Google introduces WeatherNext 3 for high-resolution global forecasting

Google DeepMind and Google Research have introduced WeatherNext 3, a flagship global weather forecasting model that replaces traditional physics simulations with direct learning from real-world observations. Traditional numerical weather prediction models rely on complex supercomputer-driven simulations that carry a six-hour data lag, which can introduce biases into fast-changing variables like rain and surface temperature.

Real-time satellite data ingestion

WeatherNext 3 breaks from traditional training methods by ingesting a continuous mosaic of live global geostationary satellite data alongside sparse weather station observations. This allows the system to update every hour and provide a richer, continuously updated view of the atmosphere without the data lag inherent in older models.

Resolution and performance improvements

The model generates forecasts at multiple spatial resolutions, producing a global weather picture roughly five times sharper than its predecessor, WeatherNext 2. Surface variables such as temperature and moisture are visualized at a 5-kilometer resolution, other surface variables at 10 kilometers, and atmospheric variables like wind speed at 25 kilometers.

Precipitation and clean energy variables

Global weather models have historically struggled to accurately predict precipitation due to fast-moving cloud processes. By training on NASA satellite-based precipitation data and internal global precipitation reanalysis, WeatherNext 3 achieves a Continuous Ranked Probability Score improvement of up to 60 percent against baseline measurements.

In addition to rain and snow tracking, the model introduces specialized predictions for renewable energy production. It forecasts 100-meter wind speeds for turbine-height output, alongside cloud cover and sun radiation levels, helping grid operators and clean energy developers match asset generation with consumer demand.