Chemical Tagging in Galactic Archaeology

• Chemical tagging is a technique in Galactic archaeology that uses precise elemental abundance patterns to identify and reconstruct the birth clusters of stars.
• It employs advanced statistical metrics and clustering algorithms to distinguish low intra-cluster scatter from higher inter-cluster differences in chemical space.
• Empirical studies show that strong chemical tagging recovers about 30–40% of birth groups, highlighting challenges like contamination and limited chemical resolution.

Chemical tagging is a methodology in Galactic archaeology that seeks to reconstruct the dispersed stellar birth environment of field stars using their detailed elemental-abundance patterns. The underlying premise is that stars born from the same molecular cloud possess a distinctive, nearly invariant “chemical fingerprint,” and that, by measuring high-dimensional abundance vectors, one might reassemble dissolved clusters or associations even after phase mixing and dynamical dispersal. Strong chemical tagging—the recovery of individual birth clusters solely by chemistry—has motivated the development of multi-element spectroscopic surveys and sophisticated statistical frameworks. The overall viability of this approach, its quantitative limitations, and its current frontiers have been critically examined by high-resolution observational campaigns, cosmological simulations, and the evolution of data-driven clustering algorithms.

1. Chemical Homogeneity and Theoretical Basis

The founding assumption of chemical tagging is that co-natal stars are chemically homogeneous to a high precision across a suite of elements tracing diverse nucleosynthetic channels (α-process, iron-peak, s/r-process, etc.), while the ensemble of star-forming regions exhibits measurable inter-cluster chemical diversity. Quantitatively, chemical-tagging efficacy is governed by the ratio of intra-cluster to inter-cluster abundance scatter:

• Intra-cluster scatter for element $X$:

$$\sigma_{\mathrm{intra}}(X) = \sqrt{ \frac{1}{N_*} \sum_{i=1}^{N_*} ( [X/H]_i - \langle [X/H] \rangle )^2 }$$

• Inter-cluster dispersion:

$$\sigma_{\mathrm{inter}}(X) = \sqrt{ \frac{1}{N_{\mathrm{cl}}} \sum_{j=1}^{N_{\mathrm{cl}}} ( \langle [X/H] \rangle_j - \langle\langle [X/H] \rangle\rangle )^2 }$$

A complementary metric is the star–star “chemical distance” in abundance space:

$d_{nn'}^2 = \sum_{i=1}^I ( x_{n,i} - x_{n',i} )^2$

where $x_{n,i}$ is the abundance of element $X_i$ in star $n$. Chemical tagging is only effective if intra-cluster distances are substantially smaller than inter-cluster distances, i.e., clusters are both internally tight and externally separated in chemical space (Bhattarai et al., 2024).

Simulations of Milky Way analogs using the FIRE-2 model demonstrate that open clusters (OCs) exhibit internal abundance scatter $\sigma_{\mathrm{intra}} \lesssim 0.02$ dex for nine light elements, closely matching observed OCs. However, the mean chemical patterns between clusters are often not unique, especially in the absence of elements strongly sensitive to rare nucleosynthetic processes (Bhattarai et al., 2024). Networks of chemical similarity thus tend to show high internal homogeneity but poor inter-cluster discrimination unless the chemical dimensionality and intrinsic variation are both sufficiently high.

2. Statistical Metrics and Empirical Clustering Probabilities

Chemical tagging leverages various empirical and theoretical metrics to quantify the likelihood of two stars sharing a common birth site. Mitschang et al. introduced a pairwise mean absolute abundance difference:

$$\delta_{ij} = \frac{1}{N_C} \sum_{C=1}^{N_C} |A_C^i - A_C^j|$$

where $N_C$ is the number of shared elements, and $A_C^i$ is the abundance of element $C$ in star $i$ (Mitschang et al., 2012).

From calibration samples of open clusters, the distribution of $\delta_{ij}$ for intra-cluster and inter-cluster pairs enables the construction of a “cluster probability function” $P_{\rm cluster}(\delta)$, representing the empirical chance that a given abundance separation arises from co-natal stars:

$$P_{\rm cluster}(\delta) = \frac{\eta_{\rm intra}(\delta)}{\eta_{\rm intra}(\delta) + \eta_{\rm inter}(\delta)}$$

where $\eta_{\rm intra}$ and $\eta_{\rm inter}$ are the (normalized) density of intra- and inter-cluster pairs at separation $\delta$.

The separation between these distributions (i.e., the location where $P_{\rm cluster}=0.5$) sets the practical precision required for strong chemical tagging. Studies demonstrate that at $N_C\gtrsim 9-12$ elements and $\sigma \lesssim 0.05$ dex, chemical differentiation reaches its practical floor, and the probability of correctly identifying conatal pairs saturates (Mitschang et al., 2012). Greater precision or more discriminating elements (particularly n-capture species) incrementally improve this ceiling.

3. Methodological Approaches and Clustering Algorithms

The central challenge is discovering clustered structures in high-dimensional, noisy abundance data. Various clustering and dimensionality-reduction algorithms have been applied:

• Density-/hierarchical-based methods: DBSCAN, HDBSCAN, and OPTICS detect overdensities without a fixed number of clusters, identifying structure in chemical space that is robust to measurement noise (Spina et al., 2022, Price-Jones et al., 2019, Casamiquela et al., 2021).
• Partitioning methods: $k$-means and its variants, while computationally efficient, assume (hyper)spherical clusters and can suffer from over-fragmentation in correlated chemical spaces (Hogg et al., 2016, Blanco-Cuaresma et al., 2018).
• Manifold learning: t-SNE and UMAP provide nonlinear, density-preserving mapping from high-dimensional abundance vectors to 2D/3D for visualization; clustering in these projections has successfully recovered known clusters in survey data (Kos et al., 2017).
• Phylogenetic analysis: Maximum Parsimony and Neighbor-Joining trees, when applied to optimized subsets of tracers (e.g., Al, Ba, Co, Fe, Mg, Mn, Sc, Ti), outperform conventional clustering by capturing “family tracks” in chemical space (Blanco-Cuaresma et al., 2018).
• Deep and data-driven methods: Graph attention auto-encoders integrate chemistry with action and age proximity, learning “informed” chemical spaces where clustering becomes more efficient (Spina et al., 22 Sep 2025). Conditional auto-encoders disentangle stellar-parameter variation from abundance signals in raw spectra, enabling abundance-free chemical tagging (Mijolla et al., 2021).
• Latent-factor mixture models: MCFA leverages a nucleosynthetic factor model with mixtures-of-Gaussians in latent space, formally incorporating missing data and allowing for joint chemical and cluster assignments (Casey et al., 2019).

Clustering performance is typically quantified via homogeneity (purity), completeness (recall), the $V$-measure, and the fraction of true clusters recovered at specified thresholds. Even in idealized, high-precision datasets, current frameworks recover no more than $\sim30-40\%$ of open clusters with significant purity and completeness, with the majority of groups consisting of members from multiple true clusters (Casamiquela et al., 2021, Spina et al., 2022).

4. Empirical Results and Quantitative Benchmarks

Multiple empirical studies reveal key quantitative results and limits of strong chemical tagging:

Metric	Typical Value	Reference
Intra-cluster scatter	$\lesssim$0.02 dex (light el.)	(Bhattarai et al., 2024, Casamiquela et al., 2021)
Inter-cluster dispersion	0.03–0.12 dex (light el.)	(Bhattarai et al., 2024, Mitschang et al., 2012, Casamiquela et al., 2021)
Recovery fraction (RF)	$\lesssim$ 30–40%	(Casamiquela et al., 2021, Spina et al., 2022, Price-Jones et al., 2019)
False-positive contamination	$>$40–70% for blind groups	(Casamiquela et al., 2021, Bhattarai et al., 2024)
Heavy element differentiation	up to 0.5 dex inter-cluster	(Bhattarai et al., 2024, Mitschang et al., 2012, Casamiquela et al., 2021)

For example, in the FIRE-2 simulations, $\sigma_{\rm intra} \lesssim 0.02$ dex is observed, yet the mean abundance patterns of OCs are not unique: 86% of intra-cluster star pairs and 73% of inter-cluster pairs exhibit $d^2 < 2\times10^{-4}$ in chemical space, producing a $\sim$40–65% contamination at small distances (Bhattarai et al., 2024). Large observational surveys (APOGEE, GALAH, Gaia-ESO, etc.) replicate these findings at comparable or slightly higher scatter.

The chemical “cell” volume analysis in 10D space (with $\sim$0.05 dex element precision) yields only $\sim$500 independent “chemical volume elements” in the $\alpha$-enhanced disk, so only the most massive and/or chemically remote birth groups are detectable above the noise [1507.