Low-resolution starlight reveals 133 candidate chemical families in the Milky Way

The Milky Way’s high-alpha disk preserves clues to an early chapter of the Galaxy’s formation. Its stars have relatively high proportions of alpha elements, such as magnesium, compared with iron. Astronomers want to know whether this population is chemically smooth or contains distinguishable groups that record different histories of enrichment.

In a preprint, Tianmin Wu and colleagues report 133 candidate chemical families containing 4,642 unique stars, or 5.3% of their sample. The arXiv landing record lists the work as accepted for publication in The Astrophysical Journal. Its central finding is not the recovery of 133 original star clusters: the families may share similar enrichment histories without having formed in the same cloud.

Looking beyond iron

The researchers used chemical measurements derived from low-resolution spectra collected by the LAMOST telescope. Spectra encode information about composition through the imprints of elements on starlight. At low resolution, individual features often blend together, making precise measurements harder.

The underlying catalog uses DD-Payne, a data-driven neural-network approach constrained by physical information, to extract abundances from those spectra. The team selected red giants with sufficiently good observations and restricted temperatures, surface gravities and iron abundances. A further selection based on magnesium relative to iron isolated 87,275 high-alpha stars.

Each star was represented by 18 chemical measurements, including iron abundance and ratios involving lighter elements and eight neutron-capture elements. Carbon and nitrogen were combined into one measurement to reduce the effects of changes to their surface abundances as stars evolve. Oxygen, sodium and strontium were excluded because their measurements were less reliable.

Finding neighborhoods in chemical space

The team applied DBSCAN, an algorithm that identifies dense neighborhoods in a dataset without requiring a predetermined number of groups. Here, proximity meant chemical resemblance, not physical closeness in the Galaxy.

They standardized each chemical measurement and used a distance calculation that gave less weight to elements with larger measurement uncertainties. DBSCAN then searched for concentrations controlled by two settings: how far a neighborhood extends and how many nearby stars are needed to establish a dense region.

Those choices matter. Small neighborhoods select compact groups; larger ones can connect otherwise distinguishable structures through intermediate stars. Across the tested settings, the procedure produced 447 cluster detections, including repeated or overlapping structures.

Before consolidating these detections, the researchers excluded the 120 largest, using an empirical boundary in the cluster-size distribution to screen out potentially overconnected structures. They then merged chemically similar detections across settings, requiring all constituent clusters within a family to satisfy their similarity threshold. The resulting 133 families included 59 assembled from multiple detections and 74 represented by single detections. Some stars belong to more than one family because individual abundance errors allow overlapping assignments.

Testing whether the groups are meaningful

One check compared measurements for 44,657 stars observed by both LAMOST and the higher-resolution APOGEE survey. With appropriately adjusted neighborhood sizes, the surveys produced comparable numbers of clusters and fractions of assigned stars. That supports the usefulness of the lower-resolution measurements, but does not establish identical membership for individual groups.

Another check used a mock population in which every star had the same intrinsic composition, with independent Gaussian measurement noise added. DBSCAN usually found one cluster or none; multiple clusters appeared only in a narrow parameter range, with at most four.

The mock test therefore argues against measurement noise alone routinely fragmenting a homogeneous population into numerous groups. However, it removed both intrinsic chemical variation and correlations between elements, so it could not isolate the contribution of those correlations. The authors interpret coherent multi-element patterns as an important source of the method’s discriminatory power.

Chemical diversity, not confirmed ancestry

The families show distinct abundance patterns even at nearly identical iron abundances. Differences are especially prominent among neutron-capture elements, including barium and europium. Such variations may reflect different contributions from enrichment sources, including supernovae and asymptotic giant branch stars, but this analysis does not establish a unique origin for each family.

Orbital information reinforces the need for caution: most families span broad ranges of angular momentum. Chemically similar stars need not have closely matching orbits or a common birthplace.

Nor does the 5.3% detection fraction measure the full extent of chemical structure. Larger abundance errors can hide additional families, while the catalog depends on screening and clustering choices. Existing red-giant ages also have substantial random and systematic errors, preventing firm conclusions about when each family formed. More precise abundances and independent ages are needed to turn these candidate chemical groupings into a clearer history of the early Galactic disk.