Classification and diagnostics of the GALAH dataset with t-SNE reduction of spectral information

Peculiar spectra and objects to which they belong are relatively abundant among targets of general all-sky surveys such as GALAH. Detection of such objects is important because the automatic evaluation of their stellar and chemical properties might turn out to be very challenging and therefore introduces a complication in achieving scientific goals such as galactic archaeology. On the other hand, the nature of peculiar objects can be quite intriguing and might include various types of stars such as cataclysmic variables, symbiotic stars, stars with massive outflows or inflows, stars exhibiting chromospheric emission, different binary systems, and others.

We employ the novel dimensionality reduction technique t-SNE (t-Distributed Stochastic Neighbor Embedding) to alleviate the discovery and overview of distinct morphological types that occur among the tens of thousands of spectra in GALAH. Using all available spectral information, t-SNE projection map exposes recognizable morphological groups, detecting also diverse influences and glitches occuring in the spectral reduction pipelines. We assign classification categories to these distinct groups of GALAH spectra, making use of a utility called t-SNE Explorer, which provides all available supplementary information about observed objects to the visually very compelling projection map.