How digital twins work

01 A MODEL WITH A LIVE COUNTERPART

A static CAD file describes a designed object. A simulation explores what a model does under assumed conditions. A digital twin adds a third ingredient: a continuing relationship with a particular physical system. Measurements from the real asset update the computational counterpart, and the counterpart produces information that can guide operation, maintenance, or redesign.

The word digital is less important than the word twin. The model must be about an intended or actual object, process, or system, and it must preserve enough identity and context to remain useful across time. A generic jet-engine model can answer design questions; a twin of one engine can incorporate that unit’s vibration history, repairs, and operating environment.

02 THE DIGITAL THREAD IS THE MECHANISM

The twin stays alive through a digital thread: the identifiers, data pipelines, interfaces, and records that connect the physical asset to the model. Sensors measure temperature, pressure, vibration, flow, position, or energy use. Gateways timestamp and transport those values. Software cleans them, aligns them with a common clock, and maps them to the variables used by the model.

Without the thread, the twin becomes an ordinary simulation. The distinction is operational, not cosmetic. A model that is updated once a year may be valuable, but it is not answering the same question as a model updated every second. The required update rate depends on the system’s time scale: building energy may tolerate minutes, while a control loop may require milliseconds.

03 STATE ESTIMATION FILLS THE GAPS

Sensors do not expose every internal variable. They are noisy, they drift, and they are placed where installation is practical rather than where mathematics would prefer. A twin therefore estimates hidden state. The model predicts how the system should evolve, then a filter combines that prediction with the newest measurement.

The Kalman filter is a canonical example. It maintains an estimate and an uncertainty, advances the estimate through a system model, and corrects it when observations arrive. Extensions handle nonlinear dynamics or represent multiple hypotheses. The result is not a perfect mirror; it is a continuously revised belief about the physical system, with confidence attached.

04 PHYSICS AND MACHINE LEARNING SHARE THE JOB

A twin can contain first-principles equations, empirical relationships, machine-learning models, or a hybrid of all three. A thermal twin may use heat-balance equations. A structural twin may use finite-element analysis. A maintenance twin may learn how vibration signatures precede a failure. The choice depends on what is known, how quickly the answer is needed, and what evidence is available.

Physics gives a model a shape outside the data it has seen. Machine learning can capture effects that are difficult to write down, but it can also fail when operating conditions move beyond its training distribution. Hybrid models use physical constraints to limit learned behavior and use data to correct what the equations omit. This is often more useful than a contest between “AI” and “simulation”: a twin is a system for combining evidence.

05 FROM PREDICTION TO ACTION

A twin becomes operationally valuable when it changes a decision. It might estimate remaining useful life, recommend a maintenance window, test a control setting, or compare energy strategies. In a jet engine, a prediction about component wear is useful if it can be connected to inspection planning and spare-parts logistics. In a factory, a prediction is useful if the line can respond before a stoppage.

The action loop must be designed with care. A model can recommend an intervention without being authorized to execute it. Automatic control may be appropriate for a bounded, reversible adjustment; it is not automatically appropriate for a safety-critical change. The twin should expose assumptions, confidence, and the time horizon of its forecast so that a human or control system can act proportionately.

06 UNCERTAINTY IS PART OF THE OUTPUT

Every layer adds uncertainty. Sensors have noise; the model has approximations; parameters are estimated from incomplete data; the future operating environment is unknown. A forecast without uncertainty is a polished guess. A serious twin reports a range, probability, or decision threshold that tells users how much reliance is justified.

Uncertainty quantification can use repeated sampling, probabilistic filters, ensembles, or surrogate models. The method matters less than the discipline: uncertainty should be measured, propagated, and validated against outcomes. If a twin repeatedly claims high confidence and misses failures, the problem is not just a bad prediction. It is a broken calibration relationship between confidence and reality.

07 WHY TWINS ARE HARD TO SCALE

One asset can be instrumented carefully. A fleet, city, or supply chain is harder. Identities must remain consistent across databases, sensor formats must be interoperable, and changes to the physical asset must be reflected in the model. Cybersecurity becomes part of model integrity because a compromised data stream can cause the twin to diverge while still appearing internally coherent.

The practical roadmap is therefore incremental. Start with a decision that has a measurable value, instrument only the variables that decision needs, validate the twin against held-out events, and make its limits visible to operators. The promise of digital twins is not a perfect virtual universe. It is a maintained chain of evidence from physical state to better choice.