Photometric [Fe/H] of RRab stars in the $G$ and $K_s$ bands by deep learning

TL;DR

This paper introduces a deep learning method to accurately estimate the metallicity of RR Lyrae stars across multiple photometric bands, enabling large-scale mapping of stellar metallicity distributions.

Contribution

A novel deep learning approach using recurrent neural networks to transfer spectroscopic metallicity calibration across photometric bands for RR Lyrae stars.

Findings

Achieved a mean absolute error of 0.1 dex in metallicity estimation.

Produced a catalog of over 60,000 RR Lyrae stars with photometric metallicities.

Mapped the metallicity distribution of RR Lyrae stars across the sky.

Abstract

RR Lyrae stars are useful chemical tracers thanks to the empirical relationship between their heavy-element abundance and the shape of their light curves. However, the consistent and accurate calibration of this relation across multiple photometric wavebands has been lacking. We have devised a new method for the metallicity estimation of fundamental-mode RR Lyrae stars in the Gaia optical $G$ and near-infrared VISTA $K_s$ wavebands by deep learning. First, an existing metallicity prediction method is applied to large photometric data sets, which are then used to train long short-term memory recurrent neural networks for the regression of the [Fe/H] to the light curves in other wavebands. This approach allows an unbiased transfer of our accurate, spectroscopically calibrated $I$-band formula to additional bands at the expense of minimal additional noise. We achieve a low mean absolute error of $0.1$ dex and high $R^2$ regression performance of $0.84$ and $0.93$ for the $K_s$ and $G$ bands, respectively, measured by cross-validation. The resulting predictive models are deployed on the Gaia DR2 and VVV inner-bulge RR Lyrae catalogs. We estimate mean metallicities of $-1.35$ dex for the inner bulge and $-1.7$ for the halo, which are significantly less than values obtained by earlier photometric prediction methods. Using our results, we establish a public catalog of photometric metallicities of over 60,000 Galactic RR Lyrae stars, and provide an all-sky map of the resulting RR Lyrae metallicity distribution. The software code used for training and deploying our recurrent neural networks is made publicly available in the open-source domain.