Summary

In 2025, machine learning continues to revolutionize various scientific fields, with major implications in physics, particularly photonics. The integration of advanced machine learning algorithms has enabled significant advances in the design and control of photonic systems, improving their efficiency and performance. These advances are essential for the development of communication, imaging and quantum computing technologies. However, physics research presents many challenges that go beyond simple performance measurements: identifying patterns and building analytical models is often just as crucial.

In this thesis, we apply computational intelligence tools, in particular heuristic optimization and neural networks, to develop data-driven approaches to solving various tasks in physics. Although data-centric, our approach remains rooted in physics, always seeking to understand the physical phenomena underlying the algorithms. The results of this thesis cover a wide range of applications, from the design of complex metasurfaces and diffraction gratings to the analysis and interpretation of spectral data. We have also successfully developed an optimizer capable of learning and adapting to the problems encountered, particularly in physics. This key tool in our arsenal outperforms state-of-the-art methods in our applications. In particular, it has enabled the design of a coronagraphic phase plate for exoplanet imaging, with a performance 25% better than the best previous solutions.

We have also designed a compact, all-dielectric beam deflection device, operating efficiently for all polarizations, reaching a maximum efficiency of 90%. Starting from a purely data-driven design, we were able to extract and validate an analytical model based on the behavior of an echelle lattice, providing a physical understanding of its operation. In addition to simulation-based tasks, we also processed experimental data, developing an animal origin classifier for scrolls, capable of distinguishing three animal species with 90% accuracy. This tool offers a non-invasive method for conservators and historians wishing to analyze fragile historical materials.

Jury members

  • Prof. Michaël LOBET (UNamur), President
  • Prof. Alexandre MAYER (UNamur), Secretary
  • Dr. Charlotte BEAUTHIER (CENAREO)
  • Prof. Benoît FRENAY (UNamur)
  • Prof. Olivier DEPARIS (UNamur)
  • Prof. Denis LANGEVIN (Université de Clermont Auvergne)
  • Prof. Hai Son NGUYEN (Ecole Centrale de Lyon)