Joshua T. Roth

This work focuses on building reproducible machine-learning aided survey pipelines that can move from raw light curves to ranked planet candidates while keeping diagnostics inspectable. This allows the vetting of tens of millions of light curves in a reasonable amount of time, and the discovery of thousands of new planet candidates in TESS (spacecraft, shown in top plot) data. These candidates can be found here. The lower plot shows an example vetting plot for a (now confirmed) TESS exoplanet candidate identified by this search. It includes various diagnostic plots such as the folded light curve, periodogram, model fits, and image stacks.

KH 15D is a fascinating pre-main-sequence binary system with a circumbinary disk that occults the stars in a changing pattern, which led to this system being called the "winking star". I study the orbital evolution of this system and its spectroscopic signatures to understand the dynamics of young binaries and their circumstellar environments, and to study whether the winking star may have a companion. The plot shows radial-velocity measurements of KH 15D over time, along with model fits and residuals. The data shown here was collected over two observing campaigns, with over 20 years of separation between them, which is why each star has two fits, corresponding the two observing periods. The unique geometry of the system and the precession of the disk lead to complex radial-velocity signatures that need to be accounted for in order to accurately model the system's orbital parameters and dynamics.

While the T16 planet hunt is machine learning aided, it is not yet fully automated. I am exploring ways to combine multiple data products and machine learning algorithms like convolutional neural networks, transformers, and graph neural networks to improve classification and vetting of candidates. The goal is to create a fully automated vetting pipeline which can be applied to other surveys and datasets, while still providing interpretable diagnostics for human review.