PaperlayerView exact revision on arXiv

Source-linked AI summary

A new method for unveiling Open Clusters in Gaia: new nearby Open Clusters confirmed by DR2

A. Castro-Ginard, C. Jordi, X. Luri, F. Julbe, M. Morvan, L. Balaguer-Núñez, T. Cantat-Gaudin

arXiv:1805.03045v2astro-ph.GA

TL;DR

The paper addresses automated open-cluster detection in rapidly growing astronomical catalogues, where manual analysis is difficult. It combines five-dimensional DBSCAN clustering with ANN classification, then checks TGAS candidates using Gaia DR2. Manual validation confirmed around 70% of 31 candidates, with 100% success among cases with Nfound > 5.

  • Problem

    Astronomical catalogues are growing beyond the reach of usual manually supervised analysis, motivating automated methods for finding open clusters in large astrometric datasets.

    Source: §1 · p. 1

  • Method

    The method uses DBSCAN to find groups in five astrometric dimensions and an ANN to distinguish real open clusters from spurious statistical clusters.

    Source: §6 · p. 10

  • Results

    Around 70% of 31 TGAS candidates were confirmed as open clusters using Gaia DR2, with 100% success for candidates with Nfound > 5.

    Source: §6 · p. 10

  • Takeaways & Limitations

    The authors report that the method reliably detects open clusters, while TGAS scanning-law artefacts and its bright magnitude limit constrain detection.

    Source: §6 · p. 10

    TGAS preprocessing removes stars with proper-motion components above 30 mas · yr−1, which may remove cluster members and prevent a cluster from being detected.

    Source: §4.1 · p. 6

Abstract

from arXiv · show

The publication of the Gaia Data Release 2 (Gaia DR2) opens a new era in Astronomy. It includes precise astrometric data (positions, proper motions and parallaxes) for more than $1.3$ billion sources, mostly stars. To analyse such a vast amount of new data, the use of data mining techniques and machine learning algorithms are mandatory. The search for Open Clusters, groups of stars that were born and move together, located in the disk, is a great example for the application of these techniques. Our aim is to develop a method to automatically explore the data space, requiring minimal manual intervention. We explore the performance of a density based clustering algorithm, DBSCAN, to find clusters in the data together with a supervised learning method such as an Artificial Neural Network (ANN) to automatically distinguish between real Open Clusters and statistical clusters. The development and implementation of this method to a $5$-Dimensional space ($l$, $b$, $\varpi$, $\mu_{\alpha^*}$, $\mu_\delta$) to the Tycho-Gaia Astrometric Solution (TGAS) data, and a posterior validation using Gaia DR2 data, lead to the proposal of a set of new nearby Open Clusters. We have developed a method to find OCs in astrometric data, designed to be applied to the full Gaia DR2 archive.

1. Introduction

Gaia’s expanding catalogues make manual analysis of stellar data increasingly difficult, while open-cluster catalogues remain incomplete and disagree on some memberships. The paper develops an automated, multidimensional search using TGAS and checks its candidate clusters with Gaia DR2.

Motivation

  • Gaia DR2 offers precise five-parameter astrometry for more than 1.3 billion stars, making manual-supervision techniques impractical and motivating data-mining and machine-learning methods for automatic exploration and open-cluster detection.Source: §1 · p. 1
  • Open clusters inform studies of star formation, Galactic-disc chemical enrichment, and stellar evolution, but catalogues disagree on which stellar groups count as clusters.The paper notes about 2,500 known clusters and incomplete knowledge beyond 1–2 kpc.Source: §1 · pp. 1–2

Contribution

  • The method searches multiple astrometric dimensions with DBSCAN, classifies candidates using CMD patterns in an ANN, and then validates candidates by hand with Gaia DR2 photometry.The authors apply the method initially to TGAS in preparation for use on the full Gaia DR2 archive.Source: §1 · p. 2

Paper scope

  • The paper tunes the method with simulations, applies it to TGAS to produce 31 candidates, and reports that Gaia DR2 validation confirms most of them.The section outline describes the candidate list and its subsequent manual validation.Source: §1 · p. 2

2. Methods

The method combines preprocessing, five-dimensional DBSCAN clustering, and an ANN classifier to search TGAS for open-cluster candidates. Its ANN is trained on CMD examples, with performance estimated using a held-out test set.

2. Methods

  • DBSCAN groups nearby sources in five-dimensional astrometric space, and an ANN classifies the resulting clusters as real open clusters or statistical detections.The ANN uses 2MASS photometry to construct color-magnitude diagrams (CMDs).Source: §2 · p. 2

2.1. Preprocessing

  • The search is restricted mainly to the Galactic disk, removes extreme proper motions and parallaxes, and divides the sky into shifted regions to manage density variation and border cases.Parameters are standardized so their weights in the five-dimensional distance calculation are equalized.Source: §2.1 · pp. 2–3

2.2. DBSCAN

  • DBSCAN uses a distance across position, parallax, and proper motion, with minPts and ϵ controlling membership; ϵ is chosen automatically from observed and randomized nearest-neighbor distances.The search uses k = minPts−1, and simulations are used to optimize L and minPts.Source: §2.2 · p. 3, Determination of the ϵ and minPts… · p. 3

2.3. Identification of Open Clusters

  • After clustering, CMD isochrone patterns are used to distinguish physical open clusters from random statistical accumulations.The ANN automates a distinction that is otherwise commonly made by visual inspection.Source: §2.3 · p. 3

2.3.1. Artificial Neural Networks

  • The multilayer perceptron passes input features through a hidden layer with weighted inputs and hyperbolic-tangent activation before the output layer classifies them.The figure depicts one hidden layer; its x_i labels denote inputs and h_i labels denote hidden neurons.Source: §2.3.1 · p. 4, Figure 4 · p. 4

2.3.2. Data Preparation

  • Training uses CMDs from known open clusters as positive examples and non-isochrone cluster detections as negatives, represented as density maps.Subsets of known cluster members are randomly selected to create training examples resembling DBSCAN detections.Source: §2.3.2 · p. 4

2.3.3. Performance of the classification

  • 97.95% is the test-set classification precision for the 296 CMD images, but the authors treat it as an upper-limit estimate because training includes the known clusters.The training set contains 19 nearby clusters, ages 40–850 Myr, and no significant differential extinction; broader training data are needed for Gaia DR2 application.Source: §2.3.3 · p. 4

3. Simulations

TGAS-like simulated data were used to test the clustering method and tune its parameters for detecting clusters while limiting false positives. The selected parameter pairs balance noise and efficiency.

3. Simulations

  • The simulation tests the clustering method and parameter choices for detecting as many clusters as possible while minimizing false positives.Source: §3 · p. 4
  • The ANN receives color-coded pixel values from CMD density maps, using Coma Berenices members as positive examples and a DBSCAN cluster labeled noise as a negative example.Source: Figure 5 · p. 5
  • The TGAS-like simulation combines Tycho-2 astrometry with errors generated to approximate TGAS errors and distributions.Nearby-star parallaxes use weighted averages of deconvolved Hipparcos parallaxes; more distant-star parallaxes use photometric values.Source: §3 · p. 5
  • Open clusters are added to the simulated data using Gaia Object Generator values, with observational errors added to their true astrometric parameters.Stars fainter than G = 12 are filtered to match TGAS's limiting magnitude.Source: §3 · p. 5

Choice of the parameters

  • Noise around ∼0.25 adds 1 spurious cluster per 4 real clusters; efficiency ∼0.3 misses one out of four real clusters.The selected ranges are L ∈[12, 16] and minPts ∈[5, 9].Source: Choice of the parameters · p. 5

4. Results

The method produced 31 open-cluster candidates from TGAS, then classified detections with an ANN and manually validated candidates using Gaia DR2 photometry. The candidate list includes overlaps with earlier nearby-cluster discoveries and is subject to incompleteness and possible spurious detections.

4. Results

  • DBSCAN detections overlapping known MWSC clusters within a 2 deg × 2 deg box were discarded; the remaining clusters were classified by an ANN using a greater-than-90% isochrone-confidence threshold.The ANN uses CMDs built from 2MASS and TGAS photometry.Source: §4 · p. 5

4.1. Open Cluster Candidates

  • 31 Open Cluster candidates resulted from applying the algorithms, and their list is sorted by how often each cluster appears across 25 parameter pairs.Nfound records the number of parameter pairs in which a cluster was found.Source: §4.1 · p. 5
  • UBC1 and UBC12 overlap with the previously published RSG4 and RSG3 clusters, respectively, which were excluded from the known-cluster cross-match.Source: §4.1 · p. 5, §4 · p. 5
  • The candidate list is incomplete, particularly at very low Galactic latitudes, where known clusters are denser.The positional 2 deg × 2 deg matching rule also discards candidates near known clusters; proper motions and parallaxes are not required to match.Source: §4.1 · p. 6
  • TGAS scanning-law inhomogeneities may produce spurious clusters: UBC7 shares proper motions and parallax with Collinder 135 but was detected separately near its center.The text notes that UBC7 lies where two scanning-law stripes cross; Collinder 135 was not detected by DBSCAN.Source: §4.1 · p. 6

5. Validation using Gaia DR2

Gaia DR2 astrometry and photometry were used to validate TGAS open-cluster candidates. The authors confirmed about 70% of the proposed candidates and found clearer isochrone patterns with Gaia photometry.

5. Validation using Gaia DR2

  • Gaia DR2's improved astrometric precision and photometry supported candidate validation, while magnitude and parallax cuts helped contrast clusters against field stars and reduce uncertainties.The validation reran the method around each candidate, using a 1- or 2-degree cone search; the cuts also removed distant stars.Source: §5 · p. 6
  • Table 1 lists 31 TGAS open-cluster candidates, which were subsequently checked against Gaia DR2 data.The table reports member means and how often each cluster was found across the explored parameter pairs.Source: Table 1 · p. 7, §5 · p. 7
  • For UBC1, Gaia DR2 members formed a clearer isochrone pattern in the color-magnitude diagram than the same stars plotted with 2MASS photometry.The comparison used the same stars for G ≤ 12.Source: §5 · pp. 6–7
  • ∼70% of the proposed candidates were confirmed, within the expected performance limits from simulations with around 25% and 50% noise.The confirmed fraction was 100% for candidates found at least five times across parameter pairs, versus 59% for candidates found fewer than five times.Source: §5 · p. 7

5.1. Comments on some of the confirmed clusters

The confirmed candidates are generally nearby Galactic-disk clusters, but individual cases show possible associations, uncertain memberships, and distinct groups with similar ages.

5.1.1. UBC1 and UBC12

  • UBC1 and UBC12 correspond to RSG4 and RSG3 at about 330 and 430 pc; RSG4's proper motions differ by 12σ from the reported values.RSG3's proper motions agree rather well with the reported values.Source: §5.1.1 · p. 7

5.1.2. UBC3

  • UBC3 is a sparse cluster at about 1.7 kpc, the farthest confirmed candidate; two available radial velocities disagree, and one star is also discordant in the CMD.Red-clump stars indicate an intermediate-age cluster; one discordant star could be a non-member.Source: §5.1.2 · p. 8

5.1. Comments on some of the confirmed clusters

  • Most confirmed candidates are within 1 kpc and tend to lie at |b| > 5°; UBC3, UBC6, and UBC27 are farther away.UBC3, UBC6, and UBC27 have parallaxes of 0.58 ± 0.04, 0.67 ± 0.01, and 0.88 ± 0.03 mas, respectively.Source: §5.1 · p. 7, Figure 8 · p. 8, Figure 10 · p. 10

5.1.3. UBC4, UBC19 and UBC31

  • UBC19 and UBC31 may be Per OB2 substructures if it spans more than 8°; UBC19's proper motions differ from Alessi Teustch 10, while UBC4 lies about 570 pc away.A deep study over a larger area is needed to assess a possible Per OB2 relationship.Source: §5.1.3 · pp. 8–9

5.1.4. UBC7 and Collinder 135

  • UBC7 and Collinder 135 are distinct, similarly aged groups, with mean parallaxes of 3.563 ± 0.006 and 3.310 ± 0.004 mas and different proper motions.Their overlapping CMD sequences support similar ages, while the parallax difference is judged to be real rather than a regional-systematics artefact.Source: §5.1.4 · p. 9

5.1.5. UBC10

  • UBC10 is sparse in the 1°-radius analysis and may relate to Cep OB2; several subgroups appear in the wider region, requiring broader analysis to confirm or reject the link.Source: §5.1.5 · p. 9

5.1.6. UBC17

  • A 2° search around UBC17 identifies ACCC19, Collinder 170, and σ Ori; the wider Ori OB1 region needs global analysis beyond this paper's scope.Source: §5.1.6 · pp. 9–10

5.1.7. UBC32

  • UBC20's TGAS candidate at about 0.5 mas was not recovered in DR2; instead, DR2 shows a sparse, offset detection at 3.5 mas.Source: §5.1.7 · p. 10

6. Conclusions

The paper presents an automated astrometric open-cluster search and reports 31 new TGAS candidates, most of which were confirmed using Gaia DR2. It also identifies reliability constraints involving the ANN training set and DBSCAN’s distance-dependent density scale.

6. Conclusions

  • The automated system combines DBSCAN in five astrometric dimensions with an ANN that distinguishes real open clusters from spurious statistical clusters using color-magnitude diagrams.It is designed for large datasets, including Gaia DR2, with minimal manual intervention.Source: §6 · p. 10
  • Around 70% of 31 new TGAS candidates were confirmed as open clusters with Gaia DR2, with 100% success for Nfound > 5.The confirmation step also revealed richer structures in some regions for further study.Source: §6 · p. 10
  • TGAS tests support reliable cluster detection, but its scanning-law artefacts are expected to be reduced, not eliminated, in Gaia DR2.TGAS’s bright limiting magnitude also prevented detection of distant, faint clusters that deeper Gaia DR2 data may detect.Source: §6 · p. 10
  • ANN discrimination depends on representative CMD training data, while DBSCAN’s ε must be adapted because cluster compactness varies with distance.The authors call for a wider, more realistic OC CMD training set for Gaia DR2.Source: §6 · p. 10

Appendix A: Color-Magnitude Diagrams of the found OCs

The appendix figures visualize identified cluster members against field stars in sky-position and proper-motion spaces, alongside color-magnitude diagrams of the members.

Appendix A: Color-Magnitude Diagrams of the found OCs

Loading 1805.03045v2…