Recovery of TESS Stellar Rotation Periods with Convolutional Neural Networks

A poster for the Cool Stars 20.5 virtual conference.

Abstract:

Complete pictures of stellar structure, magnetism, activity, and evolution are not possible without understanding rotation. While the Kepler mission revolutionized stellar astrophysics with tens of thousands of rotation period estimates, TESS is poised to increase the number of detected periods by an order of magnitude. However, systematics of the TESS mission have complicated period searches. While there are several efforts to solve this problem by removing systematics, standard methods of data reduction have shown little success. We present a method to predict rotation periods from TESS full-frame image light curves using a convolutional neural network. This method relies on a set of simulated light curves convolved with TESS galaxy light curves to emulate the instrumental noise observed in stellar signals. The simulations include a full treatment of latitudinal differential rotation, star spot evolution, and activity level to make the light curves as realistic as possible. Our approach allows the network to learn the difference between rotation signals and TESS systematics. With these tools, we can reliably estimate periods as well as a credence metric to determine the regions of parameter space the network predictions are most reliable. Using the most credible half of the test data, we predict 60% of simulation periods to within 10% accuracy, on par with some of the methods used to estimate periods from Kepler data.