Fast-moving clouds can change the output of a solar farm before a conventional forecast is refreshed. Nvidia’s 21 October account of European Earth-2 projects highlights StormScope EU, which uses 16 Meteosat satellite channels sampled every ten minutes and produces forecasts at ten-minute intervals. An Earthmover demonstration turns the output into frequently updated map tiles.[1]
The announcement brings together several uses of the same model family. Short-term cloud prediction supports immediate weather decisions; simulated cyclone seasons support insurance risk analysis; regional adaptations target air quality and solar generation. Their data, time horizons and stages of deployment differ.[1]
Cadence is one part of a useful forecast
Ten minutes describes the sampling and output interval, not a guaranteed warning time before hazardous weather. Forecast horizon, processing delay and predictive skill remain separate measures. A frequently refreshed cloud map is useful when its local predictions improve the decision an operator has to make.[1]
Developers can combine short-term StormScope models with CorrDiff, which adds spatial detail to coarser weather information. Earth2Studio supplies tooling for building forecasting workflows around these components. Organisations can adapt a model to regional data and operational requirements instead of constructing the entire forecasting system from scratch.[1][2][3]
Earth-2 applications announced in Europe
| Organisation | Application | Reported stage / limit |
|---|---|---|
| Earthmover / Nvidia | StormScope EU map-tile demo | Live demonstration; ten-minute updates |
| AXA | Cyclone scenarios across >1,000 simulated seasons | Risk analysis, not real historical seasons |
| Vaisala Xweather | Lightning-enhanced short-term forecasts | Combining products and proprietary data |
| University of Manchester | UK air-quality modelling with CorrDiff | Model adaptation and local execution |
| TotalEnergies | Solar-radiation nowcasts at two sites | Evaluation, not a stated commercial outcome |
Different customers ask different questions
AXA’s use of FourCastNet 3 spans more than 1,000 simulated cyclone seasons. Such ensembles can explore combinations absent from a short observed record. The relevant question is whether the generated extremes represent risk well enough for the insurer’s analysis, rather than whether a model has forecast that many real seasons.[1]
The University of Manchester example shows another part of the workflow: researchers trained their UK air-quality adaptation of CorrDiff on Isambard-AI, then ran it locally on DGX Spark systems. Large-scale model development and smaller-scale execution can serve different stages of the same regional application.[1]
The applications also supply different tests of usefulness. At TotalEnergies’ two photovoltaic evaluation sites, the question concerns solar-radiation nowcasts; at AXA, it concerns the realism of rare-event scenarios. A common model toolkit does not remove those application-specific demands.[1][3]
The shift is towards forecasts that fit a particular place and decision. Ten-minute cloud updates, local execution and reusable model components make that adaptation easier. The results still have to earn their value in the forecast errors, risk estimates or operating decisions each user measures.
Sources & further reading
- Pioneers Across Europe Deploy NVIDIA Earth-2 Open Models to Power Weather and Climate Prediction NVIDIA Blog · 21 Oct 2026 · Accessed 26 Oct 2026
- NVIDIA Earth-2 models and tools NVIDIA · Accessed 26 Oct 2026
- Earth2Studio forecasting examples NVIDIA / Earth2Studio · Accessed 26 Oct 2026