Method Study For Searching Barium Giants From LAMOST Spectra Based On Machine Learning

Barium giants are G-K type chemically peculiar stars which exhibit enhancement of slow neutron-capture elements,especially at Sr Ⅱ 4077 A and Ba Ⅱ 4554 (?).Chemical abundances analysis of barium stars can provide observational constraints for models of neutron-capture nucleosynthesis of asympotic giant branch stars and crucial clues for the study of chemical evolution of the Galaxy.However,the number of barium stars which have been confirmed is far from enough for further research.In recent years,with the increasing quantity and quality of spectra provided by large astronomical observation equipment,and the development of machine-learning,data supporting and new methods are provided for the classification and chemical abundance analysis of barium giantsIn this thesis,we propose five machine-learning algorithms to search for barium stars from low-resolution spectra of LAMOST(The Large Sky Area Multi-Object Fiber Spectroscopic Telescope).Two classifiers based on different features are constructed for the determination of barium stars with strontium enhancement and those with barium enhancement.The Light Gradient Boosting Machine(LightGBM)algorithm with features in the whole spectrum performs best,for the sample with strontium enhancement,precision is 98.98%and recall is 99.53%;for the sample with barium enhancement,precision is 99.92%and recall is 99.84%.In the estimation of chemical abundance for barium giants,the predictor based on LightGBM algorithm also achieves good accuracy in the prediction of[Sr/Fe]and[Ba/Fe].The first contribution of this paper is to apply LightGBM algorithm to classify barium stars and estimate the chemical abundance of them,verify the performance of different machine-learning algorithms in different feature bands,and finally build a reliable classification model and abundance prediction model.The work and innovations of this paper mainly include the following four parts:(1)Sample data obtained and data preprocessingThe 6543 low-resolution G-K type stellar spectra from LAMOST DR8 are used in the experiment.Barium star data come from Li et al.’s work(810 spectra)and the star list provided by Norfolk et al.(867 spectra).Stellar atmosphere parameters were provided by ULySS package calculation,and data are uniformly normalized and preprocessed with reducing the influence of radial velocity.(2)Feature selectionSr Ⅱ 4077 (?) and 4215 (?) are used for Sr classification model,and Ba Ⅱ 4554 (?) 4934 (?)and 6496 A are used for Ba classification model.Then,the model effects of 20 (?) feature band,70 A feature band and full spectrum as feature input are compared.(3)Classification model constructionWe choose SVM,a machine learning algorithm suitable for small samples,and LightGBM,which has strong generalization ability,to train two classification models respectively:Sr enhanced classification model and Ba enhanced classification model.Linear interpolation method is used to amplify the number of positive samples in the training set,which is also one of the innovations in this paper.In addition,the performance of KNN,SVM,RF,XGBoost and LightGBM algorithms in barium giant classification is compared.(4)Prediction model constructionTwo prediction models are constructed based on LightGBM algorithm to estimate the[Sr/Fe]and[Ba/Fe]of barium giant samples,respectively.We find that[Ba/Fe]estimated from Ba Ⅱ line at 4554 (?) has smaller dispersion than that from Ba Ⅱ line at 4934 (?),MAE4554(?)=0.07,σ4554(?)=0.12.[Sr/Fe]estimated from Sr Ⅱ line at 4077 (?) performs better than that from Sr Ⅱ line at 4215 (?),MAE4077(?)=0.09,σ4077(?)=0.16.This work is a foundation for searching barium giant candidates in large spectrum data sets provided by LAMOST DR8 and providing the reference values of[Ba/Fe]and[Sr/Fe]abundance.