The bcc coating of Lennard-Jones crystal nuclei vanishes with a change of local structure detection algorithm

TL;DR

This study examines how different local structure detection algorithms influence the characterization of Lennard-Jones crystal nuclei, revealing that the choice of method significantly affects observed polymorph compositions and the apparent presence of bcc coatings.

Contribution

The paper compares seven local structure detection algorithms and demonstrates that incorporating outlier detection with machine learning alters the perceived crystal structures.

Findings

Polymorph composition varies with the detection algorithm used.

Outlier detection improves confidence in local structure classification.

The bcc coating nearly disappears with machine learning and outlier detection.

Abstract

Since the influential work of ten Wolde, Ruiz-Montero, and Frenkel [Phys. Rev. Lett. 75, 2714 (1995)], crystal nucleation from a Lennard-Jones fluid has been regarded as a paradigmatic example of metastable crystal ordering at the surface of a critical nucleus. We apply seven commonly used local structure detection algorithms to characterize crystal nuclei obtained from transition path sampling simulations. The polymorph composition of these nuclei varies significantly depending on the algorithm used. Our results indicate that one should be very careful when characterizing the local structure near solid-solid and solid-fluid interfaces. Particles near such interfaces exhibit a local structure distinct from that of bulk fluid or bulk crystal phases. We argue that incorporating outlier detection into the local structure detection method is beneficial, leading to greater confidence in the classification results. Interestingly, the bcc coating nearly disappears when adopting a machine learning method with outlier detection.