Unsupervised identification of crystal defects from atomistic potential descriptors

TL;DR

This paper compares unsupervised algorithms like PCA, UMAP, and PaCMAP for identifying crystal defects from atomistic potential descriptors, offering a system-independent approach for analyzing molecular dynamics simulations.

Contribution

It introduces a systematic comparison of three unsupervised algorithms for defect identification, highlighting PaCMAP's robustness and UMAP's effectiveness in class-imbalanced scenarios.

Findings

PaCMAP outperforms PCA and UMAP in classification robustness.

UMAP excels in cases with significant class imbalance.

Both UMAP and PaCMAP successfully identify nuclei in supercooled water.

Abstract

Identifying crystal defects is vital for unraveling the origins of many physical phenomena. Traditionally used order parameters are system-dependent and can be computationally expensive to calculate for long molecular dynamics simulations. Unsupervised algorithms offer an alternative independent of the studied system and can utilize precalculated atomistic potential descriptors from molecular dynamics simulations. We compare the performance of three such algorithms (PCA, UMAP, and PaCMAP) on silicon and water systems. Initially, we evaluate the algorithms for recognizing phases, including crystal polymorphs and the melt, followed by an extension of our analysis to identify interstitials, vacancies, and interfaces. While PCA is found unsuitable for effective classification, it has been shown to be a suitable initialization for UMAP and PaCMAP. Both UMAP and PaCMAP show promising results overall, with PaCMAP proving more robust in classification, except in cases of significant class imbalance, where UMAP performs better. Notably, both algorithms successfully identify nuclei in supercooled water, demonstrating their applicability to ice nucleation in water.