GraphCast: Data-Driven Weather Forecasting

License: arXiv preprint (copyright held by authors) · Updated 2026-02-26

GraphCast predicts atmospheric state autoregressively on an icosahedral mesh using a graph neural network, trained on decades of reanalysis data (ERA5).

Reported results: Better than the operational deterministic baseline on the large majority of verified variables and lead times; a 10-day forecast completes in under a minute on a single accelerator, versus hours on a supercomputer.

Industrial relevance: The clearest evidence that learned surrogates can complement - and in some regimes beat - first-principles numerical solvers at a fraction of the compute cost.

Tags: WeatherGraph Neural NetworkForecastingNWPLarge Scale
Related entries