Digital twins of the Earth

Also known as Climate digital twins, Earth system digital twins, Destination Earth, Earth-2

Digital twins of the Earth are high-resolution, regularly updated simulations of the climate and weather system, built to test 'what if' scenarios and support planning. The EU's Destination Earth Climate Change Adaptation Digital Twin runs global simulations at about 5 to 10 km, and NVIDIA's Earth-2 platform adds AI emulators whose speed and energy claims come from the company and have no published benchmark.

Pilot and early commercialReported, unverifiedReviewed 21 Oct 2026
AI illustration · Illustrative scene, not a photograph of the specific device or experiment.

In plain language

A digital twin of the Earth is a computer model of the climate and weather system that is detailed enough to show local effects, such as rainfall over one valley or a storm track along one coastline. Climate models split the atmosphere and oceans into a grid of boxes. Making the boxes smaller, down to a few kilometres, shows much more local detail, but it needs far more computing power.

Destination Earth is a European programme run by ECMWF, the European Space Agency and EUMETSAT under European Commission leadership. It aims to provide these fine-scale simulations so that planners can ask questions such as how heat or flood risk might change in their area. NVIDIA's Earth-2 platform takes a different route, using AI models that try to imitate detailed simulations much faster. Independent checks against real observations are still limited in the sources reviewed, and most performance figures come from the projects and companies themselves.

Technical explanation

The Destination Earth Climate Change Adaptation Digital Twin is described as running the atmosphere and land at about 5 km, and the ocean and sea ice at 5 to 10 km (Destination Earth programme page). The paper gives about 4.5 km for the IFS atmosphere and land, 5 km for ICON, and 5 to 8 km for the ocean and sea ice. It reports hourly atmosphere and land output, and daily ocean and sea-ice output (Doblas-Reyes and colleagues, 2026). The programme page describes two kinds of run: multi-decadal climate projections under a given future scenario, and storyline simulations that replay past, present and warmer-climate versions of past weather events. The paper describes storyline runs for 1950, present-day and two-degree warming levels.

The computing cost is high. A 20-year 5 km ICON simulation needs about 0.6 million GPU hours, and a 20-year 5 km IFS-NEMO simulation needs about 23 million CPU core hours. Throughput at 5 km is about 0.5 simulated years per day. The paper says that higher-frequency output requested by some applications is not yet possible.

The paper describes quality control built into the workflow. Monitoring tools provide real-time quality control of data and model outputs, and online model evaluation runs through the AQUA and OBSALL frameworks, which compare outputs with observational networks. AQUA is described as an automated quality-assessment framework for high-resolution global simulations. In a preliminary LUMI test on one month of data, Polytope data access took 153 seconds against 44 seconds for direct FDB access. The AQUA paper does not report validation scores against observations, and comparison with CMIP and CORDEX datasets is listed as ongoing development (Nurisso and colleagues, 2026).

NVIDIA's Earth-2 is a cloud platform for AI emulation, announced on 18 March 2024 at GTC. NVIDIA says its APIs can deliver warnings and updated forecasts in seconds rather than the minutes or hours of CPU-based modelling. The CorrDiff preprint predicts 2 km data from a regional weather model over Taiwan, conditioned on a 25 km global reanalysis. It reports encouraging skill on bulk MAE and CRPS, but the abstract gives no values (Mardani and colleagues, 2023). NVIDIA's cBottle model, announced in June 2025, is in early access. NVIDIA says it can reduce data volumes by up to 3,000 times for an individual weather sample, and the announcement reports no quantitative accuracy evaluation.

What the evidence says

Each statement is typed by the strength of its evidence and linked to the source that supports it.

Demonstrated Shown in a published experiment, test or shipping product.

The Destination Earth Climate Change Adaptation Digital Twin is described as running the atmosphere and land at about 5 km, and the ocean and sea ice at 5 to 10 km. The paper reports hourly atmosphere and land output and daily ocean and sea-ice output.

Sources: [1], [3]

A 20-year 5 km ICON simulation needs about 0.6 million GPU hours, a 20-year 5 km IFS-NEMO simulation needs about 23 million CPU core hours, and throughput at 5 km is about 0.5 simulated years per day.

Sources: [3]

Higher-frequency output requested by some applications is not yet possible in the Climate DT.

Sources: [3]

The AQUA paper describes an automated quality-assessment framework for high-resolution global climate simulations. It does not report validation scores against observations, and comparison with CMIP and CORDEX datasets is listed as ongoing development.

Sources: [4]

Reported, unverified Stated by a company or team but not independently confirmed.

The Climate DT programme page names three models, ICON, IFS-FESOM and IFS-NEMO, and gives a simulation period of 1990 to 2049. The paper says projections run up to 2050 and gives historical runs from 1990 to 2014, so the period is not fully reconciled.

Sources: [1], [3] · medium confidence

The Climate DT paper states that monitoring tools provide real-time quality control of data and model outputs, and that online model evaluation runs through the AQUA and OBSALL frameworks.

Sources: [3] · medium confidence

NVIDIA's 2024 Earth-2 release says CorrDiff generates images 1,000 times faster and 3,000 times more energy-efficiently than current numerical models, and it gives no benchmark for either figure.

Sources: [5] · medium confidence

NVIDIA's June 2025 cBottle announcement describes a kilometre-scale generative model in early access, claims it can reduce petabytes of data by up to 3,000 times for an individual weather sample, and reports no quantitative accuracy evaluation against observations.

Sources: [6]

The CorrDiff preprint trains a model to predict 2 km data from a regional weather model over Taiwan, conditioned on a 25 km global reanalysis. It reports encouraging skill on bulk MAE and CRPS, but its abstract gives no values for these scores.

Sources: [7] · medium confidence

Limitations and open problems

  • Km-scale climate simulation is expensive. The Climate DT paper gives about 0.6 million GPU hours for a 20-year ICON run at 5 km and a throughput of about 0.5 simulated years per day, which limits how many long runs and scenarios can be produced.
  • Higher-frequency output that some applications request is not yet possible (Doblas-Reyes and colleagues, 2026).
  • The Climate DT paper describes online evaluation against observations, but the sources reviewed report no validation scores for km-scale outputs against observations. The AQUA paper does not claim validation, and comparison with CMIP and CORDEX datasets is still in development (Nurisso and colleagues, 2026).
  • The Earth-2 speed claim (1,000 times faster) and energy claim (3,000 times more energy-efficient) are company claims in a 2024 press release, which gives no benchmark. The cBottle data-reduction figure (up to 3,000 times) is also a company claim.
  • The Climate DT paper acknowledges access to the EuroHPC supercomputer LUMI, so long-term operation depends on continued access to public high-performance computers.

Possible future uses

Speculative

Ideas, not predictions. None of these has been demonstrated.

  • Wind energy and wildfire planning using kilometre-scale projections. The Destination Earth programme page names onshore and offshore wind energy and wildfires among its impact sectors.
  • Testing how a past extreme event might have unfolded in a warmer climate, using storyline simulations of the kind the Climate DT describes.

Timeline

  1. 18 Mar 2024NVIDIA announces Earth-2 as a climate digital twin cloud platform, with cloud APIs on DGX Cloud, at its GTC conference. Sources: [5]
  2. 10 Jun 2025NVIDIA announces cBottle, a generative AI model for kilometre-scale climate simulation, available in early access. Sources: [6]
  3. 1 Feb 2026The European Commission confirms the third phase of Destination Earth, running from June 2026 to June 2028, with ECMWF, ESA and EUMETSAT as implementing partners. Sources: [2]
  4. 14 Apr 2026Geoscientific Model Development publishes the paper describing the Destination Earth digital twin for climate change adaptation, including its 5 km simulations. Sources: [3]
  5. 20 Aug 2026Geoscientific Model Development publishes the AQUA v1.0.0 paper, a quality-assessment framework for the Climate Change Adaptation Digital Twin. Sources: [4]

Futures this could help bring about

AI illustration
How far away?

A full digital twin of the Earth

Pilots · 2 of 8 steps done

Related breakthroughs

Glossary

Digital twin
A computer model that is updated with observations and used to test scenarios for a real system.
Downscaling
Producing finer-resolution detail from a coarser model or dataset.
Storyline simulation
A simulation that replays a specific past weather event under different climate conditions, such as a warmer climate.
GPU hour
One graphics processing unit running for one hour, a common unit of computing cost.

Sources

  1. other
    Climate Change Adaptation Digital Twin
    Destination Earth (DestinE), ECMWF, 2026. Accessed 21 Oct 2026.
  2. news
    Third phase of Destination Earth confirmed
    ECMWF, 1 Feb 2026. Accessed 21 Oct 2026.
  3. peer reviewed paper
    The Destination Earth digital twin for climate change adaptation
    Francisco J. Doblas-Reyes. Geoscientific Model Development, 14 Apr 2026. DOI: 10.5194/gmd-19-2821-2026. Accessed 21 Oct 2026.
  4. peer reviewed paper
    AQUA v1.0.0: The Application for QUality Assessment for the Climate Change Adaptation Digital Twin
    Matteo Nurisso, Jost von Hardenberg, Marco Cadau et al.. Geoscientific Model Development, 20 Aug 2026. DOI: 10.5194/gmd-19-7725-2026. Accessed 21 Oct 2026.
  5. company announcement
    NVIDIA Announces Earth Climate Digital Twin
    NVIDIA, 18 Mar 2024. Accessed 21 Oct 2026.
  6. company announcement
    Clear Skies Ahead: New NVIDIA Earth-2 Generative AI Foundation Model Simulates Global Climate at Kilometer-Scale Resolution
    Timothy Costa. NVIDIA Blog, 10 Jun 2025. Accessed 21 Oct 2026.
  7. preprint
    Residual Corrective Diffusion Modeling for Km-scale Atmospheric Downscaling
    Morteza Mardani, Noah Brenowitz, Yair Cohen et al.. arXiv, 24 Sept 2023. Accessed 21 Oct 2026.

Revision history

3 changes since this profile was created.

Changes on 21 Oct 2026
Published on owner instruction (content loop iteration 19)dcc0da0
Sources and claims re-checked against originalsfbb434c
Initial profile drafted from primary sources7e66249