Simulation and artificial intelligence (AI) are two very different options for simulating reality using mathematical methods. Until now, there have been very few joint approaches, but successful current projects are revealing outstanding new opportunities for combining the two technologies.

Until now, artificial intelligence (AI) experts and simulation specialists at Siemens have been most likely to run into each other in the cafeteria or café only. That isn’t surprising, since the two expert groups solve their problems using totally different mathematical approaches. But in the future, they’ll probably be seeing more of one another at joint meetings. “We’re discovering more and more opportunities for simulation and artificial intelligence to complement one another,” says Dirk Hartmann, a simulation expert at Siemens. “Both approaches have their strengths and weaknesses. If we can manage skillfully combining them, we’ll become faster while remaining as reliable as ever.”

Background: An example that shows how differently simulation and AI operate.

The unknown machine: 

Let’s assume we have an unknown machine with lots of buttons and a few levers. We’re looking for a program that simulates this machine’s behavior. Both simulation and AI experts can solve this problem, although in different ways.


What simulation experts do

Simulation experts let the machine run for a while and observe how it responds to different inputs. Afterwards, they’ll probably open it up, analyze its internal components and circuits, and measure the components. In this way, the experts gradually come to understand the machine’s functionality and special characteristics. Based on this knowledge – and basic mathematical expertise – they build a model (digital twin) that simulates the machine using mathematical functions.

Advantages/disadvantages of simulation

  • The simulation model can be explained and reconstructed by other experts – meaning that it can be verified as correct.
  • Because the experts have attained an in-depth understanding of the technological context, the simulation model can generally also be used for ranges not explicitly tested (data can be extrapolated).
  • However, creating this type of model is extremely costly and time-consuming.
  • Only well-trained experts – within the domain and mathematical knowledge – are able to create such a model.
  • Certain effects, e.g. outage of machines due to errors are difficult to model analytically.

What AI experts do

AI experts train a suitable neural network (NN) using sample data, meaning datasets that show which sequence of input data leads to which outputs. If the network has been trained with sufficient data (order of magnitude 105), then it will be capable of simulating the functionality.

Advantages/disadvantages of artificial intelligence:

  • With sufficient training data, an AI model can be created relatively quickly and easily.
  • However, it’s often difficult to obtain enough training data (or it first has to be generated with some extra effort).
  • AI doesn’t require a precise understanding of the machine.
  • Sometimes AI can even recognize patterns in data that are not recognizable by humans or that are yet unknown.
  • A neural network can’t be interpreted – meaning that humans generally can’t reconstruct the mathematical context presented by the neural network.
  • That’s why important decisions may require additional human supervision.
  • For this same reason, neural networks yield reliable predictions only for the data ranges for which they were trained (limited ability to extrapolate).

Approaches with different strengths and weaknesses

Which technology – simulation or AI – is used when depends on the technical task. “When we succeed in simulating reality through a model – in other words, we’re able to describe the relationships between the various relevant, physical variables using mathematical equations – then a simulation is generally the better approach, because it provides us with a coherent model of reality,” explains Hartmann. “At the other extreme are problems that can’t easily be expressed in mathematical formulas, such as the question: Does the image show a dog or a cat? There would be no point in trying to answer this question with a simulation, but a neural network that’s been trained with enough training images of dogs and cats will recognize the difference (even if we don’t understand exactly how it does it). In practice, however, we often have problems that we can’t satisfactorily solve using either of the two technologies, and that’s why we combine them. The subsequent examples are only the beginning. We’re seeing a lot more opportunities for combining AI and simulation.”

The best of both worlds

AI expert Thomas Runkler from Siemens Corporate Technology adds: “For many years we have been using data to model processes, machines and plants. Now the latest AI methods allow us  to combine machine learning with analytical models like differential equations, Navier Stokes equations or Hamiltonian dynamics. We call that ‘physics informed machine learning ‘.  Thus, we combine the best of both worlds and build precise models by using data and expert knowledge efficiently.”

Supplement standard simulation model with trained AI 

For example, a simulation model can be supplemented with a (previously trained) neural network. This neural network has learned relationships that are difficult to model in the simulation. At Siemens Corporate Technology, for example, a simulation/AI model was recently developed that can predict the remaining service life of small electric motors. In this case, the neural network had learned relationships between sensor data and signs of wear and was used to supplement a standard motor simulation.  In a project in India, a similar approach is being used  to monitor pole-mounted transformers and prevent power failures. 

A simulation model is basic knowledge for AI

On the other hand, a simulation model can also turn a neural network into a fast learner. Initially, untrained neural networks are “dumb” and have to learn even obvious relationships in many individual steps. This requires a tremendous amount of training data. But when neural networks already have a sort of basic knowledge when they start training, they need much less data to be trained. “In the industrial environment in particular, there’s often insufficient training data available. Thus to train neural networks, these data have to be explicitly generated – which can be extremely resource-intensive,” explains Hartmann. “Just to give an idea of the scale, we recently generated training data for a CFD simulation, which involved calculating 20,000 optimal CFD meshes. Just doing that required 50 years of computing time on the cluster at the Technical University of Munich.”

Two-phase problem solving:

“What’s particularly interesting are applications in which we can divide the initial problem into an AI task and a simulation task that together deliver one result. It’s as reliable as a simulation but faster,” Hartmann continues. “Neural networks handle the computationally intensive subproblems. But because we must consider that they might not deliver optimal results, we combine this with a simulation that guarantees the correctness of our result.”

CFD simulation

“In a joint project with the Technical University of Munich, for example, we successfully used this approach to calculate flow rates – such as how the air flows around a car in a wind tunnel. When you calculate flow rates, you first store the geometry using mesh points – the selected mesh has a tremendous influence on the amount of computing effort that will then be required to calculate the flow rate – in any case, it’s already extremely computationally intensive to determine an optimal mesh. This initial step of calculating the optimal mesh is now performed by artificial intelligence, making the CFD simulation much faster. If AI occasionally suggests a less than optimal CFD mesh, it’s okay because it affects only the computing time and not the result.”

Generatives Design

A combination of AI and simulation can also speed up product design. Experts call this generative design. A typical design task looks something like this: The customer specifies requirements, such as maximum weight and minimum stability. From this information, the designer has to derive a product layout that’s as inexpensive as possible. “When we develop a design using only simulation, we generally have to start by simulating several recommended designs – and we know how resource-intensive that is – before we find a design that meets the requirements,” says Hartmann. “With the support of AI, it goes faster if we train a neural network to derive recommended designs from typical customer requirements. It doesn’t matter if a bad design is occasionally recommended, because this design is then tested in a simulation to determine whether it meets all the requirements. If the design is bad, we’ll find the errors. Normally – when AI has offered a good recommendation – we just have to simulate it once, which goes much faster.”

Aenne Barnard  - April 2020

Subscribe to our Newsletter

Stay up to date at all times: everything you need to know about electrification, automation, and digitalization.