Google DeepMind Releases WeatherNext 3

Google DeepMind and Research have released WeatherNext 3, a next-generation weather forecasting model that uses raw satellite and ground observation data as direct input. To address the delays and biases of the analysis fields relied upon by existing models, it enables hourly initialization by using both geostationary satellite mosaics and ECMWF analysis fields as parallel inputs.

Key Technical Features

• Improved Resolution and Speed: Compared to existing AI models, temporal resolution is improved by 6 times and spatial resolution by approximately 5 times, forecasting surface temperature and dew point at a 5km (0.05°) resolution.
• Application of FGN Architecture: Through the FGN (Flow-based Generative Network) transformer, it generates 64 ensemble members with a single forward pass without iterative sampling, which distinguishes it from GenCast.
• Multi-Resolution Output: It simultaneously outputs three resolutions: 0.05° (station head), 0.1° (gridded surface variables), and 0.25° (pressure levels).

Performance and Application

• Accuracy Improvement: It showed CRPS (Continuous Ranked Probability Score) improvements of up to 30% compared to WeatherNext 2 and up to 40% compared to ECMWF ENS for 2m temperature forecasts. In the early lead times for precipitation forecasts, an error reduction of up to 60% was confirmed based on IMERG.
• Service Integration: It is being sequentially applied to Google Search, Gemini, Maps, and Earth Engine, with medium-range precipitation forecast accuracy expected to improve by up to 50%.
• Data Accessibility: While model weights are private, output data is provided via BigQuery, Earth Engine, and Cloud Storage. Real-time data arrives with a delay of approximately 7–8 hours.