MAchine LeArning bY StochastIc Approaches: application to water clusters

Members

• Casula Michele (Sorbonne - IMPMC - CNRS UMR 7590) -- Physicist
• Ludovic Goudenège (FDM - CNRS FR 3487) Mathématicien - Responsable Equipe Partenaire
• Saitta Marco (Sorbonne - IMPMC - CNRS UMR 7590) -- Physicist - Porteur du projet
• Vuilleumier Rodolphe (ENS - PASTEUR - CNRS UMR 8640 -- Chemistry

Résumé

Thanks to recent progress, machine learning (ML) methods have shown extraordinary predictive capabilities in a great variety of domains, including chemistry, physics, and quantitative finance. In this project, we address the specific target of combining ML methods with stochastic approaches. Randomness can affect the learning process in many ways. We will build new ML schemes which are not only robust against noise-dominated training datasets, but could also exploit randomness to accelerate the learning process, by exploiting the stochastic occurrence of events obtained from unbiased sampling.
We will lay down the mathematical bases of this stochastic semi-supervised framework, and we will apply it first to simulate water clusters, a notoriously challenging system in quantum chemistry, where appropriate training sets will be generated by both classical and quantum Monte Carlo methods. Then we will extend the methodology to DNA bases aiming at a precise description of canonical and non-canonical hydrogen bond pairing as well as the van der Waals interactions between bases.