YOUNG Star detrending for Transiting Exoplanet Recovery (YOUNGSTER) – II. Using self-organizing maps to explore young star variability in sectors 1–13 of TESS data

Abstract

Young exoplanets and their corresponding host stars are fascinating laboratories for constraining the timescale of planetary evolution and planet-star interactions. However, because young stars are typically much more active than the older population, in order to discover more young exoplanets, greater knowledge of the wide array of young star variability is needed. Here Kohonen Self Organising Maps (SOMs) are used to explore young star variability present in the first year of observations from the Transiting Exoplanet Survey Satellite (TESS), with such knowledge valuable to perform targeted detrending of young stars in the future. This technique was found to be particularly effective at separating the signals of young eclipsing binaries and potential transiting objects from stellar variability, a list of which are provided in this paper. The effect of pre-training the Self-Organising Maps on known variability classes was tested, but found to be challenging without a significant training set from TESS. SOMs were also found to provide an intuitive and informative overview of leftover systematics in the TESS data, providing an important new way to characterise troublesome systematics in photometric data-sets. This paper represents the first stage of the wider YOUNGSTER program, which will use a machine-learning-based approach to classification and targeted detrending of young stars in order to improve the recovery of smaller young exoplanets.

Machine-learning Morphological Classification of TESS Light Curves of T Tauri Stars

“…More specifically, machine-learning techniques on TESS data have been used satisfactorily to study brightness variability associated with different physical processes. For example, automatic detection and classification of planets candidates (Pearson 2019;Yu et al 2019;Olmschenk et al 2021;Rao et al 2021;Battley et al 2022;Ofman et al 2022), stellar flares (Feinstein et al 2020;Vida et al 2021), episodic bright dimming (dippers; Tajiri et al 2020), and oscillating red giant stars (Hon et al 2021) have been performed using machine-learning in several available sectors of TESS with different cadences. Moreover, Claytor et al (2022) used a convolutional neural network and synthetic light curves modulated by stellar spots to infer stellar rotation periods automatically from a TESS sample.…”

Multiscale entropy analysis of astronomical time series

“…Kuszlewicz et al (2020) went into more detail and classified red giants according to their evolutionary states. Battley et al (2022) again used self-organizing maps to differentiate the TESS light curves of young eclipsing binaries and transiting objects from other types of variability. The TESS Data for Asteroseismology (T'DA) working group combined multiple separate classifiers into one large classifier to classify TESS lights curves accord-Article number, page 1 of 14 arXiv:2206.13529v1 [astro-ph.SR] 27 Jun 2022 ing to their high-level variability types (Audenaert et al 2021).…”

“…Most of the research has focused on using supervised learning to classify light curves (with some exceptions being, e.g., Valenzuela & Pichara 2018;Modak et al 2020) or on using unsupervised methods to create a latent space and then subsequently apply (or plan to apply) supervised methods to classify the data based on their positions in the latent space (see, e.g., Armstrong et al 2016;Battley et al 2022). Although supervised learning is ideal for structuring large amounts of data, it might be less efficient for smaller data sets where only lower amounts of labeled data are available.…”