Particle hydrodynamics with accurate gradients: a comparison of different formulations

TL;DR

This paper compares various modern SPH formulations focusing on gradient accuracy, revealing that reproducing kernels perform best for instabilities, while shock dissipation approaches excel in shock tests, with a trade-off in computational cost.

Contribution

It provides a comprehensive comparison of SPH methods emphasizing gradient accuracy, highlighting the effectiveness of reproducing kernels and aLE gradients in different simulation scenarios.

Findings

Reproducing kernel gradients outperform others in instability simulations.

Shock dissipation methods are robust in shock tests but less so in instabilities.

Riemann solver approaches show resistance to instability growth at low resolution.

Abstract

We compare here several modern versions of SPH with a particular focus on the impact of gradient accuracy. We examine specifically an approximation to the "linearly exact" gradients (aLE) with standard SPH kernel gradients and with linearly reproducing kernels (RPKs) that fulfill the lowest order consistency relations exactly by construction. Most of the explored SPH formulations use shock dissipation (i.e. artificial viscosity and conductivity) with slope-limited reconstruction and parameters that trigger on both shocks and noise. We also compare with a recent particle hydrodynamics formulation that uses both RPKs and Roe's approximate Riemann solver instead of shock dissipation. Not too surprisingly, we find that the shock tests are rather insensitive to the gradient accuracy, but whenever instabilities are involved the gradient accuracy plays a crucial role. The reproducing kernel gradients perform best, but they are closely followed by the much simpler aLE gradients. The Riemann solver approach has some (minor) advantages in the shock tests, but shows some resistance against instability growth and at low resolution the corresponding Kelvin-Helmholtz simulations show substantially slower instability growth than the best shock dissipation approaches. Based on the battery of benchmark tests performed here, we consider a shock dissipation approach with reproducing kernels (our versions $V_3$ and $V_5$) as best, but closely followed by a similar version ($V_2$) that uses the simpler and computationally cheaper aLE gradients.