Algorithmic comparisons of decaying, isothermal, supersonic turbulence

TL;DR

This study compares various particle-based and grid-based numerical techniques for modeling decaying supersonic turbulence, highlighting differences in dissipation, resolution effects, and computational costs across methods.

Contribution

It provides a comprehensive comparison of multiple numerical codes for supersonic turbulence, revealing how implementation details affect dissipation and resolution.

Findings

Grid codes are less dissipative than SPH codes.

Artificial viscosity implementation impacts dissipation levels.

All codes show similar statistical quantities at scales larger than 1/32 of the resolution.

Abstract

Contradicting results have been reported in the literature with respect to the performance of the numerical techniques employed for the study of supersonic turbulence. We aim at characterising the performance of different particle-based and grid-based techniques on the modelling of decaying supersonic turbulence. Four different grid codes (ENZO, FLASH, TVD, ZEUS) and three different SPH codes (GADGET, PHANTOM, VINE) are compared. We additionally analysed two calculations denoted as PHANTOM A and PHANTOM B using two different implementations of artificial viscosity. Our analysis indicates that grid codes tend to be less dissipative than SPH codes, though details of the techniques used can make large differences in both cases. For example, the Morris & Monaghan viscosity implementation for SPH results in less dissipation (PHANTOM B and VINE versus GADGET and PHANTOM A). For grid codes, using a smaller diffusion parameter leads to less dissipation, but results in a larger bottleneck effect (our ENZO versus FLASH runs). As a general result, we find that by using a similar number of resolution elements N for each spatial direction means that all codes (both grid-based and particle-based) show encouraging similarity of all statistical quantities for isotropic supersonic turbulence on spatial scales k<N/32 (all scales resolved by more than 32 grid cells), while scales smaller than that are significantly affected by the specific implementation of the algorithm for solving the equations of hydrodynamics. At comparable numerical resolution, the SPH runs were on average about ten times more computationally intensive than the grid runs, although with variations of up to a factor of ten between the different SPH runs and between the different grid runs. (abridged)