AMIES Award 2025 | Antoine De Mathelin: developing machine learning models to optimize industrial product design

Sponsored by the learned societies Société Française de Statistique (SFDS), Société de Mathématiques Appliquées et Industrielles (SMAI), and Société Mathématique de France (SMF), the 2025 thesis prize was awarded during the 14^th edition of the Forum Entreprises & Mathématiques on Tuesday, October 7, 2025.

• Thesis title: Towards reliable machine learning under domain shift and costly labeling, with applications to engineering design
• University of issuance: Paris-Saclay University
• Thesis supervisor: Prof. Mathilde Mougeot (Borelli Center ENS Paris-Saclay and ENSIIE), co-supervised by Prof. Nicolas Vayatis (Borelli Center ENS Paris-Saclay).
• Industrial supervisor: Dr. François Deheeger (Michelin)
• Company: Michelin
• Currently: Postdoctoral Research Fellow at the Sloan Kettering Institute for Cancer Research (New York)

What motivated you to do a thesis related to the socio-economic world?

I was doing an R&D internship at Michelin when a partnership with ENS Paris-Saclay began. The opportunity to continue with a thesis quickly presented itself after the internship. I was interested in the research topic, the team was motivating, and I had confidence in the scientific direction of the project.

Can you tell us about your thesis topic?

My thesis is directly motivated by concrete applications at Michelin. We began developing machine learning models to improve the industrial design of tires and optimize their performance. However, we found that models trained on historical test data did not generalize well to new products: this is a problem known as “domain shift.”

In industrial design, this point is crucial: we rely on model predictions to design new product ranges. My work therefore consisted of proposing solutions to correct these domain shifts, in particular by adapting the models or delimiting their area of validity. Overall, my thesis work proposes tools and methods for the reliable use of machine learning models in the context of industrial product design.

What were the main challenges you encountered during your research?

One of the main challenges was finding the right balance between academic research and industrial expectations. Writing scientific articles takes time: you have to master the state of the art, build relevant benchmarks, and develop the theory. At the same time, we had to produce high-quality code that could be used by teams, with a level of reliability close to that of production. This dual requirement was a great learning experience.

How has your work benefited the socio-economic world today, or how will it benefit it in the future?

On the business side, some of the elements developed during the thesis have been integrated into the production and validation chain for machine learning models. More broadly, the work has raised awareness among teams about the challenges of domain change and uncertainty quantification, providing a better understanding of the underlying mechanisms.

On the scientific front, the published contributions represent advances in the fields of transfer learning, active learning, and uncertainty quantification. Today, these topics play an important role in the development of machine learning, both in industry and in bioinformatics.

What advice would you give to young people who want to focus their research in mathematics on practical applications?

I would advise them to learn how to program well and produce high-quality code. Much of the application of mathematics today involves computer science. AI helps a lot, of course, but to become an expert in your field, you need to understand what is going on behind the tools. This technical skill is very useful for transforming a mathematical idea into a concrete solution.

How do you see the role of mathematics in solving current societal issues?

Even if mathematics simplifies reality, it offers a better understanding of the world and its underlying laws. I think that this knowledge is already a good thing in itself. Furthermore, when it comes to solving today's concrete problems, fields such as health, agriculture, industry, and leisure benefit from advances in mathematics, if only through the optimization of certain processes.

That said, I believe that some of today's challenges, such as isolation and marginalization in our Western societies, cannot be solved by technology alone. They require more profound societal changes. These are not problems that can simply be “put into an equation.”

As a young researcher, how do you perceive the evolution of the link between academic research in mathematics and the socio-economic world?

I think that today, with advances in AI, the socio-economic world is increasingly linked to academic research. We are seeing that certain theoretical developments are very quickly finding practical applications. There are more and more partnerships between academic research groups and companies. Some private companies are also contributing to scientific activity, particularly through the funding of conferences or the publication of research work.

What are your plans for the future?

I am currently doing postdoctoral research in the field of health, where I am developing methods similar to those explored in my thesis, but applied to the design of new cancer therapies. I may continue in this field or return to industrial applications. I am also interested in teaching.