UC Berkeley

There is immense desire to discover new materials to address many modern societal challenges—think of stronger alloys or more efficient solar panels. Historically, new materials have been discovered purely by physical experimentation. However, with the advances of physical theories, alongside modern computing hardware and software, computational techniques for predicting the properties of materials have become widespread. These computational techniques seek to complement physical experiments; as physical experiments are costly and time-consuming, one hopes to guide such experiments toward promising materials predicted by simulations like those presented here.My work focuses on expanding the capabilities of atomistic calculations and improving their efficiency. Atomistic simulations allow us to model fundamental properties of materials. Density functional theory (DFT) is a common choice, enabling nearly unparameterized modelling of materials from quantum mechanical theory. However, DFT simulations are computationally expensive and are typically limited to materials with up to hundreds of atoms and hundreds of picoseconds. Hence, my dissertation work seeks to expand the capabilities of such simulations to larger length- and time-scales than were previously possible.The modelling of random solid solutions (e.g. many alloys) is challenging due to their lack of translational symmetry. Chapter 2 explores the Small Set of Ordered Structures (SSOS) approach to modelling the thermodynamic and mechanical properties of high entropy alloys. The SSOS approach is an approximation wherein one can use many small structure simulations to approximate the properties of a solid solution by matching weighted correlation functions. We demonstrate how binary and ternary cells can be used to model the properties of two example quinary alloys.More recently, Machine Learning Interatomic Potentials (MLIPs) have become increasingly common for such atomistic simulations. Typically based on graph neural networks, MLIPs are trained on many DFT calculations and can capture the DFT-predicted potential energy surface at linear scaling. Such MLIPs are most often used to predict properties that can be derived from energies, forces, and stresses (e.g. relative stability, reaction barriers, elastic properties, diffusivity). Chapter 3 describes the creation of a new dataset (MP-ALOE) for "Universal" MLIPs (UMLIPs), which include data spanning most elements on the periodic table (in this case, 89 elements). The presented MP-ALOE dataset differs from previously published datasets in two distinct ways. First, it is calculated to include a large proportion of off-equilibrium structures to increase generalizability. Second, it is calculated using the accurate r2SCAN functional, which makes MP-ALOE more accurate than most other bulk crystalline datasets relative to experiment. Benchmark calculations of a UMLIP trained on this new MP-ALOE dataset show improved performance under high temperatures and pressures, implying increased stability and reliability.These projects both demonstrate advances in data-driven methods for predicting material properties via atomistic simulation.