EmulART: Emulating Radiative Transfer -- A pilot study on autoencoder based dimensionality reduction for radiative transfer models

TL;DR

This paper introduces EmulART, a novel autoencoder-based method that efficiently emulates high-resolution radiative transfer models of dust in space, significantly reducing computational costs while maintaining accuracy.

Contribution

It presents a new autoencoder and Bayesian inference approach for spectral and spatial emulation of radiative transfer models, enabling faster simulations with minimal information loss.

Findings

Successfully emulates reference models with less than 1% input information

Median residuals below 15% spectral, below 48% spatial-spectral

Achieves up to 50x speedup over traditional high-resolution simulations

Abstract

Dust is a major component of the interstellar medium. Through scattering, absorption and thermal re-emission, it can profoundly alter astrophysical observations. Models for dust composition and distribution are necessary to better understand and curb their impact on observations. A new approach for serial and computationally inexpensive production of such models is here presented. Traditionally these models are studied with the help of radiative transfer modelling, a critical tool to understand the impact of dust attenuation and reddening on the observed properties of galaxies and active galactic nuclei. Such simulations present, however, an approximately linear computational cost increase with the desired information resolution. Our new efficient model generator proposes a denoising variational autoencoder (or alternatively PCA), for spectral compression, combined with an approximate Bayesian method for spatial inference, to emulate high information radiative transfer models from low information models. For a simple spherical dust shell model with anisotropic illumination, our proposed approach successfully emulates the reference simulation starting from less than 1% of the information. Our emulations of the model at different viewing angles present median residuals below 15% across the spectral dimension, and below 48% across spatial and spectral dimensions. EmulART infers estimates for ~85% of information missing from the input, all within a total running time of around 20 minutes, estimated to be 6x faster than the present target high information resolution simulations, and up to 50x faster when applied to more complicated simulations.