Dynamical-generative downscaling of climate model ensembles

Area of Science:

• Climate Science
• Artificial Intelligence
• Computational Modeling

Background:

• High-resolution regional climate projections are vital for sectors like agriculture and hydrology.
• Current dynamical downscaling methods using regional climate models (RCMs) driven by Earth System Models (ESMs) are computationally prohibitive for large ensembles.
• Existing statistical downscaling methods often lack accuracy in capturing complex meteorological field characteristics.

Purpose of the Study:

• To develop a cost-effective approach for generating high-resolution climate projections from large ensembles.
• To improve the accuracy of uncertainty estimates in downscaled climate data.
• To combine the strengths of physics-based modeling with the efficiency of generative AI.

Main Methods:

• A two-stage downscaling process: first, dynamical downscaling using an RCM driven by ESM output to an intermediate resolution.
• Second, a generative diffusion model refines the resolution to the target scale.
• Evaluation against Coupled Model Intercomparison Project 6 (CMIP6) dynamical downscaling and statistical downscaling methods.

Main Results:

• The dynamical-generative approach provides more accurate uncertainty bounds compared to smaller ensembles or traditional statistical methods.
• Significantly lower errors were observed compared to popular statistical downscaling techniques.
• The method accurately captures the spectra, tail dependence, and multivariate correlations of meteorological fields.

Conclusions:

• The proposed dynamical-generative framework offers a flexible, accurate, and efficient solution for downscaling large climate projection ensembles.
• This approach makes high-resolution climate projections from extensive datasets computationally accessible.
• It represents a significant advancement over pure dynamical downscaling and conventional statistical methods for regional climate analysis.