Entdecken Sie eBooks

Kategorien

Entdecken Sie Hörbücher

Kategorien

Entdecken Sie Zeitschriften

Kategorien

Entdecken Sie Dokumente

Kategorien

0 Bewertungen0% fanden dieses Dokument nützlich (0 Abstimmungen)

87 Ansichten195 Seitens

Heimann_justus CFD Based Optimization of the Wave-making Characteristics of Ship Hulls

© © All Rights Reserved

PDF, TXT oder online auf Scribd lesen

s

© All Rights Reserved

Als PDF, TXT **herunterladen** oder online auf Scribd lesen

0 Bewertungen0% fanden dieses Dokument nützlich (0 Abstimmungen)

87 Ansichten195 Seitens

© All Rights Reserved

Als PDF, TXT **herunterladen** oder online auf Scribd lesen

Sie sind auf Seite 1von 195

Wave-Making Characteristics of Ship Hulls

Wave-Making Characteristics of Ship Hulls

vorgelegt von

Diplom-Ingenieur

Justus Heimann

aus Berlin

der Technischen Universitt Berlin

zur Erlangung des akademischen Grades

Doktor der Ingenieurwissenschaften

Dr.-Ing.

genehmigte Dissertation

Promotionsausschuss:

Vorsitzender:

Berichter:

Berichter:

Berichter:

Prof. Dr.-Ing. Gnther F. Clauss

Prof. Dr.-Ing. Gerhard Jensen

Prof. Dr.-Ing. Lothar Birk

Berlin 2005

D 83

Wave-Making Characteristics of Ship Hulls

Approved by the

Faculty of Mechanical Engineering and Transport Systems

of the Technical University Berlin

in partial fulfillment of the requirements of the degree of

Doktor der Ingenieurwissenschaften

Dr.-Ing.

Doctoral committee:

Prof.

Prof.

Prof.

Prof.

Dr.-Ing.

Dr.-Ing.

Dr.-Ing.

Dr.-Ing.

Gnther F. Clauss Technical University Berlin

Gerhard Jensen Technical University Hamburg-Harburg

Lothar Birk University of New Orleans

Berlin 2005

D 83

Die Deutsche Bibliothek verzeichnet diese Publikation in der

Deutschen Nationalbibliografie; detaillierte bibliografische Daten

sind im Internet ber <http://dnb.ddb.de> abrufbar.

ISBN 3-89820-445-6

Dissertation, Technische Universitt Berlin, 2005

Dieses Werk ist urheberrechtlich geschtzt.

Alle Rechte, auch die der bersetzung, des Nachdruckes und der Vervielfltigung

des Buches, oder Teilen daraus, vorbehalten. Kein Teil des Werkes darf ohne

schriftliche Genehmigung des Verlages in irgendeiner Form reproduziert oder

unter Verwendung elektronischer Systeme verarbeitet, vervielfltigt oder

verbreitet werden.

This document is protected by copyright.

No part of this document may be reproduced in any form by any means without

prior written authorization of Mensch & Buch Verlag.

info@menschundbuch.de www.menschundbuch.de

Preface

Several years ago when I was an enthusiastic regatta sailor the foundation for the present research

work was laid. Many times during the races occurred the challenging question how to improve my

own performance and that of my Laser dinghy very likely, I wasnt the only one!

My tactical and sailing skills improved over the years. So I refined techniques like roll tacking and

jibing, sailing close-hauled and upright on upwind courses, steering through waves with minimal speed

loss, reducing the wetted hull surface on downwind courses or speeding up in planning conditions. But

how to improve the hull hydrodynamics? Since the Laser dinghy is the ultimate one design there is no

room left for hull modifications. Nevertheless, my friends and I did or at least we thought we did

our best to optimize the hull hydrodynamics, mostly within the rules. For instance we polished the

underwater hull with sandpaper to generate a micro air cushion, we shaped the leading and trailing

edges of the centreboard to increase the lift-drag ratio and we fitted a little (invisible) flapper to

the transom edge to improve the planning performance. However, all this hydrodynamic fine tuning

remained rather a matter of faith than of measurable benefit.

Nowadays, after several years of study and research think I have come a bit closer to the answer of how

to improve the hull hydrodynamics measurably at least the size of the dinghies has improved a lot.

The present research work was undertaken at the Division of Naval Architecture and Ocean Engineering of the Technical University Berlin. Preparation and finalization of the thesis were undertaken at the

premises and with the generous support of FRIENDSHIP SYSTEMS GmbH. For a long educational

period I was supervised by Prof. Dr.-Ing. Dr. h.c. Horst Nowacki in whose team I had the pleasure to

work as a research and teaching scientist from 1995 to 2000 in the field of ship geometric modelling,

numerical hydrodynamics and optimization. I wish to express my sincere thanks to Prof. Nowacki for

his valuable advice, inspiration, criticism and support and his companionship of the present work.

I like to express my sincere thanks to Prof. Dr.-Ing. Gnther F. Clauss for guiding my first scientific

steps and for his commitment in the doctoral examination to become thesis supervisor. I am very grateful to Prof. Dr.-Ing. Gerhard Jensen and Prof. Dr.-Ing. Lothar Birk for taking interest in my work and

for being thesis supervisors. Furthermore, I would like to thank Prof. Dr.-Ing. Klaus Brie for chairing

the doctoral examination.

I would like to say thank you to Prof. Dr. Carl-Erik Janson from Chalmers University of Technology and

FLOWTECH Int. AB, Gteborg, Sweden, for his continuous support, many fruitful discussions and for

customizing the CFD system SHIPFLOW according to my needs. Many thanks goes to Prof. Dr. Lars

Larsson and the whole FLOWTECH team for kindly making available the SHIPFLOW system.

Very special thanks I owe to my friend Dr.-Ing. Stefan Harries for being an excellent colleague for

many years and for his invaluable support and encouragement in finalizing this work.

A nice and valuable time I spent with my friends and colleagues, particularly Claus Abt, Frauke

Baumgrtel, zgr Baskaya, Dr.-Ing. Carl-Uwe Bttner, Jrn Hinnenthal, Dr.-Ing. Karsten Hochkirch,

Dr.-Ing. Katja Jacobsen, Bernd L. Kther, Bernhard Krger, Dr.-Ing. Yeon-Seung Lee, Prof. Dr.Ing. Detlef Schulze and, last but not least, Dr.-Ing. Geir Westgaard.

Thanks for being close friends in every respect to Dr. Uwe Laude, Stephan Meyer and Suse Pfeiffer.

I wish to express my deep gratitude to my parents Prof. Ingrid and Prof. Gerhard Heimann and to my

brother Felix Heimann. Finally, without the love, the steadfast support, endless patience and continuous

pushing of the woman I love, Antje Kmmerer, this work would not have come to its happy end.

Justus Heimann Potsdam, March 2005

Abstract

Ship optimization based on computer simulations has become a decisive factor in the development of

new, economically efficient and environmentally friendly ship hull forms. An important task at an early

design stage is the optimization of the wave-making characteristics of the ship hull since for fast ships a

considerable resistance component stems from the steady ship wave system. Moreover, the ship waves

cause adverse effects in the far field as the wash hits the shore line or other vessels.

A novel hull form optimization approach has been developed and implemented. It builds on a CFD

(Computational Fluid Dynamics) based evaluation of the nonlinear ship wave pattern, on realization of

the cause and effect relation of hull variations and their impact on wave formation, which is accessed

by a perturbation approach, and on wave cut analysis (WCA). WCA yields an excellent assessment

of the wave-making characteristics of a hull form in terms of its free wave spectrum and the wave

associated pattern resistance. These features are highly integrated and controlled by a fully automated

optimization scheme. The scheme is specific in the way it tackles the optimization problem:

The hull adaptation is driven directly by hydrodynamics, avoiding prerequisites to the shape

representation and to the hull modification method.

Wave cut analysis yields the objective function of optimization in terms of the free wave spectrum

and the wave pattern resistance. This considerably improves the system identification allowing

a focused optimization.

The hull optimization is carried out directly for the effective wetted hull portion of the advancing

ship including the effects of dynamic trim and sinkage and wave formation along the hull.

The optimization process is established in terms of an iterative marching scheme of successive

sub-optimization loops. In each loop a region of the solution space is mapped to a simplified

convex quadratic image which merely possesses a single minimum determined by the active

constraints. This enables a straightforward solution procedure and a simultaneous treatment of

a large number of locally acting optimization variables, introducing much freedom to the hull

variation.

All aspects of the proposed optimization approach are presented. Applications of the optimization

scheme to practical ship hull forms show the following:

The wave-making characteristics can be considerably improved with a tangible reduction of the

wave (pattern) resistance.

Hull variations are driven by the optimization to the expected optimum hull shapes with the hull

geometry fully self-adjusting according to the optimization requirements.

Optimal interferences of local wave trains are enforced both in amplitude and phase which results

in a beneficial cancelation of the wave trains.

The specific fingerprints of locally confined hull variations (e.g., at the bulbous bow) can be

effectively traced in the wave spectrum and in the wave pattern resistance to allow a focused

minimization of particular adverse wave components.

The system identification can be significantly improved by utilizing additional information in

terms of the wave spectrum and the distribution of the wave pattern resistance over the waves

range.

Kurzfassung

Die Entwicklung konomisch und kologisch effizienter Schiffe erfordert im groem Mae den Einsatz moderner, rechnergesttzter Verfahren. Eine ausschlaggebende Rolle hierbei spielt die auf Computersimulationen basierende hydrodynamische Optimierung des Schiffsrumpfes, die in einem frhen

Entwurfsstadium erfolgen sollte. Ein erheblicher Teil des Rumpfwiderstandes und damit des Leistungsbedarfs wird bei schnellen Schiffen durch Energieverluste infolge der Ausbildung des stationren

Schiffswellensystems verursacht. Zudem gefhrden Schiffswellen Kstenschutzmanahmen, Uferbebauungen, die Ufervegetation, Hafenanlagen und andere Schiffe sowie Badende.

Im Rahmen der Arbeit wurde ein neuartiger Optimierungsansatz entwickelt und implementiert, der

sich auf die Wellenbildungseigenschaften von Schiffsrmpfen konzentriert. Das Verfahren beruht auf

der rechnergesttzten Simulation des nichtlinearen, stationren Schiffswellenfeldes, auf der Betrachtung des Ursache-Wirkungs-Zusammenhanges von Rumpfvariationen und deren Auswirkung auf die

Wellenbildung sowie auf der Wellenschnittmethodik, die ein ntzliches Instrument fr die Bewertung

der Wellenbildungseigenschaften von Rumpfgeometrien in Form von freien Wellenspektren sowie des

Wellen(bild)widerstands liefert. Diese Bausteine wurden zu einem voll automatischen Optimierungssystem integriert. Das Problem der Wellenwiderstandsminimierung wird hierbei auf eine spezielle Weise angegangen:

Rumpfformadaptionen werden direkt ber Strmungsgren gesteuert, unabhngig von einer

speziellen Geometriedarstellung sowie einem speziellen Verfahren zur Geometrievariation.

Mit Hilfe der Wellenschnittmethodik lsst sich eine aussagekrftige Zielfunktion in Form von

freien Wellenspektren und des Wellen(bild)widerstands ber den Einzelwellenkomponenten definieren, was eine erheblich verbesserte Systemidentifikation und damit eine zielgerichtete Optimierung ermglicht.

Die Optimierung erfolgt direkt fr den benetzten Rumpf unterhalb der Wellenkontur des fahrenden Schiffes in seiner dynamischen Schwimmlage.

Der Optimierungsprozess baut auf aufeinander folgenden Sub-Optimierungsschleifen, in denen

jeweils ein Ausschnitt des Lsungsraums vereinfacht auf einen konvex, quadratischen Raum

abgebildet wird, von dem man wei, dass er ein einziges Minimum aufweist, das von den aktiven

Nebenbedingungen bestimmt wird. Dadurch wird das Lsungsverfahren beschleunigt und die

Verwendung einer groen Zahl von Optimierungsvariablen mit jeweils lokalem Einflussbereich

ermglicht. Dies generiert ein substantielles Ma an Freiheit fr die Formvariation.

Die vorliegende Arbeit stellt alle Aspekte des Optimierungsansatzes ausfhrlich dar. Die Anwendung

des Verfahrens auf praxisnahe Rumpfformen weist folgendes nach:

Die Wellenbildung lsst sich deutlich verringern. Dies ist mit einer erheblichen Reduktion des

Wellen(bild)widerstand verbunden.

Rumpfgeometrievariationen sind entsprechend den Erfordernissen der Optimierung so einstellbar, dass sich erwartete optimale Geometrienderungen ergeben. Freiheiten der Rumpfvariation

knnen dort in das System eingebracht werden, wo im spezifischen Entwurfszusammenhang

noch Freirume vorliegen.

Optimale berlagerungen verschiedener Einzelwellensysteme in Amplitude und Phase entlang

des Rumpfes stellen sich so ein, dass die Einzelwellensysteme sich gegenseitig auslschen.

Der Fingerabdruck lokal begrenzter Rumpfvariationen (z.B. im Bugwulstbereich) lsst sich przise bis in die Wellenspektren und bis zum Wellen(bild)widerstand verfolgen, was eine zielgerichtete Minimierung besonders schdlicher Wellenanteile ermglicht.

Mit den Wellenspektren und der Verteilung des Wellen(bild)widerstands ber den Einzelwellenkomponenten werden wertvolle Informationen bereit gestellt, die eine erheblich verbesserte

Systemidentifikation und damit eine zielgerichtete Optimierung erlauben.

Contents

1

2

Introduction

Hydrodynamic optimization approach

11

2.1

2.2

2.3

11

13

14

3.2

CFD simulation . . . . . . . . . . . .

3.1.1 Boundary value problem . . .

3.1.2 Discretization . . . . . . . . .

3.1.3 Numerical solution . . . . . .

3.1.4 The CFD system SHIPFLOW

3.1.5 Application aspects . . . . . .

Wave cut analysis . . . . . . . . . . .

3.2.1 Boundary value problem . . .

3.2.2 Longitudinal wave cut analysis

3.2.3 Truncation correction . . . . .

3.2.4 Application aspects . . . . . .

19

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

Sensitivity analysis

4.1

4.2

4.3

4.4

4.5

4.6

4.7

4.8

5

Cause and effect chain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Process flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Hydrodynamic analysis

3.1

Design variables . . . . . . . . . . . . . . . . . . . . . . . . .

Perturbation approach . . . . . . . . . . . . . . . . . . . . . .

4.2.1 Perturbation of the discretized boundary value problem

4.2.2 Perturbation magnitudes . . . . . . . . . . . . . . . .

Panel relocation scheme . . . . . . . . . . . . . . . . . . . .

4.3.1 Joint relocation of hull panels . . . . . . . . . . . . .

4.3.2 Relocation range . . . . . . . . . . . . . . . . . . . .

Process flow of sensitivity analysis . . . . . . . . . . . . . . .

Impact on hull form and main form parameters . . . . . . . .

Impact on wave elevation and flow quantities . . . . . . . . .

Impact on free wave spectra and wave pattern resistance . . .

Evaluation of gradients . . . . . . . . . . . . . . . . . . . . .

43

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

5.1

5.2

19

20

24

25

26

28

28

30

32

35

36

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

43

45

45

51

53

56

59

70

71

79

84

89

95

Optimization functional . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

5.2.1 Objective function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

Contents

5.3

5.4

6

Applications

6.1

6.2

6.3

7

Solution scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

Hull form adaptation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

117

Wigley hull . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

FantaRoRo ferry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

Conclusion

7.1

7.2

161

Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

7.2.1 Improvement and validation . . . . . . . . . . . . . . . . . . . .

7.2.2 Coupling to high level geometric modelling and variation schemes

7.2.3 Coupling to seakeeping optimization methods . . . . . . . . . . .

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

Lists

165

Bibliography . . . . .

Figures . . . . . . . .

Tables . . . . . . . . .

Symbols, Abbreviations

161

162

162

163

164

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

165

173

179

181

1 Introduction

In recent time ship optimization based on computer simulations has become a decisive factor in the

development of new, economically efficient and environmentally friendly ship hull forms. An important task at an early design stage is the optimization of the hydrodynamic properties of the ship hull.

For instance, the reduction of the power consumption of a medium-sized ferry typical of fast short-sea

shipping by 1% already reduces the annual fuel costs by about $50.000 and the annual greenhouse gas

emissions by about 350 t.

A considerable resistance component stems from the steady ship wave system even when sailing in

relatively calm water. Fast displacement type ships consume 50% and above of their installed power

to overcome the wave resistance. The waves generated also cause adverse effects in the far field as the

wash hits the shore line or other vessels. It therefore is mandatory to reduce wave-making as much as

possible.

In order to do so early in the design process a novel hull form optimization approach has been developed

and implemented. The optimization process is fully automated requiring no user interaction, however,

permitting it. The steady wave system of a ship moving through otherwise calm water is approximated

by means of CFD (Computational Fluid Dynamics) simulation applying the state of the art nonlinear

free surface Rankine panel module of the SHIPFLOW system. The proposed optimization method

aims at an improvement of the wave-making characteristics, i.e., the minimization of the energy losses

associated with the wave formation along the hull. A valuable measure of the energy losses is attained

by wave cut analysis (WCA) which allows to determine the free wave spectrum and the wave pattern

resistance. Wave cut analysis yields the wave pattern resistance while by convention an integration

of the longitudinal pressure components over the hull provides the wave resistance.

The present optimization approach substantially differs from other approaches to wave resistance minimization in the way it tackles this objective. The substantial differences are:

Hull variations are driven directly by hydrodynamics, i.e., the hull shape is controlled via a

selected flow quantity along the hull surface. Neither, are prerequisites to the hull geometry

representation and its characteristic features nor to the hull variation method imposed. The optimization variables are flow related, constituting a link between the hull geometry and their

hydrodynamic properties.

The objective function of optimization is directly related to the steady ship wave systems in

terms of the free wave spectrum and the wave pattern resistance as determined by a longitudinal

wave cut analysis. This improves the system identification by tracing hull variations up to their

respective fingerprints in the wave spectral distribution and in the wave pattern resistance. Thus,

waves can be influenced and reduced more systematically and with higher control.

The hull optimization is carried out for the effective floating position of the advancing ship

including trim and sinkage. Variations are directly applied to the wetted hull portion under (and

slightly above) the actual wavy free surface level.

1 Introduction

The optimization process is established in terms of an iterative marching scheme of successive

sub-optimization loops which built on top of each other. In each sub-optimization the boundary

value problem is accessed by a linear perturbation approach around the current base point yielding a convex quadratic image of the solution space in a trust-region, a sub-optimization space

around the current base point. This simplified image of the solution space is known to possess a

single minimum determined by the active constraints. Thus, fast and straightforward evaluation

of the gradients of the objective function and the constraints is enabled, allowing a simultaneous

treatment of numerous, locally acting optimization variables which introduces great freedom to

the hull variation.

This particular optimization concept was conceived to pursue the following goals:

Optimization of the wave-making characteristics, i.e., tangible reduction of the wave pattern

resistance of the hull at its dynamic floating position.

Self-adjusting hull geometry variation, driven directly and solely by the hydrodynamic optimization.

Concerted reduction of particular adverse wave components.

Improvement of the system identification, i.e., correlation of cause and effect of local hull variations and their respective fingerprints in the wave spectrum and in the wave pattern resistance.

All aspects of the proposed optimization approach are presented. Chapter 2 introduces the optimization

approach and outlines the process flow. Chapter 3 elaborates the hydrodynamic analysis in terms of

CFD simulation of the nonlinear free surface flow and it represents the wave cut analysis technique

as applied to the computed wave pattern. Chapter 4 substantiates the particular choice of flow related

optimization variables. A perturbation approach is derived to access the boundary value problem in

a straightforward manner. The impact of variations on the hull form, on the flow field, the free surface wave elevation and finally the free wave spectrum and the wave pattern resistance is presented.

The compilation of gradients matrices for the objective function and the constraints is discussed. In

Chapter 5 the constrained minimization problem is introduced, the optimization functional is set-up,

the necessary conditions for a minimum are applied, and the solution by an iterative weighting scheme

is outlined. The chapter closes with a discussion of the hull form adaptation scheme. In chapter 6

the capabilities of the present optimization approach are illustrated by means of one academic and one

practical ship optimization application. The former refers to the standard Wigley hull, the latter to a

state of the art twin screw ferry featuring a bulbous bow and tunnelled transom stern. Finally, in chapter 7 the main achievements and contributions of the present work are summarized and an outlook to

further research work is given.

10

The basic ideas underlying the presented hydrodynamic ship hull optimization approach are

outlined. The pursued goals are stated. The optimization concept and the process flow is

described.

Over the last decade ship optimization based on computer simulations has become an indispensable

factor in the development of new hull forms. Optimization applications comprise all engineering tasks

involved in the design and construction process of a ship. However, an important task at an early design

stage is the optimization of the hydrodynamic properties of the ship hull. In this respect the reduction

of the ship resistance and, hence, the power consumption pays off economically for the shipowner

and/or the operator in terms of savings of operating costs. An additional benefit is less environmental

impact due to reduced emissions of exhaust-gas and noise, less wastage of fossil fuel and a protection

of the shore lines as a consequence of a reduction of the waves emitted by the ship.

A considerable resistance component stems from the steady ship wave system. This effect amplifies 1

as the ship speed increases. Fast displacement type ships consume 50% and above of their installed

power to overcome the wave resistance. It therefore is mandatory to reduce wave-making as much

as possible. In order to do so early in the design process a focused hull form optimization approach

has been developed and implemented in terms of a MATHEMATICA (Wolfram (2003)) based optimization environment. It controls the optimization process flow, it comprises a sensitivity analysis, the

optimization kernel and a longitudinal wave cut analysis (WCA) scheme, and it establishes the connection to the CFD (Computational Fluid Dynamics) system SHIPFLOW. The entire optimization process

is fully automated. Nevertheless, user interaction is supported if need be.

An extensively studied field of ship hydrodynamics is the determination of the wave pattern and the

wave resistance of a ship moving with constant speed through otherwise calm water. This is explained

by the importance of the wave flow for the design of the entire ship hull. Hence, hydrodynamic ship hull

optimization has been challenging for numerous researchers since decades. A comprehensive compilation of literature on geometric ship hull modelling and CFD based hull optimization since the outgoing

19th century until the state of the art in 1998 is given by Harries (1998). The reader is referred to this

publication for additional references. Milestones in the development of wave resistance optimization

methods are in chronological order: 2 Michell (1898), Taylor (1915), Weinblum et al. (1957), Inui

(1962), Pien and Moore (1963), Lin et al. (1963), Wigley (1963) (the collected papers of Sir Thomas

1

This holds true for displacement type ships; planning or semi-planning hulls experience a resistance reduction after transition to the planning condition.

2 this list neither claims to be complete nor representative

11

Method

Flow

analysis

Objective

function

Evaluation of

objective function

Free

variables (N)

Constraints

Experimental

study

EFD

Residuary

resistance,

wave height

Froudes hypothesis

(k-factor method),

wave probes testing

Form parameters,

max. dimensions,

stability...

Interactive /

Automatic

optimization

CFD

inviscid

Wave

(pattern)

resistance

Hull pressure

integration, wave

pattern analysis

Distinctive

hull features

(N 101 )

Present

automatic

optimization

CFD

inviscid

Wave

pattern

resistance

Wave pattern

analysis

(wave spectrum)

Flow related

quantities

(N > 103 )

floating position)

Hull shape

parameters

(N 102 )

- dito -

Havelock on hydrodynamics), Sharma (1966), Kracht (1978), Hsiung (1981), Janson and Larsson

(1996), Nowacki (1997), Sding (1997), and Harries (1998).

Due to the availability of fast and reliable CFD codes, in particular free surface Rankine panel codes,

advanced CASHD (Computer Aided Ship Hull Design) tools for the geometric hull modelling and

a wide range of suitable optimization methods, hydrodynamic ship hull optimization has become

a rapidly developing field of both research and practical application. Tools range from fully automatic implementations over semi-interactive environments to highly interactive optimization processes.

The latter are still favoured by many shipyards and model basins and often involve EFD (Experimental Fluid Dynamics). Recent CFD based wave resistance optimizations are reported, e.g. by

Baskaya (1997), Hirayama et al. (1998), Huan and Huang (1998), Huang et al. (1998), Harries and Abt

(1999), Birk and Harries (2000), Harries et al. (2001), Hendrix et al. (2001), Percival et al. (2001),

Peri et al. (2001), Sding (2001a), Tuck et al. (2002), Birk and Harries (2003), Heimann and Harries

(2003), Maisonneuve et al. (2003), Valdenazzi et al. (2003), Valorani et al. (2003), and Grigoropoulos

(2004).

The formation of waves generated by a ship moving with constant speed through unrestricted calm

water is determined by its hull shape and speed. 3 The design speed, normally, is fixed very early by

the prospective shipowner and/or the operator. A fleet, service and/or route optimization might have

been conducted to determine an appropriate service speed. Hence, a minimum requirement to ship

design is that the ship reaches the demanded service speed. However, more critical is the requirement

of an optimal hydrodynamic performance at (or around) the service speed. In terms of resistance and

power consumption the latter requirement is accomplished solely by a favourable hull design.

The present optimization approach aims at an improvement of the wave-making characteristics, i.e.,

the minimization of the energy losses associated with the wave formation along the hull which can be

measured in terms of the free wave spectrum and the wave pattern resistance. This improves the system

identification by tracing hull variations up to their respective fingerprints in the spectral distribution

and in the wave pattern resistance. Hull variations are driven directly by hydrodynamics, i.e., the hull

shape is controlled via a selected flow quantity along the hull. The motivation of addressing the wave

resistance problem by an optimization approach which focuses on wave cut analysis and, hence, on the

3

12

free wave spectrum and the wave pattern resistance may be best given by citation of selected historical

references:

"[...] The above important equation [editorially: integral equation for the wave pattern resistance determined from the free wave spectrum] not only gives a method of calculating wave-making resistance

but also clearly indicates that the wave-making resistance can be reduced only by reducing the amplitude of each individual elementary wave. Hence, the main effort of our research work in this field is to

derive a method of reducing these amplitudes. [...]"

Pien and Moore (1963)

"[...] Since the free-wave spectrum suffices to describe the dominant part of the entire wave field,

we can approximately predict the wave flow in a major part of the fluid, once we have measured it

at a few places. Further, it should be noted that within the linearized approach [...] the free-wave

spectrum is a linear function of the generating source distribution whereas the wave resistance is not.

Hence we can predict the wave-making properties of certain linear combinations of systems which have

been analyzed individually. For example, if we only know the wave resistance of a certain hull form,

we cannot estimate rationally the wave-resistance of a catamaran, consisting of a pair of such hulls.

However, if we know the free-wave spectrum of the single hull, we can reasonably predict the wave

resistance of any parallel configuration of two or more such hulls in the horizontal plane. Evidently,

this principle has potential applications to design work. [...]"

Sharma in Eggers et al. (1967)

"[...] The use of the wave-resistance integral [editorially: integral equation for the wave pattern resistance determined from the free wave spectrum] with experimental measurements of the amplitude

function A() to determine the total wave resistance provides a direct measurement of the wave resistance without recourse to Froudes hypothesis. This approach is known as wave pattern analysis.

A significant feature of this technique is that the measured quantity A() is linear in the ships disturbance, as compared to the quadratic wave resistance. By relating changes in the wave amplitude

to changes in hull form, a linear optimization of the hull shape can be achieved more directly than is

possible by studying the effects of shape on the total wave resistance. [...]"

Newman (1977)

The present optimization approach substantially differs from other approaches to wave resistance minimization in the way it tackles this objective. Table 2.1 gives an outline of the present optimization

approach in comparison to EFD and other CFD based wave resistance optimization techniques. Differences to related methods comprise the

Objective function: Wave pattern resistance (free wave spectrum)

Evaluation of the objective function: Wave pattern analysis of the computed ship wave pattern

Free variables: Flow related quantities along the hull surface (N > 103 )

Scope: Optimization of confined hull regions or the entire hull in its dynamic floating position

The cause and effect chain from the variation of an optimization variable up to its impact in the wave

spectrum and the wave pattern resistance is, in its proper sense, of complicated and nonlinear nature.

13

In the proposed method the whole chain is simplified, i.e., linearized by linear perturbation of the

boundary value problem around its actual state. The boundary value problem is that of potential theory dealing with the nonlinear free surface conditions. By application of the perturbation approach

the nonlinear relations are transferred to explicit equations depending in a linear and straightforward

manner on the optimization variables.

As a consequence, the whole cause and effect chain starting with the variation of the optimization

variables, resulting in a hull form variation, affecting the flow quantities like the free surface velocities

and the wave formation, and finally bringing about a variation of the free wave spectrum and the wave

pattern resistance, as the latter determined by longitudinal wave cut analysis, can be consistently traced.

This opens the track to an improved system identification and optimization since hull variations can be

explicitly traced all the way to their respective fingerprints in the spectral distribution and in the wave

pattern resistance. In the present scheme the cause and effect relations can be simply represented by

sensitivity terms or gradients.

The optimization process flow is shown in figure 2.1. A full sub-optimization task is performed by

processing the flow chart once from the initial quality assessment down to the improvement assessment.

Normally, a series of sub-optimization tasks is conducted by progressively reentering the loop.

The flow simulations are conducted by means of the nonlinear free surface potential flow module of

the CFD system SHIPFLOW. SHIPFLOW is a Rankine panel solver requiring a panelization of the

hull and the free surface. On the basis of the waves computed with SHIPFLOW a detailed performance

assessment is carried out in terms of the free wave spectrum and the wave pattern resistance. A longitudinal wave cut analysis utilizing a novel implementation of the longitudinal wave cut method in the

code SWASH, see chapter 3, is undertaken.

Within the optimization environment five stages are highly integrated setting up a complete suboptimization task:

Initial quality assessment

Initialization

Sensitivity analysis

Hull improvement

Improvement assessment

Initial quality assessment

The optimization process commences with the initial 4 quality assessment. At this stage the process

control parameters and the hull geometry enter the process. Control parameters concern the global

process flow, the perturbation and sensitivity analysis, the free surface flow simulation, the wave cut

analysis and the set-up and solution of the minimization problem. The hull geometry is given in terms

4

14

Parent / interim

hull form

(CFD)

(process flow)

Performance assessment

(WCA)

Modification of

control parameter set

(Restart)

Initialization

(Re-)Initialization of

optimization task

(variables, constr., bounds)

Systematic hull

perturbation analysis

Transfer matrices

Initial quality

assessment

Systematic variation of

design variables

(flow quantity related)

Sensitivity analysis

(inverse perturb. solution)

(wave elevat., flow quant.)

Impact on

wave energy distribution

(free wave spectrum, RWP)

Evaluation of gradients

(MoM, constraints, bounds)

Minimization of objective

function implying constraints

(iterative scheme)

Generation of

improved hull form &

prediction of gains / changes

(CFD)

further processing

Performance assessment

(WCA)

Automatic adaptation of

optimization task

Stop

Gradients matrices

15

of an offset or panel input point mesh. The offset geometry is used to perform an initial nonlinear

free surface CFD computation with free trim and sinkage which is forced to firm convergence. The

computed free surface elevations and the panel mesh geometry are then passed to a longitudinal wave

cut method which provides the wave spectrum and the wave pattern resistance. 5 Details are given in

chapter 3. The flow solution from SHIPFLOW along with the wave spectrum and the wave pattern

resistance from the wave cut analysis using SWASH serve as the base point (base solution) for the

sub-optimization task.

Initialization

At the second stage the optimization variables (flow related quantities in terms of the hull panel source

densities) are set according to the control parameters. The constraints (bounds) are determined. A

systematic hull perturbation analysis is conducted which correlates the optimization variables, the (discretized) hull geometry, the flow velocities, the wave formation, the wave spectrum and the wave

pattern resistance. This yields transfer matrices which serve as the basic instrument for the sensitivity

analysis. Details are given in chapter 4.

Sensitivity analysis

Once the transfer matrices are assembled the systematic variation of the optimization variables (i.e., the

hull panel source densities) starts. The variation of a single optimization variable has an impact on the

hull form, on the flow field, on the wave formation and, finally, on the free wave spectrum and the wave

pattern resistance. All these influences are determined within the present environment for all variables

(by parallel solving) and are compiled in terms of gradients matrices for the objective function and the

constraints (bounds). Details are given in chapter 4.

Hull improvement

The gradients matrices are directly utilized by the hull improvement instance. The optimum the

minimum of the objective function or measure of merit (MoM) is determined within the limits of

the sub-optimization task. The governing linear equation system is repeatedly solved for the seeked

optimization variables in terms of an iteration scheme which is used to isolate the active from the

passive constraints. Finally, an improved hull form is derived applying the gradients matrices. The hull

form is obtained in terms of a relocated hull panel input point mesh which possesses the same topology

as the initial mesh. Details are given in chapter 5.

Improvement assessment

The final stage of the sub-optimization task serves as the base point (base solution) for the successive

sub-optimization loop. The improved hull geometry is sent again to SHIPFLOW to compute the fully

nonlinear free surface flow. The longitudinal wave cut method then provides the wave spectrum and

the wave pattern resistance.

5

The evaluation of the free surface elevations and their bidirectional interpolation which yields the longitudinal wave cut is

performed outside SHIPFLOW by the wave cut method.

16

The optimization process is terminated by three alternative stopping criteria. Details are given in

section 6.1. If none of the stopping criteria apply the optimization enters the next sub-optimization

loop. The optimization process is fully automated requiring no user interaction. However, the user

may manually interrupt the process at certain points. This allows, for instance, to restart from any

previous solution with a modified or adapted set of control parameters.

Formally, each sub-optimization task is composed of a prediction and a correction step. The prediction

step, essentially, comprises the optimization kernel, i.e., the initialization, the sensitivity analysis and

the hull improvement stage. In the correction step the predicted improvements are assessed by a full

flow and wave cut analysis. Here, the predicted optimization results are corrected for nonlinearities, for

the new dynamic floating position and for effects caused by the reconstruction of a coherent hull surface. The solution of the correction step then serves as the base point (base solution) for the subsequent

sub-optimization task.

17

18

3 Hydrodynamic analysis

The hydrodynamic flow analysis comprises the evaluation of the nonlinear free surface wave

flow by a CFD simulation and the analysis of the energy components contained within the

wave pattern. The first task is accomplished by the free surface potential flow module XPAN

of the CFD system SHIPFLOW. The second task is performed by longitudinal wave cut analysis utilizing a new tool called SWASH, developed by the author. The potential theory based

CFD simulation as well as the wave cut analysis method are presented in this chapter.

An extensively studied field of ship hydrodynamics is the determination of the wave pattern and the

wave resistance of a ship moving with constant speed through otherwise calm water. This is explained

by the importance of the wave flow for the design of the entire ship hull. Nowadays, advanced, reliable

and fast CFD tools are available for the numerical evaluation of the free surface wave flow. Methods

based on inviscid potential theory are widespread and are the first choice in practical wave resistance

computations for a wide range of ship types and flow cases. Moreover, potential flow methods, or

rather free surface Rankine panel methods, are extremely useful in the CFD based wave resistance

optimization, for they are comparatively straightforward in handling, they are fast and they generate

reliable results in terms of detailed flow information for a variety of applications.

Most of the CFD codes utilize integral resistance values derived from hull pressure integration. In this

way, however, information is lost on where beneficial and adverse effects originate. A more detailed

examination is attainable on the basis of the spectral distribution of wave energy along the components

of the steady ship wave system. Hence, the wave pattern evaluated by means of a free surface Rankine

panel method are thoroughly analysed by a longitudinal wave cut analysis. In the present context wave

cut analysis is used both in the sense of system identification and to provide the objective function of

optimization. Wave cut analysis yields the wave pattern resistance RW P while by convention an

integration of the longitudinal pressure components over the hull provides the wave resistance RW .

The steady ship wave system is determined by means of a CFD simulation. In the present work the

nonlinear free surface Rankine panel module of the SHIPFLOW system is applied, see section 3.1.4.

The boundary value problem of potential theory is solved numerically. Boundary value problems are

addressed by boundary element methods (BEM) which require the discretization of all boundaries of

the flow domain, here the hull and the wavy free surface. The boundaries are discretized either by

first order flat or higher order panels. Potential functions are introduced by distributing singularities

(usually sources and/or dipoles) along each panel. The boundary conditions on the hull or the wavy

free surface are enforced for each panel at its control point. The solution of the boundary value problem

leads to a linear(ized) system of equations for the unknown panel source densities. Once the source

19

3 Hydrodynamic analysis

densities are known the velocities and the pressure forces are evaluated. Either linear or nonlinear free

surface conditions are applied. Nonlinear methods have been found superior to linear methods with

respect to predicting the wave amplitudes, the phases and the wave resistance, see Jensen (1988), Raven

(1996) and Janson (1997). For a numerical treatment the nonlinear free surface condition is linearized

around a known base solution. Since the free surface condition is to be applied at the a priori unknown

free surface an iterative solution scheme is used. In the iteration scheme the flow quantities and the

free surface position are updated in an alternating manner. The dynamic floating position is iteratively

updated by a force and momentum balance.

At present it is not possible to compute the ship resistance by CFD methods, whether viscous or inviscid, to the same accuracy as in a towing tank test by means of EFD (Experimental Fluid Dynamics).

CFD tools still cannot compete with EFD when it comes to power prediction. However, even if CFD

cannot provide the absolute power consumption to a high level of accuracy, CFD methods are well

applicable in comparing different hull shapes or variants in terms of their hydrodynamic performance.

This particularly holds true for nonlinear free surface Rankine panel solvers. Compared to EFD, application of CFD is

fast,

relatively inexpensive,

not sensitive to measurement uncertainties,

reproducible and

able to provide detailed flow information.

These properties are especially valuable for CFD based hull optimization. Hence, nonlinear free surface

Rankine panel solvers are frequently used by ship model basins, shipyards and consultants for manual

and sometimes semi-automated or even fully automated hull shape improvement.

Solvers that fall in the category of nonlinear free surface Rankine panel methods and which have

been successfully applied in ship design and optimization are, for instance, KELVIN by Sding

(2000), RAPID by MARIN, see Raven (1996) and Raven (2004), SHALLO/-SHALLO by HSVA,

see Jensen (1988) and Marzi (2004) and the potential flow module XPAN of the SHIPFLOW system

by FLOWTECH, see section 3.1.4.

The boundary value problem of potential free surface flow is thoroughly discussed and documented in

the literature. For instance see Dawson (1977), Ni (1987), Jensen (1988), Jensen et al. (1989), Nowacki

(1995), Raven (1996) and Janson (1997) for literature on this topic. In this section only an outline of

the boundary value problem is given. For a detailed discussion the reader is referred to the given

references.

Reference coordinate system and normalization

The flow is described in an Eulerian sense, i.e., the reference coordinate system is fixed to the ship

having the same speed but does not follow its dynamic trim and sinkage. A right-handed Cartesian

20

system is defined with the origin located at the fore perpendicular in the undisturbed free surface level,

x pointing downstream and z pointing upward, see figure 3.2.

~ | and/or the length

All quantities are used in non-dimensional form, normalized by the ship speed |U

between perpendiculars LPP .

Laplace equation

The governing equation of potential free surface flow is the Laplace equation which is a linear, homogeneous partial differential equation derived from mass continuity

~2 = 0 .

(3.1)

Velocity potential

Assuming an inviscid (ideal) incompressible fluid with irrotational flow, the velocity components can

be determined from the gradient of a scalar potential function

~ = |U

~ | ~ .

U

(3.2)

The total velocity potential comprises the potential of the undisturbed onset flow and the potential

= + .

(3.3)

The velocity potential has to satisfy the Laplace equation (3.1) and it has to comply with the boundary

conditions.

Kinematic and dynamic hull boundary conditions

For a well posed boundary value problem boundary conditions need to be defined. The kinematic hull

boundary condition implies that the flow velocity must have a known component in the hull normal

direction 1

=F ,

n

(3.4)

where, usually F = 0, since no fluid particle is supposed to pass through the surface of a rigid body. In

terms of the disturbance potential the kinematic hull boundary condition reads

b ~n + F ,

= U

n

(3.5)

b = U

~ /|U

~ | indicates the unit velocity vector of the undisturbed onset flow. An additional

where U

dynamic hull boundary condition may be imposed, claiming equilibrium between the hydrodynamic

and hydrostatic pressure forces on the wetted hull portion and the ships weight distribution. This

integral condition primarily determines the dynamic trim and sinkage.

1

21

3 Hydrodynamic analysis

Kinematic and dynamic free surface boundary conditions

At the free surface two more boundary conditions are applied, both of which need to be fulfilled at the

initially unknown wavy free surface. The kinematic free surface condition implies that the flow must

be tangential to the free surface, i.e., no fluid particle is supposed to leave the surface = f (x, y)

+

=0,

x x y y z

(3.6)

where refers to the distance (normalized) of the free surface from the undisturbed initial plane.

therefore describes the wave elevation (normalized). The second condition, the dynamic free surface

boundary condition, states that the static pressure, expressed through Bernoullis equation, must be

constant (atmospheric) at the free surface

#

"

1

2

2

2

+

+

+

1 = 0 .

(3.7)

Fn2 2

x

y

z

The kinematic (3.6) and dynamic (3.7) free surface boundary conditions are given in normalized form.

Fn in (3.7) indicates the Froude number as defined by

VS

Fn =

.

gL

where VS the ship speed and g the acceleration due to gravity. Here the characteristic length L is taken

as the length between perpendiculars LPP .

Linearization of the free surface boundary condition

The free surface problem is nonlinear since the free surface boundary conditions, (3.6) and (3.7), are

nonlinear and are to be imposed at the initially unknown free surface. Hence, the flow quantities depend

nonlinearly on the location of the free surface. Solution methods for the fully nonlinear free surface

problem have been proposed since the 1980s. Present methods linearize the free surface boundary

conditions around a known base solution and solve the problem in an iterative manner. In the iteration scheme the flow quantities and the free surface elevation are updated in alternating steps, see

section 3.1.3. A common approach is to linearize the free surface boundary conditions in a first order

Taylor series expansion around the known base solution and to introduce perturbations due to the hull

and/or the waves. Thus, higher order contributions are assumed to be small. Reviews of different

linearizations are presented, e.g., in Newman (1976) and Raven (1996, 1997).

The majority of the presently available nonlinear free surface Rankine panel solvers adopt the so called

Dawson linearization of the free surface boundary conditions, as originally suggested by Dawson

(1977). In Dawsons method the linearization is conducted around the so called double-body flow 2

as base solution. The kinematic and dynamic boundary conditions are linearized in a perturbation

sense. The perturbation is due to the waves. The free surface boundary conditions, actually, are to be

satisfied at the initially unknown free surface. But, so as to allow a direct solution the conditions are

transferred to and solved at the known free surface. Consistency would require to incorporate transfer

terms to account for this transfer of the boundary conditions. Dawsons linearization is known to be an

2

In a double-body flow the water surface is treated as a symmetry plane at which the underwater portion of the hull is

mirrored.

22

inconsistent linearization since it neglects these transfer terms. However, practical applications have

shown that Dawsons method performs equally well or even more stable than other more consistent

linearizations, see Raven (1996) and Janson (1997) for a thorough investigation. At convergence the

inconsistency of the linearization is no longer relevant.

The free surface elevation and the velocity potential are expanded to first order in Taylor series

around the known base solution . The expansions are then introduced to the free surface boundary

conditions neglecting higher order contributions in and or mixed higher order terms. Finally, the

linearized kinematic and dynamic boundary conditions are combined to give a formulation in known

and unknown velocities and their derivatives. The linearization of the free surface boundary conditions

is described in detail by Janson (1997).

Once the new velocity potential is determined the free surface elevation is obtained from the linearized free surface boundary condition 3

"

2 2 2

#

1 2

(x, y) = Fn 1 +

+

+

2

+

+

, (3.8)

2

x

y

z

x x y y z z

with being the velocity potential of the known base solution.

Radiation and transom conditions

Further to the above conditions a radiation condition is required which implies that the disturbance

velocity due to the presence of the hull vanishes as the distance from the hull r approaches infinity

lim |~ | = 0 .

(3.9)

In the steady state solution a further radiation condition is needed to prevent waves from propagating

upstream of the hull. However, this condition is not formulated explicitly in Rankine panel methods.

It is enforced implicitly by the numerical method, for instance by applying an upstream differencing

scheme and/or an upstream shift of the free surface collocation (control) points in conjunction with the

so called raised panel method. 4

If a ship possesses a transom stern a special transom condition is required to model the flow separation

at the transom edge (according to the Kutta condition) and the flow behaviour aft of the transom stern.

Further boundary conditions are to be applied in case of restricted waters at the bottom and (if present)

at canal-walls. In inviscid flows typically a Neumann condition, i.e., a zero normal flow condition, is

applied at these boundaries.

Singularity distribution

Potential functions are introduced by distributing singularities on the boundaries of the flow domain,

essentially the hull and the wavy free surface. 5 The singularities (e.g., sources or dipoles) themselves

3

In the SHIPFLOW code both an upstream differencing operator is applied longitudinally for the evaluation of the numerical

derivatives and, in addition, the free surface collocation points are shifted upstream relative to the corresponding free

surface panels, see Janson (1997) for details.

5 Unless the wave Green function is applied, known as Havelock singularity, which inherently fulfills the linearized free

surface boundary condition and only requires a singularity distribution on the hull.

23

3 Hydrodynamic analysis

already satisfy the Laplace equation (3.1), see e.g. Newman (1977) and Wendt (1996). Usually, source

singularities are distributed along the boundaries of the flow domain and are integrated to give the

disturbance potential

ZZ

q

dS ,

(3.10)

r pq

SD

where q is the source density per unit area at point q on the boundary surface SD and r pq is the distance

from point q where the source is located to an arbitrary point p where the potential is to be computed.

Introducing (3.10) to the kinematic hull boundary condition (3.5) yields the Fredholm integral equation

of the second kind which serves as an explicit relation for the unknown hull source densities. Introducing (3.10) to the combined, linearized free surface condition (if the free surface is present) yields

additional relations at the unknown free surface. A simultaneous solution of both relations then yields

the seeked a priori unknown source densities on the hull and the free surface.

3.1.2 Discretization

However, a solution of the boundary value problem in closed form for practical ship application cases

is infeasible. Hence, the boundary value problem is solved numerically. A numerical solution requires

the discretization of the boundaries of the flow domain, essentially the hull and a reasonable portion of

the wavy free surface. Panel methods discretize the boundaries by quadrilateral or triangular panels.

Panel methods for computing the potential flow around arbitrary three dimensional bodies were invented by Hess and Smith (1962). They introduced flat quadrilateral panels with a constant singularity

distribution to approximate the boundary value problem. In their notation the flat quadrilateral panels

are generated from so called input points which constitute a coherent point mesh along the hull. From

a topological point of view input points are equivalent to offset points which often are arranged along

hull sections and profile curves and are located directly on the hull surface. However, it is impossible

to connect the input points by flat quadrilateral panels along a curved hull surface without introducing

gaps between the panel edges. In the original Hess & Smith method the flat quadrilateral panels are

determined from the input points so that the panel corner points are at least a close approximation of the

guiding input points and that the panel normal direction is a good approximation of the actual surface

normal. The boundary conditions are enforced for each panel at a panel control point. 6 Later Hess

(1972, 1979) extended the method to higher order curved panels with a linear singularity distribution

across the panels and to lifting surfaces.

As a consequence of the discretization of the boundary value problem the integral equations are transformed to sum equations, which yield an explicit relation for the unknown panel source densities j :

NF S

j Ai j = Bi

j=1

i = 1, , NFS ,

(3.11)

where Ai j indicates the influence coefficient from panel j to the velocity at panel i, Bi is the inhomogeneous term at panel i. NFS accounts for the total number of panels, including the hull and free surface

6

In the original Hess & Smith method the so called null point is suggested where the panel itself induces no velocity in

its own plane. However, the determination of the null point requires the solution of additional simultaneous nonlinear

equations, so, usually, it is simply approximated by the centroid of the panel.

24

panels. The influence coefficients Ai j are determined from the induced velocities 7 Xi j , Yi j , Zi j in the

respective directions and, in case of the free surface portion, from the velocities and velocity derivatives

of the base solution. Details are given in Janson (1997).

The free surface domain is discretized by so called Rankine panels. The linearized boundary condition

is satisfied in a discrete sense at the free surface panel collocation points. Accordingly, the free surface

elevation is evaluated in a discrete sense with the elevation at collocation point i given by virtue of the

linearized free surface boundary condition (3.8):

i =

1 2

Fn 1 + 2x i + 2y i + 2z i 2 (x i x i + y i y i + z i z i )

2

i = NH + 1, , NFS , (3.12)

with x = /x etc. and NH < i NFS for the free surface panels.

3.1.3 Numerical solution

The numerical solution of the nonlinear free surface boundary value problem is thoroughly described

by Jensen (1988), Raven (1996) and Janson (1997).

State of the art CFD systems make use of the so called raised panel method. The free surface panels are

raised by a certain 8 distance above the free surface collocation points possibly in conjunction with an

upstream collocation point shift, Jensen et al. (1986). In this way the radiation condition is implicitly

enforced and convergence is improved. As a further numerical feature the velocity derivatives at the

free surface are determined along longitudinal panel rows by an upstream differencing scheme and

transversally along rows with constant x by a central difference scheme, as originally suggested by

Dawson (1977).

Equation (3.11) provides a linear system of equations for the unknown hull and free surface panel

source densities. But, since the free surface condition is to be applied at the a priori unknown free

surface an iterative solution scheme is used where (3.11) is solved successively. In each iteration the

boundary value problem is linearized with respect to the solution of the previous iteration. The setup

is as follows:

1. Compute a base solution, e.g., by adopting a Neumann or a double-body condition

undisturbed free surface.

at the

2. Use the velocities from the previous step to compute the free surface elevation from the linearized

dynamic free surface condition (3.8) at the undisturbed free surface.

3. Move the free surface panels to the new free surface computed at the previous step.

4. Move the latest known flow solution to the adjusted free surface panelization of the previous

step, without being recomputed.

5. Solve the linearized boundary value problem according to (3.11) at the actual free surface.

7

Xi j , Yi j , Zi j are pure geometrical quantities and represent the induced portion from panel j to the velocity at panel i in the

x-, y- and z-direction respectively.

8 carefully selected

9 Applying the double-body condition yields the Bernoulli wave.

25

3 Hydrodynamic analysis

Fig. 3.1: Zonal approach of the CFD system SHIPFLOW (courtesy of FLOWTECH Int. AB.)

6. Apply the solution of the previous step to adjust the free surface elevation according to the

linearized dynamic free surface condition (3.8).

7. Continue with step 3 until convergence.

Firm convergence is requested for the maximum wave change, for the maximum change of the source

densities and, if allowed to freely adjust, for the trim and sinkage of the hull in successive iterations.

The dynamic floating position is updated via a force and momentum balance at each iteration.

The wave resistance is computed from the converged solution by integration of the longitudinal pressure components over the hull. For further details be referred to Janson (1997).

3.1.4 The CFD system SHIPFLOW

In the present work the CFD simulation of the steady ship wave system is conducted by means of the

free surface Rankine panel module XPAN of the CFD system SHIPFLOW 10 , see e.g. Larsson et al.

(1989, 1990, 2004a,b), Larsson (1993) and Janson (1997).

SHIPFLOW solves the boundary value problem of potential theory numerically. The boundaries of the

flow domain, essentially the hull and the wavy free surface, are discretized either by flat or, as in the

higher order method, by bi-parabolic quadrilateral panels. Source singularities 11 are distributed along

each panel either with a constant source density or, as in the higher order method, with a linear distribution of the source density along the panel. The boundary conditions on the hull and the wavy free

surface are for each panel enforced at a control point defined as the centroid of the panel. Raised panels

are utilized at the free surface in conjunction with an upstream collocation point shift. Tangential flow

separation at the edge of a transom stern (if present) is modelled by an additional free surface transom

group which extents downstream of the transom. So called extra hull panels above the free surface level

might be used in case of ships with transom sterns or complicated geometries to stabilize the convergence history. The solution of the boundary value problem leads to a linear system of equations for the

unknown panel source densities. Once the source densities are known, the velocities and the pressure

forces are evaluated. SHIPFLOW adopts either a linear or the nonlinear free surface condition. For a

10

11

and dipoles on lifting surfaces and on the wake sheet

26

numerical treatment the nonlinear free surface condition is linearized around a known base solution.

Since the free surface condition is to be applied at the a priori unknown free surface contour an iterative solution scheme is used. In the iteration scheme the flow quantities, i.e., the source densities and

velocities, and the free surface position are updated in an alternating manner, see section 3.1.3. The

dynamic floating position is iteratively updated by a force and momentum balance.

Prominent features of SHIPFLOW/XPAN are:

Wave resistance from pressure integration over the wetted hull surface

Wave pattern resistance from a transverse wave cut method 12

Linear and nonlinear wave pattern

Longitudinal and transverse wave cuts

Flow velocities

Pressure distribution

Dynamic trim and sinkage

Shallow water effects

Multihull option

Automated mesh generation

Propeller representation by an actuator disk model

Lifting surfaces

Induced resistance and lift

The SHIPFLOW code is specifically designed for steady ship hydrodynamics. In addition to the inviscid potential free surface flow, SHIPFLOW also offers modules to compute boundary layers, viscous

flow effects and drag. The way SHIPFLOW handles these unique problems is by splitting the flow

into three zones, which allows an efficient simulation of the governing flow phenomena. The zonal

approach is outlined in figure 3.1. A full investigation comprises the following three stages which are

processed successively:

Zone 1 Potential flow calculation. Rankine panel method with linear or nonlinear free surface bound-

Zone 2 Boundary layer calculation. Momentum integral method for laminar and turbulent boundary

Zone 3 RANSE calculation. Finite difference based solution of the Reynolds-Average Navier-Stokes

12

27

3 Hydrodynamic analysis

SHIPFLOW is robust in application and extensively validated. It comes with an automated mesh

generation utility, it provides a variety of control mechanism and I/O capabilities to the user, supplying

detailed results. SHIPFLOW Release 2.8.10 and higher provides additional, specifically customized 13

interface capabilities like the I/O of the panel geometry, of selected influence coefficients and certain

mechanism for the SHIPFLOW execution control. Moreover, the SHIPFLOW system is widely used

by ship model basins, shipyards and consultants both for hydrodynamic analysis in the preliminary

design stage and for hull shape improvement.

The SHIPFLOW system was therefore selected for the CFD computations in the present work.

3.1.5 Application aspects

The inherent neglect of the viscosity by potential flow methods implies that wave-viscous interactions

cannot be captured. A thorough investigation, both numerically and experimentally, of wave-viscous

interactions including propeller effects is reported by Nowacki and Sharma (1972). Wave-viscous interactions occur mainly in the stern region where wave amplitudes tend to be overpredicted by CFD.

However, the differences are small for the ship types and speed range particularly addressed by the

present optimization approach: relatively fast and slender hulls like ferries, RoPax, container vessels

etc.. Potential flow codes also assume the boundary layer to be sufficiently thin. This is a valid assumption for the forebody of a displacement ship and for the entire hull of a fast ship where the boundary

layer detaches the transom edge without considerable thickening. CFD validation studies comparing

computed with measured wave pattern showed the excellent ability of modern nonlinear free surface

Rankine panel methods to model the steady wave systems of the ship types and speed range 14 relevant

for the present work, see e.g. Raven (1996), Janson (1997), Raven and Prins (1998a,b), Nowacki et al.

(1999), Heimann (2000), Valdenazzi et al. (2003), Heimann and Harries (2003). Furthermore, Rankine

panel methods proved to be sensitive even to small and locally confined hull form changes.

A well known and extensively studied feature of CFD methods, whether viscous or inviscid, is their

dependence upon the grid or mesh on which the discretized flow equations are solved numerically.

This property also applies to Rankine panel methods. Further uncertainties are introduced by the

raised panel method. These effects are attributed to numerical dispersion (wavelength error) and in

particular to numerical damping (amplitude error). Within the scope of the SHIPFLOW solver griddependence and CFD validation studies were reported, e.g., by Janson (1997), Harries and Schulze

(1997), Schulze and Harries (1997) and Lee (2003). Janson (1997) provides a valuable guideline for

the panel mesh generation and the selection of certain control parameters.

Raven (1996) and Janson (1997) report comprehensive studies to the application of the nonlinear versus

the linear free surface conditions. In general, nonlinear free surface Rankine panel codes have been

found superior in predicting the wave amplitudes, the phases and the wave resistance.

The steady ship wave system, as determined by means of a CFD simulation, is further assessed by a

wave cut analysis. In the present scheme the longitudinal wave cut method is applied by means of

13

14

Professor Dr. Carl-Erik Janson, Chalmers University of Technology and FLOWTECH Int. AB, Gteborg, Sweden

0.2 Fn 0.4 (potentially higher Froude numbers)

28

Q=0

Q>0

z, z

x

yWC

VS

xE

Fig. 3.2: Schematic sketch of the ship wave system (only one symmetric half is considered).

the MATHEMATICA based tool called SWASH (Ship Waves Analysis Light) developed by the author.

SWASH is an implementation of the wave cut method following Sharma (1963, 1966) and Eggers et al.

(1967).

Wave cut analysis (WCA) in the sense of system identification has shown to be a powerful tool

in CFD based optimization of the wave-making characteristics of ships, e.g. Heimann (2000) and

Heimann and Harries (2003). Most of the CFD codes utilize integral resistance values derived from

pressure integration. In this way, however, information is lost on where beneficial and adverse effects

originate. A more detailed examination is attainable on the basis of the spectral distribution of wave

energy along the components of the steady ship wave system by utilizing the so called free wave spectrum. This improves the system identification, i.e., the energy loss associated with the transverse and

diverging waves can be studied more clearly. Furthermore, the impact of local hull variations on the

individual components of wave formation can be traced by investigating the changes in the wave spectrum with respect to the associated hull variations. This then allows a reduction of particular adverse

wave components. However, the primary aim of the present optimization approach is to minimize the

total energy losses associated with the wave formation along the hull which can be measured in terms

of the wave pattern resistance RW P as determined by the wave cut analysis.

As an additional benefit computed wave pattern and hence wave cut analysis based on them prove

to be less sensitive to the flow field discretization than hull pressure integration. This is a tentative

29

3 Hydrodynamic analysis

result of related studies concerning that topic, e.g. Janson (1997), Raven and Prins (1998a,b) and

Heimann and Harries (2003).

3.2.1 Boundary value problem

The boundary value problem of the steady linear free surface potential flow is thoroughly discussed and

documented in the literature, see Michell (1898), Lord Kelvin (1906), Wehausen and Laitone (1960),

Eggers (1962), Inui (1962), Newman (1963), Wigley (1963) (the collected papers of Sir Thomas Havelock on hydrodynamics) and Sharma (1965) for a condensed selection of literature on this topic. In this

section only a brief outline of the problem is given. For an introduction see also Newman (1977).

Reference coordinate system and normalization

The flow is described in an Eulerian sense, i.e., the reference coordinate system is fixed to the ship.

A right-handed Cartesian system is defined with the origin located at the fore perpendicular in the

undisturbed free surface level, x pointing downstream and z pointing upward, see figure 3.2. Infinite

water depth is assumed, i.e., the water depth h is greater than half the fundamental wavelength 0 :

h > 0 /2.

All quantities are used in non-dimensional form. Length related measures are normalized by the basic

wavenumber

k0 =

g

,

VS2

(3.13)

where g is the acceleration due to gravity and VS the ship speed. VS is assumed constant. In other

words: The problem described is of steady-state nature and all dependence on time is blended out.

The wave pattern resistance RW P is obtained in non-dimensional form by

k03

,

g

RW P = RWP

(3.14)

where the bold expression RWP stands for the dimensional resistance component and is the water

density.

Dispersion relation, wavelength and wavenumbers

Prior to the further derivations some important quantities should be defined. The dispersion relation

for infinite water depth connects the wavenumber k, the wavelength of the plane progressive wave

moving at an oblique angle with respect to the x-axis and the speed of the steady wave system. The

dispersion relation is given in non-dimensional form (normalized by k0 ) by

k() =

2

= sec2 () .

()

(3.15)

() =

30

2

k()

(3.16)

and the wavenumber k is defined according to (3.15) in non-dimensional form (normalized by k0 ) by

k() = sec2 () .

(3.17)

s() = k() cos() = sec()

(3.18)

u() = k() sin() = sec() tan()

(3.19)

are derived. The component wave direction , the longitudinal s and transverse u wavenumbers are

defined within the integration bounds 15

0 /2 ,

1s<,

0u<.

, s and u are fully convertible into one another, and so are the relations based on them. 16

Linearized free surface boundary condition

A non-viscous (ideal) incompressible fluid with irrotational flow is assumed. The kinematic (3.6) and

dynamic (3.7) free surface boundary conditions are linearized around the undisturbed free surface level

with respect to the undisturbed onset flow. Their combination yields the well known Neumann-Kelvin

boundary condition which is given in normalized form by

2

x

=0.

z

(3.20)

A velocity potential which satisfies the Laplace equation (3.1) and the linearized free surface boundary

condition (3.20) can be derived following Havelock (1934a) at a large distance from the disturbance

(the ship) as

Z/2

=

e(z sec

())

(3.21)

/2

where AS and AC denote the sine (odd) and cosine (even) component, respectively, of the so called free

wave spectrum. The velocity potential (3.21) satisfies the linearized boundary condition at the free

surface in the far field of the ship.

For modelling the whole wave system, including the local, bounded waves in the vicinity of the hull,

a special form of the velocity potential can be utilized which is made up by another Greens function,

as given in Havelock (1932). This inherently satisfies the linearized free surface boundary condition

(3.20) in the whole flow domain, see e.g. Eggers et al. (1967) for details.

The wave pattern resistance can be determined from the far field waves represented by the velocity

potential (3.21) alone, as shown below.

15

16

The lower integration bound of s = 1 conforms by virtue of (3.18) with = 0 , i.e., no waves are longer than those

determined by the fundamental wavelength.

31

3 Hydrodynamic analysis

3.2.2 Longitudinal wave cut analysis

The present optimization approach aims at the minimization of the wave pattern resistance which is

determined by a longitudinal wave cut analysis. In case of flows symmetric to the ship center plane, a

single infinitely long wave cut parallel to the ship suffices to perform the whole analysis.

The longitudinal wave cut method is preferred to other wave analysis methods since it is comparatively fast, relatively robust in application and it has a broad application range. This points are further

discussed in section 3.2.4.

The used notation and axes orientation follows to a large extent Sharma (1963). Complementary material is selected from Sharma (1966) and Eggers et al. (1967).

Free wave system

Out of the double infinite manifold of all possible plane progressive waves, varying both in wavenumber k and in component wave direction , it is only a single infinite set of so called free waves which

contribute to the wave pattern resistance, see e.g. Eggers et al. (1967). These free waves satisfy condition (3.15), i.e., wavenumber and wavelength of each component of the free wave system are determined one-to-one by its wave direction. Furthermore, outside a confined region around the disturbance

by the ship it is the free waves which constitute the wave pattern. Whereas, the local disturbance

arising from the presence of the nonfree waves vanish with the square of the reciprocal distance from

the disturbance.

Assuming a control surface located parallel to the course of the ship at infinite distance 17 to the ship,

the free wave system can be derived from the velocity potential (3.21) and from the Neumann-Kelvin

linearization of the dynamic free surface condition. The elevation of the free waves expressed in terms

of the component wave direction is given by

Z/2

(x, y) = [AS () sin(s x + u y) AC () cos(s x + u y)] d ,

(3.22)

Z

(x, y) =

1

d

ds .

ds

(3.23)

After rearranging in terms of s, relation (3.23) results in a form suitable for further treatment

Z

(x, y) =

1

[AS () sin(u y) AC () cos(u y)] cos(s x)}

17

d

ds .

ds

(3.24)

Here the control surface at y is adopted for practical reasons in order to conform with the SHIPFLOW internal

representation; an analogous derivation can be conducted for the control surface at y +, see Sharma (1963).

32

Relations (3.22) to (3.24) can be interpreted as a superposition of elementary plane progressive waves

of all feasible wave directions and varying wave amplitudes in compliance with the steadiness condition

for the phase velocity, see relation (3.17) and compare e.g. Havelock (1934b). This results in the well

known Kelvin wave pattern which is confined in a bisector to the x-axis of 19 28 on deep water.

The waves follow the ship. The transverse wave component according to the basic wavenumber and

the basic wavelength travels in the ships advance direction, i.e., with = 0 . The transverse waves

possess transversally aligned wave crests and are composed of component waves up to = 35 16 .

Larger values of correspond to the so called diverging waves which have shorter wavelength and

converge toward the origin, compare e.g. Newman (1977).

Fourier transformation

The continuous Fourier integral is applied to transform between the free waves and the related free

wave spectrum, as it is represented further below. The Fourier transformation is given by

1

(x, y) =

Z

[S(s) sin(s x) +C(s) cos(s x)] ds ,

(3.25)

Z

(x, y) sin(s x) dx ,

S(s) =

(x, y) cos(s x) dx ,

C(s) =

(3.26)

where S and C denote the sine (odd) and cosine (even) components, respectively, of the Fourier transforms. In terms of the longitudinal wave cut analysis (x, y = const.) in the Fourier transforms (3.26)

represents a longitudinal wave cut either determined from a CFD simulation or from wave probe measurements in a towing tank. The longitudinal wave cut is given as wave elevations along varying

x-positions for a constant transverse cutting plane or offtrack position y = const..

Free wave spectrum

The Fourier transforms and the components of the free wave spectrum can be correlated by equating

(3.25) with the expression for the free wave pattern (3.24) followed by a separation of variables. By

virtue of the Fourier transforms (3.26) this yields

Z

S(s) =

d

,

ds

d

.

ds

C(s) =

(3.27)

33

3 Hydrodynamic analysis

Resolving (3.27) for AS and AC finally results in

1

[S(s) cos(u y) +C(s) sin(u y)]

1

AC () = [S(s) sin(u y) C(s) cos(u y)]

AS () =

ds

,

d

ds

,

d

(3.28)

with AS and AC denoting the sine (odd) and cosine (even) components, respectively, of the free wave

spectrum. For the sake of simplicity AS and AC are referred to as free wave spectra from now on.

The notation of the free wave spectra in terms of A suggests a wave amplitude spectrum. Hence,

AS and AC are frequently called wave amplitude functions, e.g. Newman (1977). However, this terminology is slightly misleading, since AS and AC do not actually have the physical meaning of wave

amplitudes. In fact, the free wave spectra are better denoted as wave energy equivalents, i.e., terms

which express the energy losses associated with the formation of the individual components of the free

wave system. The wave spectra correlate with the real and imaginary part of the degenerated form of

the Kochin function. A detailed derivation and review is given by Eggers et al. (1967).

The free wave spectra are pure functions of the hull shape and the ship speed assuming infinite water depth. They yield a direct measure of the wave energy distribution along the components of the

ship wave system. This is why they are extremely valuable in ship hull optimization. In the present

optimization scheme the free wave spectra are directly utilized for decision making in terms of the objective function. However, the free wave spectra ought to be applied with some care since the Fourier

transforms and the free wave spectra are not invariant against coordinate transformations, like a shift

of the origin, whereas the integrand of the wave pattern resistance, equations (3.30) and (3.32) below,

is invariant against coordinate transformations.

Wave pattern resistance

Formally, the wave pattern resistance is deduced from energy considerations, i.e., the conservation of

momentum in a control volume which contains the steady ship wave system composed of a superposition of plane progressive waves of all feasible wave directions and varying wave amplitudes in

compliance with the steadiness condition for the phase velocity. The wave pattern resistance can be

evaluated directly from the Fourier transforms without reference to the free wave spectra which yields

(in non-dimensional form)

r

Z

1 2

1

1

RW P =

S (s) +C2 (s)

1 2 ds .

(3.29)

s

s

1

A different notation, going back to Havelock, adopts the free wave spectra and is given (in nondimensional form) by

RW P =

Z/2

0

A2S () + AC2 () cos3 () d .

(3.30)

Both expressions for the wave pattern resistance are equivalent and fully convertible into one another.

Expression (3.30) is favoured in the present work because the notation in terms of the free wave spectra and the component wave direction is more intuitive both for the optimization application (objective

34

function) and for the interpretation of the results. For instance, application of (3.30) enables the designer to trace the cause and effect of hull variations and wave pattern changes. The factor cos3 () implies that the predominant portion of the wave pattern resistance originates from the transverse waves,

which explains why ship hull optimizations often focus on the transverse waves regime. However, due

to the positive definite nature of expression (3.30) it is impossible to fully neutralize the wave pattern

resistance, e.g., by favourable interference effects. 18

From the wave pattern resistance coefficient 19

RW P

CW P =

=

0.5 Sre f k02

Z/2

CW P () d ,

(3.31)

an extremely useful and descriptive chart, as will become evident in chapter 6, is obtained by plotting

the integrand of (3.31) as a function of the wave direction

CW P () =

2

0.5 Sre f k0

(3.32)

with Sre f denoting the wetted hull surface area for normalization. Expression (3.32) yields a resistance

function which measures the wave energy distribution along the components of the ship wave system.

SWASH conducts the numerical integrations by means of the trapezoidal rule assuming a dense sampling which proved to be robust and straightforward.

In practical applications due to limited computer resources (CFD) or restrictions due to tank wall reflections in experiments (EFD) longitudinal wave cuts need to be truncated at a finite distance downstream

of the hull. In order to retain the wave energy contained within the wave pattern downstream of the

truncation point SWASH treats wave cuts of finite length with a special truncation correction. At a sufficiently large distance aft of the stern the wave cut is usually dominated by transverse waves following

the ship. This gives rise to utilizing an analytic truncation function composed of harmonic waves decaying to infinity. A truncation correction based on an analytical asymptotic extension of the wave cut

is applied here.

Sharma (1966) and Eggers et al. (1967) suggested the following approach for an analytical asymptotic

extension of the wave cut infinitely downstream

e

(x, y)

xE+

c1 cos(x) c2 sin(x)

c3 + x

x + ,

(3.33)

where the coefficients c1 , c2 and c3 are determined by fitting (3.33) to the tail end of the wave cut

signal up to the truncation point xE . The unknown coefficients are determined by regression analysis. 20

Because of the strong asymptotic decay of both the wave elevation and the slope at x , according

to Sharma, practically no correction is needed to account for the truncation of the wave cut upstream.

18

note the non-dimensional form of RW P in (3.14)

20 In the present implementation the Levenberg-Marquardt method is applied.

19

35

3 Hydrodynamic analysis

Sharma suggested to drop the coefficient c3 from (3.33) for the sake of simplicity if the wave cut signal

extends far downstream xE , see e.g. Eggers et al. (1967). Lalli et al. (1998, 2000) suggested to drop

c3 not only for the sake of simplicity, but also because it stabilized their results. Lalli systematically

applied longitudinal wave cut analysis to measured wave signals of a Series 60 model and a fast hardchine catamaran. As a consequence of dropping c3 the singularity of the correction function (3.33) is

inevitably fixed to the origin. Otherwise, retaining c3 in the regression analysis causes the singularity

to self-adjusting, promising a better fit due to the additional degree of freedom. The latter conforms

with the experiences gained by the author.

In order to evaluate the analytical asymptotic extension of the longitudinal wave cut weighted Fourier

transforms are introduced. The weighted Fourier transforms are given by virtue of (3.26) as

SW FT (s) =

ZxE p

s2 1

s2 1

ZxE

CW FT (s) =

(x, y) sin(s x) dx +

Z p

xE

(x, y) cos(s x) dx +

s2 1 e

(x, y) sin(s x) dx ,

Z p

xE

s2 1 e

(x, y) cos(s x) dx ,

(3.34)

where SW FT and CW FT denote the sine (odd) and cosine (even) component, respectively, of the

weighted Fourier transforms, the weight being s2 1. The respective first integrals over (, xE ]

in (3.34) are evaluated by numerical integration as before with the longitudinal wave cut (x, y) truncated at xE . The respective second integrals over [xE , ) in (3.34) are evaluated in closed analytical

form by means of the complementary Fresnel integrals, e.g. Press et al. (1992) and Bronstein et al.

(1993), which arise in the description of diffraction phenomena. The closed analytical evaluation is

elaborated in detail in Sharma (1966) and Eggers et al. (1967) and not repeated here.

Introducing the weighted Fourier transforms (3.34) to the free wave spectra (3.28) yields

s

[SW FT (s) cos(u y) +CW FT (s) sin(u y)] ,

s

AC () = [SW FT (s) sin(u y) CW FT (s) cos(u y)] .

AS () =

(3.35)

Finally the wave pattern resistance is determined by (3.30) with the free wave spectra from (3.35).

3.2.4 Application aspects

The longitudinal wave cut method is favoured over other wave analysis methods. Advantages of the

longitudinal cut method are:

In case of flows symmetric to the ship center plane a single wave cut parallel to the ship suffices

to perform the whole analysis. No wave slopes need to be determined.

Asymmetric flows can be treated smoothly.

The tendency of over-predicting the stern waves in non-viscous CFD simulations or the viscous

wake flow phenomena in EFD does not affect the analysis as much as in case of transverse wave

cuts.

36

The longitudinal wave cut method is comparatively less affected by lateral stretching of the free

surface panelization in CFD, provided the longitudinal wave cut stays in the neighbourhood of

the ship.

The method is well applicable to wave probe measurements in a towing tank which allows easy

comparison between CFD and EFD.

Disadvantages of the longitudinal cut method are:

The method is error-prone in case of an improper application of the truncation correction which,

usually, is avoidable by monitoring the fitting process.

Proper application of the truncation correction demands for comparatively large free surface

domains in CFD, extended fairly downstream, which is costly both in terms of computation time

and memory consumption.

A bidirectional interpolation of the free surface mesh is required, first transversally then longitudinally, in order to determine the longitudinal wave cut from computed discrete free surface

elevations. This introduce further inaccuracies.

Basically, inverting the advantages of the longitudinal wave cut method to disadvantages and vice

versa yields the equivalents for the transverse cut method. A comprehensive overview of the underlying theory of the different methods is given in Eggers (1976). Advocates of the transverse wave

cut method are for instance Nakos (1991) and Raven and Prins (1998a,b). Whereas, recent studies to

the applicability of the longitudinal wave cut method, especially focusing on measured wave cuts of

fast and high-speed crafts, were reported by Dumez and Cordier (1996), Brizzolara et al. (1998) and

Lalli et al. (1998, 2000). Spinney (2002) compared four different longitudinal and transverse wave cut

techniques in a thorough numerical study applying the linear and nonlinear Rankine panel module of

the SHIPFLOW code. In terms of the longitudinal wave cut method he utilized an implementation of

the method following Michelsen and Uberoi (1971) and the SWASH code. Moreover, implementations

of the single and multiple transverse wave cut method following Eggers et al. (1967) were utilized. He

found that the multiple transverse wave cut method is robust in application, however, some care is

required in selecting appropriate transverse cutting planes, see also Raven and Prins (1998a,b). In his

study the longitudinal wave cut method, particularly the truncation correction, appeared to be sensitive

to the signal length and the lateral position of the cutting plane.

According to the experience of the author, improper application of the truncation correction in the

longitudinal wave cut method can be avoided by a proper setup of the fitting process. An alternative

approach of dealing with truncated wave cuts is the application of the so called Fourier Transformation

Proper (FTP) as proposed by Schmiechen (1999). He claims that his method permits precisely to

identify spectra from truncated records, i.e., from a finite set of sampled data of transient signals and

that it overcomes the problem of classical uncertainty.

Sding (2001b) proposed a different strategy of wave pattern analysis which is independent of longitudinal and transverse wave cuts. In his method the coefficients of the harmonic wave components of all

feasible directions which constitute the free wave system are determined directly by a least-square fitting procedure. The fitting is conducted with respect to the computed (by CFD) wave elevations at the

discrete free surface collocation points sampled in a rectangular region aft of the ship. Sdings wave

analysis method requires special care in the setup of the free surface discretization and the selection of

the fitting region.

37

3 Hydrodynamic analysis

In the present work the longitudinal wave cut method with truncation correction following Sharma

(1966) and Eggers et al. (1967) is applied for the wave pattern analysis. Practical application of longitudinal wave cut analysis mainly depends on the ships speed, on the lateral position of the cutting plane

yWC and on the utilizable length of the wave signal up to the truncation point xE , i.e., the downstream

edge of the computational free surface domain. The latter two influences are in opposite direction. In

order to avoid serious analysis errors and to fulfill the requirement of a free wave system the cut is to

stay outside a region close to the hull where local flow effects are predominant. However, the greater

the lateral distance, the shorter is the utilizable wave signal until truncation, which is a detrimental

effect of the Kelvin angle and the limited extension of the computational domain. Various studies

indicate that satisfactory results are obtained for medium speed and fast monohulls at lateral cut positions in the range 0.2 y/LPP 0.3 provided the cut extends far enough downstream to allow for

a reasonable truncation correction. The cut should feature several times the fundamental wavelength.

For details see Lalli et al. (1998, 2000), Nowacki et al. (1999), Heimann (2000) and Valdenazzi et al.

(2003).

The effects of the lateral position of the cutting plane yWC and the signal length up to the truncation

point xE on the wave pattern resistance from longitudinal wave cut analysis are subsequently demonstrated by means of a theoretical and a practical example.

In the theoretical example the linearized Neumann-Kelvin wave pattern of an even submerged dipole

was investigated by wave cut analysis. The dipole moment was adjusted so that the problem of a flow

about a sphere of unit radius, located at the coordinate origin directly beneath the free surface level, was

simulated. A basic wavenumber of k0 = 1 was adopted which conforms to a fundamental wavelength

of 0 = 2 . The exact theoretical, non-dimensional wave resistance is RW = 1.528, see e.g. Sharma

(1966). The linearized wave pattern was determined from the expressions in the original publication by

Havelock (1928). A representation of the wave system, computed with the aid of the MATHEMATICA

package, is given in figure 3.3. Longitudinal wave cuts were evaluated for the center plane (including

the local wave system) and for three transversal positions at yWC /0 = 0.5, 1, 2, with a signal length of

up to xE /0 = 10. The wave cuts off the center plane were analysed with SWASH with and without

correction of the truncation error as well as for three different signal lengths xE /0 = 5, 7, 10. A

comparison of the wave pattern resistance from wave cut analysis with the exact theoretical value

is shown in figure 3.4 and 3.5. Figure 3.4 reveals that the wave pattern resistance approaches the

exact theoretical value with increasing signal length. Whereas figure 3.5 illustrates the immanent

effect of shortening the utilizable wave signal length by shifting the wave cut transversally. A large

error is introduced to RW P at the outmost wave cut. The truncation correction proves to be capable

of recovering the wave energy contained within the wave pattern downstream of the truncation point

fairly well. Generally, in order to allow a truncation correction a minimum fitting length (tail end of the

wave cut up to the truncation point xE ) of 0.5 0 is required. However, in terms of a stable regression

analysis a larger fitting length of about one fundamental wavelength is advisable. The total effect of

the fitting length on the wave pattern resistance is negligible as various tests have shown.

In the second example the wave pattern of the well-known Hamburg test case (alias Ville de Mercure) was studied. According to Bertram et al. (1992) the Fn = 0.24, LPP = 153.7 m, TF = 9.2 m

and TA = 10.3 m, CB = 0.652. The single-screw container vessel features a bulbous bow, a stern

bulb and a transom stern. Longitudinal wave cuts for varying transversal positions in the range

0.262 yWC /LPP 0.574 and signal lengths in the range 2 xE /LPP 4 were determined by a

nonlinear free surface CFD simulation with SHIPFLOW allowing trim and sinkage to adjust freely.

A comparatively dense hull and free surface panelization was adopted, extending 4 LPP downstream,

38

Fig. 3.3: Linearized wave pattern of an even submerged dipole beneath the origin, k0 = 1 and 0 = 2 .

comprising the Kelvin wedge, with a longitudinal and lateral panel stretching towards the hull. The

computed wave cuts were analysed with SWASH and the results in terms of the wave pattern resistance coefficient 21 CW P were compared to the corresponding results from measurements. The latter

were attained from wave probe measurements conducted by the Hamburg Ship Model Basin (HSVA),

Bertram et al. (1992). No tank wall reflections were present in the measured wave cuts up to a signal

length of xE /LPP = 4. Both the computed and the measured wave cuts were corrected for truncation

errors. Figure 3.6 shows that already a signal length of xE /LPP = 2.5 yields almost the same value for

CW P as the longer wave cuts for the innermost position. The wave pattern resistance from the com21

39

3 Hydrodynamic analysis

100

90

80

wave cut position : yWC / 0 = 0.5

truncation correction

70

no truncation correction

60

5

7

8

9

Normalized signal length xE / 0 [-]

10

Fig. 3.4: Related wave pattern resistance of the even submerged dipole, influence of the signal length

xE and the truncation correction.

100

90

80

signal length : xE / 0 = 10

truncation correction

70

no truncation correction

60

0.5

0.75

1

1.25

1.5

1.75

Normalized wave cut position yWC / 0 [-]

Fig. 3.5: Related wave pattern resistance of the even submerged dipole, influence of the wave cut position yWC and the truncation correction.

puted wave pattern exceeds those of the measured wave cuts. This phenomenon is attributed to a slight

over-prediction of the stern waves in the CFD simulation which in turn is due to the absence of viscous

damping. However, more significant is the decrease of the resistance values with increasing transverse

cut position in figure 3.7. This effect is considerably amplified in case of the computed wave pattern.

But, it is not fully explained by the shortening of the utilizable wave signal length as the comparatively

gentle decrease of the experimental curve suggests.

40

0.2

0.15

0.1

SHIPFLOW

0.05

Experiment

0

2

2.5

3

3.5

Normalized signal length xE / LPP [-]

Fig. 3.6: Wave pattern resistance of the Hamburg test case determined from computed (SHIPFLOW)

wave cuts and wave probe measurements (HSVA), influence of the signal length xE .

0.2

0.15

0.1

SHIPFLOW

0.05

Experiment

0

0.25

0.3

0.35

0.4

0.45

0.5

Normalized wave cut position yWC / LPP [-]

0.55

Fig. 3.7: Wave pattern resistance of the Hamburg test case determined from computed (SHIPFLOW)

wave cuts and wave probe measurements (HSVA), influence of the wave cut position yWC .

This effect is attributed to numerical damping (amplitude error) inherent in the CFD simulation. The

effects of numerical damping are thoroughly studied by Raven (1996) and by Janson (1997), the latter

by means of the SHIPFLOW code. The main error sources are the discretization of the continuous free

surface source distribution and the difference scheme used for evaluating the velocities and their derivatives in the free surface boundary condition. Tests suggest that the lateral panel stretching, inevitable

in the majority of cases, is further impairing the numerical damping. Hence, lateral stretching of the

free surface panelization ought to be avoided or at least kept small. Moreover, a sufficient number

of panels per fundamental wavelength ought to be placed uniformly in longitudinal direction. Studies

41

3 Hydrodynamic analysis

suggest the number of panels per fundamental wavelength 0 should be at least 25 in order to capture

most of the shorter wave components. The highest wave frequency that can be resolved by the free

surface panelization is defined by the Nyquist cut-off frequency which is proportional to the width of

the outermost lateral panel strip, if lateral panel stretching is used.

While computed wave pattern, and hence wave cut analysis based on them, prove to be less sensitive to

flow field discretization than hull pressure integration, the wave pattern resistance is affected by cumulative numerical damping effects. Consequently the wave pattern resistance deduced from computed

wave pattern typically is smaller than the wave resistance from pressure integration. Nevertheless,

in terms of the optimization application the error due to numerical damping in the CFD simulation

is of minor concern since only relative differences of hull variants are compared without explicitly

addressing the absolute resistance value.

It should be emphasized that the combination of a nonlinear free surface CFD simulation with a wave

cut analysis approach, which itself is based upon linearized theory assuming a vanishing effect of

nonlinear contributions, is inconsistent from a mathematical point of view. However, this approach

seems justified by its successes in optimization applications and in the application of wave cut analysis to measured wave cuts, as the pertinent literature sources given throughout this chapter suggest.

Moreover, a numerical superposition study conducted by Heimann and Harries (2003) indicates that

the linear wave cut analysis is capable of accurately tracing, on a relative basis, the nonlinear impact

of locally confined hull form variations 22 in the nonlinear wave pattern.

22

42

4 Sensitivity analysis

The proposed hydrodynamic optimization scheme utilizes gradients or sensitivity terms for

the optimizations. The impact of hull variation on the system state quantities is expressed

in terms of gradients. Special flow related optimization variables are introduced. The nonlinear boundary value problem is accessed by a linear perturbation approach allowing a

straightforward and fast computation of the gradients. The process flow of the sensitivity

analysis is outlined.

The present chapter focuses on the evaluation of the gradients, i.e., the sensitivity terms 1 of the objective function and the constraints as needed for the optimization scheme. Nonlinear relations between

the system state quantities and the optimization variables are accessed by a linear perturbation approach. This allows a straightforward and fast computation of the gradients. Moreover, it enables a

consistent tracing of the impact of the variation of a single optimization variable on the system state

quantities like, e.g., the free wave spectra and the wave pattern resistance. In addition, the variation

impact on certain hull form parameters, some of which are constrained within the optimization, is

explained.

The main part of this chapter focuses on the theoretical background of the gradients evaluation. In

section 4.8 practical aspects of the gradients evaluation are discussed and an illustration of the gradients

with respect to local hull variations is given.

The derivation of the gradients is based on discrete relations as predetermined by the discrete view of

the underlying boundary element flow solver as outlined in section 3.1. Hence, the reader is recommended to first catch the basic concept of the underlying hydrodynamic analysis which is outlined in

chapter 3 before continuing with the present chapter.

All quantities defined or derived in this chapter are given in non-dimensional form, normalized by the

ship speed and/or the length between perpendiculars.

Normally, the design variables in ship hull form optimization are directly related to the hull geometry.

For instance, nearly all recent optimization approaches attain control of the hull form variation by

utilizing either hull offsets, explicit polynomial functions, B-Spline basis functions or even higher level

geometric quantities like hull form parameters. Other methods adopt shift functions in order to modify

basic ship curves like sections and/or waterlines. A comprehensive overview of hull form variation

approaches can be found in Nowacki et al. (1995).

1

43

4 Sensitivity analysis

Hull form variation by shifting hull offsets in general gives the highest degree of freedom for form

variation, but increases the computational costs since a vast amount of free variables, i.e., offset coordinates, need to be controlled. In contrast the design variables are effectively condensed and thus

reduced in higher level hull modelling and variation approaches like in form parameter based schemes,

see Birk (1998) and Harries (1998). Nevertheless, the more general the design variables the more

global is their influence on the hull form, thus restraining the ability of locally confined hull variations. In either case, the individual choice of the hull form variation method and the particular design

variables very much depends on the underlying hull geometry representation.

The basic concept of the present approach particularly differs in the sense that the design variables

ought to be independent of the hull geometry representation and its characteristic features. Within

the current context the variation of the hull geometry should be understood as an implicit and derived

property in the optimization since actual control of hull variation is achieved by a non-geometric quantity.

The present choice of the design variables, more precisely called optimization variables, is based on

the following ideas:

The optimization variables ought to be self-explanatory and self-consistent.

They ought to be independent of the hull geometry representation and its characteristic features.

The optimization variables constitute a link between the hull geometry and their hydrodynamic

properties in terms of wave formation and wave pattern resistance.

Their influence should be locally confined.

As outlined in chapter 3 the computation of the steady ship wave system is carried out with the aid of the

nonlinear free surface Rankine panel module XPAN of the CFD system SHIPFLOW. The underlying

boundary value problem is solved by discretization of the boundaries of the flow domain, essentially,

the hull and the wavy free surface. In the first order panel method the boundaries are approximated

by flat quadrilateral panels. Potential functions are introduced by distributing source singularities of

constant density over each panel. 2 The boundary conditions on the hull and the wavy free surface are

enforced for each panel at a control point defined as the panels centroid. The direct solution of the

boundary value problem leads to a linear system of equations for the unknown panel source densities.

See section 3.1 for a closer description of the panel method.

In view of the above considerations the author concluded that the hull panel source densities ideally

meet the demands on the optimization variables. On the one hand the source densities are the principal

parameters in the hydrodynamic analysis which fulfill the governing Laplace equation and lead to the

free surface elevation and, thus, to the free wave spectra and to the wave pattern resistance. On the other

hand they are linked to the hull geometry, even if only in discretized form, since the hull panels are

employed to satisfy 3 the kinematic hull boundary condition. However, the hull panel source density

distribution is independent of the hull geometry representation and its characteristic features.

In this chapter perturbation relations are introduced which substitute the essentially nonlinear boundary

value problem by explicit, linear relations. In fact, provided the variations are small the perturbation

2

Alternatively a higher order method can be selected which applies bi-parabolic panels with a linear source density distribution along each panel.

3 in a discrete sense

44

relations establish an explicit and linear link between the source density distribution and the crossconnected boundaries of the flow domain in terms of the location and orientation of the hull and free

surface panels. The perturbation approach is presented in section 4.2.

From a theoretical point of view all hull panel source densities may serve as free optimization variables.

However, in practical applications, as presented in chapter 6, only specific hull regions are usually subject to an optimization like the bulbous bow, the entrance and/or run while other regions are required to

stay untouched. The latter is accomplished in the present scheme by freezing whole hull panel regions

throughout the optimization or even by activating or deactivating single hull panels. The influence of a

single hull panel source density variation is considered a locally traceable effect.

In the present optimization approach hull variation is controlled by a large number of variables in terms

of hull panel source densities. The optimization scheme is based on gradients. Two problems arise:

At first, an explicit relation between the optimization variables and the hull geometry itself is required.

Furthermore, the evaluation of the gradients by means of direct CFD simulations rapidly generates

high computational costs in view of the large number of optimization variables. In order to overcome

these problems the present optimization method is based on a perturbation approach. The nonlinear

boundary conditions are transferred to explicit equations depending in a linear and straightforward

way on the optimization variables. In other words, the nonlinear boundary value problem is linearized

around the actual state of the system. Thus, the whole cause and effect chain starting with the variation

of the optimization variables, resulting in a hull geometry 4 variation, finally affecting the wave pattern,

the free wave spectra and the wave pattern resistance can be consistently traced.

4.2.1 Perturbation of the discretized boundary value problem

The unknown source densities are related to the boundary conditions on the hull and the free surface

in a nonlinear manner, compare section 3.1. However, the present optimization approach is based on

explicit, linear perturbation relations for the gradients evaluation in order to

access the inverse design problem in a straightforward way, i.e., seeking a disturbance of the hull

boundary (and free surface boundary) which matches a desired, predefined hull source density

distribution,

bypass costly direct numerical solutions of the boundary value problem by means of CFD simulations for a large set of variables.

The perturbation method is considered an approximation procedure permitting a straightforward treatment of the nonlinear boundary value problem. The fundamental idea of the perturbation approach is

as follows: A small parameter is introduced into the problem such that an already known 0th -order

solution corresponds to = 0 and that all higher order perturbations of this basic solution are traceable

by their specific power of . Perturbation approaches are asymptotic power series expansions with the

property that all higher order series terms approach zero faster than all other preceding series terms

as approaches zero. The benefit of this approach is that any nonlinear problem can be decomposed

4

45

4 Sensitivity analysis

into a sequence of linear problems with ascending order by sorting in powers of . The linear subproblems can then be solved separately one after the other. Yet, a restriction to the 1st -order problem

already yields a meaningful approximation of the fully nonlinear problem, presuming the perturbation

magnitude is kept small. The approximation is known to get closer the smaller gets. A detailed

representation of the perturbation theory can be found in Cole (1968) and van Dyke (1975).

Practical applications of perturbation methods frequently deal with the inverse design problem, i.e.,

searching for a design which meets one or more desired, predefined performance criteria. A systematic

perturbation study is carried out for a parent design which then gets adapted in a subsequent step in

order to satisfy the desired, predefined criteria. Other implementations of perturbation methods are

aiming at an acceleration of the optimization process. Expensive direct numerical evaluations of the

cost function are by-passed by simple, explicit perturbation relations. The present work benefits from

both these practical aspects of the perturbation method as will be pointed out shortly.

Performance criteria in inverse design applications are flow related state variables like a desired distribution of the dynamic hull pressure, the flow velocities along the hull or in the wake. Geometrical

quantities are used in terms of offset coordinates or local inclinations with respect to the incident flow

direction. The geometry is varied either by simply shifting offset points or spline vertices, by distortion

approaches or the variation of high-level form parameters. In case of discretized geometries, cells,

patches or panels are either shifted or inclined or both. In inverse design applications the decisive flow

and geometrical quantities are both expanded in terms of perturbation expressions which are applied to

the governing equations of the flow problem. By utilization of series expansions and sorting in powers

of the perturbation parameter the nonlinear boundary value problem is then transferred to mostly linear relations linking flow and geometry. Practical applications of the perturbation method to 2-D and

axis-symmetric inviscid flow problems are reported, e.g., in Schmidt (1976), Bristow (1974, 1980) and

Favre et al. (1987). Kraus (1989) extended the application range to 3-D problems. Huan and Huang

(1998) formulated both a perturbation and an adjoint equations method which they applied to the inviscid nonlinear free surface flow problem. They successfully validated both techniques by a comparison

with direct CFD solutions and experimental data for the flow about the Salvesen hydrofoil below a

nonlinear free surface. An application to an inverse hydrofoil design problem is also considered in their

work. Schulze (1996) reported a comprehensive investigation of the practicability of locally confined

perturbations for both inviscid and viscous flows about 2-D profiles.

The perturbation relations derived in this chapter are based on discretized boundary equations as the

discrete view of the underlying boundary element flow solver implicates. This does not imply a loss

of generality since the discretized relations do approach their continuous counterparts with increasing

mesh refinement. The governing equation for the panel source densities is given in discretized form by

(3.11)

NF S

j Ai j = Bi

j=1

i = 1, , NFS .

NFS accounts for the total number of panels which includes the hull panels NH and the free surface

panels N f s , i.e., NFS = NH + N f s . Flat quadrilateral panels are used with a constant source density

distribution across the panel.

The perturbation relations are derived by expanding both flow and geometrical quantities in terms of

perturbation expressions which are then applied to the boundary value problem. The panel source

densities represent the unknowns in the flow solution and at the same time are selected as optimization

46

variables. Expanding the normalized (by U ) panel source densities per unit area in a perturbation

expression and truncating after the linear term yields

(0)

(1)

n = n + n + O(2 ) ,

(1)

n = n + O(2 )

n = 1, , N NH ,

with (0) indicating the 0th -order solution, i.e., the basic solution. Superscript

solution, i.e., the perturbation expression.

(4.1)

(1)

The corresponding perturbation of the hull geometry is realized by perturbating the ship hull surface in

its normal direction. The hull surface is approximated by flat quadrilateral panels. Hence, the panels

are relocated in their normal directions in- or outwards of the hull panel mesh. The panel relocation

technique is described in section 4.3. Expressing the normalized (by LPP ) hull panel normal relocations

in terms of a perturbation notation and truncating after the linear term yields

~Pk = ~P(0) + ~n(1) + O(2 ) ,

k

k

(1)

~Pk ~nk = ~nk + O(2 )

k = 1, , NK NH .

(4.2)

~Pk may be any point on the flat quadrilateral panel k. ~Pk will be later associated with the panel control

point. NK is the number of active hull panels applied in the panel normal relocation scheme. The

(1)

(1)

magnitude of the panel shift in panel normal direction ~nk is composed of the shift amplitude nk and

the unit vector normal nk

(1)

(1)

~nk = nk nk

~nk = nk nk .

The hull part (i = 1, , NH ) of the influence coefficients Ai j and the inhomogeneous term Bi in (3.11)

are pure geometrical quantities solely depending on the geometry, the relative position and the orientation of the panels. The free surface part (i = NH + 1, , NFS ) in addition includes the flow velocities,

generally, the free surface velocities and their derivatives from a previous step in a nonlinear free surface iteration. However, these velocities and derivatives are treated as constant, known terms in the current evaluation of the influence coefficients. Hence, the influence coefficients and the inhomogeneous

term are actually affected only by the geometry perturbations in terms of the panel normal relocations

(1)

nk . Since this dependency is implicit, both the influence coefficients and the inhomogeneous term

need to be expanded in Taylor series which may be truncated after the linear terms

(0)

Ai j = Ai j +

NK

(0)

(1)

nk

k=1

(0)

Bi = Bi +

NK

(1)

nk

k=1

Ai j

nk

+ O(2 ) ,

(0)

Bi

+ O(2 )

nk

i, j = 1, , NFS ,

(4.3)

in order to obtain explicit and linear relations in the perturbation parameter . Kraus (1989) verified the

validity of the Taylor expansions by means of systematic perturbation studies with a 2-D profile and a

47

4 Sensitivity analysis

rotationally symmetric Rankine ovoid, both deeply submerged. He computed the change of the influence coefficients Ai j due to a y-perturbation with varying magnitude of selected panel corner points.

For small y-perturbations in the range 10% of the initial y-offset Kraus found good agreement of the

Taylor expansions (4.3) with results from exact computations applying the Hess & Smith relations for

the Ai j . A similar study conducted by the author for the 3-D case confirmed the results of Kraus. As a

test case the flow around the standard Wigley hull at Fn = 0.3 including the free surface was studied.

Selected hull panels were systematically shifted in panel normal direction in- and outwards of the hull

surface. The resulting change in the influence coefficients Ai j , Xi j , Yi j , Zi j and the inhomogeneous term

Bi were computed by means of three different methods: SHIPFLOW/XPAN, a test program prepared

by the author and the linearized relations (4.3). A comparison of the the results approve the applicability of the linearizations (4.3) for moderate hull panel normal relocations in the range of 10% of a

typical hull panel size. 5 Larger perturbation values are feasible but should be verified beforehand for

the particular application case.

The linearization of the boundary relations can be conducted now by introducing the perturbation

approaches (4.1) and (4.2) together with the Taylor expansions (4.3) into the boundary value problem

(3.11)

NFS

j=1

(0) (0)

(1) (0)

(0)

j Ai j + j Ai j + j

(0)

NK

k=1

(1)

nk

Ai j

nk

(1)

+ 2 j

NK

k=1

(1)

nk

(0) )

Ai j

nk

(0)

= Bi +

NK

(1)

nk

k=1

i = 1, , NFS .

(0)

Bi

nk

(4.4)

Neglecting higher order contributions of O(2 ) and sorting in orders of yields for the 0th -order

NF S

(0) (0)

Ai j

j=1

(0)

i = 1, , NFS ,

= Bi

(4.5)

which constitutes the basic equation system. For the 1st -order terms one gets

NF S

j=1

NK

(0) )

(1) (0)

(0)

(1) Ai j

j Ai j + j nk

nk

k=1

NK

(1)

nk

k=1

(0)

Bi

nk

i = 1, , NFS ,

(4.6)

which represents the contribution due to the perturbation. After rearrangement, (4.6) can be written in

compact matrix notation

A(0) ~(1) = T(0) ~n(1)

,

[NFS NFS ] [NFS ] = [NFS NK ] [NK ]

(4.7)

T

5

6

(0)

(0)

Tik

(0)

(0)

NFS

B

(0) Ai j

= i + j

nk

nk

j=1

i = 1, , NFS

.

k = 1, , NK

(4.8)

e.g., typical hull panel size in terms of the average square root of the hull panel areas, as will be discussed in section 4.2.2

(0)

A hull panel normal relocation does not affect the inhomogeneous terms, i.e., in this case Bi /nk = 0 for all k.

48

Equation (4.7) represents the fundamental linear perturbation relation relating the source density variations ~(1) and the panel normal relocations ~n(1) . The coefficient matrix A(0) in (4.7) is already known

from a base flow solution. The evaluation of the transfer matrix T(0) is the expensive part since the

partial derivatives of the coefficients with respect to the hull panel normal relocations have to be calculated. This can be done either analytically or numerically as will be addressed in section 4.8. As

expression (4.8) reveals the transfer matrix comprises the variation effect on both the hull and the free

surface portion. However, during the evaluation of the transfer matrix the free surface panelization

itself is temporarily frozen.

The hull panel source densities are selected as the free variables in the optimization process. Nevertheless, the optimization aims at an improvement of the wave-making characteristics which finally is

the result of a favourable adaptation of the hull form. Relation (4.7) provides the link between the

source density variations and the corresponding hull panel adaptations. However, the inverse relation

is required as will be derived below.

The inverse solution

The inverse solution yields a variation of the hull form which matches the predefined variation of

the hull source densities. The inverse relation is obtained by inversion of the transfer matrix T(0) in

(4.7) and solving for the hull panel normal relocations ~n(1) . For the inverse solution all rows and

columns of the vector and matrix elements in (4.7) which are associated with the free surface portion

(NK,H < i, j NFS ) can be omitted, temporarily, since only the hull source densities ~(1) are considered

as variables in the optimization, thus

~n(1) = T(0)1 A(0) ~(1)

.

[NK ] = [NK NK ] [NK NK ] [NK ]

(4.9)

In all inverse cases utilized in the optimization the transfer matrix is non-singular, diagonally dominant

and, as addressed above, temporarily square. Hence, the number of hull panel relocations equals the

number of equations which in turn permits a unique solution. 7 In the optimization context the equation

system (4.9) is to be solved N times for N optimization variables, i.e., N hull source density variations

~(1) . Since each time only the hull source densities on the right hand side vary equation (4.9) is solved

numerically in a straightforward manner by LU-decomposition and back substitution, for details of the

technique be referred to Press et al. (1992).

For the sake of completeness the over-determined case may be discussed briefly. In the over-determined

cases the number of unknowns, i.e., the hull panels included in the panel normal relocation scheme, is

less than the number of equations. For instance, one allows only a subset of all hull panels to adjust

by moving normally in- and outwards of the mesh (NK < NH ), whereas, at the same time a predefined,

desired variation of the hull source densities is to be met at all hull panels. This may be the case, e.g.,

when all panels which are part of the flat of side (FOS) and/or flat of bottom (FOB) are frozen (not

allowed to move) in order to maintain those specific hull characteristics in the optimization. However,

at the same time the same panels are forced to actively adjust a specific -variation which, for instance,

could be a zero source density variation.

7

In particular, as shown by Cramers rule, there is a unique solution if T(0) has a matrix inverse.

49

4 Sensitivity analysis

A solution of the over-determined system is possible only in a mean sense, i.e., by a minimization of

the residuals. For this purpose the perturbation relation (4.9) is reformulated in terms of a functional

expression which is to be minimized in a least-square sense

!2

NH

F=

i=1

NFS

(1) (0)

NK

(1)

(0)

j Ai j + nk Tik

j=1

= min .

(4.10)

k=1

F

(1)

nl

=0

l = 1, , NK < NH ,

which yield

NH

(0)

Til

i=1

NFS

(1) (0)

Ai j +

j

j=1

NK

(1)

nk

k=1

(0)

Tik

=0

l = 1, , NK < NH .

(4.11)

Finally, the inverse relation for the over-determined case may be written in compact matrix notation

e (0)1 T(0)T A(0) ~(1)

~n(1) = T

,

[NK ] = [NK NK ] [NK NH ] [NH NH ] [NH ]

(4.12)

with

NH

T

kk

ik

ik

i=1

k = 1, , NK < NH .

Equation system (4.12) represents the normal equations of the linear least-square problem (4.10). For

NK = NH the linear system for the over-determined case (4.12) reduces to the special case of (4.9).

The direct solution

Contrary to the inverse problem in the direct approach one seeks to determine the source density variations for a preceding variation of the hull form. Again the perturbation relation (4.7) is applied but this

time it is solved for the source density variations which result from a preceding variation of the hull

panelization, i.e., a normal relocation of the hull panels. Thus, in the direct case the source densities no

longer serve as optimization variables as in the inverse approach. Rather, this time the variation of the

source densities, particularly, those at the free surface result from a preceding hull perturbation. Since

it is searched for the variation of all panel source densities (hull and free surface) this time the free

surface portion in (4.7) must be taken into account, thus giving

~ (1) = A(0)1 T(0) ~n(1)

.

[NFS ] = [NFS NFS ] [NFS NK ] [NK ]

(4.13)

Equation system (4.13) is to be solved repeatedly, generally N times according to the number of

optimization variables, with varying right hand sides. This calls for a numerical treatment by LUdecomposition and back substitution.

50

In practical applications the free surface domain needs to extend far enough up- and downstream and

sidewise in order to allow for a proper wave cut analysis. At the same time the panelization has to

be dense enough to capture much of the wave information. This in turn results in large systems of

equations with a huge coefficient matrix 8 which is costly both in computation time and in memory

consumption. This gives rise to reducing the portion of the free surface domain considered in the

solution of the perturbation relation (4.13). 9 A validation study conducted by the author has shown

that a truncation of the free surface panelization sidewise is practicable to a certain extent as long as

the perturbations are kept small.

Alternatively, the source density variations may be calculated only in a mean sense. At the same time

the influence of the whole free surface domain is kept. Assuming that the number of source panels just

including a subset of the free surface domain, e.g., truncated sidewise, is NFS

f with (NFS

f < NFS ), then,

analogous to (4.10), the direct problem (4.13) can be expressed in terms of a functional expression

which is minimized in a least-square sense. This yields

with

~ (1) = A

,

NFS

f = NFS

f NFS

f NFS

f NFS [NFS NK ] [NK ]

(4.14)

NFS

e (0) A

e(0) = A(0) A(0) = A(0) T A(0)

A

jj

ij

ij

i=1

j = 1, , NFS

f < NFS .

Equation system (4.14) represents the normal equations of the linear least-square problem. For NFS

f =

NFS the linear system (4.14) becomes identical to (4.13).

4.2.2 Perturbation magnitudes

As outlined in the preceding section a perturbation scheme is applied to linearize the boundary value

problem. The basic assumption is that the magnitude of the perturbation is small. Since the hull

panel source densities and the hull panel normal relocations are selected as perturbation quantities

their magnitudes both need to be bounded in a perturbation sense. In order to keep the scheme robust

and straightforward the same perturbation magnitude is adopted for all hull panels.

The perturbation magnitude for the normalized hull panel source densities, in (4.1), is chosen as a

fixed small value in order to comply with the perturbation assumptions. The perturbation magnitude

may be simply related to a characteristic velocity U (e.g. the incident flow velocity)

0 < 1 ,

n = = U

(4.15)

to a typical source density value, for instance the average absolute value of all hull panel source

densities

NH

(0)

|i |

n = = i=1

NH

8

9

0 < 1 ,

(4.16)

Note: The underlying CFD simulation by SHIPFLOW still includes the whole free surface domain in the solution.

51

4 Sensitivity analysis

or to the average of the weighted absolute value of all hull panel source densities. If the hull panel

source densities are weighted by their panel areas Si the perturbation magnitude may be defined as

NH

(0)

Si |i |

n = = i=1 N

H

Si

0 < 1 .

(4.17)

i=1

Since the numerator in (4.17) can be viewed as the perturbation of the volume flow through hull panel

i, with

(0)

Vi = Si |i |

0 < 1 ,

definition (4.17) can be equivalently interpreted as an increment of the average volume flow through

all hull panels. The advantage of the definition (4.17) for the perturbation magnitude is that it accounts

for the specific hull panelization and the flow case in question which allows a more intuitive and

appropriate choice of .

Similarly, the perturbation magnitude for the hull panel normal relocations, nk in (4.2), can be selected. For instance, the perturbation magnitude may be simply related to a characteristic length L (like

the length between perpendiculars or the waterline length)

0 < ~n 1 ,

nk = n = ~n L

(4.18)

a typical hull panel size, e.g., the average hull panel diagonal or the average square roots of the hull

panel areas Si

NH

nk = n = ~n

i=1

NH

Si

0 < ~n 1 .

(4.19)

Again the advantage of the more elaborate definition is that it accounts for the specific hull panelization,

particularly, in terms of the panel mesh resolution 10 which allows a more intuitive and appropriate

choice of n.

Many more choices of the perturbation magnitudes are conceivable. Which of the alternatives are

eventually utilized to define the perturbation magnitudes is a practical decision by the user. For the

perturbation scheme it is of minor importance as along as the perturbation magnitudes are kept small

in a perturbation sense.

In the present optimization scheme the more general definitions (4.17) and (4.19) for the perturbation

magnitudes are implemented. In either definition the perturbation magnitudes are already related to the

specific hull panelization and the flow case in question. This allows a more general determination of

the small perturbation parameters and ~n so as to cover a broad range of hull shapes, flow conditions

and hull panelizations at the same time. From a practical application point of view the perturbation

magnitudes and n should correspond. They must still be selected large enough in order to produce

detectable variations 11 of the system state quantities and at the same time they need to be small enough

10

11

in terms of variations in a significant digit which is required to be at least some order of magnitude higher than the adopted

mantissa length

52

z

0.5

2

0

0

y

1

1

x

2

3

Fig. 4.1: Artificial bicubic surface patch and panelization by 3 3 flat quadrilateral panels (right).

in order to comply with the perturbation approach. A systematic study conducted by the author with

different ship applications and with varying panel mesh resolutions determined a value of = ~n = 0.1

for the perturbation parameters as most suitable when defining the perturbation magnitudes according

to (4.17) and (4.19).

Whether the wave-making characteristics of a ship hull are considered good or bad is basically determined by the hull shape. Hence, the final outcome of an optimization is a new hull shape or a

favourable hull shape variation.

The present optimization scheme is based on a discrete view of the hull surface, i.e., a hull panelization

by flat quadrilateral Hess & Smith panels with a constant source density across each panel. The panel

patch-work rendering the hull surface is G0 -discontinuous in this case. That means, small gaps will

arise between adjacent panel edges. The method of formation of flat quadrilateral Hess & Smith panels

from a regularly ordered input or offset point mesh is briefly discussed in section 3.1.2. A detailed

description is given in the original publication by Hess and Smith (1962).

The hull modification originates from a variation of the hull source densities. Different choices for

the modification of the hull panel mesh are feasible. Two alternatives should be addressed. A simple

but effective approach is to relocate the underlying input point mesh which constitutes the hull panels.

From a topological point of view input points are equivalent to offset points which often are arranged

along hull sections and profile curves and are located directly on the hull surface. Different schemes

exist. Still popular and simple is to shift hull points in lateral (mostly in y-direction) direction in order

to attain a hull form variation, e.g. see Kraus (1989). In this case the panel mesh adjusts indirectly. An

alternative technique is to shift single hull panels directly in- and outwards of the mesh. Unless special

precautions are taken considerable gaps may develop between adjacent hull panels which requires a

certain control mechanism.

53

4 Sensitivity analysis

Input Points Flat H&S Panels Transformations

Fig. 4.2: Relocation of the corner points of the center panel in panel normal direction (left), effect of

the equivalent relocation of the four central input points (right).

The advantages and disadvantages of either approach may be summarized briefly, starting with the

input/offset point relocation scheme:

Easy application points are simply shifted by their offset coordinates,

no special treatment of the panels is necessary, they are self-adjusting, accordingly,

the initial topology of the input point mesh is kept at all time,

points are free to move since no dependency on the neighbourhood needs to be maintained, 12

constraints for the input points like, e.g., a symmetry condition at the midships plane are relatively easy to enforce, but

on the other hand the forced relocation of the hull panels is a dependent transformation only

indirectly controllable by the transformation of the underlying input points,

panel shape and area considerably change and hence are not invariant against the transformation,

at the same time panel rotations are enforced,

special effort is required to maintain geometric boundary conditions and,

since in practice the number of optimization variables does not match the number of the underlying input points an over-determined system is to be solved which leads to an averaging in a

least-square sense with somehow ambiguous results.

Advantages and disadvantages of the hull panel normal relocation scheme are:

It can be considered the discrete version of a continuous hull surface variation scheme,

the variation of the hull panels can be controlled directly,

12

unless the initial topology of the input point mesh is violated, e.g., by point intersections

54

Input Points Flat H&S Panels Transformations

Fig. 4.3: Homogeneous panel relocation (left), effect of the joint relocation (right).

no panel rotations are enforced, 13

panel shape and area are invariant against the panel transformation, 13

since the (variable) hull source densities and the hull panel relocations can be mapped one-toone, typically a unique solution of a determined linear system is expected, but

on the other hand a more expensive implementation is required, because panel-connectivities as

well as relocations of both the panel corner and the input points have to be traced,

special effort is necessary to maintain geometric boundary conditions,

the topology of the input point mesh is violated since the inverse formation of a coherent input

point mesh from independently displaced panels is impossible (contradicting demands),

relocation of adjacent panels in different directions or with different magnitudes cause considerable gaps in the panel mesh

which requires a special treatment of neighbouring panels and a reconstruction scheme in order

to restore a unique hull surface from the emerging incoherent panel cluster.

With the above considerations in mind it is meaningful to combine both approaches so as to combine the

advantages of either approach and at the same time to control or even to neutralize the disadvantages.

Consequently, the following hull panel relocation scheme has been developed which, in some sense,

performs both approaches successively.

13

55

4 Sensitivity analysis

Input Points Flat H&S Panels Transformations

Fig. 4.4: Stepped panel relocation (left), effect of the joint relocation (right).

4.3.1 Joint relocation of hull panels

The idea behind the joint relocation scheme is simple: first, let the hull panels shift separately in their

normal directions according to the optimization requirements and, in a subsequent step, redistribute the

panel normal relocations back to the associated input point mesh by a vector averaging. In the present

work this scheme is supported by an inequality constraint in the optimization which keeps track of

adjacent panel relocations in order to ensure smooth transitions in the distorted panel mesh, see section

5.2.2.

The relocation scheme is illustrated by means of an artificial bicubic surface patch (figures 4.1 to

4.4), a generic ferry-type stern hull surface patch possessing y-symmetry (figures 4.5 to 4.8) and by a

SHIPFLOW panelization example featuring the transition area from the bulb to the bow (figures 4.9 to

4.13). The surface patches in the examples are panelized by flat quadrilateral panels according to the

Hess & Smith method. Initial small gaps, i.e., G0 -discontinuities, can be observed between adjacent

panel edges as is immanent in the method.

Figure 4.2 illustrates the effect of moving a single panel outwards of the mesh in its normal direction,

see picture to the left. The panel normal relocation is achieved by simply shifting the four panel corner

points homogeneously in panel normal direction. Apparently, large gaps arise between the panel edges

of the shifted center panel and the yet untouched adjacent panels. 14 Alternatively to the relocation of

the panel corner points the associated four input points may be shifted in the panel normal direction by

the same amount. Even though, in both cases, the final shifted position of the center panel is identical,

in the latter case the adjacent panels are forced to follow the movement of the input points and, thus,

panel gaps have been reduced considerably, see picture to the right. Unfortunately, this property is

purchased by the penalty of an unwanted panel rotation of the adjacent panels.

14

Note: For illustration purposes the panel relocations in the examples are exaggerated, compared to the panel perturbations

allowed in the hull optimization.

56

z 1

0.5

0

0

1

-1

2

x

3

4

-2

Fig. 4.5: Generic ferry-type stern hull surface patch (upside down), y-symmetry, and panelization by

4 4 flat quadrilateral panels (right).

So far only a single panel has been touched at a time. In view of the perturbation scheme this is a

somehow artificial case never arising in practical applications. In the variation scheme all hull panels

or at least a panel cluster will be adapted at the same time. Figure 4.3 shows a homogeneous panel

relocation of the whole patch and in figure 4.4 a stepped panel relocation is presented. Apparently,

as can be observed in the pictures to the left, shifting each panel separately ignoring the neighbourhood causes considerable gaps since the panels start detaching or intersecting which is amplified with

increasing panel displacement. Obviously, a technique is required to resize the gaps to normal size

as implied by the Hess & Smith method and to maintain the initial topology of the input point mesh.

However, at the same time the characteristic of the panel relocation must be kept as best as possible.

The idea is to redistribute the panel normal relocations back to the associated input point mesh by a

vector averaging in a subsequent step.

Let IPm denote the vector representation of the m-th input point and Nm the number of adjacent panels 15

with 1 Nm 8. In order to reconstruct a coherent input point mesh with a well defined topology the

(1)

panel normal relocations ~nk have to be redistributed back to the associated input points. This can be

achieved by vector averaging

(0)

1

(0)

= IPm +

Nm

Nm

(1)

~nk(i

im =1

m ,m)

m = 1, , NM

.

1 Nm 8

(4.20)

The final adaptation of the panels then is deduced from the thus reconstructed input point mesh, again

by means of the Hess & Smith method. Examples for the joint relocation scheme are presented in the

pictures on the right hand side in figures 4.3 and 4.4 and in the following figures.

15

In case of quadrilateral panels a maximum of four panels are linked to a single input point. Since in case of symmetrical

ship hulls the mirror panels across the symmetry plane are also considered in the relocation scheme, topologically twice

the number of panels may be linked to a single input point at maximum.

57

4 Sensitivity analysis

Input Points Flat H&S Panels Transformations

Fig. 4.6: Homogeneous panel normal relocation (left), effect of the joint relocation (right).

The joint relocation scheme is applied to construct a smooth input point mesh with a well defined

topology. The point mesh can be further processed by surface interpolation or approximation and fairing techniques, see e.g. Nowacki et al. (1998), Nowacki and Kaklis (1998) and Westgaard (2000). In

the present work the computer aided ship hull design (CASHD) tool MultiSurf from AeroHydro Inc.,

Letcher et al. (1999), is utilized as a geometrical postprocessor to the optimization scheme. MultiSurf

is applied in a script technique in order to generate higher-level B-Spline hull surface representations

from the input point mesh data in an automated procedure, compare section 5.4. Moreover, this simplifies the export of the hull geometries to other systems and/or applications via portable and freely

convertible formats (like IGES).

Properties of the joint relocation scheme

Taking into account the exaggeration of the panel relocations in the examples and considering that

sign and magnitude of the panel relocations in ship application will be somehow smoothly distributed

across the hull the following properties of the joint relocation scheme can be concluded:

The technique is considered a discrete version of a continuous hull surface variation scheme,

the initial panel orientations are kept by the scheme,

panel shape and area are changed only slightly,

a coherent input point mesh with a well defined topology, the initial mesh topology, is kept,

gaps between adjacent panel edges remain small, as is typical of the Hess & Smith method,

the variable hull source densities and the panel normal relocations can be mapped one-to-one

which leads to unique solution of the perturbation relations, but,

undesired yet unavoidable panel rotations are introduced due to the averaging procedure and

58

Input Points Flat H&S Panels Transformations

Fig. 4.7: Upstream stepped panel normal relocation (left), effect of the joint relocation (right).

compared to the simple relocation of input points the joint relocation scheme is more expensive

in implementation.

In the perturbation scheme the variation of the hull form only is a derived quantity since actual control

of the hull variation is attained by the variation of the hull source densities. Nevertheless, like in all

engineering tasks topological and geometrical constraints have to be imposed in order to achieve valid

and meaningful hull form variations.

In the joint panel relocation scheme the hull panel mesh is gradually moved in surface normal direction

which can be considered a topology-preserving form variation. 16 Topological bounds are imposed to

avoid prohibited panel interactions, like panels approaching each other to close or even crossing each

other which may cause singular value problems. Moreover, topological bounds are required to avoid

panel relocations violating the initial panel alignment or violating symmetry conditions, for instance

panels crossing the midship plane. Another focus is the preservation of topological hull features like

the flat of side and bottom or the preservation of other self contained hull regions.

In this context geometrical bounds are imposed which introduce a permissible range for each panel

relocation. Typically, in ship application the hull surface is forced to lie within the bounding box

defined by the maximum length, the maximum (half-)breadth and the maximum draft.

The relocation bounds are applied to restrict the hull panel normal relocations at the first stage of the

joint relocation scheme. Whereas, the subsequent reconstruction of a coherent input point mesh by

vector averaging does not demand any further constraints handling. The reason is, that the averaging

procedure forces the hull panels to stay within the envelope of the preceding panel normal relocation.

16

presuming the magnitudes of the panel relocations are, in a perturbation sense, small and that the curvature distribution

of convex-concave hull areas, like the transition from bulbous-bow to the bow, is moderate

59

4 Sensitivity analysis

Input Points Flat H&S Panels Transformations

Fig. 4.8: Downstream stepped panel normal relocation (left), effect of the joint relocation (right).

Via the perturbation approach, see equation (4.9), both the hull panel source density variations and the

derived hull panel normal relocations are linearly linked to each other. Hence, the bounds imposed

on the hull panel relocations can be directly mapped into the range of the hull panel source density

variations.

All subsequent definitions and derivations in this section are related to the SHIPFLOW coordinate

system and to the ships initial floating position at rest. 17 Recalling the SHIPFLOW coordinate system,

a right-handed Cartesian system is defined with the origin located at the fore perpendicular in the

undisturbed free surface level, x pointing downstream and z pointing up. Since, in the present work it

is focused on laterally symmetrical ship hulls it always is referred to the positive, i.e., to the starboard

side. Furthermore, all definitions and derivations are related to the panel control points CT . The reason

is to keep the scheme unique but comprehensive and to comply with the boundary conditions of the

panel method which, likewise, are satisfied at the panel control points.

Topological bounds

Topological bounds are imposed to avoid prohibited panel interactions, to avoid panel relocations violating the initial panel alignment or violating the symmetry condition at midships plane and to preserve

topological hull features like the flat of side and flat of bottom.

Midship plane

The present work is focused on laterally symmetrical flow cases. Since only one half of the hull

is considered 18 in the flow simulation special precautions are required to avoid panels crossing or

17

Since the perturbation is applied to the dynamic ship condition considering the wetted hull panelization under the wavy

free surface and the effective floating position of the advancing ship including trim and sinkage, the hull virtually is

transformed back to its initial floating position at rest.

18 the starboard side

60

Input Points Flat H&S Panels Transformations

2

1

-1

0

1

-1

0

y

-1

-2

Fig. 4.10: Cut-out of the panelization showing the intersection region at the fore perpendicular FP and

the design waterline DW L (upper edge of the transition area from the bulb to the bow).

detaching the symmetry plane and panels touching the plane at obtuse angles. 19 Otherwise, in case of

laterally asymmetric flows where the hull has to be taken on the whole the panel relocation scheme can

be applied, in general, without special precautions of this kind.

19

Panels which are in close vicinity and even parallel to the symmetry plane are prohibited in potential theory methods since

the mutually induced velocities at the symmetrical counter-panel usually causes singular value problems.

61

4 Sensitivity analysis

Fig. 4.11: Panel overlapping at design waterline as a consequence of a homogenous panel normal relocation (left), effect of the joint relocation (right) considering symmetry at the midships

plane.

Figure 4.5 shows a generic ferry-type stern hull patch (upside down) with y-symmetry at midships

plane. To the left the corresponding panelization by 4 4 flat quadrilateral panels is given. If the

panels next to the symmetry plane start to move in their normal directions they inevitably detach or

cross the symmetry plane as illustrated by the pictures on the left hand side in figures 4.6 to 4.8. A

simple solution to prevent the panels next to the symmetry plane either from detaching or crossing the

plane during panel relocations is to simply freeze the component of their panel normal which points

perpendicular to the symmetry plane. 20 As a result the corresponding panels are forced to stay in a

plane parallel to the midship plane, thus, ensuring a closed input point mesh. Even though this approach

is easy to implement, yet, it restricts a free panel relocation.

An alternative approach, as followed in the present work, takes into account not only the panels next

to the symmetry plane but also their mirror images. In terms of the relocation scheme that means

to simply include the mirror panels across the symmetry plane in the vector averaging of the joint

relocation scheme. In this way the input point mesh is forced to remain at the symmetry plane while

the panels next to the symmetry plane are, at the same time, free to move. This particular behaviour of

20

62

Fig. 4.12: Stem panels crossing the midships plane as a consequence of an inward normal relocation

of the upper bow panels (left), effect of the joint relocation (right) considering symmetry at

the midships plane.

the joint relocation scheme is illustrated by the pictures to the right in figures 4.6 to 4.8.

The above outlined procedure is a straightforward way to maintain a coherent input point mesh. However, problems arise if arbitrary hull panels intend to cross the midship plane. This case is illustrated by

means of a typical hull panelization of a ferry-hull, as presented in figure 4.9 and 4.10. Let us focus on

the intersection region of the fore perpendicular FP and the design waterline DW L (upper edge of the

transition area bulbous-bow bow). Figure 4.12 shows the special case of a hull panel relocation with

the stem panels crossing the symmetry plane (see the enlarged views at the bottom of the figure). The

application of the joint relocation scheme, alone, is not sufficient in this case in order to reconstruct a

coherent input point mesh. Apparently, another relocation constraint has to be imposed a minimum

distance between the hull panels and the midships plane needs to be preserved, too.

Specific hull features

In hull optimization distinct hull features are often considered as fixed due to requirements such as

water depth restrictions, harbour or canal restrictions, cargo handling and storage, accommodation of

63

4 Sensitivity analysis

Fig. 4.13: Lengthwise stepped panel normal relocation (left), effect of the joint relocation (right) considering symmetry at the midships plane.

the propulsion system, specific stern shapes with respect to propulsive efficiency etc..

Specific hull features or self contained hull regions may be frozen during the hull optimization process.

If this gets translated into the notation of the panel relocation scheme this is equivalent to suppressing

the panel normal relocations of specific hull panels. Mathematically this is achieved by employing

equality constraints

(1)

~nk p = 0

k p = 1, , Nk p

,

Nk p < NK NH

(4.21)

with the index k p indicating frozen or passive hull panels. In terms of the present implementation

those panels do not appear in the equation systems right from the beginning. Instead of equalities,

alternatively, inequality constraints

(1)

nk p max k~nk p k 0

(4.22)

can be employed to allow for a little leeway, i.e., small changes of the passive panels around their

initial positions. However, the absolute maximal leeway of the passive panels, defined as nk p max , is

small compared to the maximum allowed relocation of the active hull panels, nka max .

64

e

Cas

e

Cas

e

Cas

CT j( 0 )

y

(0)

(0)

k

CT

Initial panel alignment

CT k j

CT j(1)

(1)

CT k j

Panel relocation Case 1

CTk(1)

Fig. 4.14: Generic 2-D panel relocation example examining two panels arranged in a blunt angle (e.g.

the transition from FOB to the bilge radius) permitted and prohibited (crossed out) relocation cases are compared.

The benefit of allowing all hull panels to adjust freely, even though the passive panels, however, only

on a small scale, is that the overall relocation of the panel mesh promises to be more smooth since no

abrupt breaks in the relocation pattern are expected.

Panel alignment

In the present work both the CFD simulation and the vector averaging of the joint relocation scheme

depend on a regularly aligned panel mesh. Therefore, it is mandatory for the panel relocation scheme

to preserve the initial panel alignment or topology. From an application point of view this means

avoiding all events where panels approach each other too close or even crossing each other during

panel relocation.

In SHIPFLOW the hull panels are basically aligned length- and sectionwise forming a regular mesh as

required by the CFD system. A typical panelization is shown in figure 4.9. Problems with panel overlapping as a consequence of panel relocations are prevailing in case of hull regions with pronounced

convex-concave shape characteristics, for instance at the intersection region of fore perpendicular FP

and design waterline DW L, see figure 4.10. Figure 4.11 demonstrate the possible panel overlapping at

the design waterline following an outwardly directed panel normal relocation. Although, the degree

65

4 Sensitivity analysis

of panel overlapping in figure 4.11 is still within permissible limits a more pronounced normal relocation would cause serious problems within the mesh reconstruction scheme and, hence, needs to be

prevented beforehand.

Undesired panel collisions can be avoided by defining appropriate bounds for the panel normal relocations. Figure 4.14 outlines permitted and prohibited relocation cases by means of a simple 2-D panel

relocation example. For the sake of clarity the problem is reduced to two dimensions, but, all conclusions drawn are valid also in three dimensions. In the initial panel alignment (filled grey) two connected

panels are arranged in a blunt angle. Both panels get relocated independently in their normal directions.

Relocation case 1 and case 2 are permitted since the initial panel alignment is maintained, although,

both panels already start to detach or to cross each other, respectively. The relocations in case 3 and

case 4 (dashed) are prohibited and have to be prevented. Although the initial panel alignment is still

preserved in relocation case 3 the panel control points already approached each other below a certain

limit. In case 4 the panel alignment is reversed and, hence, the initial mesh topology is violated.

The degree of panel alignment during panel relocations can be measured by the change of the distance

vector,

(1) (1) (1)

CT k j = CT j CT k ,

(0)

in relation to the initial 21 distance vector, CT k j . It represents the vector pointing from the control

point of panel k to the control point of panel j, see figure 4.14.

On closer examination of figure 4.14 it becomes apparent that not the modified 22 (due to perturbation)

distance vector itself but rather that part of the vector in direction of the corresponding initial distance

vector may serve as a suitable measure of the degree of panel alignment. In other words, the projection

of the modified vector onto the unit initial distance vector,

c (0)

CT

kj =

(0)

CT k j

,

(0)

kCT k j k

(1) c (0)

CT k j CT

kj ,

(1) c (0)

(0)

CT k j CT

k j PA kCT k j k 0 ,

(4.23)

with PA being a slack parameter controlling the panel alignment. PA is a user predefined parameter

(0)

to control the panel approaching in terms of the initial control point distance kCT k j k. Equation (4.23)

21

22

indicated by (0)

indicated by (1)

66

(1)

may be recasted in terms of the panel normal relocations ~nk, j in the form

0 < PA < 1

k = 1, , NK

f

j = 1, , NH , , NFS

(0)

(0)

(1)

(1)

c

(~n j ~nk ) CT k j + (1 PA) kCT k j k 0

.

NK NH

nk n j

k 6= j

kj 6= { j k}T

(4.24)

The inequality constraint (4.24) is utilized to bound the hull panel relocations in order to preserve a

coherent panel alignment, avoiding serious panel collisions and panel crossings. The parameter PA can

be estimated by analysing the most hazardous relocation case of two adjacent panels directed against

each other or are arranged in some sharp angle. Tests with ship applications conducted by the author

identified a value of PA 0.5 or slightly higher to be appropriate.

Monitoring of the panel alignment may concentrate particularly on the direct neighbourhood of a hull

panel. But, to be on the safe side and to keep the scheme comprehensive the inequality constraint (4.24)

is evaluated for all hull panels against all other hull panels, no matter how far the panels are located

from each other. Furthermore, the worst case is assumed in monitoring of two active hull panels. That

is, both panels approaching each other with the same magnitude during panel normal relocation. In

addition the panel alignment is monitored for the active hull panels against a subset (NFS

f ) of the free

23

surface panels in the close vicinity to the hull.

Geometrical bounds

Geometrical bounds are utilized to preserve a permissible range for the hull panel relocations. Typically, in ship application the hull surface is forced to lie within the bounding box defined by the

maximum length, the maximum beam and the maximum draft. In addition, in order to comply with

the perturbation approach the panel relocations are required to be small with respect to a reference

length.

Maximum panel relocation

As described above the present optimization approach utilizes gradients in the optimization. The gradients are evaluated by means of explicit linear relations. A perturbation scheme is applied to linearize

the boundary value problem. The basic assumption is that the magnitude of the perturbation is small.

Since the hull panel normal relocations are selected 24 as perturbation quantities their relocation magnitude need to be reasonably bounded.

Similar to the treatment of the passive hull panels the maximum allowed normal relocation of the

active 25 hull panels can be formulated in terms of an inequality constraint

(1)

nk max k~nk k 0

k = 1, , NK .

(4.25)

23

in addition to the hull panel source densities

25 active hull panels are indicated by index k

24

67

4 Sensitivity analysis

Different choices of the maximum panel normal relocation nk max are conceivable. Following the definitions of the perturbation magnitudes in section 4.2.2 the maximum panel relocation may be simply

related to a characteristic length L (e.g. the length between perpendiculars or the waterline length)

0 < ~nmax 1 ,

(4.26)

controlled by a parameter ~nmax . Alternatively, the maximum panel relocation may be related to a

typical hull panel size, e.g., the average square roots of the hull panel areas Si

NH

i=1

Si

0 < ~nmax 1 .

NH

(4.27)

Which of the two expressions is eventually utilized to define the maximum normal panel relocation

nmax is a practical decision by the user. For the perturbation scheme it is of minor importance as long

as the maximum panel relocation is kept small.

Experiences gained from the optimization applications suggest to determine a perturbation magnitude

n according to (4.19) which at the same time acts as the maximum panel normal relocation, i.e.,

nmax = n.

Longitudinal bounds

In ship optimization the ship length most often is given as a fixed or at least only slightly modifiable

quantity. Accordingly, the longitudinal panel relocation, here in x-direction, may be bounded at both

bow and stern.

For the ships longitudinal direction this requires for the panel relocations

(0)

(1)

k = 1, , NK ,

(4.28)

with CTk x indicating the x-component of the k-th panel control point. xBow and xStern are the fore- and

aftmost hull point, respectively. Equation (4.28) can be recast as follows:

(0)

(1)

(0)

(1)

(CTk x + nk x ) + xStern + Stern LPP

0 Bow 1 ,

0 Stern 1 .

(4.29)

(4.30)

Introducing slack parameters Bow,Stern the longitudinal panel relocations at the bow and stern can be

variably controlled, respectively. For instance, small values of the slack parameters allow for a little

leeway, i.e., the active hull panels are permitted to move a small distance beyond the extents of the hull.

However, if the panels are to stay within the extents then the slack parameters would have to be set to

zero.

Lateral bounds

Similar to the longitudinal extension of the ship the hull often has to comply with a maximum breadth.

Accordingly, the lateral panel relocation, here in y-direction, may be bounded by the symmetry plane

and the maximum breadth.

68

For the ships lateral direction this requires for the panel relocations

(0)

(1)

k = 1, , NK ,

(4.31)

with CTk y indicating the y-component of the k-th panel control point. yB max represents the maximum

lateral hull extension and yMS the y-coordinate of the midship plane in the SHIPFLOW coordinate

system, usually yMS = 0. Equation (4.31) can be interpreted in terms of inequality constraints for the

midship plane and the maximum breadth, respectively,

B

(0)

(1)

CTk y + nk y (yMS + MS ) 0

2

B

(0)

(1)

(CTk y + nk y ) + yB max + SS 0

2

0 < MS 1 ,

(4.32)

0 SS 1 ,

(4.33)

with B denoting the ships breadth. By means of the slack parameters MS,SS the lateral panel relocations at the midship plane and the shipside can be variably controlled. For instance, a small value of SS

enables the active hull panels to move a small distance beyond the half-breadth of the hull. Otherwise,

to keep the panels exactly within the half-breadth then SS would be set to zero. However, the lower

bound associated with MS must be chosen large enough to keep the panels away from the midship

plane.

Vertical bounds

Finally, the hull design must not violate the maximum permitted draft. Like the ships breadth the draft

is usually limited. In order to avoid the upmost hull panels (at the level of the wavy free surface) to

move above a certain waterline level an upper vertical bound might be defined. The appropriate zW FS

bound corresponds to the z-coordinate of the control point of the upmost wetted hull panel typically

located in the wavy free surface level (WFS) along the hull side.

For the ships vertical direction, the z-direction in the SHIPFLOW coordinate system, this requires for

the panel relocations

(0)

(1)

zT max CTk z + nk z zW FS

k = 1, , NK ,

(4.34)

with CTk z indicating the z-component of the k-th panel control point. zW FS represents a certain waterline

above the wavy free surface level and zT max the z-coordinate of the maximum vertical extension at

keel. Equation (4.34) can be interpreted in terms of inequality constraints for the maximum vertical

extensions

(0)

(1)

0 Keel 1 ,

(4.35)

(CTk z + nk z ) + zW FS + W FS T 0

0 W FS 1 ,

(4.36)

(0)

(1)

with T denoting the ships draft. Again using slack parameters, here Keel,W FS , the vertical panel

relocations can be variably controlled. For instance, a small value of Keel permit the active hull panels

to move a small distance below the draft of the hull. Otherwise, to keep the panels exactly within the

draft Keel would be set to zero.

69

(Re-)Initialization of

optimization task

(variables, constr., bounds)

Transfer matrices

Systematic hull

perturbation analysis

Initialization

4 Sensitivity analysis

(inverse perturb. solution)

(wave elevat., flow quant.)

Impact on

wave energy distribution

(free wave spectrum, RWP)

Gradients matrices

Sensitivity analysis

Systematic variation of

design variables

(flow quantity related)

Evaluation of gradients

(MoM, constraints, bounds)

Fig. 4.15: Process flow of hull perturbation and sensitivity analysis (detail magnification of the process

flow, figure 2.1).

The process flow of the sensitivity analysis is depicted in figure 4.15. Figure 4.15 gives a detail of the

general optimization process flow, see figure 2.1.

The sensitivity analysis is the central unit of the optimization scheme. The sensitivity analysis comprises the cause and effect chain from the variation of the optimization variables (the cause) to the

variation of the wave pattern, the free wave spectra and the wave pattern resistance (the effects). The

sensitivity analysis is based upon a preceding flow solution and wave cut analysis. In the present

scheme a converged nonlinear free surface CFD simulation by SHIPFLOW is utilized as a base flow

solution.

In a preparation step the perturbation scheme is initialized by evaluation of the transfer matrices which

serve as the basic instrument for the sensitivity analysis. The partial derivatives of the influence coefficients with respect to the hull panel normal relocations in the transfer matrices are numerically

evaluated. This is done by subsequently shifting each active hull panel by the perturbation magnitude

n, as defined by (4.19), in its panel normal direction. Then, in principle, the influence coefficients

are evaluated for the initial, undisturbed and the relocated panel position and the derivatives in the

transfer matrices are numerically determined by forward differencing. The transfer matrices comprise

the variation effect on both the hull and the free surface portion. However, at the stage when the

70

transfer matrices are evaluated the free surface panelization itself is temporarily frozen. This process

continues for all active hull panels until the transfer matrices are completely assembled. However,

since the normal relocation of a single hull panel only affects a single row and a single column in the

influence coefficient matrices a fast and memory conserving scheme was developed for the successive

construction of the transfer matrices.

Once the transfer matrices are assembled the systematic variation of the optimization variables may

start. A single optimization variable, i.e., a single hull source density, is varied at a time. The source

density variation essentially affects the neighbouring hull panels. In order to restore the hull boundary

condition 26 the adjacent panels are forced to adjust their positions, accordingly. The panel adjustment

by panel normal relocations is controlled by virtue of the transfer matrices and the inverse perturbation

relation (4.9). Hence, the variation of a single source density results, at first instance, in a normal

relocation of the neighbouring hull panels. This local deformation of the panel mesh again causes a

variation of the hull form parameters.

The locally confined panel normal relocation in turn has an impact on the flow quantities in terms

of the flow potential and the wave elevation, as will be seen in section 4.6. The impact on the free

surface elevation determined from (4.79) considering the hull panel normal relocations from (4.9), the

perturbation potentials from the direct perturbation relation (4.13) and the transfer matrices (4.80).

Details are given in section 4.6. Prior to further processing the perturbation wave elevations at the free

surface collocations points are interpolated in order to yield the perturbation wave cut at a predefined

offtrack position.

Finally, the variation of the free wave spectra and the impact on the wave pattern resistance is determined by a longitudinal wave cut analysis of the perturbation wave cut. The reader is referred to

section 4.7 for details. The cause and effect chain is thus completed up to the objective function of

optimization, i.e., the wave pattern resistance.

Principally, the above procedure is to be repeated for all hull panel source densities which are considered as optimization variables. However, the formulation of the perturbation relations allows for a joint

treatment of all optimization hull panels in one single step. Hence, the above steps need to be processed

only once in order to perform the whole sensitivity analysis. By far the most resource intensive part is

the evaluation of the transfer matrices in the preceding preparation step.

For further processing the sensitivity relations are utilized in terms of gradients. The gradients are

numerically evaluated by means of forward differences. Details of the gradient evaluation are given in

section 4.8.

The present optimization scheme is based upon locally confined hull variations. Hence, the variation

of a single hull source density is supposed to result in a locally restricted hull panel normal relocation.

Whereas, the global hull variation results from a superposition of all local variation effects with the

objective of minimizing the wave pattern resistance. In this section an approach is outlined how to

confine the hull variations locally.

26

in a perturbation sense

71

4 Sensitivity analysis

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

......... 4

......... 3

......... 2

......... 1

Level 0

Fig. 4.16: Schematic sketch of the hull panel influence segment of a single source density variation at

level 0, edge cases (left) and central location (right).

In ship hull optimization the main hull form parameters, generally, are considered as fixed given quantities or are kept within lower and upper bounds. The reason is to maintain, for instance, a certain

predefined deadweight, to maintain a specific floating position or to comply with stability requirements. Normally, many criteria are in the catalog of requirements. In the present optimization scheme

the main hull form parameters, like the displacement, the waterline area, the center of buoyancy and

the longitudinal center of flotation are considered.

The present approach is based on discretized hull boundaries. A hull panel normal relocation naturally

causes a variation of the hull form parameters. This implicates the evaluation of the form parameter

variations on a discrete basis, too. This practice coincides with the linear perturbation approach. 27

The hull optimization is carried out for the effective floating position of the advancing ship including

trim and sinkage. Consistently, the form parameter variations are evaluated for the advancing ship

considering the wetted hull panelization under the wavy free surface and the effect of trim and sinkage. 28

Again, all definitions and derivations in this section are related to the SHIPFLOW coordinate system

(compare section 4.3.2).

Hull form

Since the present optimization scheme is based upon locally confined hull variations the variation of

a single hull source density is supposed to result in a locally restricted hull panel normal relocation.

Mathematically this is expressed by the inverse perturbation relation (4.9). A single source density

variation in (4.9) is realized by a single nonzero entry in the right hand side vector ~(1) with otherwise

zero components. In other words, it is searched for a hull panel normal relocation which complies with

a single nonzero source density variation, but zero variations at the remaining hull panels. However,

27

Comparisons with results from continuous integration of the form parameters confirm the validity of evaluating the form

parameter variations on a discrete basis.

28 This practice somewhat differs from the common approach to assign the form parameters and hydrostatic coefficients to

the ships floating position at rest. However, both approaches are realizable within the present scheme, alternatively.

72

this constraint is too strict for the hull panel relocation as tests with different ship applications and

different variation scenarios revealed. Mathematically the inverse relation (4.9) provides a unique

solution for this case. However, from an application point of view the resulting hull panel relocation

can not be considered local since undesired, awkward panel movement can be observed at considerable

distance from the origin of the source density variation. The undesired panel movement mainly occurs

at the (flat of) bottom and (flat of) side of the ship where the panels are almost aligned with the main

flow direction. Due to their vanishing angle of incidence the disturbance velocities are small or even

vanishing at those panels. This, normally, is accompanied by a comparatively small source density

value. As a consequence, these hull panels are highly sensitive to any, yet remote, perturbation of the

flow boundaries. In order to overcome this problem a slight modification of the above procedure should

be introduced:

A local panel relocation can be easily attained by restricting the solution of (4.9) solely to the hull

panels in the neighbourhood of the source density variation. This is a somewhat practical approach and

it proves to be very effective in application. Moreover, this approach is substantiated by the special

property of a single spacial source singularity whose main influence effect is locally confined, since

the velocity induced by the source singularity vanishes rapidly with 1/r3 , r being the distance to the

singularity.

In figure 4.16 the influence segment of the source density variation is sketched. For example an influence segment up to level 1 means that solely the hull panels up to level 1 are shifted according to the

inverse relation (4.9) in order to adjust a given source density variation at level 0 and, simultaneously,

preserving the source densities at the adjacent hull panels up to level 1. At the same time all panels

exterior to the segment are untouched.

From an application point of view the influence segment must be locally confined, thus, allowing a

locally traceable variation effect. A too large influence segment may also cause considerable trouble

in the optimization when preserving the geometrical constraints. On the other hand the influence

segment must be large enough so that the variation effects are of sufficient magnitude to be utilized

in terms of gradients. From a systematic optimization study an influence segment up to level 1

(or 2) was identified as a suitable compromise. In the special case of an influence segment which

is completely reduced to level 0, i.e., to the variation hull panel itself, there is effectively no more

difference between varying the hull source densities or relocating the hull panels directly. Optimization

studies with a degenerated influence segment (to level 0) showed a tendency to produce somewhat

bumpy shape variations.

Finally, the global hull variation results from a superposition of all local variation effects with the

objective of minimizing the wave pattern resistance.

Displacement volume

The effective displacement volume may be represented in terms of the perturbation notation as the sum

of the basic displacement V (0) and its variation V 29

V = V (0) + V .

29

(4.37)

73

4 Sensitivity analysis

The displacement variation resulting from the normal relocation of the single hull panel k can be calculated on a discrete basis by the panel area Sk multiplied by the magnitude of the panel normal relo(1)

cation nk

(1)

(1)

Vk = Sk ~nk nk = Sk nk 1

k = 1, , NK .

(4.38)

Evidently, the displacement variation due to the simultaneous normal relocation of all active hull panels

is given by the sum of the Vk

NK

V =

Vk ,

(4.39)

k=1

Generally, the displacement variation is bounded or even set to zero. In the present optimization scheme

the variation of the derived hull form parameters is controlled by variable lower and upper bounds. The

bounds for the displacement variation (4.39) can be expressed in terms of two inequality constraints

Vlow Vinit + V 0

Vup Vinit V 0

0 < Vlow 1 ,

0 < Vup 1 ,

(4.40)

(4.41)

with Vlow/up being parameters controlling the displacement variation and Vinit is the initial displacement.

In the present scheme Vinit is selected as the displacement of the effective submerged hull portion under

the wavy free surface of the advancing ship. Vinit is deduced from a CFD simulation for the initial hull

considering the wavy free surface and the effect of trim and sinkage.

Center of buoyancy

The effective longitudinal (x-direction) and vertical (z-direction) center of buoyancy may be repre(0)

(0)

sented in terms of the perturbation notation by the basic location LCB , VCB varied by LCB , VCB

(0)

(0)

(4.42)

The variation of the center of buoyancy due to the normal relocation of the single hull panel k can be

calculated on a discrete basis by

1 (1)

(0)

LCB k = CTk x + nk x ,

2

1 (1)

(0)

VCB k = CTk z + nk z

2

k = 1, , NK .

(4.43)

The variation of the center of buoyancy LCB , VCB due to the simultaneous normal relocation of all

active hull panels can be expressed in the form

(0)

(0)

74

(4.44)

with LCB init indicating the initial center of buoyancy. The center of buoyancy in (4.44) is evaluated

by

NK

(0)

k=1

LCB =

V (0) + V

NK

(0)

k=1

VCB =

V (0) + V

(4.45)

Apparently, in (4.45) both the numerator and the denominator are linear functions of the perturbation

quantities. Thus, the rational functions LCB and VCB depend in a nonlinear manner on the perturbation

quantities. Unfortunately, this violates the basic assumptions of the linear perturbation approach. In

order to avoid rational expressions in the optimization scheme the 1st -order statical moments may be

used instead whose variations are given by virtue of the relations (4.44) by

(0)

(0)

(4.46)

NK

(0)

k=1

(0)

NK

k=1

NK

MSL =

k=1

NK

MSV =

(4.47)

k=1

The sum terms in (4.47) are still of higher order in the perturbation quantities. In order to comply with

the linear perturbation assumption equation (4.47) requires a consistent linearization. The perturbation

terms may be temporarily written by introducing the small perturbation parameter . Then applying

(4.38), (4.39) and (4.43) to (4.47) yields

NK

MSL =

1 (1)

(0)

(CTk x + 2 nk x ) Vk LCB init

k=1

NK

MSV =

(0)

(CTk z

k=1

1 (1)

+ nk z ) Vk VCB init

2

NK

Vk ,

k=1

NK

Vk .

(4.48)

k=1

75

4 Sensitivity analysis

Finally, the terms in parenthesis in (4.48) are expanded and higher order contributions associated with

O(2 ) are neglected. 30 This results in linear expressions for the variation of the 1st -order statical

moments of buoyancy

NK

(0)

(0)

CTk x

MSL =

k=1

NK

CTk z

MSV =

k=1

(4.49)

Presuming the hull perturbations to be small, the variation of the 1st -order statical moments of buoyancy

(4.49) can be considered an equivalent to the variation of the center of buoyancy (4.44). 31

According to the displacement variation variable lower and upper bounds for the variation of the 1st order statical moments of buoyancy (4.49) can be defined in terms of inequality constraints

MSL low MSL init + MSL 0

(4.50)

(4.52)

0 < MSL up 1 ,

0 < MSV up 1 ,

(4.51)

(4.53)

with MSL low/up and MSV low/up being parameters controlling the variation of the statical moments. MSL init

and MSV init are the initial 1st -order statical moments of buoyancy which are deduced from the effective

submerged hull portion under the wavy free surface of the advancing ship, with

MSL init = LCB init Vinit 6= 0 ,

Waterline area

The effective waterline area may be represented in terms of the perturbation notation as the sum of the

(0)

basic waterline area AW L and its variation AW L

(0)

AW L = AW L + AW L .

(4.54)

In the present context the waterline area is associated with the effective wavy free surface contour along

the hull side of the advancing ship. That is, the waterline area is put up by horizontal cuts through the

hull along the effective wavy free surface contour at the hull side. The variation of the waterline area

due to the normal relocation of a single active waterline panel k is deduced on a discrete basis from the

length of the upper panel edge multiplied by the panel normal relocation. Both the upper panel edge

and the panel normal relocation are, beforehand, projected onto the horizontal xy-plane. The upper

panel edge in the SHIPFLOW frame of reference is for each hull panel defined by the upper panel

30

31

A study, conducted by the author, with realistic perturbation cases revealed a very close correlation of the variation of the

center of buoyancy and the variation of the 1st -order statical moments.

76

corner points (CP) number #2 and #3. 32 Thus, the variation of the waterline area due to the normal

relocation of the single active waterline panel k can be calculated on a discrete basis by

(0)

(0)
(1)

AW L k =

k = 1, , NK NK ,

(4.55)

CP 3 k xy CP 2 k xy
nk knk xy k

with

(0)

(0)

CP3 k x CP2 k x

(0)

(0)

(0)

(0)

CP

CP

=

CP

CP

2 k xy

3 k xy

2 k y

3 k y

0

and

nk x

knk xy k =
nk y
.

0

The variation of the waterline area AW L due to the simultaneous normal relocation of all active waterline panels is given by the sum of the AW L k

AW L =

NK

AW L k

k =1

(4.56)

Variable bounds for the variation of the waterline area (4.56) can be expressed in terms of inequality

constraints

AW L low AW L init + AWL 0

0 < AW L low 1 ,

AW L up AW L init AWL 0

0 < AW L up 1 ,

(4.57)

(4.58)

with AW L low/up being the parameters controlling the variation of the waterline area. AW L init is the initial

waterline area which is deduced from the effective wavy free surface contour along the hull side of the

advancing ship.

Longitudinal center of flotation

The effective longitudinal center of flotation may be represented in terms of the perturbation notation

(0)

by the basic location LCF varied by LCF

(0)

(4.59)

The variation of the longitudinal center of flotation can be derived according to the center of buoyancy.

The variation of the longitudinal center of flotation due to the normal relocation of the single active

waterline panel k can be calculated on a discrete basis by

1 (1)

(0)

LCF k = Pk + nk x

2

k = 1, , NK NK .

(4.60)

with

(0)

Pk =

32

1

(0)

(0)

(CP2 k x +CP3 k x ) .

2

(4.61)

For an observer looking from outside in positive y-direction into the hull the panel corner points are for each hull panel

oriented clockwise with the first corner point located in the lower left panel corner.

77

4 Sensitivity analysis

The variation of the center of flotation LCF due to the simultaneous normal relocation of all active

waterline panels can be expressed in the form

(0)

(4.62)

with CF init indicating the initial center of flotation. The center of flotation in (4.62) is determined by

(0)

NK

(0)

LCF AW L + LCF k AW L k

k =1

(0)

AW L + AWL

LCF =

(4.63)

Similar to the derivation of the center of buoyancy, in order to avoid rational expressions, not the center

of flotation (4.63) is used in the optimization scheme but the 1st -order statical moment whose variation

is given by virtue of relation (4.62) as

(0)

(0)

(4.64)

LCF AW L =

(0) (0)

LCF AW L +

NK

LCF k

k =1

AW L k ,

MSLCF =

NK

LCF k

k =1

AW L k LCF init AW L .

(4.65)

The sum in (4.65) is still of higher order in the perturbation quantities. In order to comply with the linear

perturbation assumption equation (4.65) requires a consistent linearization. The perturbation terms may

be temporarily expressed by introduction of the small perturbation parameter . Then applying (4.55),

(4.56) and (4.60) to (4.65) yields

MSLCF =

NK

(0)

(Pk

k =1

1 (1)

+ nk x ) AW L k LCF init

2

NK

AW L k

k =1

(4.66)

Finally, the term in parenthesis in (4.66) is expanded and higher order contributions associate with

O(2 ) are neglected. 33 This results in a linear expression for the variation of the 1st -order statical

moment of flotation

MSLCF =

NK

(0)

Pk

k =1

(4.67)

Variable bounds for the variation of the 1st -order statical moment of flotation (4.67) can be defined in

terms of inequality constraints

MSLCF low MSLCF init + MSLCF 0

MSLCF up MSLCF init MSLCF 0

33

0 < MSLCF up 1 ,

(4.68)

(4.69)

78

Fig. 4.17: Outward normal shift of a panel segment at the bow (to the right) of the Wigley hull (top

side view). Predicted change of the hull source density distribution by the direct perturbation

approach (middle) compared to the result of a SHIPFLOW recomputation (bottom).

with MSLCF low/up being the parameters controlling the variation of the statical moment. MSLCF init indicates the initial 1st -order statical moment of flotation which is deduced from the effective wavy free

surface area of the advancing ship, with

MSLCF init = LCF init AW L init 6= 0 .

The variation of a single optimization variable, i.e., a single hull source density, results in a locally

confined hull form variation, as discussed in section 4.5. The local hull form variation in turn has an

impact on the flow quantities in terms of the flow potential and the wave elevation.

Source density distribution

In the direct perturbation approach 34 (4.13) it is searched for a modification of the hull and, particularly,

of the free surface source density distribution subject to a preceding hull form variation. The preceding

hull form variation again is the result of a variation of the optimization variables at the first instance of

the perturbation scheme.

Figure 4.17 exemplarily illustrates the effect of an imaginary outward normal relocation of a 3 3

panel segment 35 at the bow of the standard Wigley hull featuring a double symmetric mathematical

hull form with a rectangular center line contour (see section 6.2 for details of the Wigley hull). The

panel perturbation is based upon a preceding free surface flow solution at a Froude number of 0.3 (the

34

35

stepped relocation of a 3 3 panel segment with the maximum outward shift at the middle panel

79

4 Sensitivity analysis

free surface is not pictured). A comparison, on a qualitative basis, of the predicted change of the hull

source density distribution by the direct perturbation relation (4.13) with the result of a SHIPFLOW

recomputation shows an excellent agreement. In figure 4.17 it can be observed that at the shifted

panel segment itself the magnitudes of the panel source densities decreased, whereas in the immediate

vicinity they increased.

Wave elevation and wave cut

The normalized 36 free surface wave elevation is determined by the linearized dynamic free surface

boundary condition (3.8)

"

#

2 2 2

1 2

+

+

.

(x, y) = Fn 1 +

+

+

2

2

x

y

z

x x y y z z

Here the SHIPFLOW notation is used in which is the flow potential of the previous iteration step 37

whereas corresponds to the current flow solution, e.g. the converged solution.

The effective wave elevation may be represented in terms of the perturbation notation as the sum of the

basic wave elevation (0) and its variation

(x, y) = (0) (x, y) + (x, y) .

(4.70)

The basic wave elevation (0) in (4.70) corresponds to the latest flow solution which is determined by

a nonlinear free surface CFD simulation. This flow solution is considered fully converged with respect

to the wave formation, the source density distribution and the freely adjusted ships floating position.

The free surface domain is discretized by Rankine panels. The boundary condition is satisfied in

a discrete sense at the free surface panel collocation points. Accordingly, the wave elevations are

evaluated in a discrete sense with the wave elevation at collocation point i given by (3.12)

i =

1 2

Fn 1 + 2x i + 2y i + 2z i 2 (x i x i + y i y i + z i z i )

2

i = NH + 1, , NFS ,

with x = /x etc. and NH < i NFS for the free surface panels. The velocities according to the

previous SHIPFLOW iteration step (1) 38 in (3.12) are determined from

NFS

x i = Ux i =

j=1

NFS

y i = Uy i =

(1) (1)

j +U y

Yi j

j=1

NF S

z i = Uz i =

(1) (1)

j +U x

Xi j

(1) (1)

j +U z

Zi j

j=1

i = NH + 1, , NFS .

(4.71)

36

by LPP

In the 1st step of a nonlinear free surface iteration is selected as the double-body flow or simply as the undisturbed flow

potential.

38 the last but one SHIPFLOW iteration step related to the converged solution

37

80

The hull perturbation is carried out around the base flow solution which corresponds to the last converged SHIPFLOW solution (0) . Hence, the effective velocities at the free surface collocation points

are composed of the basic velocities and the perturbation effects

NF S

x i = ux i =

Xi j j +U x ,

j=1

NF S

y i = uy i =

Yi j j +U y ,

j=1

NFS

z i = uz i =

Zi j j +U z

i = NH + 1, , NFS .

j=1

(4.72)

The influence coefficients Xi j Zi j in (4.72) are pure geometrical quantities depending on the geometry, the relative position and the orientation of the panels. The influence coefficients are affected by the

(1)

perturbation, in particular by the panel normal relocations nk . Since this dependency is implicit the

influence coefficients are expanded in Taylor series which are truncated after the linear terms

(0)

NK

(0)

Xi j = Xi j +

(1)

nk

Xi j

nk

k=1

Yi j =

(0)

Yi j +

NK

k=1

Zi j =

(0)

Zi j +

NK

k=1

+ O(2 ) ,

(0)

(1)

nk

Yi j

nk

+ O(2 ) ,

(0)

(1)

nk

Zi j

nk

+ O(2 )

i, j = NH + 1, , NFS ,

(4.73)

where the perturbation terms are expressed by the small perturbation parameter . The effective source

density distribution is composed of the base solution (4.5) and the source density variation which is

determined by the solution of the direct perturbation relation (4.13)

(0)

(1)

j = j + j

+ O(2 )

j = 1, , NFS .

(4.74)

Now, if the relations (4.73) and (4.74) are introduced to the free surface velocities (4.72) one gets

NF S

x i =

(0)

Xi j +

j=1

NF S

j=1

k=1

(0) (0)

Xi j j +

y i = ,

z i =

NK

(1)

nk

NFS

j=1

(0) !

Xi j

nk

(0)

(1)

j + j

(0) (1)

Xi j j +

NF S

j=1

(0)

j

NK

k=1

+U x

(1)

nk

(0)

Xi j

nk

NF S

j=1

(1)

j

NK

k=1

i = NH + 1, , NFS .

(0)

(1)

nk

Xi j

nk

+U x ,

(4.75)

81

4 Sensitivity analysis

In the sense of a consistent linearization, higher order contributions of O(2 ) in the perturbation parameter are neglected in (4.75) which yields

NFS

x i =

NFS

(0) (0)

Xi j j +

j=1

NFS

y i =

j=1

NFS

z i =

j=1

(0) (0)

Yi j j +

(0) (0)

Zi j j +

(0) (1)

Xi j j

NFS

(0)

NK

(0)

j=1

j=1

k=1

NF S

NFS

NK

j=1

NFS

j=1

(0) (1)

Yi j j +

(0) (1)

Zi j j +

j=1

NFS

j=1

(0)

j

(0)

j

Xi j

(1)

nk

k=1

k=1

+U x + O(2 ) ,

(0)

(1)

nk

NK

nk

Yi j

nk

+U y + O(2 ) ,

(4.76)

(0)

(1)

nk

Zi j

nk

+U z + O(2 )

i = NH + 1, , NFS .

Finally, the expressions for the free surface velocities from the previous, i.e., the next to last

SHIPFLOW iteration (4.71) and the velocities according to the current, converged solution comprising

the perturbation effects (4.76) are introduced to the linearized dynamic free surface boundary condition

(3.12). After sorting in orders of this yields for the 0th -order the basic wave elevation according to

(3.12) where the flow potential = (0) equals the base flow potential according to the current, converged solution before perturbation. If (3.12) is formulated in terms of the free surface velocities the

basic wave elevation follows as

(0)

i =

i

1 2h

(0)

(0)

(0)

Fn 1 +Ux2i +Uy2i +Uz2i 2 Ux i ux i +Uy i uy i +Uz i uz i

2

i = NH + 1, , NFS . (4.77)

The 1st -order variation of the wave elevation at the i-th free surface collocation point results in

i =

Fn2

"

NFS

Ux i

j=1

(0) (1)

Xi j j +

NF S

Uy i

j=1

(0)

j

NFS

NK

k=1

(1)

nk

(0) !

Xi j

nk

(0) !

NK

(0) (1)

(0)

(1) Yi j

Yi j j + j nk nk

j=1

j=1

k=1

NF S

Uz i

NF S

j=1

(0) (1)

Zi j j +

NFS

j=1

(0)

j

NK

k=1

(1)

nk

(0) !#

Zi j

nk

i = NH + 1, , NFS .

(4.78)

h

~ x i X(0) ~ (1) + TX (0) ~n(1) +

~ = Fn2 U

~ y i Y(0) ~ (1) + TY (0) ~n(1) +

U

i

~ z i Z(0) ~ (1) + TZ(0) ~n(1)

U

[N f s ] =

82

(4.79)

with N f s being the number of free surface panels excluding the hull panels. TX TZ in (4.79) are

transfer matrices for the variation of the free surface wave elevation with

TX

TY

TZ

(0)

(0)

(0)

(0)

TX ik

(0)

TY ik

(0)

TZ ik

NF S

j=1

NFS

j=1

NFS

j=1

(0)

(0)

j

Xi j

nk

(0)

(0)

j

Yi j

nk

(0)

(0)

j

Zi j

nk

,

i = NH + 1, , NFS

.

k = 1, , NK

(4.80)

All the transfer matrices, TX TZ and T in equation (4.8), are evaluated simultaneously in the preparation step where the perturbation scheme is initialized, see section 4.4.

As already mentioned, the basic quantities indicated by (0) in the above relations correspond to the last

flow solution which is determined by a nonlinear free surface iteration utilizing SHIPFLOW. This flow

solution can be considered fully converged, particularly, with respect to the wave formation so that the

maximum wave change at convergence is at least one order less than the expected variations due to the

hull perturbation. To be on the safe side, for the wave elevation a convergence criterium 39 of 107

to 1010 is selected in the CFD simulation. An advantageous side effect of the small convergence criterium is that the actual nonlinear free surface boundary conditions, (3.6) and (3.7), can be considered

fully satisfied on a discrete, numerical basis with

= (0) . Therefore, the variation of the free surface

wave elevation as an effect of the hull perturbation takes place, practically, around the actual nonlinear

free surface.

In practical applications the free surface domain needs to extend far enough up- and downstream and

sidewise in order to allow for a proper wave cut analysis. At the same time the panelization has to be

dense enough to capture much of the wave information. This in turn causes large equations systems

which are costly both in computation time and in memory consumption. This gives rise to reduce

the portion of the free surface domain considered in the evaluation of the variation impact on the

free surface source densities and wave elevations. As already addressed in section 4.2.1, a validation

study conducted by the author revealed that a truncation 40 of the free surface panelization sidewise

at approximated 2/3 (or even 1/2) of the regular 41 lateral extension is acceptable due to a negligible

influence on the results as long as the perturbations are kept small.

Prior to further processing, by means of a longitudinal wave cut analysis, the discrete wave elevations including the variation effects from (4.79) are interpolated laterally and longitudinally at the free

surface collocation points in order to yield a longitudinal wave cut at a predefined lateral position,

compare section 3.2.2. At first the wave elevations are interpolated laterally in y-direction for constant

x-positions by means of a piecewise cubic spline interpolation. The curve mesh, thus constructed,

then is interpolated longitudinally in x-direction for a constant lateral wave cut position. In the present

scheme a linear interpolation is utilized longitudinally. A comparison study applying experimental

results indicated this interpolation to be stable, also compare Janson (1997).

39

The underlying CFD simulation by SHIPFLOW still includes the whole free surface domain in the solution.

41 at minimum covering the Kelvin wedge

40

83

4 Sensitivity analysis

VS

# 76 (10% LPP)

# 68 (10% LPP)

Fig. 4.18: Panelization of the Wigley hull. Source density perturbations at selected panels (for the sake

of clarity only the panels at the starboard side are marked).

The effective longitudinal wave cut including the effect of variation may be represented in terms of the

perturbation notation as the sum of the basic wave cut (0) and its variation

(x, y) = (0) (x, y) + (x, y) .

(4.81)

Formally the perturbation relation (4.81) is equivalent to relation (4.70), except that (4.81) applies to a

constant transverse cutting plane or offtrack position y = const..

The variation of the free wave spectra and the impact on the wave pattern resistance is determined by

a wave cut analysis of the longitudinal wave cut, including the variation effects. The cause and effect

chain is thus completed up to the objective function of optimization, the wave pattern resistance.

Free wave spectra

The sine (odd) and cosine (even) components of the free wave spectrum, in non-dimensional form, are

determined as functions of the component wave direction from the weighted Fourier transforms of a

longitudinal wave cut according to (3.35)

s

[SW FT (s) cos(u y) +CW FT (s) sin(u y)] ,

s

AC () = [SW FT (s) sin(u y) CW FT (s) cos(u y)] .

AS () =

84

prediction ~ single restart step | | | fully converged solution | |

Wave pattern resist. change CWP 1000 [-]

0.1

0.05

-0.05

-0.1

-0.15

0

-0.02

-0.04

-0.06

x / LPP [-]

Free wave spectra change AC 1000 [-]

2

-1

-2

20

40

60

Component wave direction []

-1

-2

20

40

60

Component wave direction []

20

40

60

Component wave direction []

Fig. 4.19: Improvement of the wave making characteristics of the Wigley hull at Fn = 0.3. The result

of the first sub-optimization loop of an optimization process, considering all but the center

line hull source densities as free variables, is shown.

The weighted Fourier transforms of a longitudinal wave cut are given by virtue of (3.34)

SW FT (s) =

ZxE p

s2 1

s2 1

ZxE

CW FT (s) =

(x, y) sin(s x) dx +

Z p

xE

(x, y) cos(s x) dx +

s2 1 e

(x, y) sin(s x) dx ,

Z p

xE

s2 1 e

(x, y) cos(s x) dx .

In practical applications, due to limited computer resources (CFD) or restrictions due to tank wall

reflections (EFD), longitudinal wave cuts (x, y) need to be truncated at a finite distance downstream

of the hull. In order to retain the wave energy contained within the wave pattern downstream of the

truncation point a special truncation correction is applied. An analytical function is introduced e

(x, y)

and fitted to the tail end of the wave cut signal up to the truncation point. For details see section 3.2.3.

So as to keep the following derivation clear and straightforward it may be temporarily assumed that

the wave cut signal continues far downstream avoiding the need for a truncation correction. Hence,

85

4 Sensitivity analysis

prediction ~ single restart step | | | fully converged solution | |

Wave pattern resist. change CWP 1000 [-]

0.0004

0.0002

0

-0.0002

-0.0004

-0.00005

-0.0001

-0.00015

-0.0002

0

20

40

60

Component wave direction []

x / LPP [-]

Free wave spectra change AC 1000 [-]

0.006

0.004

0.002

0

-0.002

-0.004

-0.006

0

20

40

60

Component wave direction []

0.006

0.004

0.002

0

-0.002

-0.004

-0.006

0

20

40

60

Component wave direction []

Fig. 4.20: Change of the wave pattern characteristic due to variation of one single hull panel source

density in its optimum gradient direction until the perturbation bound is reached. Here panel

# 68, located above keel level 10% LPP downstream of the bow, is selected.

without loss of generality, the weighted Fourier transforms simplify to

SW FT (s) =

Z p

s2 1 (x, y) sin(s x) dx ,

s2 1 (x, y) cos(s x) dx .

CW FT (s) =

(4.82)

If the longitudinal wave cut, including the effect of variation (4.81), is introduced to (4.82) then the

(0)

(0)

weighted Fourier transforms follow as the sum of the basic transforms SW FT and CW FT and their variations SW FT and CW FT

(0)

(0)

86

(4.83)

prediction ~ single restart step | | | fully converged solution | |

Wave pattern resist. change CWP 1000 [-]

0.02

0.01

-0.01

-0.002

-0.004

-0.006

-0.008

-0.01

0

x / LPP [-]

Free wave spectra change AC 1000 [-]

0.2

-0.2

-0.4

0

20

40

60

Component wave direction []

20

40

60

Component wave direction []

0.2

-0.2

-0.4

0

20

40

60

Component wave direction []

Fig. 4.21: Change of the wave pattern characteristic due to variation of the hull panel source density

# 76, located directly beneath the wavy free surface level 10% LPP downstream of the bow.

with SW FT and CW FT determined from

SW FT (s) =

Z p

s2 1 (x, y) sin(s x) dx ,

s2 1 (x, y) cos(s x) dx .

CW FT (s) =

(4.84)

(0)

(0)

Accordingly, this yields for the free wave spectra the sum of the basic wave spectra AS and AC and

their variations AS and AC

(0)

AS () = AS () + AS () ,

(0)

AC () = AC () + AC () .

(4.85)

s

[ SW FT (s) cos(u y) + CW FT (s) sin(u y)] ,

s

AC () = [ SW FT (s) sin(u y) CW FT (s) cos(u y)] .

AS () =

(4.86)

87

4 Sensitivity analysis

prediction ~ single restart step | | | fully converged solution | |

Wave pattern resist. change CWP 1000 [-]

0.001

0.0005

0

-0.0005

-0.001

-0.0015

0

0.0004

0.0002

-0.0002

-0.0004

0

x / LPP [-]

Free wave spectra change AC 1000 [-]

0.03

0.02

0.01

0

-0.01

-0.02

-0.03

0

20

40

60

Component wave direction []

20

40

60

Component wave direction []

0.03

0.02

0.01

0

-0.01

-0.02

-0.03

0

20

40

60

Component wave direction []

Fig. 4.22: Change of the wave pattern characteristic due to variation of the hull panel source density

# 252, located directly beneath the wavy free surface level 40% LPP downstream of the bow.

The wave pattern resistance serves as the objective function of the hydrodynamic hull optimization.

The non-dimensional form of the wave pattern resistance, in a notation going back to Havelock, is

given by (3.30)

RW P =

Z/2

0

A2S () + AC2 () cos3 () d .

Introducing the free wave spectra from (4.85) to the above expression yields

Z/2

2

2

(0)

(0)

RW P =

AS () + AS () + AC () + AC ()

cos3 () d .

(4.87)

This expression for RW P contains square terms of the variations of the free wave spectra. In a consistent

linearization the square terms, A2S () and AC2 (), are dropped from (4.87) since they are of higher

order and small in terms of the perturbation parameter . However, as will be discussed in section 5.2,

88

prediction ~ single restart step | | | fully converged solution | |

Wave pattern resist. change CWP 1000 [-]

0.02

0.01

0

-0.01

-0.02

-0.03

0

0.002

0

-0.002

-0.004

-0.006

-0.008

-0.01

0

x / LPP [-]

Free wave spectra change AC 1000 [-]

0.6

0.4

0.2

0

-0.2

-0.4

-0.6

0

20

40

60

Component wave direction []

0.6

0.4

0.2

0

-0.2

-0.4

-0.6

20

40

60

Component wave direction []

20

40

60

Component wave direction []

Fig. 4.23: Change of the wave pattern characteristic due to variation of the hull panel source density

# 538, located directly beneath the wavy free surface level 10% LPP upstream of the stern.

it is preferred to set-up and to solve the objective functional in a least-square approach. Hence, it is

advantageous to keep expression (4.87) in its original form which inherently is of least-square type.

On closer examination of (4.87) it becomes apparent that a reduction of RW P can be achieved only by

a minimization of the free wave spectra, weighted by cos3/2 (), i.e.,

h

i2

(0)

3/2

min

AS () + AS () cos ()

h

i2

(0)

3/2

min

AC () + AC () cos ()

and

for all .

(4.88)

The objective function of optimization will be defined and derived in detail in section 5.2.

The present optimization scheme is based on gradients. The impact of the variation of the optimization

variables, i.e., the hull panel source densities, on the hull form, on the wave pattern, the free wave

spectra and the wave pattern resistance is measured in terms of gradients. Two principal types of

89

4 Sensitivity analysis

0.03

0.02

0.01

0

-0.01

-0.02

-0.03

-3

-2

-1

0

1

Source density ratio Ds68 / Ds68 max [-]

Fig. 4.24: Change of the wave pattern resistance due to variation of the single hull panel source density

# 68 (an abscissa value of +1 means that the source density variation is pushed to its upper

bound allowed in optimization).

gradients characterize the sensitivity analysis. One is related to the perturbation approach, in particular,

to the set-up of the transfer matrices

(0)

~X1

,

~n

(4.89)

and the second specifies the sensitivities of the objective function and the constraints with respect to

the optimization variables, i.e., the hull source densities

(0)

~X2

,

~

(4.90)

(0)

with ~X1,2 being place holders for any of the system state quantities.

In principle, the partial derivatives of the state variables can be evaluated either analytically or numerically. In the analytical approach the partial derivatives of the system state quantities need to be

evaluated in closed analytical form. This rapidly becomes a highly complex task since the boundary

value problem depends in a complicated nonlinear manner on the optimization variables and on the

hull panel locations. Yet the influence coefficients Ai j , Xi j , Zi j are nonlinear functions of the panel

geometry. A derivation of the analytical relations may be based on the Hess & Smith relations. 42 However, the present scheme is based on a numerical flow solution. This calls for a numerical evaluation

of the partial derivatives of the influence coefficients with respect to the hull panel normal relocations

(4.89). Therefore, the gradient (4.89) is approximated here by forward differences in a straightforward

way, see further below.

42

Kraus (1989) formulated closed analytical expressions for the gradients of the influence coefficients for 2-D flows about

NACA profiles omitting the free surface. In his work the hull variation was conducted by shifting the panel input points

in lateral direction.

90

0.5

-0.5

-1

-1.5

-3

-2

-1

0

1

Source density ratio Ds76 / Ds76 max [-]

Fig. 4.25: Change of the wave pattern resistance due to variation of hull panel source density # 76.

In their initial form the objective function and the constraints are related in complicated nonlinear

manner to the optimization variables. However, in the present scheme the cause and effect chain is

linearly interlinked since the nonlinear boundary value problem is transferred to explicit equations

depending linearly on the optimization variables. Hence, the gradients of the objective function and

the constraints with respect to the optimization variables (4.90) simply turn into the slopes of the

corresponding straight-line equations. The slopes can be determined simply by forward differences

(4.92) too.

In the limiting case when the perturbation quantities approach zero the forward differences become the

partial derivatives

(0)

(0)

(0)

~X1 ({CT}

~X

+ ~n) ~X1 ({CT}(0) ) ~X1

=

= 1

lim

~n0

~n

~n

~n

(0)

(0)

(0)

(0)

~X2 (~ + ~) ~X2(~ ) ~X2

~X

lim

=

= 2 .

~

~

~

~0

(4.91)

(4.92)

The gradients of the influence coefficients with respect to the hull panel normal relocations (4.91) are

utilized in the evaluation of the transfer matrices, (4.8) and (4.80). The gradients are approximated by

forward differences according to (4.91) by successively shifting each active hull panel by the perturbation magnitude n, as defined by (4.19), in its panel normal direction. Since the normal relocation of

a single hull panel only affects a single row and a single column in each influence coefficient matrix

a fast and memory conserving scheme was developed for the successive construction of the transfer

matrices. Applying a marching technique, the evaluation of the gradients due to each single hull panel

normal relocation only requires the computation of two rows and two columns for each influence coefficient matrix. Nevertheless, the transfer matrices and the coefficient matrix A(0) (utilized in the direct

perturbation relation (4.13)) need to be processed and temporarily stored in memory. In practical applications relatively large free surface domains are handled which require an appropriate number of

91

4 Sensitivity analysis

0.04

0.02

-0.02

-0.04

-3

-2

-1

0

1

2

Source density ratio Ds252 / Ds252 max [-]

Fig. 4.26: Change of the wave pattern resistance due to variation of hull panel source density # 252.

panels. This in turn causes large equations systems with a large coefficient matrix. To overcome the

problem the portion of the free surface domain which is considered in the evaluation of the transfer

matrices and in the solution of the direct perturbation relation (4.13) might be reduced, as outlined in

section 4.6.

The second type of gradients (4.92) characterizes the sensitivities of the objective function and the

constraints with respect to the optimization variables. These gradients constitute the essential result of

the sensitivity analysis and are directly utilized by the optimization scheme, see chapter 5. They are

evaluated at each stage of the cause and effect chain.

Figures 4.19 to 4.27 show a representative perturbation study. The gradients of the objective function,

i.e., the free wave spectra and the wave pattern resistance, are investigated by means of a numerical

study for the standard Wigley hull at Fn = 0.3. Figure 4.19 shows the optimization result 43 of the first

sub-optimization loop of an optimization process. All but the center line hull source densities were

selected as free variables in the optimization. Figures 4.20 to 4.23 show the change of the wave pattern

characteristic due to the variation of selected panel source densities along the Wigley hull, see sketch

in figure 4.18. The selected source density values were pushed in their respective optimum gradient

directions up to their individual variation bounds allowed in optimization. The variation bounds were

defined so that the hull panel normal relocations did not exceed the perturbation magnitude n as

computed from (4.19), with ~n = 0.1. Finally, in figures 4.24 to 4.27 the change of the wave pattern

resistance due to the variation of the selected source density values is represented. Each source density

variation was normalized by its variation bound (for instance, an abscissa value of +1 means that the

source density variation was pushed to its upper bound allowed in the optimization).

The perturbation relations derived in this chapter were applied to predict the gradients of the objective

function. The predicted gradients (prediction) were compared to results from direct SHIPFLOW

43

The longitudinal wave cut at the offtrack position y/LPP = 0.25 was considered for the wave cut analysis.

92

0.5

-0.5

-1

-3

-2

-1

0

1

2

Source density ratio Ds538 / Ds538 max [-]

Fig. 4.27: Change of the wave pattern resistance due to variation of hull panel source density # 538.

recomputations. 44 In the SHIPFLOW recomputations two cases were considered: a single restart step

from the converged base solution and a fully nonlinear free surface iteration until convergence. As

long as the perturbations stay in line with the perturbation bounds a good correspondence was found

for both the continuous distributions (figures 4.19 to 4.23) and the scalar wave pattern resistance values

(figures 4.24 to 4.27). The results are excellent and prove the validity of the method.

44

In this case, the SHIPFLOW runs were conducted directly with the individually relocated, gaping hull panels as provided

by the perturbation scheme. No joint panel relocation was applied.

93

4 Sensitivity analysis

94

characteristics

The developed hydrodynamic optimization scheme aims at an improvement of the wavemaking characteristics and a reduction of the wave (pattern) resistance by a favourable

variation of the ship hull form. The optimization scheme utilizes gradient information of

both the objective function and the constraints with respect to the optimization variables.

The definition and evaluation of the gradients was given in chapter 4. The present chapter

focuses on the set up and the solution of the constrained minimization problem. An outline

of the method for the adaptation of the new, improved hull form completes the chapter.

The developed optimization scheme aims at an improvement of the wave-making characteristics and

a reduction of the wave (pattern) resistance of ship hulls. As already outlined it differs from related

schemes in four substantial aspects:

Hull variations are controlled by a flow related quantity, i.e., the hull panel source densities.

Neither, are essential prerequisites to the hull representation and its characteristic features nor

to the hull variation method imposed. Moreover, the optimization variables constitute a link

between the hull geometry and their hydrodynamic properties.

The objective function of optimization is directly related to the steady ship wave systems in

terms of free wave spectra and the wave pattern resistance as determined by a longitudinal wave

cut analysis. This improves the system identification since waves can be influenced and reduced

more systematically and with higher control.

The hull optimization is carried out for the most realistic case of the advancing ship in its effective

floating position including trim and sinkage. Variations are directly applied to the wetted hull

portion under (and slightly above) the actual wavy free surface level.

The optimization process is established in terms of an iterative marching scheme of successive

sub-optimization loops which built on top of each other. At each sub-optimization stage the

boundary value problem is simplified by a linear perturbation approach around the current base

point yielding a convex quadratic image of the solution space in a trust-region, i.e., the suboptimization space around the current base point. This simplified image is known to possess a

single minimum determined by the active constraints. Thus, fast and straightforward evaluation

of the gradients of the objective function and the constraints is enabled, allowing a simultaneous

treatment of numerous, locally acting optimization variables (of order 103 ) which introduces

great freedom to the hull variation.

The present chapter focuses on the hull improvement stage, as a part of the optimization process,

compare figure 5.1. The constrained minimization problem is discussed, the optimization functional is

95

Minimization of objective

function implying constraints

(iterative scheme)

Generation of

improved hull form &

prediction of gains / changes

Hull improvement

Fig. 5.1: The hull improvement stage (detail magnification of the process flow, figure 2.1).

defined, the normal equations are derived and a special, iterative solution scheme is introduced. Finally,

the adaptation of the new, improved hull form is explained.

In this section a brief overview of constrained minimization problems is given with a focus on the

present minimization task. The reader is referred to the comprehensive literature to this topic for a

detailed survey. Following literature resources may serve as a starting point: Nowacki (1976, 1994),

Nowacki et al. (1995), Birk and Harries (2003), The NEOS Guide (2004), Mittelmann and Spellucci

(2004).

Engineering design problems can normally be formulated in the format of optimization problems. For

a mathematical treatment by an optimization strategy the real world is first described by a physical

model which then is transferred to an appropriate mathematical model. The latter usually is solved

by a suitable numerical algorithm. Most engineering optimizations have the format of constrained

minimization problems which can be generally stated as:

Find the vector of free optimization variables ~x = (x1 , x2 , , xN )T in the N-dimensional space,

bounded by

ximin xi ximax

i = 1, , N ,

!

subject to the inequality constraints

g j (~x ,~p) 0

j = 1, , NG

hk (~x ,~p) = 0

k = 1, , NH .

~x indicates the vector of free variables of the minimum design. ~p represents the vector of optimization parameters. Optimization parameters are necessary to describe the design problem but they are

kept constant during optimization. More precisely, optimization parameters are understood as passive

96

optimization variables, whereas active optimization variables are allowed to freely adjust during the

optimization. If g j (~x ,~p) 0 and hk (~x ,~p) = 0 holds true for all j and k, then the vector of free variables ~x is feasible and represents a valid design candidate. Whereas, if any of the inequality constraints

has a negative value or if just one equality constraint has a non-zero value, the design candidate is

infeasible. The set of all feasible design candidates constitutes the design space.

Engineering design problems usually are either of linear (LP) or nonlinear (NLP) programming type. A

linear programming problem exists if both the objective function and the constraints are linear functions

of the free variables. In case of an LP-problem the minimum is always governed by one or more

constraints. In nonlinear programming problems either the objective function, a constraint or both

together are nonlinear functions of the free variables. The majority of engineering design problems are

of nonlinear type. In section 5.2 the present optimization problem is discussed in this respect.

In the above formulation of the constrained minimization problem the constraints are specified explicitly. Special algorithms are required to deal with explicit constraints. Search methods for instance often

implement special algorithms which adopt a certain strategy as soon as the pattern search intends to

violate a constraint. However, constrained minimization problems can be re-written as unconstrained

problems with a modified objective function. This allows a treatment of the constrained problem by

unconstrained optimization algorithms. The modification of the objective function is done by adding

penalty functions. Penalty function methods are subdivided into interior, or barrier penalty function

methods, and exterior penalty function methods. For a detailed description of penalty function methods the reader is referred to the pertinent literature, e.g., see the cross references at The NEOS Guide

(2004) and Mittelmann and Spellucci (2004). The constrained minimization problem can be recasted

in an equivalent unconstrained form in terms of a free variational problem based on the Euler-Lagrange

approach

NH

NG

k=1

j=1

(5.1)

The objective functional (5.1) still depends on the constant optimization parameters ~p, but for the

sake of clarity they are omitted in the above representation. ~ and ~ are the vectors of a priori unknown Lagrange multipliers. Hence, the constrained minimization problem has been transferred to

an unconstrained one at the cost of introducing new additional unknowns. Frequently, the inequality

constraints are transformed beforehand to equality constraints by introducing so called slack variables

~s, i.e., g j (~x) 0 is transformed to g j (~x) s2j = 0. However, the slack variables introduce even further

unknowns to the problem.

In engineering design problems the number of constraints often becomes rather large. This is especially

true if certain positional constraints have to be maintained. Then the number of additional unknowns

in terms of necessary, but undesired Lagrange multipliers and slack variables may rapidly exceed the

actual number of free optimization variables. In this case the Lagrangian approach may be ineffective

(CPU-intensive and memory consuming). For this reason the constraints handling is performed in a

different way in the present scheme, see section 5.2.

97

First order necessary conditions

The first order necessary conditions for a local extremum of the objective functional (5.1) require that

the gradient equals the null vector, i.e., the partial first order derivatives with respect to all variables

vanish

! ~

~x F(~x ,~ ,~ ) =

0

!

~ F(~x ,~ ,~ ) = ~0

! ~

~ F(~x ,~ ,~ ) =

0.

(5.2)

As a further condition the free variables vector ~x at the extremum point is supposed to lie on the active

constraints. Equality constraints are always active. All inequality constraints which are inactive at the

extremum point vanish by j = 0 for j

/ active set and, thus, do not affect the result. The vector ~x

which satisfies the first order necessary conditions is called a stationary point.

Sufficient conditions are introduced in order to classify the extremum, i.e., whether it is a minimum,

a maximum or an inflection point. Generally, the sufficient conditions are associated with the second

order derivatives of the objective function 1 with respect to all variables. The sufficient condition for ~x

to be a minimum is fulfilled, provided the necessary conditions are satisfied, if the Hessian matrix of

the second order derivatives, including the mixed derivatives, is positive definite, i.e., all its eigenvalues

are positive. However, in practical application the proof of the sufficient conditions is an expensive task

and, hence, the sufficient conditions are often not checked explicitly.

Solution space

In the previous section the mathematical model of the constrained minimization problem was outlined.

However, for a proper choice of a suitable solution strategy the specific shape of the solution space is

of particular importance.

In cases where the objective function and the constraints can be expressed explicitly as functions of

one single free variable (univariate problem) or a small set of variables, a picture of the shape of the

solution space is often known a priori or can be simply deduced by sampling of the respective functions.

However, hydrodynamic optimization problems are of complicated nonlinear nature with several free

variables (multi-variate problem) and constraints and frequently with multiple competing objective

functions (multi-objective problem). Moreover, this class of problems usually represent multi-stage

calculation processes involving different analysis tool for the evaluation of the objective function(s)

and the constraints. Here it is difficult or even impossible to draw a picture of the solution space. Even

the representation of the objective function over a single free variable is too expensive in most cases

since the evaluation of the objective function even for a single design candidate may be very costly.

For the moment it might be assumed that the solution space is known or its shape can be approximated by a response surface. Of particular importance in the classification of optimization problems is

whether the shape of the solution space is convex or non-convex. If the solution space is convex both

the objective function and the constraints need to be convex functions of the free variables. Convex optimization problems are characterized by a solution space with a single global optimum and are called

1

i.e., the objective function and the constraints have to be twice differentiable in the whole solution space

98

uni-modal. Otherwise, if the objective function or a constraint is non-convex the optimization problem

is called multi-modal. Here the feasible space is composed of one or more local optima, one of which

being the global optimum. Hydrodynamic optimization problems fall into the latter category. In recent

time several algorithms have been invented which are particularly designed to deal with multi-modal

cases. Their aim is to find the global optimum at a justifiable expense. Multi-modal optimization tasks

are start-point sensitive since the selection of the starting point significantly affects the course the algorithm takes and the quality of the optimum it finds. Normally, in hydrodynamic hull optimization

an existing, approved parent hull form is selected as starting point. Any significant improvement of

the parent hull by means of an optimization strategy indicates a successful application, independently

of the character of the optimum reached (typically a local optimum or just an improvement is encountered).

The present optimization process, as presented in section 2.3, is established in terms of an iterative

marching scheme of single successive sub-optimization loops which built on top of each other. At

each sub-optimization stage the boundary value problem is accessed by a linear perturbation approach

around the actual state of the system, i.e., the current base point 2 , see section 4.2. Both the objective

function and the constraints are formulated as quadratic expressions and are combined in a functional

formulation in terms of a convex sum. That way, the simplified convex quadratic image of the solution

space in the sub-optimization space around the current base point is attained which is known to possess

only one single minimum 3 determined by the active constraints. The perturbation approach enables a

fast and straightforward evaluation of the gradients of the objective function and the constraints with

respect to the optimization variables in the sub-optimization space. The acceleration of the evaluation

process is of particular importance in the present scheme since the number of optimization variables in

terms of hull panel source densities, naturally, tends to be comparatively large (of order 103 ).

Each sub-optimization task only renders a sub-optimization space, i.e., a confined region of the full

solution space. From the point of view of the sub-optimization task the actual solution space appears

as a simplified cut-out which is temporarily bounded, hiding away the global picture. The upcoming

sections address the set-up and the solution of the sub-optimization problem.

In this section a modification of the general optimization functional (5.1) is introduced which is used in

the present optimization approach. First, a motivation for the modification of the functional is given.

The present optimization scheme is based on local hull variations which are controlled by a variation of

the hull panel source densities (the optimization variables), compare chapter 4. An average hull panelization for a potential flow CFD simulation comprises approximately 1000 panels. For a calculational

example it may be assumed that all hull panels are allowed to actively adjust their positions by panel

normal relocations. This implicates the following constraints:

The hull source density variations are bounded by a lower and an upper bound which are determined by the perturbation approach approximately 1000 constraints.

2

3

the actual state of the system is determined by a converged nonlinear free surface flow solution for the current hull variant

This is only valid for the approximation of the objective function in the vicinity of the current base point.

99

Each single hull panel normal relocation is forced to stay within a lower and an upper bound

which are determined by the perturbation approach and/or topological as well as geometrical

bounds approximately 1000 constraints.

Relocation of adjacent panels in different directions or with different magnitudes cause considerable large gaps in the panel mesh, unless the gaps between adjacent panels are constrained to

a permissible range approximately 3000 constraints.

The variations of the main form parameters are bounded, e.g. displacement, center of buoyancy,

water plane area, etc. 5 to 10 constraints.

All constraints are of inequality type. If they are considered in terms of a Lagrange approach, slack

variables have to be introduced to transform the inequality to equality constraints. Hence, the objective

functional F(~x,~,~s) represents a function of the Lagrange multipliers, the slack variables and the free

optimization variables which are all considered as unknowns. Applying the first order necessary conditions for a local minimum (5.2) to the above outlined example roughly results in an equation system

with 104 linearly independent equations.

Equation systems of this size are about an order of magnitude smaller then what commonly are called

large-scale linear systems. 4 However, in the present hull optimization applications the coefficient

matrix was found to be of dense structure, frequently ill-conditioned or very close to singular. Considerable problems were initially encountered when attempting to solve this equation system on a state

of the art PC for typical hull optimization applications. Both CPU and memory consumption were far

beyond justifiable limits. 5

Apparently, a different approach is required. Even though in hydrodynamic hull optimization the

objective function and the constraints are nonlinear functions of the optimization variables, in the

present approach these relations are linearized via the perturbation approach in a sub-optimization

space around the current state of the system. Hence, the actual solution space may be realized as

a linearized image which is temporarily bounded. This gives rise to applying a linear programming

algorithm, though only in the sub-optimization space around the current system state. Nevertheless,

the application of LP was discarded for the following reasons:

The integrand in the expression for the wave pattern resistance (3.30), i.e., the sine and cosine

free wave spectra to the square, inherently is of least-square type and, hence, is predestinated for

a formulation of the optimization functional in a least-square sense.

In the present scheme a perturbation approach is introduced in order to access the inverse design

problem and to accelerate the computation of the gradients of the objective function and the

constraints with respect to the optimization variables. The perturbation approach is truncated

after the linear term. However, one may later want to include higher order contributions in the

perturbation approach. Since this would prohibit the application of an LP algorithm, a different

solution scheme is therefore better suited for extensions.

Moreover, it is straightforward to combine the objective function and the constraints in an optimization functional as weighted convex sum of quadratic least-square terms. An appropriate

4

Large-scale linear systems are encountered for instance in the simulation of ocean currents or in weather simulations and

may posses 105 to 107 unknowns.

5 For the numerical solution different decomposition algorithms of the MATHEMATICA package, Wolfram (2003), like the

Cholesky decomposition or the singular value decomposition (SVD), were applied.

100

(- )

C (j 0 )

C j av

C j org

center

C j shift

DC j org

DC j max

active constraint

r

C j (x )

(+ )

DC j max

active constraint

inactive constraint

Fig. 5.2: Generalized inequality constraint bounded by a lower (-) and an upper (+) bound.

weight distribution is determined by an iteration scheme. In this way, the number of equations is

drastically reduced and a greater flexibility is introduced to the solution scheme.

Nevertheless, an important feature of the LP formulation also applies to the devised optimization approach: the solution space of the sub-optimization task is always convex, hence, the problem is unimodal with one single minimum which is determined by the active constraints.

In the present approach a modified version of the Lagrangian functional formulation (5.1) is used. All

terms related to equality constraints are dropped from the functional in favour of inequalities. Inequalities introduce freedom to the problem since boundary conditions do not need to be met exactly but

rather they have been released to a certain band, i.e., bounded by a lower and an upper bound. Nevertheless, if required, equality conditions maybe recovered by gradually shrinking the freedom (i.e.,

the band-width). On the other hand, by tangible releasing the bounds inequality constraints are effectively enabled. As another feature of the modified functional formulation the Lagrangian multipliers ~

have been exchanged by positive weights. The weights are treated as parameters, i.e., the optimization

functional is not an explicit function of the weights. The functional expression combines the objective

function and the (inequality) constraints in terms of a positively weighted convex sum of least-square

terms

NG

w0 > 0 , w j 0 .

j=1

(5.3)

w0 may be any non-zero positive scaling factor and ~w are either positive scalar weights indicating

active constraints, or are equal to zero in case of inactive constraints. A suitable weight distribution is

determined by an iteration scheme which is described in detail in section 5.3.

The objective function f (~x) in (5.3) is directly derived from the non-dimensional expression for the

wave pattern resistance (3.30), i.e.,

f (~x) := fRW P (~x) .

(5.4)

The inequality constraints g j (~x) in (5.3) are defined in general form as

(0)

(5.5)

In figure 5.2 the generalized inequality constrained with lower and upper bound is sketched. C j max

(0)

being the maximum allowed tolerance or the lower and upper bound of the j-th constraint C j (~x), C j

(0)

indicates the base value and C j shi f t the center-shift of the j-th constraint. The base value C j

of the

101

-1

-2

-1

-0.5

0

0.5

1

Generalized optimization variable x / D xmax [-]

Fig. 5.3: Characteristic of the generalized constraint function subject to different weights w1 .

current sub-optimization problem is required to be inside the bounds. An initial value or an origin C j org

of the constraint may be defined. Depending on the kind of constraint C j org can either take an initial

constant value, or it may be the average point C j av of the lower and upper bound or it may be equal to

the base value of the sub-optimization problem. For a consistent treatment of all constraints within the

solution scheme the generalized inequality constraint (5.5) is related (indicated by an over-bar) to its

bound C j max . Moreover, since the optimization functional is set up and minimized in a least-square

sense, expression (5.5) is reformulated in a suitable form by an equivalent transformation

!2

(0)

C j (~x) C j C j shi f t

0

C j max > 0 for j = 1, , NG .

(5.6)

g j (~x) := 1

C j max

The constraints terms may be alternatively used in terms of penalty functions, since this terminology

better suits their actual meaning, as the following discussion of the weight distribution suggests. In

the functional expression (5.3) the objective function and the terms accounting for the (inequality)

constraints are weighted against each other. Inactive constraints are dropped from the functional by

vanishing weights. High weights for the active constraints generate a stiff problem with less freedom

for the variation, where the constraints dominate the problem at the cost of a poor optimization result.

On the other hand, small weights generate a loose problem with much leeway for the minimization but

with a tendency of constraints violation. Active constraints directly lie on a bound or are outside the

bounds

C j (~x) C(0) C j shi f t

j

C j max > 0 for j active set .

(5.7)

1

C j max

C j (~x) C(0) C j shi f t

j

C j max > 0 for j

/ active set ,

<1

C j max

102

(5.8)

compare figure 5.2.

The influence of different weights on the characteristic of the constraint function is shown in figure 5.3.

In the simplified example it is adopted that the generalized optimization variable x itself is constraint

by a lower (1) and an upper (+1) bound with the base point and the origin located at the average

point. Hence, the generalized constraint can be expressed here simply as

#

"

2

x

w1 g1 (x) = w1 g1 (x) = w1

1

w1 0 , xmax > 0 .

xmax

Figure 5.4 illustrates the principle of application of the constraints weights within the optimization

functional. For this purpose the above example was extended by an objective term so that it features a

single objective and a single constraint term which are joined together in the generalized optimization

functional F(x) = w0 f (x) + w1 g1 (x). Since the same generalized constraint, as introduced above, is

utilized the permissible variation range is limited to 1 x/ xmax +1. For the objective f (x) term

a simple function is taken for illustration which depicts, like the wave pattern resistance, a quadratic

function in the optimization variable

"

#

2

x

w0 f (x) = w0

2 +1

w0 > 0 , xmax > 0 .

xmax

From an observation of the objective f (x) function in figure 5.4 it is apparent that the global minimum

is expected at the active upper constraint bound. But how is the bound detected by means of a de facto

unconstrained functional approach? This task is accomplished by the constraint term which works,

in this respect, in the same way as a penalty function. The amount of penalty is controlled by the

associated weight w1 . Figure 5.4 reveals how different weights w1 of the constraint term are utilized to

change the characteristic of the functional expression F(x). A too small weight (w1 = 0.5) causes an

infeasible minimum of the optimization functional at x/ xmax = +4/3, since it violates the constraint

1 x/ xmax +1, compare the lower picture in figure 5.4. On the other hand a too large weight

(w1 = 2) shifts the minimum of F(x) inside the feasible domain at x/ xmax = +2/3, as the upper

picture in figure 5.4 illustrates. Although this solution is feasible, yet it is not the seeked minimum

of the objective function f (x). Apparently, the true minimum is located at the active upper constraint

bound x/ xmax = +1 which is correctly detected by a proper selection of the weight (w1 = 1), see

middle picture in figure 5.4. This simple example reveals that the appropriate choice of the weights

is of particular importance in the present optimization scheme. Therefore, in section 5.3 an iteration

scheme is introduced for the iterative adaptation of the weight distribution. The scheme keeps track of

the active set of constraints.

Based on the preceding discussion the generalized optimization functional, finally, is obtained by introducing (5.4) and (5.6) to the functional formulation (5.3) which yields

NG

j=1

(0)

C j (~x) C j C j shi f t

C j max

!2

The formulation (5.9) combines the objective function and the inequality constraints in terms of a

positively weighted convex sum of quadratic least-square terms. That way a convex solution space

(of the sub-optimization task) is constituted which is characterized by a single minimum (uni-modal

103

12

10

8

6

4

Minimum

2

0

-2

-1

-0.5

0

0.5

1

1.5

Generalized optimization variable x / D xmax [-]

12

10

8

6

4

2

Minimum

0

-2

-1

-0.5

0

0.5

1

1.5

Generalized optimization variable x / D xmax [-]

12

10

8

6

4

2

Minimum

0

-2

-1

-0.5

0

0.5

1

1.5

Generalized optimization variable x / D xmax [-]

Fig. 5.4: Influence of the weight w1 on the characteristic of an optimization functional by means of an

example featuring a single objective w0 f (x) and a single constraint term w1 g1 (x).

104

problem). Moreover, the minimum is determined by the active constraints. This suggests a scheme

which keeps track of the active set of constraints (and free variables).

The first order necessary conditions for a local minimum of the optimization functional (5.9) require

that the gradient equals the null vector, i.e., the partial first order derivatives with respect to the optimization variables vanish

! ~

~x F(~x ) =

0.

(5.10)

The optimization functional (5.9) is of type of a linear least-squares problem. Hence, the application of

the first order necessary conditions (5.10) results in the so called normal equations of the linear leastsquares problem, see e.g. Press et al. (1992) and Bjrck (1996). In the upcoming sections the normal

equations are derived for the objective function and the constraints expressions.

The wave pattern resistance RW P , in the non-dimensional form (3.30), serves as the objective function

of the hydrodynamic hull optimization. For instance, a normalization according to the wave pattern

resistance coefficient (3.31)

CW P =

RW P

,

0.5 Sre f k02

is unsuited as objective function since it is normalized by the wetted hull surface area Sre f which may

considerably vary in the optimization process. 6 Likewise, the ratio of wave pattern resistance to the

displacement (buoyant) force according to

RW P RW P

=

,

g

is not used as objective function in the present implementation. In the opinion of the author a practical

optimization task is the optimization of the wave-making characteristics for a specified transportation

capacity. Naturally, wave-making and wave (pattern) resistance are crucial in case of medium speed

and fast volume critical 7 cargo ships like container carriers or ferries. However, for this type of ships

transportation capacity is difficult to measure correctly in terms of displacement. Hence, the resistancedisplacement ratio in this case does not provide an appropriate measure of merit. Moreover, from a

practical implementation point of view the resistance-displacement ratio is not directly usable in the

above fractional form since both the numerator and the denominator are functions of the optimization

variables.

Hence, a different strategy is followed. The wave pattern resistance is applied as the objective function

of optimization in the non-dimensional form of (3.30)

Z/2

RW P =

A2S () + AC2 () cos3 () d ,

0

6

7

unless CW P is related to a fixed wetted hull surface area, e.g., those of the initial hull

opposite to weight critical cargo ships, e.g. bulker

105

while the variations of selected main form parameters, like the displacement, are kept within bounds

by additionally defined inequality constraints, see section 5.2.2.

The free wave spectra AS and AC in (3.30) are derived in the previous chapter according to equation

(4.85) as linearized functions of the perturbation quantities. The free wave spectra can be expressed

(1)

explicitly in terms of the optimization variables by expanding AS and AC in all free variables n . This

(0)

(0)

yields the tangent plane in hyperspace attached to the basic point AS and AC

N

(0)

AS () = AS () +

n=1

N

(0)

AC () = AC () +

(0)

(1) AS ()

(0)

(1) AC ()

n=1

+ O(2 ) ,

+ O(2 ) .

(5.11)

The partial derivatives of the free wave spectra in (5.11) are evaluated numerically by forward differences as discussed in section 4.8. Introducing the expansions (5.11) to the wave pattern resistance

integral (3.30) results in 8

!2

!2

Z/2

(0)

(0)

N

N

A

()

A

()

(0)

(1)

(0)

(1)

S

C

cos3 () d . (5.12)

RW P = AS () + n

+ AC () + n

n

n

n=1

n=1

0

In view of equation (5.4), the objective function of optimization can be expressed directly in terms of

the wave pattern resistance (5.12) with

(1)

fRW P (~ ) =

Z/2"

(0)

AS () +

(0)

(1) AS ()

n

n

n=1

Z/2"

(0)

AC () +

(0)

(1) AC ()

n

n

n=1

cos

3/2

cos

3/2

#2

d +

#2

d ,

()

()

(5.13)

Before applying the first order necessary conditions to the objective function, the continuous integral

over the component wave directions in (5.13) is restated for numerical treatment. Since the free wave

spectra are evaluated numerically by a longitudinal wave cut analysis, see section 3.2, they are already

sampled at discrete -values in the range 0 < /2. According to section 3.2 the continuous integral

over is evaluated numerically by means of the trapezoidal rule. This yields for the objective function

(5.13)

fRWP (~(1) ) =

NM

wem

m=0

8

(0)

(0)

(1) AS m

A S m + n

n

n=1

106

!2

NM

wem

m=0

(0)

(0)

(1) AC m

AC m + n

n

n=1

!2

, (5.14)

with the numerical integration factors w

em defined as 9

(

0.5 : for m = 0 and NM

3

w

em = cos (m )

.

1 : otherwise

In the majority of optimization applications the entire wave (pattern) resistance is minimized. Occasionally only finite components of the steady ship wave system are the objective of optimization. For

instance, one aims at a reduction of the transverse waves which may generate an adverse wake flow

for propeller and rudder operation. Furthermore, fast and high-speed vessels may be prevented from

emitting steep divergent waves which cause adverse effects in the far field. In such particular cases it

may be advantageous to focus on a specific wave range, for instance the transverse or divergent waves

alone. In the present optimization approach this is achieved in a straightforward way by simply limiting

the evaluation of the objective function (5.14) to the desired wave range, i.e., the integration in (5.13)

or (5.14) is carried out solely for the particular interesting component wave directions: min max

rather then 0 /2. 10

The first order necessary conditions for a minimum require that the partial derivatives of the optimization functional with respect to the optimization variables vanish. According to (5.10) the partial

derivatives of the objective function (5.14) with respect to the optimization variables yields

fRW P

(1)

NM

= 2

m=0

NM

w

em

wem

m=0

(0)

AS m +

(0)

(1) AS m

n

n

n=1

(0)

AC m +

(0)

(1) AC m

n

n

n=1

N

(0)

AS m

+

l

(0)

AC m

l

l = 1, , N .

(5.15)

AS

(0)

(0)

AS nm

(0)

A

= Sm ,

n

AC

(0)

(0)

AC nm

(0)

A

= Cm

n

n = 1, , N

,

m = 0, , NM

em into a diagonal matrix (applying the identity matrix)

w

e0 0 0

e1 0

0 w

f

emm =

W W

m = 0, , NM .

.. . .

..

..

. .

.

.

0

w

em

Then, after interchanging the summation order, equation (5.15) can be written in compact matrix notation

(0)

~ (1) fR = AS (0) f

W AS (0) T ~(1) + AS (0) f

W ~AS +

WP

9

10

(0)

AC (0) f

W AC (0) T ~(1) + AC (0) f

W ~AC ,

(5.16)

Note: The wave pattern resistance, the free wave spectra and the gradients are still determined for the whole wave directions

range.

107

as a system of N equations. The constant factor 2 in (5.15) might be dropped since it is considered

implicitly by the leading weighting factor w0 of the objective function, compare (5.3).

Expression (5.16) is the normal equation form of the objective function. In the upcoming section a

similar derivation is carried out for the penalty functions.

The optimization functional (5.9) is composed of the objective function and the terms accounting for

the inequality constraints. The constraints are treated via penalty functions which introduce a certain

penalty to the objective function as soon as a constraint becomes active. The constraints or penalty

terms are formulated in a least-squares sense related to their respective maximum allowed bounds.

According to (5.6) the general form of a constraint term is expressed by

(0)

gj (~(1) ) = g j (~(1) ) =

C j (~(1) ) C j C j shi f t

C j max

!2

In the previous chapter the constraints terms were linearized by means of a perturbation approach. The

constraints can be expressed explicitly in terms of the optimization variables by expanding the C j in all

(1)

(0)

free variables n . This yields the tangent plane in hyperspace attached to the basic point C j

(0)

Cj = Cj +

(0)

(1) C j

n

n

n=1

N

+ O(2 )

j = 1, , NG .

(5.18)

The partial derivatives of the constraints functions in (5.18) are evaluated numerically by forward

differences as discussed in section 4.8. If the expansion (5.18) is introduced to (5.17) the general form

of the constraint term is obtained 11

gj (~(1) ) =

(0)

(1) C j

C j max 2

n=1

C j shi f t

!2

j = 1, , NG .

(5.19)

(0)

Since the current basic value C j of the constraint function is not necessarily identical to the average

point C j av of the lower and upper constraint bound, a center point shift C j shi f t is introduced, see figure 5.2. Depending on the type of constraint function a center point shift may be required or not as will

be discussed below.

According to (5.10) the partial derivatives of the constraints terms (5.19) with respect to the optimization variables yield

gj

(1)

l

11

C j max

(0)

(1) C j

n

n

n=1

N

2

2

C j shi f t

(0)

C j

l = 1, , N

.

j = 1, , NG

(5.20)

108

Moreover, according to the optimization functional (5.9) the sum of all constraints terms multiplied

by their respective weights is to be composed. The summation of (5.20) for all constraints terms NG

results in

! (0)

(0)

NG

NG

N

C j

gj

wj

(1) C j

l = 1, , N ,

(5.21)

w j (1) = C 2 n n C j shi f t l

j max

n=1

j=1

j=1

l

whereas the multiplication factor 2 in (5.20) is considered implicitly by the weights w j . Equation

(5.21) can be written in compact matrix notation by introducing the abbreviations 12

(0)

(0) C (0) =

C

nj

C j

1

,

C j max n

C j shi f t

~

C

shi f t C j shi f t =

C j max

n = 1, , N

,

j = 1, , NG