Youth analysis of near infrared spectra of young low-mass stars and brown dwarfs

TL;DR

This study develops a machine learning-based method using near-infrared spectra to accurately identify young low-mass stars and brown dwarfs, focusing on gravity-sensitive features and spectral indices.

Contribution

It introduces new spectral indices and demonstrates that machine learning on NIR spectra effectively distinguishes young objects from field stars.

Findings

Achieved spectral typing precision below 1 subtype

Developed a new gravity-sensitive index (TLI-g) with superior performance

H-band shape and specific absorption features are key for gravity classification

Abstract

We aim at building a method that efficiently identifies young low-mass stars and brown dwarfs from low-resolution near-infrared spectra, by studying gravity-sensitive features and their evolution with age. We built a dataset composed of all publicly available ($\sim$2800) near-infrared spectra of dwarfs with spectral types between M0 and L3. First, we investigate methods for the derivation of the spectral type and extinction using comparison to spectral templates, and various spectral indices. Then, we examine gravity-sensitive spectral indices and apply machine learning methods, in order to efficiently separate young ($\lesssim$10 Myr) objects from the field. Using a set of six spectral indices for spectral typing, including two newly defined ones (TLI-J and TLI-K), we are able to achieve a precision below 1 spectral subtype across the entire spectral type range. We define a new gravity-sensitive spectral index (TLI-g) that consistently separates young from field objects, showing a performance superior to other indices from the literature. Even better separation between the two classes can be achieved through machine learning methods which use the entire NIR spectra as an input. Moreover, we show that the H- and K-bands alone are enough for this purpose. Finally, we evaluate the relative importance of different spectral regions for gravity classification as returned by the machine learning models. We find that the H-band broad-band shape is the most relevant feature, followed by the FeH absorption bands at 1.2 $\mu m$ and 1.24 $\mu m$ and the KI doublet at 1.24 $\mu m$.