Opening the datasets behind our AI weather models
Google DeepMind's WeatherNext team announced the open release of two datasets, SatWeave and PARDIG, that were used to build its recently launched WeatherNext 3 weather model. Historical archives and live data feeds for both are now available through Google Cloud Storage, and PARDIG's historical catalogs are additionally offered in Google Earth Engine. The stated goal is to help outside researchers and developers build weather, hydrology and climate applications, not just to document the model.
SatWeave is a global mosaic of geostationary satellite imagery. It pulls from five satellites covering different regions: GOES-East and GOES-West, Meteosat 0 and Meteosat Indian Ocean, and Himawari. Combining them is nontrivial because each satellite uses different map projections, file formats, sensors and scan schedules that range from 10 to 30 minutes. Google says most existing global composites handle this by discarding incompatible aspects, offering four or fewer channels or refreshing only every few hours. SatWeave instead blends observations from six geostationary satellites across 17 visible, water vapor and infrared channels.
The mosaic updates hourly, which matters for forecasting. WeatherNext 3 can therefore be initialized with a near-live view of developing clouds and storms, whereas models initialized on conventional analysis data see refreshes only every six hours. Google is releasing both the hourly live feed and a historical archive spanning more than a decade. The company used SatWeave for medium-range forecasts but flagged solar power forecasting and wildfire monitoring as areas it wants the wider community to explore.
PARDIG, short for Precipitation AI Reanalysis: Densely Inferred GPM, tackles global rainfall estimation. Precipitation drives food and water supply and shapes flood and drought risk, but ground radars and dense rain gauge networks are unevenly distributed and unavailable in many regions. PARDIG uses an AI architecture similar to WeatherNext 3 and is trained directly on sparse spaceborne radar observations from NASA's Global Precipitation Measurement (GPM) Core Observatory. Its inputs are SatWeave imagery and atmospheric fields from ECMWF.
The output is a global precipitation map at 15-minute time steps and roughly 10-kilometer spatial resolution, with an average latency of about five hours. Google evaluated it against three continental radar networks in the US, Europe and Japan, plus more than 36,000 weather stations worldwide, and reports that PARDIG beats leading global baselines across all analyzed rain rates, including NASA's IMERG-Final, a research-quality product widely used in applications. On hourly precipitation detection skill, PARDIG improves on IMERG-Final by more than 30% for light-to-moderate rainfall, according to the announcement, and Google says it was particularly effective at extreme precipitation, which existing global products can struggle to detect.
Because the model learns from satellite observations rather than depending primarily on ground measurements, Google expects reliable accuracy essentially everywhere, including places without dense gauges or radar. For a developer, the practical distinction is resolution and cadence: a 10-kilometer, 15-minute product is far finer than daily or coarse-grid precipitation datasets, though the roughly five-hour latency means it suits near-real-time monitoring rather than instant alerting at minute scale. The comparison examples in the announcement contrast PARDIG against NOAA's ~1 km MRMS radar during the July 2025 Texas Hill Country flash floods, a reminder that ground radar remains finer where it exists.
Google is releasing PARDIG historical archives covering 1999 to 2026 alongside a live feed. WeatherNext 3 used PARDIG for medium-range forecasting; Google says it wants partners to explore flood, landslide and drought models, particularly in regions lacking dense radar coverage. The datasets are described as the latest addition to Google's Earth AI portfolio. The work involved teams across Google DeepMind and Google Research with Google Cloud and Google Geo, and credits open data from ECMWF, NASA, JAXA, NOAA, EUMETSAT, KMA, JMA, KNMI and Synoptic Data.
A likely trade-off is that the two products are tuned for the model they were built for: SatWeave's value depends on how well a downstream system can consume its 17 channels and hourly cadence, and PARDIG's skill numbers come from Google's own benchmark choices. Researchers comparing against IMERG-Final should verify performance in their own region and rain-rate regime. Access through Cloud Storage and Earth Engine also means usage is tied to Google's cloud tooling, an inference about convenience and lock-in rather than a stated policy.
Why it matters: AI weather teams have increasingly competed on models while treating training data as proprietary, so publishing the observation and precipitation layers gives outside researchers a way to reproduce and extend WeatherNext 3-style work. Hydrology, insurance, agriculture and disaster-response teams in radar-poor regions gain an hourly satellite mosaic and a decade-plus precipitation archive they can query directly. For developers, the near-term work is building downstream pipelines, such as flood or solar forecasting, on these feeds, though the real test will be independent validation of PARDIG's claimed gains beyond Google's benchmarks.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Foundational precipitation and geostationary satellite datasets to observe the planet clearly.