“GraphCast: Learning skillful medium-range global weather forecasting,” by Remi Lam, Alvaro Sanchez-Gonzalez and colleagues at Google DeepMind, was posted to arXiv in December 2022 and later published in Science. It describes a machine-learning weather model that takes the current and recent state of the atmosphere and predicts how it will evolve, rather than solving the physics equations a conventional forecasting system integrates step by step.
GraphCast is an autoregressive model built on graph neural networks over a multi-scale mesh of the globe. Trained on ECMWF’s ERA5 reanalysis archive of historical weather, it produces a 10-day forecast of hundreds of surface and atmospheric variables at 0.25-degree resolution in under a minute on a single machine. The paper reports that GraphCast is more accurate than ECMWF’s HRES, the leading deterministic operational system in the world, on 90 percent of the 2,760 variable-and-lead-time combinations the authors evaluated, and that it also predicts severe events such as cyclone tracks and atmospheric rivers well.
The result mattered because it showed a learned model could beat a system that meteorologists had refined over decades, while running orders of magnitude faster and cheaper. That combination of speed and accuracy is why national weather services and private forecasters have moved quickly to adopt machine-learning models, and why GraphCast became a flagship example of the broader “AI for science” pattern.