Barrett, Christopher D.
Dickel, Doyl E
Kadiri, Haitham El
Date of Degree
Graduate Thesis - Open Access
Doctor of Philosophy (Ph.D)
James Worth Bagley College of Engineering
Department of Mechanical Engineering
Machine learning techniques using artificial neural networks (ANNs) have proven to be effective tools to rapidly mimic first principles calculations. These tools are capable of sub meV/atom accuracy while operating with linear scaling with respect to the system size. Here novel interatomic potentials are constructed based on the rapid artificial neural network (RANN) formalism. This approach generates precise force fields for various metals that have historically been difficult to describe at the atomic scale. These force fields can be utilized in molecular dynamics simulations to provide new physical insights. The RANN formalism, which is incorporated into a LAMMPS molecular dynamics package, utilizes fingerprints inspired by the modified embedded atom method (MEAM) formalism and angular screening which enables shorter neighbor lists and faster computations. It has been shown that this implementation can replicate speeds comparable to traditional models while maintaining high agreement (~1meV/atom) with DFT. This formalism has been used to predict correct slip modes in Mg and successfully model the challenging structure of zinc for the first time. Also RANN potentials for titanium and zirconium accurately predict the phase diagrams and triple points with high accuracy as computed by relative free energy calculation. New Ti and Zr potential successfully predict the dislocation core structures and slip planes for high pressure phase of these materials. The formalism's precision and transferability enable the construction of a binary Ti-Al system with DFT accuracy at MD speed. Due to the RANN's great fidelity to DFT data and predictive capability, these potentials might be helpful in the future for investigating behavior and interaction in large-scale atomistic simulation.
Nitol, Mashroor Shafat, "Predictive computational materials Modeling with machine learning: creating the next generation of atomistic potential using neural networks" (2021). Theses and Dissertations. 5373.