Harsh Apurva Sharma is an assistant professor of mechanical engineering whose research sits at the intersection of computational science and mechanical/aerospace engineering, with a focus on the design, analysis and control of complex engineering dynamic systems. Sharma develops data-driven methods that leverage domain knowledge (about the underlying physics, for example) to construct efficient and predictive surrogate models that support data-informed analysis and decision-making for large-scale dynamical systems. His work emphasizes digital twins. In this Q&A, Sharma talks about these digital replicas and how they can help optimize everything from factory floors to our built environment to human health.
Q: What is a digital twin?
A digital twin is a digital representation of a physical system or asset. That could be an aircraft, a power plant, wind turbine or even an energy grid. You usually begin with a computer-based model of the real physical system. As the physical asset operates in the real world, you have data coming from the physical asset that you can use to update the digital representation, with the ultimate goal being that you can make data-informed decisions that can help you enhance the physical asset’s performance. That could be predictive maintenance or some sort of optimization to increase efficiency.
There’s a bi-directional interaction between a physical asset and its digital twin. As technologies improve, we’re able to collect vast amounts of data from sensors on a physical asset and use it to update the digital twin more frequently. The other side of this interaction is deciding what you do with the information coming in from that physical asset. On the digital side, we can test out what we call “what if” scenarios to look at how the system might behave under different conditions or how it may evolve in the future.
For example, let’s say you have a digital twin of an aircraft wing on flights between New York and Chicago. After 10 flights, there might not be much change. But after 100 flights, say there might be some structural degradation in the wing. After each of those flights, you are getting information from the sensors on the aircraft wing that can be reflected in a digital twin to monitor the wing’s health. That data can help inform decisions like what type of maintenance to perform.
Q: How are digital twins useful today?
The term “digital twin” was coined by NASA in 2010, but the concept goes back to the Apollo 13 mission in 1970. While preparing for the mission, NASA created digital representations of the spacecraft. During the mission, an oxygen tank exploded and damaged the spacecraft, disabling its electrical and life support systems. It wasn’t “real-time” as we think of it today, but even back in 1970, NASA had data coming back from the spacecraft to Mission Control in Houston. NASA ran a lot of “what if” scenarios because it had this digital twin. Because engineers had that data and could use it to make informed decisions, they were able to make adjustments for that unexpected scenario to get the crew back to Earth.
To see digital twins in use today, we can go back to the aircraft example. Engine manufacturers like GE and Rolls Royce create corresponding digital twins for their engines. They’ve built on-board sensors that track important metrics for the engines and constantly feed that data back to the digital twin. If you think about all of the aircraft these engines are on, that could be hundreds or thousands of flights every day providing operational data and helping to schedule maintenance.
We now see companies like BMW creating digital twins of manufacturing facilities to optimize manufacturing throughput. That allows them to virtually implement changes to test how those changes might impact a facility’s efficiency before doing it in the real world.
Even looking at the climate, there’s ongoing work to use digital twins to make predictions for “what if” scenarios. That could apply to hurricane season or arctic blasts during the winter. There is work on creating digital twins that can be used for forest fires to help predict where fires might occur, how they might spread when they start, and what actions to take to prevent or minimize damage.
Q: People might be the ultimate complex system. Could we have digital twins, too?
Yes, and personalized medicine is now being recognized as a key area where digital twins can have a great impact. In this example, there might be a case where a patient has a condition like heart disease, cancer or diabetes—with a digital twin of the affected organ or area of their body.
Let’s say a patient goes into a clinic for some tests. That data will feed into their digital representation, which can, in turn, help the doctor make decisions based on the most updated representation of the patient’s health. And the idea is that every visit to the doctor can update and improve the digital twin to accurately reflect patient health. That facilitates personalized medicine, where the predictions from that digital twin would be used to create a specific treatment for that specific patient. There’s a lot of research happening in this area, especially regarding heart disease and cancer.
As the technology continues to evolve, it could one day get to the point where every person, regardless of their health, has a digital twin. Those could be used to help you decide what to eat, or what type of exercises you should or should not do.
Q: What’s the future of digital twins?
There’s a lot of potential in areas like smart cities, personalized medicine, or for spacecraft and in planning space missions. There’s ongoing work to create digital twins of Earth, which could help understand and predict climate. As we improve our ability to create accurate digital representations, digital twins will become closer to their real-world counterparts, which will only enhance their performance.
Having said that, we are not all the way there yet. There are a few key challenges we still have to solve, and that’s what excites me about working in this field.
One challenge is that in order to have a good digital representation of a physical asset, we have data coming in that we use to infer that asset’s health or physical state. But what happens in application is that typically we want to know things we can’t measure. So we have to rely on what we call indirect observations.
For example, if we look at an aircraft wing, we can’t just break the wing open and see what’s going on in terms of its structural health. That is where we use indirect observations from data we can measure to make decisions. Sometimes, that data can be “noisy” and it can be sparse, both in terms of how frequently it comes in as well as spatially—for example, depending on where sensors are placed on an aircraft wing. So even though we think we have a lot of data, sometimes that data might not be what we need to infer the correct state or what we want to know, and that’s a key challenge.
Beyond that, I think the real challenge for digital twins is making them truly predictive. Predictive capability depends on the application, but at its core it means being able to accurately forecast scenarios that we have not seen before. For example, consider a digital twin based on a bridge that has never had any structural issues. Now, if some structural degradation begins in the bridge, can a digital twin accurately predict what will happen to a bridge or how long it will remain structurally sound once there has been some damage? Being able to answer questions like these, especially outside the range of available data, is what I mean by truly predictive digital twins. It is, I would say, the grand scientific challenge for this field.
This is where I see some of my own work coming into play, focusing on making digital twins truly predictive by bringing together data-driven methods with physics-based modeling and domain knowledge. Data is good, but we have seen that when we get scenarios beyond training data, relying on data alone isn’t enough. By combining data-driven methods with physics-based insight, I can see us overcoming this challenge and using digital twins to make predictions that are accurate and reliable.
Featured photo caption: Harsh Sharma discusses a digital twin research project with a student in his office. Sharma, an assistant professor of mechanical engineering, studies complex engineering systems, with an emphasis on using digital twins. Photo: Joel Hallberg.