Thesis Committee: Ian Crossfield (Chair), Sara Seager, Saul Rappaport and Nevin Weinberg

Abstract:
Giant planets play a fundamental role in shaping the architecture of planetary systems and the formation and evolution of smaller planets. Over the last few decades, over 1000 giant planets have been discovered outside the Solar System, yet many open questions remain about their formation and evolution histories. How do close-in giant planets — the so-called “hot Jupiters” — reach orbits as short as ~0.01 AU, where they could not have accreted their gaseous envelopes? How do many hot Jupiters attain radii larger than predicted by standard models of planetary structure? How do giant planets form, and what determines their final masses? To answer these questions, we need to amass a large and diverse population of giant planets that will allow us to uncover their evolution histories through both case studies and statistical analyses.

In this thesis, I focus on the discovery and characterization of unusual systems that may shed light on the pasts of giant planets. In particular, I present detailed analyses of systems that have not been subject to the overwhelming tidal forces capable of erasing many traces of orbital evolution. These include young systems that may still be actively undergoing planetary migration, giant planets with wider orbits than most hot Jupiters, and planets orbiting predominantly radiative stars, which exert weak tidal forces. Some of these newly discovered planets have spent their entire lifetimes near the stellar irradiation threshold at which giant planets become larger than expected, and are valuable in constraining planet inflation models. Some are also favorable targets for transit spectroscopy to study the atmospheres and chemical composition of giant planets.

I also develop methods and tools to further expand our collection of known giant planets using data from the K2 and TESS space missions. After demonstrating the traditional human vetting approach to planet candidate identification, I present two convolutional neural networks, AstroNet-Triage and AstroNet-Vetting, capable of automatically performing triage and vetting on TESS light curves. These are the first machine learning-based classifiers to be trained and tested on realTESS data, and can rapidly and accurately eliminate false positives. These models not only allow humans to focus on the strongest planet candidates instead of false positives, but also identify candidates in an unbiased, homogeneous manner so as to facilitate occurrence rate calculations.