On November 14, 2023, DeepMind published GraphCast, a machine-learning weather forecasting system, alongside a paper in the journal Science. GraphCast predicts global weather up to 10 days ahead at 0.25-degree resolution (about 28 km at the equator), and it is trained directly from decades of historical reanalysis data rather than encoding the physics of the atmosphere by hand.
The headline result, stated in the paper “GraphCast: Learning skillful medium-range global weather forecasting” by Remi Lam and colleagues, is that GraphCast “significantly outperforms the most accurate operational deterministic systems on 90% of 1380 verification targets.” The benchmark it beat was the High Resolution Forecast (HRES) from the European Centre for Medium-Range Weather Forecasts, the long-standing industry standard.
The speed difference was as striking as the accuracy. GraphCast produces a 10-day forecast in under a minute on a single Google TPU v4, where conventional numerical weather prediction takes hours on a supercomputer with hundreds of machines. The model also improved prediction of severe events such as tropical cyclones and atmospheric rivers.
GraphCast is part of a broader pattern in which learned models match or beat hand-built physical simulators. It generalized the AlphaFold lesson - that deep learning could crack a hard scientific problem - into a domain, weather, where the conventional method was already extremely good. The Science page is bot-protected; the claims here are corroborated by DeepMind’s own announcement and the openly readable arXiv preprint (2212.12794).