Google DeepMind’s weather AI can forecast extreme weather faster and more accurately

This year the Earth has been hit by a record number of unpredictable extreme weather events made worse by climate change. Predicting them faster and with greater accuracy could enable us to prepare better for natural disasters and help save lives. A new AI model from Google DeepMind could make that easier. 

In research published in Science today, Google DeepMind’s model, GraphCast, was able to predict weather conditions up to 10 days in advance, more accurately and much faster than the current gold standard. GraphCast outperformed the model from the European Centre for Medium-Range Weather Forecasts (ECMWF) in more than 90% of over 1,300 test areas. And on predictions for Earth’s troposphere—the lowest part of the atmosphere, where most weather happens—GraphCast outperformed the ECMWF’s model on more than 99% of weather variables, such as rain and air temperature 

Crucially, GraphCast can also offer meteorologists accurate warnings, much earlier than standard models, of conditions such as extreme temperatures and the paths of cyclones. In September, GraphCast accurately predicted that Hurricane Lee would make landfall in Nova Scotia nine days in advance, says Rémi Lam, a staff research scientist at Google DeepMind. Traditional weather forecasting models pinpointed the hurricane to Nova Scotia only six days in advance.

“Weather prediction is one of the most challenging problems that humanity has been working on for a long, long time. And if you look at what has happened in the last few years with climate change, this is an incredibly important problem,” says Pushmeet Kohli, the vice president of research at Google DeepMind.  

Traditionally, meteorologists use massive computer simulations to make weather predictions. They are very energy intensive and  time consuming to run, because the simulations take into account many physics-based equations and different weather variables such as temperature, precipitation, pressure, wind, humidity, and cloudiness, one by one. 

GraphCast uses machine learning to do these calculations in under a minute. Instead of using the physics-based equations, it bases its predictions on four decades of historical weather data. GraphCast uses graph neural networks, which map Earth’s surface into more than a million grid points. At each grid point, the model predicts the temperature, wind speed and direction, and mean sea-level pressure, as well as other conditions like humidity. The neural network is then able to find patterns and draw conclusions about what will happen next for each of these data points. 

For the past year, weather forecasting has been going through a revolution as models such as GraphCast, Huawei’s Pangu-Weather and Nvidia’s FourcastNet have made meteorologists rethink the role AI can play in weather forecasting. GraphCast improves on the performance of other competing models, such as Pangu-Weather, and is able to predict more weather variables, says Lam. The ECMWF is already using it.

When Google DeepMind first debuted GraphCast last December, it felt like Christmas, says Peter Dueben, head of Earth system modeling at ECMWF, who was not involved in the research. 

“It showed that these models are so good that we cannot avoid them anymore,” he says. 

GraphCast is a “reckoning moment” for weather prediction because it shows that predictions can be made using historical data, says Aditya Grover, an assistant professor of computer science at UCLA, who developed ClimaX, a foundation model that allows researchers to do different tasks relating to modeling the Earth’s weather and climate. 

DeepMind’s model is “great work and extremely exciting,” says Oliver Fuhrer, the head of the numerical prediction department at MeteoSwiss, the Swiss Federal Office of Meteorology and Climatology. Fuhrer says that other weather agencies, such as the ECMWF and the Swedish Meteorological and Hydrological Institute, have also used the graph neural network architecture proposed by Google DeepMind to build their own models. 

But GraphCast is not perfect. It still lags behind conventional weather forecasting models in some areas, such as precipitation, Dueben says. Meteorologists will still have to use conventional models alongside machine-learning models to offer better predictions. 

Google DeepMind is also making GraphCast open source. This is a good development, says UCLA’s Grover. 

“With climate change on the rise, it’s very important that big organizations, which have had the luxury of so much compute, also think about giving back [to the scientific community],” he says.