Chris Chipot
Unité Mixte de Recherche, CNRS/UL 7565, Université de Lorraine, Vandoeuvre-les-Nancy, France
At the atomic scale, simulating transitions between stable states of the free-energy landscape, often impeded by sluggish molecular processes, constitutes a daunting challenge. One common approach to overcoming these barriers and speeding up dynamics is supplied by importance-sampling algorithms, which require well-defined reaction-coordinate models using compact sets of collective variables (CVs). While traditional methods rely on intuition for reducing dimensionality, recent advancements in machine learning (ML) offer robust alternatives. We compare two data-driven ML techniques, the state-free reversible variational approach for Markov process networks (SRVs) and variational committor-based neural networks (VCNs), to identify the slowest decorrelating CV and committor probability in transitions between stable states. Using simple model systems, both methods demonstrate the ability to identify relevant descriptors and can be adapted for importance-sampling schemes using a reweighting algorithm that approximates kinetic transition properties.
