Estimating $\left[ \alpha / \text{Fe} \right]$ from Gaia low-resolution BP/RP spectra using the ExtraTrees algorithm

TL;DR

This paper explores using machine learning, specifically the ExtraTrees algorithm, to estimate alpha element to iron abundance ratios from Gaia's low-resolution BP/RP spectra, enabling stellar property analysis at unprecedented scale.

Contribution

It demonstrates the feasibility of estimating [α/H] from low-resolution Gaia spectra using machine learning models trained on synthetic and observed data.

Findings

Models can estimate [α/H] for cool stars from low-resolution spectra.

Estimates rely on correlations with other stellar parameters.

Models are limited in distinguishing stars with similar [α/H] values.

Abstract

Gaia Data Release 3 will contain more than a billion sources with positions, parallaxes, and proper motions. In addition, for hundreds of millions of stars, it will include low-resolution blue photometer (BP) and red photometer (RP) spectra. Obtained by dispersing light with prisms, these spectra have resolutions that are too low to allow us to measure individual spectral lines and bands. However, the combined BP/RP spectra can be used to estimate some stellar properties such as $T_{\text{eff}}$, $\log{g}$, and $\left[ \text{M} / \text{H} \right]$. We investigate the feasibility of using the ExtraTrees algorithm to estimate the alpha element to iron abundance ratio $\left[ \alpha / \text{H} \right]$ from low-resolution BP/RP spectra. To infer $\left[ \alpha / \text{H} \right]$ from the spectra, we created regression models trained on two samples: a set of synthetic spectra and a set of observed spectra from stars that have known $\left[ \alpha / \text{H} \right]$ since they have been observed using HERMES. We applied each model to the other sample and to a larger observed sample to assess the performance of the models. In addition, we used our models to analyse stars from the Gaia-Enceladus structure. We find that models using the ExtraTrees algorithm can be used to estimate $\left[ \alpha / \text{H} \right]$ from low-resolution BP/RP spectra of cool stars. However, they do this by exploiting correlations between $\left[ \alpha / \text{H} \right]$ and other parameters, rather than the causal effect of $\left[ \alpha / \text{H} \right]$ on the spectrum. Hence, they are unlikely to be useful in studies that attempt to distinguish stars that only differ in $\left[ \alpha / \text{H} \right]$.