WorksPhysics and Astronomy
An implicit gas-kinetic scheme for internal and external flows
Abstract
The gas-kinetic scheme(GKS) is a promising computational fluid dynamics (CFD) method for solving the Navier-Stokes equations. It is based on the analytical solution of the BGK equation, which enables accurate and robust simulations. While GKS has demonstrated excellent properties (e.g., unified treatment of inviscid and viscous fluxes, inherent adaptive dissipation control), its application to classical engineering problems, such as aerodynamic flows and fluid machinery, remains underdeveloped compared to conventional CFD methods. This study bridges this gap by advancing GKS capabilities for real-world engineering challenges. First, the GKS is extended to a rotating coordinate frame, enabling efficient simulations of internal flows in turbomachinery. Second, the computational inefficiency of explicit GKS is addressed through an implicit time discretization using the generalized minimal residual method. The Jacobian matrices for inviscid/viscous fluxes are approximated using the first-order kinetic flux vector splitting scheme and the thin shear layer approximation to enhance robustness and computational efficiency further. Third, the shear-stress transport turbulence model is coupled to expand GKS's applicability to industrial turbulent flows. Numerical tests, including internal compressor rotor flow and external flow over a 3-D wingbody, validate the proposed method's accuracy and efficiency. Via our implicit scheme, the force coefficients of the 3-D wing-body flow with about five million mesh elements can converge after 500 steps. This work represents a practical advancement of GKS, demonstrating its potential to compete with established CFD solvers in high-Reynolds-number external and internal turbulent flows.
*Corresponding author
*Email addresses: [email protected] (Yue Zhang), [email protected] (Xing Ji), [email protected] (Kun Xu)
Introduction
The advancement of the aviation industry has led to an increasing demand for enhanced aerodynamic performance, improved propulsion efficiency, and heightened safety in aircraft design. Computational fluid dynamics (CFD) has emerged as a fundamental technology in this field, enabling engineers to simulate fluid flow characteristics numerically. By providing high-resolution data in complex flow scenarios, CFD significantly reduces the cost and duration associated with traditional wind tunnel testing. It addresses challenges such as external aerodynamics in aircraft shape design, wing-engine interactions, and internal flow issues in compressor and turbine components. This work aims to develop the gas-kinetic scheme (GKS) method specifically for aviation applications.
The GKS is a numerical approach rooted in kinetic theory, designed to solve the Euler and Navier–Stokes equations [1]. Unlike conventional solvers that rely on Riemann problems to compute inviscid flux and the central difference method for viscous fluxes, GKS operates at the mesoscopic scale, utilizing kinetic theory to capture the underlying physical processes with greater accuracy. By integrating both normal and tangential gradients of flow variables around cell interfaces during the time evolution of the gas distribution function, GKS effectively incorporates multi-dimensional effects [2]. The time-accurate surface distribution function allows for direct updates of flow variables without additional equations, facilitating the implementation of compact reconstruction methods. A third-order compact scheme was first developed by combining cell-averaged flow variables with pointwise variables at cell interfaces [3]. HWENO-type methods and two-step multi-resolution WENO reconstructions have been developed within the GKS framework to achieve high-order spatial accuracy [4, 5]. Furthermore, the compact GKS (CGKS) has been successfully implemented on both structured and unstructured meshes for two- and three-dimensional problems [6, 7, 8, 9, 10, 11, 12], which shows the potential of GKS in high-order accuracy and computational efficiency.
Accurate and efficient predictions of steady-state flow are of vital importance in design and optimization problems within aviation applications. However, explicit time integration methods can be computationally expensive when applied to steady problems, particularly when dealing with small-scale meshes for resolving boundary layers. To improve computational efficiency, implicit schemes are often employed for temporal discretization. Two widely used implicit methods are the lower-upper symmetric Gauss-Seidel (LU-SGS) and generalized minimal residual (GMRES) methods. The LU-SGS method, introduced by Yoon [13], decomposes the coefficient matrix into lower triangular, diagonal, and upper triangular components, solving the system through forward and backward sweeps. While LU-SGS is stable and straightforward to implement, it often exhibits slow convergence. In contrast, the GMRES method [14, 15], a Krylov subspace technique, minimizes the residual norm within a Krylov subspace, offering enhanced robustness and superior performance for ill-conditioned systems, albeit with higher memory requirements. Various preconditioners have been applied to GMRES to improve convergence further [18, 19] [16, 17]. Significant research has focused on enhancing the efficiency of GKS solvers, including the development of LU-SGS-based [20, 21] and GMRES-based [22, 23, 24] methods. In this paper, the GMRES method is employed to enhance convergence. As demonstrated in [18, 25], using an accurate Jacobian matrix can significantly improve the performance of the GMRES method. Therefore, this study utilizes a more precise Jacobian matrix than previous research for implicit GKS.
In addressing challenges in aviation applications, it is important to consider typical rotating components such as compressors and turbines in engines. A non-inertial rotating coordinate system is well-suited for analyzing the intricate flow dynamics associated with these components. In this framework, the effects of centrifugal and Coriolis forces become significant and must be incorporated into the governing kinetic equations [26]. Our previous work developed the GKS for rotating mesh problems using this approach [27]. The corresponding macroscopic governing equations are systematically derived from the BGK equation via the Chapman-Enskog expansion, facilitating a rigorous transition from kinetic to continuum descriptions. This study focuses on steady-state problems, ignoring the acceleration term in the gas distribution function to simplify flux calculations.
The structure of this paper is organized as follows: In Section 2, the gas-kinetic scheme in a rotating coordinate frame is derived. Section 3 presents the reconstruction methods for non-equilibrium and equilibrium distribution functions. Section 4 discusses the implicit time discretization using the GMRES method in detail. Section 5 introduces the shear-stress transport (SST) turbulence model and its discretization. In Section 6, the numerical results are presented, and finally, Section 7 concludes the paper.
Gas-Kinetic Finite Volume Scheme
BGK equation in rotating framework
In this section, the gas-kinetic BGK equation in a rotating frame is derived to deal with the rotation problem. The system returns to the inertial frame when the rotation speed equals zero. The gas-kinetic BGK equation in a rotating frame is
where is the gas distribution function, is the corresponding equilibrium state, and is the collision time. is the particle velocity in the rotating frame. And the acceleration in the rotating frame is
where is the angular velocity of the rotating frame, and is a position vector from the origin of rotation to the particle’s position. is centrifugal force and is Coriolis force. Denote as the particle velocity in the absolute inertial reference frame; the relationship among velocities is
where is convection velocity due to the frame rotation. According to , the acceleration term in the rotating frame can be expressed as . Then the BGK equation becomes
where can be defined by absolute velocity , such as . The collision term in the above equation describes the evolution process from a non-equilibrium state to an equilibrium one, with the satisfaction of the compatibility condition
where and ( is the number of internal degree of freedom, i.e. for three-dimensional diatomic gas). Based on the Chapman-Enskog Expansion [27], the Euler and N-S equations in the rotating frame can be obtained. The N-S equations in a rotating frame are
where , , , , , and are the density, absolute velocity, pressure, temperature, energy, enthalpy, and viscosity stress of fluid. With the gradient of temperature and viscosity stress equal to zero, the N-S equations become the Euler Equations.
Finite volume discretization
The finite volume method is used to discretize the system. The whole domain is discretized into small cells
For each cell , the boundary faces can be expressed as
Taking moments of the BGK equation (1) and integrating over the cell , the semi-discretized form of the finite volume scheme can be written as
where is the cell average conservative value, is the volume of cell , is the flux passing through the fixed element surface, presents the flux due to the mesh motion, is the normal direction of cell surface and is the source term due to frame rotation.
The integration of flux can be approximated (the index is omitted )
where is the cell surface area, and is the coordinate of the cell surface center. The flux through the surface is expressed as
There is a conflict that the particle velocity across the cell surface is the relative velocity, but the momentum and energy that particles carry is about the absolute velocity. To solve this problem, the transformation from the flux with relative velocity to the flux with absolute velocity is used here. The components of are determined through the following transformation relations:
The flux is defined by the relative velocity, and can be evaluated through the following integral formulation:
where denotes the origin of the local coordinate system with x-direction aligned with , and , and denotes the rotation matrix, with
and reduces to when . The microscopic velocities in the local coordinate system are obtained through the transformation .
Gas evolution model
In order to construct the numerical fluxes at , the integral solution of the BGK equation Eq. (1) is
with the trajectory
In Eq. (5), is the initial gas distribution function, and is the corresponding equilibrium state. are the initial position by tracing back particles at time back to .
Before constructing the initial distribution function and equilibrium state , first denote
In the following derivation, quadratic terms of time will be ignored directly. With the consideration of possible discontinuity at an interface, the initial distribution is constructed as
where is the Heaviside function. and are the initial gas distribution functions on the left and right sides of the interface, which are determined by corresponding initial macroscopic variables and their spatial derivatives. With the second-order accuracy, is constructed by Taylor expansion around
for . Based on the Chapman-Enskog expansion, is given by
where is the equilibrium distribution function defined by the macroscopic variables at both sides of a cell interface, are defined by the spatial derivatives of
and
are determined by the compatibility condition
Substituting Eq. (7) and (8) into (6), the initial gas distribution has following form
Then, the equilibrium distribution is defined by the Taylor expansion
where and are determined from the reconstruction of macroscopic flow variables presented in section 3.3, and is obtained by the compatibility condition (9).
In a smooth flow region, the collision time for viscous flow is determined by , where is the dynamic viscosity coefficient and is the pressure at the cell interface, and for inviscid flow, the collision time should be set as zero. In order to properly capture the unresolved shock structure, additional numerical dissipation is needed. The physical collision time in the exponential function part can be replaced by a numerical collision time
where and denote the pressure on the left and right sides of the cell interface.
By substituting Eq. (10) and Eq. (11) into Eq. (5) with and , the second-order time-dependent gas distribution function at a cell interface becomes
The fluxes in Eq. (4) can be obtained by taking the moments of the above distribution function. The calculation of moments can be found in [1].
In order to account for the Prandtl number not equal to 1, additional correction heat flux should be added to the energy flux:
where can be obtained through the viscous flux . The viscous flux can be obtained by
Spatial Reconstruction
In our work, the Green-Gauss reconstruction is used to reconstruct the macroscopic flow variables, which is more robust than the least square method when dealing with a large aspect ratio mesh. Furthermore, the discontinuity feedback factor is introduced to enhance the shock-capturing capabilities.
Green-Gauss reconstruction
The target cell is denoted as , with the cell center represented by . The neighboring cells are denoted as for . The cell-averaged values of cell are denoted as . The classical Green-Gauss reconstruction with only cell-averaged values is adopted to provide the linear polynomial for the sub-stencil.
where is the area of the cell’s surface and is the surface’s normal vector.
Discontinuity feedback factor
A discontinuity feedback factor (DF) is introduced to modify the reconstruction process to enhance shock-capturing capabilities adaptively. This approach analyzes the reconstructed interface values to identify regions where discontinuities will likely propagate in subsequent timesteps, allowing for more accurate shock tracking. It was first proposed in [11], and several improvements have been made in [27]. Denote as discontinuity feedback factor at targeted cell
where is the DF obtained at the th Gaussian point of the interface around cell , which can be calculated by
where is pressure, and are the Mach numbers defined by normal and tangential velocity, and superscript , denote the left and right values of the Gaussian points. Then, the Green-Gauss reconstruction is modified as
Reconstruction of equilibrium states
After reconstructing the non-equilibrium state, a kinetic weighted average method can be used to get the equilibrium states and their derivatives [28],
For the normal derivatives, the above solution is further modified according to the idea in linear diffusive generalized Riemann problem (dGRP) [29]
where and are the coordinates of the left and right cell centroid, and is the normal vector of the interface. By adding a penalty term in Eq. (12), the whole scheme is essentially free from the odd-even decoupling phenomenon [30].
Solution updates and temporal discretization
Implicit scheme for the fluid governing equations
According to the semi-discretization Eq. (2), the back Euler method can be write as
With the approximation of
The scheme can be
The spatial residual is the time-averaged value of the flux residual over the global time step of CFL number 0.5. Moreover, the GKS flux through flux translation ( Eq. (3) and Eq. (4) ) is used to calculate the flux residual. To obtain the , the flux Jacobian matrix is calculated by the first-order KFVS scheme for inviscid flux, and thin shear layer (TSL) approximation for viscous flux. The derivation of the Jacobian matrix of the face flux is shown in Appendix § 9. And the source term Jacobian matrix is given by
Treatment of boundary conditions
The ghost cells are used to realize boundary conditions. The ghost cell values and inner cell values has the relationship
Take a differential on both sides
e.g.
The is obtained by the chain rule. In this case, the implicit treatment for the boundary condition is constructed.
GMRES method
The complete linear system for steady-state flow calculations is described by Eq. (13) and Eq. (14). The generalized minimal residual (GMRES) approach can be employed to solve this system, which belongs to the family of Krylov subspace techniques. The detailed implementation steps of our GMRES solver are outlined in Algorithm 1.
Algorithm 1 Implementation of GMRES method
Input:
Output:
for do
for do
communicate to other processors
end for
solve
end for
return
The implementation incorporates preconditioning techniques to enhance convergence. Specifically, a Lower-Upper Symmetric Gauss-Seidel (LUSGS) scheme is utilized as the left preconditioner, while an identity matrix is maintained as the right preconditioner. For a given matrix , a decomposition into is performed, where , , and represent the lower triangular, diagonal, and upper triangular components, respectively. The left preconditioner is defined as:
To solve the linear system , it can be expressed as:
The parallelization of the code is based on the Message Passing Interface (MPI). The physical domain is partitioned using the METIS library to achieve load balancing across processors. Recursive bisection is employed through METIS_PartGraphRecursive() for partitioning with eight or fewer subdomains. The multilevel k-way partitioning algorithm METIS_PartGraphKway() is utilized for larger numbers of partitions. This partitioning strategy minimizes communication overhead while maintaining balanced computational loads across processors. The recursive bisection method is preferred for smaller partition counts as it produces more geometrically compact subdomains, while k-way partitioning is more effective for larger numbers of partitions. In the context of MPI-based parallelization, the GMRES algorithm can be naturally extended to parallel computing. However, the LUSGS preconditioning method cannot achieve parallel consistency. Therefore, off-diagonal elements in the parallel buffer are ignored, and the LUSGS method is applied independently within each processor core, requiring only one communication per iteration. Although this approach does not maintain serial-parallel consistency, experimental results indicate that the method’s convergence remains unaffected. Additionally, to enhance the efficiency of LUSGS, Reverse Cuthill-McKee (RCM) reordering is employed.
Turbulence Models
Shear-stress transport (SST) turbulence model
The Shear-Stress Transport (SST) turbulence model, developed by Menter and first presented in [31], marks a notable improvement in turbulence modeling. This model refines the original Menter model by modifying the formulation of turbulent eddy viscosity and adjusting one constant (). The governing equations, expressed in conservation form, can be represented as a coupled system:
The viscous stress tensor is given by:
The turbulent eddy viscosity is calculated as:
Constants are blended from inner (1) and outer (2) regions using:
where and are the constants for the inner and outer regions, respectively. Additional functions are defined as follows:
where represents the density, is the turbulent kinematic viscosity, is the molecular dynamic viscosity, is the distance from the field point to the nearest wall, and denotes the vorticity magnitude, defined as
By [32], a production limiter is introduced by modifying the production term in the -equation; thus, the modified SST model is given by
where and denote the production and destruction terms in the -equation, and , , and denote the production, destruction, and cross production terms in the equation.
The model also requires specific boundary conditions. For farfield or inlet conditions, the following constraints apply:
Wall boundary conditions are defined as:
The complete set of model constants is as follows:
After determining the turbulent viscosity , the gas-kinetic scheme is integrated with the SST model by adjusting the collision time as follows:
where denotes the pressure. The heat flux correction term is expressed as:
with representing the heat flux, where 0.72 and 0.9 are the Prandtl numbers for air and turbulence, respectively.
Discrete form of the turbulence model
The advection, diffusion, and source terms of the turbulence model are treated separately in this work. A simple first-order upwind scheme is utilized for inviscid flux, with the velocity at the cell interface obtained from the kinetic equilibrium state at that interface. For viscous flux, a second-order central scheme is applied.
The -equation and -equation are addressed separately in the two-equation SST turbulence model. The Jacobian of the inviscid term and viscous term can be derived directly by ignoring the derivative of the conservative variables. For the source term, implicit treatment is applied to the negative part to ensure the positive definiteness of the turbulence variables, while explicit treatment is used for the positive part. The Jacobian of the source term is approximated by
Test case
The time step is the time step defined in each cell
where is the CFL number, and , , and represent the magnitude of related velocities, sound speed and kinematic viscosity coefficient of cell , respectively.
All simulations were conducted on the SUGON computation platform using dual AMD EPYC 7285 processors, each equipped with 32 cores at a base frequency of 2.5 GHz. The system architecture leverages the Zen microarchitecture, and the computation network is supported by a 100Gb InfiniBand (IB) connection. The code is compiled using GCC version 7.3.1 with the -O3 optimization flag and is linked to Open MPI version 4.1.5.
The turbulent flow past the Onera M6 wing
This test case is the three-dimensional turbulent transonic flow at a Mach number of Ma = 0.8395 over an Onera M6 wing. The Reynolds number based on the mean aerodynamic chord () of the wing, is , and the angle of attack (AoA) is . This case serves as a benchmark CFD validation problem for external flows due to its complex flow features, including local supersonic regions, shock waves, and turbulent boundary layer separation, while maintaining a relatively simple geometry.
The website provides the computational mesh [33]. The total mesh number used in this test case is 2,608,200. The first layer of cells is located at a distance of from the wall. The surface mesh is shown in Figure 1a. The surface distribution defined by the distance of the cell center and face is shown in Figure 1b, which shows the first layer of cells is located at the wall with a distance of less than 1. The parallel computing efficiency of

(a) The surface mesh

(b) The surface distribution
Figure 1. The surface mesh of the M6 wing with 2,608,200 cells
the GMRES solver with 100 iterations in the M6 wing test case is shown in Table 1. The results show that the GMRES solver can remain 90% parallel efficiency up to 256 cores. The convergence history of the M6 wing test case is shown in Figure 2(a). The CFL number
| Cores | Wall time (s) | Actual speedup | Parallel efficiency |
| 64 | 732 | 1.00 | 100.00% |
| 128 | 380 | 1.93 | 96.32% |
| 256 | 203 | 3.61 | 90.15% |
| 512 | 132 | 5.55 | 69.32% |
Table 1. Wall clock time and parallel computing efficiency of the GMRES solver with 100 iterations in the M6 wing test case

Figure 2. (a) The convergence history of the M6 wing test case (b) The coefficient of pressure contour of flow , , 3.06 degrees attack angle past the M6 wing
is set as an exponential function increasing from 1 to 100 during the first 100 steps, and remains at 100 afterwards. The convergence history shows the residual of the M6 wing test case is reduced to after 600 steps, and reduces slowly afterwards. Finally, 2000 steps are run for this case. The coefficient of pressure distribution of the M6 wing is shown in Figure 2(b). The coefficient of pressure distribution at different spanwise positions of the M6 wing
is shown in Figure 3, which shows the pressure distribution is similar to the experimental data [34] and can accurately capture the position of the shock wave.
Figure 3. The coefficient of pressure distribution of the M6 wing at different spanwise positions with a Reynolds number of 5.0E+06 and a Mach number of 0.85. Moreover, the reference temperature is 310.92K, which is set as the incoming total temperature. Seven different angles of attack are considered, which are 2.5°, 2.75°, 3°, 3.25°, 3.5°, 3.75°, and 4°. The medium point-matched grid (L3, 5, 111, 808 hexahedral elements) provided by the DPW Committee is used in this test, and the surface grid is shown in Figure 4a. The first layer of cells is located at the wall with a distance of from the wall, while the contour is shown in Figure 4b.
NASA Common Research Model (CRM) Wing-Body buffet study
The buffet study presented in this paper corresponds to test case 2 of the 5th AIAA Drag Prediction Workshop. This case is a genuine guide to test the simulation capacity of the CFD solver in transonic flow with complex geometry. The geometry is still the Common Research Model (CRM) Wing-Body configuration. The incoming flow is a subsonic flow
The parallel computing efficiency of the GMRES solver with 100 iterations in the CRM Wing-Body Buffet Study is shown in Table 2. The results show that the GMRES solver can remain about 80% parallel efficiency up to 512 cores.
| Cores | Wall time (s) | Actual speedup | Parallel efficiency |
| 64 | 1507 | 1.00 | 100.00% |
| 128 | 784 | 1.92 | 96.11% |
| 256 | 412 | 3.66 | 91.44% |
| 512 | 244 | 6.18 | 77.20% |
| 1024 | 181 | 8.33 | 52.04% |
Table 2. Wall clock time and parallel computing efficiency of the GMRES solver with 100 iterations in the CRM Wing-Body Buffet Study
The CFL number is set as an exponential function increasing from 1 to 100 during the first 100 steps, and remains at 100 afterwards. The characteristic of the CRM Wing-Body Buffet Study is shown in Figure 5. The results are compared with the NTF test data [35] and the ONERA-elsA solver [36] results. The results show that the lift coefficient of our solver is closer to the NTF test data than the ONERA-elsA solver. The drag characteristics are plotted in terms of the idealized profile drag defined as , where is the aspect ratio, is the drag coefficient, and is the lift coefficient. The results show that the drag coefficient of our solver is lower than the NTF test data, and the ONERA-elsA solver is higher than the NTF test data.

(a) The lift coefficient vs. angle of attack

(b) The drag coefficient vs. lift coefficient
Figure 5. The characteristic of the CRM Wing-Body Buffet Study of Reynolds number 5.0E+06 and Mach number 0.85

(a) The surface mesh

(b) The surface distribution
Figure 4. The computational mesh of the CRM Wing-Body configuration
The convergence history of the CRM Wing-Body Buffet Study at angle of attack is shown in Figure 6. The results show that the lift and drag coefficients converge after 500 steps. A total of 2000 steps are run to reach the convergence.

Figure 6. The convergence history of the CRM Wing-Body Buffet Study at angle of attack
The pressure coefficient distribution at station 10 () at different angles of attack is shown in Figure 7. The results show that the pressure coefficient distribution is similar to the experimental data [35] and can capture the position of the shock wave.

Figure 7. The wing pressure coefficient distribution at station 10 () at different angle of attack
The contour of Mach number at different spanwise positions at angle of attack is shown in Figure 8. The contour shows the interaction between the shock wave and the boundary layer separation. Moreover, the contour of pressure coefficient and the surface streamlines are shown in Figure 9, indicating a mass flow separation starts at about 50% of the chord.

(a) The station 4 ()

(b) The station 6 ()

(c) The station 8 ()

(d) The station 10 ()
Figure 8. The contour of Mach number at different spanwise positions at angle of attack

Figure 9. The pressure coefficient distribution and the surface streamlines at angle of attack
NASA Rotor 67
NASA Rotor 67 is a typical transonic rotor blade, which is widely used in the field of turbomachinery. The geometry database and experimental data can be found in [37, 38]. The basic information of the blade is shown in Table 3.
| Parameter | Value |
| Number of blades | 22 |
| Design rotational speed (rpm) | 16043 |
| Design mass flow rate (kg/s) | 33.25 |
| Design pressure ratio | 1.63 |
| Tip speed (m/s) | 429 |
| Tip clearance at design speed (cm) | 0.1061 |
| Relative Mach number at tip inlet | 1.38 |
| Aspect ratio (mean blade height/root axial chord) | 1.56 |
| Solidity at root | 3.11 |
| Solidity at tip | 1.29 |
| Tip diameter at inlet (cm) | 51.4 |
| Tip diameter at outlet (cm) | 48.5 |
| Hub-to-tip ratio at inlet | 0.375 |
| Hub-to-tip ratio at outlet | 0.478 |
Table 3. Basic information of NASA Rotor 67 [37]
The computational mesh of NASA Rotor 67 is shown in Figure 10. The positions of the inlet and outlet are set at the experimental station. The total number of cells is 941,824. The H-O-H type of topology is used to mesh the blade surface, and the first layer of cells is located at the blade surface with a distance of to the blade surface. As shown in Figure 10(b), the tip clearance is set as 0.6 times of the design tip clearance (0.1061 cm) by considering the vena contraction effect, instead of refining the mesh near the blade tip.
Figure 10. (a) The mesh at blade surface; (b) The mesh projection on S2 surface; (c) The mesh projection on S1 surface. The computational mesh of NASA Rotor 67
At the inlet, a total pressure of 101325 Pa and a total temperature of 288.15 K are specified, with absolute flow direction parallel to the rotation axis. Static pressure is prescribed at the outlet at the shroud, with radial equilibrium determining pressures at other positions. Performance characteristics are computed using outlet pressures ranging from 105,000 Pa to 141,000 Pa. The blade and hub walls rotate at 16043 rpm with adiabatic wall conditions, while periodic boundaries are used in the circumferential direction. The shroud is set as a static wall. For initialization, the lowest pressure case starts with zero velocity and inlet to-
tal conditions, while subsequent cases use the converged solution from the previous pressure level.
The computed choked mass flow rate is 34.67 kg/s, while the corresponding experimental result is 34.96 kg/s, with a relative error of 0.84%. The computed performance characteristics with normalized mass flow rate as the parameter are shown in Figure 11. The adiabatic efficiency is defined as
Figure 11. (a) Total pressure ratio; (b) Adiabatic efficiency. Performance characteristics of NASA Rotor 67
where and are the outlet and inlet total temperature respectively, is the outlet total temperature that would be achieved through an isentropic process starting from the inlet total pressure and ending at the outlet total pressure, and are the outlet and inlet total pressure respectively, and is the specific heat ratio. The peak efficiency and near-stall points can be well captured and agree with the experimental data.
The computation time for 1000 steps is about 1506s with 128 cores. For each working point, 2000 steps are run to reach the convergence. Figure 12 shows the convergence history of mass flow rate, total pressure ratio, and adiabatic efficiency at the peak efficiency point. The CFL number is an exponential function increasing from 1 to 100 during the first 100 steps, and remains at 100 afterwards. The results show that the performance parameters converge after approximately 300 steps.

Figure 12. (a) The convergence history of mass flow rate; (b) The convergence history of total pressure ratio; (c) The convergence history of adiabatic efficiency. The convergence history of NASA Rotor 67 at peak efficiency point
The comparison of relative Mach number contours between computed results and experimental data at the peak efficiency point and near the stall point is shown in Figure 13 at different spanwise positions. The computed results can capture the position of the shock wave from the experimental data. However, at the near stall point, at 30% span, the computed results show flow separation on the suction surface. In contrast, the experimental data show separation on the pressure surface, indicating that the computed results have a larger angle of attack corresponding to a lower axial velocity of the incoming flow. The comparison of circumferentially averaged outlet parameters between computed results and experimental data at the peak efficiency point and near the stall point is shown in Figure 14. The results of total pressure and total temperature are normalized by the inlet total pressure and total temperature, respectively. At the peak efficiency point, the computed results show good agreement with the experimental data. However, at the near stall point, the computed results show higher total pressure and total temperature than the experimental data between 20% and 40% span, which agrees with the result of relative Mach number contours at 30% span. The higher total pressure and total temperature indicate more work input per unit mass flow due to the lower mass flow rate in this region.
Figure 13. Comparison of relative Mach number contours for NASA Rotor 67: numerical results and experimental contours [37] at peak efficiency point (left two) and near stall point (right two)
Figure 14. comparison of circumferentially averaged outlet parameters of NASA Rotor 67 between computed results and experimental data [37]
NASA Rotor 37
NASA Rotor 37 was initially designed for an experimental study of a four-stage axial compressor. The design parameters of this four-stage compressor were chosen to be typical of the inlet stage of an aircraft engine’s high-pressure compressor, with experimental results available in reference [39]. Later, NASA conducted more detailed measurements on one of the rotors in isolation (without inlet guide vanes and outlet stator) and named it Rotor 37. In 1994, the International Gas Turbine Institute of the American Society of Mechanical Engineers (ASME/IGTI) used NASA Rotor 37 as a blind test case to evaluate turbomachinery CFD codes [40], aiming to assess the capability of CFD programs in predicting compressor flow fields at that time. NASA Rotor 37 has attracted considerable numerical simulation studies [41] [42, 43, 44] not only because of its comprehensive experimental data and simple isolated rotor flow with accurately definable boundary conditions allowing fair and detailed evaluation of turbomachinery CFD, but also due to its representativeness among transonic rotors making it essential to understand its flow structure and loss mechanisms. The basic geometric and aerodynamic parameters of NASA Rotor 37 are shown in Table 4.
| Parameter | Value |
| Number of blades | 36 |
| Design rotational speed (rpm) | 17188.7 |
| Design mass flow rate (kg/s) | 20.19 |
| Design pressure ratio | 2.106 |
| Tip speed (m/s) | 454.14 |
| Tip clearance at design speed (cm) | 0.0356 |
| Relative Mach number at tip inlet | 1.48 |
| Relative Mach number at hub inlet | 1.13 |
| Aspect ratio (mean blade height/root axial chord) | 1.19 |
| Solidity at tip | 1.29 |
| Hub-to-tip ratio at inlet | 0.7 |
Table 4. Basic information of NASA Rotor 37 [40]
The computational mesh of NASA Rotor 37 is shown in Figure 15. The positions of the inlet and outlet are set at the experimental station. The total number of cells is 972,800. The H-O-H type of topology is used to mesh the blade surface, and the first layer of cells is located at the blade surface with a distance of to the blade surface. As in the case of NASA Rotor 67, the tip clearance is set as 0.6 times the design tip clearance (0.0356 cm) by considering the vena contraction effect, instead of refining the mesh near the tip of the blade.

Figure 15. The computational mesh of NASA Rotor 37
The boundary conditions and numerical settings are the same as those of NASA Rotor 67. The computed choked mass flow rate is 20.96 kg/s, while the corresponding experimental result is 20.93 0.14 kg/s, which falls within the experimental error band. The computed performance characteristics with normalized mass flow rate as the parameter are shown in Figure 16. The comparison shows that our computed efficiency is slightly lower than
Figure 16. Performance characteristics of NASA Rotor 37
the experimental value, and the stall margin is smaller than the experimental result. The
convergence history at the 98% choked mass flow rate point of mass flow rate, total pressure ratio, and adiabatic efficiency is shown in Figure 17. The results show that the performance parameters converge after approximately 4000 steps. It is also noted that significant variations only occur in the first 500 steps, while the subsequent convergence process is relatively smooth with variations within 0.5%.

(a) The convergence history of mass flow rate

(b) The convergence history of total pressure ratio

(c) The convergence history of adiabatic efficiency
Figure 17. The convergence history of NASA Rotor 37 at 98% choked mass flow rate point
The relative Mach number contours of NASA Rotor 37 at 98% choked mass flow rate are shown in Figure 18. The contours show the flow structure of NASA Rotor 37 at 98% choked mass flow rate. Figure 19 shows the circumferentially averaged outlet parameters (total pressure, total temperature, adiabatic efficiency, and flow angle) of NASA Rotor 37 at 98% choked mass flow rate. At 98% choked mass flow rate, the computed total temperature and total pressure under the 30% span are higher than the experimental values, which is related to the boundary condition settings. The entire hub is set as a rotating wall, while in the experiment, only the rotor section of the hub is rotating, resulting in higher computational results. Between 40% and 80% span, the total pressure matches well with the experiment, but the total temperature is slightly higher than the experimental value, thus corresponding to lower efficiency. Meanwhile, the computed absolute flow angle is slightly smaller than the experimental value, corresponding to larger separation and higher losses.

(a) 90% span

(b) 70% span

(c) 30% span
Figure 18. The relative Mach number contours of NASA Rotor 37 at 98% choked mass flow rate
Figure 19. The circumferentially averaged outlet parameters of NASA Rotor 37 at 98% choked mass flow rate
Conclusion
In this paper, an implicit GMRES solver based on the gas-kinetic scheme was developed. First, the gas-kinetic scheme was extended to accommodate a rotating frame of reference. Subsequently, the implicit GMRES solver was formulated, with the Jacobian matrix derived using the kinetic flux vector splitting (KFVS) approximation for inviscid fluxes and the thin-layer approximation for viscous fluxes. To address turbulent flow, the shear-stress transport turbulence model was employed. The solver was validated through four test cases, which included external flows (Onera M6 wind and CRM Wing-Body Buffet Study) and internal flows (NASA Rotor 67 and NASA Rotor 37). The results demonstrated that the solver effectively captures the characteristics of these test cases. For most of the test cases, the convergence is achieved within 1000 steps. For the external flow test cases, our solver can accurately capture the position of the shock wave and the separation region. For the internal flow test cases, our solver can give a good prediction of the performance parameters. In the future, to enhance the solver’s capability for practical engineering problems, the implicit GMRES solver will be extended to handle unsteady flow problems using the dual timestepping method. Additionally, more advanced turbulence models, such as Detached Eddy Simulation (DES) and Large Eddy Simulation (LES), will be implemented to better address turbulent flow simulation.
Acknowledgements
We thank Mr. Junzhe Cao and Mr. Yuze Zhu for their assistance in mesh generation and discussions regarding the test cases, as well as Dr. Yajun Zhu for the Jacobian matrix derivation. This work was supported by the National Key R&D Program of China (Grant No. 2022YFA1004500), the National Natural Science Foundation of China (Nos. 12172316, 92371107, 12302378 and 92371201), the Hong Kong Research Grant Council (Nos. 16208021, 16301222, and 16208324) and the Funding of National Key Laboratory of Computational Physics, and the Natural Science Basic Research Plan in Shaanxi Province of China (No. 2025SYS-SYSZD-070).
Jacobian Matrix for Inviscid Flux and Viscous Flux
Jacobian matrix for inviscid flux
At the cell face, consider the approach of flux calculation by using the moments of relative velocity (Eq. (3) and Eq. (4)). The Jacobian matrix of the inviscid flux is given by
where and are the flux and the conservative variables of relative velocity, respectively. According to Eq. (3), the Jacobian matrix of with respect to is given by
where is the velocity of mesh motion. And the Jacobian matrix of with respect to is given by
The flux Jacobian matrix is given by
where is the transformation matrix from local coordinate to global coordinate. The flux is given by
where and are calculated by the left and right cells with the same calculation method
where , and for the half integration is used and for the half integration is employed. The flux Jacobian matrix about gives
where
The transformation between primitive and conservative flow variables gives
where
Jacobian matrix for viscous flux
The thin shear layer (TSL) [45, 46] approximation is used for viscous flux with the assumption that the viscous effect is only significant in the normal direction of the cell surface. The viscous flux is approximated as
where is the normal vector of the cell surface. To obtain the viscous stress tensor, the derivatives are approximated by the normal derivatives via
where is the variable such as , and are the cell-averaged values of at the right and left of the cell surface, respectively, and is the distance between the cell centers of left and right cells. The viscous stress tensor and heat flux are given by
and is the thermal conductivity. Thus, the Jacobian matrix of the viscous flux is given by
and , where
and
References
- [18]Y. Xiaoquan, J. Cheng, H. Luo, Q. Zhao, Robust implicit direct discontinuous Galerkin method for simulating the compressible turbulent flows, AIAA Journal 57 (3) (2019) 1113–1132.
- [19]C. Zhou, Z. J. Wang, An evaluation of implicit time integration schemes for discontinuous high order methods, in: 21st AIAA Computational Fluid Dynamics Conference, 2013, p. 2688.DOI
- [20]G. Cao, H. Su, J. Xu, K. Xu, Implicit high-order gas kinetic scheme for turbulence simulation, Aerospace Science and Technology 92 (2019) 958–971.arxiv.org/abs/1811.08005
- [21]H. Liu, X. Ji, Y. Mao, Y. Ding, K. Xu, A compact gas-kinetic scheme with scalable geometric multigrid acceleration for steady-state computation on 3d unstructured meshes, arXiv preprint arXiv:2402.02075 (2024).arxiv.org/abs/2402.02075
- [22]S. Tan, Q. Li, Time-implicit gas-kinetic scheme, Computers & Fluids 144 (2017) 44–59.DOI
- [23]Y. Yang, L. Pan, K. Xu, Implicit high-order gas-kinetic schemes for compressible flows on three-dimensional unstructured meshes I: Steady flows, Journal of Computational Physics 505 (2024) 112902.arxiv.org/abs/2304.09485
- [24]Y. Yang, L. Pan, K. Xu, Implicit high-order gas-kinetic schemes for compressible flows on three-dimensional unstructured meshes II: Unsteady flows, Journal of Computational Physics 521 (2025) 113534.arxiv.org/abs/2403.00482
- [25]Y. Jia, N. Zhengyu, Y. Xiaoquan, D. Jue, T. Xiaolong, W. Peifeng, A robust compact least-squares reconstruction method for compressible turbulent flow simulations of complex configurations, Chinese Journal of Aeronautics 36 (12) (2023) 113–138.
- [26]D. Zhou, A three-dimensional gas-kinetic BGK scheme for simulating flows in rotating machinery, Advances in Applied Mathematics and Mechanics 11 (1) (2019) 168–196.DOI
- [27]Y. Zhang, X. Ji, K. Xu, A high-order compact gas-kinetic scheme in a rotating coordinate frame and on sliding mesh, International Journal of Computational Fluid Dynamics (2023) 1–20.arxiv.org/abs/2210.17385
- [28]X. Ji, F. Zhao, W. Shyy, K. Xu, A HWENO reconstruction based high-order compact gas-kinetic scheme on unstructured mesh, Journal of Computational Physics 410 (2020) 109367.DOI
- [29]G. Gassner, F. Lörcher, C.-D. Munz, A contribution to the construction of diffusion fluxes for finite volume and discontinuous Galerkin schemes, Journal of Computational Physics 224 (2) (2007) 1049–1063.DOI
- [30]J. Blazek, Chapter 4 - structured finite-volume schemes, in: J. Blazek (Ed.), Computational Fluid Dynamics: Principles and Applications (Third Edition), third Edition, Butterworth-Heinemann, Oxford, 2015, pp. 73–120.
- [31]F. R. Menter, Two-equation eddy-viscosity turbulence models for engineering applications, AIAA journal 32 (8) (1994) 1598–1605.DOI
- [32]F. Menter, Fully new two equation k−ω turbulence models for aerodynamic flows, in: 23rd fluid dynamics, plas madynamics, and lasers conference, 1993, p. 2906.
- [33]J. Slater, Onera m6 wing: Study 1, accessed: 2020-09-01 (2008). URL https://www.grc.nasa.gov/www/wind/valid/m6wing/m6wing01/m6wing01.html
- [34]V. Schmitt, Pressure distributions on the ONERA M6-wing at transonic mach numbers, experimental data base for computer program assessment, AGARD AR-138 (1979).
- [35]D. W. Levy, K. R. Laflin, E. N. Tinoco, J. C. Vassberg, M. Mani, B. Rider, C. L. Rumsey, R. A. Wahls, J. H. Morrison, O. P. Brodersen, et al., Summary of data from the fifth computational fluid dynamics drag prediction workshop, Journal of Aircraft 51 (4) (2014) 1194–1213.DOI
- [36]D. Hue, Fifth drag prediction workshop: Computational fluid dynamics studies carried out at ONERA, Journal of Aircraft 51 (4) (2014) 1295–1310.DOI
- [37]A. J. Strazisar, J. R. Wood, M. D. Hathaway, K. L. Suder, Laser anemometer measurements in a transonic axial-flow fan rotor, Tech. rep. (1989).
- [38]L. FOTTNER, Test cases for computation of internal flows in aero engine components (1990).
- [39]L. Reid, R. D. Moore, Design and overall performance of four highly loaded, high speed inlet stages for an advanced high-pressure-ratio core compressor, Tech. rep. (1978).
- [40]D. Wisler, Rotor 37 blind test case, in: ASME/IGTI Int. Gas Turbine Conf., 1994.
- [41]T. Arima, T. Sonoda, M. Shiratori, A. Tamura, K. Kikuchi, A numerical investigation of transonic axial compressor rotor flow using a low-reynolds-number k=ε turbulence model (1999).






