Digital twins are an interesting idea that, like "cognitive computing", is easily abused by marketing, and will probably rake in a lot of consulting fees for people like the authors of this piece (Accenture Research) and companies like IBM.
The essence of a digital twin is a simulation complex enough to be useful in making predictions in the real world. (That's the "twin" part.) Making complex simulations, as you might imagine, is difficult. It requires effort, deep domain knowledge (rare talent), good feedback mechanisms with the real situation in question, and some means of managing that complexity.
Digital twins do exist in deployment. What differentiates them from, say, any old machine-learning model you might use for predictions is that a "digital twin" is probably used for a more complex task than just classification. That is, it's probably used to direct the actions of a system. The words imply a larger solution.
So one thing you see is simulations that embed machine-learning models and predict what actions to take in a given state. Think of it like AlphaGo applied to business scenarios.
What are the pitfalls? Real-world data in these environments is non-stationary and messy, so signal may be low, or the ways you find signal might change over time.
To make the "digital twin" useful you are probably integrating with large software systems not entirely in your control, which may be hard to reason about (ERP systems like SAP).
The digital twin idea, insofar as it includes large parametric models that depend on algorithms like deep reinforcement learning, matters now, because those models are able to find structure in complexity, and make ever more accurate predictions about what to do. That is, we're able to identify optimal actions in more complex situations, with techniques more sophisticated than expert systems.
All that aside, this sort of thing is already getting deployed under the right circumstances, and you could argue that it is the future of a lot of business operations in supply chain and manufacturing.
Our models of system are only a pale reproduction of how those systems work, rendering designs that are tuned by those "evil twins" bound by our current understanding of the system which can be lacking sometimes.
Nevertheless, in engineering, notably aerospace for example, those digital twins are crucial to make better parts, as thrusters for example, do not allow for continuous measurements of temperature (everything burns there).
It rendered hardware system design far more agile.
Yet the concepts to build a novel hardware is not encoded in this simulation.
Another example of "future is past, repeated" is machine learning. Carrying encoded stereotypes in predictions.
As we try to model everything with some notable failures, such as economy, let's just be aware of the limitations of those models.