NVIDIA has added a Europe-specific nowcasting model to its open Earth-2 weather family. StormScope EU, announced at GTC Berlin on October 22, 2026, is a regional version of the StormScope model for short-term forecasting, trained on European weather and climate data. NVIDIA says it uses EUMETSAT satellite observations to predict evolving cloud conditions across 16 Meteosat channels at 10-minute intervals. The details are in NVIDIA's Earth-2 Europe announcement.
If you build with weather data for energy, insurance, shipping or research, this is worth understanding for two reasons: the models are described as open and customizable, and the post names real organizations using them. It sits alongside Ai2's HiRO-ACEv2 extreme-weather emulator and Microsoft's open-source bias correction for subseasonal forecasts in a fast-moving open weather-AI field.
TL;DR
| Question | Answer |
|---|---|
| What is new? | StormScope EU, a regional StormScope trained on European data |
| What does it predict? | Evolving cloud conditions across 16 Meteosat channels |
| How often? | Every 10 minutes, updated as new satellite observations arrive |
| Data source | EUMETSAT satellite observations |
| Is it open? | NVIDIA describes Earth-2 as open models and tools customizable with your own data |
| Tools | Earth2Studio for forecasts, open source PhysicsNeMo for training |
| Live demo | At GTC Berlin with Earthmover, served as map tiles |
| Named users | AXA, Vaisala Xweather, University of Manchester, TotalEnergies, Space42 |
What StormScope EU does
A nowcast predicts the next minutes to hours, not days. Global weather models give a broad forecast, but, as NVIDIA's post puts it, decisions about solar farms, shipping routes or insured property depend on what happens locally. StormScope EU targets that gap.
Instead of numerical physics, the model works from satellite imagery. Meteosat observes Earth in multiple spectral channels; StormScope EU takes the 16 channels and predicts how cloud conditions evolve. Because new satellite observations arrive every ten minutes, the system can refresh at the same cadence.
NVIDIA is showing a live pipeline at GTC Berlin built with Earthmover. The demo uses Earth2Studio to run forecasts and Earthmover's data platform to manage and serve results as map tiles, so visitors can explore predictions across Europe without downloading whole datasets. That serving pattern, tiles instead of bulk files, is a practical template for anyone publishing weather-AI output in a web app.
The Earth-2 toolkit in plain terms
NVIDIA names several pieces, and it helps to separate their jobs.
| Component | Role according to NVIDIA |
|---|---|
| StormScope / StormScope EU | Short-term forecasting, regional version trained on Europe |
| CorrDiff | Adds local detail to coarser weather data |
| FourCastNet 3 | Used by AXA to simulate regional cyclone risk beyond the historical record |
| Earth2Studio | Runs forecasts and example workflows |
| PhysicsNeMo | Open source framework for training and customization |
Developers can combine models: StormScope for the near-term evolution and CorrDiff to sharpen resolution. They can also reuse the architectures for new environmental problems such as air-quality modeling, which is exactly what one university team did.
Who is using it and for what
NVIDIA's post lists five users. These are NVIDIA's descriptions, not independent evaluations.
- AXA uses FourCastNet 3 to study regional cyclone risk beyond the historical record across more than 1,000 simulated seasons, generating insights for risk assessment and prevention.
- Vaisala Xweather combines Earth-2 with its proprietary observations, lightning data and forecasting expertise, adapting StormScope with lightning data to extend short-term forecast lead time for government, energy, maritime, transportation and insurance customers.
- University of Manchester adapted CorrDiff for UK air-quality modeling, trained it on the Isambard-AI supercomputer, and ran it locally on DGX Spark systems.
- TotalEnergies is exploring StormScope-based solar-radiation nowcasts at two photovoltaic sites and building subhourly models.
- Space42 built SARDiff4Height with CorrDiff and PhysicsNeMo to estimate surface heights from a single radar image and produce uncertainty maps.
The Manchester example matters for developers: training happened on a national supercomputer but inference ran on a desktop-class machine. The post does not give accuracy numbers for that model.
What NVIDIA does not say
The announcement is a platform and customer story, not a benchmark paper. Several things readers will want are missing from the post:
- Skill scores. No verification metrics against persistence or physical nowcasting baselines for StormScope EU.
- Lead time. The post gives the 10-minute update interval but not how far ahead the model forecasts.
- Licensing. "Open" is NVIDIA's description; check each model's license on its page before commercial use.
- Precipitation. The model predicts cloud conditions; the post does not claim rainfall or severe-storm warnings.
- Hardware needs. Beyond the Manchester DGX Spark example, there are no requirements listed.
For anyone planning production use, ask for verification against a standard nowcasting baseline over your region and season. Cloud forecasts are a means to an end such as solar irradiance, and errors compound when you convert one to the other.
Nowcasting, forecasting and downscaling: three different jobs
Newcomers often blur these terms, and the Earth-2 pieces map onto them cleanly.
- Forecasting predicts the atmosphere hours to days ahead, usually on a coarse global grid. AI models in this class learn from decades of reanalysis data and run far faster than physics solvers.
- Nowcasting predicts the next minutes to a couple of hours, usually from observations such as satellite and radar. Speed of refresh matters more than global coverage. This is where StormScope EU sits.
- Downscaling takes a coarse field and adds local detail, for example turning a 25 km grid into one that resolves a valley. CorrDiff is the tool NVIDIA names for this.
A realistic solar-energy workflow chains all three: a coarse forecast sets the day, a nowcast tracks the cloud front that will cross the plant in the next hour, and downscaling resolves the site. TotalEnergies' described work, evaluating StormScope-based solar-radiation nowcasts at two photovoltaic sites, is the nowcast step in that chain.
Which industries should pay attention
| Sector | Why a 10-minute cloud nowcast helps | What to validate first |
|---|---|---|
| Solar and grid operators | Output swings as cloud banks pass; better short-term estimates reduce balancing cost | Cloud-to-irradiance error at your sites |
| Insurers | Fast-forming storms drive claims; nowcasts support alerts and triage | Lead time and false-alarm rate |
| Shipping and aviation | Local cloud and storm development affects routing | Coverage near coasts and over the sea |
| Researchers | Open models and tools allow retraining on new variables such as air quality | Licensing and reproducibility |
| Weather-data companies | Combine satellite nowcasts with proprietary sensors such as lightning | Whether the added data actually extends lead time |
The last row is the Vaisala Xweather approach. NVIDIA says the company is adapting StormScope with lightning data to advance short-term forecasting lead time, which is a concrete example of the open-model value proposition: a specialist adds data nobody else has, and gets a product competitors cannot copy by downloading the same base model.
How to try it
- Start with Earth2Studio. NVIDIA points to its forecasting examples as the way to run models end to end.
- Pull the model into your own workflow. Combine StormScope with CorrDiff when you need local detail.
- Fine-tune with your data. PhysicsNeMo is the open source framework for training and customization. Vaisala's lightning-data adaptation is the pattern to copy.
- Serve results as tiles. The Earthmover demo shows how to expose forecasts without distributing large files.
- Check satellite data access. EUMETSAT is the source of the Meteosat observations; confirm access terms for operational feeds on the EUMETSAT site.
A note on the Space42 and Manchester examples
Two of the named projects are not forecasts at all, which shows how far the toolkit stretches. Space42's SARDiff4Height estimates surface height from a single radar image and outputs uncertainty maps, so analysts can see where the estimate deserves a second look. That uncertainty output is the part to copy: a diffusion model that reports where it is unsure is far more useful in operations than one that returns a single confident field. Manchester's air-quality adaptation shows the other trick, reusing a weather architecture for a neighboring physical problem with a different training set.
Why this matters
Weather AI has moved from global medium-range forecasting to specialized, locally tuned products. A 10-minute European cloud nowcast is directly useful for solar balancing, where a passing cloud bank changes output in minutes, and for insurers watching fast-developing convective storms. NVIDIA's pitch is that open models plus customization let regional specialists build products on top instead of relying on a single national model.
It is also part of NVIDIA's broader European push this week. Our coverage of the GTC Berlin keynote, of NVIDIA's Europe AI-in-science pledge, and of its UK agentic-safety compute investment shows a consistent strategy: seed open models and datasets, then sell the compute they run on. Treat the customer claims as marketing until validated, and watch for peer-reviewed scores on StormScope EU.
Frequently asked questions
What is StormScope EU? A regional nowcasting model on EUMETSAT data predicting cloud conditions across 16 Meteosat channels every 10 minutes.
Is it open? NVIDIA calls Earth-2 open; verify licenses.
Who uses it? AXA, Vaisala Xweather, University of Manchester, TotalEnergies and Space42, per NVIDIA.
Does it replace physics models? The post does not claim that; it positions the models as complementary tools.
Related reading
- Ai2 HiRO-ACEv2 3 km extreme weather emulator
- Microsoft probabilistic bias correction for subseasonal weather
- GTC Berlin 2026 keynote: everything announced
- NVIDIA Europe AI-in-science pledge
- NVIDIA UK agentic safety compute investment
- NVIDIA Cosmos data factory at GTC Berlin
- Official: NVIDIA: Earth-2 open models across Europe · EUMETSAT
Details are from NVIDIA's blog post of October 22, 2026; model availability and licenses may change.