INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN FLUIDS

Int. J. Numer. Meth. Fluids 2007; 55:917–943

Published online 4 October 2007 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/fld.1469

High-order ENO and WENO schemes for unstructured grids

W. R. Wolf1 and J. L. F. Azevedo2, ∗, †

1 Instituto Tecnológico de Aeronáutica, CTA/ITA, São José dos Campos, São Paulo 12228-900, Brazil
2 Instituto de Aeronáutica e Espaço, CTA/IAE, São José dos Campos, São Paulo 12228-903, Brazil

SUMMARY
This work describes the implementation and analysis of high-order accurate schemes applied to high-
speed flows on unstructured grids. The class of essentially non-oscillatory schemes (ENO), that includes
weighted ENO schemes (WENO), is discussed in the paper with regard to the implementation of third-
and fourth-order accurate methods. The entire reconstruction process of ENO and WENO schemes is
described with emphasis on the stencil selection algorithms. The stencils can be composed by control
volumes with any number of edges, e.g. triangles, quadrilaterals and hybrid meshes. In the paper, ENO
and WENO schemes are implemented for the solution of the dimensionless, 2-D Euler equations in a
cell centred finite volume context. High-order flux integration is achieved using Gaussian quadratures.
An approximate Riemann solver is used to evaluate the fluxes on the interfaces of the control volumes
and a TVD Runge–Kutta scheme provides the time integration of the equations. Such a coupling of
all these numerical tools, together with the high-order interpolation of primitive variables provided by
ENO and WENO schemes, leads to the desired order of accuracy expected in the solutions. An adaptive
mesh refinement technique provides better resolution in regions with strong flowfield gradients. Results
for high-speed flow simulations are presented with the objective of assessing the implemented capability.
Copyright q 2007 John Wiley & Sons, Ltd.

Received 25 July 2006; Revised 27 January 2007; Accepted 30 January 2007

KEY WORDS: ENO schemes; WENO schemes; unstructured grids; aerodynamics; compressible flow

1. INTRODUCTION

In the present paper, the authors continue the effort of implementing high-order accurate schemes
for the solution of hyperbolic conservation laws in an unstructured grid context [1]. The motivation

∗ Correspondence to: J. L. F. Azevedo, Instituto de Aeronáutica e Espaço, CTA/IAE/ASA-L, São José dos Campos,
São Paulo 12228-903, Brazil.
†
E-mail: [email protected], [email protected]

Contract/grant sponsor: Fundação de Amparo à Pesquisa do Estado de São Paulo, FAPESP; contract/grant number:
03/10047-2
Contract/grant sponsor: Conselho Nacional de Desenvolvimento Cientı́fico e Tecnológico, CNPq; contract/grant
number: 501200/2003-7

Copyright q 2007 John Wiley & Sons, Ltd.

918 W. R. WOLF AND J. L. F. AZEVEDO

for the work is the need for accurate simulations of high Mach number aerodynamic flows with
strong discontinuities. In recent years, several efforts have been made by the Computational Fluid
Dynamics (CFD) group of Instituto de Aeronáutica e Espaço (IAE) [2–4] for the development
of computational tools which are able of accurately capturing discontinuities such as the shock
waves appearing in the aerodynamic flows of interest. Some upwind schemes such as the van Leer
flux vector splitting scheme [5], the Liou AUSM+ flux vector splitting scheme [6] and the Roe
flux difference splitting scheme [7] have been implemented and tested for second-order accuracy
considering a TVD–MUSCL reconstruction [8]. However, Reference [4] shows that nominally
second-order schemes presented results with an order of accuracy smaller than the expected one
in the solutions for unstructured grids. Aside from this fact, it is well known that TVD schemes
have their order of accuracy reduced to first order in the presence of shocks due to the effect of
limiters [9].
Actually, this last observation initiated the development and implementation of essentially non-
oscillatory (ENO) schemes, introduced by Harten et al. [10], in which oscillations up to the order
of the truncation error are allowed to overcome the drawbacks and limitations of TVD schemes.
Subsequent on the development of ENO schemes, the weighted ENO (WENO) schemes were
introduced by Liu et al. [11] with the purpose of presenting better convergence rate for stationary
cases, better smoothing for the flux vectors and better accuracy using the same stencils as the ENO
schemes.
In this work, the ENO schemes and the WENO schemes are implemented in a cell centred finite
volume context for unstructured meshes. The two-dimensional Euler equations are considered to
represent the flows of interest. The ENO and WENO schemes have been developed with the
purpose of accurately capturing discontinuities appearing in problems governed by hyperbolic
conservation laws. In the high-speed flows of interest in the present paper, these discontinuities are
mainly represented by shock waves and contact discontinuities. The entire reconstruction process
of ENO and WENO schemes is described in detail for quadratic and cubic polynomials and, hence,
for third- and fourth-order accurate numerical methods. The extension to arbitrary orders of spatial
accuracy is straightforward from the formulation presented in the paper. For the ENO schemes,
interpolation polynomials of one order less than the order of accuracy expected in the solution are
computed and these polynomials are used to reconstruct a good approximation to the values of the
flow variables within the cells. These polynomials interpolate primitive variable values in Gauss
quadrature points using stencils determined by a von Neumann neighbourhood [12]. The control
volume moments and the mean values of primitive variables in the cells are used to compute the
polynomial coefficients and, hence, one can compute the oscillation indicators for the polynomials
and select the smoothest among them.
While the ENO schemes use the smoothest polynomial, the WENO schemes use all ENO com-
puted polynomials for the stencils and, therefore, they construct only one polynomial. Non-negative
weights, which must add up to one, are computed for every ENO polynomial through oscillation
indicators and the WENO polynomial is constructed by the sum of all the ENO polynomials
multiplied by the respective weights [11, 13]. The weights attributed to the polynomials can be of
the order of accuracy desired in the solution if these polynomials are in the discontinuous regions
of the mesh or they can be of order one if the polynomials are in the smooth regions of the flow.
This paper is organized as follows. In Section 2, the authors recall the necessary theoretical
formulation concerning the governing equations, the flux computation across the interfaces of the
cells, and the time evolution of the numerical scheme. Section 3 is concerned with the imple-
mentation of ENO and WENO schemes with considerations to the basic formulation, selection

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
 HIGH-ORDER ENO AND WENO SCHEMES FOR UNSTRUCTURED GRIDS 919

of the oscillation indicator, reconstruction of third-order accurate schemes, and reconstruction of

fourth-order accurate schemes. Particular attention is paid to the stencil selection algorithms. In
Section 4, one can see a brief description of the adaptive mesh refinement technique used in the
computations. In Section 5, some numerical results are presented and, finally, in Section 6, several
general discussions appear in the conclusion of the paper. The results presented in Section 5 include
some well documented numerical test cases such as: the Ringleb flow; the flow in a channel with
a forward-facing step; the flow in a typical hypersonic inlet; and the flow over a 2-D blunt body.

2. THEORETICAL FORMULATION

2.1. Governing equations

In the present work, the 2-D Euler equations are solved in their integral form as



*
Q dV + (∇ · P) dV = 0 (1)
*t V V

where P = E ı̂ + F jˆ. The application of the divergence theorem to Equation (1) yields



*
Q dV + (P · n) dS = 0 (2)
*t V S

where V represents the control volume, S represents the surface of the control volume and n is
the unit normal vector to the surface, S, positive outward. The vector of conserved variables, Q,
and the convective flux vectors, E and F, are given by
⎧ ⎫ ⎧ ⎫ ⎧ ⎫
⎪
⎪ ⎪
⎪
u ⎪ ⎪ ⎪
⎪
v ⎪ ⎪
⎪ ⎪
⎪ ⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎨
u ⎬ ⎨
u + p ⎬
2 ⎨
uv ⎪ ⎬
Q= , E= , F= (3)
⎪
⎪
v ⎪
⎪ ⎪
⎪
uv ⎪ ⎪ ⎪
⎪
v 2 + p ⎪
⎪
⎪
⎪ ⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎩ ⎭ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎩ ⎭ ⎩ ⎭
e (e + p)u (e + p)v

In the above equations,
represents the density, p is the pressure, u and v represent the Cartesian
components of the velocity, and e is the total energy per unit of volume. The system is closed by
the equation of state for a perfect gas

p = ( − 1)[e − 12
(u 2 + v 2 )] (4)

where the ratio of specific heats, , was chosen as 1.4 for all computations in this work.
Equation (2), discretized in a cell centred finite volume context, can be rewritten for the ith
control volume as


*Q i 1
=− (P · n) dS (5)
*t Vi Si

where Q i is the mean value of Q at time t over the ith control volume, Vi .

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
920 W. R. WOLF AND J. L. F. AZEVEDO

2.2. Numerical flux evaluation

The spatial discretization is concerned with finding a discrete approximation to the integral in the
right-hand side of Equation (5). In the present section, this discretization is discussed with respect
to the formulation of the numerical flux computation. The subject of obtaining interface values for
the flow properties is deferred to a latter section. The control volumes considered in this work are
triangles and quadrilaterals and their boundaries can be decomposed into a finite number of line
segments,
j . One should observe that the control volumes could be composed by any type of
polygon, because the really important aspect is that their bounding contours could be decomposed
into a finite number of line segments. Hence, it is possible to write

Si =
j (6)

Thus, the boundary integral from Equation (5) can be decomposed into

(P · n) dS = (P · n) dS (7)
Si j
j

and, as n is constant on each line segment,
j , the right side of Equation (7) is discretized using
N -point Gaussian integration formulae with degree 2N − 1 and order of accuracy 2N . Using such
integration formulae, one obtains the following approximation:




N
(P · n) dS ≈ |
j | w P(Q(G  ), t) · n (8)
Si j =1

where G  and w are, respectively, the Gaussian points and the weights on the
j line segment.
For the third- and fourth-order schemes, two Gaussian points are necessary along each line
segment. Given the coordinates of the end points, z 1 and z 2 , one can obtain the Gaussian points as
√  √  √  √ 
3+1 3+1 3+1 3+1
G 1 = √ z1 + 1 − √ z 2 and G 2 = √ z 2 + 1 − √ z1 (9)
2 3 2 3 2 3 2 3

and the respective weights, w1 and w2 , are chosen as w1 = w2 = 12 .

Using the method described above, one can compute values of Q i in some instant, t, and then,
from these mean values, one can reconstruct polynomials that represent the primitive variables
,
u, v and p. Finally, it is possible to compute values of the conserved variables in the Gaussian
points. Due to the discontinuity of the reconstructed values of the conserved variables across the
cell boundaries, one must use a numerical flux function to approximate these flux values along such
boundaries. In the present work, the authors have used the Roe flux difference splitting method [7]
to compute such approximations. Hence, the flux evaluation in the right-hand side of Equation (8)
is computed as

P(Q j (G  ), t) · n ≈ P(Q i (G  ), Q ng (G  ), t; n)

= 12 [P(Q i (G  )) + P(Q ng (G  ))] · n

− 12 |A(Q i (G  ), Q ng (G  ), t; n)|[Q ng (G  ) − Q i (G  )] (10)

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
 HIGH-ORDER ENO AND WENO SCHEMES FOR UNSTRUCTURED GRIDS 921

In Equation (10), the |A| matrix is a positive semidefinite matrix formed from the flux Jacobian
matrix in the sense defined by Roe [7]. In the same equation, Q i (G  ) is the vector of conserved
variables computed at the quadrature point, G  , using the polynomial reconstruction associated
with the ith control volume. Similarly, Q ng (G  ) is the vector of conserved variables computed at
the same quadrature point, G  , but using the polynomial reconstruction associated with the ngth
control volume. Here, it is being assumed that ‘ng’ represents the neighbour of the ith control
volume, which shares the
j line segment with the ith cell.

2.3. Temporal discretization

The temporal discretization is concerned with solving a system of ordinary differential equations
in time. Hence, an ODE problem can be written as
dQ i 1
= − C(Q i ) (11)
dt Vi
In the present work, a third-order fully explicit TVD Runge–Kutta scheme is used to advance the
solution of the governing equations in time for the spatially third- and fourth-order schemes. It
can be written as
(0)
Q i = Q in

(1) (0) ti (0)

Qi = Qi − 1 C(Q i )
Vi
(2) (0) (1) ti (1)
Q i = 2 Q i + 3 Q i − 3 C(Q i ) (12)
Vi
(3) (0) (2) ti (2)
Q i = 4 Q i + 5 Q i − 5 C(Q i )
Vi
(3)
Q in+1 = Q i

where the  coefficients are 1 = 1, 2 = 34 , 3 = 14 , 4 = 13 , 5 = 23 , as suggested in

Reference [14].

3. ENO AND WENO RECONSTRUCTIONS

3.1. General considerations

The reconstruction procedure for the ENO schemes is based on the approximation of mean values
of the primitive variables for each cell in the mesh by polynomials of one order less than the
expected spatial order of accuracy. For the construction of polynomials of th order, one must use
N () cells, where N () = (+1)(+2)/2. The first step in obtaining the polynomial reconstruction
for each cell is to define the possible set of cells, called a stencil, that will be used. In a finite
volume cell centred scheme, the stencils can be selected in a von Neumann neighbourhood for
a linear polynomial reconstruction (second-order accuracy). This approach can be extended to
higher orders through the use of von Neumann neighbourhoods of the primary neighbours already

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
922 W. R. WOLF AND J. L. F. AZEVEDO

selected for the linear reconstruction. In the present work, the cells are triangles or quadrilaterals
and the p(x, y) polynomials can be calculated as

p(x, y) = r1 2 (x − xc )1 (y − yc )2 (13)

||


Here, || = 1 +2 , with i ∈ {0, 1, 2, . . .}, xc and yc are the Cartesian coordinates of the barycentre
of the control volume for which the reconstruction is being performed, and r1 2 are unknown
coefficients which are some approximations to the derivatives of the primitive variables.
By imposing that the integration of p(x, y) along each cell of the stencil returns the mean value
for the corresponding primitive variable in that cell, one can write a linear system, [R]{r } = {u},
of N () equations for the N () unknowns, r1 2 . Here, [R] is the matrix of control volume
moments, as in Gooch [15], computed using the scaling technique proposed by Friedrich [13] to
circumvent a poorly conditioned matrix. Moreover, {r } is the vector of unknown coefficients that
must be found and {u} is the vector composed by the mean values for each primitive variable.
The stencil is considered admissible if the [R] matrix is invertible. There are other approaches for
the calculation of the polynomials, as the one proposed by Abgrall [16], which uses interpolation
through barycentric coordinates, or the one proposed by Hu and Shu [17], which is based in
reconstructing high-order polynomials using only linear polynomials. These alternate approaches,
however, are beyond the scope of the present paper.
The control volume moments that form the [R] matrix are defined as


1
x  y ≡ (x − xc ) (y − yc ) dV (14)
Vi Vi

and they are evaluated using Gaussian quadrature formulae following the same procedure as the
one used in the flux computations. To avoid computing moments of each control volume for
all the stencils, one can replace the terms (x − x c ) and (y − yc ) by (x − xc j ) + (xc j − xc ) and
(y − yc j )+(yc j − yc ), respectively [13, 15]. The new terms appearing in the equations are simply the
split between the distances of the centroids of the V j and Vi control volumes plus the initial terms
evaluated using the centroids of the V j volume instead of the centroids of the Vi volume. Finally,
the control volume moments are calculated, saving computational cost and memory storage, by
the following general form:


1 
x  y ij ≡ [(x − xc j ) + (xc j − xc )] (y − yc j ) + (yc j − yc ) dV
Vi Vi
  



 
= (x − xc j ) (y − yc j ) x − y − (15)
=0 =0 

After the polynomial reconstructions are performed for each cell, the next step is to verify
which polynomial is the least oscillatory to use in the ENO scheme. The level of oscillation is
computed using some indicator that assesses the smoothness of p(x, y). Taking into consideration
the results presented in the literature [1, 13], the oscillation indicator used in the present work
is the one proposed by Jiang and Shu [18]. The formulation for this oscillation indicator can be

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
 HIGH-ORDER ENO AND WENO SCHEMES FOR UNSTRUCTURED GRIDS 923

expressed as
⎡  2 ⎤1/2

| |

* p(x, y)
OIJS ( p(x, y)) = ⎣ h ||−1 dx dy ⎦ (16)
1
||
 Vi *x 1 *y 2

where h is the mesh size, considered in the present computations as the diameter of the control
volume inscribed circle.
Differently from the ENO schemes, the WENO schemes use all the calculated polynomials.
These polynomials are added together through the use of weights which are computed for each
one of the polynomials as inversely proportional to its respective oscillation indicator. The main
idea in the WENO reconstruction is to attribute the computed
weights for each polynomial with
the aim of reconstructing a new polynomial as p(x, y) = m k=1
k pk (x, y). The weights are of
order one in the smooth regions of the flow and are of the order of the desired accuracy in the
solution in the regions with discontinuities. The weights can be computed as

[ + OI( pk (x, y))]−

k =
m −
(17)
k  =1 [ + OI( pk  (x, y))]

where  is a small real number used to avoid division by zero and  is a positive integer. The
WENO schemes have the property of being very smooth and stable in smooth regions of the flow,
but this property is lost if  is chosen too large. In that case, the scheme tends to behave like the
ENO schemes. In the present work, the  term is chosen as  = 2, because this yielded the best
convergence rates in previous computations by the authors [1].

3.2. Third-order implementation

For the implementation of the third-order accurate schemes one must use six cells for the polynomial
reconstructions. Five unknown coefficients are computed for each polynomial through the solution
of a 5-equation linear system. The linear system matrix is composed by the control volume
moments, and the primitive variables of the cells considered in the stencil make up the forcing
term. The algorithm implemented is based on the ideas proposed by Abgrall [16]. The first step
consists in finding the stencil that presents the smoothest polynomial among the possible stencils
for a linear reconstruction using only the primary neighbours of the main control volume [1]. The
main control volume is labelled ‘MAIN’ in all pertinent figures throughout the paper. The algorithm
uses this concept for both the ENO and WENO schemes for meshes composed by triangles and
quadrilaterals. After the smoothest three-cell stencil is chosen, one can start the procedure for
the third-order accuracy reconstruction. If the control volume is a triangle, one has a maximum
of four secondary neighbours for the reconstruction and, therefore, four possibilities of stencils
for the selection. However, if the control volume is a quadrilateral, one has a maximum of six
secondary neighbours for the reconstruction and, therefore, 20 possibilities of stencils for selection.
In Figure 1 one can see a typical neighbourhood for the implementation of the algorithm with the
hatched volumes representing a possible stencil.
Although the case in which one has 20 stencils for selection can occur, in most cases, one has
only five secondary neighbours for the reconstruction in quadrilaterals and, then, 10 possibilities
of stencils for selection. Such a situation happens because, when selecting the stencil with the
primary neighbours, one has, in most cases, a stencil that is not centred as can be viewed in

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
924 W. R. WOLF AND J. L. F. AZEVEDO

NG3
NG3
NG4 NG1

MAIN
NG5 NG4 NG1 MAIN NG2 NG5
NG2

Figure 1. Typical neighbourhood for a third-order accuracy reconstruction. The hatched volumes exemplify
a possible stencil for the reconstruction.

2ndNG

2ndNG
NG1 MAIN

2ndNG NG2 2ndNG

2ndNG

Figure 2. Non-centred neighbourhood for a third-order accuracy reconstruction in a mesh composed

exclusively by quadrilaterals.

Figure 2 for the control volumes labelled NG1 and NG2. In such a case, one can see that there
are five secondary neighbours (the control volumes labelled as 2nd NG). In triangular control
volumes, one can have the presence of three secondary neighbours for the reconstruction, which is
a bad case for reconstruction on a triangular mesh. An example of this situation can be observed in
Figure 3, where the hatched control volumes NG1 and NG2 compose the smoothest stencil for a
second-order reconstruction and the hatched control volumes, NG3, NG4 and NG5, are the control
volumes used for the third-order reconstruction. Due to the lack of possibilities to switch the cells
and, therefore, have more stencils for a third-order accuracy reconstruction, this case allows for
the construction of only one stencil. If this stencil is in a discontinuous region of the flow, it will
produce oscillations, and it will lead to a non-physical solution. Hence, the procedure adopted to
avoid the problem where one has a stencil such this, in a discontinuous region of the flow, is to
reduce the order of accuracy of the scheme in these cells.

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
 HIGH-ORDER ENO AND WENO SCHEMES FOR UNSTRUCTURED GRIDS 925

NG4

NG3 NG1
NG2
NG5

MAIN

Figure 3. Bad case for a third-order reconstruction on a triangular mesh.

The algorithm here implemented has presented good results for most cases analysed in the present
work. As expected, the WENO schemes presented better results with respect to the convergence
ratio and with respect to the smoothness of the solution in comparison with the ENO schemes. As
mentioned in the papers of Abgrall [16] and Harten and Chakravarthy [19], one has to reduce the
order of accuracy in some cells to maintain the stability of the method for both ENO and WENO
schemes. The technique used for the reduction of the order of accuracy is the same presented by
Abgrall [16]. Such a technique computes an auxiliary term as


= |r1 2 ||(x − xc )1 (y − yc )2 | (18)
=2 1 +2

and verifies if  is greater than the mean value of the primitive variable in the cell times a  factor.
In the affirmative case, one must reduce the order of accuracy of the scheme. The  factor used in
the present computations is  = 0.95, as suggested in Abgrall [16]. It should be pointed out that
the authors have studied several algorithms for stencil selection in Reference [1]. The algorithm
which is being used in the present paper was the best compromise between computational cost
and quality of the solutions.

3.3. Fourth-order implementation

For the implementation of the fourth-order accurate schemes, one must use 10 cells for the
polynomial reconstructions. Nine unknown coefficients are computed for each polynomial through
the solution of a 9-equation linear system. As previously discussed, the linear system matrix is
composed by the control volume moments, and the primitive variables of the cells considered in the
stencil form the forcing term. The algorithm implemented is again based on the ideas proposed by
Abgrall [16], and it is an extension of the algorithm presented for the third-order method. Starting
from the smoothest third-order stencil among the possible stencils for a quadratic reconstruction,
one can initiate the procedure for the fourth-order accuracy reconstruction. One must amplify
the third-order stencil adding the neighbour cells that share edges with the secondary neighbours
already selected in the quadratic reconstruction. This is the cheapest algorithm to implement, in
accordance with the approaches considered in Reference [1]. However, the resulting reconstruction
procedure can still be considered very costly. The implemented algorithm uses this concept for
both the ENO and WENO schemes for meshes composed by any type of control volume.
As an example, for a triangular control volume mesh, one can have, in the worst case, six new
cells, which are the tertiary neighbours of the main control volume, for the fourth-order scheme.

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
926 W. R. WOLF AND J. L. F. AZEVEDO

NG6

NG7 NG6 NG7 NG3

NG4 NG3
NG8 NG1 NG8 NG4 NG1

NG5 MAIN NG9

NG9 NG2 NG5 NG2 MAIN

Figure 4. Typical neighbourhood for a fourth-order accuracy reconstruction. The hatched volumes exemplify
a possible stencil for the reconstruction.

Such a situation provides 15 different stencils for selection. For a mesh composed by quadrilateral
control volumes, the worst case leads to a reconstruction procedure that is very expensive in terms
of computational cost. This worst case for a quadrilateral grid contains eight new tertiary cells for
selection and, in this case, 70 different stencils could be constructed. Although the case in which
one has six new cells for selection in a triangular grid, and the case in which one has eight new cells
for selection in a quadrilateral grid can occur, in most cases, one has five tertiary neighbours for
the reconstruction in triangles and six tertiary neighbours for the reconstruction in quadrilaterals.
Then, there are five possibilities of stencils for selection in triangular grids and 15 possibilities in
quadrilateral grids. In Figure 4 one can see a typical neighbourhood for the implementation of the
algorithm with the hatched volumes representing a possible stencil.

4. ADAPTIVE MESH REFINEMENT

In this work, grid adaptation is used to improve the quality of the solution by increasing the number
of grid points in regions where shock and expansion waves are present. Hence, a property sensor
based on the solution obtained in the original mesh is computed and it is used to determine in
which regions additional computational points are necessary. The general definition of the sensor,
following the work in References [2, 3], can be expressed as
  
|∇m |
(sensor)i = max (19)
m m max − m min
i

where m can be any conserved or primitive variable and m max and m min are the maximum and
the minimum values of the m property in the flowfield. All volumes in which the sensor exceeds
some specified threshold value are refined. Additional information about the implementation of
the adaptive mesh refinement procedure can also be found in References [2, 3].

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
 HIGH-ORDER ENO AND WENO SCHEMES FOR UNSTRUCTURED GRIDS 927

5. RESULTS

The present section discusses computational results for the third- and fourth-order schemes, which
are formulated for meshes composed by triangles and quadrilaterals. The objective of the present
study is to verify the newly implemented capability and to assess its advantages/disadvantages
with regard to the discretization methods previously available in the code [1]. Hence, the test cases
here analysed are selected among those for which analytical solutions and/or well documented,
independent data are available in the literature.

5.1. Ringleb flow

The first test case presents a study of the order of accuracy for the schemes analysed here using
the Ringleb flow problem [20]. The Ringleb flow is basically a potential flow in a strongly bent
duct and can be described analytically. At any point in the domain, the flow is characterized by
two parameters, q and k. The parameter q is the non-dimensional velocity that is constant along
isoline circles and the parameter k is constant along the streamlines. The domain used in the
present computations has taken the streamline with k = 0.8 and 0.4 for the inner and outer walls
of the duct, respectively. The choice of these parameters guarantees that the flow is everywhere
subsonic. The inlet and the outlet of the duct are chosen as being the velocity isoline q = 0.3 and
the boundary conditions are set in agreement with Reference [21]. An intermediate triangular mesh
used in the present computations, the exact solution obtained in terms of density contours and the
numerical solution obtained with the third-order WENO scheme can be observed in Figure 5. This
is a steady state case and, hence, the Courant–Friedrichs–Lewy (CFL) number is set as a constant
value. The CFL value used in this test case was CFL = 0.3 for both the third- and fourth-order
schemes.
For this test case, the order of accuracy of the ENO and WENO schemes is measured through
the computation of the L2 norm of the error in the density distribution as a function of the mesh
refinement. The solutions are carried out with the third-order ENO and WENO schemes and the
fourth-order ENO and WENO schemes using the Jiang and Shu [18] oscillation indicator. All the
calculations performed for this test case considered the Roe flux difference splitting method for
the numerical flux computation. The slope of the best-fit least-square line, through a plot of the

4 4 4

0
Y

0 0
Y

-4 -4 -4

0 5 0 5 0 5
(a) X (b) X (c) X

Figure 5. Computational grid and solutions for the Ringleb flow problem: (a) intermediate mesh used in
the Ringleb flow computations; (b) density contour lines of the exact solution; and (c) density contour
lines for the numerical solution obtained with the third-order WENO method.

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
928 W. R. WOLF AND J. L. F. AZEVEDO

Table I. Order of accuracy for Ringleb flow computation

obtained with a least-square line fit.
Method Order of accuracy

Third-order ENO 1.96

Third-order WENO 2.66
Fourth-order ENO 1.91
Fourth-order WENO 2.16

3rd order ENO 3rd order WENO

3rd order slope 3rd order slope
10-2 2nd order slope 2nd order slope
10-3
|| rho ||

|| rho ||

10-3
10-4

0.2 0.4 0.2 0.4

(a) h (b) h

Figure 6. Discrete L2 norm of density error as a function of mesh size, h, in log–log

plot: (a) results obtained for third-order ENO-Roe reconstruction and (b) results obtained for
third-order WENO-Roe reconstruction.

logarithm of the L2 norm of the density error as a function of the logarithm of the mesh size gives
a measure of the true order of accuracy of the method.
The results of such study are shown in Table I for all cases considered here. The third-order
methods are shown in Figure 6 together with the slopes that indicate ideal second- and third-order
accuracy. One can observe in Table I that the third-order ENO scheme achieved only second-order
accuracy, while the third-order WENO scheme presented the best results in terms of order of
accuracy. The actual order of accuracy obtained with the third-order WENO scheme is 2.66, which
is only slightly lower than the theoretical one. In Figure 7, one can see the results for the fourth-
order schemes, together with the slopes representative of second- and third order of accuracy. The
fourth-order schemes did not achieve the expected order of accuracy. While the fourth-order ENO
scheme presented almost second order of accuracy, the fourth-order WENO presented an actual
order of accuracy equal to 2.16. One should observe that all the computations performed in this
test case considered meshes composed exclusively by triangular control volumes.
The results for the third- and fourth-order schemes are directly related to the reconstructions
close to the boundaries. In such regions, the algorithm for stencil selection does not have sufficient
neighbours to reconstruct the high-order schemes and the order of accuracy decreases to second
order. The results for the third-order ENO scheme are in agreement with the results shown in

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
 HIGH-ORDER ENO AND WENO SCHEMES FOR UNSTRUCTURED GRIDS 929

4th order ENO 10-2

4th order WENO
3rd order slope 3rd order slope
2nd order slope 2nd order slope
-2
10
|| rho ||

|| rho ||
-3
10

-3
10

10-4
0.2 0.3 0.4 0.5 0.2 0.3 0.4 0.5
(a) h (b) h

Figure 7. Discrete L2 norm of density error as a function of mesh size, h, in log–log plot:
(a) results obtained for fourth-order ENO-Roe reconstruction and (b) results obtained for
fourth-order WENO-Roe reconstruction.

Reference [13], in terms of order of accuracy. The third-order WENO scheme presented the
highest value of order of accuracy, and the fourth-order schemes presented results with lower
orders of accuracy than the respective third-order schemes. Again, one must emphasize that these
results are strongly related to the reconstructions close to the boundary regions. Due to the lack
of sufficient control volumes near the boundaries, the algorithms for stencil selection cannot find
suitable stencils to maintain fourth order in such regions.

5.2. Forward-facing step

The second test case is a compressible flow in a channel with a forward-facing step that is well
documented in Reference [22]. The channel length is 3, with height 1, in dimensionless units.
The step is located in a position with 0.6 dimensionless units from the channel entrance. It has a
height of 0.2 dimensionless units. Three different meshes are used to test the schemes. The first
mesh considered in this test case has 16 384 nodes and 16 064 quadrilateral control volumes. The
complete view and a blown-up view of the step region can be seen in Figure 8. The second and third
meshes considered in this test case are composed exclusively by triangular control volumes and
have 7022 and 13 256 nodes, and 13 608 and 25 850 control volumes, respectively. The complete
view and the blown-up view of the step region for both meshes can be seen in Figure 9 for
the coarse mesh, and in Figure 10, for the fine mesh. An uniform Mach number of 3.0 is set
as inflow at t = 0 dimensionless time units. Density is made dimensionless with respect to the
entrance condition and pressure is made dimensionless with respect to the entrance density times
the entrance speed of sound squared. One should observe that this test case considers unsteady
flow. The time step used in this test case is t = 0.0004 for all the schemes.
The more interesting structures of the flow develop at time equal to t = 4.0 dimensionless time
units. Results at this instant of time appear in Figures 11–13, for the computation with WENO
schemes. In Figures 11(a)–(c) one can see the solution for the second-, third- and fourth-order
WENO schemes, respectively. These solutions presented in Figures 11(a)–(c) are obtained for the
mesh composed exclusively by quadrilateral control volumes. In Figures 12(a) and 13(a) one can

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
930 W. R. WOLF AND J. L. F. AZEVEDO

2.5
0.6

0.4
1.5
Y

Y
1 0.2

0.5
0
0
0 0.5 1 1.5 2 2.5 3 0.2 0.4 0.6 0.8 1
(a) X (b) X

Figure 8. Mesh composed exclusively by quadrilateral control volumes used for the compu-
tation of the forward-facing step problem: (a) entire mesh for the channel and (b) detail of
the mesh in the step corner region.

2.5
0.5

2 0.4

1.5 0.3
Y

0.2
1
0.1
0.5
0

0
0 0.5 1 1.5 2 2.5 3 0.4 0.5 0.6 0.7 0.8 0.9 1
(a) X (b) X

Figure 9. Coarse mesh composed exclusively by triangular control volumes used for the
computation of the forward-facing step problem: (a) entire mesh for the channel and
(b) detail of the mesh in the step corner region.

see the solutions for the second-order WENO scheme, in Figures 12(b) and 13(b) one can see
the solutions for the third-order WENO scheme, and in Figures 12(c) and 13(c) one can see the
solutions for the fourth-order WENO scheme. The solutions presented in Figure 12 are obtained for
the coarse triangular mesh and the solutions in Figure 13 are obtained for the fine triangular mesh.
For the fine triangular mesh, one can also observe the solution for a second-order MUSCL scheme
in Figure 14. Roe’s flux difference splitting method is used in all the computations performed for
this test case, and the minmod limiter is used in the TVD–MUSCL solution.

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
 HIGH-ORDER ENO AND WENO SCHEMES FOR UNSTRUCTURED GRIDS 931

2.5 0.5

0.4
2

0.3
1.5

Y
Y

0.2
1
0.1

0.5 0

-0.1
0
0 0.5 1 1.5 2 2.5 3 0.4 0.5 0.6 0.7 0.8 0.9 1
(a) X (b) X

Figure 10. Fine mesh composed exclusively by triangular control volumes used for the
computation of the forward-facing step problem: (a) entire mesh for the channel and
(b) detail of the mesh in the step corner region.

At the instant of time presented in the solutions, a detached shock evolves to a lambda shock
that reflects in the upper channel wall. A contact discontinuity is created past the lambda shock
and both the reflected shock and the contact discontinuity move downstream along the channel.
The reflected shock is again reflected at the lower wall as well as at the upper wall near the end
of the channel. The contact discontinuity interacts with the shock that reflects in the lower wall.
This interaction occurs in the region near the end of the channel, just upstream of the last shock
reflection. At the corner region, there is also a weak oblique shock wave that ends the expansion
region due to a Prandtl–Meyer expansion fan. This weak shock interacts with the first reflected
shock near the lower wall of the channel. This typically leads to a flow structure that resembles
a non-physical shock–boundary layer interaction in the region where the second shock reflection
occurs. The solution in this region is very dependent on the treatment applied to the corner of the
step. In the present work, no special treatment is applied to the corner of the step, contrary to the
procedure adopted in the original reference [22]. One can see that the results presented here are
similar to those in References [12, 16], where, as in this work, no special treatment is applied to
the corner of the step.
The results obtained in Figure 11 for the mesh composed by quadrilateral control volumes
present a very sharp capture of the first detached shock and of the lambda shock that reflects on
the upper wall of the channel. The contact discontinuity was also well captured for the second-,
third- and fourth-order WENO schemes. The first shock reflection, as well as the second reflection,
were well defined in the solutions for all schemes. For the mesh in Figure 8, one can observe
that the second reflection of the shock, which occurs in the lower wall, was affected by the
non-physical shock–boundary layer interaction created downstream of the corner region. Such a
behaviour is very pronounced for the solutions obtained with the second- and third-order WENO
schemes and the quadrilateral mesh. The fourth-order WENO scheme presented a solution with
a less pronounced shock–boundary layer interaction as can be seen in the Y shock formation in
Figure 11(c). One can see that the WENO schemes capture the expansion fan near the corner
region. However, the schemes are not able to capture the weak oblique shock created downstream

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
932 W. R. WOLF AND J. L. F. AZEVEDO

0.5

0
0 0.5 1 1.5 2 2.5 3
(a) X

0.5

0
0 0.5 1 1.5 2 2.5 3
(b) X

0.5

0
0 0.5 1 1.5 2 2.5 3
(c) X

Figure 11. Mach number contours at t = 4.0 dimensionless time units for the mesh in Figure 8: (a) Mach
number contours for the second-order WENO scheme; (b) Mach number contours for the third-order
WENO scheme; and (c) Mach number contours for the fourth-order WENO scheme.

of the expansion region. Figure 11 also shows a separation bubble, in the post-shock region, in
the lower wall of the channel. For the present computations, this circulation zone appears only for
the calculations with the quadrilateral mesh. The region where the contact discontinuity and the
second shock reflection meet and the third reflection of the shock, near the end of the channel, are
better resolved by the third- and fourth-order WENO schemes.
The results in Figure 12, obtained for the coarse mesh composed by triangular control volumes,
present a difference regarding the height of the lambda shock position. One can see that the lambda
shock forms closer to the wall for the second-order WENO scheme and this affects the position of
the contact discontinuity. In Figure 12, one can see that all the shock reflections and the contact
discontinuity are better captured for the third- and fourth-order schemes. The second-order WENO
scheme presents the non-physical shock–boundary layer interaction created past the corner region,
in the lower wall of the channel, although such effect is much less pronounced in this case than
in the solutions with the quadrilateral meshes. In other words, the second reflection of the shock
presents a Y shock instead of a simple reflection at the wall. The results for the second-order
scheme, in the region where the interaction between the contact discontinuity and the last shock

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
 HIGH-ORDER ENO AND WENO SCHEMES FOR UNSTRUCTURED GRIDS 933

0 .5

0
0 0 .5 1 1 .5 2 2 .5 3
(a) X

0 .5

0
0 0 .5 1 1 .5 2 2 .5 3
(b) X

0 .5

0
0 0 .5 1 1 .5 2 2 .5 3
(c) X

Figure 12. Mach number contours at t = 4.0 dimensionless time units for the mesh in Figure 9: (a) Mach
number contours for the second-order WENO scheme; (b) Mach number contours for the third-order
WENO scheme; and (c) Mach number contours for the fourth-order WENO scheme.

reflection occurs, are poor in comparison with the results obtained by the third- and fourth-order
schemes. The third-order WENO scheme presents a poor capture of the weak shock near the
corner region. However, the fourth-order scheme is the only scheme which captured it well for
the triangular coarse grid. These schemes indicate a much less pronounced shock–boundary layer
interaction-type phenomenon in the shock reflection along the lower wall of the channel. All shock
reflections have a better resolution for the third- and fourth-order WENO schemes and they are in
the correct positions, when compared to Reference [22].
The results obtained with the WENO schemes for the fine triangular mesh present the more
accurate description of the flow features for this test case, as one can see in Figure 13. All the
flow features are more sharply captured due to the mesh refinement. One should note that for
this triangular fine grid, the Jiang and Shu [18] oscillation indicator did not present stability
for the simulation with the fourth-order scheme, and, then, the oscillation indicator used in this
test case, for the fourth-order scheme, was the one proposed by Harten and Chakravarthy [19].
For the second- and third-order schemes, the Jiang and Shu oscillation indicator [18] is used in
the computations. For this fine mesh, the second-order WENO scheme has captured the contact
discontinuity and the shock wave generated in the corner of the step differently from the results for

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
934 W. R. WOLF AND J. L. F. AZEVEDO

0.5

0
0 0.5 1 1.5 2 2.5 3
(a) X

0.5

0
0 0.5 1 1.5 2 2.5 3
(b) X

0.5

0
0 0.5 1 1.5 2 2.5 3
(c) X

Figure 13. Mach number contours at t = 4.0 dimensionless time units for the mesh in Figure 10: (a) Mach
number contours for the second-order WENO scheme; (b) Mach number contours for the third-order
WENO scheme; and (c) Mach number contours for the fourth-order WENO scheme.

0.5

0
0 0 .5 1 1 .5 2 2 .5 3
X

Figure 14. Mach number contours for the second-order MUSCL scheme at t = 4.0 dimensionless time
units for the mesh in Figure 10.

the coarse triangular mesh. In Figure 14, one can observe the solution for the second-order MUSCL
scheme. This scheme presents a solution much more dissipative than the solution obtained for the
second-order WENO scheme. The shock reflections are poorly captured in comparison with the
second-order WENO scheme, and the slip line has almost disappeared in the downstream region of
the channel. This scheme does not capture the weak shock in the step corner region and, then, one

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
 HIGH-ORDER ENO AND WENO SCHEMES FOR UNSTRUCTURED GRIDS 935

can see the Y shock formation in the lower wall. If one compares the solution for the second-order
WENO scheme, obtained for the triangular coarse mesh, with the solution for the second-order
MUSCL scheme, obtained with the triangular fine mesh, one can observe the strong similarity
between such solutions, at least in terms of Mach number contours.
For the third- and fourth-order scheme solutions in Figure 13, the shock wave reflections and
the contact discontinuity are sharper than for the second-order scheme solution. However, as one
can also observe in Figure 13, the third-order scheme presents more accurate results than the
second- and the fourth-order schemes considering the resolution of the weak shock near the step
corner region. One can further observe that the non-physical shock–boundary layer interaction has
almost disappeared for the second- and fourth-order schemes and it has completely disappeared
for the third-order scheme solution. This fact is related with the computation of the weak shock
near the corner of the step. For the refined triangular mesh, the fourth-order solution is similar
to the second-order one, regarding the appearance of the Y shock and the shock–boundary layer
interaction. The authors believe that the better resolution of the step corner region flow obtained
with the third-order scheme, compared with that obtained with the fourth-order scheme, is related
to the use of the different oscillation indicators. The literature [13] presents results showing that the
Jiang and Shu oscillation indicator obtains more accurate results than the Harten and Chakravarthy
one. The fact the physical flow properties are better captured for this case is related to the mesh
refinement and to the order of accuracy of the schemes. A refined mesh and/or a higher order
scheme tend(s) to diminish the appearance of the non-physical shock–boundary layer interaction.
The correct capture of the shock wave created at the corner of the step can be seen only for the
triangular fine mesh, considering the second- and third-order schemes. The fourth-order scheme
was able to capture the weak shock even in the coarse triangular grid. Hence, it is clear that the
appearance of the non-physical shock–boundary layer interaction along the lower wall is directly
related to the resolution of the corner flow, and the solution can always be improved with adequate
mesh refinement. The increase in the order of accuracy of the scheme is also a determinant key
ingredient to solve the problem concerning the appearance of the non-physical shock–boundary
layer interaction. It is shown in the paper that, in most cases here analysed, the order of accuracy
increase presents better definition for the solution. Only for the fine triangular grid, the third-order
scheme presents a more accurate solution for the problem than the fourth-order one. However,
as already discussed, these two solutions use different oscillation indicators and, therefore, such
direct comparison is somewhat hindered.

5.3. Flow in a hypersonic inlet

The third test case discusses the computational results of the cold gas hypersonic flow in a
2-D inlet configuration which is representative of some proposed inlet geometries for typical
transatmospheric vehicles. This test case is well documented in Reference [2]. For the present
simulations, the fluid is treated as a perfect gas with constant specific heat and no chemistry is
taken into account. The purpose of the simulations is to compare the solutions of the implemented
non-oscillatory schemes in high Mach number flow conditions. Such simulations are aimed at
verifying if these schemes are able to represent all flow features, such as strong shocks, shock
reflections and interactions, and expansion regions. Moreover, there is interest in verifying whether
the schemes can avoid oscillations in the presence of such strong discontinuities. The adaptive
refinement process previously described is used in the computations, and the sensor is based on
density gradients.

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
936 W. R. WOLF AND J. L. F. AZEVEDO

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(a) X (b) X

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(c) X (d) X

Figure 15. Complete view of the inlet meshes: (a) initial coarse mesh with 1217 volumes; (b) mesh, after
one refinement pass in the initial coarse mesh, with 3091 volumes; (c) mesh, after two refinement passes
in the initial coarse mesh, with 8958 volumes; and (d) initial fine mesh with 4838 volumes.

Table II. Numbers of nodes and volumes for the meshes

used in the simulations of the flow in the inlet.
Mesh Nodes Volumes

Initial (coarse) 681 1217

After one refinement pass 1649 3091
After two refinement passes 4631 8958
Initial (fine) 2563 4838

The results considering an inlet entrance Mach number Me = 4 are discussed in the present work.
The simulations are performed for the initial fine mesh, without adaptive refinement, observed
in Figure 15(d), and for the adaptively refined mesh, after two refinement passes, observed in
Figure 15(c). The adaptively refined mesh is obtained using the flow solution computed by the
third-order WENO scheme. All meshes used in the present simulations are composed exclusively
by triangular control volumes. Results for quadrilateral meshes can be seen in Reference [23].
Table II indicates the characteristics of the initial and refined meshes. Moreover, in Figure 15, one
can observe a complete view of the meshes. The CFL number used for all the computations in
this test case is CFL = 0.2.
The comparison among the second-, third- and fourth-order ENO and WENO scheme results is
presented in Figures 16–18. One can observe in Figure 16 the results in terms of density contours
for the third-order WENO scheme. The results in Figure 16(a) are obtained using the mesh in
Figure 15(d), whereas those in Figure 16(b) are achieved with the mesh in Figure 15(c). Second-
and fourth-order results for density contours are not presented here due to the similarity of the
solutions in such overall view of the flowfield. However, the comparison among the schemes is

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
 HIGH-ORDER ENO AND WENO SCHEMES FOR UNSTRUCTURED GRIDS 937

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
X X

Figure 16. Inlet flow solutions for the third-order WENO scheme: (a) flow solution in terms of density
contours for Me = 4, 0, for the mesh in Figure 15(d) and (b) flow solution in terms of density contours
for Me = 4, 0, for the mesh in Figure 15(c).

0.8

0.7 ENO 2nd order ENO 2nd order

WENO 2nd order 1.5 WENO 2nd order
0.6 WENO 3rd order WENO 3rd order
WENO 4th order WENO 4th order
0.5
1
Cp

0.4
Cp

0.3

0.2 0.5

0.1

0 0
0.4 0.6 0.8 0.2 0.4 0.6 0.8 1
(a) X (b) X

Figure 17. C p distributions along the inlet walls for entrance Mach number Me = 4, 0,
obtained with the mesh of Figure 15(d): (a) C p distribution along the lower wall and
(b) C p distribution along the upper wall.

achieved by pressure coefficient (C p ) distributions along the inlet walls. Figures 17 and 18 present
the C p distributions obtained with the various schemes.
In Figure 16, one can observe the strong oblique shocks formed along the walls of the inlet
entrance. Such shock waves produce a shock–shock interaction at the dimensionless longitudinal
coordinate x = 0.58. An expansion fan occurs in the lower wall and interacts with a shock wave
resulting from the shock–shock interaction at x = 0.66. The shock wave which interacts with the
expansion fan reflects in the lower wall of the inlet, and the other shock wave resulting from the
shock–shock interaction reflects in the upper wall. Downstream of this region, one can observe
an expansion fan formed in the upper wall corner and another expansion fan in the second lower
wall corner. Furthermore, another shock–shock interaction occurs near the exit region of the inlet.
Finally, there is a last shock reflection in the lower wall, near the exit region of the inlet.
One can observe that the physical characteristics of the flow are better defined for the simulation
performed with the mesh obtained by adaptive refinement. In this test case, the two different
meshes are used to verify the influence of the adaptive refinement in the solutions. The simulations

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
938 W. R. WOLF AND J. L. F. AZEVEDO

0.6 ENO 2nd order ENO 2nd order

WENO 2nd order 1.5 WENO 2nd order
ENO 3rd order ENO 3rd order
WENO 3rd order WENO 3rd order
WENO 4th order WENO 4th order
0.4
1

Cp
Cp

0.2
0.5

0
0
0.4 0.6 0.8 0.2 0.4 0.6 0.8 1
(a) X (b) X

Figure 18. C p distributions along the inlet walls for entrance Mach number Me = 4, 0,
obtained with the mesh of Figure 15(c): (a) C p distribution along the lower wall and
(b) C p distribution along the upper wall.

performed in the present work have shown that higher than second-order ENO schemes are very
sensitive to the meshes. The second-order ENO and WENO schemes maintain stability for both
meshes used in this test. Third- and fourth-order WENO schemes present stability for both meshes,
similarly to the second-order schemes. However, the third- and fourth-order ENO schemes present
more sensitivity with respect to the meshes. While the third-order ENO scheme presents results
only for the adaptively refined mesh, the fourth-order ENO scheme does not maintain stability
with any of the inlet meshes here analysed.
Figure 17 shows the C p distributions along the lower and upper walls of the inlet for the mesh
appearing in Figure 15(d). The curves for the second-order ENO and WENO schemes and for the
third- and fourth-order WENO schemes are shown. The third- and fourth-order ENO schemes have
not converged for this computation. In the lower wall curve, one can clearly identify the entrance
shock, which is followed by a C p plateau, corresponding to the entrance ramp. This plateau is
ended by the expansion region associated with the first kink in the lower wall. Just downstream
the expansion region, there is another shock wave which increases the C p values. Finally, one
can observe the presence of an expansion fan and a last shock wave reflected in the wall. For the
upper wall, one can also see the presence of the entrance shock wave, followed by a fairly long C p
plateau associated with the long entrance ramp. The C p plateau is ended by the reflection of the
shock wave, which results from the shock–shock interaction at approximately x = 0.58. Finally,
one can note the appearance of an expansion fan, reducing the C p values until the exit section of
the inlet.
The results in Figure 17 indicate that the second-order ENO scheme does not capture correctly
the first expansion region in the lower wall of the inlet. The shock wave which impinges upon the
wall just downstream of the expansion region is more smeared for the solution with this scheme.
Therefore, the scheme cannot resolve the expansion region and the shock wave. The downstream
expansion region is also captured with a lower definition, when compared to the results of the
other schemes tested. The results for the second-, third- and fourth-order WENO schemes are very
similar, as far as capturing the discontinuities appearing in this flow is concerned. However, along

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
 HIGH-ORDER ENO AND WENO SCHEMES FOR UNSTRUCTURED GRIDS 939

the lower wall, the fourth-order WENO scheme captures the first expansion region and the second
shock wave more accurately.
Figure 17 also shows that, in the upper wall, the ENO scheme presents results with lower
resolution when compared to the other schemes. The more accurate results are obtained by the
third- and fourth-order schemes, as one can see in the second shock wave resolution. The second-
order WENO scheme does not present the same peak, after the second shock reflection, for the
C p distribution as the one achieved by the third- and fourth-order WENO schemes, as can be
observed in Figure 17(b).
The results presented in Figure 18 consider the C p distributions along the lower and upper walls
for the mesh appearing in Figure 15(c). All the schemes present a small oscillation downstream
the first entrance shock wave. The second-order ENO scheme presents a more smeared solution
to the expansion regions in the lower inlet wall, and the third-order ENO scheme presents strong
oscillations in the first expansion region and smaller oscillations immediately downstream the
second shock wave. The WENO schemes present similar solutions regarding the expansion regions
and the second shock wave. In the last shock reflection region, in the lower wall, near the exit section
of the inlet, the third-order scheme presents the highest value for the C p , while the second-order
ENO scheme presents the lowest value.
In Figure 18, for the C p distributions along the upper wall, the schemes present an oscillation
inside the first shock wave. This fact is related to the mesh definition in the region where the shock
occurs. Immediately upstream the second shock region, ending the C p plateau, the fourth-order
WENO scheme presents a small oscillation. One can observe in Figure 18(b) that the second-order
ENO scheme presents the more diffusive results among the schemes studied in this test case. For
all schemes tested, some oscillations can also be observed in the second C p plateau, downstream
of the second shock.

5.4. Flow over a blunt body

The last test case presented considers the cold gas hypersonic flow over a 2-D blunt body with
freestream Mach number M∞ = 20.0 and angle of attack  = 0◦ . Again, the perfect gas hypothesis
is assumed, despite the high Mach numbers appearing in the studied flow. The aim of this test is to
verify the behaviour of the high-order ENO and WENO schemes with relation to the appearance of
the carbuncle phenomenon [24]. The carbuncle phenomenon is characterized by the appearance of
a spurious disturbance behind a detached shock wave that forms over blunt bodies. The carbuncle
problem was first reported in 1988 in the work of Peery and Imlay [24]. Such a numerical problem
presents a stable solution which includes a recirculation bubble situated in front of the stagnation
region of the flow.
In the present test case, a mesh with 1668 nodes and 3127 control volumes is used. This mesh
can be viewed in Figure 19(a). The solution for this simulation is shown in Figure 19(b) in terms of
pressure contours for the third-order WENO scheme. The CFL number used in the computations
for this test case is CFL = 0.2. The results for the second- and fourth-order ENO and WENO
schemes and for the third-order ENO scheme are not shown in the paper since they are very
similar to the third-order WENO scheme results. The computations for this test case use the Roe
flux difference splitting scheme as numerical flux function.
One can observe in the detail of the stagnation region in Figure 19(c) that the third-order
WENO scheme results do not present the typical recirculation bubble characteristic of the carbuncle
phenomenon. The other ENO and WENO scheme results also do not present the recirculation bubble

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
940 W. R. WOLF AND J. L. F. AZEVEDO

6 6

5 5

4 4

3 3
Y

Y
2 2

1 1

0 0
-2 -1 0 1 2 -2 -1 0 1 2
(a) X (b) X

1 1

0.5 0.5
Y

0 0

-1.5 -1 -0.5 0 -1.5 -1 -0.5 0

(c) X (d) X

Figure 19. Pressure contours for the M∞ = 20.0 flow over the 2-D blunt body: (a) mesh for the 2-D
blunt body; (b) third-order WENO scheme flow solution; (c) detail of the stagnation region for the flow
solution obtained with the third-order WENO scheme; and (d) detail of the stagnation region for the
flow solution obtained with the TVD–MUSCL scheme.

for this simulation. In Figure 19(d), one can observe the appearance of the recirculation bubble
present in the stagnation region of the blunt body for a solution obtained with the TVD–MUSCL
Roe scheme for flux evaluation.

5.5. Computational cost

In this paper, algorithms for third- and fourth-order stencil selection were implemented for the
ENO and WENO schemes. The results in terms of the averaged computational cost are shown in
Table III. The data appearing in the table were obtained by time measurements of the simulations
considering a pre-defined number of iterations. Then, all the time measurements were divided by
the time of the MUSCL simulation for a mesh composed exclusively of triangular control volumes.

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
 HIGH-ORDER ENO AND WENO SCHEMES FOR UNSTRUCTURED GRIDS 941

Table III. Comparison of the averaged computational cost of MUSCL,

ENO and WENO schemes.
Averaged iteration cost Averaged iteration cost
Scheme per triangular volume per quadrilateral volume

MUSCL second-order 1.0 1.4

ENO second-order 1.4 2.1
WENO second-order 1.9 2.9
ENO third-order 4.7 10.5
WENO third-order 5.1 11.9
ENO fourth-order 7.6 37.8
WENO fourth-order 8.7 39.2

One can observe in Table III that the second-order ENO and WENO schemes present a compu-
tational cost between 1.4 and 2.9 times the cost of the MUSCL scheme. The higher computational
cost of the non-oscillatory schemes is related to the use of the algorithms for stencil selection and,
in the case of the WENO schemes, to the computation of the weights and the WENO polynomials.
The third-order ENO and WENO schemes present computational costs of 3.3 and 2.7 times the
costs of the respective ENO and WENO second-order schemes for a triangular mesh. One can
observe, however, that the relation of these costs for a quadrilateral mesh is higher than for a
triangular mesh. The costs increasing for quadrilateral meshes are directly related to the larger
number of admissible stencils for third-order schemes when compared to second-order schemes
in quadrilateral meshes.
As shown in Table III, the fourth-order schemes present the highest computational costs. Again,
one can observe the higher costs of the schemes for quadrilateral meshes when compared to
triangular meshes. Beyond the larger number of admissible stencils of the third- and fourth-order
schemes, one should mention that for some test cases the order of the schemes must be reduced.
The entire process of reducing the order of accuracy of the schemes is another factor that turns
the computational cost higher. The authors also point out that the implementation of the ENO and
WENO schemes in the code can be optimized. However, this effort is beyond the scope of the
present paper.
Finally, one should mention that despite the fact that third- and fourth-order schemes present
higher computational costs than second-order schemes, the higher order schemes yield solutions
with a better definition of the flow features. These flow features are not captured by schemes of
lower order of accuracy. For lower order schemes, it is necessary to use more refined meshes
and, hence, the computational costs are higher. As one can observe, the forward-facing step case
appearing in a previous section of the paper presents a comparison of solutions for schemes with
different orders of accuracy using equal meshes. One can observe the sharp definition of the flow
features for the third- and fourth-order schemes compared to the solutions of the second-order
schemes. However, a more complete study of this type is beyond the scope of this paper.

6. CONCLUSIONS

The reconstruction of third- and fourth-order essentially non-oscillatory (ENO) schemes and
weighted essentially non-oscillatory (WENO) schemes is presented in this work. Although, in

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
942 W. R. WOLF AND J. L. F. AZEVEDO

this paper, only third- and fourth-order accurate schemes are actually implemented and assessed,
the formulation of ENO and WENO reconstruction is treated in a generic framework that allows
the construction of polynomials of any order and, hence, of schemes with an arbitrary order of
accuracy. The flux difference splitting scheme of Roe is used as the numerical flux function in the
simulations performed for this paper, and an adaptive refinement technique is employed in order to
yield a sharper representation of the strong gradients appearing in the flows. The test cases studied
in this work include the Ringleb flow, the supersonic flow in a channel with a forward-facing step,
the flow in a typical hypersonic inlet, and the high speed flow over a 2-D blunt body.
The analysis of the effective order of accuracy of the schemes is presented for the computation
of the Ringleb flow problem. The third-order WENO scheme was able to obtain an actual order
of accuracy of 2.66, which is fairly close to the nominal one. The third-order ENO scheme and
the fourth-order ENO and WENO schemes only reached second-order accuracy. These results are
strongly related to the reconstructions close to the boundary regions.
The solutions for the quadrilateral mesh in the forward-facing step flow present a non-physical
shock–boundary layer interaction with a recirculation bubble after the Y shock. The solutions for
the triangular meshes show that the fourth-order WENO scheme is the only method able to capture
the weak shock, in the step corner region, for the coarse triangular grid. However, for the fine
triangular grid, the third-order WENO scheme captured the weak shock with a better resolution
than the fourth-order scheme. The second-order WENO scheme presents a solution with much
more definition for this case than the second-order MUSCL scheme.
For the flow in the hypersonic inlet, the WENO schemes have proven to be more robust than the
ENO schemes. The third- and fourth-order ENO schemes have shown to be more mesh dependent
for this case. One can observe that the C p distributions along the inlet walls are very similar for
the WENO schemes. The authors believe that this fact has to do with the decrease in the order
of accuracy in the boundaries. Such a decrease is caused by the lack of a sufficient number of
cells to reconstruct the third- and fourth-order scheme polynomials. With regard to the flow over
the blunt body, one can observe that the ENO and WENO solutions do not present the carbuncle
problem for the mesh and configuration used in the present work. However, for the TVD–MUSCL
scheme, the carbuncle phenomenon is observed in the solution.

ACKNOWLEDGEMENTS
The authors gratefully acknowledge the support of Fundação de Amparo à Pesquisa do Estado de São Paulo
(FAPESP) through a Masters Scholarship for the first author under the FAPESP Grant No. 03/10047-2.
The authors also acknowledge the partial support of Conselho Nacional de Desenvolvimento Cientı́fico
e Tecnológico (CNPq) under the Integrated Project Research Grant No. 501200/2003-7. The authors are
also thankful to Mr Edson Basso for the very fruitful discussions during the course of the present research.

REFERENCES
1. Wolf WR, Azevedo JLF. High-order unstructured grid ENO and WENO schemes applied to aerodynamics flows.
Proceedings of the 17th AIAA Computational Fluid Dynamics Conference, AIAA Paper No. 2005-5115, Toronto,
Canada, 2005.
2. Azevedo JLF, Korzenowski H. Comparison of unstructured grid finite volume methods for cold gas hypersonic
flow simulations. Proceedings of the 16th AIAA Applied Aerodynamics Conference, AIAA Paper No. 98-2629,
Albuquerque, New Mexico, 1998.
3. Figueira da Silva LF, Azevedo JLF, Korzenowski H. Unstructured adaptive grid flow simulations of inert and
reactive gas mixtures. Journal of Computational Physics 2000; 160(2):522–540.

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld
 HIGH-ORDER ENO AND WENO SCHEMES FOR UNSTRUCTURED GRIDS 943

4. Azevedo JLF, Figueira da Silva LF, Strauss D. Order of accuracy study of unstructured grid finite volume
upwind schemes. Journal of the Brazilian Society of Mechanical Sciences and Engineering 2007, accepted for
publication.
5. van Leer B. Flux–vector splitting for the Euler equations. Proceedings of the 8th International Conference on
Numerical Methods in Fluid Dynamics. Lecture Notes in Physics, vol. 170. Springer: Berlin, 1982; 507–512.
6. Liou MS. A sequel to AUSM:AUSM+. Journal of Computational Physics 1996; 129(2):364–382.
7. Roe PL. Approximate Riemann solvers, parameter vectors, and difference schemes. Journal of Computational
Physics 1981; 43(2):200–212.
8. Anderson WK, Thomas JL, van Leer B. A comparison of finite volume flux vector splittings for the Euler
equations. AIAA Journal 1986; 24(9):1453–1460.
9. Hirsch C. Numerical Computation of Internal and External Flows, vol. 2. Wiley: New York, 1990.
10. Harten A, Osher S, Engquist B, Chakravarthy SR. Uniformly high-order accurate essentially non-oscillatory
schemes III. Journal of Computational Physics 1987; 71(2):231–303.
11. Liu XD, Osher S, Chan T. Weighted essentially non-oscillatory schemes. Journal of Computational Physics 1994;
115(1):200–212.
12. Sonar T. On the construction of essentially non-oscillatory finite volume approximations to hyperbolic conservation
laws on general triangulations: polynomial recovery, accuracy and stencil selection. Computer Methods in Applied
Mechanics and Engineering 1997; 140(2):157–181.
13. Friedrich O. Weighted essentially non-oscillatory schemes for the interpolation of mean values on unstructured
grids. Journal of Computational Physics 1998; 144(1):194–212.
14. Shu CW, Osher S. Efficient implementation of essentially non-oscillatory shock-capturing schemes. Journal of
Computational Physics 1988; 77(2):439–471.
15. Gooch CFO. High order ENO schemes for unstructured meshes based on least-squares reconstruction. Report
No. P631-1296, Argonne National Laboratory, Mathematics and Computer Science Division, 1997.
16. Abgrall R. On essentially non-oscillatory schemes on unstructured meshes: analysis and implementation. Journal
of Computational Physics 1994; 114(1):45–58.
17. Hu C, Shu CW. Weighted essentially non-oscillatory schemes on triangular meshes. Journal of Computational
Physics 1999; 150(1):97–127.
18. Jiang GS, Shu CW. Efficient implementation of weighted ENO schemes. Journal of Computational Physics 1996;
126(1):77–99.
19. Harten A, Chakravarthy SR. Multi-dimensional ENO schemes for general geometries. ICASE Report No. 91-76,
1991.
20. Shapiro AH. The Dynamics and Thermodynamics of Compressible Fluid Flow, vol. 2. The Ronald Press: New
York, 1954.
21. Vankeirsbilck P, Deconinck H. Higher order upwind finite volume schemes with ENO-properties for general
unstructured meshes. AGARD Report No. 787, 1992.
22. Woodward P, Colella P. The numerical simulation of two dimensional fluid flow with strong shocks. Journal of
Computational Physics 1984; 54(1):115–173.
23. Wolf WR, Azevedo JLF. Essentially non-oscillatory schemes on cold gas hypersonic flow simulations. Proceedings
of the XXVI Iberian Latin–American Congress on Computational Methods in Engineering, Guarapari, ES, Brazil,
2005.
24. Perry KM, Imlay ST. Blunt body flow simulations. 24th AIAA Joint Propulsion Conference, AIAA Paper No.
88-2904, Boston, MA, 1988.

Copyright q 2007 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2007; 55:917–943
DOI: 10.1002/fld