Application of Deep Learning in Analysis of Stellar Spectra

Abstract

In recent years, the analysis of high-resolution stellar spectra has become increasingly important for estimating key stellar parameters such as effective temperature (Teff ), surface gravity (log g), metallicity ([M/H]), and rotational velocity (v sin i). Traditional methods often rely on manual calibration or spectrum synthesis, which can be time-consuming and error-prone, especially for M dwarfs whose spectra are dense with molecular features. In this study, we investigate the use of convolutional neural networks (CNNs) to automate the estimation of stellar parameters using synthetic and observed data.We adopt a StarNet-like CNN architecture trained on synthetic spectra generated from the PHOENIX-ACES model grid, and evaluate its performance on real observations from the CARMENES survey. Separate models are developed for each parameter, allowing for dedicated tuning and improved prediction accuracy. The training process includes flux normalization and data augmentation across multiple spectral windows. To assess performance, we compare predicted parameters with literature values and find strong agreement, especially for Teff and log g, with significantly reduced mean squared error.This approach demonstrates the effectiveness of deep learning in spectroscopic analysis, offering a scalable solution for stellar parameter estimation. The results reinforce the potential of CNNs to support large-scale stellar surveys and contribute to more accurate stellar characterization.