Friday, July 5th, 15h, Salle des thèses 580F, Bâtiment de la Halle aux Farines, 10 rue Françoise Dolto, 75013 Paris.
PhD defense of Federico Ghimenti
Irreversible sampling of glassy systems.
supervised by Frédéric van Wijland
Abstract:
In silico sampling of the Boltzmann distribution for a system of many interacting degrees of freedom is a ubiquitous task in natural sciences, and it becomes challenging in some relevant fields, from protein folding to machine learning. Amorphous systems are one of the hardest instances of this problem, due to the impressive slowdown of their dynamics upon mild variation of temperature or density.
This theses addresses how and to what extent sampling in glassy systems can be accelerated by resorting to a specific class of nonequilibrium dynamics, which are unphysical, but tailored to sample the Boltzmann distribution in their steady state. We quantify the speedup and characterize the dynamical pathways uncovered by these dynamics in a variety of models, from mean field spin glasses to structural glasses in two, three and infinite dimensions. We discuss how these ideas can be put in practice to outperform current state-of-the art algorithms for structural glass formers.