On the accuracy of nite dierence solutions for nonlinear

water waves
Harry B. Bingham

(hbb@mek.dtu.dk)
Mechanical Engineering, Technical University of Denmark, DK-2800 Lyngby,
Denmark
Haiwen Zhang (hazh@cowi.dk)
Marine & Foundation Eng., COWI A/S, DK-2800, Lyngby, Denmark
Abstract. This paper considers the relative accuracy and eciency of low- and
high-order nite dierence discretisations of the exact potential ow problem for
nonlinear water waves. The method developed is an extension of that employed
by [1] to allow arbitrary order nite dierence schemes and a variable grid spacing.
Time-integration is performed using a fourth-order Runge-Kutta scheme. The linear
accuracy, stability and convergence properties of the method are analysed and high-
order schemes with a stretched vertical grid are found to be advantageous relative
to second-order schemes on an even grid. Comparison with highly accurate periodic
solutions shows that these conclusions carry over to nonlinear problems and that
the advantages of high-order schemes improve with both increasing nonlinearity and
increasing accuracy tolerance. The combination of non-uniform grid spacing in the
vertical and fourth-order schemes is suggested as optimal for engineering purposes.
Keywords: Nonlinear waves, nite dierence methods, convergence, accuracy, sta-
bility.
1. Introduction
Marine, coastal and ocean engineers require a means for predicting the
propagation of nonlinear water waves and their subsequent interaction
with xed or oating structures. The physics of most such problems
is well described by the incompressible Navier-Stokes equations, but
direct numerical solution of these equations is limited to very small
Reynolds number ows (Re = O(1000).) The introduction of Reynolds
averaging to obtain the Reynolds Averaged Navier Stokes Equations
(RANSE) and a turbulence closure model allows solutions to be ob-
tained at realistic Reynolds numbers. Although progress has been rapid
over the past several decades, the computational eort required by
RANSE solvers still imposes severe limitations on domain size and
resolution. The next level of approximation is to neglect viscosity and
assume an irrotational ow to obtain a potential ow problem governed

Support from the Danish Centre for Scientic Computing is acknowledged.

c 2006 Kluwer Academic Publishers. Printed in the Netherlands.
BinghamZhang.tex; 9/06/2006; 11:05; p.1
2 Bingham & Zhang
by the Laplace equation (see e.g. [2] for a detailed statement of the
problem).
A large number of methods exist for solving the exact Laplace prob-
lem for surface gravity waves or some approximation to it. The most
widespread method for treating wave-body interaction problems is the
Boundary Integral Equation Method (BIEM) which uses Greens theo-
rem to project the three-dimensional (3D) problem onto the boundary
surface of the uid volume. The most mature of these methods are
based on a perturbation expansion in wave steepness and the linear
free-surface Green function which reduces the computational domain
to the structure surface and a small nearby portion of the free surface.
Such methods routinely provide second-order solutions which are in
widespread use for the design of oshore structures (see e.g. [3, 4]).
Higher order in nonlinearity can in principle be obtained using the free-
space Rankine Green function and larger portions of the free surface
(e.g. [5, 6, 7, 8]).
Numerous other methods have been devised for projecting the prob-
lem onto some portion of the uid boundary and thereby reducing
the computational dimension of the problem by one. Some popular
examples are pseudo-spectral methods e.g. [9, 10]; and Boussinesq-type
methods which assume a polynomial expansion in the vertical e.g. [11,
12]. Such methods are relatively ecient, but less exible in treating
wave-body interactions or domains with complicated boundaries.
An obvious alternative to surface projection methods is a direct
numerical solution of the 3D Laplace problem, which is perhaps less
common, but on the increase. The nite volume methodology which
is widely used by RANSE solvers is readily applied to the simpler
Laplace problem, either on its own or as an outer domain surrounding
an interior RANSE domain (see e.g. [13]). The nite element method
has also been applied to the problem e.g. by [14], and a nite dierence
based solution was developed by [1].
This paper is motivated by two observations related to the solu-
tion of the exact Laplace problem for nonlinear wave-body interaction.
First, projection methods exchange degrees of freedom (discretising the
internal uid volume) for complexity (expansions, or Green function
interactions). This exchange is clearly advantageous for perturbation
solutions, but the advantage is less obvious for the fully nonlinear prob-
lem. In this case so much of the boundary requires discretisation that it
is possible for 3D methods to be competitive due to their relative sim-
plicity and inherent ease of obtaining an optimal scaling of the solution
eort. The second observation is that most, if not all, existing direct
Laplace solvers are second-order accurate (at best), while signicant
evidence exists for the advantage of using higher-order schemes.
BinghamZhang.tex; 9/06/2006; 11:05; p.2
Accuracy of nite dierence wave solutions 3
Thirty years ago, Kreiss & Oliger [15] investigated nite dierence
solutions to the linear wave equation and showed that fourth-order
was optimal in some sense. Experience with nite dierence solutions
to Boussinesq-type equations [16] also shows a signicant advantage
to using fourth-order schemes. Here we extend the method of [1] as
follows. Arbitrary order nite dierence schemes are used to discre-
tise the continuous derivatives; and while we retain a structured grid,
we allow for a non-uniform grid spacing to allow clustering of grid
points where desired. A fourth-order Runge-Kutta scheme is used for
the time integration. The implementation is only in two-dimensions
(2D) at this point, and here we nd that the Generalised Minimal
RESidual (GMRES) Krylov subspace iterative method preconditioned
by the linearised, second-order version of the matrix leads to optimal
scaling of the solution eort (i.e. O(N) where N is the total number
of grid points). Reduction of the residual by seven orders of magnitude
is achieved in 10 iterations nearly independently of problem size and
the order of the nite dierence schemes used. In 3D we expect that
some form of multigrid will be optimal and this is the topic of ongoing
research.
After a description of the theory and implementation of the method
in 2 and 3, the linear convergence, accuracy and stability properties
are established by standard Fourier analysis in 4. In 5 the nonlinear
accuracy of the method is quantied by comparison with periodic solu-
tions based on stream function theory [17]. From this we conclude that
high-order schemes and a stretched vertical grid are more ecient than
second-order schemes to achieve a target accuracy of solution. This
advantage improves with increased nonlinearity and with decreased
error tolerance. Fourth-order schemes are suggested as being an optimal
balance of accuracy and complexity for engineering purposes and some
sample calculations for wave-bottom interaction are presented in 6.
Conclusions are drawn in 7.
2. Formulation
Consider the irrotational ow of an incompressible inviscid uid. A
Cartesian coordinate system is adopted, with origin on the still water
plane and the z-axis pointing vertically upwards, x = [x, y] is a hor-
izontal vector and t is time. The uid domain is bounded by the sea
bottom at z = h(x) and the free-surface at z = (x, t). Both and
h are assumed to be single valued functions of x. Following [18], we
express the kinematic and dynamic free surface boundary conditions in
terms of the velocity potential and the vertical component of velocity
BinghamZhang.tex; 9/06/2006; 11:05; p.3
4 Bingham & Zhang
evaluated directly on the free-surface:

= (x, , t), and w =

z

z=

t
=

+ w(1 + ) (1a)

t
= g
1
2

+
1
2
w
2
(1 + ). (1b)
Here = [/x, /y] is the horizontal gradient operator, g the grav-
itational acceleration and partial dierentiation is indicated when the
independent variables appear as subscripts. These relations can be
viewed as evolution equations for and

to be integrated forward
in time from initial conditions. The horizontal gradients appearing on
the right hand sides can be immediately computed, but obtaining the
vertical component of velocity w, requires a means of satisfying the
Laplace equation throughout the depth of the uid along with the
kinematic bottom boundary condition:

2
+
zz
= 0, h < z < (2a)

z
+h = 0, z = h. (2b)
For a direct solution of this Laplace problem, it is convenient to
make a change of variables in the vertical coordinate (the widely used
transformation) dened by
(x, z, t) =
z + h(x)
(x, t) + h(x)
=
z + h(x)
d(x, t)
, (3)
where the total thickness of the uid layer d = + h has been intro-
duced. This transformation converts the Laplace problem to

2
+
2

+ 2

+
_
+
2
z
_

= 0, 0 < < 1(4a)

(
z
+h )

+h = 0, = 0, (4b)
where the derivatives of can be expressed as
= (1 )
h
d

d
(5)

2
=
1
d
_

2
h
h h
d
_

d
_

2


d
_
(6)

1 2
d
2
h

d
(h +)

z
=
1
d
(7)
which illustrates the specic derivatives of h and required for their
evaluation. After solving (4) for the potential (x, ), the vertical com-
ponent of uid velocity on the free surface is given by
w =
1
d

|
=1
(8)
BinghamZhang.tex; 9/06/2006; 11:05; p.4
Accuracy of nite dierence wave solutions 5
which allows (1) to be stepped forward in time, closing the problem.
If the internal kinematics of the ow are desired they can also be
computed from (x, ), for example the uid velocities are given by
u(x, z) =
x
(x, z) =
x
(x, ) +
x

(x, ) (9)
v(x, z) =
y
(x, z) =
y
(x, ) +
y

(x, ), (10)
w(x, z) =
z
(x, t) =
1
d

(x, ). (11)
3. Numerical solution
For the time-integration of (1) we employ the classical explicit four-
stage, fourth-order Runge-Kutta scheme (see e.g. [19]). For the spatial
discretisation, a grid of N
x
points is dened along the x-axis at which
the time-stepping variables and

are to be evolved. Periodic or Neu-
mann (specied normal component of velocity) conditions are imposed
at the horizontal boundaries. For the solution of (4a) & (4b), N
z
points
are dened in the vertical under each horizontal grid point, arbitrarily
spaced between = 0 and = 1 (the same vertical spacing is however
chosen under each horizontal grid point.) Choosing r nearby points,
allows order (r 1) nite-dierence schemes for the rst and second
derivatives in x and to be developed in the standard way by means
of Taylor series expansion and the inversion of a small linear system at
each grid point.
3.1. Finite-difference discretisations
For the one-dimensional rst- and second-derivatives in x and , r =
+ + 1 points are used where indicates the number of points to
the right (or top) and the number of points to the left (or bottom)
of the point of interest. With periodic conditions all x-derivatives are
centrally discretised with = , while for Neumann conditions only
points within the computational domain are used leading to o-centred
schemes near the boundaries. For -derivatives o-centred schemes are
used near the bottom and free-surface. For the mixed x-derivative,
a full square stencil of r
2
-points is used and the schemes are thus
o-centred near non-periodic boundaries. In this way, all derivatives
are formally accurate to O(x
r1

) where x

is the maximum grid

spacing, and this is veried in 4.
The resultant discretisation of (4a) is then applied at all non-boundary
grid points, while the appropriate boundary conditions are imposed
BinghamZhang.tex; 9/06/2006; 11:05; p.5
6 Bingham & Zhang
at boundary points. This gives a rank N = N
x
N
z
linear system of
equations
A[] = [b] (12)
where A is the matrix of coecients, [] a vector of the unknown at
each grid point and [b] a vector holding zeros, except at those points
corresponding to inhomogeneous boundary conditions.
3.2. Iterative solution of the linear system
The matrix Ain (12) is sparse and contains a maximum of r
2
N nonzero
elements. A direct solution to this system using sparse matrix tech-
niques is eective at small values of N but is not competitive (even
in 2D) as N increases. This is especially true for higher order dis-
cretisations which give more non-zeros per row and more spreading in
the entries thus leading to extensive ll-in when factoring A, and this
quickly becomes prohibitive both in terms of memory use and computa-
tional eort. An iterative solution, properly preconditioned, turns out
to be much more eective. We employ here the GMRES (Generalised
Minimal RESidual) method of [20], preconditioned on the left using the
linearised, second-order accurate version of A. Specically, we set = 0
in (3) which makes independent of time and removes all functions of
from the derivatives in (5-7). Second-order nite-dierence schemes
(r = 3) are then used to discretise the system as described above to
produce the preconditioning matrix M. The preconditioning step then
consists of solving a linear system of the form M[q] = [s].
Since M is time-constant, it is built and LU factored only once after
which preconditioning requires only a back-substitution step which
is generally 10-100 times faster than a factorisation for this prob-
lem. As shown in 5, the iteration count to achieve a relative con-
vergence tolerance of 10
7
when using this preconditioner is generally
5-15 and independent of N and kh, although it increases slowly with
r (approximately a 50% increase for r = 3 to r = 5). Factorisation
and back-substitution is performed using the MA41 package from the
Harwell Subroutine Library. These routines are a potentially parallel
sparse multi-frontal variant of Gaussian elimination, which is particu-
larly eective on matrices with a nearly symmetric pattern. The method
chooses pivots from the diagonal using the approximate minimum de-
gree algorithm of [21]. When solving systems with a single right hand
side the routine also makes ecient use of level 2 Basic Linear Algebra
Subprograms (BLAS), which have been optimised using the Automat-
ically Tuned Linear Algebra Software (ATLAS, see e.g. [22])). For
BinghamZhang.tex; 9/06/2006; 11:05; p.6
Accuracy of nite dierence wave solutions 7
further details on this routine see [23] and references therein. In this
work we have only used the serial version of the code.
An example of the scaling of the solution eort for both the direct
and the iterative solutions is plotted in Figure 1, which shows the
calculation time to propagate a steady nonlinear wave of approximately
80% of the limiting steepness at kh = for 100 time-steps. The log of
0
1
2
3
4
5
2 2.5 3 3.5 4 4.5 5
l
o
g
1
0
(
C
P
U
)
log
10
(N)
Direct, r=3
Direct, r=5
Direct, r=7
Iterative
Figure 1. Scaling of the computational eort for direct and iterative solutions.
the calculation time on a 3.2 GHz Pentium processor with 1 gigabyte of
RAM memory is plotted vs. the log of N for the direct solution using
second-, fourth-, and sixth-order discretisations as well as the sixth-
order discretisation solved iteratively. Using the iterative solution, the
curves for dierent order discretisations are indistinguishable on the
scale of this plot. While the direct solutions can be seen to tend towards
a super-linear scaling of the eort, the iterative solution scales with
precisely N.
While this preconditioning strategy is expected to work equally well
in three-dimensions as far as iteration counts go, it is possible that the
increased complexity of the linearised matrix in that case will result in
a less than optimal scaling when compared to a multigrid solver. This
topic is under investigation and will be discussed in a follow-up paper.
4. Linear accuracy and stability of the discrete solution
The rst step in quantifying the performance of a given discretisa-
tion scheme is to consider the linearised version of the problem on a
horizontal bottom:

t
= w (13)
BinghamZhang.tex; 9/06/2006; 11:05; p.7
8 Bingham & Zhang

t
= g (14)

2
+
zz
= 0, h < z < 0 (15)

z
= 0, z = h (16)
where for the linear problem

= (x, 0, t) and w = w(x, 0, t). This
problem has the well known travelling wave solution
=
_
H
2
e
i(kxt)
_
(17)
=
_
i g

H
2
cosh [k(z + h)]
cosh (kh)
e
i(kxt)
_
(18)
where H is the wave height, L = 2/k the wave length, T = 2/
the wave period, and and k are related by the dispersion relation

2
= gk tanh (kh).
4.1. Convergence
The essential ingredient of the discrete solution scheme is the evaluation
of w from

. To verify the convergence of the method, we compare the
computed value to the exact result w = k tanh (kh)

, for increasingly
ne discretisations. To simplify the analysis, we consider a periodic
domain in x with a uniform grid spacing x. This makes all x-grid
points equivalent and allows this part of the problem to be expressed
in closed form as a function of the x-resolution (number of grid points
per wavelength.) For example, the second-order scheme
1
x
2
(
i1

2
i
+
i+1
), where
i
= (x = ix) becomes

2
x
2
N
2
x
_
2 + 2 cos
_
2
N
x
__
(19)
where we have taken L = N
x
x so that N
x
represents the number of
grid points per wavelength. This is standard von Neumann (Fourier)
analysis (e.g. [24]). Choosing a convenient wavelength, L = 1, xes
the wavenumber and choosing a value of kh then determines the depth
h. Since the discretisation in the vertical is on a (possibly) variable
grid and non-periodic, the rest of the problem is solved numerically.
Choosing N
z
points in the vertical thus produces a reduced version of
(12) which is of rank N
z
B[] = [b]. (20)
Numbering the grid points from the bottom to the free-surface, [b]
is a vector of zeros except for the last entry which holds the known
BinghamZhang.tex; 9/06/2006; 11:05; p.8
Accuracy of nite dierence wave solutions 9
magnitude of

. B is the matrix holding the coecients for /z in the
rst row, those for
2
/z
2
for rows 2 : N
z
1 (plus the discrete version
of
2
/x
2
on the diagonal) and nally a 1.0 on the last diagonal. After
solving this system for at each vertical grid point, the vertical rst-
derivative scheme is applied to get the approximation for w which can
be compared to the exact result.
Figure 2 shows the convergence of the calculations for several choices
of discretisation scheme at a relative water depth of kh = 4. The left
column plots results for a uniform grid using second-, fourth-, and sixth-
order schemes (r = 3, 5, & 7 points respectively) while the right hand
column shows the same order schemes applied to a stretched grid which
clusters points towards the free-surface. The stretched grid used here
is dened by
j
= [sin ((j 1))] where = /[2(N
z
1)]. Each
plot shows the relative error (| w w
e
|/ w
e
) versus N
z
on a log-log scale
with w
e
the exact result. The dierent lines represent dierent values
of N
x
as shown in the legend on the rst plot. An approximate value
for the asymptotic slope of the nest x-resolution (N
x
= 1000) line
also appears on each plot, and this has been computed using a least
squares t to the last four data points on that line. Two points are
notable from these plots: First, the estimated asymptotic convergence
rate of each method is close to the expected second-, fourth-, or sixth-
order rate. Secondly, and perhaps of more practical signicance, is the
dierence in magnitude of the errors for the dierent order schemes.
Of particular interest is the resolution required by each method to
achieve a given accuracy of solution, and especially notable are the large
gains in accuracy obtained by a fourth-order scheme on a stretched grid
relative to the widely used second-order scheme on an even grid. We
will return to this point in the next section.
The general trend of the calculations is the same at other values of
kh but, as might be expected, the errors get better with smaller kh
and worse with larger kh reecting the transition from a linear to an
exponential behaviour of (z).
4.2. Stability and accuracy
Fructus et al. [10] have shown how the linear part of the time stepping
problem can be integrated exactly by means of a Fourier transform,
allowing much larger time steps to be taken for the nonlinear contribu-
tion. We have chosen not to do so here because our goal is a model which
can treat problems with an abitrary horizontal boundary and including
xed and oating bodies of complicated form within the computational
domain. In this case it is not clear how to apply the required Fourier
transform in a robust way.
BinghamZhang.tex; 9/06/2006; 11:05; p.9
10 Bingham & Zhang
The linear accuracy and stability of the method can be quantied
by considering the semi-discrete form of the time-stepping equations
(13) & (14)

t
_

_
=
_
0 J
12
g 0
_ _

_
, (21)
applied to the wave solution of (17). Here and

are the Fourier
amplitudes of and

, while the factor J
12
is the discrete approximation
for k tanh (kh) discussed in the previous section (i.e. w = J
12

). This
is a method of lines approach (see e.g. [19, 25]) and is valid as long as
the above 2 2 matrix is amenable to an eigenvalue decomposition,
in which case stability is governed by the largest of these and the
stability region of a particular time-stepping scheme. For the discreti-
sations discussed here, the eigenvalues are purely imaginary, with the
largest occurring at the Nyquist mode. Figure 3 plots the magnitude
of the maximum eigenvalue (normalised by the Nyquist frequency) as
a function of relative depth for the Nyquist wave. These curves are for
second-order discretisations (r = 3), showing the eect of increasing
vertical resolution. A uniform vertical spacing is shown to the left and
a stretched grid to the right. Figure 4 plots the same quantities but
with a xed vertical resolution of N
z
= N
x
and using second-, fourth-,
and sixth-order discretisations. These plots can be used to nd the
stability limit for any desired time-stepping scheme. For a given dis-
cretisation, stability is insured by keeping the quantity
max
t within
the stability region of the time-stepping scheme of interest (e.g. for
fourth-order Runge-Kutta, |
max
| t 2

2). Notable from these plots

is the relatively minor role played by both the vertical discretisation
and the order of the scheme on the stability limit of the method.
Choosing now to focus on the fourth-order Runge-Kutta method
which will be used in practise, the stability analysis discussed above
can be extended to determine the overall accuracy of a given scheme.
As with the convergence analysis, we consider L = 1 so that N
x
= 1/x
represents the resolution of the wave (number of grid points per wave-
length). For a given relative water depth kh and a vertical resolution
N
z
, the eigenvalues of the system can then be computed. Applied to a
single equation, the fourth-order Runge-Kutta method can be expressed
by the complex amplication factor
(t) = 1 + t +
(t)
2
2
+
(t)
3
6
+
(t)
4
24
(22)
where denotes an eigenvalue of the system. Equation (22) expresses
the evolution of the solution in the eigenvector basis for a single time
step, including both a relative amplication given by |(t)|, and a
BinghamZhang.tex; 9/06/2006; 11:05; p.10
Accuracy of nite dierence wave solutions 11
phase shift of = arg((t)). Clearly for a stable solution we must
have |(t)| 1. As measures of overall accuracy we dene:
A |(t)|
Nt
, N
t
=
T
t
(23)
Q
c
num
c
=

C
r
, =
2
N
x
(24)
i.e. the relative amplitude and phase error over one complete wave
period. Here the hyperbolic Courant number is C
r
= ct/x with c the
exact phase speed and the expression for Q comes from the numerical
wave period T
num
= 2t/, which leads to a discrete phase speed of
c
num
= L/T
num
= x/(t ).
Figure 5 plots the amplitude and phase errors of the second-order
model with a vertical resolution of N
z
= 10, and at the two Courant
numbers 1 and 0.5. These plots are for a relative water depth of kh = 4,
results for an evenly spaced grid appear to the left and for a vari-
able grid to the right. These plots illustrate two general features of
the error: First, dispersion (phase) errors dominate, and are typically
several orders of magnitude larger than diusion (amplitude) errors,
and second; reducing the Courant number improves the diusion error
but not the dispersion error. Figure 6 plots the errors at Cr = 0.5
with a vertical resolution of N
z
= 10 showing the eect of increasing
the order of method from second to sixth. This plot makes it clear
that increasing the order of the spatial derivatives has no signicant
eect on the diusion errors but dramatically improves the dispersion
error. Figure 7 plots dispersion errors for the same conditions but with
N
z
= 15 and N
z
= 20.
Finally, Figure 8 collects the dispersion errors for second-, fourth-
and sixth-order discretisations on both a uniform and a stretched verti-
cal grid as a function of grid resolution. These plots are for C
r
= 1 and
kh = 4, and as found during the convergence calculations the trend is
very similar at other relative water depths but the errors increase with
larger kh and decrease with smaller kh for the same vertical resolution.
Table I collects the approximate resolution required by each method
to obtain relative phase errors of 10
3
, 10
4
, and 10
5
.
5. Nonlinear accuracy and eciency
To quantify the accuracy and eciency of the schemes discussed above
on nonlinear problems, we consider a series of highly-accurate periodic
solutions computed using the stream function theory method of [17].
BinghamZhang.tex; 9/06/2006; 11:05; p.11
12 Bingham & Zhang
Table I. Approximate resolution required to obtain a given accu-
racy in the linear w at kh = 4. n.a. indicates that the values are
o the scale of Figure 2.
Model Approximate Nx x Nz for an error of
10
3
10
4
10
5
r = 3, even 100 x 80 n.a. n.a.
r = 3, cosine 50 x 32 100 x 80 n.a.
r = 5, even 16 x 16 32 x 32 50 x 60
r = 5, cosine 12 x 10 20 x 16 50 x 25
r = 7, even 10 x 10 10 x 16 15 x 24
r = 7, cosine 7 x 9 10 x 10 15 x 15
Each wave is characterised by a height H, a length L, the water depth
h, and a mean Eulerian velocity u
E
which is set to zero to model
waves in an innite domain. To cover the full range of interest, we have
chosen the conditions shown in Table II for testing the model. Three kh
values appear corresponding to shallow, intermediate and deep water
conditions, and for each of these there are two levels of nonlinearity
at approximately 10% and 90% of the theoretical limiting steepness as
found by [26]. (See also [27] where a convenient rational t to the data
is given.)
Table II. Periodic nonlinear wave conditions used to test the model.
kh 0.5 (shallow) 2 (intermediate) 2 (deep)
H/L 10%(H/L)max .0059 .011 .0135
H/L 90%(H/L)max .053 .10 .12
For each of the conditions of Table II the four models dened in
Table III have been run. Each test was run for a total of ve periods
after which the relative error per wave period was computed from
Error =

5

2
5
e

2
,
2
two norm, (25)
where
5
is the computed surface elevation after exactly ve wave
periods and
e
is the target result from stream function theory. A
Courant number of C
r
= 1 was used for all cases. The results are
collected in Figure 9 with the mildly nonlinear case to the left and
the strongly nonlinear case to the right, kh values increase from top to
BinghamZhang.tex; 9/06/2006; 11:05; p.12
Accuracy of nite dierence wave solutions 13
Table III. The four models tested.
r Order of accuracy Vertical grid
3 2
nd
even
3 2
nd
cosine
5 4
th
cosine
7 6
th
cosine
bottom. The point values correspond to the grid resolutions: (N
x
, N
z
) =
[(16, 9), (32, 9), (64, 12), (128, 24)].
The behaviour of the errors for the 10% case are consistent with
the predictions made in 4. A relative residual tolerance of 10
7
was
used in the iterative scheme for these calculations, so relative errors of
less than 10
5
to 10
6
can not be expected. For the highly-nonlinear
cases the errors are typically larger, but follow a similar trend. The
case at kh = 0.5 resembles a very high solitary wave (very sharp and
narrow peak) and required the application of a lter to obtain stable
results. The lter used was a simple truncation of the Fourier space at
a cut-o frequency of half the Nyquist frequency (i.e. FFT, truncate,
FFT back). Apart from this case, no smoothing or ltering was used
in these calculations.
Figure 10 plots the iterations required by each model vs. N for the
90% steepness case in deep and shallow water. The 10% case required
approximately half as many iterations.
These results are consistent with the linear analysis and demonstrate
that if results at a given accuracy are desired, a signicant improvement
in eciency can be obtained by moving from second-order to fourth-
order in the spatial dierencing scheme and stretching the vertical
grid. It is also clear that decreasing accuracy tolerance and increasing
nonlinearity favour higher-order methods.
6. Sample applications to shoaling wave problems
The analysis described above suggests that a fourth-order model with a
stretched vertical grid of around ten points and a horizontal resolution
of fteen to twenty points per wavelength will be suitable for general
applications of the method. We include here calculations using such a
model for two wave-bottom interaction test cases: Linear shoaling up
a beach; and nonlinear harmonic generation over a submerged bar.
BinghamZhang.tex; 9/06/2006; 11:05; p.13
14 Bingham & Zhang
6.1. Linear shoaling
The derivatives of appearing in the sigma-transformed Laplace prob-
lem where validated by the nonlinear periodic test cases described
above, but the terms involving derivatives of h were not. To validate
these terms, and to test the variable horizontal grid spacing features of
the code we consider the problem of a linear wave shoaling up a beach.
We use a bathymetry dened by
h(x) = h
0

h
0
h
1
2
_
1 + tanh
_
sin (
x
l
)
1 (
2x
l
)
2
__
,
l
2
x
l
2
(26)
with h
0
= 0.5, h
1
= .025, and l = 12 (see Figure 11). This gives a mean
slope of h
x
0.05. A linear wave of period T = .8014s and deep water
length L
0
= 1m is generated at the left end of the domain and allowed
to propagate until a steady state has been reached everywhere. This
produces relative conditions from deep to shallow water, corresponding
to 0.4 kh .
Wave generation and absorption is achieved by relaxing the numer-
ical solution towards a specied analytic solution over regions near the
boundaries of the domain. This is an idea which is often applied in ocean
circulation modelling (see e.g. [28] who modied the idea proposed by
[29]). Relaxation zones for simultaneous generation and absorption of
waves are created by dening a relaxation coecient 0 c
r
(x) 1,
and an exact desired solution (
e
,

e
). After each stage of the time
integration the solution within a relaxation zone is then redened to
be
(x, t) = [1 c
r
(x)] (x, t) + c
r
(x)
e
(x, t)

(x, t) = [1 c
r
(x)]

(x, t) + c
r
(x)

e
(x, t). (27)
To avoid reections, c
r
should be smooth near the junction between
the relaxation zone and the rest of the computational domain. For
pure absorption, the exact solution is simply taken to be zero. Zones of
two wave lengths near both ends have been used here which generally
results in overall reected wave heights of less than 0.1%.
Figure 11 plots the amplitude envelope of the converged steady state
time signals for several choices of horizontal discretisation, along with
the exact result predicted by energy conservation:
a(x)
a
0
=
_
_
k(x)
_
1 +
2k
0
h
0
sinh (2k
0
h
0
)
_
k
0
_
1 +
2k(x)h(x)
sinh (2k(x)h(x))
_
_
_
1
2
. (28)
BinghamZhang.tex; 9/06/2006; 11:05; p.14
Accuracy of nite dierence wave solutions 15
For all four discretisations, N
z
= 9 points in the vertical have been
used and the time step is chosen to give a deep water C
r
= 1. Results
are shown using a coarse and a ne grid (N
x
= 257 & 513 points)
and with both a uniform and a stretched horizontal grid spacing. The
uniform grids provide resolutions of 16 and 32 points per wavelength
respectively at the deep end of the domain but at the shallow end, where
the wavelength has been reduced to 0.4m, the resolution is much
coarser. The stretched grids keep the resolution approximately constant
along the length of the domain by dening x
i
= l tanh (C s
i
)/ tanh (C),
s
i
= (i 1)ds, ds = 1/(N
x
1), with l the length of the domain and
C a constant which controls the stretching. For this case, l = 16 and
C = 1.5 were chosen. While the three other results can be seen to
have converged to the exact solution, the uniformly spaced coarse grid
results show signicant errors due to the under-resolution towards the
shallow end of the domain.
6.2. Harmonic generation over a submerged bar
As a nal example testing all the features of the model, we consider the
transformation of a mildly-nonlinear, monochromatic wave as it travels
up and over a submerged bar. As the wave shoals up the front face of
the bar it steepens dramatically, accumulating higher harmonics which
are released on the down slope producing an irregular pattern behind
the bar. This phenomenon is widely used as a strenuous test case for
nonlinear models of surface wave propagation on a variable bottom.
Details of the experiment can be found in [30], which is usually scaled
by a factor of two, to make it equivalent to [31].
The top plot in Figure 12 shows the bathymetry of the experiment
and a snapshot of the steady-state surface elevation scaled by a factor
of three to make it more visible. The lower plot in this gure shows
a harmonic analysis of the time series after a steady-state has been
reached over the entire domain for the long-wave case (T = 2.01s).
The calculations were made using N
x
= 513, N
z
= 9 and t = .025s
(deep water C
r
0.75.) Relaxation zones of length 2L at both ends
of the domain were used to generate and absorb the waves. Stream
function theory was used for the incident wave. The harmonic analysis
was performed by making a least-squares t of a sum of nine harmonics
to the time series (measured or computed) at a number of points along
the tank length. This gave a residual of at least four orders of magnitude
smaller than the rst harmonic, indicating that level of accuracy. Time
series of the elevations at the points indicated in Figure 12 are shown
in Figure 13. The comparison with the measured results is quite good.
BinghamZhang.tex; 9/06/2006; 11:05; p.15
16 Bingham & Zhang
7. Conclusions
The accuracy and eciency of a nite dierence based solution to the
exact potential ow problem for nonlinear waves has been established
in 2D. The model is exible in terms of the order of the numerical
schemes and the grid spacing; with a computational eort which scales
with N the total number of grid points used. A fourth-order accurate
model with a stretched vertical grid of approximately ten points, a
resolution of 15-20 points per wavelength and a Courant number of
one has been shown to be adequate for general purpose applications.
This model is readily extended to 3D and the inclusion of xed and
oating bodies is conceptually straightforward, which are the topics of
ongoing development.
References
1. B. Li and C.A. Fleming. A three-dimensional multigrid model for fully
nonlinear water waves. Coastal Engineering, 30:235258, 1997.
2. J. N. Newman. Marine Hydrodynamics. The MIT Press, Cambridge,
Massachusetts, 1977.
3. J. N. Newman. Ecient hydrodynamic analysis of very large oating
structures. Marine Structures, 18:169180, 2005.
4. J. N. Newman and C.-H. Lee. Boundary-element methods in oshore structure
analysis. Journal of Oshore Mechanics & Arctic Engineering, 124:8189, 2002.
5. P. D. Sclavounos and H. Borgen. Seakeeping analysis of a high-speed monohull
with a motion-control bow hydrofoil. J. Ship Res., 48:77117, 2004.
6. V. Bertram, H. Yasukawa, and G. Thiart. Evaluation of local pressures in
ships using potential ow models. In Oceans 2005 - Europe (IEEE Cat. No.
05EX1042), pages 11317 Vol.1. IEEE, 2005.
7. B. B uchmann, P. Ferrant, and J. Skourup. Run-up on a body in waves and
current. fully nonlinear and nite-order calculations. Applied Ocean Res.,
22:349360, 2000.
8. S.T. Grilli, P. Guyenne, and F. Dias. A fully non-linear model for three-
dimensional overturning waves over an arbitrary bottom. Int. J. Num. Methods
in Fluids, 35:829867, 2001.
9. D. G. Dommermuth and D. K. P. Yue. A high-order spectral method for the
study of nonlinear gravity waves. J. Fluid Mech., 184:267288, 1987.
10. D. Fructus, D. Clamond, J. Grue, and . Kristiansen. An ecient model for
three-dimensional surface wave simulations. part i: Free space problems. J.
Comput. Phys., 205:665685, 2005.
11. D.R. Fuhrman, H.B. Bingham, and P.A. Madsen. Nonlinear wave-structure
interaction with a high-order Boussinesq model. Coastal Engineering, 52:655
672, 2005.
12. F. Shi, J.T. Kirby, R.A. Dalrymple, and Q Chen. Wave simulations in ponce
de leon inlet using Boussinesq model. J. of Waterway Port Coastal and Ocean
Eng., 129:124135, 2003.
BinghamZhang.tex; 9/06/2006; 11:05; p.16
Accuracy of nite dierence wave solutions 17
13. H.-C. Chen and S.-K. Lee. Rans/laplace calculations of nolinear waves induced
by surface-piercing bodies. J. Engineering Mech., 125(11):12311242, 1999.
14. G.X. Wu, Q.W. Ma, and R. Eatock-Taylor. Numerical simulation of sloshing
waves in a 3d tank based on a nite element method. Applied Ocean Res.,
20:337355, 1998.
15. H. O. Kreiss and J. Oliger. Comparison of accurate methods for the integration
of hyperboloic equations. Tellus Series A, 24(3):199215, 1972.
16. D.R. Fuhrman and H.B. Bingham. Numerical solutions of fully nonlinear
and highly dispersive Boussinesq equations. Int. J. Num. Methods in Fluids,
44(3):231255, 2004.
17. J.D. Fenton. The numerical solution of steady water wave problems. Comput.
Geosci., 14(3):35768, 1988.
18. V. E. Zakharov. Stability of periodic waves of nite amplitude on the surface
of a deep uid. J. Appl. Mech. Tech. Phys., 9:190194, 1968.
19. A. Iserles. A rst course in the numerical analysis of dierential equations.
Cambridge University Press, 1996.
20. Y. Saad and M. H. Schultz. GMRES: a generalized minimal residual algorithm
for solving nonsymmetric linear systems. SIAM J. Sci. Stat. Comput., 7:856
869, 1986.
21. P.R. Amestoy, T.A. Davis, and I.S. Du. An approximate minimum degree
ordering algorithm. SIAM J. on Mat. Anal. Appl., 17:886905, 1996.
22. R.C. Whaley, A. Petitet, and J.J. Dongarra. Automated empirical opti-
mization of software and the atlas project, 2000. Available: http://math-
atlas.sourceforge.net/, 2000.
23. I.S. Du. Multiprocessing a sparse matrix code on the alliant fx/8. J. Comput.
Appl. Math., 27:229239, 1989.
24. Hirsch. Numerical computation of internal and external ows. Volume 1:
Fundamentals of numerical discretizaton. John Wiley & Sons, New York, 1988.
25. L.N. Trefethen. Finite Dierence and Spectral Methods
for Ordinary and Partial Dierential Equations. Available:
http://web.comlab.ox.ac.uk/oucl/work/nick.trefethen/pdetext.html,
Unpublished text, 1996.
26. J. M. Williams. Limiting gravity waves in water of nite depth. Phil. Trans.
Roy. Soy. Lond. A, 302:139188, 1981.
27. J.D. Fenton. Nonlinear wave theories. In B. Le Mehaute and D.M. Hanes,
editors, The Sea, pages 325. John Wiley & Sons., 1990.
28. T.G. Jensen. Open boundary conditions in stratied ocean models. J. Marine
Systems, 16:297322, 1998.
29. J. Larsen and H. Dancy. Open boundaries in short wave simulations - a new
approach. Coastal Engineering, 7:285297, 1983.
30. H.R. Luth, B. Klopman, and N. Kitou. Projects 13g: Kinematics of waves
breaking partially on an oshore bar; ldv measurements for waves with and
without a net onshore current. Technical Report H1573, Delft Hydraulics,
Delft, The Netherlands, 1994.
31. S. Beji and J.A. Battjes. Experimental investigation of wave propagation over
a bar. Coastal Engineering, 19:151162, 1993.
BinghamZhang.tex; 9/06/2006; 11:05; p.17
18 Bingham & Zhang
Even grid Cosine grid
10
0
10
1
10
2
10
4
10
3
10
2
10
1
N
z
E
r
r
o
r
r=3
slope=1.9812
N
x
=7
N
x
=10
N
x
=15
N
x
=20
N
x
=50
N
x
=100
N
x
=300
N
x
=1000
10
1
10
2
10
5
10
4
10
3
10
2
10
1
N
z
E
r
r
o
r
slope=2.0421
r=3
10
1
10
2
10
6
10
4
10
2
N
z
E
r
r
o
r
r=5
slope=3.9202
10
1
10
2
10
8
10
6
10
4
10
2
N
z
E
r
r
o
r
slope=4.0132
r=5
10
1
10
2
10
8
10
6
10
4
10
2
N
z
E
r
r
o
r
r=7
slope=5.878
10
1
10
2
10
10
10
8
10
6
10
4
10
2
N
z
E
r
r
o
r
slope=6.4604
r=7
Figure 2. Convergence of the linear w calculated from

at kh = 4. The slope
is the asymptotic slope of the line for Nx = 1000. Nz point values are at
[9,12,16,24,32,48,100].
BinghamZhang.tex; 9/06/2006; 11:05; p.18
Accuracy of nite dierence wave solutions 19
10
0
10
5
0
0.5
1
1.5
k
N
h

m
a
x

/

(
g

k
N
)
1
/
2
Even grid
N
z
=N
x
/ 2
N
z
=N
x
N
z
=2 N
x
10
0
10
5
0
0.5
1
1.5
k
N
h

m
a
x

/

(
g

k
N
)
1
/
2
Cosine grid
N
z
=N
x
/ 2
N
z
=N
x
N
z
=2 N
x
Figure 3. Maximum eigenvalues (r = 3) vs. relative Nyquist depth. Variation with
vertical grid renement shown by the dierent lines.
10
0
10
5
0
0.5
1
1.5
k
N
h

m
a
x

/

(
g

k
N
)
1
/
2
Even grid
r=3
r=5
r=7
10
0
10
5
0
0.5
1
1.5
k
N
h

m
a
x

/

(
g

k
N
)
1
/
2
Cosine grid
r=3
r=5
r=7
Figure 4. Maximum eigenvalues using Nz = Nx grid points in the vertical vs. relative
Nyquist depth. Variation with increasing order shown by the dierent lines.
BinghamZhang.tex; 9/06/2006; 11:05; p.19
20 Bingham & Zhang
0 10 20 30 40
10
8
10
6
10
4
10
2
10
0
N
x
/L
1

A
Cr=0.5
Cr=1
0 10 20 30 40
10
8
10
6
10
4
10
2
10
0
N
x
/L
1

A
Cr=0.5
Cr=1
0 10 20 30 40
10
2
10
1
10
0
10
1
N
x
/L
|
1

Q
|
Cr=0.5
Cr=1
0 10 20 30 40
10
4
10
3
10
2
10
1
10
0
10
1
N
x
/L
|
1

Q
|
Cr=0.5
Cr=1
Figure 5. Relative amplitude and phase errors over one period as a function of
resolution for two Courant numbers. 2nd-order model (r = 3) with Nz = 15 at
kh = 4.
BinghamZhang.tex; 9/06/2006; 11:05; p.20
Accuracy of nite dierence wave solutions 21
0 10 20 30 40
10
8
10
6
10
4
10
2
10
0
N
x
/L
1

A
r=3
r=5
r=7
0 10 20 30 40
10
8
10
6
10
4
10
2
10
0
N
x
/L
1

A
r=3
r=5
r=7
0 10 20 30 40
10
4
10
3
10
2
10
1
10
0
N
x
/L
|
1

Q
|
r=3
r=5
r=7
0 10 20 30 40
10
5
10
4
10
3
10
2
10
1
10
0
N
x
/L
|
1

Q
|
r=3
r=5
r=7
Figure 6. Relative amplitude and phase errors over one period as a function of
horizontal resolution with Nz = 10 using 2nd- through 6th-order discretisations.
For Cr = 0.5, Nz = 10, and kh = 4.
BinghamZhang.tex; 9/06/2006; 11:05; p.21
22 Bingham & Zhang
0 10 20 30 40
10
5
10
4
10
3
10
2
10
1
10
0
N
x
/L
|
1

Q
|
r=3
r=5
r=7
0 10 20 30 40
10
6
10
4
10
2
10
0
N
x
/L
|
1

Q
|
r=3
r=5
r=7
0 10 20 30 40
10
6
10
4
10
2
10
0
N
x
/L
|
1

Q
|
r=3
r=5
r=7
0 10 20 30 40
10
8
10
6
10
4
10
2
10
0
N
x
/L
|
1

Q
|
r=3
r=5
r=7
Figure 7. Relative phase errors over one period as a function of horizontal resolution
with Nz = 15 & 20 using 2nd- through 6th-order discretisations. For Cr = 0.5,
Nz = 10, and kh = 4.
BinghamZhang.tex; 9/06/2006; 11:05; p.22
Accuracy of nite dierence wave solutions 23
Even grid Cosine grid
10
0
10
1
10
2
10
4
10
3
10
2
10
1
|
1

Q
|
N
z
r=3
N
x
=7
N
x
=10
N
x
=15
N
x
=20
N
x
=50
N
x
=100
N
x
=300
N
x
=1000
10
1
10
2
10
5
10
4
10
3
10
2
10
1
N
z
|
1

Q
|
r=3
10
1
10
2
10
6
10
4
10
2
|
1

Q
|
N
z
r=5
10
1
10
2
10
8
10
6
10
4
10
2
N
z
|
1

Q
|
r=5
10
1
10
2
10
10
10
8
10
6
10
4
10
2
|
1

Q
|
N
z
r=7
10
1
10
2
10
10
10
8
10
6
10
4
10
2
N
z
|
1

Q
|
r=7
Figure 8. Linear dispersion errors for 2nd-, 4th- & 6th-order discretisations as a
function of resolution. The Courant number Cr = 1, and kh = 4.
BinghamZhang.tex; 9/06/2006; 11:05; p.23
24 Bingham & Zhang
10% Max. H/L 90% Max. H/L
10
2
10
3
10
4
10
5
10
0
N
R
e
l
a
t
i
v
e

e
r
r
o
r

p
e
r

w
a
v
e

p
e
r
i
o
d
kh=0.5
r=3, Even
r=3, Cos
r=5, Cos
r=7, Cos
10
2
10
3
10
4
10
3
10
2
10
1
10
0
kh=0.5
N
R
e
l
a
t
i
v
e

e
r
r
o
r

p
e
r

w
a
v
e

p
e
r
i
o
d
10
2
10
3
10
4
10
5
10
0
kh=2
N
R
e
l
a
t
i
v
e

e
r
r
o
r

p
e
r

w
a
v
e

p
e
r
i
o
d
10
2
10
3
10
4
10
6
10
4
10
2
10
0
kh=2
N
R
e
l
a
t
i
v
e

e
r
r
o
r

p
e
r

w
a
v
e

p
e
r
i
o
d
10
2
10
3
10
4
10
6
10
4
10
2
10
0
kh=6.2832
N
R
e
l
a
t
i
v
e

e
r
r
o
r

p
e
r

w
a
v
e

p
e
r
i
o
d
10
2
10
3
10
4
10
6
10
4
10
2
10
0
kh=6.2832
N
R
e
l
a
t
i
v
e

e
r
r
o
r

p
e
r

w
a
v
e

p
e
r
i
o
d
Figure 9. Nonlinear errors (25) for the four versions of the model listed in Table III.
All plots follow the legend shown.
BinghamZhang.tex; 9/06/2006; 11:05; p.24
Accuracy of nite dierence wave solutions 25
kh = 0.5 kh = 2
0 1000 2000 3000
0
5
10
15
20
N
A
v
e
.

i
t
e
r
a
t
i
o
n
s
r=3, even
r=3, cos
r=5, cos
r=7, cos
0 1000 2000 3000
0
5
10
15
20
N
A
v
e
.

i
t
e
r
a
t
i
o
n
s
Figure 10. Average iteration counts for the 90% steepness case.
0.7
0.8
0.9
1
1.1
1.2
1.3
2 4 6 8 10 12 14
-0.5
-0.25
0
a
(
x
)
/
a
0
-
h
(
x
)
/
L
x/L
Linear theory
Nx=513, uniform
Nx=513, stretched
Nx=257, uniform
Nx=257, stretched
-h(x)
Figure 11. Harmonic analysis of a linear wave shoaling up a beach.
BinghamZhang.tex; 9/06/2006; 11:05; p.25
26 Bingham & Zhang
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0 5 10 15 20 25
z

x
12.5 14.5 17.3 21
0
0.2
0.4
0.6
0.8
1
1.2
1.4
0 5 10 15 20 25 30 35 40
a
(
x
)
/
a
1
(
0
)
x
1st harmonic
Exp
2nd harmonic
Exp
3rd harmonic
Exp
4th harmonic
Exp
5th harmonic
Exp
Figure 12. Top: Bar test geometry with a snapshot of the surface elevation (scaled
by a factor 3) along with the measurement locations. Bottom: Harmonic analysis of
the time series at steady-state.
BinghamZhang.tex; 9/06/2006; 11:05; p.26
Accuracy of nite dierence wave solutions 27
-0.02
-0.01
0
0.01
0.02
0.03
25 26 27 28 29 30

t
x=12.5m
-0.02
-0.01
0
0.01
0.02
0.03
25 26 27 28 29 30

t
x=14.5m
-0.02
-0.01
0
0.01
0.02
0.03
30 31 32 33 34 35

t
x=17.3m
-0.02
-0.01
0
0.01
0.02
0.03
35 36 37 38 39 40

t
x=21m
Figure 13. Time series of wave elevations over the bar. Solid line - Calculations;
dots - Measurements.
BinghamZhang.tex; 9/06/2006; 11:05; p.27
BinghamZhang.tex; 9/06/2006; 11:05; p.28