🔭 astrophysics

ERGO-ML: A continuous organization of the X-ray galaxy cluster population in TNG-Cluster with contrastive learning

This paper demonstrates that Nearest Neighbour Contrastive Learning (NNCLR) can effectively map the continuous physical and dynamical evolution of galaxy clusters in the TNG-Cluster simulation into a low-dimensional representation space using X-ray emission maps, enabling the prediction of key properties and the identification of causal relationships without relying on traditional summary statistics.

Original authors: Urmila Chadayammuri, Lukas Eisert, Annalisa Pillepich, Katrin Lehle, Mohammadreza Ayromlou, Dylan Nelson

Published 2026-05-27

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Urmila Chadayammuri, Lukas Eisert, Annalisa Pillepich, Katrin Lehle, Mohammadreza Ayromlou, Dylan Nelson

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: Organizing a Messy Library

Imagine you have a library with thousands of books, but none of them have titles, authors, or summaries on the spine. They are just a chaotic pile. Traditionally, astronomers have tried to sort galaxy clusters (huge groups of galaxies held together by gravity) by looking at a few specific numbers, like "total weight" or "how bright the center is." It's like trying to sort that library by only looking at the thickness of the book.

This paper introduces a new way to organize these "books" (galaxy clusters) using a type of artificial intelligence called Contrastive Learning. Instead of just counting numbers, the AI looks at the actual "pictures" (X-ray maps) of the gas inside these clusters and learns to arrange them on a map based on how similar they look.

The Ingredients: The "TNG-Cluster" Simulation

To train this AI, the researchers didn't use real photos from telescopes (which are often blurry or incomplete). Instead, they used a super-powerful computer simulation called TNG-Cluster.

Think of this simulation as a "video game universe" where the laws of physics are perfectly programmed. The researchers generated 352 different galaxy clusters and took snapshots of them at different times in their history (from when the universe was younger to now). They took pictures of these clusters from three different angles, resulting in about 8,000 images.

These images show the hot gas (the "intracluster medium") that fills the space between galaxies. This gas glows in X-rays, revealing structures like shockwaves from collisions, smooth bubbles, or swirling storms.

The Method: Teaching the AI to "See" Similarity

The researchers used a technique called Nearest Neighbour Contrastive Learning (NNCLR). Here is how it works, using an analogy:

Imagine you are teaching a child to recognize different types of clouds.

The Augmentation (The Trick): You show the child a picture of a fluffy cloud. Then, you show them the same picture, but you zoom in, rotate it, blur it slightly, or make it look like it's taken through a foggy window. You tell the child, "These are the same cloud, just seen differently."
The Lesson: You also show them a picture of a storm cloud and tell them, "This is different."
The Result: The AI learns to ignore the "tricks" (rotation, blurring, zooming) and focuses on the true shape of the cloud. It builds a mental map where similar clouds are close together, and different clouds are far apart.

In this paper, the "clouds" are galaxy clusters. The AI learned to ignore things like how the cluster is rotated or how bright the image is, and instead focused on the actual structure: Is it a calm, smooth ball? Is it two clusters crashing into each other? Is the center dense or flat?

The Findings: A Smooth Map of the Universe

When the researchers plotted the AI's "mental map" (called a UMAP) on a 2D screen, something beautiful happened. The clusters didn't form random dots; they formed a continuous landscape.

The Top of the Map: Shows clusters that are currently crashing into each other (merging). They look messy and chaotic.
The Bottom of the Map: Shows "relaxed" clusters that are calm, isolated, and smooth.
The Right Side: Shows clusters with a very dense, bright center (like a bullseye).
The Left Side: Shows clusters with a flatter, more spread-out center.

The most exciting part is that this map isn't just about looks. The researchers found that if you know where a cluster sits on this map, you can guess its physical properties:

Mass: Heavier clusters tend to be in certain areas.
History: Clusters that had a massive crash recently are grouped together.
Black Holes: The activity of the giant black holes in the center correlates with where the cluster sits on the map.

Why This Matters (According to the Paper)

It's a Better Way to Sort: Instead of forcing every cluster into a simple "Yes/No" box (e.g., "Is it a cool core? Yes/No"), this method shows a smooth gradient. It acknowledges that nature is a spectrum, not a set of discrete categories.
It Handles "Bad" Data: The researchers tested the AI by deliberately blurring the images or cutting off the faint parts (simulating what happens when a telescope doesn't have enough time to take a clear photo). The AI still organized the clusters correctly. This means this method could work on real telescope data from different instruments, even if the data quality varies.
Finding "Twins": If astronomers find a weird, messy cluster in the real sky, they can use this map to find its "twins" in the simulation. By looking at the twins in the simulation, they can learn the history of the real cluster (e.g., "Oh, this messy one probably just collided with another galaxy 2 billion years ago").

The Limits

The paper is honest about what the AI can't do perfectly yet.

Small Details: The AI is great at seeing the big picture (the whole cluster), but it struggles to predict very specific, tiny details (like the exact temperature of the very center) because the images used for training weren't zoomed in enough to see those tiny pixels clearly.
Prediction: While the AI can organize the clusters beautifully, predicting exact numbers (like "this cluster has exactly 1.5 trillion solar masses") from the map is only accurate to within a small percentage for mass, but less accurate for more complex things like star formation rates.

Summary

In short, this paper shows that by using a smart AI to look at X-ray pictures of galaxy clusters, we can organize the entire population of the universe into a smooth, continuous map. This map reveals the hidden history of these clusters—how they merged, how heavy they are, and how active their central black holes are—without needing to reduce them to simple numbers first. It's like turning a chaotic pile of puzzle pieces into a clear picture of the universe's evolution.

Technical Summary: ERGO-ML: A continuous organization of the X-ray galaxy cluster population in TNG-Cluster with contrastive learning

Problem Statement
The physical properties of the intracluster medium (ICM) encode signatures of gravitational potentials, mergers, and feedback from supernovae and supermassive black holes (SMBHs). Traditionally, galaxy clusters have been characterized using summary statistics (e.g., halo mass, X-ray luminosity, cool-core state) or radial profiles. However, reducing complex 2D X-ray images to scalar metrics discards significant information content and introduces scatter in scaling relations. Furthermore, inferring physical conditions from X-ray emission is challenging due to the complexity of cooling channels, non-thermal processes, and instrumental limitations such as varying spatial resolution and exposure times. While machine learning, specifically Convolutional Neural Networks (CNNs), has been applied to infer cluster masses, these models often lack invariance to common astronomical transformations (rotation, translation, dynamic range) and do not naturally organize the population into a continuous, interpretable space.

Methodology
The authors employ Nearest Neighbour Contrastive Learning (NNCLR) to construct a low-dimensional representation space for X-ray emission maps of galaxy clusters. The study utilizes the TNG-Cluster suite, a cosmological magnetohydrodynamical (MHD) simulation of 352 massive galaxy clusters ( $M_{200c} \ge 10^{14.3} M_\odot$ ).

Data Generation: Intrinsic X-ray emission maps were generated for 352 clusters at eight snapshots within the redshift range $0 \le z < 1$ . Projections were taken along three orthogonal axes ( $\hat{x}, \hat{y}, \hat{z}$ ), resulting in a dataset of approximately 8,000 images. Each map covers a field of view of $4R_{200c}$ with a line-of-sight depth of $2R_{200c}$ , normalized such that the 99th percentile pixel in the central 10% is set to 1.
Augmentation Strategy: To ensure the learned representation is invariant to observational variations, the training pipeline applies random augmentations including:
- Geometric transformations: Zoom (up to 10%), rotation ( $\pm 90^\circ$ ), affine translation (up to 25%), and flipping.
- Photometric/Instrumental effects: Gaussian blurring (simulating Point Spread Function variations), Gaussian noise, and intensity clipping (simulating varying exposure times and surface brightness limits).
Architecture: The augmented images are processed through a Residual Network (ResNet) acting as a feature extractor. The network maps inputs to an embedding space. The authors experiment with embedding dimensions ( $N=2$ to $256$), selecting $N=8$ as the default.
Loss Function: The NNCLR algorithm minimizes the distance between the embedding of an augmented view and its nearest neighbor in a "support set" (positive pair), while maximizing the distance to negative pairs (embeddings of different images).
Visualization and Analysis: The high-dimensional embeddings are projected into 2D using Uniform Manifold Approximation and Projection (UMAP) for visualization. The authors analyze the distribution of physical properties (masses, merger times, SMBH activity, dynamical state indicators) within this space and test the ability to predict these properties via simple linear fits to the UMAP coordinates.

Key Contributions and Results

Continuous Representation Space: The NNCLR method successfully organizes the X-ray maps into a continuous distribution rather than discrete clusters. The 2D UMAP visualization reveals smooth transitions:
- From merging systems with well-separated components (top of the map) to relaxed, isolated objects (bottom).
- From flat emission profiles (bottom left) to centrally peaked (cuspy) profiles (bottom right).
- The representation captures trends with redshift, halo mass, gas/stellar/SMBH mass, time since last major merger, and dynamical state indicators.
Invariance to Observational Effects: The inclusion of clipping and blurring augmentations ensures that the representation remains robust. The study demonstrates that clusters with different exposure times or spatial resolutions (simulating instruments like Chandra vs. eROSITA) occupy similar regions in the representation space if their underlying physical structures are analogous.
Property Prediction:
- High Accuracy: Mass-related properties (halo mass, stellar mass, gas mass, SMBH mass) and instantaneous black hole accretion rates (BHAR) can be predicted from the 2D UMAP coordinates with a few percent error (linear fit).
- Moderate Accuracy: Time since the last major merger is predicted within an order of magnitude.
- Limitations: Cool-core indicators and the time since the Brightest Cluster Galaxy (BCG) merger are not well-predicted by the global representation. The authors attribute this to the resolution of the training images, where core-scale features ( $\sim 0.01-2$ pixels) may be under-sampled.
Analogue Detection: The representation space enables the identification of visual analogues. The five nearest neighbors of a target cluster in the embedding space share similar physical properties and morphological states (e.g., merging vs. relaxed) more closely than the general population. This method outperforms traditional searches based solely on merger tree parameters (mass ratios and impact parameters), which can fail to capture the visual appearance of a merger due to projection effects.

Significance and Claims
The paper claims that contrastive learning is a powerful tool for characterizing galaxy clusters directly from their images, bypassing the need for summary statistics. By organizing the population into a continuous, low-dimensional manifold, the method allows for:

The derivation of constraints on physical properties and formation histories using cosmological hydrodynamical simulations.
The identification of correlations between physical properties that may suggest causal relationships.
A robust framework applicable to future observational surveys (e.g., eROSITA, Athena) by training on mocks that account for instrument-specific limitations.

The authors emphasize that while the current analysis uses idealized simulations, the methodology is designed to be transferable to real observations, provided models are trained on realistic mocks that include instrumental effects. The work serves as a foundation for the ERGO-ML series, demonstrating that self-supervised learning can extract meaningful physical information from complex astrophysical images.