AI for Dating Stars: A Benchmarking Study for Gyrochronology

Andres Moya, Jarmi Recio-Martinez, Roberto J. Lopez-Sastre; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021, pp. 1971-1981

In astronomy, age is one of the most difficult stellar properties to measure, and gyrochronology is one of the most promising techniques for the task. It consists in dating stars using their rotational period and empirical linear relations with other observed stellar properties, such as stellar effective temperature, parallax, and/or photometric colors in different passbands, for instance. However, these approaches do not allow to reproduce all the observed data, resulting in potential significant deviations in age estimation. In this context, we propose to explore the stellar dating problem using gyrochronology from the AI perspective. Technically, we replace other linear combinations and traditional techniques with a machine learning regression approach. For doing so, we introduce a thorough benchmarking study of state-of-the-art AI regression models trained and tested for stellar dating using gyrochronology. Our experiments reveal promising results, where some models report a mean average error <0.5 Gyr, which can be considered as an outstanding breakthrough in the field. We also release a dataset and propose a set of simple assessment protocols to aid research on AI for dating stars as part of this study. Code and data to reproduce all our results are available at https://github.com/gramuah/ai4datingstars