Google DeepMind has made public the two "raw materials" behind the WeatherNext 3 weather model, called SatWeave and PARDIG. Researchers and developers can get real-time data streams and historical archives.

SatWeave solves a very trivial but fatal problem. Global coverage relies on five geostationary weather satellites: GOES East and GOES West in the United States, two Meteosats in Europe, and Himawari in Japan. Each satellite has different map projections, file formats, sensors, and scanning rhythms. In the past, anyone who wanted to use the global cloud map had to spend a lot of time doing the alignment.

This data is now updated hourly and has been archived for more than ten years. WeatherNext 3 relies on it to provide hour-by-hour global forecasts. Traditional analysis data is only updated every six hours.

PARDIG is a reanalysis of global precipitation data, itself made by AI: using a spherical grid Transformer, taking radar observations from NASA precipitation measurement satellites as the answer, and inputting satellite images from SatWeave plus atmospheric data from the European Center for Medium-Range Weather Forecasts. The spatial resolution is about 10 kilometers, the temporal resolution is 15 minutes, and the average delay is about 5 hours.

Google compared ground-based radar networks in the United States, Europe, and Japan with more than 36,000 weather stations around the world, saying that it is better than mainstream data sets such as NASA's IMERG-Final in all levels of precipitation intensity. It can identify hourly light rain and moderate rain by more than 30%, and it can also capture extreme precipitation more accurately.

This data may be most useful in places where radar is not available. Many areas in Africa and South America lack ground radars, and in the past, precipitation could only be estimated by crude satellites. PARDIG learns directly from satellite observations and can be used for flood warnings, agricultural insurance, and photovoltaic output forecasting in these areas.