betaFeedback

WorksPhysics and Astronomy

A pressure-based coupled general synthetic iterative scheme for rarefied gas flow simulation

Abstract

This work develops a pressure-based coupled (PBC) GSIS. Within each GSIS iteration, a conventional kinetic solver advances the velocity distribution function, from which non-equilibrium constitutive relations are extracted. A PBC macroscopic solver then solves the steady macroscopic synthetic equations: continuity and momentum equations are solved as a coupled pressure-velocity system under fixed temperature, followed by a segregated temperature solve. A boundary treatment is constructed to maintain strict consistency between kinetic boundary fluxes and macroscopic boundary conditions throughout the macroscopic iterations. The proposed method is validated against direct simulation Monte Carlo (DSMC) data across multiple three-dimensional test cases, including tube-to-vacuum flows, a dynamic gas lock configuration for EUV lithography, divertors in nuclear fusion, and sphere flow at Mach numbers up to 10. Numerical results demonstrate that our GSIS-PBC preserves asymptotic-preserving properties and yields satisfactory accuracy across continuum-to-transitional flow regimes. For low-speed near-continuum internal flows, GSIS-PBC delivers wall clock time speed-ups of up to two orders of magnitude compared with the conventional iterative scheme, and substantially outperforms the original density-based GSIS for challenging low-Mach internal flow problems. For nuclear-fusion divertor flows, GSIS-PBC is faster than the DSMC by four orders of magnitude.

Introduction

Rarefied internal gas flows underpin a broad spectrum of modern engineering applications, ranging from extreme ultraviolet (EUV) lithography in semiconductor manufacturing to divertor systems in nuclear fusion [1, 2]. Unlike hypersonic external flows, these complex internal flows are characterized by low Mach numbers rather than strong compressibility. Once nonequilibrium effects become significant, the Navier-Stokes-Fourier (NSF) equations are no longer adequate, which motivates the use of the mesoscopic Boltzmann kinetic equations [3, 4]. Solving such kinetic equations, however, can be computationally prohibitive. For example, the direct simulation Monte Carlo (DSMC) method needs approximately one week of computation time on 4,096 CPUs to reach steady state for each pumping-probability configuration in three-dimensional divertor simulations [5].

This high computational cost arises from three distinct sources. The first is statistical sampling: at low Mach numbers, the mean flow signal is weak relative to thermal fluctuations, so extensive sampling is required to extract it, thereby increasing the computational burden [6]. Deterministic kinetic solvers overcome this challenge [7]. The second source is the absence of the asymptotic-preserving (AP) property [8]: in the near-continuum regime, the collision time and mean free path are much smaller than the characteristic flow scales, which incurs substantial computational time and memory overhead. Over the past decades, a wide range of multiscale numerical methods have been developed to address this issue, including hybrid NSF-DSMC coupling approaches [9], the unified gas-kinetic scheme and its variants [10, 11, 12], implicit-explicit Runge-Kutta methods [13], stochastic Fokker-Planck methods [14], the unified stochastic particle method [15], and the general synthetic iterative scheme (GSIS) [16]. The third source is that conventional time-marching algorithms are ill-suited to steady-state problems. Since steady-state solutions constitute the primary objective in most applications, a numerical method should avoid redundant transient evolution and converge rapidly toward the steady state. Nevertheless, the multiscale properties alone do not guarantee fast steady-state convergence: explicit implementations remain constrained by time-step stability limits, while implicit acceleration hinges on effective treatment of the coupling between macroscopic variables and numerical fluxes [17].

In addition to being noise-free and possessing the AP property [18], the GSIS directly addresses the problem of slow convergence to steady state by accelerating kinetic iterations through macroscopic synthetic equations that retain nonequilibrium corrections. The original concept was first proposed in the diffuse synthetic acceleration method for solving the neutron transport equation, as reviewed in [19]. This method was subsequently extended to linear rarefied gas flows with flow velocities perpendicular to the computational domain [20, 21, 22, 23], including typical cases such as linear Couette and Poiseuille flows. Finally, the approach was further generalized to general nonlinear rarefied gas flow problems [16, 24, 25]. In GSIS, because the NSF constitutive relations are explicitly embedded within the synthetic equation, its diffuse-type operator enables efficient exchange of flow information across the entire computational domain, and fast steady-state convergence is achievable within dozens of kinetic iterations1. In our previous GSIS method, the macroscopic synthetic equations are solved using density-based NSF solvers. For hypersonic flow around an X38-like vehicle, GSIS outperforms the unified gas-kinetic wave-particle method by a factor of approximately 146 in cell-normalized CPU-hour costs at a Knudsen number of 0.00443; also for internal flows with simple geometry, e.g., the cavity flow, GSIS outperforms DSMC by a factor of approximately 3.6×1043.6 \times10^{4} in CPU-hour costs at a Knudsen number of 0.005 [27]. The advantages of GSIS become even more pronounced as the Knudsen number decreases further. It should be noted that the high-order/low-order method can also accelerate convergence toward the steady-state solution [28, 29]. In this approach, the stress tensor and heat flux appearing in the macroscopic synthetic equations are directly evaluated from the numerical solution of kinetic equation, supplemented by additional consistency correction terms. However, because the NSF constitutive relations are not explicitly incorporated, it suffers from stability limitations for large time-step sizes [30, 31].

Nevertheless, for the dynamic gas lock [32] and tokamak divertor [5] cases, our density-based GSIS suffers from poor convergence performance. Owing to the substantial separation of acoustic and convective time scales characteristic of low-Mach internal flows, considerable stiffness mitigation is required – a challenge exacerbated for long-duct geometries. This identical issue manifests in continuum flows described by the NSF equations. Within such continuum flow settings, pressure-based coupled (PBC) methods tackle this limitation by enforcing mass and momentum conservation in a single implicit system, wherein pressure and velocity are solved jointly and the dependence of face mass fluxes on these variables is retained [33, 34]. Optimized for the weakly compressible limit, PBC formulations alleviate the convergence difficulty caused by timescale disparity in low-speed regions [35, 36].

Based on this observation and the crucial role of macroscopic synthetic equations in driving the rapid steady-state evolution of kinetic equations, this work develops a GSIS-PBC method for efficient steady-state simulations of complex rarefied internal flows. Three novel treatments are introduced. First, the macroscopic solver first couples pressure and velocity through the mass and momentum equations, and subsequently updates the temperature field independently, while retaining the nonequilibrium stress and heat flux corrections supplied by the kinetic solution. Second, to accommodate the PBC, the GSIS framework is redesigned for improved stability. Third, although the synthetic equation accelerates the exchange of flow information within the bulk region, a new boundary treatment is developed to fully exploit the capabilities of GSIS. This framework incorporates the response of boundary fluxes to variations in macroscopic states into the coupled-macroscopic-mesoscopic solution procedure, rigorously enforces consistency between mesoscopic and macroscopic quantities, and permits synchronous adjustment of boundary states and interior flow fields. The proposed method aims to achieve several orders of magnitude speedup in wall-clock time compared with the conventional density-based GSIS for internal rarefied flow problems.

The remainder of this paper is organized as follows. The governing equations are presented in Section 2, followed by a detailed exposition of the GSIS-PBC numerical framework and associated boundary treatment in Section 3. Section 4 validates the performance of

their power. Only recently have Liu and Xu incorporated this concept into their wave-particle decomposition method, achieving speedups of several orders of magnitude [26]. GSIS-PBC via a suite of three-dimensional test cases encompassing dynamic gas lock and tokamak divertor flows. Finally, conclusions are summarized in § 5.

Governing equations

The gas kinetic model

Let f(t,x,ξ)f(t,\boldsymbol{x},\boldsymbol{\xi}) be the velocity distribution function and f dx dξf\,\mathrm{d}\boldsymbol{x}\,\mathrm{d}\boldsymbol{\xi} be the gas mass contained in the corresponding phase-space element, where tt denotes time, x∈R3\boldsymbol{x} \in\mathbb{R}^{3} is the spatial coordinate, and ξ∈R3\boldsymbol{\xi} \in\mathbb{R}^{3} is the molecular velocity. For a monatomic gas, the evolution of ff can be described by the kinetic equation following the relaxation-time approximation:

∂f∂t+ξ⋅∇xf=fS−fτ,(1)\frac{\partial f}{\partial t}+\boldsymbol{\xi}\cdot\nabla_{\boldsymbol{x}}f=\frac{f^{S}-f}{\tau}, \tag*{(1)}

where τ\tau is the collision relaxation time. The target distribution fSf^{S} in Eq. (1) is specified by the Shakhov model [37] as

fS=feq[1+(1−Pr⁡)c⋅q5pRT(∣c∣2RT−5)].(2)f^{S}=f^{\mathrm{eq}}\left[1+(1-\operatorname{Pr})\frac{\boldsymbol{c}\cdot\boldsymbol{q}}{5pRT}\left(\frac{\lvert\boldsymbol{c}\rvert^{2}}{RT}-5\right)\right]. \tag*{(2)}

Here RR is the specific gas constant and Pr⁡\operatorname{Pr} is the Prandtl number. For the monatomic ideal gas considered in this work, Pr⁡=2/3\operatorname{Pr}=2/3. The peculiar molecular velocity is c=ξ−u\boldsymbol{c}=\boldsymbol{\xi}-\boldsymbol{u}, where u\boldsymbol{u} is the bulk velocity. The local Maxwell equilibrium distribution is

feq=ρ(2πRT)3/2exp⁡(−∣c∣22RT).(3)f^{\mathrm{eq}}=\frac{\rho}{(2\pi RT)^{3/2}}\exp\left(-\frac{\lvert\boldsymbol{c}\rvert^{2}}{2RT}\right). \tag*{(3)}

The density ρ\rho, temperature TT, thermodynamic pressure pp, and heat-flux q\boldsymbol{q} appearing in Eqs. (2) and (3), together with u\boldsymbol{u}, are obtained from the moments of ff defined in § 2.2.

Macroscopic equations

Introduce the collision-invariant vector

ψ(ξ)=(1,ξ,∣ξ∣22)T.(4)\boldsymbol{\psi}(\boldsymbol{\xi})=\left(1,\boldsymbol{\xi},\frac{\lvert\boldsymbol{\xi}\rvert^{2}}{2}\right)^{\mathsf{T}}. \tag*{(4)}

The conservative variables are the corresponding velocity moments,

U=(ρρuρE)=∫R3ψ(ξ)f(t,x,ξ) dξ,(5)\boldsymbol{U}=\begin{pmatrix}\rho\\ \rho\boldsymbol{u}\\ \rho E\end{pmatrix}=\int_{\mathbb{R}^{3}}\boldsymbol{\psi}(\boldsymbol{\xi})f(t,\boldsymbol{x},\boldsymbol{\xi})\,\mathrm{d}\boldsymbol{\xi}, \tag*{(5)}

where EE is the total energy per unit mass. For monatomic gas, the energy moment gives E=3RT/2+∣u∣2/2E=3RT/2+\lvert\boldsymbol{u}\rvert^{2}/2, while the scalar pressure is p=ρRTp=\rho RT.

The pressure tensor Π\boldsymbol{\Pi} and heat-flux vector q\mathbf{q} follow from the higher-order moments of the velocity distribution function:

Π=∫R3ccTf dξ=pI+σ,\boldsymbol{\Pi}=\int_{\mathbb{R}^{3}}\mathbf{c}\mathbf{c}^{\mathrm{T}}f\,\mathrm{d}\boldsymbol{\xi}=p\mathbf{I}+\boldsymbol{\sigma},
q=∫R3∣c∣22cf dξ,(6)\mathbf{q}=\int_{\mathbb{R}^{3}}\frac{\lvert\mathbf{c}\rvert^{2}}{2}\mathbf{c}f\,\mathrm{d}\boldsymbol{\xi}, \tag*{(6)}

where I\mathbf{I} is the identity tensor and σ\boldsymbol{\sigma} is the deviatoric part of the pressure tensor.

Taking the moments of Eq. (1) with respect to ψ\boldsymbol{\psi} therefore yields the exact macroscopic balance equations

∂ρ∂t+∇⋅(ρu)=0,\frac{\partial\rho}{\partial t}+\nabla\cdot(\rho\mathbf{u})=0,
∂(ρu)∂t+∇⋅(ρuuT+pI+σ)=0,\frac{\partial(\rho\mathbf{u})}{\partial t}+\nabla\cdot\left(\rho\mathbf{u}\mathbf{u}^{\mathrm{T}}+p\mathbf{I}+\boldsymbol{\sigma}\right)=0,
∂(ρE)∂t+∇⋅[(ρE+p)u+σ⋅u+q]=0,(7)\frac{\partial(\rho E)}{\partial t}+\nabla\cdot\left[(\rho E+p)\mathbf{u}+\boldsymbol{\sigma}\cdot\mathbf{u}+\mathbf{q}\right]=0, \tag*{(7)}

which contains no constitutive approximation: the stress and heat flux are only the velocity moments in Eq. (6), rather then expressed in terms of the conservative variables and their spatial derivatives.

The degree of gas rarefaction is characterized by the Knudsen number

Kn=λLref=μpLrefπRT2,(8)\mathrm{Kn}=\frac{\lambda}{L_{\mathrm{ref}}}=\frac{\mu}{pL_{\mathrm{ref}}}\sqrt{\frac{\pi RT}{2}}, \tag*{(8)}

where LrefL_{\mathrm{ref}} is the characteristic length, λ\lambda is the molecular mean free path, and μ\mu is the dynamic viscosity. In the near-continuum regime, where Kn≪1\mathrm{Kn}\ll1, a first-order Chapman–Enskog expansion of Eq. (1) recovers the NSF constitutive relations [38],

σNSF=−μ[∇u+(∇u)T−23(∇⋅u)I],\boldsymbol{\sigma}^{\mathrm{NSF}}=-\mu\left[\nabla\mathbf{u}+(\nabla\mathbf{u})^{\mathrm{T}}-\frac{2}{3}(\nabla\cdot\mathbf{u})\mathbf{I}\right],
qNSF=−κ∇T.(9)\mathbf{q}^{\mathrm{NSF}}=-\kappa\nabla T. \tag*{(9)}

The corresponding transport coefficients satisfy

μ=pτ,κ=52μRPr.(10)\mu=p\tau,\qquad\kappa=\frac{5}{2}\frac{\mu R}{\mathrm{Pr}}. \tag*{(10)}

The first relation connects the relaxation time to the viscosity, whereas the Shakhov correction enforces the prescribed Prandtl number and thereby fixes κ\kappa relative to μ\mu. The temperature dependence of the viscosity is modeled by the power law

μ(T)=μref(TTref)ω,(11)\mu(T)=\mu_{\mathrm{ref}}\left(\frac{T}{T_{\mathrm{ref}}}\right)^{\omega}, \tag*{(11)}

where TrefT_{\mathrm{ref}} is the reference temperature, μref=μ(Tref)\mu_{\mathrm{ref}}=\mu(T_{\mathrm{ref}}) is the reference viscosity, and ω\omega is the viscosity index.

Pressure-based coupled GSIS

The GSIS accelerates steady kinetic calculations by coupling the kinetic iteration with macroscopic synthetic equations [16, 27]. In the present method, a PBC solver advances the macroscopic synthetic equations by treating the continuity and momentum equations as a coupled pressure–velocity system and then solving the temperature equation. To accommodate the PBC solver, the GSIS framework is modified for numerical stability, followed by the detailed PBC formulation and the kinetic-macroscopic boundary treatment.

The GSIS framework

As illustrated in Figure 1, GSIS-PBC alternates between kinetic and macroscopic updates. Multiple kinetic updates are performed between successive macroscopic corrections to reduce numerical oscillations in the velocity distribution function and improve the stability of the coupled iteration. At the nnth GSIS iteration, the conventional iterative scheme (CIS) performs NCISN_{\mathrm{CIS}} kinetic iterations to advance the velocity distribution function from fnf^n to fn+1/2f^{n+1/2}, from which the macroscopic state, high-order terms (HoTs), and kinetic boundary fluxes are evaluated. These quantities supply the initial state, closure, and boundary information for the PBC macroscopic solver. After the PBC iterations, the updated macroscopic state Wn+1W^{n+1} is used to correct the equilibrium part of fn+1/2f^{n+1/2}, yielding fn+1f^{n+1}. The process is repeated until convergence.

GSIS-PBC solution procedure

Figure 1. GSIS-PBC solution procedure. Here nn and mm index GSIS and PBC iterations, respectively; NCISN_{\mathrm{CIS}} is the number of CIS steps per GSIS iteration, and MM is the maximum number of PBC iterations. The primitive state is W=(p,u,T)TW=(p,u,T)^{\mathrm T}; ΔWpu\Delta W_{\mathrm{pu}} and ΔT\Delta T are the pressure–velocity and temperature corrections. The symbols R\mathcal{R}, JJ, and MM denote residual vectors, approximate Jacobians, and pseudo-time matrices, respectively. Here feqf^{\mathrm{eq}} denotes the Maxwellian equilibrium distribution.

For cell ii, the macroscopic quantities supplied to the PBC solver are obtained by taking velocity moments of fin+1/2f_i^{n+1/2} according to Eqs. (5) and (6), giving the conservative state Uin+1/2U_i^{n+1/2}, stress σin+1/2\boldsymbol{\sigma}_i^{n+1/2}, and heat flux qin+1/2\boldsymbol{q}_i^{n+1/2}. The associated primitive state is denoted by Win+1/2=(pin+1/2,uin+1/2,Tin+1/2)T\mathbf{W}_i^{n+1/2}=(p_i^{n+1/2},\boldsymbol{u}_i^{n+1/2},T_i^{n+1/2})^{\mathrm T}.

The NSF stress and heat flux are evaluated from the resulting spatial field Wn+1/2\mathbf{W}^{n+1/2} using Eq. (9). The HoTs, which characterize the non-equilibrium constitutive relations beyond the NSF level, are obtained by subtracting these NSF contributions from the corresponding kinetic moments:

HoTσ,in+1/2=σin+1/2−[σNSF(Wn+1/2)]i,HoTq,in+1/2=qin+1/2−[qNSF(Wn+1/2)]i,(12)\begin{aligned} \mathrm{HoT}_{\sigma,i}^{n+1/2} &= \boldsymbol{\sigma}_i^{n+1/2} - \left[\boldsymbol{\sigma}^{\mathrm{NSF}}\left(\mathbf{W}^{n+1/2}\right)\right]_i,\\ \mathrm{HoT}_{q,i}^{n+1/2} &= \mathbf{q}_i^{n+1/2} - \left[\mathbf{q}^{\mathrm{NSF}}\left(\mathbf{W}^{n+1/2}\right)\right]_i, \tag*{(12)} \end{aligned}

where [⋅]i\left[\cdot\right]_i denotes the cell-wise NSF value evaluated using gradients reconstructed from neighboring cells. Combining the NSF contributions with the HoTs and omitting the time derivative yields the steady-state macroscopic synthetic equations:

∇⋅[(ρuρuuT+pI+σNSF(W)+HoTσn+1/2(ρE+p)u+[σNSF(W)+HoTσn+1/2]⋅u+qNSF(W)+HoTqn+1/2)]=0.(13)\nabla\cdot\left[ \begin{pmatrix} \rho\mathbf{u}\\ \rho\mathbf{u}\mathbf{u}^{\mathrm{T}} + p\mathbf{I} + \boldsymbol{\sigma}^{\mathrm{NSF}}(\mathbf{W}) + \mathrm{HoT}_{\sigma}^{n+1/2}\\ (\rho E+p)\mathbf{u} + \left[\boldsymbol{\sigma}^{\mathrm{NSF}}(\mathbf{W})+\mathrm{HoT}_{\sigma}^{n+1/2}\right]\cdot\mathbf{u}+\mathbf{q}^{\mathrm{NSF}}(\mathbf{W})+\mathrm{HoT}_{q}^{n+1/2} \end{pmatrix} \right] = \mathbf{0}. \tag*{(13)}

The PBC solver advances these equations with the HoTs held fixed, while updating the NSF stress and heat flux with the current macroscopic state. The resulting state Wn+1\mathbf{W}^{n+1} is then used to update the kinetic distribution by replacing its equilibrium contribution. At velocity point kk, this gives

fi,kn+1=fi,kn+1/2−feq(Win+1/2;ξk)+feq(Win+1;ξk),(14)f_{i,k}^{n+1}=f_{i,k}^{n+1/2}-f^{\mathrm{eq}}\left(\mathbf{W}_i^{n+1/2};\boldsymbol{\xi}_k\right)+f^{\mathrm{eq}}\left(\mathbf{W}_i^{n+1};\boldsymbol{\xi}_k\right), \tag*{(14)}

where feq(W;ξk)f^{\mathrm{eq}}(\mathbf{W};\boldsymbol{\xi}_k) denotes the Maxwell equilibrium distribution at velocity point kk. Since this correction does not guarantee non-negativity, the corrected distribution is checked cell-wise. In practice, if min⁡kfi,kn+1<−10−8ρin+1/(RTin+1)3/2\min_k f_{i,k}^{n+1}<-10^{-8}\rho_i^{n+1}/(RT_i^{n+1})^{3/2}, all velocity-space values in cell ii are replaced by feq(Win+1;ξk)f^{\mathrm{eq}}(\mathbf{W}_i^{n+1};\boldsymbol{\xi}_k); otherwise, the correction in Eq. (14) is retained. Across the numerical cases considered in this work, negative values arose primarily following the first GSIS correction, were seldom observed after the second correction, and were absent in the converged distribution.

Once GSIS has converged, the macroscopic state remains unchanged between successive GSIS iterations, and hence σNSF\boldsymbol{\sigma}^{\mathrm{NSF}} and qNSF\mathbf{q}^{\mathrm{NSF}} no longer change. Consequently, the converged solution of the synthetic equation (13) is exactly the moment equation (7) loyal to the kinetic equation (1) in the steady state.

The CIS iteration

At the nnth GSIS iteration, the CIS advances fnf^n to the intermediate distribution fn+1/2f^{n+1/2}. The distribution after ℓ\ell kinetic updates is denoted by fℓf^\ell, with f0=fnf^0=f^n and fnCIS=fn+1/2f^{n_{\mathrm{CIS}}}=f^{n+1/2}. Let Ωi\Omega_i be a control volume of volume ViV_i and centroid xi\mathbf{x}_i, and let NiN_i denote the set of face-neighbouring cells of cell ii, including the ghost cells associated with physical boundary faces. The common face of cells ii and jj is denoted by ijij, with centroid xij\mathbf{x}_{ij} and area vector Sij=AijnijS_{ij} = A_{ij}\mathbf{n}_{ij} directed from ii to jj. At velocity point kk, the finite-volume form of Eq. (1) is

ViΔt(fi,kℓ+1−fi,kℓ)+∑j∈Ni(ξk⋅Sij)fij,kℓ+1=Viτiℓ(fi,kS,ℓ−fi,kℓ+1),(15)\frac{V_i}{\Delta t}\left(f_{i,k}^{\ell+1}-f_{i,k}^{\ell}\right)+\sum_{j\in N_i}\left(\boldsymbol{\xi}_k\cdot\mathbf{S}_{ij}\right)f_{ij,k}^{\ell+1}=\frac{V_i}{\tau_i^{\ell}}\left(f_{i,k}^{S,\ell}-f_{i,k}^{\ell+1}\right), \tag*{(15)}
fij,k={fi,k+φi,k∇fi,k⋅(xij−xi),ξk⋅Sij≥0,fj,k+φj,k∇fj,k⋅(xij−xj),ξk⋅Sij<0,f_{ij,k}= \begin{cases} f_{i,k}+\varphi_{i,k}\nabla f_{i,k}\cdot(\mathbf{x}_{ij}-\mathbf{x}_i), & \boldsymbol{\xi}_k\cdot\mathbf{S}_{ij}\ge0,\\ f_{j,k}+\varphi_{j,k}\nabla f_{j,k}\cdot(\mathbf{x}_{ij}-\mathbf{x}_j), & \boldsymbol{\xi}_k\cdot\mathbf{S}_{ij}<0, \end{cases}

Here Δt\Delta t is a pseudo-time step introduced to enhance numerical stability, and φi,k∈[0,1]\varphi_{i,k}\in[0,1] is the Venkatakrishnan limiter [39] applied to the cell-centred gradient. The upwind reconstructed value is selected according to the sign of ξk⋅Sij\boldsymbol{\xi}_k\cdot\mathbf{S}_{ij}.

For a physical boundary face bb of area AbA_b adjacent to cell ii, let nb\mathbf{n}_b be the unit normal pointing out of the gas domain and set

Ξb±={ξ:±ξ⋅nb>0}.(16)\Xi_b^{\pm}=\{\boldsymbol{\xi}:\pm\boldsymbol{\xi}\cdot\mathbf{n}_b>0\}. \tag*{(16)}

The boundary distribution used by the CIS is

fb,k={fi,krec,ξk∈Ξb+,fb,kbc,ξk∈Ξb−,(17)f_{b,k}= \begin{cases} f_{i,k}^{\mathrm{rec}}, & \boldsymbol{\xi}_k\in\Xi_b^{+},\\ f_{b,k}^{\mathrm{bc}}, & \boldsymbol{\xi}_k\in\Xi_b^{-}, \end{cases} \tag*{(17)}

where fi,krecf_{i,k}^{\mathrm{rec}} is reconstructed from the adjacent cell with the same limited reconstruction as in Eq. (15), while fb,kbcf_{b,k}^{\mathrm{bc}} is specified by the prescribed mass-flow, stationary fully diffuse wall, pressure inlet or outlet, or far-field boundary condition. Their discrete constructions are given in Appendix A.

PBC solution of the macroscopic synthetic equations

The PBC formulation is used to efficiently solve the steady macroscopic synthetic equations (13) when large disparities between acoustic and convective time scales dominate. Let m=0,…,M−1m=0,\ldots,M-1 denote the PBC iteration index and Wm=(p,u,T)T\mathbf{W}^{m}=(p,\mathbf{u},T)^{\mathrm{T}} the current primitive state, where MM is the prescribed maximum number of PBC iterations. The continuity and momentum equations are first solved as a coupled pressure–velocity system [33, 34], with temperature held fixed. After pressure and velocity are updated, the face quantities are reevaluated and the scalar temperature equation is solved to obtain Wm+1\mathbf{W}^{m+1}. Figure 1 summarizes this sequence.

The following subsections construct the internal and boundary face fluxes, derive the coupled pressure–velocity and temperature systems, and present their pseudo-time-stabilized solution.

Finite-volume discretization of internal-face fluxes

For an internal face ijij, let nij\mathbf{n}_{ij} be the unit normal directed from cell ii to cell jj, and let un=u⋅niju_n=\mathbf{u}\cdot\mathbf{n}_{ij}. The face flux corresponding to Eq. (13) is

Fij=Fijadv+Fijvis,(18)\mathbf{F}_{ij}=\mathbf{F}_{ij}^{\mathrm{adv}}+\mathbf{F}_{ij}^{\mathrm{vis}}, \tag*{(18)}

where

Fijadv=Aij[ρunρunu+pnij(ρE+p)un],(19)\boldsymbol{F}_{ij}^{\mathrm{adv}}=A_{ij} \begin{bmatrix} \rho u_n\\ \rho u_n\boldsymbol{u}+p\boldsymbol{n}_{ij}\\ (\rho E+p)u_n \end{bmatrix}, \tag*{(19)}

and

Fijvis=Aij[0[σNSF+HoTσn+1/2]⋅nij{[σNSF+HoTσn+1/2]⋅nij}⋅u+[qNSF+HoTqn+1/2]⋅nij].(20)\boldsymbol{F}_{ij}^{\mathrm{vis}}=A_{ij} \begin{bmatrix} 0\\ \left[\boldsymbol{\sigma}^{\mathrm{NSF}}+\mathrm{HoT}_{\sigma}^{n+1/2}\right]\cdot\boldsymbol{n}_{ij}\\ \left\{\left[\boldsymbol{\sigma}^{\mathrm{NSF}}+\mathrm{HoT}_{\sigma}^{n+1/2}\right]\cdot\boldsymbol{n}_{ij}\right\}\cdot\boldsymbol{u} +\left[\boldsymbol{q}^{\mathrm{NSF}}+\mathrm{HoT}_{q}^{n+1/2}\right]\cdot\boldsymbol{n}_{ij} \end{bmatrix}. \tag*{(20)}

In Eqs. (19) and (20), the face index is retained only on the fluxes and geometric quantities; the unindexed flow variables denote their face values. The numerical construction below prevents pressure–velocity decoupling associated with the collocated variable arrangement, applies upwinding to transported quantities, and accounts for mesh non-orthogonality in the NSF stress and heat flux.

For the advective flux Fijadv\boldsymbol{F}_{ij}^{\mathrm{adv}}, evaluating AijunA_{ij}u_n requires the normal velocity at the face. With both pressure and velocity stored at cell centres, direct velocity interpolation can leave a checkerboard pressure mode undetected: the discrete pressure gradients cancel in the cell-centred momentum equations. The velocity field then does not respond to the pressure jump pj−pip_j-p_i between adjacent cells, and the interpolated face volume flow, and hence the mass flux in the continuity equation, contains no direct coupling between their pressures. Following Rhie and Chow [40], this missing coupling is restored by adding a compact face-flux correction involving pj−pip_j-p_i. Accordingly, AijunA_{ij}u_n in (19) is replaced by

V˙ijRC=uij⋅Sij−βijRCΔijp,(21)\dot{V}_{ij}^{RC}=\boldsymbol{u}_{ij}\cdot\boldsymbol{S}_{ij}-\beta_{ij}^{RC}\Delta_{ij}p, \tag*{(21)}

where uij=wiui+wjuj\boldsymbol{u}_{ij}=w_i\boldsymbol{u}_i+w_j\boldsymbol{u}_j. The cell-centre displacement and its normal projection are

dij=xj−xi,dn,ij=nij⋅dij>0,(22)\boldsymbol{d}_{ij}=\boldsymbol{x}_j-\boldsymbol{x}_i,\qquad d_{n,ij}=\boldsymbol{n}_{ij}\cdot\boldsymbol{d}_{ij}>0, \tag*{(22)}

and the interpolation weights are

wj=min⁡{1,max⁡[0,nij⋅(xij−xi)dn,ij]},wi=1−wj.(23)w_j=\min\left\{1,\max\left[0,\frac{\boldsymbol{n}_{ij}\cdot(\boldsymbol{x}_{ij}-\boldsymbol{x}_i)}{d_{n,ij}}\right]\right\},\qquad w_i=1-w_j. \tag*{(23)}

In (21), the first term is the volume flow obtained from the interpolated velocity, and the second is the Rhie–Chow pressure correction. The part of the two-cell pressure difference that is not represented by the reconstructed pressure gradient is

Δijp=(pj−pi)−∇‾pij⋅dij,(24)\Delta_{ij}p=(p_j-p_i)-\overline{\nabla}p_{ij}\cdot\boldsymbol{d}_{ij}, \tag*{(24)}

where ∇‾pij=wi(∇p)i+wj(∇p)j\overline{\nabla}p_{ij}=w_i(\nabla p)_i+w_j(\nabla p)_j. For a linear pressure field, the two terms in (24) are equal and the correction vanishes. For an odd–even pressure mode, their difference provides the missing direct coupling between pip_i and pjp_j. The correction coefficient in (21) is

βijRC=Aijdn,ij(wiDiRC+wjDjRC).(25)\beta_{ij}^{RC}=\frac{A_{ij}}{d_{n,ij}}\left(w_iD_i^{RC}+w_jD_j^{RC}\right). \tag*{(25)}

where DiRCD_i^{RC} is the momentum mobility of cell ii. To motivate its construction, consider a cell-local approximation of the steady momentum balance [40], Au,iδui≃−Viδ(∇p)i\mathbf{A}_{u,i}\delta\mathbf{u}_i \simeq-V_i\delta(\nabla p)_i, where Au,i\mathbf{A}_{u,i} represents the local momentum response to a velocity perturbation. Approximating this local response by an isotropic scalar scale au,ia_{u,i} gives δui≃−(Vi/au,i)δ(∇p)i\delta\mathbf{u}_i \simeq-(V_i/a_{u,i})\delta(\nabla p)_i, which motivates the mobility DiRC=Vi/au,iD_i^{RC}=V_i/a_{u,i}. Let Niint\mathcal{N}_i^{\mathrm{int}} denote the set of cells sharing an internal face with cell ii, and let Fib\mathcal{F}_i^{\mathrm{b}} denote its physical boundary faces. The scalar response is constructed as

au,i=∑j∈Niint[max⁡(m˙^ij,0)+4μijAij3dn,ij]+∑b∈Fib∥Bu,ib∥2,(26)a_{u,i}=\sum_{j\in\mathcal{N}_i^{\mathrm{int}}}\left[\max\left(\widehat{\dot{m}}_{ij},0\right)+\frac{4\mu_{ij}A_{ij}}{3d_{n,ij}}\right]+\sum_{b\in\mathcal{F}_i^{\mathrm{b}}}\left\|\mathbf{B}_{u,ib}\right\|_2, \tag*{(26)}

where μij=(μi+μj)/2\mu_{ij}=(\mu_i+\mu_j)/2, and m˙^ij\widehat{\dot{m}}_{ij} is the most recently evaluated shared face mass flow rate [41], including the RC pressure correction and the selected spatial reconstruction. When no preceding value is available, it is initialized as m˙^ij 0=(ρi+ρj)[(ui+uj)⋅nij]Aij/4\widehat{\dot{m}}_{ij}^{\,0}=(\rho_i+\rho_j)[(\mathbf{u}_i+\mathbf{u}_j)\cdot\mathbf{n}_{ij}]A_{ij}/4. The first internal-face term estimates the convective self-response with the lagged flow held fixed and its sign determining the donor. The second is the largest eigenvalue of the two-point viscous velocity block in Eq. (45), representing its maximum directional response. For a physical boundary face, Bu,ib=[Jb,ipu]mom,u\mathbf{B}_{u,ib}=[\mathbf{J}_{b,i}^{pu}]_{\mathrm{mom},u} is the 3×33\times3 velocity subblock of the boundary Jacobian. It is evaluated at the current cell-centred state with pressure, temperature, and the boundary consistency corrections held fixed, while retaining the derivatives of the boundary-condition-dependent incoming-flux mapping. Its matrix 22-norm is the largest singular value and gives the maximum magnitude of the linearized momentum-flux response over all unit velocity perturbations. The corresponding face-integrated boundary Jacobians are given in Appendix § 11.

The advected quantities are evaluated by donor-cell upwinding. The donor cell and mass flow rate are

iup={i,V˙ijRC≥0,j,V˙ijRC<0,m˙ij=ρijV˙ijRC.(27)i_{\mathrm{up}}= \begin{cases} i, & \dot{V}_{ij}^{RC}\geq0,\\ j, & \dot{V}_{ij}^{RC}<0, \end{cases} \qquad \dot{m}_{ij}=\rho_{ij}\dot{V}_{ij}^{RC}. \tag*{(27)}

Here, ρij=piup/(RTiup)\rho_{ij}=p_{i_{\mathrm{up}}}/(RT_{i_{\mathrm{up}}}), and the subscript up denotes a quantity evaluated in the donor cell iupi_{\mathrm{up}}. The pressure in the momentum flux is interpolated as pij=wipi+wjpjp_{ij}=w_ip_i+w_jp_j, and the total enthalpy per unit mass is H=52RT+∣u∣22H=\frac{5}{2}RT+\frac{|\mathbf{u}|^2}{2}. With the Rhie–Chow volume flow and donor-cell upwinding, the advective flux in Eq. (19) is evaluated numerically as

Fijadv=[m˙ijm˙ijuup+pijSijm˙ijHup],(28)\mathbf{F}_{ij}^{\mathrm{adv}}= \begin{bmatrix} \dot{m}_{ij}\\ \dot{m}_{ij}\mathbf{u}_{\mathrm{up}}+p_{ij}\mathbf{S}_{ij}\\ \dot{m}_{ij}H_{\mathrm{up}} \end{bmatrix}, \tag*{(28)}

For the viscous flux Fijvis\mathbf{F}_{ij}^{\mathrm{vis}}, the cell-centred velocity and temperature gradients are reconstructed by the least-squares method. On a non-orthogonal mesh, dij\mathbf{d}_{ij} is not generally aligned with nij\mathbf{n}_{ij}, so direct interpolation of the cell-centred gradients does not necessarily reproduce the difference between the adjacent cell values. For a velocity component or temperature, denoted by ϕ\phi, the resulting face gradient is evaluated using the standard non-orthogonal treatment [42, 43],

∇ϕij=∇ϕ‾ij+(ϕj−ϕi)−∇ϕ‾ij⋅dijdn,ijnij,∇ϕ‾ij=wi∇ϕi+wj∇ϕj.(29)\nabla\phi_{ij}=\overline{\nabla\phi}_{ij}+\frac{(\phi_j-\phi_i)-\overline{\nabla\phi}_{ij}\cdot\mathbf{d}_{ij}}{d_{n,ij}}\mathbf{n}_{ij}, \qquad \overline{\nabla\phi}_{ij}=w_i\nabla\phi_i+w_j\nabla\phi_j. \tag*{(29)}

The added normal component enforces ∇ϕij⋅dij=ϕj−ϕi\nabla\phi_{ij}\cdot\mathbf{d}_{ij}=\phi_j-\phi_i without altering the reconstructed tangential gradient. The component gradients form the face velocity-gradient tensor used in the NSF stress, while the temperature gradient gives the NSF heat flux through Eq. (9). The face thermal conductivity is κij=(κi+κj)/2\kappa_{ij}=(\kappa_i+\kappa_j)/2.

The HoTs are interpolated with the same geometric weights,

HoTσ,ijn+1/2=wiHoTσ,in+1/2+wjHoTσ,jn+1/2,HoTq,ijn+1/2=wiHoTq,in+1/2+wjHoTq,jn+1/2.(30)\mathrm{HoT}^{n+1/2}_{\sigma,ij}=w_i\mathrm{HoT}^{n+1/2}_{\sigma,i}+w_j\mathrm{HoT}^{n+1/2}_{\sigma,j}, \qquad \mathrm{HoT}^{n+1/2}_{q,ij}=w_i\mathrm{HoT}^{n+1/2}_{q,i}+w_j\mathrm{HoT}^{n+1/2}_{q,j}. \tag*{(30)}

The face-integrated traction and heat flow rate are defined by

tij=[σijNSF+HoTσ,ijn+1/2]⋅Sij,Qij=[qijNSF+HoTq,ijn+1/2]⋅Sij.(31)\mathbf{t}_{ij}=\left[\boldsymbol{\sigma}^{\mathrm{NSF}}_{ij}+\mathrm{HoT}^{n+1/2}_{\sigma,ij}\right]\cdot\mathbf{S}_{ij}, \qquad Q_{ij}=\left[\mathbf{q}^{\mathrm{NSF}}_{ij}+\mathrm{HoT}^{n+1/2}_{q,ij}\right]\cdot\mathbf{S}_{ij}. \tag*{(31)}

Using the centred face velocity u‾ij=(ui+uj)/2\overline{\mathbf{u}}_{ij}=(\mathbf{u}_i+\mathbf{u}_j)/2, the numerical viscous flux is

Fijvis=[0tijtij⋅u‾ij+Qij].(32)\mathbf{F}^{\mathrm{vis}}_{ij}= \begin{bmatrix} 0\\ \mathbf{t}_{ij}\\ \mathbf{t}_{ij}\cdot\overline{\mathbf{u}}_{ij}+Q_{ij} \end{bmatrix}. \tag*{(32)}

For clarity, ρij\rho_{ij} in Eq. (27) and uup\mathbf{u}_{\mathrm{up}} and HupH_{\mathrm{up}} in Eq. (28) are evaluated from the donor-cell state. For a higher-order evaluation of the advective flux, the primitive variables (p,u,T)(p,\mathbf{u},T) may instead be reconstructed from cells ii and jj to face ijij using a slope-limited reconstruction. The sign of V˙ijRC\dot{V}^{RC}_{ij} selects the reconstructed upwind state, from which ρij\rho_{ij}, uup\mathbf{u}_{\mathrm{up}}, and HupH_{\mathrm{up}} are evaluated. The Rhie–Chow volume flow, the centred pressure pijp_{ij}, and the viscous flux in Eq. (32) retain the forms given above. Together, Eqs. (28) and (32) define the numerical flux on an internal face.

Consistent mesoscopic–macroscopic boundary flux

While the macroscopic synthetic equations boost flow information exchange within the bulk region, a consistent mesoscopic–macroscopic boundary would fully leverage GSIS to improve convergence toward steady state. Following the work of Zhang et al. [25], the macroscopic boundary flux is updated during the PBC iterations while remaining consistent with the kinetic boundary condition imposed in the CIS. In the half-range integrals below, the superscripts ++ and −- indicate integration over the outgoing and incoming velocity half-spaces Ξb+\Xi_b^+ and Ξb−\Xi_b^-, respectively, as defined in Eq. (16). The same signs label the corresponding face-integrated boundary-flux components. The continuous equilibrium half-range flux per unit area is

Feq±(W;nb)=∫Ξb±(ξ⋅nb)ψ(ξ)feq(W;ξ) dξ.(33)F_{\mathrm{eq}}^{\pm}(\mathbf{W};\mathbf{n}_b)=\int_{\Xi_b^{\pm}}(\boldsymbol{\xi}\cdot\mathbf{n}_b)\boldsymbol{\psi}(\boldsymbol{\xi})f^{\mathrm{eq}}(\mathbf{W};\boldsymbol{\xi})\,\mathrm{d}\boldsymbol{\xi}. \tag*{(33)}

The corresponding face-integrated continuous outgoing flux is

F~b+(W)=AbFeq+(W;nb).(34)\widetilde{F}_b^+(\mathbf{W})=A_bF_{\mathrm{eq}}^+(\mathbf{W};\mathbf{n}_b). \tag*{(34)}

At the start of the PBC stage, set Wi0=Win+1/2\mathbf{W}_i^0=\mathbf{W}_i^{n+1/2}. The face-integrated outgoing and incoming kinetic boundary fluxes are stored as

Fb+,m=0=Ab∑ξk∈Ξb+ϖk(ξk⋅nb)ψ(ξk)fb,kn+1/2,Fb−,m=0=Ab∑ξk∈Ξb−ϖk(ξk⋅nb)ψ(ξk)fb,kn+1/2,(35)\begin{aligned} F_b^{+,m=0}&=A_b\sum_{\boldsymbol{\xi}_k\in\Xi_b^+}\varpi_k(\boldsymbol{\xi}_k\cdot\mathbf{n}_b)\boldsymbol{\psi}(\boldsymbol{\xi}_k)f_{b,k}^{n+1/2},\\ F_b^{-,m=0}&=A_b\sum_{\boldsymbol{\xi}_k\in\Xi_b^-}\varpi_k(\boldsymbol{\xi}_k\cdot\mathbf{n}_b)\boldsymbol{\psi}(\boldsymbol{\xi}_k)f_{b,k}^{n+1/2}, \tag*{(35)} \end{aligned}

where ϖk\varpi_k is the velocity-space quadrature weight. A tilde denotes a face-integrated continuous half-space flux, whereas Fb±,mF_b^{\pm,m} denotes the corresponding PBC boundary flux. The difference between each stored discrete flux and its continuous counterpart at m=0m=0 is retained as the boundary consistency correction during the PBC iterations.

During PBC iteration mm, the outgoing flux is written as the current continuous equilibrium half-range flux plus the fixed boundary consistency correction:

Fb+,m=F~b+(Wi,bm)+[Fb+,m=0−F~b+(Wi,b0)].(36)F_b^{+,m}=\widetilde{F}_b^+(\mathbf{W}_{i,b}^m)+\left[F_b^{+,m=0}-\widetilde{F}_b^+(\mathbf{W}_{i,b}^0)\right]. \tag*{(36)}

Here Wi,bm\mathbf{W}_{i,b}^m is interpolated to boundary face bb from the adjacent cell ii, using the same limited reconstruction as in Eq. (15). Let F~b−(Wi,b,Fb+)\widetilde{F}_b^-(\mathbf{W}_{i,b},F_b^+) denote the boundary-condition-dependent mapping for the continuous incoming half-space flux. The incoming flux is constructed analogously as

Fb−,m=F~b−(Wi,bm,Fb+,m)+[Fb−,m=0−F~b−(Wi,b0,Fb+,m=0)].(37)F_b^{-,m}=\widetilde{F}_b^-(\mathbf{W}_{i,b}^m,F_b^{+,m})+\left[F_b^{-,m=0}-\widetilde{F}_b^-(\mathbf{W}_{i,b}^0,F_b^{+,m=0})\right]. \tag*{(37)}

The resulting boundary flux used in the PBC residual is

Fbm=Fb+,m+Fb−,m.(38)F_b^m=F_b^{+,m}+F_b^{-,m}. \tag*{(38)}

At a common fixed point of the kinetic and macroscopic iterations, the state returned by the PBC solver satisfies Win+1=Wi0\mathbf{W}_i^{n+1}=\mathbf{W}_i^0; consequently, the reconstructed boundary state used in the final PBC flux coincides with Wi,b0\mathbf{W}_{i,b}^0. Equations (36) and (37) then recover the stored discrete reference values of both half-space fluxes. The total macroscopic boundary flux therefore coincides with the discrete kinetic boundary flux.

The closed-form expressions for Feq±F_{\mathrm{eq}}^{\pm} and the boundary-specific forms of F~b−\widetilde{F}_{b}^{-} for prescribed mass-flow boundaries, a stationary fully diffuse wall, pressure inlets and outlets, and far-field boundaries are given in Appendix § 10.

With the internal and boundary face fluxes specified, the conservative-variable residual of cell ii is

Ri=[Rρ,iRmom,iRE,i]=∑j∈Niint(Fijadv+Fijvis)+∑b∈FibFb.(39)\mathcal{R}_i = \begin{bmatrix} \mathcal{R}_{\rho,i}\\ \mathcal{R}_{\mathrm{mom},i}\\ \mathcal{R}_{E,i} \end{bmatrix} = \sum_{j\in\mathcal{N}_i^{\mathrm{int}}} \left(F_{ij}^{\mathrm{adv}}+F_{ij}^{\mathrm{vis}}\right) + \sum_{b\in\mathcal{F}_i^b}F_b . \tag*{(39)}

Coupled pressure–velocity solution

The pressure–velocity residual is formed from the mass and momentum components of Eq. (39):

Ripu≡[Rρ,iRmom,i],(40)\mathcal{R}_i^{pu}\equiv \begin{bmatrix} \mathcal{R}_{\rho,i}\\ \mathcal{R}_{\mathrm{mom},i} \end{bmatrix}, \tag*{(40)}

where the subscripts ρ\rho and mom\mathrm{mom} select the mass and momentum components, respectively. The global vector Rpu\mathcal{R}^{pu} is assembled from the cell residuals Ripu\mathcal{R}_i^{pu}. Let Wpu\mathbf{W}_{pu} collect the cell states Wpu,i=(pi,ui)T\mathbf{W}_{pu,i}=(p_i,\mathbf{u}_i)^{\mathrm{T}}, and let ΔWpu\Delta\mathbf{W}_{pu} collect the corresponding corrections ΔWpu,i=(Δpi,Δui)T\Delta\mathbf{W}_{pu,i}=(\Delta p_i,\Delta\mathbf{u}_i)^{\mathrm{T}}. At fixed temperature, the approximate linearization is

JpuΔWpu=−Rpu,Jpu=Jintpu+Jbndpu≈∂Rpu∂Wpu.(41)\mathbf{J}^{pu}\Delta\mathbf{W}_{pu}=-\mathcal{R}^{pu}, \qquad \mathbf{J}^{pu}=\mathbf{J}_{\mathrm{int}}^{pu}+\mathbf{J}_{\mathrm{bnd}}^{pu} \approx\frac{\partial\mathcal{R}^{pu}}{\partial\mathbf{W}_{pu}}. \tag*{(41)}

The nonlinear residual uses the selected spatial reconstruction, whereas the approximate Jacobian is evaluated from the first-order cell-centred flux representation. Here Jintpu\mathbf{J}_{\mathrm{int}}^{pu} and Jbndpu\mathbf{J}_{\mathrm{bnd}}^{pu} denote the contributions to the approximate Jacobian from internal and boundary faces, respectively; these two contributions are constructed below.

For the internal-face part Jintpu\mathbf{J}_{\mathrm{int}}^{pu}, the mass and momentum components of Eqs. (28) and (32) give the cellwise Jacobian action

(JintpuΔWpu)i=∑j∈Niint∑s∈{i,j}[Jρ,ij,sJmom,ij,s]ΔWpu,s,(42)\left(\mathbf{J}_{\mathrm{int}}^{pu}\Delta\mathbf{W}_{pu}\right)_i = \sum_{j\in\mathcal{N}_i^{\mathrm{int}}} \sum_{s\in\{i,j\}} \begin{bmatrix} J_{\rho,ij,s}\\ J_{\mathrm{mom},ij,s} \end{bmatrix} \Delta\mathbf{W}_{pu,s}, \tag*{(42)}

where the 1×41\times4 continuity row and 3×43\times4 momentum block are

Jρ,ij,s=∂m˙ij∂Wpu,s,Jmom,ij,s=∂∂Wpu,s(m˙ijuup+pijSij+tij).(43)J_{\rho,ij,s}=\frac{\partial\dot{m}_{ij}}{\partial\mathbf{W}_{pu,s}}, \qquad J_{\mathrm{mom},ij,s} = \frac{\partial}{\partial\mathbf{W}_{pu,s}} \left(\dot{m}_{ij}\mathbf{u}_{\mathrm{up}}+p_{ij}\mathbf{S}_{ij}+\mathbf{t}_{ij}\right). \tag*{(43)}

The continuity row is evaluated with the face weights, βijRC\beta_{ij}^{RC}, and the reconstructed pressure gradient fixed at the current nonlinear state. Keeping iupi_{\mathrm{up}} fixed during the linearization gives

∂m˙ij∂pi=ρijβijRC+δi,iupV˙ijRCRTi,∂m˙ij∂pj=−ρijβijRC+δj,iupV˙ijRCRTj,\frac{\partial\dot{m}_{ij}}{\partial p_i} = \rho_{ij}\beta_{ij}^{RC} + \delta_{i,i_{\mathrm{up}}}\frac{\dot{\mathcal{V}}_{ij}^{RC}}{RT_i}, \qquad \frac{\partial\dot{m}_{ij}}{\partial p_j} = -\rho_{ij}\beta_{ij}^{RC} + \delta_{j,i_{\mathrm{up}}}\frac{\dot{\mathcal{V}}_{ij}^{RC}}{RT_j},
∂m˙ij∂ui=ρijwiSijT,∂m˙ij∂uj=ρijwjSijT.(44)\frac{\partial\dot{m}_{ij}}{\partial\mathbf{u}_i} = \rho_{ij}w_i\mathbf{S}_{ij}^{\mathrm{T}}, \qquad \frac{\partial\dot{m}_{ij}}{\partial\mathbf{u}_j} = \rho_{ij}w_j\mathbf{S}_{ij}^{\mathrm{T}}. \tag*{(44)}

where δs,iup\delta_{s,i_{\mathrm{up}}}, with s∈{i,j}s \in\{i,j\}, is the Kronecker delta. The NSF part of the face-integrated stress tij\mathbf{t}_{ij} in Eq. (31) is evaluated using the non-orthogonally corrected gradient in Eq. (29). Its approximate linearization retains the two-point normal response

ΔtijNSF≃Cμ,ij(Δui−Δuj),Cμ,ij=μijAijdn,ij(I3+13nijnijT).(45)\Delta\mathbf{t}_{ij}^{\mathrm{NSF}} \simeq C_{\mu,ij}(\Delta\mathbf{u}_i-\Delta\mathbf{u}_j), \qquad C_{\mu,ij}=\frac{\mu_{ij}A_{ij}}{d_{n,ij}}\left(I_3+\frac{1}{3}\mathbf{n}_{ij}\mathbf{n}_{ij}^{\mathrm{T}}\right). \tag*{(45)}

Here I3I_3 is the 3×33 \times3 identity matrix. The non-orthogonal part, transport-coefficient derivatives, and HoTσ,ijn+1/2\mathrm{HoT}_{\sigma,ij}^{n+1/2} remain in the nonlinear residual but are not differentiated in this approximate Jacobian. To expand the momentum block, define

Eu=[03×1  I3],epT=[1  0  0  0].(46)E_u=\left[0_{3\times1}\ \ I_3\right], \qquad e_p^{\mathrm{T}}=\left[1\ \ 0\ \ 0\ \ 0\right]. \tag*{(46)}

Here EuE_u selects the three velocity components of a pressure–velocity increment, while epTe_p^{\mathrm{T}} selects its pressure component. With ζi=1\zeta_i=1, ζj=−1\zeta_j=-1, the 3×43 \times4 derivative of the oriented momentum flux with respect to cell s∈{i,j}s \in\{i,j\} is

Jmom,ij,s=uupJρ,ij,s+m˙ijδs,iupEu+wsSijepT+ζsCμ,ijEu.(47)J_{\mathrm{mom},ij,s}=\mathbf{u}_{\mathrm{up}}J_{\rho,ij,s}+\dot{m}_{ij}\delta_{s,i_{\mathrm{up}}}E_u+w_sS_{ij}e_p^{\mathrm{T}}+\zeta_sC_{\mu,ij}E_u. \tag*{(47)}

The four terms represent, respectively, the response of the mass flow, the donor velocity, the face pressure, and the two-point NSF stress.

For the boundary part JbndpuJ_{\mathrm{bnd}}^{\mathrm{pu}}, the consistency corrections in Eqs. (36) and (37) are fixed during the PBC iterations. The boundary residual retains the reconstructed face state Wi,bW_{i,b}, whereas the approximate Jacobian is evaluated with Wi,b=WiW_{i,b}=W_i. Omitting the iteration superscript, the cellwise boundary contribution is

(JbndpuΔWpu)i=∑b∈FibJb,ipuΔWpu,i.(48)\left(J_{\mathrm{bnd}}^{\mathrm{pu}}\Delta W_{\mathrm{pu}}\right)_i=\sum_{b\in\mathcal{F}_i^b}J_{b,i}^{\mathrm{pu}}\Delta W_{\mathrm{pu},i}. \tag*{(48)}

where the 4×44 \times4 boundary-face block is

Jb,ipu=[(I5+∂F~b−(Wi,Fb+)∂Fb+)∂F~b+(Wi)∂Wpu,i+∂F~b−(Wi,Fb+)∂Wpu,i]ρ,mom.(49)J_{b,i}^{\mathrm{pu}}=\left[\left(I_5+\frac{\partial\widetilde{F}_b^{-}(W_i,F_b^{+})}{\partial F_b^{+}}\right)\frac{\partial\widetilde{F}_b^{+}(W_i)}{\partial W_{\mathrm{pu},i}}+\frac{\partial\widetilde{F}_b^{-}(W_i,F_b^{+})}{\partial W_{\mathrm{pu},i}}\right]_{\rho,\mathrm{mom}}. \tag*{(49)}

Here I5I_5 is the 5×55 \times5 identity matrix, and the subscript (ρ,mom)(\rho,\mathrm{mom}) retains the mass and momentum rows. Detailed expressions for the boundary-specific derivatives are given in Appendix C. Together, the internal and boundary face sums form JpuJ^{\mathrm{pu}} in Eq. (41).

Temperature solution

Following the coupled pressure–velocity solve, pp and u\mathbf{u} are held fixed during a separate temperature update, consistent with the segregated energy treatment in pressure-based coupled methods [44]. At fixed pressure, however, the equation of state couples TT and ρ\rho, so a temperature correction also changes the density and the bulk kinetic-energy density. To isolate the thermal balance, the total-energy residual RER_E is combined with the mass and momentum residuals to form the temperature residual RT\mathcal{R}_T, in line with related conservative-residual transformations used for enthalpy updates [45]. The relation ρE=3ρRT/2+ρ∣u∣2/2\rho E = 3\rho RT/2+\rho|\mathbf{u}|^2/2 gives

32ρR dT=d(ρE)−u⋅d(ρu)+(∣u∣22−32RT)dρ.(50)\frac{3}{2}\rho R\,\mathrm{d}T = \mathrm{d}(\rho E) - \mathbf{u}\cdot\mathrm{d}(\rho\mathbf{u}) + \left(\frac{|\mathbf{u}|^2}{2}-\frac{3}{2}RT\right)\mathrm{d}\rho. \tag*{(50)}

Using the conservative-variable transformation in Eq. (50), with an additional correction for the discrete kinetic-energy defect, the cell temperature residual is defined as

RT,i=RE,i−ui⋅Rmom,i+(Ki−32RTi)Rρ,i−∑j∈NiintKi,ij.(51)\mathcal{R}_{T,i} = \mathcal{R}_{E,i} - \mathbf{u}_i\cdot\boldsymbol{\mathcal{R}}_{\mathrm{mom},i} + \left(K_i-\frac{3}{2}RT_i\right)\mathcal{R}_{\rho,i} - \sum_{j\in\mathcal{N}^{\mathrm{int}}_i}\mathcal{K}_{i,ij}. \tag*{(51)}

Here Ki=∣ui∣2/2K_i=|\mathbf{u}_i|^2/2. The conservative-to-temperature residual transformation introduces the discrete kinetic-energy defect Ki,ij=m˙ij∣uup−ui∣2/2\mathcal{K}_{i,ij}=\dot{m}_{ij}|\mathbf{u}_{\mathrm{up}}-\mathbf{u}_i|^2/2, where uup\mathbf{u}_{\mathrm{up}} is the velocity used in the numerical advective flux. This defect is subtracted in Eq. (51) to exclude this discretization-induced contribution from the thermal balance.

Applying the corrected residual definition in Eq. (51) to the internal-face fluxes in Eqs. (28) and (32), and to the complete mesoscopic–macroscopic boundary flux in Eq. (38), gives, respectively,

FT,ij=m˙ij(52RTup−32RTi)−pijui⋅Sij+tij⋅(uij−ui)+Qij,FT,i,b=FE,b−ui⋅Fmom,b+(Ki−32RTi)Fρ,b.(52)\begin{aligned} F_{T,ij} &= \dot{m}_{ij}\left(\frac{5}{2}RT_{\mathrm{up}}-\frac{3}{2}RT_i\right)-p_{ij}\mathbf{u}_i\cdot\mathbf{S}_{ij}+\mathbf{t}_{ij}\cdot(\mathbf{u}_{ij}-\mathbf{u}_i)+Q_{ij},\\ F_{T,i,b} &= F_{E,b}-\mathbf{u}_i\cdot\mathbf{F}_{\mathrm{mom},b}+\left(K_i-\frac{3}{2}RT_i\right)F_{\rho,b}. \tag*{(52)} \end{aligned}

Let RT\boldsymbol{\mathcal{R}}_T, T\mathbf{T}, and ΔT\Delta\mathbf{T} collect the cell residuals, temperatures, and temperature corrections, respectively. At fixed pressure and velocity, the approximate Jacobian JT≈∂RT/∂TJ_T\approx\partial\mathcal{R}_T/\partial T is assembled cellwise as

(JTΔT)i=∑j∈NiintJT,ijint(ΔTi,ΔTj)T+∑b∈FibJT,i,bΔTi=−RT,i.(53)(J_T\Delta\mathbf{T})_i=\sum_{j\in\mathcal{N}^{\mathrm{int}}_i}J^{\mathrm{int}}_{T,ij}(\Delta T_i,\Delta T_j)^{\mathsf{T}}+\sum_{b\in\mathcal{F}^b_i}J_{T,i,b}\Delta T_i=-\mathcal{R}_{T,i}. \tag*{(53)}

To construct JT,ijintJ^{\mathrm{int}}_{T,ij} in Eq. (53), all quantities in FT,ijF_{T,ij} other than TiT_i and TjT_j are frozen at the start of the temperature update, including the mass flow, donor-cell selection, face traction, transport coefficients, HoTs, and non-orthogonal heat-flux correction. With the same first-order approximation as in the pressure–velocity system, the resulting 1×21\times2 row JT,ijint=∂FT,ij/∂(Ti,Tj)J^{\mathrm{int}}_{T,ij}=\partial F_{T,ij}/\partial(T_i,T_j) is

JT,ijint=Rm˙ij{[10],m˙ij≥0,[−3232],m˙ij<0+Cκ,ij[1−1],(54)J^{\mathrm{int}}_{T,ij}=R\dot{m}_{ij} \begin{cases} \left[1\quad0\right], & \dot{m}_{ij}\geq0,\\ \left[-\frac{3}{2}\quad\frac{3}{2}\right], & \dot{m}_{ij}<0 \end{cases} +C_{\kappa,ij}\left[1\quad-1\right], \tag*{(54)}

where Cκ,ij=κijAij/dn,ijC_{\kappa,ij}=\kappa_{ij}A_{ij}/d_{n,ij}. For the boundary face bb, the corresponding Jacobian is

JT,i,b=∂FE,b∂Ti−ui⋅∂Fmom,b∂Ti+(Ki−32RTi)∂Fρ,b∂Ti−32RFρ,b.(55)J_{T,i,b}=\frac{\partial F_{E,b}}{\partial T_i}-\mathbf{u}_i\cdot\frac{\partial\mathbf{F}_{\mathrm{mom},b}}{\partial T_i}+\left(K_i-\frac{3}{2}RT_i\right)\frac{\partial F_{\rho,b}}{\partial T_i}-\frac{3}{2}RF_{\rho,b}. \tag*{(55)}

Here ∂Fb/∂Ti\partial\mathbf{F}_b/\partial T_i is the final column of the full boundary-flux Jacobian in Eq. (C.1). Its boundary-specific forms are given by Eqs. (C.4), (C.7), and (C.8) in Appendix § 11.

Pseudo-time stabilization and linear solution

The spatial systems derived above are stabilized by block-diagonal local pseudo-time terms, following the general idea of pseudo-transient continuation [46]. For cell ii, the pseudo-time step is

Δtipse=CFLΛispecVi,Λispec=∑j∈NiAij[∣uij⋅nij∣+χa,ijcs,ij+8μij3ρidn,ij],χa,ij=min⁡[1,max⁡(0.01,Uijcs,ij)],Uij=∣ui∣2+∣uj∣22,cs,ij=cs,i+cs,j2.(56)\begin{aligned} \Delta t_i^{\mathrm{pse}} &= \frac{\mathrm{CFL}}{\Lambda_i^{\mathrm{spec}}}V_i,\\ \Lambda_i^{\mathrm{spec}} &= \sum_{j\in N_i} A_{ij}\left[|u_{ij}\cdot n_{ij}|+\chi_{a,ij}c_{s,ij}+\frac{8\mu_{ij}}{3\rho_i d_{n,ij}}\right],\\ \chi_{a,ij} &= \min\left[1,\max\left(0.01,\frac{U_{ij}}{c_{s,ij}}\right)\right],\qquad U_{ij}=\sqrt{\frac{|\mathbf{u}_i|^2+|\mathbf{u}_j|^2}{2}},\qquad c_{s,ij}=\frac{c_{s,i}+c_{s,j}}{2}. \tag*{(56)} \end{aligned}

Here Δtipse\Delta t_i^{\mathrm{pse}} is the local pseudo-time step, CFL is the prescribed Courant number, and Λispec\Lambda_i^{\mathrm{spec}} is the corresponding local spectral-radius estimate. The sound speed is cs,i=γRTic_{s,i}=\sqrt{\gamma RT_i}, with γ=5/3\gamma=5/3. The density in the viscous term remains the cell density ρi\rho_i, consistent with the momentum inertia of cell ii. At a physical boundary face, the velocity, sound speed, and viscosity are evaluated from the adjacent interior cell. For a boundary face, dn,ij=2∣xij−xi∣d_{n,ij}=2|\mathbf{x}_{ij}-\mathbf{x}_i|. The prescribed lower bound of 0.01 retains a nonzero acoustic contribution as the local velocity vanishes. The corresponding cell-local pseudo-time blocks for the pressure–velocity and temperature systems are

Mpu,i=ViΔtipse[00T0ρiI3],MT,i=3ρiR2ΔtipseVi.(57)M_{pu,i}=\frac{V_i}{\Delta t_i^{\mathrm{pse}}} \begin{bmatrix} 0 & \mathbf{0}^{\mathrm{T}}\\ \mathbf{0} & \rho_i I_3 \end{bmatrix}, \qquad M_{T,i}=\frac{3\rho_i R}{2\Delta t_i^{\mathrm{pse}}}V_i. \tag*{(57)}

Assembling the cell-local blocks in Eq. (57) into the global block-diagonal matrices Mpu\mathbf{M}_{pu} and MT\mathbf{M}_T gives the stabilized linear systems

(Jpu+Mpu)ΔWpu=−Rpu,(JT+MT)ΔT=−RT,(58)\left(\mathbf{J}^{pu}+\mathbf{M}_{pu}\right)\Delta\mathbf{W}_{pu}=-\mathbf{R}^{pu}, \qquad \left(\mathbf{J}_T+\mathbf{M}_T\right)\Delta T=-\mathbf{R}_T, \tag*{(58)}

The resulting linear systems can be solved using standard iterative methods, either implemented directly or provided by existing numerical libraries.

Numerical examples

This section assesses the accuracy, robustness, and computational efficiency of GSIS-PBC through four three-dimensional test cases. The tube flow case examines the performance of GSIS-PBC over a wide range of gas rarefaction. The dynamic gas lock tests convergence under prescribed mass-flow boundary conditions, while the divertor case examines large-scale simulation in a complex internal geometry. The sphere flow case evaluates the capability of GSIS-PBC for high-speed flows.

Note that the axisymmetric geometries used here are chosen solely to facilitate validation against reference solutions. All GSIS-PBC simulations use full three-dimensional meshes, and the method is applicable to both axisymmetric and non-axisymmetric configurations. The DSMC simulations performed in this work use the two-dimensional axisymmetric formulation in SPARTA [47] with the variable-hard-sphere collision model.

Convergence is reached when the relative change between two consecutive iterations satisfies

ϵ=max⁡Q∈{ρ,u,T}[∑i∣Qin−Qin−1∣2Vi∑i∣Qin−1∣2Vi]1/2<10−5.(59)\epsilon= \max_{Q\in\{\rho,u,T\}} \left[\frac{\sum_i \left|Q_i^n-Q_i^{n-1}\right|^2 V_i}{\sum_i \left|Q_i^{n-1}\right|^2 V_i}\right]^{1/2} < 10^{-5}. \tag*{(59)}

A CFL number of 500 is used for the PBC pseudo-time step in all cases. The sphere and tube simulations are performed on an Intel Xeon Silver 4208 [email protected] GHz platform, while the dynamic gas lock and divertor simulations are performed on an AMD EPYC 7452 [email protected] GHz platform. The reported GSIS-PBC wall-clock times include initialization and iterations. The cost per GSIS iteration varies with the number of linear-solver iterations required for the macroscopic systems in Eq. (58).

Flow through a circular tube into vacuum

To assess the applicability of GSIS-PBC over a wide range of gas rarefaction, we consider helium flow through a circular tube into vacuum [48], see Figure 2. The tube radius is taken the LrefL_{\mathrm{ref}}, giving a tube length of 10. The upstream reservoir has a radius and length of 8, while the corresponding dimensions of the downstream reservoir are both 4. The molecular mass of helium is m=6.65×10−27 kgm = 6.65 \times10^{-27}\,\mathrm{kg}. Its viscosity at Tref=T0=273.15 KT_{\mathrm{ref}} = T_0 = 273.15\,\mathrm{K} is μ0=μ(T0)=1.916×10−5 Pa s\mu_0 = \mu(T_0) = 1.916 \times10^{-5}\,\mathrm{Pa\,s}, and the viscosity index is ω=0.66\omega= 0.66. At the upstream end, the temperature is fixed at T0T_0, and the inlet pressure pinp_{\mathrm{in}} is adjusted to obtain the prescribed rarefaction parameter δrp\delta_{\mathrm{rp}}. A vacuum boundary is applied at the downstream end. All remaining boundaries are stationary, fully diffuse walls maintained at T0T_0. The parameter δrp\delta_{\mathrm{rp}} is related to the Knudsen number in Eq. (8), evaluated using the inlet reservoir state and the tube radius, by

Geometry and dimensions of the tube-flow configuration

Figure 2. (a) Computational domain. (b) Dimensionless geometry. Figure 2: Geometry and dimensions of the tube-flow configuration. All dimensions are normalized by the tube radius.

δrp=π2Kn.(60)\delta_{\mathrm{rp}} = \frac{\sqrt{\pi}}{2\mathrm{Kn}}. \tag*{(60)}

Four cases with δrp=1\delta_{\mathrm{rp}} = 1, 10, 100, and 1,000 are considered. All four cases use the same three-dimensional spatial mesh containing 135,000 cells and a uniform Cartesian velocity grid of 30×24×2430 \times24 \times24 points over ξ/RT0∈[−10,10]×[−8,8]×[−8,8]\boldsymbol{\xi}/\sqrt{RT_0} \in[-10,10] \times[-8,8] \times[-8,8]. GSIS-PBC and CIS use the same spatial mesh, velocity grid, and convergence criterion. Each GSIS-PBC iteration consists of five kinetic iterations followed by ten PBC iterations.

Figure 3 compares the centreline pressure and axial velocity obtained with GSIS-PBC and CIS at δrp=1\delta_{\mathrm{rp}} = 1, 10, and 100, with DSMC results also shown for δrp=100\delta_{\mathrm{rp}} = 100. GSIS-PBC and CIS agree closely at δrp=1\delta_{\mathrm{rp}} = 1 and 10. At δrp=100\delta_{\mathrm{rp}} = 100, GSIS-PBC agrees well with DSMC, whereas CIS exhibits noticeable deviations due to the large numerical dissipation when coarse spatial grids are used, particularly in the axial velocity.

Profiles of pressure and axial velocity along the tube centreline

Figure 3. Profiles of pressure and axial velocity along the tube centreline from the inlet (X=0X = 0) to the outlet (X=10X = 10) at δrp=1\delta_{\mathrm{rp}} = 1, 10, and 100. Lines denote GSIS-PBC, black circles denote DSMC at δrp=100\delta_{\mathrm{rp}} = 100, and the other symbols denote CIS. Here X=x/LrefX = x/L_{\mathrm{ref}} and P=p/pinP = p/p_{\mathrm{in}}, and the axial velocity is normalized by 2RT0\sqrt{2RT_0}.

The dimensionless mass flow rate M˙\dot{M} is obtained by normalizing the total mass flow rate by πLref2pin/2RT0\sqrt{\pi}L_{\mathrm{ref}}^{2}p_{\mathrm{in}}/\sqrt{2RT_{0}}. Table 1 compares the predicted mass flow rates with the DSMC and NS reference results and summarizes the iteration counts and total wall-clock times of GSIS-PBC and CIS. The relative mass-flow deviations of GSIS-PBC from DSMC are at most 1.45% for δrp=1\delta_{\mathrm{rp}} = 1, 10, and 100, while the deviation from the NS result is 0.63% at δrp=1,000\delta_{\mathrm{rp}} = 1,000. In contrast, CIS underpredicts the reference mass flow rates by 33.52% and 42.46% at δrp=100\delta_{\mathrm{rp}} = 100 and 1,000, respectively. These cases span inlet Knudsen numbers from 8.86×10−48.86 \times10^{-4} to 8.86×10−18.86 \times10^{-1}. At small Knudsen numbers, traditional kinetic solvers lacking AP properties may introduce substantial numerical dissipation on coarse meshes, leading to underpredicted mass flow rates. For example, at δrp=100\delta_{\mathrm{rp}} = 100, the largest cell size is approximately 57 times the inlet mean free path, so the CIS becomes inaccurate. The GSIS, however, can give accurate predictions due to its AP property.

δrp\delta_{\mathrm{rp}}M˙ref\dot{M}_{\mathrm{ref}}\multicolumn{4}{c}{GSIS-PBC}\multicolumn{4}{c}{CIS}Speedup
M˙\dot{M}Error (%)Iter.tt (h)M˙\dot{M}Error (%)Iter.tt (h)
10.1970.1980.51170.200.1980.511890.432.2×
100.3300.3261.2190.110.3485.454841.059.5×
1000.8980.8851.4570.090.59733.522,7966.2068.9×
1,0001.2671.2750.6360.080.72942.467,84716.03200.4×

Table 1. Mass-flow comparison and computational costs for tube flow into vacuum. The reference values are DSMC results for δrp=1\delta_{\mathrm{rp}} = 1, 10, and 100, and a NSF result computed by FLUENT for δrp=1,000\delta_{\mathrm{rp}} = 1,000. The two-dimensional axisymmetric DSMC simulations use approximately 0.225, 0.899, and 4.2 million spatial cells for δrp=1\delta_{\mathrm{rp}} = 1, 10, and 100, respectively, with an average of 44 simulation particles per cell. Using 640 cores, the corresponding wall-clock times are 0.55, 2.19, and 10.2 h. Relative mass-flow errors are ∣M˙−M˙ref∣/M˙ref|\dot{M}-\dot{M}_{\mathrm{ref}}|/\dot{M}_{\mathrm{ref}}, reported as percentages. GSIS-PBC iteration counts refer to GSIS iterations, and tt denotes the total wall-clock time. The speedup of GSIS-PBC relative to CIS is estimated by dividing the reported total wall-clock time of CIS by that of GSIS-PBC.

Using 96 cores, GSIS-PBC converges in 6–17 iterations, with total wall-clock times ranging from 0.08 to 0.20 h. For δrp=1\delta_{\mathrm{rp}} = 1, 10, 100, and 1,000, the wall-clock speedups over CIS are approximately 2.2, 9.5, 68.9, and 200.4, respectively. The increasing speedup as the inlet Knudsen number decreases indicates a growing computational advantage of GSIS-PBC over CIS for the tested cases.

Flows in dynamic gas lock

The computational cost of dynamic gas lock (DGL) simulations motivates the development of efficient kinetic solvers [32]. This case examines how rapidly GSIS-PBC brings the boundary state and the interior flow field to convergence under prescribed mass-flow boundary conditions. The geometry is adapted from Teng et al. [1]. The three-dimensional computational domain is constructed by revolving the meridional half-section through 360° about its symmetry axis, as shown in Figure 4. Two coaxial cylindrical chambers, each 200 mm in diameter and 100 mm in height, are connected by a tapered central passage. Helium is supplied radially through a 3 mm-high annular inlet between the chambers. A circular wafer is located in the lower chamber. The inlet gas and all walls, including the wafer, are maintained at T0=300T_0 = 300 K. Three inlet flow rates of 10, 100, and 1,000 sccm are considered. The upper and lower outlet pressures are fixed at 7 and 1 Pa.

Geometry and boundary conditions of the dynamic gas lock

Figure 4. Geometry and boundary conditions of the dynamic gas lock. All dimensions are in millimetres.

The three-dimensional spatial mesh contains 850,080 cells, and the molecular velocity domain ξ/RT0∈[−8,8]3\boldsymbol{\xi}/\sqrt{RT_0} \in[-8, 8]^3 is discretized using a uniform Cartesian grid of 24×24×2424 \times24 \times24 points.

Figure 5 compares the horizontal and axial velocity profiles at inlet flow rates of 10 and 100 sccm. The GSIS-PBC predictions agree well with the DSMC results for both cases. The corresponding pressure distributions and streamlines are shown in Figure 6. Table 2 summarizes the net mass flow rates through the upper outlet, with positive values denoting flow out of the computational domain. At 10 sccm, the streamlines indicate inflow through the upper opening, consistent with the negative net mass flow rate, whereas increasing the inlet flow rate to 100 sccm produces net outflow. The maximum relative deviation of GSIS-PBC from the reference mass flow rates is approximately 1.25%.

Velocity profiles in the Z = 0 plane of the dynamic gas lock

Figure 5. Velocity profiles in the Z=0Z = 0 plane of the dynamic gas lock at inlet flow rates of 10 and 100 sccm: (a) horizontal velocity along Y=95Y = 95; (b) axial velocity along X=0X = 0. Coordinates are in millimetres, and U=ux/2RT0U = u_x/\sqrt{2RT_0}, V=uy/2RT0V = u_y/\sqrt{2RT_0}. Lines denote GSIS-PBC, and squares denote DSMC.

Pressure distributions and streamlines in the DGL at 10 sccmPressure distributions and streamlines in the DGL at 100 sccm

Figure 6. Pressure distributions and streamlines in the DGL.

Inlet (sccm)Reference (sccm)GSIS-PBC (sccm)Error (%)
10-18.73-18.660.37
10057.4556.731.25
1,000762.70768.220.72

Table 2. Net mass flow rates through the upper outlet, expressed in sccm (standard cubic centimeter per minute). The reference values are obtained from DSMC for inlet flow rates of 10 and 100 sccm, and from Fluent for 1,000 sccm. Errors are relative to these reference values.

The DGL simulations are performed using 640 cores. The GSIS with a density-based macroscopic solver [25, 27], denoted GSIS-DB, serves as the baseline for the convergence comparison. Each GSIS-DB iteration consists of 2 kinetic iterations followed by 400 density-based macroscopic iterations, whereas each GSIS-PBC iteration consists of 10 kinetic iterations followed by 10 PBC iterations. For inlet flow rates of 10 and 100 sccm, GSIS-PBC converges in 16 and 20 iterations, requiring 0.35 and 0.44 h, respectively. In contrast, GSIS-DB requires at least 2,818 iterations and 13.23 h for both cases. The corresponding wall-clock speedups of GSIS-PBC are approximately 60 and 30, respectively. At 1,000 sccm, GSIS-PBC converges in 9 iterations and 0.21 h. For reference, the two-dimensional axisymmetric DSMC simulations use 225,000 spatial cells, with 12 and 18 million simulation particles for 10 and 100 sccm, and require 4.12 and 12.55 h, respectively. These results demonstrate the rapid convergence of GSIS-PBC for the present DGL flows under prescribed mass-flow boundary conditions.

Flows in a divertor in nuclear fusion

This case examines the convergence and computational efficiency of GSIS-PBC for large-scale three-dimensional gas flow in a complex divertor geometry is derived from the divertor component of the publicly available simplified ITER tokamak model.2 The full three-dimensional domain is shown in Figure 7(a), and the cross-sectional geometry and boundary labels are given in Figure 7(b). Helium is supplied through two pressure inlets, I1 and I2, at 10 Pa and 420 K. Taking Lref=1L_{\mathrm{ref}} = 1 cm, the Knudsen number based on the inlet state is Kn=0.298\mathrm{Kn} = 0.298. All solid surfaces are stationary, fully diffuse reflecting walls maintained at 420 K. The pumping probability is set to zero, so the pumping surface is also treated as a wall. Vacuum boundary conditions are imposed at the leakage openings O1–O4. The spatial mesh, shown in Figure 7(c), contains 12,248,344 polyhedral cells and 80,495,318 faces. The molecular velocity space contains 10,648 discrete points.

Computational domain, right-hand portion of the $y=0$ cross-section, polyhedral mesh, and local pressure field and streamlines

Figure 7. (a) Computational domain. (b) Right-hand portion of the y=0y = 0 cross-section. (c) Polyhedral mesh. (d) Local pressure field and streamlines. Computational domain, mesh, and local flow in the divertor. Green and red mark the pressure inlets and vacuum leakage outlets, respectively; the remaining boundaries are walls.

The streamlines in Figure 7(d) show how the inlet gas spreads through the central cavity into the connected leakage passages. The pressure remains relatively high in the cavity and decreases markedly towards the vacuum openings, as shown in Figure 8(a). The flow is predominantly at low Mach numbers, with Ma∼10−2\mathrm{Ma} \sim10^{-2} over most of the domain, while the velocity increases appreciably near the vacuum openings.

Cross-sectional pressure and z-velocity distributions in the divertor

Figure 8. Cross-sectional pressure and zz-velocity distributions in the divertor. Coordinates are in millimetres; pressure and velocity are in Pa and m/s, respectively.

After initialization with 50 macroscopic and five kinetic iterations, each GSIS iteration comprises five kinetic and ten PBC iterations. Using 3,200 CPU cores, GSIS-PBC converges in 11 iterations with a total wall-clock time of 0.34 h. The rapid convergence obtained for this predominantly low-Mach-number internal flow demonstrates the effectiveness of the pressure–velocity coupled macroscopic solver. In Ref. [5], DSMC simulations consumed approximately 6.9×1056.9 \times10^{5} CPU core-hours to simulate a 10.8∘10.8^\circ sector of the DTT divertor.

By exploiting toroidal symmetry [49], the GSIS-PBC cost for the same toroidal extent is estimated at approximately 33 CPU core-hours, giving a cross-case CPU-cost ratio of approximately 2×1042 \times10^{4}. Moreover, under the same computational configuration as GSIS-PBC, GSIS-DB requires at least 23.7 h to achieve inlet–outlet mass-flow balance, approximately 70 times the total wall-clock time required for GSIS-PBC to converge.

At convergence, the net inflow rates predicted by GSIS-PBC through I1 and I2 are 6.338×10−46.338 \times10^{-4} and 1.237×10−31.237 \times10^{-3} kg/s, respectively. The leakage outflow rates through O1–O4 are 1.143×10−41.143 \times10^{-4}, 3.945×10−43.945 \times10^{-4}, 9.522×10−49.522 \times10^{-4}, and 4.097×10−44.097 \times10^{-4} kg/s, respectively. Under the present boundary conditions, O3 accounts for approximately half of the total leakage flow and therefore warrants particular attention in studies of neutral-gas leakage control.

Flow around a sphere

Finally, to assess the applicability of GSIS-PBC at high flow speeds, we consider argon flow around a sphere. The computational domain is bounded by a sphere of diameter D=Lref=1D = L_{\mathrm{ref}} = 1 m and a concentric spherical far-field boundary of radius 5 m, as shown in Figure 9. The molecular mass of argon is m=6.63×10−26m = 6.63 \times10^{-26} kg. Its viscosity follows Eq. (11), with μref=2.117×10−5\mu_{\mathrm{ref}} = 2.117 \times10^{-5} Pa s at Tref=273.15T_{\mathrm{ref}} = 273.15 K and ω=0.81\omega= 0.81. The freestream temperature and pressure are T∞=273.15T_{\infty} = 273.15 K and p∞=0.1266p_{\infty} = 0.1266 Pa, respectively, corresponding to Kn = 0.05 based on the sphere diameter. The freestream is directed along the positive xx-axis, with Ma∞=2\mathrm{Ma}_{\infty} = 2, 5, and 10, corresponding to velocities of 615.8, 1,539.5, and 3,079.0 m/s, respectively. The sphere surface is maintained at Tw=T∞T_w = T_{\infty}, with fully diffuse reflection imposed.

Computational domain

(a) Computational domain.

Cross-section of the spatial mesh

(b) Cross-section of the spatial mesh.

Figure 9. Geometry and spatial mesh for flow around a sphere. The sphere diameter is 1 m, and the spherical outer boundary has a radius of 5 m.

The GSIS-PBC computations use a common spatial mesh containing 360,000 cells, with a first-layer thickness of 0.005 m at the sphere surface. The molecular velocity space is discretized by uniform Cartesian grids with the resolutions listed in Table 3. Each calculation starts with 100 macroscopic iterations and five kinetic iterations to establish the initial flow field. The subsequent GSIS iterations use ten PBC iterations per macroscopic update, with six kinetic updates between successive macroscopic updates for Ma∞=2\mathrm{Ma}_{\infty} = 2 and 5, and ten for Ma∞=10\mathrm{Ma}_{\infty} = 10.

Ma∞\mathrm{Ma}_{\infty}MethodSpatial cellsVelocity grid or total particlesCoresIterationsTime (h)
2GSIS-PBC360,00030×24×2430 \times24 \times2428880.13
2DSMC100,3563,709,747200–0.27
5GSIS-PBC360,00030×24×2430 \times24 \times24288100.18
5DSMC100,3563,710,463200–0.39
10GSIS-PBC360,00040×36×3640 \times36 \times36480100.32
10DSMC182,6497,156,819200–1.15

Table 3. Computational settings and costs for sphere flow at Kn=0.05\mathrm{Kn} = 0.05. GSIS-PBC uses a three-dimensional domain, whereas DSMC uses a two-dimensional axisymmetric formulation. The fourth column gives the molecular velocity grid resolution for GSIS-PBC and the total number of simulation particles for DSMC. The iteration counts refer to GSIS iterations; times are total wall-clock times.

Figure 10 compares the streamwise velocity and temperature fields predicted by GSIS-PBC with the DSMC results. Good overall agreement is observed at all three Mach numbers. GSIS-PBC predicts a slightly earlier temperature rise than DSMC on the upstream side of the shock at Ma∞=5\mathrm{Ma}_{\infty} = 5, and this difference becomes more pronounced at Ma∞=10\mathrm{Ma}_{\infty} = 10. At Ma∞=2\mathrm{Ma}_{\infty} = 2, the difference is less evident. At higher Mach numbers, the Shakhov model’s velocity-independent collision frequency may allow energetic molecules to travel too far upstream, exaggerating the pre-shock temperature rise [50]. The comparisons support the applicability of GSIS-PBC to the present argon flows at freestream velocities up to Mach 10.

Comparison of streamwise velocity for argon flow around a sphere at $\mathrm{Ma}_{\infty}=2$

(a) Ma∞=2\mathrm{Ma}_{\infty}=2, uxu_x.

Comparison of temperature for argon flow around a sphere at $\mathrm{Ma}_{\infty}=2$

(b) Ma∞=2\mathrm{Ma}_{\infty}=2, TT.

Comparison of streamwise velocity for argon flow around a sphere at $\mathrm{Ma}_{\infty}=5$

(c) Ma∞=5\mathrm{Ma}_{\infty}=5, uxu_x.

Comparison of temperature for argon flow around a sphere at $\mathrm{Ma}_{\infty}=5$

(d) Ma∞=5\mathrm{Ma}_{\infty}=5, TT.

Comparison of streamwise velocity for argon flow around a sphere at $\mathrm{Ma}_{\infty}=10$

(e) Ma∞=10\mathrm{Ma}_{\infty}=10, uxu_x.

Comparison of temperature for argon flow around a sphere at $\mathrm{Ma}_{\infty}=10$

(f) Ma∞=10\mathrm{Ma}_{\infty}=10, TT.

Figure 10. Comparisons of streamwise velocity (left column) and temperature (right column) for argon flow around a sphere at Kn=0.05\mathrm{Kn}=0.05, at the z=0z=0 plane. Filled contours show the DSMC results, and black solid lines denote GSIS-PBC. The spatial coordinates are in metres.

The computational settings and costs are summarized in Table 3. The GSIS-PBC calculations use 288 cores for Ma∞=2\mathrm{Ma}_{\infty} = 2 and 5, and 480 cores for Ma∞=10\mathrm{Ma}_{\infty} = 10; the DSMC calculations use 200 cores. Steady macroscopic fields are obtained within ten GSIS iterations for all three cases, demonstrating rapid convergence over the tested Mach-number range. In terms of CPU core-hours, the three-dimensional GSIS-PBC solver runs marginally faster than its two-dimensional axisymmetric counterpart at Kn=0.05\mathrm{Kn}=0.05. If the axisymmetric GSIS formulation is adopted, the simulation runtime will be reduced by approximately one order of magnitude [51]. As the Knudsen number decreases, the required simulation mesh size and computational time increase dramatically for DSMC, whereas the GSIS scheme retains the same order of computational cost.

Conclusions

This work presents a pressure–based coupled general synthetic iterative scheme (GSIS–PBC) for efficient steady–state multiscale simulations of rarefied gas flows. By embedding a pressure–velocity coupled macroscopic solver within the GSIS framework and developing a consistent mesoscopic–macroscopic boundary treatment, the proposed method effectively mitigates the convergence deterioration exhibited by the original density–based GSIS for low–Mach–number internal flows, where large disparities between acoustic and convective time scales dominate.

The accuracy and computational performance of GSIS–PBC are systematically validated against DSMC references across a suite of three–dimensional benchmark cases. Numerical results demonstrate that GSIS–PBC inherits the asymptotic–preserving property of GSIS and maintains satisfactory prediction accuracy for both internal low–speed flows and external high–Mach–number rarefied flows. For near–continuum internal flows, GSIS–PBC achieves up to two orders of magnitude wall–clock speed–up compared with the conventional iterative kinetic solver. For challenging internal problems such as dynamic gas–lock and fusion divertor flows, GSIS–PBC substantially outperforms the original density–based GSIS and yields a speed–up of approximately four orders of magnitude relative to DSMC.

Future research will integrate the GSIS–PBC solver with adjoint shape optimization techniques [52] to facilitate multiscale design–oriented simulations for practical engineering systems, such as EUV lithography chambers, dynamic gas locks, and neutral–gas exhaust components in fusion facilities.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This work is supported by the National Natural Science Foundation of China (125B2052). The authors acknowledge the computing resources from the Center for Computational Science and Engineering at the Southern University of Science and Technology.

Data availability

Data will be made available on request.

Kinetic boundary implementation

This appendix specifies the incoming distribution fb,kbcf_{b,k}^{\mathrm{bc}} in Eq. (17) for the boundary conditions used in this work. The boundary normal and the half–spaces Ξb±\Xi_b^{\pm} follow Eq. (16). The outgoing distribution on Ξb+\Xi_b^{+} is reconstructed from the adjacent cell, whereas the formulas below prescribe the incoming distribution on Ξb−\Xi_b^{-}.

Prescribed mass-flow and fully diffuse wall boundaries

Let Γ\Gamma be a boundary on which the nonnegative inward mass-flow rate M˙Γ\dot{M}_{\Gamma} and the incoming Maxwellian temperature TΓT_{\Gamma} are prescribed. The total area and uniform inward mass-flux density are

AΓ=∑b∈ΓAb,j˙Γ=M˙ΓAΓ≥0.(61)A_{\Gamma} = \sum_{b\in\Gamma} A_b,\qquad\dot{j}_{\Gamma} = \frac{\dot{M}_{\Gamma}}{A_{\Gamma}} \ge0. \tag*{(61)}

The unit-density equilibrium at TΓT_{\Gamma} is represented by WΓ(1)=(RTΓ,0,TΓ)T\mathbf{W}_{\Gamma}^{(1)} = (RT_{\Gamma},\mathbf{0},T_{\Gamma})^{\mathrm{T}}. Define the discrete outgoing mass-flux density and the incoming mass-flux density of this unit-density equilibrium by

jb+=∑ξk∈Ξb+ϖ‾k(ξk⋅nb)fi,krec,j˙Γ,b−,(1)=∑ξk∈Ξb−ϖ‾k(ξk⋅nb)feq(WΓ(1);ξk).(62)j_b^{+} = \sum_{\boldsymbol{\xi}_k\in\Xi_b^{+}} \overline{\varpi}_k(\boldsymbol{\xi}_k\cdot\boldsymbol{n}_b)f_{i,k}^{\mathrm{rec}},\qquad\dot{j}_{\Gamma,b}^{-,(1)} = \sum_{\boldsymbol{\xi}_k\in\Xi_b^{-}} \overline{\varpi}_k(\boldsymbol{\xi}_k\cdot\boldsymbol{n}_b)f^{\mathrm{eq}}(\mathbf{W}_{\Gamma}^{(1)};\boldsymbol{\xi}_k). \tag*{(62)}

The incoming Maxwellian amplitude and distribution are

ρΓ,b=−j˙Γ−jb+j˙Γ,b−,(1),fb,kbc=ρΓ,bfeq(WΓ(1);ξk),ξk∈Ξb−.(63)\rho_{\Gamma,b} = \frac{-\dot{j}_{\Gamma}-j_b^{+}}{\dot{j}_{\Gamma,b}^{-,(1)}},\qquad f_{b,k}^{\mathrm{bc}} = \rho_{\Gamma,b}f^{\mathrm{eq}}(\mathbf{W}_{\Gamma}^{(1)};\boldsymbol{\xi}_k),\qquad\boldsymbol{\xi}_k\in\Xi_b^{-}. \tag*{(63)}

Equation (A.3) enforces the prescribed facewise mass-flux density in the discrete velocity space. All walls considered in this work are stationary; fully diffuse reflection at temperature TwT_w is obtained by setting j˙Γ=0\dot{j}_{\Gamma}=0 and TΓ=TwT_{\Gamma}=T_w.

Pressure inlet and outlet

The pressure boundaries use an incoming reservoir Maxwellian constructed from the reconstructed interior state Wi,b=(pi,b,ui,b,Ti,b)T\mathbf{W}_{i,b}=(p_{i,b},\boldsymbol{u}_{i,b},T_{i,b})^{\mathrm{T}}. The subscript (i,b)(i,b) denotes a quantity reconstructed from the adjacent cell ii to the interior side of boundary face bb; ρi,b\rho_{i,b} and cs,i,bc_{s,i,b} denote the density and local speed of sound evaluated from this state, respectively. For an outlet with prescribed pressure poutp_{\mathrm{out}}, the corrected normal velocity and reservoir state are

Δuout=pi,b−poutρi,bcs,i,b,Wbout=(pout,ui,b+Δuoutnb,Ti,b)T.(64)\Delta u_{\mathrm{out}}=\frac{p_{i,b}-p_{\mathrm{out}}}{\rho_{i,b}c_{s,i,b}},\qquad\mathbf{W}_{b}^{\mathrm{out}}=(p_{\mathrm{out}},\boldsymbol{u}_{i,b}+\Delta u_{\mathrm{out}}\boldsymbol{n}_b,T_{i,b})^{\mathrm{T}}. \tag*{(64)}

For an inlet with prescribed pressure pinp_{\mathrm{in}} and temperature TinT_{\mathrm{in}}, the corresponding quantities are

Δuin=pi,b−pinρi,bcs,i,b,Wbin=(pin,ui,b+Δuinnb,Tin)T.(65)\Delta u_{\mathrm{in}}=\frac{p_{i,b}-p_{\mathrm{in}}}{\rho_{i,b}c_{s,i,b}},\qquad\mathbf{W}_{b}^{\mathrm{in}}=(p_{\mathrm{in}},\boldsymbol{u}_{i,b}+\Delta u_{\mathrm{in}}\boldsymbol{n}_b,T_{\mathrm{in}})^{\mathrm{T}}. \tag*{(65)}

With Wb\mathbf{W}_b selected from Eqs. (A.4) and (A.5), the incoming distribution is

fb,kbc=feq(Wb;ξk),ξk∈Ξb−.(66)f_{b,k}^{\mathrm{bc}}=f^{\mathrm{eq}}(\mathbf{W}_b;\boldsymbol{\xi}_k),\qquad\boldsymbol{\xi}_k\in\Xi_b^{-}. \tag*{(66)}

Far-field boundary

For the prescribed state W∞=(p∞,u∞,T∞)T\mathbf{W}_{\infty} = (p_{\infty}, \mathbf{u}_{\infty}, T_{\infty})^{\mathsf{T}}, the incoming distribution is

fb,kbc=feq(W∞;ξk),ξk∈Ξb−.(67)f^{\mathrm{bc}}_{b,k} = f^{\mathrm{eq}}(\mathbf{W}_{\infty}; \boldsymbol{\xi}_k), \qquad\boldsymbol{\xi}_k \in\Xi_b^{-}. \tag*{(67)}

Kinetic-consistent macroscopic boundary implementation

This appendix gives the closed-form evaluation of the Maxwellian half-space moments in Eq. (33) and the boundary-specific expressions for Fb−\mathbf{F}^{-}_{b} in Eq. (37).

Maxwellian half-space moments

Complete the outward unit normal nb\mathbf{n}_b with two orthogonal unit tangents t1,b\mathbf{t}_{1,b} and t2,b\mathbf{t}_{2,b} to form a local orthonormal frame. For a molecular velocity ξ\boldsymbol{\xi}, let

ξn=ξ⋅nb,ξt1=ξ⋅t1,b,ξt2=ξ⋅t2,b.(68)\xi_n = \boldsymbol{\xi} \cdot\mathbf{n}_b, \qquad\xi_{t_1} = \boldsymbol{\xi} \cdot\mathbf{t}_{1,b}, \qquad\xi_{t_2} = \boldsymbol{\xi} \cdot\mathbf{t}_{2,b}. \tag*{(68)}

In this frame, the local representation of the equilibrium half-space moment in Eq. (33) is

Feq,loc±(W;nb)=∫Ξb±ξn[1, ξn, ξt1, ξt2, 12∣ξ∣2]Tfeq(W;ξ) dξ.(69)\mathbf{F}^{\pm}_{\mathrm{eq,loc}}(\mathbf{W}; \mathbf{n}_b) = \int_{\Xi_b^{\pm}} \xi_n \left[1,\ \xi_n,\ \xi_{t_1},\ \xi_{t_2},\ \frac{1}{2} \lvert\boldsymbol{\xi}\rvert^2\right]^{\mathsf{T}} f^{\mathrm{eq}}(\mathbf{W}; \boldsymbol{\xi}) \,\mathrm{d}\boldsymbol{\xi}. \tag*{(69)}

This integral has the closed form

Feq,loc±=ρ[I1±I2±ut1I1±ut2I1±12[I3±+(ut12+ut22+2a2)I1±]].(70)\mathbf{F}^{\pm}_{\mathrm{eq,loc}} = \rho \begin{bmatrix} I_1^{\pm} \\ I_2^{\pm} \\ u_{t_1} I_1^{\pm} \\ u_{t_2} I_1^{\pm} \\ \frac{1}{2}\left[I_3^{\pm} + (u_{t_1}^{2} + u_{t_2}^{2} + 2a^{2})I_1^{\pm}\right] \end{bmatrix}. \tag*{(70)}

Here a=RTa = \sqrt{RT}, un=u⋅nbu_n = \mathbf{u} \cdot\mathbf{n}_b, and utr=u⋅tr,bu_{t_r} = \mathbf{u} \cdot\mathbf{t}_{r,b} for r=1,2r = 1, 2. Define η=un/(2a)\eta= u_n/(\sqrt{2}a) and

C±(η)=12[1±erf⁡(η)],G(η)=exp⁡(−η2)2π.(71)C_{\pm}(\eta) = \frac{1}{2}\left[1 \pm\operatorname{erf}(\eta)\right], \qquad G(\eta) = \frac{\exp(-\eta^{2})}{\sqrt{2\pi}}. \tag*{(71)}

The required normal moments are

I1±=unC±±aG,I_1^{\pm} = u_n C_{\pm} \pm aG,
I2±=(un2+a2)C±±unaG,I_2^{\pm} = (u_n^{2} + a^{2})C_{\pm} \pm u_n aG,
I3±=(un3+3una2)C±±a(un2+2a2)G.(72)I_3^{\pm} = (u_n^{3} + 3u_n a^{2})C_{\pm} \pm a(u_n^{2} + 2a^{2})G. \tag*{(72)}

Let the five-moment rotation from the local frame to the global Cartesian frame be

P=[100000nxt1xt2x00nyt1yt2y00nzt1zt2z000001].(73)P = \begin{bmatrix} 1 & 0 & 0 & 0 & 0 \\ 0 & n_x & t_{1x} & t_{2x} & 0 \\ 0 & n_y & t_{1y} & t_{2y} & 0 \\ 0 & n_z & t_{1z} & t_{2z} & 0 \\ 0 & 0 & 0 & 0 & 1 \end{bmatrix}. \tag*{(73)}

Thus, the global half-space moment defined in Eq. (33) is recovered as

Feq±(W;nb)=PFeq,loc±(W;nb).(74)\mathbf{F}^{\pm}_{\mathrm{eq}}(\mathbf{W}; \mathbf{n}_b) = P\mathbf{F}^{\pm}_{\mathrm{eq,loc}}(\mathbf{W}; \mathbf{n}_b). \tag*{(74)}

Prescribed mass-flow and fully diffuse wall boundaries

For the prescribed mass-flow boundary in Appendix A.1, the target outward mass flux through face bb is −jΓAb-j_{\Gamma}A_b. Given the outgoing boundary flux F~b+\widetilde{F}_b^{+}, the incoming Maxwellian density and flux are

ρ~Γ,b=jΓAb+(F~b+)ρAbRTΓ/(2π),F~b−(Wi,b,F~b+)=Abρ~Γ,bFeq−(WΓ(1);nb).(75)\widetilde{\rho}_{\Gamma,b}=\frac{j_{\Gamma}A_b+(\widetilde{F}_b^{+})_{\rho}}{A_b\sqrt{RT_{\Gamma}/(2\pi)}}, \qquad\widetilde{F}_b^{-}(W_{i,b},\widetilde{F}_b^{+})=A_b\widetilde{\rho}_{\Gamma,b}F_{\mathrm{eq}}^{-}(W_{\Gamma}^{(1)};\boldsymbol{n}_b). \tag*{(75)}

Here the subscript ρ\rho denotes the mass-flux component. By construction, the total boundary flux defined in Eq. (38) satisfies the prescribed facewise mass-flow condition, (Fbm)ρ=−jΓAb(F_b^{m})_{\rho}=-j_{\Gamma}A_b, at every PBC iteration.

For a stationary fully diffuse wall, setting jΓ=0j_{\Gamma}=0 and TΓ=TwT_{\Gamma}=T_w in Eq. (B.8) gives the incoming diffuse-reflection flux and enforces zero normal mass flow.

Pressure inlet and outlet

The pressure-outlet and pressure-inlet reservoir states WboutW_b^{\mathrm{out}} and WbinW_b^{\mathrm{in}} are defined in Eqs. (A.4) and (A.5), respectively. The continuous incoming flux is

F~b−(Wi,b,F~b+)=Ab{Feq−(Wbout;nb),pressure outlet,Feq−(Wbin;nb),pressure inlet.(76)\widetilde{F}_b^{-}(W_{i,b},\widetilde{F}_b^{+})=A_b \begin{cases} F_{\mathrm{eq}}^{-}(W_b^{\mathrm{out}};\boldsymbol{n}_b), & \text{pressure outlet},\\ F_{\mathrm{eq}}^{-}(W_b^{\mathrm{in}};\boldsymbol{n}_b), & \text{pressure inlet}. \end{cases} \tag*{(76)}

Far-field boundary

With the prescribed state W∞W_{\infty} defined in Appendix A.3, the continuous incoming flux is

F~b−(Wi,b,F~b+)=AbFeq−(W∞;nb).(77)\widetilde{F}_b^{-}(W_{i,b},\widetilde{F}_b^{+})=A_bF_{\mathrm{eq}}^{-}(W_{\infty};\boldsymbol{n}_b). \tag*{(77)}

It is independent of the interior state and remains fixed during the PBC iterations. Equations (B.8), (B.9), and (B.10) give the boundary-specific incoming fluxes in Eq. (37). Together with Eq. (B.7), they complete the boundary update in Eqs. (36)–(38).

Analytical Jacobians of the boundary fluxes

Only the current continuous half-space fluxes in Eqs. (36) and (37) are differentiated; the bracketed discrete consistency corrections remain fixed during the PBC iterations. Consistent with the cell-centred boundary linearization in § 3.3.3, Wi,b=WiW_{i,b}=W_i is used in the Jacobian. The full-state form of Eq. (49) is

Jb,i=(I5+∂F~b−(Wi,F~b+)∂F~b+)∂F~b+(Wi)∂Wi+∂F~b−(Wi,F~b+)∂Wi∈R5×5.(78)J_{b,i}=\left(I_5+\frac{\partial\widetilde{F}_b^{-}(W_i,\widetilde{F}_b^{+})}{\partial\widetilde{F}_b^{+}}\right)\frac{\partial\widetilde{F}_b^{+}(W_i)}{\partial W_i}+\frac{\partial\widetilde{F}_b^{-}(W_i,\widetilde{F}_b^{+})}{\partial W_i}\in\mathbb{R}^{5\times5}. \tag*{(78)}

Its upper-left 4×44\times4 block gives the pressure–velocity Jacobian in Eq. (49), and its final column is used in Eq. (55).

Maxwellian half-space Jacobian

Each boundary Jacobian uses the Maxwellian half-space derivative. Let h±\mathbf{h}^{\pm} denote the column vector in Eq. (B.3), so that Feq,loc±=ρh±\mathbf{F}^{\pm}_{\mathrm{eq,loc}}=\rho\mathbf{h}^{\pm}. Differentiating Eqs. (B.3) and (B.5) gives

hun±=[C±2I1±ut1C±ut2C±12[3I2±+(ut12+ut22+2a2)C±]],hut1±=[00I1±0ut1I1±],hut2±=[000I1±ut2I1±],hT±=[I1,T±RC±ut1I1,T±ut2I1,T±12[5RI1±+(ut12+ut22+2a2)I1,T±]](79)\mathbf{h}^{\pm}_{u_n}= \begin{bmatrix} C_{\pm}\\ 2I^{\pm}_1\\ u_{t1}C_{\pm}\\ u_{t2}C_{\pm}\\ \frac{1}{2}\left[3I^{\pm}_2+\left(u_{t1}^{2}+u_{t2}^{2}+2a^{2}\right)C_{\pm}\right] \end{bmatrix}, \qquad \mathbf{h}^{\pm}_{u_{t1}}= \begin{bmatrix} 0\\ 0\\ I^{\pm}_1\\ 0\\ u_{t1}I^{\pm}_1 \end{bmatrix}, \qquad \mathbf{h}^{\pm}_{u_{t2}}= \begin{bmatrix} 0\\ 0\\ 0\\ I^{\pm}_1\\ u_{t2}I^{\pm}_1 \end{bmatrix}, \qquad \mathbf{h}^{\pm}_{T}= \begin{bmatrix} I^{\pm}_{1,T}\\ RC_{\pm}\\ u_{t1}I^{\pm}_{1,T}\\ u_{t2}I^{\pm}_{1,T}\\ \frac{1}{2}\left[5RI^{\pm}_{1}+ \left(u_{t1}^{2}+u_{t2}^{2}+2a^{2}\right)I^{\pm}_{1,T}\right] \end{bmatrix} \tag*{(79)}

where I1,T±=±aG/(2T)I^{\pm}_{1,T}=\pm a\mathcal{G}/(2T). Combining these derivatives with ρ=p/(RT)\rho=p/(RT) and the rotation P\mathbf{P} in Eq. (B.6) gives the half-space Jacobian in global Cartesian coordinates:

Jeq±(W;nb)=Pρ[h±p  hun±  hut1±  hut2±  hT±−h±T]PT.(80)\mathbf{J}^{\pm}_{\mathrm{eq}}(\mathbf{W};\mathbf{n}_b)=\mathbf{P}\rho\left[\frac{\mathbf{h}^{\pm}}{p}\ \ \mathbf{h}^{\pm}_{u_n}\ \ \mathbf{h}^{\pm}_{u_{t1}}\ \ \mathbf{h}^{\pm}_{u_{t2}}\ \ \mathbf{h}^{\pm}_{T}-\frac{\mathbf{h}^{\pm}}{T}\right]\mathbf{P}^{\mathsf{T}}. \tag*{(80)}

It follows from Eq. (34) that ∂F~b+/∂Wi=AbJeq+(Wi;nb)\partial\widetilde{\mathbf{F}}_b^{+}/\partial\mathbf{W}_i=\mathbf{A}_b\mathbf{J}^{+}_{\mathrm{eq}}(\mathbf{W}_i;\mathbf{n}_b).

Prescribed mass-flow and fully diffuse wall boundaries

Let eρT=[1  0  0  0  0]\mathbf{e}_{\rho}^{\mathsf{T}}=[1\ \ 0\ \ 0\ \ 0\ \ 0] select the mass-flux component. Differentiating Eq. (B.8) gives the incoming flux derivatives and the resulting full boundary Jacobian as

∂F~b−∂F~b+=2πRTΓFeq−(WΓ(1);nb)eρT,∂F~b−∂Wi=0,\frac{\partial\widetilde{\mathbf{F}}_b^{-}}{\partial\widetilde{\mathbf{F}}_b^{+}}= \sqrt{\frac{2\pi}{RT_{\Gamma}}}\mathbf{F}_{\mathrm{eq}}^{-}(\mathbf{W}_{\Gamma}^{(1)};\mathbf{n}_b)\mathbf{e}_{\rho}^{\mathsf{T}}, \qquad \frac{\partial\widetilde{\mathbf{F}}_b^{-}}{\partial\mathbf{W}_i}=0,
Jb,i=[I5+2πRTΓFeq−(WΓ(1);nb)eρT]AbJeq+(Wi;nb).(81)\mathbf{J}_{b,i}=\left[\mathbf{I}_5+\sqrt{\frac{2\pi}{RT_{\Gamma}}}\mathbf{F}_{\mathrm{eq}}^{-}(\mathbf{W}_{\Gamma}^{(1)};\mathbf{n}_b)\mathbf{e}_{\rho}^{\mathsf{T}}\right]\mathbf{A}_b\mathbf{J}_{\mathrm{eq}}^{+}(\mathbf{W}_i;\mathbf{n}_b). \tag*{(81)}

The stationary fully diffuse wall has the same Jacobian with TΓ=TwT_{\Gamma}=T_w.

Pressure inlet and outlet

For either pressure boundary, let Δu\Delta u denote the corresponding normal-velocity correction in Eqs. (A.4) and (A.5). The required derivatives are

∂Δu∂pi=1ρics,i−Δupi,∂Δu∂Ti=Δu2Ti,∂Δu∂ui=0.(82)\frac{\partial\Delta u}{\partial p_i}=\frac{1}{\rho_i c_{s,i}}-\frac{\Delta u}{p_i}, \qquad \frac{\partial\Delta u}{\partial T_i}=\frac{\Delta u}{2T_i}, \qquad \frac{\partial\Delta u}{\partial u_i}=0. \tag*{(82)}

Accordingly, the two pressure-reservoir Jacobians are

Jb,outres=∂Wbout∂Wi=[001×30nb∂Δuout∂piI3nb∂Δuout∂Ti001×31],(83)\mathbf{J}^{\mathrm{res}}_{b,\mathrm{out}}=\frac{\partial\mathbf{W}^{\mathrm{out}}_b}{\partial\mathbf{W}_i}= \begin{bmatrix} 0 & \mathbf{0}_{1\times3} & 0\\ n_b\frac{\partial\Delta u_{\mathrm{out}}}{\partial p_i} & \mathbf{I}_3 & n_b\frac{\partial\Delta u_{\mathrm{out}}}{\partial T_i}\\ 0 & \mathbf{0}_{1\times3} & 1 \end{bmatrix}, \tag*{(83)}
Jb,inres=∂Wbin∂Wi=[001×30nb∂Δuin∂piI3nb∂Δuin∂Ti001×30].\mathbf{J}^{\mathrm{res}}_{b,\mathrm{in}}=\frac{\partial\mathbf{W}^{\mathrm{in}}_b}{\partial\mathbf{W}_i}= \begin{bmatrix} 0 & \mathbf{0}_{1\times3} & 0\\ n_b\frac{\partial\Delta u_{\mathrm{in}}}{\partial p_i} & \mathbf{I}_3 & n_b\frac{\partial\Delta u_{\mathrm{in}}}{\partial T_i}\\ 0 & \mathbf{0}_{1\times3} & 0 \end{bmatrix}.

The incoming pressure-boundary flux is independent of Fb+\mathbf{F}_b^{+}, i.e., ∂F~b−/∂Fb+=0\partial\widetilde{\mathbf{F}}_b^{-}/\partial\mathbf{F}_b^{+}=0. Substitution of the corresponding reservoir Jacobian gives

Jb,i=Ab{Jeq+(Wi;nb)+Jeq−(Wbout;nb)Jb,outres,pressure outlet,Jeq+(Wi;nb)+Jeq−(Wbin;nb)Jb,inres,pressure inlet.(84)\mathbf{J}_{b,i}=\mathbf{A}_b \begin{cases} \mathbf{J}_{\mathrm{eq}}^{+}(\mathbf{W}_i;\mathbf{n}_b)+\mathbf{J}_{\mathrm{eq}}^{-}(\mathbf{W}^{\mathrm{out}}_b;\mathbf{n}_b)\mathbf{J}^{\mathrm{res}}_{b,\mathrm{out}}, & \text{pressure outlet},\\ \mathbf{J}_{\mathrm{eq}}^{+}(\mathbf{W}_i;\mathbf{n}_b)+\mathbf{J}_{\mathrm{eq}}^{-}(\mathbf{W}^{\mathrm{in}}_b;\mathbf{n}_b)\mathbf{J}^{\mathrm{res}}_{b,\mathrm{in}}, & \text{pressure inlet}. \end{cases} \tag*{(84)}

Far-field boundary

The far-field incoming flux is independent of both Wi\mathbf{W}_i and Fb+\mathbf{F}_b^{+}. Therefore,

∂F~b−∂Wi=0,∂F~b−∂Fb+=0,Jb,i=AbJeq+(Wi;nb).(85)\frac{\partial\widetilde{\mathbf{F}}_b^{-}}{\partial\mathbf{W}_i}=0,\qquad \frac{\partial\widetilde{\mathbf{F}}_b^{-}}{\partial\mathbf{F}_b^{+}}=0,\qquad \mathbf{J}_{b,i}=\mathbf{A}_b\mathbf{J}_{\mathrm{eq}}^{+}(\mathbf{W}_i;\mathbf{n}_b). \tag*{(85)}

References

  1. [1]S. Teng, M. Hao, J. Liu, X. Bian, Y. Xie, K. Liu, Pollutant inhibition in an extreme ultraviolet lithography machine by dynamic gas lock, Journal of Cleaner Production 430 (2023) 139664. doi:10.1016/j.jclepro.2023.139664.
  2. [2]N. Vasileiadis, G. Tatsios, S. Misdanitis, D. Valougeorgis, Modeling of complex gas distribution systems operating under any vacuum conditions: Simulations of the ITER divertor pumping system, Fusion Engineering and Design 103 (2016) 125–135. doi:10.1016/j.fusengdes.2015.12.033.DOI
  3. [3]H. Alsmeyer, Density profiles in argon and nitrogen shock waves measured by the absorption of an electron beam, Journal of Fluid Mechanics 74 (3) (1976) 497–513. doi:10.1017/S0022112076001912.DOI
  4. [4]T. Ewart, P. Perrier, I. A. Graur, J. G. Méolans, Mass flow rate measurements in a microchannel, from hydrodynamic to near free molecular regimes, Journal of Fluid Mechanics 584 (2007) 337–356. doi:10.1017/S0022112007006374.DOI
  5. [5]C. Tantos, S. Varoutis, V. Hauer, C. Day, P. Innocente, 3D numerical study of neutral gas dynamics in the DTT particle exhaust using the DSMC method, Nuclear Fusion 64 (1) (2024) 016019. doi:10.1088/1741-4326/ad0c80.DOI
  6. [6]L. L. Baker, N. G. Hadjiconstantinou, Variance reduction for Monte Carlo solutions of the Boltzmann equation, Physics of Fluids 17 (5) (2005) 051703. doi:10.1063/1.1899210.DOI
  7. [7]V. Aristov, Direct Methods for Solving the Boltzmann Equation and Study of Nonequilibrium Flows, Springer, 2001.DOI
  8. [8]Jin S., Asymptotic-preserving schemes for multiscale physical problems, Acta Numerica 31 (2022) 415–489.arxiv.org/abs/2112.05920
  9. [9]Q. Sun, I. D. Boyd, G. V. Candler, A hybrid continuum/particle approach for modeling subsonic, rarefied gas flows, Journal of Computational Physics 194 (1) (2004) 256–277.DOI
  10. [10]K. Xu, J.-C. Huang, A unified gas-kinetic scheme for continuum and rarefied flows, Journal of Computational Physics 229 (2010) 7747–7764.
  11. [11]Z. L. Guo, K. Xu, R. J. Wang, Discrete unified gas kinetic scheme for all Knudsen number flows: Low-speed isothermal case, Physical Review E 88 (2013) 033305.pubmed.ncbi.nlm.nih.gov/24125383
  12. [12]C. Liu, Y. J. Zhu, K. Xu, Unified gas-kinetic wave-particle methods I: Continuum and rarefied gas flow, Journal of Computational Physics 401 (2020) 108977.arxiv.org/abs/1811.07141
  13. [13]G. Dimarco, L. Pareschi, Asymptotic preserving implicit-explicit Runge–Kutta methods for nonlinear kinetic equations, SIAM Journal on Numerical Analysis 51 (2013) 1064–1087.arxiv.org/abs/1205.0882
  14. [14]M. H. Gorji, P. Jenny, An efficient particle Fokker–Planck algorithm for rarefied gas flows, Journal of Computational Physics 262 (2014) 325–343.DOI
  15. [15]F. Fei, J. Zhang, J. Li, Z. Liu, A unified stochastic particle Bhatnagar-Gross-Krook method for multiscale gas flows, Journal of Computational Physics 400 (2020) 108972.arxiv.org/abs/1808.03801
  16. [16]W. Su, L. H. Zhu, P. Wang, Y. H. Zhang, L. Wu, Can we find steady-state solutions to multiscale rarefied gas flows within dozens of iterations?, Journal of Computational Physics 407 (2020) 109245.DOI
  17. [17]Y. J. Zhu, C. W. Zhong, K. Xu, Implicit unified gas-kinetic scheme for steady state solutions in all flow regimes, Journal of Computational Physics 315 (2016) 16–38.DOI
  18. [18]W. Su, L. H. Zhu, L. Wu, Fast convergence and asymptotic preserving of the general synthetic iterative scheme, SIAM Journal on Scientific Computing 42 (2020) B1517–B1540.arxiv.org/abs/2003.09958
  19. [19]M. L. Adams, E. W. Larsen, Fast iterative methods for discrete-ordinates particle transport calculations, Progress in Nuclear Energy 40 (2002) 3–159.DOI
  20. [20]D. Valougeorgis, S. Naris, Acceleration schemes of the discrete velocity method: gaseous flows in rectangular microchannels, SIAM J. Sci. Comput. 25 (2) (2003) 534–552.DOI
  21. [21]S. Naris, D. Valougeorgis, F. Sharipov, D. Kalempa, Discrete velocity modelling of gaseous mixture flows in MEMS, Superlattices Microstruct. 35 (3) (2004) 629–643.DOI
  22. [22]J. Lihnaropoulos, S. Naris, D. Valougeorgis, Formulation and stability analysis of rapidly convergent iteration schemes for the 2-d linearized bgk equation, Transp. Theory Stat. Phys. 36 (4–6) (2007) 513–528.DOI
  23. [23]L. Szalmás, D. Valougeorgis, A fast iterative model for discrete velocity calculations on triangular grids, J. Comput. Phys. 229 (11) (2010) 4315–4326.DOI
  24. [24]L. H. Zhu, X. C. Pi, W. Su, Z. H. Li, Y. H. Zhang, L. Wu, General synthetic iterative scheme for nonlinear gas kinetic simulation of multi-scale rarefied gas flows, Journal of Computational Physics 430 (2021) 110091.DOI
  25. [25]Y. Zhang, R. Yuan, L. Luo, L. Wu, An efficient treatment of heat-flux boundary conditions in GSIS for rarefied gas flows, Computers & Fluids 315 (2026) 107113. doi:10.1016/j.compfluid.2026.107113.DOI
  26. [26]C. Liu, K. Xu, Wave–particle decomposition for kinetic equations II: Full Boltzmann equation, arXiv preprintArXiv:2609.35894v1. arXiv:2609.35894.arxiv.org/abs/2609.35894
  27. [27]Y. Zhang, J. Zeng, R. Yuan, W. Liu, Q. Li, L. Wu, Efficient parallel solver for rarefied gas flow using GSIS, Computers & Fluids 281 (2024) 106374.arxiv.org/abs/2310.18916
  28. [28]L. Chacón, G. Chen, D. A. Knoll, C. Newman, H. Park, W. Taitano, J. A. Willert, G. Womeldorff, Multiscale high-order/low-order (holo) algorithms and applications, J. Comput. Phys. 330 (2017) 21–45.DOI
  29. [29]W. T. Taitano, D. A. Knoll, L. Chacón, J. M. Reisner, A. K. Prinja, Moment-based acceleration for neutral gas kinetics with BGK collision operator, J. Comput. Theor. Transp. 43 (2014) 83–108.DOI
  30. [30]J. Zeng, W. Su, L. Wu, General synthetic iterative scheme for unsteady rarefied gas flows, Commun. Comput. Phys. 34 (2023) 173–207.arxiv.org/abs/2104.00370
  31. [31]C. D. Hauck, M. P. Laiu, S. R. Schnake, On high-order/low-order and micro-macro methods for implicit time-stepping of the BGK model, SIAM Journal on Scientific Computing 47 (6) (2025) A3566–A3593.arxiv.org/abs/2410.00678
  32. [32]T. He, W. Gao, Z. Zhang, Y. Liu, L. Zhao, Z. Ma, Rarefied gas simulation in the dynamic gas lock of EUV lithography: A comparative study of DSMC, ESBGK, ESBGK-DSMC, and ESFP method, International Communications in Heat and Mass Transfer 172 (2026) 110219. doi:10.1016/j.icheatmasstransfer.2025.110219.
  33. [33]M. Darwish, I. Sraj, F. Moukal-led, A coupled finite volume solver for the solution of incompressible flows on unstructured grids, Journal of Computational Physics 228 (1) (2009) 180–201. doi:10.1016/j.jcp.2008.08.027.DOI
  34. [34]M. Darwish, F. Moukalled, A fully coupled Navier-Stokes solver for fluid flow at all speeds, Numerical Heat Transfer, Part B: Fundamentals 65 (5) (2014) 410–444. doi: 10.1080/10407790.2013.869102.DOI
  35. [35]E. Turkel, Preconditioned methods for solving the incompressible and low speed compressible equations, Journal of Computational Physics 72 (2) (1987) 277–298. doi:10.1016/0021-9991(87)90084-2.DOI
  36. [36]C.-N. Xiao, F. Denner, B. G. M. van Wachem, Fully-coupled pressure-based finite-volume framework for the simulation of fluid flows at all speeds in complex geometries, Journal of Computational Physics 346 (2017) 91–130. doi:10.1016/j.jcp.2017.06.009.DOI
  37. [37]E. Shakhov, Approximate kinetic equations in rarefied gas theory, Fluid Dynamics 3 (1) (1968) 112–115.DOI
  38. [38]S. Chapman, T. G. Cowling, The mathematical theory of non-uniform gases, Cambridge University Press, 1990.
  39. [39]V. Venkatakrishnan, Convergence to steady state solutions of the Euler equations on unstructured grids with limiters, Journal of Computational Physics 118 (1) (1995) 120–130. doi:10.1006/jcph.1995.1084.DOI
  40. [40]C. M. Rhie, W. L. Chow, Numerical study of the turbulent flow past an airfoil with trailing edge separation, AIAA Journal 21 (11) (1983) 1525–1532. doi:10.2514/3.8284.DOI
  41. [41]F. Denner, F. Evrard, B. G. M. van Wachem, Conservative finite-volume framework and pressure-based algorithm for flows of incompressible, ideal-gas and real-gas fluids at all speeds, Journal of Computational Physics 409 (2020) 109348. doi:10.1016/j.jcp.2020.109348.DOI
  42. [42]A. Syrakos, S. Varchanis, Y. Dimakopoulos, A. Goulas, J. Tsamopoulos, A critical analysis of some popular methods for the discretisation of the gradient operator in finite volume methods, Physics of Fluids 29 (12) (2017) 127103. doi:10.1063/1.4997682.DOI
  43. [43]S. R. Mathur, J. Y. Murthy, A pressure-based method for unstructured meshes, Numerical Heat Transfer, Part B: Fundamentals 31 (2) (1997) 195–215. doi:10.1080/10407799708915105.DOI
  44. [44]L. Mangani, M. Darwish, F. Moukalled, Development of a pressure-based coupled CFD solver for turbulent and compressible flows in turbomachinery applications, in: Proceedings of ASME Turbo Expo 2014, ASME, 2014, p. V02BT39A019, paper GT2014-25967. doi:10.1115/GT2014-25967.DOI
  45. [45]A. Miettinen, T. Siikonen, Application of pressure- and density-based methods for different flow speeds, International Journal for Numerical Methods in Fluids 79 (5) (2015) 243–267. doi:10.1002/fld.4051.DOI
  46. [46]C. T. Kelley, D. E. Keyes, Convergence analysis of pseudo-transient continuation, SIAM Journal on Numerical Analysis 35 (2) (1998) 508–523. doi:10.1137/S0036142996304796.DOI
  47. [47]S. J. Plimpton, S. Moore, A. Borner, A. Stagg, T. Koehler, J. Torczynski, M. Gallis, Direct simulation Monte Carlo on petaflop supercomputers and beyond, Physics of Fluids 31 (8) (2019) 086101. doi:10.1063/1.5108534.DOI
  48. [48]S. Varoutis, D. Valougeorgis, O. Sazhin, F. Sharipov, Rarefied gas flow through short tubes into vacuum, Journal of Vacuum Science & Technology A 26 (2) (2008) 228–238. doi:10.1116/1.2830639.DOI
  49. [49]W. Li, Y. Zhang, J. Zeng, L. Wu, Multiscale simulation of rarefied gas flows in simplified divertor tokamak test facility particle exhaust, Commun. Comput. Phys. 39 (2026) 1536–1558.DOI
  50. [50]R. Yuan, L. Wu, Capturing the influence of intermolecular potential in rarefied gas flows by a kinetic model with velocity-dependent collision frequency, Journal of Fluid Mechanics 942 (2022) A13. doi:10.1017/jfm.2022.350.DOI
  51. [51]Y. Wen, L. Wu, A general synthetic iterative solver for axisymmetric rarefied gas and electrostatic charged-particle flows, arXiv:2607.21277.arxiv.org/abs/2607.21277
  52. [52]Y. Zhang, R. Yuan, L. Wu, A fast-converging and asymptotic-preserving adjoint shape optimization of rarefied gas flows, Journal of Computational Physics (2026) 114960.DOI

Paper details

vlnc:W_xksffqyy
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents