📊 statistics

Bayesian Cosmic Void Finding with Graph Flows

This paper introduces a Bayesian cosmic void finding method that utilizes a deep graph neural network with flow-matching to sample stochastic mappings from galaxy catalogs to void definitions, thereby capturing probabilistic uncertainties and outperforming deterministic approaches in extracting cosmological information.

Original authors: Leander Thiele

Published 2026-04-20

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Leander Thiele

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: Finding the "Empty" Spaces in the Universe

Imagine the universe isn't just a random scattering of stars and galaxies. Instead, think of it like a giant, 3D spiderweb. The galaxies are the shiny beads on the web, but most of the space is actually empty. These giant, empty bubbles are called Cosmic Voids.

Scientists love studying these voids because they hold secret clues about how the universe works (like dark energy and the Big Bang). However, finding them is tricky. We can't see the "empty space" directly; we can only see the galaxies floating around it. It's like trying to map the shape of a cave by only looking at the bats flying near the entrance.

The Problem: The "One Right Answer" Trap

For a long time, scientists used computer programs to find these voids. These programs were deterministic, meaning if you gave them the same data, they would always give you the exact same map of voids.

The author of this paper, Leander Thiele, argues that this is a mistake. Because our view of the universe is "sparse" (we only see a few galaxies, not the whole picture), there isn't just one correct map of the voids. There are many possible maps that could all be true.

The Analogy: Imagine you are trying to guess the shape of a hidden sculpture in a dark room by feeling only a few scattered marbles on the floor.

Old Method: You guess one specific shape and say, "This is definitely it!"
New Method: You realize there are many shapes that could fit those marbles. Instead of picking one, you generate a whole cloud of possible shapes, each slightly different, to show the full range of possibilities.

The Solution: "Test Particles" on a Flow

To solve this, the author created a new AI method called Bayesian Cosmic Void Finding. Here is how it works, using a creative metaphor:

1. The "Test Particles" (The Drifters)

Imagine you drop thousands of tiny, invisible balloons (called "test particles") into the universe. These balloons don't know where the voids are yet. They are just floating randomly.

2. The "Graph Neural Network" (The Invisible Wind)

The AI acts like a smart, invisible wind. It looks at where the galaxies are and pushes the balloons.

If a balloon is near a cluster of galaxies, the wind pushes it away (because voids are empty).
If a balloon is in a big empty space, the wind guides it to the center of that space.

3. The "Flow" (The Journey)

The AI doesn't just snap the balloons into place instantly. It simulates a journey over "fictitious time." The balloons drift, swirl, and settle into the centers of the voids. This process is called Flow Matching. It's like watching a river carry leaves; the AI learns the current that naturally carries the leaves to the calm pools (the voids).

4. The "Graph" (The Neighborhood Chat)

The AI is a Graph Neural Network. Think of the galaxies and balloons as people at a party.

Each balloon talks to its nearest neighbors (other balloons and nearby galaxies).
They pass messages like, "Hey, I'm near a galaxy, I should move left," or "I'm surrounded by empty space, I should stay here."
By chatting with neighbors, the whole group figures out where the voids are without needing a central boss telling them what to do.

The Surprise: Being "Messy" is Better

Here is the most interesting part of the paper. The AI was trained using a "teacher" (a standard, perfect void-finding program). You would expect the AI to just copy the teacher perfectly.

But the AI didn't. It became stochastic (random).

The Teacher: Always draws the exact same map.
The AI: Draws a slightly different map every time you run it.

Why is this good?
The author calls this Regularization. Because the AI explores many possible maps instead of just one, it avoids getting stuck on weird, specific details that might be wrong. It's like a student who studies a concept by exploring many different examples rather than memorizing one single answer.

The Result: When the author tested the AI's "messy" maps against real physics, the AI actually found more useful information about the universe than the perfect, rigid teacher did. The "cloud" of possibilities contained more truth than a single "perfect" line.

The "Dark Matter" Challenge

The author also tried to use this method to find voids in Dark Matter (the invisible stuff that holds galaxies together) using only the visible galaxies as a guide.

The Result: It was much harder. It's like trying to map the shape of a mountain range by only looking at the birds flying over it. The connection is weak.
The Lesson: The AI struggled, but it didn't fail randomly. It produced a "cloud" of uncertainty that honestly reflected how hard the problem was. This is a sign of a smart system: it knows when it's guessing.

Why Does This Matter?

Better Science: By accepting that there are many possible answers, we get a more honest and accurate picture of the universe.
Future Proofing: This method can be applied to any definition of a void, even ones we haven't invented yet.
Efficiency: It's a fast way to generate thousands of possible universe maps, which helps scientists test their theories faster.

In a Nutshell

The author built a smart AI that doesn't just look for one answer to "Where are the cosmic voids?" Instead, it generates a whole family of possible answers. By embracing the uncertainty and randomness of the universe, this AI actually learns the secrets of the cosmos better than the old, rigid methods. It's a shift from asking "What is the answer?" to asking "What are all the possible answers, and what do they teach us?"

1. Problem Statement

The identification of cosmic voids (underdense regions in the large-scale structure) is crucial for extracting cosmological information and testing astroparticle physics. However, finding genuine matter underdensities in sparse galaxy surveys is an underconstrained inverse problem.

The Limitation of Traditional Methods: Current void finders (e.g., VIDE, ZOBOV) are deterministic. They map a sparse galaxy catalog to a single void catalog, effectively ignoring the probabilistic nature of the problem.
Instability: Deterministic methods are unstable; small perturbations in galaxy positions (e.g., due to redshift-space distortions or measurement errors) can lead to significantly different void catalogs.
The Goal: The paper aims to solve for the Bayes-optimal mapping from observed galaxy catalogs to a probability distribution over void catalogs, $P(\{v_\alpha\} | \{g_\beta\})$ , rather than a single deterministic output. This allows for the quantification of uncertainty and potentially better cosmological constraints.

2. Methodology: Graph Flow Matching

The author proposes a generative machine learning framework based on Flow Matching and Graph Neural Networks (GNNs).

A. Core Concept: Test Particles

Instead of classifying density fields directly, the method evolves "test particles" to represent voids.

State Definition: Each test particle $v_\alpha(t)$ is a tuple containing a position $\vec{r}_\alpha(t)$ and attributes (e.g., void volume) $\vec{a}_\alpha(t)$ .
Flow Matching: The method learns a velocity field $f_\phi$ that transports particles from a simple base distribution (Poisson noise at $t=0$ ) to the target distribution of void catalogs (at $t=1$ ).
Objective: The network minimizes the difference between the predicted velocity and the velocity of a linear interpolant between the base and target samples.

B. Architecture: Graph Neural Networks

The velocity field is parameterized by a GNN that operates on two types of nodes:

Galaxies ( $G$ ): The observed sparse tracers.
Test Particles ( $P$ ): The evolving void candidates.

Message Passing: The network updates particle features by aggregating messages from:
- Nearby galaxies (k-nearest neighbors).
- Other test particles (to capture void-void correlations and exclusion).
Training Strategy:
- Oversampling: Since the number of voids is unknown, the model uses $N_p > N_v$ test particles.
- Optimal Transport Assignment: To pair test particles with target voids during training, the authors minimize the sum of squared Euclidean distances between initial particle positions and target void positions (solved via the Hungarian algorithm).
- Clustering: During inference, the evolved particles are clustered into unique voids using a Friends-of-Friends (FoF) algorithm.

C. Training Data

Simulations: Quijote simulations ( $1 h^{-1}$ Gpc box, $512^3$ particles).
Cosmology: Varied 5-parameter $\Lambda$ CDM models.
Galaxy Populations: Generated via Halo Occupation Distribution (HOD) to simulate sparse galaxy catalogs ( $N_g \approx 10^5$ ).
Targets:
1. Galaxy Voids: VIDE voids identified directly in the galaxy field (deterministic teacher).
2. Dark Matter (DM) Voids: VIDE voids identified in the full particle field (stochastic teacher, as galaxies are sparse tracers of DM).

3. Key Contributions

Probabilistic Void Finding: The first method to sample from the full posterior distribution of void catalogs conditioned on sparse galaxy data, moving beyond deterministic point estimates.
Bayes-Optimal Mapping: The framework is general enough to target any void definition (e.g., defined by density, velocity, or simulation particles), providing the theoretically optimal mapping given the available information.
Regularization via Stochasticity: The authors demonstrate that the stochasticity inherent in the trained model acts as a beneficial regularizer.
Graph Flow Architecture: A novel application of flow matching using GNNs to handle the relational nature of cosmic structures (galaxies and voids).

4. Results

A. Galaxy Voids (Deterministic Teacher)

Performance: The model successfully reproduces the Void Size Function (VSF) and correlation functions (void-galaxy and void-void) of the VIDE teacher.
Stochasticity: The model produces a distribution of void catalogs. While individual samples show variance, the mean of the generated catalogs matches the ground truth statistics closely.
Cosmological Information: Crucially, the mean of the graph flow predictions contains more cosmological information (higher Fisher information) than the deterministic VIDE catalog. The model effectively "denoises" the deterministic output through regularization.

B. Dark Matter Voids (Stochastic Teacher)

Challenge: Predicting DM voids from sparse galaxies is highly underconstrained.
Performance: The model learns a mapping with high stochasticity. While it cannot perfectly reconstruct specific DM void locations (low cross-correlation with ground truth), it successfully learns the statistical properties (VSF) and does not simply output random noise.
Interpretation: The model behaves as a Bayes-optimal estimator under limited information, producing a population of "failed" trajectories for uncertain regions while converging on high-confidence voids.

C. Stability and Generalization

The method is robust to small perturbations in galaxy positions, unlike deterministic finders which show large catalog variations under similar jitter.
The use of traditional message aggregation (mean/max) outperformed attention mechanisms, likely due to the smooth nature of cosmological fields.

5. Significance and Future Outlook

Cosmological Inference: By providing a probabilistic void catalog, this method allows for more robust cosmological parameter estimation, accounting for the uncertainty in void identification. The finding that the "regularized" mean outperforms the deterministic teacher is a significant insight for machine learning in cosmology.
Flexibility: The framework is not limited to current void definitions. It can be trained to find voids defined by velocity fields or other simulation-level quantities that are inaccessible to observers, bridging the gap between theory and observation.
Generalizability: The "test particle under a conditioned flow" approach is applicable to other cosmic structure problems, such as filament finding or populating low-resolution simulations with halos.

Conclusion:
Thiele presents a paradigm shift from deterministic void finding to Bayesian sampling. By leveraging Graph Flow Matching, the method captures the inherent uncertainty of sparse galaxy surveys, producing void catalogs that are not only statistically consistent with ground truth but also yield superior cosmological constraints through effective regularization.