Antoine de Mathelin: Reliable Machine Learning for Product Design

A PhD graduate from the Centre Borelli (CNRS/ENS Paris-Saclay), Antoine de Mathelin was awarded in 2025 the second prize in the “Maths, Business & Society” thesis competition for work situated at the intersection of applied mathematics, machine learning, and design engineering. His dissertation, “Towards Reliable Machine Learning under Domain Shift and Costly Labeling Constraints: Application to Product Design,” was conducted within the Hadamard Doctoral School under the supervision of Mathilde Mougeot and Nicolas Vayatis, as part of a CIFRE partnership with Michelin.

A framework for robust predictive models

The research focuses on the development of machine learning models intended to serve as surrogate models for the design of industrial systems, in contexts where data are costly to generate and may reflect conditions that differ from those of actual use. The dissertation examines in particular the handling of domain shift and the use of active learning strategies to selectively and efficiently acquire the new evaluations required.

From a methodological standpoint, the author combines tools for uncertainty quantification, domain adaptation, and sequential optimization, with the objective of delivering predictions that can be effectively integrated into design processes subject to industrial constraints. These contributions are supported by a software implementation, notably through a Python library for transfer learning made available to R&D teams.

Industrial applications

The methods developed were evaluated on use cases provided by the thesis partners. At Michelin, they are applied to datasets derived from different testing campaigns, with a view to adapting models to new contexts without fully repeating measurement processes. At Renault, they are integrated into design pipelines in order to reduce the number of simulations required to explore the design space.

The prize jury highlighted the diversity of the contributions (theoretical, experimental, and software-based) as well as the effective deployment of the methods within industrial environments.

An example of laboratory–industry collaboration

This dissertation illustrates the collaborative dynamic between the Centre Borelli, ENS Paris-Saclay, and their socio-economic partners around issues of modeling and decision-making under uncertainty. By addressing questions of model reliability and data acquisition frugality, it also aligns with the laboratory’s research priorities in the optimization of experimental design and the sustainability of scientific computing practices.