The full text of this item is not available at this time because the student has placed this item under an embargo for a period of time. The Libraries are not authorized to provide a copy of this work during the embargo period, even for Texas A&M users with NetID.
Data Driven Interatomic Potential Development
Abstract
The advancements in materials have driven significant progress in humanity which is largely enabled by a deeper understanding of fundamental materials science. For materials, the potential energy surface of the atomistic systems governs both their thermodynamics and kinetic properties. The exact potential energy surface lives in an extremely high-dimensional space where electronic degrees of freedom pose particular challenges due to the inherent quantum-mechanical wave nature of electrons which makes them computationally expensive to simulate. Therefore, many thermodynamics and kinetic properties are impractical to calculate using a brute-force quantum mechanical simulation. Instead, interatomic potentials, i.e. force fields, are introduced to map the complex quantum system onto a classical system using few-body interatomic interactions. These interatomic potentials facilitate large-scale molecular dynamics simulations, allowing for the calculation of thermodynamics and kinetic properties and providing atomistic insights into the behavior of the materials. This thesis focuses on the development of interatomic potentials using data-driven methods.
Chapter 1 discusses the general background and concepts for interatomic potentials. In this chapter, we introduce the general form and requirements for an interatomic potential. We then delve into both classical and machine learning interatomic potentials. Three classical interatomic potentials (Lennard-Jones Potential, Tersoff Potential, and Embedded Atom Method) and three machine learning interatomic potentials (Spectral Neighbor Analysis Potential, DeePMD, and POET) will be discussed in detail. We also derive the forces and stress using the Lennard-Jones interatomic potential as an example.
Chapters 2 and 3 discuss the generation and selection of structures for training machine learning interatomic potentials. Chapter 2 focuses on selecting structures to train machine learning potentials using a batch active learning method. The batch active learning method selects structures using the structure’s dissimilarity and the potential’s predictive capability, which allows for the generation of accurate and consistent machine learning interatomic potentials using fewer structures. Chapter 3 focuses on the efficient generation of structures for training machine learning interatomic potentials across a wide temperature range, which necessitates the inclusion of an-harmonic effects. To achieve this goal, we develop a self-consistent phonon-based method for generating atomic structures at finite temperatures which enables faster and cheaper training and validation of machine learning interatomic potentials.
Chapter 4 develops a method for the efficient prediction of material properties by fitting classical interatomic potentials. Specifically, we develop a data-driven approach for optimizing classical interatomic potentials using only a few structures calculated by density functional theory. These classical interatomic potentials are then used to predict material properties accurately at a much smaller cost than the brute-force density functional theory approach.
Chapter 5 summarizes the major research findings in this dissertation and elaborates on future research directions.
Citation
Wilson, Nathan Robert (2022). Data Driven Interatomic Potential Development. Doctoral dissertation, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /198470.