Machine learning estimation of physical properties of S0 Galaxies from their optical spectra

Abstract

In this work, we show that it is possible to infer precise information about some of the main physical properties of lenticular galaxies from the shape of their entire optical spectrum. We study this methodology as an alternative to the more conventional way of individually analyzing the most important emission and/or absorption lines in this frequency band. By using neural networks trained with high signal-to-noise spectra ranging from 400 nm to 800 nm, we have determined the accuracy of the predictions for the following interesting properties: the equivalent width (EW) of the emission lines Hα, Hβ, [O III] and [NII]; the D4000 break, the specific star formation rate, sSFR, and the stellar mass to light ratio in the SDSS r-band, M∗/Lr. We provide a comparison of the performance of this method using as input, on the one hand, all the dimensionality available in the spectra and, on the other hand, only their first principal components (PC). We conclude that the latter procedure produces better results when predicting the selected variables. We have also inferred that 5 is the ideal number of PCs to compute the values of these variables and identified the most dominant ones to determine which and how many eigenspectra are required for a minimal optimal prediction. Finally, we have tested the performance of our methodology as a WHAN activity classifier, also obtaining encouraging results