Theses and Dissertations

ORCID

https://orcid.org/0009-0006-1810-5814

Advisor

Barrett, Christopher D.

Committee Member

Dickel, Doyl E.

Committee Member

Oppedal, Andrew L.

Date of Degree

5-16-2025

Original embargo terms

Immediate Worldwide Access

Document Type

Graduate Thesis - Open Access

Major

Mechanical Engineering

Degree Name

Master of Science (M.S.)

College

James Worth Bagley College of Engineering

Department

Michael W. Hall School of Mechanical Engineering

Abstract

Machine learning derived interatomic potentials have proven to be effective tools in simulation that mimic the accuracy of ab-initio calculations to the sub meV/atom level. They offer the advantage of computational speed operating at linear scaling with respect to the system’s size making them more efficient than classical potentials like MEAM (modified embedded atom methods). In this work, interatomic potentials are created based on the rapid artificial neural network (RANN) formalism for magnesium (Mg), Aluminum (Al) and their alloys. From previous works, the RANN formalism produces high-fidelity atomic models with accurate force fields for several metals, offering new insights into their physical properties. The RANN formalism is implemented with the LAMMPS MD package using a training database from DFT calculations, fingerprints inspired from modified embedded atom method formalism (MEAM) and angular screening to optimize atomic neighbor interactions and computational time. The generated RANN potential is a minimized single hidden layer architecture shown to replicate speeds of classical potentials, and its high accuracy is successfully validated with first-principle physical results of elastic, structural, bulk, defects, and thermal properties. This confirms the significance of these potentials for efficient large-scale molecular dynamics simulation.

Share

COinS