Model Research On Estimation Of Stellar Parameters Based On Photometric Data And Low-Resolution Spectral Data

The stellar atmospheric physical parameters(effective temperature Teff,surface gravity log g,metallicity[Fe/H])and element abundance([C/Fe],[Mg/Fe],[N Fe])are very important to study the structure,composition and chemical evolution of the Milky Way galaxy.With the use of large telescopes all over the world,astronomers can obtain massive photometric data and spectral data.How to accurately measure stellar parameters(stellar atmospheric physical parameters and element abundance)from these observation data is the most concerned problem of astronomers.With the development of machine learning and artificial intelligence,the use of machine learning and deep learning to estimate stellar parameters from survey data is a very promising and worth exploring direction.The survey data of the optical band includes photometric data and spectral data.Photometric data refers to the use of telescopes to measure the weak energy of stars at different wavelengths,which is used to represent the illuminance of stars,and its value is called magnitude.Such as Large Synoptic Survey Telescope(LSST),Stellar Abundance and Galactic Evolution(SAGE)and Beijing-Arizona Sky Survey(BASS)mainly provide photometric data.Spectral data refers to the flux distribution of monochromatic light arranged by wavelength when stellar light passes through a dispersive system,such as Large Sky Area Multi-Object Fibre Spectroscopy Telescope(LAMOST)provides spectral data.The photometric data contains less information and a large amount of data,but the spectral data contains more information and a small amount of data.Some survey projects include both photometric data and spectral data,such as the Sloan Digital Sky Survey(SDSS)sky survey in the United States.In 2024,the China Space Station Telescope(CSST)will obtain billions of photometric data and hundreds of millions of spectral data,which will provide a rich data source for expanding the stellar parameter space.CSST provides photometric data in NUV,u,g,r,i and z bands and spectral data with a resolution of 200.Accurate estimation of stellar parameters based on these data is of key significance to the success of CSST project.For photometric data,a large number of studies are based on photometric data to construct color index feature to estimate Teff and[Fe/H].However,the strong collinearity between color indexes leads to model instability and affects the final prediction accuracy.For spectral data,the current research mainly focuses on medium and high resolution spectra,and there are few research focus on estimating stellar parameter from low resolution spectra similar to sepctra obtained with CSST.Based on discussion above,this thesis focus on the problem of determining the stellar parameters from photometric data and spectral data.The main content are as follows:(1)We use LightGBM model to establish the mapping of photometric data,color index,principal components and stellar atmospheric physical parameters respectively.The experimental results show that the principal component overcomes the instability of the model.The prediction errors of Teff,log g and[Fe/H]are 90 K,0.40 dex and 0.20 dex respectively,which are lower than the prediction errors of color index by 6%,11%and 13%.In addition,it is found that the second principal component and log g can be used to draw HR diagram and analyze the evolution history of stars.By comparing with machine learning models such as XGBoost,Random Forest and Linear regression,we found that the prediction error of LightGBM model is lower,and the speed of the model is 4 to 40 times faster than that of other models.Therefore,principal component+LightGBM can be used to explore the stellar atmospheric physical parameters of CSST massive photometric data.(2)We develop a method to determining the element abundance from CSST low-resolution spectra similar to spectra obtained with CSST.We use residual artificial neural network model to estimate element abundance.The experimental results show that the residual artificial neural network has stability and interpretability.The errors of[C/Fe],[Mg/Fe]and[N/Fe]in the test set are 0.06 dex,0.05 dex and 0.11 dex,which are close to the accuracy of high-resolution spectra.By comparing with StartNet,convolutional neural network and other models,we found that the prediction error of residual artificial neural network model is lower than that of other deep learning models,which is more suitable for regression prediction of low-resolution spectral.Therefore,residual artificial neural network can be used in the study of CSST spectral data.