Machine learning applications in exoplanet host star recommendation and determination

Abstract

In the past three decades, parallel to the detection and characterisation of extrasolar planets, the link between a planet and its host star has been explored extensively. The presence of planetary companions has been linked to the star’s chemical, physical and galactic properties, such that particular configurations may potentially lead to a greater likelihood of formation. If these correlations are truly astrophysical, these characteristics can be leveraged to recommend potential host stars based solely on their stellar properties. This work aims to use the power of ML to develop, train and test recommendation engines capable of discriminating between the stellar host and comparison star samples. The project explores two approaches with different applications. The first is a spectral classification model, which provides a score for the likelihood of a star hosting a giant planet based on its spectrum. An exploratory phase extensively validates the use of machine vision and the architecture used for this application. This leads to the development of two upgraded designs to enhance performance and overall stability. The project establishes this method as a credible tool which can be implemented in spectrograph pipelines to recommend follow-up observations. The second application is to develop a tool to aid in analysing incomplete stellar abundance data in differentiating between hosts and comparison stars by leveraging supervised and unsupervised learning, as well as MOO, to allow for reliable imputation. The complete dataset could then be applied to comparison tests and host star recommendation strategies. This project presents the first implementation of the MOO algorithm, constraining its usage for binary classification of giant-planet hosts and multi-label classification for giant, low-mass and multiple-planet hosts by evaluating its performance on a stellar abundance catalogue. Performance in both approaches is then analysed to direct future work.