GraphCast

What it is

GraphCast is a machine-learning model for global weather forecasting, built by Google DeepMind and described in a paper published in Science; the company announced it on 14 November 2023^[1]. It produces forecasts up to 10 days ahead at a latitude/longitude resolution of 0.25 degrees^[1]. At that resolution the model covers the whole planet on a regular grid rather than a single region, which is the setting in which national weather services normally run their global forecast systems.

How it works

GraphCast is based on graph neural networks^[1]. Instead of simulating the atmosphere by numerically solving physical equations, it learns from past weather how the state of the atmosphere tends to evolve. It was trained on four decades of ECMWF ERA5 weather reanalysis data, the record of past atmospheric conditions reconstructed by the European Centre for Medium-Range Weather Forecasts^[1].

The model works in steps. It predicts the weather six hours ahead, and that prediction is then fed back in and rolled forward in six-hour steps to build a forecast reaching up to 10 days^[1]. At each step it outputs five Earth-surface variables and six atmospheric variables at each of 37 altitude levels^[1].

Results

Speed is the most visible difference from conventional forecasting. A 10-day forecast takes less than a minute on a single Google TPU v4 machine^[1].

Accuracy was measured against ECMWF's high-resolution forecast, HRES. GraphCast was more accurate than HRES on more than 90% of 1,380 test variables and forecast lead times^[1]. The margin was clearest in the troposphere, the layer between roughly 6 and 20 km up where most weather happens: there the model outperformed HRES on 99.7% of test variables for future weather^[1]. Google DeepMind said GraphCast is the most accurate 10-day global weather forecasting system in the world^[1].

The company also pointed to a specific event: it said the model predicted Hurricane Lee's landfall in Nova Scotia about nine days ahead^[1].

Why it matters

Weather forecasting has long depended on large physics-based simulations run on supercomputers. A learned model that reaches comparable or better accuracy while running in under a minute on one accelerator changes the cost and speed at which forecasts can be produced, and makes it practical to run many forecasts or to iterate quickly on them.

Google DeepMind open-sourced the model code, which is available on GitHub at google-deepmind/graphcast^[1]. Releasing the code let researchers and weather services examine, reproduce and build on the approach rather than treating it as a closed product.

Where it sits

GraphCast belongs to a wave of machine-learning approaches to medium-range weather prediction that train on reanalysis data and emulate what physics-based models compute directly. Its reported comparison against HRES, a widely used operational high-resolution forecast, gave that approach a clear benchmark, and the open-source release made it a common starting point for later work.