Machine Learning Anisotropic Coarse-Grained Simulation Models of Small-Molecule and Polymeric Organic Semiconductors

Abstract

A set of machine learning workflows have been developed to automate the generation of accurate anisotropic coarse-grained models and interaction potentials for small molecules and polymers as well as to analyze the aggregate structure of dilute semiflexible polymers with anisotropic monomers. The multiscale coarse-graining method for isotropic coarse-grained particles has been extended to anisotropic coarse-graining of small molecules and polymers using a mixture of machine learning tools and classical simulation methods. The resulting coarse-grain interaction potentials derived from the machine-learned forcematching approach are flexible and scalable with respect to the type of molecules, the size of the simulation, and the simulation conditions. The robust deep-learning models were specifically used to construct coarse-grained interaction potentials for single-site anisotropic modeling of organic molecules and have shown the capability of reproducing the liquid crystal phase behavior of organic semiconductors. An autoencoder machine learning approach has been used to automate the encoding of atomistic trajectories into unique anisotropic coarse-grained sites. This automated procedure allows for the creation of a simplified representation of organic polymers with the added feature of an accurate back mapping to the atomistic trajectories using the decoder network. Machine learning tools are also developed in this work to analyze and predict the aggregation tendencies of small anisotropic molecules and organic semiconducting polymers in either the liquid or solution phase. A practical deep-learning framework, for the anisotropic coarse-graining of polymers and anisotropic macromolecules, was implemented alongside an automated workflow to predict the polymers’ key aggregation behaviors based on their structure, flexibility, and the simulation condition.