Survey of Surveys. II. Stellar parameters for 23 millions of stars

TL;DR

This paper compiles stellar parameters for 23 million stars from multiple surveys, using machine learning to improve accuracy, especially for metal-poor stars, and provides a publicly accessible catalog.

Contribution

It introduces a new multi-survey catalog of stellar parameters and a machine learning approach that enhances precision and accuracy over previous methods.

Findings

Significant improvement in parameter estimates for metal-poor stars.

Validation shows better accuracy compared to other machine learning catalogs.

Public release of a comprehensive stellar parameter catalog for 23 million stars.

Abstract

In the current panorama of large surveys, the vast amount of data obtained with different methods, data types, formats, and stellar samples, is making an efficient use of the available information difficult. The Survey of Surveys is a project to critically compile survey results in a single catalogue, facilitating the scientific use of the available information. In this second release, we present two new catalogs of stellar parameters (Teff, logg, and [Fe/H]). To build the first catalog, SoS-Spectro, we calibrated internally and externally stellar parameters from five spectroscopic surveys (APOGEE, GALAH, Gaia-ESO, RAVE, and LAMOST) and externally on the PASTEL database. The second catalog, SoS-ML catalog, is obtained by using SoS-Spectro as a reference to train a multi-layer perceptron, which predicts stellar parameters based on two photometric surveys, SDSS and SkyMapper. As a novel approach, we build on previous parameters sets, from Gaia DR3 and Andrae et al. (2023), aiming to improve their precision and accuracy. We obtain a catalog of stellar parameters for around 23 millions of stars, which we make publicly available. We validate our results with several comparisons with other machine learning catalogs, stellar clusters, and astroseismic samples. We find substantial improvements in the parameters estimates compared to other Machine Learning methods in terms of precision and accuracy, especially in the metal-poor range, as shown in particular when validating our results with globular clusters. We believe that there are two reasons behind our improved results at the low-metallicity end: first, our use of a reference catalog, the SoS-Spectro, which is calibrated using high-resolution spectroscopic data; and second, our choice to build on pre-existing parameter estimates from em Gaia and Andrae et al., rather than attempting to obtain our predictions from survey data alone.