GraphCast beats conventional weather forecasting on speed and accuracy

Google DeepMind published GraphCast, a machine-learning weather model that generates a 10-day global forecast at 0.25-degree resolution — temperature, wind, pressure and humidity across the Earth’s surface and 37 atmospheric levels — in under a minute on a single Google TPU v4 machine, work also published in Science the same week. Conventional numerical weather prediction produces forecasts of comparable scope only after hours of computation on a dedicated supercomputer.

DeepMind reported that GraphCast beat the European Centre for Medium-Range Weather Forecasts’ HRES system, widely treated as the deployed gold standard, on more than 90% of 1,380 tested variables and altitudes, with a particularly large margin in the troposphere. The model is built on a graph neural network, a machine-learning architecture suited to data — like the interlocking grid of the atmosphere — with irregular, spatially connected structure rather than the fixed grid a convolutional network expects.

The result did not emerge from nowhere: GraphCast was trained on four decades of ECMWF’s ERA5 reanalysis dataset, itself produced by blending observations with a traditional physics-based model to fill gaps where direct measurements were unavailable. In that sense, the AI system did not replace physical modelling so much as compress decades of it into a much faster, if less interpretable, form. Conventional forecasting also retains advantages GraphCast was not built to match, including physically grounded uncertainty estimates and performance in rare extreme-weather regimes underrepresented in the training data.

GraphCast became one of the clearer public demonstrations that deep learning could outperform an established, heavily engineered scientific pipeline on its own home turf, using orders of magnitude less compute at inference time, and it fed a wave of subsequent AI-weather efforts from other labs and national forecasting agencies.