Google DeepMind has released PARDIG and SatWeave, the two datasets behind its WeatherNext 3 forecasting model, as open resources for researchers and developers. SatWeave is a global mosaic of geostationary satellite imagery that gives the model a live view of developing clouds and storms, while PARDIG is an AI-built global precipitation dataset used to train the model’s rainfall forecasts. Both are now available with historical archives and live feeds on Google Cloud Storage, DeepMind announced on 13 November.

SatWeave: a live view of the whole planet
SatWeave stitches together imagery from six geostationary satellites — GOES-East, GOES-West, two Meteosat spacecraft and Himawari — into one consistent global mosaic across 17 visible, water vapour and infrared channels. Each satellite uses different projections, formats and scan schedules, so forming a coherent global picture from them has been a persistent problem.
The mosaic updates hourly, which lets WeatherNext 3 issue a new global forecast every hour from the latest satellite observations. Models that start from traditional analysis data refresh only every six hours, so fast-moving storms can already be hours old by the time a forecast runs. DeepMind is also releasing a historical SatWeave archive spanning more than a decade.
PARDIG: rainfall, mapped by AI
PARDIG, short for Precipitation AI Reanalysis: Densely Inferred GPM, is an AI model trained on sparse radar observations from NASA’s Global Precipitation Measurement core observatory, using SatWeave imagery and atmospheric fields from the European Centre for Medium-Range Weather Forecasts as inputs. The result is a precipitation map at 15-minute intervals and roughly 10-kilometre resolution, delivered with five hours of latency.
DeepMind tested PARDIG against three radar networks in the United States, Europe and Japan, plus more than 36,000 weather stations worldwide. It beat NASA’s IMERG-Final, the research-grade precipitation product, at every rain rate analysed, improving hourly detection of light-to-moderate rain by more than 30 percent. It was also strongest at catching extreme rainfall, the kind global products most often miss.
The team illustrates the point with the July 2025 Texas Hill Country flash floods, where near-real-time PARDIG tracked the hourly rainfall far more closely than IMERG-Final when both were compared with ground radar. Because PARDIG learns from satellites rather than ground instruments, DeepMind expects it to hold up in regions with no dense radar network at all.
Our Read
The interesting move is what DeepMind is giving away. WeatherNext 3 stays a Google product, but the data behind it is now open, with feeds and archives back to 1999. The bet is that an open weather-data layer grows the field, while the model, the compute and the distribution in Search, Maps and the Gemini app stay in Google’s hands.
There are limits. PARDIG’s five-hour latency means it is a research and planning tool, not a live warning system, and the PARDIG paper is a preprint, not yet replicated by others. The 30 percent figure is DeepMind’s own evaluation. Even with those caveats, an open rainfall record at this resolution is a genuine step for flood, drought and climate work.
Why It Matters
Rainfall is the hardest part of the water cycle to measure well, and the gaps are worst exactly where floods and droughts hit hardest — regions without dense radar or gauge networks. A freely available global precipitation record, refreshed continuously and stretching back to 1999, gives researchers, insurers, farmers and disaster planners the same baseline DeepMind’s own forecasters use.
It also sets a pattern. DeepMind has now opened the data behind two of its flagship systems in the same week — AlphaGenome 2’s DNA model arrived days earlier — and the weather release shows the same logic: open the ingredients, keep the product. If other labs follow, weather and climate modelling could get its first genuinely shared data layer.
Background
WeatherNext 3 itself launched on 3 September as Google’s most advanced global weather model, producing hourly forecasts at up to 5-kilometre resolution across 15-day, 64-member ensembles. It is already integrated into Google Search, the Gemini app, Google Maps and the Maps Platform Weather API.
Google has spent the autumn putting its newest models straight into Search, and Gemini 4 Hydrogen became the default AI Mode model this week. This release extends that pattern from search answers to the open data underneath them, with the datasets now sitting on Google Cloud Storage for anyone to use.
Source: DeepMind’s PARDIG and SatWeave announcement and AlphaSignal’s report


0 Comments