Unsupervised machine learning techniques for chemical analysis in spectroscopic stellar surveys

Abstract

In this work, we use the dimensionality reduction technique UMAP (Uniform Manifold Approximation and Projection) and a clustering algorithm (HDSCAN) on a large sample of stellar abundance ratios from a high-quality sample of the APOGEE DR16 survey (16000 red clump stars). We are able to reliably differentiate groups of stars corresponding with the chemical thick disk and thin disc, as well as a group corresponding to high α metal rich stars, and groups with anomalous abundances of certain elements, some of which are due to low precision on the abundances of P, Co, and Na determined by the pipeline