Magritte, a modern software library for 3D radiative transfer: II. Adaptive ray-tracing, mesh construction and reduction

TL;DR

This paper introduces Magritte's adaptive ray-tracing and mesh optimization techniques that significantly reduce computational costs in 3D radiative transfer simulations while maintaining accuracy.

Contribution

It presents new adaptive ray-tracing and mesh reduction methods in Magritte, improving efficiency in 3D radiative transfer modeling.

Findings

Adaptive ray-tracing adapts to spatial mesh and uses hierarchical HEALPix discretization.

Mesh generation with Gmsh allows tailored high-quality meshes for Magritte.

Mesh reduction can decrease element count by an order of magnitude with minimal accuracy loss.

Abstract

Radiative transfer is a notoriously difficult and computationally demanding problem. Yet, it is an indispensable ingredient in nearly all astrophysical and cosmological simulations. Choosing an appropriate discretization scheme is a crucial part of the simulation, since it not only determines the direct memory cost of the model but also largely determines the computational cost and the achievable accuracy. In this paper, we show how an appropriate choice of directional discretization scheme as well as spatial model mesh can help alleviate the computational cost, while largely retaining the accuracy. First, we discuss the adaptive ray-tracing scheme implemented in our 3D radiative transfer library Magritte, that adapts the rays to the spatial mesh and uses a hierarchical directional discretization based on HEALPix. Second, we demonstrate how the free and open-source software library Gmsh can be used to generate high quality meshes that can be easily tailored for Magritte. In particular, we show how the local element size distribution of the mesh can be used to optimise the sampling of both analytically and numerically defined models. Furthermore, we show that when using the output of hydrodynamics simulations as input for a radiative transfer simulation, the number of elements in the input model can often be reduced by an order of magnitude, without significant loss of accuracy in the radiation field. We demonstrate this for two models based on a hierarchical octree mesh resulting from adaptive mesh refinement (AMR), as well as two models based on smoothed-particle hydrodynamics (SPH) data.