Research article Special Issues

The instability of liquid films with temperature-dependent properties flowing down a heated incline

  • Received: 10 October 2018 Accepted: 07 January 2019 Published: 23 December 2019
  • MSC : 74E99

  • We investigate the stability of free-surface flow on a heated incline. We develop a complete mathematical model for the flow which captures the Marangoni effect and also accounts for changes in the properties of the fluid with temperature. We apply a linear stability analysis to determine the stability of the steady and uniform flow. The associated eigenvalue problem is solved numerically by means of a spectral collocation method. Comparisons with nonlinear simulations are also made and the agreement is good.

    Citation: Jean-Paul Pascal, Serge D’Alessio, Syeda Rubaida Zafar. The instability of liquid films with temperature-dependent properties flowing down a heated incline[J]. AIMS Mathematics, 2019, 4(6): 1700-1720. doi: 10.3934/math.2019.6.1700

    Related Papers:

    [1] Andriy A. Avramenko, Igor V. Shevchuk . Renormalization group analysis of heat transfer in the presence of endothermic and exothermic chemical reactions. Mathematical Biosciences and Engineering, 2019, 16(4): 2049-2062. doi: 10.3934/mbe.2019100
    [2] Weidong Zhao, Mobeen Munir, Ghulam Murtaza, Muhammad Athar . Lie symmetries of Benjamin-Ono equation. Mathematical Biosciences and Engineering, 2021, 18(6): 9496-9510. doi: 10.3934/mbe.2021466
    [3] Ivan Tsyfra, Wojciech Rzeszut . On reducing and finding solutions of nonlinear evolutionary equations via generalized symmetry of ordinary differential equations. Mathematical Biosciences and Engineering, 2022, 19(7): 6962-6984. doi: 10.3934/mbe.2022328
    [4] Salvador Chulián, Álvaro Martinez-Rubio, María Luz Gandarias, María Rosa . Lie point symmetries for generalised Fisher's equations describing tumour dynamics. Mathematical Biosciences and Engineering, 2021, 18(4): 3291-3312. doi: 10.3934/mbe.2021164
    [5] Heiko Enderling, Alexander R.A. Anderson, Mark A.J. Chaplain, Glenn W.A. Rowe . Visualisation of the numerical solution of partial differential equation systems in three space dimensions and its importance for mathematical models in biology. Mathematical Biosciences and Engineering, 2006, 3(4): 571-582. doi: 10.3934/mbe.2006.3.571
    [6] Mario Lefebvre . An optimal control problem without control costs. Mathematical Biosciences and Engineering, 2023, 20(3): 5159-5168. doi: 10.3934/mbe.2023239
    [7] Cristian Morales-Rodrigo . A therapy inactivating the tumor angiogenic factors. Mathematical Biosciences and Engineering, 2013, 10(1): 185-198. doi: 10.3934/mbe.2013.10.185
    [8] Avner Friedman, Harsh Vardhan Jain . A partial differential equation model of metastasized prostatic cancer. Mathematical Biosciences and Engineering, 2013, 10(3): 591-608. doi: 10.3934/mbe.2013.10.591
    [9] M. B. A. Mansour . Computation of traveling wave fronts for a nonlinear diffusion-advection model. Mathematical Biosciences and Engineering, 2009, 6(1): 83-91. doi: 10.3934/mbe.2009.6.83
    [10] Xiaobing Feng, Tingao Jiang . Mathematical and numerical analysis for PDE systems modeling intravascular drug release from arterial stents and transport in arterial tissue. Mathematical Biosciences and Engineering, 2024, 21(4): 5634-5657. doi: 10.3934/mbe.2024248
  • We investigate the stability of free-surface flow on a heated incline. We develop a complete mathematical model for the flow which captures the Marangoni effect and also accounts for changes in the properties of the fluid with temperature. We apply a linear stability analysis to determine the stability of the steady and uniform flow. The associated eigenvalue problem is solved numerically by means of a spectral collocation method. Comparisons with nonlinear simulations are also made and the agreement is good.



    The numerical methods for many scientific and engineering problems ultimately lead to the numerical methods for the 2×2 block structure linear systems, see for example, finite element or finite difference methods discretization of some partial differential equations [5,8,9,20], numerical methods for solving weighted least squares problems [23], augmented immersed interface method for Stokes and Darcy or Navier-stokes and Darcy coupling equations [15] and so on. In this paper, we consider the 2×2 block structure linear system arising from mixed finite element discretization Navier-Stokes equation [8], which is called the saddle point problem and has the form

    Ax=[ABB0][xy]=[fg]=b, (1.1)

    where ACn×n is a non-Hermitian positive definite matrix, BCn×m is a matrix with rank(B)=r, here and in the sequence, rank() denotes the rank of a given matrix, fCn and gCm are given vectors, xCn and yCm are unknown vectors, with mn. Note that when r=m, (1.1) is the nonsingular saddle point problem [2] and has a unique solution, and when r<m, (1.1) is the singular saddle point problem, in this case, we suppose that the singular saddle point problem (1.1) is consistent, i.e., brange(A), the range of A.

    Because the block matrices A and B are usually large and sparse, (1.1) is suitable to be solved by the iterative methods. Efficient numerical methods for solving nonsingular and singular saddle point problems have been studied in the literatures, see [13,25] and the references therein. Due to the advantage of the computational efficiency, Uzawa method [1] for solving (1.1) received wide attention and obtained considerable achievements in recent years. The iteration scheme of Uzawa method can be described as

    {xk+1=A1(fByk),yk+1=yk+τ(Bxk+1g),

    where τ is a positive parameter. The main computation costs of Uzawa method lies in the solution of the linear system Ax=fBy at each step, one prefer to approximate its solution by iteration method since the matrix A is always large and sparse.

    Using different iteration methods to approximate xk+1 lead to variant forms of the Uzawa method, see [10,13,16,21,24,26] for examples. In particular, splitting A as

    A=H+S, (1.2)

    where H=12(A+A) and S=12(AA), and approximating xk+1 by the efficient HSS method [4], then the Uzawa-HSS method [21,22] is proposed, the iteration scheme of the Uzawa-HSS method [21,22] is defined as follows

    {xk+1=xk+2α(αI+S)1(αI+H)1(fAxkByk),yk+1=yk+τQ1(Bxk+1g),

    where α and τ are two positive constants, and QCm×m is a given Hermitian positive definite matrix. Approximating xk+1 by the SHSS method [17], we have the Uzawa-SHSS method [16]

    {xk+1=xk+(αI+H)1(fAxkByk),yk+1=yk+τQ1(Bxk+1g).

    Splitting matrix A into its positive definite and skew-Hermitian parts as Ap+As, and approximating xk+1 by the PSS method [3], then the iteration scheme of the Uzawa-PSS method [10] can be defined as

    {xk+1=xk+2α(αI+As)1(αI+Ap)1(fAxkByk),yk+1=yk+τQ1(Bxk+1g),

    where Ap=DH+2LH,As=LHLH+S, DH and LH being the diagonal part and strictly lower triangular part of H. The iteration scheme of the MLHSS method [14] is

    {xk+1=xk+(P+H)1(fAxkByk),yk+1=yk+Q1(Bxk+1g),

    where PCn×n and QCm×m are Hermitian positive definite matrices.

    Recently, Wang et al. [19] proposed the single-step iteration (SSI) method for solving the non-Hermitian positive definite linear systems. In this paper, we will use the SSI method to approximate xk+1 in the first step of the Uzawa method and propose a new Uzawa-type method, named Uzawa-SSI method, to solve the non-Hermitian saddle point problem (1.1).

    The rest of this paper is organized as follows. In Section 2, the Uzawa-SSI method for solving the non-Hermitian saddle point problems (1.1) is proposed. The convergence for nonsingular saddle point problem and semi-convergence for singular case of the Uzawa-SSI method are discussed in Section 3. Numerical examples are given in Section 4 to show the effectiveness of the proposed method for solving non-Hermitian saddle point problems (1.1). Finally, in Section 5, we make a brief conclusion of this paper.

    Let PCn×n be a given Hermitian positive definite matrix, based on the splitting (1.2), the SSI method for solving the non-Hermitian positive linear system Az=c is defined as [19]

    (P+H)xk+1=(PS)xk+c. (2.1)

    Theoretical as well as numerical results in [19] stated that the SSI method is a more efficient method for solving the non-Hermitian positive definite linear systems. With different choices of the matrix P, the SSI method covers several other methods, for example, if P=αI, then the SSI method reduces to the single-step HSS (SHSS) iteration method. The matrix P can also be taken as P=αH,P=Λ (where Λ=diag(d1,d2,,dn),di>0,i=1,2,,n) or other different Hermitian matrices. Note that (1, 1) block A in saddle point matrix A is non-Hermitian positive definite, if we adopt the SSI iteration scheme (2.1) to approximate xk+1, we can derive the following Uzawa-SSI method for solving non-Hermitian saddle point problems (1.1).

    Method 2.1 (THE UZAWA-SSI METHOD) Given initial vectors x0Cn,y0Cm, and positive relaxation parameter τ. For k=0,1,2,, until the iteration sequence converges, compute

    {xk+1=xk+(P+H)1(fAxkByk),yk+1=yk+τQ1(Bxk+1g), (2.2)

    where PCn×n and QCm×m are Hermitian positive definite matrices, τ>0 is a constant.

    The iteration scheme of Uzawa-SSI method (2.2) can be rewritten in matrix-vector form as

    xk+1=Γxk+M1b, (2.3)

    where

    Γ=[P+H0B1τQ]1[PSB01τQ]

    is the iteration matrix of the Uzawa-SSI method, and

    M=[P+H0B1τQ].

    The Uzawa-SSI method possesses similar iteration schemes as the Uzawa-HSS method [21,22], the Uzawa-SHSS method [16] and MLHSS method [14]. In fact, the Uzawa-SSI method can be regarded as a parameterized MLHSS method with the parameter 1τ in the (2, 2) block of the splitting matrix. The Uzawa-SSI method reduces to the MLHSS method when τ=1, we can choose appropriate parameter τ to make the Uzawa-SSI method have better numerical results. The propsed method (2.2) use iteration (2.1) to approximate xk+1, the solution of the shift skew-Hermitian subsystem is avoided compare with the Uzawa-HSS method. Moreover, if we let P=αI, then the Uzawa-SSI method reduces to the Uzawa-SHSS method, we may choose some Hermitian positive definite matrix P instead of αI to improve the computation efficiency of the Uzawa-SSI method.

    In this section, we will study the convergence and semi-convergence properties of the Uzawa-SSI method when it is used to solving the nonsingular and singular non-Hermitian saddle point problems (1.1), respectively. For this purpose, the following notations, definition and results are needed. σ(E) and ρ(E) denote the spectral set and the spectral radius of a square matrix E, respectively. The smallest nonnegative integer i such that rank(Ei) = rank(Ei+1) is called the index of E, and is denoted by index(E). The range and the null spaces of E are denoted by R(E) and N(E), respectively.

    Lemma 3.1. [6] Both roots of the complex quadratic equation λ2ϕλ+ψ=0 have modulus less than one if and only if

    |ϕ¯ϕψ|+|ψ|2<1,

    where ¯ϕ is the conjugate complex number of ϕ.

    Definition 3.1. [7] The iteration method (2.3) is semi-convergent if for any initial guess [x0,y0], the iteration sequence [xk,yk] produced by (2.3) converges to a solution [x,y] of linear systems Ax=b. Moreover, it holds

    [xy]=(IΓ)Dc+(IE)[x0y0],withE=(IΓ)(IΓ)D,

    where I is the identity matrix and (IΓ)D denotes the Drazin inverse of IΓ.

    Following lemma describes the sufficient and necessary semi-convergence conditions of the iteration scheme (2.3).

    Lemma 3.2. [7] The iteration scheme (2.3) is semi-convergent if and only if the following two conditions hold true:

    1. index (IΓ)=1, or equivalently, rank (IΓ)2 = rank (IΓ),

    2. ϑ(Γ)<1, where ϑ(Γ)=max{|λ|,λσ(Γ),λ1}<1 is called the pseudo-spectral radius of the iteration matrix Γ.

    Let λ be an eigenvalue of the Uzawa-SSI iteration matrix Γ and (u,v)Cn+m be the corresponding eigenvector, in terms of the expression of Γ, we have

    [PSB01τQ][uv]=λ[P+H0B1τQ][uv],

    or equivalently,

    {λ(P+H)u=(PS)uBv,λτBu=(λ1)Qv. (3.1)

    For nonsingular saddle point problem (1.1), the Uzawa-SSI method (2.3) is convergent if and only if the spectral radius of the iteration matrix Γ is less than 1, i.e., ρ(Γ)<1.

    For the convergence of Uzawa-SSI, we have the following results.

    Lemma 3.3. Let A be non-Hermitian positive definite and B be of full column rank. If λ is an eigenvalue of iteration matrix Γ, and [u,v] is the corresponding eigenvector with uCn and vCm, then λ1 and u0.

    Proof. We prove the conclusions by contradiction. If λ=1, then it follows from (3.1) that

    {Au+Bv=0,Bu=0.

    It is know that the saddle point matrix

    [ABB0]

    is nonsingular when A is non-Hermitian positive definite and B has full column rank, hence we have u=0 and v=0, which contradicts the assumption that [u,v] is an eigenvector of the iteration matrix Γ, so λ1.

    If u=0, then the first equality in (3.1) reduces to Bv=0. Because B is a matrix of full column rank, we can obtain v=0, which is a contradiction.

    Theorem 3.1. Let ACn×n be non-Hermitian positive definite and BCn×m be of full column rank, PCn×n and QCm×m are Hermitian positive definite. Then the Uzawa-SSI method (2.3) is convergent if and only if the parameter τ satisfy

    t22t212t1<1and0<τ<2t1(t21+2t1t22)t3(t21+t22), (3.2)

    where t1=uHuuPu,it2=uSuuPu,t3=uBQ1BuuPu, here i is the imaginary unit.

    Proof. It follows from Lemma 3.3 that λ1. Using the nonsingularity of A and solving v from the second equality of (3.1), we have

    v=λτλ1Q1Bu.

    Substituting it into the first equality of (3.1) yields

    (PS)uλτλ1BQ1Bu=λ(P+H)u. (3.3)

    From Lemma 3.3, we known that u0. Multiplying u/(uPu) to the both sides of (3.3) from left gives

    u(PS)uuPu+λτ1λuBQ1BuuPu=λu(P+H)uuPu.

    Denote t1=uHuuPu,it2=uSuuPu,t3=uBQ1BuuPu. After simple computation, we get λ2ϕλ+ψ=0, with

    ϕ=2+t1it2τt31+t1,ψ=1it21+t1.

    After some careful calculations, we have

    |ϕ¯ϕψ|+|ψ|2=1+t22+[(2+t1τt3)t1t22]2+(τt2t3)2(1+t1)2.

    Therefore, according to Lemma 3.1, |ϕ¯ϕψ|+|ψ|2<1 if and only if

    {1+t22(1+t1)2<0,[1+t22(1+t1)2]2>[(2+t1τt3)t1t22]2+(τt2t3)2. (3.4)

    Solving (3.4) gives

    t22t212t1<1andτ<2t1(t21+2t1t22)t3(t21+t22).

    From the above discussion, we obtain (3.2).

    Corollary 3.1. Assume the condition in Theorem 3.1 are satisfied. Then the Uzawa-SSI method for solving saddle-point problem (1.1) is convergent if the following conditions hold:

    σ2maxλ2min2λmin<1and0<τ<2λmin(λ2min+2λminσ2max)λmax(Q1BP1B)(λ2min+σ2max), (3.5)

    where λmin is the smallest eigenvalue of matrix ˜H and σmax is the largest singular-value of matrix ˜S with ˜H=P12HP12 and ˜S=P12SP12.

    In particular, when A is Hermitian positive definite, the above conditions become

    0<τ<2λmin+2λmax(Q1BP1B). (3.6)

    Proof. Note that the upper bound of τ in (3.2) can be written as the product of ζ1(t1,t2,t3) and ζ2(t1,t2,t3) with

    ζ1(t1,t2,t3)=2t21t3(t21+t22)andζ2(t1,t2,t3)=t21+2t1t22t1.

    It is easy to see that both ζ1(t1,t2,t3) and ζ2(t1,t2,t3) are monotonic increasing with respect to the variable t1, and both ζ1(t1,t2,t3) and ζ2(t1,t2,t3) are monotonic decreasing with respect to t22.

    Let z=p12u, we have

    t3maxZCnzP12BQ1BP12zzz=λmax[(P12BQ12)(Q12BP12)]=λmax[(Q12BP12)(P12BQ12)]=λmax(Q1BP1B).

    Thus, a lower bound of the product of ζ1(t1,t2,t3) and ζ2(t1,t2,t3) is

    2t1(t21+2t1t22)t3(t21+t22)2λ2min(λ2min+2λminσ2max)λmax(Q1BP1B)λmin(λ2min+σ2max)=2λmin(λ2min+2λminσ2max)λmax(Q1BP1B)(λ2min+σ2max).

    By making use of Theorem 3.1, inequalities (3.5) and (3.6), it can be seen that the Uzawa-SSI iteration method is convergent if τ satisfy (3.5). Specifically, if A is Hermitian positive definite, the condition (3.6) can be directly from (3.5) as H=A and S=0 in this case.

    In this subsection, we will analyze the semi-convergence of the Uzawa-SSI method for solving singular saddle point problem (1.1) when rank(B)=r<m. Actually, we only need to verify that the iteration scheme (2.3) satisfies the two conditions in Lemma 3.2.

    In the first place, considering the condition index (IΓ)=1, we have the following result.

    Theorem 3.2. Let ACn×n be non-Hermitian positive definite and BCn×m be rank deficient. Suppose that P and Q are Hermitian positive definite, and Γ is the iteration matrix of Uzawa-SSI method, parameter τ>0. Then, rank (In+mΓ)2 = rank (In+mΓ).

    Proof. Note that Γ=In+mM1A, hence, rank(In+mΓ)2 = rank(In+mΓ) hold if

    null((M1A)2)=null(M1A).

    Obviously, null((M1A)2) null(M1A) holds, so we only need to prove

    null((M1A)2)null(M1A).

    Let x=[x1,x2]Cn+m, with x1Cn and x2Cm, satisfy (M1A)2x=0, and denote y=M1Ax, then we only need to demonstrate y=0. Let y=[y1,y2]. After simple calculation, we have

    y=[y1y2]=[(P+H)1Ax1+(P+H)1Bx2τQ1B(P+H)1Ax1τQ1Bx1+τQ1B(P+H)1Bx2]. (3.7)

    Since, M1Ay=(M1A)2x=0 and M is invertible, (M1A)2x=0 is equality to Ay=0, i.e.,

    {Ay1+By2=0,By1=0. (3.8)

    Since A is positive definite, from the first equation of (3.8) we can easily get y1=A1By2. Then, substituting this relationship into the second equality of (3.8), we obtain BA1By2=0. Owing to the positive definiteness of the matrix A1, it has By2=0. Taking By2=0 into y1=A1By2, we get y1=0. Therefore, the first equality of (3.7) becomes Bx2=Ax1, using the second equality of (3.7), we have

    y2=τQ1B(P+H)1Ax1τQ1Bx1τQ1B(P+H)1Ax1=τQ1Bx1.

    Furthermore, using By2=0 and τ>0, we have x1BQ1Bx1=0, that is Bx1=0, therefore y2=0. Hence y=[y1,y2]=0.

    In the following, we verify the iteration scheme (2.3) satisfy ϑ(Γ)<1 of Lemma 3.2. Let B=U[Br,0]V be the singular value decomposition of the matrix B, where

    Br=[Σr0]Cn×r,Σr=diag(σ1,σ2,,σr),

    where UCn×n and VCm×m are two unitary matrices and σi(i=1,2,,r) is a singular value of B.

    We partition the matrix V as V=[V1,V2] with V1Cm×r,V2Cm×(mr) and define

    Ω=[U00V].

    It is obvious that Ω is a (n+m)×(n+m) unitary matrix and the iteration matrix Γ is unitary similar to the matrix ˆΓ=ΩΓΩ. Hence, the matrix Γ has the same spectrum with the matrix ˆΓ. Thus, we only need to analyze the pseudo-spectral radius of the matrix ˆΓ now.

    Denoting ˆP=UPU,ˆH=UHU,ˆS=USU, some careful calculations yields

    ˆΓ=[^Γ10ˆLInr], (3.9)

    where

    ^Γ1=[(ˆP+ˆH)1(ˆPˆS)(ˆP+ˆH)1BrτV1Q1V1Br(ˆP+ˆH)1(ˆPˆS)IrτV1Q1V1Br(ˆP+ˆH)1Br]

    and

    ˆL=[τV2Q1V1Br(ˆP+ˆH)1(ˆPˆS),τV2Q1V1Br(ˆP+ˆH)1Br].

    Then, from Eq (3.9), ϑ(ˆΓ)<1 hold if and only if ρ(^Γ1)<1. Note that V1Q1V1 is Hermitian positive definite, comparing with the iteration matrix of Uzawa-SSI, ^Γ1 can be viewed actually as the iteration matrix of (2.3) used for solving nonsingular saddle point problem

    [ˆABrBr0][ˆxˆy]=[ˆfˆg],

    with the preconditioning matrix V1QV1.

    Let ˆλ be the eigenvalue of ^Γ1 and [ˆu,ˆv]Cn+r be the eigenvector of ^Γ1 corresponding to the eigenvalue ˆλ. Denote

    ^t1=ˆuˆHˆuˆuˆPˆu,i^t2=ˆuˆSˆuˆuˆPˆu,^t3=ˆuBrV1Q1V1BrˆuˆuˆPˆu, (3.10)

    where i is the imaginary unit. Then from Theorem 3.1, we can derive the following result.

    Theorem 3.3. Let A be non-Hermitian positive definite, B be of rank deficient, P and Q are Hermitian positive definite. Then the pseudo-spectral radius of matrix Γ is less than 1, i.e., ϑ(Γ)<1 if and only if parameters τ satisfy

    ^t22^t122^t1<1and0<τ<2^t1(^t12^t22+2^t1)^t3(^t12+^t22),

    where ^t1,^t2 and ^t3 are defined in (3.10).

    Corollary 3.2. Assume the conditions in Theorem 3.3 are satisfied. Then, the Uzawa-SSI method for singular saddle-point problem(1.1) is semi-convergent if the parameter τ satisfy:

    σ2maxλ2min2λmin<1and0<τ<2λmin(λ2min+2λminσ2max)λmax(Q1BP1B)(λ2min+σ2max),

    where λmin is the smallest eigenvalue of matrix ˜H and σmax is the largest singular-value of matrix ˜S with ˜H=P12HP12 and ˜S=P12SP12. In particular, if A is Hermitian positive definite, the above conditions are

    0<τ<2λmin+2λmax(Q1BP1B).

    In this section, we will verify the efficiency of the Uzawa-SSI method when it is used to solve nonsingular and singular saddle point problem (1.1). The Uzawa-HSS [21,22], the Uzawa-SHSS [16], the Uzawa-PSS [10] and modified local HSS(MLHSS)[14] methods are compared with the Uzawa-SSI method from the aspects of the number of iteration steps (denoted by ‘IT’) and the elapsed CPU times in seconds (denoted by ‘CPU’).

    In the implementation, for the matrices P and Q of the tested methods, we choose P=αI in the MLHSS method, P=H in the Uzawa-SSI method, Q=diag(BD1B) in the all tested methods, where D=diag(A). We have proved that the convergence conditions in Corollaries 1 and 2 are satisfied for the matrices P=H and Q=diag(BD1B) in the following examples. In addition, all the involved sub-linear system are solved by Cholesky or LU factorization in combination with AMD reordering. The parameters τ for the Uzawa-SSI method, α for the MLHSS method, (α,τ) for the Uzawa-HSS method, Uzawa-SHSS method, and Uzawa-PSS method, are chosen to be the ones resulting in the least iteration step. All the tested iteration methods are started from zero vector and terminated when the current iteration solution xk satisfies

    RES=||bAxk||||b||<106

    or the iteration steps exceed kmax=1500 (results in this case are denoted by ‘-’). In addition, all runs are performed in MATLAB 2016a on a personal computer with Intel(R) Celeron(R) 3205U @ 1.50GHz (8G RAM) Windows 8 system. Using experimentally found optimal parameters in the interval (0,6000]. In actual computations, we choose right-hand-side vector[f,g]R3q2 for nonsingular case and [f,g]R3q2+2 for singular case such that the exact solution of (1.1) is x with all elements 1.

    Example 4.1. [10]. Consider the linearized steady Navier-Stokes equation, i.e., the steady Oseen equation of the following form

    {νu+(w)u+p=f,inΩ.divu=g,inΩ.u=g,onΩ.

    Where ν>0 is the kinematic viscosity (inversely proportional to the Reynolds number), is the Laplacian operation, is the gradient and div is the divergence, the "wind" w is the velocity field obtained from the previous Picard iteration step.

    The test problem is a "leaky" two-dimensional lid-driven cavity problem on the unit square domain. We use the IFISS software package developed by Elman et al. [12] with Q2Q1 mixed finite element on uniform grids to generate linear system corresponding to 16×16,32×32 and 64×64 meshes. The resulting linear system for the discrete solution has the form

    Ax=[ABB0][xy]=[fg]=b, (4.1)

    where ACn×n is the discretization of the diffusion and convection terms, BCn×m is the discrete gradient and B is the negative discrete divergence. In (4.1), A is nonsymmetric positive definite matrix and B is rank deficient, so the A is singular. The test nonsingular problem is constructed by dropping the last row of the matrix B [11]. For each case, we test viscosity value ν=0.1.

    In Table 1, we list the numerical results of the Uzawa-SSI method, the Uzawa-HSS method, the Uzawa-SHSS method, the Uzawa-PSS method and MLHSS method for nonsingular saddle point problems with ν=0.1. From Table 1, we can observe that for ν=0.1, compared with other four methods, the Uzawa-SSI method is the most efficient one, when reaching the stop criterion, it needs the least iteration steps and CPU times.

    Table 1.  Numerical results for nonsingular case of Example 4.1.
    Method Parameters IT CPU RES
    q=16 Uzawa-SSI 0.4 42 0.0469 6.1802e-07
    Uzawa–HSS (0.51, 0.1) 153 0.0938 8.2989e-07
    Uzawa-SHSS (0.042, 0.31) 45 0.0625 9.9330e-07
    Uzawa–PSS (1.672, 1.08) 323 0.1563 9.7113e-07
    MLHSS 0.56 219 0.1094 9.7869e-07
    q=32 Uzawa-SSI 0.37 40 0.0781 7.9691e-07
    Uzawa–HSS (0.32, 0.06) 325 0.9219 8.3458e-07
    Uzawa-SHSS (0.036, 0.31) 84 0.2188 9.7044e-07
    Uzawa–PSS (1.83, 1.06) 992 2.2344 9.4148e-07
    MLHSS 0.58 701 1.5781 8.0074e-07
    q=64 Uzawa-SSI 0.31 41 0.6719 9.3715e-07
    Uzawa–HSS (0.42, 0.08) 1190 23.5 9.3018e-07
    Uzawa-SHSS (0.022, 0.25) 162 2.8438 9.0700e-07
    Uzawa–PSS
    MLHSS

     | Show Table
    DownLoad: CSV

    In Table 2. We list the numerical results of the Uzawa-SSI method, the Uzawa-HSS method, the Uzawa-SHSS method, the Uzawa-PSS method and MLHSS method for singular saddle point problems with ν=0.1. We can observe that Uzawa-SSI method is most efficient one, which use least iteration steps and CPU times than the Uzawa-HSS, the Uzawa-SHSS, the Uzawa-PSS and MLHSS methods to achieve stopping criterion.

    Table 2.  Numerical results for singular case of Example 4.1.
    Method Parameters IT CPU RES
    q=16 Uzawa-SSI 0.39 41 0.0469 9.1261e-07
    Uzawa–HSS (0.51, 0.1) 153 0.1094 8.3426e-07
    Uzawa-SHSS (0.05, 0.32) 46 0.0781 7.9389e-07
    Uzawa–PSS (1.66, 1.08) 322 0.1406 9.8122e-07
    MLHSS 0.56 219 0.0938 7.0314e-07
    q=32 Uzawa-SSI 0.37 40 0.0781 7.6433e-07
    Uzawa–HSS (0.41, 0.1) 369 1.1094 9.0712e-07
    Uzawa-SHSS (0.01, 0.22) 58 0.7969 9.9234e-07
    Uzawa–PSS (1.64, 1.02) 940 1.6094 9.5791e-07
    MLHSS 0.58 701 0.8594 7.9512e-07
    q=64 Uzawa-SSI 0.35 41 0.4844 9.6804e-07
    Uzawa–HSS (0.39, 0.1) 1072 22.0938 9.4074e-07
    Uzawa-SHSS (0.01, 0.2) 89 1.5469 8.7736e-07
    Uzawa–PSS
    MLHSS

     | Show Table
    DownLoad: CSV

    In what follows, we consider two artificially constructed examples.

    Example 4.2. Let us consider the nonsingular saddle-point problem (1.1) has the following coefficient sub-matrices:

    A=[IT+TI00IT+TI]R2q2×2q2,
    B=[IFFI]R2q2×q2,

    and

    T=νh2tridiag(1,2,1)+12htridiag(1,0,1)Rq×q,
    F=1htridiag(1,1,0)Rq×q.

    Here, denotes the Kronecker product, ν is a parameter and h=1q+1 is the discretization meshsize, see [18].

    In Table 3, we report the numerical results for Example 4.2, we list the numerical results of the Uzawa-SSI method, the Uzawa-HSS method, the Uzawa-SHSS method, the Uzawa-PSS method and MLHSS method for nonsingular saddle point problems with ν=1. From Table 3, we can observe that for ν=1, compared with other four methods, the Uzawa-SSI method is the most efficient one, when reaching the stop criterion, it needs the least iteration steps and CPU times.

    Table 3.  Numerical results for Example 4.2 with ν=1.
    Method Parameters IT CPU RES
    q=16 Uzawa-SSI 2.2 40 0.0156 8.7237e-07
    Uzawa-HSS (740, 0.58) 162 0.0781 9.9965e-07
    Uzawa-SHSS (35.5, 1.44) 62 0.0625 9.9653e-07
    Uzawa–PSS (560, 0.84) 126 0.0469 9.9214e-07
    MLHSS 0.1 79 0.0469 9.3330e-07
    q=32 Uzawa-SSI 3.34 44 0.0469 8.5020e-07
    Uzawa-HSS (910, 0.2) 623 1.2031 9.9177e-07
    Uzawa-SHSS (20.2, 1.4) 99 0.0938 9.9719e-07
    Uzawa–PSS (1860, 0.8) 247 0.2969 9.8611e-07
    MLHSS 0.11 123 0.3594 9.9713e-07
    q=64 Uzawa-SSI 4.35 70 0.3750 8.9753e-07
    Uzawa-HSS (4000, 0.2) 1087 11.7344 9.9075e-07
    Uzawa-SHSS (20.2, 1.448) 147 1.1719 9.9804e-07
    Uzawa–PSS
    MLHSS 0.1 189 1.9531 9.7499e-07

     | Show Table
    DownLoad: CSV

    Example 4.3. Let us consider the singular saddle-point problem (1.1) has the following coefficient sub-matrices:

    A=[IT+TI00IT+TI]R2q2×2q2,
    B=[ˆBb1b2]R2q2×(q2+2),

    with

    ˆB=[IFFI]R2q2×q2,b1=ˆBT[e0],b2=ˆBT[0e],
    e=[1,1,,1]Rq2/2,

    and

    T=νh2tridiag(1,2,1)+12htridiag(1,0,1)Rq×q,
    F=1htridiag(1,1,0)Rq×q.

    Here, denotes the Kronecker product, ν is a parameter and h=1q+1 is the discretization meshsize, see [22,27].

    This problem is a technical modification of Example 4.2. Here, matrix B is an augmentation of the full rank matrix ˆB with two linearly independent vectors b1 and b2. As b1 and b2 are linear combinations of the columns of the matrix ˆB, B is a rank-deficient matrix.

    In Table 4, we list the numerical results of the Uzawa-SSI method, the Uzawa-HSS method, the Uzawa-SHSS method, the Uzawa-PSS method and MLHSS method for singular saddle point problems with ν=1. We can observe that Uzawa-SSI method is most efficient one, which use least iteration steps and CPU times than other four methods to achieve stopping criterion.

    Table 4.  Numerical results for Example 4.3 with ν=1.
    Method Parameters IT CPU RES
    q=16 Uzawa-SSI 0.41 40 0.0156 8.0268e-07
    Uzawa–HSS (258, 0.14) 129 0.0625 9.9724e-07
    Uzawa-SHSS (13.4, 0.27) 58 0.0313 9.9910e-07
    Uzawa–PSS (230, 0.14) 146 0.1563 9.4478e-07
    MLHSS
    q=32 Uzawa-SSI 0.295 66 0.0625 8.9038e-07
    Uzawa–HSS (606, 0.093) 247 0.4844 9.9569e-07
    Uzawa-SHSS (35.4, 0.214) 82 0.1094 9.6742e-07
    Uzawa–PSS (510, 0.082) 279 3.3594 9.7925e-07
    MLHSS
    q=64 Uzawa-SSI 0.16 114 0.5156 9.8078e-07
    Uzawa–HSS (484, 0.024) 591 6.2031 9.9879e-07
    Uzawa-SHSS (2.4, 0.06) 120 1.0781 9.9344e-07
    Uzawa–PSS (1020, 0.04) 545 48.6719 9.9242e-07
    MLHSS

     | Show Table
    DownLoad: CSV

    From the numerical results, it can be observed that the Uzawa-SSI method is the best choice as it outperforms other four methods for solving the non-Hermitian nonsingular and singular saddle-point problems.

    In this paper, based on the SSI method for non-Hermitian positive definite linear system, we have proposed the Uzawa-SSI method for solving non-Hermitian nonsingular and singular saddle point problems (1.1). Compared with the Uzawa-HSS, the Uzawa-SHSS, the Uzawa-PSS and the MLHSS methods, the proposed method has more simple convergence conditions, which are easy to be satisfied. Numerical results verified the efficiency of the Uzawa-SSI method.

    However, the Uzawa-SSI method involved a parameter τ. It is formidable to find a optimal parameter τ in the actual calculation, therefore, we did not discuss the choice of the parameter τ in this paper. Consider that the validity of the new method depends on the selection of parameters, how to find a easy calculated parameter should be a direction for future research.

    This work was supported by the National Natural Science Foundation of China (Grant No. 11861059) and the Natural Science Foundation of Northwest Normal University (No. NWNU-LKQN-17-5).

    The authors declare there is no conflict of interest.



    [1] P. L. Kapitza, Wave flow of thin layers of a viscous fluid: II. Fluid flow in the presence of continuous gas flow and heat transfer, Zh. Eksp. Teor. Fiz., 18 (1948), 19-28.
    [2] P. L. Kapitza and S. P. Kapitza, Wave flow of thin layers of a viscous fluid: III. Experimental study of undulatory flow conditions, Zh. Eksp. Teor. Fiz., 19 (1949), 105-120.
    [3] T. B. Benjamin, Wave formation in laminar flow down an inclined plane, J. Fluid Mech., 2 (1957), 554-574. doi: 10.1017/S0022112057000373
    [4] C.-S. Yih, Stability of liquid flow down an inclined plane, Phys. Fluids, 6 (1963), 321-334. doi: 10.1063/1.1706737
    [5] D. J. Benney, Long waves on liquid films, J. Math. and Phys., 45 (1966), 150-155.
    [6] B. Gjevik, Occurrence of Finite-Amplitude Surface Waves on Falling Liquid Films, Phys. Fluids, 13 (1970), 1918-1925.
    [7] S. Kalliadasis, E. A. Demekhin, C. Ruyer-Quil, et al. Falling liquid films, Springer-Verlag, London, 2012.
    [8] S. Kalliadasis, E. A. Demekhin, C. Ruyer-Quil, et al. Thermocapillary instability and wave formation on a film flowing down a uniformly heated plane, J. Fluid Mech., 492 (2003), 303-338. doi: 10.1017/S0022112003005809
    [9] C. Ruyer-Quil, B. Scheid, S. Kalliadasis, et al. Thermocapillary long waves in a liquid film flow. part 1. low-dimensional formulation, J. Fluid Mech., 538 (2005), 199-222.
    [10] B. Scheid, C. Ruyer-Quil, S. Kalliadasis, et al. Thermocapillary long waves in a liquid film flow. part 2. linear stability and nonlinear waves, J. Fluid Mech., 538 (2005), 223-244.
    [11] P. Trevelyan, B. Scheid, C. Ruyer-Quil, et al. Heated falling films, J. Fluid Mech., 592 (2007), 295-334.
    [12] S. Lin, Stability of a liquid flow down a heated inclined plane, Lett. Heat Mass Transfer, 2 (1975), 361-369. doi: 10.1016/0094-4548(75)90002-8
    [13] Y. O. Kabova and V. V. Kuznetsov, Downward flow of a nonisothermal thin liquid film with variable viscosity, J. Appl. Mech. Tech. Phys., 43 (2002), 895-901.
    [14] D. A. Goussis and R. A. Kelly, Effects of viscosity on the stability of film flow down heated or cooled inclined surfaces: Long-wavelength analysis, Phys. Fluids, 28 (1985), 3207-3214.
    [15] C.-C. Hwang and C.-I. Weng, Non-linear stability analysis of film flow down a heated or cooled inclined plane with viscosity variation, Int. J. Heat Mass Transfer, 31 (1988), 1775-1784.
    [16] K. A. Ogden, S. J. D. D'Alessio, J.P. Pascal, Gravity-driven flow over heated, porous, wavy surfaces, Phys. Fluids, 23 (2011), 122102.
    [17] J. P. Pascal, N. Gonputh and S. J. D. D'Alessio, Long-wave instability of flow with temperature dependent fluid properties down a heated incline, Intl J. Eng. Sci., 70 (2013), 73-90. doi: 10.1016/j.ijengsci.2013.05.003
    [18] S. J. D. D'Alessio, C. J. M. P. Seth and J. P. Pascal, The effects of variable fluid properties on thin film stability, Phys. Fluids, 26 (2014), 122105.
    [19] M. Chhay, D. Dutykhb, M. Gisclonb, et al. New asymptotic heat transfer model in thin liquid films, Appl. Math. Model., 48 (2017), 844-859.
    [20] J. H. Spurk and N. Aksel, Fluid Mechanics, 2nd edition, Springer-Verlag, Berlin, 2008.
    [21] C. Ruyer-Quil and P. Manneville, Improved modeling of flows down inclined planes, Eur. Phys. J. B, 15 (2000), 357-369.
    [22] C. Ruyer-Quil and P. Manneville, Further accuracy and convergence results on the modeling of flows down inclined planes by weighted-residual approximations, Phys. Fluids, 14 (2002), 170-183.
    [23] N. Gonputh, J. P. Pascal and S. J. D. D'Alessio, Inclined flow of a heated fluid film with temperature dependent fluid properties, In: Proceedings of the 2012 International Conference on Scientific Computing, pp. 247-253, CSREA Press, 2012.
  • Reader Comments
  • © 2019 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(5153) PDF downloads(492) Cited by(3)

Figures and Tables

Figures(9)

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog