A full digital twin of the Earth

How far are we from an interactive digital twin of the Earth that is used for policy decisions?

The EU's Destination Earth Climate Change Adaptation Digital Twin runs global simulations at about 4.5 to 5 km for the atmosphere and land, and about 5 to 8 km for the ocean and sea ice. Its paper describes an operational system, and the third programme phase is confirmed to run from June 2026 to June 2028. The sources reviewed do not show validation scores for these outputs against observations, a named policy decision that used them, or a benchmark for NVIDIA's AI emulator speed and energy claims. Interactive, on-demand answers to policy questions are a stated design goal, not a shown result.

AI illustration · Illustrative scene, not a photograph of the specific device or experiment.
WHERE WE AREPilots

Close. Early pilots are running in the real world.

The Climate DT is described as an operational, quality-controlled workflow for multi-decadal global climate projections at about 5 km. The sources reviewed do not show a named decision that used its outputs, and the AI emulators that would speed up the system have no published benchmark. Pilot fits this evidence. It is not scaling, because no sourced evidence shows routine use by decision-makers.

The tech tree

Everything recorded on Vrulum that this future needs, oldest foundations at the top and the goal at the bottom. Lines show what each step requires.

What it takes: A full digital twin of the Earth Unlocked Being researched Locked
  1. 2023AI weather forecastingPilot and early commercial Unlocked.
  2. 2024Digital twins of the EarthPilot and early commercial Unlocked.
  1. 2024Funded second phase of Destination EarthDone Unlocked.
  2. 2025AI emulators at kilometre scaleIn progress Being researched.
  3. 2026Global km-scale climate simulationsDone Unlocked.
  4. 2026Funded third phase of Destination EarthIn progress Being researched.
  5. Date not recordedMulti-decade climate projections and storyline runsIn progress Being researched.
  6. Date not recordedOutput frequency and throughputIn progress Being researched.
  7. Date not recordedQuality control and validation against observationsIn progress Being researched.
  8. Date not recordedInteractive, on-demand use for policy decisionsNot started Locked.
  1. GoalA full digital twin of the EarthGoalRequires: Digital twins of the Earth, AI weather forecasting, Global km-scale climate simulations, Funded second phase of Destination Earth, Funded third phase of Destination Earth, Multi-decade climate projections and storyline runs, Output frequency and throughput, Quality control and validation against observations, AI emulators at kilometre scale, Interactive, on-demand use for policy decisions. Locked.

The road so far

2 of 8 steps done, 5 in progress.

  1. Done · 14 Apr 2026Global km-scale climate simulations

    The Climate DT paper describes global simulations at about 4.5 to 5 km for the atmosphere and land, and about 5 to 8 km for the ocean and sea ice. The paper says the workflow ran on the LUMI and MareNostrum 5 pre-exascale supercomputers. [4][1]

  2. Done · Jun 2024Funded second phase of Destination Earth

    ECMWF's announcement of 13 December 2023 gives a 24-month phase starting in mid-June 2024. ECMWF, ESA and EUMETSAT implement it under the leadership of the European Commission's DG CNECT. [2]

  3. In progress · Jun 2026Funded third phase of Destination Earth

    The European Commission confirmed phase three on 1 February 2026. It runs from June 2026 to June 2028. The announcement gives no funding amounts and no resolution target for the system as a whole. [3]

  4. In progressMulti-decade climate projections and storyline runs

    The programme page describes multi-decadal projections under future scenarios, and storyline runs that replay past events in a warmer climate. The paper gives projections for 2015 to 2040 under SSP3-7.0 and historical runs from 1990 to 2014. The programme page gives a simulation period of 1990 to 2049. The sources do not fully reconcile these periods. [1][4]

  5. In progressOutput frequency and throughput

    Atmosphere and land output is hourly, and ocean and sea-ice output is daily. The paper says higher-frequency output requested by some applications is not yet possible. Throughput at 5 km is about 0.5 simulated years per wall-clock day, against a target of about 1. [4]

  6. In progressQuality control and validation against observations

    The paper says validation needs to be treated as a statistical process, including comparison with reality whenever possible. Its AQUA tool compares outputs with reanalyses or benchmark runs, and OBSALL compares them with observed climatologies. Figure 4 shows example biases against surface stations and ERA5. The sources reviewed give no validation scores for km-scale outputs against observations. [4]

  7. In progress · 10 Jun 2025AI emulators at kilometre scale

    NVIDIA's CorrDiff preprint predicts 2 km data over Taiwan from a 25 km global reanalysis, and reports encouraging skill on MAE and CRPS, but its abstract gives no values. NVIDIA's 2024 release claims CorrDiff produces 12.5x higher resolution images than current numerical models 1,000x faster and 3,000x more energy efficiently, with no benchmark. cBottle is in early access, and NVIDIA's 2025 post claims accuracy without giving any metrics. [7][5][6]

  8. Not startedInteractive, on-demand use for policy decisions

    The paper's abstract says the system is designed to support on-demand responses to policy questions, with quantified uncertainty. The sources reviewed do not describe a planning or policy decision that used its outputs. [4]

Why it matters

A working replica could let planners test local heat, hydrology, wildfire and wind-energy risks before they commit to costly adaptation. Whether its outputs can be trusted at this scale depends on validation that the sources do not yet report. The gap between a simulation that runs and a tool that decision-makers rely on is the main thing to watch.

What stands in the way

  • Each 20-year 5 km run is costly. The paper gives about 0.6 million GPU hours for an ICON run and about 23 million CPU core hours for an IFS-NEMO run, at a throughput of about 0.5 simulated years per day.
  • Higher-frequency output that some applications request is not yet possible.
  • Validation of km-scale outputs against observations is not reported with scores in the sources reviewed.
  • The funded phases in the sources end in June 2028, and no funding amounts are given. Continued operation beyond that is not shown.
  • The AI emulator claims on speed, energy and data reduction have no published benchmark, and accuracy is not quantified in the sources reviewed.
  • The programme's simulation period (1990 to 2049) and the paper's projection period (2015 to 2040) are not reconciled, so users cannot yet be sure which time horizon is supported.

Signals to watch

  • Published validation scores for km-scale outputs against observations.
  • A named planning or policy decision that cites Destination Earth outputs.
  • Funding announced beyond June 2028 for the Climate Change Adaptation Digital Twin.
  • Throughput of about 1 simulated year per day at 5 km, the paper's stated target.
  • Independent benchmarks of NVIDIA's emulator claims for speed, energy and accuracy.
  • Higher-frequency output made available to applications that currently request it.

Sources

  1. other
    Climate Change Adaptation Digital Twin
    Destination Earth (DestinE), ECMWF, 2026.
  2. news
    Phase two of Destination Earth confirmed
    ECMWF (DestinE), 13 Dec 2023.
  3. news
    Third phase of Destination Earth confirmed
    ECMWF, 1 Feb 2026.
  4. peer reviewed paper
    The Destination Earth digital twin for climate change adaptation
    Francisco J. Doblas-Reyes. Geoscientific Model Development, 14 Apr 2026.
  5. company announcement
    NVIDIA Announces Earth Climate Digital Twin
    NVIDIA, 18 Mar 2024.
  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.
  7. preprint
    Residual Corrective Diffusion Modeling for Km-scale Atmospheric Downscaling
    Morteza Mardani. arXiv, 24 Sept 2023.

Last reviewed 21 Oct 2026.