200,000+ Deep Learning-inferred Periods of Stellar Variability from the All-Sky Automated Survey for Supernovae

Abstract

Stars exhibit a range of variability periods that depend on their mass, age, and evolutionary stage. For space-based photometric data, convolutional neural networks (CNNs) have demonstrated success in recovering and measuring periodic variability from photometric missions like Kepler and TESS. All-sky ground-based surveys can have similar if not longer baselines than space-based missions; however, these datasets are more challenging to work with due to irregular sampling, more complex systematics, and larger data gaps. In this work, we demonstrate that CNNs can be used to derive variability periods from ground-based surveys. From the All-Sky Automated Survey for Supernovae, we recover 208,260 variability periods between 1 and 30 days, approximately 60% of which are new detections. We recover periods for active RSCVn, anomalous sub-subgiants, and cool dwarfs that are consistent with previously measured rotation periods, while periods for stars above the Kraft break are generally spurious. We also identify periodic signals in tens of thousands of giant stars that correspond to frequencies of stellar oscillations rather than rotation. Our results highlight that CNNs can be used on sparsely sampled ground-based photometry to recover periodicity. We conclude that the findings of our work are very promising for the potential recovery of hundreds of thousands of stellar rotation periods in data from the Vera C. Rubin Observatory's Legacy Survey of Space and Time and the Nancy Grace Roman Space Telescopes Galactic Bulge Time Domain Survey.