PhD Student in Applied Mathematics
Machine learning for the control of complex dynamical systems — Centre Borelli, ENS Paris-Saclay, in collaboration with Framatome.
Office 3S28 · ENS Paris-Saclay · Gif-sur-Yvette, France
I am a 3rd year PhD Student in Applied Mathematics at Centre Borelli (ENS Paris-Saclay) under supervision of Nicolas Vayatis. I am working in collaboration with Framatome, a world leader in civil nuclear industry.
Before starting my PhD, I graduated from the MVA Master’s program at ENS Paris-Saclay, the Master’s in Mathematics of Modeling at Sorbonne Université, and earned a Civil Engineering degree from Mines Nancy.
I am also a Teaching Assistant (TA) for the MVA course Introduction to Statistical Learning for the 2026-2027 academic year.
My work focuses on the control of complex dynamical systems under real-world constraints, combining machine learning and model-based approaches to improve both computational efficiency and robustness. During my PhD, I develop ML surrogate models for industrial dynamical systems and integrate them into nonlinear MPC pipelines with formal safety guarantees. My work bridges surrogate dynamics learning, differential algebraic equations, and optimal control — with a focus on sample efficiency and stability.
All my publications are listed here and on my Google Scholar. For any request, feel free to contact me by email.
Structuring the latent space by timescale makes behavior cloning of an NMPC expert for nuclear load-following more accurate and more often feasible — and, used to warm-start the NMPC, it recovers full feasibility while cutting computation time by 15%.
A summary of my work on applying Machine Learning methods to Optimal Control of Nuclear Power Plants.
Research internship supervision.
Research internship supervision.
In long-horizon nonlinear MPC, we show that learned surrogate dynamics can significantly accelerate computation while preserving safety.
This paper develops a ML surrogate simulation scheme for fast integration of Differential Algebraic Equations modeling a nuclear reactor core.
This paper develops two major improvements on the global optimization method LIPO from Malherbe & Vayatis, 2017: an empirical stopping criterion and a decaying exploration rate.
Unifying the views of AdaBoost, in order to better understand its dynamics.