Date of Award

Summer 8-20-2021

Level of Access Assigned by Author

Open-Access Thesis

Degree Name

Doctor of Philosophy (PhD)

Department

Physics

Advisor

David Batuski

Second Committee Member

Dale Kocevski

Third Committee Member

Neil Comins

Additional Committee Members

Andre Khalil

Liping Yu

Abstract

Galaxy mergers are dynamic systems that offer us a glimpse into the evolution of the cosmos and the galaxies that constitute it. However, with the advent of large astronomical surveys, it is becoming increasingly difficult to rely on humans to classify the vast number of astronomical images collected every year and find the images that capture these systems. In recent years, researchers have increasingly relied on machine learning and computer vision classifiers, and while these techniques have proven useful for classifying broad galaxy morphologies, they have struggled to identify galaxy mergers. A random forest classifier was applied to a subset of galaxies from the Cosmic Assembly Near-infrared Extragalactic Legacy Survey (CANDELS) to classify merger and non-merger events. 283 merging and 283 non-merging galaxies were selected from the five CANDELS fields, totaling a combined 566 galaxies for training and validation. The classifier was trained on a set of parameters measured for each galaxy, including mass, star formation rate, galactic half-light radius, as well as Concentration and Asymmetry measurements. The classifier performed with a mean accuracy of 92.31% and a precision of 0.9332 on the validation dataset. Additionally, a computer vision convolutional neural network was trained to analyze and classify images of merger and non-merger events in the same fields. Due to the small number of merger events present in the CANDELS fields, data augmentation was utilized to increase the dataset significantly and boost performance. The computer vision classifier performed with an accuracy of 87.87% and a precision of 0.8683 on validation data. The pre-trained convolutional neural network was then used to predicted classes for a dataset containing active galactic nuclei (AGN) hosting galaxies and a control sample, although no correlation was found between predicted classes and whether the galaxy hosts an AGN.

Included in

Physics Commons

Share