Finding Quasars behind the Galactic Plane. IV. Candidate Selection from Chandra with Random Forest

TL;DR

This work tackles the scarcity of quasars behind the Galactic plane by leveraging Chandra CSC 2.1 X-ray sources cross-matched with Gaia DR3 and CatWISE2020, and applying a Random Forest classifier to separate quasars from Galactic contaminants. It uses a carefully constructed training set and Gaia-based proper-motion filtering to identify 6,286 quasar candidates, including 863 GPQ candidates and 514 high-confidence GPQs, extending to fainter fluxes than previous GPQ catalogs. Photometric redshifts are estimated with an RF regressor that combines optical, infrared, and X-ray features, enabling population studies even in regions with limited spectroscopy. Pilot spectroscopy confirms two GPQs at redshifts $z=1.2582$ and $z=1.1313$, validating the method and highlighting GPQs as valuable tools for astrometric reference frames and Milky Way absorption studies.

Abstract

Quasar samples remain severely incomplete at low Galactic latitudes because of strong extinction and source confusion. We conduct a systematic search for quasars behind the Galactic plane using X-ray sources from the Chandra Source Catalog (CSC 2.1), combined with optical data from Gaia DR3 and mid-infrared data from CatWISE2020. Using spectroscopically confirmed quasars and stars from data sets including DESI, SDSS, and LAMOST, we apply a Random Forest classifier to identify quasar candidates, with stellar contaminants suppressed using Gaia proper-motion constraints. Photometric redshifts are estimated for the candidates using a Random Forest regression model. Applying this framework to previously unclassified CSC sources, we identify 6286 quasar candidates, including 863 Galactic Plane Quasar (GPQ) candidates at |b|<20°, of which 514 are high-confidence candidates. Relative to the previously known GPQ sample, our selected GPQs reach fainter optical and X-ray fluxes, improving sensitivity to low-flux GPQs. In addition, both the GPQ candidates and known GPQs display harder X-ray spectra than the all-sky quasar sample, consistent with increased absorption through the Galactic plane. Pilot spectroscopy confirms two high-confidence GPQ candidates as quasars at spectroscopic redshifts of z=1.2582 and z=1.1313, and further spectroscopic follow-up of the GPQ sample is underway. This work substantially improves the census of GPQs and provides a valuable target sample for future spectroscopic follow-up, enabling the use of GPQs to refine the reference frames for astrometry and probe the Milky Way interstellar and circumgalactic media with the absorption features of GPQs.

Figures (8)

• Figure 1: Normalized confusion matrix of the RF classifier computed on the validation set. The matrix is color-coded by the number of sources in each cell. Diagonal entries show the fraction of correctly classified objects (i.e., recall or completeness) for each class, while off-diagonal entries indicate the misclassification rates.
• Figure 2: Distribution of RF-based photometric redshifts for the quasar candidates. The red histogram shows the full-sky candidate sample, and the blue histogram shows candidates at low Galactic latitude ($|b|<20^\circ$). The ordinate shows source counts (logarithmic scale).
• Figure 3: Distribution of the all-sky candidates in the $G$ versus W1-W2 diagram as classified by the RF model. Red and blue points denote quasar and stellar candidates, respectively; the top and right panels show the corresponding marginal distributions in $G$ and W1-W2. The $G$ magnitudes shown on the horizontal axis are corrected for Galactic extinction. We apply a uniform magnitude filter to all candidates, excluding sources with observed $G>21$ mag (prior to extinction correction) and removing very bright objects with $G<10$ mag.
• Figure 4: Distributions of $\log(f_{\mathrm{PM}0})$, the probability density at zero proper motion derived from Gaia astrometry (see Eq. \ref{['eq:fpm0']}), for spectroscopically confirmed quasars (green), spectroscopically confirmed stars (yellow), and the GPQ candidates (white). The vertical dashed line marks the adopted threshold $\log(f_{\mathrm{PM}0})=-4$ used to suppress stellar contaminants; note that $f_{\mathrm{PM}0}$ is a probability density (not a probability) and therefore can exceed unity.
• Figure 5: Spatial distribution of GPQ candidates ($|b|<20^\circ$) in Galactic coordinates (Mollweide projection). Points are color-coded by the RF-derived quasar membership probability, $P_{\rm QSO}$ (color bar).