GraphCast weather model is published in Science

Background

Weather forecasting is a large computational and scientific problem. Operational centers combine observations with physics-based models of the atmosphere and run those models on supercomputers to estimate future temperature, wind, pressure, precipitation, and other conditions. The approach is powerful, but each forecast requires substantial computing time, and forecast quality depends on both the initial observations and the model. Researchers have therefore explored machine-learning models that learn from historical weather data and can generate a forecast quickly.

Announcement and paper

On 14 November 2023, Google DeepMind announced GraphCast, with a paper published in Science ^[1]. The paper appeared with DOI 10.1126/science.adi2336 ^[1]. The announcement described an AI model for global weather forecasting and gave results on accuracy, speed, and cyclone tracking ^[1].

Benchmark results

GraphCast was compared with ECMWF HRES. GraphCast beat HRES on more than 90% of 1,380 test variables and lead times ^[1]. This is a broad test: it covers many forecast variables and many forecast horizons rather than a single weather case. A model can perform well on one variable but poorly on another, so a wide comparison is more informative than a narrow one. The result is therefore a claim about consistency across a large test set, not just a demonstration on one map or one storm.

Speed

A 10-day forecast with GraphCast takes under a minute on one TPU v4 machine ^[1]. Forecast speed matters because a model can then be run repeatedly, for example to test different assumptions or to compare scenarios. It also changes the practical rhythm of experimentation: a result that arrives in under a minute can be checked and revised much more quickly than a long supercomputer job. The speed figure is separate from the accuracy comparison, but together they define the practical appeal of the model.

Cyclones and Hurricane Lee

Google DeepMind highlighted better tracking of cyclones ^[1]. One concrete example was Hurricane Lee. The live GraphCast model on ECMWF's website predicted about nine days ahead that Hurricane Lee would make landfall in Nova Scotia ^[1]. Traditional forecasts locked in on Nova Scotia only about six days ahead ^[1]. That is a difference of about three days of lead time ^[1]. Extra lead time can matter for warnings and preparations, but one storm is a case study rather than a full measure of average performance.

Open source

The model code was released as open source ^[1]. Open code allows researchers to inspect the method, reproduce reported results, and test changes; it also lets forecasters and developers evaluate the model on cases they care about. The release makes GraphCast part of a shared research record rather than only a published description.

Why it matters

The announcement matters because it combined a published comparison with operational-style evidence. The paper reported results over 1,380 test variables and lead times, with GraphCast ahead of ECMWF HRES on more than 90% of them ^[1]. It also reported a concrete speed figure: under a minute for a 10-day forecast on one TPU v4 ^[1]. And it connected the model to a live forecast: the GraphCast model on ECMWF's website and the Hurricane Lee landfall prediction ^[1]. Those details make the work legible to readers who want to know what the model did, how it was tested, and where it was made available.

Place in development

GraphCast belongs to a growing effort to apply machine learning to weather prediction. Its results are notable, but they do not by themselves settle how such models should be used operationally. A fast model and an early correct forecast are useful signals; operational use still requires validation across regions, seasons, and rare events. The open-source release gives other researchers a way to examine and extend the work ^[1]. In that sense, the announcement is both a result and a starting point for further testing.