Identifying local structures in dipolar colloid-polymer mixtures using machine learning

Abstract

A colloid-polymer mixture with an applied electric field is subject to two categories of forces, the induced dipole-dipole interactions and depletion forces due to the polymer. The combination of these forces with different polymer concentrations and external electric fields results in the formation of various structures with different local orders. To study these structures in more detail, the colloid-polymer system can be replicated computationally using the molecular dynamics method which enables us to calculate particle features such as bond order parameters computationally and further investigate these features using advanced methods such as machine learning. In this thesis, we apply a multi-step machine learning algorithm to dipolar-depletion systems and identify the local structure of atoms in the simulation box using the algorithm. The machine learning algorithm is a combination of multiple cutting-edge machine learning techniques including autoencoders, Gaussian mixture models, and a cluster merging technique. These algorithms are combined to create a multi-step process that can identify different structures of matter in any molecular dynamics simulation output. This algorithm utilizes unsupervised machine learning which does not require labeled data and is applicable to known and unknown local structures. The machine learning model can identify different local structures in colloid-polymer systems and the identified clusters of atoms in the systems is in complete agreement with our understanding of the systems.