Research On Automatic Measurement Algorithm Of Stellar Atmospheric Physical Parameters

With the continuous development and implementation of astronomical research projects(such as SDSS,GAIA and LAMOST),the amount of spectral data obtained has increased rapidly,making the automatic measurement of the physical parameters of the stellar atmosphere based on the spectrum gradually become one of the main scientific research directions of the future astronomical spectral analysis.At present,deep learning algorithm is a hot topic in the field of artificial intelligence.Its algorithm is outstanding in the fast processing and abstract representation of complex big data.Based on the LAMOST DR8 celestial stellar spectral data,the convolutional neural network and regression algorithm are combined to measure the physical parameters of the stellar atmosphere.The main contents are as follows:1.Pretreatment of spectral data.In the experiment,20,000 pieces of low-resolution spectral data from LAMOST DR8 were used,and all the original spectral data were interpolated into the same wavelength range for feature extraction of spectral flow.Then the normalized one-dimensional spectrum is converted into a two-dimensional matrix form by folding and input into the model for training and testing.2.Application of machine learning algorithm in measurement of stellar atmospheric parameters.Firstly,the basic architecture of classical machine learning algorithm is derived and four machine learning algorithms are introduced:Classification And Regression Trees(CART),Support Vector Regression(SVR),Random Forest(RF)and Deep Forest,(DF),respectively successfully applied to the stellar spectral data set,proving that these four algorithms are feasible in parameter measurement and have good prediction effect.Experimental results and error analysis show that DF algorithm can automatically estimate stellar atmospheric parameters more accurately than other three regression algorithms.3.The Convolutional neural network(CNN)and the classical regression network model in machine learning are combined to construct the model.The selection of CNN network model was theoretically analyzed,and then the CNN model and different regression networks were combined for parameter measurement.The final experimental results show that the two-dimensional spectrum retains more information,reduces the information loss caused by dimensionality reduction of the information spectrum in higher dimensions,and thus greatly improves the accuracy of parameters.In addition,it also shows that CNN needs to be combined with a strong nonlinear fitting regression to obtain a higher prediction accuracy.