Bulk classification and analysis of TESS γ Doradus stars using machine learning methods

Supervised machine learning was used to classify γ Doradus stars present in the TESS-SPOC data pipeline, to investigate the efficacy of fast bulk classification of pulsating stars in minimally-processed data. A fully-connected neural network was set up and trained as a binary classifier using built catalogues of previously confirmed pulsators of four types present in the γ Doradus instability strip, as well as known non-variable stars. During validation, the model obtained a 94.4% precision score. The trained network was then input with binned Lomb-Scargle periodograms of 173,398 stars within the ranges Teff = 6500 - 7500 K and Tmag = 9.0 - 12.0. The total time for the network to classify all candidates was 11.1 minutes, with a pre-processing time of ∼5 ms per lightcurve. The probability distributions and HR diagram positions of the output classifications were analysed and a small set of the results visually verified. It was found that a classifier confidence threshold of 77.4% was most suitable and yielded 7,749 potential γ Doradus candidates, representing 4.47% of the analysed set. Of 100 of these visually checked, only seven were misclassified EB stars, and three likely rotational variables. Eight of the classifications showed evidence of p-mode pulsations suggestive of γ Doradus and δ Scuti hybrids. This investigation shows a way in which only minimal treatment of TESS lightcurve data is necessary for high quality classifications of pulsators, allowing for quick identification in large datasets. This is important, as a large and diverse pool of candidates is necessary for thorough investigation and testing of stellar evolution models.