Physics-informed variational autoencoders for cosmological field reconstruction and parameter inference

Abstract

In modern cosmology, predicting cosmological parameters is key to understanding the fundamental physical laws of the universe and dictating how cosmic structures are formed, evolve, and are observed. Parameters such as the total matter density (Ωm) and the amplitude of matter fluctuations (σ8) cannot be measured directly; instead, they have to be predicted based on complicated, high-dimensional simulation maps or observational data. Since, due to the complexity of the high-dimensional maps, traditional deep learning models often fail to give meaningful results, as they often learn shortcuts to statistical patterns that may appear correct but completely ignore the actual laws of physics. In this work, a Physics Informed Variational Autoencoder (PI-VAE) has been proposed as a combined framework for learning the compact representations of cosmological fields while directly imposing fundamental physical constraints to accurately reconstruct multi-channel cosmological fields, and directly infer key cosmological parameters from the learned latent space. By attaching a lightweight parameter regression head with VAE, the research looks into how physical information stored in the latent representation can be used for parameter predictions. To conduct this research, the CAMELS (Cosmology and Astrophysics with Machine Learning Simulations) dataset has been used, which comprises thousands of hydrodynamical simulations intended to systematically vary cosmological and astrophysical parameters. Our findings, supported by the ablation experiments, highlight the potential of physics-guided deep generative models for cosmological analysis by showing that physics-informed latent representations can simultaneously achieve meaningful cosmological parameter inference and accurate field reconstruction.