Abstract
The paper investigates the applicability of machine learning (ML) to
weather prediction by building a low-resolution ML model for global
weather prediction. The forecast performance of the ML model is assessed
by comparing it to that of persistence and a numerical physics-based
model, whose prognostic state variables and resolution are identical to
those of the ML model. The ML model typically provides realistic
prediction of the weather for the entire globe for about five forecast
days. For the first three forecast days, the ML model outperforms
persistence in the extratropics. While the relative performance of the
ML model compared to the physics-based model is mixed, the ML forecasts
are more accurate for the specific humidity in the extratropics and the
specific humidity and temperature in the tropics. These results suggests
that ML has a potential to improve the prediction of state variables
most affected by parameterized processes in numerical models.