Englisch

AI, Engineering Expertise, Training Data and the Value of Uncertainty

Picture of Alexander Kunz Alexander Kunz 11.03.2025

Digitalization Insights

CADFEM meets ... Kevin Cremans, one of the founders of the AI startup PI Probaligence and the brains behind the AI solution STOCHOS. In the interview, he provides insights into what is important when companies want to use artificial intelligence in the field of simulation and digital engineering. In this part, he takes the perspective of the user, the simulation engineer. Kevin  outlines what distinguishes the STOCHOS AI solution and how it is being introduced to productive use in the field of simulation. A conversation about the need for engineering knowledge, implementation paths and the advantages of probabilistic methods.

Kevin, ChatGPT has triggered an AI hype. Suddenly AI has become something tangible. How do you feel about this?

As you know, ChatGPT is an AI that anyone can use, even in private settings. I was just as enthusiastic about it as everyone else. What many people don't know is that the foundations for what we now call AI were laid back in the 1950s, and work on neural networks has been going on ever since. These methods have gained momentum with increasing hardware performance and the enormous amount of data generated by a digitalized world.

The keyword is data: This is crucial for AI and machine learning. How do you obtain valid training data, especially in simulation?

That’s a difficult question, because it very much depends on the task at hand. But what I can say about simulation is that the domain expert, i.e. the simulation professional for mechanics or fluid dynamics, has to take a very close look when generating data and data sets. If, for example, components change significantly and the mesh no longer fits, the results quickly become implausible. This is then often blamed on the AI, despite it not being able to do anything about it. Incidentally, this also makes AI a good tool for checking data quality. If it cannot learn a correlation, this is an indication of inadequate training data.

You therefore need to be very clear about the use case of a simulation that you want to solve with AI. When getting started, consultation is essential to ensure that AI delivers reliable results after implementation, isn’t it?

Yes, I can confirm that. In discussions with customers, it sometimes becomes clear that AI is still sometimes misunderstood. This might also be due to the fact that many opinions are based on what people have experienced with ChatGPT. This is why it is important to talk to each other: What is possible with AI, what is not? Where are the limits? What can be expected? And as you mentioned, it's important to set specific goals beforehand so that success can be measured afterwards.

Optical simulation

Precision & efficiency combined. Find out in our webinars and further information how you can develop your optical designs with maximum precision.

Learn more

Optics_LP

Ansys Startup Program: simulation for startups

Take advantage of the best in simulation technology, right from the start of your business, with a reduced investment.

More info

Teaser_Ansys-Startup-program

AI insights for users: Generate and validate data

Data is the foundation of machine learning. How can users generate high-quality and valid training data? And how can they verify the results of an already trained AI? What are the possibilities for integrating AI solutions into existing workflows?

Watch the video

So, does it help to start small? In other words, do you set up a simple use case and then scale it up?

Small steps are recommended, especially in this area. Simple use cases allow you to see what works and get a feel for what is possible. You can then build on this. This approach is also good for gaining acceptance of AI within the company.

Can Machine Learning be used when generating data records?

Yes, there are methods for this that we have integrated into STOCHOS. We let the AI itself define which data points help it learn the most. This is different from traditional DoE, where lots of data points are generated at once, half of which would often not have been necessary. For example, where you realize that parameters have been forgotten and therefore have to start all over again.

With the adaptive approach, our model says: “I need this geometry for the design because I still have gaps in my knowledge here. Please generate the following data point, which will help me a lot”. In this way, we only generate as many data points that are required for a good model. This only works with probabilistic models. But that's exactly what makes our STOCHOS solution so special. We have models that can predict uncertainties. We can ask the model: where are you most uncertain? We can then generate a data point there.

So, the AI has been trained. How can the results be checked?

This is where the domain expert comes into play again. They should play around with the model a bit and see whether what the model predicts is plausible. Of course, there are also techniques for this, i.e. methods such as cross-validation, where the training data is divided several times into training and test data points and then used to train the model. Test data points are points that were not given to the model for training and are therefore well suited to checking how good it is.

This is also where the probabilistic comes into play again. This is because STOCHOS outputs a confidence interval for every prediction, even later during productive use. If a model was subsequently modified and makes a prediction that significantly deviates from the trained values, our model would say: “I'm very unsure about the prediction here. It would be better if you simulated it again”.

We can even show on the component where, for example, the stress fields are particularly uncertain. This is very important because with classic (non-probabilistic) machine learning methods, such as neural networks, you simply get a prediction and don't know how reliable the result is. Of course, that can be problematic.

Once I see that my model is good and that it works, and I want to use it productively, how do I implement it?

Once you are sure that it works, then it's time to deploy it. Depending on the task, there are several options for this. This is where those of you at CADFEM and your digital engineering expertise come into play. For example, this can be done using a WebApp or direct integration into an existing workflow. We ourselves offer a Python library that can be integrated relatively easily into existing processes. A WebApp can be exported from Stochos with just one line of code.

AI for Engineers in Simulation and Product Development

Learn about the current AI technologies, emerging trends, and how to apply AI to your simulation tasks.

To eLearning offer

ai-for-engineers-in-simulation-and-product-development-21202[1]

Ansys HFSS

Industry standard for determining the radiation and transmission behavior of high-frequency electromagnetic fields.

More info

Ansys_HFSS_Thumb

Editorial

Alexander Kunz

CADFEM Germany GmbH

+49 (0)8092 7005-889
akunz@cadfem.de

Editorial

Klaus Kuboth

CADFEM Germany GmbH

+49 (0)8092 7005-279
kkuboth@cadfem.de