Published
Few terms in industrial technology have been stretched further. A digital twin can mean a CAD model on a screen, a live dashboard of sensor readings, a physics simulation calibrated against a real machine, or a system that predicts a failure before it happens. These are not variations on one idea; they differ in what they cost and in what they can be trusted to do.
A test that separates them
Ask what decision the model is allowed to make on its own. If the answer is none — it visualises, and a person interprets — then it is a representation. Useful, often worth building, but its value is in communication rather than prediction.
If the answer is that you would change a setpoint, delay a maintenance window, or approve a process change on the strength of what it says, then it needs to have been validated against the real system, and someone needs to be able to state how far it can be trusted before it stops being reliable.
Validation is the expensive part
Building a model is comparatively tractable. Establishing that its predictions match reality across the operating range, and continuing to establish that as equipment wears and materials change, is the sustained cost — and it is the part most often underestimated at the point the project is scoped.
A model that was accurate at commissioning and has not been checked since is not a twin of anything current. It is a twin of the machine as it was, which is a different and much less useful object.
Fidelity should be argued for, not maximised
There is a persistent instinct to make the model as detailed as possible. But fidelity costs computation, calibration effort and maintenance, and past a point it buys nothing for the decision at hand. A model that answers one question well is generally a better investment than one that answers many questions approximately.
The practical version of this is to start from the decision and work backwards to the minimum model that supports it, rather than starting from the asset and modelling everything about it.
Why this matters for scoping
The reason to be precise about the term is that it sets expectations that are later measured against. A project delivered as a visualisation, sold as a predictive twin, is judged as a predictive twin — and fails, despite having produced something genuinely useful.
Saying plainly which one is being built is not a lesser claim. It is the one that survives contact with the plant.