What a digital twin actually is, once you strip the marketing

#A term doing too much work

I have seen four different things called a digital twin in the same month: a three-dimensional model of a facility, a dashboard over sensor readings, a physics simulation calibrated once against real measurements, and a system that adjusts equipment based on predictions. Those are not variations on one idea.

The concept goes back to Grieves in the early 2000s, formalised in a 2014 white paper as a physical object, its virtual counterpart, and the data connecting them [1]. The systematic reviews since then have found the same definitional spread I have run into in practice [3].

#The distinction that is actually useful

Kritzinger and colleagues propose a classification based purely on how data moves, and it is the most practically useful framing I have found [2].

Digital model. No automatic data flow in either direction. A CAD model is a digital model. Changing the real thing does not change it, and changing it does not change the real thing.

Digital shadow. Automatic flow from the physical object to the digital one, and manual flow back. Most industrial monitoring is a digital shadow. Sensors update the model, and any action on the equipment is taken by a person.

Digital twin. Automatic flow in both directions. The digital object changes state in response to the physical one, and the physical one changes in response to the digital.

By that definition, most products sold as digital twins are digital shadows. That is not a criticism. Shadows are valuable, and they are considerably safer, because closing the loop on industrial equipment means software making decisions with physical consequences.

#Why the loop is hard to close

Writing back into a physical process pulls in safety certification, control system integration and liability. Existing control systems already have safety instrumented functions with their own assurance requirements, and a twin that writes setpoints becomes part of that argument [6]. Most organisations correctly conclude that the shadow gives them most of the value for a fraction of the risk.

#What I care about in the plumbing

Independent of which category you are building, the same infrastructure problems appear [4] [5].

• Identity: the same physical asset appears in the maintenance system, the historian and the engineering drawings under different identifiers.
• Time: observations arrive late, duplicated and out of order, and state has to be reconstructable for any past moment rather than only the present.
• Topology: what connects to what, which is a graph and not a hierarchy, however much the asset register wants it to be a tree.
• Provenance: for any value, which observation produced it and when, because an operator will ask.

Those four are where I have spent nearly all of my time on this problem, and none of them are specific to any industry.

#The honest summary

A digital twin is worth building when a decision depends on the state of something you cannot observe directly and continuously. If the decision is made weekly and somebody can walk out and look, a report is cheaper and more reliable.