Variational autoencoder for generating realistic $N$-body simulations for dark matter halos

TL;DR

This paper demonstrates that a variational autoencoder can efficiently generate realistic dark matter density field images, matching theoretical predictions and enabling scalable cosmological data synthesis.

Contribution

It introduces a VAE-based method for compressing and generating cosmological simulation images, improving scalability and realism over previous approaches.

Findings

Generated images match the power spectra of real simulations.

The VAE provides a fast, scalable data generation method.

Results align with theoretical b1CDM predictions.

Abstract

In this paper, we explore the use of a variational autoencoder (VAE), a deep generative model, to compress and generate images of dark matter density fields from $\Lambda$CDM like cosmological simulations. The VAE learns a compact, low-dimensional representation of the large-scale structure, enabling both accurate reconstruction and generation of statistically realistic samples. We evaluated the generated images by comparing their power spectra to those of real simulations and the theoretical $\Lambda$CDM prediction, finding strong agreement with the state-of-the-art simulations. In addition, the VAE provides a fast and scalable method for generating synthetic cosmological data, making it a valuable tool for data augmentation. These capabilities can accelerate the development and training of more advanced machine learning models for cosmological analysis, particularly in scenarios where large-scale simulations are computationally expensive. Our results highlight the potential for generative artificial intelligence as a practical bridge between physical modeling and modern deep learning in cosmology.