Abstract

This dissertation details the generation of a kilonova simulation training library and the training of subsequent machine learning emulators used for multidimensional interpolation. We begin with a Gaussian process interpolation methodology and apply that emulator for Bayesian inference with the AT 2017gfo light-curve dataset. We then expand on our emulator interpolation methodology, replacing the scalar Gaussian process emulator with a vector neural network emulator for light-curve generation. Our neural network emulator results are localized to a much more narrow region of the parameter space, prompting a detailed supportive investigation of the systematic uncertainty associated with our inference approach. We then broaden our initial examination of only electromagnetic observations to include gravitational-wave observations which inform us about the progenitor binary parameters. These binary parameters, when used in conjunction with fits to general-relativistic simulations of neutron star mergers, serve as a separate channel for predicting ejecta parameters such as mass and velocity. We examine the inferred ejecta predictions when using only electromagnetic observations, only gravitational wave observations, and joint electromagnetic and gravitational wave data. Our analysis shifts gears slightly, focusing our inference on examining which nuclear mass and fission models produce elemental abundances most closely aligned with the abundance patterns observed in nature. We use the most closely-matching abundance pattern in conjunction with our simulated abundances to identify a preferred ratio of wind-to-dynamical ejecta, thus introducing an additional observational prior for our standard light-curve inference. Finally, we conclude with a return to exploring machine learning emulators, this time focusing on a random forest emulator for kilonova spectra. Expanding on model limitations identified using light-curve inferences, we propose an additional kilonova component to rectify an underluminosity in our model, finding that the proposed solution in fact exacerbates the issue. The dissertation concludes with a discussion about future work, with an appendix which briefly highlights relevant unpublished results.

Publication Date

8-2024

Document Type

Dissertation

Student Type

Graduate

Degree Name

Astrophysical Sciences and Technology (Ph.D.)

Department, Program, or Center

Physics and Astronomy, School of

College

College of Science

Advisor

Richard O'Shaughnessy

Advisor/Committee Member

Nathan Cahill

Advisor/Committee Member

Joshua Faber

Campus

RIT – Main Campus

Download

Share

COinS