Stellar parameter estimation in O-type stars using artificial neural networks

This work presents the results of the implementation of a deep learning system capable of estimating the effective temperature and surface gravity of O-type stars. The proposed system was trained with a database of 5,557 synthetic spectra computed with the stellar atmosphere code CMFGEN that covers stars with $T_{\mathrm{eff}}$ from $\sim$20,000 K to $\sim$58,000 K, $log(L/L_{\odot })$ from 4.3 to 6.3 dex, log $g$ from 2.4 to 4.2 dex, and mass from 9 to 120 $M_{\odot }$. Important advantages proposed in this paper include using a set of equivalent width measurements over the optical region of the stellar spectra, which avoids processing the full spectra with the inherent computational cost and allows it to apply the same trained system over different spectra resolutions. The validation of the system was performed by processing a sample of twenty O-type stars taken from the IACOB database, and a subgroup of eleven stars of those twenty taken from The Galactic O-Star Spectroscopic Catalog (GOSC) with lower resolution. As complementary work, we show the results of a synthetic spectra fitting process with the aim of simplifying the comparison with other estimations and parameter fitting from the literature.