Seminar by Joachim Bona-Pellissier: "PIKS: Physics-informed kernel methods"
Monday, November 2, 2026
at 2pm
Room B023
Laboratoire Hubert Curien
Joachim Bona-Pellissier, Postdoctoral researcher, MaLGa in Genoa, Italy.
Abstract
Physics-informed machine learning incorporates physical principles—often expressed through differential operators—into data-driven models. In this talk, I will introduce Physics-Informed Kernel methodS (PIKS), a framework for physics-informed learning in reproducing kernel Hilbert spaces (RKHS). I will discuss their asymptotic behaviour in the misspecified setting, where the target function does not need to belong to the hypothesis space. These results extend classical kernel learning guarantees to settings where the target’s smoothness is unknown: under suitable universality assumptions on the kernel, satisfied by Gaussian and sufficiently smooth Matérn kernels, PIKS consistently learns both the target function and its differential structure.
I will then discuss convergence rates in the well-specified setting (where the target belongs to the RKHS), focusing on how much physical information can accelerate learning. The analysis reveals two regimes: with limited differential data, the prediction rate depends jointly on the numbers of value and differential observations; beyond a problem-dependent threshold, it matches the rate achievable when the physical constraint is enforced exactly.
This seminar will be held in English.
