Google Upgrades Weather Forecasting With Real-Time Satellite AI

Google Research and DeepMind have introduced WeatherNext 3, an AI model that replaces traditional physics-based simulations with direct processing of real-time satellite data. This approach eliminates the six-hour delay associated with numerical weather prediction models, allowing for hourly forecast updates.

The new model offers a resolution of up to five kilometers, which is five times sharper than the previous WeatherNext 2 version. It processes temperature and humidity at five-kilometer resolution, surface variables at ten kilometers, and atmospheric values like wind speed at 25 kilometers.

Training data includes NASA's IMERG dataset and Google's proprietary global precipitation analysis. The model shows performance improvements of up to 60 percent over IMERG and 30 percent over MRMS in medium-range forecasts. It also estimates wind farm and solar output to assist grid operators.

WeatherNext 3 is currently integrated into Google Search, Gemini, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine. Data is accessible to developers and businesses via BigQuery, Earth Engine, or bulk downloads from Google Cloud Storage.

The model utilizes the Functional Generative Network architecture, previously used in WeatherNext 2, which was introduced in November 2025. Google notes that while precipitation accuracy is expected to increase by up to 50 percent, national weather services remain the primary source for safety advisories.