Abstract
High-order numerical methods are essential for reducing numerical dispersion and dissipation in Computational Fluid Dynamics (CFD), yet they often suffer from instabilities in under-resolved or high-gradient regimes. This paper investigates the robustness and efficiency of entropy stable collocated discontinuous Galerkin methods. By utilizing Summation-By-Parts (SBP) operators and Simultaneous-Approximation-Terms (SATs), we demonstrate that entropy stable schemes (ES-C) provide the necessary mathematical framework to simulate complex turbulent flows and shocks where conventional methods fail. Our results show that while ES-C schemes incur higher per-step computational costs, their inherent stability and high-order accuracy make them more efficient for reaching target error thresholds in challenging 3D benchmarks.
Introduction
In the pursuit of high-fidelity Computational Fluid Dynamics, researchers face a persistent trade-off between high-order accuracy and numerical robustness. While high-order methods like Discontinuous Galerkin (DG) or collocation methods offer superior error reduction properties, they are notoriously fragile. When simulations encounter under-resolved turbulence or physical discontinuities such as shocks, these solvers often experience non-physical oscillations leading to numerical divergence or "blow-ups."
Traditional remedies—including artificial dissipation, spectral filtering, or limiters—frequently require problem-specific manual tuning and can inadvertently degrade the accuracy of the solution. This work explores an alternative: the development of solvers that are inherently stable by design. By mimicking continuous entropy physics at the discrete level, entropy stable methods aim to provide a "hands-off" reliability for the compressible Navier–Stokes equations.
Numerical Framework
We evaluate three distinct spatial discretization strategies within a unified framework based on SBP operators and SATs, implemented using the PETSc library:
- Entropy Stable (ES-C): This approach utilizes a two-point entropy conservative flux (Chandrashekar) paired with interface dissipation. It is mathematically proven to be nonlinearly stable by satisfying a discrete entropy inequality, effectively mimicking the continuous entropy analysis.
- Split Form (SF-KG): A middle-ground approach using the Kennedy and Gruber flux. While it preserves kinetic energy and improves stability over standard forms, it lacks a rigorous proof of entropy stability.
- Conventional Discontinuous Collocation (DC): The standard high-order approach where physical fluxes are evaluated directly at each node. This method is the most computationally efficient per degree of freedom but lacks inherent stability guarantees.
Temporal integration is handled via high-order schemes, specifically the 5th-order Dormand-Prince Runge–Kutta method, to ensure that time-stepping errors do not dominate the spatial discretization analysis.
Benchmarking Robustness
The methods were subjected to five increasingly rigorous 3D benchmarks to test the limits of their stability and accuracy:
- Isentropic Vortex & Viscous Manufactured Solution: Used to verify basic accuracy and confirm that the code achieves theoretical convergence rates.
- Taylor–Green Vortex (TGV): A standard test for the transition to turbulence at $Re = 1,600$. On coarse meshes, the conventional DC method failed for almost all polynomial degrees $p > 2$. In contrast, both SF-KG and ES-C remained stable up to $p=15$.
- Decaying Homogeneous Isotropic Turbulence (HIT): A high-stress test ($Re_\lambda = 194$, $Ma_t = 0.62$) involving the formation of "shocklets." This benchmark proved to be the breaking point for the SF-KG method, which failed at $p=1$ for several resolutions. Only the ES-C scheme successfully completed the simulation across all tested configurations.
- Supersonic Flow Past a Square Bar: An engineering-style problem ($Re_\infty = 10,000$, $Ma_ \infty = 1.5$) featuring complex bow shocks and flow separation. The DC method crashed almost immediately ($t \approx 0.0013$). While the SF-KG method lasted longer, it eventually failed as the shock stabilized. The ES-C method was the only scheme to run to completion ($t=100$), reliably capturing the complex shock-wake interactions.
Performance and Efficiency
A critical finding of this study is that high-order accuracy is not merely a theoretical advantage but a practical one. For a specific target error level (e.g., $10^{-4}$), high-order polynomials ($p=7$ or $p=15$) reached the solution in significantly less wall-clock time than low-order methods.
However, robustness comes at a price. The ES-C method is more computationally expensive per time step than the DC method due to the requirement of computing logarithmic averages and additional flux evaluations. Despite this overhead, the ES-C scheme's ability to remain stable on coarser meshes—where other methods would require extreme refinement or fail entirely—renders it a highly efficient choice for complex physics. The "sweet spot" for efficiency in these SBP-SAT methods typically occurs between polynomial degrees $p=3$ and $p=7$.
Conclusion
Entropy stable collocated discontinuous Galerkin methods represent a significant step toward reliable, high-order CFD. The ES-C scheme demonstrated unparalleled robustness, surviving every benchmark including supersonic shocks and under-resolved turbulence. While the computational kernels for entropy stability are more complex, the trade-off is a solver that finishes the simulation without the need for manual stabilization tuning. Future work will focus on optimizing these kernels for exascale architectures and integrating these robust methods into $hp$-adaptive mesh refinement frameworks.
Summary updated