938,720 giants from LAMOST I: Determination of stellar parameters and α, C, N abundances with deep learning

Extracting accurate atmospheric parameters and elemental abundances from stellar spectra is crucial for studying the Galactic evolution. In this paper, a deep neural network architecture named StarNet is used to estimate stellar parameters (T, log g, [M/H]), α-elements as well as C and N abundances from LAMOST spectra, using stars in common with APOGEE survey as training data set. With the spectral signal-to-noise ratio (S/N) in g band (S/N ) larger than 10, the test indicates our method yields uncertainties of 45 K for T, 0.1 dex for log g, 0.05 dex for [M/H], 0.03 dex for [α/M], 0.06 dex for [C/M] and 0.07 dex for [N/M]. Because of few stars with [M/H] < −1.0 dex in the training set, the uncertainties are dominated by stars with [M/H] > −1.0 dex. Based on test results, we think StarNet is valid for measuring parameters from low-resolution spectra of the LAMOST survey. The trained network is then used to predict parameters for 938,720 giants from LAMOST DR5. Within the range of stellar parameters 4000 K < T <5300 K, 0 dex < log g < 3.8 dex and −2.5 dex < [M/H] < 0.5 dex, the comparisons with high-precision measurements (e.g., PASTEL, asteroseismic log g) yield uncertainties of 100 K for T, 0.10 dex for log g, 0.12 dex for [M/H]. Our estimations are consistent with values from the high-precision measurements. In this research, a deep neural network is successfully applied on the numerous spectra from LAMOST. The deep neural network shows an excellent performance, which demonstrates that deep learning can effectively reduce the inconsistencies between parameters measured by the individual survey pipelines.