Bridging the gap between quantum mechanics and experiments with atomistic materials simulations using machine learning
Author(s)
Chapman, James Eric
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
As the exploration of materials trends further towards the atomic scale, understanding the dynamic processes that occur at such domains becomes increasingly important. These processes include the nucleation of voids, the coarsening of grains, as well as the growth and melting of surfaces and particles. As these processes are all governed by individual atomic-level interactions, any strategy that aims to probe these regimes, be it experimental or computational, must be capable of accurately capturing the evolution of atomic-level processes. Over the past century, computational techniques have been instrumental in the study of materials at these time and length scales, and have been widely used to explain a multitude of atomistic processes. However, even with the successes of modern computational methods, they are hindered by either the computational cost associated with a particular method (quantum mechanics), limiting the time and lengths scales that can be studied, or by a given model's accuracy (semi-empirical), restricting the types of phenomena that can be simulated. In this thesis, data-driven (machine learning (ML)) methods are utilized to bridge the gap between these two limitations, by combining the accuracy of quantum mechanics with the efficiency of semi-empirical methods. In this thesis, ML models for potential energy, atomic forces, and the total stress tensor are independently constructed for each materials system using generated density functional theory data as their reference. Atomic/nano-scale phenomena, for three elemental systems (Al, Pt, and Li), such as the diffusion of defects on surfaces and within bulk environments, the temperature dependence of lattice, mechanical, the growth of surfaces and grains in large-scale systems containing hundreds-of-thousands of atoms, and the structural properties of liquids and complex defect environments are predicted, using atomistic simulations, to show the breadth of the capabilities of the ML models in bridging the gap between the quantum world and the observable one. The work presented in this thesis highlights the ability of machine learning methods to accurately simulate both small and large-scale phenomena, connecting simulations with experiments, at the accuracy of quantum mechanics.
Sponsor
Date
2020-06-09
Extent
Resource Type
Text
Resource Subtype
Dissertation