Fifth-order compact gas-kinetic scheme cuts simulation cost by nearly tenfold

Computational fluid dynamics has long faced a stubborn trade-off: second-order numerical schemes are cheap and robust but blur the fine details of turbulence, while high-order schemes resolve those details at a price that has kept them out of most industrial software. A new study published in Advances in Aerodynamics reports a systematic head-to-head comparison suggesting that a fifth-order compact gas-kinetic scheme can deliver the accuracy of a second-order scheme at roughly an order of magnitude lower computational cost, a result the authors describe as the first clear verification of the advantages of high-order compact gas-kinetic schemes for viscous flows containing discontinuities such as shock waves.

The study, carried out by Yaqing Yang, Fengxiang Zhao and Kun Xu, evaluates two representative members of the gas-kinetic scheme family on structured meshes: a newly developed fifth-order compact gas-kinetic scheme, abbreviated CGKS-5th, and a conventional second-order gas-kinetic scheme, GKS-2nd. Both schemes operate within a finite-volume framework in which numerical fluxes at cell interfaces are computed from a time-accurate solution of the Bhatnagar-Gross-Krook model of the gas distribution function. This kinetic formulation allows the method to compute inviscid and viscous fluxes in a single unified expression and to shift adaptively between equilibrium-state fluxes in smooth regions and non-equilibrium transport fluxes near discontinuities.

The two schemes differ fundamentally in how they reconstruct the flow solution inside each cell. GKS-2nd builds a linear polynomial from data in the target cell and its six face-neighboring cells, coupling the result with a Venkatakrishnan limiter to keep the solution stable near shocks. CGKS-5th instead assembles a quartic polynomial using a compact stencil that adds the twelve edge-neighboring cells, drawing on cell-averaged values, cell-averaged gradients and line-averaged derivatives that the gas-kinetic evolution provides simultaneously at each time step. Near discontinuities, the scheme applies a generalized ENO nonlinear reconstruction that adaptively falls back to a second-order, essentially non-oscillatory form, balancing high-order accuracy in smooth regions with robustness where the solution breaks down.

The comparison followed two complementary strategies. The first asked how much computation each scheme needs to reach a prescribed target resolution. The second fixed the computational budget and asked how much accuracy and resolution each scheme could deliver within it. Test cases ranged from subsonic to supersonic turbulent flows, covering both smooth solutions and flows with strong discontinuities, a regime where the authors note that the advantages of high-order methods had previously been verified mainly for smooth flows and where nonlinear reconstructions or limiters could in principle erode the high-order benefit.

The verification began with a three-dimensional advection of a density perturbation, a standard accuracy test with a known exact solution. Both schemes achieved their expected orders of accuracy, and the nonlinear versions behaved consistently with the linear ones. To reach the same error level, CGKS-5th required markedly fewer mesh degrees of freedom than GKS-2nd, an early indication of the efficiency gap that the later turbulent cases would quantify.

The first major benchmark was the subsonic Taylor-Green vortex at Mach number 0.1 and Reynolds number 1600, a canonical case for assessing how numerical methods capture turbulence generation, energy dissipation and vortex evolution. The linear CGKS-5th ran on a uniform mesh of 96 cubed cells, while the linear GKS-2nd needed 296 cubed cells to reach comparable resolution, a configuration the authors verified by direct testing. Comparing the time histories of volume-averaged kinetic energy and integrated enstrophy against a spectral solution computed on a 512 cubed mesh, the two schemes achieved equivalent resolution. In wall-clock terms, measured with multi-GPU parallel computing, CGKS-5th was nearly an order of magnitude faster for the same result.

The study then pushed the benchmark into territory that has rarely been used for this purpose: a supersonic Taylor-Green vortex at Mach number 2, which produces pronounced shock-vortex interactions. Because the classical fluctuation-form initial condition does not extend directly to high-speed regimes, the authors proposed a new initial condition tailored to supersonic conditions. Here CGKS-5th ran on 116 cubed cells against 384 cubed cells for GKS-2nd, with results from CGKS-5th on a 256 cubed mesh serving as the reference. The kinetic energy and enstrophy histories again matched between the two schemes at their respective meshes. Despite the higher algorithmic complexity introduced by the nonlinear reconstruction in the fifth-order scheme, the efficiency advantage over the limiter-based second-order scheme remained close to an order of magnitude.

The final and most demanding test fixed the computational budget instead of the resolution. For isothermal jet simulations at Mach numbers 0.9 and 1.5 with a Reynolds number of 400,000, the team designed a structured mesh of 352 by 176 by 176 cells for CGKS-5th and a mesh of 528 by 264 by 264 cells for GKS-2nd, with cell sizes scaled so that the total computation times differed by less than five percent over the simulated interval. At this matched cost, the differences between the schemes became stark. In the transonic jet, CGKS-5th resolved visibly more intricate turbulent structures, and its mean and root-mean-square axial velocities along the centerline agreed closely with experimental measurements from previous studies and with a large-eddy simulation using a sixth-order compact scheme on a grid of 24.5 million cells. The second-order scheme showed significant discrepancies, particularly in the downstream region between four and ten jet diameters where turbulent structures are most complex, and its predicted potential core length agreed less well with the reference data.

The supersonic jet at Mach 1.5 reinforced the pattern. CGKS-5th captured clearer small-scale structures and sharper internal shock features within the jet, and the oscillations in its computed mean axial velocity aligned with the locations of internal shock waves identified in the shadowgraph visualizations, a signature the second-order scheme failed to reproduce. The authors conclude that at matched computational cost, the fifth-order scheme delivers substantially higher resolution for turbulent flows involving shocks and small-scale vortical structures, which they argue underscores its practical value for scale-resolving turbulence simulations.

To make large-scale applications feasible, both schemes were implemented on multiple graphics processing units using CUDA for on-device parallel computation and MPI for communication between processes, with numerical experiments run in double precision on NVIDIA GeForce RTX 4090 GPUs. Scaling tests on the transonic jet case showed that CGKS-5th maintained a parallel efficiency above 95 percent when moving from four GPUs to eight, indicating that the scheme can exploit additional hardware effectively. Taken together, the results support the authors’ broader argument: the obstacles that have limited high-order methods in industrial practice, namely robustness near discontinuities and computational expense, may be less prohibitive for compact gas-kinetic schemes than commonly assumed, at least for the structured-mesh, viscous compressible flow problems examined here. The study frames its contribution as closing a validation gap rather than as a prescription, and the reported speedups apply to the specific benchmark configurations tested.

Subject of Research: Performance comparison of fifth-order compact and second-order gas-kinetic schemes for compressible flow simulations

Article Title: Performance evaluation of high-order compact and second-order gas-kinetic schemes in compressible flow simulations

Article References: Yang, Y., Zhao, F., & Xu, K. (2026). Performance evaluation of high-order compact and second-order gas-kinetic schemes in compressible flow simulations. Advances in Aerodynamics, 8(1), Article 31. https://doi.org/10.1186/s42774-025-00250-3

Image Credits: AI Generated

DOI: 10.1186/s42774-025-00250-3

Keywords: gas-kinetic scheme, high-order methods, computational fluid dynamics, compressible flow, turbulence simulation, shock-vortex interaction, Taylor-Green vortex, jet simulation, finite volume method, GPU computing, numerical accuracy, shock capturing