On November 13, 2026, Google DeepMind began pointing developers to two datasets behind its weather model: PARDIG and SatWeave. Its WeatherNext page now lists "Access PARDIG" and "Access SatWeave" under datasets. This post covers what the page and the paper confirm, and what is only claimed on social media.
The release matters for anyone who builds on weather data: energy traders, farm and logistics tools, climate researchers. It also matters for AI-for-science readers, because PARDIG is itself an AI-made data product, not a raw measurement.
TL;DR: claimed vs verified
| Claim | Status |
|---|---|
| PARDIG: 10 km global precipitation maps every 15 minutes | Verified on DeepMind's page |
| SatWeave: mosaic of five geostationary satellites, every 60 minutes, 17 channels | Verified on DeepMind's page |
| PARDIG is an AI model trained on sparse GPM radar data | Verified in the WeatherNext 3 paper |
| PARDIG beats IMERG against radar and rain gauges | Stated in the paper, by the authors |
| Precipitation detection skill up 30 percent | Not verified: only in a relay of social posts |
| Met Office scientists are reviewing WeatherNext 3 | Not verified: no official page found |
What PARDIG and SatWeave are
DeepMind describes PARDIG as "our AI-driven global precipitation reanalysis delivering 10km precipitation maps every 15 minutes." It describes SatWeave as "our seamless global mosaic of five geostationary satellites updated every 60 minutes across 17 channels."
Geostationary satellites sit fixed over one part of the Earth, so each sees a region continuously. Stitching five into one global picture is hard because the instruments differ. SatWeave is the stitched result. Weather models can then read cloud and storm structure hourly instead of waiting for the six-hourly analysis that older systems use.
PARDIG answers a different problem: rain is hard to measure globally. Radar covers only some land areas, and gauges are sparse. PARDIG is an AI estimate of rainfall built to fill that gap.
How they fit in WeatherNext 3
The WeatherNext 3 paper (arXiv 2609.03582) explains the role of each dataset.
- SatWeave's role. The model takes two analysis frames and the 12 most recent geostationary mosaic frames. The mosaic arrives with about one hour of latency, which the authors say is at least five hours fresher than the latest analysis. Feeding satellite data directly lets the model use observations instead of only analysis data, and avoid inheriting the analysis's biases.
- PARDIG's role. The paper says WeatherNext 3 adds two precipitation products to what WeatherNext 2 had. One is PARDIG, "a novel experimental precipitation estimate" based on a separate AI model trained to predict sparse space-borne radar data from NASA's GPM core observatory. The other is NASA's IMERG Final product. Both are at 0.1 degrees. PARDIG is available from 1999, IMERG from June 2000.
The authors say PARDIG matches independent ground radar and rain gauge observations more closely than IMERG. That is the authors' statement, made in the paper. It is a reasonable claim to test with your own region, and it is the reason to read the dataset documentation before relying on it.
Where the numbers come from
The paper evaluates precipitation with CRPS, a score for probabilistic forecasts, against IMERG, the US radar product MRMS and rain gauges. It also reports Brier scores at several rain thresholds, and reliability diagrams that show WeatherNext 3 is better calibrated than earlier global models, which are overconfident in their rain predictions. It also admits artifacts: the paper shows example artifacts in 24-hour PARDIG-head forecasts for single ensemble members.
We did not find a "more than 30 percent" precipitation detection figure in the parts of the paper we read. It circulated in a summary of DeepMind's posts. Until DeepMind states its baseline, region and threshold, treat it as a claim. For the same reason, read any headline gain as dependent on the metric.
The Met Office claim
Social-media summaries say the release comes with a partnership with the UK Met Office, whose scientists are reviewing WeatherNext 3 to blend AI analysis with physics-based forecasting. The quoted line is attributed to a Met Office science account: the model "analyses existing data and weather patterns for insight that complements our physics-based approach."
We could not find an official Met Office or DeepMind page that confirms a formal partnership, and DeepMind's WeatherNext page does not mention it. A scientist reviewing a model is a weaker statement than a partnership. Wait for an official post before you treat this as a deployment.
Background: what came before
WeatherNext 3 was introduced in early September. Our earlier post on the hourly 5 km global forecasts covers the launch: hourly forecasts at 5 km for key surface variables, against a 25 km grid for WeatherNext 2. We also covered DeepMind's cyclone forecasting work, the first place the lab shared more of the stack.
The model page also highlights its use for the National Hurricane Center's forecast of Hurricane Melissa's landfall in Jamaica, and for renewable energy: wind and solar operators can use radiation and cloud cover variables. The datasets are new because they open up the inputs and intermediate products, not only the forecasts.
Others are working on the same problem. Ai2's HiRO-ACEv2 simulates extreme weather at 3 km at a fraction of the cost. Google's own Earth AI geospatial agent uses weather and geospatial data for reasoning tasks. And public-data efforts like Hugging Face's Project Protostar push for AI-ready scientific datasets.
Why AI weather models want raw observations
Most AI weather models learn from analysis data: a gridded best guess of the atmosphere that agencies produce every six hours. The WeatherNext 3 paper names two problems with that. Analyses carry known biases, especially for precipitation and surface temperature. And they are produced every six hours, with observations assimilated up to three hours past the analysis time, plus about three hours to publish. A forecast that starts from them begins already several hours old.
Satellite mosaics fix the second problem. An hourly picture of clouds, water vapor and temperature with about an hour of latency is much fresher. The paper's design feeds the model satellite frames offset by zero to five hours from the analysis, so it can see what changed since the analysis was made. The authors say this lets the model run forecasts every hour of the day. The paper also compares hourly with six-hourly initializations using a latency-adjusted CRPS, which asks a fair question: given real publishing delays, which forecast is actually newest when a decision is made?
Rain is the variable where this matters most, because it changes fast and is badly measured. That is the case for PARDIG: a precipitation target trained on space-borne radar, which measures rain more directly than infrared cloud tops do. It also explains why the paper keeps IMERG predictions in the model. IMERG is widely used, so users can compare against a product they already know.
The trade-off is dependency. A model fed by a satellite mosaic is only as good as the mosaic, and a model trained on an AI-made rain target inherits that target's errors. Treat the datasets as a way to inspect and test the pipeline, not as ground truth.
How to try them
- Open the WeatherNext page and follow "Access PARDIG" and "Access SatWeave."
- Read the license and citation terms on each page. For the earlier WeatherNext 3 Earth Engine data, Google asks users to cite DeepMind's machine learning models, and the terms are CC BY 4.0 style.
- Start with a region where you have ground truth, such as a network of rain gauges. Compare PARDIG with IMERG, and with your own measurements.
- For SatWeave, check the channel list and the time stamps. Hourly updates and roughly one hour of latency mean it suits nowcasting more than long-range work.
- Record the dataset version in your results. These are experimental products.
What this means for builders
- Rain data you can query at 15-minute steps helps flood alerts, logistics routing and crop planning, if PARDIG's accuracy holds in your region.
- AI-made reference data carries model risk. PARDIG is itself a model output. Errors in it can flow into anything trained on it. Keep a gauge-based check in your pipeline.
- A mosaic of five satellites gives a common input for models other than Google's. That lowers the barrier for researchers who could not stitch the sources themselves.
What is still open
- Dataset licenses and file formats, which we could not read from the page text alone.
- The size of PARDIG's gain over IMERG in numbers, by region and rain threshold.
- Whether a Met Office collaboration exists beyond scientist review.
- Whether other agencies will adopt the products operationally.
Related reading
- WeatherNext 3: hourly 5 km global forecasts
- DeepMind WeatherNext cyclone forecasting
- Ai2 HiRO-ACEv2 extreme weather emulator
- Google Earth AI geospatial reasoning agent
- Hugging Face Project Protostar: AI-ready science data
- Anthropic's guide to strong safety evaluations, a useful lens for reading any benchmark claim
- Sources: DeepMind WeatherNext 3 page, WeatherNext 3 paper, arXiv 2609.03582
Details are accurate as of November 13, 2026. Specs are from DeepMind's page and paper; unverified items are labeled.
