Spatial field reconstruction with INLA: Application to simulated galaxies

TL;DR

This paper introduces a novel Bayesian inference-based method combining PCA, NMF, and INLA to efficiently reconstruct detailed spatial structures in astrophysical images, significantly reducing computational costs.

Contribution

It presents a new post-processing technique that enhances low-quality MCRT simulation outputs using advanced statistical methods for faster, accurate galaxy image reconstruction.

Findings

Reproduces high-quality galaxy images approximately 5 times faster.

Achieves median residuals below 20% in reconstructed images.

Effectively handles missing data and low signal-to-noise in astrophysical images.

Abstract

Aims. Monte Carlo Radiative Transfer (MCRT) simulations are a powerful tool for understanding the role of dust in astrophysical systems and its influence on observations. However, due to the strong coupling of the radiation field and medium across the whole computational domain, the problem is non-local and non-linear and such simulations are computationally expensive in case of realistic 3D inhomogeneous dust distributions. We explore a novel technique for post-processing MCRT output to reduce the total computational run time by enhancing the output of computationally less expensive simulations of lower-quality. Methods. We combine principal component analysis (PCA) and non-negative matrix factorization (NMF) as dimensionality reduction techniques together with Gaussian Markov random fields and the Integrated nested Laplace approximation (INLA), an approximate method for Bayesian inference, to detect and reconstruct the non-random spatial structure in the images of lower signal-to-noise or with missing data. Results. We test our methodology using synthetic observations of a galaxy from the SKIRT Auriga project - a suite of high resolution magneto-hydrodynamic Milky Way-sized galaxies simulated in cosmological environment by 'zoom-in' technique. With this approach, we are able to reproduce high photon number reference images $\sim5$ times faster with median residuals below $\sim20\%$.