Google Unveils Next-Generation Weather AI 'WeatherNext 3' Trained on Real-Time Satellite Data

Google DeepMind and Google Research have unveiled WeatherNext 3, a next-generation global weather AI model that directly learns from real-time observational data to provide high-resolution forecasts. The model is currently rated as the most accurate on the live leaderboard of independent evaluation agency Brightband.

High-Resolution and Real-Time Forecasting Capabilities

While existing models relied on Numerical Weather Prediction (NWP) data with a 6-hour delay, WeatherNext 3 uses live geostationary satellite mosaics and sparse observation station data as direct inputs. This allows it to continuously update atmospheric states, generating hourly forecasts with 5km resolution for surface variables and 25km for atmospheric variables. This represents an approximately 5x improvement in the clarity of the global weather picture compared to the previous version, WeatherNext 2.

Significant Improvement in Precipitation Forecast Accuracy

To accurately model fine-scale cloud processes, the model was trained using NASA's IMERG data and proprietary satellite radar-based reanalysis data. As a result, medium-range global precipitation accuracy improved by up to 60% compared to IMERG, and rainfall gauge measurements improved by 10% at initial lead times. It particularly captures local variations better in regions where topographic details are critical, such as coastal or mountainous areas.

Renewable Energy and Service Integration

The model introduces features to predict wind speed at 100m turbine height for wind power generation and to estimate high-resolution cloud cover and solar irradiance for solar power generation. Starting today, the model is applied to Google Search, the Gemini app, Google Maps, and Google Earth Engine. Developers can query hourly updated high-resolution forecast data via BigQuery or Earth Engine.