Accelerated Sampling of Rare Events with Entropy-Based Collective Variables

Understanding phase transitions in materials science is essential for rationalizing and tailoring their design and mechanisms. Enhanced sampling methods such as umbrella sampling or metadynamics try to accelerate the sampling of novel configurations within molecular dynamics (MD) simulations beyond oversampled energy minima. However, its efficiency depends on the choice of collective variables (CVs) that can effectively distinguish between different phases and span the configurational space. Defining these CVs usually requires prior knowledge about the states of interest, which limits their use in exploring arbitrary potential energy surfaces (PESes). In this study, we introduce an information theory approach to perform enhanced sampling simulations utilizing information contents as CV. The entropy values account for the similarity between the instantaneous simulation frame and a reference frame by comparing their atomic environments, and are used to push the system towards new distributions based on these quantifiable overlaps. This enables a blind exploration of metastable states connected by transition pathways consistent with the underlying PES. The methods are implemented using the QUESTS (Quick Uncertainty and Entropy via STructural Similarity) package and are demonstrated for several examples, from simple molecular systems to solid-state phase transitions. This approach can enable a systematic understanding and exploration of phase transformations in materials using atomistic simulations.