Bayesian Cosmic Void Finding with Graph Flows

2026-04-20 · Source: stat.ML updates on arXiv.org · Field: Science & Research — Space Science & Astronomy, Artificial Intelligence & Machine Learning, Research Methodology & Innovation · Depth: Expert, extended

Summary

Leander Thiele introduces a novel Bayesian method for identifying cosmic voids in sparse galaxy surveys, addressing the underconstrained and probabilistic nature of this problem. Traditional void-finding algorithms produce deterministic catalogs, neglecting inherent uncertainties. The new algorithm employs a deep graph neural network to evolve "test particles" using a flow-matching objective, sampling from the stochastic mapping between galaxy catalogs and arbitrary void definitions. Trained on 12,000 Quijote simulations with a side length of 1 h^-1 Gpc, the model demonstrates effective performance, even outperforming its deterministic VIDE teacher in terms of cosmological information extracted from void catalogs. The method is shown to emulate existing void finders with useful regularization and can identify Bayes-optimal mappings for various void definitions, including those based on simulated matter density and velocity fields.

Key takeaway

For astrophysicists and cosmology researchers working with large-scale structure data, this Bayesian void-finding approach offers a significant advancement over deterministic methods. You should consider integrating this graph neural network and flow-matching technique to extract more robust cosmological information from galaxy surveys. This method provides a more accurate, probabilistic representation of cosmic voids, potentially leading to improved constraints on cosmological parameters and a deeper understanding of the cosmic web.

Key insights

A graph neural network using flow matching probabilistically identifies cosmic voids, outperforming deterministic methods in cosmological information.

Principles

• Cosmic void finding is inherently probabilistic.
• Stochasticity can act as a useful regularization.
• Bayes-optimal mapping improves cosmological constraints.

Method

The method evolves "test particles" via a deep graph neural network using a flow-matching objective, trained on cosmological simulations to sample void catalogs from galaxy data.

In practice

• Use oversampling (Np > Nv) for unknown void counts.
• Employ the Hungarian algorithm for optimal particle-void assignment.
• Apply Friends-of-Friends (FoF) for clustering final test particles.

Topics

• Bayesian Void Finding
• Graph Neural Networks
• Flow Matching
• Cosmic Voids
• Cosmological Simulations

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• #AAAI2026 invited talk: machine learning for particle physics — Machine learning, especially graph neural networks, can overcome limiting assumptions in particle tracking to discover new physics.
• Uncertainty-Aware Prediction of Lung Tumor Growth from Sparse Longitudinal CT Data via Bayesian Physics-Informed Neural Networks — BPINNs can predict lung tumor growth with calibrated uncertainty from sparse CT data.
• FAIR Universe Weak Lensing ML Uncertainty Challenge: Handling Uncertainties and Distribution Shifts for Precision Cosmology — A new benchmark dataset and challenge addresses ML uncertainties and distribution shifts in weak lensing cosmology.
• Operator learning for solving Fokker-Planck equations with various initial conditions — A conditional normalizing flow PINN efficiently solves Fokker-Planck equations for diverse initial conditions by leveraging SDEs and time-weighted loss.
• Local graph estimation with pathwise false discovery control — Local graph estimation with pathwise feature selection (PFS) infers interpretable substructures around target variables.
• The Bayesian Workflow book is coming! — The "Bayesian Workflow" book offers practical guidance for applying Bayesian methods, addressing common challenges in statistical and computational practice.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.