Google introduces WeatherNext 3, its AI weather model

Google DeepMind and Google Research have introduced WeatherNext 3, an artificial intelligence model that updates its weather forecasts every hour and reaches a resolution of up to 5 kilometers. According to independent real-time evaluations by Brightband, it is the company’s most advanced and accurate global model to date.

The improvement targets a specific problem: general forecasts often miss important details when weather changes quickly or terrain affects local conditions. A storm, for example, can affect a coastal city, a valley, and a mountainous area just a few kilometers away in very different ways.

More detail and hourly updates

WeatherNext 3 generates hourly forecasts at several scales. It can represent temperature and humidity on a 5-kilometer grid, other surface variables at 10 kilometers, and atmospheric variables such as wind at 25 kilometers.

Overall, Google says it provides an image of the weather that is approximately five times more detailed than WeatherNext 2, which worked with 25-kilometer grids and updated every six hours.

The difference can be noticeable in everyday situations. A higher-resolution forecast is more likely to distinguish between rain in a mountainous area and dry weather a few kilometers away, or to better detect how heat is distributed across a region.

The model uses real-world weather data

AI weather models are typically trained on data from traditional forecasting systems. Those systems simulate the atmosphere using physics and supercomputers, but some data sources can be delayed by up to six hours.

WeatherNext 3 incorporates global mosaics of geostationary satellite images updated every hour. It also uses data from weather stations and precipitation observations gathered by satellites and radar.

This means each new forecast starts with a more recent picture of what is happening in the atmosphere. The advantage is especially important when storms, fronts, or heavy rain develop suddenly.

The model is also designed for areas that have historically had less access to high-resolution local forecasts, especially in Latin America, Africa, and Asia-Pacific. Building regional models of this kind with traditional methods requires substantial computing power.

Greater accuracy for rain and snow

Precipitation remains one of the most difficult variables to forecast. Clouds and storms form through fast, small-scale processes, so models can produce maps that are too blurry or fail to mark the boundaries of a severe storm accurately.

To improve this area, WeatherNext 3 was trained with precipitation data from IMERG, NASA’s system for combining measurements from multiple satellites, along with a global radar-based reconstruction developed by Google.

In its evaluations of medium-range forecasts, Google reports improvements of up to:

• 60% compared with IMERG.
• 30% compared with MRMS, another precipitation reference system.
• 10% compared with rain-gauge measurements at initial forecast horizons.

These figures use CRPS, a metric that compares the quality of a probabilistic prediction with what ultimately happened. They do not mean that every forecast is 60% more accurate in every place or situation.

What changes for you

Starting today, WeatherNext 3 will power Google’s weather experiences in Google Search, the Gemini app, Google Maps, Google Maps Platform Weather API, and Google Earth Engine.

Google says that when planning a day or more in advance, users will be able to receive precipitation forecasts that are up to 50% more accurate, with the greatest improvements in regions where forecasts have been less reliable. In practice, this could help you decide when to plan an outdoor activity, prepare for a trip, or figure out whether you need to bring a raincoat.

The system also provides forecasts for renewable energy: wind at 100 meters above ground, approximately the height of many turbines, as well as cloud cover and solar radiation. This information can help grid operators and energy companies estimate how much electricity their wind and solar farms will produce.

Google will make global data available to researchers, developers, and businesses through BigQuery, Earth Engine, and downloads from Google Cloud Storage, without requiring them to configure the model themselves.

The advance does not eliminate weather uncertainty or replace official alerts. Local weather agencies will still be needed for emergencies and extreme events. The important point is that digital forecasts are getting closer to what is actually happening in each area, especially where detailed, up-to-date data has been lacking.