nuGAN: Generative Adversarial Emulator for Cosmic Web with Neutrinos

TL;DR

This paper introduces nuGAN, a deep learning-based generative adversarial network that efficiently emulates the cosmic web influenced by neutrinos, enabling rapid generation of accurate 2D cosmological maps across various neutrino masses.

Contribution

The study presents a novel GAN model capable of generating realistic 2D cosmic web maps for different neutrino masses, offering a fast alternative to traditional simulations.

Findings

nuGAN generates maps with less than 5% error in power spectrum within certain scales.

The generated maps are statistically independent and closely resemble true cosmological data.

The approach is effective for mildly non-linear scales, with potential for extension to higher resolutions.

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 $\nu$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 $\nu$GAN are accurate within a 5% error on power spectrum between k=0.01 to k=0.5 h/Mpc. 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.