2020-11-08, 13:10–13:10, Posters
This work presents a spectral fitting method based on machine learning.
Galaxy spectra at optical wavelength contain much information of the physical properties such as star-formation rate, stellar mass, metallicity and their history.
Conventional methods of spectral fitting are based on the least chi-square, Bayesian statistics, etc.
However, these methods require extremely long time and may not be appropriate for a future large spectroscopic survey.
In contrast, a machine-learning based method enables us to perform a spectral fitting much faster than conventional methods.
Our work investigates the feasibility of this method.
We constructed a neural network to decompose the input spectra into single stellar population (SSP) spectra.
We produced training and test data set using an SSP library (CB07) with various ages, velocity dispersions and dust extinction.
A 1.0 % Gaussian noise is added both to the training and test spectra in this study.
We found that our neural network accurately reproduces both the output physical parameters and spectra with extremely short time, that is, less than one second for 10000 spectra.
This result indicates that the spectral fitting based on the machine-learning is appropriate for a future large spectroscopic survey.