The traditional weather prediction models improve slowly, while machine learning models struggle with extreme weather and fine details. To address these gaps, we developed an online correction system that leverages a machine learning model's skillful large-scale circulation to guide a physical model. This hybrid model enhances large-scale skill while preserving small-scale features, providing a viable pathway for improving operational weather forecasting.

This study represents the first application of CorrDiff, a diffusion-based model, to achieve 3 km resolution statistical downscaling over China by integrating various variables at multiple levels. Through structural improvements to CorrDiff, our methods produce forecasts that outperform the China Meteorological Administration Mesoscale Model for almost all target variables. Our results show that for radar reflectivity prediction, CorrDiff yields more realistic outputs than regression models.

Weather forecasts sometimes show high-frequency noise degrading predictions. Our study reveals stripe patterns arise from mismatches between dynamic and physical calculations in models. Simplified experiments demonstrate that adjusting their connection eliminates stripes. This advances numerical weather prediction understanding, aiding forecasters and the public. Our diagnostic methods provide a framework for solving this global meteorological modeling challenge.

In order to investigate the feedbacks of chemical data assimilation on meteorological forecasts, we developed a strongly coupled aerosol–meteorology four-dimensional variational (4D-Var) assimilation system, CMA-GFS-AERO 4D-Var, based on the framework of the incremental analysis scheme of the China Meteorological Administration Global Forecasting System (CMA-GFS) operational global numerical weather model. The results show that assimilating BC (black carbon) observations can generate analysis increments not only for BC but also for atmospheric variables.

This work explores the use of global storm-resolving model (GSRM) simulation data to enhance global climate modeling (GCM) through a machine learning–based model physics suite. Stable multiyear climate simulations with improved precipitation characteristics are achieved by using 80-day GSRM data.