Date of Award

2022-05-01

Degree Name

Master of Science

Department

Physics

Advisor(s)

Jorge A. Munoz

Abstract

First principles-based simulations have allowed us to explore emerging phenomena in a variety of systems. Its steadfast practicality has led to an increase in molecular and materials design ranging from drug discovery to planetary formation. However ubiquitous in its field, one of its biggest drawbacks is its computational cost, notably so in molecular dynamics simulations. To counter this setback, there have been many leading efforts in machine learning methods, whether it be in algorithms or network architectures. Our contribution uses an active learning algorithm paired with a tensor field network, e3nn. By steadily feeding new data points to our model, based on data it struggles with, our model gains a better idea of the system. First, using Born-Oppenheimer molecular dynamics, we simulate body-centered cubic zirconium with a (4 x 4 x 4) supercell of 128 atoms at 1500K for 800 time steps. This DFT model is subsequently fed into the network where we define energies as our input and forces as our output. The active learning runs that follow reshuffle the top 5% errors of the test set into the train set. Several runs later, we analyze the accuracy of our trained model by generating phonon dispersions and comparing their structure to that of the original DFT data.

Language

en

Provenance

Received from ProQuest

File Size

75 p.

File Format

application/pdf

Rights Holder

Vanessa Judith Meraz

Share

COinS