China regional 3 km downscaling based on residual Corrective Diffusion model

Abstract

A fundamental challenge in numerical weather prediction is efficiently producing high-resolution forecasts. A widely adopted solution is to downscale coarser global model outputs. This study focuses on statistical downscaling, which uses historical data to learn mappings between low- and high-resolution meteorological fields. Deep learning has become a key tool for this, enabling powerful super-resolution models like diffusion models and Generative Adversarial Networks for downscaling applications. Herein, we leverage CorrDiff, a diffusion-based downscaling framework, with three key enhancements relative to its original implementation. First, the study domain is expanded to nearly 40 times the spatial extent of the original version. Second, the scope of target downscaling variables is extended beyond surface variables to incorporate upper-air variables across six pressure levels. Third, a global residual connection is integrated to further boost prediction accuracy. To produce 3 km resolution forecasts for the China region, the trained CorrDiff model is applied to 25 km global grid forecasts derived from two sources: a conventional global numerical weather prediction model and a deep learning-based weather model developed using Spherical Fourier Neural Operators. The China Meteorological Administration Mesoscale Model (CMA-MESO) is selected as the baseline for comparative evaluation. Experimental results demonstrate that the downscaled forecasts generated by our framework outperform the direct forecasts of CMA-MESO for almost all target variables, as quantified by the Mean Absolute Error metric. Specifically, forecasts of radar composite reflectivity reveal that CorrDiff, as a generative model, is capable of capturing fine-grained meteorological details, yielding more physically realistic predictions than deterministic regression-based downscaling models.