Digital Twins: Simulating the Physical World in Software

A dashboard shows you what a machine is doing. A digital twin shows you what it will do if you change something. That distinction — a live, data-fed model you can run forward — is what separates the real thing from the marketing term, and it's why digital twins have become one of the most useful ideas in industrial software even as the phrase gets attached to any 3D rendering with a sensor feed.

What makes a twin a twin

Three ingredients have to be present:

• A model of the physical asset or system — geometry, but more importantly behaviour: the physics of a turbine, the thermodynamics of a building, the flow rules of a supply chain, the traffic dynamics of a city.
• A live data connection from sensors, control systems and business systems that keeps the model synchronised with the real state, at whatever cadence matters — milliseconds for a machine, hours for a port.
• The ability to run scenarios: simulate a change, predict an outcome, optimise a decision, and — in closed-loop deployments — push the result back to the physical system.

Miss the first and you have monitoring. Miss the second and you have a simulation. Miss the third and you have a very expensive diagram.

Where they deliver

• Rotating machinery and energy assets. The original use case. GE, Siemens and others twin turbines, generators and compressors to predict failures, schedule maintenance on condition rather than calendar, and tune operation for efficiency. Savings from avoided unplanned downtime are the clearest ROI in the field.
• Manufacturing. A twin of a production line lets you test a new layout, a changed cycle time or a new product introduction virtually before touching the floor. BMW, Siemens and others run entire factories in simulation, increasingly on platforms like NVIDIA Omniverse that support physically accurate, multi-vendor models.
• Buildings and infrastructure. Twins of hospitals, airports and data centres optimise energy and space; twins of bridges and tunnels track structural health against sensor data. Construction uses them to compare as-built against design in near real time.
• Cities and utilities. Singapore, Helsinki and others maintain city-scale twins for planning, flood modelling and traffic; water and electricity networks use them to locate leaks and balance load.
• Healthcare. Patient-specific twins — a model of an individual's heart built from imaging, used to plan an intervention — are moving through clinical validation. Organisational twins of hospitals help with capacity planning.

How they're built

The stack has settled into recognisable layers. At the bottom, IoT data ingestion from PLCs, sensors and SCADA systems, often via an edge gateway that filters and buffers. Above that, a data platform that stores time series and asset hierarchies. The model layer combines physics-based simulation (finite-element, computational fluid dynamics, discrete-event) with machine-learning models trained on the asset's own history — hybrid modelling is the norm because pure physics is too slow to run live and pure ML doesn't extrapolate to conditions it hasn't seen. On top sit the applications: visualisation, alerting, optimisation and what-if tools. Interoperability standards — Asset Administration Shell in manufacturing, OPC UA for industrial data, BIM/IFC for buildings, and the Digital Twin Consortium's reference work — are what make twins composable rather than bespoke.

What goes wrong

Digital-twin projects fail in predictable ways. Data integration is underestimated every time: sensor feeds are messy, asset registries are wrong, and the twin is only as good as the ground truth it syncs to. Model fidelity is mismatched to the decision — a twin built for design detail is too slow to run operationally; a lightweight operational twin can't answer engineering questions. Ownership falls between operations, IT and engineering. And security is an afterthought: a twin with write access to physical control systems is an attack path, and twins aggregate exactly the operational data an adversary would want — see IoT Security.

Where it's going

Two convergences are changing the field. The first is AI: foundation models trained on physics and sensor data are making twins cheaper to build and better at generalising, and language interfaces are letting operators query a twin in plain English. The second is spatial computing: viewing a twin overlaid on the real asset through AR closes the gap between model and reality for the people doing maintenance. The destination is a twin that's less a project and more a standard layer of any serious physical operation — the way a CMDB became standard for IT.