Neural Concept, a small 4-year-old Swiss AI company, is undertaking a long-delayed task: putting the “aided” into computer-aided design. Pierre Baqué, founder and CEO, plans to do this by substituting a system trained on a company’s data to be able to predict a product’s performance without doing a simulation. This has the potential to eliminate much of the simulations being done currently and, in some cases, drastically reduce design cycle time.
An extensive background in deep learning (PhD, École Polytechnique Fédérale de Lausanne) has prepared Baqué for his mission, which is nothing less than computer-aided design in the truest sense. Companies, their design cycles purged of time-consuming validation by simulation and their analysts freed from the drudgery of routine simulations, can do deeper design exploration and make more innovative products.
Neural Concept has found a way to apply deep learning to simulation results. When a simulation returns a successful result—whether it be a part that is strong enough or streamlined enough—Baqué’s application takes note. After careful study of many cases, from mapping design requirements to passing performance using previously finite element analysis (FEA) and computational fluid dynamics (CFD) data, Neural Concept is able to suggest a solution for the next design without doing a simulation.
Neural Concept’s AI system could be a design revolution in the offing. We learn that Baqué is visiting San Francisco and, anxious to know more, we rush over to question him.
When did you get started with deep learning?
I started my PhD in deep learning at EPFL in 2014. It was a time when deep learning was getting very popular for computer vision, image recognition. A new wave of AI started there with these deep convolutional neural networks. I was quite lucky to be in this environment. I wrote a paper on using that technology for 3D.
What was your education before your PhD?
Applied mathematics with computer science on the side.
How did this lead to engineering applications?
I had some friends who were working in engineering doing simulation and I learned the fundamental technology. I thought of what I was doing with deep learning and 3D geometry and how I could apply it to engineering. By the end of my PhD, I was working on applications for engineering.
When did you start Neural Concept?
I founded the company in 2018 and started operating it in early 2019.
What made you decide to start a company instead of working for a company?
When I have an idea, I like to carry it through. I could only do that with my own company.
Your company would offer a product or a service?
I was leaning towards a product, but I didn’t have a clear idea of what it would look like. It was more of a set of technologies at the beginning. But how to actually make it into a product that people can use and in what context? That was going to come from our research.
How did you get your funding?
I was part of a research lab and my PhD adviser offered some funding that had come in for a technology transfer.
Are you the only founder?
Two professors were initially cofounders but they are not actively working in the company. One of them is a board member.
What came first? Customers or funding?
We had customers before we were funded. We were working with Airbus initially and with PSA when we launched. We were working with an automotive company in the U.S., too.
What does Neural Concept make and do?
We are bringing simulation and optimization to design. When I say optimization, it is more like performance-driven design suggestions for designers. This is to accelerate the loop or remove the long loops between design and simulation design simulation. We want to give a tool to the designer that does the part of the job that the simulation engineer has done so that the designer can work independently, have much more feedback quickly and get direction automatically.
How does it work?
It is quite simple. The simulation engineer or someone with a similar background, say, a data scientist, will take the data of the company, put it inside our software, do a bit of scientific validation and train the model with the data so it can predict the physics that produce the results. From there, it can proceed directly to results without having to create the mesh and do the numerical solution.
The neural network will learn to make this correspondence between the CAD design and the simulation results without meshing, solving and post-processing.
Are you cutting the simulation engineer out of the loop?
No. The deep learning model has its own uncertainty. It can be used in a design space that it has not been trained for. It is the job of the simulation engineer to validate the model. Not the specific simulation, but to validate that the model is correctly applied in the design space it is trained for and to make sure the model’s limitations are well communicated to the design team.
The simulation engineer is [needed] to validate the correspondence between the CAD design and the simulation results. We provide a set of tools for that validation. Also, the simulation engineer will interact with the AI system from time to time. He will be able to monitor what’s going on and get some feedback on what worked and what didn’t work and be able to retrain this system. Instead of being someone who will react on demand to the designer, the simulation engineer is working much further upstream by training the model, better helping the design engineer.
After the validation, the design team deploys the system and from then on, they can iterate by themselves instead of having to involve the simulation engineer each time.
To be clear, you are going directly from the input to the output, skipping everything in between, including the meshing and actual simulation?
Exactly.
For structural and fluid simulation.
Yes, both. Structural and fluid now, but more and more thermal as well.
What would prevent anyone from using your AI-assisted design system?
There isn’t enough self-change in the industry. People like to keep doing things the same way they’re doing them now. And also, there is some setup. We can’t have it working the next day. One has to make sure they have the right data, train the model, set up the processes to make sure that the model can be retrained. The simulation needs to set the right boundaries around the domain. It is a tool that is efficient, that is usable with repeated simulation and design tasks.
What kind of company would get the most benefit from Neural Concept?
A perfect example is an automotive supplier that is working on one type of design day after day with thousands of modifications. Each has its own request for simulation and possibly a redesign. This company has lots of data for these different designs, which would be of great value in training our system.
Where would Neural Concept be of little benefit?
Let’s say you have a small team designing a crazy aircraft from scratch and without much experience in simulation. That would be a bit more difficult for us. There are no examples to learn from.
But if it is a known problem, solved many times before, your system would be ideal?
Exactly. In my view, simulation should be reserved for the exploration of completely new design concepts. If I’m designing a really new and different system and I am validating the design with experiments, then use simulation. But if you are using analysis during production, use our system instead.