Compact Gas-Kinetic Scheme (CGKS)

• CGKS is a high-order numerical method based on the BGK model that evolves a time-dependent gas distribution function to compute both interface fluxes and conservative variables.
• It achieves compactness by simultaneously updating cell-averaged values and gradients using a time-accurate kinetic solution, minimizing stencil size while handling inviscid and viscous effects in a unified manner.
• The scheme demonstrates robust performance and efficient convergence across various mesh types and dynamic flow regimes, validated on a range of benchmark compressible flow problems.

Searching arXiv for recent and foundational CGKS papers to ground the article. Compact gas-kinetic scheme (CGKS) denotes a family of high-order finite-volume, and in later work also finite-difference, methods for compressible Euler and Navier–Stokes equations that are built directly on kinetic theory rather than on a macroscopic Riemann solver. In CGKS, the central numerical object is a time-dependent gas distribution function obtained from the BGK model; its moments provide not only interface fluxes but also time-dependent interface conservative variables, which in turn permit simultaneous evolution of cell-averaged flow variables and their cell-averaged gradients. This coupling is the basis of the scheme’s compactness: high-order reconstruction is achieved with only immediate neighbors, while inviscid and viscous effects are handled in a unified kinetic formulation (Pan et al., 2014).

1. Definition and kinetic formulation

CGKS is rooted in the BGK model of the Boltzmann equation, written in three-dimensional form as

$f_t + u\cdot\nabla f = \frac{g-f}{\tau},$

where $f$ is the gas distribution function, $g$ is the local Maxwellian equilibrium, and $\tau$ is the collision time. The collision term satisfies the compatibility condition

$$\int \frac{g-f}{\tau}\,\boldsymbol{\psi}\,d\Xi = 0, \qquad \boldsymbol{\psi}=(1,u,\tfrac12(u^2+\xi^2))^T$$

or its corresponding multidimensional extension with full velocity components. Macroscopic conservative variables and fluxes are velocity moments of $f$,

$$W(x,t)=\int \boldsymbol{\psi}\, f\, d\Xi, \qquad F(x,t)=\int u \boldsymbol{\psi}\, f\, d\Xi,$$

and the first-order Chapman–Enskog truncation

$$f \approx g - \tau\left(u\cdot\nabla g + g_t\right)$$

recovers the compressible Navier–Stokes equations, with $\tau=\mu/p$ and $\Pr=1$ in the standard formulation (Ji et al., 2020).

The interface flux in CGKS is not produced by a piecewise-constant Riemann solution. Instead, it is obtained from the integral solution of the BGK equation along characteristics, which yields a time-dependent interface distribution function. A widely used second-order interface expression combines a collision-driven central part and kinetic left/right contributions with Heaviside splitting, and introduces a numerical collision time $f$0 to add dissipation near shocks. For smooth regions, a simplified “smooth” solver replaces the full discontinuity-resolving expression and reduces dissipation (Ji et al., 2021).

A persistent misconception is that “gas-kinetic scheme” only refers to a flux formula. In CGKS, the kinetic interface solution is used more broadly: it supplies both the interface flux and the interface conservative state at later times. This is the crucial structural distinction from classical Riemann-solver-based finite-volume methods and is what makes compact high-order reconstruction practical on short stencils (Liu et al., 2024).

2. Compactness, gradients, and high-order reconstruction

In CGKS, compactness means that high-order spatial reconstruction uses only the target cell and its immediate von Neumann neighbors. The mechanism is the direct evolution of both cell-averaged conservative variables and cell-averaged gradients. After the interface distribution function is evolved to $f$1, interface conservative values $f$2 are obtained by moments of $f$3, and the cell-averaged gradients are updated by Gauss–Green: $f$4 Thus each cell carries $f$5 and $f$6, and these quantities are evolved from the same kinetic interface solution rather than from separate governing equations (Liu et al., 2024).

Third-order CGKS on structured and unstructured meshes commonly uses a quadratic compact reconstruction. On 3D structured hexahedral meshes, the reconstruction polynomial is constrained by exact matching of cell averages in the target cell and its neighbors, together with least-squares matching of cell-averaged gradients in neighboring cells. On 3D hybrid unstructured meshes, the same principle is implemented with least-square-constrained reconstruction over tetrahedra, prisms, pyramids, and hexahedra, again using only von Neumann neighbors (Ji et al., 2020).

To treat discontinuities, CGKS incorporates HWENO-type or multi-resolution WENO reconstructions. The typical pattern is to combine a high-order compact polynomial with lower-order polynomials constructed on smaller sub-stencils, and to replace linear weights by nonlinear weights based on smoothness indicators. In later hybrid-mesh work, the lower-order part can even be built from single-cell information because cell-averaged gradients are already evolved and stored. This suggests that compactness in CGKS is not merely a geometric property of the stencil; it is a consequence of using kinetic evolution to generate additional local degrees of freedom that are reconstruction-ready (Ji et al., 2021).

Very-high-order CGKS extends this idea further. Sixth- and eighth-order compact GKS were constructed on fully compact stencils with spectral-like resolution, and later arbitrarily high-order compact reconstructions were organized through adaptive stencil extension governed by a discontinuity feedback factor. In that formulation, higher-order derivative content is multiplied by a feedback factor $f$7, so the polynomial collapses to first order near discontinuities while retaining full order in smooth regions. The stated advantage is that this avoids expensive high-order smoothness indicators on large stencils while preserving CFL numbers above 0.5 even up to ninth order (Zhao et al., 2019).

3. Time integration and space–time coupling

A defining feature of CGKS is that temporal accuracy is obtained from the same time-dependent interface evolution that produces the flux. Third-order compact schemes use a single-step second-order update for cell averages,

$f$8

with both $f$9 and $g$0 extracted from the gas-kinetic solver. Later high-order formulations adopt two-stage fourth-order multi-stage multi-derivative time stepping, which uses flux time derivatives available from the BGK evolution and achieves fourth-order time accuracy with only two stages (Ji et al., 2020).

This space–time coupling differs materially from standard explicit Runge–Kutta applied to a time-independent numerical flux. In CGKS, the interface distribution function is explicitly modeled as a function of time over the entire step, and the time derivative of the flux is available analytically or semi-analytically from the same kinetic solution. A plausible implication is that the favorable CFL properties reported for CGKS are tied not only to compact spatial reconstruction but also to this time-accurate flux construction (Zhao et al., 2019).

The compact finite-difference realization introduced later preserves the same idea in a different discretization form. There, a dual-grid strategy updates conservative variables on both a primary grid and an identical dual grid offset by half a mesh spacing. Numerical fluxes are reconstructed from physical fluxes at nodal and interfacial locations, while averaged spatial derivatives are updated from time-accurate interface states supplied by gas-kinetic evolution. The resulting framework is still compact and conservative, but its implementation is recast in finite-difference form on structured grids (Zhao et al., 16 Jun 2026).

Another misconception is that CGKS is only a spatial discretization paired with an arbitrary time integrator. The literature instead treats the method as a space–time coupled construction: the interface gas distribution function, its moments, and their time derivatives are part of the core algorithm, not an optional embellishment (Pan et al., 2014).

4. Meshes, geometry, and moving-domain extensions

CGKS first appeared as a compact third-order gas-kinetic scheme for compressible Euler and Navier–Stokes equations, then was extended to unstructured meshes, three-dimensional structured hexahedral meshes, and later three-dimensional hybrid unstructured meshes. On hybrid unstructured meshes, cells may be tetrahedra, prisms, pyramids, or hexahedra; triangular faces use 3 Gaussian points and quadrilateral faces use 4 Gaussian points. Iso-parametric or isoparametric mappings are used for geometric integration on distorted elements and faces (Pan et al., 2016).

For moving-domain problems, CGKS has been formulated in arbitrary Lagrangian–Eulerian form. In that setting, arbitrary cells are subdivided into tetrahedra and quadrilateral faces into triangles, the BGK equation is written in a moving frame, and the flux takes the ALE form involving $g$1. The method is coupled with a multi-stage multi-derivative time discretization, HWENO-type nonlinear reconstruction, and gradient compression factors, and it is constructed to satisfy the geometric conservation law to machine precision on moving meshes (Zhang et al., 2024).

A separate extension treats rotating coordinate frames and sliding meshes. There the kinetic equation includes centrifugal and Coriolis acceleration, rotating and stationary domains are coupled through a sliding interface with mortar elements and polygon clipping, and the compact third-order reconstruction uses cell averages and cell-averaged gradients on each side of the interface. The compact stencil is especially advantageous here because reconstruction near the sliding interface does not require large cross-interface neighborhoods (Zhang et al., 2022).

These developments indicate that compactness in CGKS is not restricted to static Cartesian settings. The same core ingredients—time-dependent kinetic interface evolution, direct gradient update, and compact reconstruction—have been transplanted to distorted hexahedra, hybrid polyhedra, rotating domains, and moving ALE meshes without changing the underlying kinetic logic (Zhang et al., 2024).

5. Robustness, efficiency, and acceleration

CGKS robustness in discontinuous flows is typically supported by three coupled ingredients: a full gas-kinetic interface solver with numerical collision time, nonlinear reconstruction in characteristic variables, and a mechanism that suppresses excessive gradient content near shocks. One widely used device is the gradient compression factor, defined from pressure jumps and normal and tangential Mach-number differences at face Gaussian points; updated cell-averaged gradients are multiplied by a cellwise factor $g$2, so that reconstruction becomes more dissipative in nonsmooth regions while remaining effectively unchanged in smooth flow (Ji et al., 2021).

For steady-state computation, two acceleration routes have been developed. The first is a two-level p-multigrid strategy in which the high-order level uses CGKS with cell averages and gradients, while the low-order level uses a first-order finite-volume scheme with only cell averages and an implicit backward Euler smoother. Restriction and prolongation act only on cell averages, because gradients do not participate in level transfer. In the reported tests, speedups range from about $g$3 to more than $g$4, with the paper summarizing the overall gain as typically “one order of magnitude” in convergence rate (Ji et al., 2021).

The second route is geometric multigrid on 3D unstructured meshes. There, mesh coarsening is performed by a two-step parallel agglomeration algorithm using a random hash for cell interface selection and a geometric skewness metric for deletion confirmation. A three-level V-cycle with explicit forward Euler on coarse levels and first-order KFVS residuals yields convergence-rate improvements up to roughly an order of magnitude and CPU-time reductions of about $g$5–$g$6, while the paper reports that coarse-grid LU-SGS offers only slight additional speedup and poorer parallel scalability (Liu et al., 2024).

A different efficiency bottleneck is memory usage in third-order CGKS on 3D unstructured meshes. The original HWENO reconstruction stores a coefficients matrix for the quadratic polynomial, which is expensive in memory. A later memory-reduction reconstruction splits the quadratic least-square regression into several linear least-square regressions based on the idea that second-order derivatives are “derivatives of derivatives.” In the hexahedral case, the reconstruction-module memory per cell drops from 276 doubles to 60 doubles, matrix assembly time drops from 1.3 s to 0, and overall solver CPU time is reduced by about 20 to 30 percent (Liu et al., 2024).

These acceleration and efficiency studies clarify an important point: the CGKS literature does not present compactness only as an accuracy device. Compactness also reduces communication, memory traffic, and implementation complexity for multigrid, matrix-free relaxation, and large-scale unstructured simulation (Liu et al., 2024).

6. Applications, validation range, and later developments

CGKS has been validated across a wide span of regimes and geometries. Early third-order compact schemes demonstrated accurate solutions for 1D and 2D Riemann problems, double Mach reflection, shock–vortex interaction, front-step flows, viscous shock tubes, lid-driven cavity flow, and smooth vortex propagation. Later 3D formulations covered sinusoidal-wave advection, circular cylinder and sphere flows, Taylor–Green vortex, homogeneous isotropic turbulence, ONERA M6 wing, and strong bow shocks around spheres (Ji et al., 2020).

On hybrid unstructured meshes, the reported range extends from incompressible-like boundary layers and low-Re cylinder and sphere flows to transonic wings and hypersonic vehicles. The gradient-compression-based compact scheme was tested on cases from incompressible to hypersonic flow and was explicitly described as broadly applicable. In subsonic and transonic cases, it matched reference drag, lift, separation angle, and pressure distributions; in supersonic and hypersonic cases, it preserved robustness and avoided numerical breakdown on irregular mixed-element meshes (Ji et al., 2021).

Recent developments also target application-specific regimes. A wall-modeled fifth-order CGKS uses a pressure-gradient-based non-equilibrium wall model coupled to the outer CGKS-5th solver. In the reported cylinder and transonic RAE 2822 airfoil cases, the wall model markedly improves near-wall predictions, especially the skin-friction coefficient, while adding less than 1% runtime overhead in a multi-GPU implementation. For the RAE 2822 case, the wall-normal direction is coarsened by a factor of twenty relative to wall-resolved requirements, with comparable coarsening in other directions (Yang et al., 29 Jun 2026).

The broad trajectory of CGKS suggests a methodological pattern rather than a single fixed algorithm. Foundational third-order compact GKS established the kinetic interface evolution and compact reconstruction principle; later work extended it to sixth-, eighth-, and ninth-order compact formulations, finite-difference realizations, ALE and rotating frames, multigrid acceleration, memory-reduced reconstruction, and wall modeling (Pan et al., 2014). A plausible implication is that “CGKS” now names a design paradigm: evolve interface distributions kinetically, update both averages and gradient-like information from that same evolution, and exploit those quantities to obtain compact high-order accuracy with strong robustness across smooth, shocked, steady, unsteady, static, and moving-mesh compressible flows.