Artificial intelligence (AI) has emerged as a promising alternative to traditional numerical weather prediction (NWP) models. However, the mismatch between training and operational input data limits its practical application in real-time forecasting. This study overcomes this challenge by fine-tuning the FuXi model using the Global/Regional Assimilation and Prediction Enhanced System (GRAPES) reanalysis data, resulting in an adapted version termed FuXi-GRAPES (FuXi-G). FuXi-G was evaluated over the full calendar year 2021. Compared to its baseline configuration, FuXi-G demonstrates marked improvements in forecast skill, particularly in the Southern Hemisphere and the tropics, achieving temporal and spatial accuracy that rivals or slightly surpasses that of NCEP operational forecasts. The FuXi-G model also exhibited a notable reduction in error propagation, outperforming NCEP forecasts by Day 5 and alleviating seasonal forecast biases in the Southern Hemisphere. A case study of Typhoon Khanun underscored the model’s enhanced ability to predict high-impact weather events, including a critical directional shift of the typhoon associated with the breakdown of the subtropical high. These results suggest that fine-tuning AI models can substantially improve forecasting accuracy while avoiding the computational burden of retraining on new datasets, providing a scalable approach for implementing AI models in operational weather forecasting.

Atmospheric reanalysis is a process that utilizes numerical weather prediction (NWP) models and data assimilation systems to integrate historical meteorological observations, thereby reconstructing datasets that represent past atmospheric states. This paper presents the China Meteorological Administration’s (CMA) latest global atmospheric reanalysis product, CRA1.5, which employs the Global Spectral Model (GSM) and the Gridpoint Statistical Interpolation (GSI) Data Assimilation system. CRA1.5 features a model spectral resolution of T_L1534 (equivalent to approximately 13 km in horizontal grid spacing), a product time interval of 1 h, a model top at 0.27 hPa, and a hybrid-4DEnVar scheme along with an Ensemble Kalman Filter for data assimilation. The temporal coverage of CRA1.5 extends from 1979 to near real-time. During the reanalysis production, we assimilated a substantial volume of reprocessed satellite data and extensive conventional observations. These observations include dense networks of conventional observations within China (comprising approximately 120 sounding stations and 2400 meteorological stations) as well as satellite observations from China’s Fengyun, Haiyang, and Yunyao satellite series. Compared to CRA-40, the root-mean-square error (RMSE) of geopotential height at 500 hPa is reduced by 22.46%, the RMSE of temperature at 200 hPa is reduced by 21.7%, and the RMSE of zonal wind at 850 hPa is reduced by 14.41%. CRA1.5 outperforms ERA5 in 100-m wind speed over China (RMSE: 3.21–3.36 m s⁻¹ vs. 3.46–3.59 m s⁻¹) and achieves comparable accuracy to ERA5 in atmospheric precipitable water (RMSE: 3.10 mm vs. 3.13 mm). CRA1.5’s 2-m temperature trend aligns closely with HadCRUT5, capturing global warming accurately. With higher spatiotemporal resolution (0.1° × 0.1°, hourly) and advanced assimilation, CRA1.5 provides a critical dataset for climate monitoring, NWP, and AI-based meteorological research. CRA1.5 has been implemented for real-time operation at the CMA. Updates are available up to 6 h after the analysis time, with a final update provided at a 3-day lag.