Diffusion modeling for high-resolution weather downscaling over the Hawaiian Islands

Abstract

High-resolution weather simulations are essential for capturing the complex terraindriven processes that shape precipitation and wind patterns across the Hawaiian Islands. However, dynamical downscaling using models like WRF at kilometer-scale resolution is computationally expensive. In this study, we explore a generative machine learning approach using a probabilistic score-based diffusion model to statistically downscale ERA5 reanalysis data from 27 km to 1.5 km resolution. The model is trained on hourly ERA5 reanalysis from 2002 to 2009 with WRF simulations as the target, and tested on the 2010–2012 period. We evaluate its performance across key surface variables, including 2 meter temperature, 10-meter wind components, and accumulated precipitation. Evaluation metrics include CRPS, power spectrum analysis, PDF comparisons, and case studies. The model performs well in recreating mesoscale features induced by the topography of the Hawaiian Islands, but struggles to fully capture variability driven by large-scale synoptic forcing. Once trained, the diffusion model produces a high-resolution downscale ensemble in seconds, offering a computationally efficient alternative to dynamical models.