
  9. Validation
  10. Testing
  11. Extension to km-scale downscaling
  12. Methodology extension
  13. Experimental protocol
  14. Testing
  15. Conclusion

Figures (5)

  • Figure 1: Schematic of the TAUDiff model
  • Figure 2: (a) European region used for training the downscaling models, with select locations used for evaluating performance. Comparison of model predictions: (b) vorticity snapshot at UTC: 2023-12-31 21:00, (c) spatial spectrum, and (d) temporal spectra at select locations shown in (a).
  • Figure 3: Assessment of downscaling performance on bias corrected CAM4 data: (a) Vorticity contours at a representative time instance, (b) temporal spectrum, (c) vorticity distributions, and (d) local storm counts.
  • Figure 4: Schematic of the km-scale downscaling pipeline.
  • Figure 5: Assessment of ERA5 to CERRA downscaling performance: (a) Vorticity contours at UTC: 2010-11-10 21:00, (b) spatial spectrum, and (c) temporal spectra at select locations as shown in \ref{['fig:modelcompare']}(a).