Google introduces WeatherNext 2, its weather AI

Google DeepMind and Google Research have introduced WeatherNext 2, an artificial intelligence model that generates weather forecasts up to 8 times faster and at a resolution of one forecast per hour.

The system does not offer a single version of the future. Starting from the same initial conditions, it can generate hundreds of possible weather scenarios, including less likely but more important ones to prepare for, such as extreme rainfall, heat waves, or sudden changes in wind.

From a single forecast to hundreds of possibilities

Most weather-related decisions depend on more than knowing whether it will rain tomorrow. An airline needs to assess alternative routes, an energy company must anticipate production from a wind farm, and a city has to prepare for different levels of heat or flooding.

WeatherNext 2 is designed to work with that range of possibilities. Each prediction takes less than a minute on a single specialized processing unit known as a TPU. Google says generating the same volume of scenarios with physics-based weather models could take hours on a supercomputer.

The model uses an approach called Functional Generative Network, or FGN. Instead of adding random variations at the end of the process, it incorporates that noise directly into the model's architecture. This is intended to produce different scenarios that remain consistent with one another and physically realistic.

Greater accuracy across connected variables

WeatherNext 2 can predict individual weather elements, such as temperature, humidity, or wind speed. But its value also lies in connecting them to represent complete systems.

For example, estimating the temperature at each point in a region is not enough. To plan an electricity grid, it is important to know how temperature, wind, and humidity combine across the entire area and how that combination could affect energy consumption or production.

According to Google's tests, WeatherNext 2 outperforms the previous generation on 99.9% of the variables and forecast lead times evaluated, across horizons of 0 to 15 days. The company presents this result as a performance comparison, not as a guarantee that every individual forecast will be accurate.

Google has also used this technology in experimental cyclone forecasts, where analyzing several scenarios can help weather agencies make decisions before a storm's path becomes clear.

Where it will be available

The company is already moving the model beyond the research environment. WeatherNext 2 data is coming to:

• Google Earth Engine, for analyzing geospatial and environmental information.
• BigQuery, for querying weather data at scale.
• Vertex AI, through an early access program that allows companies to create custom inferences, meaning they can run the model within their own workflows.

Google has also updated forecasts in Search, Gemini, Pixel Weather, and the Google Maps weather API with WeatherNext technology. In the coming weeks, this information will also begin powering weather features in Google Maps.

For you, the most visible change will be gradual: more detailed and frequently updated forecasts within services you already use. The more significant impact, however, may be behind the screen, in decisions involving flights, energy, transportation, agriculture, and emergencies.

The next step will be seeing how the system performs outside demonstrations and in real-world weather conditions. Generating more scenarios does not eliminate weather uncertainty, but it makes it easier to plan for, which is exactly where a weather forecast can make a difference.