Denoising medium resolution stellar spectra with neural networks

TL;DR

This paper demonstrates that neural network autoencoders can effectively denoise medium resolution stellar spectra, improving analysis speed and accuracy for astronomical observations.

Contribution

It introduces neural network denoisers capable of reconstructing stellar spectra with high accuracy across varying conditions, enhancing spectral analysis methods.

Findings

Achieved less than 1% relative error in spectrum reconstruction at SNR of 10.

Networks adapt to different extinction, fluxing, and Doppler shifts.

Denoised spectra facilitate faster initial parameter estimation.

Abstract

We trained denoiser autoencoding neural networks on medium resolution simulated optical spectra of late-type stars to demonstrate that the reconstruction of the original flux is possible at a typical relative error of a fraction of a percent down to a typical signal-to-noise ratio of 10 per pixel. We show that relatively simple networks are capable of learning the characteristics of stellar spectra while still flexible enough to adapt to different values of extinction and fluxing imperfections that modifies the overall shape of the continuum, as well as to different values of Doppler shift. Denoised spectra can be used to find initial values for traditional stellar template fitting algorithms and - since evaluation of pre-trained neural networks is significantly faster than traditional template fitting - denoiser networks can be useful when a fast analysis of the noisy spectrum is necessary, for example during observations, between individual exposures.