Understanding material properties and performance enabled by molecular simulations and machine learning potentials

Abstract

Machine learning potentials (MLPs) are being rapidly adopted to describe complex potential energy surfaces and to solve emerging materials science and engineering problems. Trained against quantum mechanical datasets, MLPs are versatile mathematical surrogates to describe various chemical and physical phenomena originating from atomistic interactions. MLPs are notably superior over classical interatomic force fields owing to their high training flexibility, efficiency, and accuracy. In this thesis, artificial neural network (ANN), sparse Gaussian process (SGP), and neural equivariant MLPs were developed and applied in the investigations of mechanical and thermodynamic behaviors of metals, main-group semiconductors, and perovskites. In Chapter 3, a high-dimensional NN potential (HDNNP) was developed for the condensed phase nickel. This HDNNP trained using the geometry and force data extracted directly from ab initio molecular dynamics can predict the melting point of face-centered cubic nickel within a few Kelvins of the true value. In Chapter 4, the thermodynamic stabilities of icosahedral boron allotropes, its phase diagram were predicted with a SGP MLP trained using an on-the-fly active learning scheme. In Chapter 5, a neural equivariant interatomic potential was employed to tackle challenges associated with the variations of elemental configurations in a high-performance air electrode perovskite (i.e., PrNi1-δCoδO3) for protonic ceramic electrochemical cell (PCEC) applications. MLP-based phonon calculations suggest that the Ni/Co occupancy affects lattice thermal and chemical expansions differently, impacting a tradeoff between PCEC performance and stability. In Chapter 6, the neural equivariant MLP was used to study BaZr1-δYδO3 (BZY), a proton-conducting electrolyte used in PCEC assembly, paving the way to provide solutions to improve scalability of future PCEC manufacturing.