Blog · Field notes · Oct 21, 2026 · 5 min
Europe's grid operators, its cities and its climate programme all moved digital twins from plans to infrastructure in 2025-26. The architecture is now well specified; the step that decides whether a twin can be relied on is a prediction scored against what it never measured.
A digital twin used to be a promise in a strategy document. In 2025 and 2026 it became infrastructure. Europe's transmission and distribution operators defined the use cases for a twin of the electricity grid and chose an architecture for it. Cities received an open toolbox for building local twins. The continent's climate twin entered a third phase of routine operation. The engineering questions, how twins connect, who owns the data and which standards they speak, are now well answered. The question that decides whether any of them can be relied on is a different one: does the twin predict what it has not measured, over a period it has not seen?
Europe's electricity network is taking on new loads and new generation in more places than it was planned for. A twin of the grid is meant to let operators see, and plan for, what they cannot measure directly.
In January 2026 ENTSO-E and the EU DSO Entity published a joint report on TSO-DSO digital twin use cases, the second deliverable of their joint task force on the EU's digitalisation of energy action plan. It sets out four: consumer-centric flexibility, network planning and hosting capacity, planning for high-impact low-frequency events, and coordinated security assessment. The report describes itself as a shift "from identifying challenges to defining concrete actions".
A month later ENTSO-E set out the architecture: a federation of independently operated twins forming a system of systems, interoperable across borders "without centralisation or loss of data sovereignty". It calls for common data foundations, lifecycle-aligned modelling, open standards and shared ontologies. The choice is sound. A grid is owned by many parties, and a federated twin matches how it is actually run.
For cities, the European Commission's Local Digital Twins programme reached a milestone in June 2026 with the launch of the EU LDT Toolbox in Valencia: an open-source infrastructure for interoperable local twins with 11 functional tools and a European building database, handed to the LDT CitiVERSE consortium for long-term governance.
Industry has moved in parallel. In June 2025 NVIDIA announced an Omniverse Blueprint for smart city AI, a reference workflow that builds a photorealistic, simulation-ready city twin from aerial, satellite or map data, generates synthetic data to post-train vision models, and deploys agents that monitor camera and sensor feeds.
At planetary scale, Destination Earth moved into its third phase in June 2026, running to June 2028. ECMWF reports that its second phase matured and put into operation the climate adaptation and weather extremes twins, and consolidated workflows, evaluation and uncertainty quantification. Its climate twin replays record heatwaves and floods to show how they would unfold in a world 2°C warmer.
Read side by side, these programmes specify architecture, data governance, interoperability and use cases in detail. Validation is the part they leave for later. ENTSO-E's federation report, on its own summary page, does not discuss how a twin's predictions are to be checked. Destination Earth stands apart because it comes from weather forecasting, a field that scores its predictions against observations every day.
A twin earns trust through four properties, and none of them is architectural.
It predicts what was not measured. A twin that reproduces the meters it was fitted to has shown that it can fit. The test is a feeder without a meter, a district held out, or a year that had not happened when the twin was frozen.
It is scored once, against a rule set in advance. If the comparison is chosen after the result is known, almost any twin can be made to look good.
It contains the people. Grid load, building heat and street traffic are driven by who is home and what they are doing. A twin of the wires without the lives that draw power through them reproduces yesterday's pattern and misses the change that a new policy or a heatwave brings.
It states its range. Every model holds within some range of conditions and fails beyond it. A twin that says where it was tested, and declines to answer outside that, is more useful than one that answers everything.
Twins are being written into how Europe plans grid reinforcement and hosting capacity, as the operators' own use cases show. A twin that is wrong in a consistent direction moves investment to the wrong feeders. A twin that is right, and can show it was right on data it had not seen, lets a small number of measurements go much further.
In Living Worlds, a city is its streets, buildings and the power, water, heat and air that run through them, driven by who is home and what they do. Each place model is checked against real measurements before it is relied on, and used within the range where it held.
Power through 40 real low-voltage feeders that had no meter was predicted every fifteen minutes for a year the model had never seen, from the metered feeders around them: ahead of calibrated physics on 38 of 40 feeders and of a learned model on 30 of 40, and ahead of both in every season (Met). A model that first learned from Living Worlds' simulation and then from a handful of real measurements predicted 30 unmeasured feeders better than the same model trained on real measurements alone, 0.712 normalised error against 0.880 with five measured feeders, level with eighty; taught by a deliberately wrong simulation, it did worse than either (Met). That is one region and one operator, and the study says so.
Inside buildings, daylight on six desks of a real office was predicted every ten minutes for eleven unseen days with an error of 58.1 lux, against 115.3 for calibrated physics (Met). Thermal comfort across 13,451 votes in 15 countries came closer than the standard design index, but a learned classifier came out slightly ahead, so the registered bar was missed (Not met). Each was registered before the data was read and scored once.
Sources
Blog
A feeder model that learned from Living Worlds' checked simulation first, then from a handful of real measurements, predicted 30 unmeasured low-voltage feeders better than the same model on real measurements alone; power through 40 unmetered feeders beat physics and a learned model; thermal comfort in fifteen countries came close but was not met.
Living Worlds predicted how real rooms sound in arrangements and positions nobody had measured, and how daylight falls on office desks on days nobody had read: met for a room never seen (20 of 20), four rooms never seen (4 of 4), sound beyond what was measured and daylight at the desk; not met in a second room.
World models now generate explorable places that hold together for minutes. The next frontier is behaviour over years, many lives at once, and results checked against a reality the model has never seen.