INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN FLUIDS

Int. J. Numer. Meth. Fluids 2009; 00:124 Prepared using dauth.cls [Version: 2002/09/18 v1.01]
High-order nite dierence schemes for incompressible ows
H. Fadel
1,
, M. Agouzoul
1
, P.K. Jimack
2
1
ERD3M, Departement Mecanique, Ecole Mohammadia dIngenieurs, BP765 Rabat-Agdal Morocco
2
School of Computing, University of Leeds, Leeds LS2 9JT UK
SUMMARY
This paper presents a new high order approach to the numerical solution of the incompressible Stokes
and Navier-Stokes equations. The class of schemes developed is based upon a velocity-pressure-pressure
gradient formulation which allows: (i) high order nite dierence stencils to be applied on non-
staggered grids; (ii) high order pressure gradient approximations to be made using standard Pade
schemes, and; (iii) a variety of boundary conditions to be incorporated in a natural manner. Results
are presented in detail for a selection of two-dimensional steady-state test problems, using the fourth
order scheme to demonstrate the accuracy and robustness of the proposed methods. Furthermore,
extensions to higher orders and time-dependent problems are illustrated, whilst the extension to
three-dimensional problems is also discussed. Copyright c 2009 John Wiley & Sons, Ltd.
key words: Incompressible ow; Steady ow; Velocity-Pressure-Pressure gradient formulation;
Finite dierence methods; High-order accuracy; Pade schemes
1. INTRODUCTION
Over recent years, the development of high-order accurate schemes has become an attractive
approach to improving the eciency of numerical computations for more and more complex
problems. Indeed, these schemes can be used either to reduce the computational costs for a
given accuracy by reducing the mesh size or to increase the accuracy for a xed mesh size.
The motivation for this work is to propose a new class of high-order nite dierence schemes
for steady ows of incompressible Newtonian uids governed by Stokes or Navier-Stokes
equations. In particular, our goal is to obtain high-order accurate solutions for both the velocity
and the pressure, as well as the pressure derivatives, by choosing an appropriate formulation,
without the need for any other approximations than the discretization of the equations
governing the ow modelled. Although the numerical schemes and examples described here
are all in two dimensions, the extension to three dimensional cases is straightforward.

Correspondence to: [email protected]

Contract/grant sponsor: Morocco CNRST; contract/grant number: a1/038
Contract/grant sponsor: Marie Curie; contract/grant number: MEST-CT-2005-020327
Copyright c 2009 John Wiley & Sons, Ltd.
2 H. FADEL ET AL.
The problem can be stated as follows. Let be the problem domain in
2
, then the classical
steady-state Navier-Stokes (NS) equation system in primitive variable form is written as:

2
u +
1

p + (u )u = f (1a)
u = 0 (1b)
where u = (u, v)
T
denotes the velocity vector, p the static pressure eld, the kinematic
viscosity, the uid density, f = (f
x
, f
y
)
T
an external force vector and x = (x, y)
T
the
Cartesian coordinate variable. We refer to equations (1a) as the momentum equations and to
equation (1b) as the continuity equation.
To specify a unique solution, we consider the NS problem (1) together with the appropriate
boundary conditions on =
D

N
:
u = u
D
on
D
, (2a)

n
u
n
p = f
n
on
N
, (2b)

n
u

= f

on
N
. (2c)
Here the subscripts D and N denote a boundary of Dirichlet and Neumann type respectively,
n is the normal component and the tangential one. Equation (2a) is referred to as a Dirichlet
boundary condition and equations (2b) and (2c) as open outow boundary conditions.
Solving the system (1) under the boundary conditions (2), in the given form, using nite
dierence methods on a single grid leads to spurious pressure solutions. To avoid this a
range of dierent approaches have been developed. The approach that we adopt in this
work is to consider a velocity-pressure formulation obtained by replacing the incompressibility
condition by a pressure Poisson equation. This equation is derived by taking the divergence
of the momentum equations to obtain a consistent pressure equation and using the continuity
equation to eliminate the divergence of the velocity to obtain a further simplication. A detailed
discussion on this subject and related issues can be found in [1, 2]. We can now reformulate the
problem by replacing the continuity equation in the Navier-Stokes system (1) by the following
simplied pressure equation:

2
p = ( (u u) + f ) . (3)
Knowing that :
(u u) = u : u + u ( u) (4)
and further use of the zero velocity divergence leads to the following form [3]:

2
p = (2(
x
u
y
v
y
u
x
v) + f ) . (5)
In order to complete the mathematical formulation, it is necessary to specify appropriate
boundary conditions for the pressure. This is complicated by the fact that, a priori, we have no
explicit information about the pressure at the boundary. Nevertheless, such a condition may
be obtained by taking the scalar product of the momentum equations with the unit normal
vector n at the boundary. On its own this is not sucient however since the momentum
equations are already satised on the boundary, so these conditions do not in themselves
add any new information and the system remains undetermined. Moreover, the velocity-
pressure formulation can not be equivalent to the primitive variable one without enforcing
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
HIGH-ORDER FINITE DIFFERENCE SCHEMES FOR INCOMPRESSIBLE FLOWS 3
the divergence-free condition on the boundaries [1, 4, 5, 6]. Hence, by using this condition and
the following identity:

2
u = ( u) u , (6)
the derived curl-curl consistent boundary condition for the pressure is:

n
p = n (u (u )u +f ) on
N
. (7)
Therefore, and when properly normalised, our reference problem is now expressed by:

1
Re

2
u +p + (u )u = f (8a)

2
p = 2(
x
u
y
v
y
u
x
v) + f , (8b)
with the appropriate boundary conditions:
u = u
D
, on
D
, (9a)
1
Re

n
u
n
p = f
n
on
N
, (9b)
1
Re

n
u

= f

on
N
, (9c)

n
p = n
_

1
Re
u (u )u +f
_
on
N
. (9d)
Here, the Reynolds number is dened by Re = U
0
L
0
/ for some characteristic velocity U
0
and
length scale L
0
.
Now that the mathematical formulation is set, we propose to investigate a velocity-pressure-
pressure gradient formulation to ensure a high order accuracy of the obtained results, especially
for the pressure. This formulation is simply dened by the fact that the pressure gradient is
considered as unknown in the previous velocity-pressure formulation. In order to approximate
these additional variables we use the Pade schemes [7, 8, 9, 10], often referred to in the literature
as compact schemes. Shih et al, [11], proposed a similar approach (scheme (i) in [11]): using
the primitive variable formulation but with the velocity and pressure rst derivatives also
considered as unknowns and approximated using a Pade 4
th
order scheme within the domain
and a 4
th
order one-sided Taylor approximation on the boundaries. In our approach, we study
the use of the 4
th
order Pade scheme only for the pressure derivative and the same order Pade
non-central scheme on the boundaries.
An outline of the remainder of this paper is as follows. In sections 2 and 3 we introduce the
numerical methods together with the solution strategy for Stokes and Navier-Stokes equations
respectively. In section 4 we then discuss results obtained for dierent test cases. Finally, our
conclusions and perspectives on future work are presented in section 5.
2. NUMERICAL METHODS FOR STOKES EQUATIONS
2.1. Discretization
In the particular case where it is possible to assume that convection eects may be safely
neglected and Re 0, the problem (8-9) may be reformulated to the following Stokes
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
4 H. FADEL ET AL.
equations:

2
u +p = f (10a)

2
p = f , (10b)
together with the appropriate boundary conditions:
u = u
D
, on
D
, (11a)

n
u
n
p = f
n
on
N
, (11b)

n
u

= f

on
N
, (11c)

n
p = n(u +f ) on
N
. (11d)
To discretize the systems momentum equations (10a) we express them as Poisson equations,
this allows the system (10) to be written as:

2
w = g on (12)
where
w =
_
_
u
v
p
_
_
and g =
_
_

x
p f
x

y
p f
y

x
f
x
+
y
f
y
_
_
. (13)
The solution w and the forcing function g are assumed to be suciently smooth and have the
required continuous partial derivatives.
The solution of equation (12) can be approximated using the standard fourth-order scheme
derived by Rosser [12, 13] and represented here by the molecular display:
_
_
1 b 1
a 20 a
1 b 1
_
_
w = c
_
_
0 1 0
1 8 1
0 1 0
_
_
g , (14)
with
a = 2
(5y
2
x
2
)
x
2
+ y
2
, b = 2
(5x
2
y
2
)
x
2
+ y
2
, c =
(x
2
y
2
)
x
2
+ y
2
such that x and y are the grid size in the x-direction and the y-direction respectively.
The last point remaining to complete the discretization of our reference problem is the
treatment of the Neumann boundary conditions (11b), (11c) and (11d). These include rst
order, second order and mixed second order partial derivatives that are approximated using
standard Taylor series, consistent with the order of our scheme (see appendix A for full details).
2.2. Pade Schemes
As stated previously, the schemes developed in this work are based on a velocity-pressure-
pressure gradient formulation. To dene the additional equations required for the pressure
gradient, that is considered as unknown, we propose to consider the standard Pade schemes.
For a function f, these schemes are dened by :
f

i2
+ f

i1
+ f

i
+ f

i+1
+ f

i+2
= c
f
i+3
f
i3
6h
+ b
f
i+2
f
i2
4h
+ a
f
i+1
f
i1
2h
, (15)
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
HIGH-ORDER FINITE DIFFERENCE SCHEMES FOR INCOMPRESSIBLE FLOWS 5
such that i indexes the nodes uniformly dividing a mesh of size h, and f

i
is an approximation
of the rst derivative of f at the node i. The coecients a, b, c, and (note that a, b, and
c are dierent from teh previous subsection) are derived by developing all the variables in the
Taylor series for the mesh node i and requiring all the coecients of the resulting expansion
to vanish up to the required order. The rst non-zero coecient sets the formal truncation
error, and hence the order of accuracy, of the scheme [10].
To obtain fourth-order accuracy, the relation between these coecients is given by:
a + 2
2
b + 3
2
c = 2
3!
2!
( + 2
2
) . (16)
Here we are interested in the family of tridiagonal 4
th
order schemes that are dened by:
= 0, a =
2
3
( + 2), b =
1
3
(4 1), c = 0. (17)
By choosing = 1/4, we recover the classical Pade scheme cited by Collatz [8]:
f

i1
+ 4f

i
+ f

i+1
=
3
h
(f
i1
+ f
i+1
) , (18)
with a leading truncation error of (2/5!)h
4
f
(5)
. However, this scheme is central and therefore
it can only be applied inside the domain . We therefore need a boundary closure formula for
nodes on the boundary. Such a 4
th
order forward formula, at a grid node i = 0, is given by:
f

i
+ 3f

i+1
=
1
6h
(17f
i
+ 9f
i+1
+ 9f
i+2
f
i+3
) , (19)
with a leading truncation error of (6/5!)h
4
f
(5)
.
2.3. Solution strategy
It is now clear that our problem contains ve unknowns, rather than three, at each non-
Dirichlet node. These are the velocity components u and v, the pressure eld p and the pressure
derivatives
x
p and
y
p. To summarise, the scheme developed for the Stokes ow is based on
the following discretization strategy.
Velocity solution:
In : 9-point Poissons Formula (14)
On
D
: u values treated as known values
On
N
: Taylor approximations for u and v derivatives.
Pressure solution:
In : 9-point Poissons Formula (14)
On
D
: one p value specied to avoid the hydrostatic pressure mode
On
N
: Taylor approximations for u, v and p derivatives.
Pressure gradient solution:
In : Standard Pade scheme (18)
On : Pade boundary closure (19).
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
6 H. FADEL ET AL.
The list of unknowns and dierence equations at each node is detailed in appendix A. Due
to the linearity of the system of discretized equations they may be written in matrix form.
This leads to an unsymmetric matrix and, in the case of a problem with Dirichlet BC for the
velocity components and Newmann BC for the pressure, the system takes the following form:
Ax =
_

_
A
1
0 0 A
2
0
0 A
1
0 0 A
2
A
3
A
4
A
5
0 0
0 0 A
6
A
7
0
0 0 A
8
0 A
9
_

_
_
_
_
_
_
_
u
v
p

x
p

y
p
_
_
_
_
_
_
= F . (20)
The sparsity patterns of each of the sub-matrices A
1
to A
9
are plotted for a mesh size [1010]
in gure 1.
3. NUMERICAL METHODS FOR NAVIER-STOKES EQUATIONS
3.1. Discretization
For the Navier-Stokes problem (8-9), as for the Stokes case, we consider the momentum
equations in the Poisson equation form (18), where
w =
_
_
u
v
p
_
_
and g = g
1
+ g
2
, (21)
with
g
1
=
_
_
Re(
x
p f
x
)
Re(
y
p f
y
)

x
f
x
+
y
f
y
_
_
and g
2
=
_
_
Re(u
x
u + v
y
u)
Re(u
x
v + v
y
v)
2(
y
u
x
v
x
u
y
v)
_
_
. (22)
Here g
1
is equal to g in (13) up to a constant but, due to the nonlinearity of g
2
, the discretization
is not so straightforward as for the Stokes problem. Details of the linearization techniques used
are therefore considered in the next subsection.
3.2. Linearization Techniques
3.2.1. Picard method Picard iteration, also known as successive substitution, is a simple
method to implement and typically converges from a large basin of attraction. It consists of
taking one of the vectors u from the previous iteration level whilst the other is taken at the
new level. This allows us to rewrite the vector g
2
in (22) as:
g
2
=
_
_
Re(u
k

x
u
k+1
+ v
k

y
u
k+1
)
Re(u
k

x
v
k+1
+ v
k

y
v
k+1
)
2(
y
u
k

x
v
k+1

x
u
k+1

y
v
k
)
_
_
, (23)
where k denotes the previous iteration level and k + 1 the new one. A similar linearization is
applied to the Neumann boundary condition for the pressure (9d):

n
p = n
_
u
k+1
(u
k
)u
k+1
+ f
_
. (24)
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
HIGH-ORDER FINITE DIFFERENCE SCHEMES FOR INCOMPRESSIBLE FLOWS 7
Figure 1. Sparsity patterns of the sub-matrices in the system (20)
Unfortunately, however, the convergence of the Picard method is generally linear and
typically rather slow. Since it has good convergence properties it is often used to provide
a suitable initial guess for the Newton method, which is introduced in the next section, and
which converges much more rapidly provided a suciently good initial iterate is used.
3.2.2. Newton method Newtons method is another well-known procedure for the solution
of nonlinear systems of algebraic equations. It generally achieves quadratic convergence when
the current iteration value u
k
is suciently close to the true solution u. The Newton iteration
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
8 H. FADEL ET AL.
process is now described.
Assuming that the system to be solved is of the form
N(u) = 0 , (25)
the rst step consist of evaluating the Jacobian J
k
of N at u
k
:
J
k
=
N(u
k
)
u
. (26)
The linearised system to be solved is then given by:
J
k
u
k
= N(u
k
) , (27)
and the updated solution is
u
k+1
= u
k
+ u
k
. (28)
For the nite dierence approach described above, the linearization is obtained using a
Taylor series expansions in two variables such that the series are truncated after the rst
derivative terms. This is illustrated as follows for an arbitrary function f(x
1
, x
2
):
f(x
k+1
1
, x
k+1
2
) f(x
k
1
, x
k
2
) +
x1
f(x
k
1
, x
k
2
)(x
k+1
1
x
k
1
) +
x2
f(x
k
1
, x
k
2
)(x
k+1
2
x
k
2
) (29)
f(x
k+1
1
, x
k
2
) + f(x
k
1
, x
k+1
2
) f(x
k
1
, x
k
2
) . (30)
Applying this approach to each of the nonlinear terms in the vector g
2
in (22) leads to:
g
2
= g
21
+ g
22
, (31)
where
g
21
=
_
_
Re(u
k

x
u
k+1
+ u
k+1

x
u
k
+ v
k

y
u
k+1
+ v
k+1

y
u
k
)
Re(u
k

x
v
k+1
+ u
k+1

x
v
k
+ v
k

y
v
k+1
+ v
k+1

y
v
k
)
2(
y
u
k

x
v
k+1
+
y
u
k+1

x
v
k

x
u
k+1

y
v
k

x
u
k

y
v
k+1
)
_
_
(32)
and
g
22
=
_
_
Re(u
k

x
u
k
+ v
k

y
u
k
)
Re(u
k

x
v
k
+ v
k

y
v
k
)
2(
x
u
k

y
v
k

y
u
k

x
v
k
)
_
_
. (33)
Finally, applying the same approach to the Neumann boundary condition for the pressure (9d)
leads to:

n
p = n
_
u (u
k
)u
k+1
(u
k+1
)u
k
+ (u
k
)u
k
+ f
_
. (34)
3.3. Solution strategy
Three dierent solutions strategies, based upon the above linearization schemes, have been
adopted in this work. For small Reynolds number the Newton method is applied, with an
initial guess typically obtained by the solution to a corresponding Stokes problem. If this fails
to converge then a hybrid method consisting of few Picard iterations as an initial guess for the
Newton method is then applied.
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
HIGH-ORDER FINITE DIFFERENCE SCHEMES FOR INCOMPRESSIBLE FLOWS 9
For high Reynolds numbers even this strategy may be very inecient and so an approach
based upon numerical continuation is adopted. Here, in order to obtain an initial guess that is
suciently close to the true solution for the Newton algorithm to converge, the problem (21)
is solved with:
g = g
1
+ g
2
(35)
for gradually increasing values of such that 0 1.
4. NUMERICAL RESULTS
In this section, we present a selection of numerical results obtained using the schemes developed
above for the solution of four typical test problems. The rst two problems are selected to
illustrate the numerical properties of the solver based upon known analytical solutions. It
should be noted that for even simpler test problems with low order polynomial solutions,
such as classical Poiseuille ow or the colliding ow suggested in [14], our solution schemes
demonstrated superconvergent properties that recovered the exact solution on all meshes.
The second pair of test cases are selected to represent more realistic uid mechanics
applications characterized, from a physical point of view, by simple geometry and complex
physics. The rst problem represents a conned ow with Dirichlet BCs for the velocity, namely
a leaky lid-driven cavity at Re=100, 400 and 1000. This case exhibits some interesting physical
features, specically singularities at the cavity top corners due to discontinuous BCs, and a
recirculation zone driven by the moving top wall. The second problem is an inow/outow
test case, that is, a backward-facing step at Re=800. This ow is physically characterized by
two recirculation zones, which requires appropriate outow BCs in order to be successfully
simulated.
In all four cases, the iterative solver used was the restarted GMRES [15]: with the restart
dimension set to 30 and using only a simple diagonal preconditioner. This is clearly far from an
optimal preconditioning strategy and its improvement will be the subject of further research,
along with the use of alternative solvers for non-symmetric matrices.
4.1. Stokes ow test case
As a rst test case for the Stokes ow, we chose a mathematical problem [16] with a known
exact solution in order to assess the performance and the accuracy of the scheme developed.
Let = [0, 1] [0, 1] be the ow domain and the exact solution given by:
u = 2x
2
(1 x)
2
sin(y) cos(y) (36a)
v = 2x(x 1)(2x 1) sin
2
(y) (36b)
p = sin(x) cos(y) . (36c)
Based upon this solution, the right-hand side of the system to be solved (10) is given by:
f
x
= 2(1 + 6x + 2(
2
3)x
2
4
2
x
3
+ 2
2
x
4
)sin(2y) + cos(x) cos(y) (37a)
f
y
= 4
2
x(1 3x + 2x
2
) cos(2y) 12(1 2x) sin(y)
2
sin(x) sin(y) (37b)
f = 2 sin(x) cos(y) . (37c)
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
10 H. FADEL ET AL.
n
x
n
y
u Number of iterations
2020 8.14e
10
31
4040 5.23e
10
96
8080 1.13e
9
310
160160 6.32e
10
1023
320320 1.07e
9
2001
Table I. Grid renement, continuity and the number of linear iterations for the Stokes test case (36)
n
x
n
y
E
u

Order u E
v

Orderv
2020 1.24 10
4
- 1.24 10
4
-
4040 7.58 10
6
3.98 7.84 10
6
3.98
8080 4.95 10
7
3.99 4.88 10
7
4.00
160160 3.11 10
8
3.99 3.07 10
8
3.99
320320 1.95 10
9
4.00 1.92 10
9
4.00
Table II. Accuracy order of velocity components for the Stokes test case (36)
n
x
n
y
E
p

Order p E

x
p

Order
x
p
_
_
E

y
p
_
_

Order
y
p
2020 6.27 10
3
- 3.76 10
2
- 3.07 10
2
-
4040 3.88 10
4
4.01 3.26 10
3
3.53 2.71 10
3
3.50
8080 2.41 10
5
4.01 2.83 10
4
3.52 3.37 10
4
3.51
160160 1.51 10
6
4.00 2.43 10
5
3.54 2.07 10
5
3.52
320320 9.39 10
8
4.01 2.09 10
6
3.54 1.81 10
6
3.52
Table III. Accuracy order of pressure eld and derivatives for the Stokes test case (36)
The boundary conditions for this test case are constrained to those given by the solution
(36). The values of u are prescribed on all boundaries while the hydrostatic pressure mode is
eliminated by xing the pressure value at a single grid node. The tolerance of GMRES iteration
is set to 10
14
(which is clearly excessively small however our goal here is not to assess the
eciency of the linear solver, rather the eciency of the discretization scheme) and in table I
we present the resulting maximum continuity check along with the number of linear iterations
required on each grid used. As is typical in the absence of problem-specic preconditioning the
iteration count approximately doubles each time the mesh spacing is halved.
In the tables II and III we present a grid renement and accuracy analysis of the scheme
developed. This clearly shows fourth order convergence in the maximum error for the velocity
components and the pressure. For the pressure derivatives the order of accuracy is 3.5, which is
a very good result knowing that the present Pade scheme is dispersive, and that no boundary
conditions are specied. The convergence order was obtained using the following formula :
Order() =
log(E
n

/ E
2n

)
log 2
. (38)
Additionally, the results obtained with the present scheme, denoted in the following by fds1,
are compared with the ones obtained using three other schemes, dened as follows. The choice
of these schemes allows comparisons to be made against more conventional nite dierence
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
HIGH-ORDER FINITE DIFFERENCE SCHEMES FOR INCOMPRESSIBLE FLOWS 11
and nite element discretizations.
Scheme fds2 : modied version of fds1 where the 9-points Poisson scheme used to
discretize the equation (12) is replaced by the classical 5-points Poisson scheme [12]
and second order Taylor approximations are used to approximate rst and second order
derivatives.
Scheme fds3 : modied version of fds1 where the pressure derivatives are not considered
as unknowns. These are then discretized in the momentum equation (right hand side
g in (13)) by 4th order Taylor approximations while the Pade schemes are no longer
necessary.
Scheme fes1 : classical mixed nite element solver for Stokes equations based on
triangular P2-P1 elements [14].
The simulations are performed under the same conditions described above and each nite
dierence scheme uses the same linear solver than fds1, namely GMRES with a simple diagonal
preconditioner, while the nite element scheme uses the MinRes linear solver [17] with a similar
preconditioner. In table V we show average convergence orders obtained by running the same
set of simulations performed previously for the four schemes. It should be noted that for the
last two schemes, the pressure derivative errors and accuracy orders were post-computed by
using 4
th
order Taylor approximations in fds3 and by considering the gradient at each node
as the average of each surrounding cell values in fes1. Our scheme, fds1, clearly shows superior
convergence rates for p and its derivatives. It is also important to look at the actual errors and
performance on a particular grid however. Table IV shows these results for a [40 40] mesh,
and it is clear that the errors in the pressure and its derivatives are already superior (or at
least as good in the nite element case) as the other schemes, which converge more slowly. For
completeness of the comparison we also include the the number of non-zero matrix elements
and the CPU times required by each scheme. These latter gures should be treated with great
care however since we have not attempted to optimize any of the implementations developed.
Scheme E
u

E
v

E
p

x
p

_
_
E

y
p
_
_

cpu T(s) A size

fds1 7.58 10
6
7.84 10
6
3.88 10
4
3.26 10
3
2.71 10
3
61 75703
fds2 1.76 10
3
1.29 10
3
6.13 10
2
5.41 10
1
4.98 10
1
52 43635
fds3 2.49 10
6
2.56 10
6
8.71 10
4
4.32 10
3
3.76 10
3
50 88986
fes1 7.50 10
7
3.62 10
7
2.28 10
4
1.61 10
2
1.47 10
2
190 142746
Table IV. Uniform Errors and other simulation characteristics for a [40 40] mesh
Scheme O u O v O p O
x
p O
y
p
fds1 3.99 3.99 4.01 3.53 3.51
fds2 1.97 1.98 1.99 1.35 1.34
fds3 3.76 3.77 3.55 3.10 3.03
fes1 3.87 3.73 2.63 1.34 1.25
Table V. Accuracy orders (average)
For the present test case, we discuss below the advantages and drawbacks of the compared
schemes.
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
12 H. FADEL ET AL.
fds2 : as expected, this scheme is 2
nd
order accurate for the velocity components and the
pressure, whilst the pressure derivatives accuracy is only 1.35 despite the use of the 4
th
order Pade scheme, which leads us to conclude that the global accuracy of our solution
strategy is predominately driven by the Laplacian discretization.
fds3 : this scheme shows better performances in term of the time required for the
simulation but it is slightly lower order than the scheme fds1. It should also be noticed
that although the matrix size is smaller that the one obtained using fds1, the number of
non-zero elements and therefore the matrix storage size is bigger. These inferior results
for the pressure approximations also allow us to see the improving eect of the Pade
scheme on the accuracy in these terms.
fes1 : this scheme shows poorer performance than the previous schemes, except for fds2,
in term of accuracy, computing time and matrix storage size. This reects the additional
overhead associated with an unstructured nite element solver (that we would be unlikely
to use on a simple geometry such as this).
4.2. Kovasznay Flow
Kovasznay ow [18] is an analytical solution of 2D steady-state Navier-Stokes equations, with
no forcing term, that is similar to the laminar ow over a periodic array of cylinders. Unlike
Poiseuille ow, this ow incorporates nonlinear eects and it is therefore a good test for the
full Navier-Stokes solution algorithm. Let = [0.5, 1.0] [0.5, 1.5] be the ow domain. The
analytical solution has the form:
u(x, y) = 1 e
x
cos(2y) (39a)
v(x, y) =

2
e
x
sin(2y) (39b)
p(x, y) =
1
2
(1 e
2x
) , (39c)
where the parameter is given in terms of the Reynolds number Re by:
=
Re
2

_
Re
2
4
+ 4
2
. (40)
As with the previous test case, the boundary conditions are given by the solution (39).
The values of u are prescribed on all the boundaries while the hydrostatic pressure mode is
eliminated by xing the pressure value at a single grid node. This problem was solved for a
value of Re equal to 40 and the tolerance of the GMRES iteration was again set to 10
14
.
For this test case the Newton linearization technique converged after 5 nonlinear iterations,
the stopping criteria being set to 10
12
based on the nonlinear residual, (without the need
for the Picard steps to nd an initial iterate) for all mesh sizes considered. This is due to the
relatively small value of Reynolds number.
Tables VI and VII show a grid renement and accuracy analysis of the scheme. As for
the Stokes test case, fourth order accuracy is conrmed for the velocity components and the
pressure, whilst an order of accuracy of 3.5 is observed for the pressure derivatives.
4.3. Lid-driven cavity ow
The lid-driven cavity ow is a classical and simple test case consisting of a conned uid ow
in a cavity driven by the moving upper wall, while the other walls are static. The problem to
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
HIGH-ORDER FINITE DIFFERENCE SCHEMES FOR INCOMPRESSIBLE FLOWS 13
n
x
n
y
E
u

Order u E
v

Order v
3020 2.6 10
3
- 2.51 10
3
-
6040 1.65 10
4
3.98 1.58 10
4
3.99
12080 1.04 10
5
3.99 9.87 10
5
4.00
240160 6.48 10
7
4.00 6.13 10
7
4.01
480320 4.05 10
8
4.00 3.83 10
8
4.00
Table VI. Accuracy orders of velocity components for Kovasznay ow test case (39)
n
x
n
y
E
p

Order p E

x
p

Order
x
p
_
_
E

y
p
_
_

Order
y
p
3020 8.08 10
3
- 1.26 10
2
- 1.31 10
2
-
6040 5.16 10
4
3.97 1.12 10
3
3.49 1.15 10
3
3.51
12080 3.24 10
5
3.99 9.98 10
5
3.49 1.01 10
4
3.52
240160 2.06 10
6
3.99 8.82 10
6
3.50 8.74 10
6
3.52
480320 1.26 10
7
4.02 7.80 10
7
3.50 7.56 10
7
3.53
Table VII. Accuracy order of pressure eld and derivatives for Kovasznay ow test case (39)
be solved, in a unit square, is governed by the Navier-Stokes system of equations (8) together
with the following Dirichlet BCs for the velocity components:
u(0, y) = 0 , v(0, y) = 0 (41a)
u(1, y) = 0 , v(1, y) = 0 (41b)
u(x, 0) = 0 , v(x, 0) = 0 (41c)
u(x, 1) = 1 , v(x, 1) = 0 . (41d)
Our simulations were performed on a [80 80] uniform mesh (x = y = 0.0125) for
Reynolds numbers of 100, 400 and 1000 such that the characteristic velocity and length are
chosen to be equal to unity. Both GMRES and nonlinear residual tolerances were set to 10
9
.
Using coarser meshes good asymptotic convergence was not achieved due to the poor resolution
of the near wall ow, and leading to results dierent from those of Ghia [19]. We also noticed
that without the use of the Taylor approximations (58b) and (58d) for the pressure boundary
condition discretization, and only for this test case, the scheme failed to converge because of
the singularities at the top corners of the domain.
Concerning the linearization technique, we have used the Newton method with 2 Picard
iterations to obtain an initial guess. For Re = 100, no continuation procedure was necessary
and only 5 Newton iterations assured convergence. For the higher Reynolds number test cases
use of the continuation procedure was necessary: such that was incremented in steps of 0.25
for Re = 400 and 0.1 for Re=1000. The Newton iterations converged quickly in all cases, with
only 4 to 5 iterations needed at each during the continuation.
In the gure 2 we show the u-velocity proles along a vertical line passing through the
geometric centre of the cavity for the three Reynolds numbers considered, whilst gure 3
shows the v-velocity proles along a horizontal line, also passing through the geometric centre
of the cavity. For both gures, the solid lines represent the numerical results obtained using our
proposed scheme on a uniform grid, whilst symbols indicate the solutions reported in [19]. In
order to visualize the streamlines, the stream function and the vorticity are obtained by solving
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
14 H. FADEL ET AL.
0
0.2
0.4
0.6
0.8
1
-0.4 -0.2 0 0.2 0.4 0.6 0.8 1
y
u-velocity
Present
Re=100 in [19]
Re=400 in [19]
Re=1000 in [19]
Figure 2. u-velocity at vertical line x=0.5 for the lid-driven cavity with Re=100, 400 and 1000
the linear system using the 9-point Poisson scheme (14). The calculations are performed as
a post-processing step at the end of the ow simulations and gure 4 shows the computed
streamlines for a Reynolds number of 1000.
The mesh dependency of the solution for a Reynolds number of 100 was examined for ve
mesh congurations, beginning from [8080] and increasing by 20 in both directions. In table
(VIII) we show the velocity components values at centre node (0.5, 0.5) compared with values
obtained in [19] that indicates clearly that the ner mesh leed to mesh-independent solutions.
Quantity [80 80] [100 100] [120 120] [140 140] [160 160] Ghia [19]
u(0.5, 0.5) 0.208033 0.210836 0.210892 0.210894 0.210895 0.2109
v(0.5, 0.5) 0.053973 0.054487 0.054529 0.054537 0.054539 0.05454
Table VIII. Mesh dependency check at (0.5, 0.5) for the lid-driven cavity at Re = 100
4.4. Backward facing step
The backward-facing step is a another standard benchmark problem. For this particular
implementation the domain is a rectangular channel such that the re-entrant backward-facing
step is simulated by imposing a parabolic velocity prole along the upper half of the inlet
boundary and zero inow velocity along the lower half of the inow boundary. An open
boundary condition is applied at the outow and the no-slip velocity condition is imposed
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
HIGH-ORDER FINITE DIFFERENCE SCHEMES FOR INCOMPRESSIBLE FLOWS 15
-0.6
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0 0.2 0.4 0.6 0.8 1
v
-
v
e
l
o
c
i
t
y
x
Present
Re=100 in [19]
Re=400 in [19]
Re=1000 in [19]
Figure 3. v-velocity at horizontal line y=0.5 for the lid-driven cavity with Re=100, 400 and 1000
Figure 4. Streamlines for the lid-driven cavity ow at Re=1000
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
16 H. FADEL ET AL.
on the remaining (top and bottom) boundaries. Specically and as shown in gure 5, the
domain = [0, 30] [0.5, 0.5], the governing equations the same as for the previous test
case, and the imposed boundary conditions are as follows:
u(0, 0 y 0.5) = 24y(0.5 y) , v(0, 0 y 0.5) = 0 (42a)
u(0, 0.5 y < 0) = 0 , v(0, 0.5 y < 0) = 0 (42b)
u(x, 0.5) = 0 , v(x, 0.5) = 0 (42c)
u(x, 0.5) = 0 , v(x, 0.5) = 0 (42d)
Re
1

x
u(30, y) p = 0 , Re
1

x
v(30, y) = 0 . (42e)
Figure 5. Backward facing step geometry and boundary conditions [21]
This problem was simulated on a [20600], equally spaced, mesh (x = y = 0.05) and for
a Reynolds number of 800 such that the characteristic velocity is the average inow velocity
and equal to unity and characteristic length in the channel height and also equal to unity.
The GMRES and nonlinear residual tolerances were set to 10
9
. The resulting ow has been
demonstrated to be steady and stable [20], and our computational results are compared with
those provided by Gartling [21]. As with the previous test case, simulations on a coarser mesh
than that used for the results presented here fail to capture all of the features of the resulting
uid ow. Also as for the previous test case, we have used the continuation approach to obtain
a converged solution at this Reynolds number: this time based upon increments of equal
to 0.125. Each of these nonlinear problems was solved with Newtons method using 2 Picard
iterations to obtain the initial guess.
We show in gure 6 the u-velocity proles along the two vertical lines x = 7 and x = 15 and
in gure 7 the v-velocity proles along the same vertical lines. In both gures, the solid lines
represent the numerical results obtained using our proposed scheme, while symbols indicate
the solutions reported in [21].
At a Reynolds number of 800 the ow studied here is characterized by two separation
regions. The rst of these starts at the step corner and continues downstream for a distance
of approximately 12 step heights, whilst the second is on the upper wall and occupies a region
from approximately 10 to 20 step heights downstream. To illustrate these, gure 8 shows plots
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
HIGH-ORDER FINITE DIFFERENCE SCHEMES FOR INCOMPRESSIBLE FLOWS 17
-0.4
-0.2
0
0.2
0.4
-0.2 0 0.2 0.4 0.6 0.8 1 1.2
y
u-velocity
Present
x=7 in [21]
x=15 in [21]
Figure 6. u-velocity for the backward-facing step at vertical lines x=7 and x=15
of the pressure and stream function contours for the rst ten step heights downstream of the
inow boundary: these plots show a good match with the equivalent results presented in [21].
Additionally, table IX shows a comparison of separation and reattachment point locations in
the recirculating regions obtained using the present scheme versus those obtained in [21, 20].
Quantity Present study Gartling[21] Gresho and al.[20]
Lower eddy reattachment x location 6.10 6.10 6.10
Upper eddy separation x location 4.85 4.86 4.85
Upper eddy reattachment x location 10.50 10.48 10.49
Table IX. Comparison of present results with Gartling[21] and Gresho and al.[20]
5. EXTENSIONS
5.1. Higher order Schemes :
Numerous extensions to the work presented here are possible. For example, the framework
that we have described naturally permits the use of higher order stencils and higher order
Pade schemes. This may be conrmed by extending the accuracy to 6th order, and in tables X
and XI we present the results obtained by such an extension for the Stokes test case of section
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
18 H. FADEL ET AL.
-0.4
-0.2
0
0.2
0.4
-0.02 -0.015 -0.01 -0.005 0 0.005
y
u-velocity
Present
x=7 in [21]
x=15 in [21]
Figure 7. v-velocity for the backward-facing step at vertical lines x=7 and x=15
Figure 8. Pressure and stream function contours for the ow over a backward-facing step
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
HIGH-ORDER FINITE DIFFERENCE SCHEMES FOR INCOMPRESSIBLE FLOWS 19
4.1. Our numerical strategy remains the same but we now use the 6
th
order Poisson equation
scheme introduced by Rosser [12] and the same order Pade schemes and Taylor approximations.
The grid renement sequences shown in tables X and XI demonstrate the expected sixth order
accuracy for velocity and pressure, and an order of accuracy of 5.5 for the pressure gradient.
Extension to even higher orders is of course possible in a similar manner.
n
x
n
y
E
u

Order(u) E
v

Order(v)
2020 4.54 10
5
- 4.57 10
5
-
4040 7.25 10
7
5.97 7.20 10
7
5.99
8080 1.15 10
8
5.99 1.12 10
8
6.00
160160 1.79 10
10
6.00 1.73 10
10
6.02
320320 2.80 10
12
6.00 2.65 10
12
6.03
Table X. Accuracy order of velocity components for Stokes test case (36) using 6
th
order scheme
n
x
n
y
E
p

Order(p) E

x
p

Order(
x
p)
_
_
E

y
p
_
_

Order(
y
p)
2020 8.24 10
4
- 9.72 10
3
- 9.51 10
3
-
4040 1.31 10
5
5.98 2.27 10
4
5.42 2.25 10
4
5.40
8080 2.05 10
7
5.99 5.27 10
6
5.43 5.30 10
6
5.41
160160 3.21 10
9
6.00 1.22 10
7
5.43 1.23 10
7
5.43
320320 4.98 10
11
6.01 2.83 10
9
5.43 2.87 10
9
5.42
Table XI. Accuracy order of pressure eld and derivatives for Stokes test case (36) using 6
th
order
scheme
5.2. The Unsteady Case :
We consider in the following the unsteady Stokes equation under appropriate boundary and
initial conditions:

t
u
2
u +p = f (43a)

2
p = f , (43b)
The spatial discretization strategy remains the same as described in section 2, leading to a
semi-discretized problem that can be expressed in the following matrix form:
M
dx
dt
= Ax + F (44)
The matrix A, the vector F and the unknown vector x are dened as in (20), whilst the
matrix M given by:
M =
_

_
M
1
0 0 0 0
0 M
1
0 0 0
M
2
M
3
0 0 0
0 0 0 0 0
0 0 0 0 0
_

_
(45)
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
20 H. FADEL ET AL.
t n
x
n
y
E
u

Order(u) E
v

Order(v)
1/10 1010 9.33 10
7
- 5.08 10
6
-
1/20 2020 6.01 10
8
3.96 3.15 10
7
4.01
1/40 4040 3.83 10
9
3.97 1.94 10
8
4.02
1/80 8080 2.38 10
10
4.01 1.21 10
9
4.01
1/160 160160 1.48 10
11
4.00 7.49 10
11
4.01
Table XII. Accuracy order of velocity components for unsteady Stokes test case (46)
t n
x
n
y
E
p

Order p E

x
p

Order
x
p
_
_
E

y
p
_
_

Order
y
p
1/10 1010 1.26 10
4
- 3.58 10
4
- 3.01 10
4
-
1/20 2020 8.08 10
6
3.96 3.21 10
5
3.48 3.19 10
5
3.24
1/40 4040 5.12 10
7
3.98 2.93 10
6
3.45 3.44 10
6
3.21
1/80 8080 3.26 10
8
3.97 2.72 10
7
3.44 3.69 10
7
3.22
1/160 160160 2.08 10
9
3.98 2.51 10
8
3.44 3.86 10
8
3.26
Table XIII. Accuracy order of pressure eld and derivatives for unsteady Stokes test case (46)
To perform the temporal dierencing, we use here the Backward Dierence Formula (BDF)
[22] familly of methods.
In order to validate our unsteady version of the scheme for the Stokes ow, lets consider the
following exact test case dened on = [0, 1] [0, 1] such that the solution is given by:
u = sin(x) sin(y + t) (46a)
v = cos(x) cos(y + t) (46b)
p = cos(x) sin(y + t) . (46c)
Based upon this solution, the right-hand side of the system to be solved (43) is given by:
f
x
= sin(x)(cos(y + t) + sin(y + t)) (47a)
f
y
= cos(x)(3 cos(y + z) sin(y + z)) (47b)
f = 2 cos(x) sin(y + z) . (47c)
The boundary conditions for this test case are dened by the solution (46) as well as the initial
conditions with t = 0. The values of u are given on all boundaries while the pressure value is
xed at a single grid node. The tolerance of GMRES iteration is set to 10
14
and the solutions
were computed up to time 10.
In this work the 4
th
order BDF scheme is used, to be compatible with 4
th
order spatial
discretizations, using exact solutions as given by (46) for the rst three steps. The numerical
results showing the order of accuracy, computed using formula (38), are displayed in tables
XII and XIII. As predicted, it demonstrate clearly that the scheme is fourth-order accurate in
time for all the variables, whilst the pressure derivatives have an order approaching 3.5.
6. CONCLUSIONS
In this paper we have presented a new class of high-order nite dierence schemes for
the accurate and stable solution of the incompressible Stokes and Navier-Stokes equations.
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
HIGH-ORDER FINITE DIFFERENCE SCHEMES FOR INCOMPRESSIBLE FLOWS 21
Our approach is based upon a velocity-pressure-pressure gradient formulation, where all
components are approximated on a single mesh using high order nite dierence stencils for
the velocity and pressure, and corresponding Pade schemes for the pressure gradient. The
solution strategy for the new class of schemes presented in this paper can be summarised by
three major points:
1. The momentum and pressure equations are written in a Poisson equation form, such
that schemes developed for this equation could be used.
2. The pressure derivatives are considered as unknown, which leads us to use Pade schemes
to discretize them.
3. Other derivatives involved in the equations are discretized using Taylor approximations.
A detailed derivation of the 4th order spatial discretization has been presented and numerical
tests, that conrm the 4th order accuracy for the velocity and pressure components, have been
described. Furthermore, these numerical tests show that the order of accuracy obtained for
the pressure derivatives is 3.5.
There are many other potential extensions which we have yet to implement. In principle,
at least, the extension from two dimensions to three dimensions should be straightforward.
Similarly, the extension to time-dependent problems using high order implicit time-stepping
has been shown to be feasible without substantial further development: so long as the time-
step size is not too large the Newton solver should converge using an initial iterate based upon
a simple extrapolation of the solution at the previous time levels. This is the subject of our
current research, where we seek to apply time-dependent Stokes and Navier-Stokes equations to
uid-structure interaction problems. The results of this work will be the subject of subsequent
papers treating both small and large structure deformations.
Of course, further research is also needed to improve the eciency of the schemes practical
implementation. Specically, the current linear solver is far from optimal as it is based upon
a very simple diagonal preconditioner, which fails to control the growth in the number of
iterations required as the mesh size is rened. Optimizing (or at least substantially improving)
this preconditioner is an important next step. Furthermore, we currently solve the linear
systems to the same accuracy at each Newton iteration: further eciency improvements should
be made by reducing signicantly the level of accuracy required at the rst and second Newton
iterations. Perhaps more challenging, but also of great potential value, would be the extension
of these higher order schemes to non-uniform grids. This would be desirable for ows, often
occurring at higher Reynolds number, which exhibit boundary layers, or similar phenomena.
REFERENCES
1. Gresho PM and Sani RL. Incompressible Flow and Finite Element Method, Volume 2, Isothermal Laminar
Flow. Wiley Chichester, England, 1998.
2. Gresho PM and Sani RL. On pressure boundary-condition for the incompressible Navier-Stokes equations,
International Journal for Numerical Methods in Fluids 1987; 7: 11111145. Chichester, 1998.
3. Roache PJ Comutational Fluid Dynaimcs. Hermosa Publishers, Albuquerque, New Mexico, USA, 1992.
4. Karniadakis GE, Israeli M and Orszag SA. High-order splitting methods for the incompressible NavierStokes
equations, Journal of Computational Physics 1991; 97: 414443.
5. Petersson NA. Stability of Pressure Boundary Conditions for Stokes and NavierStokes Equations, Journal
of Computational Physics 2001; 172: 4070.
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
22 H. FADEL ET AL.
6. Henshaw WD and Petersson NA. A split-step scheme for the incompressible Navier-Stokes equations, in
Numerical simulations of incompressible ows (Half Moon Bay, CA, 2001), 108125, World Sci. Publishing,
River Edge, NJ, 2003.
7. Hirsh RS. Higher Order Accurate Dierence Solutions of Fluid Mechanics Problems by a Compact
Dierencing Technique, Journal of Computational Physics 1975; 19 : 90109.
8. Collatz L. The Numerical Treatement of Dierential Equations,Springer-Verlag, New York, 2005.
9. Collatz L. Hermitian Methods for Initial Value Problems in Partial Dierential Equations, Topics in
Numerical Analysis, Proceedings of the Royal Irish Academy Conference on Numerical Analysis 1972 :
4161.
10. Lele SK Compact Finite Dierence Schemes with Spectral-like Resolution, Journal of Computational
Physics 1992; 103: 1642.
11. Shih TM, Tan CH and Hwang BC. Eect of Grid Staggering on Numerical Schemes, International Journal
for Numerical Methods in Fluids 1989; 9: 193212.
12. Rosser JB. Nine-point dierence solution for Poissons equation, Computers and Mathematics with
Applications 1975; 1: 351360.
13. Strikwerda JC. Finite Dierence Schemes and Partial Dierential Equations, Wadsworth & Brooks/Cole,
Pacic Grove, CA, 1989.
14. Elman HC, Silvester DJ and Wathen AJ. Finite Elements and Fast Iterative Solvers with Applications in
Incompressible Fluid Dynamics, Oxford University Press, Oxford, 2005.
15. Saad Y and Schultz MH. GMRES: a generalized minimal residual method for solving nonsymmetric linear
systems, SIAM Journal on Scientic and Statistical Computing 1986; 7: 856869.
16. Bochev PB, Gunzburger MD and Lehoucq RB. On stabilized nite element methods for the Stokes problem
in the small time step limit. International Journal for Numerical Methods in Fluids 2007; 53: 573597.
17. Paige CC and Saunders MA. Solution of sparse indenite systems of linear equations. SIAM Journal on
Numerical Analysis 1975; 12:617629.
18. Kovasznay LIG. Laminar ow behind a two-dimensional grid. Proc. Cambridge Philos. Soc. 1948; 44:58-62.
19. Ghia U, Ghia KN and Shin CT. High-Re solutions for incompressible ow using the Navier-Stokes equations
and a multigrid method Journal of Computational Physics 1982; 48: 387411.
20. Gresho PM, Gartling DK, Torczynski JR, Clie KA, Winters KH, Garratt TJ, Spence A and Goodrich
JW. Is the steady viscous incompressible two-dimensional ow over a backward-facing step at Re = 800
stable? International Journal for Numerical Methods in Fluids 1993; 17:501541.
21. Gartling DK. A test problem for outow boundary conditions - Flow over a backward facing step.
International Journal for Numerical Methods in Fluids 1990; 11:953-967.
22. Brenan KE, Campbel SL, and Petzold LR. The Numerical Solution of Initial Value Problems in Dierential
Algebraic Equations, Elsevier Science Publishing Co., New York, 1989.
7. APPENDIX A
This appendix provides a fully-expanded description of the discretised equations that we solve
in the case of the Stokes problem. In the domain , we use the 9-point Poisson scheme (14)
for both the pressure and the velocity components, and the Pade scheme (18) for the pressure
derivatives. These lead to the following dicretisations:
For the x-velocity u:
a(u
i1,j
+ u
i+1,j
) + b(u
i,j1
+ u
i,j+1
) + u
i1,j1
+ u
i+1,j1
+ u
i1,j+1
u
i+1,j+1
20u
i,j
c(8
x
p
i,j
+
x
p
i1,j
+
x
p
i+1,j
+
x
p
i,j1
+
x
p
i,j+1
) (48)
= c(8f
xi,j
+ f
xi1,j
+ f
xi+1,j
+ f
xi,j1
+ f
xi,j+1
) .
For the y-velocity v:
a(v
i1,j
+ v
i+1,j
) + b(v
i,j1
+ v
i,j+1
) + v
i1,j1
+ v
i+1,j1
+ v
i1,j+1
v
i+1,j+1
20v
i,j
c(8
y
p
i,j
+
y
p
i1,j
+
y
p
i+1,j
+
y
p
i,j1
+
y
p
i,j+1
) (49)
= c(8f
yi,j
+ f
yi1,j
+ f
yi+1,j
+ f
yi,j1
+ f
yi,j+1
) .
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
HIGH-ORDER FINITE DIFFERENCE SCHEMES FOR INCOMPRESSIBLE FLOWS 23
For the pressure p:
a(p
i1,j
+ p
i+1,j
) + b(p
i,j1
+ p
i,j+1
) + p
i1,j1
+ p
i+1,j1
+ p
i1,j+1
p
i+1,j+1
20p
i,j
= c(8
x
f
xi,j
+
x
f
xi1,j
+
x
f
xi+1,j
+
x
f
xi,j1
+
x
f
xi,j+1
(50)
+8
y
f
yi,j
+
y
f
yi1,j
+
y
f
yi+1,j
+
y
f
yi,j1
+
y
f
yi,j+1
) .
For the pressure partial derivatives
x
p:

x
p
i1,j
+ 4
x
p
i,j
+
x
p
i+1,j
+
3
x
(p
i1,j
p
i+1,j
) = 0 . (51)
For the pressure partial derivatives
y
p:

y
p
i,j1
+ 4
y
p
i,j
+
y
p
i,j+1
+
3
y
(p
i,j1
p
i,j+1
) = 0 . (52)
On the boundary
N
, the pressure boundary condition (11d) and its partial derivatives
needs to be discretised.
For the pressure p:
at the corners
n
x
(
x
p
i,j
+
2
xy
v
i,j

2
yy
u
i,j
) + n
y
(
y
p
i,j
+
2
xy
u
i,j

2
xx
v
i,j
) = n
x
f
xi,j
+ n
y
f
yi,j
(53)
with
n
x
= (n e
x
)
x
_
x
2
+ y
2
and n
y
= (n e
y
)
y
_
x
2
+ y
2
;
where n
x
and n
y
are the components of the outward normal vector n while e
x
and
e
y
are versors of the two-dimensional Cartesian coordinate system.
along horizontal boundaries

y
p
i,j
+
2
xy
u
i,j

2
xx
v
i,j
= f
yi,j
; (54)
along vertical boundaries

x
p
i,j
+
2
xy
v
i,j

2
yy
u
i,j
= f
xi,j
. (55)
For the pressure partial derivatives
x
p:

x
p
i,j
+ 3
x
p
i+1,j
+
1
6x
(17p
i,j
9p
i+1,j
9p
i+2,j
+ p
i+3,j
) = 0 . (56a)
For the pressure partial derivatives
y
p:

y
p
i,j
+ 3
y
p
i,j+1
+
1
6y
(17p
i,j
9p
i,j+1
9p
i,j+2
+ p
i,j+3
) = 0 . (57a)
In the above equations we use the appropriate 4
th
order Taylor approximation in (58) for rst
order derivatives
x
p
i,j
and/or
y
p
i,j
, a combination of the appropriate approximations in (58)
for mixed second order derivatives
2
xy
u
i,j
and/or
2
xy
v
i,j
, and the appropriate approximation
in (59) for second order derivatives
2
xx
v
i,j
and/or
2
yy
u
i,j
:
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls
24 H. FADEL ET AL.
i = 0, f

i
=
1
12h
(25f
i
+ 48f
i+1
36f
i+2
+ 16f
i+3
3f
i+4
) (58a)
i = 1, f

i
=
1
12h
(3f
i1
10f
i
+ 18f
i+1
6f
i+2
+ f
i+3
) (58b)
1 < i < n 2, f

i
=
1
12h
(f
i2
8f
i1
+ 8f
i+1
f
i+2
) (58c)
i = n 2, f

i
=
1
12h
(3f
i+1
+ 10f
i
18f
i1
+ 6f
i2
f
i3
) (58d)
i = n 1, f

i
=
1
12h
(25f
i
48f
i1
+ 36f
i2
16f
i3
+ 3f
i4
) ; (58e)
i = 0, f

i
=
1
12h
2
(45u
i
154u
i+1
+ 214u
i+2
156u
i+3
+ 61u
i+4
10u
i+5
) (59a)
i = n 1, f

i
=
1
12h
2
(45u
i
154u
i1
+ 214u
i2
156u
i3
+ 61u
i4
10u
i5
) .(59b)
Copyright c 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 00:124
Prepared using dauth.cls