Digital twin is the least precise term in industrial technology, which makes it easy to sell and difficult to buy.
Four different things with one name
Separating them is the most useful thing a buyer can do before any conversation with a vendor.
A 3D visualisation. A model of the asset you can look at and navigate. Useful for communication and training, and not a twin in any meaningful sense because it does not represent behaviour.
A simulation. A model that represents behaviour and can be run with different inputs. Genuinely useful for design and planning, and disconnected from the real asset.
A connected model. A simulation receiving live data from the asset, so its state tracks reality. This is where the term starts to mean something.
A validated predictive model. A connected model that has been demonstrated to predict the asset's behaviour accurately enough to base decisions on. This is the version that changes anything, and it is rare.
The first two are widely sold as the fourth. The distinction is whether anyone has demonstrated that the model predicts what the machine actually does.
What the direction covers
The scope: digital copies of real equipment, sensors, predictive analytics and simulation.
Four areas.
Modelling. Building a representation that captures the behaviour that matters, at a fidelity appropriate to the decision.
Connectivity. Feeding the model real data, which depends on the Industry 4.0 layer.
Validation. Demonstrating the model matches reality, and continuing to demonstrate it.
Application. Using it to make a decision that would otherwise be made worse.
The maintenance problem nobody costs
The reason most twins degrade into expensive visualisations.
A physical asset changes continuously. Components wear, parts are replaced with slightly different ones, modifications are made, control settings are adjusted, the product mix changes and sensors drift or fail.
The model does not change unless someone changes it.
Four consequences.
Divergence is gradual and invisible. There is no alarm when the twin stops matching the asset. It simply becomes progressively less accurate until someone notices a prediction was wrong.
Modifications are the main driver. Every undocumented change to the real equipment is a divergence, which makes twin maintenance dependent on change control discipline, connecting to the CAD and PLM direction.
Revalidation is a recurring cost. Periodically confirming the model still predicts correctly, which is work nobody budgeted for after the implementation project closed.
An unmaintained twin is worse than none. Because people trust it, and a trusted model that is quietly wrong produces confident bad decisions. This is the same pattern as the medical AI article: a system that degrades silently while retaining authority.
For workforce planning this is direct. A twin needs an owner with time, permanently, and the business case that omitted that role has understated the cost by the largest line in it.
Where this sits in the domain
Industrial IoT and digital twins is the eighth of ten directions in Astra Trainer's advanced manufacturing domain, depending on smart manufacturing and Industry 4.0 for the data layer, on product design and PLM for the model definition, and on maintenance and asset management for condition information.
It also connects to robotics and autonomous systems, where simulation and offline programming use the same capability, and to the AI, data and computing domain for the modelling and infrastructure half. Partners implementing twins usually need the data layer working first, which is the most common sequencing error. You can see the ten directions here.
Where twins genuinely earn their keep
Five cases, stated as cases because the general claim is too vague to evaluate.
Virtual commissioning. Testing control logic against a simulated machine before the real one is built or during installation. This reduces commissioning time measurably and is one of the most proven applications, particularly in automation projects where commissioning is the schedule risk.
Operator training. Training on a simulated plant, including failure scenarios that would be dangerous or costly to create. Genuinely valuable in process industries and in anything where a mistake is expensive.
Scenario testing. Answering what happens if we change this, without changing it. Useful for layout changes, product mix changes and capacity questions.
Process optimisation. Finding better operating parameters in the model rather than by experimenting on production.
Understanding a problem. Reproducing a fault in a model to test hypotheses about its cause.
Notice what these have in common: each replaces something expensive, slow or unsafe to do on the real asset. That is the test. A twin that duplicates something you could simply look at is a cost.
The roles, named
Simulation engineers. Building and validating models.
Virtual commissioning engineers. A specific and valuable specialism in automation projects.
Industrial data engineers. Supplying the connection, shared with the Industry 4.0 direction.
Model owners and maintainers. The role that determines whether a twin survives, and the one most often unfilled.
Process engineers using twins for optimisation.
Controls engineers testing logic against models.
Training developers building simulator-based training.
Asset engineers using twins for condition and life assessment, connecting to maintenance.
Who can be trained into it
Controls engineers. The strongest conversion, particularly for virtual commissioning, because they already know the control logic being tested and the equipment it runs on.
Simulation engineers from design. Already build models, needing the live connection and validation layers.
Process engineers. Know the behaviour the model must reproduce, which is the hardest input to obtain.
Operators. The reality check. They know what the machine actually does, including the behaviours that are not in any specification, and a model built without that input reproduces the documentation rather than the equipment.
Data engineers. Into the connectivity layer.
Training staff. Into simulator-based training, which is a growing and practical application.
Do not connect a model to control. A digital twin used for decision support is different from one that writes back to a control system. Any system that can influence plant control is subject to the same functional safety and change control requirements as the control system itself, and a model that has not been validated to that standard must not have that capability. Training builds modelling and validation understanding. It does not authorise connection to, or modification of, any control or safety system.
What to take from this
Four different things share the name, and only a validated connected model justifies the cost. Ask what has been demonstrated, not what is displayed.
Twins diverge from reality continuously and silently, which makes maintenance a permanent role rather than a project task.
An unmaintained twin is worse than none, because it retains authority while becoming wrong.
The strong cases replace something expensive, slow or unsafe on the real asset. Virtual commissioning is the most proven.
And operators supply the behaviours that are not in any specification, which is what separates a model of the documentation from a model of the machine.
What actually counts as a digital twin?
The term covers visualisations, simulations, live-connected models and validated predictive models. Only the last two justify the cost, and the test is whether anyone has demonstrated the model predicts what the machine does.
Why do twins degrade?
Because the asset changes and the model does not. Wear, replacement parts, modifications, control adjustments and sensor drift all cause divergence, and there is no alarm when it happens.
Is an out-of-date twin harmless?
No. It is worse than having none, because people still trust it. A quietly wrong model with retained authority produces confident bad decisions.
Where do twins clearly pay back?
Virtual commissioning, operator training including dangerous scenarios, scenario testing, process optimisation and fault reproduction. Each replaces something expensive, slow or unsafe on the real asset.
Who converts into this work?
Controls engineers for virtual commissioning, design simulation engineers needing the live and validation layers, process engineers who know the behaviour, and operators who supply the reality check.