CHAPTER 5

Mathematical Modeling and

Algorithm Development
In this chapter the mathematical representation of some selected models is
presented.

5.1 NAVIER-STOKES EQUATIONS

The basic relation for a general description of fluid flow is a nonlinear
equation of motion which is based fundamentally on the principle of
balance of inertia and external forces represented by volume and surface
forces, as follows (Batchelor, 1967; Foias et al., 2001):

 
@ρui @p @ui @uj 2 @ul
1 rUðρuui Þ 5 2 1 rj Uμ 1 2 δij 1 ρfi (5.1)
@t @xi @xj @xi 3 @xl

where ui are the components of the velocity vector, p is the static pres-
sure, ρ is density, μ is the viscosity and ρfi are the external body force
components.
The conservation of mass in the flow field is also expressed by the
continuity equation as follows:
@ρ
1 rUðρuÞ 5 0 (5.2)
@t
For incompressible fluids (ρ 5 constant), the basic equations are
reduced to:
@ui @p
ρ 1 ρuUrui 5 2 1 ρνΔui 1 ρfi (5.3)
@t @xi
and
rUu 5 0 (5.4)
where ν 5 μ/ρ is the kinematic viscosity.

Numerical Models for Submerged Breakwaters. © 2016 Elsevier Ltd.

All rights reserved. 77
78 Numerical Models for Submerged Breakwaters

These equations are called Navier-Stokes equations.

Considering the turbulence in the flow field, the velocity vector
can be decomposed to a mean value and a dynamic component as fol-
lows (Foias et al., 2001):

u 5 u 1 u0 (5.5)

where u is:
ðT
1
u5 udt (5.6)
T
0

The other parameters in the equation can also be averaged (Reynolds

averaging). Thus, the equation is:



@ρui @p @ui @uj 2 @ul  
1rUðρuu i Þ52 1rj Uμ 1 2 δij 1ρfi 2rU ρu0 u0i
@t @xi @xj @xi 3 @xl
(5.7)
For most applications in coastal and marine engineering, and particu-
larly for submerged breakwaters, it is unnecessary to resolve the details of
the turbulent fluctuations. Therefore, a turbulence model is needed. A
turbulence model approximates the unknown averaged products of turbu-
lent fluctuations. Reynolds stress is basically approximated. The two-
equation turbulence models (i.e. the standard k-ε model) are the most
common models used to approximate the mean flow characteristics for
the numerical problem. The first variable is called turbulent kinetic
energy (k) and the second variable is its turbulent dissipation (ε). The
standard k-ε model is for fully turbulent flows and includes two partial
differential equations (PDEs); one is for the turbulent kinetic energy k
which is derived from the exact equation and one for its dissipation rate ε
derived empirically as discussed before (Foias et al., 2001).
The finite volumes method can be applied to discretize the equations.
Using the finite volume method, the equations can be integrated in a
finite volume. The divergence theorem can be applied and then the inte-
grals can be reduced to a finite number of points. Thus, the derivative
terms are replaced using approximation differences. Other equations can
also be approximated similarly.
The general form of the Navier-Stokes equations in three dimensions
is presented here.
 Mathematical Modeling and Algorithm Development 79

The continuity equation is:

@ρ @ðρuÞ @ðρvÞ @ðρwÞ
1 1 1 50 (5.8)
@t @x @y @z
The momentum equation is:
@ui @ui @p @τ ij @u0i u0j
ρ 1 ρuj 1 2 ρgi 1 1ρ 50 (5.9)
@t @xj @xi @xj @xj
where V 5 ui^ 1 vj^ 1 w k^ is the velocity vector and gi is the body forces and


@ui @uj
τ ij 5 2 μ 1 (5.10)
@xj @xi
where


k2 @ui @uj 2
u0i u0j 5 2cμ 1 1 kδij (5.11)
ε @xj @xi 3
!
@Ui @Uj 2 2
2 ρu0i u0j 5 μt 1 2 ρkδij 5 2μt Eij 2 ρkδij
@xj @xi 3 3 (5.12)
δij 5 1 if i 5 j and δij 5 0 if i 6¼ j
The unknowns include velocities (u, v, w)(x, y, z, t), pressure P(x, y, z, t),
surface η(x, y, t) and the domain (t). When we neglect diffusive terms in the
Navier-Stokes equations, we get the Euler equations for conservation of
mass and momentum (Batchelor, 1967; Foias et al., 2001). The three-
dimensional partial differential equations for viscous, compressible flow are:


@ρ @ρ @ρ @ρ @U @V @W
1U 1V 1W 52ρ 1 1 (5.13)
@t @x @y @z @x @y @z


@U @U @U @U 1 @P μ @2 U @2 U @2 U
1U 1V 1W 52 1 gx 1 1 2 1 2
@t @x @y @z ρ0 @x ρ0 @x2 @y @z
(5.14)


@V @V @V @V 1 @P μ @2 V @2 V @2 V
1U 1V 1W 52 1 gy 1 1 1
@t @x @y @z ρ0 @y ρ0 @x2 @y2 @z2
(5.15)

2 
@W @W @W @W 1 @P μ @ W @2 W @2 W
1U 1V 1W 52 1gz 1 1 2 1 2
@t @x @y @z ρ0 @z ρ0 @x2 @y @z
(5.16)
80 Numerical Models for Submerged Breakwaters

However, if the flow is incompressible, the conservation of mass

reduces to:
@U @V @W
1 1 50 (5.17)
@x @y @z

5.2 THE TURBULENT MODEL

Among the existing turbulent models the standard two-equation k-ε
models are the most used methods to estimate the eddy viscosity. In the
k-ε model the transport equations for k and ε are solved. The turbulent
viscosity is computed from (Foias et al., 2001):
k1=2 k2
μt 5 ρcμ 5 ρcμ (5.18)
l ε
or
μt k1=2 k2
νt 5 5 cμ 5 cμ (5.19)
ρ l ε
where k is the kinetic energy, ε is the dissipation rate and cμ is a parameter
that depends on the k-ε turbulence model:
cμ 5 0:09 (5.20)
The turbulent kinetic energy for three-dimensional flows is given by:
1 S S S
k 5 Uðu 2 1 v 2 1 w 2 Þ (5.21)
2
The governing transport equations for this model can be written as:
k-equation:


 
@k @ðui kÞ @ μt @k
1 5 μ1 1P 2ε (5.22)
@t @xi @xi σk;s @xi


 
@k @ðukÞ @ðvkÞ @ðwkÞ @ μt @k
1 1 1 5 μ1
@t @x @y @z @x σk;s @x


 

 
@ μ @k @ μ @k
1 μ1 t 1 μ1 t 1
@y σk;s @y @z σk;s @z
"


2
2
2
2
2
μt @u @v @w @u @v @u @w
2 12 12 1 1 1 1
ρ @x @y @z @y @x @z @x
!2 
@v @w
1 1 2ε (5.23)
@z @y
 Mathematical Modeling and Algorithm Development 81

ε-equation:


 
@ðεÞ @ðui εÞ @ μ @ε ε ε2
1 5 μ1 t 1 c1;ε P 2 c2;ε (5.24)
@t @xi @xi σε;s @xi k k

@ε @ðuεÞ @ðvεÞ @ðwεÞ

1 1 1 5
@t @x @y @z


 

 

 
@ μt @ε @ μt @ε @ μt @ε
μ1 1 μ1 1 μ1 1
@x σε;s @x @y σε;s @y @z σε;s @z
"


2
2
2
2
2
ε μt @u @v @w @u @v @u @w
c1;ε 2 12 12 1 1 1 1
k ρ @x @y @z @y @x @z @x

2 
@v @w ε2
1 1 2 c2;ε (5.25)
@z @y k

The unknown constants cμ, c1,ε, c2,ε, σk and σε are assumed to be con-
stant for all types of flows. The cμ constant was determined previously.
The empirical constants ck , cε ,c1;ε , c2;ε are given in Table 5.1. For further
information see Batchelor, 1967; Foias et al., 2001.

 
0 0 @ui μt @ui @uj 2 @ui
P 5 2 ui uj 5 1 2 kδij
@xj ρ @xj @xi 3 @xj


2
2
2  (5.26)
μt @u @v @u @v
5 2 12 1 1
ρ @x @y @y @x


 
@ui μt @ui @uj 2 @ui
P 5 2 u0i u0j 5 1 2 kδij
@xj ρ @xj @xi 3 @xj


2
2
2
2
2
μt @u @v @w @u @v @u @w
5 2 12 12 1 1 1 1
ρ @x @y @z @y @x @z @x

2 
@v @w
1 1
@z @y
(5.27)

Table 5.1 Parameters for the standard k-ε model

Model parameters
σk 5 1 σε 5 1:314 c1;ε 5 1:44 c2;ε 5 1:92
82 Numerical Models for Submerged Breakwaters

Therefore, the general form of the Navier-Stokes equations in three

dimensions is:


@u @u @u @u
ρ 1ρ u 1v 1w 5
@t @x @y @z

2 

@p @ u @2 u @2 u @ @u
ρgx 2 1μ 1 2 1 2 1ρ 2vt (5.28)
@x @x2 @y @z @x @x





@ @u @v @ @u @w
1ρ vt 1 1ρ vt 1
@y @y @x @z @z @x


@v @v @v @v
ρ 1ρ u 1v 1w 5
@t @x @y @z

2 

@p @v @2 v @2 v @ @v
ρgy 2 1 μ 1 1 1 ρ 2vt (5.29)
@y @x2 @y2 @z2 @y @y





@ @u @v @ @v @w
1ρ vt 1 1ρ vt 1
@x @y @x @z @z @y


@w @w @w @w
ρ 1ρ u 1v 1w 5
@t @x @y @z

2 

@p @w @2 w @2 w @ @w
ρgz 2 1μ 1 2 1 2 1ρ 2vt (5.30)
@z @x2 @y @z @z @z





@ @u @w @ @v @w
1ρ vt 1 1ρ vt 1
@x @z @x @y @z @y

5.3 INITIAL AND BOUNDARY CONDITIONS

The initial and boundary conditions are also required in numerical
modeling of submerged breakwaters. Depending on the problem, these
conditions may vary and either one or both may be applied in the prob-
lem. For example, if we are dealing with infinite domains, we are not
concerned about the boundary conditions and the problem reduces to an
initial value problem. In the case of the steady conditions, the problem
then reduces to a boundary value problem. The boundary conditions
 Mathematical Modeling and Algorithm Development 83

may be specified by variable value, gradient of the variable or relationship

of the variable and its gradient. In these problems, the number of bound-
ary equations is determined based on the order of the highest spatial
derivatives in the governing equation for each coordinate space.
Obviously, for an unsteady problem with finite domain, both initial and
boundary conditions are needed. These problems are called initial-
boundary value problems.

5.4 SHALLOW WATERS

In the model of long waves in shallow waters, the vertical velocity and
acceleration of the fluid particles are assumed to be zero. Therefore, all
terms containing W in the governing equations are omitted. The hori-
zontal plane in the still water surface level is taken as the Cartesian coor-
dinate system. Therefore, the water depth and water surface variations are
determined according to this reference plane (for further information, see
Mader, 2005).
Vertically averaged velocity components can be introduced by
(Hansen, 1956):
ðη
1
U5 udz (5.31)
ðh 1 ηÞ
2h

and
ðη
1
V5 vdz (5.32)
ðh 1 ηÞ
2h

where U is the velocity in the x direction, V is the velocity in the

y direction, η is the wave height above mean water level and h is depth
(Mader, 2005; Vreugdenhil, 1994).

5.5 THE EXTENDED MILD-SLOPE EQUATION

The mild-slope equation has been improved as the influence of steep
slopes and bottom curvatures has been included. (Chamberlain and
Porter, 1995; Massel, 1993, 1996; Porter and Staziker, 1995).
According to the expression by Massel (1989), the following restric-
tion exists for wave propagation over mild slopes:
84 Numerical Models for Submerged Breakwaters

rh ra
U {1 (5.33)
kh ka
where r is the horizontal gradient operator and h, k and a are water
depth, wave number and wave amplitude, respectively.
The mild-slope equation is relatively accurate for slopes to a maxi-
mum ratio of 1:3 (Booij, 1983); however, beyond this ratio the classical
mild-slope equation by Berkhoff (1972) is no longer very accurate.
Furthermore, for arbitrary topographies and particularly submerged
breakwaters, bottom curvature ( r2 h) and the square of bottom slope
(ðrhÞ2 ) should be considered (Massel, 1993).
Massel (1993, 1996) developed an extended refraction-diffraction
equation for the steep slopes based on the Galerkin-Eigenfunction
method. This extended equation includes the higher-order terms.
The parameters related to the higher-order bottom effects are deter-
mined based on the expressions by Massel (1993). Some expressions given
by Massel (1993) were corrected later by Suh et al. (1997).

5.6 BOUSSINESQ EQUATIONS

Boussinesq equations (Boussinesq, 1871, 1872) are a set of nonlinear
partial differential equations. Boussinesq equations incorporate frequency
dispersion while the shallow water equations are not frequency-dispersive.
Boussinesq equations are widely applied in numerical modeling in the
field of coastal engineering to model water waves in shallow seas and
harbors. The Boussinesq equations are applicable for fairly long waves.
Although wave simulation in such cases is perfectly described by the
Navier-Stokes equations, currently it is extremely difficult to solve the
full three-dimensional equations in complicated models. Therefore,
approximate models such as the Boussinesq equations can be applied.
They can be used to reduce three-dimensional problems to two-
dimensional states. This is typically done by assuming a linear vertical
distribution of the flow field with nonhydrostatic effects which is different
from the nonlinear shallow water equations (see Dingemans, 1997;
Hamm et al., 1993; Kirby, 2003).
The Boussinesq equations are suitable for the simulation of wave fields
in the presence of coastal structures and harbors. Finite-difference, finite-
volume or finite-element techniques are applicable for the discretization
of the governing equations.
 Mathematical Modeling and Algorithm Development 85

5.7 SMOOTHED PARTICLES HYDRODYNAMICS

Smoothed particles hydrodynamics (SPH) (Gingold and Monaghan
1977; Lucy 1977) is a meshless Lagrangian method, where the fluid is
discretized using a finite set of physical particles within the body of the
fluid. The required physical quantities are computed directly and the
free surface of the flow can be tracked. Each particle with given mass,
density and volume can move and carry some value of ui. SPH interpo-
lation for field values of any scalar value uðx; tÞ is (Dalrymple and
Rogers 2006; Liu and Liu, 2003; Monaghan, 1994):
X mi X
uðx; tÞ 5 ui W ðx 2 xi ; hÞ 5 uðxi ÞVi W ðx 2 xi ; hÞ (5.34)
i
ρi i

where mi and ρi are the mass and the density of particle i, so the volume
of the particle i is Vi 5 mi =ρi and we have:

ρi 5 ρðxi Þ (5.35)

ui 5 uðxi Þ (5.36)

W is also a kernel function on the points in the neighborhood of x.

For W parametrized by its smoothing length h:
ð
W ðx; hÞdx 5 1
Rd (5.37)
lim W ðU; hÞ 5 δ
h-0

where δ is the Dirac’s delta distribution.

The approximation of δ with the smoothing kernel W with length
h goes to zero as h-0. Depending on the distance between the parti-
cles Δp, the error may vary significantly with zero for Δp-0.
The SPH method is consistent if Δp=h-0 as h-0. In practice this
ratio depends on the smoothing kernel used and is between about 0.25
and 0.5 (Dalrymple and Rogers 2006; Liu and Liu, 2003; Monaghan,
1994). For further information about Smoothed particles hydrodynam-
ics method and its application in the field of coastal engineering, see
Dalrymple and Knio (2001); Dalrymple and Rogers (2006); Gingold
and Monaghan (1977); Hoover (2006); Liu and Liu (2003); Lucy
(1977); Monaghan, (1994).
86 Numerical Models for Submerged Breakwaters

5.8 ARTIFICIAL NEURAL NETWORKS

5.8.1 MLP Model
Algorithm Derivation
Considering the steepest descent algorithm, the Newton’s algorithm and
Gauss-Newton’s algorithm, a brief description on the derivation of the
Levenberg-Marquardt (LM) algorithm (Levenberg, 1944; Marquardt,
1963) is presented here (Yu and Wilamowski, 2011).
For the LM algorithm, the performance index to be optimized is the
sum of the squares of the deviations defined to evaluate the training pro-
cess. It is calculated for all training patterns and network outputs, by
1 XN
SðwÞ 5 p51
ðyp 2 y^p Þ2 (5.38)
2
and using the equation:
1 XN 2
SðwÞ 5 ½e 
p51 p
(5.39)
2
where w consists of all weights of the network, p is the index of patterns,
from 1 to N, where N is the number of patterns, and yp and y^p are
respectively the desired and the actual output vector when applying
pattern p (Gavin, 2011; Yu and Wilamowski, 2011).
The LevenbergMarquardt algorithm is basically a combination of
the steepest descent algorithm and the GaussNewton algorithm. The
updating rule is (Yu and Wilamowski, 2011):

wk11 5 wk 2 ðJkT Jk 1λIÞ21 Jk ek (5.40)

where λ is a non-negative damping factor and I is the identity matrix
(Gavin, 2011; Marquardt, 1963; Yu and Wilamowski, 2011) and
8 9
> @e1 @e1 @e1 >
>
> ? >
>
>
> @w @w @w >
>
>
>
1 2 M >
>
>
> >
>
>
> @e2 @e2 @e2 >>
>
< ? >
=
J5 @w1 @w 2 @w M (5.41)
>
> >
>
> ^
> ^ & ^ > >
>
> >
>
>
> @e >
>
>
> @e @e >
>
>
>
P P
?
P >
>
: @w @w @w ;
1 2 M
 Mathematical Modeling and Algorithm Development 87

where vector w includes weights w1 ; w2 ; . . .; wm which are linear and error

vector e has the form
8 9
>
> e1 >
< > =
e2
e5 (5.42)
>
> ^>
: > ;
eP
Therefore, the Levenberg-Marquardt algorithm will be more similar to
the steepest descent algorithm when the combination coefficient λ is very
large or the parameters are far from their desirable value. The Levenberg-
Marquardt algorithm will be more similar to the Gauss-Newton method
when the combination coefficient λ is very small or parameters are close
to their desirable value (see Gavin, 2011; Haykin, 1999; Marquardt, 1963;
Sinha and Gupta, 1999; Yu and Wilamowski, 2011).

Transfer Function
The transfer function in an artificial neural network is an essential element
of its structure. A transfer function limits the response amplitude of the
unit (also known as a squashing function). It squashes the allowed ampli-
tude of the response signal to a finite value. The most commonly used
transfer functions include identity, step, binary step, unipolar sigmoid,
bipolar sigmoid, hyperbolic tangent and radial basis functions (Karlik and
Olgac, 2010). Usually, three conventional differentiable and monotonic
transfer functions are employed for the evolution of the MLP architecture
along with the LM learning algorithm. These proposed well-known and
effective activation functions are bipolar sigmoid, unipolar sigmoid, and
hyperbolic tangent, and are discussed next (Karlik and Olgac, 2010).
The radial basis function will be discussed later, as it will be used for the
RBF model.
The unipolar sigmoid transfer function is given as (Karlik and Olgac,
2010):
1
yðxÞ 5 (5.43)
1 1 e2x
A graphic of the sigmoid function can be seen in Figure 5.1 (Karlik
and Olgac, 2010).
This function is particularly beneficial for BP algorithm as it mini-
mizes the computational capacity for the training. The sigmoid function
maps the inputs onto (0,1) (Karlik and Olgac, 2010).
88 Numerical Models for Submerged Breakwaters

Figure 5.1 Unipolar sigmoid function.

Figure 5.2 Bipolar sigmoid function.

The bipolar sigmoid transfer function is given by (Karlik and Olgac, 2010):
1 2 e2x
yðxÞ 5 (5.44)
1 1 e2x
The function is analogous to the sigmoid function. However, in the
Bipolar sigmoid function output values are in the range of [ 2 1, 1]
(Figure 5.2).
The function is basically the ratio between the hyperbolic sine and
cosine functions. It can be also rewrite in the form of exponential func-
tions as follows (Karlik and Olgac, 2010):
sin hðxÞ ex 2 e2x
yðxÞ 5 tan hðxÞ 5 5 x (5.45)
cos hðxÞ e 1 e2x
The hyperbolic tangent function is similar to the sigmoid function but
with an output range between 21 and 1 (as seen in Figure 5.3).
 Mathematical Modeling and Algorithm Development 89

Figure 5.3 Hyperbolic tangent function.

To conduct a modeling using the neural networks, to select an appro-

priate transfer function for the model, all the transfer functions mentioned
here should be tested using the same data and compared together.
A various number of nodes in the hidden layer for all the transfer func-
tions should be selected (for example: a number of nodes like 5, 10, 15,
20, 25, etc.). The number of iterations is typically the same for all transfer
functions and the training samples are set in such a way as to be
completely random for an improved learning performance. The neurons
in the different models are arranged to learn at the same rate by assigning
a learning rate parameter (Sinha and Gupta, 1999). Past research demon-
strate that the artificial neural network performance might be improved
by choosing an effective transfer function. In many cases, having com-
pared the performances of functions, simulation results illustrate that the
hyperbolic tangent (Tanh) and logistic sigmoid transfer functions offer
better recognition accuracy than the other functions for various sets of
hidden neurons (Karlik and Olgac, 2010). An antisymmetric hyperbolic
tangent transfer function often have a slightly better performance in LM
learning in comparison with an unsymmetrical logistic transfer function
(Sinha and Gupta,1999). However, a serious issue to be resolved in
regression problems is extrapolation effects. To resolve this problem, the
transfer function in the output layer is replaced by a linear function, pro-
viding an unchanged activation level in the output layer. The linear trans-
fer function does not saturate, so it can extrapolate further. However, a
low learning rate, lower than 0.1, must be applied (Haykin, 1999; Sinha
and Gupta, 1999). Depending on the cases, the model often computed
better results when Tanh-Linear (combination of transfer functions for
nodes in hidden and output layers, respectively) is used compared with
Tanh-Tanh or Tanh-Sigmoid (Haykin, 1999; Sinha and Gupta, 1999).
90 Numerical Models for Submerged Breakwaters

Figure 5.4 Gaussian function.

5.8.2 RBF Model

A brief summary of RBFs is presented here (for a more thorough
description see Broomhead and Lowe, 1988; Haykin, 1999; Howlett and
Jain, 2001a,b). A radial basis function is defined as:

jjxi2ck jj2
2
ϕðjjxi 2 ck jjÞ 5 e 2σ2
k (5.46)

where ϕ is a Gaussian radial basis function, xi is the input vector, ck and

σk represent respectively the center and the width of the kth RBF unit
(Haykin, 1999).
The Gaussian function can be seen in Figure 5.4. Its output is deter-
mined by the distance between xi and ck . The function produces very low
or high values respectively if the input is outside the range of training
data and close to a neuron. The output layer of the RBF network is
linear. The outputs of the hidden layer are weighted by fwk gmk51 finally
providing the output function y^ðxi Þ (Haykin, 1999):
Xm
y^ðxi Þ 5 k51 k
w ϕðjjxi 2 ck jjÞ 1 b (5.47)

where w1 ; w2 ; . . .; wm are weight factors and b is bias value.

REFERENCES
Batchelor, G.K., 1967. An Introduction to Fluid Dynamics. Cambridge University Press.
Berkhoff, J.C.W., 1972. Computation of combined refraction-diffraction, Proceedings of
the 13th International Conference on Coastal Engineering, vol. 1. ASCE.
Booij, N., 1983. A note on the accuracy of the mild-slope equation. Coastal Eng. 7,
191203.
 Mathematical Modeling and Algorithm Development 91

Boussinesq, J., 1871. Théorie de l’intumescence liquide, applelée onde solitaire ou de

translation, se propageant dans un canal rectangulaire. C. R. Acad. Sci. 72, 755759.
Boussinesq, J., 1872. Théorie des ondes et des remous qui se propagent le long d’un canal
rectangulaire horizontal, en communiquant au liquide contenu dans ce canal des
vitesses sensiblement pareilles de la surface au fond. J. Math. Pures Appl. Deuxiéme
Série 17, 55108.
Broomhead, D.S., Lowe, D., 1988. Multivariable functional interpolation and adaptive
networks. Complex Syst. 2, 321355.
Chamberlain, P.G., Porter, D., 1995. The modified mild-slope equation. J. Fluid Mech.
291, 393407.
Dalrymple, R., Rogers, B., 2006. Numerical modeling of water waves with the SPH
method. Coast. Eng. 53 (23), 141147.
Dalrymple, R.A., Knio, O., 2001. SPH modelling of water waves. In: Hanson, H.,
Larson, M. (Eds.), ASCE Conf. Proc, vol. 260. ASCE, Lund, Sweden, p. 80.
Dingemans, M.W., 1997. Wave propagation over uneven bottoms. Advanced Series on
Ocean Engineering 13. World Scientific, Singapore, ISBN 981-02-0427-2. See Part
2, (Chapter 5).
Foias, C., Manley, O., Rosa, R., Temam, R., 2001. Navier-Stokes Equations and
Turbulence. Cambridge University Press, Cambridge, UK.
Gavin, H., 2011. The Levenberg-Marquardt Method for Nonlinear Least Squares Curve-
Fitting Problems. Duke University.
Gingold, R.A., Monaghan, J.J., 1977. Smoothed particle hydrodynamics: theory and
application to non-spherical stars. Mon. Not. R. Astron. Soc. 181, 375389.
Hamm, L., Madsen, P.A., Peregrine, D.H., 1993. Wave transformation in the nearshore
zone: a review. Coast. Eng. 21 (13), 539.
Hansen, W., 1956. TheoriezurErrechnung des Wasserstandes und der Strömungen in
TandmeerennebstAnwendungen. Tellus 8 (3).
Haykin, S., 1999. Neural Networks: A Comprehensive Foundation, second ed. Prentice-
Hall, Englewood Cliffs, NJ.
Hoover, W.G., 2006. Smooth Particle Applied Mechanics: The State of the Art. World
Scientific.
Howlett, R., Jain, L., 2001a. Radial basis function networks 1- Recent developments in
theory and applications. Studies in Fuzzyness and Soft Computing. Physica-Verlag,
New York.
Howlett, R., Jain, L., 2001b. Radial basis function networks 2 - New advances in design.
Studies in Fuzzyness and Soft Computing. Physica-Verlag, New York.
Karlik, B., Olgac, A., 2010. Performance analysis of various activation functions in gener-
alized MLP architectures of neural networks. Int. J. Artif. Intell. Expert Syst. (IJAE) 1
(4), 111122.
Kirby, J.T., 2003. Boussinesq models and applications to nearshore wave propagation, surf-
zone processes and wave-induced currents. In: Lakhan, V.C. (Ed.), Advances in
Coastal Modeling. Elsevier Oceanography Series, 67. Elsevier, pp. 141.
Levenberg, K., 1944. A Method for the solution of certain non-linear problems in least
squares. Q. Appl. Math. 2, 164168.
Liu, G.R., Liu, M.B., 2003. Smoothed Particle Hydrodynamics: A Meshfree Particle
Method. World Scientific, Singapore.
Lucy, L.B., 1977. A numerical approach to the testing of the fission hypothesis. Astron. J.
82, 10131024.
Mader, C.L., 2005. Numerical Modeling of Water Waves, second ed. CRC PRESS.
Marquardt, D., 1963. An algorithm for least-squares estimation of nonlinear parameters.
SIAM J. Appl. Math. 11 (2), 431441.
Massel, S.R., 1989. Hydrodynamics of Coastal Zones. Elsevier Science Publ., Amsterdam.
92 Numerical Models for Submerged Breakwaters

Massel, S.R., 1993. Extended refractiondiffraction equation for surface waves. Coastal
Eng. 19, 97126.
Massel, S.R., 1996. Ocean Surface Waves: Their Physics and Prediction. World Scientific,
Singapore.
Monaghan, J.J., 1994. Simulating free surface flows with SPH. J. Comput. Phys. 110,
399406.
Porter, D., Staziker, D.J., 1995. Extensions of the mild-slope equation. J. Fluid Mech.
300, 367384.
Sinha, N.K., Gupta, M.M., 1999. Soft Computing and Intelligent Systems: Theory and
Applications. Academic Press, New York, A volume of 25 chapters, 614 p.
Vreugdenhil, C.B., 1994. Numerical Methods for Shallow-Water Flow. Kluwer Academic
Publishers.
Suh, K.D., Lee, C., Park, W.S., 1997. Time-dependent equations for wave propagation
on rapidly varying topography. Coastal Eng. 32, 91117.
Yu, H., Wilamowski, B., 2011. second ed. Levenberg-Marquardt Training, vol. 5.
Industrial Electronics Handbook.