Fig. 3: Comparison of 10-day forecast performance across models over a 1-year testing period, spanning July 03, 2023–June 30, 2024.

The figure presents the globally-averaged and latitude-weighted anomaly correlation coefficient (ACC) for forecasts generated by the FuXi model and ECMWF HRES in 10-day forecasts. FuXi forecasts are initialized using analysis fields produced by FuXi-DA with (red solid and green dashed lines) and without (black) background (bg) forecasts. The analysis includes five variables: relative humidity (RH), temperature (T), geopotential (Z), u component of wind (U), and v component of wind (V), at three pressure levels (300, 500, and 850 hPa). FuXi forecasts (red and black lines) are verified against ERA5, and also against FuXi-DA analyses. The five rows and three columns correspond to five variables and three pressure levels, respectively. When FuXi (green dashed lines) and ECMWF HRES (blue) forecasts are evaluated against their respective initialization time series, they inherently exhibit higher ACC in early lead times. Red dots indicate time steps where FuXi Weather significantly outperforms ECMWF HRES, based on the t-test at the 95% confidence level. The performance change on day 4 arises from the model transition from FuXi-Short to FuXi-Medium.