Source-linked AI summary
Hunting for open clusters in \textit{Gaia} DR2: $582$ new OCs in the Galactic disc
A. Castro-Ginard, C. Jordi, X. Luri, J. Álvarez Cid-Fuentes, L. Casamiquela, F. Anders, T. Cantat-Gaudin, M. Monguió, L. Balaguer-Núñez, S. Solà, R. M. Badia
TL;DR
Open-cluster samples remain incomplete, while earlier searches were constrained by region selection and computational limits. This paper combines astrometric DBSCAN searches with ANN classification of CMD isochrone patterns in a Big Data environment. It reports 582 previously unknown clusters and a 45% increase in the known-cluster count.
Problem
Earlier open-cluster searches focused on particular regions or divided the Galactic disc according to computational limits, motivating a Big Data approach to blind searches.
Source: §1 · p. 2Method
The method uses DBSCAN to find five-dimensional astrometric overdensities and an ANN to recognize OC isochrone patterns in their CMDs.
Source: §5 · p. 9Results
582 previously unknown OCs were found in the Galactic disc up to G = 17, increasing the number of previously known OCs by 45%.
Source: §5 · p. 10Takeaways & Limitations
The method can identify substructures in rich regions and features such as tidal tails, and could be fine-tuned to study individual objects.
Source: §5 · p. 9The density-based detection strategy is limited to the most compact objects in the search field.
Source: §4.1.1 · p. 4
Abstract
from arXiv · showhide
Open clusters are key targets for both Galaxy structure and evolution and stellar physics studies. Since \textit{Gaia} DR2 publication, the discovery of undetected clusters has proven that our samples were not complete. Our aim is to exploit the Big Data capabilities of machine learning to detect new open clusters in \textit{Gaia} DR2, and to complete the open cluster sample to enable further studies on the Galactic disc. We use a machine learning based methodology to systematically search in the Galactic disc, looking for overdensities in the astrometric space and identifying them as open clusters using photometric information. First, we use an unsupervised clustering algorithm, DBSCAN, to blindly search for these overdensities in \textit{Gaia} DR2 $(l,b,\varpi,\mu_{\alpha^*},\mu_\delta)$. After that, we use a deep learning artificial neural network trained on colour-magnitude diagrams to identify isochrone patterns in these overdensities, and to confirm them as open clusters. We find $582$ new open clusters distributed along the Galactic disc, in the region $|b| < 20$. We can detect substructure in complex regions, and identify the tidal tails of a disrupting cluster UBC~$274$ of $\sim 3$ Gyr located at $\sim 2$ kpc. Adapting the methodology into a Big Data environment allows us to target the search driven by physical properties of the open clusters, instead of being driven by its computational requirements. This blind search for open clusters in the Galactic disc increases in a $45\%$ the number of known open clusters.
1. Introduction
Open clusters support studies of Galactic structure, evolution, and stellar physics, motivating a complete and homogeneous census. Earlier searches were limited by region selection or computational constraints, so this work adapts a machine-learning search to a Big Data environment.
1. Introduction
- A complete, homogeneous open-cluster census is needed to support studies of stellar evolution and Galactic-disc processes.Open clusters also help constrain stellar structure and evolutionary models.Source: §1 · p. 1
- Gaia DR2 prompted extensive cluster-census work, including searches using machine learning, Gaussian mixtures, visual inspection, and friends-of-friends methods.Prior studies reported new clusters in all-sky TGAS data, the Galactic anticentre, Perseus, and broader catalogue searches.Source: §1 · p. 1
- Earlier searches focused on particular regions or divided the disc according to algorithmic and hardware limits, constraining blind all-sky searches.The paper identifies Big Data processing as a way to define search regions by target structures rather than computational limits.Source: §1 · p. 2
- The paper combines DBSCAN searches for five-dimensional astrometric overdensities with an ANN that recognizes isochrone patterns in candidate CMDs.The authors adapt the approach to run in a Big Data environment.Source: §1 · p. 2
2. Methodology
The methodology prepares homogeneous sky regions, searches them repeatedly with DBSCAN, and classifies candidate clusters using CMD patterns. Distributed computation enables a disc-wide search, while simulated and real examples support ANN training and evaluation.
2. Methodology
- DBSCAN searches standardized five-dimensional astrometry in locally homogeneous L × L regions, with shifted grids and parameters optimized on Gaia-like simulations.The grid is shifted by L/3 and 2L/3 to account for border clusters; repeated runs across 35 parameter pairs assess candidate reliability.Source: §2 · p. 2
2.1. Distributed computation of DBSCAN
- 18 hours sequentially became about 1.4 hours on one 48-core node and under 18 minutes on four nodes for the same TGAS configuration.PyCOMPSs distributes DBSCAN across regions, allowing the full disc analysis to process data too large for one machine’s memory; it averages 8.27 hours per parameter pair on four nodes.Source: §2.1 · p. 2
2.2. OCs validation with Deep Learning
- A deep ANN classifies DBSCAN overdensities by learning OC isochrone patterns from CMDs, using convolutional layers followed by fully connected layers.The network receives a two-dimensional histogram in GBP − GRP versus G.Source: §2.2 · p. 3
2.2.1. Building the training set
- Because confirmed OC CMDs alone were insufficient for training, the authors augment real examples with simulated isochrones and field-star CMDs.The simulation spans log(age) 6.6–10.3 dex and distances 0.4–4 kpc; some known clusters with too few stars or dispersed isochrones were excluded.Source: §2.2.1 · p. 3
2.2.2. Performance of the classification
- The ANN architecture is selected by minimizing test loss on an 80%/20% training-test split, then tuned on anticentre candidates to retain at least 80% of confirmed OCs.Applying the final model to disc candidates recovers that 80% requirement for known OCs.Source: §2.2.2 · p. 4
3. Data
The search uses Gaia DR2 astrometry and photometry across the Galactic disc, with a magnitude limit chosen to maintain usable measurement precision. The latitude range captures most clusters in prior catalogues.
3. Data
- The search covers 0° ≤ l ≤ 360° and −20° ≤ b ≤ 20°, where prior catalogues place 94–99% of their known OCs.The latitude range is selected because the expected concentration of OCs is greatest there.Source: §3 · p. 4
- The Gaia DR2 sample is limited to G = 17, yielding 122,727,809 stars with the stated astrometric and photometric uncertainty levels.Median parallax uncertainty is 0.094 mas; proper-motion uncertainties are 0.158 and 0.137 mas yr−1 for µα∗ and µδ, respectively.Source: §3 · p. 4
4. Results
Applying the methodology to the Galactic disc produced a catalogue of 2,213 possible open-cluster candidates, including known and newly discovered objects.
4. Results
- The disc-wide search yielded 2,213 possible OC candidates, including previously known and newly discovered clusters.The passage reports candidate counts, not the number confirmed as new clusters.Source: §4 · p. 4
4.1. Comparison with existing catalogues
The authors cross-match their detections with existing catalogues to identify known groups, using positional and astrometric compatibility checks where available. The comparisons also surface candidates associated with known star-forming regions.
4.1.3. Dias et al. (2002) and Kharchenko et al. (2013)
- 5 OCs meet the position and proper-motion compatibility criteria for tentative matches with Dias et al. (2002) or Kharchenko et al. (2013).The positional screen uses a 0.5° radius; when needed, proper motions are compared with a Welch t-test, while parallaxes are not compared because those catalogues report distances without uncertainties.Source: §4.1.3 · p. 5
- The method also finds objects related to σ-Ori, Collinder 228, Bochum 10, NGC 1980, NGC 1981, NGC 6514, NGC 6530, and NGC 6604.The two catalogues classify some of these groups differently, including as moving groups, associations, or nebulosities.Source: §4.1.3 · p. 5
4.1.4. Bica et al. (2019)
- 45 additional coincidences with Bica et al. (2019) are globular clusters, and repeated GC detections provide a diagnostic of DBSCAN and the ANN's ability to recognise old isochrones.ω-Cen was found most often by DBSCAN; simulated CMDs help the ANN recognise isochrones older than 10 Gyr.Source: §4.1.4 · p. 6
4.1.5. Sim et al. (2019) and Liu et al. (2019)
- 8 OC candidates match UPK objects; proper motions agree within 1σ, while parallaxes agree within at most 1.91σ, partly because Sim et al. (2019) report estimated distances.The authors consider the objects matched despite the larger parallax discrepancy.Source: §4.1.5 · p. 6
- 45 new OC candidates are compatible with Liu & Pang (2019) objects, alongside four coincidences with clusters already reported by CG18 and CG19.The four previously reported matches are Liu & Pang cluster IDs 1973, 2143, 2230, and 2385.Source: §4.1.5 · p. 7
4.2. Newly found OCs
The search yielded 582 classified open clusters, grouped by their astrometric and colour–magnitude properties. The section describes their distance and Galactic distributions, highlights the disrupting cluster UBC 274, and reports subgroups associated with star-forming regions.
4.2. Newly found OCs
- The 676 recurrent candidates yielded 582 classified OCs: 245 class A, 236 class B, and 101 class C, with class A typically showing concentrated astrometry and a clean isochrone.Class A also includes 115 clusters (47%) with stars evolved beyond the main sequence.Source: §4.2 · p. 7
- Candidates were visually classified using CMDs and astrometric distributions, with radial velocities when available; mean radial-velocity dispersions were 5.47, 6.59, and 11.81 km·s−1 for classes A, B, and C.The authors describe these measurements as an external check on possible contaminant stars; part of the dispersions may be due to multiplicity.Source: §4.2 · p. 7
4.2.1. Comments on the new OCs
- Inverting mean parallaxes gives estimated distances of 387 pc to about 11 kpc; the new OCs increase the census by 18% below 1 kpc, 54% at 1–2 kpc, and 49% beyond 2 kpc.The authors caution that parallax inversion is unsuitable for objects with large relative parallax uncertainties; Fig. 4 compares the parallax distributions of new and previously known OCs.Source: §4.2.1 · p. 7, Figure 4 · p. 8
- 83.5% of new OCs lie at |b| < 5°; the section also reports few new objects at longitudes 140°–210° and none at high |ZGal| in the inner disc (RGal < 7 kpc).Figures 3, 5, and 6 show the census in l–b, X–Y, and Galactocentric R–Z projections, respectively.Source: §4.2.1 · p. 7, Figure 3 · p. 8, Figure 5 · p. 8
4.2.2. Specific remarks on UBC 274
- UBC 274 lies at about 2 kpc and b ≈ −12.8°, has 365 stars, and shows 13 of 15 measured radial velocities compatible with a mean of −22.92 km·s−1.Its concentrated astrometric clump and clean isochrone imply an age of about 3 Gyr; its spatial distribution is elongated in the proper-motion direction.Source: §4.2.2 · p. 7
- UBC 274’s elongation suggests it is another example of disruption, alongside tidal tails previously detected in the Hyades, Praesepe, and Coma Berenices.Fewer than 20% of previously known clusters are older than 1 Gyr, and only 5% are older than 2 Gyr.Source: §4.2.2 · p. 8
4.2.3. Substructure in star forming regions
- The search identified seven groups related to the Carina Nebula, illustrating that several detected objects may be associated with one star-forming structure.Figure 9 shows their spatial distribution.Source: §4.2.3 · p. 8
- Collinder 228 and UBC 505 share sky coordinates but were distinguished as separate objects by their parallaxes of 0.42 and 0.29 mas, respectively.Figure 9 marks stars of the new UBC clusters in different colours and related known clusters with dashed circles.Source: §4.2.3 · p. 8, Figure 5 · p. 8
5. Conclusions
The paper applies a blind, two-stage machine-learning search to Gaia DR2, detecting astrometric overdensities with DBSCAN and confirming cluster candidates through ANN recognition of CMD isochrones. The whole-disc search finds 582 previously unknown open clusters and identifies substructure, including tidal tails around UBC 274.
5. Conclusions
- DBSCAN detects overdensities in five-dimensional astrometric space, then an ANN identifies CMD isochrone patterns to confirm open clusters.Source: §5 · p. 9
- PyCOMPSs deployment on MareNostrum enables searches targeted to cluster properties rather than computational limits, including in high-density regions.Source: §5 · p. 9
- The ANN improves photometric confirmation over a simple multi-layer perceptron when two-dimensional correlations are present in CMDs.Source: §5 · p. 9
- The blind search detects substructure in rich regions and features such as tidal tails; finer parameter tuning could support more detailed studies of individual clusters.Source: §5 · p. 9
- 582 previously unknown open clusters were found across the Galactic disc at |b| < 20° and G ≤ 17, increasing the known-cluster count by 45%.The reported increases were 18% within 1 kpc, 54% between 1 and 2 kpc, and 49% beyond 2 kpc.Source: §5 · p. 10
- UBC 274, about 3 Gyr old at roughly 2 kpc, has an elongated outskirts shape associated with disruption by tidal tails.Source: §5 · p. 10, Figure 8 · p. 10