🔭 astrophysics

Constructing a Mock Galaxy Catalog for the All-sky SPECtroscopic Survey of Nearby Galaxies (A-SPEC) Using the Machine-assisted Semi-Simulation Model

This paper presents a machine-assisted semi-simulation model trained on IllustrisTNG to construct a mock galaxy catalog for the A-SPEC survey, achieving high prediction accuracy for baryonic properties and successfully reproducing luminosity-dependent clustering.

Original authors: Dongkok Kim, Yongseok Jo, Ho Seong Hwang, Ji-hoon Kim, Juhan Kim, Jaehyun Lee, Hyeonguk Bahk, Young-Man Choi, Moo-Young Chun, Sang-Hyun Chun, Haeun Chung, Kim Dachan, Sungwook E. Hong, Minhee Hyun, Do

Published 2026-07-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dongkok Kim, Yongseok Jo, Ho Seong Hwang, Ji-hoon Kim, Juhan Kim, Jaehyun Lee, Hyeonguk Bahk, Young-Man Choi, Moo-Young Chun, Sang-Hyun Chun, Haeun Chung, Kim Dachan, Sungwook E. Hong, Minhee Hyun, Donghui Jeong, Jae-Woo Kim, Kang-Min Kim, Yunjong Kim, Jongwan Ko, Minseong Kwon, Ho-Gyu Lee, Jong Chul Lee, Yongseok Lee, Hyunho Lim, Heeyoung Oh, Changbom Park, Hyunmi Song, Mingyeong Yang, Yongmin Yoon

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

The Big Picture: Building a "Cosmic Simulator" for a New Telescope

Imagine astronomers are about to launch a massive new telescope survey called A-SPEC. Its job is to take pictures and spectra (fingerprints of light) of millions of nearby galaxies across the entire sky. Before they turn the telescope on, they need to know what to expect. They need a "test drive" to see if their plans work.

To do this, the team built a Mock Galaxy Catalog. Think of this not as a real photo, but as a highly detailed, computer-generated "fake universe." It contains millions of fake galaxies that look, act, and cluster together just like the real ones they hope to find.

The paper explains how they built this fake universe using a clever mix of physics simulations and machine learning.

1. The Problem: The "Ghost" in the Machine

To make a realistic fake galaxy, you usually need to simulate the physics of gas, stars, and black holes. This is like trying to simulate a thunderstorm in a computer: you need to track every drop of rain, every gust of wind, and the temperature.

The Issue: Doing this for the entire sky is too expensive. It would take a supercomputer thousands of years to run.
The Shortcut: The team used a "Dark Matter Only" simulation. Imagine a skeleton of the universe made of invisible, heavy stuff (Dark Matter). It's easy to simulate because you only have to track gravity. But a skeleton without flesh (stars and gas) isn't a galaxy.

The Solution: They needed a way to put "flesh" on the "bones." They needed a way to guess what the stars and gas look like just by looking at the invisible Dark Matter skeleton.

2. The Teacher: Learning from a "Perfect" Universe

To teach the computer how to guess, they used a "teacher" simulation called IllustrisTNG.

The Teacher: This is a super-detailed simulation that does include all the gas, stars, and physics. It's like a high-resolution movie of a galaxy forming.
The Lesson: The team looked at the Teacher's movie and asked: "If I only show you the Dark Matter skeleton, can you guess what the stars and gas look like?"

They trained a Machine Learning Model (a type of AI) on this data. The AI learned the secret recipes connecting the invisible skeleton to the visible galaxy.

3. The New Ingredients: Adding "Spice" to the Recipe

The authors didn't just use the old recipe; they added new ingredients to make the AI smarter. They realized that just knowing the mass of a galaxy isn't enough. They added:

Shape and Spin: How twisted or stretched the galaxy is (like a spinning top).
The Neighborhood: Who lives next door? Is the galaxy in a crowded city (a cluster) or a lonely farm (a void)?
History: How fast was the galaxy eating up new matter recently?

The Result: With these new ingredients, the AI became very good at guessing the Stellar Mass (how much star stuff there is) and Gas Mass. It got about 96% accurate for stars and 90% for gas. However, it was still a bit shaky at guessing exactly how fast stars are being born (Star Formation Rate) or how "metal-rich" the gas is.

4. The Construction Site: The "Zoom-In" Universe

Now that they had a smart AI, they needed a place to apply it. They couldn't just use the small "Teacher" universe because it wasn't big enough to cover the whole sky.

They built a custom N-body simulation (a giant Dark Matter skeleton) called NASIM.

The Analogy: Imagine a map of the world. Usually, a map has the same level of detail everywhere. But this map is special.
- The "center" of the map (where our local universe is) is zoomed in super high-definition.
- As you move further away, the map zooms out, becoming less detailed.
Why? The telescope survey is "flux-limited," meaning it sees bright, close galaxies clearly, but faint, far-away galaxies only if they are huge. The simulation mimics this by being super sharp nearby and slightly blurrier far away. This saved them massive amounts of computing power.

5. The Final Product: The "Fake" Catalog

They took their giant, custom-built Dark Matter skeleton and ran their AI over it. The AI "painted" baryonic properties (stars, gas, metallicity) onto every Dark Matter clump.

Did it work?

Yes, mostly. When they compared their fake galaxies to real observations, the fake ones matched the real ones in two key ways:
1. Brightness: The fake galaxies had the right mix of bright and dim stars.
2. Clustering: The fake galaxies hung out in groups and clusters just like the real ones do.

The Catch: The AI still struggles a bit with the "extremes." It's good at the average galaxy, but it sometimes misses the very quiet, gas-poor galaxies or the super-active star-bursting ones. It's like a weather forecaster who is great at predicting "sunny with a chance of rain" but bad at predicting a sudden tornado.

Summary

The paper describes building a virtual reality universe for a new telescope survey.

They trained an AI on a high-detail physics simulation.
They gave the AI new clues (like shape and neighborhood) to make it smarter.
They built a custom-sized universe that is detailed where the telescope looks and less detailed where it doesn't need to be.
They used the AI to paint stars and gas onto this universe, creating a "Mock Catalog."

This catalog is now a free resource for astronomers to test their theories and prepare for the real data that A-SPEC will collect. It's a "flight simulator" for the cosmos.

Technical Summary: Constructing a Mock Galaxy Catalog for A-SPEC Using the Machine-assisted Semi-Simulation Model

Problem Statement
The All-sky SPECtroscopic survey of nearby galaxies (A-SPEC) aims to compile physical properties and redshifts for galaxies brighter than $K_S = 13.75$ across the entire sky. While current surveys increasingly target high-redshift universes, the local Universe (within hundreds of megaparsecs) remains incompletely mapped. Accurate mock galaxy catalogs are essential for validating survey designs, interpreting observational data, and estimating covariances between cosmological parameters. However, generating realistic mocks that reproduce both the statistical clustering and diverse baryonic properties (e.g., stellar mass, gas mass, star formation rate) of target galaxies is challenging. Purely data-driven methods often lack individual galaxy properties, while full hydrodynamical simulations covering the necessary survey volume ( $\sim 2 \text{ Gpc}^3$ ) are computationally prohibitive. Furthermore, existing machine learning (ML) approaches often struggle to accurately predict gas-related properties and star formation rates due to the complex, degenerate nature of astrophysical processes.

Methodology
The authors employ a Machine-assisted Semi-Simulation Model (MSSM) to bridge the gap between dark-matter-only (N-body) simulations and baryonic physics. The methodology involves three primary stages:

Training Data Generation: The model is trained on the IllustrisTNG cosmological magnetohydrodynamical simulation suite, specifically utilizing the high-resolution TNG100-1 run. To address the scarcity of massive halos in TNG100-1, subhalos from the lower-resolution TNG300-1 are included and scaled using constant factors derived from the high-resolution run. The training sample is restricted to subhalos with sufficient particle counts ( $N_{DM} \ge 40, N_{star} \ge 40$ ) and non-zero gas properties.
Machine Learning Architecture: The MSSM utilizes an Extremely Randomized Trees (ERT) regressor. The model operates in a two-stage learning process:
- First, it predicts stellar magnitudes in eight photometric bands ( $U, B, V, K, u, g, r, z$ ).
- Second, these predicted magnitudes are added to the input features to predict remaining baryonic properties: stellar mass ( $M_{star}$ ), gas mass ( $M_{gas}$ ), star formation rate (SFR), and gas metallicity ( $Z_{gas}$ ).
- Input Features: The model ingests dark matter (DM) subhalo properties (mass, circular velocity, velocity dispersion, spin), environmental parameters (local density, neighbor counts, tidal anisotropy), and subhalo anisotropy parameters (Bullock spin, position/velocity offsets, axis ratios). Crucially, the authors introduce particle-based environment measures (density contrast $\delta_R$ and tidal anisotropy $\alpha_R$ ) and a pseudomass accretion rate (pMAR) derived from tracing the most bound particle, compensating for the lack of full merger tree information in the target N-body simulation.
Application to N-body Simulation: A custom N-body simulation (NASIM) is constructed using a zoomed-in approach with four nested resolution levels. This design ensures sufficient resolution to resolve subhalos down to the mass limit required by the A-SPEC flux limit ( $K_S \le 13.75$ ) across the redshift range $z \approx 0$ to $0.2$, while maintaining a large survey volume. The MSSM is applied to the subhalo catalog of NASIM to "paint" baryonic properties onto the DM subhalos. To mitigate resolution mismatches between the hydrodynamic training data and the N-body application, the authors apply mass-dependent scaling to specific structural inputs ( $v_{circ}$ and $\sigma_v$ ).

Key Contributions and Results

Improved Predictive Accuracy: By introducing subhalo anisotropy parameters and particle-based environmental features, the authors improved the model's performance. The final model achieves coefficients of determination ( $R^2$ $R^{2}$ ) of:
- $0.96$ for Stellar Mass ( $M_{star}$ )
- $0.90$ for Gas Mass ( $M_{gas}$ )
- $0.70$ for Star Formation Rate (SFR)
- $0.79$ for Gas Metallicity ( $Z_{gas}$ )
  The model successfully reproduces the luminosity-dependent clustering of target galaxies when tuned to match number density.
Feature Importance Analysis: Using SHapley Additive exPlanations (SHAP), the study quantifies feature contributions. While $M_{subhalo}$ , $v_{circ}$ , and $\sigma_v$ dominate predictions for stellar properties, environmental features contribute significantly (20–40%) to gas-related properties. Specifically, density contrast ( $\delta$ ) measured at $z=0.2$ within a window proportional to the subhalo size was found to be the most influential environmental metric for predicting gas mass.
Handling Unresolved Populations: To address the bias introduced by excluding subhalos with low gas/SFR (due to resolution limits), the authors implemented a pseudo-random assignment strategy for unresolved baryonic properties based on scaling relations with stellar components. This approach qualitatively improved the recovery of the distribution functions for SFR and gas mass, particularly at the low-end tails.
Mock Catalog Construction: The resulting mock catalog reproduces the observed $K_S$ -band luminosity function and the two-point correlation functions of the target galaxies. The authors note that while stellar properties are well-recovered, gas-related properties (SFR and metallicity) still exhibit systematic deviations at the distribution tails, indicating that current DM features do not fully encode the physics of star formation and metal enrichment.

Significance and Claims
The paper claims to provide a reference mock galaxy catalog for the A-SPEC survey that includes a comprehensive set of predicted baryonic properties. The significance of this work lies in:

Validation Tool: The catalog serves as a critical testbed for the A-SPEC survey design, allowing for the estimation of parameter covariances and the validation of selection functions without the prohibitive cost of full hydrodynamical simulations over the survey volume.
Methodological Advancement: The study demonstrates that machine learning models trained on high-resolution hydrodynamical simulations can be successfully transferred to lower-resolution N-body simulations with appropriate feature engineering (e.g., anisotropy parameters, particle-based environments) and input scaling.
Resource Release: The authors release the mock catalog, which will serve as a foundational resource for upcoming A-SPEC science programs, including the construction of complete catalogs of nearby galaxy groups and clusters.

The authors remain modest regarding the model's limitations, acknowledging that the predictive performance for SFR and metallicity is not yet sufficient to fully capture the diversity of starburst and quiescent populations, and that future improvements will require better handling of merger histories and more precise scaling of subhalo properties across different simulation resolutions.