DeepMind's GraphCast predicts 10-day weather in under a minute

GraphCast is Google DeepMind's open-source AI system for global medium-range weather forecasting. Running on a single Google TPU v4, it generates a high-resolution 10-day weather forecast for the entire planet in less than a minute.

Published in Science, the model was tested head-to-head against the industry gold standard: the High Resolution Forecast (HRES) from the European Centre for Medium-Range Weather Forecasts (ECMWF). GraphCast outperformed HRES on more than 90% of 1,380 test variables and lead times. In the troposphere—the 6 to 20 km atmospheric layer nearest Earth's surface where weather actually hits us—GraphCast beat HRES on 99.7% of variables.

What it does

GraphCast splits the globe into a 0.25-degree latitude/longitude grid, translating to roughly 28 km by 28 km squares at the equator. That adds up to over one million grid points covering Earth.

At every single point, it predicts 5 surface variables—including temperature, wind speed and direction, and mean sea-level pressure—alongside 6 atmospheric variables across 37 altitude levels, covering specific humidity, temperature, and wind.

Beyond day-to-day conditions, GraphCast shines at spotting extreme weather early, despite never being explicitly trained to track disasters. Paired with a basic cyclone tracker, it tracks storm tracks more reliably than traditional physics models.

During a live trial on ECMWF's site in September 2023, GraphCast predicted Hurricane Lee's landfall in Nova Scotia roughly 9 days in advance. Conventional models were all over the place and only locked onto the landfall about 6 days out.

It also models atmospheric rivers to forecast flood risks and flags when regional heat will smash historical maximums.

How it works and hardware requirements

Traditional Numerical Weather Prediction (NWP) computes physics equations on massive supercomputer clusters over several hours. It is scientifically proven but slow and brutally expensive.

GraphCast bypasses step-by-step physical calculations using Graph Neural Networks (GNNs), which map spatial data over the globe naturally. DeepMind trained it on 40 years of ECMWF's ERA5 historical reanalysis data—a massive archive combining satellite imagery, radar, and ground weather stations with NWP fills.

While training was computationally intensive, running inference is practically instantaneous. A full 10-day global forecast takes under 60 seconds on a single TPU v4. Compare that to hours on hundreds of supercomputer nodes for HRES.

DeepMind has released the code on GitHub, and ECMWF is already running live operational experiments with it.

Why it matters

For meteorologists, emergency response teams, and weather-dependent industries like renewables and logistics, GraphCast shows that data-driven AI can outperform decades of thermodynamic modeling in pure forecasting accuracy. What used to require a national supercomputing center can now run in seconds on modest cloud hardware.

That said, it does not kill traditional meteorology. GraphCast depends entirely on ERA5 data produced by conventional systems. We expect the immediate future to be hybrid: physics models handling data assimilation and ground truth checks, while AI handles the high-speed heavy lifting. The code is available now for researchers looking to build regional forecasting pipelines.