CorrDiff, part of NVIDIA Earth-2, uses generative AI to downscale coarse climate projections (like CMIP6) into high-resolution, bias-corrected data. The model combines spatial super-resolution (11x), temporal downscaling (daily to hourly), and bias correction in one workflow by training on paired datasets—CanESM5 assimilated
•12m read time• From developer.nvidia.com
Table of contents
Why downscaling climate projections is key to risk assessmentInput and target datasets for CorrDiff trainingHow to train CorrDiff for climate downscalingRunning inference and evaluating downscaled output for CanESM5 SSP585Applications, takeaways, and next steps for climate downscalingGet startedSort: