Stellar Populations With Optical Spectra: Deep Learning vs. Popular Spectrum Fitting Codes

TL;DR

This study compares traditional spectrum fitting codes and a deep learning model for recovering stellar population properties from simulated galaxy spectra, highlighting the superior accuracy and speed of the deep learning approach when properly trained.

Contribution

It demonstrates that a convolutional neural network significantly outperforms traditional codes in speed and accuracy for stellar population analysis, emphasizing the importance of training data quality.

Findings

StarNet outperforms all traditional codes in speed and accuracy.

pPXF is the fastest among non-ML codes and best at recovering properties.

Errors depend on true values, S/N ratio, and mock spectrum construction methods.

Abstract

We compare the performance of several popular spectrum fitting codes (Firefly, starlight, pyPipe3D and pPXF), and a deep-learning convolutional neural network (StarNet), in recovering known stellar population properties (mean stellar age, stellar metallicity, stellar mass-to-light ratio M*/L_r and the internal E(B-V)) of simulated galaxy spectra in optical wavelengths. Our mock spectra are constructed from star-formation histories from the IllustrisTNG100-1 simulation. These spectra mimic the Sloan Digital Sky Survey (SDSS) through a novel method of including the noise, sky residuals and emission lines taken directly from SDSS. We find that StarNet vastly outperforms all conventional codes in both speed and recovery of stellar population properties (error scatter < 0.08 dex, average biases < 0.02 dex for all tested quantities), but it requires an appropriate training set. Of the non-machine-learning codes, pPXF was a factor of 3-4 times faster than the other codes, and was the best in recovering stellar population properties (error scatter of < 0.11 dex, average biases < 0.08 dex). However, the errors and biases are strongly dependent on both true and predicted values of stellar age and metallicity, and signal-to-noise ratio. The biases of all codes can approach 0.15 dex in stellar ages, metallicities and log M*/L_r , but remain < 0.05 for E(B-V). Using unrealistic Gaussian noise in the construction of mock spectra will underestimate the errors in the metallicities by a factor of two or more, and mocks without emission lines will underestimate the errors in stellar age and M*/L_r by a factor of two.