betaFeedback

WorksMathematics

Development of A Hermite Weighted Compact Nonlinear Scheme based on the Two-Stage Fourth-Order Temporal Accurate Framework

Abstract

Improved five-point low dissipation nonlinear schemes are proposed in this paper within the framework of weighted compact nonlinear schemes (WCNSs) \cite{Deng2000}. Particularly we follow the work of Li and Du \cite{Li2016} on the two-stage fourth-order temporal accurate discretization scheme, which is developed based on the Lax-Wendroff method.

Introduction

Fundamentals of the numerical methods

Consider the time-dependent hyperbolic conservation law in one dimension together with its initial condition, given by

∂u∂t+∂f(u)∂x+∂g(u)∂y=0,(x,y)∈R×R,t>0,\frac{\partial\mathbf{u}}{\partial t}+\frac{\partial\mathbf{f}(\mathbf{u})}{\partial x}+\frac{\partial\mathbf{g}(\mathbf{u})}{\partial y}=0,\quad(x,y)\in\mathbb{R}\times\mathbb{R},\quad t>0,
u(x,0)=u0(x),x∈R,(1)\mathbf{u}(x,0)=\mathbf{u}_0(x),\quad x\in\mathbb{R}, \tag*{(1)}

where u\mathbf{u} denotes the vector of conservative variables, and f(u)\mathbf{f}(\mathbf{u}) the vector of flux terms.

(∂u∂t)i+1/2,j=−A+(ui+1/2,jL)(∂u∂x)i+1/2,jL−RI+L(∂g(ui+1/2,jL)∂y)i+1/2,j\left(\frac{\partial\mathbf{u}}{\partial t}\right)_{i+1/2,j} =-\mathbf{A}^{+}(\mathbf{u}_{i+1/2,j}^{L})\left(\frac{\partial\mathbf{u}}{\partial x}\right)_{i+1/2,j}^{L} -\mathbf{R}\mathbf{I}^{+}\mathbf{L}\left(\frac{\partial\mathbf{g}(\mathbf{u}_{i+1/2,j}^{L})}{\partial y}\right)_{i+1/2,j}
−A−(ui+1/2,jR)(∂u∂x)i+1/2,jR−RI−L(∂g(ui+1/2,jR)∂y)i+1/2,j.(2)-\mathbf{A}^{-}(\mathbf{u}_{i+1/2,j}^{R})\left(\frac{\partial\mathbf{u}}{\partial x}\right)_{i+1/2,j}^{R} -\mathbf{R}\mathbf{I}^{-}\mathbf{L}\left(\frac{\partial\mathbf{g}(\mathbf{u}_{i+1/2,j}^{R})}{\partial y}\right)_{i+1/2,j}. \tag*{(2)}
Ωij≡(xi−1/2,xi+1/2)×(yi−1/2,yi+1/2)(3)\Omega_{ij}\equiv(x_{i-1/2},x_{i+1/2})\times(y_{i-1/2},y_{i+1/2}) \tag*{(3)}
f(u)=1Δx∫x−Δx/2x+Δx/2h(ξ,y,t) dξ(4)\mathbf{f}(\mathbf{u})=\frac{1}{\Delta x}\int_{x-\Delta x/2}^{x+\Delta x/2}\mathbf{h}(\xi,y,t)\,\mathrm{d}\xi \tag*{(4)}
g(u)=1Δy∫y−Δy/2y+Δy/2h(x,ξ,t) dξ(5)\mathbf{g}(\mathbf{u}) = \frac{1}{\Delta y}\int_{y-\Delta y/2}^{y+\Delta y/2} \mathbf{h}(x,\xi,t)\,\mathrm{d}\xi \tag*{(5)}
ddtui(t)=R(ui)=−hi+1/2,j−hi−1/2,jΔx−hi,j+1/2−hi,j−1/2Δy(6)\frac{\mathrm{d}}{\mathrm{d}t}\mathbf{u}_i(t) = \mathcal{R}(\mathbf{u}_i) = -\frac{\mathbf{h}_{i+1/2,j}-\mathbf{h}_{i-1/2,j}}{\Delta x}-\frac{\mathbf{h}_{i,j+1/2}-\mathbf{h}_{i,j-1/2}}{\Delta y} \tag*{(6)}
ddt(ddtui(t))=∂∂tR(ui)=−hi+1/2,j′−hi−1/2,j′Δx−hi,j+1/2′−hi,j−1/2′Δy(7)\frac{\mathrm{d}}{\mathrm{d}t}\left(\frac{\mathrm{d}}{\mathrm{d}t}\mathbf{u}_i(t)\right) = \frac{\partial}{\partial t}\mathcal{R}(\mathbf{u}_i) = -\frac{\mathbf{h}'_{i+1/2,j}-\mathbf{h}'_{i-1/2,j}}{\Delta x}-\frac{\mathbf{h}'_{i,j+1/2}-\mathbf{h}'_{i,j-1/2}}{\Delta y} \tag*{(7)}
hi+1/2,j′≈3640(∂f∂t)i−3/2,j−29480(∂f∂t)i−1/2,j+1067960(∂f∂t)i+1/2,j−29480(∂f∂t)i+3/2,j+3640(∂f∂t)i+5/2,j(8)\mathbf{h}'_{i+1/2,j} \approx\frac{3}{640}\left(\frac{\partial\mathbf{f}}{\partial t}\right)_{i-3/2,j}-\frac{29}{480}\left(\frac{\partial\mathbf{f}}{\partial t}\right)_{i-1/2,j}+\frac{1067}{960}\left(\frac{\partial\mathbf{f}}{\partial t}\right)_{i+1/2,j}-\frac{29}{480}\left(\frac{\partial\mathbf{f}}{\partial t}\right)_{i+3/2,j}+\frac{3}{640}\left(\frac{\partial\mathbf{f}}{\partial t}\right)_{i+5/2,j} \tag*{(8)}
(∂f∂t)i+1/2,j=(∂f∂u(ui+1/2,j))(∂u∂t)i+1/2,j(9)\left(\frac{\partial\mathbf{f}}{\partial t}\right)_{i+1/2,j}=\left(\frac{\partial\mathbf{f}}{\partial\mathbf{u}}(\mathbf{u}_{i+1/2,j})\right)\left(\frac{\partial\mathbf{u}}{\partial t}\right)_{i+1/2,j} \tag*{(9)}

Spatial discretization of Eq. (1) is performed on an equally-spaced grid with distance between two adjacent grid nodes denoted by hh. At each node ii, we define xi=ihx_i=ih, i=1,⋯ ,Ni=1,\cdots,N, and ui=u(xi,t)\mathbf{u}_i=\mathbf{u}(x_i,t). The midpoints associated to the node ii are defined as xi±12=xi±h2x_{i\pm\frac{1}{2}}=x_i\pm\frac{h}{2}, which indicate the cell interfaces across which the fluxes are evaluated.

Evaluation of Eq. (1) at each node yields a system of ordinary differentiation equations, given in the form of

dui(t)dt=Ri(u)=−∂f(u)∂x∣x=xi,i=1,⋯ ,N.(10)\frac{\mathrm{d}\mathbf{u}_i(t)}{\mathrm{d}t}=\mathcal{R}_i(\mathbf{u})=-\left.\frac{\partial\mathbf{f}(\mathbf{u})}{\partial x}\right|_{x=x_i},\qquad i=1,\cdots,N. \tag*{(10)}

1

where we define Ri(u)\mathcal{R}_i(\mathbf{u}) as the spacial operator.

A semi-discrete solution in the conservative finite difference form can be further obtained as

dui(t)dt=−hi+1/2−hi−1/2h,(11)\frac{\mathrm{d}\mathbf{u}_i(t)}{\mathrm{d}t}=-\frac{\mathbf{h}_{i+1/2}-\mathbf{h}_{i-1/2}}{h}, \tag*{(11)}

where the primitive function or the often so-called numerical flux function h(x,t)\mathbf{h}(x,t) is implicitly defined by

f(u)=1h∫x−h/2x+h/2h(ξ,t) dξ.(12)\mathbf{f}(\mathbf{u})=\frac{1}{h}\int_{x-h/2}^{x+h/2}\mathbf{h}(\xi,t)\,\mathrm{d}\xi. \tag*{(12)}

In the framework of weighted compact nonlinear schemes (WCNSs) [1], the flux hi+1/2\mathbf{h}_{i+1/2} is numerically approximated using the linear combination of mid-point flux terms. The approximated flux, denoted by h^i+1/2\widehat{\mathbf{h}}_{i+1/2}, is evaluated by

h^i+1/2=3640f^i−3/2−29480f^i−1/2+1067960f^i+1/2−29480f^i+3/2+3640f^i+5/2,(13)\widehat{\mathbf{h}}_{i+1/2}=\frac{3}{640}\widehat{\mathbf{f}}_{i-3/2}-\frac{29}{480}\widehat{\mathbf{f}}_{i-1/2}+\frac{1067}{960}\widehat{\mathbf{f}}_{i+1/2}-\frac{29}{480}\widehat{\mathbf{f}}_{i+3/2}+\frac{3}{640}\widehat{\mathbf{f}}_{i+5/2}, \tag*{(13)}

and Taylor series expansion of h^i+1/2\widehat{\mathbf{h}}_{i+1/2} indicates that

h^i+1/2=hi+1/2+O(h5).(14)\widehat{\mathbf{h}}_{i+1/2}=\mathbf{h}_{i+1/2}+O\left(h^5\right). \tag*{(14)}

The unknown mid-point flux terms on the right hand side of Eq. (13) are computed using numerical upwind flux functions, which are given in a generic form

f^i+12=12[(f(uR,i+12)+f(uL,i+12))−∣A^i+12∣(uR,i+12−uL,i+12)],(15)\widehat{\mathbf{f}}_{i+\frac{1}{2}}=\frac{1}{2}\left[\left(\mathbf{f}(\mathbf{u}_{R,i+\frac{1}{2}})+\mathbf{f}(\mathbf{u}_{L,i+\frac{1}{2}})\right)-\left|\widehat{\mathbf{A}}_{i+\frac{1}{2}}\right|\left(\mathbf{u}_{R,i+\frac{1}{2}}-\mathbf{u}_{L,i+\frac{1}{2}}\right)\right], \tag*{(15)}

where the high-order interpolated flow variables on the left- and right-hand sides of the mid-point xi+12x_{i+\frac{1}{2}} are denoted by the subscripts, LL and RR, respectively, and A^i+12\widehat{\mathbf{A}}_{i+\frac{1}{2}} denotes the approximate Jacobian matrix of the flux function with respect to the conservative variables, i.e., A^i+12≈∂f∂u(ui+12n)\widehat{\mathbf{A}}_{i+\frac{1}{2}} \approx\frac{\partial\mathbf{f}}{\partial\mathbf{u}}\left(\mathbf{u}_{i+\frac{1}{2}}^n\right).

Before proceeding to the section 2.3 for the detailed discussion of nonlinear interpolation of aforementioned flow variables in Eq. (15), we first review the time marching algorithm used to integrate Eq. (10), since it is the core building block of the numerical methods in the proposed work. Instead of using the traditional third-order strongly stable Runge–Kutta method [2], we follow the work of Li and Du [3] on the two-stage fourth-order temporal accurate discretization scheme, which is developed based on the Lax-Wendroff method.

The basic idea in our reconstruction is to use only rr stencils to reconstruct the point-wise values of solutions and spatial derivatives for the 2r−12r-1-order ADER scheme in one dimension, while in two dimensions, the dimension-by-dimension sub-cell reconstruction approach for spatial derivatives is employed.

The hyperbolic conservation laws may develop discontinuities in its solution even if the initial conditions are smooth.

An important advantage or feature of the WCNS method is that the variable interpolation is performed on the primitive/conservative/characteristic variables from the solution points to the ux points.

Two-stage fourth-order temporal accurate scheme

Given the flow state at time step nn, the two-stage temporal scheme performed to integrate Eq. (10) to the next time level n+1n+1 is summarized as follows.

Step 1. Define a vector of intermediate node values uin+12\mathbf{u}_i^{n+\frac{1}{2}}, which are calculated by

uin+12=uin−k2h[h^i+12∗−h^i−12∗],h^i+12∗=h^i+12n+k4(∂h^∂t)i+12n.(16)\begin{aligned} \mathbf{u}_i^{n+\frac{1}{2}} &= \mathbf{u}_i^n-\frac{k}{2h}\left[\widehat{\mathbf{h}}_{i+\frac{1}{2}}^*-\widehat{\mathbf{h}}_{i-\frac{1}{2}}^*\right],\\ \widehat{\mathbf{h}}_{i+\frac{1}{2}}^* &= \widehat{\mathbf{h}}_{i+\frac{1}{2}}^n+\frac{k}{4}\left(\frac{\partial\widehat{\mathbf{h}}}{\partial t}\right)_{i+\frac{1}{2}}^n. \tag*{(16)} \end{aligned}

where kk is the time step size.

Step 2. Advance the solution to the time state tn+1=tn+kt^{n+1}=t^n+k by

uin+1=uin−kh[h^i+124th−h^i−124th],h^i+124th=h^i+12n+k6(∂h^∂t)i+12n+k3(∂h^∂t)i+12n+12.(17)\begin{aligned} \mathbf{u}_i^{n+1} &= \mathbf{u}_i^n-\frac{k}{h}\left[\widehat{\mathbf{h}}_{i+\frac{1}{2}}^{4th}-\widehat{\mathbf{h}}_{i-\frac{1}{2}}^{4th}\right],\\ \widehat{\mathbf{h}}_{i+\frac{1}{2}}^{4th} &= \widehat{\mathbf{h}}_{i+\frac{1}{2}}^n+\frac{k}{6}\left(\frac{\partial\widehat{\mathbf{h}}}{\partial t}\right)_{i+\frac{1}{2}}^n+\frac{k}{3}\left(\frac{\partial\widehat{\mathbf{h}}}{\partial t}\right)_{i+\frac{1}{2}}^{n+\frac{1}{2}}. \tag*{(17)} \end{aligned}

Remark 1. In Step 1, the flux h^i+12n\hat{h}_{i+\frac{1}{2}}^n can be readily determined from Eq. (13), and its derivative with respect to time is also approximated using Eq. (13), such that

(∂h^∂t)i+12n=3640(∂f^∂t)i−32n−29480(∂f^∂t)i−12n+1067960(∂f^∂t)i+12n−29480(∂f^∂t)i+32n+3640(∂f^∂t)i+52n.(18)\left(\frac{\partial\hat{h}}{\partial t}\right)_{i+\frac{1}{2}}^n = \frac{3}{640}\left(\frac{\partial\hat{f}}{\partial t}\right)_{i-\frac{3}{2}}^n - \frac{29}{480}\left(\frac{\partial\hat{f}}{\partial t}\right)_{i-\frac{1}{2}}^n + \frac{1067}{960}\left(\frac{\partial\hat{f}}{\partial t}\right)_{i+\frac{1}{2}}^n - \frac{29}{480}\left(\frac{\partial\hat{f}}{\partial t}\right)_{i+\frac{3}{2}}^n + \frac{3}{640}\left(\frac{\partial\hat{f}}{\partial t}\right)_{i+\frac{5}{2}}^n . \tag*{(18)}

The flux derivative (∂h^∂t)i+12n+12\left(\frac{\partial\hat{h}}{\partial t}\right)_{i+\frac{1}{2}}^{n+\frac{1}{2}} in Step 2 can be obtained in the same way but at the intermediate time level. We discuss the evaluation of the variable (∂f^∂t)i+12n\left(\frac{\partial\hat{f}}{\partial t}\right)_{i+\frac{1}{2}}^n in the following subsection.

Generalized Riemann problem solver

While the flux term f^i+12\hat{f}_{i+\frac{1}{2}} can be computed by a general Riemann solver, for instance, the Rusanov scheme [4], or the low dissipation hybrid Rusanov-Roe scheme [5], the evaluation of (∂f^∂t)i+12\left(\frac{\partial\hat{f}}{\partial t}\right)_{i+\frac{1}{2}} plays a key role, which is conducted following the Cauchy-Kovalevskaya procedure, such that

(∂f^∂t)i+12=(∂f^∂u(ui+12n))(∂u∂t)i+12=A^i+12(∂u∂t)i+12,(19)\left(\frac{\partial\hat{f}}{\partial t}\right)_{i+\frac{1}{2}} = \left(\frac{\partial\hat{f}}{\partial u}\left(u_{i+\frac{1}{2}}^n\right)\right)\left(\frac{\partial u}{\partial t}\right)_{i+\frac{1}{2}} = \hat{A}_{i+\frac{1}{2}}\left(\frac{\partial u}{\partial t}\right)_{i+\frac{1}{2}}, \tag*{(19)}

and the Generalized Riemann problem solver is used for the term (∂u∂t)i+12\left(\frac{\partial u}{\partial t}\right)_{i+\frac{1}{2}}, given by

(∂u∂t)i+12=−A^i+12+ΔuL,i+12−A^i+12−ΔuR,i+12,(20)\left(\frac{\partial u}{\partial t}\right)_{i+\frac{1}{2}} = -\hat{A}_{i+\frac{1}{2}}^{+}\Delta u_{L,i+\frac{1}{2}}-\hat{A}_{i+\frac{1}{2}}^{-}\Delta u_{R,i+\frac{1}{2}}, \tag*{(20)}

where the operator Δ\Delta denotes the spacial derivative, such that ΔuL/R,i+12=(∂u∂x)L/R,i+12\Delta u_{L/R,i+\frac{1}{2}}=\left(\frac{\partial u}{\partial x}\right)_{L/R,i+\frac{1}{2}}. The matrix A^i+12\hat{A}_{i+\frac{1}{2}} can be diagonized in the form of

A^i+12=RΛL,(21)\hat{A}_{i+\frac{1}{2}}=\mathcal{R}\Lambda\mathcal{L}, \tag*{(21)}

where R\mathcal{R}, and L\mathcal{L} denotes the right and left eigenmatrix, respectively, and Λ\Lambda is the diagonal matrix consisting of eigenvalues λk\lambda_k. Then two unknown matrices in Eq. (20) are defined as

A^i+12±=RΛ±L,(22)\hat{A}_{i+\frac{1}{2}}^{\pm}=\mathcal{R}\Lambda^{\pm}\mathcal{L}, \tag*{(22)}

where Λ+\Lambda^{+} consists of eigenvalues λk+=∣λk∣\lambda_k^{+}=|\lambda_k|, and Λ−\Lambda^{-} of eigenvalues λk−=−∣λk∣\lambda_k^{-}=-|\lambda_k|.

As aforementioned, we will focus on the high-order interpolation of flow variables uL/R,i+12u_{L/R,i+\frac{1}{2}} in section 2.3, and its first spacial derivative ΔuL/R,i+12\Delta u_{L/R,i+\frac{1}{2}} in section 2.4.

Nonlinear interpolation of flow variables based on the Hermite polynomial

In the proposed work we employ the Hermite polynomial instead of the Lagrangian approach for the higher-order approximation of the flow variables at the mid-point. The Hermite polynomial enjoys the advantage of its compact stencil, since it has been demonstrated that only half of the grid points are required to derive the polynomial of the same degree when compared to the Lagrangian approach. In the following work, we mainly consider the evaluation of variables on the left-hand side of xi+12x_{i+\frac{1}{2}}, i.e., uL,i+12n{u}^{n}_{L,i+\frac{1}{2}} for the purpose of simplicity, and uR,i+12n{u}^{n}_{R,i+\frac{1}{2}} is readily obtained from a symmetrical form of uL,i+12{u}_{L,i+\frac{1}{2}}. 45 We further drop the subscript LL and the supscript nn for the rest of the discussion, unless noted otherwise.

As shown in Fig. ??, given the node value ui\mathbf{u}_i, and its spatial derivative Δui\Delta\mathbf{u}_i, only a three-point full stencil, which is denoted by Si+12={xi−1,xi,xi+1}S_{i+\frac{1}{2}}=\{x_{i-1},x_i,x_{i+1}\}, is necessary for the construction of a fourth-degree Hermite polynomial p(x)\mathbf{p}(x), such that ui+12=p(xi+12)\mathbf{u}_{i+\frac{1}{2}}=\mathbf{p}\left(x_{i+\frac{1}{2}}\right). It can be determined by satisfying

p(xk)=uk,k=i−1,i,i+1,(23)\mathbf{p}(x_k)=\mathbf{u}_k,\quad k=i-1,i,i+1, \tag*{(23)}

Then we can obtain a fifth-order spacial accurate interpolation for ui+12\mathbf{u}_{i+\frac{1}{2}}, which takes the form of

ui+12=p(xi+12)=−18ui−1+916ui+916ui+1−3h64(Δui−1+3Δui+1).(24)\mathbf{u}_{i+\frac{1}{2}}=\mathbf{p}\left(x_{i+\frac{1}{2}}\right)=-\frac{1}{8}\mathbf{u}_{i-1}+\frac{9}{16}\mathbf{u}_i+\frac{9}{16}\mathbf{u}_{i+1}-\frac{3h}{64}\left(\Delta\mathbf{u}_{i-1}+3\Delta\mathbf{u}_{i+1}\right). \tag*{(24)}

On the other hand, the full stencil can be split into three sub-stencils, namely, Si+12,1={xi−1,xi}S_{i+\frac{1}{2},1}=\{x_{i-1},x_i\}, Si+12,2={xi,xi+1}S_{i+\frac{1}{2},2}=\{x_i,x_{i+1}\}, and Si+12,3={xi−1,xi,xi+1}S_{i+\frac{1}{2},3}=\{x_{i-1},x_i,x_{i+1}\}. Over each sub-stencil a specific Hermite polynomial can be constructed, which is subject to the following condition

p1(xk)=uk,k=i−1,i,Δp1(xi−1)=Δui−1p2(xk)=uk,k=i,i+1,Δp2(xi+1)=Δui+1,p3(xk)=uk,k=i−1,i,i+1(25)\begin{aligned} \mathbf{p}_1(x_k)&=\mathbf{u}_k,\quad k=i-1,i,\quad\Delta\mathbf{p}_1(x_{i-1})=\Delta\mathbf{u}_{i-1}\\ \mathbf{p}_2(x_k)&=\mathbf{u}_k,\quad k=i,i+1,\quad\Delta\mathbf{p}_2(x_{i+1})=\Delta\mathbf{u}_{i+1},\\ \mathbf{p}_3(x_k)&=\mathbf{u}_k,\quad k=i-1,i,i+1 \tag*{(25)} \end{aligned}

respectively. Once these three quadratic polynomials are determined, their corresponding interpolated mid-point flow variables are obtained as

ui+12,1=p1(xi+12)=−54ui−1+94ui−3h4Δui−1ui+12,2=p2(xi+12)=14ui+34ui+1−h4Δui+1ui+12,3=p3(xi+12)=−18ui−1+34ui+38ui+1(26)\begin{aligned} \mathbf{u}_{i+\frac{1}{2},1}&=\mathbf{p}_1\left(x_{i+\frac{1}{2}}\right)=-\frac{5}{4}\mathbf{u}_{i-1}+\frac{9}{4}\mathbf{u}_i-\frac{3h}{4}\Delta\mathbf{u}_{i-1}\\ \mathbf{u}_{i+\frac{1}{2},2}&=\mathbf{p}_2\left(x_{i+\frac{1}{2}}\right)=\frac{1}{4}\mathbf{u}_i+\frac{3}{4}\mathbf{u}_{i+1}-\frac{h}{4}\Delta\mathbf{u}_{i+1}\\ \mathbf{u}_{i+\frac{1}{2},3}&=\mathbf{p}_3\left(x_{i+\frac{1}{2}}\right)=-\frac{1}{8}\mathbf{u}_{i-1}+\frac{3}{4}\mathbf{u}_i+\frac{3}{8}\mathbf{u}_{i+1} \tag*{(26)} \end{aligned}

Linear combination of ui+12,k\mathbf{u}_{i+\frac{1}{2},k} yields

ui+12=∑k=13dkui+12,k,(27)\mathbf{u}_{i+\frac{1}{2}}=\sum_{k=1}^{3}d_k\mathbf{u}_{i+\frac{1}{2},k}, \tag*{(27)}

where dkd_ks denote the linear weight, having values d1=1/16d_1=1/16, d2=9/16d_2=9/16, and d3=3/8d_3=3/8, respectively.

The interpolations in Eq. (26) can also be expressed in a generic form using (approximated) nthn^{th} derivatives (n=1,2n=1,2) in space, which are derived from the Taylor series, given by

ui+12,k=ui+ui,k(1)h2+ui,k(2)h28,k=1,2,3,(28)\mathbf{u}_{i+\frac{1}{2},k}=\mathbf{u}_i+\mathbf{u}^{(1)}_{i,k}\frac{h}{2}+\mathbf{u}^{(2)}_{i,k}\frac{h^2}{8},\quad k=1,2,3, \tag*{(28)}

where the first- and second-order derivatives are approximated by

ui,0(1)=1h(−2ui−1+2ui−hΔui−1),ui,1(1)=1h(−2ui+2ui+1−hΔui+1),ui,2(1)=12h(−ui−1+ui+1),(29)\begin{aligned} \mathbf{u}^{(1)}_{i,0}&=\frac{1}{h}\left(-2\mathbf{u}_{i-1}+2\mathbf{u}_i-h\Delta\mathbf{u}_{i-1}\right),\\ \mathbf{u}^{(1)}_{i,1}&=\frac{1}{h}\left(-2\mathbf{u}_i+2\mathbf{u}_{i+1}-h\Delta\mathbf{u}_{i+1}\right),\\ \mathbf{u}^{(1)}_{i,2}&=\frac{1}{2h}\left(-\mathbf{u}_{i-1}+\mathbf{u}_{i+1}\right), \tag*{(29)} \end{aligned}

and

ui,0(2)=1h2(−2ui−1+2ui−2hΔui−1),\mathbf{u}_{i,0}^{(2)} = \frac{1}{h^2}\left(-2\mathbf{u}_{i-1}+2\mathbf{u}_i-2h\Delta\mathbf{u}_{i-1}\right),
ui,1(2)=1h2(2ui−2ui+1+2hΔui+1),\mathbf{u}_{i,1}^{(2)} = \frac{1}{h^2}\left(2\mathbf{u}_i-2\mathbf{u}_{i+1}+2h\Delta\mathbf{u}_{i+1}\right),
ui,2(2)=1h2(ui−1−2ui+ui+1),(30)\mathbf{u}_{i,2}^{(2)} = \frac{1}{h^2}\left(\mathbf{u}_{i-1}-2\mathbf{u}_i+\mathbf{u}_{i+1}\right), \tag*{(30)}

respectively. The smooth indicator, βk\beta_k, is defined as [1]

βk=(hui,k(1))2+(h2ui,k(2))2,k=1,2,3.(31)\beta_k = \left(h\mathbf{u}_{i,k}^{(1)}\right)^2+\left(h^2\mathbf{u}_{i,k}^{(2)}\right)^2,\quad k=1,2,3. \tag*{(31)}

Nonlinear weight, ωk\omega_k, is used to replace the linear weight, dkd_k, in Eq. (27) in order to alleviate non-physical oscillations when any sub-stencil is deemed crossed by a discontinuity. For instance, the nonlinear weight of Jiang and Shu [6] can be used

ωk=αk∑k=02αk,αk=dk(βk+ϵ)2,(32)\omega_k=\frac{\alpha_k}{\sum_{k=0}^{2}\alpha_k},\quad\alpha_k=\frac{d_k}{(\beta_k+\epsilon)^2}, \tag*{(32)}

where the small parameter ϵ=10−6\epsilon=10^{-6} is used to prevent division by zero. It suggests that the corresponding JS weight can adaptively approach 0 for a substencil crossed by discontinuities, thus diminishing possible numerical oscillations, and continuously approximate the optimal linear weight in smooth regions, therefore achieving high-order accuracy.

Nonlinear interpolation of the first spacial derivative of the flow variables

As shown in Eq. (20), the GRP solver requires the solution of the first spacial derivative in advance, which is the major focus of this section. Although a set of Hermite polynomials for the flow variables over the full stencil and the three sub-stencils have already been determined in the previous section, we do not recommend to simply differentiate them with respect to space and use the resulting solutions for the derivation of the first spacial derivatives involved; since it leads into one-order of accuracy loss in space, and even more concerning is that the convex combination strategy from the WENO concept cannot be used. A remedy is to construct a new set of Hermite polynomials for the computation of the first spacial derivative [7, 8]. First of all, a fifth-degree Hermite polynomial p(x)\mathbf{p}(x) over the same three-point full stencil, Si+12={xi−1,xi,xi+1}S_{i+\frac{1}{2}}=\{x_{i-1},x_i,x_{i+1}\}, is required to construct. It can be formulated by satisfying the particular conditions

p(xk)=uk,k=i−1,i,i+1,\mathbf{p}(x_k)=\mathbf{u}_k,\quad k=i-1,i,i+1,
Δp(xl)=Δul,l=i−1,i,i+1.(33)\Delta\mathbf{p}(x_l)=\Delta\mathbf{u}_l,\quad l=i-1,i,i+1. \tag*{(33)}

Then a fifth-order spacial accurate interpolation for Δui+12\Delta\mathbf{u}_{i+\frac{1}{2}} can be obtained

Δui+12=Δp(xi+12)=1h(364ui−1−32ui+9364ui+1)+164(Δui−1−12Δui−15Δui+1).(34)\Delta\mathbf{u}_{i+\frac{1}{2}}=\Delta\mathbf{p}\left(x_{i+\frac{1}{2}}\right)=\frac{1}{h}\left(\frac{3}{64}\mathbf{u}_{i-1}-\frac{3}{2}\mathbf{u}_i+\frac{93}{64}\mathbf{u}_{i+1}\right)+\frac{1}{64}\left(\Delta\mathbf{u}_{i-1}-12\Delta\mathbf{u}_i-15\Delta\mathbf{u}_{i+1}\right). \tag*{(34)}

Over each of the three sub-stencils, namely, Si+12,1={xi−1,xi}S_{i+\frac{1}{2},1}=\{x_{i-1},x_i\}, Si+12,2={xi,xi+1}S_{i+\frac{1}{2},2}=\{x_i,x_{i+1}\}, and Si+12,3={xi−1,xi,xi+1}S_{i+\frac{1}{2},3}=\{x_{i-1},x_i,x_{i+1}\}, a specific Hermite polynomial of third degree can be constructed, which is subject to the following condition

p1(xk)=uk,Δp1(xk)=Δuk,k=i−1,i,p2(xk)=uk,Δp2(xk)=Δuk,k=i,i+1,p3(xk)=uk,k=i−1,i,i+1,Δp3(xi)=Δui,(35)\begin{aligned} \mathbf{p}_1(x_k)&=\mathbf{u}_k,\quad\Delta\mathbf{p}_1(x_k)=\Delta\mathbf{u}_k,\quad k=i-1,i,\\ \mathbf{p}_2(x_k)&=\mathbf{u}_k,\quad\Delta\mathbf{p}_2(x_k)=\Delta\mathbf{u}_k,\quad k=i,i+1,\\ \mathbf{p}_3(x_k)&=\mathbf{u}_k,\quad k=i-1,i,i+1,\quad\Delta\mathbf{p}_3(x_i)=\Delta\mathbf{u}_i, \tag*{(35)} \end{aligned}

respectively. The first spacial derivatives at the mid-point can be obtained thereafter

Δui+12,1=Δp1(xi+12)=92h(ui−1−ui)+74Δui−1+154Δui,Δui+12,2=Δp2(xi+12)=32h(−ui+ui+1)−14Δui−14Δui+1,Δui+12,3=Δp3(xi+12)=18h(ui−1−8ui+7ui+1)+14Δui.(36)\begin{aligned} \Delta\mathbf{u}_{i+\frac{1}{2},1}&=\Delta\mathbf{p}_1\left(x_{i+\frac{1}{2}}\right)=\frac{9}{2h}(u_{i-1}-u_i)+\frac{7}{4}\Delta u_{i-1}+\frac{15}{4}\Delta u_i,\\ \Delta\mathbf{u}_{i+\frac{1}{2},2}&=\Delta\mathbf{p}_2\left(x_{i+\frac{1}{2}}\right)=\frac{3}{2h}(-u_i+u_{i+1})-\frac{1}{4}\Delta u_i-\frac{1}{4}\Delta u_{i+1},\\ \Delta\mathbf{u}_{i+\frac{1}{2},3}&=\Delta\mathbf{p}_3\left(x_{i+\frac{1}{2}}\right)=\frac{1}{8h}(u_{i-1}-8u_i+7u_{i+1})+\frac{1}{4}\Delta u_i. \tag*{(36)} \end{aligned}

As is the case for ui+12,k\mathbf{u}_{i+\frac{1}{2},k} in Eq. (27), a convex combination can also be found for Δui+12,k\Delta\mathbf{u}_{i+\frac{1}{2},k}, such that

Δui+12=∑k=13dˉkΔui+12,k,(37)\Delta\mathbf{u}_{i+\frac{1}{2}}=\sum_{k=1}^{3}\bar{d}_k\Delta\mathbf{u}_{i+\frac{1}{2},k}, \tag*{(37)}

where the linear weight dˉk\bar{d}_k is given by dˉ1=1/112\bar{d}_1=1/112, dˉ2=15/16\bar{d}_2=15/16, and dˉ3=3/56\bar{d}_3=3/56, respectively.

The interpolations in Eq. (26) can also be expressed in a generic form using (approximated) nthn^{\mathrm{th}} derivatives (n=2,3n=2,3), given by

Δui+12,k=Δui+ui,k(2)h2+ui,k(3)h28,k=1,2,3,(38)\Delta\mathbf{u}_{i+\frac{1}{2},k}=\Delta\mathbf{u}_i+\mathbf{u}^{(2)}_{i,k}\frac{h}{2}+\mathbf{u}^{(3)}_{i,k}\frac{h^2}{8},\quad k=1,2,3, \tag*{(38)}

where the second- and third-order derivatives are approximated by

ui,0(2)=1h2(6ui−1−6ui+2hΔui−1+4hΔui),ui,1(2)=1h2(−6ui+6ui+1−4hΔui−2hΔui+1),ui,2(2)=1h2(ui−1−2ui+ui+1),(39)\begin{aligned} \mathbf{u}^{(2)}_{i,0}&=\frac{1}{h^2}(6u_{i-1}-6u_i+2h\Delta u_{i-1}+4h\Delta u_i),\\ \mathbf{u}^{(2)}_{i,1}&=\frac{1}{h^2}(-6u_i+6u_{i+1}-4h\Delta u_i-2h\Delta u_{i+1}),\\ \mathbf{u}^{(2)}_{i,2}&=\frac{1}{h^2}(u_{i-1}-2u_i+u_{i+1}), \tag*{(39)} \end{aligned}

and

ui,0(3)=1h3(12ui−1−12ui+6hΔui−1+6hΔui),ui,1(3)=1h3(12ui−12ui+1+6hΔui+6hΔui+1),ui,2(3)=1h3(−3ui−1+3ui+1−6hΔui),(40)\begin{aligned} \mathbf{u}^{(3)}_{i,0}&=\frac{1}{h^3}(12u_{i-1}-12u_i+6h\Delta u_{i-1}+6h\Delta u_i),\\ \mathbf{u}^{(3)}_{i,1}&=\frac{1}{h^3}(12u_i-12u_{i+1}+6h\Delta u_i+6h\Delta u_{i+1}),\\ \mathbf{u}^{(3)}_{i,2}&=\frac{1}{h^3}(-3u_{i-1}+3u_{i+1}-6h\Delta u_i), \tag*{(40)} \end{aligned}

respectively. The smooth indicator, βˉk\bar{\beta}_k, is defined as

βˉk=(h2ui,k(2))2+(h3ui,k(3))2,k=1,2,3.(41)\bar{\beta}_k=\left(h^2\mathbf{u}^{(2)}_{i,k}\right)^2+\left(h^3\mathbf{u}^{(3)}_{i,k}\right)^2,\quad k=1,2,3. \tag*{(41)}

The nonlinear weight of Jiang and Shu [6] are also used in this case, taking the form of

ωˉk=αk∑k=02αk,αk=dk(βˉk+ϵ)2,(42)\bar{\omega}_k=\frac{\alpha_k}{\sum_{k=0}^{2}\alpha_k},\quad\alpha_k=\frac{d_k}{(\bar{\beta}_k+\epsilon)^2}, \tag*{(42)}

where the small parameter ϵ=10−6\epsilon=10^{-6} is used to prevent division by zero.

Δui+12=116h(3ui−1−24ui+21ui+1+hΔui−1−3hΔui+1)(43)\Delta u_{i+\frac{1}{2}}=\frac{1}{16h}\left(3u_{i-1}-24u_i+21u_{i+1}+h\Delta u_{i-1}-3h\Delta u_{i+1}\right) \tag*{(43)}

for the the full stencil, and

Δui+12,1=1h(−3ui−1+3ui−2hΔui−1)Δui+12,2=1h(−ui+ui+1)Δui+12,3=1h(−ui+ui+1)(44)\begin{aligned} \Delta u_{i+\frac{1}{2},1}&=\frac{1}{h}\left(-3u_{i-1}+3u_i-2h\Delta u_{i-1}\right)\\ \Delta u_{i+\frac{1}{2},2}&=\frac{1}{h}\left(-u_i+u_{i+1}\right)\\ \Delta u_{i+\frac{1}{2},3}&=\frac{1}{h}\left(-u_i+u_{i+1}\right) \tag*{(44)} \end{aligned}

Numerical results

Sod and Lax shock tube problems

Riemann initial-value problems of Sod [9] and Lax [10] are used to further evaluate shock-capturing capability of the proposed schemes employing discretization of the 1-D Euler equations

∂ρ∂t+∂(ρu)∂x=0,\frac{\partial\rho}{\partial t}+\frac{\partial(\rho u)}{\partial x}=0,
∂(ρu)∂t+∂(ρu2)∂x=−∂p∂x,\frac{\partial(\rho u)}{\partial t}+\frac{\partial(\rho u^2)}{\partial x}=-\frac{\partial p}{\partial x},
∂E∂t+∂(uE)∂x=−∂(up)∂x,\frac{\partial E}{\partial t}+\frac{\partial(uE)}{\partial x}=-\frac{\partial(up)}{\partial x},

where E=e+12u2E=e+\frac{1}{2}u^2 is the total energy per unit mass, and ee is internal energy. The dependent variables are related through the perfect gas equation of state given by p=(γ−1)ρep=(\gamma-1)\rho e with γ=1.4\gamma=1.4, thus closing the Euler equations system.

The Sod shock tube problem involves a right-moving shock of Mach number 1.7, while for the Lax shock tube problem, the right-moving shock has Mach number 2.0. Initial conditions for the Sod problem are

(ρ,u,p)={(1,0,1)x∈[0,0.5],(0.125,0,0.1)x∈(0.5,1],(45)(\rho,u,p)= \begin{cases} (1,0,1) & x\in[0,0.5],\\ (0.125,0,0.1) & x\in(0.5,1], \end{cases} \tag*{(45)}

and the results at t=0.2t=0.2 are given by solving the problem on an evenly-distributed grid of N=101N=101 points.

Initial conditions for the Lax shock-tube problem are

(ρ,u,p)={(0.445,0.698,3.528)x∈[0,0.5],(0.5,0,0.571)x∈(0.5,1].(46)(\rho,u,p)= \begin{cases} (0.445,0.698,3.528) & x\in[0,0.5],\\ (0.5,0,0.571) & x\in(0.5,1]. \end{cases} \tag*{(46)}

This case is also simulated on an evenly distributed grid of N=101N=101 points, and the results at t=0.14t=0.14 are shown in Figure 2 for density and velocity distributions.

Density

(a)

Velocity

(b)

Figure 2. Lax problem numerical and exact solutions at t=0.14t = 0.14; (a) density, and (b) velocity.

Osher Shu

Density

(a)

Velocity

(b)

Figure 1. Numerical and exact solutions for Sod problem at t=0.2t = 0.2; (a) density, (b) zoom in of density, (c) velocity.

Shock/density-wave interaction problem

Figure 3. Shock/density-wave interaction problem; numerical solutions and the exact solution at t=5t=5; (a) full spatial domain, and (b) zoom in on high-amplitude region.

Two-dimensional Riemann problems

Configuration 6

The equations of motion are the same as those used for configuration 3, the 2-D Euler equations, but now with initial conditions given by

(ρ,u,v,p)={(1,0.75,−0.5,1)(x,y)∈[12,1]×[12,1],(2,0.75,0.5,1)(x,y)∈[0,12]×[12,1],(1,−0.75,0.5,1)(x,y)∈[0,12]×[0,12],(3,−0.75,−0.5,1)(x,y)∈[12,1]×[0,12].(47)(\rho,u,v,p)= \begin{cases} (1,0.75,-0.5,1) & (x,y)\in\left[\frac{1}{2},1\right]\times\left[\frac{1}{2},1\right],\\ (2,0.75,0.5,1) & (x,y)\in\left[0,\frac{1}{2}\right]\times\left[\frac{1}{2},1\right],\\ (1,-0.75,0.5,1) & (x,y)\in\left[0,\frac{1}{2}\right]\times\left[0,\frac{1}{2}\right],\\ (3,-0.75,-0.5,1) & (x,y)\in\left[\frac{1}{2},1\right]\times\left[0,\frac{1}{2}\right]. \end{cases} \tag*{(47)}

Boundary conditions are the same as in the preceding test case.

Configuration 6 of 2-D Riemann problems

Figure 4. Configuration 6 of 2-D Riemann problems in [11] computed on a grid of 1024×10241024 \times1024 points; 40 density contour lines ranging from 0.1 to 2.9 at t=0.3t = 0.3.

Acknowledgments

This work is supported by the open fund from the State Key Laboratory of Aerodynamics (Grant No. SKLA20180302), and the third author is partially supported by the National Natural Science Foundation of China (Grant No. 11872144).

Conclusions

Acknowledgments

This work was financially supported by the National Numerical Wind Tunnel Project (No. NNW2019ZT5-B15) and the National Key Project (No. GJXM92579).

References

  1. [1]X. Deng, H. Zhang, Developing high-order weighted compact nonlinear schemes, Journal of Computational Physics 165 (1) (2000) 22 – 44. doi:10.1006/jcph.2000.6594.DOI
  2. [2]S. Gottlieb, C. Shu, E. Tadmor, Strong stability-preserving high-order time discretization methods, SIAM Review 43 (1) (2001) 89–112. doi:10.1137/S003614450036757X.DOI
  3. [3]J. Li, Z. Du, A two-stage fourth order time-accurate discretization for lax–wendroff type flow solvers in hyperbolic conservation laws, SIAM Journal on Scientific Computing 38 (5) (2016) A3046–A3069. doi:10.1137/15M1052512.DOI
  4. [4]V. V. Rusanov, Calculation of interaction of non-stationary shock waves with obstacles, Zh. Vychisl. Mat. Mat. Fiz. 1 (2) (1961) 267–279.DOI
  5. [5]D. Wang, X. Deng, G. Wang, Y. Dong, Developing a hybrid flux function suitable for hypersonic flow simulation with high-order methods, International Journal for Numerical Methods in Fluids 81 (5) (2016) 309–327. doi:10.1002/fld.4186.DOI
  6. [6]G.-S. Jiang, C.-W. Shu, Efficient implementation of weighted ENO schemes, Journal of Computational Physics 126 (1) (1996) 202–228. doi:10.1006/jcph.1996.0130.DOI
  7. [7]H. Liu, J. Qiu, Finite difference hermite weno schemes for hyperbolic conservation laws, Journal of Scientific Computing 63 (2) (2015) 548–572. doi:10.1007/s10915-014-9905-2. URL https://doi.org/10.1007/s10915-014-9905-2
  8. [8]H. Liu, J. Qiu, Finite difference hermite weno schemes for conservation laws, ii: An alternative approach, Journal of Scientific Computing 66 (2) (2016) 598–624. doi:10.1007/s10915-015-0041-4. URL https://doi.org/10.1007/s10915-015-0041-4
  9. [9]G. A. Sod, A survey of several finite difference methods for systems of nonlinear hyperbolic conservation laws, Journal of Computational Physics 27 (1) (1978) 1–31. doi:10.1016/0021-9991(78)90023-2.DOI
  10. [10]P. D. Lax, Weak solutions of nonlinear hyperbolic equations and their numerical computation, Communications on Pure and Applied Mathematics 7 (1) (1954) 159–193. doi:10.1002/cpa.3160070112.DOI
  11. [11]P. Lax, X. Liu, Solution of two-dimensional riemann problems of gas dynamics by positive schemes, SIAM Journal on Scientific Computing 19 (2) (1998) 319–340. doi:10.1137/S1064827595291819.DOI

Paper details

vlnc:W_y3n6uma4
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

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

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents