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

TL;DR

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.

Abstract

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.

Figures (36)

• Figure 1: Schematic of the FuXi Weather system. Satellite radiance observations are brought in through machine learning data assimilation (DA) coordinated with the FuXi forecast model.
• Figure 1: Typical data coverage from observations collected by FengYun-3E (blue), Meteorological Operational Polar Satellite - C (Metop-C) (red), National Oceanic and Atmospheric Administration - 20 (NOAA-20) (green), and Global Navigation Satellite System (GNSS) radio occultation (RO) (yellow). This represents data spanning the period from 3 hours before to 4 hours after 12 UTC on June 1, 2023. These data are utilized to generate analysis fields for 12 UTC on the same date.
• Figure 1: Satellite data availability for FengYun-3E, Meteorological Operational Polar Satellite-C (Metop-C), and National Oceanic and Atmospheric Administration-20 (NOAA-20) for the 1-year testing period. The colors denote the number of observations at the temporal resolution of 1 hour.
• Figure 2: Comparison of forecast performance among various models over a 1-year testing period, spanning from July 03, 2023, to June 30, 2024. The figure presents the globally-averaged, latitude-weighted root mean square error (RMSE) for forecasts generated by the FuXi model and ECMWF HRES (blue) in 10-day forecasts. FuXi forecasts are initialized using analysis fields produced by FuXi-DA with (red) and without (black) background forecasts. The analysis includes 5 variables: relative humidity (${\textrm{R}}$), temperature (${\textrm{T}}$), geopotential (${\textrm{Z}}$), u component of wind (${\textrm{U}}$), and v component of wind (${\textrm{V}}$), at three pressure levels (300 hPa, 500 hPa, and 850 hPa). The five rows and three columns correspond to five variables and three pressure levels, respectively. As ECMWF HRES is evaluated against its own initialization time series, it inherently exhibits lower RMSE in early lead times.
• Figure 2: Comparison of analysis fields produced by FuXi Weather and 42-hour FuXi forecasts over a 1-year testing period from July 03, 2023, to June 30, 2024. The time series show the globally-averaged, latitude-weighted root mean square error (RMSE) for the analysis fields of FuXi Weather with (solid red lines) and without (solid black lines) background (bg) forecasts, along with 42-hour FuXi forecasts (dashed blue lines). The comparison includes five variables: relative humidity (${\textrm{R}}$), temperature (${\textrm{T}}$), geopotential (${\textrm{Z}}$), u component of wind (${\textrm{U}}$), and v component of wind (${\textrm{V}}$), at three pressure levels (300 hPa, 500 hPa, and 850 hPa). The five rows and three columns correspond to five variables and three pressure levels, respectively. To improve clarity, the original data are shown with reduced opacity, while solid lines represent smoothed values using a 12-point moving average.