ACHIEVEMENTS.AI

GraphCast forecasts ten days of global weather in under a minute

Google DeepMind published GraphCast in Science on 14 November 2023. Trained on about forty years of reanalysis data, it predicts hundreds of atmospheric variables ten days ahead at quarter-degree resolution, and beat the European Centre's operational forecast on 90 per cent of 1,380 verification targets. A forecast that takes hours on a supercomputer takes it under a minute.

Climate and weather Capability threshold Independently validated

Background

Weather forecasting works by solving the equations of atmospheric physics forward in time, on a grid, on a supercomputer. The method is sixty years old and has been refined continuously, and the European Centre’s system is the one everybody else is measured against.

Machine-learned forecasting had been tried. FourCastNet arrived in 2022 and Huawei’s Pangu-Weather after it, both promising and neither displacing the physics.

What happened

Google DeepMind published GraphCast in Science on 14 November 2023. It was trained on about forty years of recorded atmospheric conditions and predicts hundreds of variables ten days ahead, at quarter-degree resolution, in six-hour steps.

It beat the European Centre’s operational forecast on 90 per cent of 1,380 verification targets. Restricted to the troposphere, where most weather happens, it did better on 99.7 per cent of variables. Against Pangu-Weather it won on 98.8 per cent of the 252 targets that model had reported.

The speed is the part that is hard to take in. A ten-day global forecast takes under a minute on a single machine, against hours on a supercomputer.

One limitation belongs in the summary and is usually left out. The paper explicitly excludes precipitation from its claims. Rain is what most people mean by weather, and this result does not cover it.

What followed

GraphCast also identified cyclone tracks, atmospheric rivers and the onset of extreme temperatures earlier than the physical models, without having been trained to look for any of them. DeepMind reports that it placed Hurricane Lee in Nova Scotia nine days before landfall, three days earlier than conventional methods.

The code was released, and the European Centre, whose system GraphCast was measured against, began experimenting with it in the same year.

Why it mattered

Numerical weather prediction solves the equations of atmospheric physics step by step, and has been refined for sixty years. GraphCast solves nothing: it was trained on what the atmosphere did before. That a learned model beat the physics on most measures, and did it thousands of times faster, is the finding, and the European Centre began experimenting with it in the same year.

Sources

Cite this page

AI Achievements. (2023). GraphCast forecasts ten days of global weather in under a minute. Retrieved 2026-08-29, from https://achievements.ai/milestone/graphcast-weather-forecasting 

@misc{achievements_graphcast_weather_forecasting,
  title  = {GraphCast forecasts ten days of global weather in under a minute},
  author = {{AI Achievements}},
  year   = {2023},
  url    = {https://achievements.ai/milestone/graphcast-weather-forecasting}
}