ν \nu GAN: A Generative Adversarial Emulator for Cosmic Web with Neutrinos

Neerav Kaushal, Elena Giusarma, Mauricio Reyes · 2025

Read ν \nu GAN: A Generative Adversarial Emulator for Cosmic Web with Neutrinos by Neerav Kaushal, Elena Giusarma, Mauricio Reyes (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-03936.txt.

Introduction

Neerav Kaushal, Elena Giusarma, and Mauricio Reyes build a 2025 emulator for the cosmic web that includes neutrinos. Neutrinos are light particles that stream far and shape how cosmic structure forms. Running full simulations for many neutrino masses is slow and costly, yet surveys need such forecasts to read their data.

Their model, a generative network set against a critic, makes two dimensional maps of matter for masses from zero to 0.4 electron volts. The maps look like true maps and do not copy the training set. Key measures match within about five percent over a wide range of scales. The authors note that finer scales will need stronger models, such as diffusion based ones.

Catalog record

CREATORNeerav Kaushal, Elena Giusarma, Mauricio Reyes
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE36 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-03936

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

ν \nu GAN: A Generative Adversarial Emulator for Cosmic Web with Neutrinos Neerav Kaushal, Elena Giusarma, Mauricio Reyes Abstract Understanding the impact of neutrino masses on the evolution of Universe is a crucial aspect of modern cosmology. Due to their large free streaming lengths, neutrinos significantly influence the formation of cosmic structures at non-linear scales. To maximize the information yield from current and future galaxy surveys, it is essential to generate precise theoretical predictions of structure formation. One approach to achieve this is by running large sets of cosmological numerical simulations, which is a computationally intensive process. In this study, we propose a deep learning-based generative adversarial network (GAN) model to emulate the Universe for a variety of neutrino masses. Our model called ν \nu GAN (for neutrino GAN) is able to generate 2D cosmic webs of the Universe for a number of neutrino masses ranging from 0.0 eV to 0.4 eV. The generated maps exhibit statistical independence, lack correlations with training data, and very closely resemble the distribution of matter in true maps. We assess the accuracy of our results both visually and through key statistics used in cosmology and computer vision analyses. Our results indicate that samples generated by ν \nu GAN are accurate within a 5% error on power spectrum between k = 0.01 k=0.01 to k = 0.5 k=0.5 h ​ Mpc − 1 h\text{Mpc}^{-1} . Although this accuracy covers the mildly non-linear scales, consistent with other works and observations, achieving higher accuracy at fully non-linear scales requires more sophisticated models, such as diffusion models. Nevertheless, our work opens up new avenues for building emulators to generate fast and massive neutrino simulations, potentially revolutionizing cosmological predictions and analyses. This work serves as a proof-of-concept, paving the way for future extensions with higher-resolution 3D data and advanced generative models. I Introduction Neutrinos are among the most abundant particles in the Universe, with number densities only slightly lower than those of photons. In the early Universe, neutrinos were relativistic, contributing to the radiation energy density during that epoch. Unlike photons, neutrinos have now become non-relativistic and possess rest mass, thereby contributing to the total matter density of the Universe. This indicates that relic neutrinos can have significant effects on cosmological observables, particularly influencing the background evolution, the spectra of matter perturbations, and the anisotropies in the Cosmic Microwave Background (CMB). Next-generation large-scale structure (LSS) surveys, including Dark Energy Spectroscopic Instrument (DESI) [ 2 ] , eROSITA [ 48 ] , Euclid [ 5 ] , Nancy Grace Roman Space Telescope (WFIRST) [ 73 ] , and Rubin Observatory’s Legacy Survey of Space and Time (LSST) [ 44 ] , will map vast cosmological volumes with high precision, producing datasets that span a wide range of redshifts and cosmic epochs. These surveys are expected to significantly improve constraints on cosmological parameters, including the sum of neutrino masses, thereby enhancing our understanding of the fundamental role neutrinos play in the evolution of the Universe. Recent analyses from DESI and the Atacama Cosmology Telescope (ACT) [ 1 , 15 ] have already refined cosmological constraints on \sum m_{\nu} . Current upper bounds from joint analyses of CMB and LSS data lie in the range \sum m_{\nu}<0.06 – 0.08 eV at the 95% confidence level, assuming a minimal \Lambda CDM+ \sum m_{\nu} framework. Extracting the full scientific potential of these upcoming surveys requires theoretical predictions of comparable precision. In particular, accurate comparison between observations and theory demands rigorous modeling of the spatial distribution of matter and luminous tracers, especially in scenarios involving massive neutrinos. Analytic tools such as perturbation theory [ 13 ] , can provide reliable predictions on quasi-linear scales, [ 82 , 65 , 29 , 39 , 11 , 10 , 32 , 77 , 22 , 46 , 18 , 3 , 8 , 9 , 30 , 24 , 19 , 31 , 78 , 38 , e.g.] , but most of the cosmological information resides in the fully non-linear regime, which requires more sophisticated approaches. In the absence of a comprehensive analytical model, numerical simulations with massive neutrinos are crucial for studying these non-linear scales and validating observations. In particular, N-body simulations are among the most crucial tools for evolving cosmological matter fluctuations under gravity alone, allowing for direct comparisons with theoretical predictions. These simulations are instrumental in generating mock galaxy catalogs, computing covariance matrices, and optimizing observational strategies. Recent developments in N-body simulations incorporating massive neutrinos have significantly advanced our understanding of their impact on clustering at fully non-linear scales in real space [ 14 , 40 , 81 , 80 , 58 ] , the clustering and abundance of halos and cosmic voids [ 17 , 16 , 47 ] , and matter clustering in real space [ 79 , 12 ] . However, a significant drawback of cosmological N-body simulations is their high computational cost. A single simulation demands extensive computational resources and can take days or even weeks to complete. This computational bottleneck restricts the amount of information that can be extracted from observational data and tested against theoretical models. Consequently, there is a pressing need for faster methods to generate cosmological simulations that maintain both accuracy and reliability, thereby accelerating the process and enhancing our ability to extract and analyze cosmological information. Over the past decade, deep learning has emerged as a powerful tool for emulating high-resolution numerical simulations and accelerating cosmological predictions. Convolutional neural networks (CNNs) have been widely used for parameter inference, weak lensing map generation, and super-resolution of N-body simulations [ 52 , 49 , 63 , 34 , 4 , 37 , 55 , 57 , 41 , 54 , 85 , 84 , 67 , 64 ] . CNNs have also been successfully applied to model the impact of massive neutrinos on structure formation [ 25 , 26 ] , demonstrating their effectiveness in extracting cosmological information from complex datasets. These models can efficiently learn to emulate the evolution of cosmic structures from high-fidelity simulations, significantly reducing computational costs. More recently, generative models such as Generative Adversarial Networks (GANs) [ 63 , 74 , 59 , 6 ] , normalizing flows [ 21 , 42 ] , and diffusion models [ 52 , 50 , 51 , 69 ] have demonstrated remarkable success in generating synthetic cosmological data that closely match high-fidelity simulations. While CNNs and generative models have been extensively applied to a range of cosmological tasks, GAN-based emulators specifically tailored for massive neutrino cosmologies remain relatively unexplored. Our work aims to fill this gap by developing a GAN framework conditioned on neutrino mass, enabling rapid generation of cosmic web realizations across different neutrino mass scenarios. This approach builds on the success of generative models for cosmic structure emulation while addressing the unique challenges posed by massive neutrino effects on non-linear structure formation. In particular, we employ deep GANs to generate 2D cosmic web realizations of the Universe, conditioned on a range of neutrino masses. We assume a scenario of three degenerate neutrinos, where the electron, muon, and tau neutrino species share the same mass. The choice of GANs is motivated by their ability to effectively learn the complex probability distributions underlying the data, enabling

Read the complete text