Exploring the ex-situ components within Gaia DR3

Technical Summary: Exploring the ex-situ components within Gaia DR3

Problem Statement
The hierarchical merger hypothesis posits that large galaxies, including the Milky Way (MW), form through the accretion of smaller satellite galaxies. While high-quality astrometric data from the Gaia mission has revolutionized the study of Galactic structure, a significant discrepancy exists between the sample sizes of astrometric surveys (providing 6D phase space information for millions of stars) and spectroscopic surveys (providing chemical abundances for only $\sim$1 million stars). Chemical abundance data is crucial for distinguishing in-situ stars (formed within the MW) from ex-situ stars (accreted from dwarf galaxies), as they exhibit distinct chemical evolution patterns. However, relying solely on kinematics without chemical data has been challenging due to the lack of labeled training data and the complexity of high-dimensional phase space. Existing unsupervised clustering methods struggle to guarantee that identified structures are related to ancient merger events rather than local disruptions. This study addresses the need for a data-driven methodology that utilizes kinematic data alone to identify ex-situ stars with high purity, leveraging the large sample size of Gaia DR3.

Methodology
The authors developed a two-phase deep learning framework to classify stars as in-situ or ex-situ based on 6D phase space parameters and orbital actions.

1. Base Model (NN_FIRE):

   • Architecture: A fully connected neural network (NN) with five hidden layers (three with 128 nodes, two with 64 nodes), utilizing ReLU activation and batch normalization.
   • Training Data: Synthetic data from the FIRE-2 cosmological hydrodynamic simulations (specifically the m12i galaxy from the Latte suite). The training set consisted of 47.7 million synthetic stars with "truth-level" labels indicating their ex-situ status (defined as originating >30 kpc from the primary galaxy center).
   • Inputs: 3D position and velocity in Galactocentric Cartesian coordinates.
   • Loss Function: Focal Loss was employed to address the extreme class imbalance (where ex-situ stars comprise <1% of the sample), focusing the model on hard-to-classify samples.
   • Performance: On the test set, NN_FIRE achieved an Area Under the Curve (AUC) of 0.98.

2. Parallel Model (NN_parallel):

   • Architecture: A parallel structure where the weights of the pre-trained NN_FIRE are frozen. A sub-network processes orbital actions ($J_R, J_z, J_\phi$) derived from observational data. The outputs of both networks are concatenated and processed by additional dense layers.
   • Training Data: Observational data from Gaia DR3, cross-matched with APOGEE DR17 and LAMOST DR8 VAC to obtain chemical abundances.
   • Labeling Strategy: Instead of using truth-level labels (unavailable for real data), the authors used a chemical tagging criterion in the $[Mg/Mn]$-$[Al/Fe]$ plane to label training samples as ex-situ or in-situ. This criterion was designed to be independent of dynamics to avoid bias.
   • Target Sample: 27,085,748 stars from Gaia DR3 with available radial velocities and photo-astrometric distances, filtered for quality (RUWE < 1.4, radial velocity error < 20 km/s, distance error < 30%).
   • Preprocessing: A Monte Carlo approach was used to propagate errors, and stars in regions prone to distance overestimation (near the Galactic center) were excluded.

Key Results

• Classification: The NN_parallel model identified 160,146 ex-situ stars from the target sample of ~27 million.
• Sample Composition:
  • Metallicity: The metallicity distribution peaks around -1.3 to -1.4 dex, consistent with the Gaia-Sausage-Enceladus (GSE) population, though the sample comprises multiple components.
  • Identified Structures: The model successfully isolated member stars of the Large Magellanic Cloud (LMC), Small Magellanic Cloud (SMC), and the Sagittarius (Sgr) dwarf galaxy (both core and stream).
  • Globular Clusters: 20 globular clusters were identified as ex-situ components.
  • Substructures: The authors selected member stars for six known substructures: GSE, Thamnos, Sequoia, Helmi streams, Wukong (LMS-1), and Pontus.
• Spatial Distribution:
  • The ex-situ fraction increases with vertical height ($|z|$) from the Galactic disc and with Galactocentric distance ($R$).
  • Component Fractions: The estimated proportions of ex-situ stars are:
    • Thin Disc ($|z| < 0.6$ kpc): 0.1%
    • Thick Disc ($0.6 \le |z| < 4$ kpc): 1.6%
    • Halo ($4 \le |z| \le 10$ kpc): 63.2%

Significance and Claims
The paper positions this work as a preliminary study introducing a novel algorithm and providing a large, reliable sample of ex-situ stars for the community. The authors claim that their deep learning approach offers a significant advantage over traditional linear segmentation methods by better balancing precision and recall in high-dimensional data. Specifically, the method successfully eliminates the majority of disc stars while capturing known merger debris and substructures without relying on clustering algorithms for the initial selection.

The authors explicitly state that their results regarding the proportions of ex-situ stars in the thin disc, thick disc, and halo are preliminary estimations subject to Gaia selection effects and the performance of the NN. They do not claim these figures as definitive conclusions but rather as a basis for statistical exploration. The study serves as a "prior work" intended to demonstrate the viability of using supervised deep learning on kinematic data to uncover merger history, with future work planned to incorporate simulations, chemical abundance analyses, and clustering algorithms for deeper investigation.