Unsupervised machine learning methods for polymer nanocomposites data via molecular dynamics simulation

Abstract

ABSTRACT Molecular dynamics (MD) simulation has been an invaluable tool for polymer nanocomposites (PNCs) research. MD simulation investigations provide massive data about the movements of PNCs beads on the micro-level. However, there are many challenges involved in correctly understanding and predicting the performance and properties of molecules from these massive MD simulation data. Traditional data-driven techniques are limited due to computational expense, required accuracy, and the availability of high-dimensional PNCs structures. Machine learning can give an effort to process massive data, but not all machine learning methods are available for all occasions; domain knowledge or manual intervention is required. Here, we aim to investigate unsupervised machine learning algorithms to discover clustering and shaping movements of PNCs beads under tension. Cluster Envelope Shaping Method (CESM) was proposed to analyze the clustering of beads and the shaping of clusters based on MD simulation data. The matrix-free PNC based elastomer simulation and Single-Chain Polymer Nanoparticles simulation, which demonstrate two types of deformations, verified the validity of this method.