Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

High-Resolution Climate Projections Using Diffusion-Based Downscaling of a Lightweight Climate Emulator

Haiwen Guan, Moein Darman, Dibyajyoti Chakraborty, Troy Arcomano, Ashesh Chattopadhyay, Romit Maulik

TL;DR

The proposed approach is able to preserve the coarse-grained dynamics from LUCIE while generating fine-scaled climatological statistics at ~28km resolution, and is able to preserve the coarse-grained dynamics from LUCIE while generating fine-scaled climatological statistics at ~28km resolution.

Abstract

The proliferation of data-driven models in weather and climate sciences has marked a significant paradigm shift, with advanced models demonstrating exceptional skill in medium-range forecasting. However, these models are often limited by long-term instabilities, climatological drift, and substantial computational costs during training and inference, restricting their broader application for climate studies. Addressing these limitations, Guan et al. (2024) introduced LUCIE, a lightweight, physically consistent climate emulator utilizing a Spherical Fourier Neural Operator (SFNO) architecture. This model is able to reproduce accurate long-term statistics including climatological mean and seasonal variability. However, LUCIE's native resolution (~300 km) is inadequate for detailed regional impact assessments. To overcome this limitation, we introduce a deep learning-based downscaling framework, leveraging probabilistic diffusion-based generative models with conditional and posterior sampling frameworks. These models downscale coarse LUCIE outputs to 25 km resolution. They are trained on approximately 14,000 ERA5 timesteps spanning 2000-2009 and evaluated on LUCIE predictions from 2010 to 2020. Model performance is assessed through diverse metrics, including latitude-averaged RMSE, power spectrum, probability density functions and First Empirical Orthogonal Function of the zonal wind. We observe that the proposed approach is able to preserve the coarse-grained dynamics from LUCIE while generating fine-scaled climatological statistics at ~28km resolution.

High-Resolution Climate Projections Using Diffusion-Based Downscaling of a Lightweight Climate Emulator

TL;DR

The proposed approach is able to preserve the coarse-grained dynamics from LUCIE while generating fine-scaled climatological statistics at ~28km resolution, and is able to preserve the coarse-grained dynamics from LUCIE while generating fine-scaled climatological statistics at ~28km resolution.

Abstract

The proliferation of data-driven models in weather and climate sciences has marked a significant paradigm shift, with advanced models demonstrating exceptional skill in medium-range forecasting. However, these models are often limited by long-term instabilities, climatological drift, and substantial computational costs during training and inference, restricting their broader application for climate studies. Addressing these limitations, Guan et al. (2024) introduced LUCIE, a lightweight, physically consistent climate emulator utilizing a Spherical Fourier Neural Operator (SFNO) architecture. This model is able to reproduce accurate long-term statistics including climatological mean and seasonal variability. However, LUCIE's native resolution (~300 km) is inadequate for detailed regional impact assessments. To overcome this limitation, we introduce a deep learning-based downscaling framework, leveraging probabilistic diffusion-based generative models with conditional and posterior sampling frameworks. These models downscale coarse LUCIE outputs to 25 km resolution. They are trained on approximately 14,000 ERA5 timesteps spanning 2000-2009 and evaluated on LUCIE predictions from 2010 to 2020. Model performance is assessed through diverse metrics, including latitude-averaged RMSE, power spectrum, probability density functions and First Empirical Orthogonal Function of the zonal wind. We observe that the proposed approach is able to preserve the coarse-grained dynamics from LUCIE while generating fine-scaled climatological statistics at ~28km resolution.
Paper Structure (15 sections, 7 equations, 9 figures, 2 tables)

This paper contains 15 sections, 7 equations, 9 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Data and Methods
  3. LUCIE
  4. Data
  5. SFNO-SR
  6. UNet-SR
  7. Elucidating Diffusion Models (EDM)
  8. Conditional Downscaling
  9. Posterior Sampling Downscaling
  10. Computational Cost
  11. Results
  12. Climatology
  13. Power spectrum and extreme events
  14. EOF of zonal wind
  15. Discussion and Conclusions

Figures (9)

  • Figure 1: Schematic of the downscaling framework. (Left) Training phase: The downscaling model is trained on T30 gridded ERA5 data to learn the transformation from a coarse resolution to a high resolution. (Right) Inference phase: The trained downscaling model is applied to the coarse-resolution output from the LUCIE climate emulator to generate high-resolution climate projections.
  • Figure 2: Super-resolved climatological snapshots averaged over the 2010–2018 period for 2m temperature, zonal wind, meridional wind, and precipitation. Coarse-grid dynamics for this period were generated by the LUCIE emulator and then downscaled with different super-resolution algorithms corresponding to the different rows. ERA5 reanalysis (our assumed ground truth) is provided in the first row from the top.x
  • Figure 3: Climatological mean of ERA5, Bicubic interpolation, SFNO-SR, Conditional EDM, and EDM with Posterior Sampling in December, January, and February of temperature in CONUS area from 2010 to 2018.
  • Figure 4: Climatological mean of ERA5, Bicubic interpolation, SFNO-SR, Conditional EDM, and EDM with Posterior Sampling in June, July, and August of temperature in CONUS area from 2010 to 2018.
  • Figure 5: Temporal mean of precipitation over March, April, and May from 2010 to 2018 in India.
  • ...and 4 more figures