GraphCast & Learned Forecasts

The ML approach is audaciously simple: take about forty years of ERA5 reanalysis — the ECMWF's best reconstruction of historical weather — and learn the direct map from 'the atmosphere now' to 'the atmosphere in six hours.' Then feed the output back in and repeat: an autoregressive rollout to day 10.

FourCastNet used Fourier neural operators; Pangu-Weather (2023, Nature) used a 3D transformer and became the first learned model to beat the ECMWF's operational forecast in standard verification; GraphCast (2023, Science, Learning skillful medium-range global weather forecasting) pushed further with graph networks.