A multi-stage machine learning-based method to estimate wind parameters from Hα lines of massive stars

Abstract

This work presents a multi-stage method for estimating wind parameters in the domain of massive stars. We use the H$\alpha$ non-rotating synthetic spectral lines from the ISOSCELES database’s $\delta$-slow solutions to train a Gaussian Mixture Model-based cluster method and a deep neural network classifier. Then, the observed H$\alpha$ line profiles are deconvolved and classified into a class that provides a reduced subset of line profiles defined in ISOSCELES. This allows us to accurately and rapidly identify the closest line profile within the selected subset and obtain the wind parameters: $v_{\infty }$ and $\dot{M}$. Compared to traditional methods, this multi-stage proposal significantly reduces the computation time required to determine the wind parameters and gives more accurate and objective results. Interesting results of this work include evaluating the method for a sample of 12 B-supergiants, offering a notable improvement in the fitting of the line profiles, as it allows for a better approximation of the shape of the P Cygni lines for both components, absorption, and emission.