betaFeedback

WorksPhysics and Astronomy

Multiple-distribution-function finite-difference lattice Boltzmann method for incompressible Navier-Stokes equation

Abstract

In this paper, a multiple-distribution-function finite-difference lattice Boltzmann method (MDF-FDLBM) is proposed for the convection-diffusion system based incompressible Navier-Stokes equations (NSEs). By Chapman Enskog analysis, the convection-diffusion system based incompressible NSEs can be recovered from MDF-FDLBM. Some quantities, including the velocity gradient, velocity divergence, strain rate tensor, shear stress and vorticity, can be computed locally by the first-order moment of the non-equilibrium distribution function. Through the von Neumann analysis, we conduct the stability analysis for the MDF-FDLBM and incompressible finite-difference lattice Boltzmann method (IFDLBM). It is found that the IFDLBM will be more stable than that of MDF-FDLBM with small kinematic viscosity, and the MDF-FDLBM will be more stable than that of IFDLBM with large Courant-Friedrichs-Lewy condition number. Finally, some simulations are conducted to validate the MDF-FDLBM. The results agree well with the analytical solutions and previous results. Through the numerical testing, we find that the MDF-FDLBM has a second-order convergence rate in space and time. The MDF-FDLBMcombined with non-uniform grid also works well. Meanwhile, compared with IFDLBM, it can be found that MDF-FDLBM offers higher accuracy and computational efficiency, reducing computation time by more than 36%.

Introduction

As a mesoscopic numerical method, the lattice Boltzmann method (LBM) has received great attention in the past 30 years. It has a great performance in simulating complex fluid flows, including the thermal flows [1, 2], microscale flows [3, 4, 5], porous media flows [6, 7, 8], turbulent flows [9, 10], multiphase flows [11, 12, 13, 14, 15, 16], multicomponent flow and other fluid flows. And the finite difference lattice Boltzmann method (FDLBM), which combining LBM with finite-difference scheme, also has attracted much attention. Compared with LBM, FDLBM can promote the geometrical flexibility, due to the discrete-velocity being decoupled with lattice and time steps. This feature makes non-uniform meshes can be used to FDLBM to improve computational efficiency.

In 1995, a finite-difference lattice Boltzmann equation for the simulation of the incompressible Navier-Stokes equations was proposed by Reider and Sterling [17]. Subsequently, by using body-fitted coordinates with non-uniform grids, Mei and Shyy explored a FDLBM in curvilinear coordinates [18]. Based on above mentioned works, Guo et al. [19] proposed a FDLBM with a mixed difference scheme dealing with the advection term. Then, Wang et al. [20] extend the FDLBM to solve nonlinear convection-diffusion equations (NCDEs). Guo et al. [21] developed a new FDLBM by using the second-order Lax Wendroff scheme and first-order Euler’s formula to discrete space and time derivatives, respectively. Wang et al. [22] explored a high-order FDLBM with a third-order implicit-explicit Runge-Kutta scheme and a fifth-order weighted essentially non-oscillatory scheme for time and space discretization. Chen et.al. [23] develop second-order FDLBM for the NSEs with a mixed difference scheme for space discretization. Then the second-order FDLBM has been combined with the multiple-relaxation-time model to solve the NCDEs [24]. In addition, other high-order accuracy FDLBMs are also developed [25, 26], and these models have been used to simulate complex flow problems, including three-dimensional incompressible flows [27], two-phase liquid-vapor flows [28], natural convection in some special geometries [29, 30], and blood flow [31]. The immersed boundary method was also incorporated into the FDLBM to simulate the incompressible flows [32]. Besides, there are many scholars studied the stability of FDLBMs [24, 33, 34, 35]. El-Amin et.al. used the Fourier expansion to analyze a new explicit FDLBM [34]. As a common stability analysis method, the Von Neumann method has been used to study the stability of three-level finite-difference-based lattice Boltzmann schemes [33], the FDLBM with upwind differences schemes [35] and MRT-FDLBM [24].

For incompressible NSEs, there are two kinds of FDLBM. One is the standard FDLBMs [19, 22], where the equilibrium distribution function is defined by density and velocity. The other is incompressible FDLBM [36], where the equilibrium distribution function is defined by pressure and velocity. Recently, Chai et el. [37] developed a multiple-distribution-function LBM (MDF-LBM) for convection-diffusion-system based on incompressible NSEs. The authors converted incompressible NSE into a convection-diffusion-system. Instead of solving incompressible NSEs directly, MDF-LBM is used to solve convection-diffusion system. Based on this work, we can also introduce this idea to the FDLBM, and develop a MDF-FDLBM with a second-order convergence rate in space and time for the incompressible NSEs. The MDF-FDLBM has some advantages as follows.

  • Compared with the MDF-LBM, MDF-FDLBM can adopt non-uniform grid simulation to improve the computational efficiency of the model.

  • Compared to IFDLBM and standard FDLBM, one can use fewer discrete velocities to construct the MDF-FDLBM. Take the two-dimensional problems as an example, the D2Q4 or D2Q5 lattice model is enough for the MDF-FDLBM, while the D2Q9 model needs to be used for the IFDLBM and standard FDLBM.

  • Some physical variables, including the velocity gradient, velocity divergence, strain rate tensor, shear stress and vorticity, can be computed locally by the first-order moment of the non-equilibrium distribution function. But, for the IFDLBM and standard FDLBM, it is difficult to develop a local computation scheme for the velocity gradient and vorticity, and the computation of strain rate tensor need to use the second-order moments of the nonequilibrium distribution function.

The rest of the paper is organized as follows. In Section 2, the incompressible NSEs is transformed into the convection-diffusion-system, and a MDF-FDLBM for the convection-diffusion-system based incompressible NSEs is developed. And then, the stability of IFDLBM and MDF-FDLBM are studied in Section 3. In Section 4, some numerical simulations are conducted. Finally, we summarizes the results and concludes in Section 5.

Physical model and governing equation

The incompressible Navier-Stokes equations

The incompressible fluid flows can be described by the incompressible Navier-Stokes equations, which the NSEs can be written as

∇⋅u=R,∂u∂t+∇⋅(uu)=−∇P+∇⋅(ν∇u)+F,(1)\nabla\cdot\boldsymbol{u} = R, \\ \frac{\partial\boldsymbol{u}}{\partial t} + \nabla\cdot(\boldsymbol{u}\boldsymbol{u}) = -\nabla P + \nabla\cdot(\nu\nabla\boldsymbol{u}) + \boldsymbol{F}, \tag*{(1)}

where RR is the source term, u\boldsymbol{u} is a dd-dimensional velocity vector, PP is the pressure, ν\nu is the kinematic viscosity and F\boldsymbol{F} is the dd-dimensional external force term. As noted in Ref [37], the above NSEs can also be rewritten as a coupled convection-diffusion system,

∂uˉα∂t+∇⋅(uˉαu+PEα)=∇⋅(ν∇uˉα)+Fˉα,α=0,1,…,d(2)\frac{\partial\bar{u}_{\alpha}}{\partial t} + \nabla\cdot(\bar{u}_{\alpha}\boldsymbol{u} + P\boldsymbol{E}_{\alpha}) = \nabla\cdot(\nu\nabla\bar{u}_{\alpha}) + \bar{F}_{\alpha}, \qquad\alpha= 0,1,\ldots,d \tag*{(2)}

where uˉ0=ρ0\bar{u}_0=\rho_0 and uˉα=uα(α=1−d)\bar{u}_\alpha=u_\alpha(\alpha=1-d), Fˉ0=R\bar{F}_0=R and Fˉα=Fα(α=1−d)\bar{F}_\alpha=F_\alpha(\alpha=1-d), E0=0E_0=0 and Eα(α=1−d)E_\alpha(\alpha=1-d) is the unit vector in dd-dimensional space.

It can be noticed that the NSEs can be reformulated as (d+1)(d+1) CDEs. And the NSEs (1) are equivalent to the NCDEs (2). However, the FDLBM is more effective to solve the NCDEs, due to fewer discrete velocity is needed for the NCDEs.

The MDF-FDLBM for the convection-diffusion-system based NSEs

In the FDLBM, there are two kinds of model, including the single-relaxation-time FDLBM and multiple-relaxation-time FDLBM. Here, we use the MRT-FDLBM to ensure the higher stability and accuracy. Based on this work [24], the evolution equation of the MDF-FDLBM for the convection-diffusion-system based NSEs (2) can be deduced

f^j,α+(x,t+Δt)=f^j,α+(x,t)−Δtcj⋅∇fj,α(x,t+12Δt)+Δt(Fj,α+Gj,α)(x,t)+Δt22(∂tFj,α+∂tGj,α)(x,t),(3)\hat{f}^{+}_{j,\alpha}(x,t+\Delta t)=\hat{f}^{+}_{j,\alpha}(x,t)-\Delta t c_j\cdot\nabla f_{j,\alpha}\left(x,t+\frac{1}{2}\Delta t\right)+\Delta t(F_{j,\alpha}+G_{j,\alpha})(x,t)+\frac{\Delta t^2}{2}\left(\partial_tF_{j,\alpha}+\partial_tG_{j,\alpha}\right)(x,t), \tag*{(3)}

where fj,α(x,t)f_{j,\alpha}(x,t) is the distribution function at position xx and time tt along the discrete velocity cjc_j. Δt\Delta t is the time step and ∇\nabla is the spatial gradient operator. The Fj,αF_{j,\alpha} and Gj,αG_{j,\alpha} are the source terms. The f^\hat{f} and the f^+\hat{f}^{+} are the combination of distribution functions and equilibrium distribution functions. They can be defined as

f^=f+Δt2Λ~(f−feq),(4)\hat{f}=f+\frac{\Delta t}{2}\tilde{\Lambda}(f-f^{eq}), \tag*{(4)}

and

f^+=f−Δt2Λ~(f−feq).(5)\hat{f}^{+}=f-\frac{\Delta t}{2}\tilde{\Lambda}(f-f^{eq}). \tag*{(5)}

In Eqs. (4) and (5), Λ~=(Λ~jk)\tilde{\Lambda}=(\tilde{\Lambda}_{jk}) is a q×qq\times q invertible collision matrix, and qq is the number of the discrete velocities. According to the MRT-FDLBM [24], for the collision matrix Λ~\tilde{\Lambda}, some appropriate requirement need to be satisfied:

∑jejΛ~jk=s0ek,∑jcjΛ~jk=S~ck,∀k=1,2,…,q,(6)\sum_j e_j\tilde{\Lambda}_{jk}=s_0e_k,\qquad\sum_j c_j\tilde{\Lambda}_{jk}=\tilde{S}c_k,\quad\forall k=1,2,\ldots,q, \tag*{(6)}

where e=(1,1,…,1)∈Rqe=(1,1,\ldots,1)\in\mathbb{R}^q, S~\tilde{S} is an invertible d×dd\times d relaxation matrix corresponding to the kinematic viscosity ν\nu. The equilibrium distribution function should be defined as

fj,αeq=ωj[uˉα+cj⋅(uˉαu+PEα)cs2].(7)f^{eq}_{j,\alpha}=\omega_j\left[\bar{u}_\alpha+\frac{c_j\cdot(\bar{u}_\alpha\boldsymbol{u}+PE_\alpha)}{c_s^2}\right]. \tag*{(7)}

And the source terms are defined as

Fj,α=ωjFα,(8)F_{j,\alpha}=\omega_jF_\alpha, \tag*{(8)}
Gj,α=ωjcj⋅∂t(uˉαu+PEα)cs2.(9)G_{j,\alpha}=\omega_j\frac{c_j\cdot\partial_t(\bar{u}_\alpha\boldsymbol{u}+PE_\alpha)}{c_s^2}. \tag*{(9)}

In addition, the second evolution equation used to evaluate distribution function fj,α(x,t+12Δt)f_{j,\alpha}(x,t+\frac{1}{2}\Delta t) can be written as

fˉj,α(x,t+h)=fˉj,α+(x−cjh,t)+h[Fj,α+Gˉj,α+h2∂tFj,α+h2∂tGˉj,α](x−cjh,t),(10)\bar{f}_{j,\alpha}(x,t+h)=\bar{f}^{+}_{j,\alpha}(x-c_jh,t)+h\left[F_{j,\alpha}+\bar{G}_{j,\alpha}+\frac{h}{2}\partial_tF_{j,\alpha}+\frac{h}{2}\partial_t\bar{G}_{j,\alpha}\right](x-c_jh,t), \tag*{(10)}

where h=Δt/2h=\Delta t/2, and

fˉj,α=fj,α−h2(−Λ~jk(fk,α−fk,αeq)),(11)\bar{f}_{j,\alpha}=f_{j,\alpha}-\frac{h}{2}\left(-\tilde{\Lambda}_{jk}\left(f_{k,\alpha}-f^{eq}_{k,\alpha}\right)\right), \tag*{(11)}
fˉj,α+=fj,α+h2(−Λ~jk(fk,α−fk,αeq)).(12)\bar{f}_{j,\alpha}^{+} = f_{j,\alpha} + \frac{h}{2}\left(-\tilde{\Lambda}_{jk}\left(f_{k,\alpha} - f_{k,\alpha}^{eq}\right)\right). \tag*{(12)}

The discrete source terms Gˉj,α\bar{G}_{j,\alpha} is given by

Gˉj,α(x,t)=ωjcj⋅(I+h2Λ~)−1∂t(uˉα+PEα)cs2(13)\bar{G}_{j,\alpha}(x,t) = \omega_j \frac{c_j \cdot\left(I+\frac{h}{2}\tilde{\Lambda}\right)^{-1}\partial_t\left(\bar{u}\alpha+PE_{\alpha}\right)}{c_s^2} \tag*{(13)}

Applying Taylor expansion to (10) and ignoring the term O(h2)O(h^2), we have

fˉj,α(x,t+h)=fˉj,α+(x,t)−hcj⋅∇fˉj,α+(x,t)+h[Fj,α(x,t)+Gˉj,α(x,t)+h2∂tFj,α(x,t)+h2∂tGˉj,α(x,t)].(14)\bar{f}_{j,\alpha}(x,t+h) = \bar{f}_{j,\alpha}^{+}(x,t) - hc_j\cdot\nabla\bar{f}_{j,\alpha}^{+}(x,t) + h\left[F_{j,\alpha}(x,t)+\bar{G}_{j,\alpha}(x,t) +\frac{h}{2}\partial_tF_{j,\alpha}(x,t)+\frac{h}{2}\partial_t\bar{G}_{j,\alpha}(x,t)\right]. \tag*{(14)}

The gradient terms ∇fj\nabla f_j in Eq. (3) and ∇fˉj+\nabla\bar{f}_j^{+} in Eq. (13) can be discretized by a mixed difference scheme [19],

∇Πj∗=∂Πj∗∂χα∣m=η∂Πj∗∂χα∣c+(1−η)∂Πj∗∂χα∣u,(15)\nabla\Pi_j^{*}=\left.\frac{\partial\Pi_j^{*}}{\partial\chi_{\alpha}}\right|_{m}=\eta\left.\frac{\partial\Pi_j^{*}}{\partial\chi_{\alpha}}\right|_{c}+(1-\eta)\left.\frac{\partial\Pi_j^{*}}{\partial\chi_{\alpha}}\right|_{u}, \tag*{(15)}

where Πj∗\Pi_j^{*} represents fjf_j or fˉj+\bar{f}_j^{+}, and the parameter η∈[0,1]\eta\in[0,1]. The terms ∂Πj∗∂χα∣u\left.\frac{\partial\Pi_j^{*}}{\partial\chi_{\alpha}}\right|_{u} and ∂Πj∗∂χα∣c\left.\frac{\partial\Pi_j^{*}}{\partial\chi_{\alpha}}\right|_{c} represent second up-wind difference and central-difference schemes, and are given by

∂Πj∗∂χα∣c=Πj∗(χα+Δχα,t)−Πj∗(χα−Δχα,t)2Δχα,(16)\left.\frac{\partial\Pi_j^{*}}{\partial\chi_{\alpha}}\right|_{c}=\frac{\Pi_j^{*}(\chi_{\alpha}+\Delta\chi_{\alpha},t)-\Pi_j^{*}(\chi_{\alpha}-\Delta\chi_{\alpha},t)}{2\Delta\chi_{\alpha}}, \tag*{(16)}
∂Πj∗∂χα∣u={3Πj∗(χα,t)−4Πj∗(χα−Δχα,t)+Πj∗(χα−2Δχα,t)2Δχα,if ciα≥0,−3Πj∗(χα,t)+4Πj∗(χα+Δχα,t)−Πj∗(χα+Δχα,t)2Δχα,if ciα<0.(17)\left.\frac{\partial\Pi_j^{*}}{\partial\chi_{\alpha}}\right|_{u}= \begin{cases} \frac{3\Pi_j^{*}(\chi_{\alpha},t)-4\Pi_j^{*}(\chi_{\alpha}-\Delta\chi_{\alpha},t)+\Pi_j^{*}(\chi_{\alpha}-2\Delta\chi_{\alpha},t)}{2\Delta\chi_{\alpha}}, & \text{if } c_{i\alpha}\geq0,\\ \frac{-3\Pi_j^{*}(\chi_{\alpha},t)+4\Pi_j^{*}(\chi_{\alpha}+\Delta\chi_{\alpha},t)-\Pi_j^{*}(\chi_{\alpha}+\Delta\chi_{\alpha},t)}{2\Delta\chi_{\alpha}}, & \text{if } c_{i\alpha}<0. \end{cases} \tag*{(17)}

In the calculation, the time step Δt\Delta t is given by the CFL condition number,

Δt=CFLΔxc,(18)\Delta t=CFL\frac{\Delta x}{c}, \tag*{(18)}

where Δx\Delta x is the minimum grid scale, and c=∣cj∣c=|c_j|. Courant-Friedrichs-Lewy (CFLCFL) condition number is an important parameter to evaluate the stability and convergence of method. The evolution process of the MDF-FDLBM is shown in Figure 1, and can be listed as follows.

The implementation process of the MDF-FDLBM

Figure 1. The implementation process of the MDF-FDLBM.

Step (1): Calculate f^+(x,t)\hat{f}^{+}(x,t) by (19), and calculate fˉ+(x,t)\bar{f}^{+}(x,t) by (20).

f^+=(I−Λ)f^+Λfeq(19)\hat{f}^{+}=(I-\Lambda)\hat{f}+\Lambda f^{eq} \tag*{(19)}

where Λ=I−(I−Δt2Λ~)(I+Δt2Λ~)−1\Lambda=I-\left(I-\frac{\Delta t}{2}\tilde{\Lambda}\right)\left(I+\frac{\Delta t}{2}\tilde{\Lambda}\right)^{-1}.

fˉ+=(I−3Δt4Λ)fˉ+3Δt4Λfeq=34fˉ++14fˉ.(20)\bar{f}^{+}=\left(I-\frac{3\Delta t}{4}\Lambda\right)\bar{f}+\frac{3\Delta t}{4}\Lambda f^{eq}=\frac{3}{4}\bar{f}^{+}+\frac{1}{4}\bar{f}. \tag*{(20)}

Step (2): Conducting the second evaluation function (14).

Step (3): Evaluate spatial gradient term c⋅∇f(x,t+h)c\cdot\nabla f(x,t+h) by Eqs. (19) and (14).

fˉ(x,t+h)→(19)f(x,t+h)→(14)∇f(x,t+h),\bar{f}(x,t+h)\xrightarrow{(19)}f(x,t+h)\xrightarrow{(14)}\nabla f(x,t+h),
f=(I+h2Λ~)−1(fˉ+h2Λ~feq).(21)f=\left(I+\frac{h}{2}\tilde{\Lambda}\right)^{-1}\left(\bar{f}+\frac{h}{2}\tilde{\Lambda}f^{eq}\right). \tag*{(21)}

Step (4): Compute f^(x,t+Δt)\hat{f}(x,t+\Delta t) by (3).

The Chapman Enskog analysis of the MDF-FDLBM

First, the CE analysis is conducted to the main evolution equation (3) to recover the NCDE. The distribution function fj,αf_{j,\alpha}, the source terms Fj,α,Gj,αF_{j,\alpha},G_{j,\alpha}, the time and space derivatives can be expanded as,

fj,α=fj,α(0)+εfj,α(1)+ε2fj,α(2),∂t=ε∂t1+ε2∂t2,∇=ε∇1,f_{j,\alpha}=f_{j,\alpha}^{(0)}+\varepsilon f_{j,\alpha}^{(1)}+\varepsilon^{2}f_{j,\alpha}^{(2)},\quad\partial_t=\varepsilon\partial_{t_1}+\varepsilon^{2}\partial_{t_2},\quad\nabla=\varepsilon\nabla_1,
G~j,α=εG~j,α(1)+ε2G~j,α(2),Fj,α=εFj,α(1)+ε2Fj,α(2).(22)\widetilde{G}_{j,\alpha}=\varepsilon\widetilde{G}_{j,\alpha}^{(1)}+\varepsilon^{2}\widetilde{G}_{j,\alpha}^{(2)},\quad F_{j,\alpha}=\varepsilon F_{j,\alpha}^{(1)}+\varepsilon^{2}F_{j,\alpha}^{(2)}. \tag*{(22)}
∂tfj,α+Δt2∂t2fj,α+cj⋅∇fj,α+cj⋅∇Δt2∂tfj,α=[−Λ~jk(fk,α−fk,αeq)]+Δt2∂t[−Λ~jk(fk,α−fk,αeq)]+Fj,α+Gj,α+Δt2(∂tFj,α+∂tGj,α).(23)\begin{aligned} \partial_t f_{j,\alpha}+\frac{\Delta t}{2}\partial_t^2f_{j,\alpha}+\boldsymbol{c}_j\cdot\nabla f_{j,\alpha}+\boldsymbol{c}_j\cdot\nabla\frac{\Delta t}{2}\partial_t f_{j,\alpha}=\left[-\widetilde{\Lambda}_{jk}(f_{k,\alpha}-f_{k,\alpha}^{\mathrm{eq}})\right] \\ &+\frac{\Delta t}{2}\partial_t\left[-\widetilde{\Lambda}_{jk}(f_{k,\alpha}-f_{k,\alpha}^{\mathrm{eq}})\right]+F_{j,\alpha}+G_{j,\alpha}+\frac{\Delta t}{2}(\partial_tF_{j,\alpha}+\partial_tG_{j,\alpha}). \tag*{(23)} \end{aligned}

Substituting Eq. (22) into Eq. (23) yields

O(ε0): −Λ~jk(fk(0)−fkeq)=0⟺fj(0)=fjeq(24)O(\varepsilon^0):\ -\widetilde{\Lambda}_{jk}(f_{k}^{(0)}-f_{k}^{\mathrm{eq}})=0\Longleftrightarrow f_j^{(0)}=f_j^{\mathrm{eq}} \tag*{(24)}
O(ε1): ∂t1fj(0)+cj⋅∇1fj(0)=−Λ~jkfk(1)+Gj(1)+Fj(1),(25)O(\varepsilon^1):\ \partial_{t_1}f_j^{(0)}+\boldsymbol{c}_j\cdot\nabla_1f_j^{(0)}=-\widetilde{\Lambda}_{jk}f_k^{(1)}+G_j^{(1)}+F_j^{(1)}, \tag*{(25)}
O(ε2): ∂t2fj(0)+∂t1fj(1)+Δt2∂t12fj(0)+cj⋅∇1fj(1)+cj⋅∇1Δt2∂t1fj(0)=−Λ~jkfk(2)+Δt2∂t1(−Λ~jkfk(1))+Gj(2)+Fj(2)+Δt2(∂t1Fj(1)+∂t1Gj(1)).(26)\begin{aligned} O(\varepsilon^2):\ &\partial_{t_2}f_j^{(0)}+\partial_{t_1}f_j^{(1)}+\frac{\Delta t}{2}\partial_{t_1}^{2}f_j^{(0)}+\boldsymbol{c}_j\cdot\nabla_1f_j^{(1)}+\boldsymbol{c}_j\cdot\nabla_1\frac{\Delta t}{2}\partial_{t_1}f_j^{(0)} \\ ={}&-\widetilde{\Lambda}_{jk}f_k^{(2)}+\frac{\Delta t}{2}\partial_{t_1}(-\widetilde{\Lambda}_{jk}f_k^{(1)})+G_j^{(2)}+F_j^{(2)}+\frac{\Delta t}{2}(\partial_{t_1}F_j^{(1)}+\partial_{t_1}G_j^{(1)}). \tag*{(26)} \end{aligned}

According to Eqs. (7), (8) and (9), one can determine the moment conditions of fi,αeqf_{i,\alpha}^{\mathrm{eq}}, Gi,αG_{i,\alpha} and Fi,αF_{i,\alpha},

∑ifi,αeq=uˉα,∑icifi,αeq=uˉαu+PEα,∑icicifi,αeq=uˉαcs2I(27)\sum_i f_{i,\alpha}^{\mathrm{eq}}=\bar{u}_{\alpha},\qquad\sum_i\boldsymbol{c}_i f_{i,\alpha}^{\mathrm{eq}}=\bar{u}_{\alpha}\boldsymbol{u}+PE_{\alpha},\qquad\sum_i\boldsymbol{c}_i\boldsymbol{c}_i f_{i,\alpha}^{\mathrm{eq}}=\bar{u}_{\alpha}c_s^2\boldsymbol{I} \tag*{(27)}
∑iGi,α=0,∑iciGi,α=∂t(uˉαu+PEα),(28)\sum_i G_{i,\alpha}=0,\qquad\sum_i\boldsymbol{c}_iG_{i,\alpha}=\partial_t(\bar{u}_{\alpha}\boldsymbol{u}+PE_{\alpha}), \tag*{(28)}
∑iFi,α=Fˉα,∑iciFi,α=0.(29)\sum_i F_{i,\alpha}=\bar{F}_{\alpha},\qquad\sum_i\boldsymbol{c}_iF_{i,\alpha}=0. \tag*{(29)}

Based on the moment conditions, we can get the following equations through summing Eqs. (22b) and (22c) over jj,

∂t1uˉα+∇1⋅(uˉαu+PEα)=Fα(1),(30)\partial_{t_1}\bar{u}_{\alpha}+\nabla_1\cdot(\bar{u}_{\alpha}\boldsymbol{u}+PE_{\alpha})=F_{\alpha}^{(1)}, \tag*{(30)}
∂t2uˉα+Δt2∂t12uˉα+∇1⋅∑jcjfj(1)+Δt2∇1⋅∂t1(uˉαu+PEα)=Fα(2)+Δt2∂t1Fα(1).(31)\partial_{t_2}\bar{u}_{\alpha}+\frac{\Delta t}{2}\partial_{t_1}^{2}\bar{u}_{\alpha}+\nabla_1\cdot\sum_j\boldsymbol{c}_jf_j^{(1)}+\frac{\Delta t}{2}\nabla_1\cdot\partial_{t_1}(\bar{u}_{\alpha}\boldsymbol{u}+PE_{\alpha})=F_{\alpha}^{(2)}+\frac{\Delta t}{2}\partial_{t_1}F_{\alpha}^{(1)}. \tag*{(31)}

According to Eq. (26a), one can obtain

∂t12uˉα=∂t1Fα(1)−∂t1∇1⋅(uˉαu+PEα).(32)\partial_{t_1}^{2}\bar{u}_{\alpha}=\partial_{t_1}F_{\alpha}^{(1)}-\partial_{t_1}\nabla_1\cdot(\bar{u}_{\alpha}\boldsymbol{u}+PE_{\alpha}). \tag*{(32)}

If the derivatives of time and space are interchangeable, Eq. (26b) can be rewritten as

∂t2uˉα=Fα(2)−∇1⋅∑jcjfj(1).(33)\partial_{t_2}\bar{u}_\alpha=F_\alpha^{(2)}-\nabla_1\cdot\sum_j c_jf_j^{(1)}. \tag*{(33)}

With the aid of Eq. (6), one can get the following equation through multiplying the Eq. (22b) by cjc_j and summing it over jj,

∂t1(uˉαu+PEα)+∇1⋅cs2uˉαI=−S~∑jcjfj(1)+M1,G(1),(34)\partial_{t_1}(\bar{u}_\alpha u+PE_\alpha)+\nabla_1\cdot c_s^2\bar{u}_\alpha I=-\widetilde{S}\sum_jc_jf_j^{(1)}+M_{1,G}^{(1)}, \tag*{(34)}

where M1,G(1)=∑cjGj(1)M_{1,G}^{(1)}=\sum c_jG_j^{(1)}, and S~\widetilde{S} is a D-dimensional diagonal matrix, which ia related to the diffusion coefficient ν\nu. The relationship are S~=sI\widetilde{S}=sI and cs2s=ν\frac{c_s^2}{s}=\nu. From Eq. (34), one can obtain

∑cjfj(1)=−S~−1[∂t1(uˉαu+PEα)+∇1⋅cs2uˉαI−M1,G(1)],(35)\sum c_jf_j^{(1)}=-\widetilde{S}^{-1}\left[\partial_{t_1}(\bar{u}_\alpha u+PE_\alpha)+\nabla_1\cdot c_s^2\bar{u}_\alpha I-M_{1,G}^{(1)}\right], \tag*{(35)}

and

∂t2uˉα=Fα(2)+∇1⋅cs2S~−1∇1⋅uαI+∇1⋅S~−1[∂t1(uˉαu+PEα)−M1,G(1)].(36)\partial_{t_2}\bar{u}_\alpha=F_\alpha^{(2)}+\nabla_1\cdot c_s^2\widetilde{S}^{-1}\nabla_1\cdot u_\alpha I+\nabla_1\cdot\widetilde{S}^{-1}\left[\partial_{t_1}(\bar{u}_\alpha u+PE_\alpha)-M_{1,G}^{(1)}\right]. \tag*{(36)}

Because

M1,G=∂t1(uˉαu+PEα),(37)M_{1,G}=\partial_{t_1}(\bar{u}_\alpha u+PE_\alpha), \tag*{(37)}

if νI=cs2S~−1\nu I=c_s^2\widetilde{S}^{-1} or ν=cs2/s\nu=c_s^2/s, with the help of Eq. (36), one can deduce

∂t2uˉα=Fα(2)+∇1⋅ν∇1⋅uαI.(38)\partial_{t_2}\bar{u}_\alpha=F_\alpha^{(2)}+\nabla_1\cdot\nu\nabla_1\cdot u_\alpha I. \tag*{(38)}

Through combining the results at ε\varepsilon and ε2\varepsilon^2 scales, i.e., Eqs. (26a) and (38), we can recover the convection-diffusion-system based Navier-Stokes-equations (2) correctly.

Next, we will recover the convection-diffusion-system based NSEs (2) from the second evolution equation (10) with the CE analysis. With the Taylor expansion, Eq. (10) can be expressed as

Djfj,α+h2Dj2fj,α=−Λ~jkfk,αneq+h2(−Λ~jkDkfk,αneq)+Gˉj,α+Fj,α+h2(∂tFj,α+∂tGˉj,α).(39)D_jf_{j,\alpha}+\frac{h}{2}D_j^2f_{j,\alpha}=-\widetilde{\Lambda}_{jk}f_{k,\alpha}^{\mathrm{neq}}+\frac{h}{2}(-\widetilde{\Lambda}_{jk}D_kf_{k,\alpha}^{\mathrm{neq}})+\bar{G}_{j,\alpha}+F_{j,\alpha}+\frac{h}{2}(\partial_tF_{j,\alpha}+\partial_t\bar{G}_{j,\alpha}). \tag*{(39)}

Then we can expand Eq. (39) at different orders of ε\varepsilon,

O(ε0): −Λ~jk(fk,α(0)−fk,αeq)=0, ⇔ fj,α(0)=fj,αeq,(40)O(\varepsilon^0):\ -\widetilde{\Lambda}_{jk}(f_{k,\alpha}^{(0)}-f_{k,\alpha}^{\mathrm{eq}})=0,\ \Leftrightarrow\ f_{j,\alpha}^{(0)}=f_{j,\alpha}^{\mathrm{eq}}, \tag*{(40)}
O(ε1): D1jfj,α(0)=−Λ~jkfk,α(1)+Gˉj,α(1)+Fj,α(1),(41)O(\varepsilon^1):\ D_{1j}f_{j,\alpha}^{(0)}=-\widetilde{\Lambda}_{jk}f_{k,\alpha}^{(1)}+\bar{G}_{j,\alpha}^{(1)}+F_{j,\alpha}^{(1)}, \tag*{(41)}
O(ε2): ∂t2fj,α(0)+D1jfj,α(1)+h2D1j2fj,α(0)=−Λ~jkfk,α(2)+h2(−Λ~jkD1kfk,α(1))+Gˉj,α(2)+Fj,α(2)+h2(∂t1Fj,α(1)+∂t1Gˉj,α(1)).(42)\begin{aligned} O(\varepsilon^2):\ \partial_{t_2}f_{j,\alpha}^{(0)}+D_{1j}f_{j,\alpha}^{(1)}+\frac{h}{2}D_{1j}^2f_{j,\alpha}^{(0)} ={}&-\widetilde{\Lambda}_{jk}f_{k,\alpha}^{(2)}+\frac{h}{2}(-\widetilde{\Lambda}_{jk}D_{1k}f_{k,\alpha}^{(1)})+\bar{G}_{j,\alpha}^{(2)}+F_{j,\alpha}^{(2)}\\ &+\frac{h}{2}(\partial_{t_1}F_{j,\alpha}^{(1)}+\partial_{t_1}\bar{G}_{j,\alpha}^{(1)}). \tag*{(42)} \end{aligned}

With the help of Eq. (35b), Eq. (35c) can be rewritten as

∂t2fj,α(0)+D1jfj,α(1)+h2(∇1⋅cjFj,α(1)+∇1⋅cjGˉj,α(1))=−Λ~jkfk,α(2)+Gˉj,α(2)+Fj,α(2).(43)\partial_{t_2}f_{j,\alpha}^{(0)}+D_{1j}f_{j,\alpha}^{(1)}+\frac{h}{2}(\nabla_1\cdot c_jF_{j,\alpha}^{(1)}+\nabla_1\cdot c_j\bar{G}_{j,\alpha}^{(1)})=-\widetilde{\Lambda}_{jk}f_{k,\alpha}^{(2)}+\bar{G}_{j,\alpha}^{(2)}+F_{j,\alpha}^{(2)}. \tag*{(43)}

Summing Eqs. (43) and (35b) over jj, one can obtain

∂t1uˉα+∇1⋅(uˉαu+PEα)=Fα(1),(44)\partial_{t_1}\bar{u}_\alpha+\nabla_1\cdot(\bar{u}_\alpha u+PE_\alpha)=F_\alpha^{(1)}, \tag*{(44)}
∂t2uˉα+∇1⋅∑jcjfj,α(1)=Fα(2)−h2∇1⋅M1,Gˉ(1)(45)\partial_{t_2}\bar{u}_{\alpha}+\nabla_1\cdot\sum_j\boldsymbol{c}_j f_{j,\alpha}^{(1)}=F_{\alpha}^{(2)}-\frac{h}{2}\nabla_1\cdot\boldsymbol{M}_{1,\bar{G}}^{(1)} \tag*{(45)}

where M1,Gˉ(1)\boldsymbol{M}_{1,\bar{G}}^{(1)} is the first-order moment condition of source term Gˉ\bar{G}. Multiplying the Eq. (35b) by cj\boldsymbol{c}_j and summing it over jj, we can derive

∑jcjfj,α(1)=−S~−1[∂t1(uˉαu+PEα)+∇1⋅uˉαcs2I−M1,Gˉ(1)](46)\sum_j\boldsymbol{c}_j f_{j,\alpha}^{(1)}=-\tilde{\boldsymbol{S}}^{-1}\left[\partial_{t_1}(\bar{u}_{\alpha}\boldsymbol{u}+P E_{\alpha})+\nabla_1\cdot\bar{u}_{\alpha}c_s^2\boldsymbol{I}-\boldsymbol{M}_{1,\bar{G}}^{(1)}\right] \tag*{(46)}

where Eq. (6) has been used. Substituting above equation and M1,Gˉ(1)=(I+h2S~)−1∂t1(uˉαu+PEα)\boldsymbol{M}_{1,\bar{G}}^{(1)}=\left(\boldsymbol{I}+\frac{h}{2}\tilde{\boldsymbol{S}}\right)^{-1}\partial_{t_1}(\bar{u}_{\alpha}\boldsymbol{u}+P E_{\alpha}) into Eq. (37b), we have

∂t2uˉα=Fα(2)+∇1⋅ν∇1⋅uˉαI(47)\partial_{t_2}\bar{u}_{\alpha}=F_{\alpha}^{(2)}+\nabla_1\cdot\nu\nabla_1\cdot\bar{u}_{\alpha}\boldsymbol{I} \tag*{(47)}

where νI=S~−1cs2\nu\boldsymbol{I}=\tilde{\boldsymbol{S}}^{-1}c_s^2 and ν=cs2/s\nu=c_s^2/s. According to the results at ε\varepsilon and ε2\varepsilon^2 scales, i.e., Eqs. (37a) and (47), Eq. (2) can be recovered exactly.

Then we will discuss the calculation the pressure PP. According to Eq. (32), one can obtain

∑ici,αfi,αne=∑ici,αfi,α−∑ici,αfi,αeq=∑ici,αfi,α−(uˉαuα+P)=−cs2s∇αuˉα,(48)\begin{aligned} \sum_i c_{i,\alpha}f_{i,\alpha}^{ne}=\sum_i c_{i,\alpha}f_{i,\alpha}-\sum_i c_{i,\alpha}f_{i,\alpha}^{eq} \\ &=\sum_i c_{i,\alpha}f_{i,\alpha}-(\bar{u}_{\alpha}u_{\alpha}+P) \\ &=-\frac{c_s^2}{s}\nabla_{\alpha}\bar{u}_{\alpha}, \tag*{(48)} \end{aligned}

if we sum Eq. (48) over α\alpha and adopt the continuity equation in Eq. (1), one can obtain

∑α=1d∑ici,αfi,α−(∣u∣2+dP)=−cs2sR,(49)\sum_{\alpha=1}^{d}\sum_i c_{i,\alpha}f_{i,\alpha}-(|\boldsymbol{u}|^2+dP)=-\frac{c_s^2}{s}R, \tag*{(49)}

From Eq. (49), we can get

P=1d(∑α=1d∑ici,αfi,α+cs2s1R−∣u∣2).(50)P=\frac{1}{d}\left(\sum_{\alpha=1}^{d}\sum_i c_{i,\alpha}f_{i,\alpha}+\frac{c_s^2}{s_1}R-|\boldsymbol{u}|^2\right). \tag*{(50)}

Remark 1: In the above work, if Eq. (49) is ensured, the continuity equation (1) must be satisfied. It indicated that continuity equation (1) can not be considered and the evolution equations (3) and (10) with α=0\alpha=0 can be ignored.

Remark 2: Because the source terms Gj,αG_{j,\alpha} and Gˉj,α\bar{G}_{j,\alpha} are on the order of O(Ma2)O(\mathrm{Ma}^2), and the coefficient of the Gj,αG_{j,\alpha} and Gˉj,α\bar{G}_{j,\alpha} the is Δt\Delta t and hh in the two evolution equations. Thus the source terms ΔtG\Delta tG and hGhG are on the order of O(δtMa2)O(\delta t\mathrm{Ma}^2), which can be neglected in the computation. In Eq. (10), it can be found that the terms h22∂tFˉj,α\frac{h^2}{2}\partial_t\bar{F}_{j,\alpha} and h22∂tGˉj,α\frac{h^2}{2}\partial_t\bar{G}_{j,\alpha} are on the order of O(δt2)O(\delta t^2), thus, they can be omitted. In this case, the evolution equations (3) and (10) can be rewritten as

f^j,α(x,t+Δt)=f^j,α+(x,t)−Δtcj⋅∇fj,α(x,t+12Δt)+ΔtFj,α(x,t)+Δt22∂tFj,α(x,t),(51)\hat{f}_{j,\alpha}(\boldsymbol{x},t+\Delta t)=\hat{f}_{j,\alpha}^{+}(\boldsymbol{x},t)-\Delta t\boldsymbol{c}_j\cdot\nabla f_{j,\alpha}\left(\boldsymbol{x},t+\frac{1}{2}\Delta t\right) +\Delta tF_{j,\alpha}(\boldsymbol{x},t)+\frac{\Delta t^2}{2}\partial_tF_{j,\alpha}(\boldsymbol{x},t), \tag*{(51)}

and

fˉj,α(x,t+h)=fˉj,α+(x−cjh,t)+hFj,α.(52)\bar{f}_{j,\alpha}(\boldsymbol{x},t+h)=\bar{f}_{j,\alpha}^{+}(\boldsymbol{x}-\boldsymbol{c}_jh,t)+hF_{j,\alpha}. \tag*{(52)}

The computation of some physical quantities

For MDF-FDLBM, some quantities can be directly computed by the first-order moment of the non-equilibrium distribution function. According to Eq. (48), the velocity gradient can be obtained

∇βuα=−s1cs2∑ici,βfi,αne,(α≠0).(53)\nabla_{\beta}u_{\alpha}=-\frac{s_1}{c_s^2}\sum_i c_{i,\beta}f_{i,\alpha}^{ne},\qquad(\alpha\ne0). \tag*{(53)}

The velocity divergence ∇⋅u\nabla\cdot\boldsymbol{u}, strain rate tensor S\boldsymbol{S}, shear tress σ\boldsymbol{\sigma} and the antisymmetric velocity gradient tensor Ω\boldsymbol{\Omega} can be derived

∇⋅u=∑α=1d∇αuα=−s1cs2∑α=1d∑ici,αfi,αne,(54)\nabla\cdot\boldsymbol{u}=\sum_{\alpha=1}^{d}\nabla_{\alpha}u_{\alpha}=-\frac{s_{1}}{c_{s}^{2}}\sum_{\alpha=1}^{d}\sum_{i}c_{i,\alpha}f_{i,\alpha}^{\mathrm{ne}}, \tag*{(54)}
Sαβ=12(∇βuα+∇αuβ)=−s12cs2∑i(ci,βfi,αne+ci,αfi,βne),(α≠0,β≠0),(55)S_{\alpha\beta}=\frac{1}{2}\left(\nabla_{\beta}u_{\alpha}+\nabla_{\alpha}u_{\beta}\right)=-\frac{s_{1}}{2c_{s}^{2}}\sum_{i}\left(c_{i,\beta}f_{i,\alpha}^{\mathrm{ne}}+c_{i,\alpha}f_{i,\beta}^{\mathrm{ne}}\right),\quad(\alpha\ne0,\beta\ne0), \tag*{(55)}
σαβ=2ρ0νSαβ=−ρ0∑i(ci,βfi,αne+ci,αfi,βne),(α≠0,β≠0),(56)\sigma_{\alpha\beta}=2\rho_{0}\nu S_{\alpha\beta}=-\rho_{0}\sum_{i}\left(c_{i,\beta}f_{i,\alpha}^{\mathrm{ne}}+c_{i,\alpha}f_{i,\beta}^{\mathrm{ne}}\right),\quad(\alpha\ne0,\beta\ne0), \tag*{(56)}
Ωαβ=12(∇βuα−∇αuβ)=−s12cs2∑i(ci,βfi,αne−ci,αfi,βne),(α≠0,β≠0).(57)\Omega_{\alpha\beta}=\frac{1}{2}\left(\nabla_{\beta}u_{\alpha}-\nabla_{\alpha}u_{\beta}\right)=-\frac{s_{1}}{2c_{s}^{2}}\sum_{i}\left(c_{i,\beta}f_{i,\alpha}^{\mathrm{ne}}-c_{i,\alpha}f_{i,\beta}^{\mathrm{ne}}\right),\quad(\alpha\ne0,\beta\ne0). \tag*{(57)}

Stability analysis of the MDF-FDLBM

In this section, we will discuss the stability of the MDF-FDLBM. To compare the stability of the MDF-FDLBM and IFDLBM [36], the linear stability analysis is also implemented to the IFDLBM. And the stability region of MDF-FDLBM and IFDLBM are compared under different conditions.

The stability analysis of the MDF-FDLBM

As a common tool to evaluate the numerical stability of LBMs, the Von Neumann method will be used for the MDF-FDLBM. For simplicity, the source terms will be neglected. There are d+1d+1 evolution equations in MDF-FDLBM. However, as it mentioned in Remark 2, only the evolution equations (α=1,2,…,d\alpha=1,2,\ldots,d) affects the stability. Because the evolution equation of α=0\alpha=0 do not influence the calculation of PP and u\boldsymbol{u}. Therefore, only the evolution equations of α=1,2,…,d\alpha=1,2,\ldots,d need to be analyzed. For brevity, we consider the D2Q5 model for example,

cj=(010−100010−1)c,(58)\boldsymbol{c}_{j}=\begin{pmatrix}0&1&0&-1&0\\0&0&1&0&-1\end{pmatrix}\boldsymbol{c}, \tag*{(58)}
wj=0−5=1/5.(59)w_{j=0-5}=1/5. \tag*{(59)}

Starting from the following two evolution equations,

fj(x,t+Δt)+0.5ωjk[fk−fkeq](x,t+Δt)=fj(x,t)−0.5ωjk[fk(x,t)−fkeq(x,t)]−Δtcj⋅∇fj(x,t+Δt2),(60)f_{j}(\boldsymbol{x},t+\Delta t)+0.5\omega_{jk}\left[f_{k}-f_{k}^{\mathrm{eq}}\right](\boldsymbol{x},t+\Delta t)=f_{j}(\boldsymbol{x},t)-0.5\omega_{jk}\left[f_{k}(\boldsymbol{x},t)-f_{k}^{\mathrm{eq}}(\boldsymbol{x},t)\right]-\Delta t\boldsymbol{c}_{j}\cdot\nabla f_{j}\left(\boldsymbol{x},t+\frac{\Delta t}{2}\right), \tag*{(60)}
gj(x,t+Δt)+0.5ωjk[gk−gkeq](x,t+Δt)=gj(x,t)−0.5ωjk[gk(x,t)−gkeq(x,t)]−Δtcj⋅∇gj(x,t+Δt2),(61)g_{j}(\boldsymbol{x},t+\Delta t)+0.5\omega_{jk}\left[g_{k}-g_{k}^{\mathrm{eq}}\right](\boldsymbol{x},t+\Delta t)=g_{j}(\boldsymbol{x},t)-0.5\omega_{jk}\left[g_{k}(\boldsymbol{x},t)-g_{k}^{\mathrm{eq}}(\boldsymbol{x},t)\right]-\Delta t\boldsymbol{c}_{j}\cdot\nabla g_{j}\left(\boldsymbol{x},t+\frac{\Delta t}{2}\right), \tag*{(61)}

where the ωjk=ΔtΛ~jk\omega_{jk}=\Delta t\widetilde{\Lambda}_{jk}. It is related to the relaxation time matrix S=(s0,s1,s2,s3,s4)\boldsymbol{S}=(s_{0},s_{1},s_{2},s_{3},s_{4}). Then we introduce the following linear expansion,

fj(x,t)=fjeq‾(x,t)+fj′(x,t),gj(x,t)=gjeq‾(x,t)+gj′(x,t),(62)f_{j}(\boldsymbol{x},t)=\overline{f_{j}^{\mathrm{eq}}}(\boldsymbol{x},t)+f_{j}^{\prime}(\boldsymbol{x},t),\qquad g_{j}(\boldsymbol{x},t)=\overline{g_{j}^{\mathrm{eq}}}(\boldsymbol{x},t)+g_{j}^{\prime}(\boldsymbol{x},t), \tag*{(62)}

where gj′(x,t)g_{j}^{\prime}(\boldsymbol{x},t) and fj′(x,t)f_{j}^{\prime}(\boldsymbol{x},t) are the fluctuating quantities. The global equilibrium distribution gjeq‾(x,t)\overline{g_{j}^{\mathrm{eq}}}(\boldsymbol{x},t) and fjeq‾(x,t)\overline{f_{j}^{\mathrm{eq}}}(\boldsymbol{x},t) are the constants and they will not change over time and space. Substituting Eq. (53) into Eqs. (52) and (51), then adding the generation equations, one can get

[(1+ωjk2)δjk+ωjk2Γ1jk]fj′(x,t+Δt)+[(1+ωjk2)δjk+ωjk2Γ2jk]gj′(x,t+Δt)=[(1−ωjk2)δjk+ωjk2Γ1jk]fk′(x,t)−δjkΔtck⋅∇fj′(x,t+Δt2)+[(1−ωjk2)δjk+ωjk2Γ2jk]gk′(x,t)−δjkΔtck⋅∇gj′(x,t+Δt2).(63)\begin{aligned} \left[\left(1+\frac{\omega_{jk}}{2}\right)\delta_{jk}+\frac{\omega_{jk}}{2}\Gamma_{1jk}\right]f'_j\left(\boldsymbol{x},t+\Delta t\right)+\left[\left(1+\frac{\omega_{jk}}{2}\right)\delta_{jk}+\frac{\omega_{jk}}{2}\Gamma_{2jk}\right]g'_j\left(\boldsymbol{x},t+\Delta t\right) \\ &=\left[\left(1-\frac{\omega_{jk}}{2}\right)\delta_{jk}+\frac{\omega_{jk}}{2}\Gamma_{1jk}\right]f'_k(\boldsymbol{x},t)-\delta_{jk}\Delta t\boldsymbol{c}_k\cdot\nabla f'_j\left(\boldsymbol{x},t+\frac{\Delta t}{2}\right) \\ &+\left[\left(1-\frac{\omega_{jk}}{2}\right)\delta_{jk}+\frac{\omega_{jk}}{2}\Gamma_{2jk}\right]g'_k(\boldsymbol{x},t)-\delta_{jk}\Delta t\boldsymbol{c}_k\cdot\nabla g'_j\left(\boldsymbol{x},t+\frac{\Delta t}{2}\right). \tag*{(63)} \end{aligned}

where the Γ1jk\Gamma_{1jk} and Γ2jk\Gamma_{2jk} can be calculated by

Γ1jk=∂fjeq(x,t)∂fk(x,t)+∂gjeq(x,t)∂fk(x,t),Γ2jk=∂fjeq(x,t)∂gk(x,t)+∂gjeq(x,t)∂gk(x,t).(64)\Gamma_{1jk}=\frac{\partial f_j^{\mathrm{eq}}(\boldsymbol{x},t)}{\partial f_k(\boldsymbol{x},t)}+\frac{\partial g_j^{\mathrm{eq}}(\boldsymbol{x},t)}{\partial f_k(\boldsymbol{x},t)},\qquad \Gamma_{2jk}=\frac{\partial f_j^{\mathrm{eq}}(\boldsymbol{x},t)}{\partial g_k(\boldsymbol{x},t)}+\frac{\partial g_j^{\mathrm{eq}}(\boldsymbol{x},t)}{\partial g_k(\boldsymbol{x},t)}. \tag*{(64)}

The Γ1jk\Gamma_{1jk} and Γ2jk\Gamma_{2jk} contain the derivative of the equilibrium distribution function, which is difficult to calculate directly. The chain rule can be used to convert the derivative. For the D2Q5 model, u=(u1,u2)\boldsymbol{u}=(u_1,u_2). Then we can deduce

∂fjeq∂fk=∂fjeq∂u1∂u1∂fk=wj(1+cj⋅u+12cj,1cj,1cs2),\frac{\partial f_j^{\mathrm{eq}}}{\partial f_k}=\frac{\partial f_j^{\mathrm{eq}}}{\partial u_1}\frac{\partial u_1}{\partial f_k}=w_j\left(1+\frac{\boldsymbol{c}_j\cdot\boldsymbol{u}+\frac{1}{2}c_{j,1}c_{j,1}}{c_s^2}\right),
∂fjeq∂gk=∂fjeq∂u2∂u2∂gk=wjcs2(cjβuα−cjαuβ+12cjαcjβ),\frac{\partial f_j^{\mathrm{eq}}}{\partial g_k}=\frac{\partial f_j^{\mathrm{eq}}}{\partial u_2}\frac{\partial u_2}{\partial g_k}=\frac{w_j}{c_s^2}\left(c_{j\beta}u_\alpha-c_{j\alpha}u_\beta+\frac{1}{2}c_{j\alpha}c_{j\beta}\right),
∂gjeq∂fk=∂gjeq∂u1∂u1∂fk=wj(1+cj⋅u+12cjβcjβcs2),\frac{\partial g_j^{\mathrm{eq}}}{\partial f_k}=\frac{\partial g_j^{\mathrm{eq}}}{\partial u_1}\frac{\partial u_1}{\partial f_k}=w_j\left(1+\frac{\boldsymbol{c}_j\cdot\boldsymbol{u}+\frac{1}{2}c_{j\beta}c_{j\beta}}{c_s^2}\right),
∂gjeq∂gk=∂gjeq∂u2∂u2∂gk=wjcs2(cjαuβ−cjβuα+12cjαcjβ).(65)\frac{\partial g_j^{\mathrm{eq}}}{\partial g_k}=\frac{\partial g_j^{\mathrm{eq}}}{\partial u_2}\frac{\partial u_2}{\partial g_k}=\frac{w_j}{c_s^2}\left(c_{j\alpha}u_\beta-c_{j\beta}u_\alpha+\frac{1}{2}c_{j\alpha}c_{j\beta}\right). \tag*{(65)}

Let Sk′(t)=[Fk′(t+Δt2),Gk′(t+Δt2),Fk′(t),Gk′(t)]TS'_k(t)=[F'_k(t+\frac{\Delta t}{2}),G'_k(t+\frac{\Delta t}{2}),F'_k(t),G'_k(t)]^{T}, where Fk′(κ,t)=∫fk′(x,t)exp⁡(−iκ⋅x)dxF'_k(\boldsymbol{\kappa},t)=\int f'_k(\boldsymbol{x},t)\exp(-i\boldsymbol{\kappa}\cdot\boldsymbol{x})d\boldsymbol{x} and Gk′(κ,t)=∫gk′(x,t)exp⁡(−iκ⋅x)dxG'_k(\boldsymbol{\kappa},t)=\int g'_k(\boldsymbol{x},t)\exp(-i\boldsymbol{\kappa}\cdot\boldsymbol{x})d\boldsymbol{x}. Applying the Fourier transform to Eq. (63), the following equation is obtained

Sk′(κ,t+12Δt)=G~Sk′(κ,t),(66)S'_k\left(\boldsymbol{\kappa},t+\frac{1}{2}\Delta t\right)=\widetilde{G}S'_k(\boldsymbol{\kappa},t), \tag*{(66)}

where the G~\widetilde{G} is the growth matrix. Its definition is

G~=diag⁡[I+0.5ω(I−Γ1),I+0.5ω(I−Γ2),I,I]−1\widetilde{G}=\operatorname{diag}\left[I+0.5\omega(I-\Gamma_1),I+0.5\omega(I-\Gamma_2),I,I\right]^{-1}
diag⁡[I−rT,I−rT,I−0.5ω(I+Γ1),I−0.5ω(I+Γ2)].(67)\operatorname{diag}\left[I-rT,I-rT,I-0.5\omega(I+\Gamma_1),I-0.5\omega(I+\Gamma_2)\right]. \tag*{(67)}

where κ=(κx,κy)\boldsymbol{\kappa}=(\kappa_x,\kappa_y) is the wave number, r=Δt/Δxr=\Delta t/\Delta x and T=diag⁡(T0,T1,…Tq)T=\operatorname{diag}(T_0,T_1,\ldots T_q). For the mixed difference scheme,

Tj=i(1−η)(sin⁡ϑjx+sin⁡ϑjy)+η2[6−4exp⁡(−ϑjx)−4exp⁡(−ϑjy)+exp⁡(−2iϑjx)+exp⁡(−2iϑjy)].(68)T_j=i(1-\eta)(\sin\vartheta_{jx}+\sin\vartheta_{jy})+\frac{\eta}{2}\left[6-4\exp(-\vartheta_{jx})-4\exp(-\vartheta_{jy})+\exp(-2i\vartheta_{jx})+\exp(-2i\vartheta_{jy})\right]. \tag*{(68)}

It should be noted that the growth matrix G~\widetilde{G} is a 2q×2q2q\times2q diagonal matrix. For the D2Q5 model, the size of G~\widetilde{G} is 10×1010\times10. Now, the stability of Eq. (60) is equal to the stability of the linear system Eq. (66). According to the von Neumann stability condition, if the spectral radius of growth matrix $\widetilde{G}$ is less than 1, the linear system Eq. (57)` is stable. However, The growth matrix contains many variables, so it is necessary to determine some variables to analyze the numerical stability of the MDF-FDLBM.

Now, we consider the stability analysis of the linear system Eq. (66). First of all, we set u=(0.5,0.5)T\boldsymbol{u}=(0.5,0.5)^T and c=1.0c=1.0. For the multi-relaxation model, the relaxation parameters are set as s0=s3=s4=s′s_0=s_3=s_4=s' and s1=s2=ss_1=s_2=s for MDF-FDLBM. Then the stability regions are presented in Figure 2 and Figure 3. It can be found that the largest area of the stability region appears when s′=ss'=s. When s′s' increases or decreases, the area of stability region of MDF-FDLBM will decreases. Besides, the velocity cc is taken as 1, so the CFL condition number is equal to rr. We also analysis the relationship between the stability region and the CFL condition number with s′=1.5ss'=1.5s. As shown in Figure 3, the stability region areas decrease with the increase of rr.

Stability regions of MDF-FDLBM for four values of $s'$

Figure 2. The stability region of MDF-FDLBM with different s′s'

Stability region for $r = 0.25$

(a) r=0.25r = 0.25

Stability region for $r = 0.5$

(b) r=0.5r = 0.5

Stability region for $r = 0.75$

(c) r=0.75r = 0.75

Stability region for $r = 0.9$

(d) r=0.9r = 0.9

Figure 3. The stability region of MDF-FDLBM with different CFL condition number

The stability analysis of the IFDLBM

The von Neumann method is conducted to the IFDLBM. For simplicity, the source term is neglected and D2Q9 model is adopted.

D2Q9:

cj=(010−101−1−110010−111−1−1)c,(69)\mathbf{c}_j=\begin{pmatrix}0&1&0&-1&0&1&-1&-1&1\\0&0&1&0&-1&1&1&-1&-1\end{pmatrix}c, \tag*{(69)}
w0=4/9,wj=1−4=1/9,wj=5−9=1/36.(70)w_0=4/9,\quad w_{j=1-4}=1/9,\quad w_{j=5-9}=1/36. \tag*{(70)}

According to the Ref. [36], the evolution equation of IFDLBM without source term can be written as

fj(x,t+Δt)+0.5ωjk[fk−fkeq](x,t+Δt)=fj(x,t)−0.5ωjk[fk(x,t)−fkeq(x,t)]−Δtcj⋅∇fj(x,t+Δt2).(71)f_j(\boldsymbol{x},t+\Delta t)+0.5\omega_{jk}\left[f_k-f_k^{\mathrm{eq}}\right]\left(\boldsymbol{x},t+\Delta t\right)=f_j(\boldsymbol{x},t)-0.5\omega_{jk}\left[f_k(\boldsymbol{x},t)-f_k^{\mathrm{eq}}(\boldsymbol{x},t)\right]-\Delta t\boldsymbol{c}_j\cdot\nabla f_j\left(\boldsymbol{x},t+\frac{\Delta t}{2}\right). \tag*{(71)}

Then the linear expansion fj(x,t)=fjeq(x,t)+fj′(x,t)f_j(\boldsymbol{x},t)=f_j^{\mathrm{eq}}(\boldsymbol{x},t)+f'_j(\boldsymbol{x},t) is conducted to Eq. (71), we can obtain

[(1+ωjk2)δjk+ωjk2Γjk]fj′(x,t+Δt)=[(1−ωjk2)δjk+ωjk2Γjk]fk′(x,t)−δjkΔtck⋅∇fj′(x,t+Δt2),(72)\left[\left(1+\frac{\omega_{jk}}{2}\right)\delta_{jk}+\frac{\omega_{jk}}{2}\Gamma_{jk}\right]f'_j(\boldsymbol{x},t+\Delta t)=\left[\left(1-\frac{\omega_{jk}}{2}\right)\delta_{jk}+\frac{\omega_{jk}}{2}\Gamma_{jk}\right]f'_k(\boldsymbol{x},t)-\delta_{jk}\Delta t\boldsymbol{c}_k\cdot\nabla f'_j\left(\boldsymbol{x},t+\frac{\Delta t}{2}\right), \tag*{(72)}

where the term Γjk=∂fjeq(x,t)/∂fk(x,t)\Gamma_{jk}=\partial f_j^{\mathrm{eq}}(\boldsymbol{x},t)/\partial f_k(\boldsymbol{x},t). It should be noticed that the equilibrium distribution function of IFDLBM is different from that of MDF-FDLBM. This is also the main reason for the different stability regions of the two models.

For the IFDLBM, the Γjk\Gamma_{jk} can be calculated by

∂fjeq∂fk={wjcj⋅cjcs2+wj(cjcj−cs2I):(cj⋅∇uu)2cs4,if k=0,−λjλ0+λjω0λ0(cj⋅u)+wjcj⋅cjcs2+wj(cjcj−cs2I):(cj⋅∇uu)2cs4,if k≠0.(73)\frac{\partial f_j^{\mathrm{eq}}}{\partial f_k}= \begin{cases} w_j\frac{\boldsymbol{c}_j\cdot\boldsymbol{c}_j}{c_s^2}+w_j\frac{(\boldsymbol{c}_j\boldsymbol{c}_j-c_s^2\boldsymbol{I}):(\boldsymbol{c}_j\cdot\nabla\boldsymbol{u}\boldsymbol{u})}{2c_s^4}, & \text{if } k=0,\\ -\frac{\lambda_j}{\lambda_0}+\frac{\lambda_j\omega_0}{\lambda_0}(\boldsymbol{c}_j\cdot\boldsymbol{u})+w_j\frac{\boldsymbol{c}_j\cdot\boldsymbol{c}_j}{c_s^2}+w_j\frac{(\boldsymbol{c}_j\boldsymbol{c}_j-c_s^2\boldsymbol{I}):(\boldsymbol{c}_j\cdot\nabla\boldsymbol{u}\boldsymbol{u})}{2c_s^4}, & \text{if } k\ne0. \end{cases} \tag*{(73)}

Let Qk′(t)=[Fk′(t+Δt2),Fk′(t)]TQ'_k(t)=[F'_k(t+\frac{\Delta t}{2}),F'_k(t)]^{\mathrm{T}}. Applying the Fourier transform to Eq. (72), the following equation is obtained

\begin{pmatrix} I+0.5\omega-0.5\omega\Gamma&0\\ 0&I \end{pmatrix} Q'_k(\boldsymbol{\kappa},t+\frac{1}{2}\Delta t)= \begin{pmatrix} -r^{\mathrm{T}}&I-0.5\omega+0.5\omega\Gamma\\ I&0 \end{pmatrix} Q'_k(\boldsymbol{\kappa},t), \tag*{(74)}\end{pmatrix}

and we have

Sk′(κ,t+12Δt)=G^Sk′(κ,t),(75)S'_k\left(\boldsymbol{\kappa},t+\frac{1}{2}\Delta t\right)=\widehat{G}S'_k(\boldsymbol{\kappa},t), \tag*{(75)}

From Eq. (74), the growth matrix G^\widehat{G} can be determined as

G^=(I+0.5ω−0.5ωΓ00I)−1(−rTI−0.5ω+0.5ωΓI0).(76)\widehat{G}= \begin{pmatrix} I+0.5\omega-0.5\omega\Gamma&0\\ 0&I \end{pmatrix}^{-1} \begin{pmatrix} -r^{\mathrm{T}}&I-0.5\omega+0.5\omega\Gamma\\ I&0 \end{pmatrix}. \tag*{(76)}

It should be noted that, for IFDLBM with D2Q9 model, the growth matrix G^\widehat{G} is an 18×1818\times18 matrix and it is no longer a diagonal matrix. If the spectral radius of growth matrix G^\widehat{G} is less than 1, the linear system Eq. (74) is stable. However, too many free parameters make it difficult to analyze matrix G^\widehat{G}. Therefore, we set u=(0.5,0.5)T\boldsymbol{u}=(0.5,0.5)^{\mathrm{T}} and c=1.0c=1.0. Then it is clear that the spectral radius of the matrix G^\widehat{G} is the function of parameters ω\omega and rr. As we know c=1.0c=1.0, so rr is equals to the CFL condition number. For the multi-relaxation model, the relaxation parameters are taken as s0=s1=s2=s4=s6=s7=s8=s′s_0=s_1=s_2=s_4=s_6=s_7=s_8=s' and s3=s5=ss_3=s_5=s for IFDLBM.

Figure 4 shows the stability domains of IFDLBM with different s′s'. When s′=1.5ss'=1.5s, the area of the stable region is the largest, and when s′s' is less than ss, the area of the stable region will reduce significantly. Compared with the stability region of MDF-FDLBM, we can find the IFDLBM will be more stable with high ω\omega. Because ω=Δts\omega=\Delta t s and cs2/s=νc_s^2/s=\nu, it can be deduced that the IFDLBM will be more stable under the condition of low ν\nu. Besides, we also fixed s′=1.5ss'=1.5s and presented the stability domains of IFDLBM with different rr in Figure 5. It is clear that the stability region areas decrease with the increase of rr. This phenomenon is similar to the results of some stability analyses. Compared with Figure 3, we can find that MDF-FDLBM is more stable when the large CFL condition number is taken.

Stability region for $s' = 0.5s$

(a) (a) s′=0.5ss' = 0.5s

Stability region for $s' = 1.0s$

(b) (b) s′=1.0ss' = 1.0s

Stability region for $s' = 1.5s$

(c) (c) s′=1.5ss' = 1.5s

Stability region for $s' = 2.0s$

(d) (d) s′=2.0ss' = 2.0s

Figure 4. The stability region of IFDLBM with different s′s'

Stability region for $r = 0.1$

(a) r=0.1r = 0.1

Stability region for $r = 0.25$

(b) r=0.25r = 0.25

Stability region for $r = 0.5$

(c) r=0.5r = 0.5

Stability region for $r = 0.7$

(d) r=0.7r = 0.7

Figure 5. The stability region of IFDLBM with different CFL condition number

::::

Numerical example

In this section, four different numerical examples are used to test the present MDF-FDLBM, including the four-roll mill problem, the periodic flow, the poiseuille flow and the lid driven flow. Unless otherwise specified, the distribution function is initialized by the equilibrium distribution function, i.e., f^j=fjeq\hat{f}_j=f_j^{eq}. We use the D2Q5 lattice model and s′=1.5ss'=1.5s for all MDF-FDLBM in this section. To test the accuracy of the MDF-FDLBM, the following global relative error (GRE) is adopted

GRE=∑i∣ϕ(xi,t)−ϕ∗(xi,t)∣∑i∣ϕ∗(xi,t)∣,(77)GRE=\frac{\sum_i|\phi(\boldsymbol{x}_i,t)-\phi^*(\boldsymbol{x}_i,t)|}{\sum_i|\phi^*(\boldsymbol{x}_i,t)|}, \tag*{(77)}

where ϕ(x,t)\phi(\boldsymbol{x},t) and ϕ∗(x,t)\phi^*(\boldsymbol{x},t) denote numerical and analytical solutions, respectively.

The two-dimensional four-roll mill problem

To test the validity of the MDF-FDLBM, we use the MDF-FDLBM to simulate the two-dimensional four-roll mill problem. The physical domain is fixed in [0,2π]×[0,2π][0,2\pi]\times[0,2\pi], and the periodic boundary condition is used for four boundaries. The force term can be given by

F1=U02sin⁡(x)cos⁡(x)+2νU0sin⁡(x)cos⁡(y),F2=U02sin⁡(y)cos⁡(y)+2νU0sin⁡(y)cos⁡(x).(78)\begin{aligned} F_1&=U_0^2\sin(x)\cos(x)+2\nu U_0\sin(x)\cos(y),\\ F_2&=U_0^2\sin(y)\cos(y)+2\nu U_0\sin(y)\cos(x). \tag*{(78)} \end{aligned}

The analytical solution of velocity is defined as,

u1=U0sin⁡(x)cos⁡(y),u2=−U0sin⁡(y)cos⁡(x).(79)u_1=U_0\sin(x)\cos(y),\qquad u_2=-U_0\sin(y)\cos(x). \tag*{(79)}

According to the analytical solutions of velocity, we can get the analytical solutions of velocity gradient, velocity divergence, stain rate tensor,

∂u1∂x=−∂u2∂y=U0cos⁡(x)cos⁡(y),∂u1∂y=−∂u2∂x=−U0sin⁡(x)sin⁡(y),\frac{\partial u_1}{\partial x}=-\frac{\partial u_2}{\partial y}=U_0\cos(x)\cos(y),\qquad\frac{\partial u_1}{\partial y}=-\frac{\partial u_2}{\partial x}=-U_0\sin(x)\sin(y),
∇⋅u=∂u1∂x+∂u2∂y=0,\nabla\cdot u=\frac{\partial u_1}{\partial x}+\frac{\partial u_2}{\partial y}=0,
Sxx=−Syy=U0cos⁡(x)cos⁡(y),Sxy=Syx=0.S_{xx}=-S_{yy}=U_0\cos(x)\cos(y),\qquad S_{xy}=S_{yx}=0.
ω=∂u2∂x−∂u1∂y=2U0sin⁡(x)sin⁡(y).(80)\omega=\frac{\partial u_2}{\partial x}-\frac{\partial u_1}{\partial y}=2U_0\sin(x)\sin(y). \tag*{(80)}

In our simulation, the lattice grids is set as 64×6464\times64, and we take U0=0.0001U_0=0.0001, ν=0.01\nu=0.01, CFL=0.5CFL=0.5. Figure 6, Figure 7, Figure 8 and Figure 9 show the numerical solutions of velocity, velocity gradient, stain rate tensor and vorticity of MDF-FDLBM at different locations, respectively. It can be found that the numerical results of these physical quantities agree well with the analytical solutions. The results show that MDF-FDLBM is effective to solve the convection-diffusion system based NSEs, and the calculation schemes of these physical quantities are correct.

Two-panel plots of velocity at different positions

Figure 6. The MDF-FDLBM numerical and analytical solution of velocity at different positions [symbol:numerical solution, solid line: analytical solution].

Four-panel plots of velocity gradient at different positions

Figure 7. The MDF-FDLBM numerical and analytical solution of velocity gradient at different positions [symbol:numerical solution, solid line: analytical solution].

MDF-FDLBM numerical and analytical solution of strain rate tensor at different positions

Figure 8. The MDF-FDLBM numerical and analytical solution of strain rate tensor at different positions [symbol:numerical solution, solid line: analytical solution].

MDF-FDLBM numerical and analytical solution of vorticity at different positions

Figure 9. The MDF-FDLBM numerical and analytical solution of vorticity at different positions [symbol:numerical solution, solid line: analytical solution].

In addition, we need to measure the convergence rate of the MDF-FDLBM. The use of periodic boundary conditions in the four-roll mill problem ensures that the convergence rate of the MDF-FDLBM is not affected by boundaries. In theory, the MDF-FDLBM has a second-order convergence rate in space and time, so the error is O(δt2+δx2)O(\delta t^2+\delta x^2). Some numerical simulations are carried out at different lattice grids (N×N=60×60N\times N=60\times60, 70×7070\times70, 80×8080\times80, 90×9090\times90, 100×100100\times100, 1100×1101100\times110) with dt=dxdt=dx. The results are shown in Figure 10. From the figure, we can find that the MDF-FDLBM has second-order accuracy in space and time, which is consistent with theoretical accuracy.

GRE of MDF-FDLBM in different lattice space and time step

Figure 10. The GRE of MDF-FDLBM in different lattice space and time step

The periodic flow

Now, we are going to use the periodic flow to verify the performance of the MDF-FDLBM. It is a steady state problem and the external force term can be expressed by

Fx(x,y)=8π2νu0sin⁡2πxsin⁡2πy,F_x(x,y)=8\pi^2\nu u_0\sin2\pi x\sin2\pi y,
Fy(x,y)=8π2νu0cos⁡2πxcos⁡2πy.(81)F_y(x,y)=8\pi^2\nu u_0\cos2\pi x\cos2\pi y. \tag*{(81)}

where the kinetic viscosity ν\nu is set to be 0.01 and u0=0.1u_0 = 0.1. Under the proper initial and boundary conditions, the analytical solutions can be obtained,

p(x,y)=14ρ0u02(cos⁡4πx−cos⁡4πy),u(x,y)=u0sin⁡2πxsin⁡2πy,v(x,y)=u0cos⁡2πxcos⁡2πy,(82)\begin{aligned} p(x,y) &= \frac{1}{4}\rho_0u_0^2(\cos4\pi x-\cos4\pi y),\\ u(x,y) &= u_0\sin2\pi x\sin2\pi y,\\ v(x,y) &= u_0\cos2\pi x\cos2\pi y, \tag*{(82)} \end{aligned}

where ρ0=1.0\rho_0 = 1.0, u=(u,v)\boldsymbol{u} = (u,v) is the velocity and pp is the presure. We implemented a simulation with a fixed computational region of [0,1]×[0,1][0,1] \times[0,1], utilizing periodic boundary conditions on the four boundaries. The CFL condition number is set to 0.5 and the weight coefficient of mixed difference scheme η=0.8\eta= 0.8. The uniform and non-uniform grids are used to simulate this problem, respectively. The results of velocity and pressure with grid 32×3232 \times32 and 64×3264 \times32 are shown in Figure 11, which demonstrates good agreement with the analytical solution. Some physical quantities are also calculated through the non-equilibrium distribution function. And the numerical results are presented in the Figure 12. The numerical results of those physical quantities are in good agreement with the analytical solution. This suggests that computational schemes utilizing the non-equilibrium distribution function are effective, as is the MDF-FDLBM combined with non-uniform grid when solving a convection-diffusion system based on NSEs.

Three plots comparing uniform-grid, non-uniform-grid, and analytical solutions for velocity components $u$ and $v$, and pressure $p$

Figure 11. The MDF-FDLBM numerical and analytical solutions of the velocity and pressure.

The numerical solution and analytical solutions of some physical quantity

Figure 12. (a) ∂u/∂x\partial u/\partial x; (b) ∂v/∂y\partial v/\partial y; (c) ∇⋅u\nabla\cdot u; (d) SxxS_{xx}; (e) SyyS_{yy}; (f) Ωxy\Omega_{xy}. Figure 12: The numerical solution and analytical solutions of some physical quantity.

Besides, the computational efficiency of the MDF-FDLBM is also a key characteristic that we focus on. Therefore, we recorded the computing time and GREs of IFDLBM and MDF-FDLBM respectively. In order to effectively compare the computational efficiency of the two models, the evolution of both models are stopped at 10000 steps, and the periodic flow tends to be stable at this time. The computational time and GREs of the two models are shown in the Table 1. From the Table 1, it can be found the GREs of MDF-FDLBM are smaller than that of IFDLBM. At the same time, the computational time of MDF-FDLBM is smaller than that of IFDLBM. Compared to IFDLBM, MDF-FDLBM saves more than 35% in computation time. There are two reasons for this phenomenon. Firstly, D2Q9 model is used in IFDLBM, while the MDF-FDLBM adopts the D2Q5 model. There are 9×9=819 \times9 = 81 cycles are involved in the IFDLBM evolution process. For the MDF-FDLBM, two evolution equations are calculated in simulation, and there are 2×5×5=502 \times5 \times5 = 50 cycles in one evolutionary step. Secondly, the MDF-FDLBM adopts a linear equilibrium distribution function, while the equilibrium distribution function in IFDLBM contains quadratic terms. Obviously, the calculation of equilibrium distribution function in MDF-FDLBM is simpler than that in IFDLBM. Therefore, the reduction of the cycles number results in the reduction of the computation time. For this problem, the computational efficiency of MDF-FDLBM is higher than that of IFDLBM, and the MDF-FDLBM is more accurate than that of IFDLBM.

ν\numodel16×1616 \times1632×3232 \times3248×4848 \times4864×6464 \times64
ν=0.005\nu= 0.005GRE of IFDLBM3.1899×10−23.1899 \times10^{-2}6.3791×10−36.3791 \times10^{-3}4.9637×10−34.9637 \times10^{-3}4.2204×10−34.2204 \times10^{-3}
time of IFDLBM2.58s2.58s9.53s9.53s21.01s21.01s37.24s37.24s
GRE of MDF-FDLBM2.3754×10−22.3754 \times10^{-2}6.1662×10−46.1662 \times10^{-4}2.2326×10−42.2326 \times10^{-4}2.7564×10−42.7564 \times10^{-4}
time of MDF-FDLBM1.52s1.52s5.93s5.93s12.31s12.31s22.93s22.93s
decrease percentage41.09%41.09\%37.78%37.78\%41.41%41.41\%38.43%38.43\%
ν=0.001\nu= 0.001GRE of IFDLBM1.0719×10−11.0719 \times10^{-1}1.0994×10−21.0994 \times10^{-2}2.0904×10−32.0904 \times10^{-3}3.5852×10−43.5852 \times10^{-4}
time of IFDLBM2.6s2.6s9.53s9.53s21.00s21.00s37.02s37.02s
GRE of MDF-FDLBM1.6495×10−21.6495 \times10^{-2}6.1275×10−36.1275 \times10^{-3}4.2685×10−44.2685 \times10^{-4}7.8567×10−57.8567 \times10^{-5}
time of MDF-FDLBM1.51s1.51s5.94s5.94s12.29s12.29s23.37s23.37s
decrease percentage41.92%41.92\%37.67%37.67\%41.48%41.48\%36.87%36.87\%

Table 1. Example 1: A comparison of the GREs and CPU time between the IFDLBM and MDF-FDLBM.

The two-dimensional Poiseuille flow

The two-dimensional Poiseuille flow is taken account for testing the stability of the MDF-FDLBM. This problem is driven by a constant external force (F1=1.0×10−6F_1 = 1.0 \times10^{-6}) in a channel, and its analytical solution of velocity, velocity

gradient, velocity divergence, strain rate tensor and vorticity can be expressed as

u1=F1H22ν[yH−(yH)2],u2=0,u_1=\frac{F_1H^2}{2\nu}\left[\frac{y}{H}-\left(\frac{y}{H}\right)^2\right],\qquad u_2=0,
∂u1∂x=∂u2∂x=∂u2∂y=0,∂u1∂y=F1H2ν[1−2yH],\frac{\partial u_1}{\partial x}=\frac{\partial u_2}{\partial x}=\frac{\partial u_2}{\partial y}=0,\qquad\frac{\partial u_1}{\partial y}=\frac{F_1H}{2\nu}\left[1-2\frac{y}{H}\right],
∇⋅u=∂u1∂x+∂u2∂y=0,(83)\nabla\cdot u=\frac{\partial u_1}{\partial x}+\frac{\partial u_2}{\partial y}=0, \tag*{(83)}
Sxx=Syy=0,Sxy=Syx=F1H4ν[1−2yH],S_{xx}=S_{yy}=0,\qquad S_{xy}=S_{yx}=\frac{F_1H}{4\nu}\left[1-2\frac{y}{H}\right],
ω=∂u2∂x−∂u1∂y=−F1H2ν[1−2yH].\omega=\frac{\partial u_2}{\partial x}-\frac{\partial u_1}{\partial y}=-\frac{F_1H}{2\nu}\left[1-2\frac{y}{H}\right].

In our simulation, the lattice size is set as 32×3232 \times32 for the computational domain [0,1]×[0,1][0,1]\times[0,1]. The periodic boundary condition is adopted in xx direction and the non-equilibrium extrapolation scheme is applied to treat the top and bottom walls. Figure 13,Figure 14,Figure 15,Figure 16 show the numerical results of MDF-FDLBM with ν=0.001\nu=0.001. It can be found that the results of velocity uu, velocity gradient (∂u1/∂x\partial u_1/\partial x and ∂u1/∂y\partial u_1/\partial y), velocity divergence ∇⋅u\nabla\cdot u, strain rate tensor (SxxS_{xx} and SxyS_{xy}) and the vorticity ω\omega agree well with the analytical solution.

The MDF-FDLBM numerical and analytical solution of velocity at different positions

Figure 13. The MDF-FDLBM numerical and analytical solution of velocity at different positions [symbol:numerical solution, solid line: analytical solution].

Velocity gradient at different positions

Figure 14. The MDF-FDLBM numerical and analytical solution of velocity gradient at different positions [symbol:numercial solution, solid line: analytical solution].

Strain rate tensor at different positions

Figure 15. The MDF-FDLBM numerical and analytical solution of strain rate tensor at different positions [symbol:numercial solution, solid line: analytical solution].

Vorticity at different positions

Figure 16. The MDF-FDLBM numerical and analytical solution of vorticity at different positions [symbol:numercial solution, solid line: analytical solution].

We also conduct some comparation between the MDF-FDLBM, IFDLBM and MDF-LBM. The results are presented in Table 2. It can be noticed that the stability of MDF-FDLBM is better than the MDF-LBM while the errors of both models are similar. Besides, from Table 2, we can find that the MDF-FDLBM is more stable with large ν\nu, while the IFDLBM is more stable with small ν\nu. This numerical result is consistent with the conclusion obtained from the stability analysis. In addition, it is observed that the GREs of MDF-FDLBM are smaller than that of IFDLBM.

F=10−1F = 10^{-1}F=10−2F = 10^{-2}F=10−3F = 10^{-3}F=10−4F = 10^{-4}F=10−5F = 10^{-5}F=10−6F = 10^{-6}
ν=0.01\nu= 0.01MDF-FDLBM3.4607×10−33.4607 \times10^{-3}6.1900×10−46.1900 \times10^{-4}6.2142×10−46.2142 \times10^{-4}6.2145×10−46.2145 \times10^{-4}6.2145×10−46.2145 \times10^{-4}6.2145×10−46.2145 \times10^{-4}
IFDLBM--2.7500×10−22.7500 \times10^{-2}2.7488×10−22.7488 \times10^{-2}2.7488×10−22.7488 \times10^{-2}2.7488×10−22.7488 \times10^{-2}2.7488×10−22.7488 \times10^{-2}
ν=0.001\nu= 0.001MDF-FDLBM--1.3363×10−11.3363 \times10^{-1}6.4287×10−36.4287 \times10^{-3}4.7497×10−44.7497 \times10^{-4}4.7447×10−44.7447 \times10^{-4}4.7446×10−44.7446 \times10^{-4}
IFDLBM----1.5170×10−21.5170 \times10^{-2}1.5337×10−21.5337 \times10^{-2}1.5339×10−21.5339 \times10^{-2}1.5339×10−21.5339 \times10^{-2}
ν=0.0001\nu= 0.0001MDF-FDLBM--------3.4607×10−33.4607 \times10^{-3}2.9603×10−42.9603 \times10^{-4}
IFDLBM------6.4776×10−26.4776 \times10^{-2}6.4639×10−26.4639 \times10^{-2}6.4639×10−26.4639 \times10^{-2}

Table 2. Example 3: A comparison of the GREs between the MDF-FDLBM and IFDLBM.

In order to further verify the conclusion of stability analysis, we use MDF-FDLBM and IFDLBM to simulate this example under different CFL condition number. The external force term FF is set to 10−410^{-4}, the grid size is set to 32×3232\times32 and c=1.0c=1.0. The simulation results of different CFL condition number are shown in Table 3. It can be found that the stability of MDF-FDLBM is better than IFDLBM when the CFL condition number is large, and the error of MDF-FDLBM is still small with large CFL condition number. In addition, the GRE of MDF-FDLBM is smaller than IFDLBM, indicating that MDF-FDKBM is more accurate than IFDLBM in this example. Meanwhile, the calculation time of the two models are also recorded in Table 3. It can be seen that the calculation time of MDF-FDKBM is obviously less than that of IFDLBM. This phenomenon is the same as in Table 1. The reason for this phenomenon are the same. In this example, the reduction time of MDF-FDLBM is more than 56%. This also indicates that the computational efficiency of MDF-FFLBM is higher than that of IFDLBM.

CFL=0.1CFL = 0.1CFL=0.3CFL = 0.3CFL=0.5CFL = 0.5CFL=0.7CFL = 0.7CFL=0.9CFL = 0.9
ν=0.01\nu= 0.01
GRE of MDF-FDLBM4.0605×10−34.0605 \times10^{-3}4.0336×10−34.0336 \times10^{-3}3.9957×10−33.9957 \times10^{-3}4.2918×10−34.2918 \times10^{-3}6.8811×10−36.8811 \times10^{-3}
time of MDF-FDLBM116.47s116.47s38.84s38.84s23.35s23.35s16.79s16.79s13.01s13.01s
GRE of IFDLBM2.7488×10−22.7488 \times10^{-2}3.2652×10−23.2652 \times10^{-2}3.7857×10−23.7857 \times10^{-2}4.3119×10−24.3119 \times10^{-2}4.8433×10−24.8433 \times10^{-2}
time of IFDLBM266.89s266.89s90.52s90.52s56.09s56.09s38.33s38.33s29.89s29.89s
time reduction ratio56.36%56.36\%57.39%57.39\%58.37%58.37\%56.20%56.20\%56.47%56.47\%
ν=0.001\nu= 0.001
GRE of MDF-FDLBM2.0447×10−32.0447 \times10^{-3}1.7724×10−31.7724 \times10^{-3}1.5278×10−31.5278 \times10^{-3}1.4421×10−31.4421 \times10^{-3}1.6319×10−31.6319 \times10^{-3}
time of MDF-FDLBM115.91s115.91s38.65s38.65s23.37s23.37s16.79s16.79s13.01s13.01s
GRE of IFDLBM5.0468×10−25.0468 \times10^{-2}1.1061×10−11.1061 \times10^{-1}1.8049×10−11.8049 \times10^{-1}2.6347×10−12.6347 \times10^{-1}--
time of IFDLBM280.08s280.08s102.99s102.99s61.61s61.61s43.99s43.99s--
time reduction ratio58.62%58.62\%62.29%62.29\%62.07%62.07\%61.83%61.83\%--
ν=0.0005\nu= 0.0005
GRE of MDF-FDLBM4.1387×10−34.1387 \times10^{-3}3.8526×10−33.8526 \times10^{-3}3.6170×10−33.6170 \times10^{-3}3.4875×10−33.4875 \times10^{-3}3.4832×10−33.4832 \times10^{-3}
time of MDF-FDLBM116.16s116.16s38.75s38.75s23.35s23.35s16.66s16.66s13.10s13.10s
GRE of IFDLBM7.4985×10−27.4985 \times10^{-2}2.0969×10−12.0969 \times10^{-1}3.9927×10−13.9927 \times10^{-1}7.6692×10−17.6692 \times10^{-1}--
time of IFDLBM279.60s279.60s93.71s93.71s56.17s56.17s40.22s40.22s--
time reduction ratio58.45%58.45\%58.65%58.65\%58.43%58.43\%58.58%58.58\%--

Table 3. Example 3: A comparison of the GREs between the MDF-FDLBM and IFDLBM with different CFL condition number.

The two-dimensional lid-driven cavity flow

As a bench mark problem, the lid-driven cavity flow is simulated to test the capacity of the MDF-FDLBM, which is driven by a constant velocity u0=0.1u_0 = 0.1 of the top wall. The length of the square cavity is taken as 1. It is a complicated problem because the fluid mechanical phenomena is rich and there is no analytical solution available. The initial velocity and press are set to be u=0u = 0 and p=0p = 0 and the four wall is treated by the non-equilibrium extrapolation scheme.

In this part, we would conduct some simulations by the MDF-FDLBM at different Reynolds number (Re=Lu0/νRe = Lu_0/\nu). In our simulation, the lattice size is chosen to be 128 ×\times 128 for Re=400Re = 400 and 256 ×\times 256 for Re=1000Re = 1000, 3200, 5000. We take κ=1.0\kappa= 1.0 and CFL=0.5CFL = 0.5. The simulations results are represented in Figure 17. As we can seen, there are four vortices appear in the cavity when Re≤1000Re \le1000, the first one is a primary vortex at the center of the cavity, the others are the secondary vortices at the lower left and lower right corners and a third level vortex at the lower right corner. When Re number is up to 3200 or 5000, we can observe a secondary vortex appears in the upper left corner. And a new third level vortex appears at the lower left corner. Besides, the center of the primary vortex approaches the center of the cavity as ReRe increases. These phenomena are consistent with previous work [23]. Compared with the simulation results of MDF-LBM [37], MDF-FDLBM can capture more flow details even for Nx×Ny=256×256N_x \times N_y = 256 \times256. The profiles of velocity along vertical and horizontal lines are displayed in Figure 18. It is obviously that the results are in good agreement with the available results. In Table 10, the locations of the vortices are also consistent with the previous work.

Streamlines of lid-driven flow at $Re=400$

(a) Re=400Re=400

Streamlines of lid-driven flow at $Re=1000$

(b) Re=1000Re=1000

Streamlines of lid-driven flow at $Re=3200$

(c) Re=3200Re=3200

Streamlines of lid-driven flow at $Re=5000$

(d) Re=5000Re=5000

Figure 17. The streamline of Lid driven flow under different ReRe number

Velocity profiles along the centerline at different $Re$

Figure 18. Velocity profiles along the centerline at different ReRe, (o) is reference date [23]. [(a) Re=400Re=400, 1000, 3200, 5000 from left to right; (b) Re=400Re=400, 1000, 3200, 5000 from bottom to top.]

Conclusions

In this paper, the incompressible NSEs is transformed in to a convection-diffusion system. A MDF-FDLBM with the MRT model is proposed for the convection-diffusion system based NSEs. Through the CE analysis, the incompressible NSEs can be recovered correctly from the MDF-FDLBM. Additionally, some local scheme for the velocity gradient, velocity divergence, strain rate tensor, shear stress, and vorticity are proposed. We also analyze

Re=400Re = 400Re=1000Re = 1000Re=3200Re = 3200Re=5000Re = 5000
primary vortexMDF-FDLBMxx0.55680.53300.51760.5144
yy0.60810.56790.54030.5344
Ref[38]xx0.55470.53130.51650.5176
yy0.60550.56260.54690.5373
primary vortexMDF-FDLBMxx0.05000.08300.08080.0720
yy0.04860.07810.11860.1355
Ref[38]xx0.05080.08590.08590.0784
yy0.04690.07810.10940.1373
primary vortexMDF-FDLBMxx0.88790.86620.81780.7937
yy0.12270.11300.08500.0827
Ref[38]xx0.89060.89060.81250.8078
yy0.12500.12500.08590.0745

Table 4. Example 3: A comparison of the GREs between the MDF-FDLBM and MDF-LBM.

the stability of the MDF-FDLBM and IFDLBM. It is found that the IFDLBM will be more stable than that of MDF-FDLBM with smaller ν\nu, and the MDF-FDLBM will be more stable than that of IFDLBM with large CFL condition number.

Some classic fluid problems are conducted to test the performance of the MDF-FDLBM, including the two-dimensional four-roll mill problem, the periodic flow, the two-dimensional Poiseuille flow and the two-dimensional lid-driven cavity flow. The results make great agreements with the analytical solution or previous works. This shows that MDF-FDLBM and the local scheme of these physical quantities are valid. Through the numerical testing, the MDF-FDLBM has a second-order convergence rate in space and time. Besides, the stability of MDF-FDLBM and IFDLBM was compared by different examples. The numerical results show that IFDLBM is more stable under small kinematic viscosity, while MDF-FDLBM is more stable under large CFL condition number. This is consistent with the conclusion of theoretical analysis. Meanwhile, compared with IFDLBM, it can be found that the simulation results of MDF-FDLBM are more accurate, and the calculation efficiency of MDF-FDLBM is higher.

In light of the aforementioned advantages, it is reasonable to suggest that the MDF-FDLBM can be extended for investigating thermal flows and multiphase fluid systems subject to the incompressible NSEs and CDE, which will be considered in the future work.

Acknowledgments

This work is supported by the National Natural Science Foundation of China (Grants No. 51836003 and No. 12072127), the Graduates Innovation Fund, Huazhong University of Science and Technology (No. 2020yjsCXCY034), and the Nature Science Foundation of Hubei province (Grant No. 2020CFB384).

References

  1. [1]S. Succi, The lattice Boltzmann equation: For fluid dynamics and beyond, Clarendon Press (2001).
  2. [2]M. Sheikholeslami, D. D. Ganji, Entropy generation of nanofluid in presence of magnetic field using lattice Boltzmann method, Physica A 417 (2015) 273–286.
  3. [3]C. Lim, C. M. Chu, S. Xu, Y. Chew, Application of lattice Boltzmann method to simulate microchannel flows, Phys. Fluids 14 (7) (2002) 2299–2308.
  4. [4]D. Raabe, Overview of the lattice Boltzmann method for nano-and microscale fluid dynamics in materials science and engineering, Model. Simul. Mater. Sci. Eng. 12 (6) (2004) R13.
  5. [5]S. Tao, Z. Guo, Boundary condition for lattice Boltzmann modeling of microscale gas flows with curved walls in the slip regime, Phys. Rev. E 91 (4) (2015) 043305.
  6. [6]F. J. Higuera, S. Succi, R. Benzi, Lattice gas dynamics with enhanced collisions, Europhysics Letters 9 (4) (1989) 345–349.
  7. [7]Z. Guo, T. Zhao, Lattice Boltzmann model for incompressible flows through porous media, Phys. Rev. E 66 (3 Pt 2B) (2002) 036304.
  8. [8]A. Montessori, P. Prestininzi, M. L. Rocca, S. Succi, Lattice Boltzmann approach for complex nonequilibrium flows, Phys. Rev. E 92 (2015) 043308.
  9. [9]G. Strumolo, V. Babu, New directions in computational aerodynamics, Phys. World 10 (8) (1997) 45.
  10. [10]H. Chen, S. Kandasamy, S. Orszag, R. Shock, S. Succi, V. Yakhot, Extended Boltzmann kinetic equation for turbulent flows, Science 301 (5633) (2003) 633–636.
  11. [11]X. Shan, H. Chen, Lattice Boltzmann model for simulating flows with multiple phases and components, Phys. Rev. E 47 (3) (1993) 1815–1819.
  12. [12]M. R. Swift, W. R. Osborn, J. M. Yeomans, Lattice Boltzmann simulation of nonideal fluids, Phys. Rev. Lett. 75 (5) (1995) 830–833.
  13. [13]A. J. Briant, A. J. Wagner, J. M. Yeomans, Lattice Boltzmann simulations of contact line motion. i. liquid-gas systems, Phys. Rev. E 69 (14) (2004) 031602.
  14. [14]G. Falcucci, S. Ubertini, S. Succi, Lattice Boltzmann simulations of phase-separating flows at large density ratios: the case of doubly-attractive pseudo-potentials, Soft Matter 6 (18) (2010) 4357–4365.
  15. [15]L. Biferale, P. Perlekar, M. Sbragaglia, F. Toschi, Convection in multiphase fluid flows using lattice Boltzmann methods, Phys. Rev. Lett 108 (5) (2012) 104502.
  16. [16]Z. Chai, H. Liang, R. Du, B. Shi, A lattice Boltzmann model for two-phase flow in porous media, SIAM J. Sci. Comput. 41 (4) (2019) B746–B772.
  17. [17]M. B. Reider, J. D. Sterling, Accuracy of discrete-velocity BGK models for the simulation of the incompressible Navier-Stokes equations, Comput. fluids 24 (4) (1995) 459–467.
  18. [18]R. Mei, W. Shyy, On the finite difference-based lattice Boltzmann method in curvilinear coordinates, J. Comput. Phys. 143 (2) (1998) 426–448.
  19. [19]Z. Guo, T. Zhao, Explicit finite-difference lattice Boltzmann method for curvilinear coordinates, Phys. Rev. E 67 (6) (2003) 066709.
  20. [20]H. Wang, B. Shi, H. Liang, Z. Chai, Finite-difference lattice Boltzmann model for nonlinear convection-diffusion equation, Appl. Math. Comput. 309 (2017) 334–349.
  21. [21]Z. Guo, T. Zhao, Finite-difference-based lattice Boltzmann model for dense binary mixtures, Phys. Rev. E 71 (2) (2005) 026701.
  22. [22]Y. Wang, Y. He, J. Huang, Q. Li, Implicit–explicit finite-difference lattice Boltzmann method with viscid compressible model for gas oscillating patterns in a resonator, Int. J. Numer. Methods Fluids 59 (8) (2009) 853–872.
  23. [23]X. Chen, Z. Chai, H. Wang, B. Shi, A finite-difference lattice Boltzmann method with second-order accuracy of time and space for incompressible flow, Comput. Math. Appl. 80 (12) (2020) 3066–sai2019naturalsai2019naturalsai3081.
  24. [24]X. Chen, Z. Chai, J. Shang, B. Shi, Multiple-relaxation-time finite-difference lattice Boltzmann model for the nonlinear convection-diffusion equation, Phys. Rev. E 104 (2021) 035308.
  25. [25]K. Hejranfar, E. Ezzatneshan, A high-order compact finite-difference lattice Boltzmann method for simulation of steady and unsteady incompressible flows, Int. J. Numer. Methods Fluids 75 (10) (2014) 713–746.
  26. [26]K. Hejranfar, M. H. Saadat, Preconditioned weno finite-difference lattice Boltzmann method for simulation of incompressible turbulent flows, Comput. Math. Appl. 76 (6) (2018) 1427–1446.
  27. [27]E. Ezzatneshan, K. Hejranfar, Simulation of three-dimensional incompressible flows in generalized curvilinear coordinates using a high-order compact finite-difference lattice Boltzmann method, Int. J. Numer. Methods Fluids 89 (7) (2019) 235–255.
  28. [28]K. Hejranfar, E. Ezzatneshan, Simulation of two-phase liquid-vapor flows using a high-order compact finite-difference lattice Boltzmann method, Phys. Rev. E 92 (5) (2015) 053305.
  29. [29]S. Polasanapalli, K. Anupindi, A high-order compact finite-difference lattice Boltzmann method for simulation of natural convection, Comput Fluids 181 (2019) 259–282.
  30. [30]H. R. Khakrah, P. Hooshmand, D. Ross, M. Jamshidian, Numerical analysis of free convection and entropy generation in a cavity using compact finite-difference lattice Boltzmann method, Int. J. Numer. Methods Heat Fluid Flow 30 (2) (2019) 977–995.
  31. [31]M. Sakthivel, A. Kameswararao, Asymmetric compact finite-difference lattice Boltzmann method for blood flow simulations, Phys. Rev. E 100 (4) (2019) 043307.
  32. [32]L. Kim, H. Yang, M. Ha, Z. Xu, H. Xiao, S. Lyu, Immersed boundary-finite difference lattice Boltzmann method using the feedback forcing scheme to simulate the incompressible flows, Int. J. Precis. Eng. Manuf. 17 (8) (2016) 1049–1057.
  33. [33]C. V. Krivichevich, S. A. Mikheev, Stability of three-layer finite difference-based lattice Boltzmann schemes, Vychisl. Metody Programm. 15 (2014) 211–221.
  34. [34]M. El-Amin, S. Sun, A. Salama, On the stability of the finite difference based lattice Boltzmann method, Proc. Comput. Sci. 18 (2013) 2101–2108.
  35. [35]C. V. Krivichev, S. A. Mikheev, Stability study of finite-difference-based lattice Boltzmann schemes with upwind differences of high order approximation, Vychisl. Metody Programm. 16 (2015) 196–204.
  36. [36]X. Chen, Z. Chai, Y. Zhao, B. Shi, Numerical simulation of power-law fluid flow in a trapezoidal cavity using the incompressible finite-difference lattice Boltzmann method (2023). arXiv:2306.07603.arxiv.org/abs/2306.07603
  37. [37]Z. Chai, B. Shi, C. Zhan, Multiple-distribution-function lattice Boltzmann method for convection-diffusion-system-based incompressible navier-stokes equations, Phys. Rev. E 106 (2022) 055305.
  38. [38]U. Ghia, K. N. Ghia, C. T. Shin, High-Re solutions for incompressible flow using the Navier-Stokes equations and multigrid method, J. Comput. Phys. 48 (1982) 387–411.

Paper details

vlnc:W_368gfe3v
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