Customization of 2d materials using ml techniques

Abstract

Graphene is a foundational carbon nanomaterial that has been researched extensively since its discovery in 2004 due to its exceptional thermal, electrical, optical, and mechanical properties. However, large-scale synthesis of pristine graphene is usually accompanied by the presence of defects, such as Stone-Wales defects, vacancies, doping, and functionalization. This study focuses on the covalent functionalization of graphene with hydrogen, methyl, and ethyl groups, including selected mixed functionalization cases. This thesis develops an end-to-end inverse design framework to identify the corresponding functionalized graphene layout given a user-desired set of thermomechanical properties. Due to the computational cost of molecular dynamics simulations, brute-force identification of the appropriate layout becomes prohibitive. Therefore, a machine learning surrogate modeling approach, used in tandem with an evolutionary-based optimization algorithm, is adopted.

This study starts with a systematic analysis of 600 unique graphene layouts with hydrogen or methyl functionalizations. Non-equilibrium molecular dynamics simulations with a temperature gradient, along with tensile simulations at a constant strain rate, are used to extract four key material properties: Young’s modulus, maximum stress, strain at maximum stress, and thermal conductivity. All simulations are performed using LAMMPS, an open-source software package for atomistic simulations. In detail, graphene sheets of 220 by 100 Å, with 8,528 carbon atoms in the lattice, are considered. Hydrogen and methyl groups are added with percentage coverages of up to 15% and 12%, respectively. For mixed functionalization cases, the upper limit is set to 10% for each considered functional group. High-coverage functionalization simulations involve a simulation box containing up to 17,000 atoms. Using label and bag-of-words encodings, datasets are generated by concatenating layout features with the corresponding extracted material properties. A variety of regression models are trained and tested on a hold-out test set. Gaussian process regression, k-nearest neighbors, and support vector regression show the best predictive performance, evaluated using four metrics: root mean square error, coefficient of determination, mean absolute percentage error, and normalized root mean square error. Young’s modulus, maximum stress, and thermal conductivity achieve excellent generalization with coefficients of determination greater than 0.9, while strain at maximum stress exhibits some challenges in prediction...