SWAN

SCIENTIFIC AND TECHNICAL DOCUMENTATION

SWAN Cycle III version 40.91A

SWAN SCIENTIFIC AND TECHNICAL DOCUMENTATION

by

The SWAN team Delft University of Technology Faculty of Civil Engineering and Geosciences Environmental Fluid Mechanics Section P.O. Box 5048 2600 GA Delft The Netherlands [email protected] http://www.swan.tudelft.nl

mail address :

e-mail home page

: :

Copyright (c) 1993-2013 Delft University of Technology. Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation; with no Invariant Sections, no Front-Cover Texts, and no BackCover Texts. A copy of the license is available at http://www.gnu.org/licenses/fdl.html#TOC1.

iv

Contents
1 Introduction 1.1 Historical background . . 1.2 Purpose and motivation 1.3 Readership . . . . . . . . 1.4 Scope of this document . 1.5 Overview . . . . . . . . . 1.6 Acknowledgements . . . 1 1 2 2 2 4 4 7 7 10 10 10 12 12 17 19 25 31 33 35 35 36 37 37 39 41 41 43 43 46 47

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

2 Governing equations 2.1 Spectral description of wind waves . . . . . . . . . . . 2.2 Propagation of wave energy . . . . . . . . . . . . . . . 2.2.1 Wave kinematics . . . . . . . . . . . . . . . . . 2.2.2 Spectral action balance equation . . . . . . . . . 2.3 Sources and sinks . . . . . . . . . . . . . . . . . . . . . 2.3.1 General concepts . . . . . . . . . . . . . . . . . 2.3.2 Input by wind (Sin ) . . . . . . . . . . . . . . . . 2.3.3 Dissipation of wave energy (Sds ) . . . . . . . . . 2.3.4 Nonlinear wave-wave interactions (Snl ) . . . . . 2.3.5 First- and second-generation model formulations 2.4 Wave damping due to vegetation . . . . . . . . . . . . 2.5 The inuence of ambient current on waves . . . . . . . 2.6 Modelling of obstacles . . . . . . . . . . . . . . . . . . 2.6.1 Transmission . . . . . . . . . . . . . . . . . . . 2.6.2 Reection . . . . . . . . . . . . . . . . . . . . . 2.6.3 Diraction . . . . . . . . . . . . . . . . . . . . . 2.7 Wave-induced set-up . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . in SWAN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3 Numerical approaches 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.1 Discretization in geographical space . . . . . . . . . . 3.2.2 Note on the choice of geographic propagation schemes 3.2.3 Discretization in spectral space . . . . . . . . . . . . v

vi 3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 Solution algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . An illustrative explanation of the sweeping approach . . . . . . . . . . . . Implementation of DIA within the four-sweep technique . . . . . . . . . . . Convergence-enhancing measures . . . . . . . . . . . . . . . . . . . . . . . Stopping criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Governing equations in curvi-linear co-ordinates . . . . . . . . . . . . . . . Computation of force in curvi-linear co-ordinates . . . . . . . . . . . . . . . Numerical treatment of obstacles . . . . . . . . . . . . . . . . . . . . . . . Crossing of obstacle and grid line . . . . . . . . . . . . . . . . . . . . . . . Integration over . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Transformation from relative to absolute frequency . . . . . . . . . . . . . Interpolation of spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Computation of breaking source term . . . . . . . . . . . . . . . . . . . . . On the approximation of refraction on coarse grids in large-scale SWAN applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.16.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.16.2 The problem with refraction on coarse grids and how to deal with it 3.16.3 A historical overview of limitation on c . . . . . . . . . . . . . . . 48 51 54 54 56 58 60 61 63 63 64 65 66 68 68 68 72 75 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 77 78 78 84 87 87 88 89 89 90 93 93 93 94 94 95 95 95

4 Wave boundary and initial conditions 5 Implementation of 2D wave set-up 5.1 Methods . . . . . . . . . . . . . . . . . . . . . . 5.2 Analysis and Results . . . . . . . . . . . . . . . 5.2.1 Discretization of the 2D setup equation . 5.2.2 The iterative solver for the linear system

6 Iterative solvers 6.1 Strongly Implicit Procedure (SIP) . . . . . . . . . . . . . . . . . . . . . . . 6.2 Successive Over Relaxation (SOR) technique . . . . . . . . . . . . . . . . . 7 Parallel implementation aspects 7.1 Load balancing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2 Parallelization of implicit propagation schemes . . . . . . . . . . . . . . . . 8 Unstructured mesh implementation 8.1 Description of an unstructured grid . . . . . . . . 8.1.1 Denitions . . . . . . . . . . . . . . . . . . 8.1.2 Relations between number of cells, vertices 8.1.3 Conditions imposed to the grid . . . . . . 8.2 Some notes on grid generation . . . . . . . . . . . 8.3 Numerical method . . . . . . . . . . . . . . . . . 8.3.1 Discretization procedure . . . . . . . . . . . . . . . . . . . . . . and faces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

vii 8.3.2 The sweeping algorithm . . . . Interpolation at user-dened locations . Computation of wave-induced force . . Calculation of diusion-like terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 . 100 . 102 . 104 105 107 123

8.4 8.5 8.6

9 The overall solution algorithm Bibliography Index

viii

Chapter 1 Introduction
The main goal of the SWAN model is to solve the spectral action balance equation without any a priori restrictions on the spectrum for the evolution of wave growth. This equation represents the eects of spatial propagation, refraction, shoaling, generation, dissipation and nonlinear wave-wave interactions. The basic scientic philosophy of SWAN is identical to that of WAM cycle 3. SWAN is a third-generation wave model and it uses the same formulations for the source terms. Whereas the WAM model considers problems on oceanic scales, with SWAN wave propagation is calculated from deep water to the surf zone. Since, WAM makes use of explicit propagation schemes in geographical and spectral spaces, it requires very small grid sizes in shallow water and is thus unsuitable for applications to coastal regions. For that reason, SWAN employs implicit schemes, which are more robust and economic in shallow water than the explicit ones. Note that SWAN may be less ecient on oceanic scales than WAM.

1.1

Historical background

Over the past two decades, a number of advanced spectral wind-wave models, known as third-generation models, has been developed such as WAM (WAMDI Group, 1988), WAVEWATCH III (Tolman, 1991), TOMAWAC (Benoit et al., 1996) and SWAN (Booij et al., 1999). These models solve the spectral action balance equation without any a priori restrictions on the spectrum for the evolution of wave growth. Based on the wave action balance equation with sources and sinks, the shallow water wave model SWAN (acronym for Simulating WAves Nearshore) is an extension of the deep water third-generation wave models. It incorporates the state-of-the-art formulations for the deep water processes of wave generation, dissipation and the quadruplet wave-wave interactions from the WAM model (Komen et al., 1994). In shallow water, these processes have been supplemented with the state-of-the-art formulations for dissipation due to bottom friction, triad wave-wave interactions and depth-induced breaking. SWAN is fully spectral (in all directions and frequencies) and computes the evolution of wind waves in coastal regions 1

Chapter 1

with shallow water and ambient current. SWAN is developed at Delft University of Technology and is freely available from http://www.swan.tudelft.nl. It is used by many goverment authorities, research institutes and consultants worldwide. The feedback has widely indicated the reliability of SWAN in dierent experiment and eld cases. Initially, the SWAN cycle 1 was formulated to be able to handle only stationary conditions on a rectangular grid. Later on, SWAN cycle 2 model has been developed. This is considered as the second step in the development of SWAN models. Cycle 2 of SWAN is stationary and optionally nonstationary. It can compute the wave propagation not only on a regular rectangular grid, but also on a curvi-linear grid. Previous ocial versions 30.62, 30.75, 40.01 and 32.10 belong to the cycle 2 of SWAN. This section is under preparation.

1.2

Purpose and motivation

The purpose of this document is to provide relevant information on the mathematical models and numerical techniques for the simulation of spectra of random short-crested, wind-generated waves in coastal regions. Furthermore, this document explains the essential steps involved in the implementation of various numerical methods, and thus provides an adequate reference with respect to the structure of the SWAN program.

1.3

Readership

This document is, in the rst place, addressed to those, who wish to modify and to extend mathematical and numerical models for shallow wind-wave problems. However, this material is also useful for those who are interested in the application of the techniques discussed here. The text assumes the reader has basic knowledge of analysis, partial dierential equations and numerical mathematics and provides what is needed both in the main text and in the appendices.

1.4

Scope of this document

SWAN is a third-generation wave model for obtaining realistic estimates of wave parameters in coastal areas, lakes and estuaries from given wind, bottom and current conditions. However, SWAN can be used on any scale relevant for wind-generated surface gravity waves. The model is based on the wave action balance equation (or energy balance in the absence of currents) with sources and sinks. Good introductory texts on the background of SWAN are Young (1999) and Booij et al. (1999).

Introduction The following wave propagation processes are represented in SWAN: propagation through geographic space, refraction due to spatial variations in bottom and current, diraction, shoaling due to spatial variations in bottom and current, blocking and reections by opposing currents and transmission through, blockage by or reection against obstacles. The following wave generation and dissipation processes are represented in SWAN: generation by wind, dissipation by whitecapping, dissipation by depth-induced wave breaking, dissipation by bottom friction and wave-wave interactions in both deep and shallow water.

In addition, the wave-induced set-up of the mean sea surface can be computed in SWAN. However, wave-induced currents are not computed by SWAN. In 1D cases, computation of wave-induced set-up is based on exact shallow water equations, whereas in 2D cases they need to be approximated since the eects of wave-induced currents are ignored. Diraction is modelled in a restrict sense, so the model should be used in areas where variations in wave height are large within a horizontal scale of a few wave lengths. However, the computation of diraction in arbitrary geophysical conditions is rather complicated and requires considerable computing eort. To avoid this, a phase-decoupled approach, as described in (Holthuijsen et al., 2003), is employed so that same qualitative behaviour of spatial redistribution and changes in wave direction is obtained. This approach, however, does not properly handle diraction in harbours or in front of reecting obstacles. SWAN is stationary and optionally nonstationary and can be applied in Cartesian or curvilinear (recommended only for small scales) or spherical (small scales and large scales) coordinates. The stationary mode should be used only for waves with a relatively short residence time in the computational area under consideration, i.e. the travel time of the waves through the region should be small compared to the time scale of the geophysical conditions (wave boundary conditions, wind, tides and storm surge).

Chapter 1

1.5

Overview

The remainder of this document is subdivided as follows: In Chapter 2 the action balance equations used in SWAN are presented. Next, each source term of the governing equations is in depth described. In Chapter 3 the main characteristics of the nite dierence method for the discretization of the governing equations in irregular horizontal planes are outlined. Various dierencing schemes for spatial propagation are reported. Chapter 4 is concerned with discussing several boundary conditions and their implementation. Chapter 5 is devoted to the design of the two-dimensional wave set-up of sea surface. Chapter 6 is devoted to the linear solvers for the solution of the resulted linear systems of equations. Chapter 7 deals with some consideration on parallelization of SWAN on distributed memory architectures. Chapter 8 presents an unstructured-grid procedure for SWAN. Chapter 9 concludes this document by summarizing the overall solution algorithm of SWAN. This document, however, is not intended as being complete. Although, this document describes the essential steps involved in the simulation of waves, so that the user can see which can be modied or extended to solve a particular problem properly, some issues involved in SWAN are not included. Below, a list of these issues is given, of which the information may be available elsewhere: reection, iterative solvers, overall solution algorithm.

1.6

Acknowledgements

The present SWAN team are grateful to the original authors from the very rst days of SWAN which took place at the Delft University of Technology in Delft, The Netherlands in 1993: Nico Booij, Leo Holthuijsen and Roeland Ris. We further want to acknowledge all contributors who helped us to improve SWAN, reported bugs, and tested SWAN thoroughly: Tim Campbell, John Cazes, IJsbrand Haagsma, Agnieszka Herman, Jim Kaihatu, Kees Kassels, Annette Kieftenburg, Ekaterini Kriezi, Roberto Padilla-Hernandez, Erick Rogers, Kees Vuik, Andre van der Westhuysen and Marcel Zijlema. Many thanks are due to Gerbrant van Vledder who provided the source code XNL for exact computation of four wave-wave interations. It was also the important role which SWAN played in several projects, mostly funded by the Oce of Naval Research (USA), which helped a lot to develop and maintain SWAN. The present version of SWAN is supported by Rijkswaterstaat (as part of the Ministry of

Introduction Transport, Public Works and Water Management, The Netherlands).

We are nally grateful to all those other people working on the Public Domain Software without which the development of SWAN would be unthinkable: Linux, Intel, GNU F95, A L TEX, MPICH2, Perl and many others.

Chapter 1

Chapter 2 Governing equations

2.1 Spectral description of wind waves

Wind generated waves have irregular wave heights and periods, caused by the irregular nature of wind. Due to this irregular nature, the sea surface is continually varying, which means that a deterministic approach to describe the sea surface is not feasible. On the other hand, statistical properties of the surface, like average wave height, wave periods and directions, appear to vary slowly in time and space, compared to typical wave periods and wave lengths. The surface elevation of waves in the ocean, at any location and any time, can be seen as the sum of a large number of harmonic waves, each of which has been generated by turbulent wind in dierent places and times. They are therefore statistically independent in their origin. According to linear wave theory, they remain independent during their journey across the ocean. Under these conditions, the sea surface elevation on a time scale of one hundred characterstic wave periods is suciently well described as a stationary, Gaussian process. The sea surface elevation in one point as a function of time can be described as (t) =
i

ai cos(i t + i )

(2.1)

with the sea surface elevation, ai the amplitude of the ith wave component, i the relative radian or circular frequency of the ith wave component in the presence of the ambient current (equals the absolute radian frequency when no ambient current is present) and i the random phase of the ith wave component. This is called the random-phase model. In the presence of the ambient current, it is assumed that it is uniform with respect to the vertical co-ordinate and the changes in the mean ow within a wave length are so small that they aect only negligibly the dispersion relation. The absolute radian frequency then equals the sum of the relative radian frequency and the multiplication of the wave number and ambient current velocity vectors: =+ku 7 (2.2)

Chapter 2

which is the usual Doppler shift. For linear waves, the relative frequency is given by 2 = gk tanh(kd) where g is the acceleration of gravity and d is the water depth. Ocean waves are chaotic and a description in the time domain is rather limited. Alternatively, many manipulations are more readily described and understood with the variance density spectrum, which is the Fourier transform of the auto-covariance function of the sea surface elevation: + E (f ) = C ( )e2if d (2.4)

(2.3)

with C ( ) =< (t) (t + ) > (2.5) where C ( ) is auto-covariance function, <> represents mathematical expectation of random variable and (t), (t + ) represent two random processes of sea surface elevation, represents the time lag. In the eld of ocean wave theory it is conventional to degne a spectrum E (f ) slightly dierent from the above one: E (f ) = 2E (f ) for f 0 and E (f ) = 0 for f < 0 (2.6)

The description of water waves through the dened variance density spectrum E (f ) is called spectral description of water waves. It can be proved that the variance of the sea surface elevation is given by < 2 >= C (0) =
+ 0

E (f )df

(2.7)

which indicates that the spectrum distributes the variance over frequencies. E (f ) should therefore be interpreted as a variance density. The dimensions of E (f ) are m2 /Hz if the elevation is given in m and the frequencies in Hz . The variance < 2 > is equal to the total energy Etot of the waves per unit surface area if multiplied with a properly chosen coecient: 1 Etot = w g < 2 > 2 (2.8)

The terms variance density spectrum and energy density spectrum will therefore be used indiscriminately in this document. In many wave problems it is not sucient to dene the energy density as a function of frequency alone. It is mostly required to distribute the wave energy over directions as well. This spectrum, which distributes the wave energy over frequencies and directions, will be

Governing equations

denoted with E (f, ). As the total energy density at a frequency f is dstributed over the directions in E (f, ), it follows that: E (f ) =
2 0

E (f, )d

(2.9)

The energy density spectrum E (f ) and E (f, ) are depicted in Figure 2.1. Based on the energy density spectrum, the integral wave parameters can be obtained. These parameters

E ( f , )
N storm swell wind sea two-dimensional

f
NORTH

North Sea

f
local breeze

E( f )

wind sea

swell one-dimensional

Figure 2.1: Illustrations of 1D and 2D wave spectra. (Reproduced from Holthuijsen (2007) with permission of Cambridge University Press.) can be expressed in terms of the so-called nth moment of the energy density spectrum: mn =
0

f n E (f )df

(2.10)

So, the variance of the sea surface elevation is given by m0 =< 2 >. Well-known parameters are the signicant wave height: (2.11) Hs = 4 m 0 and some wave periods: Tm01 = m0 , m1 Tm02 = m0 , m2 Tm10 = m 1 m0 (2.12)

10

Chapter 2

In SWAN, the energy density spectrum E (, ) is generally used. On a larger scale the spectral energy density function E (, ) becomes a function of space and time and wave dynamics should be considered to determine the evolution of the spectrum in space and time. For brevity, the notation E (, ) will still be used.

2.2
2.2.1

Propagation of wave energy

Wave kinematics

Using the linear wave theory and the conversion of wave crests, the wave propagation velocities in spatial space within Cartesian framework and spectral space can be obtained from the kinematics of a wave train (Whitham, 1974; Mei, 1983): 1 2| k | d k dx = (cx , cy ) = cg + u = 1 + 2 +u dt 2 sinh(2|k |d) |k | d = c = dt d 1 d = c = dt k d u + u x d c g k t s u d +k d m m

(2.13)

where cx , cy are the propagation velocities of wave energy in spatial x, y space, c and c are the propagation velocities in spectral space , space, d is water depth, s is the space co-ordinate in the wave propagation direction of and m is a co-ordinate perpendicular to s. The expression for c is presented here without diraction eects. These are treated separately in Section 2.6.3. Furthermore, k = (kx , ky ) = (|k | cos , |k | sin ) , u = ( ux , u y ) (2.14) In addition, the operator d/dt denotes the total derivative along a spatial path of energy propagation, and is dened as d = + ( c g + u) x dt t (2.15)

2.2.2

Spectral action balance equation

All information about the sea surface is contained in the wave variance spectrum or energy density E (, ), distributing wave energy over (radian) frequencies (as observed in a frame of reference moving with current velocity) and propagation directions (the direction normal to the wave crest of each spectral component). Usually, wave models determine the evolution of the action density N (x, t; , ) in space x and time t. The action density

Governing equations

11

is dened as N = E/ and is conserved during propagation in the presence of ambient current, whereas energy density E is not (Whitman, 1974). It is assumed that the ambient current is uniform with respect to the vertical co-ordinate and is denoted as U . The evolution of the action density N is governed by the action balance equation, which reads (e.g., Mei, 1983; Komen et al., 1994): N c N c N Stot + x [(cg + U )N ] + + = (2.16) t The left-hand side is the kinematic part of this equation. The second term denotes the propagation of wave energy in two-dimensional geographical x-space, with the group velocity cg = / k following from the dispersion relation 2 = g |k | tanh(|k |d) where k is the wave number vector and d the water depth. The third term represents the eect of shifting of the radian frequency due to variations in depth and mean currents. The fourth term represents depth-induced and current-induced refraction. The quantities c and c are the propagation velocities in spectral space (, ). The right-hand side contains Stot , which is the source/sink term that represents all physical processes which generate, dissipate, or redistribute wave energy. They are dened for energy density E (, ). Details are given in Section 2.3. Equation (2.16) can be recasted in Cartesian or spherical co-ordinates. For small scale applications the spectral action balance equation may be expressed in Cartesian co-ordinates as given by N cx N cy N c N c N Stot + + + + = (2.17) t x y With respect to applications at shelf sea or oceanic scales the action balance equation may be recasted in spherical co-ordinates as follows: c N c N c N c N Stot N + + + + = t (2.18)

with respect to longitude and latitude . Note that is the with action density N wave direction taken counterclockwise from geographic East. The propagation velocities are reformulated as follows. On a sphere, we have dx = R cos d dy = Rd (2.19)

with R the radius of the earth. The propagation velocities in geographic space are then given by d 1 2|k |d |k | cos 1 1 + = c = + u 2 dt R cos 2 sinh(2|k |d) |k | 1 1 2|k |d |k | sin d = c = 1 + + u 2 dt R 2 sinh(2|k |d) |k |

(2.20)

12

Chapter 2

with u and u the ambient currents in longitude and latitude direction, respectively. The propagation velocity in space remain unchanged. To rewrite the propagation velocity c in terms of spherical co-ordinates, we use the so-called Clairauts equation that states that on any geodesic, the following expression holds: R cos cos = constant (2.21)

Dierentiation of Eq. (2.21) with respect to a space co-ordinate s in wave direction gives R sin cos d d R cos sin = 0 ds ds (2.22)

Since, dy = ds sin , we have d/ds = sin /R. Substitution into Eq. (2.22) and using ds = (cx cos + cy sin )dt yields d cx cos + cy sin = cos tan dt R (2.23)

This term (2.23) accounts for the change of propagation direction relative to true North when travelling along a great circle. This holds for deep water and without currents. Hence, cx cos + cy sin cos tan (2.24) c = c R is related to the action density N in a local Cartesian frame (x, y ) through In Eq. (2.18), N dddd = N dddxdy , or N = N R2 cos . Substitution into (2.18) yields: N N c N c cos N c N c N Stot + + cos1 + + = t (2.25)

2.3

Sources and sinks

First, in Section 2.3.1 general concepts of the physical processes of generation, dissipation and nonlinear wave-wave interactions that are implemented in SWAN are outlined. Next, complete expressions for these physical processes are given in subsequent sections. Finally, for completeness, the rst- and second-generation formulations as employed in SWAN are outlined in Section 2.3.5.

2.3.1

General concepts

In shallow water, six processes contribute to Stot : Stot = Sin + Snl3 + Snl4 + Sds,w + Sds,b + Sds,br . (2.26)

These terms denote, respectively, wave growth by the wind, nonlinear transfer of wave energy through three-wave and four-wave interactions and wave decay due to whitecapping,

Governing equations

13

bottom friction and depth-induced wave breaking. First, a brief summary of the formulations is given below. Next, for each term complete expressions are outlined. Wind input Transfer of wind energy to the waves is described with a resonance mechanism (Phillips, 1957) and a feed-back mechanism (Miles, 1957). Resonance with wind-induced pressure uctations The pressure distribution induced by wind at the sea surface is random. It propagates more or less a frozen pattern over the surface with wind speed. This can be Fourier transformed to produce harmonic pressure waves that propagate with wind speed. If this harmonic pressure wave remains in phase with a free harmonic surface wave, then the wind energy is transferred from the pressure wave to the surface wave. The energy input by this mechanism, which contributes to the initial stages of wave growth, varies linearly with time. Feedback of wave-induced pressure uctations When a wave has been generated by the resonance mechanism as explained above, it will distort the wind prole just above the water surface. This distortion results in an over pressure on the wind ward side of the crest of the wave and an under pressure at the lee side of the crest. It means that when the sea surface moves up and down, the pressure also follows the same movements, therefore transfer energy to the wave. This energy transfer is proportional to the energy in the wave itself, so the wave grows more as it gets larger. This eect is found to be exponential in time. Based on the two wave growth mechanisms, wave growth due to wind commonly described as the sum of linear and exponential growth term of a wave component: Sin (, ) = A + BE (, ) (2.27)

in which A and B depend on wave frequency and direction, and wind speed and direction. The eects of currents are accounted for by using the apparent local wind speed and direction. The expression for the term A is due to Cavaleri and Malanotte-Rizzoli (1981) with a lter to avoid growth at frequencies lower than the Pierson-Moskowitz frequency (Tolman, 1992a). Two optional expressions for the coecient B are used in the SWAN model. The rst is taken from an early version of the WAM Cycle 3 model (the WAMDI group, 1988). It is due to Snyder et al. (1981), rescaled in terms of friction velocity U by Komen et al. (1984). The drag coecient to relate U to the driving wind speed at 10 m elevation U10 is taken from Wu (1982). The second expression for B in SWAN is taken from the WAM Cycle 4 model (Komen et al., 1994). It is due to Janssen (1991a) and it accounts explicitly for the interaction between the wind and the waves by considering atmospheric boundary layer eects and the roughness length of the sea surface. The corresponding set of equations is solved (as in the WAM model) with the iterative procedure of Mastenbroek et al. (1993).

14 Dissipation

Chapter 2

The dissipation term of wave energy is represented by the summation of three dierent contributions: whitecapping Sds,w , bottom friction Sds,b and depth-induced breaking Sds,br . Whitecapping is primarily controlled by the steepness of the waves. In presently operating third-generation wave models, the whitecapping formulations are based on a pulse-based model (Hasselmann, 1974), as adapted by the WAMDI group (1988): k Sds,w (, ) = E (, ) k (2.28)

denote a where is a steepness dependent coecient, k is wave number and and k mean frequency and a mean wave number, respectively (cf. the WAMDI group, 1988). Komen et al. (1984) estimated the value of by closing the energy balance of the waves in fully developed conditions. This implies that this value depends on the wind input formulation that is used. Since two expressions are used for the wind input in SWAN, also two values for are used. The rst is due to Komen et al. (1984), as in WAM Cycle 3. The second expression is an adaptation of this expression based on Janssen (1991a), as in WAM Cycle 4 (see Janssen, 1991b; G unther et al., 1992). Young and Banner (1992) and Banner and Young (1994) have shown that the results of closing the energy balance in this manner depend critically on the choice of a high-frequency cut-o frequency above which a diagnostic spectral tail is used. In SWAN, this cut-o frequency is dierent from the one used in the WAM model. Dierences in the growth rates between the WAM model and SWAN are therefore to be expected. A number of alternative whitecapping expressions have been proposed to improve the accuracy of SWAN. These range from alternative calibrations of the Komen et al (1984) expression, e.g. Rogers et al (2003), to alternative ways of calculating mean spectral steepness, e.g. Van Vledder and Hurdle (2002). In SWAN, another alternative is presented. This alternative is proposed by Van der Westhuysen et al (2007) and Van der Westhuysen (2007), based on the whitecapping expression of Alves and Banner (2003). This expression is based on experimental ndings that whitecapping dissipation appears to be related to the nonlinear hydrodynamics within wave groups. This yields a dissipation term that primarily depends on quantities that are local in the frequency spectrum, as opposed to ones that are distributed over the spectrum, as in the expression of Komen et al (1984). However, the nal whitecapping expression proposed by Alves and Banner (2003) features additional dependencies on the spectral mean wavenumber and steepness, which is problematic in situations of mixed sea and swell often encountered in the nearshore. Therefore, their whitecapping expression is applied in Van der Westhuysen (2007) without these mean spectral dependencies. This adapted whitecapping expression is used together with a wind input term that is based on that of Yan (1987).

Governing equations

15

In shallow water the orbital motions of the water particles, induced by surface waves, extend down to the sea oor. This gives rise to an interaction between te surface waves and the bottom. An overview of dierent wave-bottom interaction mechanisms and of their relative strengths is given by Shemdin et al. (1978). They are: scattering on bottom irregularities, motion of a soft bottom, percolation into a porous bottom and friction in the turbulent bottom boundary layer. The rst process results in a local redistribution of wave energy by scattering of wave components. The last three are dissipative. Their strength depends on the bottom conditions. For continental shelf seas with sandy bottoms, the dominant mechanism appears to be bottom friction (Bertotti and Cavaleri, 1994) which can generally be expressed as: Sds,b = Cb 2 E (, ) g 2 sinh2 kd (2.29)

in which Cb is a bottom friction coecient. A large number of models has been proposed since the pioneering paper of Putnam and Johnson (1949). Hasselmann et al. (1973) suggested to use an empirically obtained constant. It seems to perform well in many different conditions as long as a suitable value is chosen (typically dierent for swell and wind sea). A nonlinear formulation based on drag has been proposed by Hasselmann and Collins (1968) which was later simplied by Collins (1972). More complicated, eddy viscosity models have been developed by Madsen et al. (1988) and by Weber (1989, 1991a, 1991b). Considering the large variations in bottom conditions in coastal areas (bottom material, bottom roughness length, ripple height, etc.), there is no eld data evidence to give preference to a particular friction model (Luo and Monbaliu, 1994). For this reason, the simplest of each of these types of friction models has been implemented in SWAN: the empirical JONSWAP model of Hasselmann et al. (1973), the drag law model of Collins (1972) and the eddy-viscosity model of Madsen et al. (1988). The eect of a mean current on the wave energy dissipation due to bottom friction is not taken into account in SWAN. The reasons for this are given by Tolman (1992b) who argues that state-of-the-art expressions vary too widely in their eects to be acceptable. He found that the error in nding a correct estimate of the bottom roughness length scale has a much larger impact on the energy dissipation rate than the eect of a mean current. When waves propagate towards shore, shoaling leads to an increase in wave height. When the ratio of wave height over water depth exceeds a certain limit, waves start to break, thereby dissipating energy rapidly. In extreme shallow water (surf zone), this process becomes dominant over all other processes. The process of depth-induced wave breaking is still poorly understood and little is known about its spectral modelling. In contrast to this, the total dissipation (i.e. integrated over the spectral space) due to this type of wave breaking can be well modelled with the dissipation of a bore applied to the breaking waves in a random eld (Battjes and Janssen, 1978; Thornton and Guza, 1983). Laboratory observations (e.g., Battjes and Beji, 1992; Vincent et al. 1994; Arcilla et al., 1994 and Eldeberky and Battjes, 1996) show that the shape of initially uni-modal spectra propagating across simple (barred) beach proles, is fairly insensitive to depth-induced breaking. This

16

Chapter 2

has led Eldeberky and Battjes (1995) to formulate a spectral version of the bore model of Battjes and Janssen (1978) that conserves the spectral shape. Expanding their expression to include directions, the expression reads: Sds,br (, ) = Dtot E (, ) Etot (2.30)

in which Etot is the total wave energy and Dtot < 0 is the rate of dissipation of the total energy due to wave breaking according to Battjes and Janssen (1978). The value of Dtot depends critically on the breaker parameter = Hmax /d (in which Hmax is the maximum possible individual wave height in the local water depth d). In SWAN, both a constant value and a variable value are available. Examples of a variable breaker parameter can be found in Nelson (1987) and Ruessink et al. (2003). (Both are implemented in SWAN.) The constant value is = 0.73 found as the mean value of the data set of Battjes and Stive (1985). Nonlinear wave-wave interactions The basic properties of wave-wave interactions were discovered during the fundamental research of Phillips (1960) and Hasselmann (1960, 1962, 1963a,b). The physical meaning of the interactions is that resonant sets of wave components exchange energy, redistributing energy over the spectrum. In deep and intermediate water, four-wave interactions (so-called quadruplets) are important, whereas in shallow water three-wave interactions (so-called triads) become important. In deep water, quadruplet wave-wave interactions dominate the evolution of the spectrum. They transfer wave energy from the spectral peak to lower frequencies (thus moving the peak frequency to lower values) and to higher frequencies (where the energy is dissipated by whitecapping). In very shallow water, triad wave-wave interactions transfer energy from lower frequencies to higher frequencies often resulting in higher harmonics (Beji and Battjes, 1993). Low-frequency energy generation by triad wave-wave interactions is not considered here. A full computation of the quadruplet wave-wave interactions is extremely time consuming and not convenient in an operational wave model. Nevertheless, SWAN has an option to compute the Boltzmann integral in an exact manner. The approach is the exact method developed by Webb, Tracy and Resio (WRT) (Resio et al., 2001). This algorithm was reprogrammed by Van Vledder, bearing the name XNL (Van Vledder and Bottema, 2003). This method is also enable to capture the frequency shift and the spectral shape changes as water depth decreases. A number of techniques, based on parametric methods and approximations have been proposed to improve computational speed of computing quadruplets (see Young and Van Vledder (1993) for a review). In SWAN, the computations are carried out with the Discrete Interaction Approximation (DIA) of Hasselmann et al. (1985). This DIA has been found

Governing equations

17

to be quite successful in describing the essential features of a developing wave spectrum; see Komen et al. (1994). For uni-directional waves, this approximation is not valid. In fact, the quadruplet interaction coecient for these waves is nearly zero. For nite-depth applications, Hasselmann and Hasselmann (1981) have shown that for a JONSWAP-type spectrum the quadruplet wave-wave interactions can be scaled with a simple expression. In some cases, the DIA technique may not be accurate enough. In Hashimoto et al. (2002), it was demonstrated that the accuracy of the DIA may be improved by increasing the number of quadruplet congurations. They proposed a Multiple DIA with up to 6 wave number congurations. In very shallow water, triad wave interactions become important for steep waves. It transfers energy to higher frequencies, resulting in higher harmonics (Beji and Battjes, 1993). The energy transfer in this process can take place over relatively short distance and can dramatically change single peaked spectra into multiple peaked spectra, which has frequently been observed in the eld (Arcilla et al., 1994) and in a number of laboratory experiments with a bar-trough prole (Beji and Battjes, 1993) and a plane beach prole (Nwogu, 1994). A rst attempt to describe triad wave-wave interactions in terms of a spectral energy source term was made by Abreu et al. (1992). However, their expression is restricted to non-dispersive shallow water waves and is therefore not suitable in many practical applications of wind waves. The breakthrough in the development came with the work of Eldeberky and Battjes (1995) who transformed the amplitude part of the Boussinesq model of Madsen and Srensen (1993) into an energy density formulation and who parameterized the biphase of the waves on the basis of laboratory observations (Battjes and Beji, 1992; Arcilla et al., 1994). A discrete triad approximation (DTA) for co-linear waves was subsequently obtained by considering only the dominant self-self interactions. Their model has been veried with ume observations of long-crested, random waves breaking over a submerged bar (Beji and Battjes, 1993) and over a barred beach (Arcilla et al., 1994). The model appeared to be fairly successful in describing the essential features of the energy transfer from the primary peak of the spectrum to the super harmonics. A slightly dierent version, the so-called Lumped Triad Approximation (LTA) was later derived by Eldeberky (1996). This LTA technique is employed in SWAN.

2.3.2

Input by wind (Sin )

Sin (, ) = A + BE (, ) (2.31)

Wave growth by wind is described by: in which A describes linear growth and BE exponential growth. It should be noted that the SWAN model is driven by the wind speed at 10m elevation U10 whereas it uses the friction velocity U . For the WAM Cycle 3 formulation the transformation from U10 to U is obtained with 2 2 U = CD U10 (2.32)

18

Chapter 2

in which CD is the drag coecient from Wu (1982): CD (U10 ) = 1.2875 103 , for U10 < 7.5m/s (0.8 + 0.065s/m U10 ) 103 , for U10 7.5m/s (2.33)

For the WAM Cycle 4 formulations, the computation of U is an integral part of the source term. Linear growth by wind For the linear growth term A, the expression due to Cavaleri and Malanotte-Rizzoli (1981) is used with a lter to eliminate wave growth at frequencies lower than the PiersonMoskowitz frequency (Tolman, 1992a)1 : 0.13g 2 28U (2.34) in which w is the wind direction, H is the lter and PM is the peak frequency of the fully developed sea state according to Pierson and Moskowitz (1964) as reformulated in terms of friction velocity. A= H = exp (
PM

1.5 103 (U max[0, cos( w )])4 H , 2g 2

) 4 ,

PM =

Exponential growth by wind Two expressions for exponential growth by wind are optionally available in the SWAN model. The rst expression is due to Komen et al. (1984). Their expression is a function of U /cph : a U B = max[0, 0.25 (28 cos( w ) 1)] (2.35) w cph in which cph is the phase speed and a and w are the density of air and water, respectively. This expression is also used in WAM Cycle 3 (the WAMDI group, 1988). The second expression is due to Janssen (1989,1991a). It is based on a quasi-linear wind-wave theory and is given by: 2 a U max[0, cos( w )]2 (2.36) B= w cph where is the Miles constant. In the theory of Janssen (1991a), this constant is estimated from the non-dimensional critical height :
=
1

1.2 ln4 2 gze r e c2 ph

, 1 r = c/|U cos( w )|

(2.37)

In Eq. (10) of Tolman (1992a) the power of 105 should be 103 ; H. Tolman, personal communication, 1995.

Governing equations

19

where = 0.41 is the Von Karman constant and ze is the eective surface roughness. If the non-dimensional critical height > 1, the Miles constant is set equal 0. Janssen (1991a) assumes that the wind prole is given by: U (z ) = U z + ze z0 ln[ ] ze (2.38)

in which U (z ) is the wind speed at height z (10m in the SWAN model) above the mean water level, z0 is the roughness length. The eective roughness length ze depends on the roughness length z0 and the sea state through the wave-induced stress w and the total surface stress = a |U |U : ze = z0 1
|w | | |

z0 =

2 U g

(2.39)

The second of these two equations is a Charnock-like relation in which is a constant equal to 0.01. The wave stress w is given by: w = w
2 0 0

k BE (, ) dd k

(2.40)

The value of U can be determined for a given wind speed U10 and a given wave spectrum E (, ) from the above set of equations. In the SWAN model, the iterative procedure of Mastenbroek et al. (1993) is used. This set of expressions (2.36) through (2.40) is also used in WAM Cycle 4 (Komen et al., 1994).

2.3.3

Dissipation of wave energy (Sds )

Whitecapping: Komen et al (1984) formulation The processes of whitecapping in the SWAN model is represented by the pulse-based model of Hasselmann (1974). Reformulated in terms of wave number (rather than frequency) so as to be applicable in nite water depth (cf. the WAMDI group, 1988), this expression is: k Sds,w (, ) = E (, ) k (2.41)

denote the mean frequency and the mean wave number, respectively, and the where and k coecient depends on the overall wave steepness. This steepness dependent coecient, as given by the WAMDI group (1988), has been adapted by G unther et al. (1992) based on Janssen (1991a) (see also (Janssen, 1991b)): s k = KJ = Cds ((1 ) + ) s PM k
p

(2.42)

For = 0 the expression of reduces to the expression as used by the WAMDI group (1988). The coecients Cds , and p are tunable coecients, s is the overall wave steepness, s PM

20

Chapter 2 3.02 103 . The (2.43)

is the value of s for the Pierson-Moskowitz spectrum (1964): s P M = overall wave steepness s is dened as: Etot s = k

and the total wave energy Etot are dened The mean frequency , the mean wave number k as (cf. the WAMDI group, 1988):
1 = Etot 2 0 2 0 0 2 0 0 0

1 E (, )dd

(2.44)
2

= E 1 k tot

1 E (, )dd k

(2.45) (2.46)

Etot =

E (, )dd

The values of the tunable coecients Cds and and exponent p in this model have been obtained by Komen et al. (1984) and Janssen (1992) by closing the energy balance of the waves in idealized wave growth conditions (both for growing and fully developed wind seas) for deep water. This implies that coecients in the steepness dependent coecient depend on the wind input formulation that is used. Since two dierent wind input formulations are used in the SWAN model, two sets of coecients are used. For the wind input of Komen et al. (1984; corresponding to WAM Cycle 3; the WAMDI group, 1988): Cds = 2.36 105 , = 0 and p = 4. Janssen (1992) and also G unther et al. (1992) 5 obtained (assuming p = 4) Cds = 4.10 10 and = 0.5 (as used in the WAM Cycle 4; Komen et al., 1994). It is well-known that SWAN underestimates structurally the mean (or peak) wave periods by 10 to 20%. This has also been observed in the SWAN hindcasts as described by Rogers et al. (2003). Investigations of Rogers et al. (2003) showed that adjusting the parameter from 0 to 1 leads to an improved prediction of the wave energy at lower frequencies. Because of this, is set to 1 as default since version 40.91A. However, it should be mentioned that adapting without retuning Cds may lead to an overestimation of the fetch and duration unlimited wave heights as reported by Pierson and Moskowitz (1964). Whitecapping: saturation-based model An alternative description for whitecapping in SWAN is given by Van der Westhuysen et al (2007) and Van der Westhuysen (2007), which is an adapted form of the expression of Alves and Banner (2003). The latter is based on the apparent relationship between wave groups and whitecapping dissipation. This adaption is due to the fact that it can also be applied to mixed sea-swell conditions and in shallow water. This was done by removing the dependencies on mean spectral steepness and wavenumber in the original expression, and by applying source term scaling arguments for its calibration (see below). This led to the

Governing equations following expression for whitecapping dissipation Sds,break (, ) =

Cds

21

B (k ) Br

p/2

(tanh(kh))(2p0 )/4 gkE (, )

(2.47)

in which the density function B (k ) is the azimuthal-integrated spectral saturation, which is positively correlated with the probability of wave group-induced breaking. It is calculated from frequency space variables as follows B (k ) =
2 0

cg k 3 E (, )d

(2.48)

and Br = 1.75 103 is a threshold saturation level. The proportionality coecient is set to Cds = 5.0 105 . When B (k ) > Br , waves break and the exponent p is set equal to a calibration parameter p0 . For B (k ) Br there is no breaking, but some residual dissipation proved necessary. This is obtained by setting p = 0. A smooth transition between these two situations is achieved by (Alves and Banner, 2003); p0 p0 B (k ) p= + tanh 10 1 2 2 Br

(2.49)

In Van der Westhuysen (2007) the dissipation modes of breaking and non-breaking waves are separated, so that they are active over dierent parts of the spectrum: Sds,w (, ) = fbr ( )Sds,break + [1 fbr ( )] Sds,nonbreak , (2.50)

where Sds,break is the contribution by breaking waves (2.47), and Sds,nonbreak dissipation by means other than breaking (e.g. turbulence). The changeover between the two modes is made with a smooth transition function fbr similar to (2.49): 1 1 B (k ) fbr ( ) = + tanh 10 1 2 2 Br

(2.51)

Since relatively little is known about the dissipation mechanisms of the non-breaking lowfrequency waves, their dissipation is not modelled in detail. Instead, the expression (2.41) is used for Sds,nonbreak , to provide general background dissipation of non-breaking waves. For this component, the parameter settings of Komen et al. (1984) are applied. The wind input expression used in saturation-based model is based on that by Yan (1987). This expression embodies experimental ndings that for strong wind forcing, u /c > 0.1 say, the wind-induced growth rate of waves depends quadratically on u /c (e.g. Plant 1982), whereas for weaker forcing, u /c < 0.1 say, the growth rate depends linearly on u /c (Snyder et al, 1981). Yan (1987) proposes an analytical t through these two ranges of the form: t = D u c
2

cos( ) + E

u cos( ) + F cos( ) + H c

(2.52)

22

Chapter 2

where D,E ,F and H are coecients of the t. Yan imposed two constraints: t Snyder and
u /c

for

U5 u 1 (or 0.036) c c

(2.53)

lim t = Plant

(2.54)

in which Snyder and Plant are the growth rates proposed by Snyder et al (1981) and Plant (1982), respectively. Application of Eqs. (2.53) and (2.54) led us to parameter values of D = 4.0 102 ,E = 5.52 103 ,F = 5.2 105 and H = 3.02 104 , which are somewhat dierent from those proposed by Yan (1987). We found that our parameter values produce better fetch-limited simulation results in the Pierson and Moskowitz (1964) fetch range thant the original values of Yan (1987). Finally, the choice of the exponent p0 in Eqs. (2.47) and (2.49) is made by requiring that the source terms of whitecapping (Eq. 2.47) and wind input (Eq. 2.52) have equal scaling in frequency, after Resio et al (2004). This leads to a value of p0 = 4 for strong wind forcing (u /c > 0.1) and p0 = 2 for weaker forcing (u /c < 0.1). A smooth transition between these two limits, centred around u /c = 0.1, is achieved by the expression p0 ( ) = 3 + tanh w u 0. 1 c (2.55)

where w is a scaling parameter for which a value of w = 26 is used in SWAN. In shallow water, under strong wind forcing (p0 = 4), this scaling condition requires the additional dimensionless factor tanh(kh)1/2 in Eq. (2.47), where h is the water depth. Bottom friction The bottom friction models that have been selected for SWAN are the empirical model of JONSWAP (Hasselmann et al., 1973), the drag law model of Collins (1972) and the eddy-viscosity model of Madsen et al. (1988). The formulations for these bottom friction models can all be expressed in the following form: Sds,b = Cb 2 E (, ) g 2 sinh2 kd (2.56)

in which Cb is a bottom friction coecient that generally depends on the bottom orbital motion represented by Urms :
2 Urms = 2 0 0

2 E (, )dd sinh2 kd

(2.57)

Hasselmann et al. (1973) found Cb = CJON = 0.038m2 s3 which is in agreement with the JONSWAP result for swell dissipation. However, Bouws and Komen (1983) suggest a value of CJON = 0.067m2 s3 for depth-limited wind-sea conditions in the North Sea. This value

Governing equations

23

is derived from revisiting the energy balance equation employing an alternative deep water dissipation. Both values are available in SWAN. The expression of Collins (1972) is based on a conventional formulation for periodic waves with the appropriate parameters adapted to suit a random wave eld. The dissipation rate is calculated with the conventional bottom friction formulation of Eq. (2.56) in which the bottom friction coecient is Cb = Cf gUrms with Cf = 0.015 (Collins, 1972)2 . Madsen et al. (1988) derived a formulation similar to that of Hasselmann and Collins (1968) but in their model the bottom friction factor is a function of the bottom roughness height and the actual wave conditions. Their bottom friction coecient is given by: g Cb = fw Urms (2.58) 2 in which fw is a non-dimensional friction factor estimated by using the formulation of Jonsson (1966) cf. Madsen et al. (1988): 1 a 1 + log10 ( ) = mf + log10 ( b ) KN 4 fw 4 fw (2.59)

in which mf = 0.08 (Jonsson and Carlsen, 1976) and ab is a representative near-bottom excursion amplitude: 2 1 a2 E (, )dd (2.60) b = 2 0 0 sinh2 kd and KN is the bottom roughness length scale. For values of ab /KN smaller than 1.57 the friction factor fw is 0.30 (Jonsson, 1980). Depth-induced wave breaking To model the energy dissipation in random waves due to depth-induced breaking, the borebased model of Battjes and Janssen (1978) is used in SWAN. The mean rate of energy dissipation per unit horizontal area due to wave breaking Dtot is expressed as: H2 1 2 = BJ Qb max Dtot = BJ Qb ( )Hmax 4 2 8 = 1 in SWAN, Qb is the fraction of breaking waves determined by: 1 Qb Etot = 8 2 ln Qb Hmax (2.61)

in which BJ

(2.62)

in which Hmax is the maximum wave height that can exist at the given depth and is a mean frequency dened as:
1 = Etot
2

2 0 0

E (, )dd

(2.63)

Collins (1972) contains an error in the expression due to an erroneous Jacobian transformation. See page A-16 of Tolman (1990).

24

Chapter 2

The fraction of depth-induced breakers (Qb ) is determined in SWAN with

0,

for 0.2
2

Qb =

0 exp (Q0 1)/ Q 2 Q , for 0.2 < < 1 2 exp (Q0 1)/ 2 0

(2.64)

1,

where = Hrms /Hmax . Furthermore, for 0.5, Q0 = 0 and for 0.5 < 1, Q0 = (2 1)2 . Extending the expression of Eldeberky and Battjes (1995) to include the spectral directions, the dissipation for a spectral component per unit time is calculated in SWAN with: Dtot BJ Qb Sds,br (, ) = E (, ) (2.65) E (, ) = 2 Etot The maximum wave height Hmax is determined in SWAN with Hmax = d, in which is the breaker parameter and d is the total water depth (including the wave-induced set-up if computed by SWAN). In the literature, this breaker parameter is often a constant or it is expressed as a function of bottom slope or incident wave steepness (see e.g., Galvin, 1972; Battjes and Janssen, 1978; Battjes and Stive, 1985; Arcilla and Lemos, 1990; Kaminsky and Kraus, 1993; Nelson, 1987, 1994). In the publication of Battjes and Janssen (1978) in which the dissipation model is described, a constant breaker parameter, based on Miches criterion, of = 0.8 was used. Battjes and Stive (1985) re-analyzed wave data of a number of laboratory and eld experiments and found values for the breaker parameter varying between 0.6 and 0.83 for dierent types of bathymetry (plane, bar-trough and bar) with an average of 0.73. From a compilation of a large number of experiments Kaminsky and Kraus (1993) have found breaker parameters in the range of 0.6 to 1.59 with an average of 0.79. An alternative to the bore-based model of Battjes and Janssen (1978) is proposed by Thornton and Guza (1983). This model can be regarded as an alteration of Battjes and Janssen with respect to the description of the wave height probability density function. The total dissipation due to depth-induced breaking is formulated as B3 3 H pb (H ) dH (2.66) 8d 0 in which B is a proportionality coecient and pb (H ) is the probability density function of breaking waves times the fraction of breakers, Qb . Based on eld observations, the wave heights in the surf zone are assumed to remain Rayleigh distributed, even after breaking. This implies that all waves will break, not only the highest as assumed by Battjes and Janssen (1978). The function pb (H ) is obtained by multiplying the Rayleigh wave height probability density function p(H ), given by Dtot = 2H H p(H ) = 2 exp Hrms Hrms
2

for 1

(2.67)

Governing equations by a weighting function W (H ) dened so that 0 W (H ) 1, to yield pb ( H ) = W ( H ) p( H )

25

(2.68)

Thornton and Guza (1983) proposed the following weighting function in which the fraction of breaking waves is independent of the wave height: W ( H ) = Qb = Hrms d
n

(2.69)

with a calibration parameter n(=4) and a breaker index (not to be confused with the Battjes and Janssen breaker index!). The integral in expression (2.66) can then be simplied, as follows: 3 3 Qb Hrms (2.70) H 3 pb (H ) dH = Qb H 3 p(H ) dH = 4 0 0 Hence, Dtot = 3B 3 3 Qb Hrms 32 d (2.71)

2.3.4

Nonlinear wave-wave interactions (Snl )

Quadruplets In this section two methods are described for the computation of nonlinear interactions at deep water. The rst method is called the DIA method and is relatively crude in the approximation of the Boltzmann integral. The second one is called the XNL approach and is implemented in SWAN by G. Ph. van Vledder. DIA The quadruplet wave-wave interactions are computed with the Discrete Interaction Approximation (DIA) as proposed by Hasselmann et al. (1985). Their source code (slightly adapted by Tolman, personal communication, 1993) has been implemented in the SWAN model. In the DIA two quadruplet wave number congurations are considered, both with frequencies: 1 = 2 = , 3 = (1 + ) = + , 4 = (1 ) = (2.72) where is a coecient with a default value of 0.25. To satisfy the resonance conditions for the rst quadruplet, the wave number vectors with frequency 3 and 4 lie at an angle of 3 = 11.48o and 4 = 33.56o to the angle of the wave number vectors with frequencies 1 and 2 . The second quadruplet is the mirror image of the rst quadruplet with relative angles of 3 = + = 11.48o and 4 = = 33.56o . An example of this wave number conguration is shown in Figure 2.2. See Van Vledder et al. (2000) for further information about wave number congurations for arbitrary values of .

26

Chapter 2

2 1.5 1 0.5 0 0.5 1 1.5 2 1

= 0.25 3 = 11.48 4 = 33.56

k (1/m)
k3 k
1

0.08 0.06 0.04 0.02

k4 k
2

0 0.02 0.04 0.06 0.08

0.5

0.5

0.02 0.04 0.06 0.08 kx (1/m)

0.1

Figure 2.2: Wave number conguration for =0.25 and its position in a discrete frequencydirection spectrum (from Van Vledder et al., 2000). Within this discrete interaction approximation, the source term Snl4 (, ) for the nonlinear transfer rate is given by:
Snl4 (, ) = Snl4 (, ) + Snl4 (, )

(2.73)

where Snl4 refers to the rst quadruplet and Snl4 to the second quadruplet (the expressions for Snl4 are identical to those for Snl4 for the mirror directions). The DIA exchanges wave variance at all three wave number vectors involved in a quadruplet wave number conguration. The rate of change of wave variance due to the quadruplet interaction at the three frequency-direction bins can be written as: 2 Snl4 (, ) 2 4 + + Snl4 ( , ) = 1 Cnl4 (2 ) g 2 1 Snl4 ( , )

11

where Cnl4 = 3 107 by default. Eq. (2.74) conserves wave variance, momentum and action when the frequencies are geometrically distributed (as is the case in the SWAN model).

E ( + , + ) E ( , ) + (1 + )4 (1 )4 E (, )E ( + , + )E ( , ) 2 (1 2 )4 E 2 (, )

(2.74)

Governing equations

27

The wave variance density at the frequency-direction bins E ( + , + ) and E ( , ) is obtained by bi-linear interpolation between the four surrounding frequency-direction bins. Similarly, the rate of change of variance density is distributed between the four surrounding bins using the same weights as used for the bi-linear interpolation. In the DIA algorithm, Eq. (2.74) (and its mirror image) is applied to all spectral bins in a discrete frequency-direction spectrum. Figure 2.2 shows an example of one wave number conguration and its mirror image in a discrete spectrum. An extended spectral grid is applied to compute the interactions in the frequency range aected by the parametric spectral tail. SWAN has an option to replace bi-linear interpolation of wave variance density using the nearest bin approach using a weight equal to 1. Following the WAM group (WAMDI, 1988), the quadruplet interaction in shallow water with depth d is obtained by multiplying the deep water nonlinear transfer rate with a scaling factor R(kp d): nite depth deep water Snl4 = R(kp d)Snl4 (2.75) where R is given by: R(kp d) = 1 + Csh1 (1 Csh2 kp d)eCsh3 kp d kp d (2.76)

in which kp is the peak wave number of the frequency spectrum. WAMDI (1988) proposes the following values of the coecients: Csh1 = 5.5, Csh2 = 5/6 and Csh3 = 5/4. In the shallow water limit, i.e., kp 0 the nonlinear transfer rate tends to innity. Therefore, a lower limit of kp = 0.5 is applied, resulting in a maximum value of R(kp d) = 4.43. To increase the model robustness in case of arbitrarily shaped spectra, the peak wave number (cf. Komen et al., 1994). kp is replaced by kp = 0.75k XNL (G. Ph. van Vledder)

The second method for calculating the nonlinear interactions in SWAN is the so-called Webb-Resio-Tracy method (WRT), which is based on the original six-dimensional Boltzmann integral formulation of Hasselmann (1962, 1963a,b), and additional considerations by Webb (1978), Tracy and Resio (1982) and Resio and Perrie (1991). A detailed description of the WRT method and its implementation in discrete spectral wave models like SWAN is given in Van Vledder (2006). An overview of computational methods for computing the exact nonlinear transfer rate is given in Benoit (2005). The Boltzmann integral describes the rate of change of action density of a particular wave number due to resonant interactions between pairs of four wave numbers. To interact these wave numbers must satisfy the following resonance conditions k1 + k2 = k3 + k4 1 + 2 = 3 + 4 . (2.77)

28

Chapter 2

The rate of change of action density N1 at wave number k1 due to all quadruplet interactions involving k1 is given by N1 = t G k1 , k2 , k3 , k4 k1 + k2 k3 k4 (1 + 2 3 4 ) [N1 N3 (N4 N2 ) + N2 N4 (N3 N1 )] dk2 dk3 dk4 , (2.78)

where the action density N is dened in terms of the wave number vector k , N = N (k ). The term G is a complicated coupling coecient for which an explicit expression has been given by Herterich and Hasselmann (1980). In the WRT method a number of transformations are made to remove the delta functions. A key element in the WRT method is to consider the integration space for each (k1 , k3 ) combination N1 =2 t in which the function T is given by T k1 , k3 = G k1 , k2 , k3 , k4 k1 + k2 k3 k4 (2.80) T k1 , k3 dk3 , (2.79)

(1 + 2 3 4 ) k1 , k3 , k4 [N1 N3 (N4 N2 ) + N2 N4 (N3 N1 )] dk2 dk4 , in which k1 , k3 , k4 =

1 0

when when

k1 k3 k1 k4 k1 k3 > k1 k4

(2.81)

The delta functions in Eq. (2.80) determine a region in wave number space along which the integration should be carried out. The function determines a section of the integral which is not dened due to the assumption that k1 is closer to k3 than k2 . The crux of the Webb method consists of using a local co-ordinate system along a so-named locus, that is, the path in k space that satises the resonance conditions for a given combination of k1 and k3 . To that end the (kx , ky ) co-ordinate system is replaced by a (s, n) co-ordinate system, where s (n) is the tangential (normal) direction along the locus. After some transformations the transfer integral can then be written as a closed line integral along the closed locus T k1 , k3 = G J (k1 , k3 , k4 ) (2.82)

[N1 N3 (N4 N2 ) + N2 N4 (N3 N1 )] ds ,

in which G is the coupling coecient and J is the Jacobian term of a function representing the resonance conditions. The Jacobian term is a function of the group velocities of interacting wave numbers J = |cg,2 cg,4 |1 (2.83)

Governing equations

29

Numerically, the Boltzmann integral is computed as the nite sum of many line integrals T for all discrete combinations of k1 and k3 . The line integral (2.82) is solved by dividing the locus in typically 40 pieces, such that its discretized version is given as:
ns

T k1 , k3

G( si ) J ( si ) P ( si ) si ,
i=1

(2.84)

in which P (si ) is the product term for a given point on the locus, ns is the number of segments, si is the discrete co-ordinate along the locus, and si is the stepsize. Finally, the rate of change for a given wave number k1 is given by
n nk N (k1 ) T (k1 , k3 ) kik3 i3 t ik3 =1 i3 =1

(2.85)

where nk and n are the discrete number of wave numbers and directions in the computational spectral grid, respectively. Note that although the spectrum is dened in terms of the vector wave number k , the computational grid in a wave model is more conveniently dened in terms of the absolute wave number and wave direction (k, ) to assure directional isotropy of the calculations. Taking all wave numbers k1 into account produces the complete source term due to nonlinear quadruplet wave-wave interactions. Details of the computation of a locus for a given combination of the wave numbers k1 and k3 can be found in Van Vledder (2006). It is noted that these exact interaction calculations are extremely expensive, typically requiring 103 to 104 times more computational eort than the DIA. Presently, these calculations can therefore only be made for highly idealized test cases involving a limited spatial grid. The nonlinear interactions according to the WRT method have been implemented in SWAN using portable subroutines. In this implementation, the computational grid of the WRT method is based to the discrete spectral grid of SWAN. The WRT method uses a (k, ) grid which is based on the (, ) grid of SWAN. In addition, the WRT routines inherit the power of the parametric spectral tail as in the DIA. Choosing a higher resolution than the computational grid of SWAN for computing the nonlinear interactions is possible in theory, but this does not improve the results and is therefore not implemented. Because nonlinear quadruplet wave-wave interactions at high frequencies are important, it is recommended to choose the maximum frequency of the wave model about six times the peak frequency of the spectra that are expected to occur in a wave model run. Note that this is important as the spectral grid determines the range of integration in Eq. (2.85). The recommended number of frequencies is about 40, with a frequency increment factor 1.07. The recommended directional resolution for computing the nonlinear interactions is about 10 . For specic purposes other resolutions may be used, and some testing with other resolutions may be needed.

30

Chapter 2

An important feature of most algorithms for the evaluation of the Boltzmann integral is that the integration space can be pre-computed. In the initialization phase of the wave model the integration space, consisting of the discretized paths of all loci, together with the interaction coecients and Jacobians, are computed and stored in a binary data le. For each discrete water depth such a data le is generated and stored in the work directory. The names of these data les consist of a keyword, xnl4v5, followed by the keyword xxxxx, with xxxxx the water depth in a certain unit (meters by default), or 99999 for deep water. The extension of the binary data le is bqf (of Binary Quadruplet File). If a BQF le exists, the program checks if this BQF le has been generated with the proper spectral grid. If this is not the case, a new BQF le is generated and the existing BQF le is overwritten. During a wave model run with various depths, the optimal BQF is used, by looking at the nearest water depth dN for which a valid BQF le has been generated. In addition, the result is rescaled using the DIA scaling (2.76) according to
dN d = Snl4 Snl4

R(kp d) . R(kp dN )

(2.86)

Triads The Lumped Triad Approximation (LTA) of Eldeberky (1996), which is a slightly adapted version of the Discrete Triad Approximation (DTA) of Eldeberky and Battjes (1995) is used in SWAN in each spectral direction:
+ Snl3 (, ) = Snl3 (, ) + Snl3 (, )

(2.87)

with
+ Snl3 (, ) = max[0, EB 2ccg J 2 | sin | E 2 (/2, ) 2E (/2, )E (, ) ]

(2.88) (2.89)

and
+ Snl3 (, ) = 2Snl3 (2, )

in which EB is a tunable proportionality coecient. The biphase is approximated with = with Ursell number U r: 0. 2 + tanh( ) 2 2 Ur (2.90)

g Hs Tm01 2 (2.91) Ur = 2 d2 8 2 The triad wave-wave interactions are calculated only for 0 U r 1. The interaction coecient J is taken from Madsen and Srensen (1993): J=
2 2 k/ 2 (gd + 2c/2 ) 2 2 2 2 d2 ) gd3 k k d(gd + 15 5

(2.92)

Governing equations

31

2.3.5

First- and second-generation model formulations in SWAN

The source term Stot for the rst- and second-generation formulation (relaxation model) of SWAN is (Holthuijsen and de Boer, 1988):
Sin = A + BE ,

if E < Elim and | w | <

2 2 2

Stot =

Sds,w = 0,

where Sin and Sds,w represent input by wind and decay for over-developed sea states, respectively, A and BE are a linear and exponential growth term, respectively, E is the spectral density, Elim is the saturated spectrum, is a time scale and and w are the discrete spectral wave direction and the wind direction, respectively. The expressions for the source terms Sin and Sds,w have been modied for shallow water applications (N. Booij and L.H. Holthuijsen, personal communication, 1996) and are given below. The distinction between rst- and second-generation is only in the formulation of the saturated spectrum Elim as outlined below. Linear and exponential growth The linear growth term A is given by an expression due to Cavaleri and Malanotte-Rizolli (1981) as adapted by Holthuijsen and de Boer (1988) and Holthuijsen et al. (1996):

1 C2 2 g 2 drag a 2 w

Elim E

, if E > Elim and | w | < if E > Elim and | w | >

(2.93)

where 1 is a coecient that has been tuned to be 1 = 188, Cdrag is a drag coecient equal to Cdrag = 0.0012 and PM,d is the fully developed peak frequency including the eect of shallow water and is estimated from the depth dependent relation of the Shore Protection Manual (1973): PM PM,d = (2.95) 0.375 ) tanh(0.833d with the dimensionless depth = gd d 2 U10 The Pierson-Moskowitz (1964) frequency is PM = 0.13g 2 U10 (2.97) (2.96)

A=

(U10 max[0, cos( w )])4 , 0.7PM,d < 0.7PM,d

(2.94)

0,

32

Chapter 2

The exponential growth term BE is due to Snyder et al. (1981) rescaled in terms of U10 as adapted by Holthuijsen and de Boer (1988) and Holthuijsen et al. (1996): B = max[0, 2 5 a U10 cos( w ) 3 )] ( 2 w /k (2.98)

in which the coecients 2 and 3 have been tuned to be 2 = 0.59 and 3 = 0.12. Decay If the spectral densities are larger than the wind-dependent saturation spectrum Elim , (e.g., when the wind decreases), energy is dissipated with a relaxation model: Sds,w (, ) = where ( ) is a time scale given by ( ) = 4 2
2

Elim (, ) E (, ) ( )

(2.99)

g U10 cos( w )

(2.100)

in which the coecient 4 has been tuned to be 4 = 250. Saturated spectrum The saturated spectrum has been formulated in term of wave number with a cos2 directional distribution centred at the local wind direction w . It is essentially an adapted PiersonMoskowitz (1964) spectrum:

k3 2cg

For the rst-generation formulation, the scale factor is a constant and equals = 0.0081 (2.102)

Stot =

( ) 4 exp 5 4 PM,d

cos2 ( w ) , for | w | < for | w |

2 2

(2.101)

0,

For the second-generation formulation, the scale factor depends on the total dimensiontot,sea of the wind sea part of the spectrum and the dimensionless depth less wave energy E : d 0.223 ] = max[(0.0081 + (0.013 0.0081)ed ), 0.0023E (2.103) tot,sea tot,sea is given by: where the total dimensionless wind sea wave energy E
2 tot,sea = g Etot,sea E 4 U10

(2.104)

Governing equations with Etot,sea =

33

w +/2 w /2

0.7PM,d

E (, )dd

(2.105)

The maximum value of is taken to be 0.155. This dependency of on the local dimensionless energy of the wind sea permits an overshoot in the wave spectrum under wave generation conditions. For deep water = 0.0081 as proposed by Pierson and Moskowitz (1964).

2.4

Wave damping due to vegetation

Apart from the six processes as outlined in Section 2.3, SWAN has an option to include wave damping over a vegetation eld (mangroves, salt marshes, etc.) at variable depths. A popular method of expressing the wave dissipation due to vegetation is the cylinder approach as suggested by Dalrymple et al. (1984). Here, energy losses are calculated as actual work carried out by the vegetation due to plant induced forces acting on the uid, expressed in terms of a Morison type equation. In this method, vegetation motion such as vibration due to vortices and swaying is neglected. For relatively sti plants the drag force is considered dominant and inertial forces are neglected. Moreover, since the drag due to friction is much smaller than the drag due to pressure dierences, only the latter is considered. Based on this approach, the time-averaged rate of energy dissipation per unit area over the entire height of the vegetation is given by: 2 gk v = CD bv Nv 3 2
3

sinh3 kh + 3 sinh kh 3 H 3k cosh3 kh

(2.106)

where is the water density, CD is the drag coecient, bv is the stem diameter of cylinder (plant), Nv is the number of plants per square meter, h is the vegetation height, h is the water depth and H is the wave height. This formula was modied by Mendez and Losada (2004) for irregular waves. The mean rate of energy dissipation per unit horizontal area due to wave damping by vegetation is given by: 1 D bv Nv gk < v > = C 2 2
3

sinh3 kh + 3 sinh kh 3 Hrms 3k cosh3 kh

(2.107)

D is the bulk drag coecient that may depend on the wave height. This is the only with C calibration parameter required for a given plant type. To include wave damping due to vegetation in SWAN, Eq. (2.26) will be extended with Sds,veg based on Eq. (2.107). A spectral version of the vegetation dissipation model of Mendez and Losada can be obtained by expanding Eq. (2.107) to include frequencies and directions as follows Dtot Sds,veg (, ) = E (, ) (2.108) Etot

34 with Dtot

Chapter 2
3

1 gk = C D bv N v 2g 2

+ 3 sinh kh sinh3 kh 3 Hrms 3k cosh3 kh

(2.109)

are given by Eqs. (2.44) and (2.45), where the mean frequency , the mean wave number k 2 respectively. With Hrms = 8Etot , the nal expression reads Sds,veg k 2 2 g C D bv N v =
3

+ 3 sinh kh sinh3 kh Etot E (, ) 3k cosh3 kh

(2.110)

Apart from extending the Mendez and Losadas formulation to a full spectrum, the possibility to vary the vegetation vertically is included. The contribution of each vertical segment is calculated individually with the total energy dissipation equal to the sum of the dissipation in each layer up till the still water level. With this implementation of the differences in characteristics of each layer, plants such as mangrove trees may be conveniently input into the SWAN model. The layer-wise segmentation is implemented by integration of the energy dissipation over height as follows:
I

Sds,veg =
i=1

Sds,veg,i

(2.111)

where I the number of vegetation layers and i the layer under consideration with the energy dissipation for layer i. First, a check is performed to establish whether the vegetation is emergent or submergent relative to the water depth. In case of submergent vegetation the energy contributions of each layer are added up for the entire vegetation height. In case of emergent vegetation only the contributions of the layers below the still water level are taken into account. The implementation of vertical variation is illustrated in Figure 2.3. The energy dissipation term for a given layer i therefore becomes 3 k 2 2 Sds,veg,i = g CD,i bv,i Nv,i Etot i1 h + 3 sinh k i h sinh k i 1 h i h sinh3 k sinh3 k E (, ) 3k cosh3 kh

(2.112)

The corresponding terms for each layer can then be added and the total integrated as described earlier to obtain the energy dissipation over the entire spectrum. Here h is the total water depth and i the ratio of the depth of the layer under consideration to the total water depth up to the still water level, such that
I i=1

i 1

(2.113)

Finally, in addition to the vertical variation, the possibility of horizontal variation of the vegetation characteristics is included as well. This inclusion enables the vegetation in a given region to be varied so as to reect real density variations in the eld.

Governing equations

35

~ Sveg,2 CD,2 bv,2 Nv,2 h ~ Sveg,1 CD,1 bv,1 Nv,1

2h 1h Sveg = Sveg,1 + Sveg,2

Figure 2.3: Layer schematization for vegetation.

2.5

The inuence of ambient current on waves

Waves are subject to the inuence of ambient current, when they propagate on it. The ambient current can be tidal current, ocean current, local wind generated current, river current and wave generated current. It has been observed that current aects the growth and decay of waves (Yu, 1952; Hedges et al., 1985; Lia et al., 1989). The observations have shown that in a strong opposite current the wave steepness and wave height increase signicantly. These changes take place rapidly where the waves are blocked by the current, often accompanied with current-induced whitecapping and wave reections. Moreover, at the blocking frequency action is also partially transferred away from the blocking frequency to higher and lower frequencies by nonlinear wave-wave interactions (Ris, 1997). It was Longuet-Higgins and Stewart (1960, 1961, 1962) who founded the theoretical description of wave-current interactions. Since then, many additional results of wave-current interactions have been published. If waves propagate in the presence of ambient current, action density is conserved whereas energy density is not. Therefore, in SWAN the action balance equation has been adopted.

2.6

Modelling of obstacles

SWAN can estimate wave transmission through a (line-)structure such as a breakwater (dam). It is assumed that the obstacle is narrow compared to the grid size. If in reality the width is large compared with grid size, the feature preferably is to be modeled as a bathymetric feature. The following text refers to narrow obstacles. Such an obstacle will aect the wave eld in three ways: it will reduce the wave height of waves propagating through or over the obstacle all along its length, it will cause waves to be reected, and

36

Chapter 2 it will cause diraction around its end(s).

In irregular, short-crested wave elds, however, it seems that the eect of diraction is small, except in a region less than one or two wavelengths away from the tip of the obstacle (Booij et al., 1993). Therefore the model can reasonably account for waves around an obstacle if the directional spectrum of incoming waves is not too narrow, unless one is interested in the wave eld deep into the shadow zone. Since obstacles usually have a transversal area that is too small to be resolved by the bottom grid in SWAN, an obstacle is modelled as a line in the computational area. See Section 3.10 for the numerical implementation of obstacles in SWAN.

2.6.1

Transmission

There are several mechanisms for transmission of waves. In SWAN, the user may compute transmission of waves passing over a dam with a closed surface or may choose a constant transmission coecient. If the crest of the breakwater is at a level where (at least part of the) waves can pass over, the transmission coecient Kt (dened as the ratio of the (signicant) wave height at the downwave side of the dam over the (signicant) wave height at the upwave side) is a function of wave height and the dierence in crest level and water level. It must be noted that the transmission coecient can never be smaller than 0 or larger than 1. In SWAN, two expressions can be employed. The rst is taken from Goda et al. (1967): Kt =
1, 0,

0.5(1

(F sin( 2 Hi

< F + ))) , H i F > Hi

F Hi

(2.114)

where F = h d is the freeboard of the dam and where Hi is the incident (signicant) wave height at the upwave side of the obstacle (dam), h is the crest level of the dam above the reference level (same as reference level of the bottom), d the mean water level relative to the reference level, and the coecients , depend on the shape of the dam (Seelig, 1979) as given in Table 2.1. It should be noted that this formula is only valid for slopes more gentle than 1:0.7 (1.4:1 or 55 degrees). Expression (2.114) is based on experiments in a Table 2.1: Parameters for transmission according to Goda et al. (1967). case vertical thin wall 1.8 0.1 caisson 2.2 0.4 dam with slope 1:3/2 2.6 0.15

wave ume, so strictly speaking it is only valid for normal incidence waves. Since there are

Governing equations

37

no data available on oblique waves, it is assumed that the transmission coecient does not depend on direction. Furthermore, it is assumed that the frequencies remain unchanged over an obstacle (only the energy scale of the spectrum is aected and not the spectral shape). For an impermeable rough low-crested dam, the following expression of dAngremond et al (1996) is chosen: F Bk Kt = 0. 4 (2.115) + 0.64( )0.31 (1 e0.5p ) Hi Hi with Bk the crest width and p tan / Hi /L0p the breaker parameter. For this, the 2 slope of the breakwater must be given and L0p = gTp /2 is the deep water wave length. The restriction to eq. (2.115) is as follows: 0.075 Kt 0.9 (2.116)

In most cases, the crest width is such that Bk < 10Hi . However, if this is not the case, the following expression should be used instead of (2.115): Kt = 0.35 with the restriction: 0.05 Kt 0.006 Bk F + 0.51( )0.65 (1 e0.41p ) Hi Hi Bk + 0.93 Hi (2.117)

(2.118)

The formulas (2.115) and (2.117) give a discontinuity at Bk = 10Hi . Following Van der Meer et al (2005), for practical application, Eq. (2.115) is applied if Bk < 8Hi , Eq. (2.117) if Bk > 12Hi and in between (8Hi Bk 12Hi ), a linear interpolation is carried out.

2.6.2

Reection

Often there is reection against quays or breakwaters. Depending on the nature of the obstacle (a smooth surface or a rubble-mound breakwater) the reected wave eld can be more or less scattered. SWAN is able to diuse the reection over wave components in dierent directions.

2.6.3

Diraction

To accommodate diraction in SWAN simulations, a phase-decoupled refraction-diraction approximation is suggested (Holthuijsen et al., 2003). It is expressed in terms of the directional turning rate of the individual wave components in the 2D wave spectrum. The approximation is based on the mild-slope equation for refraction and diraction, omitting phase information. It does therefore not permit coherent wave elds in the computational domain.

38

Chapter 2

In a simplest case, we assume there are no currents. This means that c = 0. Let denotes the propagation velocities in geographic and spectral spaces for the situation without diraction as: cx,0 , cy,0 and c,0 . These are given by: cx,0 = 1 h cos , cy,0 = sin , c,0 = k k k h n (2.119)

where k is the wave number and n is perpendicular to the wave ray. We consider the following eikonal equation K 2 = k 2 (1 + ) (2.120) with denoting the diraction parameter as given by: (ccg E ) = ccg E

(2.121)

2 where E (x, y ) is the total energy of the wave eld ( Hs ). Due to diraction, the propagation velocities are given by:

cx = cx,0 , cy = cy,0 , c = c,0 where = 1+

cy,0 + cx,0 x y

(2.122)

(2.123)

In early computations, the wave elds often showed slight wiggles in geographic space with a wavelength of about 2x in xdirection. These unduly aected the estimations of the gradients that were needed to compute the diraction parameter . The wave eld was therefore smoothed with the following convolution lter:
n n 1 Ei,j = Ei,j 0.2[Ei1,j + Ei,j 1 4Ei,j + Ei+1,j + Ei,j +1 ]n1

(2.124)

where i, j is a grid point and the superscript n indicates iteration number of the convolution cycle. The width of this lter (standard deviation) in xdirection x , when applied n times is 1 x 3n x (2.125) 2 By means of computations, n = 6 is found to be an optimum value (corresponding to spatial resolution of 1/5 to 1/10 of the wavelength), so that x 2x. For the y direction, the expressions are identical, with y replacing x. Note that this smoothing is only applied to compute the diraction parameter . For all other computations the wave eld is not smoothed. Diraction is SWAN should not be used if,

Governing equations an obstacle or coastline covers a signicant part of the down-wave view, and

39

the distance to that obstacle or coastline is small (less than a few wavelengths), and the reection o that obstacle or coastline is coherent, and the reection coecient is signicant. This implies that the SWAN diraction approximation can be used in most situations near absorbing or reecting coastlines of oceans, seas, bays, lagoons and fjords with an occasional obstacle such as islinds, breakwaters, or headlands but NOT in harbours or in front of reecting breakwaters or near wall-dened cli walls. Behind breakwaters (which may be reecting), the SWAN results seem reasonable if the above conditions are met.

2.7

Wave-induced set-up

In a (geographic) 1D case the computation of the wave-induced set-up is based on the vertically integrated momentum balance equation which is a balance between the wave force (gradient of the wave radiation stress normal to the coast) and the hydrostatic pressure gradient (note that the component parallel to the coast causes wave-induced currents but no set-up). dSxx d + gH =0 (2.126) dx dx where d is the total water depth (including the wave-induced set-up) and is the mean surface elevation (including the wave-induced set-up) and Sxx = g is the radiation stress tensor. Observation and computations based on the vertically integrated momentum balance equation of Dingemans et al. (1987) show that the wave-induced currents are mainly driven by the divergence-free part of the wave forces whereas the set-up is mainly due to the rotation-free part of these forces. To compute the set-up in 2D, it would then be sucient to consider the divergence of the momentum balance equation. If the divergence of the acceleration in the resulting equation is ignored, the result is: Fx Fy + (gH ) (gH ) = 0 x y x x y y (2.128) 1 [n cos2 + n ]Edd 2 (2.127)

This approximation can only be applied to open coast (unlimited supply of water from outside the domain, e.g. nearshore coasts and estuaries) in contrast to closed basin, e.g. lakes, where this approach should not be used.

40

Chapter 2

Chapter 3 Numerical approaches

3.1 Introduction

The accuracy with which physical processes for wave growth are approximated numerically is of crucial importance in assessing the predictive realism of spectral wave models. There is a need to separate these numerical errors from errors due to physical modelling. Third-generation wave models pose a numerical diculty caused by the presence of multiple time scales. This is a reection of the physical nature of wind waves, which consist of a wide range of frequencies. The ratio of the largest to the smallest time scale of spectral components is often substantially larger than one. When this is the case, the action balance equation is called sti (Press et al., 1993)1 . Taking proper account of these time scales is a necessary condition for numerical accuracy. This would require the use of a very small time step in a numerical algorithm, which may be impractical. Moreover, the action balance equation is usually so sti that its numerical implementation combined with economically large time steps often prevent a stable solution. In this respect, nonlinear four-wave interaction usually poses the biggest problem, since this process is associated with high sensitivity to spectral change. In a number of papers concerning spectral wave computation, numerical measures are proposed to achieve stable model results economically. WAMDI Group (1988) suggest to use a semi-implicit time integration scheme with a time step that matches the time scale of lowfrequency waves. However, numerically stable solution of the resulting sytem of equations can not be guaranteed (Hargreaves and Annan, 2001). The ratio of the largest eigenvalue to the smallest eigenvalue of the sti system of equations, called the condition number, can be so large that even a fully-implicit method combined with large time steps precludes a stable solution. For counterexamples, see Hargreaves and Annan (2001). The only remedy is time step reduction or under-relaxation so that the modied system of equations has a spectrum of eigenvalues with a more favourable condition number. To guarantee numerical stability at relatively large time steps, the so-called action density
1

The equivalent situation for such an equation is to have eigenvalues of very dierent magnitudes.

41

42

Chapter 3

limiter has been introduced in WAM in the early 1980s (Hersbach and Janssen, 1999). This limiter restricts the rate of change of the energy spectrum at each time step. Because low-frequency waves carry the most energy, it is desirable to solve the balance equation in this part of the spectrum accurately without intervention by the limiter, whereas for highfrequency waves using an equilibrium level is sucient. Although this approach lacks a rigorous foundation and is not generally applicable or valid, it appears to guarantee numerical stability at relatively large time steps even when these do not match the time scales of wave growth. Moreover, it is believed that the limiter will not aect the stationary solution when convergence is reached. This assumption is widely employed as a justication for the use of limiters. For an overview, we refer to Hersbach and Janssen (1999) and Tolman (2002) and the references quoted therein. Tolman (1992) proposes an alternative to the action density limiter in which the time step is dynamically adjusted where necessary to ensure accurate wave evolution. The calculation of this optimal time step is related to the action density limiter. Further details can be found in Tolman (1992, 2002). . The steady-state solution in the SWAN model is obtained in an iterative manner, which can be regarded as a time marching method with a pseudo time step. This pseudo time step generally does not match the relatively small time scale in frequency space and consequently, divergence will occur. Therefore, SWAN makes use of the action density limiter to stabilize the iteration process (Booij et al., 1999). However, experience with SWAN has revealed that the limiter acts not only in the equilibrium space, but also in the energycontaining part of the wave spectrum. This nding is also conrmed by Tolman (2002). Furthermore, the limiter appears to be active over almost all spectra in the geographical domain and during the entire iteration process. This activity has been associated with poor convergence behaviour, such as small-amplitude oscillation in frequency space. Ris (1999) demonstrated that stationary SWAN results are inuenced by the settings of the action limiter while De Waal (2001) suspects that the limiter acts as a hidden sink in the source term balance under equilibrium conditions. The question to what extent this limiter adversely aects the stationary solution of SWAN has not been addressed previously, and is considered here. An alternative way to restrict the high rate of change at higher frequencies is underrelaxation, i.e. making smaller updates by means of a much smaller (pseudo) time step (Ferziger and Peri c, 1999). Consequently, a limiter may no longer be needed. Although this approach may be suitable to SWAN, it slows down convergence signicantly. Here, we propose a new method that nds a compromise between fast convergence on the one hand and minimizing the role of the limiter in the energetic part of the spectrum on the other. The key idea to achieve this is to link the extent of updating to the wave frequencythe larger the frequency, the smaller the update. This approach is therefore called frequencydependent under-relaxation. The second objective concerns the formulation and the use of termination criteria required by the iteration procedure in SWAN. In principle, the iterative process should be stopped

Numerical approaches

43

if the convergence error dened as the dierence between the current iterate and the stationary solution is smaller than a prescribed tolerance. At present, the stopping criteria in SWAN make use of the dierence between successive iterates as a measure of the error in the converged solution. Experiences in the application of SWAN have shown that the iteration process is often more erratic and typically much slower than reported by Booij et al. (1999). As a result, the current stopping criteria often lead to premature termination of simulations. This is characterised by the fact that, due to the relatively low rate of convergence, the convergence error is larger than the dierence between the successive iterates. A stopping criterion is proposed that uses the second derivative or curvature of the series of successive iterates of the calculated wave height. The premise is that this curvature approaches zero upon full convergence.

3.2

Discretization

Discretization of (2.17) is carried out using the nite dierence method. The homogeneous part of equation (2.17) is given by N cx N cy N c N c N + + + + . t x y (3.1)

We choose a rectangular grid with constant mesh sizes x and y in x and y direction, respectively. The spectral space is divided into elementary bins with a constant directional resolution and a constant relative frequency resolution / (resulting in a logarithmic frequency distribution). We denote the grid counters as 1 i Nx , 1 j Ny , 1 l N and 1 m N in x, y , and spaces, respectively. All variables are located at points (i, j, l, m). Time discretization takes place with the implicit Euler technique. We obtain the following approximation of (3.1): [cx N ]i+1/2 [cx N ]i1/2 n N n N n 1 |i,j,l,m + |j,l,m + t x [cy N ]j +1/2 [cy N ]j 1/2 n |i,l,m + y [c N ]l+1/2 [c N ]l1/2 n |i,j,m + [c N ]m+1/2 [c N ]m1/2 n |i,j,l ,

(3.2)

where n is a time-level with t a time step. Note that locations in between consecutive counters are reected with the half-indices.

3.2.1

Discretization in geographical space

Since, the unknown N and the propagation velocities are only given in points (i, j, l, m), further approximation is needed. A rst order upwind scheme in geographical space may

44

Chapter 3

be employed, since it fully monotone, i.e. it can not to give rise to spurious oscillations. A disadvantage of this scheme is that it is numerically diusive, which naturally degrades the accuracy of the model. This numerical diusion is caused by gradients of wave action across geographic space, e.g. due to refraction by bathymetry or currents. However, in the current SWAN version, two alternatives to this scheme are implemented, namely the second order SORDUP and the third order Stelling/Leendertse schemes. These schemes produce far less numerical diusion. First order upwind scheme; BSBT The uxes cx N at (i + 1/2, j, l, m) and cy N at (i, j + 1/2, l, m) are approximated in the following way: cx N |i+1/2,j,l,m = and cy N |i,j +1/2,l,m = cy N |i,j,l,m , cy N |i,j +1,l,m , cy |i,j,l,m > 0 . cy |i,j +1,l,m < 0 (3.4) cx N |i,j,l,m , cx N |i+1,j,l,m , cx |i,j,l,m > 0 cx |i+1,j,l,m < 0 (3.3)

The uxes at (i 1/2, j, l, m) and (i, j 1/2, l, m) are obtained from (3.3) and (3.4), respectively, by decreasing the indices by 1 in appropriate manner. Note that the combination of the time and geographic space discretizations in (3.2), (3.3) and (3.4) is also known as the rst order, backward space, backward time (BSBT) scheme. SORDUP For the SORDUP scheme which is the default scheme for stationary computations, the two terms in Eqs. (3.3) and (3.4) representing x and y derivatives are replaced by 1.5(cx N )ix 2(cx N )ix 1 + 0.5(cx N )ix 2 x and 1.5(cy N )iy 2(cy N )iy 1 + 0.5(cy N )iy 2 y
it ,n it ,n

(3.5)
iy ,i ,i

(3.6)
ix ,i ,i

See also Rogers et al. (2002). In the neighboorhood of open boundaries, land boundaries and obstacles (i.e., the last two grids adjoining such grid points for the SORDUP scheme), SWAN will revert to the rst order upwind BSBT scheme. This scheme has a larger numerical diusion but that is usually acceptable over the small distances involved. Stelling and Leendertse scheme For the Stelling and Leendertse scheme which is the default scheme for nonstationary computations,

Numerical approaches

45

the two terms in Eqs. (3.3) and (3.4) representing x and y derivatives are replaced by
5 ( c N ) ix 6 x 1 1 ( c N ) ix 1 + 2 (cx N )ix 2 12 ( c x N ) ix 3 5 4 x x it ,n

+
iy ,i ,i

(cx N )ix +1 (cx N )ix 1 4x

it 1 iy ,i ,i

(3.7) and
5 ( c N ) iy 6 y 1 1 ( c N ) iy 1 + 2 (cy N )iy 2 12 ( c y N ) iy 3 5 4 y y it ,n

(3.8) See also Stelling and Leendertse (1992). In the neighboorhood of open boundaries, land boundaries and obstacles (i.e., the last three grids adjoining such grid points for the Stelling and Leendertse scheme), SWAN will revert to the rst order upwind BSBT scheme. Usually, the numerical diusion of the Stelling and Leendertse scheme is so small that the so-called garden-sprinkler eect (GSE) may show up if propagation over very large distances is considered. This eect is due to the spectral resolution (see Booij and Holthuijsen, 1987). It can be counteracted by a diusion term that has been explicitly added to the numerical scheme. Its value depends on the spectral resolution and the propagation time of the waves. The diusion applied in the propagation direction is c2 gT Dss = 12 where T is the wave age. The diusion normal to the propagation direction is Dnn =
2 c2 g T 12

(cy N )iy +1 (cy N )iy 1 + 4y ix ,i ,i

it 1 ix ,i ,i

(3.9)

(3.10)

From these, diusion coecients are calculated as Dxx = Dss cos2 +Dnn sin2 , Dyy = Dss sin2 +Dnn cos2 , Dxy = (Dss Dnn ) cos sin (3.11)

The diusion terms are given by Dxx 2N 2N 2N D 2 D yy xy x2 xy y 2 (3.12)

and are computed at the time level it 1 as follows Dxx (N )ix +1 2(N )ix + (N )ix 1 x2 (N )iy +1 2(N )iy + (N )iy 1 y 2
it 1

(3.13)
iy ,i ,i it 1

Dyy

(3.14)
ix ,i ,i

46 Dxy

Chapter 3 (N )ix ,iy (N )ix 1,iy (N )ix 1,iy + (N )ix 1,iy 1 x y

it 1

(3.15)
i ,i

This explicit scheme is fast (having little impact on computation time) but only conditionally stable. Through mathematical analysis (not shown) it can be shown that a likely stability condition is (Fletcher, 1988): Q= max(Dxx , Dyy , Dxy )t 0. 5 min(x, y )2 (3.16)

Thus, it is credible that Eq. (3.16) holds true for the two-dimensional Stelling and Leendertse scheme with this GSE correction. In short, by adding the GSE correction, the unconditionally stable advection scheme of SWAN becomes a (likely) conditionally stable advection-diusion scheme. This restriction appears not to be severe as is commonly believed for a typical advection-diusion equation. In experiments, it was found that with Q 0.48, no instability was observed. It is readily shown that for typical ocean applications Dnn dominates the diusion and Q can be written as Q= cg 2 T t 12x2 (3.17)

The variable wave age T could be computed during the computations of SWAN but it requires the same order of magnitude of computer memory as integrating the action balance equation. Instead a constant wave age T can be used as an approximation, so that Eq. (3.17) becomes L2 (3.18) Q= 12x where the characteristic travel distance of the waves is L = cg T (e.g., the dimension of the ocean basin). For oceanic applications, the Courant number is typically 1 so that 2 o Q 0.25 for typical values of 10 and L/x 200 (the number of grid points in one direction of the grid). This implies that the Stelling and Leendertse scheme with the GSE correction is stable for typical ocean cases. For shelf sea (regional) applications, the value of = O(1) but the garden-sprinkler eect tends to be small on these scales and the diusion can and should not be used to avoid the stability problem. For small-scale (local) applications, typically = O(10 100). But such cases are usually treated as stationary and the SORDUP scheme should be used (no GSE correction is included in this scheme).

3.2.2

Note on the choice of geographic propagation schemes

The main interest of the SWAN users is in simulating wind-generated waves and combined swell-sea cases in coastal ocean waters, and it is particularly with the view to such computations that a simple and compact, but rst order, scheme was implemented in SWAN. A substantial body of experience gathered over the past 15 years on the performance of both lower and higher upwind schemes in SWAN suggests that in many circumstances, the

Numerical approaches

47

discretization of the propagation terms in geographical space is not a crucial issue. Many nearshore simulations have shown the solution for action density to be on the whole rather insensitive to the accuracy with which geographic propagation terms are approximated. This reects the tendency for the level of wave action to be dictated by source terms, while the local changes of the energy eld across geographical space is relatively weak. This is consonant with the established view that a certain amount of numerical diusion can be safely tolerated in the numerical scheme for geographic propagation, as its impact on wave parameters is negligible (Rogers et al., 2002; WISE Group, 2007). Also, broad wave spectra will tend make numerical diusion far less noticable in a wave eld. This would appear to suggest, however, that the use of higher order upwind schemes serves no useful purpose. This is probably not so since there might be some cases that are prone to diusion, where the benet of such schemes is obvious. One can think of a case of swell propagation over very long distances. While low-diusive, higher order schemes did permit long-distance swell cases to be validated, the reduced diusion was found to pose a serious diculty as the garden sprinkler eect becomes more visible, see e.g. WISE Group (2007).

3.2.3

Discretization in spectral space

The uxes in the spectral space (, ), as given in (3.2), should not be approximated with the rst order upwind scheme since, it turns out to be very diusive for frequencies near the blocking frequency2 . Central dierences should be used because of second order accuracy. However, such schemes tend to produce unphysical oscillations due to relatively large gradients in action density near the blocking frequency. Instead, a hybrid central/upwind scheme is employed: c N |i,j,l+1/2,m =
(1 0.5)c N |i,j,l,m + 0.5c N |i,j,l+1,m ,

c |i,j,l,m > 0 c |i,j,l+1,m < 0 c |i,j,l,m > 0 c |i,j,l,m+1 < 0

(3.19)

(1 0.5)c N |i,j,l+1,m + 0.5c N |i,j,l,m ,

and

where the parameters and are still to be chosen. For all values [0, 1] and [0, 1], a blended form arises between rst order upwind dierencing ( = = 0) and central dierencing ( = = 1). If large gradients in action density in the directional space are present, numerical oscillations can arise. Negative action density (which is unphysical) is removed from the twodimensional spectrum. To retain a conservative numerical scheme, rst the total action
2

c N |i,j,l,m+1/2 =

(1 0.5 )c N |i,j,l,m + 0.5c N |i,j,l,m+1 ,

(3.20)

(1 0.5 )c N |i,j,l,m+1 + 0.5c N |i,j,l,m ,

Waves can be blocked by the current at a relative high frequency.

48

Chapter 3

density for a certain frequency in a quadrant is computed, as follows: N ( ) = quadrant N (, ) d (3.21)

then all negative action density is removed and the positive action density is multiplied by a factor determined by the total action density divided by the absolute value of the total negative action density.

3.3

Solution algorithm

The discretization of the action balance equation (2.17) as described in Section 3.2 yields a system of linear equations that need to be solved. The corresponding matrix structure can take dierent forms, mainly depending on the propagation of wave energy in the geographic space. For instance, suppose that cx > 0 and cy > 0, everywhere. Then, the matrix structure has the following form:

.x.xxx.x.. ..x.xxx.x. ...x.xxx.x

....x....... . .....x..... ........x...

One recognizes that the subblocks on the main diagonal express coupling among the unknowns in the (, )space for each geographic grid point, whereas the o-diagonal subblocks represent coupling across geographical grid points. This system can be solved with a Gauss-Seidel technique in one step (Wesseling, 1992). For an illustrative explanation of this technique, see Section 3.4. Generally, the velocities cx and cy may have dierent signs in the geographical domain and hence, more steps are needed. However, it is well known that adapting the ordering of updates of the unknowns N in geographical space to the propagation direction can improve the rate of convergence of the Gauss-Seidel iterative procedure (Wesseling, 1992). This is done as follows. For each iteration, sweeping through grid rows and columns in geographical domain are carried out, starting from each of the four corners of the computational grid. After four sweeps, wave energy has been propagated over the entire geographical domain. During each sweep, only a subset of the unknown values of N are updated depending on the sign of cx and cy . For instance, the rst sweep starts at the lower left-hand corner and all grid points with cx > 0 and cy > 0 are updated.

....x....... ......x..... ........x... ........x...

.x.xxx.x.. . ..x.xxx.x. ...x.xxx.x .x.xxx.x.. ....x....... ......x..... ..x.xxx.x.

(3.22)

...x.xxx.x

Numerical approaches

49

After each propagation update at geographic grid point, an update in the spectral space is made. Since, according to (3.3) and (3.4), the wave energy at a single spatial location depends on the upwind grid points only, it is sucient to carry out the update within a 90o -quadrant of the (, )-space, as illustrated in Figure 3.1. Because of the implicit nature of the spectral propagation terms in (3.2), a system of equations must be formed. FurtherSWEEP 1 y y SWEEP 2 y SWEEP 3 y SWEEP 4

Figure 3.1: The solution procedure for wave energy propagation in geographical space with the appropriate directional quadrant (indicated by shaded area) for each of four sweeps. more, due to the fact that the source term Stot in (3.2) is nonlinear in N , linearization is required in order to nd a solution. Generally, the term Stot in each bin (l, m) is treated by distinguishing between positive and negative contributions and arranging these in the linear form (Ferziger and Peri c, 1999):
p n + Stot N, Stot = Stot

(3.23)

p n consists of positive contributions and Stot of negative ones. Both contributions where Stot are independent of the solution N at the corresponding bin (l, m). Any negative term that does not contain N as a multiplier is rst divided by N obtained from the previous n iteration level and then added to Stot . This stabilizes the iteration process. Details on the application of this principle to each source term in SWAN can be found in Booij et al. (1999).

The strongly nonlinear source term of depth-induced wave breaking is linearized by means of the Newton-Raphson iteration, as follows: S
n n 1

S E + E
n

n 1

( E n E n 1 )

(3.24)

Since, this process of depth-induced wave breaking has been formulated such that S = aStot and E = aEtot , the derivative S/E is analytically determined as Stot /Etot . Here, a

50

Chapter 3

is identical in both expressions and the total energy Etot and total source Stot are the integrals over all frequencies and directions of E (, ) and Sds,br (, ), respectively. As such, each dierence equation (3.2) using expressions (3.19), (3.20) and (3.23) provides an algebraic relation between N at the corresponding bin and its nearest neighbours: aP N P = aL N L + aR N R + aB N B + aT N T + b P , (3.25)

where P corresponds to central bin (l, m) and L(eft), R(ight), B(ottom) and T(op) correspond to (l 1, m), (l + 1, m), (l, m 1) and (l, m + 1), respectively. Furthermore, the coecients ak , k {P, L, R, B, T} arise from the discretizations of the uxes c N and c N p and bP contains the positive contributions of the source term Stot in (3.23) and the updated n uxes cx N (3.3) and cy N (3.4). Note that coecient aP includes Stot . The linear system of equations (3.25) for all bins within a directional quadrant at a particular geographical point is denoted by AN = b, (3.26)

where A I RK K contains the coecients ak , k {P, L, R, B, T} (and corresponds to a subblock on the main diagonal of (3.22)), b I RK contains the coecient bP and boundary values and N I RK denotes an algebraic vector containing the unknown action density values. Matrix A is non-symmetric. The dimension K of a directional quadrant equals N 1/4N . Note that linearization of the source term (3.23) enhances diagonal dominance of A, thereby improving numerical stability. Also note that neither A nor b depends on the unknowns. Each row in the matrix A corresponds to a bin (l, m). The main diagonal contains the coecients aP and directly to the left and right are the coecients aB and aT , respectively. The coecients aL and aR are on the diagonals that are N positions to the left and right of the main diagonal, respectively. The solution N is given by A1 b. Since, the only non-zero matrix elements are situated in ve diagonals, iterative solution methods that utilize the sparsity of A optimally are very attractive. In SWAN, the solution of (3.26) is found by means of an incomplete lowerupper decomposition method followed by an iteration process called the Strongly Implicit Procedure (SIP) (Ferziger and Peri c, 1999). This procedure is specically designed for (non-symmetric) penta-diagonal systems and is relatively fast. Note that in the absence of mean current there are no shifts in the frequency, and consequently the structure of A reduces to a tri-diagonal one, i.e. aL = aR = 0, which can be inverted eciently with the Thomas algorithm (Press et al., 1993; Ferziger and Peri c, 1999). Due to refraction and nonlinear wave energy transfer, interactions occur between the directional quadrants. To properly take these interactions into account and the fact that we employ the Gauss-Seidel technique and linearization of the source term (3.23), the quadrant sweeping and the solution of system (3.26) need to be repeated until some convergence

Numerical approaches

51

criteria are met. At present, the iteration process runs from s = 1 to s = S and is terminated if the maximum number of iterations S (usually 15) is reached or the following criteria for the signicant wave height Hm0 and mean relative wave period Tm01 , as given by 2 m0 Hm0 = 4 m0 , Tm01 = 2 , mj = j E (, )dd , (3.27) m1 0 0 are both satised in at least 98% of all wet grid points (i, j ):
s | Hm 0 (i, j )| < r H s 1 Hm0 (i, j ) s a or |Hm 0 (i, j )| < H

(3.28)

and

s | Tm 01 (i, j )| < r T s 1 Tm01 (i, j )

s a or |Tm 01 (i, j )| < T .

(3.29)

Here, Qs Qs Qs1 , with Q some quantity. The default values of the limiting criteria r a a are: r H = T = 0.02, H = 0.02 m and T = 0.2 s. The rationale behind the use of the integral wave parameters Hm0 and Tm01 in the stopping criteria is that these are the output variables typically of interest. The iterative solution procedure is accelerated by calculating a reasonable rst guess of the wave eld based on second-generation source terms of Holthuijsen and De Boer (1988).

3.4

An illustrative explanation of the sweeping approach

In the absence of a current, the direction of propagation of the wave crest is equal to that of the wave energy. For this case, the propagation velocity of energy (cx , cy ) is equal to the group velocity (cg,x , cg,y ). In presence of a current this is not the case, since the propagation velocities cx and cy of energy are changed by the current. Considering the applied numerical procedure in SWAN, it is initially more convenient to explain the basic principles of the numerical procedure in the absence of a current than in the situation where a current is present. So, rst, we shall focus on the sweeping technique in the absence of a current. After this, we shall discuss the numerical procedure in case a current is present. The computational region is a rectangle covered with a rectangular grid. One of the axes (say the xaxis) is chosen arbitrary, for instance perpendicular to the coast. The state in a grid point (xi ,yj ) in an upwind stencil is determined by its up-wave grid points (xi1 ,yj ) and (xi ,yj 1 ). This stencil covers the propagation of action density within a sector of 0o 90o , in the entire geographic space; see Figure 3.2. Hence, this procedure is called sweep 1 and encloses all wave energy propagation over the rst quadrant. By rotating the stencil over 90o , the next quadrant 90o 180o is propagated. Rotating the stencil twice more ensures propagation over all four quadrants (see Figure 3.1). This allows waves to propagate from all directions. Hence, the method is characterized as a four-sweep technique.

52

Chapter 3

90o 0o

Figure 3.2: Numerical scheme for wave propagation in geographic space with below the rst quadrant for which the waves are propagated. The gain of such a stencil is that the propagation is unconditionally stable because the wave characteristics lie within the concerning quadrant. Propagation is not subjected to a CFL criterion. In cases with current refraction or bottom refraction, action density can shift from one quadrant to another. This is taken into account in the model by repeating the computations with converging results (iterative four-sweep technique). Typically, we choose a change of less then 1% or so in signicant wave height and mean wave period in all geographic grid points to terminate the iteration (see Section 3.3). The numerical procedure as described above remains in principle the same when a current (Ux ,Uy ) is present. The main dierence is that the propagation velocities of energy are no longer equal to the group velocity of the waves but become equal to cx = cg,x + Ux and cy = cg,y + Uy . To ensure an unconditionally stable propagation of action in geographical space in the presence of any current, it is rst determined which spectral wave components of the spectrum can be propagated in one sweep. This implies that all wave components with cx > 0 and cy > 0 are propagated in the rst sweep, components with cx < 0 and cy > 0 in the second sweep, components with cx < 0 and cy < 0 in the third sweep, and nally, components cx > 0 and cy < 0 in the fourth sweep. Since the group velocity of the waves decreases with increasing frequency, the higher frequencies are more inuenced by the current. As a result, the sector boundaries in directional space for these higher

Numerical approaches

53

frequencies change more compared to the sector boundaries for the lower frequencies. In general, four possible congurations do occur (see Figure 3.3). Consider, for instance, one xed frequency propagating on a uniform current. The current propagates at an angle of 45o with the xaxis. The sign of the current vector and strength of the current are arbitrary. The shaded sectors in Figure 3.3 indicate that all the wave components that are propagating in the direction within the shaded sector, are propagated in the rst sweep (cx > 0, cy > 0).

(A)

Cy

(B)

Cy

Cx

Cx

(C)

Cy

Cy

(D)
Cx

Cx

Figure 3.3: Four possible congurations and propagation velocities cx , cy for a xed frequency in the presence of a current propagating at an angle of 45o with the xaxis. The top-left panel (A) represents a situation in which both cx and cy are negative due to a strong opposing current, i.e. wave blocking occurs. None of the wave components is propagated within the rst sweep. The top-right panel (B) represents a situation in which the current velocity is rather small. The sector boundaries in directional space are hardly changed by the current such that the sector boundaries are approximately the same as in the absence of a current. The bottom-left panel (C) reects a following current that causes the propagation velocities of the wave components in two sectors to be larger than zero. In this specic case, all the waves of the shaded sectors are propagated within the rst sweep. The bottom-right panel (D) represents a case with a strong following current for which all the action is take along with the current. For this case the fully 360o sector is propagated in the rst sweep.

54

Chapter 3

After it has been determined which wave components are propagated in one sweep, i.e., the sector boundaries in directional space have been determined for each frequency, the integration in frequency and directional space can be carried out for those wave components.

3.5

Implementation of DIA within the four-sweep technique

In SWAN, the quadruplets are integrated using the DIA, see Section 2.3.4. As a consequence of the four-sweep technique, two dierent types of methods can be used to calculate the four wave-wave interactions. 1. The rst method implies that the interactions are calculated in every iteration prior to the rst sweep. This method ensures conservation of energy density, but has the disadvantage that the spectral source term Snl4 (, ) for every grid point in geographical space has to be stored in internal memory. Such an integration method increases the amount of required memory with a factor about 2. The source term is stored in memory and is then explicitly integrated for a particular sweep. 2. The second method is slightly dierent, in which the interactions are calculated and integrated for every sweep separately. The calculation of the energy transfer for a specic quadrant requires also the calculation of the transfer within a sector of about 33o in the adjacent two quadrants. Calculating the interactions for one sweep increases the computation time for the quadruplets with a factor of about 1.66 (2 adjacent sectors of 33o times 4 sweeps divided by 360o ). Contrary to the rst method, the total rate of this energy shift is not stored in memory such that energy density is not conserved per sweep (and per iteration). However, this does not inuence the converging results. Within this second method, two dierent numerical schemes are available, namely, a semi-implicit scheme and a fully explicit scheme. The use of a fully explicit scheme is recommended because of the computational eciency. A semi-implicit scheme increases the computation time of the quadruplets with a factor of about 2.

3.6

Convergence-enhancing measures

As explained in Section 3.1, many time scales are involved in the evolution of wind waves. The high-frequency waves have much shorter time scales than the low-frequency waves, rendering the system of equations (3.26) sti. If no special measures are taken, the need to resolve high-frequency waves at very short time scales would result in extreme computational time. For economy, it is desirable that a numerical technique can be used with a large, xed time step. Moreover, we are mainly interested in the evolution of slowly changing low-frequency waves. For stationary problems, we are interested in obtaining the steady-state solution. Unfortunately, the convergence to the steady state is dominated

Numerical approaches

55

by the smallest time scale and, in the absence of remedial measures, destabilizing overand undershoots will prevent solution from converging monotonically during the iteration process. These oscillations arise because of the o-diagonal terms in matrix A, which can be dominant over the main diagonal, particularly when the ratio max /min is substantially larger than one. As a consequence, convergence is slowed down and divergence often occurs. To accelerate the iteration process without generating instabilities, appropriately small updates must be made to the level of action density. With the development of the WAM model, a so-called action density limiter was introduced as a remedy to the abovementioned problem. This action limiter restricts the net growth or decay of action density to a maximum change at each geographic grid point and spectral bin per time step. This maximum change corresponds to a fraction of the omni-directional Phillips equilibrium level (Hersbach and Janssen, 1999). In the context of SWAN (Booij et al., 1999), this is N PM , 2k 3 cg (3.30)

where 0 denotes the limitation factor, k is the wave number and PM = 8.1 103 is the Phillips constant for a Pierson-Moskowitz spectrum (Komen et al., 1994). Usually, = 0.1 (Tolman, 1992)3 . Note that when the physical wind formulation of Janssen (1989,1991a) is applied in SWAN, the original limiter of Hersbach and Janssen (1999) is employed. Denoting the total change in Ni,j,l,m from one iteration to the next after (3.2) by Ni,j,l,m , the action density at the new iteration level is given by
s 1 s Ni,j,l,m = Ni,j,l,m +

Ni,j,l,m min{|Ni,j,l,m |, N } . |Ni,j,l,m |

(3.31)

For wave components at relatively low frequencies, (3.31) yields the pre-limitation outcome of (3.2), because, for these components, the pseudo time step matches the time scale of their evolution. For high-frequency waves, however, (3.31) gives the upper limit for the spectrum to change per iteration due to the limiter (3.30). For typical coastal engineering applications, it is sucient to compute the energy-containing part of the wave spectrum accurately. In other words, action densities near and below the spectral peak should not be imposed by the limiter (3.30). However, our experiences with SWAN have shown that the limiter is active even close to the peak. Furthermore, during the entire iteration process, the limiter is typically active at almost every geographic grid point. The alternative measure to enhance the convergence of the stable iteration process considered here is so-called false time stepping (Ferziger and Peri c, 1999). Under-relaxation terms representing the rate of change are introduced to enhance the main diagonal of A and thus stabilize the iteration process. The system of equations (3.26) is replaced by the
It is noted here that the eective used in SWAN is not equivalent to that of WAM: the former is a factor 2 larger.
3

56 following, iteration-dependent system

Chapter 3

N s N s 1 + A Ns = b

(3.32)

with a pseudo time step. The rst term of (3.32) controls the rate of convergence of the iteration process in the sense that smaller updates are made due to decreasing , usually at the cost of increased computational time. To deal with decreasing time scales at increasing wave frequency, the amount of under-relaxation is enlarged in proportion to frequency. This allows a decrease in the computational cost of under-relaxation, because at lower frequencies larger updates are made. This frequency-dependent under-relaxation can be achieved by setting 1 = , where is a dimensionless parameter. The parameter will play an important role in determining the convergence rate and stability of the iteration process. Substitution in (3.32) gives (A + I ) N s = b + N s1 . (3.33)

When the steady state is reached (i.e. s ), system (3.33) solves A N = b since, N is a xed point of (3.33). Suitable values for must be determined empirically and thus robustness is impaired. For increasing values of , the change in action density per iteration will decrease in the whole spectrum. The consequence of this is twofold. Firstly, it allows a much broader frequency range in which the action balance equation (3.2) is actually solved without distorting convergence properties. Secondly, the use of the limiter will be reduced because more density changes will not exceed the maximum change (3.30). Clearly, this eect may be augmented by increasing the value of in (3.30). To allow proper calculation of the second-generation rst guess of the wave eld (see Section 3.3), under-relaxation is temporarily disabled ( = 0) during the rst iteration. Whereas this measure is important in achieving fast convergence, it does not aect stability, since the second-generation formulations do not require stabilization.

3.7

Stopping criteria

In general, the iterative method should be stopped if the approximate solution is accurate enough. A good termination criterion is very important, because if the criterion is too weak the solution obtained may be useless, whereas if the criterion is too severe the iteration process may never stop or may cost too much work. Experiences with SWAN have shown that the present criteria (3.28) and (3.29) are often not strict enough to obtain accurate results after termination of the iterative procedure. Thus, criteria (3.28) and (3.29) are necessary but not sucient. It was found that the iteration process can converge so slowly s 1 s that at a certain iteration s the dierence between the successive iterates, Hm 0 Hm0 , can be small enough to meet the convergence criteria, causing the iteration process to stop,

Numerical approaches

57

even though the converged solution has not yet been found. In particular, this happens when convergence is non-monotonic such that the process is terminated at local maxima or minima that may not coincide with the converged solution. Furthermore, it became apparent that, unlike Hm0 , the quantity Tm01 is not an eective s 1 s 1 s measure of convergence. It was found that the relative error in Tm01 , i.e. |Tm 01 Tm01 |/Tm01 , does not monotonically decrease near convergence, but keeps oscillating during the iteration process. This behaviour is due to small variations in the spectrum at high frequencies, to which Tm01 is sensitive. This behaviour is problematic when any form of stricter stopping criterion is developed based on Tm01 . Therefore, in the improved termination criterion proposed, Tm01 has been abandoned as a convergence measure and only Hm0 , which displays more monotonic behaviour near convergence, is retained. Stiness and nonlinearity of the action balance equation are found to yield less rapid and less monotone convergence. Ferziger and Peri c (1999) explain the slow convergence in terms of the eigenvalue or spectral radius of the iteration process generating the sequence {0 , 1 , 2 , ...}. They show that the actual solution error is given by s s+1 s , 1 (3.34)

where denotes the steady-state solution and is the spectral radius indicating the rate of convergence. The smaller , the faster convergence. This result shows that the solution error is larger than the dierence between successive iterates. Furthermore, the closer is to 1, the larger the ratio of solution error to the dierence between successive iterates. In other words, the lower the rate of convergence of the iteration process, the smaller this dierence from one iteration to the next must be to guarantee convergence. The stopping criterion of SWAN could be improved by making the maximum allowable relative increment in Hm0 a function of its spectral radius instead of imposing a xed allowable increment. By decreasing the allowable relative increment as convergence is neared, it would be possible to delay run termination until a more advanced stage of convergence. Such a stopping criterion was used by, e.g. Zijlema and Wesseling (1998). This criterion is adequate if the iteration process converges in a well-behaved manner and < 1 for all iterations. However, due to nonlinearities SWAN typically does not display such smooth behaviour. Therefore, this criterion may be less suited for SWAN. An alternative way to evaluate the level of convergence is to consider the second derivative or curvature of the curve traced by the series of iterates (iteration curve). Since the curvature of the iteration curve must tend towards zero as convergence is reached, terminating the iteration process when a certain minimum curvature has been reached would be a robust break-o procedure. The curvature of the iteration curve of Hm0 may be expressed in the discrete sense as
s s m s s 1 s 2 (H 0 ) = H m0 2H m0 + H m0 ,

(3.35)

58

Chapter 3

s m where H 0 is some measure of the signicant wave height at iteration level s. To eliminate s m the eect of small amplitude oscillations on the curvature measure, we dene H 0 s 1 s (Hm0 + Hm0 )/2. The resulting curvature-based termination criterion at grid point (i, j ) is then s 1 s 2 s 3 s | Hm 0 (i, j ) (Hm0 (i, j ) + Hm0 (i, j )) + Hm0 (i, j )| < C , s = 3, 4, ... , s 2H m 0 (i, j )

(3.36)

where C is a given maximum allowable curvature. The curvature measure is made nons dimensional through normalization with Hm 0 . Condition (3.36) must be satised in at least 98% of all wet grid points before the iterative process stops. This curvature requirement is considered to be the primary criterion. However, the curvature passes through zero between local maxima and minima and, at convergence, the solution may oscillate between two constant levels due to the action limiter, whereas the average curvature is zero. As safeguard against such a situation, the weaker criterion (3.28) is retained in addition to the stricter criterion (3.36).

3.8

Governing equations in curvi-linear co-ordinates

xi,j , yi,j , i = 1, ..., M , j = 1, ..., N i = 1, ..., M , j = 1, ..., N (3.37) (3.38)

A curvi-linear grid is characterized by the co-ordinates of the grid points, i.e.

The 4-sweep method is unchanged, so in the rst sweep action densities in the points (i 1, j ) and (i, j 1) are used to compute the action densities in the point (i, j ). Numerical approximations are obtained by a 2-dimensional Taylor expansion with respect to the point (xi,j , yi,j ). The dierences in quantities between neighbouring grid points in the curvi-linear grid are denoted as follows: x1 = xi,j xi1,j , and x2 = xi,j xi,j 1 , y2 = yi,j yi,j 1 , F2 = Fi,j Fi,j 1 (3.40) y1 = yi,j yi1,j , F1 = Fi,j Fi1,j (3.39)

The partial derivatives can be found from the 2-dimensional Taylor expansions: F1 = and F2 = F F x1 + y1 x y F F x2 + y2 x y (3.41)

(3.42)

Numerical approaches It follows that the partial derivatives can be approximated by F y2 F 1 y1 F 2 x [D ] and x1 F 2 x2 F 1 F y [D ] [ D ] = y2 x1 y1 x2

59

(3.43)

(3.44) (3.45)

where Thus, in curvi-linear co-ordinates the complete propagation terms (including time-derivative, but ignoring dependence on and temporarily) read: 1 + + + (Dx,1 + Dx,2 )c+ x,i,j + (Dy,1 + Dy,2 )cy,i,j Ni,j t Ni,j + Dx,1 (cx N )+ i1,j Dy,1 (cy N )i1,j t + + Dx,2 (cx N )+ i,j 1 Dy,2 (cy N )i,j 1 = Si,j where Dx,1 =

(3.46)

y2 x2 y1 x1 , Dy,1 = , Dx,2 = , Dy,2 = (3.47) [D ] [D ] [D ] [D ] Here, the superscript + denotes the new time level t, and the old time level t t. The equation for a stationary computation is found by putting 1/t to 0. Again, the marching method is stable as long as the propagation direction towards the point (i, j ) is enclosed between the lines connecting this point with its neighbours (i 1, j ) and (i, j 1). It can be shown that this is the case if Dx,1 cx + Dy,1 cy 0 and Dx,2 cx + Dy,2 cy 0 (3.48) This set of criterions enables the SWAN program to decide whether a certain spectral direction does belong in the sweep which is being processed (in this the rst sweep). In the second sweep, x1 = xi,j xi,j 1 , etc. and x2 = xi,j xi+1,j , etc. In the third sweep, x1 = xi,j xi+1,j , etc. and x2 = xi,j xi,j +1 , etc. In the fourth sweep, x1 = xi,j xi,j +1 , etc. and x2 = xi,j xi1,j , etc. Otherwise, all of the above equations and conditions remain the same. Conservation of action in the numerical approximation can be demonstrated for the triangle of which the corners are the three points (i, j ), (i 1, j ) and (i, j 1). If for each side of this triangle the ux is computed as the outer product of the average cN and the side itself, then the three uxes are exactly in balance assuming that the situation is stationary, and the source term is zero. In this case it is found that:
+ + [cx N ] + i,j (y2 y1 ) + [cx N ]i1,j (y2 ) + [cx N ]i,j 1 (y1 ) + + + [cy N ] + i,j (x1 x2 ) + [cy N ]i1,j (x2 ) + [cy N ]i,j 1 (x1 ) = 0

(3.49)

60

Chapter 3

3.9

Computation of force in curvi-linear co-ordinates

FORCE is the wave-driven stress, i.e. the force per unit surface driving the wave-driven current, expressed in N/m2 , is dened as the derivative of the radiation stresses: Sxx = g 1 n cos2 + n Edd 2 n sin cos Edd (3.50) (3.51) (3.52)

Sxy = Syx = g Syy = g

1 n sin2 + n Edd 2 Here, n is the ratio of group velocity and phase velocity, i.e. n= The force is then Fx = and Fy = cg k

(3.53) (3.54) (3.55)

Sxx Sxy x y Syx Syy x y

In order to compute the force, the derivative of the radiation stress tensor has to be taken. Let f be one of the components of the tensor. We have to derive expressions for f /x and f /y . Derivatives with respect to the computational grid co-ordinates and can easily be found. The transformation is based on: f f x f y = + x y and Hence, f = x and f = y f f x f y = + x y
f x f y y y x x

(3.56)

(3.57)

f y x y f y x x

f f + x x f f + y y f f,+1 f,1 2

(3.58)

(3.59)

Numerical approximations are quite simple: f+1, f1, f , 2 (3.60)

These expressions are also used for derivatives of x and y . On the boundaries of the computational region a one-sided approximation can be used.

Numerical approaches

61

3.10

Numerical treatment of obstacles

An obstacle is treated in SWAN as a line running through the computational grid, see Figure 3.4. When treating one grid point SWAN will rst determine whether one of the grid lines of the stencil crosses an obstacle; see Section 3.11 for the procedure to determine whether or not there is a crossing point. If there is a crossing it will fall back to the rst order upwind scheme.

Figure 3.4: An obstacle as a line in computational grid. In computing the action densities for the target grid point (point 0 in Figure 3.5), the
11111111111 00000000000 111 000 11111111111 00000000000 000 111 11111111111 00000000000 000 111 00000000000 11111111111 000 111 00000000000 11111111111 111 000 00000000000 11111111111 111 000 11111111111 00000000000 000 111 11111111111 00000000000 000 111 00000000000 11111111111 111 000 00000000000 11111111111 111 000 11111111111 00000000000 000 111 00000000000 11111111111 000 111 00000000000 11111111111 000 111 00000000000 11111111111 111 000 11111111111 00000000000 000 111 00000000000 11111111111 000 111 00000000000 11111111111 000 111 00000000000 11111111111 000 111 00 11 00000000000 11111111111 111 000 00 11 11111111111 00000000000 000 111 00 11 00000000000 11111111111 0 1 00000000000 11111111111 0 1 2

obstacle
Figure 3.5: Schematic sketch of transmission in SWAN. contribution of a neighbouring grid point (point 1 in this case) is reduced by Kt2 , if the connection between the two grid points crosses the obstacle. (Note that the power 2 comes from the denition of the transmission coecient which is in terms of wave height). The contribution from point 2 in the computation of point 0 is not reduced because there is no obstacle crossing in between; the program takes Kt = 1 for this point.

62

Chapter 3

A consequence of the above procedure is that the results are the same as long as the obstacle crosses the same grid lines. Thus the results for the situation shown would be the same if the obstacle would be longer as long as the end would be in the same mesh. Another consequence is that an obstacle has to cross at least a few grid lines in order to have a noticeable eect, see Figure 3.4. After the transmission coecient has been calculated, it is used in the propagation terms of the action balance equation. In curvi-linear coordinates, the propagation terms (including time-derivative, but ignoring dependence on and temporarily) read: 1 + + + (Dx,1 + Dx,2 )c+ x,i,j + (Dy,1 + Dy,2 )cy,i,j Ni,j t N 2 + 2 + i,j Dx,1 (cx Kt, 1 N )i1,j Dy,1 (cy Kt,1 N )i1,j t 2 + 2 + + Dx,2 (cx Kt, 2 N )i,j 1 Dy,2 (cy Kt,2 N )i,j 1 = Si,j

(3.61)

In order to simplify the procedure, a reected wave in a grid point is calculated from the incident wave components in the same grid point. This introduces numerical inaccuracies, but that is not uncommon in numerical models. A basic condition in numerical models is that the approximation approaches the correct limit as smaller and smaller step sizes are used. A reected wave component in the target grid point 0, as illustrated by the arrow pointing away from the obstacle in Figure 3.6, would get contributions from grid points 1 and 2 if

11111111111 00000000000 11111111111 00000000000 0000 1111 11111111111 00000000000 0000 1111 00000000000 11111111111 0000 1111 00000000000 11111111111 1111 0000 00000000000 11111111111 0000 1111 00000000000 11111111111 0000 1111 11111111111 00000000000 0000 1111 00000000000 11111111111 1111 0000 00000000000 11111111111 1111 0000 00000000000 11111111111 0000 1111 00000000000 11111111111 0000 1111 00000000000 11111111111 0000 1111 00000000000 11111111111 1111 0000 11111111111 00000000000 0000 1111 00000000000 11111111111 0000 1111 00000000000 11111111111 0000 1111 00000000000 11111111111 0000 1111 00 11 00000000000 11111111111 1111 0000 00 11 11111111111 00000000000 0000 1111 00 11 00000000000 11111111111 0 1 00000000000 11111111111 0 1 2

obstacle

Figure 3.6: Schematic sketch of reection in SWAN. there would not be an obstacle. If the obstacle is there in the way shown, the contribution from point 2 is unchanged, but the contribution from point 1 is reduced by a transmission coecient and

Numerical approaches partially replaced by the reection of an incoming component in point 0.

63

Reection only works when both grid points 0 and 1, as shown in Figure 3.6, are wet points. This implies that obstacle lines are only eective when bordered by wet points on both sides of the obstacle.

3.11

Crossing of obstacle and grid line

In the procedure for obstacles it is necessary to determine the crossing point of the obstacle and a grid line in the computational grid. The obstacle is composed of straight sides. Let one side have the end points x3 = (x3 , y3 ) and x4 = (x4 , y4 ). The end points of the grid line (both computational grid points) are x1 = (x1 , y1 ) and x2 = (x2 , y2 ). The crossing point must obey the following equation x1 + ( x2 x1 ) = x3 + ( x4 x3 ) where both and must be between 0 and 1. It follows that = and (x1 x3 )(y2 y1 ) (y1 y3 )(x2 x1 ) (x4 x3 )(y2 y1 ) (y4 y3 )(x2 x1 ) (3.63) (3.62)

(x1 x3 )(y4 y3 ) (y1 y3 )(x4 x3 ) (3.64) (x4 x3 )(y2 y1 ) (y4 y3 )(x2 x1 ) If the denominator in both expressions is zero, the two lines are parallel and it is assumed that there is no crossing. =

3.12

Integration over

Two methods are considered in SWAN for integration over frequency space. The rst method is the common trapezoidal rule. We consider the following integration I=
m 0

f E ( )d

(3.65)

where m is the highest spectral frequency and f is an arbitrary function. Usually, this function may be p , p or k p with p a power. We assume a discrete (logarithmic) distribution of frequencies: i , i = 1, ..., m. The approximation of (3.65) is as follows:
m

1 (fi1 i1 Ni1 + fi i Ni )(i i1 ) 2

(3.66)

The contribution by the tail needs to be added as well, as follows. The tail of the energy density is proportional to P . We have,
m

R P d = m m P 1

(3.67)

64

Chapter 3

P = f (m ). Hence, Assuming that a function f has a tail with power P , so that Rm m

f ( )d =

This integration is only valid if P > 1.

m f ( m ) P 1

(3.68)

The second technique for integration over makes use of the logarithmic discrete distribution of frequencies. We introduced two variables in SWAN: FRINTF and FRINTH. The rst is equal to ln(i+1 /i ), the latter to i+1 /i . Hence, i = ei with = ln(i+1 /i ) and can be approximated as = / . The integral over a function of , i.e. f ( ) is transformed as follows: f ( )d = f ( )ei di = f ( )di (3.69)

Thus, the integral can be approximated as f ( )d f i i (3.70)

The boundaries of a mesh in space are i / i+1 /i and i i+1 /i . Computation of the contribution by the tail is done as follows. It is assumed that in the tail the energy density is proportional to P . Furthermore, the discrete integration extends to M m , where M = 1 + / . Then the contribution by the tail is:
2 =0 M m

1P m M 1P m P = dd = R 1 Rm P P 1 (P 1)M

(3.71)

Assuming that a function f has a tail with power P , the integral over f has a tail contribution of 2 m f ( )dd = f ( m ) (3.72) (P 1)M P 1 =0 M m Since, M is close to 1, the tail factor can be approximated as m m (P 1)M P 1 (P 1)(1 + (P 1)(M 1)) (3.73)

In the SWAN program, we have M =FRINTH, P =PWTAIL(1) and m =MSC. The value of P depends on the quantity that is integrated. For instance, in the computation of k , P = P 2n 1. Note that it is required that P > 1, otherwise the integration fails.

3.13

Transformation from relative to absolute frequency

Internally, SWAN use action density as function of direction and relative (angular) frequency. Users may want to obtain results in terms of absolute frequency, if only because

Numerical approaches measurements were taken at xed positions.

65

Two modications of the SWAN model that were needed to supply the information to the users are the computation of integrated quantities such as average absolute frequency and the transformation of action or energy density. The average absolute frequency is determined as follows: = E (, )dd E (, )dd (3.74)

The transformation of action or energy density from relative frequency to absolute frequency is complicated because the mapping is not one-to-one, and therefore the Jacobian can become innite. The value of is determined by = W ( ). The transformation is designed such that the following requirements are met: If current velocities tend to zero the action densities for absolute frequency become identical with the densities with respect to relative frequency. The total energy density with respect to relative density is identical with the total energy density with respect to absolute density. Furthermore, it is assumed that the distribution of absolute frequencies is the same as the distribution of relative frequencies. In the continuous model the mapping is done by E (.) = E (, ) ( W ( ))d (3.75)

This relation is discretized whereby the energy density is assumed to be constant over intervals from i /M to M i .

3.14

Interpolation of spectra

The interpolation of spectra in SWAN, both in space and time, is a slight modication of the procedure as used in WAM. This procedure is not a simple (spectral) bin-by-bin interpolation because that would cause reduction of the spectral peak if the peaks of the original spectra do not coincide. It is an interpolation where the spectra are rst normalized by average frequency and direction, then interpolated and then transformed back. The average frequencies of the two origin spectra are determined using the frequency moments of the spectra mi,k = Ni (, ) k dd (3.76)

66

Chapter 3

with i=1,2 (the two origin spectra) and k =0,1 (the zero- and rst frequency moments of these spectra). Then mi,1 i = (3.77) mi,0 The average frequency for the interpolated spectrum is calculated as = (w1 m1,1 + w2 m2,1 )/(w1 m1,0 + w2 m2,0 ) (3.78)

where w1 is the relative distance (in space or time) from the interpolated spectrum to the rst origin spectrum N1 (, ) and w2 is the same for the second origin spectrum N2 (, ). Obviously, w1 + w2 = 1. The average directions of the two origin spectra are determined using directional moments of the spectra: mi,x = Ni (, ) cos()dd (3.79) and mi,y = Ni (, ) sin()dd mi,y ) mi,x (3.80) with i=1,2. The average direction is then i = atan( (3.81)

The average direction of the interpolated spectrum is calculated as = atan[ w1 m1,y + w2 m2,y ] w1 m1,x + w2 m2,x (3.82)

Finally the interpolated spectrum is calculated as follows: N (, ) = w1 N1 [ 1 /, ( 1 )] + w2 N2 [ 2 /, ( 2 )] (3.83)

3.15

Computation of breaking source term

2 Hmax 8

The surf breaking dissipation of Battjes and Janssen (1978) reads Dtot = BJ Qb (3.84)

The surf breaking source term for each spectral bin i is Si = with the normalized total dissipation Qb = BJ D <0 B (3.86) Dtot Ei Ei = D Etot (3.85)

Numerical approaches and 8Etot B= 2 = Hmax Hrms d

2

67

(3.87)

depends on E through B ), we Since, the source term is strongly nonlinear in E (since D apply the Newton linearisation to approximate the source term at iteration level n + 1, as follows: n S n+1 n Si DEi + (Ein+1 Ein ) (3.88) E i In SWAN, this approximation has been slightly adapted for reasons of numerical stability; in , is replaced by DE in+1 . This preserves positivity the rst term in the right-hand side, DE of energy density E , if the following inequality holds S <0 E We derive an expression for this derivative as follows. From (3.85), we have D S |i = |i Ei + D E E is a function of B which is proportional to E , so The normalized dissipation D D S |i = | i Bi + D E B Since, Qb is a function of B , we get (using the quotient rule) BJ Qb S |i = E B (3.91) (3.89)

(3.90)

(3.92) (3.93)

Since,

the derivative of Qb is found by dierentiating this with respect to B : Q b + ln Qb + Hence, Q b = B Q =0 Qb b (3.94)

1 Qb + B ln Qb = 0

ln Qb Qb Qb 1 (3.95) = 1 B /Qb B Qb B using Eq. (3.93). Now, Q b > 0, because 0 < Qb < 1 and B > Qb . Substitution in (3.92) gives S Qb 1 | i < 0 |i = D (3.96) E Qb B Finally, the approximation of the source term reads Sin+1 1 + Qb 1 =D Qb B
n i

Qb 1 | n En Ein+1 D Qb B i i

(3.97)

68

Chapter 3

3.16
3.16.1

On the approximation of refraction on coarse grids in large-scale SWAN applications

Introduction

In some large-scale applications SWAN is known to produce seemingly unstable results. One of the causes of the behaviour may be the fully implicit treatment of the refraction term. In this treatment the value of c is determined at the point where the action densities are to be updated. This is reasonable if its value does not vary too much over one (spatial) step. This is true in most coastal applications. In large scale applications, however, the depth may vary from one grid point to the next by a factor of 100 or more. Then the value of c at the shallowest grid point is not representative anymore of the interval between the two grid points, and it is justied to limit the value to one that is representative. The problem now is to nd a limitation which on the one hand guarantees smooth behaviour in large-scale applications and which has no inuence on small-scale applications.

3.16.2

The problem with refraction on coarse grids and how to deal with it
N c N c N Stot + x [(cg + U )N ] + + = . t

We consider the following action balance equation (3.98)

This equation is discretized using the Method of Lines (MOL) in which the time integration is chosen independently from the space discretization. The propagation terms are approximated employing higher order schemes in some forms of blending, i.e. upwind and central. Time integration is based on the implicit Euler scheme while the source terms are linearized according to the well-known Patankar rules. However, the propagation of action density in the geographic space is done in a carefully manner without solving a huge matrix, such that unconditional stability can be guaranteed. This approach is based upon the four-direction Gauss-Seidel iteration technique. For details, see Booij et al. (1999) and Zijlema and Van der Westhuysen (2005). It is interesting to note that the above described approach is equivalent to the application of the well-known semi-Lagrangian approximation4 . To this purpose we consider the material derivative of the action density N N N dN + cx + cy dt t x y
4

(3.99)

In fact, it is nothing more than the method of characteristics.

Numerical approaches

69

with (cx , cy ) = cg + U the propagation velocity vector in the geographical space. This material derivative indicates that the time rate of change is computed along the wave characteristics dened by the following ODEs dx = cx , dt dy = cy . dt (3.100)

For the purpose of illustration the spatial derivatives are replaced by rst order upwind dierences. If we assume cx and cy positive during the rst sweep, see Figure 3.7, then the
y

(i1,j)

(i,j)

90o 0o

(i,j1)

Figure 3.7: Numerical scheme for wave propagation in geographic space with below the rst quadrant for which the waves are propagated, and right the stencil. approximation of Eq. (3.99) is as follows
n+1 +1 n+1 n Ni,j Ni,j N n+1 Nin N n+1 Ni,j 1,j 1 + cx i,j + cy i,j . t x y

(3.101)

Note that the time integration is based on the rst order implicit Euler scheme. Furthermore, t is the time step, and x and y are the mesh spaces. This approximation can be rewritten as cx cy 1 + + t x y
n+1 Ni,j

cx n+1 cy n+1 1 n+1 Ni,j Ni1,j N t x y i,j 1

(3.102)

and this can be interpreted as an approximation of the material derivative, Eq. (3.99), as follows n+1 Nin Ni,j ,j (3.103) T

70 with and

Chapter 3 1 1 cx cy = + + T t x y i = i p , j = j q , p=

(3.104)

cx t cy t , q= . (3.105) x y Note that p and q are the Courant numbers. They are not integers, and therefore (i , j ) is not a grid point. This point, however, lies on the wave characteristic. The quantity n in (i , j ) which is being convected Nin ,j can be interpreted as the value of N at time t in (i, j ) in a lapsed time t. This value is simply obtained from the interpolation of the n+1 +1 n+1 surrounding values Ni,j , Nin 1,j and Ni,j 1 in the (t, x, y )plane. In this case, there is no restriction on time step t as the characteristic lies inside the computational stencil. An important assumption made in this consideration is that cx and cy do not much vary over a time step. This is reasonable as the wave characteristics are more or less non-curved lines (during the time step), as the propagation in the geographic space is slowly time varying. However, this is generally not the case for the spectral propagation. In particular, wave refraction due to bottom or current gradients is typically a fast time-varying process. Often the associated time scale is much shorter than the desired time step used in a numerical simulation. A consequence of this is that the energy can travel in one time or distance step over more than the length of one sweep (which in the absence of a current is 90o , see Figure 3.7). This is especially the case when the bathymetry is poorly resolved by the model as will be shown later. Thus, for physical consistency, a restriction on the time step should be imposed in order for the wave direction not to cross the boundaries of a quadrant in the spectral space. Specically, the Courant number based on T and (i.e. directional bin) should not exceed unity, that is,

| c | T 1 (3.106) with c the turning rate and a user-dened maximum Courant number. For example, consider the rst sweep, see Figure 3.7, the boundaries of the rst quadrant are the lines 0o and 90o . Next, we consider the directional sector in the spectral space associated with the considered sweep, see Figure 3.8. We shall show that condition, Eq. (3.106), is sucient to Cr

=0

=90

Figure 3.8: Directional sector associated with the rst sweep. assure that wave energy propagating in any bin in direction will not cross the boundaries of the directional sector, except for the rst and last bins. Since, by denition, d = c , (3.107) dt

Numerical approaches we may approximate this, as follows, n+1 n + c T .

71

(3.108)

For a given n, we choose an arbitrary point, n , inside the directional sector; see Figure 3.8. If n then, Eqs. (3.108) and (3.106) imply |n+1 n | , and hence the chosen point at next time step, n+1 , still lies inside the directional sector. If n < , i.e. in the rst bin, then one obtains n+1 = n + or n+1 = n 1 T [(1 n )c ( = 0o ) + n c ( = )] (3.109)

T T T n c ( = 0o ) + c ( = 0o ) + c ( = ) . (3.110) If c ( = 0o ) > 0 and c ( = ) > 0 then, the point at next step will be keep inside the directional sector, if Eq. (3.106) holds. In other cases the energy is leaving through boundary = 0o , though the change in direction is limited to the adjacent directional bin of the other quadrant. This holds also for the last bin and the corresponding right boundary = 90o . Note that Eq. (3.106) is not required for the stability of the method but contributes to improve its physical accuracy. For instance, in a large-scale application, one often applies coarse (nested) grids with a desired time step. Both the mesh width and the time step might be too large to represent wave refraction appropriately. This is readily seen as follows. We consider stationary, long-crested waves in (x, )space, propagating at an angle with the positive xaxis. There are no currents. There is bottom gradient along this axis so that refraction is present. Hence, the Courant number becomes (cf. Eq. (3.106)) Cr = | c | x | cg | (3.111)

with cg the group velocity. The turning rate is given by 1 h , (3.112) k h m where k is the wave number, is the frequency, h is the water depth and m is the coordinate along the wave crest (i.e. orthogonal to the propagation direction). On a coarse grid, both the mesh width x and the depth gradient h/m, and thereby c , can be very large. This implies that the CFL condition may be violated, i.e. Cr > 1. c = Yet the value of c in the shallowest grid point can be simply too large due to large bottom slopes over one mesh width so that wave energy will change direction over more than some directional bins or even the directional sector. To prevent this artefact a limitation on c seems to be justied. Recalling Eq. (3.106), a limitation would be | c | | cx | | cy | 1 + + t x y . (3.113)

72

Chapter 3

Often the desired time step is such that the rst term between brackets can be safely neglected. This implies a slightly more restriction on the turning rate and, in turn, a safety margin in the CFL condition, as follows | c | | cx | | cy | + x y . (3.114)

It must be stressed that this limitation may aect the solution locally depending on . In fact, we need to nd a good estimate for c which solely depends on . However, it is unlikely that this measure aects the solution nearshore or on ne grids, since the turning rate will not too much vary over one (spatial) step. This is an eective survival measure in the sense that it prevents the excessive refraction without deteriorating the solution elsewhere. In any case, the turning rate is limited by accuracy, not by stability!

3.16.3

A historical overview of limitation on c

The problem with refraction on a coarse grid showing some inaccurate results has been known for a long time. This issue had received some attention by Nico Booij for the rst time in November 1998. His basic idea to x this problem is as follows. We consider a case with parallel depth contours within one sector, see Figure 3.9. We assume that grid
(i1,j) (i,j)

(i,j1)

Figure 3.9: Geographic grid with parallel depth contours. point (i, j ) is in shallow water. The other two grid points (i 1, j ) and (i, j 1) are in deeper water. Let n be the coordinate along the wave rays. Then according to Snels law (Holthuijsen, 2007, NOTE 7A, pg. 207), we have 1 dc d = tan . dn c dn Here will be of the order of 45o . So, we get d 1 dc = . dn c dn (3.115)

(3.116)

Numerical approaches The slope at grid point (i, j ) determines the value of d/dn. In shallow water, c = d 1 dh = . dn 2h dn This may be approximated as follows 1 h hi,j d dn 2hi,j n

73 gh, so

(3.117)

(3.118)

with h the water depth in one of the neighbouring grid points (i 1, j ) and (i, j 1). In the numerical procedure d/dn is constant over a spatial step, so the change in direction over a step is h hi,j d n = . (3.119) dn 2hi,j In order to maintain stability the change of direction must remain below 90o . Consequently, we obtain h hi,j hi,j . (3.120) In the program the factor is replaced by a user-determined factor . Hence, the depths in surrounding grid points are reduced to hi,j , if they are larger than this value. It should be noted that this approach was outlined in an unpublished note. Our experience with this approach is that it seems not eective enough. Another try was given by Erick Rogers in the midsummer of 2007. His main motivation was refraction due to mud (non-rigid seaoor) which result in a dispersion relation that varies with (x, y ). Currently, wave refraction can only be caused by depth variation (or current variation); see Eq. (3.112). As an alternative, the turning rate can be formulated in terms of wave number as follows (see Holthuijsen, 2007, pg. 210, footnote 4) c = cg k . k m (3.121)

The conjecture of Erick Rogers was that Eq. (3.121) diers from Eq. (3.112) due to numerics. His experiments suggested that Eq. (3.121) with coarse resolution yields results that are similar to those using Eq. (3.121) or Eq. (3.112) with high resolution. Though this seems to be the case with the circular shoal case (matching a similar feature in the southern California Bight as described in [157]), our general impression, however, is that this is generally not the case. The reason is that k/m can be large as well, if the bottom slope is large (not necessarily of the same order). Moreover, current gradients may also lead to an excessive wave refraction (i.e. c k/|k | U /m). We have tried this approach for the Katrina hurricane at the Gulf of Mexico with an unstructured grid coupled to ADCIRC (Dietrich et al., 2011). The desired eect did not show up, i.e. the excessive turning rate of waves at diverse locations remained during the whole simulation.

74

Chapter 3

Chapter 4 Wave boundary and initial conditions

To obtain the numerical solution of the action balance equation (2.16), the wave boundary and initial conditions should be provided. The incoming wave components at the up-wave boundaries in the SWAN model are specied by a two-dimensional spectrum. Several options are available: A parametric one-dimensional spectrum with a certain imposed directional distribution. An example is a Jonswap spectrum. A discrete one-dimensional spectrum with a certain imposed directional distribution. This is often obtained from measurements. A discret two-dimensional spectrum. This may be obtained from other SWAN runs or other models, e.g. WAM and WAVEWATCH III. For the parametric one-dimensional spectrum, the following optional forms have been recommended: a Pierson-Moskowitz spectrum (Pierson and Moskowitz, 1964), a Jonswap spectrum (Hasselmann et al., 1973) and a Gaussian-shaped spectrum. The boundaries in frequency space are fully absorbing at the lowest and the highest discrete frequency. So, energy can freely propagate across these boundaries and thus total energy might not be conserved in some cases. However, a diagnostic tail f m (m = 4 or m = 5) is added above the high frequency cut-o, which is used to compute nonlinear wave-wave interactions at the high frequencies and to compute integral wave parameters. When the directional space is a closed circular, no directional boundary conditions are needed. However, for reasons of economy, SWAN has an option to compute only wave components in a pre-dened directional sector. In this case, the boundaries of this sector are fully absorbing (action density might be removed from the model by refraction). To facilitate the integration process of the action balance equation, wave boundary conditions in geographical space need to be provided. The boundaries of the computational grid in SWAN are either land or water. In case of land there is no problem. The land does not generate waves and in SWAN it absorbs all incoming wave energy. But in the case of 75

76

Chapter 4

a water boundary there is a problem. If observations are available, they can be used as inputs at the boundary. In case no wave conditions are given along the boundary, SWAN assumes that no waves enter the model and waves can leave the model freely along that boundary. This assumption results in errors. Therefore, to get reliable results, especially for such case, the model boundaries must be placed far away from the area of interest. In case of nonstationary computation, the default initial spectra are computed from the local wind velocities using the deep-water growth curve of Kahma and Calkoen (1992), cut o at values of signicant wave height and peak frequency from Pierson and Moskowitz (1964). The average (over the model area) spatial step size is used as fetch with local wind. The shape of the spectrum is default Jonswap with a cos2 () directional distribution centred around the local wind direction. The rst guess conditions of a stationary run of SWAN are default determined with the second generation mode of SWAN. It is possible to obtain an initial state by carrying out a previous stationary or nonstationary computation.

Chapter 5 Implementation of 2D wave set-up

5.1 Methods

For the present purpose ows in shallow water are suciently described by the so- called shallow-water equation which consists of one continuity equation and two equations of motion (one for the x and one for the y component). For the present project the shallow water equation has to be simplied. The possibilities for such simplication are investigated in the remainder of this section. Wave setup is usually conned to narrow zones in the immediate vicinity of the shoreline. The size of such areas is small enough that the setup process can be considered to be quasi-stationary. Wave-induced currents are usually weak compared with e.g. tidal currents. Therefore it seems acceptable to neglect all terms in the equation of motion where the current velocity appears. Comparisons of SWAN results with results from a full 2-dimensional ow model have to show under which conditions the simplied equation provides acceptable results. Such investigations are outside the scope of the present project. This report presents a few cases with low overall current velocities where apparently the computed setup is reasonably in accordance with expectations. Deleting from the equation of motion all terms involving current velocities we retain an equilibrium between the wave-induced force and the gradient of the water table, i.e. Fk + gd =0, xk (5.1)

where is the setup, d the depth and Fk is the wave-induced force. In a one-dimensional model (5.1) can be used directly to compute the setup. In two dimensions (5.1) is a set of two equations with only one unknown, the setup . In order to reduce the number of equations to one, we use the observation by Dingemans (1997) that wave-driven currents 77

78

Chapter 5

are mainly due to the divergence-free part of the wave forces whereas the setup is mainly due to the rotation-free part of the force eld. We therefore take the divergence of eq. (5.1) to obtain the following elliptic partial dierential equation for : Fk + (gd )=0 xk xk xk (5.2)

This Poisson equation needs one boundary condition in each point of the boundary of the computational domain. Two types of boundary conditions are foreseen; the rst one is used on the open boundaries and on the shoreline where the shoreline is dened as the line where the depth is zero: Fn + gd =0 (5.3) n It is not possible to use this boundary condition on all boundary points because then there remains an unknown constant. So on on point, for which we take the boundary point with the largest depth the setup is assumed to be 0: = 0. The second type of boundary condition with given value of is also used in nested models. The setup computed in the larger model is used as boundary condition in the nested model. In the nested model the setup is given in all points of the outer boundary. On shorelines inside the area again eq. (5.3) is used. The Poisson equation (5.2) together with its boundary conditions will be solved numerically on a curvi-linear grid. The next section discusses the details of the method, and presents some results of preliminary computations with the above model. The actual design of the modications of the SWAN commands is presented in Appendix A, the design of the modications of the code is presented in Appendix B. After each iteration performed in SWAN new values of the setup are being calculated and added to the depth, so that the SWAN model incorporates the eect of setup on the wave eld. An output quantity SETUP is added so that the user can be informed about the magnitude and distribution of the wave setup.

5.2
5.2.1

Analysis and Results

Discretization of the 2D setup equation

Problem denition The equation to be solved has the following form: (Fk + gd )=0, xk xk (5.4)

Implementation of 2D wave set-up with the setup, d the depth and Fk a golf-induced, time-averaged force. In order to solve (5.4), the following types of boundary conditions may be applied Fn + gd = 0 at the boundary , n

79

(5.5)

with n the outward direct normal. This is a so-called Neumann condition. The setup is xed upon an additive constant. = given at the boundary . This is boundary condition of Dirichlet type. At beaches always the Neumann condition (5.5) is applied. In order to solve (5.4) with boundary conditions (5.5) and (5.6) a boundary tted, vertex centered nite volume method is applied. The discretization is based on the method described in Van Beek et al. (1995). In the remainder of this Chapter we use k instead of gd. Discretization The physical domain is mapped onto a rectangular domain in the ( 1 , 2 ) plane, which is called the computational domain. All points of the domain are used, including the dry ones. Using the relation Zijlema (1996) (summation convection applied): 1 ( ) = ( g a ) , x g
( )

(5.6)

(5.7)

with a the components of the contravariant basevectors a() dened as (Segal et al, 1992): a ( ) = , g the Jacobian of the transformation: 2 2 1 g = a1 (1) a(2) a(1) a(2) . x . (5.8)

and

(5.9)

a() are the covariant base vectors dened by a ( ) = (5.10)

The contravariant base vectors follow immediately from the covariant ones due to: (1) 1 T = ( a2 (5.11) ga (2) , a(2) ) , (2) 1 T ga = ( a2 (5.12) (1) , a(1) ) .

80

Chapter 5

Application of (5.7) to equation (5.5) results in 1 ( ) ( ga (k + F )) = 0 . g (5.13)

Note that is a derivative in the Cartesian (x) direction and not in the direction. In the remainder we shall use the local numbering as given in Figure 5.1. The points (0,
(0, 2) (0, 1) (0, 0) (1, 0) (2, 0) (2, 2)

Figure 5.1: Local numbering in computational domain 0), (2, 0), (0, 2) and so on are the vertices of the cells. The integration cell for the nite volume method is dened by the cell (-1, 0), (0, -1), (1, 0), (0, 1). Integrating (5.13) over this cell gives 1 ( ) ( ga (k + F ))dx g ( ) ( ga (k + F ))d ga(1) (k + F )|(1,0) +
(1,0)

(5.14) ga(2) (k + F )|(0,1) ,

(0,1)

where x is the cell in the physical space and the cell in the computational domain. The four points (1, 0), (0, 1), (-1,0) and (0, -1) will be cell integration points. The covariant basis vectors a() are approximated by central dierences. a(2) |(0,1) = x(0,2) x(0,0) , a(1) |(1,0) = x(2,0) x(0,0) , (5.15) (5.16)

and by linear interpolation in other points. In these relations we have used that the step width in the computational domain is equal

Implementation of 2D wave set-up

81

to 1. The term needs special attention. Since it concerns derivatives in the x direction, whereas all derivatives in the computational domain are in the directions it is necessary to make some approximation. We approximate this term by the integration path method introduced in Van Beek et al. (1995). To that end is integrated in two independent directions 1 and 2 . This yields two equations to express x and y in values of neighbours. (x2,0 x),0 ) |(1,0) = 2,0 0,0 , (5.17) 1 1 ((x2,2 x2,2 ) + (x0,2 x0,2 )) |(1,0) = ((2,2 2,2 ) + (0,2 0,2 )) .(5.18) 2 2 (5.17), (5.18) may be considered as two sets of equations to express into values. Solution of this linear system results in: |(1,0) = |(0,0) c(1) + ( |(0,2) + |(2,2) )c(2) , with 1 2 1 2 1 (c(2) , c1 ( c2 (2) ) ; c = (1) , c(1) ) , C C 1 2 1 C = c2 c c c , (2) (1) (1) (2) c(1) = a(1) |(1,0) c(2) = a(2) |(0,1) + a(2) |(0,1) + a(2) |(2,1) + a(2) |(2,1) . c1 = (5.20) (5.21) (5.22)
(2,0) (0,2) (2,2)

(5.19)

A similar formula is applied for point (0, 1). Equation (5.14) together with expression (5.19) gives one row of the discretized equation. Treatment of the boundary conditions The boundary conditions at the outer boundary of the domain are relatively easy to implement. In case of Dirichlet boundary conditions the corresponding row of the matrix is made equal to 0 and the diagonal element is set to 1. The value of the boundary condition is lled into the right-hand side. Neumann boundary conditions are treated integrating over a half cell as sketched in Figure 5.2. In this case we get: ( ) ( ga (k + F )d 1 (1) (1,0) (0,1) ga (k + F )|(1,0) + ga(2) (k + F )|(0,0) . 2 (5.23)

Due to the Neumann boundary conditions the term in the boundary point (0, 0) vanishes.

82

Chapter 5

(0, 1) (-1, 0) (1, 0) (0, 0)

Figure 5.2: Half cell at boundary Mark that in this case we need to evaluate at the boundary. In order to do so we apply a one-sided integration path approach i.e. (x(2,0) x(1,0) ) |(1,0) = (2,0) (0,0) , ((x(2,2) x(2,0) ) + (x(0,2) x(0,0) )) |(1,0) = ((0,2) (0,0) ) + ((2,2) (2,0) ) (5.24) . Furthermore we need the values of a() in virtual cells, because we need the c() at the boundary. To that end we construct a row of virtual cells by extrapolating the co-ordinates of the boundary cells. The implementation of dry points Dry points complicate the software considerably. For the dry points itself there is no problem. In fact we make the corresponding row of the matrix, as well as the right-hand side element completely equal to zero. This is allowed since our linear solver is able to deal with zero rows. Dry points in the neighbourhood of wet points, however, also inuence the matrix for the wet point. Consider for example the integration point (1,0) in Figure 5.3. If (0, 0) is a

(0, 1) (2, 0) (-1, 0) (0, 0) (0, -1) (1, 0)

Figure 5.3: Dry point (2, 0) and wet point (0, 0) wet point and (2, 0) a dry point then we assume that at point (1,0) we have a Neumann

Implementation of 2D wave set-up

83

boundary condition. This means in fact that the contribution of the integration point (1,0) to the matrix and right-hand side is equal to zero. With respect to the evaluation of the gradient of with the integration path method one sided dierences are applied for those formulas involving (2,0) . This process is applied for all transitions from wet to dry points. As a consequence, in the case of a situation like in Figure 5.4 we make for point 2 zero. The reason is that in point 2 it is only possible to evaluate 1 and not 2 , and hence we

2 1

Figure 5.4: Wet points enclosed by a row of dry points have too few information to express in neighbour values. Building of the matrix and right-hand side With respect to the building of matrix and right-hand side we start by computing all contributions in the integration points. This is done by looping over the various integration points. Since the contribution of point (0,1) in cell (i, j ) is equal to that of point (0, -1) in cell (i 1, j ) it is sucient to loop over two sets of integration points only. Once we have computed the coecients in a set of integration points we must add these contributions, multiplied by some factor, to the matrix elements. This process is known as distribution. Hence the actual implementation is as follows: 1 Map indirect addressing to direct addressing. (subroutine swdct). 2 Compute and store a() for all integration points. First we use central dierences where possible and after that we apply linear interpolation for the remaining points. (subroutine swcova2d). 3 Compute ga() in all integration points using a() . (subroutine swjcta2d). For 2 integration directions do 4 Compute factors c() taking into account boundary eects and dry points. Compute contributions for matrix and right-hand side for integration points. (subroutine swtrad2d).

84

Chapter 5 5 Distribute contribution to correct matrix elements. (subroutine swdisdt2). 6 Fill essential boundary conditions. (subroutine swessbc). 7 Solve system of linear equations. (subroutine swsolve). This part is explained in more detail in the next Chapter. 8 Map from direct to indirect addressing. (subroutine swindct).

5.2.2

The iterative solver for the linear system

In this section we describe the mathematical methods, which are used to solve the system of linear equations as derived in Chapter 1. In Section 5.2.2 we consider the data structure used. Some properties of the matrix are given in Section 5.2.2. Due to the sparseness of the matrix we prefer an iterative solution method of Krylov subspace type. The details are described in Section 5.2.2. As is well known, Krylov subspace methods are only attractive when they are combined with suitable preconditioners (see Section ??). Finally our choices are summarised in Section ??. Data structure After the discretization of the Poisson equation in curvi-linear co-ordinates, one has to solve the following matrix vector system: Ax = f, (5.25)

where A is the discrete Poisson operator, x is an approximation of the setup (of the water level), and the right-hand-side vector f contains the eects of the boundary conditions and the forces due to the surface waves. In the solver it is very ecient to calculate with direct addressing, so dry points are included in the vector x. This implies that the dimension of x and f are xed and equal to M XC M Y C . In the discretization a 9-point stencil is used. That implies that only 9 matrix elements per row are non-zero. These elements are stored in a diagonal-wise way. So for this part N W KARR = 9. The rows corresponding to dry points are lled with zeroes except on the main diagonal where the value 1 is substituted. The value of x and f are taken equal to 0 at these points. Properties of the matrix The discrete operator is symmetric in the inner region. This means that ai,j = aj,i . Due to the boundary conditions the symmetry of the operator is lost. The reasons for this are:

Implementation of 2D wave set-up

85

When Dirichlet boundary conditions are used the known elements of x should be eliminated in order to keep the matrix symmetric. However this leads to a dierent dimension of A, x, and f , therefore the known elements are not eliminated. When dry points occur the derivation of the discrete boundary conditions is already complicated at the interface between wet and dry points. At this moment it is not clear how to discretize these conditions such that the resulting matrix is symmetric. These diculties motivate us to use a non-symmetric matrix. This is only a small drawback, because recently good methods have been developed to solve non-symmetric matrix vector systems. When Neumann conditions are used on all boundaries the resulting matrix is singular. The solution is determined up to a constant. We have to keep this in mind during the construction of the solution procedure. When Gauss elimination is used to solve equation (5.25), the zero elements in the bend of A become non-zero. This means that the required memory is equal to 2 M XC + 2 vectors. For M XC large, this leads to an unacceptable large amount of memory. Therefore we use an iterative solution method, where the total amount of memory is less than the memory used in the discretization procedure. The iterative solver In 1D cases, the wave-induced set-up is calculated in SWAN with a simple trapezoidal rule. In 2D cases, the Poisson equation of the divergence-free force eld is solved in SWAN with a modied Successive Over Relaxation (SOR) technique (Botta and Ellenbroek, 1985). The boundary conditions for this elliptical partial dierential equation are: at open boundaries: equilibrium between wave force and hydrostatic pressure gradient normal to the model boundary, at last grid points before shoreline: equilibrium between wave force and hydrostatic pressure gradient normal to the model boundary and at deepest boundary point: set-up is zero. The shoreline in SWAN moves as dictated by the wave-induced set-up. The set-up computations are available in both the recti-linear and curvi-linear grids.

86

Chapter 5

Chapter 6 Iterative solvers

6.1 Strongly Implicit Procedure (SIP)
AN = b (6.1)

We want to solve the following linear system of equations

where A is some non-symmetric penta-diagonal matrix, N is the wave action vector to be solved and b contains source terms and boundary values. The basis for the SIP method (Stone, 1968; Ferziger and Peri c, 1999) lies in the observation that an LU decomposition is an excellent general purpose solver, which unfortunately cannot take advantage of the sparseness of a matrix. Secondly, in an iterative method, if the matrix M = LU is a good approximation to the matrix A, rapid convergence results. These observations lead to the idea of using an approximate LU factorization of A as the iteration matrix M , i.e.: M = LU = A + K (6.2) where L and U are both sparse and K is small. For non-symmetric matrices the incomplete LU (ILU) factorisation gives such an decomposition but unfortunately converges rather slowly. In the ILU method one proceeds as in a standard LU decomposition. However, for every element of the original matrix A that is zero the corresponding elements in L or U is set to zero. This means that the product of LU will contain more nonzero diagonals than the original matrix A. Therefore the matrix K must contain these extra diagonals as well if Eq. (6.2) is to hold. Stone reasoned that if the equations approximate an elliptic partial dierential equation the solution can be expected to be smooth. This means that the unknowns corresponding to the extra diagonals can be approximated by interpolation of the surrounding points. By allowing K to have more non zero entries on all seven diagonals and using the interpolation mentioned above the SIP method constructs an LU factorization with the property that for a given approximate solution the product K 0 and thus the iteration matrix M is close to A by relation (6.2). 87

88

Chapter 6

To solve the system of equations the following iterations is performed, starting with an initial guess for the wave action vector N 0 an iteration is performed solving: U N s+1 = L1 K N s + L1 b (6.3)

Since the matrix U is upper triangular this equation is eciently solved by back substitution. An essential property which makes the method feasible is that the matrix L is easily invertible. This iterative process is repeated s = 0, 1, 2, ... until convergence is reached.

6.2

Successive Over Relaxation (SOR) technique

This section is under preparation. See also Botta and Ellenbroek (1985).

Chapter 7 Parallel implementation aspects

Domain decomposition methods have been successfully used for solving large sparse systems arising from nite dierence or nite volume methods in computational uid dynamics on distributed memory platforms. They are based, in essence, upon a partition of the whole computational domain in x-space into a number of contiguous, non-overlapping subdomains with each of them being assigned to a dierent processor. In this case the same algorithm performs on all available processors and on its own set of data (known as the SPMD programming model). Each subdomain can have multiple neighbors on each of its four sides. For this, a data structure is implemented to store all the information about the relationship of the subdomain and its particular neighbors. Next, each subdomain, look in isolation, is then surrounded by an auxiliary layer of one to three grid points originating from neighbouring subdomains. This layer is used to store the so-called halo data from neighbouring subdomains that are needed for the solution within the subdomain in question. The choice of one, two or three grid points depends on the use of propagation scheme in geographical space, i.e., respectively, BSBT, SORDUP or Stelling/Leendertse. Since, each processor needs data that resides in other neighbouring subdomains, exchange of data across boundaries of subdomains is necessary. Moreover, to evaluate the stopping criterion (3.28), global communication is required. These message passings are implemented by a high level communication library such as MPI standard. A popular distribution is MPICH which is free software1 and is used in the present study. Only simple point-topoint and collective communications have been employed. There are, however, some other implementation and algorithmic issues that need to be addressed.

7.1

Load balancing

The mapping of subdomains on processors should be chosen so as to distribute the computational load as equally as possible and to minimize the communication cost. Intuitively, it will be clear that we have to allocate contiguous blocks of equal numbers of grid points on each processor. However, in the context of SWAN applications to coastal areas, some dif1

Available from http://www-unix.mcs.anl.gov/mpi/mpich.

89

90

Chapter 7

culties arise. Firstly, wet and dry grid points may unevenly distributed over subdomains while no computations have to be done in dry points. Secondly, an unbalanced partition may arise during the simulation due to the tidal eect (dry points become wet and vice versa). In such a case, one may decide to adapt the partition such that it is balanced again (so-called dynamic load balancing). Finally, most end-users are not willing to determine the partitioning themselves, thus automatic support for partitioning the grids is desirable. In the present study, two well-established partition methods are applied. The rst is called stripwise partitioning in which the computational grid is cut along one direction, resulting in horizontal or vertical strips. The choice of cutting direction depends on the interface size of the strips which should be minimized. However, the communication volume, which is related to the total size of the interfaces, can be further reduced by means of recursive application of alternately horizontal and vertical bisection. This is known as Recursive Co-ordinate Bisection (RCB). Further details on these techniques and an overview on grid partitioning can be found, e.g. in Fox (1988) and Chrisochoides et al. (1994). . Within SWAN, the grid partitioning is carried out automatically on wet grid points only. The size of the subdomain equals the total number of wet points divided by the total number of subdomains. The implementation of a stripwise partitioning is as follows. First, an empty strip is created. Next, assign point-by-point to the created part until the size of that part has been reached. Thereafter, verify whether non-assigning wet points remain in the current strip. If so, these points will be assign to the same part too, otherwise create next empty strip. As a result, all strips have straight interfaces and include approximately the same number of wet grid points. Moreover, experiences with SWAN simulation have shown that the amount of computations in each wet grid point remains more or less constant during the simulation and hence, there is no need for dynamic load balancing. A nal remark has to be made considering grid partitioning. The above described methodology does not seem to have been implemented in spectral wave models before. In Tolman (20020, another way of distributing data over the processors is discussed: each pth wet grid point is assign to the same processor with p the total number of processors. The requirement of equal numbers of wet grid points per processor is provided automatically. However, it is impossible to compute the spatial wave propagation in an eective manner. The only alternative is to gather data for all grid points in a single processor before the calculation is performed. This will require a full data transpose, i.e. rearranging data distribution over separate processors. It is believed that this technique requires much more communication between processors than domain decomposition and therefore less suitable for SWAN.

7.2

Parallelization of implicit propagation schemes

Contrary to explicit schemes, implicit ones are more dicult to parallelize, because of the coupling introduced at subdomain interfaces. For example, concerning the four-sweep technique, during the rst sweep, an update of N (i, j, l, m) can be carried out as soon as N (i 1, j, l, m) and N (i, j 1, l, m) have been updated and thus it can not be performed

Parallel implementation aspects

91

in parallel. Parallelization of this implicit scheme requires modications. Ideally, the parallel algorithm need no more computing operations than the sequential one for the same accuracy. The simplest strategy to circumvent this problem consists in treating the data on subdomain interfaces explicitly, which in mathematical terms amounts to using a block Jacobi approximation of the implicit operator. In this context, we employ the RCB partition method, since it gives the required balanced, low-communication partitioning. This strategy possess a high degree of parallelism, but may lead to a certain degradation of convergence properties. However, this numerical overhead can be reduced by colouring the subdomains with four dierent colors and subsequently permuting the numbering of unknowns in four sweeps in accordance with the color of subdomains. Furthermore, each subdomain is surrounded by subblocks of dierent colors. See Figure 7.1. As a result, each coloured subdomain start with a dierent ordering of updates within the same sweep and

Figure 7.1: Four types of subblocks (red, yellow, green and black) treated dierently with respect to the ordering of updates (indicated by arrows) per sweep. thus reducing the number of synchronization points. This multicolor ordering technique has been proposed earlier, e.g. in Meurant (1988) and Van der Vorst (1989). Another strategy is based on the ideas proposed by Bastian and Horton (1991) and is referred here to as the block wavefront approach. It is demonstrated with the following example. First, we decompose the computational domain into a number of strips. In this example, we assume that these strips are parallel to y axis. Next, we start with the rst sweep. The processor belonging to the rst strip updates the unknowns N (i, 1, l, m) along the rst row j = 1. Thereafter, communication takes place between this processor and processor for strip 2. The unknowns N (i, 2, l, m) along j = 2 in strip 1 and N (i, 1, l, m) along j = 1 in strip 2 can be updated in parallel, and so on. After some start-up time all processors are busy. This is depicted in Figure 7.2. Finally, this process is repeated

92

Chapter 7

for the other three sweeps. Details can be found in the source code of SWAN. The block

Iteration 1
CPU 1 active CPU 2 CPU 3 idle idle

Iteration 2
CPU 1 active CPU 2 active CPU 3 idle

Iteration 3
CPU 1 CPU 2 CPU 3 active active active

*** *** *** *** * ** * ** ooo

*** *** *** *** * ** * ** * **

*** *** *** *** * ** * ** * **

*** *** *** *** * ** o oo + ++

*** *** *** *** * ** * ** o oo

*** *** *** *** * ** * ** * **

*** *** *** *** o oo + ++ + ++

*** *** *** *** * ** o oo + ++

*** *** *** *** * ** * ** ooo

Figure 7.2: Application of block wavefront approach for the rst 3 iterations during the rst sweep. Domain is divided into 3 vertical strips. Stars represent unknowns to be updated, circles mean that unknowns are currently updated and the plus signs indicate unknowns that have been updated. wavefront approach does not alter the order of computing operations of the sequential algorithm and thus preserving the convergence properties, but reduces parallel eciency to a lesser extent because of the serial start-up and shut-down phases (Amdahls law). This technique resembles much to the standard wavefront technique applied in a pointwise manner (unknowns on a diagonal are mutually independent and thus can be updated in parallel; for details, see Templates (1994), which has also been employed by Campbell et al. (2002) for parallelizing SWAN using OpenMP. The performance of the two discussed parallelization methods applied in the SWAN model has been discussed in (Zijlema, 2005). Numerical experiments have been run on a dedicated Beowulf cluster with a real-life application. They show that good speedups have been achieved with the block wavefront approach, as long as the computational domain is not divided into too thin slices. Moreover, it appears that this technique is suciently scalable. Concerning the block Jacobi method, a considerable decline in performance has been observed which is attributable to the numerical overhead arising from doubling the number of iterations. Furthermore, it may result in a solution that is computed to an accuracy that may not be realistic. In conclusion, parallelization with the block wavefront technique has been favoured and has been implemented in the current operational version of SWAN. A survey of other alternatives to the parallelization of the implicit schemes is given in Templates (1994).

Chapter 8 Unstructured mesh implementation

Since, the characteristic spatial scales of the wind waves propagating from deep to shallow waters are very diverse, a exible grid would be required to allow local renement of the mesh in areas of interest e.g., regions of strong bathymetry variations in estuaries and fjords, without incurring overhead associated with grid adaptation at some distance oshore. Traditionally, this can be achieved by employing a nesting technique. Although, this practise is very common for SWAN, it is generally recognized that this may lead to complicated programming with the corresponding signicant increase in computational eort. The use of unstructured grids, however, oers a good alternative to nested models not only because of the ease of local grid renement, either adaptive or xed, but also the high exibility to generate grids along coastline and around islands. The variable mesh is especially useful in coastal regions where the water depth varies greatly. Thus, the variable grid gives the highest resolution where it is most needed. Moreover, this can be automated to a large extent. Although, the CPU cost per grid point is often relative higher than cases with structured grids, this eect is probably more than oset by the reduction in the number of grid points. This chapter presents an unstructured grid procedure for SWAN. Details can also be found in (Zijlema, 2009, 2010). The numerical propagation scheme for structured grids is based on a four-direction Gauss-Seidel iteration technique and is accompanied by a fully implicit temporal discretization; see Section 3.3. Hence, SWAN is stable for any time step. Because of this nice property, this solution technique is tailored to unstructured grids.

8.1
8.1.1

Description of an unstructured grid

Denitions

We distinguish between two types of grids, namely structured and unstructured grids. A two-dimensional structured grid may contain quadrilaterals. These can be recti-linear or curvi-linear. The number of cells that meet each other in an internal vertex is always 4. 93

94

Chapter 8

In unstructured meshes this restriction is abandoned. Moreover, 2D unstructured grids usually consist of triangles or a combination of triangles and quadrilaterals, a so-called hybrid grid. The unstructured meshes that we consider in SWAN consist solely of triangles, also called cells. The edges of the triangles are called faces.

8.1.2

Relations between number of cells, vertices and faces

For a two-dimensional triangular mesh, the number of cells C , the number of boundary faces Eb and internal faces Ei are related according to: E b + 2E i = 3C (8.1)

The total number of faces E = Ei + Eb . With V the number of vertices and H the number of holes (islands), we have the following Eulers relation for a triangulation: C +V E =1H Usually, Eb << Ei and the number of holes H is negligibly small, so C 2V , E 3V (8.3) (8.2)

There are approximately twice as many cells as vertices in a triangular mesh. Therefore, it is an optimal choice to locate the action density in vertices as the number of unknowns is minimal on a given grid. Concerning the time-consuming evaluation of the physical processes representing the wave energy generation, dissipation and redistribution, this allows SWAN to save a considerable amount of computing time.

8.1.3

Conditions imposed to the grid

In order to avoid badly shaped grids, the grids must satisfy the following properties: The number of cells that meet at each vertex in the interior of the mesh must be at least 4 and at most 10. The angles inside each triangle must be smaller than a certain value. Let a and b be the tangential vectors of two faces of a triangle, then the angle between these two faces equals cos = |a||b| ab (8.4)

For safety, we do not allow for angles with cos < 0.8 or, equivalently, > 143o .

Unstructured mesh implementation

95

8.2

Some notes on grid generation

We briey outline some issues related to grid generation from a practical point of view. The process of grid generation can be dicult and time consuming. A common approach is proceeding from coarse to ne grid through renement in various ways. Generally, one would like to have an optimal grid in which areas where the bathymetry or evolution of the waves change rapidly require a higher resolution than areas where the physics or depth changes less. This goes around by having an indication how to determine the renement based on bathymetry or geometric variations through preliminary evaluations. To facilitate this procedure, a mesh generation package BatTri (Bilgili and Smith, 2003) is used. This grid generator is a public-domain, graphical Matlab interface to Triangle (Shewchuk, 1996). Triangle is a freely-distributed, two-dimensional Delaunay triangulator. An important key ingredient for the preparation of the grid for the model domain is bathymetry data. Boundary nodes, segments and holes can be created from this data with the use of the mesh editing options of BatTri. After checking and improving grid quality, the nal information on nodes and segments is forced into the triangulation of the domain. Triangle is called within BatTri to generate the actual grid. This triangulation includes only acute triangles. BatTri provides many pre-dened depth-dependent contraints for further mesh renement. From a numerical point of view, mesh renement is often directly related to properly resolve the shape of the wave, i.e. to keep the wavelength to grid size ratio relatively large. When wavelength decreases in shallower water, the grid size must decrease as well. Therefore, this criterion, called the h-renement, has the eect of using smaller cells in shallow water and larger cells in deeper water. Here, h refers to the water depth. Another useful criterion is known as the topographic length scale constraint, when one try to keep the ratio h/h less than one. Here, h equals the dierence between the maximum depth of a triangle and the minimum depth and h is the average depth. This criterion addresses the bathymetric slope and cells with a high value of h/h indicate areas of steep bottom topography that will need to be more nely resolved. When rening the grid, one must balance the need to fully meet the renement criteria with the desire to keep the triangle sizes from becoming too small. Thus, these criteria are generally imposed along with a minimum area constraint. The renement process is repeated iteratively until a nal grid with the appropriate resolution is obtained.

8.3
8.3.1

Numerical method
Discretization procedure

For the sake of clarity of the algorithm description below, we put all the terms but the time derivative and propagation term in the geographical space of Eq. (2.16) in one term F (x, , ): N + x [ c x N ] = F (8.5) t

96

Chapter 8

with cx = cg + U the geographic velocity vector. For the time being, we restrict ourselves to triangular meshes. However, other type of meshes can be employed as well, e.g. hybrid grids (consisting of both triangles and quadrilaterals). We consider a triangulation of a geographical domain in which Eq. (8.5) is solved; see Fig 8.1. Every vertex and all the triangles around this vertex are taken into

y x

Figure 8.1: An example of triangulation. account. Observe that the number of cells around a vertex can be dierent for all vertices. A vertex-based scheme is used in which the wave action N is stored at the vertices and Eq. (8.5) is solved in each vertex. We note that the values at boundary vertices are xed during the computation. For the time integration, we adopt the rst order implicit Euler scheme, as follows N n N n 1 + x [ c x N n ] = F n t (8.6)

where t is the time step and n is the time step counter. The main property of this approximation is that it does not suer from the stability restriction imposed by the CFL condition inherent in the explicit methods as employed in most spectral models. In principle, the time step is limited only by the desired temporal accuracy. This procedure, however, involves the solution of a large system of equations. A point-by-point multi-directional Gauss-Seidel iteration technique is employed for updating all grid vertices. A key feature of this technique is that it takes advantage of the newly acquired vertex values during an iteration. It is locally implicit but globally explicit. In other words, it circumvents the need to build or store large matrices and remains stable at any time steps. This means that this numerical procedure can converge to steady state much more rapidly than explicit methods without requiring too much computational work and memory as do implicit methods.

Unstructured mesh implementation

97

We consider the update of a vertex as labeled 1 in Figure 8.2. This involves looping over each cell of this vertex. We want to nd an approximation for the propagation term
1 0 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1

0 21 0 111 1 000 e(1) 000 111 111 000 000 111

1
1 0 0 1
(2) 1111 0000 0000 1111 0000 1111

y 3

x = not updated = to be updated 0 1 0 = updated 1

Figure 8.2: Update of the wave action at vertex 1 in a triangle 123 and the shaded directional sector in spectral space for which the waves are propagated. of Eq. (8.6). To this end, we employ some vector calculus. We consider a triangular cell as depicted in Figure 8.3. In vertex 1, we apply a mapping from a local coordinate system = ( , ) to the Cartesian one x = (x,y ). Based on this transformation x( ), we have the following base vectors that are tangential to the coordinate lines and , respectively, e(1) = The vectors e(1) = grad , e(2) = grad (8.8) are normal to the coordinate surface of constant and , respectively (see Figure 8.3). Moreover, they are reciprocal to the base vectors, i.e.
e ( ) e ( ) = ,

x ,

e(2) =

x .

(8.7)

, = {1, 2} ,

(8.9)

where is Kronecker delta (which is unity if = , and zero otherwise). Using Cramers rule, one can nd

e(1) =

1 1 2 (2) 1 2 1 2 1 (e(2) , e1 = ( e 2 (2) ) , e (1) , e(1) ) , D = e(2) e(1) e(1) e(2) . D D

(8.10)

Next, we expand the propagation term of Eq. (8.6): x [ c x N ] = cx N cy N + , x y (8.11)

98

Chapter 8

1 0 0 1

e(2)
1
1 0 0 1 0 1

e(2) e(1)

11 00 00 11 00 11

e(1)

y x
Figure 8.3: A triangular cell with geometrical quantities used for discretization in geographical space. Denitions of these quantities are provided in the text. where cx and cy are the x and y components of the wave propagation vector cx , respectively. Using the chain rule, we obtain x [ c x N ] = e 1
(1) cx N

+ e1

(2) cx N

+ e2

(1) cy N

+ e2

(2) cy N

(8.12)

Further, we approximate the derivatives in Eq. (8.12). The most simplest one is a one-sided rst order dierence scheme, as follows c x N1 c x N2 cx N , cy N1 cy N2 cy N , cx N c x N1 c x N3 , cy N c y N1 c y N3 , (8.13)

where the action densities at vertices 1, 2 and 3 are denoted by N1 , N2 and N3 , respectively. Here, we choose the mapping x( ) such that = = 1. The approximation is completed by substituting (8.13) in (8.12):
1 1 1 x [ c x N ] c x N | 1 2 e 1 + cx N | 3 e 1 + cy N | 2 e 2 + cy N | 3 e 2 . (1) (2) (1) (2)

(8.14)

Note that the components of the vectors e(1) and e(2) in Eq. (8.14) are given by Eqs. (8.10), while the base vectors are calculated according to e(1) = x1 x2 , e(2) = x1 x3 (8.15)

Unstructured mesh implementation

99

with xi = (xi , yi ) the position vector of vertex i in a Cartesian coordinate system. This space discretization is of lowest order accurate and conserves action. The upwind dierence scheme (8.14) is employed for two reasons. First, it enforces the propagation of wave action to follow the characteristics. Second, it is monotone (i.e. guaranteeing N > 0 everywhere) and compact (i.e. operating on one triangle only), while suciently accurate n n for nearshore applications. Given the action densities N2 and N3 at vertices 2 and 3 of triangle 123, the wave action in vertex 1 is readily determined according to 1 (2) (1) (2) (1) n + cx,1 e1 + e1 + cy,1 e2 + e2 N1 = t n 1 N1 (1) (1) (2) (2) n n + cx,2 e1 + cy,2 e2 N2 + cx,3 e1 + cy,3 e2 N3 + Fn . t

(8.16)

The wave directions between faces e(1) and e(2) enclose all wave energy propagation in between the corresponding directions 1 and 2 as indicated as a shaded sector in Figure 8.2. This sector is the domain of dependence of Eq. (8.16) in vertex 1. Since, the wave characteristics lie within this directional sector, this ensures that the CFL number used will properly capture the propagation of wave action towards vertex 1. So, propagation is not subjected to a CFL stability criterion. Next, the term F n in Eq. (8.16) is discretized implicitly in the sector considered. Since, the approximation in the spectral space and the linearization of the source terms are well explained in Section 3.3, we shall not pursue them any further. Eq. (8.16) constitute a coupled set of linear, algebraic equations for all spectral bins within the sector considered at vertex 1. The solution is found by means of an iterative solver; see Section 3.3 for details. The update of vertex 1 is completed when all surrounding cells have been treated. This allows waves to transmit from all directions. Due to refraction and nonlinear interactions, wave energy shifts in the spectral space from one directional sector to another. This is taken into account properly by repeating the whole procedure with converging results.

8.3.2

The sweeping algorithm

The solution of each vertex must be updated geographically before proceeding to the next one. For example, referring to Figure 8.2, the value in vertex 1 is determined by its two upwave vertices 2 and 3 only if they are already updated. For regular grids, the foursweep scheme based on a four-direction Gauss-Seidel relaxation is employed as outlined in Section 3.3. The grid points are ordered in a natural manner, e.g. left to right and bottom to top during the rst sweep, right to left and bottom to top during the second sweep, and so on. Hence, the updated values will be used immediately for updating the next unknown. However, in an unstructured mesh there are no distinct directions. Thus the vertices are ordered by their numbering which for an unstructured grid are quite random. As a consequence, the latest obtained solution will be not necessarily used for updating surrounding vertices. An ordering is proposed such that the solution of each vertex will tend to ensure that updated values from the surrounding vertices are used as soon as they are available. We

100

Chapter 8

introduce a reference point on the boundary where the incoming wave energy is imposed and order all the vertices according to their distances to the reference point in ascending order. The updates along this ordering of vertices can be interpreted as propagation of spherical wave fronts with a center on the upwave boundary through the domain as illustrated in Figure 8.4. It is expected that this specic ordering should result in a faster

Figure 8.4: Ordering of vertices along spherical wave fronts indicated by dierent color points. The black point in left-bottom corner is chosen as reference point. convergence than a random ordering of vertices. An algorithm is employed that consists of simply proceeding through a list of vertices that remain to be updated. This list is sorted according to the ascending distances of vertices to the chosen reference point. For a given vertex to be updated, we rst check if its upwave neighbours has already been updated. If this is the case, this vertex is updated and tagged in the list. Otherwise, the considered vertex is placed untagged and the process continues with the next vertex in the list of non-updated vertices. These updates are swept in two cycles. The rst cycle involves a forward sweep from the rst vertex in the list to the last. The second cycle moves backward from the last to the rst. As such, all directions of characteristics can be covered eciently. An iteration is completed when all vertices are updated in both geographic and spectral spaces so that wave energy from all directions has been propagated through geographical space. This numerical process is iterated until an a priori convergence condition is satised. Here, the curvature-based stopping criteria as outlined in Section 3.7 will be applied.

8.4

Interpolation at user-dened locations

All the quantities deals with in SWAN are located at the vertices. Hence, due to the user-dened locations of the wave parameters, interpolations are required. Let parameter j = (xj ) and Cartesian coordinates xj = (xj , yj ), with j {1, 2, 3} indicating the

Unstructured mesh implementation

101

vertices of cell i, be given. The vertices 1, 2 and 3 are ordered in a counterclockwise direction, see Figure 8.2. The associated 3 edges are denoted as 12, 23 and 31. Linear interpolation, with x0 inside cell i and 0 = (x0 ), is given by ( x) = 0 + ( x x0 ) 1 Ai 1 Ai 1 Ai (8.17)

where is a constant vector inside cell i. We apply Green-Gauss reconstruction, i.e.,

i

d =

nd

e ne
e

(8.18)

where Ai is the area of cell i and the summation runs over the 3 edges e {12, 23, 31} of cell i. The values e at edges are taken as averages: 1 12 = (1 + 2 ) , 2 1 23 = (2 + 3 ) , 2 1 31 = (3 + 1 ) 2 (8.19)

Furthermore, ne is the outward pointing normal at edge e and is obtained by rotating the edge over 90o in the clockwise direction. Hence, n12 = Rt12 , We also need the following identity n12 + n23 + n31 = 0 It is not dicult to show that = 1 [n12 (1 3 ) + n31 (1 2 )] 2Ai 1 = [1 n23 + 2 n31 + 3 n12 ] 2Ai (8.21) R= 0 1 1 0 , t12 = x2 x1 (8.20)

(8.22)

or

1 = [1 (y2 y3 ) + 2 (y3 y1 ) + 3 (y1 y2 )] x 2Ai 1 = [1 (x3 x2 ) + 2 (x1 x3 ) + 3 (x2 x1 )] y 2Ai

(8.23)

and

(8.24)

The area Ai of cell i is given by |t12 n13 |/2. Hence, with n13 = Rt13 = we have y3 y1 x1 x3 (8.25)

1 Ai = |(x2 x1 )(y3 y1 ) (x3 x1 )(y2 y1 )| 2

(8.26)

102

Chapter 8

Alternatively, we may interpolate using the following relation ( x) =

k

k k ( x) = 1 1 + 2 2 + 3 3

(8.27)

where k is a linear shape function with the following properties: 1. k is linear per cell and 2. k (xj ) = kj with kj is the Kronecker delta. We choose the following shape function
k k k ( x ) = ak 0 + ax x + ay y

(8.28)

and the coecients a follow from solving

2 3 a1 1 0 0 1 x1 y1 0 a0 a0 1 2 3 a a a 1 x2 y2 x x x = 0 1 0 2 3 a1 0 0 1 1 x3 y3 y ay ay

(8.29)

8.5

Computation of wave-induced force

FORCE is the wave-driven stress, i.e. the force per unit surface driving the wave-driven current, expressed in N/m2 , is dened as the gradient of the radiation stresses: Sxx = g 1 n cos2 + n Edd 2 n sin cos Edd (8.30) (8.31) (8.32)

Sxy = Syx = g Syy = g

1 n sin2 + n Edd 2 with n the ratio of group velocity and phase velocity. The force is then Fx = and Fy = Sxx Sxy x y Syx Syy x y

(8.33)

(8.34)

In order to compute the force in all internal vertices of the unstructured mesh, we consider a control volume (CV) as depicted in Figure 8.5. This CV is called centroid dual and is constructed by joining the centroids neighbouring the vertex under consideration. The set

Unstructured mesh implementation

103

x5 y5 x4 y4 x 0 y0
111111111111 000000000000 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111

x3 y3

x1 y1

x2 y2

A cv

Figure 8.5: Control volume (centroid dual) of vertex is shaded. Some notation is introduced. of CVs must ll the whole computational domain and must also be non-overlapping. In the following we use the numbering from Figure 8.5. Let be one of the radiation stresses Sxx , Sxy and Syy . The gradient of is computed as follows: 1 ACV e ne
e

(8.35)

where ACV is the area of the CV and the summation runs over the associated edges e of this CV. The values e at edges of the centroid dual are taken as averages, i.e. (0 + 1 )/2, (1 + 2 )/2, etc. Moreover, the value of the radiation stresses inside each triangle is simply the average of the radiation stresses in the associated vertices of the cell. Now, the derivatives of inside CV are 1 = x 2ACV and 1 = y 2ACV
n 1 i=0 n 1 i=0

(i + i+1 ) (yi+1 yi )

(8.36)

(i + i+1 ) (xi xi+1 )

(8.37)

with n the number of surrounding cells of the considered vertex and n = 0 , xn = x0 and yn = y0 . The area of the CV is given by ACV = 1 n 1 (xi yi+1 xi+1 yi ) 2 i=0 (8.38)

104

Chapter 8

8.6

Calculation of diusion-like terms

There are situations in which the following diusion-like term need to be computed on unstructured meshes: ( ) (8.39) in vertices with a space-varying diusion coecient (a tensor) and a scalar dened in vertices. In SWAN, these are the alleviation of the garden-sprinkler eect; see Eq. (3.12), and the computation of the diraction parameter; see Eq. (2.121). We consider the centroid dual as shown in Figure 8.5. The calculation consists of 3 steps. First, we compute the gradient of inside each surrounding cell using expressions (8.23) and (8.24). Next, this gradient is multiplied with the appropriate diusion coecient as given in the centroid. Finally, we compute the gradient of inside the CV according to Eqs. (8.36) and (8.37).

Chapter 9 The overall solution algorithm

This chapter is under preparation.

105

106

Bibliography
[1] Abreu, M., A. Larraza and E. Thornton, 1992: Nonlinear transformation of directional wave spectra in shallow water, J. Geophys. Res., 97, 15579-15589 [2] Alves, J.H.G.M. and M.L. Banner, 2003: Performance of a saturation-based dissipation-rate source term in modelling the fetch-limited evolution of wind waves, J. Phys. Oceanogr., 33, 1274-1298 [3] Arcilla, A.S. and C.M. Lemos, 1990: Surf-Zone Hydrodynamics, Centro Internacional de M etodos Num ericos en Ingenieria, Barcelona, 310 p. [4] Arcilla, A.S., J.A. Roelvink, B.A. OConnor, A.J.H.M. Reniers and J.A. Jimenez, 1994: The Delta ume 93 experiment, Proc. Coastal Dynamics Conf. 94, 488-502 [5] Banner, M.L. and I.R. Young, 1994: Modelling spectral dissipation in the evolution of wind waves. Part I: Assessment of existing model performance, J. Phys. Oceanogr., 24, No. 7, 1550-1571 [6] R. Barrett, M. Berry, T.F. Chan, J. Demmel, J. Donato, J. Dongarra, V. Eijkhout, R. Pozo, C. Romine, H. van der Vorst, Templates for the solution of linear systems: Building blocks for iterative methods, SIAM, Philadelphia, PA, 1994, accessible from http://www.netlib.org/templates/Templates.html. [7] P. Bastian, G. Horton, Parallelization of robust multi-grid methods: ILU factorization and frequency decomposition method, SIAM J. Sci. Stat. Comput. 12 (1991) 1457 1470. [8] Battjes, J.A. and J.P.F.M. Janssen, 1978: Energy loss and set-up due to breaking of random waves, Proc. 16th Int. Conf. Coastal Engineering, ASCE, 569-587 [9] Battjes, J.A. and G.Ph. van Vledder, 1984, Verication of Kimuras theory for wave group statistics, Proc. 19th Int. Conf. Coastal Engineering, ASCE, Houston, 642-648 [10] Battjes, J.A. and M.J.F. Stive, 1985: Calibration and verication of a dissipation model for random breaking waves, J. Geophys. Res., 90, No. C5, 9159-9167 [11] Battjes, J.A. and S. Beji, 1992: Breaking waves propagating over a shoal, Proc. 23rd Int. Conf. Coastal Engineering, ASCE, 42-50 107

108 [12] Battjes, J.A., 1994: Shallow water wave modelling, M. Isaacson and M. Quick (Eds.), Proc. Waves-Physical and Numerical Modelling, University of British Columbia, Vancouver, 1-24 [13] Beji, S. and J.A. Battjes 1993: Experimental investigation of wave propagation over a bar, Coastal Engineering, 19, 151-162 [14] Bender, L.C., 1996, Modication of the physics and numerics in a third-generation ocean wave model, J. Atm. and Ocean. Techn., 13, 3, 726-750 [15] Bidlot, J.-R., B. Hansen and P.A.E.M. Janssen, 1996: Wave modelling and operational forecasting at ECMWF, 1st International Conference on EuroGOOS, 7-11 October, 1996, The Hague, 206-213 [16] Benoit, M., F. Marcos and F. Becq, 1996: Development of a third-generation shallowwater wave model with unstructured spatial meshing. Proc. 25th Int. Conf. Coastal Engineering, ASCE, Orlando, 465-478 [17] Benoit, M., 2005: Evaluation of methods to compute the non-linear quadruplet interactions for deep-water wave spectra, Proc. 5th Int. Symp. WAVES 2005, Madrid, Spain [18] Berkho, J.C.W., 1972: Computation of combined refraction-diraction, Proc. 13th Int. Conf. Coastal Engineering, ASCE, 471-490 [19] Bertotti, L and L. Cavaleri, 1994: Accuracy of wind and wave evaluation in coastal regions, Proc. 24th Int. Conf. Coastal Engineering, ASCE, 57-67 [20] Bilgili, A. and Smith, K.W., 2003. BatTri: a 2-D nite element grid generator, version 11.11.03. Available from: http://wwwnml.dartmouth.edu/Publications/internal reports/NML-03-15/. [21] Booij, N., and L.H. Holthuijsen, 1987, Propagation of ocean waves in discrete spectral wave models, Journal of Computational Physics, vol. 68, 307-326. [22] Booij, N., L.H. Holthuijsen and P.H.M. de Lange, 1992, The penetration of shortcrested waves through a gap, Proc. 23rd Int. Conf. Coastal Eng, Venice 4-9 Oct., 1992, New York, 1993, 1044-1052 [23] Booij, N., L.H. Holthuijsen and R.C. Ris, 1996: The SWAN wave model for shallow water, Proc. 25th Int. Conf. Coastal Engng., Orlando, 668-676 [24] Booij, N., L.H. Holthuijsen and R. Padilla-Hernandez, 1997: A nonstationary, parametric coastal wave model, Conf. Coastal Dynamics 97, Plymouth, 99-107 [25] Booij, N., L.H. Holthuijsen and R. Padilla-Hernandez, 1997: Numerical wave propagation on a curvi-linear grid, Proceedings 3rd International Symposium on Ocean Wave Measurement and Analysis, WAVES 97, ASCE, 286-294

109 [26] Booij, N., L.H. Holthuijsen and R.C. Ris, 1998: Shallow water wave modelling, Oceanology International 98, The Global Ocean, Brighton, Conference Proceedings, 3, 483-491 [27] Booij, N., L.H. Holthuijsen and IJ.G. Haagsma, 1998: Comparing the secondgeneration HISWA wave model with the third-generation SWAN wave model, 5th International Workshop on Wave Hindcasting and Forecasting, Melbourne, Florida, 215-222 [28] Booij, N., R.C. Ris and L.H. Holthuijsen, 1999, A third-generation wave model for coastal regions, Part I, Model description and validation, J.Geoph.Research, 104, C4, 7649-7666 [29] Booij, N., L.H. Holthuijsen and J.A. Battjes, 2001: Ocean to near-shore wave modelling with SWAN, 4th International Conference on Coastal Dynamics 2001, Lund, Sweden, 335-344 [30] Booij, N., L.H. Holthuijsen and IJ.G. Haagsma, 2001: The eect of swell on the generation and dissipation of waves, 4th International Symposium on Ocean Wave Measurements and Analysis WAVES 2001, San Francisco, 501-506 [31] Botta, E.F.F. and M.H.M. Ellenbroek, 1985: A modied SOR method for the Poisson equation in unsteady free-surface ow calculations, J. Comput. Phys., 60, 119-134 [32] Bouws, E. and J.A. Battjes, 1982: A Monte-Carlo approach to the computation of refraction of water waves, J. Geophys. Res., 87, 5718-5722 [33] Bouws, E. and G.J. Komen, 1983: On the balance between growth and dissipation in an extreme, depth-limited wind-sea in the southern North Sea, J. Phys. Oceanogr., 13, 16531658 [34] Campbell, T., J. Cazes, E. Rogers, Implementation of an important wave model on parallel architectures, in: Oceans 2002 MTS/IEEE Conference, IEEE, 2002, pp. 15091514, available from http://www7320.nrlssc.navy.mil/mts2002/PDF/campbellpaper.pdf. [35] Cardone, V.J., W.J. Pierson and E.G. Ward, 1976: Hindcasting the directional spectra of hurricane-generated waves, J. Petr. Techn., 28, 385-394 [36] Cavaleri, L. and P. Malanotte-Rizzoli, 1981: Wind wave prediction in shallow water: Theory and applications. J. Geophys. Res., 86, No. C11, 10,961-10,973 [37] Cavaleri, L. and L.H. Holthuijsen, 1998: Wave modelling in the WISE group, Proc. 26th Int. Conf. Coastal Engng., Copenhagen, 498-508 [38] CERC, 1973: Shore Protection Manual, U.S. Army Corps of Engineers, Techn. Rep. No. 4, Vol. I

110 [39] Chen, Y. and H. Wang, 1983: Numerical model for nonstationary shallow water wave spectral transformations, J. Geophys. Res., 88, 9851-9863 [40] Chen, Y., R.T. Guza, and S. Elgar, 1997: Modelling of breaking surface waves in shallow water, J. Geophys. Res., 102, C11, 25035-25046 [41] Chrisochoides, N.P., E.N. Houstis, J.R. Rice, Mapping algorithms and software environment for data parallel PDE iterative solvers, J. Parallel and Distr. Comput. 21 (1994) 7595. [42] Collins, J.I., 1972: Prediction of shallow water spectra, J. Geophys. Res., 77, No. 15, 2693-2707 [43] Dalrymple, R.A., J.T. Kirby and P.A. Hwang, 1984: Wave diraction due to areas of energy dissipation, Journal of Waterways, Ports, Harbours and Coastal Engineering, 110, 67-79 [44] dAngremond, K., J.W. van der Meer and R.J. de Jong, 1996: Wave transmission at low-crested structures, Proc. 25th Int. Conf. Coastal Engng., ASCE, Orlando, 24182427 [45] Dingemans, M.W., A.C. Radder and H.J. de Vriend, 1978: Computations of the driving forces of wave-induced currents, Coastal Engng., 11, 539-563 [46] De Waal, J.P., 2001. Wave growth limit in shallow water. Proc. 4th Int. Symp. Waves 2001, pp. 560569. [47] Dingemans, M.W., 1997: Water wave propagation over uneven bottoms. Part 1 -linear wave propagation, Advanced Series on Ocean Engineering, 13, World Scientic, 471 p. [48] Eldeberky, Y., 1996: Nonlinear transformation of wave spectra in the nearshore zone, Ph.D. thesis, Delft University of Technology, Department of Civil Engineering, The Netherlands [49] Eldeberky, Y. and J.A. Battjes, 1995: Parameterization of triad interactions in wave energy models, Proc. Coastal Dynamics Conf. 95, Gdansk, Poland, 140-148 [50] Eldeberky, Y. and J.A. Battjes, 1996: Spectral modelling of wave breaking: Application to Boussinesq equations, J. Geophys. Res., 101, No. C1, 1253-1264 [51] Elgar, S. and R.T. Guza, 1986: Nonlinear model predictions of bispectra of shoaling surface gravity waves, J. Fluid Mech., 167, 1-18 [52] Elgar, S., R.T. Guza, B. Raubenheimer, T.H.C. Herbers and E.L. Gallagher, 1997: Spectral evolution of shoaling and breaking waves on a barred beach, J. Geophys. Res., 102,C7, 15797-15805

111 [53] Ewing, J.A., 1971: A numerical wave prediction method for the North Atlantic Ocean, Deutsch. Hydrogr. Z., 24, No. 6, 241261 [54] Ewing, J.A. and R.C. Hague, 1993: A second-generation wave model for coastal wave prediction, Proc. of 2nd Int. Symposium on Ocean Wave Measurement and Analysis, New Orleans, 576-589 [55] Fletcher, C.A.J., 1988: Computational Techniques for Fluid Dynamics, Parts I and II, Springer, 409+484 p. [56] Ferziger, J.H. and Peri c, M., 1999. Computational methods for uid dynamics (2nd edition). Springer-Verlag, Berlin. [57] G.C. Fox, A review of automatic load balancing and decomposition methods for the hypercube, in: M. Schulz (Ed.), Numerical Algorithms for Modern Parallel Computer Architectures, Springer, Verlag, 1988, pp. 6376, IMA volume 13. [58] Freilich, M.H. and R.T. Guza, 1984: nonlinear eects on shoaling surface gravity waves, Phil. Trans. R. Soc. London, A 311, 1-41 [59] Galvin, C.J., 1972: Wave breaking in shallow water, Waves on beaches and resulting sediment transport, Academic Press Inc., 413-455 [60] Gelci, R., H. Cazal e and J. Vassal, 1956: Utilization des diagrammes de propagation ` a la pr evision energ etique de la houle, Bulletin dinformation du Comit e Central doc eanographie et d etudes des c otes, 8, No. 4, 160-197 (in French) [61] Goda, Y., H. Takeda and Y. Moriya, 1967: Laboratory investigation of wave transmission over breakwaters, Rep. port and Harbour Res. Inst., 13 (from Seelig, 1979) [62] Golding, B.W., 1983: A wave prediction system for real-time sea state forecasting, J. R. Met. Soc., 109, 393-416 [63] Golub, G.H. and C.F. van Loan, 1986: Matrix Computations, North Oxford Academic, London, 476 p. [64] Gorman, R.M. and C.G. Neilson, 1999: Modelling shallow water wave generation and transformation in an intertidal estuary, Coastal Engineering, 36, 197-217 [65] W. Gropp, E. Lusk, A. Skjellum, Using MPI: Portable parallel programming with the Messing Passing Interface, MIT Press, Cambridge, MA, 1999. [66] G unther, H., S. Hasselmann and P.A.E.M. Janssen, 1992: The WAM model Cycle 4 (revised version), Deutsch. Klim. Rechenzentrum, Techn. Rep. No. 4, Hamburg, Germany

112 [67] Hargreaves, J.C. and Annan, J.D., 2001. Comments on Improvement of the shortfetch behavior in the wave ocean model (WAM). J. Atmos. Oceanic Techn., 18: 711715. [68] Hashimoto, N., IJ.G. Haagsma and L.H. Holthuijsen, 2002: Four-wave interactions in SWAN, Proc. 28th Int. Conf. Coastal Engng., ASCE, Cardi, 392-404 [69] Hasselmann, K., 1960: Grundgleichungen der Seegangsvoraussage, Schistechnik, 1, 191195 [70] Hasselmann, K., 1962: On the non-linear transfer in a gravity wave spectrum, part 1. General theory, J. Fluid Mech., 12, 481-500 [71] Hasselmann, K., 1963a: On the non-linear transfer in a gravity wave spectrum, part 2. Conservation theory, wave-particle correspondence, irreversibility, J. Fluid Mech., 15, 273-281 [72] Hasselmann, K., 1963b: On the non-linear transfer in a gravity wave spectrum, part 3. Evaluation of energy ux and sea-swell interactions for a Neuman spectrum, J. Fluid Mech., 15, 385-398 [73] Hasselmann, K. and J.I. Collins, 1968: Spectral dissipation of nite-depth gravity waves due to turbulent bottom friction, J. Mar. Res., 26, 1-12 [74] Hasselmann, K., T.P. Barnett, E. Bouws, H. Carlson, D.E. Cartwright, K. Enke, J.A. Ewing, H. Gienapp, D.E. Hasselmann, P. Kruseman, A. Meerburg, P. M uller, D.J. Olbers, K. Richter, W. Sell and H. Walden, 1973: Measurements of windwave growth and swell decay during the Joint North Sea Wave Project (JONSWAP), Dtsch. Hydrogr. Z. Suppl., 12, A8 [75] Hasselmann, K., 1974: On the spectral dissipation of ocean waves due to whitecapping, Bound.-layer Meteor., 6, 1-2, 107-127 [76] Hasselmann, K., D.B. Ross, P. M uller and W. Sell, 1976: A parametric wave prediction model, J. Phys. Oceanogr., 6, 200228 [77] Hasselmann, S. and K. Hasselmann, 1981: A symmetrical method of computing the non-linear transfer in a gravity-wave spectrum, Hamburger Geophys. Einzelschr., Serie A., 52, 8 [78] Hasselmann, S., K. Hasselmann, J.H. Allender and T.P. Barnett, 1985: Computations and parameterizations of the nonlinear energy transfer in a gravity wave spectrum. Part II: Parameterizations of the nonlinear transfer for application in wave models, J. Phys. Oceanogr., 15, 11, 1378-1391 [79] Hedges, T.S., 1987: Combination of waves and currents: an introduction, Proc. Instn. Civ. Engrs., Part 1, 82, 567-585

113 [80] Hersbach, H. and Janssen, P.A.E.M., 1999. Improvement of the short-fetch behavior in the wave ocean model (WAM). J. Atmos. Oceanic Techn., 16: 884892. [81] Herterich, K. and K. Hasselmann, 1980: A similarity relation for the nonlinear energy transfer in a nite-depth gravity-wave spectrum, J. Fluid Mech., 97, 215-224 [82] Holthuijsen, L.H., 1980: Methods of wave prediction, part I and II (Methoden voor golfvoorspelling, deel I en II, in Dutch), Technical Advisory Commission against Inundation (Technische Adviescommissie voor de Waterkeringen, in Dutch), Den Haag, The Netherlands [83] Holthuijsen, L.H. and S. De Boer, 1988: Wave forecasting for moving and stationary targets, Computer modelling in Ocean Engineering, Eds. B.Y. Schreer and O.C. Zienkiewicz, Balkema, Rotterdam, The Netherlands, 231-234 [84] Holthuijsen, L.H., Booij, N. and T.H.C. Herbers, 1989: A prediction model for stationary, short-crested waves in shallow water with ambient currents, Coastal Engineering, 13, 23-54 [85] Holthuijsen, L.H., N. Booij and R.C. Ris, 1993: A spectral wave model for the coastal zone, Proceedings 2nd International Symposium on Ocean Wave Measurement and Analysis, New Orleans, Louisiana, July 2528, 1993: New York, pp. 630641 [86] Holthuijsen, L.H., N. Booij and L. Bertotti, 1996: The propagation of wind errors through ocean wave hindcasts, J. Oshore Mech. And Arctic Eng., 118, 184-189 [87] Holthuijsen, L.H., N. Booij and R. Padilla-Hernandez, 1997: A curvi-linear, thirdgeneration coastal wave model, Conf. Coastal Dynamics 97, Plymouth, 128-136 [88] Holthuijsen, L.H., N. Booij, R. Ris, J.H. Andorka Gal and J.C.M. de Jong, 1997: A verication of the third-generation wave model SWAN along the southern North Sea coast, Proceedings 3rd International Symposium on Ocean Wave Measurement and Analysis, WAVES 97, ASCE, 49-63 [89] Holthuijsen, L.H., 1998: The concept and features of the ocean wave spectrum, Provision and engineering/operational application of ocean wave spectra, COST Conference, UNESCO, 21-25 sept., Paris, keynote address, 11-20 [90] Holthuijsen, L.H., 1998: Waves in shallow water, World Meteorological Organization Guide to wave analysis and forecasting, WMO-N0. 702, Chapter 7, 81-88 [91] Holthuijsen, L.H., R.C. Ris and N. Booij, 1998: A verication of the third-generation wave model SWAN, 5th International Workshop on Wave Hindcasting and Forecasting, Melbourne, Florida, 223-230

114 [92] Holthuijsen, L.H., N. Booij and IJ.G. Haagsma, 1998: Comparing 1st -, 2nd - and 3rd -generation coastal wave modelling, 26th Int. Conf. Coastal Engng., Copenhagen, 140-149 [93] Holthuijsen, L.H. and N. Booij, 2000: Oceanic and near-shore whitecapping eects in SWAN, 6th Int. Workshop on Wave Hindcasting and Forecasting, Monterey, 362-368 [94] Holthuijsen, L.H, R.C. Ris, N. Booij and E.Cecchi, 2000: Swell and whitecapping, a numerical experiment, Proc. 27th Int. Conf. Coastal Engng., Sydney, 346-354 [95] Holthuijsen, L.H., Herman, A. and Booij, N., 2003: Phase-decoupled refractiondiraction for spectral wave models, Coastal Engineering, 49, 291-305 [96] Holthuijsen, L.H., 2007: Waves in oceanic and coastal waters, Cambridge University Press. [97] Hsu, T.-W., S.-H. Ou and J.-M. Liau, 2005: Hindcasting nearshore wind waves using a FEM code for SWAN, Coastal Engineering, 52, 177-195 [98] Janssen, P.A.E.M., 1989: Wave induced stress and the drag of air ow over sea waves, J. Phys. Oceanogr., 19, 745-754 [99] Janssen, P.A.E.M., 1991a: Quasi-linear theory of wind-wave generation applied to wave forecasting, J. Phys. Oceanogr., 21, 1631-1642 [100] Janssen, P.A.E.M., 1991b: Consequences of the eect of surface gravity waves on the mean air ow, Int. Union of Theor. and Appl. Mech. (IUTAM), Sydney, Australia, 193-198 [101] Jonsson, I.G. 1966: Wave boundary layers and friction factors, Proc. 10th Int. Conf. Coastal Engineering, ASCE, 127-148 [102] Jonsson, I.G., 1980: A new approach to rough turbulent boundary layers, Ocean Engineering, 7, 109-152 [103] Jonsson, I.G., 1993: The Sea, Ocean Engineering Science, 9, Part A [104] Jonsson, I.G. and N.A. Carlsen, 1976: Experimental and theoretical investigations in an oscillatory turbulent boundary layer, J. Hydraulic Research, 14, 45-60 [105] Kahma, K.K. and C.J. Calkoen, 1992: Reconciling discrepancies in the observed growth of windgenerated waves, J. Phys. Oceanogr., 22, 13891405 [106] Kaminsky, G.M. and N.C. Kraus, 1993: Evaluation of depth-limited wave breaking criteria, Proc. of 2nd Int. Symposium on Ocean Wave Measurement and Analysis, New Orleans, 180-193

115 [107] Karlson, T, 1969: Refraction of continuous ocean wave spectra, Journal of Waterways, Ports, Harbours and Coastal Engineering, 95, 275-287 [108] Kirby, J.T., 1986: Higher-order approximation in the parabolic equation method for water waves, J. Geophys. Res., 91, C1, 933-952 [109] Komen, G.J., S. Hasselmann, and K. Hasselmann, 1984: On the existence of a fully developed wind-sea spectrum, J. Phys. Oceanogr., 14, 1271-1285 [110] Komen, G.J., Cavaleri, L., Donelan, M., Hasselmann, K., Hasselmann, S. and P.A.E.M. Janssen, 1994: Dynamics and Modelling of Ocean Waves, Cambridge University Press, 532 p. [111] Kuik, A.J., G.Ph. van Vledder and L.H. Holthuijsen, 1988: A method for the routine analysis of pitch-and-roll buoy wave data, J. Phys. Oceanogr., 18, 1020-1034 [112] Li, C.W. and M. Mao, 1992: Spectral modelling of typhoon-generated waves in shallow waters, J. Hydraulic Research, 30, 5, 611-622 [113] Lin, R.Q. and N.E. Huang, 1996: the Goddard coastal wave model. Part I: Numerical method, J. Phys. Oceanogr., 26, 833-847 [114] Liu, P.L., S.B. Yoon and J.T. Kirby, 1985: Nonlinear refraction-diraction of waves in shallow water, J. Fluid Mech., 153, 185-201 [115] Luo, W. and J. Monbaliu, 1994: Eects of the bottom friction formulation on the energy balance for gravity waves in shallow water, J. Geophys. Res., 99, C9, 18,50118,511 [116] Madsen, O.S., Y.-K. Poon and H.C. Graber, 1988: Spectral wave attenuation by bottom friction: Theory, Proc. 21th Int. Conf. Coastal Engineering, ASCE, 492-504 [117] Madsen, P.A. and O.R. Srensen, 1992: A new form of the Boussinesq equations with improved linear dispersion characteristics. Part 2: A slowly-varying bathymetry, Coastal Engineering, 18, 183-205 [118] Madsen, P.A. and O.R. Srensen, 1993: Bound waves and triad interactions in shallow water, Ocean Engineering, 20, 4, 359-388 [119] Mase, H. and J.T. Kirby, 1992: Hybrid frequency-domain KdV equation for random wave transformation, Proc. 23th Int. Conf. Coastal Engineering, ASCE, 474-487 [120] Mastenbroek, C., G. Burgers, and P.A.E.M. Janssen, 1993: The dynamical coupling of a wave model in a storm surge model through the atmospheric boundary layer, J. Phys. Oceanogr., 23, 1856-1866

116 [121] Van der Meer, J.W., R. Briganti, B. Zanuttigh and B. Wang, 2005: Wave transmission and reection at low-crested structures: design formulae, oblique wave attack and spectral change, Coast. Engng., 52, 915-929 [122] Mendez, F.J. and I.J. Losada, 2004: An empirical model to estimate the propagation of random breaking and nonbreaking waves over vegeation elds, Coast. Engng., 51, 103-118 [123] Mei, C.C., 1983: The applied dynamics of ocean surface waves, Wiley, New York, 740 p. [124] G. Meurant, Domain decomposition methods for partial-dierential equations on parallel computers, Int. J. Supercomputer Appl. and High Perform. Comput. 2 (4) (1988) 512. [125] Miles, J.W., 1957: On the generation of surface waves by shear ows, J. Fluid Mech., 3, 185204 [126] Miles, J.W., 1981: Hamiltonian formulations for surface waves, Applied Scientic Research, 37, 103-110 [127] Nelson, R.C., 1987: Design wave heights on very mild slopes: An experimental study, Civil. Eng. Trans., Inst. Eng. Aust., 29, 157-161 [128] Nelson, R.C., 1994: Depth limited design wave heights in very at regions, Coastal Engineering, 23, 43-59 [129] Nelson, R.C., 1997: Height limits in top down and bottom up wave environments, Coastal Engineering, 32, 247-254 [130] Nwogu, O., 1994: Nonlinear evolution of directional wave spectra in shallow water, Proc. 24th Int. Conf. Coastal Engineering, ASCE, 467-481 [131] OpenMP ARB, OpenMP, http://www.openmp.org. [132] Padilla-Hernandez, R., P. Osuna, J. Monbaliu and L. Holthuijsen, 1998: Intercomparing third-generation wave model nesting, 5th International Workshop on Wave Hindcasting and Forecasting, Melbourne, Florida, 102-112 [133] Peregrine, D.H., 1966: Long waves on a beach, J. Fluid Mech., 27, 4, 815-827 [134] Phillips, O.M., 1957: On the generation of waves by turbulent wind, J. Fluid Mech., 2, 417445 [135] Phillips, O.M., 1960: On the dynamics of unsteady gravity waves of nite amplitude. Part 1, J. Fluid Mech., 9, 193-217

117 [136] Phillips, O.M., 1977: The dynamics of the upper ocean, Cambridge University Press, 336 p. [137] Phillips, O.M., 1985: Spectral and statistical properties of the equilibrium range in wind-generated gravity waves, J. Fluid Mech., 156, 505-531 [138] Pierson, W.J. and L. Moskowitz, 1964: A proposed spectral form for fully developed wind seas based on the similarity theory of S.A. Kitaigorodskii, J. Geophys. Res., 69, 24, 5181-5190 [139] Piest, J., 1965: Seegangsbestimmung und Seegangsrefraktion in einem Meer mit nicht- ebenem Boden; eine theoretische Untersuchung, Deutsch. Hydrogr. Z., 18: 6774 (in German) [140] Press, W.H., Flannery, B.P., Teukolsky, S.A. and Vetterling, W.T., 1993. Numerical recipes in Fortran 77. The art of scientic computing (2nd edition). Cambridge University Press, New York (available from http://www.nr.com). [141] Putnam, J.A. and J.W. Johnson, 1949: The dissipation of wave energy by bottom friction, Trans. Am. Geoph. Union, 30, 67-74 [142] Radder, A.C., 1979: On the parabolic equation method for water-wave propagation, J. Fluid Mech., 95, 159-176 [143] Radder, A.C., 1992: An explicit Hamiltonian formulation of surface waves in water of nite depth, J. Fluid Mech., 237, 435-455 [144] Radder, A.C., 1996: Hamiltonian dynamics of water waves, Advances of Coastal and Ocean Engineering, World Scientic, 4, ??-?? [145] Resio, D. and W. Perrie, 1991: A numerical study of nonlinear energy uxes due to wave-wave interactions. Part I: Methodology and basic results, J. Fluid Mech., 223, 609-629 [146] Resio, D.T., J.H. Pihl, B.A. Tracy and C.L. Vincent, 2001: Nonlinear energy uxes and the nite depth equilibrium range wave spectra, J. Geophys. Res., 106, C4, 69857000. [147] Resio, D.T., C.E. Long and C.L. Vincent, 2004: Equilibrium-range constant in windgenerated wave spectra, J. Geophys. Res., 109, C01018. [148] Ris, R.C., L.H. Holthuijsen and N. Booij, 1994: A spectral model for waves in the near shore zone, Proc. 24th Int. Conf. Coastal Engng, Kobe, Japan, pp. 68-78 [149] Ris, R.C. and L.H. Holthuijsen, 1996: Spectral modelling of current wave-blocking. Proc. 25th Int. Conf. Coastal Engng., ASCE, 1247-1254.

118 [150] Ris, R.C. and L.H. Holthuijsen, 1997: Modelling of current induced wave-blocking in a spectral wave model, 8th International Biennal Conference on Physics of Estuaries and Coastal Seas, J. Dronkers and M.B.A.M. Scheers (eds.), The Hague, 139-144 [151] Ris, R.C., 1999. Model convergence of SWAN in the Westerschelde estuary. WL|Delft Hydraulics, Report H3496. [152] Ris, R.C., N. Booij and L.H. Holthuijsen, 1999: A third-generation wave model for coastal regions, Part II: Verication, J. Geophys. Res., 104, C4, 7667-7681 [153] Ris, R., L.H. Holthuijsen, J.M. Smith, N. Booij and A.R. van Dongeren, 2002: The ONR virtual testbed for coastal and oceanic wave models, Proc. 28th Int. Conf. Coastal Engng., ASCE, Cardi, 380-391 [154] Roache, P.J., 1972: Computational Fluid Dynamics, Hermosa Publishers, Albuquerque, 446 p. [155] Rogers, W.E., J.M. Kaihatu, H.A. H. Petit, N. Booij, and L.H. Holthuijsen, 2002: Diusion reduction in a arbitrary scale third generation wind wave model, Ocean Engng., 29, 1357-1390. [156] Rogers, W.E., P.A. Hwang and D.W. Wang, 2003: Investigation of wave growth and decay in the SWAN model: three regional-scale applications, J. Phys. Oceanogr., 33, 366-389. [157] Rogers, W.E., J.M. Kaihatu, L. Hsu, R.E. Jensen, J.D. Dykes and K.T. Holland, 2007: Forecasting and hindcasting waves with the SWAN model in the Southern California Bight, Coastal Engineering, 54, 1-15. [158] Ruessink, B.G., D.J.R. Walstra and H.N. Southgate, 2003: Calibration and verication of a parametric wave model on barred beaches, Coastal Engineering, 48, 139-149 [159] Sakai, T., M. Koseki and Y. Iwagaki, 1983: Irregular wave refraction due to current, J. of Hydr. Eng., ASCE, 109, 9, 1203-1215 [160] Sanders, J.W., 1976: A growth-stage scaling model for the wind-driven sea, Deutsch. Hydrogr. Z., 29, 136-161 [161] Seelig, W.N., 1979, Eects of breakwaters on waves: laboratory tests of wave transmission by overtopping, Proc. Conf. Coastal Structures, 79, 2, 941-961 [162] Shemdin, P., K. Hasselmann, S.V. Hsiao and K. Herterich, 1978: Nonlinear and linear bottom interaction eects in shallow water, in: Turbulent Fluxes through the Sea Surface, Wave Dynamics and Prediction, NATO Conf. Ser., V, 1, 347372 [163] Shewchuk, J.R., 1996. Triangle: a two-dimensional quality mesh generator and Delaunay triangulator, version 1.3. Available from: http://www2.cs.cmu.edu/ quake/triangle.html.

119 [164] Snyder, R.L., Dobson, F.W., Elliott, J.A. and R.B. Long, 1981: Array measurement of atmospheric pressure uctuations above surface gravity waves, J. Fluid Mech., 102, 1-59 [165] Stelling, G.S. and J.J. Leendertse, 1992: Approximation of convective processes by cyclic AOI methods, Proceeding 2nd international conference on estuarine and coastal modeling, ASCE Tampa, Florida, 771-782 [166] Stone, H.L., 1968: Iterative solution of implicit approximations of multidimensional partial dierential equations, SIAM J. of Numer. Anal., 5, 530558 [167] SWAMP group, 1985: Ocean wave modelling, Plenum Press, New York and London [168] Taylor, P.A. and R.J. Lee, 1984: Simple guidelines for estimating wind speed variations due to small-scale topographic features, Climatol. Bull., 18, 3-32 [169] Thornton, E.B. and R.T. Guza, 1983: Transformation of wave height distribution, J. Geophys. Res., 88, C10, 5925-5938 [170] Tolman, H.L., 1990: Wind wave propagation in tidal seas, Ph.D. thesis, Delft University of Technology, Department of Civil Engineering, The Netherlands [171] Tolman, H.L., 1991: A third-generation model for wind waves on slowly varying, unsteady and inhomogeneous depths and currents, J. Phys. Oceanogr., 21, 6, 782-797 [172] Tolman, H.J., 1992a: Eects of numerics on the physics in a third-generation windwave model, J. Phys. Oceanogr., 22, 10, 1095-1111 [173] Tolman, H.L., 1992b: An evaluation of expressions for the wave energy dissipation due to bottom friction in the presence of currents, Coastal Engineering, 16, 165-179 [174] Tolman, H. L., 1995: On the selection of propagation schemes for a spectral windwave model. NWS/NCEP Oce Note 411, 30 pp. + gures. [175] Tolman, H.L., 2002: Distributed-memory concepts in the wave model WAVEWATCH III, Parallel Comput., 28, 35-52. [176] Tracy, B. and D.T. Resio, 1982: Theory and calculation of the nonlinear energy transfer between sea waves in deep water. WES Report 11, US Army Corps of Engineers. [177] Van Vledder, G. Ph., J.G. de Ronde and M.J.F. Stive, 1994: Performance of a spectral wind-wave model in shallow water, Proc. 24th Int. Conf. Coastal Engineering, ASCE, 761-774 [178] Van Vledder, G.Ph., T.H.C. Herbers, R.E. Jensen, D.T. Resio and B. Tracy, 2000: Modelling of non-linear quadruplet wave-wave interactions in operational wave models. Proc. 27th Int. Conf. on Coastal Engineering, Sydney, Australia.

120 [179] Van Vledder, G. Ph. and D.P. Hurdle, 2002: Performance of formulations for whitecapping in wave prediction models, Proc. OMAE 2002 [180] Van Vledder, G. Ph. and M. Bottema, 2003: Improved modelling of nonlinear fourwave interactions in shallow water, Proc. 28th Int. Conf. Coastal Engineering, ASCE, 459-471 [181] Van Vledder, G. Ph., 2006: The WRT method for the computation of non-linear four-wave interactions in discrete spectral wave models, Coast. Engng., 53, 223-242 [182] Vincent, C.L., J.M. Smith and J. Davis, 1994: Parameterization of wave breaking in models, Proc. of Int. Symp.: Waves - Physical and Numerical Modelling, Univ. of British Columbia, Vancouver, Canada, M. Isaacson and M. Quick (Eds.), Vol. II, 753-762 [183] H.A. van der Vorst, High performance preconditioning, SIAM J. Sci. Stat. Comput. 10 (1989) 11741185. [184] WAMDI group, 1988: The WAM model a third generation ocean wave prediction model, J. Phys. Oceanogr., 18, 17751810 [185] Webb, D.J., 1978: Non-linear transfers between sea waves, Deep-Sea Res., 25, 279298 [186] Weber, S.L., 1989: Surface gravity waves and turbulent bottom friction, Ph.D. thesis, University of Utrecht, The Netherlands [187] Weber, S.L., 1991a: Bottom friction for wind sea and swell in extreme depth-limited situations, J. Phys. Oceanogr., 21, 149-172 [188] Weber, S.L., 1991b: Eddy-viscosity and drag-law models for random ocean wave dissipation, J. Fluid Mech., 232, 73-98 [189] P. Wesseling, An introduction to multigrid methods, John Wiley and Sons, Chichester, 1992. [190] Van der Westhuysen, A.J., 2007: Advances in the spectral modelling of wind waves in the nearshore, Ph.D. thesis, Delft University of Technology, Department of Civil Engineering, The Netherlands [191] Van der Westhuysen, A.J., M. Zijlema and J.A. Battjes, 2007: Nonlinear saturationbased whitecapping dissipation in SWAN for deep and shallow water, Coast. Engng., 54, 151-170 [192] Wilson, B.W., 1965: Numerical prediction of ocean waves in the North Atlantic for December 1959, Deutsch. Hydrogr. Z., 18, 3, p. 114-130

121 [193] WISE Group, 2007. Wave modelling - The state of the art. Progr. Oceanogr., 75, 603-674 [194] Whitham, G.B., 1974: Linear and nonlinear waves, Wiley, New York, 636 p. [195] Wu, J., 1982: Wind-stress coecients over sea surface from breeze to hurricane, J. Geophys. Res., 87, C12, 9704-9706 [196] Yan, L. 1987: An improved wind input source term for third generation ocean wave modelling, Scientic report WR-No 87-8, De Bilt, The Netherlands [197] Yamaguchi, M., 1986: A numerical model of nearshore currents based on a nite amplitude wave theory, Proc. 20th Int. Conf. Coastal Engineering, ASCE, 849-863 [198] Yamaguchi, M., 1988: A numerical model of nearshore currents due to irregular waves, Proc. 21th Int. Conf. Coastal Engineering, ASCE, 1113-1126 [199] Yamaguchi, M., 1990: A numerical model for refraction computation of irregular waves due to time-varying currents and water depth, Proc. 22th Int. Conf. Coastal Engineering, ASCE, 205-217 [200] Young, I.R., 1988: A shallow water spectral wave model, J. Geophys. Res., 93, C5, 5113-5129 [201] Young, I.R., 1999: Wind generated ocean waves, Eds. R. Bhattacharyya and M.E. McCormick, Ocean Engineering Series, Elsevier, Amsterdam, 288 p. [202] Young, I.R., and M.L. Banner, 1992: Numerical Experiments on the evolution of fetch limited waves, Int. Union of Theor. and Appl. Mech. (IUTAM), Sydney, Australia, 267-275 [203] Young, I.R. and G. Ph. van Vledder, 1993: A review of the central role of nonlinear interactions in wind-wave evolution, Phil. trans. R. Soc. London. A., 342, 505-524 [204] Young, I.R. and L.A. Verhagen, 1996a: The growth of fetch limited waves in water of nite depth. Part 1: Total energy and peak frequency, Coastal Engineering, 29, 47-78 [205] Young, I.R. and L.A. Verhagen, 1996b: The growth of fetch limited waves in water of nite depth. Part 2: Spectral evolution, Coastal Engineering, 29, 79-99 [206] Young, I.R. and L.A. Verhagen, 1996c: The growth of fetch limited waves in water of nite depth. Part 3: Directional spectra, Coastal Engineering, 29, 101-121 [207] Zijlema, M. and Wesseling, P., 1998. Higher-order ux-limiting schemes for the nite volume computation of incompressible ow. Int. J. Comput. Fluid Dyn., 9, 89-109

122 [208] Zijlema, M., 2005: Parallelization of a nearshore wind wave model for distributed memory architectures, in Parallel Computational Fluid Dynamics - Multidisciplinary applications, Eds. G. Winter and A. Ecer and J. Periaux and N. Satofuka and P. Fox, Elsevier Science B.V., Amsterdam, The Netherlands, 207-214 [209] Zijlema, M. and A.J. van der Westhuysen, 2005: On convergence behaviour and numerical accuracy in stationary SWAN simulation of nearshore wind wave spectra. Coastal Engineering, 52, 237-256 [210] Zijlema, M., 2009: Parallel, unstructured mesh implementation for SWAN, Proc. 31th Int. Conf. Coastal Engineering, ASCE, 470-482 [211] Zijlema, M., 2010: Computation of wind-wave spectra in coastal waters with SWAN on unstructured grids. Coastal Engineering, 57, 267-277 [212] Zijlema, M., G.Ph. van Vledder and L.H. Holthuijsen, 2012: Bottom friction and wind drag for wave models. Coastal Engineering, 65, 19-26

Index
ambient, 2, 7, 11, 12, 35, 113 frequency, 79, 11, 13, 14, 16, 1823, 2527, 29, 31, 34, 35, 4143, 47, 48, 50, 52 bathymetry, 24, 44, 70, 93, 95, 115 56, 6366, 71, 75, 76, 107, 115, 121 bottom, 13, 1315, 2224, 36, 52, 53, 70, 71, friction, 1, 3, 1315, 17, 18, 22, 23, 33, 112, 73, 95, 99, 100, 110, 112, 115120 114, 115, 117, 119, 120, 122 boundary, 3, 4, 13, 15, 50, 71, 75, 76, 78, 79, garden-sprinkler, 45, 46, 104 8185, 87, 94, 96, 100, 114, 115 breaking, 1, 3, 1317, 21, 2325, 49, 66, 107, harbour, 3, 39 110, 111, 114, 116, 120 initial, 13, 15, 30, 51, 75, 76, 88 Cartesian, 3, 1012, 80, 97, 99, 100 island, 93, 94 co-ordinate, 3, 7, 1012, 28, 29, 5860, 82, Jonswap, 75, 76 84 coastal, 1, 2, 15, 46, 55, 68, 89, 93, 108, 109, latitude, 11, 12 111, 113115, 118, 119, 122 limiter, 42, 55, 56, 58 convergence, 42, 43, 48, 50, 5558, 87, 88, longitude, 11, 12 91, 92, 100, 118, 122 Courant, 46, 70, 71 obstacle, 3, 3537, 39, 44, 45, 6163 current, 2, 3, 7, 1013, 15, 35, 38, 39, 43, 44, ocean, 1, 7, 8, 11, 35, 39, 46, 108, 112118, 47, 5053, 60, 65, 70, 71, 73, 77, 90, 120, 121 92, 102, 110, 112, 113, 117119, 121 curvi-linear, 2, 3, 5860, 62, 78, 84, 85, 93, plant, 33, 34 propagation, 14, 1012, 38, 43, 4549, 51 108, 113 53, 59, 62, 6871, 89, 90, 93, 95, 97 100, 108, 110, 111, 113, 116, 117, 119 dam, 16, 33, 3537, 113, 121, 122 diraction, 3, 10, 3639, 104, 108, 110, 114, quadruplets, 16, 54 115 diusion, 4447, 104 recti-linear, 85, 93 dissipation, 1, 3, 12, 1416, 2024, 33, 34, reection, 3, 4, 35, 37, 39, 41, 62, 63, 116 66, 67, 94, 107, 109, 110, 112, 117, refraction, 1, 3, 11, 37, 44, 50, 52, 68, 7073, 119, 120 75, 99, 108, 109, 114, 115, 118, 121 regular, 2, 4, 7, 15, 33, 36, 99, 118, 121 lter, 13, 18, 38 ow, 7, 77, 109, 114, 116, 121 set-up, 3, 4, 24, 39, 77, 85, 107 force, 33, 39, 60, 7779, 84, 85, 95, 99, 102, shoaling, 1, 3, 15, 110, 111 110 SORDUP, 44, 46, 89 123

124 spherical, 3, 11, 12, 100 stability, 41, 42, 46, 50, 56, 67, 68, 7173, 96, 99 stationary, 2, 3, 7, 4244, 46, 54, 59, 71, 76, 77, 108, 110, 113, 122 steepness, 14, 19, 20, 24, 35 swell, 14, 15, 20, 22, 46, 47, 109, 112, 120 triads, 16 triangular, 88, 94, 9698 unstructured, 4, 73, 93, 94, 99, 102, 104, 108, 122 vegetation, 3335 WAM, 1, 13, 14, 1720, 27, 41, 42, 55, 65, 75, 111113, 119, 120 WAVEWATCH, 1, 75, 119 whitecapping, 3, 12, 14, 16, 1922, 35, 112, 114, 120 wind, 13, 7, 1215, 1722, 3133, 35, 41, 4347, 49, 51, 54, 55, 61, 68, 69, 76, 84, 93, 99, 107109, 112122