betaFeedback

WorksMathematics

Investigation of Shock-Capturing with Bound-Preserving Limiters for the Nonlinearly Stable Flux Reconstruction Method

Abstract

Nonlinearly stable flux reconstruction (NSFR) combines the key properties of provable nonlinear stability with the increased time step from energy-stable flux reconstruction. The NSFR scheme has been successfully applied to unsteady compressible flows. Through the use of a bound-preserving limiter, positivity of thermodynamic quantities is preserved, and this enables the extension of NSFR to hyperbolic conservation laws. We extend the limiter of Zhang and Shu [1] to ensure robustness for the proposed scheme. The limiter is modified to consider the minimum density and pressure at the solution nodes when determining the value to scale the solution. The modifications are thoroughly tested with a suite of test cases. In addition to these modifications, this paper conducts a thorough investigation into the shock-capturing capabilities of the NSFR scheme and the advantages it presents over standard discontinuous Galerkin (DG) methods, where, on select variants of the flux reconstruction (FR) scheme, essentially oscillation-free solutions are demonstrated. Various parameters of the scheme are extensively tested and analyzed through several 1D and 2D compressible Euler tests that verify the high-order accuracy, entropy stability, time step advantage and shock-capturing capabilities of the NSFR scheme. These parameters include the two-point flux, quadrature nodes and the strength of the FR parameter. In addition to investigating the impact of the various two-point fluxes, this paper also presents numerical studies to determine the CFL condition required to maintain positivity for the two-point flux of choice. The investigation yields insightful results for all parameters, with the results pertaining to the type of FR scheme being of special interest. The tests showcase increased robustness, time step advantages and oscillation/overshoot mitigation when employing a stronger FR parameter.

Introduction

The discontinuous Galerkin (DG) method, proposed by Reed and Hill [2], combines both the key properties of finite volume and finite element schemes. As explained in Hesthaven and Warburton [3], the high-order scheme provides stability through a numerical flux function and utilizes high-order shape functions to represent the solution. An attractive extension of the DG approach leads to the computationally efficient and robust Summation-by-Parts (SBP) approach [4, 5, 6, 7, 8, 9, 10, 11]. The SBP property allows for the implementation of split forms, initially proposed in Tadmor [12], and in Gassner [4] for high-order schemes, which ensures entropy stability on extremely coarse meshes for unsteady problems.

Another variation of the DG method is Flux Reconstruction (FR), initially proposed by Huynh [13] and later presented as a class of energy stable flux reconstruction (ESFR) schemes [14, 15, 16]. Through the introduction of correction functions, which are equivalently viewed as a filtered DG correction field [17, 18, 19], ESFR recovers various high-order schemes such as the spectral difference from Liu et al. [20] and Huynh’s lumped-Lobatto scheme [13]. Ranocha et al. [9] was the first paper to merge a collocated DG split form and ESFR, although the authors only proved stability for a collocated DG scheme. For the Euler Equations, Abe et al. [21] took a different approach and were able to demonstrate entropy stability for the specific “g2g_2 lumped-Lobatto” ESFR split form, which corresponds to a collocated DG scheme on Gauss-Lobatto-Legendre (GLL) quadrature.

Recent developments by Cicchino et al. [22, 23] proved that nonlinear stability for ESFR schemes can only be satisfied if the FR correction functions are applied to the nonlinear volume terms. This led to the development of nonlinearly stable flux reconstruction (NSFR), where provable nonlinear stability was accomplished for the general FR case. In Cicchino and Nadarajah [24], the application of NSFR was extended for the Euler equations, and a low-storage, weight-adjusted formulation was applied for compressible flows in curvilinear coordinates.

Physically relevant entropic solutions satisfy the strict maximum principle [25], the TVD property [26], and positivity preservation for hyperbolic conservation laws. Unfortunately, numerically satisfying the second law of thermodynamics through entropy stability does not ensure that the numerical approximation also satisfies these additional properties. Higher-order entropy-stable methods improve the robustness of solvers but don’t address the problems of spurious oscillations or nonphysical density and pressure in the vicinity of a shock. Additionally, two-point fluxes, which are used for FR methods, don’t have an established CFL condition under which they are expected to preserve positivity for density and pressure. Positivity preservation strategies are essential to the success of entropy-stable methods and have been a key focus within the research community. There are a few common strategies for controlling oscillations and preventing nonphysical properties for entropy-stable methods, including artificial viscosity, subcell limiting, and limiter-based strategies.

Artificial viscosity (AV) methods are a classical technique that dates as far back as the 1940s as seen in [27]. The strategy involves introducing an additional term in the PDE to control the amount of dissipation added based on shock detectors that distinguish between smooth regions and shock waves. AV methods have been successfully used to develop shock-capturing methods that are compatible with discretely entropy stable split forms as seen in Wintermeyer et al. [28, 29]. While this is a popular method, there are critical disadvantages to this strategy, such as potentially changing the order of the PDE or time-step limitations introduced as a result of enforcing boundary conditions for the extra terms.

Subcell limiting, introduced by Dumbser et al. [30], is a novel idea wherein the DG cell is broken down into subcells and recomputed within the troubled cells from the old time level using a different and more robust numerical scheme on a subgrid level. Subcell limiting has seen much development in the context of entropy-stable methods wherein the strategy has been modified from a hard switch to the seamless blending of high-order and low-order approaches to eliminate oscillations while maintaining entropy stability as seen in Hennemann et al. [31], Rueda-Ramírez and Gassner [32], Lin and Chan [33], Sonntag and Munz [34], Vilar [35]. While this approach is effective, it has a rigorous implementation compared to other limiting strategies. Several components are involved, such as projecting from the cell to subcells, a detection procedure, recomputation of the solution using a different method for all troubled subcells, and a gathering procedure to establish a cell-centred polynomial on the main grid with the corrected subcells.

The class of bound-preserving limiters established by Zhang and Shu [36] are designed to preserve strict bounds for DGSEM and FV schemes while maintaining the order of accuracy of the original schemes. These limiters are easily extended to multi-dimensions and unstructured meshes. Bound-preserving limiters achieve strict bounds by scaling the values locally in cells that do not fall within the prescribed bounds. By using the bounds defined by the strict maximum principle or positivity preservation properties to determine whether a cell needs to be limited, the need for a troubled cell indicator is eliminated. Additionally, the cost of the limiter is minimal as it only involves the evaluation of the unlimited solution polynomial at a few predetermined quadrature points, and the limiter preserves the original cell average, thereby ensuring conservation. Bound-preserving limiters have been shown to work well with entropy-stable methods by Chen and Shu [37]. The limiters are simple, efficient and don’t rely on a shock detection method, unlike the aforementioned strategies; thus, we choose to incorporate bound-preserving limiters to ensure positivity preservation and enable the implementation of NSFR in a shock-capturing context. Despite all the aforementioned strengths of the positivity-preserving limiter, it still lacks robustness guarantees. To address this, modifications are proposed for the higher dimension (>1>1) implementations of the limiter.

The paper first introduces an advancement of the Zhang-Shu bound-preserving limiter [1] for the current state-of-the-art NSFR schemes; second, it demonstrates the robustness of the approach for a large suite of cases without the need for TVD limiters; third, we numerically demonstrate that the two-point fluxes abide by a CFL condition; and lastly, we show that the scheme provides essentially oscillation-free solutions with select flux reconstruction schemes. In § 2 we outline the DG scheme, the corresponding ESFR scheme, and then present the NSFR scheme for compressible flows. In Sec. 3, we describe the implementation of the positivity-preserving limiter, including modifications made to enhance robustness. In addition to the limiter, Sec. 3 also outlines other critical implementation details including the two-point flux, temporal discretization method, and the flux reconstruction parameter. Lastly, in § 4, we numerically verify that the proposed limiter modifications provide enhanced robustness and that the limiter preserves positivity and maintains a high-order of accuracy. This is achieved through a grid refinement order of accuracy study. We also assess the impact of the choice of two-point flux, quadrature nodes, and flux reconstruction parameters on problems involving strong shocks using various 1D and 2D numerical experiments. In addition to this extensive study, certain test cases are also used to numerically verify and establish a CFL condition for two-point fluxes that preserves positivity for density and pressure. The findings of this assessment are summarized in § 5.

Preliminaries

Nonlinearly Stable Flux Reconstruction Scheme

In this section, the nonlinearly stable FR (NSFR) scheme will be presented. The preliminaries, including the discontinuous Galerkin (DG) formulation and energy stable flux reconstruction (ESFR) scheme, are first discussed, followed by the NSFR discretization scheme used in this investigation. In this paper, row vector notation will be used. We will also use boldface for vectors and italic for scalar values, and the cc-superscript refers to the use of Cartesian coordinates.

DG Formulation

Consider the scalar 3D conservation law,

∂∂tu(xc,t)+∇⋅f(u(xc,t))=0,t≥0,xc:=[x,y,z]∈Ω,u(xc,0)=u0(xc),(1)\begin{aligned} \frac{\partial}{\partial t}u(\mathbf{x}^{c},t)+\nabla\cdot\mathbf{f}(u(\mathbf{x}^{c},t))&=0,\quad t\ge0,\quad\mathbf{x}^{c}:=[x,y,z]\in\Omega,\\ u(\mathbf{x}^{c},0)&=u_{0}(\mathbf{x}^{c}), \tag*{(1)} \end{aligned}

where f(u(xc,t))\mathbf{f}(u(\mathbf{x}^{c},t)) represents the flux. The computational domain, Ωh\Omega^{h} is partitioned into MM non-overlapping elements, denoted by Ωm\Omega_{m},

Ω≃Ωh:=⋃m=1MΩm.(2)\Omega\simeq\Omega^{h}:=\bigcup_{m=1}^{M}\Omega_{m}. \tag*{(2)}

The global solution can then be taken as the direct sum of each approximation within each element,

u(xc,t)≃uh(xc,t)=⨁m=1Mumh(xc,t).(3)u(\mathbf{x}^{c},t)\simeq u^{h}(\mathbf{x}^{c},t)=\bigoplus_{m=1}^{M}u_{m}^{h}(\mathbf{x}^{c},t). \tag*{(3)}

On each element, we represent the solution with NpN_p linearly independent modal or nodal polynomial basis of a maximum order of pp; where, Np=(p+1)dN_p=(p+1)^d, dd is the dimension of the problem. The solution can be represented as

umh(xc,t)=∑i=1Npχm,i(xc)u^m,i(t),(4)u_m^h(\boldsymbol{x}^c,t)=\sum_{i=1}^{N_p}\chi_{m,i}(\boldsymbol{x}^c)\hat{u}_{m,i}(t), \tag*{(4)}

where

χm(xc):=[χm,1(xc) χm,2(xc) … χm,Np(xc)]=χ(x)⊗χ(y)⊗χ(z)(5)\boldsymbol{\chi}_m(\boldsymbol{x}^c):=[\chi_{m,1}(\boldsymbol{x}^c)\ \chi_{m,2}(\boldsymbol{x}^c)\ \ldots\ \chi_{m,N_p}(\boldsymbol{x}^c)]=\boldsymbol{\chi}(x)\otimes\boldsymbol{\chi}(y)\otimes\boldsymbol{\chi}(z) \tag*{(5)}

holds the polynomial basis functions for the element.

The physical coordinates are mapped to the standard reference space, ξr:={[ξ,η,ζ]:−1≤ξ,η,ζ≤1}\boldsymbol{\xi}^r:=\{[\xi,\eta,\zeta]: -1\leq\xi,\eta,\zeta\leq1\}, on each element with the transformation mapping xmc(ξr):=Θm(ξr)=Θm,i(ξr)x^m,ic\boldsymbol{x}_m^c(\boldsymbol{\xi}^r):=\boldsymbol{\Theta}_m(\boldsymbol{\xi}^r)=\Theta_{m,i}(\boldsymbol{\xi}^r)\hat{\boldsymbol{x}}_{m,i}^c, where Θm,i(ξr)\Theta_{m,i}(\boldsymbol{\xi}^r) are the mapping shape functions, and x^m,ic\hat{\boldsymbol{x}}_{m,i}^c are the physical mapping control points. In this work, the reference coordinates, ξr\boldsymbol{\xi}_r, are chosen as Gauss-Lobatto-Legendre (GLL) nodes for the solution bases and either GLL or Gauss-Legendre (GL) nodes for the flux bases. The impact of the choice of flux nodes for shock problems is discussed in § 4.3.

Continuing with the DG formulation, the elementwise reference residual for Eq. (1) is,

Rmh(xc,t)=∂∂tumh(xc,t)+∇⋅f(umh(xc,t)).(6)\boldsymbol{R}_m^h(\boldsymbol{x}^c,t)=\frac{\partial}{\partial t}\boldsymbol{u}_m^h(\boldsymbol{x}^c,t)+\nabla\cdot\boldsymbol{f}(\boldsymbol{u}_m^h(\boldsymbol{x}^c,t)). \tag*{(6)}

Following a DG framework, we integrate the residual over the computational domain, utilizing orthogonality of the residual with the test function, chosen to be the basis function. After integrating by parts twice we arrive at the strong form,

Mmddtu^m(t)T+Sξf^1m(t)T+Sηf^2m(t)T+Sζf^3m(t)T+∑f=1Nf∑k=1Nfpχ(ξf,kr)TWf,k[n^r⋅fmc,r]=0T(7)M_m\frac{d}{dt}\hat{\boldsymbol{u}}_m(t)^T+S_{\xi}\hat{\boldsymbol{f}}_{1m}(t)^T+S_{\eta}\hat{\boldsymbol{f}}_{2m}(t)^T+S_{\zeta}\hat{\boldsymbol{f}}_{3m}(t)^T+\sum_{f=1}^{N_f}\sum_{k=1}^{N_{fp}}\boldsymbol{\chi}(\boldsymbol{\xi}_{f,k}^r)^TW_{f,k}[\hat{\boldsymbol{n}}^r\cdot\boldsymbol{f}_m^{c,r}]=\boldsymbol{0}^T \tag*{(7)}

where ξvr\boldsymbol{\xi}_v^r represents the reference coordinate at the volume quadrature nodes and ξf,kr\boldsymbol{\xi}_{f,k}^r represents the reference coordinate on the face ff at the facet cubature node kk, where k∈[1,Nfp]k\in[1,N_{fp}] with NfpN_{fp} being the number of facet cubature nodes on the surface. The mass and stiffness matrices are defined as,

(Mm)ij≈∫Ωrχi(ξr)χj(ξr)dΩr→Mm=χ(ξvr)TWJmχ(ξvr),(Sξ)ij=∫Ωrχi(ξr)∂∂ξχj(ξr)dΩr→Sξ=χ(ξvr)TWχξ(ξvr),(8)\begin{aligned} (M_m)_{ij}&\approx\int_{\Omega_r}\chi_i(\boldsymbol{\xi}^r)\chi_j(\boldsymbol{\xi}^r)d\Omega_r\to M_m=\boldsymbol{\chi}(\boldsymbol{\xi}_v^r)^TWJ_m\boldsymbol{\chi}(\boldsymbol{\xi}_v^r),\\ (S_{\xi})_{ij}&=\int_{\Omega_r}\chi_i(\boldsymbol{\xi}^r)\frac{\partial}{\partial\xi}\chi_j(\boldsymbol{\xi}^r)d\Omega_r\to S_{\xi}=\boldsymbol{\chi}(\boldsymbol{\xi}_v^r)^TW\boldsymbol{\chi}_{\xi}(\boldsymbol{\xi}_v^r), \tag*{(8)} \end{aligned}

and similarly for the other directions.

Here, w\boldsymbol{w} is a diagonal operator storing the quadrature weights of integration at the volume cubature nodes, and JmJ_m is a diagonal operator storing the determinant of the metric Jacobian evaluated at the volume cubature nodes. We let n^r\hat{\mathbf{n}}^{r} represent the outward pointing normal on the face at the facet cubature node in the reference element. Here, f^mr(t)=Pfmr(t)\hat{\mathbf{f}}_{m}^{r}(t)=P\mathbf{f}_{m}^{r}(t), where P=M−1χ(ξvr)TWP=M^{-1}\boldsymbol{\chi}(\boldsymbol{\xi}_{v}^{r})^{T}W is the L2L_{2} projection operator. That is, the modal reference flux is the modal L2L_{2} projection of the nodal reference flux. Lastly, fmC,r:=(fm∗,r)−(fmr)\mathbf{f}_{m}^{C,r}:=(\mathbf{f}_{m}^{*,r})-(\mathbf{f}_{m}^{r}) is the numerical flux minus the reference flux across the face.

Corresponding ESFR Scheme

Following an ESFR framework [13, 14], the reference flux is composed of a discontinuous and a corrected component,

fr(umh(Θm(ξr),t)):=fD,r(umh(Θm(ξr),t))+∑f=1Nf∑k=1Nfpgf,k(ξr)[n^r⋅fmC,r].(9)\mathbf{f}^{r}\bigl(\mathbf{u}_{m}^{h}(\boldsymbol{\Theta}_{m}(\boldsymbol{\xi}^{r}),t)\bigr):=\mathbf{f}^{D,r}\bigl(\mathbf{u}_{m}^{h}(\boldsymbol{\Theta}_{m}(\boldsymbol{\xi}^{r}),t)\bigr)+\sum_{f=1}^{N_{f}}\sum_{k=1}^{N_{fp}} \mathbf{g}^{f,k}(\boldsymbol{\xi}^{r})\left[\hat{\mathbf{n}}^{r}\cdot\mathbf{f}_{m}^{C,r}\right]. \tag*{(9)}

As illustrated in [18, 17, 38, 22, 23] the corresponding ESFR strong form as a filtered DG scheme is given by,

(Mm+Km)ddtu^m(t)T+Sξf^1mr(t)T+Sηf^2mr(t)T+Sζf^3mr(t)T+∑f=1Nf∑k=1Nfpχ(ξf,kr)TWf,k[n^r⋅fmC,r]=0T.(10)\left(M_{m}+K_{m}\right)\frac{d}{dt}\hat{\mathbf{u}}_{m}(t)^{T}+S_{\xi}\hat{\mathbf{f}}_{1m}^{r}(t)^{T}+S_{\eta}\hat{\mathbf{f}}_{2m}^{r}(t)^{T}+S_{\zeta}\hat{\mathbf{f}}_{3m}^{r}(t)^{T}+\sum_{f=1}^{N_{f}}\sum_{k=1}^{N_{fp}}\boldsymbol{\chi}(\boldsymbol{\xi}_{f,k}^{r})^{T}W_{f,k}\left[\hat{\mathbf{n}}^{r}\cdot\mathbf{f}_{m}^{C,r}\right]=\mathbf{0}^{T}. \tag*{(10)}

The ESFR strong form as a filtered DG scheme is also equivalently expressed as,

(Mm+Km)ddtu^m(t)T+χ(ξvr)TW∇rϕ(ξvr)⋅f^mr(t)T+∑f=1Nf∑k=1Nfpχ(ξf,kr)Wf,kn^r⋅[fm∗,r−ϕ(ξf,k)f^mr(t)]=0T(11)\begin{aligned} \left(M_{m}+K_{m}\right)\frac{d}{dt}\hat{\mathbf{u}}_{m}(t)^{T} &+\boldsymbol{\chi}(\boldsymbol{\xi}_{v}^{r})^{T}W\nabla^{r}\boldsymbol{\phi}(\boldsymbol{\xi}_{v}^{r})\cdot\hat{\mathbf{f}}_{m}^{r}(t)^{T}\\ &+\sum_{f=1}^{N_{f}}\sum_{k=1}^{N_{fp}}\boldsymbol{\chi}(\boldsymbol{\xi}_{f,k}^{r})W_{f,k}\hat{\mathbf{n}}^{r}\cdot\left[\mathbf{f}_{m}^{*,r}-\boldsymbol{\phi}(\boldsymbol{\xi}_{f,k})\hat{\mathbf{f}}_{m}^{r}(t)\right]=\mathbf{0}^{T} \tag*{(11)} \end{aligned}

The entire influence of the ESFR correction functions are stored in KmK_{m}, which is represented as,

(Km)ij≈∑s,v,wc(s,v,w)∫Ωr∂(s,v,w)χi(ξvr)∂(s,v,w)χj(ξvr) dΩr→Km=∑s,v,wc(s,v,w)∂(s,v,w)χ(ξvr)TWJm∂(s,v,w)χ(ξvr)=∑s,v,wc(s,v,w)(DξsDηvDζw)TMm(DξsDηvDζw),(12)\begin{aligned} (K_{m})_{ij}&\approx\sum_{s,v,w}c_{(s,v,w)}\int_{\Omega_{r}}\partial^{(s,v,w)}\chi_{i}(\boldsymbol{\xi}_{v}^{r})\partial^{(s,v,w)}\chi_{j}(\boldsymbol{\xi}_{v}^{r})\,d\Omega_{r}\\ &\to K_{m}=\sum_{s,v,w}c_{(s,v,w)}\partial^{(s,v,w)}\boldsymbol{\chi}(\boldsymbol{\xi}_{v}^{r})^{T}WJ_{m}\partial^{(s,v,w)}\boldsymbol{\chi}(\boldsymbol{\xi}_{v}^{r})\\ &=\sum_{s,v,w}c_{(s,v,w)}\left(D_{\xi}^{s}D_{\eta}^{v}D_{\zeta}^{w}\right)^{T}M_{m}\left(D_{\xi}^{s}D_{\eta}^{v}D_{\zeta}^{w}\right), \tag*{(12)} \end{aligned}

where ∂(s,v,w)\partial^{(s,v,w)} represents a multidimensional index spanning the pp-th order broken Sobolev-space [39], and its corresponding correction parameter,

2D:c(s,v)=c1D(sp+vp), such that s={0,p}, v={0,p}, s+v≥p,3D:c(s,v,w)=c1D(sp+vp+wp), such that s={0,p}, v={0,p}, w={0,p}, s+v+w≥p.(13)\begin{aligned} \text{2D:}\quad c_{(s,v)}&=c_{1D}^{\left(\frac{s}{p}+\frac{v}{p}\right)},\ \text{such that }s=\{0,p\},\ v=\{0,p\},\ s+v\ge p,\\ \text{3D:}\quad c_{(s,v,w)}&=c_{1D}^{\left(\frac{s}{p}+\frac{v}{p}+\frac{w}{p}\right)},\ \text{such that }s=\{0,p\},\ v=\{0,p\},\ w=\{0,p\},\ s+v+w\ge p. \tag*{(13)} \end{aligned}

The impact of the FR correction parameter is discussed in Section. § 3.3.

Nonlinearly Stable FR

Following the derivations in Chan [40, 41] [8], the general differential operator is used to recast Eqn.11 in a skew-symmetric two-point flux differencing form,

(Mm+Km)ddtu~m(t)T+[χ(ξmr)Tχ(ξfr)][(Q~−Q~T)⊙Fmr][IT+∑f=1Nf∑k=1Nfpχ(ξf,k)TWf,knr⋅ff,mr,∗]=0T(14)(\mathbf{M}_m+\mathbf{K}_m)\frac{d}{dt}\tilde{\mathbf{u}}_m(t)^T+\left[\boldsymbol{\chi}(\boldsymbol{\xi}_m^r)^T\boldsymbol{\chi}(\boldsymbol{\xi}_f^r)\right]\left[(\widetilde{\mathbf{Q}}-\widetilde{\mathbf{Q}}^T)\odot\mathbf{F}_m^r\right]\left[\mathbf{I}^T+\sum_{f=1}^{N_f}\sum_{k=1}^{N_{fp}}\boldsymbol{\chi}(\boldsymbol{\xi}_{f,k})^T\mathbf{W}_{f,k}\mathbf{n}^r\cdot\mathbf{f}_{f,m}^{r,*}\right]=\mathbf{0}^T \tag*{(14)}

where, Q~−Q~T\widetilde{\mathbf{Q}}-\widetilde{\mathbf{Q}}^T is the general hybridized skew-symmetric stiffness operator, ⊙\odot denotes a Hadamard product, and Fmr\mathbf{F}_m^r is the matrix storing the reference two-point flux values,

(Fmr)ij=fs(u~m(ξir),u~m(ξjr))(12(Cm(ξir)+Cm(ξjr))), ∀1≤i,j≤Nv+Nfp(15)(\mathbf{F}_m^r)_{ij}=f_s(\tilde{\mathbf{u}}_m(\boldsymbol{\xi}_i^r),\tilde{\mathbf{u}}_m(\boldsymbol{\xi}_j^r))\left(\frac{1}{2}\left(\mathbf{C}_m(\boldsymbol{\xi}_i^r)+\mathbf{C}_m(\boldsymbol{\xi}_j^r)\right)\right),\ \forall1\leq i,j\leq N_v+N_{fp} \tag*{(15)}

where Cm\mathbf{C}_m is the metric Jacobian cofactor matrix formulated as described in Cicchino et al. [23] and fs\mathbf{f}_s denotes the two-point flux which is described in Section. 3.2. This procedure, first proposed by Chan [8], is essential for entropy stability on uncollocated solution and flux nodes. For the detailed derivation and proofs for the NSFR discretization, we direct the reader to [42].

Methodology

This section presents the critical parameters in this investigation. The choice in positivity-preserving strategy, two-point flux, and FR correction parameter are key factors considered in extending the NSFR scheme to problems involving strong shocks. The impact of these factors is discussed in the following subsections.

A Modified Implementation of the Zhang-Shu Positivity-Preserving Limiter [43]

Consider the 2D Euler equations for perfect gas,

wt+f(w)x+g(w)y=0, t≥0, (x,y)∈R2,\mathbf{w}_t+\mathbf{f}(\mathbf{w})_x+\mathbf{g}(\mathbf{w})_y=0,\ t\geq0,\ (x,y)\in\mathbb{R}^2,
w=(ρmnE),f(w)=(mρu2+pρuv(E+p)u),g(w)=(mρuvρv2+p(E+p)v),(16)\mathbf{w}=\begin{pmatrix}\rho\\m\\n\\E\end{pmatrix},\quad\mathbf{f}(\mathbf{w})=\begin{pmatrix}m\\\rho u^2+p\\\rho uv\\(E+p)u\end{pmatrix},\quad\mathbf{g}(\mathbf{w})=\begin{pmatrix}m\\\rho uv\\\rho v^2+p\\(E+p)v\end{pmatrix}, \tag*{(16)}

with

m=ρu,n=ρv,E=ρe+12ρ(u2+v2),p=(γ−1)ρe,m=\rho u,\quad n=\rho v,\quad E=\rho e+\frac{1}{2}\rho(u^2+v^2),\quad p=(\gamma-1)\rho e,

where ρ\rho is density, uu and vv are the velocities in the xx-and yy-directions, mm and nn are the momentum in their respective directions, EE is total energy, pp is pressure, ee is internal energy and γ=1.4\gamma=1.4 (for air). Physically, it is required that density and pressure remain positive. Failure to preserve the positivity of these two properties results in the numerical algorithm diverging. High-order numerical schemes commonly used to solve systems of hyperbolic conservation laws, such as Eq. (16), do not provably preserve positivity without additional tools. In Zhang and Shu [43], techniques used to establish a maximum-principle preserving limiter [1] were generalized to derive a limiter designed to preserve positivity of density and pressure, provided density and pressure were positive at the previous timestep.

The set of admissible states for Eq. (16) is given by,

G={w=(ρmnE)|ρ>0 and p=(γ−1)(E−12ρ(u2+v2))>0}.(17)G = \left\{ \mathbf{w} = \begin{pmatrix} \rho\\ m \\ n \\ E \end{pmatrix} \middle| \rho> 0 \text{ and } p = (\gamma- 1) \left( E - \frac{1}{2}\rho\left(u^2 + v^2\right) \right) > 0 \right\}. \tag*{(17)}

Given the vector of approximation polynomials, uh(x)=(ρ(x),m(x),n(x),E(x))T\mathbf{u}^{h}(x) = \left(\rho(x),m(x),n(x),\mathbf{E}(x)\right)^{T}, and cell averages, w‾m=(ρ‾m,m‾m,n‾m,E‾m)∈G\overline{\mathbf{w}}_{m} = \left(\overline{\rho}_{m},\overline{m}_{m},\overline{n}_{m},\overline{E}_{m}\right) \in G, the positivity-preserving limiter is applied to transform uh(x)\mathbf{u}^{h}(x) to u~h(x)\widetilde{\mathbf{u}}^{h}(x) while maintaining high-order accuracy, positivity (ie. u~h(x)∈G\widetilde{\mathbf{u}}^{h}(x) \in G) and conservativity.

For dimensions greater than one, the limiter utilizes a mixed set of quadrature rules to determine the cell averages and minimum values for density and pressure. Using the polynomial representation of the local solution, uh(x)\mathbf{u}^{h}(x), we interpolate the solution to the new sets of quadrature rules:

ξr,1=ξα⊗ηβ,\xi^{r,1} = \xi^{\alpha} \otimes\eta^{\beta},
ξr,2=ξβ⊗ηα,(18)\xi^{r,2} = \xi^{\beta} \otimes\eta^{\alpha}, \tag*{(18)}

where the sets {ξβ,ηβ:−1≤ξβ,ηβ≤1}\{\xi^{\beta},\eta^{\beta} : -1 \le\xi^{\beta},\eta^{\beta} \le1\} and {ξα,ηα:−1≤ξα,ηα≤1}\{\xi^{\alpha},\eta^{\alpha} : -1 \le\xi^{\alpha},\eta^{\alpha} \le1\} denote GL and GLL quadrature points, respectively.

Using these two solution spaces, the cell averaged solution vector,

w‾m=(ρ‾m,(ρv1‾)m,(ρv2‾)m,E‾m)T(19)\overline{\mathbf{w}}_{m} = \left(\overline{\rho}_{m},\left(\overline{\rho\mathbf{v}_{1}}\right)_{m},\left(\overline{\rho\mathbf{v}_{2}}\right)_{m},\overline{E}_{m}\right)^{T} \tag*{(19)}

is computed using the quadrature rules as follows:

w‾m=a1λ1μw‾m+a2λ2μw‾m(20)\overline{\mathbf{w}}_{m} = \frac{a_{1}\lambda_{1}}{\mu}\overline{\mathbf{w}}_{m} + \frac{a_{2}\lambda_{2}}{\mu}\overline{\mathbf{w}}_{m} \tag*{(20)}
=a1λ1μ∑β=1L∑α=1Nωαωβw(ξr,1)+a2λ2μ∑β=1L∑α=1Nωαωβw(ξr,2),(21)= \frac{a_{1}\lambda_{1}}{\mu}\sum_{\beta=1}^{L}\sum_{\alpha=1}^{N}\omega_{\alpha}\omega_{\beta}\mathbf{w}\left(\xi^{r,1}\right) + \frac{a_{2}\lambda_{2}}{\mu}\sum_{\beta=1}^{L}\sum_{\alpha=1}^{N}\omega_{\alpha}\omega_{\beta}\mathbf{w}\left(\xi^{r,2}\right), \tag*{(21)}

where ω\omega denotes the quadrature weight, aa denotes maximum wavespeed, λ1=ΔtΔx\lambda_{1} = \frac{\Delta t}{\Delta x}, λ2=ΔtΔy\lambda_{2} = \frac{\Delta t}{\Delta y} and μ=a1λ1+a2λ2\mu= a_{1}\lambda_{1} + a_{2}\lambda_{2}.

The positivity-preserving limiter designed by Zhang and Shu, later improved by Wang and Shu [44], enforces positivity in two steps. In the first step, the density is limited using a scaling value, θ1\theta_{1}, that is determined with the cell average density, ρ‾m\overline{\rho}_{m}, and minimum density, ρmin\rho_{min}. In the original implementation of the limiter, ρmin\rho_{min} is determined by taking the lowest density value between the two quadrature rules. The density values in the vector of approximation polynomials are replaced with the limited density values, and the new vector of approximation polynomials is denoted as u^h(x)\widehat{\mathbf{u}}^{h}(x). The second step is to enforce the positivity of pressure through a scaling value, θ2\theta_2, that is determined by calculating the cell average pressure and pressure at the quadrature nodes. The cell average pressure is obtained using the cell averaged solution vector, r\mathbf{r} and the pressure at the quadrature nodes utilizes both mixed quadrature rules given in Eqn. (18). The values in the vector of approximation polynomials are scaled using θ2\theta_2 and the new vector of approximation polynomials is denoted as u~h(x)\widetilde{\mathbf{u}}^h(x).

From here, in the NSFR scheme with Euler forward in time, u~h(x)\widetilde{\mathbf{u}}^h(x) is used in the place of uh(x)\mathbf{u}^h(x). This limiter is applied at each stage of the RK method. It is also important to note that while the bound-preserving class of limiters provably ensures that density and pressure remain positive, the solutions are still prone to oscillations since it is not provably non-oscillatory.

The implementation of the positivity-preserving limiter described in Wang et al. [44] improves the original limiter through modification to the θ2\theta_2 scaling parameter but still lacks robustness guarantees as the values of density and pressure at the original solution nodes are not taken into account. This could result in negative density and pressure values at the solution nodes while these properties remain positive at the tensored quadrature points. Thus, to further ensure robustness, in this paper modifications are made to account for the values of density and pressure at the solution nodes, ξr,3=ξα⊗ηα\xi^{r,3} = \xi^\alpha\otimes\eta^\alpha, in addition to the other two sets of nodes when determining minimum density and pressure at the quadrature nodes. It should be noted that the cell averaged solution vector is still calculated as per Eq. (20) using the tensored quadrature points in Eqn. (18).

In the original implementation of the limiter, the θ\theta values which are affected by these changes are used to scale the solution nodes directly, and as such, this modification preserves the properties of the limiter. This modification has been verified through an extensive suite of test cases and also recovers the expected orders as highlighted in § 4.

Implementation

The implementation of the positivity-preserving limiter in 2D with the suggested modifications is briefly outlined in this section, and an algorithm is provided for clarity.

To achieve positivity preservation, first we define ε\varepsilon to be a small number greater than zero that serves as a lower bound for the average density and average pressure for all cells (ie. ρˉm>ε\bar{\rho}_m > \varepsilon and Pˉm>ε\bar{P}_m > \varepsilon). In this paper we use ε=10−13\varepsilon= 10^{-13} throughout. Next, we interpolate the solution nodes to the two sets of tensored quadrature points (uh(ξr,1)\mathbf{u}^h(\xi^{r,1}) and uh(ξr,2)\mathbf{u}^h(\xi^{r,2})) to determine the cell averaged solution vector as per Eq. (20). With the cell-averaged solution vector, the average pressure in the cell can be defined as:

Pˉm=(γ−1)(Eˉm−12mˉ2ρˉm−12nˉ2ρˉm).(22)\bar{P}_m = (\gamma- 1)\left(\bar{E}_m - \frac{1}{2}\frac{\bar{m}^2}{\bar{\rho}_m} - \frac{1}{2}\frac{\bar{n}^2}{\bar{\rho}_m}\right). \tag*{(22)}

With the cell averages determined, the solution vector can be scaled to ensure positivity. The first scaling value is determined using the cell average density (ρˉm\bar{\rho}_m) and the minimum density in the cell (ρmin⁡\rho_{\min}); where the latter is determined using three sets of nodes, that is, the solution nodes and the two tensored quadrature sets.

ρmin⁡=min⁡{ρ(ξr,1),ρ(ξr,2),ρ(ξr,3)}.\rho_{\min}=\min\left\{\rho(\boldsymbol{\xi}^{r,1}),\rho(\boldsymbol{\xi}^{r,2}),\rho(\boldsymbol{\xi}^{r,3})\right\}.

Using the cell average density and cell minimum density, the first scaling value is obtained:

θ1=min⁡{ρ‾m−ερ‾m−ρmin⁡,1}.\theta_1=\min\left\{\frac{\overline{\rho}_m-\varepsilon}{\overline{\rho}_m-\rho_{\min}},1\right\}.

which is employed to scale only density:

ρ^(x)=θ1(ρ(x)−ρ‾m)+ρ‾m.\widehat{\rho}(x)=\theta_1\left(\rho(x)-\overline{\rho}_m\right)+\overline{\rho}_m.

The original density values of the vector of approximation polynomials, wh(x)=(ρ(x),m(x),n(x),E(x))Tw^h(x)=\left(\rho(x),m(x),n(x),E(x)\right)^T are replaced with the scaled density values and the new vector of approximation polynomials is denoted as u^h(x)=(ρ^(x),m(x),n(x),E(x))T\widehat{u}^h(x)=\left(\widehat{\rho}(x),m(x),n(x),E(x)\right)^T.

The second scaling value is then determined using the cell average pressure and the cell minimum pressure across the three sets of nodes. The pressure at each node is determined, and if the pressure violates the preservation of positivity, the scaling value at that node is set to:

t=P(w‾m)P(w‾m)−P(u^h).t=\frac{P(\overline{w}_m)}{P(\overline{w}_m)-P(\widehat{u}^h)}.

or otherwise, it is set to 1. Once the scaling value at all the nodes for the three sets has been determined, the minimum scaling value across all sets of nodes is chosen as the second scaling value:

θ2=min⁡{t1,t2,t3}.\theta_2=\min\left\{t_1,t_2,t_3\right\}.

The second scaling value is then used to scale the vector of approximation polynomials with limited density thereby giving us:

u~h(x)=θ2(u^h(x)−w‾m)+w‾m.\widetilde{u}^h(x)=\theta_2\left(\widehat{u}^h(x)-\overline{w}_m\right)+\overline{w}_m.

which is a vector of approximation of polynomials that preserves accuracy, positivity, and conservativity. The presented positivity-preserving limiter based on Zhang and Shu [43] with the modifications presented above is detailed in Algorithm 1.

Two-Point Flux

In this paper, the split form convective two-point flux of Chandrashekar modified by Ranocha (CHRA_{\mathrm{RA}}) [45] is used. The (CHRA_{\mathrm{RA}}) flux is entropy conserving (EC), kinetic energy preserving (KEP), and pressure-equilibrium-preserving (PEP). The CHRA_{\mathrm{RA}} numerical two-point flux in the xx and yy directions is given by:

fc=[FCHRA#,GCHRA#].(23)\boldsymbol{f}_c=\left[F_{\mathrm{CH}_{\mathrm{RA}}}^{\#},G_{\mathrm{CH}_{\mathrm{RA}}}^{\#}\right]. \tag*{(23)}

Algorithm 1 2D Positivity-Preserving Limiter Implementation with Modifications

Require: ρ(ξr,3)>0\rho(\xi^{r,3}) > 0 and P(ξr,3)>0P(\xi^{r,3}) > 0 at previous timestep, where pressure is P=(γ−1)(E−12m2ρ−12n2ρ)P = (\gamma- 1)\left(E - \frac{1}{2}\frac{m^2}{\rho} - \frac{1}{2}\frac{n^2}{\rho}\right)

  1. Project solution to tensored quadrature nodes (ξr,1(\xi^{r,1} and ξr,2)\xi^{r,2})

  2. Determine cell averages w‾\overline{\mathbf{w}} and P‾\overline{P} using Eq. 20 and Eq. 22

  3. ρmin⁡←min⁡{ρ(ξr),ρ(ξr,2),ρ(ξr,3)}\rho_{\min} \leftarrow\min\{\rho(\xi^r), \rho(\xi^{r,2}), \rho(\xi^{r,3})\}

  4. if ρmin⁡<0\rho_{\min} < 0 then

  5. θ←min⁡{ρ‾−ϵρ‾−ρmin⁡,1}\theta\leftarrow\min\left\{\frac{\overline{\rho}-\epsilon}{\overline{\rho}-\rho_{\min}},1\right\}

  6. ρ^←θ1(ρ(ξr,3)−ρ‾)+ρ‾\widehat{\rho} \leftarrow\theta_1(\rho(\xi^{r,3})-\overline{\rho})+\overline{\rho}

  7. Replace ρ\rho with ρ^\widehat{\rho} in uh(x)u^h(x)

  8. end if

  9. for all ξr\xi^r in {ξr,1,ξr,2,ξr,3}\{\xi^{r,1},\xi^{r,2},\xi^{r,3}\} do

  10. if P(u^h(ξr))<0P(\widehat{u}^h(\xi^r)) < 0 then

  11. t(ξr)←P(w‾)P(w‾)−P(u^h(ξr))t(\xi^r) \leftarrow\frac{P(\overline{\mathbf{w}})}{P(\overline{\mathbf{w}})-P(\widehat{u}^h(\xi^r))}

  12. θ2←min⁡{θ2,t}\theta_2 \leftarrow\min\{\theta_2,t\}

  13. end if

  14. end for

  15. u~←θ2(uh(ξr,3)−w‾)+w‾\widetilde{u} \leftarrow\theta_2(u^h(\xi^{r,3})-\overline{\mathbf{w}})+\overline{\mathbf{w}}

Ensure: ρ~(ξr,3)>0\widetilde{\rho}(\xi^{r,3}) > 0 and P~(ξr,3)>0\widetilde{P}(\xi^{r,3}) > 0

where FCHRA#≡FCHRA#(UL,UR)F_{\mathrm{CHRA}}^{\#} \equiv F_{\mathrm{CHRA}}^{\#}(U_L,U_R) and GCHRA#≡GCHRA#(UL,UR)G_{\mathrm{CHRA}}^{\#} \equiv G_{\mathrm{CHRA}}^{\#}(U_L,U_R) are computed using the left state ULU_L and right state URU_R as:

FCHRA#=[ρ^u^ρ^u^2+p^1ρ^u^v^ρ^u^h^−{ ⁣{v1P} ⁣}],GCHRA#=[ρ^v^ρ^v^u^ρ^v^2+p^1ρ^v^h^−{ ⁣{v2P} ⁣}](24)F_{\mathrm{CHRA}}^{\#} = \begin{bmatrix} \hat{\rho}\hat{u} \\ \hat{\rho}\hat{u}^{2}+\hat{p}_{1} \\ \hat{\rho}\hat{u}\hat{v} \\ \hat{\rho}\hat{u}\hat{h}-\{\!\{v_{1}P\}\!\} \end{bmatrix}, \qquad G_{\mathrm{CHRA}}^{\#} = \begin{bmatrix} \hat{\rho}\hat{v} \\ \hat{\rho}\hat{v}\hat{u} \\ \hat{\rho}\hat{v}^{2}+\hat{p}_{1} \\ \hat{\rho}\hat{v}\hat{h}-\{\!\{v_{2}P\}\!\} \end{bmatrix} \tag*{(24)}
ρ^=ρln⁡,u^={ ⁣{v1} ⁣},v^={ ⁣{v2} ⁣},p^1={ ⁣{P} ⁣},(25)\hat{\rho}=\rho^{\ln},\qquad\hat{u}=\{\!\{v_{1}\}\!\},\qquad\hat{v}=\{\!\{v_{2}\}\!\},\qquad\hat{p}_{1}=\{\!\{P\}\!\}, \tag*{(25)}
h^=1p^2(γ−1)+12∑i=12(2{ ⁣{vi} ⁣}2−{ ⁣{vi2} ⁣})+2p^1ρln⁡,p^2=(ρP)ln⁡,(26)\hat{h}=\frac{1}{\hat{p}_{2}(\gamma-1)}+\frac{1}{2}\sum_{i=1}^{2}\left(2\{\!\{v_{i}\}\!\}^{2}-\{\!\{v_{i}^{2}\}\!\}\right)+\frac{2\hat{p}_{1}}{\rho^{\ln}},\qquad\hat{p}_{2}=\left(\frac{\rho}{P}\right)^{\ln}, \tag*{(26)}

where { ⁣{⋅} ⁣}\{\!\{\cdot\}\!\} denotes the mean:

{ ⁣{⋅} ⁣}=(⋅)L+(⋅)R2,(27)\{\!\{\cdot\}\!\}=\frac{(\cdot)_{L}+(\cdot)_{R}}{2}, \tag*{(27)}

and (⋅)ln⁡(\cdot)^{\ln} denotes the logarithmic mean:

(⋅)ln⁡=(⋅)L−(⋅)Rln⁡((⋅)L)−ln⁡((⋅)R).(28)(\cdot)^{\ln}=\frac{(\cdot)_{L}-(\cdot)_{R}}{\ln((\cdot)_{L})-\ln((\cdot)_{R})}. \tag*{(28)}

Due to the nature of the logarithmic mean, determining a CFL condition for the two-point flux is a challenge. This paper aims to numerically establish a CFL condition instead. The two-point flux is expected to preserve positivity for pressure if it follows the condition provided in Zhang and Shu [43], which states that for the Lax-Friedrichs flux, the CFLCFL condition can be relaxed to,

ΔtΔx∥(∣u∣+c)∥∞≤1.(29)\frac{\Delta t}{\Delta x}\|(|u|+c)\|_{\infty}\leq1. \tag*{(29)}

This condition can be used for all polynomial degrees as the Δx\Delta x is determined using the minimum distance between quadrature nodes within a given cell.

In addition to the condition given above, the two-point flux is expected to preserve positivity for density if it follows the condition in Ranocha [46],

ΔtΔx∥(∣u∣+c)∥∞≤0.5.(30)\frac{\Delta t}{\Delta x}\|(|u|+c)\|_{\infty}\leq0.5. \tag*{(30)}

It should be noted however that this condition uses cell width for Δx\Delta x which wouldn’t reflect the change in the minimum Δx\Delta x when the polynomial degree is changed. Thus, this condition is only expected to hold for p=2p=2.

A numerical study of the CFL conditions is given in § 4.2.3, § 4.3.1, and § 4.5.3.

FR Correction Parameter

The modified mass matrix seen in Eq. (14) has been shown to effectively act as a linear filter operator on the DG residual [17, 18]. The ESFR correction functions stored in KmK_m are scaled by the correction parameter, cc, which can be used to tune the linear filtering operation. The correction parameter is generally chosen within the range of c∈[c−,c+]c \in[c_-,c_+], where c−c_- and c+c_+ denote the minimum and maximum value, respectively, for which the scheme is stable [47] [18]. For values of c∈[c−,0)c \in[c_-,0), the filter will amplify the highest mode of the residual and for values of c∈(0,c+]c \in(0,c_+], the filter dampens the highest mode. When the correction parameter is chosen to be c=0c=0, it recovers the unfiltered DG method and we denote this value as cDGc_{DG}. Other notable values of cc include cSDc_{SD} which recovers a stable spectral difference scheme [18] and cHUc_{HU} which recovers Hyunh’s g2g_2 scheme [13].

Temporal Discretization

In this work, the solution is advanced in time using the strong stability preserving third-order accurate Runge-Kutta (SSPRK3) explicit time advancement method [48]. The method is presented below,

w(1)=un+ΔtL(tn,un),w^{(1)} = u^n + \Delta t\mathcal{L}(t^n,u^n),
w(2)=14(3un+w(1)+ΔtL(w(1),tn+Δt)),w^{(2)} = \frac{1}{4}\left(3u^n+w^{(1)}+\Delta t\mathcal{L}(w^{(1)},t^n+\Delta t)\right),
un+1=13(un+2w(2)+2ΔtL(w(2),tn+12Δt))u^{n+1} = \frac{1}{3}\left(u^n+2w^{(2)}+2\Delta t\mathcal{L}\left(w^{(2)},t^n+\frac{1}{2}\Delta t\right)\right)

The convective-based CFL condition used in this work is determined as follows:

Δt=CFLΔx~λmax⁡,Δx~=xmax⁡−xmin⁡(DOF)1/dim,λmax⁡=max⁡(∣v∣+c),(31)\Delta t = CFL\frac{\widetilde{\Delta x}}{\lambda_{\max}}, \qquad\widetilde{\Delta x} = \frac{x_{\max}-x_{\min}}{(DOF)^{1/dim}}, \qquad\lambda_{\max} = \max(|v|+c), \tag*{(31)}

where Δx~\widetilde{\Delta x} is the approximate grid spacing, dimdim is the dimension of the problem, and λmax⁡\lambda_{\max} is the maximum wavespeed throughout the entire domain with speed of sound c=γRTc=\sqrt{\gamma RT}.

Numerical Results

In this section, the accuracy and robustness of the NSFR method in the context of shock-capturing are demonstrated through a variety of 1D and 2D numerical tests. The positivity-preserving limiter (PPL) enables the extension of the NSFR method to problems involving strong shocks and complex features through the guarantee of positivity preservation, but as stated in Section 3.1, while positivity is preserved, the limiter does not eliminate oscillations in the solution. We investigate the stability of the NSFR scheme in the absence of either a TVD limiter or shock-capturing mechanisms such as slope limiters or artificial viscosity.

All tests use GLL quadrature points, the Chandrashekar flux with the Ranocha pressure fix (CHRA\mathrm{CH}_{\mathrm{RA}}), and Roe dissipation unless otherwise specified. The polynomial degree is varied for some tests to demonstrate the effect of increasing polynomial order on problems involving shocks. The polynomial degree is denoted with a pp followed by the degree, such as p3p3 to denote an approximation polynomial degree of 3. We also observe numerically that the two-point flux abides by the CFLCFL conditions, Eq. (29) and 30, in both one and two dimensions.

In addition to accuracy and robustness, the tests aim to provide a thorough assessment of the NSFR method by observing the effects of variants of flux reconstruction schemes for shock-dominated problems. The tests assessing the correction parameter are performed with a polynomial degree of 3 (p3p_3). Table 1 list the variants of the flux reconstruction scheme investigated in this work.

FR Parameterc
Discontinuous Galerkin, cDGc_{DG}0
Spectral-Difference, cSDc_{SD}7.44E−67.44E-6
Huynh’s [13] g2g_2 FR, cHUc_{HU}1.32E−51.32E-5
c+c_+2.87E−52.87E-5
c+×10c_+ \times102.87E−42.87E-4

Table 1. Values of the correction parameter for polynomial degree 3

Low Density Convergence Test

The first case used to verify the implementation of the PPL is the 2D Low Density case [43]. The problem is used to demonstrate that the scheme still achieves the expected order of accuracy. The problem is solved on a domain of size [0,2π]×[0,2π][0,2\pi] \times[0,2\pi] with periodic boundary conditions and initial condition,

ρ0(x,y)=1+0.995sin⁡(x+y),u0(x,y)=1,v0(x,y)=1,p0(x,y)=1.\rho_0(x,y) = 1 + 0.995\sin(x+y),\quad u_0(x,y) = 1,\quad v_0(x,y) = 1,\quad p_0(x,y) = 1.

The exact solution

ρ(x,y,t)=1+0.995sin⁡(x+y−2t),u(x,y,t)=1,v(x,y,t)=1,p(x,y,t)=1\rho(x,y,t) = 1 + 0.995\sin(x+y-2t),\quad u(x,y,t) = 1,\quad v(x,y,t) = 1,\quad p(x,y,t) = 1

is used to determine the order of convergence. The final time is t=0.1st = 0.1\mathrm{s} and Δx=Δy=2πN\Delta x = \Delta y = \frac{2\pi}{N}. The minimum expected density for this case is ρmin⁡=0.005\rho_{\min} = 0.005, thereby serving as a good benchmark case for positivity preservation of density. The results are shown in Table 2 for polynomial degrees 2 and 3 with grid size starting at 8×88 \times8 and successively refined to 512×512512 \times512. The scheme achieves the expected order of accuracy in the finer grids.

kkN×NN \times NdxL1L_1 errorOrderL2L_2 errorOrderLimiter
28×88 \times82.0833E-021.1166E-012.4880E-02✓
16×1616 \times161.0417E-021.8730E-022.585.4392E-032.19
32×3232 \times325.2083E-033.1915E-032.551.2218E-032.15
64×6464 \times642.6042E-033.6526E-043.131.4299E-043.10
128×128128 \times1281.3021E-034.5685E-053.001.9041E-052.91
256×256256 \times2566.5104E-045.6583E-063.012.3462E-063.02
512×512512 \times5123.2552E-047.0088E-073.012.8993E-073.02
38×88 \times81.5625E-022.4676E-029.7284E-03✓
16×1616 \times167.8125E-033.4557E-032.841.8439E-032.40✓
32×3232 \times323.9062E-032.8400E-043.611.7284E-043.42
64×6464 \times641.9531E-031.7344E-054.031.2619E-053.78
128×128128 \times1289.7656E-041.0449E-064.058.2709E-073.93
256×256256 \times2564.8828E-046.4863E-084.015.3026E-083.96
512×512512 \times5122.4414E-044.0222E-094.013.3402E-093.99

Table 2. Low Density convergence test with initial data 1+0.999sin⁡(x+y)1 + 0.999\sin(x+y) and exact solution 1+0.999sin⁡(x+y−2t)1 + 0.999\sin(x+y-2t), Δx=Δy=2πN\Delta x = \Delta y = \frac{2\pi}{N} and final time t=0.1st = 0.1\mathrm{s}

In this investigation, the behaviour of the scheme when using GL flux nodes is also examined. To use the limiter with GLL solution nodes and GL flux nodes, it is essential to verify that the scheme still reaches the expected order of accuracy. The same initial condition and exact solution are used to determine the order of convergence. Similar to the GLL test, the results are shown in Table 3 for polynomial degrees 2, 3 with grid size starting at 8×88 \times8 and successively refined to 512×512512 \times512. The CFL for this test is modified to account for the sensitivities of GL nodes. The test was monitored to ensure that the limiter is turned on for coarser meshes. The scheme achieves the expected order of accuracy in the finer grids.

NSFR - CHRACH_{RA} - Roe Dissipation - cDGc_{DG} - GL - CFL0.01
kkN×NN \times NdxL1L_1 errorOrderL2L_2 errorOrderLimiter
28×88 \times82.0833E-022.0711E+005.3860E-01✓
16×1616 \times161.0417E-025.5323E-025.231.5116E-025.16✓
32×3232 \times325.2083E-038.2801E-032.742.7284E-032.47
64×6464 \times642.6042E-038.8661E-043.223.8200E-042.84
128×128128 \times1281.3021E-039.3052E-053.254.3321E-053.14
256×256256 \times2566.5104E-048.8377E-063.403.4600E-063.65
512×512512 \times5123.2552E-048.8582E-073.322.7869E-073.63
38×88 \times81.5625E-022.0763E+006.0201E-01✓
16×1616 \times167.8125E-032.0856E-026.647.8696E-036.26✓
32×3232 \times323.9062E-031.8480E-033.509.0592E-043.12
64×6464 \times641.9531E-038.2785E-054.486.0332E-053.91
128×128128 \times1289.7656E-043.7260E-064.472.9894E-064.33
256×256256 \times2564.8828E-041.3273E-074.811.0178E-074.88
512×512512 \times5122.4414E-045.4963E-094.593.9881E-094.67

Table 3. Low Density convergence test with initial data 1+0.995sin⁡(x+y)1 + 0.995\sin(x + y) and exact solution 1+0.995sin⁡(x+y−2t)1 + 0.995\sin(x + y - 2t), Δx=Δy=2πN\Delta x = \Delta y = \frac{2\pi}{N} and final time t=0.1st = 0.1s

Sod Shock Tube

The first 1D test case used to investigate the shock capturing capabilities of the NSFR scheme is the Sod Shock Tube case. This case is a Riemann problem for the 1D Euler equations. The computational domain is [−0.5,0.5][-0.5, 0.5]. The problem is initialized as,

(ρ,w,p)={(1,0,1)if x<0,(0.125,0,0.1)if x≥0.(32)(\rho, w, p) = \begin{cases} (1, 0, 1) & \text{if } x < 0, \\ (0.125, 0, 0.1) & \text{if } x \ge0. \end{cases} \tag*{(32)}

The final time is t=0.2st = 0.2s. The number of cells is N=512N = 512 for the tests unless stated otherwise. The exact solution of density contains a left rarefaction wave and a right shock wave with a contact discontinuity in the middle.

Comparison with Strong DG

The SSPRK3 Strong DG method with Roe flux and GLL nodes is used with the PPL to obtain a solution to the Sod Shock Tube test case. The CFLCFL number used to run this case is 0.5. The case runs successfully without diverging; however, an entropy-violating shock is present in the solution. While the PPL preserves the positivity of density and pressure, the scheme itself does not prevent the presence of entropy-violating shocks. A plot of entropy over time is provided in Figure 1, wherein the entropy is shown to be increasing over time for this scheme.

Image not extracted

Figure 1. [Sod Shock Tube] Entropy over time for initial three test configurations (top) and density and pressure with SSPRK3 Strong DG with TVD limiter (bottom left) and NSFR (bottom right), using p=3p = 3, CFL 0.5 and the PPL at t=0.2st = 0.2\mathrm{s}

To eliminate the entropy-violating shock, the TVD limiter introduced in Chen and Shu [37] is used with the PPL. This test configuration confirms that the implemented PPL can be used in conjunction with other limiting strategies. The CFLCFL number used to run the case is 0.5. Referring back to Figure 1, we can see that the entropy decreases over time, and this is reflected in the density and pressure plots below. There is no entropy-violating shock present, and the density solution shows a left rarefaction wave, right shock wave and a contact discontinuity in the middle as expected. While the main features of the solution closely reflect that of the exact solution, the SSPRK3 Strong DG TVD solution contains some dissipative error at the head of the rarefaction wave and the contact discontinuity, with some dissipative error at the right shock as well. The solution does not accurately reflect the exact solution in these locations and is overly dissipative.

This test case is also run using the NSFR scheme to demonstrate the advantages of the scheme. As a result of the scheme’s robustness, at a suitable CFLCFL number, the NSFR scheme can be used to solve this case without the PPL, while the SSPRK3 Strong DG method fails without the limiter when tested with a CFLCFL as low as 0.01. The maximum CFLCFL number for which the NSFR scheme can be used without the PPL is 0.2, which is 20 times greater than the lowest tested CFLCFL for the SSPRK3 Strong DG method.

Using the PPL, the NSFR scheme can be used to run the test at a higher CFLCFL. For this test, the CFLCFL number is set to 0.5 so that the SSPRK3 Strong DG solutions can be compared to the NSFR solution. The solution obtained with the flux reconstruction parameter set to cDGc_{DG} is shown in Figure 1. The solution contains a left rarefaction wave and a right shock wave with a contact discontinuity in the middle. There is no entropy-violating shock present, and the shock transitions are sharp and are not overly dissipative. The NSFR solution does, however, contain some dispersive error in the wake of the shock. In terms of entropy, the NSFR scheme also has decreasing entropy over time as expected, but the decrease in entropy is less than that of the SSPRK3 Strong DG solution with the TVD limiter, suggesting that it is less dissipative.

Polynomial Degree Comparison

A comparison can also be conducted for the polynomial degree to demonstrate the ability of the scheme to remain stable without the added use of shock-capturing methods. The Sod Shock Tube case is run with polynomial degrees p=3,4,5p = 3, 4, 5 with CFLCFL numbers 0.5, 0.4 and 0.3 respectively. The number of cells for each of the tests is N=240,192,160N = 240, 192, 160 respectively to ensure the tests have an equivalent number of degrees of freedom. The three solutions for the different polynomial degrees are compared in Figure 2. All three results closely follow the exact solution and have sharp shock transitions. The p=4p = 4 and p=5p = 5 results follow closely with the p=3p = 3 results but have a significant increase in oscillations in the wake of the shock as expected.

Image not extracted

Figure 2. [Sod Shock Tube] Density with NSFR-CHRA, Roe dissipation, cDGcDG, p=3,4,5p = 3, 4, 5, CFL 0.5, 0.4, 0.3 respectively with PPL at t=0.2st = 0.2s. Plots of full domain (left) and 0≤x≤3.60 \le x \le3.6

CFL Condition for Two-Point Flux

This test is used to numerically verify that the two-point flux used for the tests in this investigation abides by a CFL condition. The number of cells for this test is N=500N = 500. The test is run for polynomial degrees p=2,3,4p = 2, 3, 4 with an increment of 0.0005s0.0005s for the constant time step. The CFL for a given configuration is calculated in two different approaches. The first method uses the minimum Δx\Delta x between the GLL quadrature nodes at the given degree for Δx\Delta x. The second method uses the cell size for Δx\Delta x. The CFL is calculated as follows:

CFL=ΔtΔxλˉmax⁡,(33)CFL = \frac{\Delta t}{\Delta x}\bar{\lambda}_{\max}, \tag*{(33)}

where λˉmax⁡\bar{\lambda}_{\max} denotes the maximum wavespeed across the entire domain and all time steps, calculated using the cell averaged solution vector, wˉ\bar{\mathbf{w}}.

For the first method of calculating the CFL, we observe that the test only fails when the CFL exceeds 1. This follows closely with the CFL condition given in Eq. (29), thus numerically verifying that the sufficient condition for the CFL holds for the two-point flux as well. For the second method, it is observed for p=2p = 2 that the test fails when the CFL exceeds 0.5, which follows closely with the observed condition given in Eq. (30). It is noted, however, that for p=3p = 3 and 4, the CFL at failure is lower than the expected condition as the change in the minimum Δx\Delta x between the different polynomial degrees is not negligible. This case numerically shows that Eq. (29) is a sufficient condition for positivity preservation for two-point fluxes, irrespective of the polynomial degree for the current case.

kkMin. Δx\Delta xCell WidthTime StepMax. λ\lambdaMin. Δx\Delta x CFLCell Width CFL
20.0010.0020.00212.320.950.47
0.002152.320.970.49
0.00222.320.990.49
0.002252.511.110.55
30.000550.0020.00142.260.860.24
0.001452.260.900.25
0.00152.270.950.26
0.001553.141.370.38
40.000350.0020.00082.230.710.12
0.000852.230.780.13
0.00092.240.840.15
0.000952.761.120.19

Table 4. Two-point flux positivity preservation test using the Sod Shock Tube setup with N=500N = 500.

Two-Point Flux Comparison

This test is also used to compare the impact of the two-point flux on the solution. In addition to the (CHRA\mathrm{CH}_{\mathrm{RA}}) flux used throughout this investigation, the following fluxes are considered:

The properties of the different two-point fluxes are highlighted in Table 5. The computational cost for all the fluxes is similar with the exception of the (KG) flux which does not require logarithmic averaging (Eq. (28)) thereby incurring a lower cost than the other fluxes. Despite the KG incurring a lower cost, it is not preferred for problems involving strong shocks. The issues with the KG flux are highlighted in § 4.5.

Two-Point FluxECKEPPEP
Ismail-Roe (IR)[49]✓
Kennedy-Gruber (KG)[50]✓
Chandrashekar (CH)[51]✓✓
CHRA\mathrm{CH}_{\mathrm{RA}}[45]✓✓✓

Table 5. Summary of two-point fluxes

The solution for each two-point flux is shown in Figure 3. The results for the different two-point fluxes follow closely with the exact solution. The shock location and magnitude are correct and the shock transitions are sharp. It is observed that the CHRA\mathrm{CH}_{\mathrm{RA}} solution is slightly more oscillatory in the wake of the shock compared to the other 3 fluxes. Additionally, the magnitude of the overshoot at the shock is slightly greater for the CHRACH_{RA} and IRIR fluxes compared to CHCH and KGKG. It is also noted that while the KGKG flux is less oscillatory in the wake of the shocks, it exhibits the greatest undershoot at the foot of the rarefaction wave out of all the fluxes.

Image not extracted

Figure 3. [Sod Shock Tube] Density and pressure with NSFR, Roe dissipation, cDGc_{DG}, p=3p = 3, CFL 0.5 with PPL at t=0.2t = 0.2s for CHRACH_{RA}, CHCH, IRIR and KGKG

Investigation of Flux Reconstruction Schemes

Here, we investigate variants of the flux reconstruction scheme. The Sod Shock Tube case is run with the cSDc_{SD}, cHUc_{HU}, and c+c_{+} schemes with p=3p = 3 and a CFL of 0.5. Using the solution obtained with the cDGc_{DG} scheme in Figure 1, the four different schemes are compared in Figure 4. The results for all four schemes follow closely with the exact solution, and the shocks are in the correct locations and have sharp transitions. As the value of the flux reconstruction parameter is increased from cDGc_{DG} to c+c_{+}, it is observed that the overshoot at the shock and the oscillations in the wake of the shock are diminished; which has the effect of filtering the highest mode of the solution. However, while filtering of nonphysical behaviour is an advantage, it is also observed that as the FR parameter is increased, there is a noticeable undershoot at the foot of the rarefaction wave.

Image not extracted

Figure 4. [Sod Shock Tube] Density with NSFR-CHRA, Roe dissipation, p=3p = 3, CFL 0.5 respectively with PPL at t=0.2t = 0.2s for different flux reconstruction parameters. Plots of full domain (left) and 0≤x≤3.60 \le x \le3.6

Overall, the Sod Shock Tube case demonstrates the NSFR method’s ability to simulate problems involving shocks without the need for a TVB/TVD limiter and the added robustness of the method also allows for a higher CFL. Even when parameters such as the polynomial degree and FR parameter are varied, the method still consistently provides a solution that accurately reflects the exact solution without the need for TVB/TVD limiting. The test also sufficiently verifies the CFL condition for two-point fluxes numerically.

Shu Osher Problem

The Shu Osher problem is a 1D Euler case with a shock propagating into an unsteady density field. This problem essentially simulates shock-turbulence interaction in 1D. The problem tests the ability of a scheme to capture a shock, the interaction of the shock with an unsteady density field and the waves that propagate downstream of the shock. The computational domain of the problem is [−5,5][-5,5]. The problem is initialized as,

(ρ,w,p)={(3.857143,2.629369,10.33333)if x<−4,(1+0.2sin⁡(5x),0,1)if x≥−4.(34)(\rho,w,p) = \begin{cases} (3.857143, 2.629369, 10.33333) & \text{if } x < -4, \\ (1 + 0.2\sin(5x), 0, 1) & \text{if } x \ge-4. \end{cases} \tag*{(34)}

The final time for this case is t=1.8t = 1.8s. The number of cells is N=128N = 128. As outlined in [52], at time 1.81.8s, the shock location is xs≈2.39x_s \approx2.39, the contact discontinuity location is xc≈0.69x_c \approx0.69, and the location of the leading acoustic wave is xa≈−2.75x_a \approx-2.75. The results are compared to a reference solution obtained using the third-order SSPRK3 Strong DG method with Lax-Friedrichs flux, GLL nodes, polynomial degree of p=0p = 0 and N=32768N = 32768.

This test is first run using the NSFR scheme without the PPL to demonstrate its robustness. The CFL number for this case is 0.01 and the results are given in Figure 5. This test cannot be run with the Strong DG scheme in the same configuration. There is a significant overshoot at the shock, but otherwise, the solution accurately reflects all the expected features, including the shock, contact discontinuity and acoustic waves.

Image not extracted

Figure 5. [Shu Osher Problem] Density solution with NSFR-CHRA, Roe dissipation, cDGc_{DG}, no PPL at t=1.8st = 1.8\mathrm{s}. Plots of full domain (left) and −2.6≤x≤2.6-2.6 \le x \le2.6 (right).

CFL Condition for Two-Point Flux

This test is also used to numerically verify that the two-point flux used throughout this study abides by a CFLCFL condition. The number of cells for this test is N=360N = 360. The test is run for polynomial degrees p=2,3,4p = 2, 3, 4 with an increment of 0.00005s0.00005\mathrm{s} for the constant time step. Once again, the CFLCFL for a given configuration is calculated in two different ways. The first method uses the minimum Δx\Delta x between the GLL quadrature nodes at the given degree for Δx\Delta x. The second method uses the cell size for Δx\Delta x. The CFLCFL is calculated using Eq. (33). The results of the numerical study are given in Table 6.

kkMin. Δx\Delta xCell WidthTime StepMax. λ\lambdaMin. Δx\Delta x CFLCell Width CFL
20.013890.027780.00215.450.820.41
0.002155.640.870.44
0.00225.700.900.45
0.002257.301.180.59
30.007680.027780.00144.970.910.25
0.001454.990.940.26
0.00154.970.970.27
0.001555.951.200.33
40.004790.027780.00084.900.820.14
0.000854.900.870.15
0.00094.930.930.16
0.000954.900.970.17

Table 6. Two-point flux positivity preservation test using the Shu Osher Problem setup with N=360N = 360. NSFR - CHRACH_{RA} - Roe Dissipation - cDGc_{DG} - N=360N = 360

Similar to the results obtained using Sod Shock Tube, this test also demonstrates that the two-point flux abides by the condition given in Eq. (29) for the first method. The p=4p = 4 case fails for a CFLCFL slightly less than 1, but it is sufficiently close enough to the bound that it is considered to satisfy the condition. The results also show that the two-point flux abides by Eq. (30) for a polynomial degree of 2. This test case further establishes Eq. (29) as a sufficient condition for positivity preservation of pressure for two-point fluxes irrespective of polynomial degree.

Investigation of Flux Reconstruction Schemes

This test is also run using the cDGc_{DG} and c+c_{+} schemes. The CFLCFL number used for both configurations is 0.5. This increased CFLCFL is a good indicator of the added robustness the PPL provides. The results obtained with the cDGc_{DG} scheme are shown in Figure 6. The cDGc_{DG} results capture all the locations of the shock, contact discontinuity and acoustic waves accurately, similar to the results in Figure 5 despite having a CFLCFL that is 50 times greater. Comparing the limited solution to the reference solution, however, it is evident that there are overshoots in the vicinity of the shock, and there are oscillations present immediately upstream of the shock.

Image not extracted

Figure 6. [Shu Osher Problem] Density solution with NSFR-CHRA, Roe dissipation, cDGc_{DG}, PPL at t=1.8st = 1.8\mathrm{s}. Plots of full domain (left) and −2.6≤x≤2.6-2.6 \le x \le2.6 (right).

The results obtained using the c+c_{+} scheme are shown in Figure 7. The plot shows that the shock, contact discontinuity, and leading acoustic wave are all accurately captured and the shock does not exhibit an overshoot as in the cDGc_{DG} scheme and the oscillations in the wake of the shock have been dampened. However, comparing the c+c_{+} solution to the reference solution, it is apparent that there is room for improvement as the dissipation added to dampen the oscillations and overshoot also diminish the magnitude of the waves upstream of the shock.

Image not extracted

Figure 7. [Shu Osher Problem] Density solution with NSFR-CHRA, Roe dissipation, c+c_{+}, PPL at t=1.8st=1.8\mathrm{s}. Plots of full domain (left) and −2.6≤x≤2.6-2.6 \le x \le2.6 (right).

The plot of the results between −2.6≤x≤2.6-2.6 \le x \le2.6 shows the mitigation of the overshoot and oscillations in the c+c_{+} results, but it also highlights the diminished magnitude of the waves downstream of the shock which is better captured by cDGc_{DG}. Using c+c_{+} to filter out the highest mode produces fewer oscillatory results but leads to diminished peaks. An adaptive application of the flux reconstruction parameter wherein FR, c+c_{+}, is used in shock locations while discontinuous Galerkin, cDGc_{DG}, is used elsewhere could lead to improved results that achieve the desired magnitudes while reducing overshoots and oscillations in the presence of shocks and would be part of a future study.

Impact of Flux Nodes

This test case is used to demonstrate the impact of the choice of flux nodes. The test is run once again using the same parameters as the cDGc_{DG} scheme, with the exception of the flux nodes, which are set to GL. As a result of GL quadrature being used, the CFL is also subsequently lowered. As discussed in [8], for quadratures such as GL, the solution can spike when evaluating the entropy-projected conservative variables at surface points, and to mitigate this effect, the CFLCFL is set to 0.05. The test is also rerun with the lowered CFLCFL for GLL flux nodes to facilitate a better comparison of the results. As shown in Figure 8, the results for the GL quadrature are more accurate than the GLL results. There is less overshoot in the wake of the shock, and there is less oscillatory behaviour, but it comes at the cost of the CFLCFL being lowered. The test can be run with a significantly higher CFLCFL using GLL flux nodes, and the issues of overshoot and oscillations can be addressed through the flux reconstruction parameter as shown by the c+c_{+} scheme, thereby making GLL flux nodes preferable over GL flux nodes.

Image not extracted

Figure 8. [Shu Osher Problem] Density solution with NSFR-CHRA, Roe dissipation, cDG, PPL at t=1.8st = 1.8\mathrm{s}. Plots of GLL nodes (top) and GL nodes (bottom). Plots of full domain (left) and −2.6≤x≤2.6-2.6 \le x \le2.6 (right).

Polynomial Degree Comparison

A polynomial degree comparison is also presented for the Shu Osher test case. The test is run for p=3p = 3 and p=4p = 4 at a CFLCFL of 0.5. The results for polynomial degrees 3 and 4 are shown in Figure 9. Both solutions closely follow the reference solution but the higher polynomial degree solution contains more oscillations in the waves downstream of the shock similar to the Sod Shock Tube polynomial degree comparison. However, unlike the Sod Shock Tube results, the higher polynomial degrees display less overshoot at the shock than the p=3p = 3 results.

Image not extracted

Figure 9. [Shu Osher Problem] Density solution with NSFR-CHRA, Roe dissipation, cDGc_{DG}, PPL at t=1.8t = 1.8s for p=3,4p = 3, 4. Plots of full domain (left) and −2.6≤x≤2.6-2.6 \le x \le2.6 (right).

Leblanc Shock Tube

The modified version of the Leblanc Shock Tube case seen in Zhang and Shu [43] is used to verify the 1D implementation of the PPL. Due to the extreme nature of the shock, this case is a good test of the robustness of the NSFR scheme in conjunction with the PPL. This case is a Riemann problem of 1D Euler equations. The computational domain is [−10.0,10.0][-10.0,10.0]. The problem is initialized as,

(ρ,w,p)={(2,0,109)if x<0,(0.001,0,1)if x≥0.(35)(\rho,w,p)= \begin{cases} (2,0,10^{9}) & \text{if } x<0,\\ (0.001,0,1) & \text{if } x\ge0. \end{cases} \tag*{(35)}

The final time for this case is t=1×10−4t=1\times10^{-4}. The number of cells is N=512N=512.

To further highlight the advantages of the NSFR method, it is noted that to run this case using the SSPRK3 Strong DG method, both the PPL and TVD limiters are required. The resulting solution is similar to those seen in § 4.2.1, where the use of the TVD limiter leads to dissipative error in the vicinity of the shock and contact discontinuity, which does not accurately reflect that which is seen in literature.

The CFL\mathrm{CFL} number used to run this case is 0.06 to best compare the results with the cDGc_{\mathrm{DG}} scheme. The cDGc_{\mathrm{DG}} scheme is used to run the Modified 1D Leblanc Shock Tube case. The maximum CFL\mathrm{CFL} number for this case using cDGc_{\mathrm{DG}} is 0.06. The solution is shown in Figure 10. The result captures the location and magnitude of the shocks very well. The solution follows closely with results seen in Zhang and Shu [43] and Xu and Shu

Image not extracted

Figure 10. [Leblanc Shock Tube] Density (left) and pressure (right) with NSFR-CHRA_{\mathrm{RA}}, Roe dissipation, cDGc_{\mathrm{DG}}, p=3p = 3, CFL 0.06 and PPL at t=0.0001st = 0.0001s

[53] This test case shows that the PPL can be used in conjunction with NSFR schemes to ensure that the positivity preservation property is satisfied.

Investigation of Flux Reconstruction Schemes

To demonstrate the impact of variants of flux reconstruction schemes, the case is also run using cSDc_{SD}, cHUc_{HU}, and c+c_{+}. The maximum CFLCFL number for each of the flux reconstruction parameters is 0.28, 0.29, 0.3 respectively. The three schemes have similar density and pressure profiles, so to compare the impact of the flux reconstruction parameter, the solution in the wake of the shock (5≤x≤8.755 \le x \le8.75) is presented for all the schemes in Figure 11. When comparing the different solutions, it can be seen that the oscillations in the wake of the shock are dampened for both density and pressure when c>0c > 0 despite being run with a CFLCFL that is nearly five times greater. As the value of cc increases, the oscillations become increasingly dampened, while preserving the magnitude and location of the shock. While the increase in the value of cc can dampen the oscillations, it also results in a slight undershoot at the foot of the rarefaction wave similar to the Sod Shock Tube results.

Image not extracted

Figure 11. [Leblanc Shock Tube] Density (left) and pressure (right) with NSFR-CHRA_{\mathrm{RA}}, Roe dissipation, p=3p = 3, using the PPL, comparing cDGc_{\mathrm{DG}}, cSDc_{\mathrm{SD}}, cHUc_{\mathrm{HU}}, c+c_{+} for 4≤x≤104 \le x \le10 at t=0.2st = 0.2s

The previous test configuration shows the CFLCFL advantage associated with an increase in the value of cc, but the results also exhibit better shock capturing capabilities as the oscillations in the vicinity of the shock are mitigated. To further show the added advantage of shock capturing, the tests for the different

ent flux reconstruction parameters are run again with a CFLCFL of 0.06 for all values of cc. The results are shown in Figure 12. While we still observe overshoot at the shock and undershoot at the foot of the rarefaction wave for the FR parameters that are greater than zero, we also see that the oscillations previously seen in the wake of the shock are almost entirely eliminated at a lower CFLCFL. This highlights the effectiveness of flux reconstruction schemes for shock capturing and demonstrates the potential of an adaptive approach, where the FR parameter is tuned to maximize timestep efficiency while suppressing oscillations.

Image not extracted

Figure 12. [Leblanc Shock Tube] Density (left) and pressure (right) with NSFR-CHRA, Roe dissipation, p=3p = 3, using the PPL, comparing cDGc_{DG}, cSDc_{SD}, cHUc_{HU}, c+c_{+} for 4≤x≤104 \le x \le10 with CFL=0.06CFL = 0.06 at t=0.0001st = 0.0001s

Strong Vortex Shock Wave Interaction

The implementation of the PPL in two dimensions is verified using the Strong Vortex Shock Wave (SVSW) interaction case [54] which is a benchmark case from the 5th international workshop on high-order CFD methods for the American Institute of Aeronautics and Astronautics conference led by Galbraith et al. [55]. This case can be run using the SSPRK3 Strong DG method with the limiter; however, the results contain severe oscillations and are not consistent with published results. To run this case with the SSPRK3 Strong DG method, both the PPL and a TVD limiter would be required. The NSFR scheme, however, can be used to run this test even without the application of the PPL. This allows us to compare the solutions obtained with and without the PPL to further confirm that the limiter does not affect the accuracy of the underlying scheme.

The computational domain is [0,2]×[0,1][0, 2] \times[0, 1] with a grid of Nx×Ny=300×150N_x \times N_y = 300 \times150. The initial condition of the problem involves a stationary shock (Ms=1.5M_s = 1.5) at x=0.5x = 0.5 and a strong vortex (Mv=0.9M_v = 0.9) centered at (xc,yc)=(0.25,0.5)(x_c, y_c) = (0.25, 0.5). The left boundary is a supersonic inlet, the right boundary is a subsonic outlet and the bottom and top boundaries are walls. The final time for this simulation is t=0.7st = 0.7s.

Verification of Limiter Accuracy

The tests are run with the maximum possible CFLCFL before failure. Without the limiter, the test can be run with the cDGcDG scheme for a CFLCFL of 0.27. With the limiter, the test can be run with a CFLCFL of 0.58 which is a little over two times greater. The result obtained using the cDGcDG scheme without the PPL is shown in Figure 13. As expected, the strong vortex is split into two separate vortices due to the compression effects when passing through the shock. The solution contains cylindrical acoustic waves downstream as expected and the shock also cuts off the sound waves centered on the vortex core, resulting in alternating expansion and compression regions.

Image not extracted

Figure 13. [SVSW] Density solution with polynomial degree 3, NSFR-CHRA, Roe dissipation, cDGcDG at t=0.07st = 0.07s (Left), t=0.15st = 0.15s (Middle) and t=0.7st = 0.7s (Right)

The result obtained with the limiter is visually identical to the results in Figure 13. The solution contains the expected features, including the two vortex cores, cylindrical acoustic waves, and cutoff sound waves. The solution closely reflects the solution obtained without the limiter, despite being run with a CFLCFL that is a little over two times greater. To better demonstrate that the limiter does not affect the underlying scheme, density line plots are shown. The first density plot is along the line y=0.4y = 0.4 and the second plot is along the line x=1.05x = 1.05. The first plot shows the stationary shock and the first vortex center, while the second plot shows the second vortex center. The plots are shown in Figure 14. In both plots, the solution with the limiter follows closely with the unlimited solution and reflects all the features accurately. The only noticeable difference is that the magnitude of the limited solution density is smaller at certain locations. As the limited solution reflects the unlimited solution closely despite having a much larger CFLCFL, it is evident that the limiter does not affect the underlying scheme and its accuracy.

Image not extracted

Figure 14. Density solution with NSFR-CHRACH_{RA}, Roe dissipation, cDGc_{DG}, with and without PPL at t=0.7t = 0.7s. Plot along y=0.4y = 0.4 (left) and x=1.05x = 1.05 (right).

Comparison with Strong DG

To highlight the difference between the SSPRK3 Strong DG method and the NSFR method, the Strong DG results obtained with the Lax-Friedrichs Flux and PPL with a CFL of 0.2 are shown in Figure 15 along with the plot of the entropy over time. The entropy over time for the cDGc_{DG} results with and without the limiter are nearly identical, whereas the SSPRK3 Strong DG solution has increasing entropy over time. This lack of entropy stability is reflected in the Strong DG results, which contain severe oscillations.

Image not extracted

Figure 15. Top: Density solution with the SSPRK3 Strong DG method at t=0.7t = 0.7s. Bottom: Plot of entropy over time for the NSFR-CHRACH_{RA}, Roe dissipation, cDGc_{DG} method with and without the PPL and SSPRK3 Strong DG method with the PPL

CFL Condition for Two-Point Flux

This test is also used to numerically verify that the two-point flux abides by the CFLCFL conditions, Eq. (29) and 30, in two dimensions. The number of cells for this test is Nx×Ny=202×101N_x \times N_y = 202 \times101. The test is run for polynomial degrees p=2,3,4p = 2, 3, 4 with an increment of 0.00005s for the constant time step. Once again, the CFLCFL for a given configuration is calculated in two different ways. The first method uses the minimum Δx\Delta x between the GLL quadrature nodes at the given degree for Δx\Delta x. The second method uses the cell size for Δx\Delta x. The CFLCFL is calculated using Eq. (33). The results of the numerical study are given in Table 7.

kkMin.Cell WidthTime StepMax. λ\lambdaMin. dx CFLCell Width CFL
20.004950.009900990.000953.900.750.37
0.0013.900.790.39
0.001053.900.830.41
0.00114.230.940.47
30.002740.009900990.000453.900.640.18
0.00053.900.710.20
0.000553.900.780.22
0.00069.372.050.57
40.001710.009900990.000253.900.570.10
0.00033.900.680.12
0.000353.900.800.14
0.00046.751.580.27

Table 7. Two-point flux positivity preservation test using the Strong Vortex Shock Wave setup with Nx×Ny=202×101N_x \times N_y = 202 \times101. NSFR - CHRACH_{RA} - Roe Dissipation - cDGc_{DG} - Nx×Ny=202×101N_x \times N_y = 202 \times101.

From the previous results, it is evident that the 1D condition can be relaxed up to a CFLCFL of 1. For higher dimensions, this condition still applies as demonstrated by this case wherein the test only fails if the CFLCFL is close to or greater than 1. For the second method, the p=2p = 2 configuration fails when the CFLCFL is in the range of 0.5 as given by Eq. (30) and the other polynomial degrees fail much earlier as observed in the 1D cases. This test case numerically validates the condition that positivity is preserved in 2D if the CFLCFL is ≤1\le1 as given in Eq. (29), irrespective of polynomial degree.

Impact of Flux Nodes

The application of the PPL for the NSFR scheme with GL flux nodes is also investigated using this test case. For this test, the CFL is lowered to 0.07 due to the sensitivity of the entropy-projected conservative variables at surface points as discussed for the Shu Osher case in § 4.3. Additionally, the grid size is changed to 302×151302 \times151 to ensure that the stationary shock is not at the boundaries of the cell. The result obtained with the GL flux nodes is shown in Figure 16, and it is visually identical to the results in Figure 13.

Image not extracted

Figure 16. [SVSW] Density solution with NSFR-CHRA\mathrm{CH}_{RA}, Roe dissipation, cDGc_{DG}, PPL using GL flux nodes at t=0.7t = 0.7s

The solution contains the expected features, including the two vortex cores, cylindrical acoustic waves, and cutoff sound waves. This case is a good verification that the limiter works as intended when GL flux nodes are used.

The density plots along the lines y=0.4y = 0.4 and x=1.05x = 1.05 are also provided for the GL results. To best compare the collocated and uncollocated schemes, the GLL test is rerun for a CFL of 0.07. The plots are shown in Figure 17. Both plots accurately reflect the solution features, but differences can be seen in magnitude for the vortex in the plot along x=1.05x = 1.05. The GLL solution also exhibits slightly more oscillatory behaviour. The GL solution has fewer oscillations as a result of the greater integration strength of GL nodes. Overall, this case demonstrates that the implementation of the PPL behaves as expected for GL flux nodes.

Image not extracted

Figure 17. [SVSW] Density solution with NSFR-CHRACH_{RA}, Roe dissipation, cDGc_{DG}, with GLL and GL flux nodes at t=1.8t = 1.8s. Plot along y=0.4y = 0.4 (left) and x=1.05x = 1.05 (right).

Two-Point Flux Comparison

This test is also used to investigate the behaviour of different two-point fluxes for strong shock problems. To best demonstrate that the behaviour of the fluxes differs in the presence of shocks, the SVSW test is modified slightly. The domain is expanded to [0,4]×[0,1][0, 4] \times[0, 1], the stationary shock is placed at x=2.0x = 2.0, and the final time is set to t=1.75t = 1.75s. This allows for the vortex to be tracked for a longer period of time before it interacts with the shock and will enable better comparison of the four fluxes. As the domain has been expanded, the grid size is also changed to 200×50200 \times50 to ensure that Δx=Δy\Delta x = \Delta y. A plot of density along y=0.425y = 0.425 is shown for 6 different time steps in Figure 18.

Image not extracted

Figure 18. [Extended SVSW] Plot of density along y=0.425y = 0.425 with NSFR, Roe dissipation, cDGc_{\mathrm{DG}} using different entropy conserving fluxes. Top row: t=0.00st = 0.00\mathrm{s} (Left), t=0.50st = 0.50\mathrm{s} (Middle) and t=1.00st = 1.00\mathrm{s} (Right). Bottom row: t=1.25st = 1.25\mathrm{s} (Left), t=1.50st = 1.50\mathrm{s} (Middle) and t=1.75st = 1.75\mathrm{s} (Right)

The plots of the density before the vortex interacts with the shock show that in the region upstream of the shock, the different fluxes produce identical results. Downstream of the shock, oscillations are observed of slightly varying magnitudes. The plots in the bottom row show the density along y=0.425y = 0.425 after the vortex interacts with the shock. The results for the IR, CH and CHRA\mathrm{CH}_{RA} fluxes are relatively similar. Comparing the KG flux to the three aforementioned fluxes, it is obvious that the results obtained with the KG flux are more oscillatory than the rest. The behaviour of KG is similar to that of the SSPRK3 Strong DG method.

To verify that this behaviour is related to entropy stability, a plot of entropy over time for all four fluxes is presented in Figure 19. The plot for the entropy over time demonstrates that the added dissipation of the KG flux is insufficient to ensure entropy stability, as expected. This leads to increasing entropy over time, which presents itself as severe oscillations in the solution similar to those seen in the Strong DG results.

Image not extracted

Figure 19. [Extended SVSW] Plot of entropy over time with NSFR, Roe dissipation, cDGc_{DG}, PPL for the different entropy conserving fluxes.

Shock Diffraction

The shock diffraction case is a 2D problem of a right-moving shock that passes a backward-facing corner. The Mach 5.09 shock diffracts at a 90° edge. The solution is expected to have several key features. These features include the incident shock, the diffracted shock, the reflected expansion wave and the vortex roll-up as highlighted in [56]. This case is a good test of the positivity-preserving strategy as it is easy to get nonphysical values below the 90° corner and to the right of the corner. The final time used is T=2.3T = 2.3. The computational domain is [0,1]×[6,11]∪[1,13]×[0,11][0,1] \times[6,11] \cup[1,13] \times[0,11]. The initial condition is

(ρ,u,v,p)={(ρleft,uleft,vleft,pleft)if x<0.5(1.4,0,0,1)if x>0.5(36)(\rho,u,v,p) = \begin{cases} (\rho_{\mathrm{left}},u_{\mathrm{left}},v_{\mathrm{left}},p_{\mathrm{left}}) & \text{if } x < 0.5 \\ (1.4,0,0,1) & \text{if } x > 0.5 \end{cases} \tag*{(36)}

Where ρleft=7.041132906907898\rho_{\mathrm{left}} = 7.041132906907898, uleft=4.07794695481336u_{\mathrm{left}} = 4.07794695481336, vleft=0v_{\mathrm{left}} = 0, pleft=30.05945p_{\mathrm{left}} = 30.05945. The conditions for the left boundary is inflow for x=0x = 0, 6≤y≤116 \le y \le11 and wall boundary for x=1x = 1, 0≤y≤60 \le y \le6. The bottom boundary is a wall for 0≤x≤10 \le x \le1, y=6y = 6 and outflow for (y=0,1≤x≤13)(y = 0, 1 \le x \le13). The right boundary, x=13x = 13, 0≤y≤110 \le y \le11, is outflow and the top boundary, y=11.0y=11.0, 0≤x≤130 \le x \le13, is a wall boundary. All the schemes are run with a CFLCFL of 0.45.

The test fails due to nonphysical values at the given CFL when using the PPL with only the modifications implemented by Wang et al. [44]. However, with the modifications introduced in this paper, we are able to obtain results that are comparable to the solution in Zhang and Shu [43], which uses a TVB limiter in addition to the original PPL. The results with the modified limiter are shown in Figure 20.

Image not extracted

Figure 20. [Shock Diffraction] Density solution with NSFR-CHRA\mathrm{CH}_{RA}, cDGc_{DG}, Roe dissipation, p=3p = 3, PPL at t=2.3t = 2.3s

Investigation of Flux Reconstruction Schemes

This test is also used to compare the impact of various flux reconstruction schemes. Four different schemes are run including cDGc_{DG}, cSDc_{SD}, cHUc_{HU}, and c+c_{+}. It is run with a grid of 416×352416 \times352. The final result for each configuration is shown in Figure 21. All four results reflect the expected features, including the vortex roll-up and Kelvin-Helmholtz instabilities (KHI). It should be noted that the cDGc_{DG} solution has more noise across the solution and a greater overshoot at the shock compared to the other three schemes.

Image not extracted

Figure 21. [Shock Diffraction] Density solution with NSFR ⁣− ⁣CHRA\mathrm{NSFR\!-\!CH}_{\mathrm{RA}}, Roe dissipation, p=3p = 3, PPL at t=2.3st = 2.3\mathrm{s}, using different correction parameters - cDGc_{\mathrm{DG}} (top left), cSDc_{\mathrm{SD}} (top right), cHUc_{\mathrm{HU}} (bottom left), c+c_{+} (bottom right).

To demonstrate the noise present in the cDGc_{DG} solution, a plot of density along the line (1,0)→(13,11)(1, 0) \to(13, 11) is provided for all the schemes in Figure 22. The plot is also zoomed in to the wake of the shock to better show the difference in the solutions. In the areas of the plot that are upstream of the shock, comparing the schemes is difficult due to the presence of the KHI. In the plot focusing on the wake of the shock, we can see that there is a significant amount of oscillations in the cDGc_{DG} solution compared to the other solutions. In addition to the oscillations, the cDGc_{DG} solution also has the greatest overshoot at the shock. The overshoot observed at the shock is mitigated as the correction parameter increases in value.

Image not extracted

Figure 22. [Shock Diffraction] Plot of density across (1,0)→(13,11)(1, 0) \to(13, 11) with NSFR, Roe dissipation, PPL at t=2.3st = 2.3\mathrm{s} using different flux reconstruction correction parameters (left) with a closer look at the wake of the shock (right)

The test configurations for this case so far have demonstrated the effects of increasing the cc value with the CFLCFL kept consistent across all runs. Next, we change the CFLCFL for the c+c_{+} configuration to show that the oscillations are mitigated even for a larger timestep. This test is run for the maximum CFL before failure for c+c_{+}, which is 0.61. The density plot along the line (1,0)→(13,11)(1, 0) \to(13, 11) is shown in Figure 23. In addition to the higher CFLCFL case, the cDGc_{DG} results from the previous run are also shown for comparison. The plots show that the increase in cc value continues to mitigate oscillations and overshoot, providing a better solution than cDGc_{DG}, despite the significant increase in the time step size. The added robustness and the mitigation of oscillations by the c+c_{+} scheme make it a very attractive implementation for strong shock problems that involve complex features.

Image not extracted

Figure 23. [Shock Diffraction] Plot of density across (1,0)→(13,11)(1, 0) \to(13, 11) with NSFR, Roe dissipation, PPL at t=2.3st = 2.3\mathrm{s} (left) with a closer look at the wake of the shock (right)

Double Mach Reflection

The Double Mach Reflection (DMR) problem introduced by Woodward and Colella involves sending a Mach 10 shock into a reflecting wall. The original physical problem has the shock moving horizontally into a wall inclined at a 30∘30^\circ angle. However, the problem is generally solved with a shock moving diagonally into a horizontal wall on the rectangular domain. The rectangular domain is [0,4]×[0,1][0,4] \times[0,1] with the initial condition,

(ρ,u,v,p)={(8,3338,−338,116.5)if y>3(x−16)(1.4,0,0,1)if y<3(x−16)(37)(\rho,u,v,p) = \begin{cases} \left(8,\frac{33\sqrt{3}}{8},-\frac{33}{8},116.5\right) & \text{if } y > \sqrt{3}\left(x-\frac{1}{6}\right) \\ (1.4,0,0,1) & \text{if } y < \sqrt{3}\left(x-\frac{1}{6}\right) \end{cases} \tag*{(37)}

The left boundary is inflow, and the right boundary is outflow. The bottom boundary corresponding to 0≤x<160 \le x < \frac{1}{6} has the post-shock condition. The bottom boundary corresponding to 16<x≤4\frac{1}{6} < x \le4 is a wall boundary. In most implementations of the DMR case, the top boundary is set to a function that follows the motion of the shock. To minimize implementation effort, the domain is extended in the yy-direction such that the top boundary doesn’t affect the domain of the test case. Thus, the computational domain is [0,4]×[0,3][0,4] \times[0,3]. The final time is t=0.2t = 0.2s. As outlined in Vevek et al. [57], the solution at the final time is a self-similar shock structure with two triple points and two slip lines. The primary triple point is where the incident shock, reflected shock, and Mach stem meet, while the secondary triple point is where the reflected shock, secondary Mach stem, and secondary reflected shock meet. The solution also contains two slip lines. The reflected shock interacts with the primary slip line and generates KHI along the slip line. A jet is also present near the wall. The complex features are best captured using a high-order scheme; thus, the case serves as a good indicator of the resolution of a scheme.

Running this test in the configurations given in the following subsections using the limiter in [44], results in the test failing due to nonphysical values. However, with the new modifications, we are able to obtain results that are comparable to solutions seen in literature. The timestep efficiency introduced by this modification is evident in this case and the Shock Diffraction case.

Polynomial Degree Comparison

The DMR case is first run with increasing polynomial orders while maintaining an equivalent number of degrees of freedom to demonstrate the ability to use higher order polynomials for cases involving strong shocks. The polynomial degrees used in this case are p=3,4,5p = 3,4,5, and a CFLCFL number of 0.15 is chosen for all degrees. These results are obtained using relatively coarse grids with 240 ×\times 60, 192 ×\times 48, and 160 ×\times 40 for the test space for each polynomial degree respectively. The results are shown in Figure 24.

Image not extracted

Figure 24. [DMR] Density solution with NSFR, Roe dissipation, cDGc_{DG}, PPL at t=0.2st = 0.2\mathrm{s}, using different polynomial degrees - 3 (top), 4 (middle), 5 (bottom)

Looking at the p=3p = 3 results first, we see all the expected solution features are present despite the coarseness of the grid. The solution contains two triple points, one located at x≈2.75x \approx2.75 and the other at x≈2.2x \approx2.2. There are also two slip lines present. The first slip line is prominent; where, the interaction of the reflected shock and the slip line is evident from the KHI that are generated. The secondary slip line is often

difficult to discern in DMR solutions, and the presence of the slip line is a good indicator of the resolution of the scheme. In the p=3p = 3 solution, the secondary slip line is present and can be identified at x≈1.85x \approx1.85 at the bottom boundary or at x≈2.6x \approx2.6 at the top boundary.

As the polynomial degree is increased, we expect that there will be an increase in oscillatory behaviour as demonstrated for the Sod Shock Tube case in Sec.4.2.2. Despite the presence of the spurious oscillations, it is expected that the solutions still contain all the expected features. In both the p=4p = 4 and p=5p = 5 solutions, it is evident that all the expected solution features are present and follow closely with the features described in the p=3p = 3 results. The secondary slip line is also clearly present in the solution despite the increase in oscillatory behaviour seen in the area under the secondary reflected shock. This test case demonstrates that higher polynomial degrees can be used for cases involving strong shocks.

Investigation of Flux Reconstruction Schemes

In addition to the study of polynomial degrees, the DMR case is also run with different flux reconstruction schemes to compare the effects of the cc value on the complex features. These tests are run with a finer grid of 720×180720 \times180 for the test space (720×540720 \times540 for the full computational domain). The CFLCFL for cDGc_{DG} is 0.5 and for c+c_{+} it is 0.58, which is the maximum CFLCFL for which the case can be run for the respective schemes. The results are shown in Figure 25.

Density solution with NSFR, Roe dissipation, PPL at $t = 0.2$s, using different flux reconstruction parameters

Figure 25. [DMR] Density solution with NSFR, Roe dissipation, PPL at t=0.2t = 0.2s, using different flux reconstruction parameters - cDGc_{DG} (top), c+c_{+} (bottom)

In both sets of results, there are two triple points and two slip lines present. The instabilities along primary slip line, as well as the jet near the wall, are well defined. When comparing the primary slip line for the cDGc_{DG} and c+c_{+} solutions, it is evident that the cDGc_{DG} slip line is much more oscillatory due to the noise in the solution. As a result, the KHI for the cDGc_{DG} solution is not as distinct as the KHI in the c+c_{+} solution, which is unaffected by noise and has a distinct slip line without significant oscillations. The difference in noise between the two sets of results leads to another notable distinction between the two solutions. In the cDGc_{DG} solution, the secondary slip line is difficult to discern due to the noise in its vicinity, whereas the slip line is distinct and clearly visible in the c+c_{+} solution due to the noise being filtered out. While both schemes are able to capture the main features of the solution accurately, the c+c_{+} solution presents a clear advantage due to the distinct features and dampened noise.

High Mach Astrophysical Jet

The High Mach Astrophysical Jet test case is an extreme benchmark to demonstrate the robustness of entropy-stable high-order schemes. This test is used to verify the robustness of the NSFR scheme with the PPL in two dimensions and demonstrates the advantages of the flux reconstruction scheme. This case was originally proposed by [58] and it simulates an astrophysical jet with Ma≈80Ma \approx80. Another version of this case is considered in [43] with an astrophysical jet of Ma≈2000Ma \approx2000. The two versions are outlined in this section, along with the results obtained using the NSFR scheme. Due to the high Mach numbers, this test is often run with either a TVB limiter as seen in [43] or with the imposition of TVD-like bounds as seen in [31]. Without the use of these strategies, running either version of this case using the SSPRK3 Strong DG method or the cDGc_{DG} scheme would require an extremely small CFLCFL number. A CFLCFL number as low as 0.0005 was tested for both, but still failed due to nonphysical values. This case demonstrates the added robustness of the flux reconstruction, c+c_{+} scheme, and the advantages of higher flux reconstruction correction parameters in extreme cases such as this.

Mach 80 Astrophysical Jet

The computational domain for the Mach 80 astrophysical jet is [0,2]×[−0.5,0.5][0, 2] \times[-0.5, 0.5] with the initial condition

(ρ,u,v,p)=(0.5,0,0,0.4127).(38)(\rho, u, v, p) = (0.5, 0, 0, 0.4127). \tag*{(38)}

The top, right and bottom boundaries are outflow boundaries. The left boundary is set to (ρ,u,v,p)=(0.5,30,0,0.4127)(\rho, u, v, p) = (0.5, 30, 0, 0.4127) for y∈[−0.05,0.05]y \in[-0.05, 0.05] and (ρ,u,v,p)=(0.5,0,0,0.4127)(\rho, u, v, p) = (0.5, 0, 0, 0.4127) otherwise. The final time is t=0.07t = 0.07, and Nx×Ny=320×160N_x \times N_y = 320 \times160 with γ=53\gamma= \frac{5}{3}.

Using the added robustness of the c+c_{+} scheme, this case was run for a CFLCFL number of 0.01. The results are shown in Figure 26. Even without the use of TVD limiting, the positivity of density and pressure is preserved. The results contain the expected solution features, including a bow shock, terminal Mach disk in the tip of the jet, and KHI.

Image not extracted

Figure 26. [Astrophysical Jet Mach 80] Density solution with NSFR-CHRACH_{RA}, Roe dissipation, c+c_{+}, PPL at t=0.07st = 0.07s.

Mach 2000 Astrophysical Jet

The computational domain for the Mach 2000 astrophysical jet is [−0.5,0.5]2[-0.5, 0.5]^2 with the initial condition

(ρ,u,v,p)=(0.5,0,0,0.4127)(39)(\rho, u, v, p) = (0.5, 0, 0, 0.4127) \tag*{(39)}

The top, right, and bottom boundaries are outflow boundaries. The left boundary is set to (ρ,u,v,p)=(0.5,800,0,0.4127)(\rho, u, v, p) = (0.5, 800, 0, 0.4127) for y∈[−0.05,0.05]y \in[-0.05, 0.05] and (ρ,u,v,p)=(0.5,0,0,0.4127)(\rho, u, v, p) = (0.5, 0, 0, 0.4127) otherwise. The hypersonic inflow condition on the left boundary corresponds to a Mach number of Ma=2156.91Ma = 2156.91. The final time is t=0.001t = 0.001, and Nx×Ny=300×300N_x \times N_y = 300 \times300 with γ=53\gamma= \frac{5}{3}.

Similar to the Mach 80 Astrophysical Jet case, due to the lack of TVD limiting, running this case with the SSPRK3 Strong DG method and cDGc_{DG} method is not feasible. Using the c+c_{+} flux reconstruction parameter also requires an extremely small CFLCFL number to run successfully. To allow for a higher CFLCFL number, the flux reconstruction parameter is increased to ten times the value of c+c_{+}. With c=3.76E−2c = 3.76E - 2 and a CFLCFL number of 0.0005, the case is successfully run to completion while preserving the positivity of density and pressure. The results are shown in Figure 27. The features of the solution are well captured, including the bow shock, terminal Mach disk and the unstable rollups. In previous tests, such as the DMR case, variants of flux reconstruction schemes were proven to have an advantage due to the dampening of noise, which produces a cleaner solution. In the Mach 80 and Mach 2000 cases, it is shown that the flux reconstruction scheme with higher values of the parameter, cc, also provides the added advantage of robustness.

Density solution for Astrophysical Jet Mach 2000

Figure 27. [Astrophysical Jet Mach 2000] Density solution with NSFR-CHRA_{\mathrm{RA}}, Roe dissipation, c+×10c_{+} \times10, PPL at t=0.001t = 0.001s.

Conclusion

In this paper, the NSFR schemes [22, 23], which have been successfully applied to unsteady compressible flows in arbitrary curvilinear coordinates, have been extended to shock-dominated problems using a modified implementation of the positivity-preserving limiter originally developed by Zhang and Shu [43]. Building on the improvements by Wang et al. [44], the limiter is further modified to detect minimum values of density and pressure at the solution nodes, addressing the limitation of the original approach, which ensured positivity only at the quadrature nodes but not at the solution nodes. The enhancements to the positivity-preserving limiter were required to simulate all two-dimensional strong shock-dominated cases. The newly modified limiter guarantees positivity for the entire duration of the test at the same grid size and time step. With the robustness enhanced, this limiter is used to investigate the shock-capturing capabilities of the NSFR scheme. In the 1D Sod Shock Tube and 2D Strong Vortex Shock Wave (SVSW) Interaction cases, the advantages of the NSFR scheme over standard DG methods is established as the scheme can be used to run these cases without the limiter while the DG method requires the use of the PPL and a TVD limiter. In the 1D Leblanc Shock Tube case, the advantages of the NSFR method are clearly demonstrated, specifically the results for the c+c_{+} FR scheme, which presents an essentially oscillation-free solution despite the extreme nature of this benchmark problem. The 2D Low Density case highlights the high-order accuracy of the scheme and verifies that the modified limiter preserves high-order accuracy. The shock-dominated cases also present a thorough analysis of the impact of two-point fluxes, quadrature nodes and variants of flux reconstruction schemes through the choice of the cc parameter. For the two-point flux investigation, the SVSW showcases the insufficient dissipation of the KG flux, which leads to increasing entropy. Additionally, the 1D Sod Shock Tube, 1D Shu Osher and 2D SVSW case are also used to numerically verify a CFL condition to preserve positivity for two-point fluxes. In the case of quadrature nodes, the GL quadrature is preferred as it has a stronger quadrature strength; however, due to the sensitivities of the entropy projections at the face, the GLL quadrature presents a significant timestep advantage over GL nodes as showcased in the 1D Shu Osher and 2D SVSW cases. For the FR parameter, all the test cases demonstrate that the timestep efficiency is maximized while suppressing oscillations as the value of the FR parameter is increased. The advantage of flux reconstruction is best highlighted in the astrophysical Mach jet case. With an FR scheme with cc being ten times greater than that of c+c_+, the flow can be resolved without the need for TVD limiting.

It has been demonstrated that FR schemes maximize the time step and increase robustness at the cost of reduced accuracy in smooth regions of the flow. This is due to the filtering of the highest mode as FR schemes with larger values of cc are considered. The DG method is certainly more accurate, and we wish to preserve this accuracy in smooth regions while also mitigating oscillations in the vicinity of shocks, maximizing the time step, and ensuring robustness. Thus, the ability to scale the FR parameter based on local discontinuities represents a key advantage of the NSFR scheme, making an adaptive implementation of this parameter a promising direction for future research. In addition, an investigation of the NSFR scheme’s performance for viscous shock problems and shock-turbulence interaction cases is part of ongoing research to further develop the approach to be suitable for a wide range of flows.

Acknowledgments

Sai Shruthi Srinivasan thanks the Vadasz Family Foundation and McGill Engineering Doctoral Award (MEDA). The authors would also like to thank Alexander Cicchino for helpful discussions throughout the course of this work. This research was enabled in part by support provided by Calcul Quebec and the Digital Research Alliance of Canada.

References

  1. [40]Journal of Computational Physics 378 (2019) 366–393.
  2. [41]J. Chan, Skew-symmetric entropy stable modal discontinuous Galerkin formulations, Journal of Scientific Computing 81 (2019) 459–485.arxiv.org/abs/1902.01828
  3. [42]A. Cicchino, Weight-adjusted nonlinearly stable flux reconstruction high-order methods for compressible flows in curvilinear coordinates (2024).DOI
  4. [43]X. Zhang, C.-W. Shu, On positivity-preserving high order discontinuous galerkin schemes for compressible euler equations on rectangular meshes, Journal of Computational Physics 229 (2010) 8918–8934.DOI
  5. [44]C. Wang, X. Zhang, C.-W. Shu, J. Ning, Robust high order discontinuous galerkin schemes for two-dimensional gaseous detonations, Journal of Computational Physics 231 (2012) 653–665.DOI
  6. [45]H. Ranocha, G. J. Gassner, Preventing pressure oscillations does not fix local linear stability issues of entropy-based split-form high-order schemes, Communications on Applied Mathematics and Computation (2021) 1–24.
  7. [46]H. Ranocha, Comparison of some entropy conservative numerical fluxes for the euler equations, J. Sci. Comput. 76 (2018) 216–242.DOI
  8. [47]P. Castonguay, High-order energy stable flux reconstruction schemes for fluid flow simulations on unstructured grids, Ph.D. thesis, Citeseer, 2012.
  9. [48]C.-W. Shu, S. Osher, Efficient implementation of essentially non-oscillatory shock-capturing schemes, Journal of computational physics 77 (1988) 439–471.DOI
  10. [49]F. Ismail, P. L. Roe, Affordable, entropy-consistent euler flux functions ii: Entropy production at shocks, Journal of Computational Physics 228 (2009) 5410–5436.DOI
  11. [50]C. A. Kennedy, A. Gruber, Reduced aliasing formulations of the convective terms within the navier–stokes equations for a compressible fluid, Journal of Computational Physics 227 (2008) 1676–1700.
  12. [51]P. Chandrashekar, Kinetic energy preserving and entropy stable finite volume schemes for compressible euler and navier-stokes equations, Communications in Computational Physics 14 (2013) 1252–1286.DOI
  13. [52]E. Johnsen, J. Larsson, A. V. Bhagatwala, W. H. Cabot, P. Moin, B. J. Olson, P. S. Rawat, S. K. Shankar, B. Sjögreen, H. Yee, X. Zhong, S. K. Lele, Assessment of high-resolution methods for numerical simulations of compressible turbulence with shock waves, Journal of Computational Physics 229 (2010) 1213–1237.DOI
  14. [53]Z. Xu, C.-W. Shu, Third order maximum-principle-satisfying and positivity-preserving lax-wendroff discontinuous galerkin methods for hyperbolic conservation laws, Journal of Computational Physics 470 (2022) 111591.DOI
  15. [54]A. Rault, G. Chiavassa, R. Donat, Shock-vortex interactions at high mach numbers, Journal of Scientific Computing 19 (2003) 347–371.DOI
  16. [55]M. Galbraith, S. Murman, C. Kim, P. Persson, K. Fidkowski, R. Glasby, K. Hillewaert, B. Ahrabi, 5th international workshop on high-order CFD methods, in: AIAA science and technology forum and exposition, 2017.
  17. [56]R. Hillier, Computation of shock wave diffraction at a ninety degrees convex edge, Shock waves 1 (1991) 89–98.DOI
  18. [57]U. Vevek, B. Zang, T. H. New, On alternative setups of the double mach reflection problem, Journal of Scientific Computing 78 (2019) 1291–1303.DOI
  19. [58]Y. Ha, C. L. Gardner, A. Gelb, C.-W. Shu, Numerical simulation of high mach number astrophysical jets with radiative cooling, Journal of Scientific Computing 24 (2005) 29–44.DOI

Paper details

vlnc:W_r7ec82mn
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.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents