Classification of atomic environments via the Gromov-Wasserstein distance

TL;DR

This paper introduces a novel method using the Gromov-Wasserstein distance to classify atomic environments in molecular simulations, accommodating various structural and compositional differences without prior assumptions.

Contribution

It proposes a flexible, metric-based approach for classifying atomic environments that surpasses existing methods in handling displacements, missing atoms, and chemical diversity.

Findings

Effective in distinguishing atomic environments with structural variations

Versatile across different material classes

Handles missing atoms and compositional differences

Abstract

Interpreting molecular dynamics simulations usually involves automated classification of local atomic environments to identify regions of interest. Existing approaches are generally limited to a small number of reference structures and only include limited information about the local chemical composition. This work proposes to use a variant of the Gromov-Wasserstein (GW) distance to quantify the difference between a local atomic environment and a set of arbitrary reference environments in a way that is sensitive to atomic displacements, missing atoms, and differences in chemical composition. This involves describing a local atomic environment as a finite metric measure space, which has the additional advantages of not requiring the local environment to be centered on an atom and of not making any assumptions about the material class. Numerical examples illustrate the efficacy and versatility of the algorithm.