🔭 astrophysics

Learning the Cosmic Web: Graph-based Classification of Simulated Galaxies by their Dark Matter Environments

This paper presents a novel graph-based machine learning classifier using a Graph Attention Network (GAT) and Delaunay triangulation metrics to accurately identify the dark matter cosmic web environments of simulated galaxies, demonstrating superior performance over traditional models and paving the way for analyzing observed galaxies from the DESI survey.

Original authors: Dakshesh Kololgi, Krishna Naidoo, Amelie Saintonge, Ofer Lahav

Published 2026-04-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dakshesh Kololgi, Krishna Naidoo, Amelie Saintonge, Ofer Lahav

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: Mapping the Invisible Skeleton of the Universe

Imagine the universe isn't just a random scattering of stars and galaxies. Instead, think of it as a giant, 3D spiderweb made of invisible dark matter. Astronomers call this the Cosmic Web.

In this web, galaxies (like our Milky Way) live in four distinct neighborhoods:

Voids: The empty, quiet suburbs with almost nothing around them.
Walls: Flat, sheet-like neighborhoods.
Filaments: Long, thin highways connecting different parts of the web.
Clusters: The bustling, crowded city centers where thousands of galaxies hang out together.

The Problem: We can see the galaxies (the houses), but we can't see the dark matter web (the roads and neighborhoods) directly. We need a way to look at where a galaxy is sitting and say, "Ah, this galaxy is definitely living in a Filament," or "That one is in a Cluster."

The Solution: The authors of this paper built a super-smart computer brain (a type of Artificial Intelligence) that looks at the shape of the galaxy neighborhood to guess which environment it lives in.

Step 1: The Training Ground (The Simulator)

You can't teach a computer to recognize a neighborhood unless you show it a million examples first. Since we can't see the real dark matter web, the scientists used a video game simulation called IllustrisTNG-300.

Think of this simulation as a perfect, virtual universe where the scientists know exactly where every piece of dark matter is. They used a mathematical ruler (called the T-Web method) to label every single galaxy in the simulation with its true address: "Void," "Wall," "Filament," or "Cluster."

This gave them a massive textbook of "Galaxy + Correct Address" pairs to teach their AI.

Step 2: Drawing the Map (The Graph)

To teach the AI, they needed a way to describe the neighborhood of each galaxy without just looking at a single point.

Imagine you are dropped in a strange city and asked to describe your surroundings. You wouldn't just say "I am here." You would say, "I am connected to three friends, the street is narrow, and the buildings are clustered tightly."

The scientists did this for every galaxy using Delaunay Triangulation.

The Analogy: Imagine stretching rubber bands between every galaxy and its nearest neighbors. This creates a mesh of triangles (a graph) that covers the whole universe.
The Metrics: From this mesh, they measured 10 specific things for each galaxy, like:
- How many neighbors does it have? (Degree)
- How far away are they? (Edge length)
- Are the neighbors clustered together like a huddle, or spread out? (Clustering)

These 10 numbers became the "fingerprint" of the galaxy's environment.

Step 3: The Brain (The AI Models)

The team tried three different types of "brains" to read these fingerprints:

The MLP (The Independent Thinker): This AI looked at the 10 numbers for a galaxy but ignored its neighbors. It was like trying to guess someone's job by looking at their resume but ignoring who they work with. It got about 68% right.
The GCN (The Group Thinker): This AI looked at the galaxy and its neighbors, averaging their features. It was like asking a person, "What do you and your friends do?" It got slightly better, around 70%.
The GAT+ (The Smart Detective): This is the star of the show. It's a Graph Attention Network.
- How it works: Instead of treating all neighbors equally, this AI learns to pay attention to the right neighbors.
- The Analogy: Imagine you are trying to guess if a person lives in a city or the countryside. If you ask their neighbors, the GCN asks everyone equally. The GAT+ asks, "Okay, this neighbor is clearly a farmer (countryside), but that one is a banker (city). I will listen more to the farmer and ignore the banker."
- The Result: By learning which connections matter most, the GAT+ model achieved 85% accuracy.

Why Did the "Smart Detective" Win?

The Cosmic Web is messy. The boundary between a "Wall" and a "Filament" isn't a sharp line; it's a fuzzy transition zone.

The GCN was too nice; it averaged out the differences, blurring the lines between neighborhoods.
The GAT+ was smart enough to realize, "Hey, this specific connection crosses a boundary, so I shouldn't let it confuse me. I'll focus on the connections that stay inside the neighborhood."

What Does This Mean for the Real Universe?

Right now, this AI is trained on a video game (the simulation). But the scientists plan to take this "Smart Detective" and apply it to real data from the DESI survey (a massive telescope project mapping millions of real galaxies).

The Goal: Once the AI is adjusted to handle real-world telescope glitches (like missing data or blurry images), it will be able to look at a galaxy in the real sky and tell us, with high confidence, "This galaxy is sitting on a cosmic filament."

Why Should We Care?

Understanding where a galaxy lives helps us understand why it looks the way it does.

Galaxies in Clusters might be old and dead because they ran out of gas.
Galaxies in Filaments might be young and growing because the filament is feeding them fresh gas.

By mapping the invisible web, we can finally understand the story of how galaxies grow, evolve, and die. This paper proves that looking at the shape of the galaxy neighborhood is a powerful new way to tell that story.

1. Problem Statement

Understanding galaxy evolution requires a robust classification of the cosmic web environments (voids, walls, filaments, and clusters) in which galaxies reside. While cosmological simulations provide full access to the dark matter field, observational surveys (like DESI) only offer biased, incomplete tracers (galaxies).

The Challenge: Traditional methods for defining the cosmic web often rely on reconstructing the full dark matter density field or using dynamical information (e.g., velocity shear) which is unavailable in redshift surveys.
The Goal: Develop a machine learning classifier that can infer the underlying dark matter environment of a galaxy using only the relative positions of galaxies, without reconstructing the full density field.

2. Methodology

The authors propose a three-stage, simulation-based framework using the IllustrisTNG-300 simulation.

A. Ground Truth Generation (T-Web Classification)

Source: The dark matter density field from the simulation.
Method: The T-Web classification scheme is applied. This uses the eigenvalues of the Hessian matrix of the gravitational potential ( $\Phi$ $Φ$ ).
- Logic: The signs of the three eigenvalues determine the environment:
  - 3 positive: Cluster (collapse on all axes).
  - 2 positive: Filament.
  - 1 positive: Wall.
  - 0 positive: Void.
Parameters: A smoothing scale of $R_s = 2 h^{-1}$ Mpc and an eigenvalue threshold of $\lambda_{th} = 0.2$ were selected to balance visual realism and physical robustness.
Data: Galaxies with stellar mass $M_* > 10^9 M_\odot$ were selected as proxies. A 10 Mpc buffer region was applied to the simulation boundaries to remove edge effects caused by the Delaunay algorithm.

B. Graph Construction

Instead of treating galaxies as independent points, the authors construct a graph to capture local geometric structure.

Graph Type: Delaunay Triangulation was chosen over Minimum Spanning Trees (MST) and Complex Graphs (nearest-neighbor).
- Reasoning: Delaunay triangulation provides a space-filling tessellation that naturally captures voids, walls, and filaments without arbitrary linking lengths. It outperformed other graph types in preliminary tests.
Features (Graph Metrics): For each galaxy (node), 10 graph metrics were extracted from the Delaunay graph connectivity:
- Topological: Degree, average degree, clustering coefficient, Katz centrality.
- Geometric: Mean edge length, min/max edge length.
- Density/Morphology: Tetrahedral density, shape eigenvalues ( $I_1, I_2, I_3$ ).
Preprocessing: Metrics were transformed (Box-Cox) to be Gaussian-like, and mutual information analysis was used to confirm feature relevance.

C. Machine Learning Architectures

Three models were trained to map the 10 graph metrics to the 4 cosmic web classes:

Multi-Layer Perceptron (MLP): A baseline model treating each galaxy's metrics as independent inputs (ignoring graph structure).
Graph Convolutional Network (GCN): Aggregates features from neighbors using fixed, degree-normalized weights.
Graph Attention Network (GAT+): The proposed state-of-the-art model.
- Mechanism: Uses an attention mechanism to learn adaptive weights for neighbor connections, allowing the model to prioritize informative neighbors and down-weight noisy or boundary-crossing connections.
- Architecture: 4 attention heads, residual connections, layer normalization, and dropout to prevent overfitting and handle the "heterophilous" nature of the graph (where neighbors often belong to different classes).

3. Key Results

Performance Comparison

The models were evaluated on a test set (20% of data) with a stellar mass cut of $10^9 M_\odot$ .

Model	Accuracy	Training Time
Random Forest / XGBoost	~70-72%	Fast
MLP (Baseline)	68%	Fast
GCN	69-70%	Moderate
GAT+ (Proposed)	85%	Moderate

GAT+ Superiority: The GAT+ model significantly outperformed all baselines (15-17% improvement over MLP/GCN).
Class Performance:
- Voids and Clusters: >90% accuracy (distinct environments).
- Walls and Filaments: ~~86-90% accuracy. These classes are frequently confused (~~10-14% misclassification) because they occupy intermediate, overlapping density regimes.

Feature Importance & Interpretability

Clustering Coefficient: Found to be the most informative metric, strongly correlated with dense regions (clusters/filaments). This contrasts with previous studies using fixed-link graphs, highlighting the superiority of the Delaunay approach.
Entropy Analysis: Galaxies in transition zones (e.g., wall-to-filament) exhibited high predictive entropy (low confidence), correctly reflecting the physical continuity of the cosmic web.
Latent Space: UMAP visualization showed that GAT+ embeddings created distinct clusters for each environment, whereas raw input features showed significant overlap.

4. Significance and Contributions

Novel Graph-Based Framework: The paper demonstrates that graph-based representations of galaxy positions are a powerful, physically meaningful alternative to density field reconstruction for inferring dark matter environments.
Superiority of Attention Mechanisms: It establishes that Graph Attention Networks (GAT) are essential for this task. Unlike GCNs, which average neighbor features (potentially blurring boundaries), GATs learn to weigh neighbors adaptively, effectively disentangling the heterogeneous local environments found in the cosmic web.
Handling Heterophily: The study addresses the challenge of "heterophilous" graphs (where connected nodes often have different labels), a common issue in cosmic web classification, by using multi-head attention and residual connections.
Pathway to Observational Data: While currently tested on simulations, the authors outline a clear path to applying this to real surveys like DESI (Dark Energy Spectroscopic Instrument). By training on simulations that include survey systematics (selection functions, fiber incompleteness), this method can provide probabilistic environmental classifications for millions of observed galaxies, enabling large-scale studies of how environment drives galaxy evolution.

Conclusion

The authors successfully developed a GAT+ classifier that achieves 85% accuracy in identifying dark matter cosmic web environments using only galaxy positions. This approach bypasses the need for density field reconstruction, offering a scalable and robust tool for the next generation of large-scale structure surveys.