FuXi Weather: A data-to-forecast machine learning system for global weather

Weather forecasting traditionally relies on numerical weather prediction (NWP) systems that integrates global observational systems, data assimilation (DA), and forecasting models. Despite steady improvements in forecast accuracy over recent decades, further advances are increasingly constrained by high computational costs, the underutilization of vast observational datasets, and the challenges of obtaining finer resolution. These limitations, alongside the uneven distribution of observational networks, result in global disparities in forecast accuracy, leaving some regions vulnerable to extreme weather. Recent advances in machine learning present a promising alternative, providing more efficient and accurate forecasts using the same initial conditions as NWP. However, current machine learning models still depend on the initial conditions generated by NWP systems, which require extensive computational resources and expertise. Here we introduce FuXi Weather, a machine learning weather forecasting system that assimilates data from multiple satellites. Operating on a 6-hourly DA and forecast cycle, FuXi Weather generates reliable and accurate 10-day global weather forecasts at a spatial resolution of $0.25^\circ$. FuXi Weather is the first system to achieve all-grid, all-surface, all-channel, and all-sky DA and forecasting, extending skillful forecast lead times beyond those of the European Centre for Medium-range Weather Forecasts (ECMWF) high-resolution forecasts (HRES) while using significantly fewer observations. FuXi Weather consistently outperforms ECMWF HRES in observation-sparse regions, such as central Africa, demonstrating its potential to improve forecasts where observational infrastructure is limited.

💡 Research Summary

FuXi Weather presents a fully integrated, end‑to‑end machine‑learning (ML) weather forecasting system that replaces the traditional numerical weather prediction (NWP) pipeline of data assimilation (DA) and dynamical modeling with a unified neural‑network framework. The system operates on a six‑hourly cycle, performing DA and forecast generation four times per day, and produces global forecasts out to ten days at a 0.25° (~27 km) spatial resolution.

The core of the system consists of two neural components: FuXi‑DA, a ML‑based DA engine, and FuXi, a forecast model derived from the previously published FuXi architecture. FuXi‑DA ingests raw brightness‑temperature observations from three polar‑orbiting satellites (FY‑3E, Metop‑C, NOAA‑20) covering five microwave channels, together with GNSS radio‑occultation (RO) profiles. A modified PointPillars point‑cloud encoder handles the heterogeneous spatial‑temporal sampling, while a binary mask marks missing data. The DA window spans eight hours (three hours before to four hours after the analysis time), and the background field is supplied by the most recent FuXi forecast. Experiments comparing runs with and without this background field demonstrate that the ill‑posed DA problem requires prior information: inclusion of the background reduces RMSE across all variables and stabilizes the analysis, especially at low‑level pressure surfaces where satellite information is sparse.

The analysis fields produced by FuXi‑DA are then used as initial conditions for the FuXi‑Short and FuXi‑Medium forecast models, which generate 10‑day predictions. Training uses ERA5 reanalysis data at the same 0.25° grid as the reference truth. To emulate operational variability, background forecasts are randomly sampled from ERA5 lead times of 6 h to 5 d during training. Because only one year of satellite data (June 2022–June 2023) is available, a replay‑based incremental learning strategy retrains FuXi‑DA each month using the previous year’s observations, allowing the system to adapt to sensor degradation or changes in data availability.

Performance is evaluated with latitude‑weighted root‑mean‑square error (RMSE) and anomaly correlation coefficient (ACC) against ERA5, and benchmarked against the ECMWF high‑resolution (HRES) model. Globally, FuXi Weather’s RMSE is initially higher than HRES but falls below it after variable‑specific lead times: for relative humidity (R) at 300 hPa, 500 hPa, and 850 hPa the skill surpasses HRES after 2.0, 3.25, and 2.25 days respectively; for geopotential height (Z) the crossover occurs around 7.5–8 days. ACC trends mirror these findings, with FuXi Weather achieving ACC > 0.6 (the skill threshold) earlier for moisture variables and later for dynamical fields.

Regional analyses highlight the system’s strength in observation‑sparse areas. In central Africa and northern South America, where surface stations are scarce, FuXi Weather consistently yields lower RMSE and higher ACC than HRES, confirming that satellite‑only DA can compensate for ground‑based data gaps.

From a computational perspective, the ML pipeline runs on modern GPUs and requires orders of magnitude fewer floating‑point operations than the supercomputer‑scale NWP models that power HRES. This translates into dramatically reduced operational costs while delivering comparable or longer skillful lead times.

In summary, FuXi Weather demonstrates that a data‑to‑forecast ML system can perform global, all‑sky, all‑surface, and all‑channel data assimilation and produce skillful medium‑range forecasts without relying on traditional NWP infrastructure. Its ability to outperform a leading operational model in data‑poor regions, coupled with its computational efficiency, marks a significant step toward more equitable and sustainable weather prediction worldwide.