Identifying structural changes with unsupervised machine learning methods

TL;DR

This paper demonstrates how unsupervised machine learning, specifically PCA and k-means clustering, can effectively identify phase transitions like melting points in molecular dynamics simulations by analyzing atomic radial distributions.

Contribution

The study introduces a novel application of PCA and clustering to detect structural phase changes in materials, providing an automated alternative to manual interpretation.

Findings

Accurately estimates melting points in metallic systems.

Comparable results to traditional methods on small systems.

Shows machine learning's potential in physical system analysis.

Abstract

Unsupervised machine learning methods are used to identify structural changes using the melting point transition in classical molecular dynamics simulations as an example application of the approach. Dimensionality reduction and clustering methods are applied to instantaneous radial distributions of atomic configurations from classical molecular dynamics simulations of metallic systems over a large temperature range. Principal component analysis is used to dramatically reduce the dimensionality of the feature space across the samples using an orthogonal linear transformation that preserves the statistical variance of the data under the condition that the new feature space is linearly independent. From there, k-means clustering is used to partition the samples into solid and liquid phases through a criterion motivated by the geometry of the reduced feature space of the samples, allowing for an estimation of the melting point transition. This pattern criterion is conceptually similar to how humans interpret the data but with far greater throughput, as the shapes of the radial distributions are different for each phase and easily distinguishable by humans. The transition temperature estimates derived from this machine learning approach produce comparable results to other methods on similarly small system sizes. These results show that machine learning approaches can be applied to structural changes in physical systems.