Google DeepMind WeatherNext 3 Release

Google DeepMind and Google Research have introduced WeatherNext 3, a global weather AI model that integrates real-time satellite observations to provide hourly, high-resolution forecasts. This model represents a significant shift from previous AI weather models by learning directly from live observations rather than relying solely on numerical weather prediction (NWP) simulations, which typically carry a six-hour data lag.

High-Resolution Spatial and Temporal Forecasting

WeatherNext 3 provides a global weather picture approximately five times sharper than its predecessor, WeatherNext 2. The model generates hourly forecasts across multiple spatial resolutions to maintain physical consistency from global patterns to local topography:

  • 5-kilometer resolution: Used for key surface variables such as temperature and moisture.
  • 10-kilometer resolution: Used for other surface variables.
  • 25-kilometer resolution: Used for atmospheric variables, including wind speed.

In contrast, WeatherNext 2 operated on a 25-kilometer grid with 6-hour increments. The new architecture utilizes a Functional Generative Network (FGN) mesh transformer that ingests live 1-hour geostationary satellite mosaics and historical analysis to output dense gridded fields, discrete cyclone tracks, and native station-level sparse coordinates.

Integration of Real-World Observation Data

To eliminate the biases associated with the six-hour lag of traditional NWP models, WeatherNext 3 ingests a mosaic of live, global geostationary satellite data. This allows for a new forecast to be generated every hour, grounded in the most recent observations.

Furthermore, the model trains directly on sparse weather station observation data. This approach enables the 5-kilometer grid to account for regional details like topography, which is critical for areas near coastlines, valleys, or mountain ranges. This capability is particularly beneficial for regions in Africa, Latin America, and the Asia-Pacific, where high-resolution forecasting was previously limited by the supercomputing costs of traditional regional models.

Breakthroughs in Precipitation Accuracy

WeatherNext 3 addresses the common AI struggle with precipitation forecasting by training on two high-quality data sources: NASA’s satellite-based Integrated Multi-satellite Retrievals for GPM (IMERG) and a proprietary global precipitation reanalysis based on satellite radar.

Evaluations against baselines for medium-range global forecasts show significant improvements in Continuous Ranked Probability Score (CRPS):

  • Up to 60% improvement against IMERG.
  • 30% improvement against MRMS.
  • 10% improvement against rain gauge measurements for early lead times.

Clean Energy and Ecosystem Integration

WeatherNext 3 introduces specific variables engineered for renewable energy production to assist grid operators and developers in matching power generation with consumer demand:

  • 100-meter wind speeds: Predicted at turbine-height for wind-energy output.
  • Cloud cover and sun radiation levels: High-resolution data to help solar farms estimate ground-level light.

The model is being integrated across the Google ecosystem, including Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine. For users planning a day or more ahead, this results in up to 50% more accurate precipitation forecasts, especially in historically underserved regions.

Sources