🔭 astrophysics

Machine Learning applications to Galaxy Clusters

This chapter reviews how artificial intelligence techniques are transforming galaxy cluster research by enabling more accurate mass estimation, dynamic state characterization, and baryonic physics emulation across diverse observational data, while emphasizing the critical role of simulations, uncertainty quantification, and interpretability for future precision cosmology.

Original authors: Gustavo Yepes, Daniel de Andrés

Published 2026-05-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gustavo Yepes, Daniel de Andrés

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 construction site. The biggest buildings on this site are galaxy clusters. These aren't just piles of stars; they are massive, gravity-bound cities containing thousands of galaxies, clouds of super-hot gas, and a huge amount of invisible "dark matter" that holds everything together.

Scientists want to know exactly how heavy these cosmic cities are because their weight tells us about the history and future of the entire universe. However, you can't put a galaxy cluster on a scale. Instead, astronomers have to guess the weight by looking at how the different parts of the city behave.

This paper is a review of how scientists are using Artificial Intelligence (AI) to become much better at guessing these weights, moving away from old, rigid rules to a smarter, more flexible approach.

Here is a breakdown of the paper's main ideas using simple analogies:

1. The Old Way: The "Rigid Rulebook"

For a long time, scientists tried to weigh galaxy clusters using a "rulebook" based on simplified physics.

The Analogy: Imagine trying to guess the weight of a messy pile of laundry just by measuring the size of the pile. You might use a formula like: "If the pile is 2 feet wide, it weighs 5 pounds."
The Problem: Real laundry piles are messy. Some have heavy coats, some have light socks, and some are just wet. The old rules assumed everything was perfectly round and calm (like a smooth ball of laundry). But galaxy clusters are messy, lumpy, and often crashing into each other. The old rules often got the weight wrong because they couldn't handle the messiness.

2. The New Way: The "Smart Detective" (Machine Learning)

The paper explains that researchers are now training AI models (specifically Deep Learning) to act like expert detectives. Instead of following a simple rulebook, the AI looks at the entire picture to make a guess.

How it learns: The AI is fed thousands of "fake" galaxy clusters created by supercomputer simulations. It studies these fake clusters, learning that a lumpy shape here or a specific gas pattern there usually means the cluster is heavy or light.
The Advantage: The AI doesn't just look at the total size; it looks at the shape, the texture, and how the different parts (gas, stars, dark matter) interact. It's like the AI realizing, "Ah, this pile of laundry has a heavy winter coat hidden in the corner, so even though it looks small, it's actually very heavy."

3. The Different "Clues" the AI Uses

The paper details how the AI uses different types of "clues" (observational data) to weigh the clusters:

The Sunyaev-Zel'dovich (SZ) Effect: This is like looking at how the cluster distorts the background light of the universe (the Cosmic Microwave Background). The AI looks at the entire map of this distortion, not just the center, to guess the mass.
X-ray Emissions: The gas in clusters is so hot it glows in X-rays. The AI looks at the images of this glowing gas. It found that the gas on the edges of the cluster is actually a better clue for the total weight than the messy center.
Galaxy Movements: By watching how fast the galaxies inside the cluster are moving, the AI can infer the gravity holding them. The AI is better at this than old methods because it understands that the galaxies might be moving in complex, non-spherical ways.
Combining Clues: The best results come when the AI looks at all the clues at once (gas, light, movement). It's like a detective using fingerprints, DNA, and a witness statement together to solve a case, rather than relying on just one.

4. The "Training School" Problem

There is a catch. The AI is only as good as the "school" it went to.

The Analogy: The AI learns from computer simulations. If the simulation teaches the AI that "all laundry piles are dry," but real laundry is often wet, the AI will make mistakes when looking at real clusters.
The Paper's Point: The authors emphasize that we need better simulations (like the "THE THREE HUNDRED" project mentioned) that accurately model the messy physics of gas and stars. If the simulation is wrong, the AI's guess will be wrong.

5. Beyond Just Weighing: Other AI Superpowers

The paper notes that AI isn't just for weighing clusters; it's being used for other detective work too:

Reconstructing the Invisible: AI can take a simulation that only shows "dark matter" and predict where the gas and stars should be, saving scientists from running expensive, slow simulations.
Spotting Crashes: AI can look at images and instantly tell if two clusters are crashing into each other (merging), which is a violent event that changes the cluster's shape.
Finding the "Ghost" Light: There is a faint, diffuse glow between galaxies called "Intracluster Light." It's very hard to see. AI can act like a high-powered filter to separate this faint glow from the background noise, helping scientists understand how the cluster was built over time.

The Bottom Line

The paper concludes that AI is a powerful new tool that is outperforming old methods. It handles the messy, complex reality of the universe better than simple formulas. However, to trust these AI "detectives," we must ensure they are trained on the most realistic computer simulations possible and that we understand why they make the guesses they do.

In short: We are swapping a simple ruler for a super-smart, pattern-recognizing AI to weigh the heaviest objects in the universe.

Technical Summary: Machine Learning Applications to Galaxy Clusters

Problem Statement
Galaxy clusters, the most massive gravitationally bound systems in the Universe, serve as critical laboratories for studying large-scale structure formation and as powerful cosmological probes. However, their total mass is not a direct observable; it must be inferred from multi-wavelength tracers (Sunyaev–Zel'dovich effect, X-ray emission, optical galaxy distributions, and dynamical data). Traditional mass estimation methods rely on simplified assumptions, such as hydrostatic equilibrium (HE), spherical symmetry, and self-similar scaling relations. These classical approaches suffer from systematic biases (e.g., the "mass bias" where HE estimates underestimate true mass) and significant intrinsic scatter due to non-linear processes, projection effects, and complex cluster morphologies that are not captured by integrated quantities. Furthermore, cosmological simulations required to calibrate these relations face a trade-off between volume (statistical representativeness) and resolution (physical fidelity), making it difficult to generate the large, well-characterized datasets needed for robust calibration.

Methodology
The chapter reviews the application of Artificial Intelligence (AI), specifically Machine Learning (ML) and Deep Learning (DL), to address these limitations. The core methodology involves training models on large datasets derived from cosmological hydrodynamical simulations, most notably the THE THREE HUNDRED project, which provides a statistically representative sample of clusters with detailed baryonic physics.

Key methodological approaches include:

Architectures: The use of Convolutional Neural Networks (CNNs), Graph Neural Networks (GNNs), Autoencoders, Random Forests, Support Distribution Machines (SDM), Normalizing Flows, and Score-based Generative Models.
Input Data: Models ingest raw or processed data from various tracers, including SZ Compton- $y$ maps, X-ray photon-count images, optical phase-space distributions (positions and velocities of member galaxies), and weak lensing shear maps.
Training Strategy: Models learn the non-linear mapping between observables and total mass directly from simulations, bypassing the need for explicit parametric assumptions about the thermodynamic state or geometry of the cluster.
Uncertainty Quantification: The review discusses the use of Simulation-Based Inference (SBI) and approximate Bayesian methods to estimate uncertainties, acknowledging that full Bayesian treatment remains computationally challenging for deep networks.

Key Contributions and Results
The paper synthesizes recent advances demonstrating that AI techniques systematically outperform classical methods in mass inference and other cluster properties:

Mass Inference from SZ Emission:
- Deep learning models (e.g., mResUNet, CNNs) trained on mock SZ maps can infer cluster masses with $\sim20\%$ uncertainty, capturing morphological features and projection effects often ignored by standard scaling relations.
- Hybrid approaches (Autoencoder + Random Forest) have successfully reconstructed full 3D mass profiles with significantly reduced scatter compared to HE methods.
- CNNs applied to real Planck data have provided independent calibrations of the hydrostatic mass bias, yielding values consistent with numerical simulations ( $1-b \sim 0.8–0.9$ ).
Mass Inference from X-ray Emission:
- CNNs applied to X-ray images (e.g., from Illustris TNG and MAGNETICUM) have reduced mass scatter to $\sim8\%$ , compared to $\sim18\%$ for standard techniques.
- Graph Neural Networks (GNNs) utilizing radially sampled ICM profiles have achieved a dispersion of $\sim7\%$ on test sets and have been successfully applied to real datasets like REXCESS and X-COP.
Mass Inference from Galaxies and Dynamics:
- ML methods (SDM, CNNs, Normalizing Flows) applied to line-of-sight velocity distributions and phase-space images have reduced mass scatter by factors of up to two compared to traditional $M-\sigma_v$ relations.
- These methods do not require assumptions about dynamical equilibrium, allowing for more robust mass estimates of non-relaxed systems.
Multi-Tracer and Multi-Wavelength Approaches:
- Combining SZ, X-ray, and optical data via supervised learning (e.g., Gradient Boosting, U-Net) breaks degeneracies inherent in single-tracer proxies, reducing scatter in integrated quantities by a factor of $\sim2$ .
- U-Net architectures have been used to reconstruct projected 2D total mass maps from multi-tracer inputs.
Beyond Mass Estimation:
- Baryonic Emulation: ML models map Dark Matter-only simulations to realistic baryonic fields (gas, stars, thermodynamics), offering a computationally efficient alternative to full hydrodynamical simulations.
- Dynamical States: CNNs have been used to identify and classify merging clusters based on SZ and X-ray morphologies.
- Intracluster Light (ICL): Deep learning (U-Net, ResNet) enables the detection and reconstruction of the diffuse ICL component, linking it to cluster assembly history.

Significance and Limitations
The paper posits that ML and DL provide a natural extension of classical methods, offering a flexible, data-driven framework that captures complex, non-linear dependencies and high-dimensional information (morphology, substructure). This is crucial for next-generation surveys (e.g., eROSITA, Euclid, LSST) where precision cosmology demands rigorous control of systematic uncertainties in mass calibration.

However, the authors emphasize significant limitations:

Simulation-to-Reality Gap: ML models are only as reliable as the physical modeling in their training simulations. Inaccuracies in feedback prescriptions or subgrid physics in simulations can lead to biased predictions in real observations.
Interpretability: While saliency maps suggest that mass inference is driven by features in cluster outskirts (less affected by complex baryonic physics) rather than cores, the "black box" nature of deep learning remains a challenge.
Uncertainty Quantification: Robust, well-calibrated uncertainty estimates are still an open problem, with current methods relying on approximations.
Standardization: The field lacks standardized benchmark datasets and metrics, making fair comparisons between different models difficult.

In conclusion, the paper argues that while AI offers a promising pathway to improved mass calibration and new physical diagnostics, its successful application to precision cosmology requires careful validation against observational data, the development of domain adaptation techniques, and the generation of diverse training datasets that marginalize over uncertainties in baryonic physics.