Google DeepMind and Google Research have openly released two foundational Earth science datasets—SatWeave and PARDIG—designed to eliminate data harmonization bottlenecks for artificial intelligence weather forecasting and climate modeling. Announced alongside an arXiv technical preprint, the release provides multi-decade historical archives and continuous near-real-time streams under an open Creative Commons Attribution 4.0 International (CC BY 4.0) license. Both datasets are hosted in cloud-native Zarr format on Google Cloud Storage and integrated into Google Earth Engine, giving university researchers unrestricted access to the observational backbones behind DeepMind’s WeatherNext 3 forecasting system.
Harmonizing geostationary satellite records with SatWeave
Atmospheric science research has long struggled with the fragmented nature of geostationary weather satellite feeds. Operational satellites operated by different international agencies—including NOAA GOES-East and GOES-West, EUMETSAT Meteosat Prime and Meteosat Indian Ocean, and the Japan Meteorological Agency’s Himawari—rely on disparate orbital geometries, sensor calibrations, file formats, and scanning intervals ranging from 10 to 30 minutes. Traditional global composites typically discard spectral information down to four or fewer bands and update only every few hours.
SatWeave resolves these discrepancies by blending observations from five operational geostationary positions into a single analysis-ready mosaic across 17 spectral channels. The grid encompasses seven visible, near-infrared, and shortwave-infrared reflectance bands alongside ten infrared and water vapor brightness temperature channels. The dataset provides global coverage at 0.1-degree (approximately 10-kilometer) spatial resolution. In addition to a ten-year historical archive spanning 2016 through 2025 sampled at 30-minute intervals, Google DeepMind is publishing an hourly live operational feed with roughly one hour of latency. The data adheres to standard CF Conventions and is accessible directly via public Google Cloud Storage buckets without authentication gates, interfacing smoothly with atmospheric reanalysis products from ECMWF.
High-resolution global rainfall mapping with PARDIG
Alongside SatWeave, DeepMind released PARDIG (Precipitation AI Reanalysis: Densely Inferred GPM), an AI-driven global precipitation reanalysis model and dataset. Ground radar networks and rain gauges provide high-fidelity precipitation records across North America, Western Europe, and parts of East Asia, but massive observation gaps persist over oceans, mountainous regions, and the Global South. While spaceborne radar on the NASA/JAXA Global Precipitation Measurement (GPM) Core Observatory captures high-precision vertical precipitation profiles, its orbital swath samples any given location on Earth only intermittently.
To bridge this observational gap, PARDIG applies a spherical graph-mesh transformer architecture trained directly on spaceborne radar retrievals from the GPM Combined Radar-Radiometer Algorithm (CORRA). Using SatWeave satellite imagery and ECMWF atmospheric fields as multi-modal inputs, PARDIG reconstructs complete global precipitation fields every 15 minutes at 0.1-degree spatial resolution. DeepMind published two variants: PARDIG-LR, an offline historical reanalysis covering 1999 to 2023, and PARDIG-NRT, an operational stream running from 2016 to the present with an average latency of approximately five hours.
In validation against continental radar networks in the United States, Europe, and Japan and more than 36,000 surface weather stations, PARDIG outperformed established global precipitation benchmarks, including NASA’s IMERG V07 (Final and Early runs), MSWEP V3, and ERA5 reanalysis. Crucially, the model improved hourly precipitation detection skill (measured by the Critical Success Index) by more than 30 percent for light-to-moderate rain while maintaining superior fidelity during extreme convective storm events. Both the historical catalog and near-real-time records are published in public cloud buckets, functioning as an open data repository for the global hydrological community.
What open observational data means for university research
The open release of SatWeave and PARDIG marks an important shift in commercial AI research practices. In August 2026, peer-reviewed evaluation of DeepMind’s cyclone forecasting model demonstrated that machine learning architectures can deliver actionable tropical storm warnings up to 24 hours earlier than traditional numerical models (see our report on the WeatherNext Nature paper). However, academic atmospheric scientists and university research offices have repeatedly warned that proprietary AI weather systems risk creating an unbridgeable computational and data divide between private tech giants and public institutions.
Training frontier weather models requires petabyte-scale data engineering pipelines that exceed the computing budgets of most university atmospheric science departments. By releasing pre-harmonized, analysis-ready satellite and precipitation grids under a permissive CC BY 4.0 license, DeepMind aligns its observational infrastructure with established FAIR data principles. Researchers can now train specialized regional models for wildfire smoke tracking, solar irradiance forecasting, flood inundation modeling, and drought monitoring without having to spend months wrangling raw satellite telemetry. Furthermore, because PARDIG historical catalogs are indexed directly in Google Earth Engine and stored in cloud-native Zarr format on Google Cloud Storage, academic teams can conduct planetary-scale analyses without downloading hundreds of terabytes to local campus servers.
Sources
- Google DeepMind: Opening the datasets behind our AI weather models (2026-11-13).
- arXiv: PARDIG: Global Precipitation Estimates Learned from Sparse Satellite Observations, Shalev et al. (2026-11-10).
- Google for Developers: WeatherNext Overview and Observational Datasets Documentation (2026-11).
