Sun et al. (2026) China regional 3 km downscaling based on residual Corrective Diffusion model

Identification

• Journal: Geoscientific model development
• Year: 2026
• Date: 2026-09-29
• Authors: Zhixiang Dai, Sa Xiao, Jian Sun, Qifeng Lu
• DOI: 10.5194/gmd-19-9235-2026

Research Groups

• State Key Laboratory of Severe Weather Meteorological Science and Technology (LaSW), CMA Earth System Modeling and Prediction Centre (CEMC), Beijing, China
• CMA Earth System Modeling and Prediction Centre (CEMC), Beijing, China
• Key Laboratory of Earth System Modeling and Prediction, China Meteorological Administration, Beijing, China
• NVIDIA, Shanghai, China
• Department of Computer Science and Technology, Tsinghua University, Beijing, China

Short Summary

This study presents a diffusion-based downscaling framework, Corrective Diffusion (CorrDiff), which is applied to the China region for producing 3 km resolution forecasts. The framework is enhanced with three key improvements: expanded domain size, extended target variables, and global residual connection.

Objective

• Investigate the application of CorrDiff on a larger spatial extent than its original implementation.
• Examine the downscaling of multiple variables, including surface and upper-air variables across six pressure levels.
• Evaluate the performance of CorrDiff in predicting radar composite reflectivity.

Study Configuration

• Spatial Scale: Regional (China region) with 3 km resolution forecasts.
• Temporal Scale: Hourly assimilation-forecast cycle.

Methodology and Data

• Models used: Corrective Diffusion (CorrDiff), a diffusion-based downscaling framework, and UNet regression models.
• Data sources: ERA5 reanalysis data as input, 3 km reanalysis data from CMA-RRA as target data.

Main Results

• CorrDiff outperforms direct forecasts of the China Meteorological Administration Mesoscale Model (CMA-MESO) for almost all target variables.
• The diffusion model generates more high-frequency details and realistic small-scale features compared to regression models.
• The corrections learned by the diffusion model are spatially organized and associated with meteorologically active regions.

Contributions

• This study presents a novel application of CorrDiff on a larger spatial extent than its original implementation, demonstrating its potential for regional downscaling applications.
• The framework's ability to learn physically meaningful corrections and generate realistic fine-scale structures is a significant contribution to the field of statistical downscaling.

Funding

• This research was supported by the National Natural Science Foundation of China (Grant No. 92161102001).
• The authors acknowledge the support from the State Key Laboratory of Severe Weather Meteorological Science and Technology (LaSW).

Citation

@article{Sun2026China,
  author = {Sun, Honglu and Jing, Hao and Dai, Zhixiang and Xiao, Sa and Xue, Wei and Sun, Jian and Lu, Qifeng},
  title = {China regional 3 km downscaling based on residual Corrective Diffusion model},
  journal = {Geoscientific model development},
  year = {2026},
  doi = {10.5194/gmd-19-9235-2026},
  url = {https://doi.org/10.5194/gmd-19-9235-2026}
}