StarStream: Automatic detection algorithm for stellar streams

TL;DR

StarStream addresses the need for automated and scalable detection of stellar streams beyond visually guided methods. It combines a physics-informed particle-spray stream model with kernel density estimation to form a minimal parametric mixture model that separates stream members from background using a single free parameter, the stream fraction $f_s$, and a likelihood-based membership probability. Validation on Gaia DR3-tailored mocks shows robust performance with purity and completeness around 65% at $|b|>30^{\circ}$, improving detection efficiency and enabling broader discovery of streams, even when extinction and background contamination are non-negligible. The method is fast, adaptable to other surveys, and benefits from additional spectroscopic and deeper photometric data, offering practical impact for mapping GC streams and constraining Galactic potential and dynamical history.

Abstract

The Gaia mission has led to the discovery of over 100 stellar streams in the Milky Way, most of which likely originated from globular clusters (GCs). As the upcoming wide-field surveys can potentially continue to increase the number of known streams, there is a growing need to shift focus from manual detection of individual streams to automated detection methods that prioritize both quality and quantity. Traditional techniques rely heavily on the visual expectation that GC streams are dynamically cold and thin. This assumption does not hold for all streams, whose morphologies and kinematics can vary significantly with the progenitor's mass and orbit. As a result, these methods are biased toward a subset of the whole stream population, with often unquantified purity and completeness. In this work, we present StarStream, an automatic stream detection algorithm based on a physics-inspired model rather than visual expectation. Our method provides a more accurate prediction of stream stars in the multi-dimensional space of observables, while using fewer free parameters to account for the diversity of streams. Applied to a mock GC stream catalog tailored for the Gaia DR3 dataset, our algorithm achieves both purity and completeness of at least 65% at Galactic latitudes |b| > 30 degree.

Figures (8)

• Figure 1: Demonstration of a test of the method on a mock stream. Top row: distributions of simulated tracer particles (magenta) and background stars (black) in position space (left), proper motion space (middle), and color--magnitude space (right). We plot the stream PDF $p_{\rm s}({\bm x})$ from Gaussian KDE as gray contours. The three contours from dark to light represent three values of $\ln p_{\rm s}({\bm x})-\ln p_{\rm s,max}({\bm x})=-0.5, -2, -4.5$, corresponding to the 1-$\sigma$, 2-$\sigma$, and 3-$\sigma$ ranges of the standard Gaussian distribution, respectively. We only show $10^4$ background stars randomly chosen from the total of $4\times 10^6$ for visual clarity. Bottom row: application of the method on this mock stream using $P_{\rm th}=0.4$. We show stars that are false positives (red circle), missed by the method (blue open circle), or correctly detected (blue solid circle) in the same subspaces as the top row. The mock stream members already have uncertainties added and are mixed with background stars from real Gaia DR3 data. We also plot the contours of $\ln p_{\rm s}({\bm x})-\ln p_{\rm s,max}({\bm x})=-4.5$ for comparison. For the left column of both rows, we show the location of the progenitor GC as the circle of tidal radius. The velocity of the GC is represented by the arrow.
• Figure 2: Detection ratio (upper panel) and completeness/purity (lower panel) of detected stream members as functions of probability threshold. The solid lines stand for the median values among all test streams, while the shaded ranges show the interquartile ranges. Detection ratio $=1$ is highlighted as the dot-dashed line. We also show the median detection ratio, completeness, and purity for streams with progenitor GCs at low Galactic latitude ($|b|<30^\circ$, dotted curves) and high Galactic latitude ($|b|>30^\circ$, dashed curves) separately. Since the purity at $P_{\rm th}=1$ is not well defined, we extrapolate the values at $P_{\rm th}=0.99$ out to 1 for visual clarity.
• Figure 3: Detection ratio (upper panel) and completeness/purity (lower panel) of detected stream members as functions of the progenitor GC's absolute Galactic latitude $|b|$ with $P_{\rm th}=0.5$. Individual streams are shown as black circles (detection ratio), blue diamonds (completeness), and red triangles (purity). The solid lines stand for the median values among all test streams, while the shaded ranges show the interquartile ranges. Detection ratio $=1$ is highlighted as the dot-dashed line. We calculate the percentiles at any $b$ using nearby streams smoothed by the Gaussian kernel, with bandwidth varying linearly from $5^\circ$ to $15^\circ$ from the Galactic plane to the poles.
• Figure 4: Number of detected stars in the null test $N_{\rm null}$ (red open circles) as a function of the progenitor GC's absolute Galactic latitude $|b|$. For comparison, we also show the true number of stream stars $N_{\rm true}$ (blue solid diamonds) and the number of detections $N_{\rm detect}$ (black solid circles) without removing the signal. Similarly to Fig. \ref{['fig:fdetect_completeness_purity_vs_b']}, we show the the median values as solid lines, while the shaded ranges represent the interquartile ranges.
• Figure 5: Same as Fig. \ref{['fig:fdetect_completeness_purity_vs_b']}, but with the alternative isochrone model $\texttt{PARSEC}$ (left column) and three alternative Galactic potential models (right column): 1) the base potential model scaled down by $20\%$ (dotted curves), 2) the base potential model scaled up by $20\%$ (dashed curves), and 3) the MilkyWayPotential2022 model from gala. We show the interquartile ranges for the base case for comparison.