FuXi: a cascade machine learning forecasting system for 15-day global weather forecast

Over recent years, machine learning (ML) weather models have surpassed ECMWF’s high-resolution forecast (HRES) for 10-day global forecasts at 0.25° resolution, yet extending skill to 15 days remains challenging due to error accumulation. This work presents FuXi, a cascaded ML weather forecasting system that generates 15-day global forecasts every 6 hours at 0.25° resolution, trained on 39 years of ERA5. FuXi employs three models specialized for lead-time ranges 0–5, 5–10, and 10–15 days to mitigate iterative error accumulation. Evaluations show FuXi achieves performance comparable to the ECMWF ensemble mean (EM) at 15 days, and extends skillful lead time (ACC > 0.6) beyond HRES: Z500 from 9.25 to 10.5 days and T2M from 10 to 14.5 days. An ensemble version of FuXi, created via perturbations to initial conditions and model parameters, provides uncertainty estimates and shows promising probabilistic skill compared with the ECMWF ensemble.

Lei Chen, Xiaohui Zhong, Feng Zhang, Yuan Cheng, Yinghui Xu, Yuan Qi, Hao Li