Google's WeatherNext 3 Drops Physics Simulations, Learns Weather Forecasting Directly From Satellite Data

Google Ditches Physics Simulations for Satellite Data

Google Research and DeepMind have released WeatherNext 3, an AI weather model that skips traditional numerical weather prediction (NWP) entirely and learns directly from real-time satellite observations. The change addresses a core limitation of prior AI weather systems: NWP simulations run on supercomputers and carry roughly a six-hour delay, according to Google, introducing errors for fast-changing variables like rainfall and temperature.

WeatherNext 3 instead processes live geostationary satellite data and generates a fresh forecast every hour based on the latest observations, at resolutions up to five kilometers.

Five Times Sharper Than WeatherNext 2

The model runs at multiple resolutions depending on variable: temperature and humidity at 5 kilometers, other surface variables at 10 kilometers, and atmospheric values like wind speed at 25 kilometers. That's roughly five times sharper than WeatherNext 2, which used a flat 25-kilometer grid updated every six hours.

Google illustrates the gap with a UK temperature map: WeatherNext 2 produces a blocky, pixelated forecast at 25 kilometers, while WeatherNext 3 resolves mountain ranges and coastlines at 5 kilometers. The model also trains on individual weather station data to sharpen forecasts around valleys and coastlines. Google says regions in Latin America, Africa, and the Asia-Pacific stand to benefit most, since high compute costs have historically left these areas underserved by regional forecasting models.

Precipitation and Renewable Energy Forecasts

Google trained the precipitation component on two sources: NASA's satellite-based IMERG dataset and Google's own global precipitation analysis built on satellite radar. For medium-range global forecasts, the Continuous Ranked Probability Score (CRPS) shows improvements of up to 60 percent over IMERG, 30 percent over MRMS, and 10 percent over rain gauges at short lead times, according to Google.

The model also predicts wind speeds at 100 meters — roughly turbine height — to estimate wind farm output, along with cloud cover and solar irradiance values for solar generation forecasting. Google says this is intended to help grid operators balance electricity supply and demand.

Availability

Forecast data updates hourly and is queryable through BigQuery and Earth Engine, or downloadable in bulk from Google Cloud Storage. WeatherNext 3 now powers weather features in Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine. Google says day-ahead precipitation forecasts should be up to 50 percent more accurate, with the largest gains in regions where predictions have historically been less reliable.

Both WeatherNext 3 and its predecessor share the same Functional Generative Network architecture. WeatherNext 2 launched in November 2025, claiming to beat the original WeatherNext on 99.9 percent of meteorological variables and forecast windows while running eight times faster. In August 2026, DeepMind open-sourced WeatherNext 2 and a companion cyclone-tracking model, WeatherNext Cyclones. Google continues to direct users to national weather services for official severe weather warnings.

What This Means

The shift from physics-based simulation to direct satellite-data learning marks a structural change in how AI weather models are built, not just an incremental accuracy gain. Removing the six-hour NWP dependency lets forecasts update hourly, which matters most for fast-moving events like convective storms and flash flooding — precisely the scenarios where legacy models lag. The renewable-energy forecasting angle also signals Google's ambition to sell this as infrastructure for grid operators, not just a consumer weather feature. Independent verification of the claimed CRPS and accuracy improvements will depend on peer-reviewed evaluation rather than Google's own benchmarking.