l=1
N
G
m=1
(F n )
i,j,l,m
+
1
A
i,j
_
V
i,j
i,j
(x, y, u)dv = R
i,j
, (2)
where each cell has N
f
=4 faces and a N
G
-point Gaussian quadrature numer-
ical integration procedure is used to evaluate the solution ux, the sum of
the advective and diusive uxes F = F
a
+F
d
= Vu u, through each
face. The variable u
i,j
is the average solution scalar, A
i,j
is the cell area, is
the quadrature weighting coecient, and n are the length of the cell face
and unit vector normal to the cell face, respectively. Analytical or numerical
integration can be performed to calculate the integral of the particular non-
linear source term,
i,j
, for cell (i, j). In this work, either a two- or four-stage
standard Runge-Kutta scheme is used to integrate the system of ordinary
dierential equations given by (2), depending on the desired accuracy.
The hyperbolic uxes, F
a
n = u V n, at each quadrature point are
determined from the left and right solution values, u
l
and u
r
, as follows:
F
a
n =
_
u
l
(V n) if V n 0,
u
r
(V n) if V n < 0.
(3)
The solution states, u
l
and u
r
, are determined by performing piecewise
k-order polynomial solution reconstruction within each computational cell.
Herein, the k-order CENO reconstruction proposed by Ivan and Groth is used
[IG07].
The high-order central ENO (CENO) method is not based on either se-
lecting or weighting reconstructions from multiple stencils. Instead, a hy-
brid solution reconstruction procedure is used that combines the high-order
k-exact least-squares reconstruction technique of Barth [Bar93] based on a
xed central stencil with a monotonicity preserving limited piecewise linear
least-squares reconstruction algorithm [Bar93]. The limited reconstruction
procedure is applied to computational cells with under-resolved solution con-
tent and the unlimited k-exact reconstruction scheme is used for cells in
which the solution is fully resolved. Switching in the hybrid procedure is
determined by a solution smoothness indicator. Note that for smooth and
hyperbolic problems, a k-order reconstruction produces a k+1-order accu-
rate spatial discretization. A detailed description of the CENO reconstruction
High-Order Central ENO Finite-Volume Scheme 445
procedure and of the smoothness indicator is given in the paper by Ivan and
Groth [IG07].
As previously mentioned, the proposed discretization seeks to obtain a
global k-order accurate scheme on arbitrary meshes. This implies the use of
a k-order accurate gradient for the diusive ux evaluation, which here is
derived from a k+1-order accurate reconstruction. For this reason, even if a
quartic (k =4) reconstruction has the potential to generate a 5th-order dis-
cretization of the hyperbolic ux in combination with at least three quadra-
ture points (N
G
=3), it will only be 4th-order accurate for the elliptic ux
discretization. Therefore, two quadrature points (N
G
= 2) are sucient to
obtain a global 4th-order accurate scheme with a piecewise quartic recon-
struction.
In a similar manner to hyperbolic uxes, numerical elliptic uxes, F
d
n =
u n , must be evaluated at each quadrature point. Having determined
the left and right solution reconstructions, u
k
l
(r) and u
k
r
(r), the solution gra-
dient at the inter-cellular face is obtained as the arithmetic mean of the
reconstruction gradients and thus, the ux at the calculation point, r, is
evaluated as
F
d
n =
_
1
2
_
u
k
l
(r) +u
k
r
(r)
_
_
n. (4)
The accuracy of (4) can be easily observed to be k-order accurate. To in-
fer other properties such as positivity (i.e. local satisfaction of a discrete
maximum principle) or non-existence of odd-even solution decoupling, it is
convenient to apply the elliptic discretization to the Laplace operator and
analyse the inuence coecient of each entry in the supporting stencil [Coi94].
Note that for a given discretization the inuence coecients depend only on
the mesh geometry and not on the actual solution. As proposed by Coirier
[Coi94], the positivity of the scheme can be characterized in terms of
0
and
min
coecients. Ideally,
0
< 0 for stability and
min
= 0 for positivity
[DABLP99].
In the current work, dierent mesh geometries were analysed including
Cartesian, stretched, and randomly disturbed quadrilateral grids. The anal-
ysis has shown that odd-even solution decoupling does not occur. In terms
of the stability and positivity, it was found that
0
< 0 but
min
< 0 for
discretizations of all order, unfortunately implying that, while stable, none of
the discretizations satisfy a discrete maximum principle. Note that for square
Cartesian meshes, values for
min
are found to be -0.823 for k=2, -0.362 for
k=3 and -0.854 for k=4 when inverse distance geometric weighting is used in
the k-exact reconstruction. However, the positivity can be improved by using
an inverse distance squared geometric weighting, for which
min
was found
to be -0.051 for k=2, -0.247 for k=3 and -0.324 for k=4. For non-Cartesian
meshes, large variations in the value of
min
are possible (5 <
min
< 0),
depending on the regularity and topology of the mesh.
446 L. Ivan and C.P.T. Groth
High-order treatment of boundary conditions (BCs) is crucial for develop-
ing high-order accurate schemes. Herein, high-order BCs have been imposed
by making use of extra rows of ghost cells or by constraining the least-squares
reconstruction in cells adjacent to boundaries as described in [OV02].
A exible block-based hierarchical data structure is used in conjunction
with the CENO scheme described above to facilitate automatic solution-
directed mesh adaptation on body-tted multi-block quadrilateral mesh. In
this work, an h-renement criterion based on the solution smoothness in-
dicator is used to control the renement of AMR mesh. The AMR CENO
algorithm and the renement criterion are described in details in [IG07].
3 Numerical Results
The accuracy of the hybrid CENO algorithm was rst assessed based on the
circular advection at constant angular velocity of the smooth inow varia-
tion u(x, 0) = e
d
sin
6
( d) if d [0 : 1], otherwise 0, where d = x 0.3. The
predicted solution distribution for this problem using the quartic CENO re-
construction on a 8080 Cartesian mesh is shown in Fig. 1. The error norms
for both 4th-order versions (k =3 and k =4) of the proposed CENO scheme
are also depicted in Fig. 1. As the mesh is rened, the slopes of the L
1
-, L
2
-
and L
-norms approach in the asymptotic limit -4.53, -4.55 and -4.58 for
the cubic reconstruction (k =3), and -4.94 in all error norms for the quartic
reconstruction (k =4), respectively, indicating that the expected theoretical
order of accuracy has been achieved in each case.
The high-order CENO scheme was also applied in conjunction with AMR
to a problem similar to the previous one. In this case the inow function
was u(x, 0) = e
2 d
sin
6
(2 d) if d [0 : 0.8], otherwise 0, where d = x 0.4.
This function exhibits two smooth extrema and a discontinuity close to the
second peak. Figure 2 shows the nal solution prole along the cross-section
x
y
0 0.5 1 1.5
0
0.5
1
1.5
u
1.63
1.40
1.17
0.95
0.72
0.49
0.26
0.03
Inflow
O
u
t
f
l
o
w
Farfield
F
a
r
f
i
e
l
d
N
1/2
1
,
2
,
1
(k=3)
2
(k=3)
(k=3)
1
(k=4)
2
(k=4)
(k=4)
5.0
4.0
Fig. 1. Predicted 4th-order (k=4) CENO solution (left) and L
1
, L
2
and L
error
norms for k=3 and k=4 CENO reconstructions (right)
High-Order Central ENO Finite-Volume Scheme 447
Distance
0 0.5 1 1.5
Exact Soln
u
x
y
0 0.5 1 1.5
0
0.5
1
1.5
A
A
Fig. 2. Comparison of 4th-order (k =3) AMR CENO solution on the nal mesh
and exact solution along A-A (left) and nal mesh with 2,911 1010 blocks (right)
A-A compared against the exact solution and the nal multi-block AMR
mesh and the location of line A-A. The results clearly show that the pro-
posed AMR scheme in conjunction with the h-renement criteria based on
the smoothness indicator of the hybrid CENO reconstruction is capable of
rening both under-resolved (inaccurate) and non-smooth regions of the so-
lution and will not unnecessarily rene resolved solution content. The smooth
peaks are all well captured whereas the solution discontinuity is well identied
by the smoothness indicator and well resolved by the AMR procedure.
The numerical scheme was also tested for solutions to the Laplace equation
for the curved boundary domain shown in Fig. 3. Dirichlet BCs were imple-
mented based on the exact solution u(x, y) = e
x
(cos(y) + 2 sin(y)),
= 1.5. A 4th-order solution to this problem and the L
1
, L
2
, and L
error
norms for cubic and quartic interpolants are also shown in Fig. 3. The slopes
of the L
1
- and L
2
-norms reach in the asymptotic limit -3.86 and -3.85 for k=3
and -3.86 and -3.81 for k=4, respectively. Even if the orders of accuracy for
these two interpolants are essentially the same and close to the theoretical
value, there is about one order of magnitude dierence between the absolute
error values, demonstrating the benets of using a quartic interpolant.
The proposed scheme is now applied to problems involving dierent ratios
of advection and diusion terms as determined by the Peclet number (Pe).
Solution of (1) with V = (v
0
, 0) and (x, y) = 0.01, for the BCs shown in
Fig. 4 is considered for Pe = 0.1, Pe = 1, and Pe = 10, depending on v
0
.
Figure 4 shows the numerical solution obtained for Pe = 10 on an 8040
Cartesian mesh and the error norms for the three Peclet numbers. The results
showthat the errors generated by the quartic polynomial are consistently lower
than those of the cubic interpolant by at least one order of magnitude for all
cases and obtain the theoretical accuracy in all error norms. Thus, the L
1
-norm
for k=4 is -4.02 (Pe=0.1), -4.30 (Pe=1.0) and -3.92 (Pe=10.0), respectively.
In the case of cubic interpolant, the L
1
-normis -3.92 (Pe=0.1), -3.88 (Pe=1.0)
448 L. Ivan and C.P.T. Groth
x
y
0 1 2 3 4 5
4
5
6
7
8
u
728.20
671.87
615.55
559.22
502.90
446.57
390.25
333.92
277.60
221.27
168.61
127.40
71.07
14.75
-4.03
-41.58
-97.91
-154.23
Quartic Soln (k=4), 40x40
N
1/2
1
,
2
,
1
(k=3)
2
(k=3)
(k=3)
1
(k=4)
2
(k=4)
(k=4)
3.0
4.0
Fig. 3. Fourth-order (k = 4) solution to the Laplace equation on a 4040 grid
(left) and L
1
-, L
2
- and L
y
)
Dirichlet, u(x,0)=0
O
u
t
f
l
o
w
,
u
(
x
,
y
)
/
y
=
0
N
1/2
1
,
2
,
1
(k=3)
2
(k=3)
(k=3)
1
(k=4)
2
(k=4)
(k=4)
Pe = 0.1
Cartesian Mesh
3.0
4.0
N
1/2
1
,
2
,
1
(k=3)
2
(k=3)
(k=3)
1
(k=4)
2
(k=4)
(k=4)
Pe = 1.0
Cartesian Mesh
3.0
4.0
N
1/2
1
,
2
,
2
(k=3)
(k=3)
1
(k=4)
2
(k=4)
(k=4)
Pe = 10.0
Cartesian Mesh
4.0
3.0
Fig. 4. Fourth-order solution for a channel ow problem with Pe=10 (upper-left)
and L
1
-, L
2
- and L