Abstract

Machine learning (ML) methods have revolutionized atomistic simulations, enabling highly accurate simulations and analyses at the fraction of the computational cost. Central to these advances is the use of atom-centered numerical representation of the atomistic system, where one transforms the coordinates and identities of each atom in a way that preserves the symmetries of the system. However, atom-centered representations, such as the popular Smooth Overlap of Atomic Positions (SOAP), are not as well suited for describing large macromolecular systems; in such cases, one would likely be more interested in understanding how groups of atoms interact with each other, either from a scientific or efficiency standpoint. To properly create a representation for groups of atoms, we introduce an anisotropic generalization of SOAP , which we deem AniSOAP. This generalized descriptor can describe the complex molecular geometries and capture orientation-dependent interactions that occur between groups of atoms. In this talk, I will present three different case studies that use AniSOAP, ranging from unsupervised analyses of liquid crystals to learning complicated benzene energetics. From these studies, AniSOAP gives us a data-driven way to observe how the molecular geometry influences the formations of certain phases or the energetics of particular configurations. I will then conclude the talk by describing how AniSOAP can be incorporated into a generalized coarse-grained simulation framework, and provide my thoughts on how it can be used to quantify information-loss incurred within coarse-graining.