Machine learning

Machine learning is an important new technique for developing computer models. In the following, we show its advantages and its limitations. But simulations based directly on physics also have their limitations. It is therefore necessary for each situation to consider which approach, or combination of approaches, is best.

Machine learning versus physics-based simulation

In the world of simulation and forecasting, most calculations are based on a scientific description of the underlying processes. This is often referred to simply as physics-based simulation, although natural sciences other than physics can also form the basis for a model.

The strength of such simulations is that they have general validity and their correctness (with respect to the underlying laws of nature) can be checked. But this is offset by significant drawbacks: they generally require a great deal of computing power and computing time. Moreover, developing and calibrating them is very labor-intensive. And for some processes we simply do not have a good scientific description.

Machine learning models are created in a completely different way. In machine learning, the model is not explicitly programmed. The algorithm learns itself to behave like the process that is being modeled. This learning, or training, can be based on large quantities of sensor data, but also by learning from a physics-based simulation. Training a machine learning model can require a lot of computing power and has its pitfalls. But once a model is properly trained, it provides answers very quickly, often with relatively little computing power as well.

Of course, there are also disadvantages. Chief among them is that the model is not inherently reliable. Outside the range for which it is trained, it can give unexpected answers. And even within the training range, it can sometimes behave non-physically.

Still, in many cases machine learning models are very useful in the technical-scientific sector.  Especially if some physical understanding is trained into the model.

In what situations is machine learning a good solution and when is it not?

Where machine learning is a good solution

Machine learning is certainly a good approach for forecasting when it comes to processes for which no good physical description exists.

A well-known example is modelling machine aging. This is needed for predictive maintenance, where one wants to be able to schedule machine maintenance at the time it is needed and convenient. So not too early and not too late.

Another example where machine learning is very useful is where it comes to human behavior. For example, the use of drinking water in a city as a function of time or the choices people make about their mode of transportation under certain circumstances.

Where machine learning is not a good solution

Although developments in machine learning are progressing very rapidly, machine learning is currently less suitable in situations where the physical correctness of the results is essential. That’s because it is often difficult to guarantee that a machine learning model obeys the laws of physics.

Machine learning is also often less appropriate if it is important that the operation of the model is properly understood. Depending on the specific technique used, it can be impossible to determine exactly how a machine learning model arrives at its answers. That makes it difficult to account for those answers.

Finally, machine learning may be less useful if there is not enough data, or if the data is of poor quality. This objection can sometimes be resolved by using physics-based models to generate synthetic data or by using physics in training. But a lack of data can make the use of machine learning less attractive.

The best of both worlds: combining machine learning with physics-based models

Combining machine learning with physics-based simulation provides the ideal solution in certain cases.

For example, a physics-based simulation can be used to train a machine learning model. This then learns to behave like the physical model, but is much faster.

Such a fast model can, for example, be used within design calculations to calculate many variants of a design and thus find an optimal design. Another application is found in Digital Twins, where a machine learning model is used to run in synchrony with the physical system, which is usually not possible with a physics-based simulation.

A machine learning model can also be used in real-time control algorithms, for which a physics-based simulation is often too slow. Such a control algorithm can also be learned directly with machine learning. During the training phase, a control algorithm based on machine learning can use physics-based simulation to learn how the system responds to control signals and how it should then itself respond.

Another example of the combination of machine learning with physics-based simulation is to calculate sub-processes within the physics-based simulation that require too much computing time or whose physics is not well known with a sub-model that is based on machine learning.

There are also many other options. Machine learning can, for example, be used to calibrate simulation models: it can learn to find the right parameters for the model given a set of measurement data. Or a machine learning algorithm can learn from a physics-based simulation to translate sensor values ​​into other quantities that cannot be directly observed, which is known as soft sensing.

Our approach

Although all the above examples are already known in literature, it is certainly not the case that all these techniques are now commonplace. Developments in the field of machine learning are moving very quickly, and new, better techniques are constantly becoming available that can be tried.

VORtech follows these developments closely. As one of the few companies in the Netherlands that is fully specialized in computing and simulation services, we are constantly working to find better methods for our customers. And we apply these new techniques for customers if this can be a solution to their challenge. In some cases, we are confident enough to guarantee a good result. But in other cases, we work together with the customer to try out a new technique or approach.

If you are interested, we would be happy to discuss the options with you. Fully developing a machine learning solution, especially when combined with physics-based simulation, can be complex. But it can bring enormous benefits. Often, a short project is enough to determine whether machine learning is a promising solution.

For more information about machine learning developments and how VORtech deals with them, see this blog series.