Climate Downscaling with Stochastic Interpolants (CDSI)

TL;DR

This work introduces a data-driven climate downscaling method based on stochastic interpolants that efficiently transforms coarse ESM output into high-resolution regional climate projections at a fraction of the computational cost of traditional RCMs.

Abstract

Global climate projections rely on computationally demanding Earth System Models (ESMs), which are typically limited to coarse spatial resolutions due to their high cost. To obtain high-resolution projections for regions of interest, it is common to use Regional Climate Models (RCMs), which are driven by data produced by ESMs as boundary conditions. While more efficient than running ESMs at fine resolution, RCMs remain expensive and restrict the size of ensemble simulations. Inspired by recent advances in probabilistic machine learning for weather and climate, we introduce a data-driven climate downscaling method based on stochastic interpolants. Our approach efficiently transforms coarse ESM output into high-resolution regional climate projections at a fraction of the computational cost of traditional RCMs. Through extensive validation, we demonstrate that our method generates accurate regional ensembles, enabling both improved uncertainty quantification and broader use of high-resolution climate information.

Figures (13)

• Figure 1: Comparison of optimal trajectories for an EDM diffusion process and a stochastic interpolant between LQ and HQ. Unlike diffusion-based approaches that generate high-resolution fields from pure noise, our cdsi constructs stochastic trajectories that evolve directly from the low-resolution input toward the target distribution, simplifying learning and improving sample realism.
• Figure 2: A qualitative comparison of the output for 2m temperature for all models. Note that the diffusion model is not able to remove all of the noise from the output despite using the same backbone architecture and training budget while the stochastic interpolant and CorrDiff models produces realistic samples. The UNET is a deterministic model and therefore the mean and member are the same and the same.
• Figure 3: Power spectra of temperature and precipitation.
• Figure 4: The RMSE over time for precipitation for the test period 2010--2014 for realization 1
• Figure 5: The SSR over time for precipitation for the test period 2010--2014 for realization 1