Two Faces of a Galactic Collision: Uncovering the Chemical Story of Gaia–Sausage–Enceladus

Understanding how the Milky Way formed requires identifying the remains of smaller galaxies that merged with it long ago. In this paper, Quandt-Rodriguez et al. investigate the stellar halo of the Milky Way, the region that preserves many of these ancient merger remnants. The authors focus on Gaia–Sausage–Enceladus (GSE), the most massive known merger event in the Galaxy’s history and ask whether detailed chemical information from stars can reveal structure that is hidden when using stellar motions alone. Their key idea is that stars formed together share similar chemical “fingerprints,” which can be used to trace their origins even after billions of years of mixing.

Limitations of Dynamical Approaches

The paper begins by outlining the limitations of traditional methods based on stellar dynamics. While quantities such as orbital energy and angular momentum are often used to identify merger debris, simulations show that these signatures can overlap between different accreted systems and even with stars formed inside the Milky Way itself. Chemical abundances, on the other hand, remain largely unchanged over a star’s lifetime and directly reflect the conditions of its birth environment. This approach, known as chemical tagging, has already been successful in identifying GSE through patterns such as low alpha-element abundances compared to Milky Way stars at similar metallicity. However, clustering stars directly in high-dimensional chemical space is challenging due to measurement uncertainties and overlapping signatures.

A Graph-Based Machine Learning Method

To overcome these challenges, the authors introduce a machine-learning framework based on graph attention networks (GATs). In their method, each star is treated as a “node” with chemical abundances as its features, while connections (“edges”) between stars encode similarity in their orbital properties. The GAT autoencoder learns a compact, denoised representation of the chemical abundance space by allowing chemically similar stars to exchange more information than dissimilar ones. This reconstructed chemical space is then clustered using an unsupervised algorithm to identify coherent stellar groups, without assuming beforehand how many such groups exist.

Data Selection and Stellar Sample

The data used in this study come from GALAH DR4, a large spectroscopic survey that provides high-quality abundances for many chemical elements, combined with precise astrometric data from Gaia. After applying strict quality cuts, the authors analyze nearly 4,000 halo stars with measurements for 13 different elements, spanning alpha elements, iron-peak elements, and neutron-capture elements. Known globular clusters are included as a calibration set, allowing the authors to test whether their method can recover well-established stellar groups.

Recovered Halo Substructures

The results show that the method successfully recovers the three most massive globular clusters in the sample and identifies a large in-situ halo population, making up about 41% of the stars. Most strikingly, stars dynamically associated with Gaia–Sausage–Enceladus separate into two chemically distinct clusters, labeled GSE 1 and GSE 2. These two groups overlap strongly in orbital space but differ systematically in their chemical abundance patterns, particularly in alpha elements and neutron-capture elements. Independent checks using principal component analysis confirm that this chemical split is already present in the original data and is not an artifact of the machine-learning reconstruction.

Interpreting the GSE Chemical Dichotomy

The authors interpret this chemical dichotomy as evidence for a metallicity gradient within the original GSE galaxy before it merged with the Milky Way. GSE 1 appears more metal-poor and alpha-enhanced, consistent with stars formed in the outer, less evolved regions of the progenitor. GSE 2 is more metal-rich and chemically evolved, tracing stars formed deeper in the progenitor’s core. A simple infall scenario explains how stars from different regions of the same galaxy could end up with different energies and chemical properties after the merger, producing the two observed components.

Conclusions and Broader Implications

In their conclusions, Quandt-Rodriguez et al. argue that combining chemical abundances with dynamical information through graph-based machine learning provides a powerful new way to study the Milky Way’s formation history. Their results strengthen the picture of Gaia–Sausage–Enceladus as a massive disk galaxy with an internal metallicity gradient and demonstrate that chemical tagging can reveal merger debris that has lost its dynamical coherence. This work highlights how modern surveys and data-driven methods are opening a new window onto the Galaxy’s ancient past.

Source: Quandt-Rodriguez