Learning macro from micro

Abstract

Machine Learning Determining atomic structural correlations in condensed-phase systems is crucial for understanding material properties and their behavior at the macroscale. It represents one of the central challenges in modern statistical mechanics because of the complex collective behavior emerging from microscopic many-body interactions. Using two classical condensed-phase models, a Lennard-Jones system and a hard-sphere fluid, Craven et al. show that machine learning methods trained on a set of optimally short molecular dynamics simulations can predict radial distribution functions with increased accuracy by an order of magnitude or even greater compared with traditional analytical approaches. The proposed methodology is general and could be applied more broadly across diverse condensed-phase systems. J. Phys. Chem. Lett. 10.1021/acs.jpclett.0c00627 (2020).