Deep Learning for Morphology Detection of Self-Assembly in Atomistic Simulation

Student

Zhengyuan Shen

Advisor

Ilja Siepmann

Abstract

Amphiphiles including block polymers and block oligomers can self-assemble into numerous ordered morphologies spanning length scales from a few to hundreds of nanometers. Molecular simulations using atomistic or coarse-grained force fields are widely used as a powerful tool to understand and predict the self-assembly phase behavior of these complex systems. In molecular simulations, screening block oligomer structures to search for promising candidates for functional materials is facilitated by straightforward but effective structure identification techniques. However, capturing ordered network structures with periodicity in all three dimensions is challenging because the size of the periodic simulation box and the number of molecules inside the box must be commensurate with the unit cell of the resulting network geometry. Simulations with incorrect system sizes can yield imperfect and distorted structures, and common morphology identification strategies including structure factors and order parameters may fail in assigning the morphology. Here we present a new structure detection approach based on PointNet, a neural network designed for computer vision applications using point clouds. Our PointNet was trained using atomic coordinates from molecular dynamics simulation trajectories and synthetic point clouds for ordered network morphologies which were absent from previous simulations. The trained PointNet model achieves an accuracy as high as 0.996 with high robustness against missing data for globally ordered morphologies formed by block oligomers of various architectures and their mixtures, and it also allows for the discovery of emerging ordered patterns from non-equilibrium systems.

Video

Deep Learning for Morphology Detection of Self-Assembly in Atomistic Simulation