BEGIN:VCALENDAR VERSION:2.0 PRODID:-//wp-events-plugin.com//6.4.7.2//EN TZID:Europe/Paris X-WR-TIMEZONE:Europe/Paris BEGIN:VEVENT UID:0-1173@lptms.universite-paris-saclay.fr DTSTART;TZID=Europe/Paris:20261106T140000 DTEND;TZID=Europe/Paris:20261106T173000 DTSTAMP:20260910T092806Z URL:http://www.lptms.universite-paris-saclay.fr/seminars/soutenance-de-the se-dimitrios-tzivrailis/ SUMMARY:Soutenance de thèse Dimitrios Tzivrailis - Centre CEA PARIS-SACLAY \, Digiteo - 6 Nov 26 14:00 DESCRIPTION:AI-Driven Monte Carlo Simulations with Learned Energy Surrogate s: Uncertainty Quantification and Energy-Based Regularization\nDimitrios T zivrailis\n \;\nLieu de la soutenance : Centre CEA PARIS-SACLAY. Digit eo. 91190 Gif-sur-Yvette.\nsalle Amphithéâtre de DIGITEO\n \;\n\nMar kov Chain Monte Carlo (MCMC) methods are the standard tool for estimating thermodynamic observables in statistical physics and computational chemist ry\, but their cost is dominated by repeated evaluations of the system's e nergy and forces — a cost that becomes prohibitive when these quantities require expensive first-principles calculations such as density functiona l theory. Machine learning surrogates promise to remove this bottleneck by replacing exact evaluations with fast\, learned approximations. This thes is shows that such surrogates introduce two distinct failure modes that un dermine the statistical correctness of the resulting simulation\, and deve lops one targeted method to address each. The first failure mode is episte mic uncertainty: a surrogate trained on a finite dataset carries predictio n noise that grows as the Markov chain explores regions poorly represented in training. Because the Metropolis acceptance rule is a nonlinear functi on of the energy difference\, even zero-mean noise breaks detailed balance and biases the sampled distribution — a failure invisible to standard r egression metrics. To address this\, we introduce the Penalty Ensemble Met hod (PEM)\, which combines a deep ensemble of surrogates with a noise-pena lty correction derived from the framework of Ceperley and Dewing\, convert ing the ensemble's predictive variance into a term that suppresses accepta nce in unreliable regions. Validated on the two-dimensional ϕ4 lattice fi eld theory across its ferromagnetic\, paramagnetic\, and critical phases\, PEM restores the correct stationary distribution while introducing only a modest computational overhead. The second failure mode is rooted in the t raining objective itself: standard mean-squared-error (MSE) training carri es no thermodynamic content and leaves the energy landscape unconstrained outside the training distribution\, allowing spurious low-energy minima to form and trap the Markov chain in unphysical configurations. To address t his\, we introduce Contrastive Regularization for MSE (CRMSE)\, a training -time correction inspired by Persistent Contrastive Divergence that augmen ts the MSE loss with a contrastive term penalizing energies at configurati ons visited by a failed surrogate-driven chain. Validated on two molecules from the rMD17 benchmark — ethanol and aspirin — CRMSE eliminates the spurious minima responsible for sampling failure\, preserves held-out pre dictive accuracy\, and recovers correct interatomic-distance distributions and dihedral free-energy profiles\, including in a data-scarce regime\; a complementary validation on the ϕ4 model\, presented in an appendix\, co nfirms that the same mechanism corrects the sampling bias in a simpler\, a nalytically tractable setting. PEM and CRMSE form a single progression: PE M establishes that an inference-time correction is viable and shows where it breaks down\, while CRMSE removes that dependency altogether and extend s the approach to real molecular systems — together making AI-accelerate d Monte Carlo sampling both computationally efficient and statistically re liable.\n\nJury : Claudio Attaccalite (rapporteurs)\, Aurélien Decelle\, Michel Ferrero (rapporteurs)\, Eiji Kawasaki (co-encadrant de thèse)\, Al berto Rosso (directeur de thèse)\,  Véronique Terras\, Julien Tranchida CATEGORIES:seminars LOCATION:Centre CEA PARIS-SACLAY\, Digiteo\, salle Amphithéâtre de DIGITE O\, 91190 Gif-sur-Yvette\, France X-APPLE-STRUCTURED-LOCATION;VALUE=URI;X-ADDRESS=salle Amphithéâtre de DIG ITEO\, 91190 Gif-sur-Yvette\, France;X-APPLE-RADIUS=100;X-TITLE=Centre CEA PARIS-SACLAY\, Digiteo:geo:0,0 END:VEVENT BEGIN:VTIMEZONE TZID:Europe/Paris X-LIC-LOCATION:Europe/Paris BEGIN:STANDARD DTSTART:20261025T020000 TZOFFSETFROM:+0200 TZOFFSETTO:+0100 TZNAME:CET END:STANDARD END:VTIMEZONE END:VCALENDAR