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

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.

💡 Research Summary

The paper presents a novel, chemistry‑driven approach to dissecting Milky Way halo substructure, focusing on the massive accretion event Gaia‑Sausage‑Enceladus (GSE). Using GALAH DR4, the authors extract 13 elemental abundances for each star and construct a graph where nodes are stars and edges encode dynamical similarity derived from integrals of motion (energy E, vertical angular momentum Lz, etc.). A graph attention network (GATv2) autoencoder is then trained: the encoder compresses the high‑dimensional abundance vectors into a five‑dimensional latent space that deliberately corresponds to the four major nucleosynthetic channels (α‑elements, iron‑peak, s‑process, r‑process) plus a free dimension, while the decoder reconstructs the original abundances. The attention mechanism allows the edge weights to be dynamically adjusted based on chemical similarity, ensuring that chemically alike stars exchange information and chemically distinct stars are effectively ignored, even if they are close in orbital space.

Training proceeds for 200 epochs with a learning‑rate schedule; the epoch that minimizes inter‑cluster edges while preserving intra‑cluster connectivity is selected. The resulting latent embeddings are fed into the OPTICS clustering algorithm, and an ensemble of ten independent runs yields robust membership probabilities for each stable cluster. This pipeline successfully recovers the three most massive globular clusters hidden in the sample, and an “in‑situ” fraction of roughly 41 % of the halo stars is inferred.

When the authors isolate stars that are dynamically associated with GSE, the chemical embeddings split into two distinct clusters. One cluster is metal‑poor (⟨