An autoencoder of stellar spectra and its application in automatically estimating atmospheric parameters

TL;DR

This paper presents an autoencoder-based method for extracting features from stellar spectra to automatically estimate key atmospheric parameters with high accuracy, demonstrated on real and synthetic data.

Contribution

It introduces a novel spectral feature extraction scheme using autoencoders combined with convolution, pooling, and BP networks for stellar parameter estimation.

Findings

Achieved low mean absolute errors on real spectra

Performed accurate parameter estimation on synthetic spectra

Demonstrated effectiveness of autoencoder-based feature extraction

Abstract

This article investigates the problem of estimating stellar atmospheric parameters from spectra. Feature extraction is a key procedure in estimating stellar parameters automatically. We propose a scheme for spectral feature extraction and atmospheric parameter estimation using the following three procedures: firstly, learn a set of basic structure elements (BSE) from stellar spectra using an autoencoder; secondly, extract representative features from stellar spectra based on the learned BSEs through some procedures of convolution and pooling; thirdly, estimate stellar parameters ($T_{eff}$, log$~g$, [Fe/H]) using a back-propagation (BP) network. The proposed scheme has been evaluated on both real spectra from Sloan Digital Sky Survey (SDSS)/Sloan Extension for Galactic Understanding and Exploration (SEGUE) and synthetic spectra calculated from Kurucz's new opacity distribution function (NEWODF) models. The best mean absolute errors (MAEs) are 0.0060 dex for log$~T_{eff}$, 0.1978 dex for log$~g$ and 0.1770 dex for [Fe/H] for the real spectra and 0.0004 dex for log$~T_{eff}$, 0.0145 dex for log$~g$ and 0.0070 dex for [Fe/H] for the synthetic spectra.