Digital twins

Overview

A digital twin is a virtual model of a real thing, such as a machine, a production line, a building, a power network or a whole city district, that is kept up to date with data from its physical counterpart. Sensors and business systems feed the twin with the current state of the real system. Engineers and software then use the twin to see what is happening, to understand why, to predict what will happen next and to test changes before making them for real.

The essential difference from an ordinary simulation or 3D model is the live connection. A design model describes how a product should be. A digital twin describes how one particular instance actually is today, with its own wear, settings and history, and it changes as that instance changes. Some twins only mirror data; more mature ones feed recommendations or control actions back to the physical system.

An everyday picture is a flight simulator, combined with the instruments of a real aircraft. The simulator lets pilots try dangerous manoeuvres safely. Now imagine it continuously loaded with the exact state of one aircraft in flight. A digital twin gives engineers that kind of safe, up-to-date copy of their factory, network or building.

What it is

Twins are often described by level of maturity:

Twins also differ in scope: a component twin (a bearing), an asset twin (a pump), a system twin (a production line) or a process twin (a supply chain). Larger twins are often composed of smaller ones.

Where it came from

The underlying idea of keeping a matched model of a physical system is old. Engineering teams in space programmes used ground-based replicas and simulators to understand and support spacecraft in flight.

The modern concept is usually credited to Michael Grieves, who presented it in 2002 at a Society of Manufacturing Engineers conference in Michigan, in the context of product lifecycle management. He described a real space, a virtual space and the flow of information between them, and used names such as the "Information Mirroring Model". His NASA colleague John Vickers later gave it the name "digital twin", which Grieves adopted; NASA technology roadmaps around 2010 helped spread the term in aerospace.

In the 2010s, cheaper sensors, the industrial internet of things, cloud computing and better simulation software brought digital twins into manufacturing, energy, buildings and infrastructure. The concept became a pillar of Industry 4.0. Standards work followed, including the ISO 23247 series, published from 2021, which describes a digital twin framework for manufacturing. Today the term is applied widely, sometimes loosely, so it is worth asking any "twin" what data it receives, how often, and what it does with them.

How it works

Step 1: define the purpose

Decide which decisions the twin should support: reducing downtime, testing a new layout, planning energy use or training operators. The purpose determines the scope, data and model fidelity. A twin built "for everything" usually helps with nothing.

Step 2: model the structure

Describe the physical system as a set of assets and relationships: which pump feeds which tank, which machines form which line. This asset model, sometimes held in a graph, gives every data point its context.

Step 3: connect the data

Stream sensor and control data through industrial protocols such as OPC UA or MQTT, often via an edge gateway, into a time-series store or event platform such as Apache Kafka. Add data from maintenance, production and enterprise systems. Synchronise clocks and identifiers so that data line up.

Step 4: build the models

Combine approaches as needed:

• Physics-based models use engineering equations and are reliable where the physics is well understood.
• Data-driven models learn behaviour from history, for example anomaly detection or forecasting.
• Hybrid models use physics for structure and machine learning to correct it, such as physics-informed neural networks.
• Discrete-event simulation represents flows of parts, people or vehicles through a process.

Step 5: calibrate continuously

Compare the twin's outputs with real measurements and adjust parameters so it stays faithful as the physical system ages or changes. A twin that drifts from reality is worse than none, because people trust it.

Step 6: use and close the loop

Show the current state, run what-if scenarios, generate forecasts for predictive maintenance, and send recommendations to planners or control systems. Start with human approval before allowing any automatic action.

Key concepts and terms

A worked example

This scenario is illustrative. A bottling plant has one filling line with six machines and frequent small stoppages. Managers want to know whether buying a second labelling machine or enlarging a buffer conveyor would raise output more.

The team builds a line-level twin. Machine states, counts and stoppage reasons stream every second from the line's controllers. A discrete-event simulation reproduces the line, calibrated so that over two weeks of history the twin's simulated output is within 2 per cent of actual output.

The twin shows that the labeller stops briefly about 40 times per shift and that the short buffer before it fills within 90 seconds, so stoppages ripple back to the filler. Running scenarios on the twin, doubling the buffer raises simulated output by 6 per cent, while a second labeller raises it by 8 per cent at roughly four times the cost. The plant enlarges the buffer first. After installation, measured output rises by 5 per cent, and the twin is recalibrated with the new layout for the next round of decisions.

Strengths and limitations

Strengths:

• Lets teams test changes safely before touching the real system.
• Brings scattered data into one consistent, contextual view.
• Supports prediction, optimisation and training from the same model.
• Can capture knowledge about a system that otherwise lives in a few experts' heads.
• Improves over time as more data and feedback accumulate.

Limitations:

• Building and maintaining a faithful twin is costly.
• Quality depends entirely on data quality and synchronisation.
• High-fidelity physics models can be slow; fast models may be too rough.
• The term is often used for simple dashboards, which sets false expectations.
• Feedback to control systems raises safety and security risks.

How it compares with alternatives

Practical guidance and best practices

1. Start from a specific decision and a measurable outcome.
2. Begin with one asset or line, prove value, then extend.
3. Invest early in the asset model and consistent identifiers.
4. Validate the twin against real measurements before using it for decisions.
5. Recalibrate on a schedule and after any physical change.
6. Use surrogate models where full simulation is too slow for interactive use.
7. Keep a human approval step before any automatic action on equipment.
8. Apply strong security to every data and control connection, following zero trust principles.

Common mistakes

• Building an impressive 3D model with no live data or decision attached.
• Trusting the twin without checking how far it has drifted from reality.
• Ignoring data latency, so the "live" twin is hours out of date.
• Modelling the whole factory at once instead of starting small.
• Allowing the twin to write to control systems without safeguards.

Frequently asked questions

Is a digital twin just a 3D model?

No. A 3D model may be part of the visualisation, but a twin is defined by its live data connection and its models of behaviour. Many valuable twins have only simple diagrams or no 3D view at all.

What is the difference between a digital twin and a simulation?

A simulation is a model that can be run with chosen inputs. A digital twin uses simulation and other models, but it is continuously fed with data from one specific real system and kept calibrated to it, so it reflects that system's current state.

Where does AI fit in?

Machine learning detects anomalies, forecasts behaviour, estimates hidden quantities from sensor data, and builds fast surrogate models of slow simulations. Optimisation and reinforcement learning can search for good settings using the twin as a safe testing ground.

How accurate does a twin need to be?

Accurate enough for the decision it supports. A twin used to compare two layouts needs to rank them correctly; a twin used to control a process needs much tighter accuracy and safety margins. Define the required accuracy up front and measure it.

Are there standards for digital twins?

Yes, increasingly. In manufacturing, the ISO 23247 series describes a reference framework for digital twins. Industrial communication standards such as OPC UA are often used to connect twins to equipment.

Glossary

• Digital twin: a live virtual model of a specific physical entity.
• Digital shadow: a one-way data copy without feedback.
• Fidelity: how realistic the model is.
• Calibration: tuning a model to match real measurements.
• Discrete-event simulation: modelling a process as a sequence of events.
• Surrogate model: a fast learned approximation of a detailed model.
• IIoT: the industrial internet of things, connected sensors and machines.
• What-if scenario: a hypothetical change tested on the twin.

Summary

A digital twin is a virtual model of a specific physical system, kept current with live data and used to monitor, explain, predict and test changes. The concept was presented by Michael Grieves in 2002 and named by NASA's John Vickers, and it spread through manufacturing and infrastructure with industrial connectivity and better simulation. Its value comes from a clear purpose, reliable data, models validated against reality and careful feedback to the real world. Start small, calibrate often and keep humans in control of actions.

Related

It belongs to Industries and applications in my stack, under Machine learning. Others in the same group: