The GraphCast AI system promises to provide weather forecasts up to ten days in advance that are more accurate and faster than current models. The app, developed by Google DeepMind, was described in a November 14 Science article.

Unlike a traditional forecast model, which uses real-time data and mathematical equations, GraphCast is trained on climate data from the last 39 years. To reconstruct the global climate record, information such as images from satellites, radars and weather stations is used.

Compared to the old method, this new system provided more accurate predictions for more than 90% of 1,380 test variables and forecast time frames. It can also be used to provide warnings about severe weather events.

This AI, which is open source, is already being applied by meteorological agencies such as the European Center for Medium-Range Weather Forecasts (ECMWF), which runs live experiments with forecasts on the website.

GraphCast predicts Earth’s surface variables, such as temperature, wind speed and direction; and atmosphere, such as specific humidity.

How are weather forecasts prepared?

Forecasts are typically based on numerical weather prediction (NWP), made from physics equations and algorithms on supercomputers. This process, in addition to being time-consuming, also requires computational resources.

While traditional predictions can take hours of computation on a supercomputer with hundreds of machines, GraphCast results take less than a minute on a single machine.

The idea is to use a deep learning method (a branch of machine learning in neural networks that attempts to mimic the human brain) based on decades of meteorological data. The goal is to establish cause and effect relationships and understand how the Earth’s climate evolves.

This allows you to predict what will happen in the future based on past and current events. Traditional NWP will continue to be used to “fill in the gaps” in the new model.

Extreme weather events

The app can also be used to identify and warn about severe weather events. Tests revealed that GraphCast can, for example, predict hurricane movement more accurately than a traditional model, and days in advance.

The system can also determine when temperatures are expected to rise above historical maximum temperatures for any location on Earth, phenomena that are likely to become increasingly common. In the long term, the system could help reduce the effects of climate disasters.