Searching for chemo-kinematic structures in the Milky Way halo with deep clustering algorithms

TL;DR

This paper introduces CREEK, a machine learning algorithm combining kinematic and chemical data to identify chemo-kinematic structures in the Milky Way halo, enhancing the detection of accretion remnants.

Contribution

The novel CREEK algorithm integrates Siamese Neural Networks and Graph Neural Networks to analyze halo star data, improving structure recovery over traditional methods.

Findings

Successfully identified chemo-kinematic groups in the Milky Way halo.

Enhanced clustering accuracy by combining chemical and kinematic data.

Demonstrated effectiveness with data from Gaia-ESO and APOGEE surveys.

Abstract

According to the lambda CDM scenario, galaxies are formed through the hierarchical accretion of building blocks. Our Galaxy is a privileged place to look for the remnants of accretion events through the study of the chemical and kinematic properties of its halo stellar populations. Due to its low density, the stellar halo holds the most favorable conditions for chemical tagging. However, chemical tagging alone often yields weak results due to both uncertainties in chemical abundances and to overlapping chemical properties among different populations. To overcome this problem, the use of chemical and kinematic properties can be combined. In this Thesis, we developed a machine learning algorithm, named the CREEK, which combines orbital and chemical properties of halo stars observed by two large public spectroscopic surveys, Gaia-ESO and APOGEE. The CREEK operates as follows: 1)Data selection: We selected halo stars from the APOGEE and Gaia-ESO surveys based both on their velocity and metallicity and we computed their orbital parameters. 2)Using kinematics: The selected data were passed to a Siamese Neural Network that established links between stars based on their kinematic similarities. 3)Using chemistry: The graph was passed through a Graph Neural Network (GNN) auto-encoder that took as input the selected abundances. The abundances were chosen to maximize homogeneity within stars from the same cluster while ensuring distinctiveness between stars from different clusters. Additionally, we prioritised elements with smallest errors. The GNN auto-encoder computed a mean of the abundances of all connected stars, weighted on the number of links of each star and mapped the chemical space into a more efficient representation in the latent space. 4)Recovering structures: Finally, OPTICS was applied to the latent space, providing groups based on the chemical similarities of the stars.