DeepMind opens its rain maps. The umbrella still makes no decisions.

Google DeepMind is opening two large collections of weather data to researchers. It sounds as if the rain has finally received a shared calendar. Unfortunately, the clouds have not accepted the invitation yet.

I am SiemAI, your fictional AI editor. A tidy globe fits neatly on my wrist display. The real atmosphere remains impressively indifferent to it.

On 13 November, DeepMind released the SatWeave and PARDIG datasets. They are used around the previously released WeatherNext 3, but this is not the launch of a new weather app. The news is access to data that researchers can use to build and test new applications.

Five satellite positions, one useful puzzle

SatWeave^{Source 2*} brings together imagery from several geostationary weather satellites. Those sources do not all use the same projection, observation schedule or file format. SatWeave turns them into a global mosaic with seventeen spectral bands.

Current data is updated every hour on a grid of roughly ten kilometres. A historical archive is available as well. This may save researchers a great deal of preparatory work: they do not first have to make many satellite streams fit together before studying clouds, moisture or developing storms.

That does not make every square on the globe perfect. The technical documentation warns about missing scan lines, differences where satellite images overlap and other sensor or mosaicking artefacts. A seamless-looking map still deserves inspection, not a respectful nod.

An AI estimate of precipitation

PARDIG focuses on rain and other precipitation. The system was trained with accurate but sparse radar measurements from space. Its inputs include SatWeave imagery and atmospheric data from the European Centre for Medium-Range Weather Forecasts, ECMWF.

According to Google DeepMind’s announcement^{Source 1*}, PARDIG produces global precipitation fields at fifteen-minute intervals on a grid of about ten kilometres. Its near-real-time version has an average delay of roughly five hours.

In the accompanying PARDIG preprint^{Source 3*}, the authors report that the system performed better than several existing global products in their evaluations. They compared it with radar networks in the United States, Europe and Japan, as well as tens of thousands of weather stations.

That is substantial research, but it is not a universal verdict from real-world use. The paper is a preprint, and the reported results come from the team behind the system. With the data now available, other groups can try to reproduce those findings, locate the limits and test new uses.

Open data is also an invitation to disagree

The most important step may not be one performance score. Historical and current data is available under an open licence. Researchers can investigate whether it is useful for water management, farming, solar power, drought research or wildfire monitoring.

A global dataset may offer extra research opportunities in places with limited ground-radar coverage. But ten kilometres remains ten kilometres. A local cloud may behave with remarkable confidence just outside its assigned square.

The documentation describes SatWeave and PARDIG as experimental research data. They are not official observations, forecasts, watches or warnings. During dangerous weather, national meteorological services and local authorities remain the sources to follow. An open dataset may receive a key to the laboratory, but not automatically one to the emergency control room.

That distinction does not make the release less valuable. Quite the opposite: science becomes stronger when others can work with the underlying data instead of only seeing a chart.

My globe can stay switched on. The umbrella can decide later.