A Robust Approach To Estimate Stellar Mass And Star Formation Rate Of Galaxies Using Empirical SED Templates And Machine Learning

On the cosmic time scale,stellar mass and star formation rate are an important basis to help understand the distribution and evolution of baryonic matter on the Galactic scale.For the study of galaxies and even the whole universe,it is very important to study star mass and star formation rate.For the estimation of star mass and star formation rate,astronomers have proposed a variety of calculation methods from different perspectives,such as using multi-band photometric data to fit the star family composition or using emission lines to estimate the number of massive stars that provide ultraviolet light required for ionized gas,to estimate star mass and star formation rate.However,the physical properties between different galaxies are extremely complex,so complex and degenerate parameters will be introduced when making relevant estimates.The main purpose of this dissertation is to propose a robust and simple method to estimate star mass and star formation rate.This dissertation proposes a method combining the empirical Galaxy template with machine learning.We selected three empirical spectral energy distribution templates with low redshift and low resolution proposed by Assef in 2010 as the reference for Galaxy SED decomposition.These include elliptical galaxy templates(representing the old star family),spiral galaxy templates(representing the middle-aged star family),and irregular galaxy templates(representing the young star family).The above template set has a longer wavelength range from the ultraviolet band to the infrared band(0.03 ?m to 30 ?m)than the previously used Galaxy template.Considering that the ultraviolet radiation of normal galaxies mainly comes from massive stars,and the estimation of star formation rate is dominated by ultraviolet data,optical and infrared photometric data are relatively important for the estimation of star mass.Therefore,if the sky survey data can cover the near/mid-infrared wavelength,it can help us obtain more reliable SED fitting results,to make the estimation of star mass and star formation rate more accurate.In this dissertation,I choose the normal galaxy sample source with photometric and spectral authentication in Sloan Digital Sky Survey as the experimental sample.Then the target galaxies can be divided into three age groups to obtain their proportions by using the astronomical multi-band sky survey data and Assef's empirical Galaxy spectral energy distribution template.We believe that the three-phased star families in the galaxy(i.e.old star family,middle-aged star family,and young star family)can represent the star evolution of the target galaxy.Therefore,the proportion of three age star families can be used to predict the star mass and star formation rate.We use the more efficient and easy to practice xgboost machine learning algorithm to predict and estimate the star mass and star formation rate.XGBoost algorithm is an algorithm based on the tree structure.Both classification prediction and regression prediction have excellent performance in computational efficiency and accuracy.We selected five features as the input of machine learning algorithm,including the proportion of old star family,the proportion of middle-aged star family,the proportion of young star family,redshift,and total luminosity.Using these five characteristics,we can predict and estimate the star mass and star formation rate respectively.It is found that our method can accurately predict the stellar mass with an uncertainty of only0.16 dex.For the prediction of star formation rate,although there is a strong linear correlation between the prediction results and the sample true value,the dispersion is larger than the stellar mass.The determination coefficient of regression prediction is only 0.69 and the uncertainty is 0.35 dex.At the same time,we also conducted a stability test.In the case of missing data,our method was not significantly affected.In Chapter 4,we also discuss the reason for the large dispersion of star formation rate.Considering the feature input of machine learning,we believe that the incomplete or missing input features will lead to the deviation of prediction results.However,when we add younger components or increase the proportion of irregular galaxies by the extinction of the template(i.e.increase the proportion of young star families),the prediction results of machine learning have not been significantly improved.This shows that the Assef template we selected has enough information about the young components of galaxies.Considering the calculation of star formation rate,we compare the calculation results between the common calculation methods of star formation rate.The results show that there is a large deviation between different calculation methods,which shows that there is large uncertainty in the calculation of star formation rate.We believe that this is the main reason for the large dispersion of machine learning predictions.Similarly,we also compare the calculation results of common different stellar mass calculation methods.For the stellar mass,there is little deviation between the results of different methods,which also proves the rationality of the results of our method.