Two faces of Gaia-Sausage-Enceladus: Mining the chemical abundance space with graph attention networks

TL;DR

This work addresses the challenge of disentangling Milky Way halo substructure by leveraging high-dimensional chemical abundances from GALAH DR4 in concert with integrals of motion. It introduces a graph attention autoencoder that denoises chemical space and identifies coherent stellar substructures, including a notable dichotomy within Gaia-Sausage-Enceladus (GSE) into two chemically distinct groups linked to their birthplace in the progenitor. The method recovers key globular clusters, estimates an in-situ fraction of about $41\%$, and reveals a broader, chemically-informed membership for GSE beyond purely dynamical associations. The approach demonstrates the value of graph-based chemical tagging for decoding galactic assembly histories and suggests strong utility for upcoming surveys and lower-resolution spectroscopic data.

Abstract

Recent studies suggest that chemical abundances hold the key to disentangling halo substructure, providing a more reliable tracer than dynamics alone. We aim to probe the Milky Way stellar halo using high-dimensional chemical abundances from GALAH DR4. By leveraging multiple nucleosynthesis channels in synergy with integrals of motion (IoM), we extract information hidden in the raw abundance space to perform chemical tagging. With a graph attention autoencoder, we reconstruct a dynamics-informed, denoised chemical space and identify coherent stellar substructures by applying ensemble clustering. Our method successfully recovers the three largest globular clusters hidden in the dataset, estimates the in-situ fraction to be approximately 41\%, and chemically characterizes several dynamical halo substructures. Strikingly, stars dynamically associated with Gaia-Sausage-Enceladus (GSE) separate into two chemically distinct clusters. By examining their abundances, energy ($E$) and angular momentum ($L_z$) distributions, together with the metallicity trend with $E$, we connect these clusters to their birthplace within the progenitor by proposing a simple infall scenario: one cluster traces the metal-poor, less evolved outskirts, while the other traces the metal-rich, chemically evolved core.

Figures (16)

• Figure 1: Schematic of the algorithm. Chemical abundances define node features and IoM the edges. The GAT autoencoder reconstructs the node features, learning which links to retain or break. Chemically similar stars remain connected, while dissimilar ones disconnect. On top, original (left) vs. reconstructed abundances (right) with two of the five latent-space dimensions in the middle. This diagram was created with draw.iodrawio.
• Figure 2: Lindblad diagram of the halo star sample used in this work. The four target streams identified in Dodd_2023 are highlighted in color, while GCs are shown as gray points for clarity, indistinguishable from field stars. Edges of the graph are represented as gray lines, in which each edge links two stars from the selected GALAH DR4 sample
• Figure 3: Top: Membership probability distribution for the different clusters. For each cluster, the light rectangle shows the top 75% memberships, and the darker rectangle spans the top 50% (median). Individual stars are overplotted as black points. Bottom: Cross-tab of cluster labels derived in this work versus those from Dodd_2023 and Lucatello et al. in prep.. The heatmap shows the number of stars in each label combination.
• Figure 4: Reachability plot of all identified clusters. Streams from Dodd_2023 are colored according to the palette used in their work. Gray points represent field stars, among which GCs are hidden. Shaded areas indicate the approximate locations of the consensus clusters identified in this work.
• Figure 5: Top: Lindblad diagram for all clusters identified in this work. All stars are shown in gray (hexbin), while clusters are overplotted using a 2D Gaussian KDE. The three highest-density (60, 80 and 100%) levels are shown with filled contours, and stars outside these regions are plotted as scatter points to highlight lower-density members and outliers. Bottom: [Mg/Fe] vs [Fe/H] abundance plane for the same clusters, using same criteria for the density contours.