FengYuan: An End-to-End Global Weather Forecasting Model Driven by Multi-Source Observational Data

Current data-driven weather forecasting models demonstrate superior performance yet remain dependent on reanalysis or analysis fields from numerical models, inheriting their systematic errors. To address this constraint, we present FengYuan, an end-to-end global weather forecasting model that directly processes multi-source observational data to generate forecasts. FengYuan employs a modular architecture with two components: a data assimilation module (FengYuan-DA) that integrates observational data from satellites, surface stations, and radiosondes with background fields, and a forecasting module (FengYuan-Forecasting) that generates medium-range global forecasts. The system uses a staged training strategy, first training the FengYuan-Forecasting module on ERA5 reanalysis data, then iteratively optimizing the FengYuan-DA module using the forecasting module outputs. Evaluation on the 2022 test set shows that the FengYuan-Forecasting module achieves performance comparable to state-of-the-art Artificial Intelligence (AI) models and significantly outperforms ECMWF’s Integrated Forecasting System (IFS), with enhanced capabilities for regional forecasting over East Asia. The FengYuan-DA module achieves substantial improvements, showing remarkable spatial consistency with ERA5 reanalysis, low global root-mean-square errors, and effective assimilation of multi-source observational signals. The complete end-to-end system of FengYuan, initialized with FengYuan-DA analysis fields, achieves forecast quality very close to the reanalysis-driven version while consistently outperforming IFS throughout the 10-day forecast period, demonstrating that FengYuan successfully bridges observations to forecasts while maintaining high accuracy.