NLR-TP-2002-300

Slip flow boundary conditions in discontinuous Galerkin discretizations of the Euler equations of gas dynamics

J.J.W. van der Vegt and H. van der Ven

Nationaal Lucht- en Ruimtevaartlaboratorium National Aerospace Laboratory NLR

NLR-TP-2002-300

Slip flow boundary conditions in discontinuous Galerkin discretizations of the Euler equations of gas dynamics

J.J.W. van der Vegt and H. van der Ven

This report is based on a presentation held at the fifth World Congress on Computational Mechanics, Vienna - Austria on 7-12 July 2002. This report may be cited on condition that full credit is given to NLR and the authors.

Customer: Working Plan number: Owner: Division: Distribution: Classification title:

National Aerospace Laboratory NLR A.1.B.7 National Aerospace Laboratory NLR Fluid Dynamics Unlimited Unclassified May 2002

-3NLR-TP-2002-300

Summary
Discontinuous Galerkin discretizations of the slip ow boundary condition at curved walls in inviscid gas dynamics are not very accurate when linear basis functions are combined with elements with straight edges at the boundary. This is particularly true when the boundary integrals are computed with a Gauss quadrature rule, but also occurs when the more accurate Taylor quadrature rule is used. The error at the solid surface results in a boundary layer which can signicantly pollute the numerical solution. In this paper sources of these problems are analyzed and demonstrated for the subsonic ow about a circular cylinder. It is shown that the use of the recently developed Taylor quadrature rule for the ux integrals in combination with superparametric elements results in more than a factor three reduction in total pressure loss at the wall in comparison with isoparametric elements and Gauss quadrature. The effects of boundary curvature can also be removed using mesh adaptation. Local mesh renement of linear isoparametric elements is very effective in reducing the error at slip ow boundaries and provides a good alternative to the use of superparametric elements. This is possible because it is demonstrated that it is not necessary to use a higher-order boundary representation.

-4NLR-TP-2002-300

Contents
1 2 3 4 5 6 7 Introduction Governing equations Discontinuous Galerkin discretization Numerical ux function Flux quadrature Slip ow boundary Effect of wall curvature on the accuracy of the discontinuous Galerkin discretization 8 9 Results Conclusions 16 19 25 5 7 8 11 12 14

7 Figures

(26 pages in total)

-5NLR-TP-2002-300

1 Introduction
Discontinuous Galerkin nite element methods (DGFEM) provide a new numerical discretization technique for the Euler equations of gas dynamics. The DGFEM is a combination of an upwind nite volume scheme and a nite element method, with as key feature the use of local polynomial basis functions to represent the test and trial functions. These basis functions are only weakly coupled to neighboring elements, which make discontinuous Galerkin methods ideally suited for mesh adaptation using local mesh renement. The discontinuous Galerkin nite element method for hyperbolic conservation laws has been pioneered by Cockburn and Shu (for a general survey see (Ref. 4, 5)). Discontinuous Galerkin methods for the Euler equations with local mesh renement are described in Baumann and Oden (Ref. 3) and Van der Vegt and Van der Ven (Ref. 8). Recently, the DGFEM has been extended by Van der Vegt and Van der Ven (Ref. 9) to basis functions which are discontinuous in both space and time. This so-called space-time DGFEM is well suited for problems with time-dependent boundaries, which require moving and deforming meshes. Aerodynamic applications, such as oscillating and deforming wings, can be found in (Ref. 9, 10). In most of the ow domain DG nite element methods result in an accurate numerical discretization, but at solid surfaces, where a slip ow boundary condition must be applied in inviscid gas dynamics, DGFEM suffers from a signicant loss in accuracy. This was rst demonstrated by Bassi and Rebay (Ref. 1) for the smooth subsonic ow about a circular cylinder, where in the near wall region and wake of the cylinder a signicant shear layer occurred, which is incorrect for inviscid ow. Bassi and Rebay showed that using linear polynomials for the test and trial functions in combination with quadratic superparametric elements signicantly reduced this effect for triangular elements, but did not really give an explanation for this phenomenon. Recently, Van der Ven and Van der Vegt (Ref. 10) proposed a new so-called Taylor quadrature rule for the ux integrals in the space-time DGFEM to improve the numerical efciency of the method. The Taylor quadrature rule results in a factor three improvement in performance, but it was noticed that it also results in a signicantly reduced numerical shear layer for the ow about a circular cylinder, (Ref. 10). It seems that not only the choice of element type, namely isoparametric or superparametric, but also the ux quadrature rule and boundary discretization, directly affect the accuracy near a slip ow boundary. The main objective of this paper is to investigate different techniques to improve the accuracy in the near wall region and wake of inviscid ows with a slip ow boundary condition. For subsonic ow the error caused by the slip ow boundary condition dominates the overall numerical error

-6NLR-TP-2002-300

and also generates unphysical vorticity and entropy near the wall. An improved discretization of the slip ow boundary condition is of great importance for many applications which require good accuracy in the near wall region, such as in aerodynamics. In this paper we will focus on the effects of the ux quadrature rule, element type, and the calculation of the wall pressure in combination with the HLLC ux proposed by Toro et al. (Ref. 6). Also, we will consider local mesh renement based on the boundary curvature as an alternative technique to reduce the errors near a slip ow boundary. Since the problems with the slip ow boundary conditions in discontinuous Galerkin methods occur both in two and three-dimensions, and are independent of the time integration technique, we limit the present discussion to simple two-dimensional steady ows. The results are, however, directly applicable to space-(time) DGFEM in three dimensions. In the remainder of this paper we rst discuss the discontinuous Galerkin discretization for the Euler equations. Next, we discuss the evaluation of the element face and volume integrals using Gauss and Taylor quadrature rules, followed by the evaluation of the pressure at the slip wall. The effect of the different techniques is demonstrated with the subsonic ow about a circular cylinder.

-7NLR-TP-2002-300

2 Governing equations
The Euler equations for inviscid gas dynamics with initial and boundary conditions in a domain
W  Q  d  r

can be expressed as:

 @ 8  6  4 2      ( &        " $     

(1)
 @

vectors, which are dened as:

u u t " i h ef f u u t h ef f

with

the ratio of specic heats at constant pressure and constant volume. In this paper we use

the summation convention on repeated indices.

an equation of state, for which we assume a calorically perfect gas:

&

"

coordinate direction

, and

the Kronecker delta symbol. The Euler equations are closed with ,

,
p

the components of the velocity vector

,
i

denote the density, pressure and specic total energy, respectively, in the Cartesian

data,

the data at the initial time

, and

represents the nal time. In addition, ,

the boundary operator, with

& W V

the prescribed boundary

"

"

"

"

&

"

&

with

the vector of conserved quantities,

&

&

&

"

4 P 4 6 & P Q D

the ux

-8NLR-TP-2002-300

3 Discontinuous Galerkin discretization

For the discontinuous Galerkin nite element discretization the domain as
t  c v   b  "  b 

is approximated with a

with:
  p  e b  f &  p   6 }

which is dened as:

  "    p  j  " p  &  V  Q j V } j Q k

(3)

with:
  b e p   p  f &  p   6 }

The discontinuous Galerkin discretization uses basis functions in each element which are discontinuous across element faces. The DG basis functions are dened through the following steps: element
j

:
b ( & &  }

&

"

&

"

through the mapping

& Q  j  f }

, (2)-(3):

Dene
 j  f

, as the space of functions

associated to functions in

"

&

Dene

as the space of polynomials

of degree

on the reference

"

and

the vertices, midpoints of the edges, and center of the element

"

&

&

"

and

the vertices of element

, or through the superparametric mapping

"

"

"

&

&

isoparametric mapping
k

: (2)

The elements

are connected to a reference element

, either through the

"

&

. In addition, the elements

2 i

satisfy the condition

if

&

&

, with
V

the radius of the smallest circle containing all elements .

&

tessellation
" P j r p n 2 l k

of quadrilateral elements

, which are open domains and satisfy the condition

-9NLR-TP-2002-300

dened as:
 b &    6 V

as:
 j 

for


dened as: Find a



relation is satised:
j  i   " $ i k      i k e & j i  i k

ow calculations no stabilization operator was necessary. More details, including the derivation of the weak formulation for the space-(time) discontinuous Galerkin nite element method, can be found in Van der Vegt and Van der Ven (Ref. 9). The discontinuous Galerkin nite element discretization is now obtained if we introduce the polyformulation (5) and conduct the element integrations:
c  b  ( &  j  i   " $ p k      p p k e & j p k }  i  i

nomial representations for the test and trial functions

"

"

"

condition

, as the initial data of a Riemann problem. For the present subsonic

and

, given by (4), into the weak

"

"

traces

and

of

in the boundary integrals of the elements

introducing the numerical ux

. This is accomplished by considering the and , which satisfy the

a summation over all elements in

. Weak continuity at the element faces

with the test functions

, integration by parts over the element

, and nally is ensured by

(6)

The weak formulation is obtained by multiplying the Euler equations (1) in each element
 2

"

, such that
j

, and for all

&

The weak form in each element

for the Euler equations of gas dynamics can now be

the following

(5)

with

the projection operator to the nite-dimensional space

2

and

"

then the vector of conservative variables

can be approximated as

(4) .

"

&

&

Dene the function space

 j  i f

as:

&

"

"

"

Dene the space of functions

j Q "

&

&

 P

, with the basis functions

& Q    " 

&

 i

   i f 2 i

- 10 NLR-TP-2002-300

The resulting equations for the polynomial coefcients are solved using a Runge-Kutta time integration method in combination with a multigrid convergence acceleration technique. In this paper we only consider steady state solutions, but for implicit time-accurate calculations a pseudo-time integration technique is used to solve the non-linear equations for the expansion coefcients. For more details see (Ref. 9).

- 11 NLR-TP-2002-300

4 Numerical ux function
The introduction of the numerical ux introduces upwinding into the nite element discretization and any monotone and consistent (approximate) Riemann solver for the Euler equations of gas dynamics can be used as numerical ux. In the present research we use the HLLC ux proposed by Toro et al. (Ref. 2, 6), see also Toro (Ref. 7). This is motivated by the fact that this approximate Riemann solver combines good accuracy with a low computational cost. In addition, the HLLC ux combines well with the Taylor quadrature rule for the uxes to be discussed in the next section. The HLLC ux in a form suitable for the Taylor quadrature integration can be summarized as:
  V e } } i  i j   e b e } i b Q  e W e f } f i  & 2 & e e  h e }  i     h } i e } e   i   e  e e e  } i   } e $ e  h 2 e e } h i   &  e } i 2 e  W  h } b } $ i &  f f V & & j   & Q h  " i $  " d  j

vector at

The intermediate pressure

is equal to:

wave speeds

and

are dened as:

 i 

tively.

The sufces

and

refer to the internal and external ow states at the element boundary, respec-

with

the speed of sound. The contact wave speed

is equal to:

&

with

the normal velocity, and the vector

is dened as

&

&

with

the normal ux vector and are dened as:

the unit outward normal

. The coefcients

e e

&

. The

- 12 NLR-TP-2002-300

5 Flux quadrature
An important aspect in the discontinuous Galerkin discretization is the calculation of the ux integrals in (6). This inuences both the accuracy and efciency of the method. In general these integrals are computed with product Gauss quadrature rules, with two quadrature points in each coordinate direction. This ensures a sufciently accurate evaluation of the integrals, but also requires the evaluation of the numerical ux function at each quadrature point. Especially in threedimensions and for the space-time discontinuous Galerkin nite element method this quickly becomes computationally very expensive. In order to alleviate this problem the Taylor quadrature rule was proposed and analyzed by Van der Ven and Van der Vegt (Ref. 10). The Taylor quadrature rule has as main benet that it requires only one ux evaluation per element face or volume integral. Apart from being computationally more efcient it was also observed in (Ref. 10) that this integration technique results in a smaller error at solid walls with a slip ow boundary condition. In this and the remaining sections we will further investigate this and propose further improvements to the numerical discretization of slip ow boundary conditions. The rst step in the evaluation of the integrals in (6) is to transform the integrals on the righthand side into integrals over the reference element:
k ~  p  p W & " p }   2 p } " $ j " &  p  p    b  b e 

The Gauss quadrature rule now approximates the surface ux integrals as:
   k ~    p  & k ~  p  p

necessarily the same since the elements at both sides of the face might have a local coordinate system with a different orientation. The evaluation of the volume integrals is completely analogous. For the Taylor quadrature rule we split the numerical ux function
e     " $     f &    

into two parts:



the quadrature points at both sides of the element face. Note,

with
2   

the quadrature weights,

the number of Gauss quadrature points, and and are not

faces of the reference element

with

if

and

if

and

  

  

 

i 

 

 " } & $ p "  }  f k & 

 

  "

 p &

 " 

 k

 b k & b &

one of the four

- 13 NLR-TP-2002-300

now evaluated using a Taylor series expansion of the ux at the element face center:
 p  P       " $  p   P $ &  $  p

The integrals on the righthand side of (7) can be evaluated analytically and result in simple exthe polynomial representation of


Galerkin discretization therefore automatically provides the necessary data for the Taylor quadrature. The integrals containing the dissipative part of the HLLC ux are evaluated analogously. It and we assume that discussed.

details. The evaluation of the volume integrals is analogous to the ux integrals and not further

and

are constant during the integration process, see (Ref. 10) for more

is, however, not necessary to compute exact derivatives of

to preserve accuracy and stability

given by (4), and are equal to:

i

&

pressions, see (Ref. 10). The derivatives of the ow state

in (7) can be computed directly from . The discontinuous

with

a vector measure which is dened as:

, and

"

&

&

"

"

"

"

where

represents the numerical dissipation of the HLLC scheme. The integrals are

(7) or .

- 14 NLR-TP-2002-300

6 Slip ow boundary
At a slip ow boundary we impose the boundary condition:

at


This implies that the pressure is the only non-zero contribution in the ux at the domain boundary velocity is equal to:
  i  h c e i h & i h

. The zero normal velocity at

can be imposed using ghostcells at the wall, where the

and is used in the Gauss quadrature of the ux at the domain boundary faces. This has as main benet that the same HLLC ux function can be used both for faces in the interior and at the domain boundary. The straightforward implementation of the Taylor ux quadrature requires, not only depend on the pressure, as is required by the exact boundary condition. In order to remove this inconsistency we use at the boundary faces the pressure which is obtained from the exact solution of the Riemann problem for a reective wall. This solution is relatively simple and consists either of two rarefaction or two shock waves, see Toro (Ref. 7), and can be summarized as: If
i (    

however, derivatives

, since

, and this results in a ux at the wall which does

(rarefaction wave): with

 d & i  b e d  b &

with:
  b b e d  &    b c &

now straightforward and signicantly improves the accuracy in the near wall region when it is combined with a sufciently accurate representation of the boundary geometry.

evaluation of the boundary ux integral

is satised, with

problem. The pressure

( & i

Here the sufx

refers to the internal ow state at the wall at the initial time of the Riemann in the Riemann problem ensures that slip ow boundary condition the velocity at the wall in the exact Riemann problem. The with the Taylor quadrature rule is

&

If

(shock wave):

 (

& c

& i (

- 15 NLR-TP-2002-300

not give any noticeable difference in comparison with the use of the pressure in the star region obtained from the HLLC ux using ghostcells.

The integration of the boundary ux

with the Gauss quadrature rule does

- 16 NLR-TP-2002-300

7 Effect of wall curvature on the accuracy of the discontinuous Galerkin discretization

The accuracy in the near wall region of the discretization of the slip ow boundary condition does not only depend on the evaluation of the boundary integrals, but also on the representation of the surface geometry. This was noticed by Bassi and Rebay (Ref. 1), where they showed that the use of superparametric elements improved the accuracy in the near wall region. In this section we will analyze the effect of the boundary representation on the accuracy of the domain boundary integrals when they are evaluated with the Taylor quadrature rule. If we expand the ux function with
 k

Galerkin discretization it is necessary to take the boundary curvature also into account. If we use is zero, but depending on the curvature


this contribution can be large and non-negligible. For , it is possible to take the boundary
k ~

superparametric elements, generated with the mapping

curvature into account and it can be expected that this results in a more accurate discretization at the boundary. This conrms the result from Bassi and Rebay (Ref. 1), which showed that using superparametric elements signicantly improved the accuracy near the boundary. The use of su-

linear elements, generated with the mapping

, then the contribution

with

. This expression shows that for a consistent boundary treatment in a discontinuous

"

"

"

now be expressed as:

 p  P     P  P  $ &  $  p

the principal normal vector

. The boundary ux integrals can

&

"

"

"

"

"

"

is dened as
V @ b  b e T Q

the boundary curvature

&

&

&

with

the tangential vector at the boundary curve

&

&

&

&

"

"

with

. Introduce the arclength

, and use the relations:

, which , and

"

"

"

"

&

"

"

&

"

or
 f 

depending on the index

of

, then we obtain:

and the mapping



in a Taylor series terms of around the point ,

"

"

&

&

&

"

 $

- 17 NLR-TP-2002-300

10

-2

10-3

maximum residual

10-4 10 10 10
-5

-6

-7

10-8 10-9 10
-10

100

200

300

400

500

600

700

time steps

perparametric elements requires, however, signicant modications to a general three-dimensional DGFEM code. An alternative is provided by using linear elements, generated with the mapping , and use local mesh renement based on the local boundary curvature. This can effectively
              e f k

reduce the contribution

vature becomes negligible. In the next section we will investigate both the use of superparametric elements and local mesh renement.

and linear isoparametric elements (

"

for the ow about a circular cylinder at

4

using a Gauss ux quadrature rule mesh).

to such a small value that the element cur-

Fig. 1 Maximum residual for all equations of the expansion coefcients

% ( &

& 

- 18 NLR-TP-2002-300

-1

-2

-3 -3 -2 -1 0 1 2 3 4

pressure 0.7750 0.7652 0.7555 0.7457 0.7359 0.7262 0.7164 0.7067 0.6969 0.6871 0.6774 0.6676 0.6579 0.6481 0.6383 0.6286 0.6188 0.6091 0.5993 0.5895 0.5798 0.5700 0.5603 0.5505 0.5407

-1

-2

-3 -3 -2 -1 0 1 2 3 4

mach 0.7273 0.6987 0.6701 0.6415 0.6129 0.5843 0.5557 0.5271 0.4985 0.4699 0.4413 0.4127 0.3841 0.3555 0.3269 0.2983 0.2696 0.2410 0.2124 0.1838 0.1552 0.1266 0.0980 0.0694 0.0408

using a Gauss ux quadrature rule and linear isoparametric elements (

mesh).

"

Fig. 2

Pressure eld and Mach number contours for the ow about a circular cylinder at
% " % 

- 19 NLR-TP-2002-300

8 Results
In order to investigate the effect of the different numerical discretizations for a slip ow boundary . This is the same problem as considered by Bassi and Rebay (Ref. 1) and is a good test case the solution is equal to the classical potential ow around a circular cylinder. The Mach number 0.38 is chosen in this study in order to avoid the well known difculties of computing low Mach number ows which are not relevant for the present study. At the inow we use the conditions the outow boundary we use the conditions in each of the equations for the coefcients
 V 5 4 5 4

we consider the ow about a circular cylinder with a radius of one at the Mach number since it is very sensitive to the correct treatment of the slip ow boundary condition. If
7 (

refers to the free stream value and

9

to the internal ow state. For all calculations the residual

f f c   

,
j

, was reduced to

multigrid convergence acceleration technique. Figure 1 shows the result for linear isoparametric elements and Gauss ux quadrature. It should be noted that the results obtained by Bassi and Rebay (Ref. 1) for this case did not converge, even after more than 100.000 Runge-Kutta steps. In order to obtain converged solutions they had to use superparametric elements. As a measure for the accuracy of the different boundary discretizations we use the total pressure loss, which is dened as:
4   b b e d  f b b e b & B @

For subsonic inviscid ow the total pressure loss should be zero and is therefore a good indicator for the numerical accuracy. It is also an important quantity in aerodynamic computational uid dynamics calculations, where it must be small otherwise the solution shows too much pollution through numerical dissipation and spurious vorticity generation. For the baseline solution we use
7 4

linear isoparametric elements, generated with the mapping

I

. The boundary ux integrals are evaluated with a two-point Gauss quadrature rule with unit
R f Q &  f

weights and quadrature points

, and a two-point product Gauss quadrature rule with

the same weights and quadrature points for the volume integration. Figure 2 shows the pressure eld and Mach number contours. For both quantities the contour lines are not very symmetric an articial wake. The effect of the boundary condition becomes more pronounced when we look at the total pressure loss shown in Figures 3 and 4. The maximum value of the total pressure The use of quadratic superparametric elements in combination with Gauss quadrature reduces the
c 

with respect to the axis

. This is particularly true for the Mach number which clearly shows

loss for linear isoparametric elements in combination with Gauss quadrature integration is

,
(

and

. The sufx using a

&

&

&

&

&

,
5

and

, with

the specic total enthalpy,

. At

& & & & i 5 h & i k ~ h

- 20 NLR-TP-2002-300

0.06

0.05

Gauss isopar-coarse Gauss superpar-coarse Gauss isopar-fine Taylor isopar-coarse Taylor superpar-coarse Taylor isopar-fine

0.04

pt-loss

0.03

0.02

0.01

-1

-0.5

0.5

Fig. 3

Comparison of the total pressure loss at the wall for the ow around a circular cylinder
# "

maximum value of the total pressure loss to 0.049, see Figure 3, which also shows the result on a signicant part of the boundary, which conrms the conclusion of Bassi and Rebay (Ref. 1) for P1Q1 elements, but not sufcient near the rear stagnation line to remove the articial wake. The in combination with linear isoparametric elements improves on a ne mesh. This contradicts the conclusion of Bassi and Rebay that a higher-order geometric approximation of curved boundaries is mandatory in DGFEM. The difference can probably be attributed to the convergence problems Bassi and Rebay experienced for this case, which resulted on a ne mesh in an unsteady ow. For linear isoparametric elements in combination with the Taylor quadrature rule the total pressure mesh. A comparison of Taylor and Gauss quadrature techniques in Figure 3 shows that
% 5 6 " " ? 5 ?

the ne

mesh. The use of superparametric elements does reduce the total pressure loss for

results on the

mesh in Figures 3 and 4 also show that the solution using Gauss quadrature

loss reduces to a maximum of

5 6 " 5

, see Figures 3 and 4, which also show the result on the ne mesh has a mesh when

the Taylor quadrature rule with linear isoparametric elements on the coarse comparable total pressure loss as linear isoparametric elements on the ne

elements on a coarse mesh (

elements).

"

on a coarse (

elements) and ne mesh (

"

"

 "

 5 5 6 " 5

) using Gauss and Taylor ux quadrature rules for isoparametric elements elements) and superparametric

- 21 NLR-TP-2002-300

-1

-2 -1 0 1 2 3 4

pt-loss 0.0620 0.0594 0.0567 0.0541 0.0514 0.0488 0.0461 0.0435 0.0408 0.0382 0.0355 0.0329 0.0302 0.0276 0.0250 0.0223 0.0197 0.0170 0.0144 0.0117 0.0091 0.0064 0.0038 0.0011 -0.0015

-1

-2

pt-loss 0.0620 0.0594 0.0567 0.0541 0.0514 0.0488 0.0461 0.0435 0.0408 0.0382 0.0355 0.0329 0.0302 0.0276 0.0250 0.0223 0.0197 0.0170 0.0144 0.0117 0.0091 0.0064 0.0038 0.0011 -0.0015

-1

-1

-2 -1 0 1 2 3 4

pt-loss 0.0620 0.0594 0.0567 0.0541 0.0514 0.0488 0.0461 0.0435 0.0408 0.0382 0.0355 0.0329 0.0302 0.0276 0.0250 0.0223 0.0197 0.0170 0.0144 0.0117 0.0091 0.0064 0.0038 0.0011 -0.0015

-1

-2

pt-loss 0.0620 0.0594 0.0567 0.0541 0.0514 0.0488 0.0461 0.0435 0.0408 0.0382 0.0355 0.0329 0.0302 0.0276 0.0250 0.0223 0.0197 0.0170 0.0144 0.0117 0.0091 0.0064 0.0038 0.0011 -0.0015

-1

-1

-2 -1 0 1 2 3 4

pt-loss 0.0620 0.0594 0.0567 0.0541 0.0514 0.0488 0.0461 0.0435 0.0408 0.0382 0.0355 0.0329 0.0302 0.0276 0.0250 0.0223 0.0197 0.0170 0.0144 0.0117 0.0091 0.0064 0.0038 0.0011 -0.0015

-1

-2

pt-loss 0.0620 0.0594 0.0567 0.0541 0.0514 0.0488 0.0461 0.0435 0.0408 0.0382 0.0355 0.0329 0.0302 0.0276 0.0250 0.0223 0.0197 0.0170 0.0144 0.0117 0.0091 0.0064 0.0038 0.0011 -0.0015

-1

quadrature rule (left) and Taylor quadrature rule (right). Top: linear isoparametric elements
? ?

mesh). Middle: quadratic superparametric elements (

5 6 " 5

linear isoparametric elements (

mesh).

"

"

Fig. 4 Total pressure loss for ow around a circular cylinder at

 % "

, using a Gauss

mesh). Bottom:

- 22 NLR-TP-2002-300

-1

-2

-3 -3 -2 -1 0

mach 0.7611 0.7330 0.7049 0.6768 0.6487 0.6206 0.5925 0.5644 0.5363 0.5082 0.4800 0.4519 0.4238 0.3957 0.3676 0.3395 0.3114 0.2833 0.2552 0.2271 0.1990 0.1709 0.1428 0.1147 0.0866

quadrature and quadratic superparametric elements, wall pressure from Riemann problem
?

elements).

Gauss quadrature is used. This difference can be attributed to the fact that in the Gauss quadrature rule the normal ux is computed at different locations in the element face. At each quadrature point we consider a one-dimensional Riemann problem and neglect the tangential variation of the solution in the element face. The tangential vectors at the quadrature points are slightly different and this results in different shear wave contributions from the quadrature points, which manifest themselves in spurious entropy generation near the wall. The Taylor quadrature rule considers the Riemann problem only at one point and therefore results in a more consistent discretization when combined with one-dimensional (approximate) Riemann solvers. The use of quadratic superparametric elements in combination with the Taylor quadrature rule further reduces the maximum total pressure loss to boundary curvature effects in combination with the Taylor quadrature rule also results in a significantly improved solution, see Figure 5 which shows the Mach number contours. The contours Figure 2.
 % >  

, see Figures 3 and 4. Incorporating the

are now nearly symmetric with respect to

and do not show the pronounced wake visible in

"

Fig. 5

Mach number contours for ow about a circular cylinder at

   % "

using Taylor

- 23 NLR-TP-2002-300

0.02

0.015

coarse grid one time adapted coarse grid two times adapted coarse grid three times adapted coarse grid fine grid coarse grid superparametric

0.01

0.005

0 -1 -0.5 0 0.5 1

Taylor quadrature with locally rened linear isoparametric elements (meshes 1536, 1710, 2439, 8358 elements, pressure from Riemann problem).

Local mesh adaptation in the vicinity of the wall can also be a very efcient technique to reduce the numerical error at slip ow boundaries. Especially, when it is related to the boundary curvature mesh with linear isoparametric elements. The ux quadrature is done with the Taylor quadrature rule. The adapted meshes contain 1536, 1710, 2439, 8358 elements. Renement is based on the generated vorticity, which should be zero in this subsonic inviscid ow. Also shown is the total pressure loss on the adapted mesh, on the coarse mesh with isoparametric and superparametric the total pressure loss is nearly reduced to zero, except close to the rear stagnation point. The ne grid result is nearly identical to the one-time adapted mesh but requires 3.6 times more elements than the one-time adapted mesh. The Mach number contours on the ne mesh are shown in Figure
 ?

as discussed in Section 7. Figure 6 shows three meshes obtained with local renement of a

elements, and on the ne mesh with

elements. Figure 6 shows that after three adaptations

7 and are nearly symmetric with respect to the axis

"

"

Fig. 6 Comparison of the total pressure at the wall of a circular cylinder at

pt-loss
using

   5 6 " 5

- 24 NLR-TP-2002-300

-1

-2

-3 -3 -2 -1 0 1 2 3 4

mach 0.8733 0.8386 0.8040 0.7694 0.7347 0.7001 0.6655 0.6308 0.5962 0.5615 0.5269 0.4923 0.4576 0.4230 0.3883 0.3537 0.3191 0.2844 0.2498 0.2152 0.1805 0.1459 0.1112 0.0766 0.0420

a locally rened mesh with 8358 linear isoparametric elements, pressure from Riemann problem.

"

Fig. 7

Mach contours for ow about a circular cylinder at

using Taylor quadrature on

- 25 NLR-TP-2002-300

9 Conclusions
Discontinuous Galerkin nite element discretizations of the Euler equations of gas dynamics can suffer from signicant numerical errors due to slip ow boundary conditions at curved walls, depending on the type of numerical quadrature and the representation of the domain boundary. The newly presented Taylor ux quadrature rule results in a signicant reduction of the total pressure losses in comparison with a two-point (product) Gauss quadrature rule. A simple analysis has shown that for a consistent boundary discretization in a discontinuous Galerkin nite element method it is necessary to take the effect of the domain boundary curvature into account. This is demonstrated with quadratic superparametric elements, which reduces the total pressure losses both for Gauss and Taylor quadrature rules. Superparametric elements are the most effective for Taylor quadrature rules where they result in more than a factor three reduction in total pressure loss in comparison with isoparametric elements combined with Gauss quadrature. The effects of boundary curvature can also be removed using mesh adaptation and it is demonstrated that it is not necessary to use a higher-order boundary representation as was stated by Bassi and Rebay (Ref. 1). Local mesh renement of linear isoparametric elements is very effective in reducing the error at slip ow boundaries and can provide an alternative to the use of superparametric elements. Acknowledgements The research of the rst author is supported in part by a research grant from the Netherlands National Aerospace Laboratory NLR, which is gratefully acknowledged. Sincere thanks are also due to C.M. Klaij (UT) for providing his DGFEM code for part of the calculations.

- 26 NLR-TP-2002-300

References
1. F. Bassi and S. Rebay, High-order accurate discontinuous nite element solution of the 2D Euler equations, J. Comput. Phys., 138, 251-285 (1997). 2. P. Batten, N. Clarke, C. Lambert and D. Causon, On the choice of wave speeds for the HLLC Riemann solver, SIAM J. Sci. and Stat. Comp., 18, 1553-1570 (1997). Navier-Stokes equations, Int. J. Numer. Meth. Fluids, 47, 61-73 (2000). 4. B. Cockburn, Devising discontinuous Galerkin methods for non-linear hyperbolic conservation laws, J. Comput. Appl. Math., 128, 187-204 (2001). 5. B. Cockburn and C.W. Shu, Runge-Kutta Discontinuous Galerkin Methods for convectiondominated problems, J. Sci. Comput., 16, 173-261 (2001). 6. E.F. Toro, M. Spruce and W. Speares, Restoration of the contact surface in the HLL-Riemann solver, Shock Waves, 4, 25-34 (1994). 7. E.F. Toro, Riemann solvers and numerical methods for uid dynamics. A practical introduction 2nd edition (Springer Verlag, 1999). 8. J.J.W. van der Vegt and H. van der Ven, Discontinuous Galerkin nite element method with anisotropic local grid renement for inviscid compressible ows, J. Comput. Phys. 141, 46-77 (1998). 9. J.J.W. van der Vegt and H. van der Ven, Space-time discontinuous Galerkin nite element method with dynamic grid motion for inviscid compressible ows. I. General formulation, to appear J. of Comput. Phys. (2002). 10. H. van der Ven and J.J.W. van der Vegt, Space-time discontinuous Galerkin nite element method with dynamic grid motion for inviscid compressible ows. II. Efcient ux quadrature, to appear Comput. Meth. Appl. Mech. Engrg. (2002).
t

3. C.E. Baumann and T.J. Oden, A discontinuous

nite element method for the Euler and the