Emulating CO Line Radiative Transfer with Deep Learning

TL;DR

This paper introduces COEmuNet, a deep learning model that accurately and efficiently emulates CO line radiative transfer in stellar atmospheres, significantly reducing computational costs for complex 3D simulations.

Contribution

The paper presents a novel 3D CNN-based surrogate model for CO radiative transfer, achieving high accuracy and 1000x speedup over traditional methods.

Findings

Median relative error of ~7% compared to classical solvers

1000 times faster than traditional radiative transfer calculations

Enables real-time visualization and large-scale simulation integration

Abstract

Modelling carbon monoxide (CO) line radiation is computationally expensive for traditional numerical solvers, especially when applied to complex, three-dimensional stellar atmospheres. We present COEmuNet, a 3D convolutional neural network (CNN)-based surrogate model that emulates CO line radiation transport with high accuracy and efficiency. It consists of an asymmetric encoder-decoder design that takes 3D hydrodynamical models as inputs and generates synthetic observations of evolved stellar atmospheres. The model is trained on data from hydrodynamic simulations of Asymptotic Giant Branch (AGB) stars perturbed by a companion. Given a set of input parameters, including velocity fields, kinetic temperature distribution, and CO molecular number densities, the COEmuNet model emulates spectral line observations with a median relative error of ~7% compared to a classical numerical solver of the radiative transfer equation, measured over seven frequency channels and arbitrary viewing directions. Besides, COEmuNet delivers a 1000 times speedup, enabling efficient model fitting to observational datasets, real-time visualization of simulations and progress toward integration in large-scale cosmological simulations.