What are digital twins, and where are they on the “hype cycle”?: part 2

Digital twins are getting lots of deserved attention, but are they the cure-all for whatever simulation issues you have?

Although digital twins are undoubtedly valuable tools along with basic simulation, modeling, HITL, and others, the associated all-good/no drawbacks promotion concerns me. When a tool, product, or other development gets a lot of “pumping up” to the extent that the message is “whatever your problem, ‘x’ is the solution” while its limitations are glossed over or ignored, it seems too good to be true.

Call me cynical, jaded, or just an analog-centric realist, but my experience (with occasional exceptions) is that “close-to-perfect” modeling doesn’t exist in the real world. You have to be very aware of acknowledging limitations, especially the most worrisome of all: the unknown unknowns.

Instead, it’s that all-purpose solution aspect of digital twins, which I am seeing so much. It’s even more worrisome when consulting firms add it to their all-purpose “buzzword bandwagon” list.

Among the many recent instances of published papers in academic journals I saw were two about better battery-charging algorithms. When I read the papers, the implication was that these strategies were thoroughly tested via the lab bench and even real pilot runs, so they were ready for mass adoption and to go into production. Imagine my surprise and disappointment when I found out these “new, better” perspectives were verified solely by digital-twin assessments (References 4 and 5).

I feel the same way about artificial intelligence (AI) with respect to hype and utility. AI certainly has a role to play in dealing with various problems and issues, but it is not a cure-all. Yet I see it posited in so many places as the key to a solution, even if the connection is tenuous or highly speculative.

For example, the September 2022 issue of IEEE Spectrum had multiple features with the primary message that AI may change everything in a given application. One article was about AI analysis reducing the stealthy attributes of submarines and making them easier to locate and track (with a lot of maybes, but it certainly sounds plausible); another was how AI might help increase the success rate of in-vitro fertilization (again, maybe); and another was how AI might reduce or even eliminate world hunger (sorry, I’m not buying that one).

I understand that advanced simulations such as digital twins are generally a good thing and often necessary; in fact, sometimes, there is little choice about using them. The engineers and scientists at the Jet Propulsion Laboratory and NASA who developed the very successful James Webb Space Telescope (JWST) (Figure 1) and Mars Perseverance Rover (Figure 2) did extensive tests on real hardware and software, but they also had to rely on digital twins for the many aspects which simply could not be replicated or tested on Earth.

Fig 1: This detailed high-level system block diagram of the James Webb Space Telescope can only begin to capture the complexity of the project and the many interactions among its constituent functional blocks; the project clearly required digital twin and HITL tools. (Image: Slideplayer)

Fig 2: Although the Mars Perseverance Rover was simpler than the JWST, it also had many unique and hard-to-verify design challenges which required advanced modeling and simulation tools. (Image: NASA)

My concern is that the results of sophisticated tools such as digital twins are assumed to be correct by definition. But in the real world, we know that creating an accurate model is hard, and the last few percent of the model you can’t or don’t capture causes the problems. That’s why I would feel more confident if advocates of digital twins were more up-front about its inevitable limitations and weaknesses rather than not saying anything about the subject.

Is HITL better than the digital twin? As is usually the answer the case for engineering questions, the answer is simple: “it depends.”

The factors on which it depends include time to create their respective models, confidence in that model, and complexity of simulating I/O. If you go online, there are proponents of DT who say HITL is “so yesterday” and no longer needed (Reference 6), as well as proponents of HITL who claim that DTs are overhyped (Reference 7), oversold, and overly dependent on the fidelity of the model versus reality. Others contend that the best solution is a hybrid of both, applied with care (Reference 8).

There is no doubt that it’s absolutely necessary to use various models, whether digital twins or HITL, Spice, RF packages, or simulation and analysis tools such as COMSOL Multiphysics, The Mathworks MATLAB and Simulink, and ANSYS HFSS. But you also have to be realistic about how “perfect” these models are, always keeping in that the model may show false precision to three, four, or even more significant figures in an environment where accuracy to a few percent is actually pretty good—and may be way off if the real world has “bumps” that the model didn’t capture.

Related EE World Content