The Galah Survey: Classification and Diagnostics with t-SNE Reduction of Spectral Information

Abstract

Galah is an ongoing high-resolution spectroscopic survey with the goal of disentangling the formation history of the Milky Way using the fossil remnants of disrupted star formation sites that are now dispersed around the Galaxy. It is targeting a randomly selected magnitude-limited (V 14) sample of stars, with the goal of observing one million objects. To date, 300,000 spectra have been obtained. Not all of them are correctly processed by parameter estimation pipelines, and we need to know about them. We present a semi-automated classification scheme that identifies different types of peculiar spectral morphologies in an effort to discover and flag potentially problematic spectra and thus help to preserve the integrity of the survey results. To this end, we employ the recently developed dimensionality reduction technique t-SNE (t-distributed stochastic neighbor embedding), which enables us to represent the complex spectral morphology in a two-dimensional projection map while still preserving the properties of the local neighborhoods of spectra. We find that the majority (178,483) of the 209,533 Galah spectra considered in this study represents normal single stars, whereas 31,050 peculiar and problematic spectra with very diverse spectral features pertaining to 28,579 stars are distributed into 10 classification categories: hot stars, cool metal-poor giants, molecular absorption bands, binary stars, Hα/Hβ emission, Hα/Hβ emission superimposed on absorption, Hα/Hβ P-Cygni, Hα/Hβ inverted P-Cygni, lithium absorption, and problematic. Classified spectra with supplementary information are presented in the catalog, indicating candidates for follow-up observations and population studies of the short-lived phases of stellar evolution.

The GALAH Survey: second data release

“…The final training set for stellar parameters of 10605 spectra consists of 21 Gaia benchmark stars, 12 stars overlapping with Bensby et al (2014), 77 stars with H parallaxes, 3807 stars with TGAS parallaxes, 915 stars with asteroseismic information, 669 open or globular cluster stars as well as 1805 stars already included in previous training sets (Martell et al 2017;Sharma et al 2018;Wittenmyer et al 2018). To ensure a sufficient coverage of parameter space for the training step, we expand the set with 1057 selected stars in the parameter range of [Fe/H] < −1.0, 1055 additional stars with −1.0 < [Fe/H] < −0.5, 654 additional giants with T eff < 5000 K, log g < 2.0 dex, and SNR > 125 in the green channel, 388 stars with projected high Li abundances based on the work by Traven et al (2017), and 145 stars with SNR > 100 or SNR > 50 at log g < 2 overlapping with APOGEE DR14. We stress that we excluded spectroscopic binaries from the training set, either based on previous automated stellar classifications, see Section 3.4.2 or via visual inspection of the training set spectra.…”

“…This is especially true in the case where the sheer quantity of collected information prevents us from manually inspecting the data as it comes in, and also when it is not possible to determine all sorts of outliers and unexpected issues a priori. Because the GALAH sample of observed spectra fits this description, it was necessary to develop a semi-automatic classification procedure, which has been presented in Traven et al (2017). Here we briefly outline the classification scheme of the spectra at rest wavelength and its recent improvements.…”

“…We no longer use the two-step procedure described in Traven et al (2017) for producing a t-SNE map of peculiar spectra, since we have overcome the computational difficulties and can now produce the projection map using all available spectral information of the whole dataset in less than a day on a 24-core Xeon node. This became possible by using the parallel multicore t-SNE code (Ulyanov 2016) and modifying it to remove the limit of overall information that we put into t-SNE.…”

“…This became possible by using the parallel multicore t-SNE code (Ulyanov 2016) and modifying it to remove the limit of overall information that we put into t-SNE. Additionally, we exclude the infrared band as it still suffers from the presence of strong spikes (see Traven et al 2017), which are now understood and accounted for. The inclusion of this band would hamper our classification significantly compared to what could be gained from information contained therein.…”

The GALAH Survey: lithium-strong KM dwarfs

“…Additionally, manually selected spectra classified as peculiars in GALAH were attached to the list. The list of peculiar spectra in GALAH -spectra with features like emission lines, broad TiO bands etc., including spectra affected by technical issues -has been prepared with the t-SNE classification technique (Traven et al 2017). As the pipelines being used to analyze GALAH data are in some cases not able to treat such spectra, objects from the peculiar list have been excluded from further consideration in the main GALAH survey.…”