Nucleation patterns of polymer crystals analyzed by machine learning models

TL;DR

This paper introduces a machine learning approach using auto-encoders and Gaussian mixture models to detect and analyze crystalline phases in polymer melts from molecular dynamics simulations, providing detailed insights into nucleation patterns.

Contribution

The study presents a novel, structure-agnostic machine learning method for identifying crystalline monomers, enabling detailed temporal analysis of nucleation in polymers.

Findings

The ML method agrees with traditional classifiers in identifying crystalline monomers.

Crystalline order can be detected before thermodynamic signatures appear.

Maximum crystallization efficiency occurs at a pre-transition point.

Abstract

We use machine learning algorithms to detect the crystalline phase in undercooled melts in molecular dynamics simulations. Our classification method is based on local conformation and environmental fingerprints of individual monomers. In particular, we employ self-supervised auto-encoders to compress the fingerprint information and a Gaussian mixture model to distinguish ordered states from disordered ones. The resulting identification of crystalline monomers agrees to a large extent with human-defined classifiers such as the stem-length-based classification scheme as developed in our previous work [C. Luo and J.-U. Sommer, Macromolecules 44 (2011), 1523], but does not require any foreknowledge about the structure of semi-crystalline polymers. Because of its local sensitivity, the method allows the resolution of detailed time patterns of crystalline order before an apparent signature of the transition is visible in thermodynamic properties such as for the specific volume. At a pre-transition point, we observe the highest crystallization efficiency using the fraction of monomers being conserved in the crystalline phase as compared to the number of monomers joining that phase.