Combining a digital twin with artificial intelligence (AI) can remove much of the guesswork and expense that comes with manufacturing a product. But what exactly is a digital twin, or virtual replica, and how does it streamline your production process in the real world?

Dr. Norbert Gaus of Siemens Corporate Technology defines the concept of a digital twin as a "digital representation of a physical product in all its aspects." Digital twin technology can speed time to market, reduce costs, and allow a company to create a much broader portfolio of products.

Dr. Gaus explains how AI-based simulations can take the place of creating multiple physical prototypes to achieve new designs. He describes how Siemens combines a digitized version of the physical product with artificial intelligence throughout the product life cycle. That product lifecycle includes design, components, manufacturing, operations, and service and maintenance.

In this video, Dr. Gaus also discusses the challenges that Siemens has faced in the last ten years bringing digital twin automation to life. Finally, he offers sage advice to business leaders considering digital twin technology for their own companies.

Transcript

Michael Krigsman: We're at the Siemens Spotlight on Innovation speaking with Dr. Norbert Gaus of Siemens. He's one of the world's experts on the digital twin concept.

What is your role at Siemens?

Dr. Norbert Gaus: I'm with Siemens Corporate Technology. My responsibilities include all technologies around digitization and automation.

That encompasses technology areas like the digital twin like simulation, like cybersecurity, artificial intelligence, Internet of things (IoT), and many, many other areas, software systems, so all technologies that bring together the world of IoT and digitalization.

What is a digital twin and how does it improve product performance?

A simple answer is, the digital twin is a digital representation of a physical product in all its aspects. Now, the last part of the sentence really is the critical one because it means it covers the whole lifecycle of a product. It describes the product in designing it, designing it from a mechanical point of view, describing it from embedded software, from a flow mechanics point of view, and in many, many other aspects, electrical aspects.

A digital twin also describes how to manufacture the product. It describes how the product behaves during operation and in service and in maintenance. It covers the whole product lifecycle.

If we have a digital twin that really represents the physical product in the digital world, then we can, where we want, perform tasks in the digital world, allowing us to do things faster. For example, simulation instead of timely and costly prototyping. We can do things much more often until we find the optimal solution. We will spend fewer resources, both monetary resources but also manpower.

Can you use a digital twin during all phases of the product development lifecycle?

Yes, in all phases. It's a kind of a simulation of many different models instead of just one. For the purpose, you then pick the ones you really need for this application.

Tell us about the machine learning models you create for the digital twin?

Typically, the lifecycle starts in designing a product or in designing a plant. This is where you pursue many different avenues. You design products. You build CAD models. You simulate those CAD models. You pick alternatives. You try to optimize. You try to get an idea about the embedded software that has to run on the model, on the physical product. You model this without really implementing and coding it in detail.

You simulate in all the different aspects, so it's models in the CAD area. It's simulation models for software. It's computational flow dynamics. It's simulating electrical circuits. It's a very broad set of models in the design space.

Then, in the next phase of, let me call it, a life of a product when it goes into manufacturing, for example, or installation if it's a larger plant, then you simulate different things. You simulate or you design how to manufacture the product, how well it's built for manufacturability. That allows you already there in the digital space to build a feedback loop designed for manufacturability by simulating how it's manufactured.

Once you have designed the manufacturing, you can engineer it. You can, for example, automatically feedforward this information and automatically generate the PLC code, for example, so it's a feedforward loop and a feedback loop you already have in these two phases.

During the manufacturing process then, you try to, of course, optimize the manufacturing. All of a sudden, you need a pretty accurate model of the manufacturing site itself so that you can optimize this part of the lifecycle.

The next phase of a lifecycle then is the longest one, which is operations. In operations, then you also want to optimally operate the system. You want to, for example, optimize the efficiency of a gas turbine or minimize the emissions of a gas turbine. You want to minimize downtime of a large motor, and so on and so forth.

During the operation, you need a model that is less complex. You cannot afford hours of CPU times if you want to make adaptions and optimizations in real time. You need to reduce the order of the model, significantly reducing the complexity. At the same time, for the critical components you want to optimize, the model has to be still pretty accurate. You reduce the order of the complexity of the model so that you can do inline and real-time kind of real-time simulation to optimize operations.

Then there's another important phase. This is service and maintenance. These are typically data-driven models. Artificial intelligence, when it comes to predictive maintenance, in a way is nothing else than just another kind of a model, driven and defined by data, but it's a model. This is the last part of the model in our lifecycle.

Important is that that it's really a combination of a feedforward loop, but you can automate some of the steps increasingly more so. It's also a feedback loop from service into manufacturing, from service into designing the next product generation, from the factory back into design, and so on and so forth.

How do you ensure the digital twin accurately represents its physical counterpart?

As a company, we started this journey 15 years ago; more than 10 years ago, I have to say, and invested a lot to ensure exactly the consistency between the very different representations to find the hooks and handles between the mechanical design and the flow design where embedded software comes in and what it does, and so on and so forth, really finding the hooks and handles and implementing them. Another very important component is the products we sell. They have a lifetime, depending on the vertical market, of between 10 and 30 years.

Now from the consumer side, we are all used to replace the devices we have in this field every two or three years. We have to ensure that what we sell from a digital twin perspective is still an accurate representation also after 10, 15, and 20 years, so we really also need to manage the lifecycle. We have to be able to update software and represent this in the digital twin. We have to ensure that the as-built digital twin is where the as-built is changing. You can imagine if you build a larger plant, what is on the drawings in the design that there will be some changes when we build this large plant; that the changes are being fed back into the plants.

These processes have been established and that has been a major part of our investment. It's not only bringing the capabilities of defining those models into our company. The big step really then has to be to integrate this as a real toolbox and suite.

What is the role of artificial intelligence in building a digital twin?

The artificial intelligence is a technology that, together with the digital twin, I think stands out when it comes to digitalization. A digital twin stands out because digitalization means having a digital representation of the physical product. Then it's artificial intelligence because digitization really is about data from products in the fields, in the factory, from wherever, and then generating value from these data. That's why these two technologies stand out.

They do come together. Actually, they have always come together because remember what I said at the beginning of our discussion is, when it comes to maintenance, especially predictive maintenance, preventive maintenance, you do build models based on data. Now, these are also models and part of the digital twin so, in a way, artificial intelligence actually has always been part of the digital twin. While I know that for most people this was a different world, and it was really a separated world, now it does come together.

It also comes together in other areas of the digital twin. We use artificial intelligence for model order reduction, as one example. While model order reduction has always been a known technology, in some cases we need higher complexity and nonlinear models where it's very difficult to, more or less, linearize and use traditional technologies, so we use a neural network to generate a model that pretty accurately represents a motor, for example.

You use AI to create the models used in the digital twin?

Exactly. We feed the network with the very complex design model, for example, out of an X with, let me say, one billion degrees of freedom. We push it through a model to cut it down to maybe a hundred degrees of freedom, and then make sure that we, from those 100 states, we still would be able to represent, to recalculate the critical states. That we do with a neural network with deep learning, helping us to, where we cannot, for example, measure critical states of large motors, simulate this in real time and still know how the motor behaves.

We also use artificial intelligence in generative design. Now, some ways of generative design is not new but, with artificial intelligence, what we try to do is--and not only mechanical design; also designing electrical circuits--first, we widen the design space in which we look. But if you widen the design space, you get many more design options.

Now, typically, all these options have to be simulated and these simulations are very time-consuming. We are not using artificial intelligence to simulate, but we use artificial intelligence to preselect options that have the highest chance, for example, in a finite element simulation to converge. We first widen the design space with artificial intelligence in exploring it and then we also use AI to reduce the number of simulations we need, which I think has huge potential. We do this in a few areas already and I'm convinced that it really can help us to be much broader.

What are the best customer applications for digital twins?

Actually, I don't think there is a"best" application. I'm convinced that digitalization is about the whole lifecycle depending on the vertical industries. Of course, in some of the vertical industries, not in all phases, you really need a digital twin. Some products will all always just work and operate and you will not need real-time inline simulation. In essence, every market, every product owner has to understand that he or she needs some kind of digital representation throughout the whole lifecycle to speed up time to market, to reduce cost, to being able to have a much broader portfolio and offering to the market, to use the data that is coming from the field, one way or another, from sensors, from service reports, from many data sources, which today, in in a lot of cases, is not really being utilized, not really understanding there is a lot of value in these data and this value needs to be feedback into the future product lines.

Again, I don't think there is a priority. There are priorities, at the same time--now, this may sound like a contradiction--depending on which market you are, the majority of your product lines, and so on and so forth. But in terms of a goal, you have to cover the whole product lifecycle.

What advice do you have for business leaders who want to start using a digital twin?

The most critical decision is how to get started on this journey. There is not a blueprint in which product, in which lifecycle phase a company should start, but it needs a careful analysis of, what are the dynamics of our markets; where do we as a company want to differentiate against our competitors; where do we differentiate today; where do we want to differentiate in the in the future for what reason; and how can digitalization, which is, basically, in essence, a digital twin, really helped us? As you define where you want to differentiate, these should be the areas where you get started.

We all have competitors in this world, and we all need to find the core markets where we want to be a leader and what's important to them. This is where we need to differentiate or want to differentiate as our product strategy. There are other parts of a corporate strategy but, from a portfolio strategy point of view, and where we want to differentiate on the portfolio, I think this will only be possible by being also leading in the digital part of the portfolio.

Michael Krigsman: We're at the Siemens Spotlight on Innovation speaking with Dr. Norbert Gaus of Siemens. He's one of the world's experts on the digital twin concept.

What is your role at Siemens?

Dr. Norbert Gaus: I'm with Siemens Corporate Technology. My responsibilities include all technologies around digitization and automation.

That encompasses technology areas like the digital twin like simulation, like cybersecurity, artificial intelligence, Internet of things (IoT), and many, many other areas, software systems, so all technologies that bring together the world of IoT and digitalization.

What is a digital twin and how does it improve product performance?

A simple answer is, the digital twin is a digital representation of a physical product in all its aspects. Now, the last part of the sentence really is the critical one because it means it covers the whole lifecycle of a product. It describes the product in designing it, designing it from a mechanical point of view, describing it from embedded software, from a flow mechanics point of view, and in many, many other aspects, electrical aspects.

A digital twin also describes how to manufacture the product. It describes how the product behaves during operation and in service and in maintenance. It covers the whole product lifecycle.

If we have a digital twin that really represents the physical product in the digital world, then we can, where we want, perform tasks in the digital world, allowing us to do things faster. For example, simulation instead of timely and costly prototyping. We can do things much more often until we find the optimal solution. We will spend fewer resources, both monetary resources but also manpower.

Can you use a digital twin during all phases of the product development lifecycle?

Yes, in all phases. It's a kind of a simulation of many different models instead of just one. For the purpose, you then pick the ones you really need for this application.

Tell us about the machine learning models you create for the digital twin?

Typically, the lifecycle starts in designing a product or in designing a plant. This is where you pursue many different avenues. You design products. You build CAD models. You simulate those CAD models. You pick alternatives. You try to optimize. You try to get an idea about the embedded software that has to run on the model, on the physical product. You model this without really implementing and coding it in detail.

You simulate in all the different aspects, so it's models in the CAD area. It's simulation models for software. It's computational flow dynamics. It's simulating electrical circuits. It's a very broad set of models in the design space.

Then, in the next phase of, let me call it, a life of a product when it goes into manufacturing, for example, or installation if it's a larger plant, then you simulate different things. You simulate or you design how to manufacture the product, how well it's built for manufacturability. That allows you already there in the digital space to build a feedback loop designed for manufacturability by simulating how it's manufactured.

Once you have designed the manufacturing, you can engineer it. You can, for example, automatically feedforward this information and automatically generate the PLC code, for example, so it's a feedforward loop and a feedback loop you already have in these two phases.

During the manufacturing process then, you try to, of course, optimize the manufacturing. All of a sudden, you need a pretty accurate model of the manufacturing site itself so that you can optimize this part of the lifecycle.

The next phase of a lifecycle then is the longest one, which is operations. In operations, then you also want to optimally operate the system. You want to, for example, optimize the efficiency of a gas turbine or minimize the emissions of a gas turbine. You want to minimize downtime of a large motor, and so on and so forth.

During the operation, you need a model that is less complex. You cannot afford hours of CPU times if you want to make adaptions and optimizations in real time. You need to reduce the order of the model, significantly reducing the complexity. At the same time, for the critical components you want to optimize, the model has to be still pretty accurate. You reduce the order of the complexity of the model so that you can do inline and real-time kind of real-time simulation to optimize operations.

Then there's another important phase. This is service and maintenance. These are typically data-driven models. Artificial intelligence, when it comes to predictive maintenance, in a way is nothing else than just another kind of a model, driven and defined by data, but it's a model. This is the last part of the model in our lifecycle.

Important is that that it's really a combination of a feedforward loop, but you can automate some of the steps increasingly more so. It's also a feedback loop from service into manufacturing, from service into designing the next product generation, from the factory back into design, and so on and so forth.

How do you ensure the digital twin accurately represents its physical counterpart?

As a company, we started this journey 15 years ago; more than 10 years ago, I have to say, and invested a lot to ensure exactly the consistency between the very different representations to find the hooks and handles between the mechanical design and the flow design where embedded software comes in and what it does, and so on and so forth, really finding the hooks and handles and implementing them. Another very important component is the products we sell. They have a lifetime, depending on the vertical market, of between 10 and 30 years.

Now from the consumer side, we are all used to replace the devices we have in this field every two or three years. We have to ensure that what we sell from a digital twin perspective is still an accurate representation also after 10, 15, and 20 years, so we really also need to manage the lifecycle. We have to be able to update software and represent this in the digital twin. We have to ensure that the as-built digital twin is where the as-built is changing. You can imagine if you build a larger plant, what is on the drawings in the design that there will be some changes when we build this large plant; that the changes are being fed back into the plants.

These processes have been established and that has been a major part of our investment. It's not only bringing the capabilities of defining those models into our company. The big step really then has to be to integrate this as a real toolbox and suite.

What is the role of artificial intelligence in building a digital twin?

The artificial intelligence is a technology that, together with the digital twin, I think stands out when it comes to digitalization. A digital twin stands out because digitalization means having a digital representation of the physical product. Then it's artificial intelligence because digitization really is about data from products in the fields, in the factory, from wherever, and then generating value from these data. That's why these two technologies stand out.

They do come together. Actually, they have always come together because remember what I said at the beginning of our discussion is, when it comes to maintenance, especially predictive maintenance, preventive maintenance, you do build models based on data. Now, these are also models and part of the digital twin so, in a way, artificial intelligence actually has always been part of the digital twin. While I know that for most people this was a different world, and it was really a separated world, now it does come together.

It also comes together in other areas of the digital twin. We use artificial intelligence for model order reduction, as one example. While model order reduction has always been a known technology, in some cases we need higher complexity and nonlinear models where it's very difficult to, more or less, linearize and use traditional technologies, so we use a neural network to generate a model that pretty accurately represents a motor, for example.

You use AI to create the models used in the digital twin?

Exactly. We feed the network with the very complex design model, for example, out of an X with, let me say, one billion degrees of freedom. We push it through a model to cut it down to maybe a hundred degrees of freedom, and then make sure that we, from those 100 states, we still would be able to represent, to recalculate the critical states. That we do with a neural network with deep learning, helping us to, where we cannot, for example, measure critical states of large motors, simulate this in real time and still know how the motor behaves.

We also use artificial intelligence in generative design. Now, some ways of generative design is not new but, with artificial intelligence, what we try to do is--and not only mechanical design; also designing electrical circuits--first, we widen the design space in which we look. But if you widen the design space, you get many more design options.

Now, typically, all these options have to be simulated and these simulations are very time-consuming. We are not using artificial intelligence to simulate, but we use artificial intelligence to preselect options that have the highest chance, for example, in a finite element simulation to converge. We first widen the design space with artificial intelligence in exploring it and then we also use AI to reduce the number of simulations we need, which I think has huge potential. We do this in a few areas already and I'm convinced that it really can help us to be much broader.

What are the best customer applications for digital twins?

Actually, I don't think there is a"best" application. I'm convinced that digitalization is about the whole lifecycle depending on the vertical industries. Of course, in some of the vertical industries, not in all phases, you really need a digital twin. Some products will all always just work and operate and you will not need real-time inline simulation. In essence, every market, every product owner has to understand that he or she needs some kind of digital representation throughout the whole lifecycle to speed up time to market, to reduce cost, to being able to have a much broader portfolio and offering to the market, to use the data that is coming from the field, one way or another, from sensors, from service reports, from many data sources, which today, in in a lot of cases, is not really being utilized, not really understanding there is a lot of value in these data and this value needs to be feedback into the future product lines.

Again, I don't think there is a priority. There are priorities, at the same time--now, this may sound like a contradiction--depending on which market you are, the majority of your product lines, and so on and so forth. But in terms of a goal, you have to cover the whole product lifecycle.

What advice do you have for business leaders who want to start using a digital twin?

The most critical decision is how to get started on this journey. There is not a blueprint in which product, in which lifecycle phase a company should start, but it needs a careful analysis of, what are the dynamics of our markets; where do we as a company want to differentiate against our competitors; where do we differentiate today; where do we want to differentiate in the in the future for what reason; and how can digitalization, which is, basically, in essence, a digital twin, really helped us? As you define where you want to differentiate, these should be the areas where you get started.

We all have competitors in this world, and we all need to find the core markets where we want to be a leader and what's important to them. This is where we need to differentiate or want to differentiate as our product strategy. There are other parts of a corporate strategy but, from a portfolio strategy point of view, and where we want to differentiate on the portfolio, I think this will only be possible by being also leading in the digital part of the portfolio.