Google shows AI model that can make ‘unprecedentedly accurate’ weather forecasts

Google DeepMind introduces GraphCast. According to the tech giant, this AI model can make an ‘unprecedentedly accurate’ weather forecast for the next ten days in less than a minute. DeepMind has made the code of the GraphCast model open source.

GraphCast is possible according to Google subsidiary DeepMind make rapid medium-term weather forecasts with ‘unprecedented accuracy’. The tech giant divides up a peer-reviewed report on Science details about the AI ​​model, which according to Google DeepMind is in many cases faster and more accurate than the HRES system of the European Center for Medium-Range Weather Forecasts, also known as the ECMWF.

According to its creators, the deep learning model can generate accurate weather forecasts for the next ten days. GraphCast can also warn earlier about extreme weather conditions, Google claims. The model can more accurately track the path of cyclones, ‘identify atmospheric rivers’ linked to flood risks and predict the onset of extreme temperatures.

To date, weather forecasts have mainly been calculated using numerical weather prediction. Weather forecasts therefore start with ‘carefully defined physics equations’, Google writes. These equations are then translated into algorithms that are run on supercomputers.

The GraphCast model is trained on “decades of historical weather data.” According to its creators, the model has learned the cause-and-effect relationships that determine how the weather develops. The model is trained on the ECMWF ERA5 dataset, which is based on weather observations such as satellite images, radars and weather stations. More traditional numerical weather predictions are used in that dataset to fill in any gaps when observations are incomplete for certain periods, Google writes.

According to Google, GraphCast makes predictions based on two data sets: the weather conditions of six hours ago and the current weather conditions. Based on this, the weather is predicted for the next six hours. This forecast can then be continued for another six hours, and so on. That can provide accurate results up to ten days into the future.

How GraphCast works. Source: Google DeepMind

GraphCast makes predictions at a resolution of 0.25×0.25 degrees in latitude and longitude. That amounts to 28x28km at the equator. In total, the model offers more than a million grid points on Earth. Five surface variables are predicted from each grid point, including temperature, wind speed and direction, and mean sea level pressure. Six atmospheric variables are also predicted at each point at 37 different altitude levels.

The GraphCast model was more accurate than the aforementioned HRES system from the European weather center on more than 90 percent of the 1,380 test variables, Google DeepMind claims. When the test was limited to the troposphereGraphCast was more accurate at predicting future weather conditions 99.7 percent of the time.

Google DeepMind also says that a 10-day GraphCast forecast can be generated in under a minute on a single Google Cloud TPU v4 system. Generating a traditional 10-day weather forecast on a supercomputer with hundreds of systems can take hours. DeepMind has made GraphCast’s code available open source and published on GitHub. The ECMWF weather center in Europe is already conducting experiments with the model.

GraphCast versus the HRES system in predicting extreme weather events. Source: Google Deepmind