Google DeepMind revolutionizes the weather and energy sector with its AI

Google DeepMind and Google Research have unveiled a new generation of their weather technologyWeatherNext 2. According to the company, this system would be capable of providing forecasts with unprecedented precision, while considerably reducing calculation time.

A more precise weather forecast with AI

Knowing the weather tomorrow or in the following days with unprecedented precision is the project of Google DeepMind and Google Research with their new AI.

Presented as their most advanced tool to date, it would offer hour-by-hour weather analyzes and deliver its results up to eight times faster than old versions, to the point of calling into question the interest of traditional forecasting platforms.

Google This is not his first attempt at weather. DeepMind has already designed several artificial intelligence models dedicated to forecasting, such as GraphCast and GenCast. The latter rely on NeuralGCM, a system capable of simulating the atmosphere.

With WeatherNext 2, the company claims to have reached a milestone, the new model outperforming the first version on almost all variables, such as temperature, wind or humidity, and on the majority of deadlines, with forecasts that can cover up to 15 days.

Each analysis would be produced in less than a minute using a single TPU (tensor processing unit). An impressive performance compared to traditional methods still requiring hours of calculation on supercomputers.

AI stronger than traditional methods

This progression is made possible thanks to a technique ofAI called Functional Generative Network. It consists of integrating a slight element of randomness into the model, in the form of what we call mathematical noise.

By changing this noise, the algorithm can propose several possible evolutions from the same weather data. This opens the way to hundreds of different scenarios, some of which include unusual or extreme phenomena.

The model starts by studying simple elements, like temperature or wind, then can predict larger phenomena, like a storm. Google also makes this data accessible on Earth Engine, BigQuery and Vertex, so that everyone can create their own forecasts.

Until now reserved for certain products, such as Pixel Weather or Google Mapsthese tools will soon be integrated directly into Maps.

Want to save even more? Discover our promo codes selected for you.