Abstract
Galaxy groups are essential for studying the distribution of matter on a large scale in redshift surveys and for deciphering the link between galaxy traits and their associated halos. In this work, we propose a widely applicable method for identifying groups through machine learning techniques in real space taking into account the impact of redshift distortion. Our methodology involves two neural networks: one is a classification model for identifying central galaxy groups, and the other is a regression model for predicting the mass of these groups. Both models input observable galaxy traits, allowing future applicability to real survey data. Testing on simulated datasets indicates our method accurately identifies over $92\%$ of groups with $\mathrm{M}_{vir} \geq 10^{11}h^{-1}\mathrm{M}_\odot$, with $80\%$ achieving a membership completeness of at least $80\%$. The predicted group masses vary by less than 0.3 dex across different mass scales, even in the absence of a priori data. Our network adapts seamlessly to expand to sparse samples with a flux limit of $m_{r} < 14$, to high redshift samples at $z=1.08$, and to galaxy samples from the TNG300 hydrodynamical simulation without further training. Furthermore, the framework can easily adjust to real surveys by training on redshift distorted samples without needing parameter changes. Careful consideration of different observational effects in redshift space makes it promising that this method will be applicable to real galaxy surveys.
Introduction
In modern cosmology, galaxies are understood to form within massive dark matter halos. Identifying these "galaxy groups"—collections of galaxies residing in a single halo—is critical for mapping large-scale structure and understanding galaxy evolution. Traditional methods like "Friends-of-Friends" (FoF) often suffer from geometric limitations, leading to "fracturing" (splitting groups) or "fusing" (merging distinct groups). Moreover, observations are complicated by Redshift Space Distortions (RSD), where galaxy velocities create "fingers of god" effects that distort spatial groupings. This paper introduces a machine-learning framework designed to overcome these hurdles using a dual-neural-network approach.
Methodology
The proposed framework utilizes two Artificial Neural Networks (ANNs) coupled with an iterative refinement process:
- Central Galaxy Identifier: A classification network that evaluates a target galaxy and its 10 nearest neighbors to determine which is the most likely "central" galaxy. It leverages observable traits such as r-band magnitude, color, and projected distances.
- Group Mass Estimator: A regression network that predicts the virial mass ($M_{vir}$) of identified group candidates based on the stellar mass of the central galaxy, total group stellar mass, and properties of the most massive satellites.
- Iterative Merging: To address fracturing, the system uses the predicted mass to calculate a characteristic radius ($R_{200}$). Galaxies or candidates within this radius are merged, and the process repeats until the catalog stabilizes.
Key Results and Performance
The model was primarily trained on the Millennium Simulation and demonstrated exceptional performance across various benchmarks. It identifies over 92% of groups with $M_{vir} \geq 10^{11} h^{-1} M_\odot$.
The "purity" of the identified groups—the fraction that correctly matches a real dark matter halo—consistently exceeds 95%. Furthermore, the mass estimation is accurate to within 0.3 dex across the entire mass range.
Visualizing Group Identification
The effectiveness of the algorithm in mapping galaxies to their parent halos can be seen in spatial projections. The iterative merging strategy successfully clusters galaxies that might otherwise be split by simpler geometric algorithms.
Generalizability and Future Outlook
A standout feature of this framework is its cross-simulation generalizability. Despite being trained on the semi-analytic Millennium Simulation, it performed excellently on the hydrodynamical TNG300 simulation without retraining. This suggests the networks are capturing fundamental galaxy-halo physics. The model also proved resilient to Redshift Space Distortions by incorporating a margin in line-of-sight distance during neighbor identification.
While the model shows a slight bias in underestimating the mass of low-richness groups and overestimating the most massive clusters due to sparse training data at extremes, its overall robustness is promising. Future work will focus on applying this method to real-world observational data from surveys like SDSS and DESI, where complex selection effects must be managed.
Summary updated