Parameterizing Stellar Spectra Using Deep Neural Networks

TL;DR

This paper presents a deep neural network approach for stellar spectrum parameterization, achieving high accuracy in estimating stellar parameters from real and synthetic spectra.

Contribution

It introduces a novel DNN scheme initialized with autoencoders and fine-tuned with gradient descent for stellar parameter estimation.

Findings

Achieved mean absolute errors of 0.0048 dex for logT_eff on SDSS data

Achieved mean absolute errors of 0.1477 dex for logg on SDSS data

Achieved mean absolute errors of 0.1129 dex for [Fe/H] on SDSS data

Abstract

This work investigates the spectrum parameterization problem using deep neural networks (DNNs). The proposed scheme consists of the following procedures: first, the configuration of a DNN is initialized using a series of autoencoder neural networks; second, the DNN is fine-tuned using a gradient descent scheme; third, stellar parameters ($T_{eff}$, log$~g$, and [Fe/H]) are estimated using the obtained DNN. This scheme was evaluated on both real spectra from SDSS/SEGUE and synthetic spectra calculated from Kurucz's new opacity distribution function models. Test consistencies between our estimates and those provided by the spectroscopic parameter pipeline of SDSS show that the mean absolute errors (MAEs) are 0.0048, 0.1477, and 0.1129 dex for log$~T_{eff}$, log$~g$, and [Fe/H] (64.85 K for $T_{eff}$), respectively. For the synthetic spectra, the MAE test accuracies are 0.0011, 0.0182, and 0.0112 dex for log$~T_{eff}$, log$~g$, and [Fe/H] (14.90 K for $T_{eff}$), respectively.