FuXi

FuXi is a cascaded machine learning weather forecasting system from Fudan University's Artificial Intelligence Innovation and Incubation Institute, published in June 2023 with an npj Climate and Atmospheric Science paper that October. Its key idea addresses error accumulation in long-range forecasts: instead of one model handling all lead times, three cascade models are optimized for 0-to-5, 5-to-10 and 10-to-15-day windows, trained on 39 years of ERA5 reanalysis at 0.25-degree resolution and 6-hour steps.

FuXi became the first ML weather system with 15-day forecast skill comparable to the ECMWF ensemble mean, extending skillful Z500 lead time from 9.25 to 10.5 days and 2-meter temperature from 10 to 14.5 days versus ECMWF HRES. A perturbed ensemble version adds forecast uncertainty estimates. Developed with the Shanghai Academy of AI for Science, it is part of the Chinese AI-weather wave alongside FengWu and Pangu-Weather. Like its peers it is a research system without public model weights, so openWeights is false.