FengWu

FengWu is a data-driven global medium-range weather forecasting system from the Shanghai Artificial Intelligence Laboratory, published in April 2023. Built with the University of Science and Technology of China, Shanghai Jiao Tong University and the Chinese Academy of Sciences, it treats forecasting as a multi-modal, multi-task problem: model-specific encoder-decoders with a cross-modal fusion transformer, trained on 39 years of ERA5 reanalysis data under an uncertainty loss, with a replay buffer to stabilize medium-range skill.

On 2018 hindcasts, FengWu beat GraphCast on 80 percent of 880 reported predictands, cutting the RMSE of 10-day 500-hPa geopotential forecasts, and extended skillful global forecasts to 10.75 days of lead time, the first system to do so. Each inference step takes about 600 milliseconds on an A100. A later FengWu-GHR upgrade pushed resolution to roughly 9km and skillful lead time to 11.25 days. The system is a research release rather than a downloadable open model, so openWeights is false; its operational value has been demonstrated through typhoon-track forecasts adopted in China.