a revolution for forecasting

GenCast is the name of the new model presented by Google DeepMind which seems able to do better than current weather forecasting tools. A look back at a phenomenon that could lead to a big change in this sector. Explanations.

How does it work?

This AI works on the principle of machine learning. To train it, researchers provided it with weather data from 1979 to 2018. It uses this to try to predict what will happen in the future. In this, GenCast differs from traditional models that rely on supercomputers and solve complex equations to try to simulate the physics of the atmosphere.

This method seems to be paying off. Thus, GenCast was compared to the ENS system used by the European Center for Medium-Range Weather Forecasts (ECMWF). It did better than the latter in 97.2% of cases based on data dating from 2019.

Among the strong points of this technology: its ability to predict the trajectories of climatic disasters, such as cyclones and their occurrence by warning further in advance. This represents a major advantage at a time when these events are likely to increase due to climate change.

GenCast can also produce its 15-day forecast in just eight minutes, whereas systems like ENS sometimes take several hours.

AI available in open source

Small downside however, GenCast was compared to an earlier version of the ENS which has since improved. That said, DeepMind claims to have evaluated its AI on data from 2020 and 2022 with similar conclusions.

These innovations are welcomed with caution by meteorologists who remain scientists interested in physics, as recalled The Verge. Aware of this problem, Google DeepMind has decided to publish the code of its AI as open source on GitHub. The idea is to let experts appropriate this technology in order to strengthen its social impact and its usefulness.

To learn more about this subject, you can reread this article where we tried to understand why the weather application on our iOS and Android smartphones makes mistakes so often. You can also find out more about Météo France's AI projects here.