🔭 astrophysics

An Enhanced Catalog of Gaia DR3 Galaxy Candidates with Spectroscopic and Machine-Learning Photometric Redshifts

This paper presents an enhanced Gaia DR3 galaxy catalog that significantly improves its scientific utility for cosmological studies by cross-matching with spectroscopic surveys and applying a Multiple-Bin Regression Neural Network to provide accurate redshift estimates for the entire sample.

Original authors: Junghyun Hwang, Ho Seong Hwang

Published 2026-07-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Junghyun Hwang, Ho Seong Hwang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the universe as a giant, cosmic library. For centuries, astronomers have been trying to catalog every book on the shelves, but the books are incredibly far away, and most of them don't have titles or authors listed on their spines. In astronomy, that "title" is called redshift. It's a measure of how much the light from a galaxy has been stretched as it travels through the expanding universe. Knowing the redshift is like knowing exactly how far away a galaxy is and how fast it's moving away from us. Without it, a galaxy is just a blurry smudge of light; with it, the galaxy becomes a specific point in space and time, allowing scientists to map the structure of the cosmos and understand how the universe evolved.

For a long time, getting these "titles" was like trying to read a book by holding it up to a single, dim lightbulb. You could get a rough idea, but it was hard to be sure. Then, a super-precise space telescope named Gaia started taking pictures of billions of stars and galaxies, creating a massive list of potential galaxies. But there was a catch: while Gaia was amazing at spotting them, it didn't always know their redshifts, and sometimes it accidentally included things that weren't galaxies at all, like nearby stars or active black holes. This left astronomers with a huge list of candidates that were exciting but scientifically tricky to use. They needed a way to clean up the list and figure out the distances for the millions of galaxies that didn't have a "title" yet.

This is where the new study by Junghyun Hwang and Ho Seong Hwang comes in. Think of their work as a massive, high-tech detective operation designed to upgrade the Gaia library. The team took the original list of 4.8 million potential galaxies from Gaia's third data release and gave them a serious makeover. First, they cross-referenced the list with other massive databases of galaxies that did already have their redshifts measured (like DESI, SDSS, and 2MRS). This was like checking a master directory to see if any of the mystery books had already been cataloged by other librarians. They successfully found matches for about 24.5% of the sources, instantly giving them precise, "spectroscopic" redshifts.

But what about the remaining 75% of the galaxies that didn't have a match? For these, the team built a machine-learning robot named MBRNN. Imagine this robot as a super-smart art critic that has studied thousands of galaxies with known distances. The robot was trained to look at the "colors" of the light coming from the mystery galaxies (combining data from Gaia with infrared data from other telescopes like 2MASS and unWISE) and guess their redshifts based on patterns it had learned. It's like the robot saying, "This galaxy looks a lot like the ones I studied that are 1 billion light-years away, so I'm pretty sure this one is too." The result is a new, enhanced catalog where every single galaxy has a redshift estimate, either the precise kind from the cross-match or the highly accurate "guess" from the robot.

The team didn't stop at just giving numbers; they also built a "trust score" for every galaxy. Since the robot is making a guess, it needs to tell you how confident it is. The researchers created a system to spot "outliers"—galaxies that look weird or don't fit the pattern. For example, if a source looks like a star but the robot thinks it's a galaxy, the system flags it. They assigned a probability score, $P_0$ , to each source. A low score means the source is almost certainly a real, distant galaxy, while a high score suggests it might be a contaminant (like a nearby star) or a galaxy where the redshift guess is likely wrong.

In their testing, the new machine-learning model proved to be more accurate and precise than previous methods that relied only on Gaia's data. They found that by adding extra data from infrared telescopes, the robot could see the "shape" of the light better, leading to fewer mistakes. While the model isn't perfect—it still struggles a bit with very distant galaxies or sources that look like stars—it significantly improves the quality of the data. The final catalog provides a reliable, all-sky map of galaxies with distances and trust scores, turning a messy list of candidates into a robust tool for cosmologists to study the universe's history. The authors suggest that this enhanced dataset will be a valuable resource for future large-scale studies, helping scientists filter out the noise and focus on the real cosmic stories hidden in the light.

Technical Summary: An Enhanced Catalog of Gaia DR3 Galaxy Candidates with Spectroscopic and Machine-Learning Photometric Redshifts

Problem Statement
The Gaia Data Release 3 (DR3) "galaxy_candidates" table provides an unprecedented all-sky sample of approximately 4.8 million potential extragalactic sources. However, the scientific utility of this catalog for cosmological and extragalactic studies is significantly constrained by two primary limitations: the absence of redshift information for the majority of sources and a relatively low source purity (estimated at 50–70%). The original selection prioritized completeness over purity, and without redshift data, distance estimation and statistical analysis are impossible for most entries.

Methodology
To address these limitations, the authors constructed an enhanced catalog by integrating spectroscopic redshifts ( $z_{spec}$ ) and machine-learning-based photometric redshifts ( $z_{phot}$ ).

Spectroscopic Redshift Integration:
The Gaia DR3 galaxy candidates were cross-matched against four external spectroscopic redshift catalogs: DESI DR1, SDSS DR17, 2MRS, and NED. Matching criteria were based on angular separation thresholds specific to each catalog. In cases of multiple matches, a priority order was enforced (DESI > SDSS > 2MRS > NED). This process assigned spectroscopic redshifts to 24.5% of the total catalog.
Photometric Redshift Estimation (MBRNN):
For sources lacking spectroscopic measurements, photometric redshifts were estimated using a Multiple-Bin Regression Neural Network (MBRNN).
- Input Features: The model utilized a multiwavelength dataset combining Gaia DR3 photometry ( $G, G_{BP}, G_{RP}$ ), 2MASS XSC (near-infrared $J, H, K_s$ ), and unWISE (mid-infrared $W1, W2$ ). To address the limited spectral coverage of Gaia alone, the input set was expanded to 24 features, including Sérsic magnitudes (as extended-source counterparts to PSF magnitudes), shape parameters (ellipticity, Sérsic index), QSO classification flags, and Galactic extinction values ( $E(B-V)$ ).
- Model Architecture: The MBRNN is a feed-forward multi-layer perceptron (MLP) with nine fully connected layers. It predicts the probability distribution of a source's redshift across 64 predefined bins (51 bins spanning $-0.082 \le z \le 1$ and 13 bins covering $1 \le z \le 7$ ).
- Training: The model was trained on the spectroscopic subsample (80% training, 10% validation, 10% test) using a logistic regression loss function with the Adam optimizer.
Outlier Scoring and Reliability:
A probabilistic outlier score ( $P_0$ ) was developed using Bayesian inference to quantify the risk of a source being a statistical outlier (either a Galactic contaminant or a catastrophic redshift failure).
- Sources were categorized into subsets: Set A ( $z_{spec} < 0.0017$ , likely Galactic stars), Set B ( $z_{phot}$ overestimation), and Set C ( $z_{phot}$ underestimation).
- Likelihood distributions were modeled in color-color and color-magnitude spaces ( $G$ vs. $W1-W2$ , $G-RP$ vs. $G$ , etc.) using 2D Gaussian fits.
- The posterior probability $P(outlier|x)$ was calculated to assign a reliability score to each source.

Key Results

Catalog Composition: The final catalog contains 4,842,342 sources. Spectroscopic redshifts are available for 1,184,906 sources (24.5%), while photometric redshifts are provided for the remaining 74.4% (3,601,132 sources). Approximately 1.2% of sources lack sufficient photometric data for $P_0$ calculation.
Performance Metrics: On the test set, the MBRNN model achieved a bias of 0.0014, a Mean Absolute Difference (MAD) of 0.0192, and a Normalized Median Absolute Deviation (NMAD) of 0.0021. The standard deviation ( $\sigma$ ) of residuals was 0.0739.
Feature Analysis: An ablation study confirmed that all 24 input features contribute to model performance; removing any single feature or reducing the set to 15 features resulted in equal or degraded performance metrics.
Outlier Identification: The outlier analysis successfully identified Galactic contaminants (Set A) and sources with large redshift errors (Sets B and C). The $P_0$ score effectively distinguishes reliable extragalactic sources from contaminants, with lower scores indicating higher reliability.

Significance and Claims
The authors claim that this work significantly enhances the scientific value of the Gaia DR3 galaxy candidates sample by providing a complete redshift dataset for nearly the entire catalog. The integration of multiwavelength photometry with the MBRNN model yields improved redshift accuracy and precision compared to previous estimates based solely on Gaia data.

Furthermore, the inclusion of the $P_0$ reliability score adds a critical layer of quality control, allowing researchers to filter out Galactic contaminants and sources with unreliable redshifts. The resulting catalog is presented as a robust resource for large-scale extragalactic and cosmological studies, enabling distance-dependent analyses that were previously limited by the lack of redshift information in the Gaia DR3 release. The authors also note that the low $P_0$ scores facilitated the identification of new, previously unconfirmed galaxies, demonstrating the catalog's utility for discovery.

More like this