Deep chemical tagging -- Identifying open clusters and moving groups in chemical space with graph attention networks

TL;DR

This work tackles chemical tagging in the thin disk by addressing chemical space homogeneity and measurement limitations. It introduces a graph attention auto-encoder (GATE) that semantically informs chemical patterns using kinematics and ages, producing an enhanced space for clustering. Applied to ~47{,}000 APOGEE thin-disk stars, the method identifies 282 groups, recovering five of six target open clusters and several moving groups, outperforming clustering in original chemical space. The approach demonstrates robustness to observational perturbations and offers a scalable framework for large-scale Galactic archaeology.

Abstract

Reconstructing the formation history of the Milky Way is hindered by stellar migration, which erases kinematic birth signatures. In contrast, stellar chemical abundances remain stable and can be used to trace stars back to their birth environments through chemical tagging. This study aims to improve chemical tagging by developing a method that leverages kinematic and age information to enhance clustering in chemical space, while remaining grounded in chemistry. We implement a graph attention auto-encoder that encodes stars as nodes with chemical features and connects them via edges based on orbital similarity and age. The network learns an ``informed'' chemical space that accentuates coherent groupings.Applied to $\sim$47,000 APOGEE thin disk stars, the method identifies 282 stellar groups. Among them, five out of six open clusters are successfully recovered. Other groups align with the known moving groups Arch/Hat, Sirius, Hyades, and Hercules. Our approach enables chemically grounded yet kinematically and age informed chemical tagging. It significantly improves the identification of coherent stellar populations, offering a framework for future large-scale stellar archaeology efforts.

Figures (8)

• Figure 1: Portion of the graph processed by the GATE . This subgraph shows all nodes belonging to Trumpler 20 and all their immediate neighbors. Edges between all nodes are shown as solid black lines. Cluster members are highlighted by a red circle. Each node is color-coded based on the stellar age. Nodes are arranged in the plane to minimize the distance between connected nodes.
• Figure 2: Top panel. Cumulative distributions of R$_{\rm mem}$ and R$_{\rm field}$ derived for the epochs 5, 38, and 100. Bottom panel. Difference between the area of the cumulative distributions calculated for R$_{\rm field}$ and that obtained for R$_{\rm mem}$, as a function of the training epoch. The orange, green and red vertical lines correspond to the distributions shown in the top panel. The peak of this curve identifies the model that better captures the chemical similarity/diversity between connected stars.
• Figure 3: Top panel. The histogram shows the cumulative fraction of the ratio R = $\omega$/$\omega_{\rm self}$ for all edges (blue) and for the edges connecting members of the same cluster from the calibration set (orange). The dashed red line indicates the threshold below which we consider the edge as non-significative and we remove it from the graph. Bottom panel. The histogram shows the cumulative fraction of the ratio between the node degrees after and before the edge breaking. All stars are shown in blue, while members of the calibration set are in orange. The dashed red line indicates the threshold below which we consider the star as "migrator."
• Figure 4: Five panels showing the fraction of migrating stars within the [$\alpha$/M]-[M/H] diagram at different L$_{\rm z}$ bins. The left panel corresponds to the inner disk, and the right panel to the outer disk. The contours represent the distribution of all stars within the diagram at all L$_{\rm z}$.
• Figure 5: Top panel. The plot shows the distribution of stars from the full dataset (shades of gray) and members of the open clusters from the target set (colored points) in the original [$\alpha$/M]-[M/H] diagram. Bottom panel. As a comparison we show the same diagram of the top panel reconstructed by the GATE. This chemical space has been selectively informed by kinematics as it is described in section \ref{['Sec:Methods']}.