Google DeepMind Opens PARDIG and SatWeave, the Weather Datasets Behind WeatherNext 3

For most of the planet, nobody knows precisely how much rain is falling right now. Oceans have no rain gauges, and much of the world lacks dense ground radar. Google DeepMind just opened up an AI-built global rainfall map that fills those gaps: a grid point roughly every 10 kilometers, a fresh map every 15 minutes, running from 1999 to today.

On Nov. 13, Google DeepMind released two datasets, PARDIG and SatWeave. They are key ingredients behind WeatherNext 3, the weather model Google launched in August. Until now they lived inside Google. Researchers and developers can now get both the historical archives and live feeds.

Start with SatWeave. It stitches imagery from six geostationary weather satellites, the ones that hover over a fixed spot on Earth, into one global mosaic with 17 visible, water vapor and infrared channels. It updates every hour and comes with more than a decade of history. That sounds like a simple collage, but speed is the point. Traditional forecast models usually start from atmospheric "analysis" data refreshed every six hours. WeatherNext 3 reads SatWeave directly, so it can issue a new global forecast every hour.

PARDIG is the more surprising piece. One of the best tools for measuring rain from space is the precipitation radar on the GPM Core Observatory, a NASA and Japan Aerospace Exploration Agency satellite. Its weakness is coverage: it sweeps a narrow strip as it circles Earth, so most places go unobserved most of the time. DeepMind trained an AI model to treat those scattered radar swaths as the answer key, while looking at SatWeave imagery and atmospheric data from the European Centre for Medium-Range Weather Forecasts. The model learns what rain on the ground looks like from the clouds above. After training, it can estimate rainfall everywhere, even where the radar isn't passing overhead.

Think of a teacher who can only grade a few students' homework each day, but grades it very carefully. Compare those spot checks with everyone's everyday work long enough, and you can estimate the whole class's real level pretty well.

Does it work? DeepMind says it tested PARDIG against ground radar networks in the US, Europe and Japan, plus more than 36,000 weather stations worldwide. Google's developer documentation says WeatherNext 3's precipitation forecasts trained on PARDIG had about 15% lower probabilistic error at short lead times than a version trained on IMERG, NASA's widely used rainfall product, and about 30% lower than WeatherNext 2. PARDIG's live feed lags reality by about five hours on average. Gauge-calibrated research-grade rainfall records often arrive months later.

What's new here is the combination. Global rainfall data has usually been either fast but rough, or accurate but months late, and it gets worse over oceans and poorer regions without ground radar. PARDIG aims for fast and accurate at once. Because it learns from a space radar instead of ground gauges, its quality holds up over oceans, the tropics and data-poor land. It's released under a CC BY 4.0 license, and anyone with a Google Cloud account can download it without filling out a request form.

You won't see a new weather app tomorrow because of this. But flood warnings, landslide alerts, drought monitoring and crop insurance all depend on one basic number: how much rain fell where. DeepMind says it wants to work with researchers on flood, landslide and drought models, especially in regions without dense radar. For those places, a free, global rainfall map refreshed every 15 minutes could be the first reliable rain record they've had.

Original source published: Google DeepMind, Nov 13, 2026, 20:00 Beijing time / 7:00 AM EST