Research on Parameterization of LAMOST Spectra

Abstract

We are now entering a new era in which a huge number of astronomical observations provide scientific research support for astronomers. Most of the physical information of stars is encoded into their spectra, such as spectral absorption line characteristics, radial velocities (RVs), stellar atmospheric parameters, and chemical abundances. This information can constrain the stellar evolution model and the dynamic model of the Milky Way. Current and upcoming surveys help us understand the structure of the Milky Way and the evolution of stars more comprehensively than ever before. The methods of extracting information from millions of spectra are needed due to the important of the stellar atmospheric parameters. In general, estimating the atmospheric parameters of stars through spectroscopy can be divided into two aspects that are either physical driven or data driven. The physical-driven methods use the theoretical models to fully or partly fit observed spectra by the nearest neighbor matching rule, and determines the stellar atmospheric parameters of the best matched theoretical spectrum as that of the observed spectrum. The physical-driven methods could provide uniform parameter coverage and reasonable interpretation of results. The flaw of these methods lies in the imperfection of the theoretical model, which leads to zero-point deviation for different theoretical model chosen. In recent years, the data driven methods were extensively proposed, mainly by assuming that the same star has the same atmospheric parameters in different band spectra. Transferring the stellar parameters measured from high-resolution spectra to the low-resolution spectra allows one to construct maps from the precise stellar parameters to observed spectral fluxes. Data-driven methods can provide high-precision parameters but are limited by the distribution of the reference sample, and the extrapolation of data-driven model always show bad performance. Our work employs machine learning methods in two different modes above which are suitable to analyze stellar spectra produced by the Large Sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST), a 4 m reflecting Schmidt telescope in China. The LAMOST spectrograph has two resolving modes with different wavelength coverages, the lowresolution mode of R∼ 1800 covering 3800–9000Å and the medium resolution mode of R∼ 7500 covering two separated bands, 4950–5350 and 6300–6800Å. The LAMOST-II medium resolution survey, LAMOST-II MRS (Liu et al. 2020), began on 2017 September 1 after the first five year regular lowresolution survey, LAMOST-I LRS (Luo et al. 2015). Our first work analyzes data from LAMOST DR5 LRS stellar spectra (Wang et al. 2019). LAMOST DR5 provides a stellar parameter catalog of late A and FGK-type stars with high-quality spectra measured by LAMOST Stellar Parameters pipeline (LASP; Luo et al. 2015). We use the full-band spectral information of the Phoenix theoretical grids (Husser et al. 2013), which is different from the reference sample data used by LASP. We also construct a generative spectrum network (GSN) as an interpolator, which is based on a fully connected artificial neural network that can portray the mapping relationship of stellar parameters to low-resolution spectral flux. GSNs are trained using 13,850 spectra and reproduce the theoretical spectrum well testing by 11,080 test sets. By Monte Carlo (MC) sampling 5000 couples of stellar parameters for a observed spectrum, we obtain corresponding 5000 spectra produced by GSN, and the posterior distribution of the stellar parameters based on the Bayes algorithm and then get the point estimation and error of the stellar parameters. In this way, we can link the spectra generated by the GSN with LAMOST-I LRS spectra and estimate their stellar parameters. 5.3 million LAMOST-I LRS stellar spectra with a signal-to-noise ratio (S/ N) higher than 10 are estimated by the methods above. With the high-performance parallel computing platform SPARK, the estimations of the stellar spectral parameters are obtained. Figure 1 show the comparison of our results with LASP stellar parameters in different S/N levels. We also show the comparison with external reference sets APOGEE, GCS, PASTEL catalog and astronomical samples, our results show good consistency (see Wang et al. 2019 for details). When the S/N of spectra are higher than 50, the precision of Teff, log g, [Fe/H], and [α/Fe] is 80 K, 0.14 dex, 0.07 dex and 0.17 dex, respectively. The second work analyzes RVs of medium-resolution spectra in LAMOST-II MRS (Wang et al. 2019). The RV is a basic physical quantity that describes the Doppler movement of the line of sight between the observer and the star, which is always the primary parameter extracted from the spectrum. The Publications of the Astronomical Society of the Pacific, 133:037001 (3pp), 2021 March https://doi.org/10.1088/1538-3873/abdd9b