Recovery of TESS Stellar Rotation Periods Using Deep Learning

TL;DR

This paper presents a deep learning approach using convolutional neural networks to accurately infer stellar rotation periods from TESS light curves, overcoming traditional limitations and systematics.

Contribution

The authors develop a neural network trained on synthetic data convolved with real TESS light curves, enabling robust period estimation and uncertainty prediction beyond previous methods.

Findings

Achieved 10%-accurate periods for 46% of targets

Recovered periods of real stars with literature measurements

Demonstrated resistance to half-period aliases

Abstract

We used a convolutional neural network to infer stellar rotation periods from a set of synthetic light curves simulated with realistic spot evolution patterns. We convolved these simulated light curves with real TESS light curves containing minimal intrinsic astrophysical variability to allow the network to learn TESS systematics and estimate rotation periods despite them. In addition to periods, we predict uncertainties via heteroskedastic regression to estimate the credibility of the period predictions. In the most credible half of the test data, we recover 10%-accurate periods for 46% of the targets, and 20%-accurate periods for 69% of the targets. Using our trained network, we successfully recover periods of real stars with literature rotation measurements, even past the 13.7-day limit generally encountered by TESS rotation searches using conventional period-finding techniques. Our method also demonstrates resistance to half-period aliases. We present the neural network and simulated training data, and introduce the software butterpy used to synthesize the light curves using realistic star spot evolution.