University of Strathclyde

Department of Naval Architecture, Ocean and Marine

Engineering

Analysis of ship performance in deep and shallow

waters using CFD

Momchil Terziev

A thesis presented in fulfilment of the requirements for the degree of Doctor of

Philosophy

2020
This thesis is the result of the author's original research. It has been composed by the
author and has not been previously submitted for examination which has led to the
award of a degree.

The copyright belongs to the author under the terms of the United Kingdom
Copyright Acts as qualified by University of Strathclyde Regulation 3.50. Due
acknowledgement must always be made of the use of any material contained in, or
derived from, this thesis.

Signed: Date:
 Acknowledgements

It is with great joy that I write this section because it not only represents the
culmination of my studies at the University of Strathclyde, but also because I
am reminded of the many individuals who have tremendously enriched my
life during this time.

I would like to begin by expressing my deepest gratitude to my supervisors,

Professor Atilla Incecik, and Dr Tahsin Tezdogan. Without a doubt, I am
indebted to your mentoring, encouragement, and help, without which I would
never have imagined reaching the stage where I am ready to submit my thesis.
I whole-heartedly appreciate the confidence you placed in me during this time.
To this day I continue admiring your professionalism, which will endure in
being an example to me. This has definitely enriched my personal worldview.
Finally, I would like to thank you for giving me opportunities for independent
work, and for including me in a wide range of activities both inside and
outside the department. These have definitely helped my development.
Undoubtedly, I have learned tremendously many things from you.

I gratefully acknowledge for the funding provided by the Faculty of

Engineering at the University of Strathclyde, which fully supported and made
my PhD studies possible.

I would also like to thank Dr Zhiming Yuan, for always having the time to
discuss my project with enthusiasm, and engaging constructively with me.

My thanks go to all administrative and academic staff at the Department. I

would like to specifically thank the postgraduate research administrator,
Susan Pawson, and the IT officer, Ross Gilmour, for always being extremely
helpful.

My thanks also go out to my friends from the postgraduate research centre at

the department, who have made my time here very enjoyable. I wish the best

i
of luck to Guangwei Zhao and Mohammed Baba Shehu in pursuing their
PhDs. Similarly, I would like to thank the members of the IEEE-OES student
branch at the Department, who occasionally provided a very welcome
distraction from my daily routine. I am in no doubt that they will continue to
run the branch successfully.

I would like to thank Dr Khaled Elsherbiny for his friendship. I would also like
to thank him for his hospitality upon my visit of his home country of Egypt.
This is not something I will ever forget! I sincerely hope we stay in touch, and
wish you a successful career at the Arab Academy for Science, Technology,
and Maritime Transport.

Additionally, I would like to thank Dilyana Hristova for proofreading part of

this thesis, and Zlatina Plamenova for helping me with life in Glasgow – I
appreciate your friendship very much, and I am sure both of you will have
brilliant successes in whatever endeavour you embark on.

My thanks also go to James Nicolson for his continuous support in all aspects
during the last several years. I have no doubt that without your patience, help,
and motivation, my life in Glasgow would not have been anywhere near as
enjoyable as it has been.

My siblings also deserve thanks for their support and for always being willing
to listen to me talk about my studies. I am very grateful to my brother for
devoting time to help me proofread this thesis. I would also like to wish my
sister the great success she deserves in her studies.

Finally, I would like to thank my parents for their support throughout my

studies. I am privileged to have been raised with a love for continuous
learning, which has undoubtedly helped me reach this moment. Thank you
for your enduring patience.

ii
 TABLE OF CONTENTS

Abstract .................................................................................................................... xvi

Introduction ....................................................................................................... 1
1.1 Background ................................................................................................. 1
1.2 Motivating factors ...................................................................................... 2
1.3 Research aims and objectives.................................................................... 5
1.4 Thesis structure ........................................................................................... 5
Critical Review .................................................................................................. 8
2.1 The origins of ship hydrodynamics ......................................................... 8
2.2 Contemporary perspectives ...................................................................... 9
2.3 Scale effects on a ship hull ....................................................................... 12
2.3.1 Shallow water effects ........................................................................ 21
2.4 Ship waves ................................................................................................. 24
2.5 Summary and conclusions ...................................................................... 28
A Geosim Analysis of Ship Resistance Decomposition and Scale Effects
with the aid of CFD ................................................................................................. 30
3.1 Introduction ............................................................................................... 30
3.2 Methodology ............................................................................................. 32
3.3 Ship geometry and conditions ................................................................ 34
3.4 Numerical set-up ...................................................................................... 36
3.4.1 Physics modelling ............................................................................. 36

iii
 3.4.2 Time step selection ............................................................................ 38
3.4.3 Mesh generation ................................................................................ 38
3.4.1 Computational domain and boundaries........................................ 39
3.4.2 Time-history of the numerical solution ......................................... 40
3.4.3 Verification study .............................................................................. 44
3.5 Results and discussion ............................................................................. 46
3.5.1 Error evaluation................................................................................. 46
3.5.2 Resistance decomposition ................................................................ 49
3.6 Summary and conclusion ........................................................................ 60
Application of Eddy-Viscosity Turbulence Models to Problems in Ship
Hydrodynamics ....................................................................................................... 62
4.1 Introduction ............................................................................................... 62
4.2 Background ............................................................................................... 65
4.2.1 Turbulence.......................................................................................... 66
4.3 Case-studies ............................................................................................... 75
4.4 Numerical implementation ..................................................................... 77
4.4.1 Physics modelling ............................................................................. 78
4.4.2 Time step selection ............................................................................ 79
4.4.3 Computational domain .................................................................... 80
4.4.4 Mesh generation ................................................................................ 81
4.5 Results and discussion ............................................................................. 83
4.5.1 Comparison against experimental data ......................................... 83
4.5.2 Numerical verification ...................................................................... 86
4.5.3 Results comparison ........................................................................... 88
4.6 Summary and conclusion ...................................................................... 103
A Numerical Assessment of the Scale Effects of a Ship Advancing
Through Restricted Waters .................................................................................. 105
5.1 Introduction ............................................................................................. 105
5.2 Methodology ........................................................................................... 107
5.2.1 Approach to the problem at hand................................................. 107

iv
 5.2.1 Numerical implementation............................................................ 109
5.3 Results and discussion ........................................................................... 112
5.3.1 Verification study ............................................................................ 112
5.3.2 Numerical results ............................................................................ 114
5.4 Summary and conclusion ...................................................................... 124
A Posteriori Error and Uncertainty Estimation in Computational Ship
Hydrodynamics ..................................................................................................... 126
6.1 Introduction ............................................................................................. 126
6.2 Background ............................................................................................. 128
6.2.1 Discretisation error.......................................................................... 129
6.2.2 Local error and uncertainty ........................................................... 134
6.3 Case-studies and numerical set-up ...................................................... 136
6.3.1 Spatial characteristics of the computational domain ................. 137
6.3.2 Physics modelling ........................................................................... 140
6.3.3 Time step selection and temporal discretisation ........................ 142
6.4 Results and discussion ........................................................................... 142
6.4.1 Local classification .......................................................................... 145
6.4.2 Error analysis and decomposition ................................................ 147
6.4.3 Error behaviour considerations..................................................... 154
6.4.4 Local uncertainty and extrapolated solutions ............................. 159
6.4.5 Effect on sampling density............................................................. 169
6.5 Conclusion and recommendations for future work .......................... 172
Virtual Replica of a Towing Tank Experiment to Determine the Kelvin
Half-Angle of a Ship in Restricted Water .......................................................... 175
7.1 Introduction ............................................................................................. 175
7.2 Case studies ............................................................................................. 179
7.3 Methodology ........................................................................................... 181
7.3.1 Numerical aspects ........................................................................... 182
7.3.2 Spectral representation of the wave field .................................... 190
7.4 Results and discussion ........................................................................... 193

v
 7.4.1 Ship resistance ................................................................................. 193
7.4.2 Spectral analysis of the numerical free surface ........................... 195
7.5 Summary and conclusion ...................................................................... 206
Modelling the Hydrodynamic Effect of Abrupt Water Depth Changes on
a Ship Travelling in Restricted Waters Using CFD .......................................... 208
8.1 Introduction ............................................................................................. 208
8.2 Case studies ............................................................................................. 210
8.3 Numerical implementation ................................................................... 213
8.3.1 The numerical environment .......................................................... 213
8.3.2 Computational domain and boundary conditions..................... 214
8.3.3 Computational mesh ...................................................................... 215
8.3.4 Time-step selection.......................................................................... 216
8.3.5 Numerical verification .................................................................... 217
8.4 Results and discussion ........................................................................... 219
8.5 Conclusion and summary ..................................................................... 237
Conclusions and Future Work .................................................................... 239
9.1 Introduction ............................................................................................. 239
9.2 Conclusions ............................................................................................. 239
9.3 Discussion ................................................................................................ 243
9.4 Future work ............................................................................................. 244
References ............................................................................................................... 246
Publications ............................................................................................................ 285
Appendix A ............................................................................................................ 287
Appendix B ............................................................................................................ 289
Appendix C ............................................................................................................ 291
C.1 Spalart-Allmaras model............................................................................. 292
C.2 The 𝒌 − 𝜺 family of turbulence models ................................................... 293
The 𝒌 − 𝝎 family of turbulence models ......................................................... 294

vi
vii
 LIST OF FIGURES

Figure 2.1. Hierarchy of fluid flow models, adapted from Witherden and
Jameson (2017). ........................................................................................................ 11
Figure 2.2. Resistance extrapolation ..................................................................... 14
Figure 2.3. Annual publications listed under the category “ship CFD”
according to Web of Science (2020) as of 26.04.2020. Extrapolated numbers
based on the trend up to 2030 are represented by empty bars. ........................ 18
Figure 3.1. Domain boundary conditions and dimensions for the typical
multiphase simulations .......................................................................................... 39
Figure 3.2. 3-D view of the generated mesh. Depicted: Full-scale ................... 40
Figure 3.3. Example time-history of resistance and sinkage for case λ=31.599
.................................................................................................................................... 41
Figure 3.4. Residual time-history. Depicted: λ=31.599....................................... 43
Figure 3.5. Wave resistance coefficients. .............................................................. 50
Figure 3.6. Wave cuts at y/L=0.1, 0.2, 0.3 for all examined scale factors. The ship
is located at 0 < x/L <1, whereas the flow is in the negative x direction. ......... 51
Figure 3.7. Relative locations of the wave cuts. Depicted: λ=75. ...................... 51
Figure 3.8. Frictional resistance coefficients ........................................................ 55
Figure 3.9. Calculated form factors ....................................................................... 56
Figure 3.10. Extrapolated total resistance coefficient ......................................... 58
Figure 4.1. Graphical representation of the use of turbulence models in marine
hydrodynamics. ....................................................................................................... 75
Figure 4.2. Graphical depiction of the adopted case-studies (not drawn to
scale): (a) refers to the study of Elsherbiny et al. (2019), (b) and (c) refer to
Mucha and el Moctar (2014). ................................................................................. 77

viii
Figure 4.3. Boundary conditions and dimensions for all computational
domains. ................................................................................................................... 81
Figure 4.4. Sample y+ value distribution along the wetted hull: (a) refers to the
study of Elsherbiny et al. (2019), (b) refers to Mucha and el Moctar (2014) – case
2. ................................................................................................................................. 81
Figure 4.5. 3-D view of the computational mesh. Depicted: case 1 (mirrored
using the central symmetry plane). ...................................................................... 82
Figure 4.6. Resistance coefficient comparison for case 1 ................................... 89
Figure 4.7. Resistance coefficient comparison for case 2. .................................. 90
Figure 4.8. Resistance coefficient comparison for case 3. .................................. 90
Figure 4.9. Resistance coefficient comparison for case 4. .................................. 91
Figure 4.10. Free surface profile along the hull for all case-studies. ................ 92
Figure 4.11. Skin friction comparison. Shapes connected via dotted lines
correspond to cases 1 and 2. Solitary coloured shapes correspond to case 3,
while the black shapes – case 4. ............................................................................ 94
Figure 4.12. Boundary layer and wake (top view), depicted: case 4. Top: iso-
surfaces for 0.9×𝑈. Bottom: iso-surfaces for 0.99×𝑈. ........................................... 95
Figure 4.13. Bivariate error plot for resistance. ................................................... 97
Figure 4.14. Bivariate error plot: sinkage vs. resistance. .................................... 98
Figure 4.15. Distances in the Error-Error plane and RMS error for case-studies
1-3. ........................................................................................................................... 100
Figure 4.16. Time per iteration for case 1. .......................................................... 101
Figure 5.1. Domain characteristics and boundary conditions ........................ 110
Figure 5.2. Full-scale mesh generated in Star-CCM+ ....................................... 110
Figure 5.3. Skin friction coefficients calculated at each scale and established
friction lines, used to demonstrate the relative difference between the shallow
water line of Zeng et al. (2019) compared to other predictions. ..................... 114
Figure 5.4. Predicted boundary layer thickness at different scales. ............... 117
Figure 5.5. Comparison of the wave elevation on the ship hull. .................... 118
Figure 5.6. Effect of viscosity and turbulence on near and far field waves
according to Brard (1970). .................................................................................... 119
Figure 5.7. Predicted wave resistance coefficients. .......................................... 121
Figure 5.8. Predicted form factors. ...................................................................... 122
Figure 5.9. Velocity distribution along a line at the aft perpendicular, z=0.5T.
Depicted: multiphase simulations, λ=75 and λ=1. ............................................ 123
Figure 6.1. Domain dimensions and boundary conditions ............................. 137
Figure 6.2. Resulting mesh ................................................................................... 139

ix
Figure 6.3. Fine, medium, coarse solutions, and convergence ratio of the free
surface ..................................................................................................................... 143
Figure 6.4. Skin friction nodes from the ship hull ............................................ 145
Figure 6.5. Observed local order of accuracy on the free surface nodes. ...... 146
Figure 6.6. Local observed order of accuracy on the ship hull ....................... 147
Figure 6.7. Free surface error constants. ............................................................ 148
Figure 6.8. Skin friction error constants. ............................................................ 149
Figure 6.9. Spatial distribution of the linear and quadratic terms of the error in
the free surface, computed via Eq. (6.11) and Eq. (6.12). ................................. 151
Figure 6.10. Spatial distribution of the linear and quadratic terms of the error
in the skin friction, computed via Eq. (6.11) and Eq. (6.12). ............................ 153
Figure 6.11. Cell aspect ratio and skewness angle on ship hull: fine solution
.................................................................................................................................. 154
Figure 6.12. Example of error reinforcement and attenuation ....................... 154
Figure 6.13. Error constants vs. convergence ratio ........................................... 157
Figure 6.14. Error constants vs. local observed local of accuracy .................. 159
Figure 6.15. Computed Factors of Safety using Eq. (6.17). .............................. 160
Figure 6.16. Local uncertainty spatial distribution on the free surface according
to all methods. ....................................................................................................... 161
Figure 6.17. Local uncertainty spatial distribution skin friction on the ship hull
according to all methods ...................................................................................... 164
Figure 6.18. Skin friction with error bars. .......................................................... 165
Figure 6.19. Extrapolated free surfaces .............................................................. 166
Figure 6.20. Extrapolated skin friction distribution on the hull. .................... 168
Figure 6.21. Error induced by sample size as predicted by Eq. (6.21). .......... 170
Figure 6.22. Sampling of the free surface. Each sampling, featuring less than
100% of all points is repeated 25×104 times. ..................................................... 171
Figure 6.23. Influence of sample size. ................................................................. 172
Figure 7.1. Kelvin half-angle of ship-generated waves in shallow waters as a
function of the depth Froude number. ............................................................... 180
Figure 7.2. Graphical depiction of the cross-section of the selected case-studies.
Top: New Suez Canal, bottom: rectangular canal. ........................................... 181
Figure 7.3. Length of the computational domain and dimensions of the overset
domain. ................................................................................................................... 184
Figure 7.4. 3D depiction of the computational mesh. ...................................... 185
Figure 7.5. The y+ distribution of the ship hull at 𝐹ℎ =0.77, sampled at 40s
physical time. ......................................................................................................... 186

x
Figure 7.6. Example time-history of the solution (depicted: 𝐹ℎ=0.57). .......... 188
Figure 7.7. Solutions to Eq. (7.8). Figure depicts the examined loci, alongside
the critical depth Froude number (𝐹ℎ=1) to demonstrate the effect of speed.
.................................................................................................................................. 192
Figure 7.8. Comparison of total resistance coefficients for all cases (R indicates
the rectangular canal case, whereas S – the Suez Canal). Subscripts refer to
depth Froude number........................................................................................... 195
Figure 7.9. Generated wave field in the rectangular canal at 𝐹ℎ=0.77........... 196
Figure 7.10. Generated wave field in the rectangular canal at 𝐹ℎ=0.77 and
corresponding half angle according to Havelock (1908). Dashed line originates
at FP, solid line originates at the nearest downstream peak, where the dashed
line is reflected. Broken line originates at the highest wave elevation on the
wall, dotted line originates at the ship coordinates representing the point
where the aft wave system is generated. ........................................................... 196
Figure 7.11. Processing of the wave field. Depicted: 𝐹ℎ=0.77 in the rectangular
canal. Top: the raw image – real space extents are 32m in the stream wise and
4.6m in the span wise directions. Bottom left: the Fourier representation of the
wave field. Bottom right: detected maxima and fit, superimposed onto the
theoretical relationship, Eq. (7.8). ....................................................................... 197
Figure 7.12. Derivatives 𝑑𝑘𝑦/𝑑𝑘𝑥 for deep water, shallow water, and
numerical shallow water cases. ........................................................................... 199
Figure 7.13. Predicted and theoretical half-angles. .......................................... 200
Figure 7.14. Splitting of near and far field components via manipulations of
the spectrum (cut-off wave number 𝑘𝑥𝑐= 4.7885). Top: original free surface, (a)
indicates the far field component, whereas (b) indicates the near-field
disturbance and their corresponding Fourier representations. Longitudinal
extent: 32 m. ........................................................................................................... 201
Figure 7.15. Computed free surface in the rectangular canal, 𝐹ℎ=0.57. (a), far
field (b), and near field (c) representations in the real and spectral space
(𝑘𝑥𝑐=9.5796). Longitudinal extent: 16.5 m. ........................................................ 202
Figure 7.16. Computed free surface in the Suez Canal, 𝐹ℎ=0.57. (a), far field (b),
and near field (c) representations in the real and spectral space (𝑘𝑥𝑐=9.5796).
Longitudinal extent: 13 m. ................................................................................... 203
Figure 7.17. Computed free surface in the Suez Canal, 𝐹ℎ=0.47. (a), far field (b),
and near field (c) representations in the real and spectral space (𝑘𝑥𝑐= 14.1833).
Longitudinal extent: 20 m. ................................................................................... 204

xi
Figure 7.18. Computed free surface in the Suez Canal, 𝐹ℎ=0.47. (a), far field (b),
and near field (c) representations in the real and spectral space (𝑘𝑥𝑐= 14.1833).
Longitudinal extent: 20 m. ................................................................................... 205
Figure 8.1. Schematic drawing of the step changes in water depth. Not drawn
to scale. .................................................................................................................... 212
Figure 8.2. Depiction of the computational domain (depicted: ℎ𝑖/ℎ𝑠=1.687).
.................................................................................................................................. 214
Figure 8.3. Close-up of the computational mesh on the free surface. ........... 216
Figure 8.4. Resistance increase resulting from the depth change. ................. 219
Figure 8.5. Wavefield for case 1, 𝐹ℎ𝑖 =0.77 and 𝐹ℎ𝑠 =1. The solution time and
increment interval unit at which the free surface is shown is based on the end
of the acceleration phase (shown in the first tile). The dashed line indicates the
position of the step change in water depth. ...................................................... 220
Figure 8.6. Wavefield for case 7, 𝐹ℎ𝑖 =0.9 and 𝐹ℎ𝑠 = 1.03. The solution time
and increment interval unit at which the free surface is shown is based on the
end of the acceleration phase (shown in the first tile). The dashed line indicates
the position of the step change in water depth. ................................................ 221
Figure 8.7. Wave probes at the step for 𝐹ℎ𝑖 =0.77............................................ 222
Figure 8.8. Wave probes at the step for 𝐹ℎ𝑖 =0.9.............................................. 223
Figure 8.9. Wavecut 1 (y/w=0.1) evolution for case 𝐹ℎ𝑖 =0.77, 𝐹ℎ𝑠 =1, made
dimensionless by the initial depth ℎ𝑖 =0.32. Maxima and minima are marked
with green and red points, respectively............................................................. 226
Figure 8.10. Wavecut 2 (y/w=0.2) evolution for case 𝐹ℎ𝑖 =0.77, 𝐹ℎ𝑠 =1, made
dimensionless by the initial depth ℎ𝑖 =0.32. Maxima and minima are marked
with green and red points, respectively............................................................. 227
Figure 8.11. Transition and reflection coefficients for cases 1-4 (𝐹ℎ𝑖 =0.77). 228
Figure 8.12. Analysis of the wavefield for 𝐹ℎ𝑖 =0.77, 𝐹ℎ𝑠 =1 along wavecuts
y/w=0.1 and y/w=0.2 .............................................................................................. 229
Figure 8.13. Predicted wavenumber (𝑘𝑖) for each depth Froude number .... 230
Figure 8.14. Wavenumber predictions for the region past the step. Depicted:
case 1, 𝐹ℎ𝑖 =0.77 and 𝐹ℎ𝑠 =1. .............................................................................. 231
Figure 8.15. Transmission and reflection coefficients based on Marshall and
Naghdi's (1990) method (𝐹ℎ𝑖 =0.77). ................................................................. 232
Figure 8.16. Viscous dissipation on a unit wave with different dispersive
properties travelling a unit distance. .................................................................. 234
Figure 8.17. Generated velocity field for case 1, 𝐹ℎ𝑖 =0.77 and 𝐹ℎ𝑠 =1 as the
ship and soliton interact with the step. .............................................................. 235

xii
Figure 8.18. Example velocity field near the step at the end of the simulation
for case 1: 𝐹ℎ𝑖 =0.77, 𝐹ℎ𝑠 =1. .............................................................................. 236

xiii
 LIST OF TABLES

Table 3.1. KCS Principal characteristics and case-studies ................................. 35

Table 3.2. Number of cells in each scale ............................................................... 40
Table 3.3. Grid convergence for trim and total resistance coefficient .............. 45
Table 3.4. Time step convergence for trim and total resistance coefficient .... 45
Table 3.5. Numerical and experimental total resistance coefficients, Fr = 0.26
.................................................................................................................................... 48
Table 3.6. Numerical and experimental trim, Fr = 0.26 ...................................... 48
Table 3.7. Numerical and experimental sinkage/length, Fr = 0.26 ................... 49
Table 4.1. Principal characteristics of the KCS .................................................... 76
Table 4.2. Case-studies description....................................................................... 76
Table 4.3. Summary of the tested turbulence models. ....................................... 78
Table 4.4. Computational domain cell properties. ............................................. 83
Table 4.5. Numerical results and error calculations ........................................... 84
Table 4.6. Convergence study for sinkage, trim, and resistance for case 1, λ=75,
h/T=2.2, 𝐹ℎ=0.303, EB model. ................................................................................. 88
Table 5.1. Case-studies.......................................................................................... 108
Table 5.2. Cell numbers for all simulations. ...................................................... 110
Table 5.3. Spatial discretisation-induced numerical uncertainty (for λ=75). The
wave resistance coefficient listed in this table was arrived at by subtracting the
double body resistance from the multiphase resistance. ................................ 113
Table 5.4. Temporal discretisation-induced numerical uncertainty (for λ=75).
The wave resistance coefficient listed in this table was arrived at by subtracting
the double body resistance from the multiphase resistance. .......................... 113
Table 6.1. KCS and computational domain principal characteristics in full-
scale ......................................................................................................................... 138

xiv
Table 6.2. Computed integral quantities and related uncertainties ............... 139
Table 7.1. Ship characteristics. ............................................................................. 182
Table 7.2. Test matrix and resultant Kelvin half-angles according to Havelock's
(1908) method. ....................................................................................................... 182
Table 7.3. Cell numbers for both canals. ............................................................ 186
Table 7.4. Grid and time independence (rectangular canal, 𝐹ℎ=0.57). EFD
result: 4.5047 N. ..................................................................................................... 190
Table 8.1. KCS principal characteristics (in model scale). ............................... 212
Table 8.2. Text matrix............................................................................................ 212
Table 8.3. Cell numbers for all four depth transitions ..................................... 216
Table 8.4. Numerical uncertainty study results (results are given for the ship
resistance). .............................................................................................................. 218
Table 8.5. Summary of wavenumbers for cases 1 ~ 4 (𝐹ℎ𝑖 =0.77). ................. 232
Table 0.1. Model coefficients for AKN and 𝑘 − 𝜀 closures. ............................. 293

xv
 ABSTRACT

Historically, the prediction of ship resistance has received its fair share of
attention by the scientific community. Although there is a significant body of
literature devoted to the study of ship hydrodynamics, several open research
questions of great practical relevance remain unanswered. Among these are
the extrapolation of ship resistance from model to full-scale in restricted, and
unrestricted waters, as well as shallow water ship flows.

Most approaches used to predict the performance of a ship have typically

relied on the assumptions inherent in potential flow theories, namely, the fluid
is treated as inviscid and irrotational. In many cases, these are justifiable
assumptions, yielding accurate predictions. However, there are equally many
occasions, in which the analyst may not obtain a correct picture of the
performance of a ship when relying on the assumptions of potential flow
theory. Predicting scale effects, and shallow water influences on ship
performance are prime examples of such cases.

Numerical techniques based on the Navier-Stokes equations can be thought of

as a solution in cases where it is important to model a greater proportion of
the physical phenomena. The numerical simulation of ship flows has evolved
into a highly practical approach in naval architecture. The main advantages of
using such an approach relate to the fact that it accounts for the action of
viscosity and turbulence, and can therefore model scale effects and shallow
water ship flows.

However, with the rapid advent of computational methods in all fields of

engineering, several areas have emerged as significant sources of ambiguity.
Amongst these are the best approach to modelling turbulence, the numerical

xvi
uncertainty induced as a result of mapping the continuous governing
equations onto a discrete grid, and boundary conditions within the
computational domain.

This thesis aims to address all of these issues using a commercially available
Reynolds averaged Navier-Stokes solver. Firstly, a detailed literature review
on the current methods and approaches to circumventing the problems
mentioned above, both numerically and through the use of potential flow
theory, is given. Then, studies on scale effects in deep and shallow waters are
performed, supplemented by investigations into turbulence modelling and
numerical uncertainty. Following these, the thesis’ focus shifts towards
shallow water phenomena. In particular, the modelling of ship flows without
the use of Galilean relativity, and the determination of the Kelvin half-angle in
restricted waters. Abrupt changes in the cross-section of the canal in which a
ship propagates are also explored, with focus on ship resistance and the
properties of the wave field.

Finally, the main results obtained from each chapter are summarised and
compared against the aims and objectives of this thesis, before
recommendations for future work are suggested.

xvii
 INTRODUCTION

This chapter will provide a brief overview of the issues and topics this thesis
aims to address. Then, the rationale behind the motivating factors for each
subsequent chapter will be given. Following this, the research aims and
objectives will be detailed. Finally, the thesis structure will be described.

1.1 Background

The prediction of ship resistance is a fundamental issue with great importance.

If a naval architect does not have an adequately accurate estimate of the
hydrodynamic forces, acting on a particular hull, the resulting decisions can
prove to be wrong, with detrimental consequences.

Several considerations should be taken into account when endeavouring to

predict the resistance of the ship. A useful starting point is to utilise potential
flow theories, which can be used to assess a great number of different designs
quickly, even on modern day standard computers. However, over-reliance on
potential flow theories should be avoided due to their limitations. The vast
majority of these methods, some of these are formulated to treat the fluid flow
as linear, irrotational, and inviscid (Beck and Reed, 2001). Although these
assumptions may well be suitable in some situations, and are to a certain
degree, able to achieve engineering accuracy, they never represent the full
picture.

The complexity of the problem is predominantly expressed in the mathematics

required to numerically describe the flow around a ship. Particularly elusive
are the vorticity and turbulence components. The latter is a very active field of
research in itself, which has made great strides in recent years. However, a
substantial number of the most state-of-the-art theories and methods are not

1
capable of providing the full mathematical description of the flow around a
ship. Instead, Computational Fluid Dynamics (CFD) is used, particularly
Reynolds Averaged Navier-Stokes (RANS) solvers. This approach allows the
analyst to solve for the fully nonlinear viscous flow around the ship.

Naturally, physical experiments are of immense value to the field of ship

hydrodynamics. The problem here lies with the cost, time, and technological
equipment which is a necessary prerequisite for such a study. A towing tank
is an expensive facility itself, however, it is an absolute necessity for the rapid
advancement of the field of hydrodynamics. Furthermore, high-accuracy
sensors and high-speed processing units are also necessary in order to detect
features such as flow velocity, free surface elevation, etc. in a towing tank
experimental setting.

The advances in modern computational processing power and speed, on the

other hand, have resulted in the favourable situation where a model-scale
RANS simulation can be performed on standard computers within a
reasonable timeframe. Of course, the computational requirements frequently
necessitate the use of high-performance computers (HPCs), especially when
full-scale simulations are performed in CFD. HPCs are not only becoming
more widely available, but the number and complexity of tasks they can
handle is increasing exponentially (Slotnick et al., 2014). It is therefore
reasonable to expect that RANS, and more widely used CFD methods will
become the norm for routine analyses and standard hydrodynamic problems
in the foreseeable future. In this thesis, a commercially available RANS solver,
Star-CCM+, is used to perform analysis of ship hydrodynamics in deep and
shallow water.

1.2 Motivating factors

The continued acceleration of the use of RANS methods calls for best practice
approaches to be identified, especially in the fields currently identified as
bottlenecks in ship hydrodynamics, and in CFD simulations of ship
hydrodynamics. The present thesis attempts to use the strengths of CFD RANS
simulations when compared to potential flow in the following areas:

• Scale effects: Due to the fact that a ship operates at the air-water
boundary interphase, there are both Reynolds number and Froude
number dependencies on the flow. Since these two components scale
differently, it is not possible to achieve full similarity experimentally,
using a scaled model. Compounding on this issue is the fact that scale

2
 effects on different components are not fully understood. There is also
strong evidence that constituent components of ship resistance might
not, strictly speaking, exist in the form currently assumed (that is,
linearly decomposable components). This may be deduced by
observing interdependencies between different components of ship
resistance, which suggest nonlinear relationships between those
components. CFD methods are well-placed to address these issues since
they are capable of modelling the full spectrum of physical phenomena.
In particular, RANS methods are fully nonlinear, inherently taking into
account vorticity and an approximate description of turbulence. The
present thesis will evaluate scale effects on a ship in deep and shallow
waters following a geosim approach. Emphasis is placed on
understanding the physical reasons behind the observed
interdependencies (Chapters 3 and 5).
• Turbulence dependence: Although there are many approaches to
modelling turbulence, with alternatives emerging on a regular basis,
the field of ship hydrodynamics is slow to adopt these new methods. A
survey of over 100 studies concerning maritime CFD revealed that
virtually all work is concentrated solely on two modelling approaches
when it comes to closing the Navier-Stokes equations. Simultaneously,
it was discovered that ship sinkage and trim have been identified as
particularly difficult to simulate accurately as seen in the findings of
several workshops focused on numerical ship hydrodynamics. The
increased relative importance of these metrics on ship performance
suggests that shallow and confined water cases require further
research. Therefore, a turbulence dependence study in restricted waters
is warranted and necessary to help settle the debate regarding which
turbulence model, if any, is superior (Chapter 4).
• Numerical uncertainty at full-scale: High Reynolds number flows are
difficult and expensive to measure experimentally, often resulting in
the lack of full-scale measurements used to validate numerical
predictions. In such cases, the analyst has no choice but to rely on best
practice approaches. One difficulty relates to constructing an adequate
grid onto which to solve the governing equations. Finite availability of
computational resources also dictates that allocating solution nodes
cannot be uniform across the entirety of the computational domain.
This results in areas within the solution which one can discretise with
different mesh densities. Parameters produced in the course of the

3
 simulation also require different levels of mesh refinement. To begin
addressing these issues, a study on local spatial numerical uncertainty
within the domain was performed. Specifically, the elevation of the free
surface and the skin friction of the ship were assessed (Chapter 6).
• Ship-generated waves in restricted waters: The impact of ship-
generated waves, particularly in restricted waters, can lead to bank
erosion and infrastructure damage. Their high degree of nonlinearity
presents a challenge for potential flow methods. To assess the capability
of CFD to model ship waves, a Fourier method was used to analyse the
wavefield and compare the numerical output with a theoretical
estimation of the Kelvin wake angle. In an attempt to reduce the
modelling assumptions used in the study and be as consistent as
possible with a physical towing tank, all open boundaries within the
computational domain were removed. These were replaced with no-
slip walls, rendering a greater degree of similarity and conceptual
consistency to the physical towing tank. The newly constructed virtual
towing tank is used to simulate a ship advancing through a rectangular
canal and the New Suez Canal at model-scale. The analysis of the ship-
induced disturbance reveals that the Kelvin half-angle is modelled with
a discrepancy of less than two degrees when compared to a linear
theory for the rectangular canal. However, the inflection wavenumber
at which the Kelvin half-angle occurs is shown to deviate from the
theory (Chapter 7).
• Depth transitions in shallow water: Restrictions in shallow water are
in fact rarely maintained constant in shape over long distances as
bathymetry may vary substantially over the track of a ship. Yet the
majority of studies in shallow water ship hydrodynamics examine
different conditions sequentially, and consequently treat them as
independent. It is therefore prudent to numerically construct a towing
tank featuring an abrupt change in the water depth, which the ship
encounters and advances over, as it propagates in the tank. The
characteristics of the wavefield were compared to analytical methods
to provide a kind of validation to the study. Transmission of the ship-
induced disturbance over the depth discontinuity was also evaluated
alongside the viscous dissipation of the generated waves (Chapter 8).

4
 1.3 Research aims and objectives

The research aim of this thesis is to evaluate the performance of ships in deep
and shallow waters via CFD, to use the RANS method’s strengths, to the
challenges, and seek to provide practical recommendations for improvement
of modern methods.

The specific objectives of this thesis have been designed to address the issues
raised previously. The objectives are as follows:

• To perform a thorough review of the literature on computational ship

hydrodynamics and identify open research questions.
• To perform a geosim analysis in deep and shallow waters.
• To investigate the scale effects on ship resistance and identify the
sources of these scale effects.
• To identify the best turbulence model to close the RANS equations,
providing consistent predictions over a range of parameters, at a small
computational cost.
• To demonstrate the application of numerical uncertainty estimators on
different parameters and their use in full-scale CFD simulations.
• To determine the performance of computational grids in terms of
proximity to the asymptotic range.
• To determine the Kelvin half-angle of a ship in restricted waters based
on a numerical free surface analysis method.
• To predict the impact of an abrupt change in the water depth on the
resistance of a ship and its wavefield in shallow waters.

1.4 Thesis structure

This thesis is organised in accordance with the motivating factors, aims and
objectives. The layout is summarised below.

• Chapter 2 (Critical Review) is dedicated to a review of the literature. To

begin with, a brief historical overview is presented, focusing on
important contributions in the wider field of hydrodynamics. Then,
recent work in the field of ship resistance prediction in deep and
shallow waters is critically reviewed. The current understanding of
scale effects is then summarised and further need for research is
identified. Finally, a review is given on the modelling of ship-generated
waves.

5
• Chapter 3 (A Geosim Analysis of Ship Resistance Decomposition and
Scale Effects with the aid of CFD) presents a study on the scale effects
of a ship in unrestricted waters. The analysis is performed within Star-
CCM+, and compared to experimental data at three different scales
model-scale factors. Following this, a full-scale simulation is
performed. The generated results suggest there is an interdependence
between frictional and wave resistance, and demonstrate a Reynolds
number influence on the latter. The form factor is shown to depend to
the scale factor at which the analysis is ran.
• Chapter 4 (Application of Eddy-Viscosity Turbulence Models to
Problems in Ship Hydrodynamics) presents a thorough investigation
into the best choice of modelling turbulence in shallow water ship
hydrodynamics. A simple statistical approach is proposed to rank the
turbulence models from 3 families to determine which candidate
performs both accurately and consistently across case studies, while
consuming the least computational resources in CFD.
• Chapter 5 (A Numerical Assessment of the Scale Effects of a Ship
Advancing Through Restricted Waters) presents an analysis of the scale
effects of ships in restricted waters. The source of scale effects in a canal
are identified as the presence of viscosity, and vorticity. These are
generalised to the case of unrestricted waters, and the consequences of
including those parameters are discussed. The results are shown to
conform to theoretical estimations of the scale effect. An examination of
the generated data in terms of frictional resistance coefficient and
boundary layer profiles reveal that lateral restrictions of the waterway
are not as significant as the level of vertical confinement for the
examined case.
• Chapter 6 (A Posteriori Error and Uncertainty Estimation in
Computational Ship Hydrodynamics) presents an assessment and
comparison of numerical uncertainty estimation procedures in CFD.
The application of a local error approach is examined in detail, as well
as the additional data, resulting from its application. A less-known
approach to estimate the numerical error is applied, and its confidence
interval is assessed. The results reveal that distinct metrics within the
domain, are located at different distances from the asymptotic range.
Therefore, the applicability and performance of uncertainty estimators
varies significantly depending on the parameter chosen for the
verification study.

6
• Chapter 7 (Virtual Replica of a Towing Tank Experiment to Determine
the Kelvin Half-Angle of a Ship in Restricted Water) presents a virtual
towing tank, physically consistent with model testing facilities.
Specifically, the numerical simulation features no open boundaries. It
is demonstrated that the approach can provide satisfactory predictions
in terms of resistance, and can model the wavefield accurately. The
wavefield is validated by means of a Fourier representation of the
numerical free surface, which is compared to theoretical relations,
derived from the linear dispersion relation. Differences between the
theoretical and numerical Kelvin half-angle are demonstrated. These
are found to stem from the magnitude and location of the wavenumber
inflection point, at which the Kelvin half-angle is evaluated.
• Chapter 8 (Modelling the Hydrodynamic Effect of Abrupt Water Depth
Changes on a Ship Travelling in Restricted Waters Using CFD) presents
an evaluation of the increase in resistance of a ship due to a step change
in the water depth. The numerical simulations, performed at model-
scale, showed that magnifications of ship resistance of up to 226% are
predicted when the depth Froude number is near the critical value past
the transition. As the velocity is increased, the wavefield is shown to
interact less with the depth discontinuity. The numerical data show
very good agreement with analytical relations of wave transmission
past a depth discontinuity and viscous dissipation of waves in canals.
A strong formation of a boundary layer at the canal bottom is observed,
which persists a significant time after the ship has passed.
• Chapter 9 ( Conclusions and Future Work) summarises the findings of
this thesis and lays out directions for future research. A discussion is
given on the findings and how they relate to the aims and objectives.

7
 CRITICAL REVIEW

This chapter is dedicated to an overview of the discipline to which this thesis

pertains, i.e. numerical ship hydrodynamics. The origins of the field are briefly
explored, before focus is shifted to the evolution of the discipline into its
present form, its successes, challenges, and future directions. Then, each
subset of problems examined in the following chapters is given a separate
section, where the relevant literature is reviewed. Naturally, prevalence is
given to the contributions made in the area of ship hydrodynamics.

2.1 The origins of ship hydrodynamics

The problem of ship resistance prediction has a long and rich history spanning
over centuries, and forms a cornerstone in the field of hydrodynamics. The
discipline itself, however, did not emerge as a result of the fundamental
questions regarding ship resistance, asked by both Newton and Euler
(Gotman, 2007). Nevertheless, so important was the study of fluid motion
around ships that Newton devoted a significant section of his Principia
Mathematica to it, where he asserted the proportionality of inertial resistance
of a body to the square of its velocity (Darrigol and Frisch, 2008). Though
Newton’s contributions are important, da Vinci is thought to have produced
the first treatise of fluid motion many years earlier (Gotman, 2007).

However, it was Daniel Bernoulli who first gave the present area of study its
formal designation (Darrigol and Turner, 2006). He discovered the universally
known law carrying his name. Hydrodynamics was then capable of explaining
highly practical phenomena that seem counter-intuitive at first. For instance,
Bernoulli’s equation described why the pressure drops in a pipe when its
cross-section is reduced. This found applications in describing blood flow and

8
advanced the understanding of blood pressure within the cardiovascular
system. Unfortunately, hydrodynamics could not maintain the perception of
being a practical science for long. This is best explained by one of Daniel
Bernoulli’s contemporaries, d’Alembert, whose contributions to science are
many, particularly as they relate to fluid flow.

Hydrodynamics ended up being seen by many researchers as a purely

theoretical study of immense complexity with little to no bearing on the real
world. This is best exemplified by d’Alembert’s paradox, stating that a body
would encounter no resistance in an incompressible, inviscid fluid. Of course,
the paradox, as interpreted at the time, contradicted everyday experience by
precluding the possibility of, for example, bird flight. This was partially
responsible for a fundamental split in the field itself, resulting in the two
sciences of hydrodynamics and hydraulics (Darrigol and Turner, 2006).

The primary reason behind the perceived fracture between the practical and
theoretical was likely driven by the fact that Euler’s equations, describing fluid
motion, presented not only the first nonlinear field theory, but remain
shrouded in mystery even today. In effect, Euler’s equations are still thought
to be sufficient to model ship-generated waves (Torsvik, 2009). This likely
contributed to the confusion regarding d’Alembert’s paradox, which can be
resolved as soon as one considers the complete description of the governing
equations.

There are many critical contributions to the field, without which

hydrodynamics would not be as rich a science. Some of the important names
include those of Navier, and Stokes, who derived the equations of motion in a
viscous fluid; Rankine, who developed the theory of sinks and sources;
Helmholtz, who founded the theory of vortex motion and introduced the
velocity potential; Reynolds, who formulated the ideas of laminar and
turbulent flow (Milne-Thomson, 1962); Froude (1874), who conceptualised the
first scaling law in ship hydrodynamics; and Michell (1898), whose integral (or
variations thereof) is still used extensively in practical and academic contexts
to predict the wave resistance of a ship, despite being overlooked by his
contemporaries.

2.2 Contemporary perspectives

The idea that it is practically relevant to obtain a good estimate of the resistance
of a ship emerged when the first machine-powered craft were built (Gotman,
2007). The mathematics of ship resistance prediction remains a highly active

9
field even today. In the last century, the linear description of the scientific
problem at hand matured significantly. For instance, Havelock (1908) and Inui
(1954, 1936a, 1936b) derived methods to predict the location of wave crests and
the corresponding kelvin half-angle in deep and shallow waters.

In recent years, the volume and pace of research in linear methods has reduced
dramatically. However, linear methods remain of high regard and use due to
their utility (Beck and Reed, 2001). Nevertheless, the 21st century, has seen an
explosion of academic literature concerned with the application of the Navier-
Stokes equations. Largely driven by the exponential growth in available
computational resources, the numerical solution of the fully nonlinear
governing equations has become commonplace. However, there are still many
problems in the application of all forms of the Navier-Stokes equations.

Ranging from purely mathematical to fundamentally practical, the Navier-

Stokes equations offer a wealth of open research questions. The problems in
applying the governing equations span many disciplines. For instance, it is
still unknown whether the Navier-Stokes equations produce unique solutions
in three dimensions (Carlson et al., 2014). Luckily, most of these issues do not
concern practical applications of the Navier-Stokes equations (or more
accurately, the Reynolds averaged form of the governing equations). From an
engineering point of view, all that is required is consistent provision of results
within some predefined measure of accuracy, which has been extensively
demonstrated for ships in academic contexts.

The remaining challenges relate to how one can guarantee that a solution
satisfies the predefined measure of accuracy. This is particularly of concern
when performing novel case studies. Efforts to improve predictions have been
ongoing for many years. Their evolution in the field of ship hydrodynamics is
best exemplified in a hierarchal form, as shown in Figure 2.1, adapted from
Witherden and Jameson (2017). The main difference between the present
interpretation of the hierarchy of fluid flow models and that of the above
reference is that Large Eddy Simulation (LES) and Direct Numerical
Simulation (DNS) have been incorporated at the top of the pyramid. This was
done in light of recent publications on the matter, which demonstrated that
LES can be used in practice, albeit subject to restrictions. An in-depth
discussion on these can be accessed in the open literature (Fureby et al., 2016;
Kornev et al., 2019; Kornev and Abbas, 2018; Liefvendahl and Fureby, 2017;
Shevchuk and Kornev, 2017). Although DNS remains outside researchers
capabilities for high Reynolds number flows at present, it can be anticipated

10
that this would not be the case indefinitely, keeping in mind the rapidly
increasing availability of computational resources and improvements in
algorithm efficiency.

Figure 2.1. Hierarchy of fluid flow models, adapted from Witherden and
Jameson (2017).

The issues not addressed by Figure 2.1 relate to potential difficulties and their
underlying sources relevant to each method. As one progresses towards the
top of the hierarchy, different conceptual issues must be overcome. To
elaborate, although the practicality of, for example, linear potential flow is not
questioned, practitioners are aware of the situations where its use is
admissible. The range of parameter definitions, computational resource and
ease of implementation suggest that it is trivial to compare solutions obtained
with linear potential flow solvers and experiments to build confidence.

Near the top end of the complexity scale, RANS-based flow predictions feature
dozens of parameters, models, sub-routines, and more generally, modelling
approaches. These range from turbulence modelling, through boundary
conditions, to numerical wave damping/definition, and interface
tracking/capturing to name but a few. Each category mentioned above has

11
emerged as a field of study in its own right, rendering it practically impossible
for the practitioner to have adequate expertise simultaneously in all fields. For
this reason, studies have emerged aimed at providing specific
recommendations to alleviate the burden. The required levels of technical
knowledge can be reduced upon familiarisation with relevant research work
for a wide spectrum of problems. For instance, Wackers et al. (2011) have given
the most in-depth overview of strategies to model free surfaces in ship CFD,
while Eca et al. (2015) provide an overview of the application of wall functions.

Although they are highly valuable, guides, best practices, and reviews are
insufficient to increase physical understanding in ship hydrodynamics. For
this reason, the majority of studies are concerned with specific phenomena, i.e.
a subset of the physics. These are typically carried out under idealised
conditions to isolate a certain aspect of the solution and analyse it suitably in
view of increasing the conceptual understanding of the underlying physics.

2.3 Scale effects on a ship hull

When designing a ship, its performance is usually assessed against a variety

of parameters. One of these is the expected value of resistance that the ship
will experience in calm waters. Having a good estimate of this value is crucial
because it determines the power delivered by the propulsion plant. In some
cases, the naval architect may choose to seek alternative hull forms if the
resistance falls within an unfavourably high range. It is therefore of critical
importance to obtain an accurate estimate of ship resistance in full-scale.
Unfortunately, at the design stage, one rarely possesses full-scale data. To
address this, towing tests are performed to predict the resistance in model-
scale, which is then extrapolated to full-scale.

The earliest procedure for ship resistance extrapolation was devised by Froude
(1874). It begins by assuming that ship resistance, in non-dimensional form,
can be decomposed into frictional and residuary components. Central to the
present argument is the latter being constant with scale, which is known not
to be correct (Toki, 2008). On the other hand, the frictional components vary
with Reynolds number (Re), and can be approximated by the skin friction of a
flat plate with an equivalent submerged surface.

The second and more widely used approach was proposed by Hughes (1954),
who suggested the form factor approach. Within this approach, the resistance
is decomposed into viscous and wave-making components. The latter is
hypothesised to remain constant for geometrically similar ships, because, at

12
any scale, the ship is expected to produce a geometrically similar wave pattern.
On the other hand, viscous resistance is further split into frictional and viscous
pressure components by use of the form factor. Again, the frictional resistance
approximated as that of an equivalent flat plate. Several problems plague this
method, the most important of which is perhaps that the form factor (1+k),
used in accounting for 3D effects, is assumed invariant with scale. To estimate
(1+k), a model is towed at a low speed, where the wave resistance is supposed
to be negligible, also known as the Prohaska test. Alternatively, the ITTC’78
method introduces a factor to be determined via regression analysis to account
for the wave resistance in the form factor calculation, shown in Eq. (2.1):

𝐶𝑇 = (1 + 𝑘)𝐶𝐹 + 𝑐𝐹𝑟𝑁 (2.1)

where c is a constant, chosen to fit as many 𝐶𝑇 (total resistance coefficient)

measurements as possible, while N normally attains a value between 4 and 6
(van Mannen and van Oossanem, 1988), and 𝐶𝐹 is the frictional resistance
coefficient.

Although the 2D and 3D extrapolation procedures, summarised graphically in

Figure 2.2, can provide a good estimate of the resistance prediction, they are
just that – estimates. Differences between the two methods are expected to
stem from a variety of sources. The obvious ones are associated with the
assumptions mentioned above: in the first case, the residuary resistance,
whereas in the second – the wave resistance, both assumed constant with scale.

The problem becomes worse because, while a linear decomposition is

assumed, it is well-known that resistance is a nonlinear problem (van Mannen
and van Oossanem, 1988). Furthermore, this falls within the category of
processes in which linearly breaking up physical phenomena and treating
each part separately cannot adequately describe reality. This is due to the
inability of linear systems to account for interactions between the different
components in a nonlinear system, yielding properties not exhibited by the
linear system (Saaty and Bram, 1964). An example of this problem is the
interaction between frictional and wave resistance.

Underpinning the field of engineering is dimensional analysis, defined by

White (2010) as the practice of reducing the number and complexity of
variables upon which a physical process depends. Consequently, dimensional
analysis rests on scientists’ ability to define proper relationships between
variables. Having established the difficulty in achieving this fully implies that
one can never predict exactly the full-scale resistance of a ship with precision,

13
using model experiments. To tackle this, towing tank facilities rely on
experience and large databases of model and full-scale data recorded during
sea trials.

It should be noted that the use of sea trial measurements for validation
purposes is used infrequently. This is the case because the equipment and cost,
required to perform such an activity tend to be prohibitive. A testament to this
fact is the scarcity of full-scale validation work in the literature. However, in
2016, the Lloyd’s Register organised a blind workshop on full-scale ship
hydrodynamics. The findings, and anonymised results from participants were
published in Ponkratov (2016). The aforementioned report begins with an
extensive summary of the 3D laser measurements of the hull, performed
during dry docking. Then, the experimental equipment is detailed, before the
aggregated results are given. These highlight the difficulty in performing full-
scale simulation, as only 3 participants achieved an error of approximately 3%
or less.

Figure 2.2. Resistance extrapolation

One of the most fundamental problems in model tests, and the subsequent
extrapolation is that the intrinsic physical properties of the medium (air and
water) have not been scaled down along with the ship. In this respect, potential
flow can be a useful starting point. Specifically, linear potential flow theories
predict no scale effect in the wave pattern, generated by a ship for a given
Froude number. To reveal scale effects, the physical phenomena not modelled
by the abovementioned approach should be taken into account, specifically,

14
the action of viscosity. Therefore, it can be stated that the action of viscosity in
the fluid is responsible for scale effects.

In his experiments, García-Gómez (2000) demonstrated changes in the form

factor of a ship as a result of scaling on several different hull forms. He also
suggested an empirical correction to account for the difference in the model
and full-scale ship form factors. It is important to note that according to García-
Gómez, (2000), scale effects are due to Reynolds number-dependency only,
and they are due to the friction line used.

It is well documented that the form factor (1+k) changes with Re, as
demonstrated by many researchers (García-Gómez, 2000; Kouh et al., 2009;
Lee et al., 2018; Terziev et al., 2019a; Zeng et al., 2019), but its use is still
endorsed as part of the ITTC extrapolation procedure (ITTC, 2017a). However,
placing all scale effects on the skin friction line is not correct, because every
component of ship resistance is expected to be associated with a scale effect in
its own right. This may be confirmed by examining the non-dimensional form
of the Navier-Stokes equations. Specifically, the effect of viscosity is
represented as 1/Re (Fox et al., 2015). Therefore, any viscous fluid will change
its behaviour with varying Reynolds number. This implies that at a change
from model to full-scale, where the Reynolds number changes order of
magnitude from O(106) to O(109) may induce palpable scale effects.

The wave resistance is also typically assumed invariant of Re, but boundary
layer physics suggests otherwise. To elaborate, Brard (1970) predicted that
1/3
viscosity and vorticity act on near-field waves as 1⁄(𝑅𝑒 × 𝐹𝑛2 ) , while on far
field waves as 1/(𝑅𝑒 × 𝐹𝑛4 ), where 𝐹𝑛 is the Froude number. Thus, rendering
the effect of near-field waves more significant. Coincidently, these are also of
greater practical importance in the low speed regime, where ships are tested
according to the Prohaska test, used to determine the form factor. This is
because ships generate predominantly near-field disturbances at low speeds.

The influence of turbulence is also known to impact on ship-generated waves

(Brard, 1970; Tatinclaux, 1970). Since it is not possible to achieve both Reynolds
and Froude similarity simultaneously in practice, one retains different flow
properties in terms of turbulence and vorticity when extrapolating from model
to full-scale. Here, it is useful to introduce the concept of an “equivalent ship”.
This is sometimes used in potential flow theories in an attempt to account for
boundary layer displacement thickness and its impact on flow properties
(Gotman, 2002; Lazauskas, 2009). In this concept, an “equivalent ship” is the

15
ship’s underwater geometry, plus the displacement thickness of the boundary
layer. Now, the displacement thickness being different at each scale, inevitably
means that this equivalent ship is different in shape at each scale.

One could consider the aforementioned statements from the classical point of
view of source strength distribution used by potential flow to model the ship
as a wave maker. If boundary layer physics are taken into account, then the
source strength, assigned to the stern is not the same at different scales. This is
true because the boundary layer is relatively thicker at model scale than at full-
scale. Moreover, a higher Reynolds number implies a broadening of the
turbulent kinetic energy spectrum (Durbin and Pettersson Reif, 2011). The net
effect of this is the presence of eddies of different characteristic lengths and
time scales. All of the above serve to point towards the existence of a viscous
effect on the wave resistance, as suggested by Brard (1970) and Tatinclaux
(1970), and therefore, scale effects.

The arguments laid out so far must also be considered in conjunction with the
fact that in each scale factor of a geosim series, the fraction of the ship over
which a laminar boundary layer may be observed is different. Furthermore,
knowledge that a thickening of the boundary layer occurs with an increase in
scale factor (decrease in linear dimension) suggests that one may expect wave
resistance coefficient to decrease as one moves up the Reynolds number scale
in a geosim series. This was one of the conclusions of Ferguson (1977), who
observed this effect experimentally. More recently, studies on ship hull
roughness demonstrated the that a thicker boundary layer, resulting from
surface roughness increases wave resistance (Song et al., 2019). However,
Brard (1970) discovered that turbulence and vorticity supress ship waves.
Indeed, the presence of turbulence is typically interpreted mathematically as
a “sink” for large scale motions (Golbraikh et al., 2013). It is therefore not
straightforward to predict whether the wave resistance will be higher or lower
a priori. This is the case because the combined effect of turbulence, vorticity,
and change of boundary layer properties on ship waves are difficult to
quantify. Moreover, the relationship between these components, and their
relative impact may alter for different shapes.

Flow separation is also known to play an important part in model-scale, but

not in full-scale (Kouh et al., 2009; Raven et al., 2008). Vortex formation is
typically delayed in full-scale, and when it occurs, vortices encounter higher
damping than in model-scale (Hochkirch and Mallol, 2000). Both of these
effects are likely related to the change in the laminar-turbulent boundary layer

16
transition location. This causes the different flow properties generated at each
scale to cascade and snowball towards the stern and into the wake itself.

At this stage, it is important to make some brief observations regarding the

frictional component of ship resistance. As stated earlier, this is assumed to be
one of the independent components of the total resistance. Its prediction has
received its fair share of attention form the scientific community, resulting in
numerous approaches.

Methods to calculate the frictional resistance coefficient can be classified into

three categories, all of which define the frictional resistance coefficient as a
function of the Reynolds number. The first category consist of correlation lines,
such as that of White (2006), as well as Prandtl-Schlichting and Schultz-
Grunow, reported in Schlichting (1979). The second category consists of
formulae, derived using integrated analytical two dimensional boundary layer
equations expressing the local frictional resistance of a flat plate, such as those
of Schoenherr (1932), Hughes (1954), Grigson (1999), Katsui et al. (2005), and
Lazauskas (2009). The third category is numerical lines, such as those of Eça
and Hoekstra (2008), who developed three formulations by fitting curves to
data obtained via a RANS solver - one based on the of Hughes (1954) line (in
rational form); and two in polynomial form (one linear and one cubic). More
recent numerical friction lines include that of Wang et al. (2015) and Korkmaz
et al. (2019).

While there is a plethora of available friction lines, it is important to keep in

mind that the vast majority do not take into account free-surface effects. The
analytically derived ones use the integral value of two-dimensional boundary
layer equations to express the friction coefficient. Even the numerical friction
lines were established with the use of double body models. However, as
pointed out by Stern (1985), the presence of a free surface causes a highly
complex three-dimensional flow. Thus, not only does one expect differences
due to the presence of a free surface in the frictional coefficient, but also, the
wave resistance is modified as a result of the different nature of the boundary
layer (Marquardt, 2009).

So far, primarily analytical and experimental studies aiming to determine the

resistance of a ship have been mentioned. However, an approach of rapidly
emerging popularity in the context of ship hydrodynamics is the use of
Reynolds averaged Navier-Stokes methods, as shown in Figure 2.3. Some of
the early work in which scale effects using a CFD approach have been

17
examined includes the work of Oh and Kang (1992). They modelled viscous
flow over the stern of a ship by invoking the double body approximation.
According to Gotman (2007), Foettinger first described the double body idea
by replacing the free surface with a symmetry boundary in 1924. Since then,
tests in wind tunnels using the underwater shape of a ship and its mirror
image have been performed, for example, by Patel and Sarda (1990). They
studied the turbulence and boundary layer characteristics of the Wigley
parabolic hull. The contribution of the abovementioned work is that it
describes features of the 3D flow used for numerical turbulence modelling.

Figure 2.3. Annual publications listed under the category “ship CFD”
according to Web of Science (2020) as of 26.04.2020. Extrapolated numbers
based on the trend1 up to 2030 are represented by empty bars.

Researchers were constrained for several years to modelling only a part of a

ship’s hull. Using the double body approach and placing the inlet amidships,
Eca and Hoekstra (2001) predicted the scale effect on a tanker by varying the
Reynolds number. One of the problems highlighted in their study is the
scarcity of experimental data for comparison. Although the problem is still
unresolved today, Kim et al. (2001) performed a systematic series of
experiments to alleviate this. Their motivation for performing the study was
specifically to provide validation data for CFD codes. Later, Tahara et al. (2002)
investigated the appropriate numerical setup applicable to full-scale ship
hydrodynamic performance prediction. Their findings include that one of the
key issues “for full-scale ship flow simulation is to maintain the accuracy in
the resolution of the flow within a viscous and turbulent boundary layer of
decreasing thickness.”

1 The data and associated fit to extrapolate the number publications to the year 2030 can be
found in Appendix A

18
In RANS solvers, turbulence modelling has been shown to be one of the main
factors in modelling scale effects. Duvigneau et al. (2003) performed a study
on hull form optimisation in both model and full-scale. Their findings indicate
the calculated parameters are highly dependent on the turbulence model
chosen. Indeed, Visonneau (2005) reported the same finding within the
European Full-scale Flow Research and Technology (EFFORT) project
(Bugalski, 2007). He also pointed towards free-surface effects as a potential
source of difficulties for RANS solvers in full scale. However, the main
problem remains the lack of full-scale data for validation.

Later, Raven et al. (2008) examined the capabilities of an inviscid solver and a
viscous solver to predict the full-scale performance of a ship. They found that
the wave resistance coefficient is up to 20% higher in full-scale when compared
to model-scale. However, the scaling of viscous resistance was computed via
a double body model. Continuing the trend of using the double body method,
Kouh et al. (2009) demonstrated the Reynolds number dependence of the form
factor. Their study featured several hull forms, including the KRISO
containership (KCS), which allowed them to draw comparisons between
different shapes. They demonstrated that the form factor decreases with
Reynolds number for all examined shapes.

Min and Kang (2010) questioned the basic assumptions of the extrapolation
procedure recommended by the ITTC (2017). Their paper represents the end
of the sequence of studies confirming that the form factor is Reynolds number-
dependent. However, they went further, suggesting the form factor be treated
as a function of Froude number as well, and called on other researchers to
investigate this. Min and Kang (2010) also provided a correction formula to be
used in the determination of the form factor.

An interesting approach to resolving the scaling problem was presented by

Guo et al. (2015), who utilised a non-geometrically similar approach, namely,
a similar ship, whose flow characteristics at model-scale match the parent hull
in full-scale. Although this method would be very useful if refined, it is yet to
be implemented elsewhere.

As recently as 2016, researchers opted to use the double body method to study
ship performance. For instance, Wang et al. (2016) chose this approach to
examine the effect of different draughts on the form factor. However, the main
event of 2016 in this respect was the publication of the Lloyd’s Register’s
workshop on ship scale hydrodynamics (Ponkratov, 2016), in which the

19
problems of CFD predictions in full-scale could be seen more easily. Although
attention is limited to the bare-hull submissions in the workshop in this thesis,
it is worth noting that only 3 participants achieved an acceptable level of
accuracy between the full-scale CFD result and the sea-trial data in terms of
total resistance and propeller rotation rate (3% deviation). For the purposes of
the workshop, sea trials were conducted post-docking of a general cargo ship
between Istanbul, Turkey and Varna, Bulgaria. The results indicated that air
resistance (cranes and superstructure) plays a much more important role in
the total resistance (7% contribution) than trim (3% contribution) for the
particular ship investigated. The Ponkratov (2016) report highlights the
differences between numerical setups. For instance, Starke et al. (2017), who
were one of the participants, showed that a level-set method for free surface
capturing is not capable of modelling the overturning bow wave.

The main obstacles to be overcome prior to the routine use of CFD in full-scale
computations require further research as suggested by the literature. One of
the issues frequently pointed out is the large number of cells. However,
Tezdogan et al. (2015) demonstrated that it is possible to carry out RANS
numerical simulations directly in full-scale. Alternatively, Haase et al. (2016)
proposed to verify a numerical grid in model-scale (based on Froude
similarity). Then, by altering the value of viscosity, the Reynolds number can
be changed to match the corresponding full-scale ship.

Recent work in the field of computational ship hydrodynamics has

concentrated on tackling the problem referred to above, directly (Niklas and
Pruszko, 2019a, 2019b). For instance, Sun et al. (2020) performed a numerical
analysis in model and full-scale and compared the latter with sea trial data.
Their work features double body and multiphase in for both model and full-
scale conditions. Incorporating a spinning propeller allowed Sun et al. (2020)
to determine that vortex formation is noticeably damped at full-scale when
compared at model-scale.

Other researchers have opted to focus on the action of turbulence in full-scale.

Turbulence modelling is typically a source of modelling errors, which are
difficult to quantify at full-scale (Bhushan et al., 2009, 2007; Duvigneau et al.,
2003; Pereira et al., 2017). Thus, alternatives to RANS techniques, which
resolve at least part of the turbulent kinetic energy spectrum, have emerged
and are rapidly gaining popularity. In this respect, Liefvendahl and Fureby
(2017) estimated that a full-scale Large Eddy Simulation (LES) for the Japan
Bulk Carrier (JCB) would require between 9.7×109 and 6.7×1012 cells,

20
depending on the approach (wall-modelled LES vs. wall-resolved LES). Such
grids are difficult to handle, even in academic contexts, demonstrating that
resolving the turbulent kinetic energy spectrum in full-scale is not currently
practical.

According to Pena et al. (2019), the bridging alternative (resolving part of the
turbulent kinetic energy spectrum), known as Detached Eddy Simulation
(DES), can be successfully employed to predict full-scale ship performance. In
their study, the aforementioned authors performed full-scale simulations on
the cargo ship investigated within the Lloyd’s Register workshop. The novelty
within their study is expressed in the fact that they opted to use the DES
approach. Although this is typically associated with considerably higher
requirements in cell numbers and lower time-steps, the authors achieved high
predictive accuracy. More importantly, they demonstrated that an accurate
solution can be achieved not only for integral properties, such as resistance,
but for the flow field near the ship as well.

Grid numbers are currently of some interest to the academic community. For
instance, Jasak et al. (2018) performed grid sensitivity studies in full-scale.
Their results suggest that full-scale simulations may be performed with a
relatively low number of cells while achieving high accuracy. Specifically,
simulations on a car carrier in the aforementioned work were performed with
approximately 6.4 million cells and achieved 0.24% deviation from the sea trial
result.

It is clear that full-scale simulations are possible even with current

computational power availability. However, best practice approaches should
be established to avoid the scenario observed in the Lloyd’s Register report,
where a relatively low number of participants obtained the desired level of
accuracy.

2.3.1 Shallow water effects

In many cases, instead of isolating a certain part of the physics to analyse it

separately, it is worthwhile to do the opposite in order to magnify certain
parameters, thereby making them stand out for easier analysis. In essence, this
results in seeking more challenging case studies which will accentuate sought
after phenomena. Shallow water studies can be thought to represent a such
reasoning. This is thought of in the sense that many phenomena, and their
interactions are highlighted due to the proximity of the seabed in the case of
shallow waters. Therefore, in carefully constructed experiments (physical or

21
numerical), one can deduce effects that may be obscured in deep waters due
to their small magnitude, for example, scale effects.

It is also important to have an appreciation that although a ship may spend

large parts of its operational life in unrestricted waters, it inevitably must enter
shallow/restricted waters. Therefore, to obtain a complete picture of a ship’s
performance, the designer must also understand how the ship will react to a
reduction in underkeel clearance (Tuck and Taylor, 1970). This may occur
whilst entering a port, or traversing one of the famous man-built waterways
(the Panama and Suez canals (Tuck, 1966)). It is known that an increase in ship
resistance can be expected when operating in shallow waters. Typically, this
is offset by a reduction in speed, also used to ensure against groundings, which
are the most frequently occurring accidents in the Suez Canal. While a
grounding at low forward speed may not damage the ship from a structural
point of view, it creates congestion. The low speed requirement has also meant
that the abovementioned waterways have become bottlenecks, restricting the
amount of freight passing through.

In fact, while the average annual vessel traffic has remained largely constant
in the past four decades, freight has increased in an exponential fashion (Suez
Canal Authority, 2018). Therefore, ship size must have increased
proportionally. This has meant that the main task of the Suez Canal Authority
is to now perform bathymetric surveys, since larger vessels require greater
under keel clearance and safety margins. Moreover, disproportionately many
incidents occur in shallow water according to EMSA (European Maritime
Safety Agency, 2019, 2018, 2017, 2016, 2015).

The principal phenomena, occurring in shallow waters are related to the

reduction in underkeel clearance and lateral extent of the waterway (in rivers
and canals, for example). These cause sinkage and trim of the ship, the
combined effect of which is termed ship squat. Unlike deep waters, waves
propagate at a single speed, √𝑔ℎ, where h is the water depth. In such
conditions, wave resistance is known to be nonlinear. Consequently, Michell's
(1898) integral, and variations thereof predict no resistance in shallow water
for subcritical speeds (Beck et al., 1975; Tuck, 1967, 1966). A ship propagates at
a subcritical speed when its velocity smaller than √𝑔ℎ. Conversely, when
U>√𝑔ℎ, the regime is called supercritical, whereas when the two quantities are
equal, the flow is termed critical.

22
By analogy to aerodynamics, and the associated wave resistance emanating as
sound when crafts break the sound barrier, Michell's (1898) integral predicts a
non-vanishing resistance for supercritical speeds. However, the sustained
generation of waves by sub-critically propagating objects in water of finite
depth suggests some energy transfer must occur from the object to the wave
system. Therefore, a result to the contrary defies physical experience, and was
part of the reason why Michell's (1898) integral was overlooked by his
contemporaries.

The reasons behind the non-existing predictions for subcritical speeds are
rooted in the linear nature of the aforementioned integral. In essence, wave
resistance in shallow water is non-linear. Obviously, such phenomena cannot
scale linearly. Therefore, understanding how wave resistance scales in shallow
water can increase knowledge of the deep-water equivalent, which contains a
substantial, if not dominant, linear component. This is true because the two
phenomena have the same origin and therefore share the physical mechanism
producing them. If one were to couple this fact with Tuck's (1978) forecast of
greater relative scale effect in shallow water, it is easy to see why the
examination of such cases is worthwhile.

The origins of scale effects in shallow water have several additional

dimensions. Firstly, the boundary layer changes in relative thickness with
scale, as discussed earlier. Its interaction with the seabed and/or sides causes
greater scale effects. For instance, Gourlay (2006) predicted that a ship’s
boundary layer may intersect the seabed. This may be the case even when the
depth-to-draught ratio (h/T) is not very small. For instance, Terziev et al.
(2019b) showed this to be the case for h/T=1.6. A study by Shevchuk et al.
(2016), and subsequently Böttner et al. (2020) examined the narrow gap
between a ship hull and seabed numerically and experimentally.

The findings of the aforementioned studies include the formation of a

boundary layer on the seabed. Knowing that a boundary layer will not scale
linearly implies that in full-scale, the ship operates under different conditions
than those in towing tanks. The extent to which this is impactful is yet to be
determined. In any event, scale effects are expected as a result of the disparity
between the ‘equivalent’ ship in different scales. Since the forces acting on the
hull will also be different, stemming from the scale-inconsistent boundary
layers, it is reasonable to forecast differences in ship squat as well. Indeed,
numerical and experimental studies have shown this assertion to be justifiable
(Duffy, 2008; Ferguson, 1977; Shevchuk et al., 2019).

23
However, in contrast to the above-mentioned studies a recent paper suggested
that scale effects on ship squat are negligible (Kok et al., 2020). Examining the
results presented in the aforementioned paper, it becomes apparent to the
author that such scale effects are actually predicted. For instance, the authors
of the previously mentioned paper state that “at h/T= 1.23 and 𝐹ℎ = 0.53, the
difference between the full scale and the model scale non-dimensional squat
is only 5.32%”. However, this may be a substantial difference when carried to
full-scale and could realistically cause grounding.

A systematic geosim series in shallow waters would likely to resolve the

questions posed above. However, when devising such an experiment, the
equivalent ship concept should be taken into account. A vessel at different
scales operates at different effective underkeel clearances due to changes in
the boundary layer thickness. It is therefore not immediately apparent how a
meaningful comparison should be carried out if the conditions are to be
maintained identical for all Reynolds numbers. The proximity of lateral
boundaries will also scale nonlinearly, inducing further effects.

Frictional resistance in shallow water is also of some importance for ships.

Studies have shown that the proximity of the seabed causes an increase in CF,
which according to the most recent research on the matter, is unique for each
ship (Zeng et al., 2019). This is caused as a result of three-dimensional effects
on the frictional resistance (Raven, 2019). In deep waters, these are relatively
mild, and cause a tolerable deviation from the well-known friction relations
used to predicting CF. Since scaling behaviour of the 3D contribution is
unknown, the use of relationships expressing CF should be scrutinised further.

Three-dimensional effects on the frictional resistance are an inherent aspect of

the tangential force itself (Dand, 1967). To the best of the author’s knowledge,
there are no methods to remove the 3D contribution of the frictional
component on a ship’s hull from the flat plate equivalent. Indeed, the
boundary layer of a ship is fundamentally different from that of a plate.
Therefore, there are many aspects in which a ship’s hydrodynamic behaviour
differs from that of a simpler body.

2.4 Ship waves

One fascinating consequence of operating at the air-water interphase is the

continuous production of waves. It is convenient to begin by examining the
deep-water case first. In deep waters, the ship generates a wave system,
confined within a semi-infinite wedge, extending aft of the ship. To gain a

24
picture of the waves generated by a surface-piercing disturbance, one must
simply place one wedge of half-angle ≈19° 28’ at the bow and one at the stern.
The entire wave system will be confined within these regions regardless of the
speed or underwater shape. This was described for the first time by Kelvin,
after whom the half-angle is named (Thomson, 1887).

The interesting phenomena occur once again in shallow water, where the half
angle which the wave system makes with the ship’s trajectory is speed
dependent. The shallow water case causes the previously described wedges to
be confined between the deep water limit (19° 28’) and 90° increasing as the
depth Froude number approaches to unity (Havelock, 1908; Inui, 1936a,
1936b). In other words, the waves fill the entire half-plane aft of the ship and
propagate in the same direction as the ship. The prediction of the Kelvin half-
angle is an active field of study even today, with regular new contributions
being published (Jiang et al., 2002; Lee and Lee, 2019; Rozman, 2009; Tunaley,
2014). That is, despite the fact that it has been over a century since the Kelvin
half-angle’s value in deep waters was first documented and subjected to
mathematical analysis (Thomson, 1887). This serves to highlight the wealth of
phenomena associated with ship-generated waves, rather than an inability to
describe the system.

Scientists have the habit of advancing hypotheses, which are either confirmed
or discredited experimentally. However, the opposite occurred in the study of
ship waves. John Scott Russel observed solitary waves in Scotland’s canals
years before they were subjected to appropriate mathematical analysis
successfully (Darrigol, 2003). Russel’s observations that a solitary wave may
cause a drop in the resistance at high speeds was belittled at first, only to be
vindicated years later.

Today, significant advances have occurred in the modelling of ship waves at

all water depths (Dias, 2014; Pethiyagoda et al., 2014; Rozman, 2009; Soomere,
2007). The methods range from Thomson's (1887) ray argument, through the
stationary phase method of Newman (1970) and Lighthill (1990), up to more
recent approaches reminiscent of Thomson's (1887) reasoning (Lee and Lee,
2019). These deep water relationships are also studied in a towing tank (Gomit
et al., 2014), and via satellite imagery (Rabaud and Moisy, 2013; Wu, 1992).

Using RANS solvers, one can simulate cases of arbitrary underwater cross
sections. For instance, an asymmetric canal was investigated by several
researchers as part of the PreSquat workshop (Mucha et al., 2014). Gourlay

25
(2014) modelled the required case studies using the slender body theory.
However, this method cannot recreate the free surface, which produces highly
intricate and complex patterns, as demonstrated by Tezdogan et al. (2016b).
Later, Terziev et al. (2018) simulated a set of horizontally unrestricted shallow
waters alongside dredged channels and a canal case. All of these studies
highlight that a ship’s wavefield will be highly influenced by any change in
the bathymetry. Even if the seabed is perfectly flat, the nonlinear effect in ship
waves is not fully understood. For this reason, many researchers have used
the RANS approach to model the problem at hand (Kinaci et al., 2016; Li, 2003;
Pacuraru and Domnisoru, 2017; Rotteveel and Hekkenberg, 2015;
Schweighofer, 2004; Song et al., 2019; Wackers et al., 2011; Wilson et al., 2006).

However, the extent to which a ship-generated wave is modelled by RANS

solvers in accordance with linear dispersion properties is yet unknown. That
is although significant demonstration of the accuracy of RANS solvers has
been demonstrated in a series of workshops on numerical ship
hydrodynamics, for example, Larsson et al. (2014). Human perception of
differences in between experimental and numerical free surfaces is notoriously
unreliable. For this reason, within this thesis, a study is performed to assess
the degree of agreement between the linear dispersion relation-based Kelvin
half-angle in shallow water and its numerical counterpart (Chapter 7). The
dispersion relation for canals of sloping sides is also of some interest in practice
due to their geometric similarity to riverbeds, used extensively for navigation
even nowadays.

Ship waves can cause erosion and infrastructure damage in many cases. This
is particularly relevant for ships operating in otherwise low-energy coastal
areas or lakes (Bourne, 2000; Hofmann et al., 2008; Schoellhamer, 1996).
Another equally important concern relates to the impact ship waves on fish
and their assemblages, a review of which is given in Wolter and Arlinghaus
(2003). More severe events are reported in Soomere (2009) where, firstly a case
where holidaymakers were being “forced to flee for their lives when enormous
waves erupted…”, and secondly, a case where ship waves caused a fatal
incident on the East coast of England. It is therefore important to understand
how ship waves transform over a sloping wall, representing a beach or a
canal/river side. This problem is reminiscent of that investigated by Tezdogan
et al. (2016b), Torsvik et al. (2009), Bechthold and Kastens (2020), Terziev et al.
(2020), Kok et al. (2020), and Elsherbiny et al. (2020, 2019a, 2019b) to name but
a few.

26
It is also important to note that RANS-based predictions of ship waves have
attracted some mixed opinions from researchers in terms of utility for the
present context. Although there does not seem to be any objection to using this
modelling approach per se, there is an awareness of the computational expense
at which such predictions come (Jiang et al., 2002). Even in cases where ship
hydrodynamics and the resulting waves are investigated, the wavefield is
destroyed at no more than 4~5 ship lengths aft of the stern, at most. This occurs
at the outlet boundary and is irreversible. From a naval architect’s perspective,
this may not be of great consequence, but in coastal engineering practice, these
waves need to be captured a long distance aft of the ship. Their interactions
with infrastructure can occur several miles from the ship track (Grue, 2017;
Soomere, 2009).

An economical approach to modelling ship waves is found in the Korteweg-

de Vries (KdV) equation, which accounts for dispersive and weakly nonlinear
effects. This has been used to model ship waves by many researchers (Cole,
1987; Hur, 2019; Katsis and Akylas, 1987, 1984). If the ambient pressure term
is retained in the formulation of the KdV equation, one obtains its forced
version (fKdV), which is only valid for waves propagating in a single direction
(Torsvik, 2009). An extension of the fKdV equation, capable of accounting for
waves propagating at a small angle relative to the ship’s track is the
Kadomstev-Petvianshili (KP) equation. This approach has also been
extensively used to model ship waves(Beji, 2018; Mathew and Akylas, 1990;
Sharma, 1995). These equations are frequently used in the present context
because they are susceptible to producing closed form solutions (Whitham,
2011). By contrast, the Euler form of the governing equations admits this for
simple, and therefore predominantly unrealistic cases (Torsvik, 2009).

Unlike the fKdV or KP equations, the Boussinesq equations, also members of

the long wave family of theories, have no known closed forms. In this respect,
they are related to the Euler equations. However, Boussinesq-type solutions
can account for motion in any direction. Moreover, they can handle
intermediate wavelengths, as opposed to solely long ones. Their derivation
also does not rely on a balance between dispersive and nonlinear phenomena,
as is the case with the two earlier types of equations. For this reason, the
Boussinesq form is frequently preferred (Dam et al., 2008; David et al., 2017;
Grue, 2017; Jiang et al., 2002; Torsvik et al., 2006; Wu and Wu, 1982). For
instance, according to Nwogu (1993), the Boussinesq equations, as derived by
Peregrine (1967) can model nonlinear transformations in shallow waters.

27
Nwogu's (1993) derivation, was used to examine ship waves recently by David
et al. (2017) to study ship waves propagating over an irregular bathymetry.
They compared their model with measurements made in the port of Hamburg
to demonstrate the validity of the Boussinesq approach.

The utility of Euler-based methods is indisputable. That is the case especially

when viscosity is of little to no importance and solutions can be obtained at
low cost. However, there are cases where the subtle action of viscosity has only
recently been uncovered. Specifically, as stated earlier, Böttner et al. (2020) and
Shevchuk et al. (2016) demonstrated the formation of a boundary layer on the
seabed in very shallow waters. It is in such cases that RANS methods can be
used, showcasing their strengths. For vessels with high block coefficients, i.e.
non-slender shapes, the action of viscosity may be significant. This would be
expressed as an influence in the near-field pressure distribution, which can
impact on the far-field waves.

On the other hand, in very shallow and/or narrow canals and rivers, the
reflection of waves would unavoidably interact with the aforementioned
boundary layer, found at the river or canal bottom, thus altering the wave-
induced velocity of water particles. In such cases, an Euler-based method
would not necessarily be suited for the modelling of ship-induced waves and
related wave fields.

2.5 Summary and conclusions

This chapter examined several, widely studied as independent, but

fundamentally linked topics. Firstly, a brief historical overview of the broader
discipline of ship hydrodynamics was given. This illustrated that as is typical
of phenomena within the field itself, the evolution of hydrodynamics was
frequently nonlinear. That is, great strides were made by, for instance, Euler
in describing the inviscid form of the governing equations. However, his
contributions present such levels of complexity, that even today, the Euler
equations carry open research questions.

Then, conceptual difficulties and challenges were summarised. Following this,

scale effects were described in some detail for deep and shallow water cases
sequentially. It was illustrated that phenomena observed in the latter cases can
exist in a generalised form within the former cases. Thus, it may be frequently
desirable to examine cases of greater complexity to enable a better
understanding of the physics. Shallow water flows past ships are

28
representative of such cases due to the amplification in relative importance of
nonlinear phenomena.

In line with the research aims and objectives, a survey of the literature was
performed. This suggested that scale effects in the context of ship resistance
originate from the nonlinear nature of the problems, which are treated
linearly. This is true for both deep and shallow waters. To overcome this
obstacle, one could perform simulations directly in full-scale. However, as
demonstrated by the literature, validation data for such cases is scarce. To
make matters worse, the computational cost associated with full-scale
simulations currently tends to be high.

Several aspects requiring further research were identified. These can be

summarised as follows, and are addressed in subsequent chapters:

• Scale effects on the wave resistance coefficient, form factorhave not

been investigated simultaneously in any previous study.
• Interactions between different components of ship resistance,
specifically wave and frictional resistance, have not been investigated
using a RANS solver in previous studies.
• Turbulence dependence is a source of ambiguity, therefore requiring a
separate investigation.
• Restricted water effects on the scaling of ship resistance have not been
examined in sufficient detail.
• Ship waves modelled via RANS solvers have not been extensively
compared against other methods.
• Cases where a ship advances over a change in the water depth have not
been investigated using a RANS solver.

29
 A GEOSIM ANALYSIS OF SHIP
RESISTANCE DECOMPOSITION
AND SCALE EFFECTS WITH THE AID
OF CFD

This chapter examines the scale effects on a ship in deep, unrestricted waters.
The analysis is performed for the KCS hull in three different model-scale
factors, forming a geosim series, which culminates in a full-scale set of
simulations. The specific condition to which the analysis is performed reflects
the operational speed of the ship. Simulations are performed in multiphase
and double body regimes with and without sinkage and trim. Moreover, the
viscous scaling approach is applied to gauge its performance when compared
to the linear scaling method.

3.1 Introduction

In 2011, the International Maritime Organisation (IMO) introduced the first

mandatory measure since the Kyoto Protocol (UNFCCC, 1998) to improve the
energy efficiency of ships and accelerate innovation in the maritime sector
(IMO, 2011). This regulation requires every ship that has been built after
January 1st, 2013 to be certified using the Energy Efficiency Design Index
(EEDI). EEDI can be broadly thought of as a measure of the energy efficiency
per tonne-mile. As such, ship resistance is one of the primary parameters used
in its calculation. Hou et al. (2019) discussed the two sources of uncertainty
that arise when calculating the EEDI of a ship: aleatory uncertainty, which
exists objectively due to the operating environment, for example; and

30
epistemic uncertainty, which relates to lack of knowledge. Here, the latter is
examined only.

The source of epistemic uncertainty lies with the complexity of estimating ship
resistance – a problem that remains unsolved despite the fact that scientists
have been attempting to overcome it for several centuries. For instance,
Gotman (2007) reported that both Newton and Euler had devised
approximations based on different mathematical approaches. Even with the
advent of Computational Fluid Dynamics (CFD) methods, one cannot
guarantee that a calculated value for the resistance of a ship will match
experimentally obtained data. Instead, the uncertainty is estimated and
reported on, for knowledge of the exact value of the error would allow us to
simply correct results accordingly. For this reason, expensive experiments are
routinely performed in towing tanks around the world.

While one is free to geometrically scale down a ship to a convenient size, the
physical properties of water do not change (Tropea et al., 2007). Therefore, the
troubles of the naval architect do not end once the experiment has run its
course. One may only keep the ratio of inertial and viscous forces (the
Reynolds number – Re), or the ratio of gravitational and inertial forces (the
Froude number – Fr) the same between model and ship (Lee et al., 2018).
Extrapolation procedures have been devised to keep these ratios, the earliest
by William Froude in the 1870s (Molland et al., 2017). Theoretically, these
allow us to predict the resistance of the full-scale ship using a model
experiment in any scale.

This chapter will attempt to establish a better understanding of the epistemic

uncertainty involved in calculating full-scale ship resistance using a Reynolds
Averaged Navier-Stokes (RANS) solver. To achieve the task at hand,
experimental data for the well-known KCS hull form in three different scale
factors were collected. At each scale factor, numerical simulations are
performed in three different ways to predict the components of bare hull
resistance. Specifically, the ship is scaled geometrically and simulated in both
double body and multiphase conditions. Additionally, by modifying the value
of dynamic viscosity only, the ship’s Reynolds number is changed to match its
respective value at higher scale factors without a change in characteristic
length.

The novelty in this part of the thesis is expressed in the unique approach
adopted to predicting ship resistance. While a plethora of researchers have

31
examined scale effects in the present context, to the best of the author’s
knowledge, none have performed this with a variety of methods incorporating
all physical phenomena, simultaneously validating numerical predictions
against experimental data. Specifically, the adopted methodology enables
prediction of interactions between the linearly decomposed components of
ship resistance and examine scale effects on each individually.

Of particular interest is wave resistance, which is typically assumed scale

invariant at a constant Froude number. Although the work of Raven et al.
(2008) suggested otherwise, this assumption is still applied widely. In this
context, it will be attempted to confirm and further examine the presence of
scale effects in wave resistance in an attempt to stimulate more research in this
area.

The remainder of chapter first justifies the adopted research methodology in

section 3.2, and gives the ship geometry and examined conditions in section
3.3. Then, the numerical modelling is presented in section 3.4, before the
resulting data is shown in section 3.5, accompanied by a discussion on its
significance. Finally, section 3.6 provides summary and conclusion.

3.2 Methodology

The adopted methodology revolves around the capabilities of Star-CCM+, an

extensively used, validated, and verified commercially available RANS solver
(Siemens, 2018). Making this choice allows the exploitation of the versatility
inherent in numerical simulations. In reality, one is limited to performing
‘multiphase experiments’, unless the underwater shape of a ship and its mirror
image are tested in a wind tunnel. The wider literature suggest that such
experiments are extremely rare, even in an academic context (Lee et al., 2003).
However, numerically, not only is this a possibility, but it is an approach
capable of substantially accelerating convergence characteristics of the
numerical solution.

In particular, some aspects of a simulation requiring attention are the decay of

ship motions (in steady state cases) and the convergence of the wave field. In
other words, the wave pattern must become invariant with respect to the ship.
Invoking the double body assumption implies replacing the free surface with
a symmetry plane as well as eliminating the rigid body motions of the ship.
This is thought to be the main reason why the earlier studies cited in Chapter
2 use the double body approach: insufficient computational power to fully
simulate the flow field and ship motions. Alternatively, researchers have

32
sought to predict the form factor via double body simulations. Naturally,
deviating from the actual physics of the problem examined renders
predictions less reliable.

Keeping in mind the inherent consequences present in performing simulations

under the double body assumption, it should also be mentioned why applying
it here is beneficial. The resistance decomposition of a ship according to the
ITTC (2017) is shown in Eq. (3.2):

𝐶𝑇 = (1 + 𝑘)𝐶𝐹 + 𝐶𝑊 (3.2)

Where (1+k) is the form factor, while 𝐶𝑇 , 𝐶𝐹 , and 𝐶𝑊 are the total, frictional, and
wave resistance coefficients, respectively. These constitute the measured force,
non-dimensionalised by 0.5𝜌𝑆𝑉 2 , where 𝜌 (997.561 kg/m3) is the freshwater
density, S is the wetted surface area in m2, and V is the ship speed in m/s.

Replacing the free surface with a non-deformable symmetry plane renders

𝐶𝑊 = 0, therefore Eq. (3.2) becomes 𝐶𝑇 = (1 + 𝑘)𝐶𝐹 = 𝐶𝑉𝑃 + 𝐶𝐹 , where 𝐶𝑉𝑃 is
the viscous pressure resistance coefficient. This matches the resistance
definition in all RANS solvers. Namely, the total resistance is the sum of
tangential and normal components. In multiphase simulations, RANS solvers
compute the total as 𝐶𝑇 = 𝐶𝑃 + 𝐶𝐹 , where 𝐶𝑃 is the pressure resistance. Clearly,
all components are interrelated and calculating each presents its own
challenges. Here, relying on methods can be avoided, whether potential (in the
case of 𝐶𝑊 ) or otherwise (for (1+k) and 𝐶𝐹 ), to predict their value by defining
the wave resistance coefficient as shown in Eq. (3.3) and form factor in Eq.
(3.4):

𝐶𝑊 = 𝐶𝑇multiphase − 𝐶𝑇double body (3.3)

(1 + 𝑘) = 𝐶𝑇double body ⁄𝐶𝐹double body (3.4)

The definitions of Eq. (3.3) and (3.4) are used in to predict scale effects on each
component of ship resistance.

Additionally, seeking to confirm the assertion that sinkage and trim do not
contribute much to the total resistance in deep waters, double body
simulations are run in both level and translated conditions, the latter matching
the multiphase orientation of the ship. To isolate Reynolds number effects on
resistance coefficients and the form factor, a single Froude number is used (Fr
= 0.26, which corresponds to the service condition), chosen to match available
experimental data.

33
Finally, an explanation is owed to the assertion that a change fluid viscosity
can be used to scale the ship without altering its dimensions. The term ‘viscous
scaling’ is used to describe the altering of the value of dynamic viscosity (μ)
that changes the Reynolds number. This approach was adopted following
Haase et al. (2016), where it was applied on a catamaran. The method relies on
verifying the mesh in model-scale versus experiments and changing the
viscosity value to match the full-scale Re. Here, this is performed in steps,
which match the examined scale factors to more accurately gauge its
performance. It is only applied on the double body simulations after the
orientation of the ship has been adjusted according to the running trim and
the sinkage, calculated using the multiphase approach.

Naturally, the most economical simulations are sought in terms of

computational power and time. For this reason, viscous scaling is applied to
the hull form in the 75th scale. Therefore, the values for μ shown in Table 3.1
correspond to those needed to make the Re (75) match Re (52.676), Re (31.599) and Re
(1), where the bracketed superscripts indicate the scale factor (Re (λ)). To account

for the geometric scaling, all results are multiplied by λ3 (Haase et al., 2016b).
To elaborate, the resistance values are multiplied by the ratio of scale factors,
raised to the third power. Specifically, using λ=75→ λ=31.599 as an example,
after modifying μ according to the value shown in Table 3.1 (μ(31.599) = 2.431×10-
4 Pa-s) each constituent component of the measured resistance would be

multiplied by (75/31.599)3=13.371 at the end of the simulation. This procedure

is applied analogously at each scale factor examined.

3.3 Ship geometry and conditions

There are many examples of research that attempt to resolve the ship scaling
problem. Some significant contributions to the field were discussed in Chapter
2. Here, the aim is to elaborate on the selected case-studies.

The best way to determine scaling effects is to test the same ship at different
scales while measuring parameters of interest. By doing this, the scope is
limited to a single hull form, and therefore, it would put into question the
generalisations that one might be tempted to make about other hulls forms.
Still, the literature on this subject suggests that incorporating multiple hull
forms is not advantageous. For instance, García-Gómez (2000), Min and Kang,
(2010), and Lee et al. (2018) showed scattering in results produced by different
hull forms, even when considering non-dimensional quantities, such as the
form factor. Although one is by no means insured against the same outcome,

34
it is maintained that a more in-depth study is possible considering a single hull
form.

The only choices worth considering in detail in terms of hull forms are those
created for numerical benchmark purposes. From the large number of
experiments conducted on them in different scale factors, the most attractive
proves to be the KCS, because resistance measurements (for Fr=0.26) have been
performed in a scale factor (λ) of 75 by Shivachev et al. (2017), 52.667 by
Simonsen et al., (2013) and 31.599 by Kim et al. (2001). Table 3.1 shows the
principal dimensions and case-studies in each scale. Here, the value of
dynamic viscosity in the 75th scale is in bold to highlight it as the default value
for the simulations. By contrast, viscous scaling has been performed in the
places where other values are listed.

Table 3.1. KCS Principal characteristics and case-studies

Quantity Symbol Value Unit
Scale factor λ 1 31.599 52.667 75 -
Length L 230 7.279 4.367 3.067 m
Beam B 32.2 1.019 0.611 0.429 m
Depth D 19 0.601 0.361 0.253 m
Draught T 10.8 0.342 0.205 0.144 m
Displacement ∇ 51990.120 1.649 0.356 0.123 m3
Block coefficient CB 0.6505 0.651 0.651 0.651 -
Wetted area with
rudder S 9539 9.553 3.439 1.696 m2
Longitudinal centre
of gravity LCG 111.603 3.532 2.119 1.488 m
Vertical centre of
gravity VCG 7.28 0.230 0.138 0.097 m
Metacentric height GMT 0.6 0.019 0.011 0.008 m
Velocity U 12.350 2.196 1.702 1.426 m/s
Froude number Fr 0.26 0.26 0.26 0.26 -
Reynolds number Re 3.188×109 1.794×107 8.342×106 4.909×106 -
Dynamic viscosity
for scaling μ 1.368×10-6 2.431×10-4 5.229×10-4 8.887×10-4 Pa-s

Selecting the KCS as the case study for this chapter and the remainder of the
thesis is justifiable for two reasons. Firstly, one ought to use realistic hull forms
for numerical investigations. For instance, if an alternative benchmarking hull
were to be selected, for example, the parabolic hull, the results may not be
equally applicable, restricting the possible findings and discussions. This is the
case due to the fact that the Wigley hull is a simple shape, and lacks some

35
prominent features many ships possess. Specifically, the bulbous bow, a
concave-convex variation in the surface curvature between the bow and stern,
as well as a flat bottom. The second aspect, worth considering is the wealth of
available experimental data. These were hinted at earlier in this chapter, and
will be detailed in the results sub-section. Moreover, the numerous
experiments performed with the KCS allow for validation potential in the
remaining chapters of this thesis as well. Therefore, the KCS is considered an
ideal case for the purposes of this thesis.

As mentioned earlier, a single Froude number is used. This approach is

adopted in order to isolate Reynolds number effects in the examined
parameters. The Reynolds number changes several orders of magnitude
between the smallest and largest scale factors used herein. It is therefore
anticipated that the Reynolds number would produce the most noticeable.
While Froude number effects may also be present in the scaling of ship
resistance, their effect is not forecast to be as significant as that stemming from
the Reynolds number change, since the Froude number solely changes by a
considerably smaller amount.

3.4 Numerical set-up

In this section, the relevant details regarding the numerical setup are
discussed. Prevalence is given to parameters likely to affect the computed
results. Where possible, a discussion is included on the potential effect each
decision could have.

3.4.1 Physics modelling

The RANS solver used, Siemens’ Star-CCM+, version 13.4.011, employs a

Finite Volume Method (FVM) to discretise the integral form of the Navier-
Stokes equations. Continuity and momentum are linked via a predictor-
corrector scheme.

As mentioned in Chapter 2, turbulence modelling is suspected to play a crucial

part in scale effects. Here, the realisable k-ε turbulence model with the all y+
wall treatment is used. There are several advantages of selecting this two-
equation turbulent kinetic energy-dissipation model. Primarily, interest lies in
accuracy and economy. The findings of Larsson et al. (2014) indicate that there
is no discernible change in resistance predictions with more complex models.
On the other hand, Salim and Cheah (2009) performed systematic RANS
simulations using ANSYS on the frictional resistance of a flat plate and

36
compared their results with a variety of turbulence models. The k-ε model was
shown to deliver better predictions than the k-ω model, provided that the
viscous sublayer is resolved. In other words, the y+ value should be lower than
1. Indeed, the accuracy of the result depends strongly on the ability to maintain
y+<1.

Eça et al. (2015) showed that values of y+ within the buffer zone (5<y+<30) may
lead to an error in the region of 10% when calculating the frictional resistance.
For grids featuring cells within the viscous sublayer (y+<1), a single difficulty
is reported by both Eça et al. (2015) and the ITTC (2011): large number of high-
aspect-ratio cells, which make convergence problematic. In other words,
resistance predictions are vastly superior in terms of accuracy when y+<1 is
maintained over the wetted area of the ship, as long as convergence is not
compromised. Stern et al. (2013) reported that a challenge for full-scale
computations would be the number of near-wall cells. Indeed, Piomelli and
Balaras (2002) predicted that wall-normal cell numbers would have to vary
with Re0.4 only within the outer boundary layer. Thus, the y+ value is allowed
to exceed unity in full-scale due to the prohibitively large cell number needed
otherwise. The impact of this choice is yet unknown, requiring further
investigation.

The k-ε turbulence model, widely used for full-scale flows (Schweighofer,
2004; Tezdogan et al., 2016b), is also advantageous due to its relative
computational economy. Quérard et al. (2008) found that a reduction of up to
25% in computational time is possible compared to more sophisticated
models. Furthermore, the simulations of Simonsen et al. (2013) were
performed using the k-ω model in Star-CCM+ and CFDSHIP-IOWA, which
will allow comparison of turbulence models.

In the case of multiphase simulations, the Volume of Fluid (VOF) method,

introduced by Hirt and Nichols (1981), is used to model the free surface, and
the movement of the water. This is done via the flat wave concept. The VOF
method describes the two phases by assigning a scalar value of 0 to air and 1
to water. The interface between the two, i.e. cells containing equal parts of air
and water attain a value of 0.5. Therefore, the inlet and outlet boundaries have
a field function associated with them. At the inlet, the velocity and direction
of the flat wave is specified, whereas the outlet is set to maintain the
hydrostatic pressure. The VOF model depends on both fluids accounting for
large parts of the domain, while maintaining a relatively small contact area.
For the double-body simulations, the movement of water cannot be modelled

37
in such a way. Instead, the velocity is defined at the inlet boundary, whereas
at the outlet, a pressure of 0 Pa is preserved.

The segregated flow model is used to solve the equations of state in an

uncoupled manner. In all simulations, the convective terms are solved via a
second order scheme, while the overall solution is obtained using a SIMPLE
algorithm. The dynamic trim and sinkage of the ship is captured using a
Dynamic Fluid-Body Interaction (DFBI) model, where only heave and pitch
modes of motion are allowed. More information on the aforementioned
choices may be found in Siemens (2018).

3.4.2 Time step selection

The Courant-Friedreichs-Lewy (CFL) number is sometimes used as a

condition to assess the convergence of simulated flows. It expresses the idea
that if a flow is moving across a discrete spatial grid, one must choose a
suitable time step 𝛥𝑡 and spacing 𝛥x to guarantee that the properties of the
fluid (velocity, pressure) are solved for at each grid point. In such a case,
CFL≤1. The ITTC (2011) recommends the use of 𝛥𝑡=0.005~0.01L/U, where L
and U are the ship’s length and speed, respectively. Tezdogan et al. (2016) and
Terziev et al. (2018) demonstrated that the use of 𝛥𝑡=0.0035L/U showed little
error due to the discretisation of the temporal term in the Navier-Stokes
equation, which was set as first order, and is therefore chosen as well.

3.4.3 Mesh generation

Mesh generation was performed using the automatic facilities of Star-CCM+,

which allows the user to make use of several operations. The trimmed cell
mesher is used to fabricate predominantly hexahedral cells. The alternative,
using tetrahedral cells, has been shown to deliver unreliable results by Jones
and Clarke (2010). The near-wall cells were generated using the prism layer
mesher, which is used to fabricate orthogonal cells near the hull surface. Star-
CCM+ automatically places these cells using a geometric progression to
determine their dimensions. The prism layer thickness is set to equal the
approximate value of the turbulent boundary layer thickness, derived by the
1/7 power law, found in White (2006).

38
 Figure 3.1. Domain boundary conditions and dimensions for the typical
multiphase simulations

3.4.1 Computational domain and boundaries

According to Date and Turnock (1999), the boundary conditions of a CFD

simulation play a critical role in both the accuracy and convergence of the
solution. Their position as relative to the ship is equally important because, in
rare cases, wave reflections may occur and that would invalidate the solution.
To insure against this, the computational domain is constructed following the
recommendations of the ITTC (2011). The resulting dimensions and boundary
conditions are shown in Figure 3.1. A numerical beach model, the VOF
damping length, is also applied to the outlet boundary equal to approximately
1.24L in each scale. This allows the wavefield to develop prior to reaching the
damping zone, and guarantees no reflections will occur.

39
 Figure 3.2. 3-D view of the generated mesh. Depicted: Full-scale

Only half of the ship is simulated, thus, a symmetry plane is applied to all case
studies to alleviate the computational load and allow a larger number of cells
to be used. The side boundary is set as symmetry because it does not allow
changes in velocity or pressure across it to occur, i.e. it approximates an
infinitely wide, deep sea. In any case, it is reasonable not to expect the Kelvin
wake to reach the side boundary. The computational domain, shown in Figure
3.2, is scaled with the ship linearly to minimise numerical scale effects. The
resulting three-dimensional grid properties are shown in Table 3.2. Here, the
cell count for the translated double body simulation in λ=75 have been
highlighted to indicate their use in the viscous scaling procedure.

Table 3.2. Number of cells in each scale

Scale 1 31.599 52.667 75
Multiphase 20,554,263 12,343,685 5,832,169 3,338,447
Number of cells Double body: level 5,166,585 3,750,965 2,703,735 1,615,244
Double body: translated 5,532,073 3,491,712 2,739,160 1,632,931

3.4.2 Time-history of the numerical solution

Calm water ship resistance is a steady state-problem. In other words, the

solution is not affected regardless of how long the computation is performed
for. However, in CFD, the ship experiences a well-known shock at the
beginning of the simulation (Mucha, 2017). This can be mitigated by
hydrostatic balancing (also referred to as ‘equilibrium’), but is not
implemented here to avoid contaminating the solution with deficiencies of

40
numerical algorithms. According to Siemens (2018), the equilibrium option is
a purely mathematical procedure that has no basis on physical rigid body
motion. The alternative offered by the RANS solver consists of ‘free motion’,
where the body is translated and rotated as a result of the forces acting on it at
each time step, and is implemented here.

To reduce the shock effect, the ship is constrained in all directions for the first
5 seconds of the simulations. In Figure 3.3, the time-history is illustrated for
the three components of resistance for λ=31.599. A ramp time of 10 seconds is
also adopted, which gradually applies the forces on the hull to help reduce
oscillatory motions. The combined effect of these two settings can be seen in
Figure 3.3.

Transient effects in CFD have a potential to invalidate the results, which is

why, as explained earlier, the domain boundaries must also be placed suitably.
Figure 3.3 also demonstrates that the placement of the boundaries in the
computational domain was successful, as well as the adequate decay of
transient solution characteristics.

Figure 3.3. Example time-history of resistance and sinkage for case λ=31.599

To evaluate the iterative errors, the procedure of Roy and Blottner (2006) was
used on the total resistance coefficients in multiphase and double body

41
regimes. The analysis showed an absolute error of approximately 5.504×10-4 %
and 2.112×10-6 % for multiphase and double body 𝐶𝑇 values, respectively,
which was representative of other cases as well. These are used to justify the
use of the numerical verification procedure, which assumes that iterative
errors must be at least two orders of magnitude smaller than discretisation
errors. In any case, all final values reported here are averaged over the last 25
seconds to ensure the final value is affected as little as possible by the iterative
error.

Convergence properties can also be judged based on residuals. The residual of

a numerical scheme can be broadly thought of as the (usually scaled)
difference between the iteratively approximated solution and the perfect
conservation of mass and momentum (ITTC, 2014). Typically, these are
required to reduce by several orders of magnitude for the simulation result to
be accepted. However, this is strictly dependent on how close the initial state
of the simulation is to satisfying perfectly the discretised form of the governing
equations (Siemens, 2018). This true is because if the initial state of the
simulation is very close to satisfying the laws of conservation perfectly, the
residuals will not reduce at all (Siemens, 2018). In other words, residual time-
histories should not be used on their own to assess for convergence, although
they are a powerful tool in any numerical analysis problem. Furthermore, the
observed reduction in residual values is highly case-specific: a large reduction
can still lead to a high validation error.

The double body residuals were found to reduce quickly by about five orders
of magnitude (to 10-5 ~ 10-6) within approximately the first 4000 iterations (the
ITTC's (2011) recommendation is a reduction of three orders of magnitude).
This is a highly attractive feature of this type of set-up. It is also the reason
why ship CFD simulations were performed in double body mode in the early
days of the field. In the case of multiphase simulations, all residuals decreased
by two orders of magnitude within the first 8000 iterations (considerably later
than in the double body case). Reducing the magnitude of residuals further
proved difficult. This is not considered problematic, as stated by the ITTC
(2011) A representative case for the residual behaviour in both multiphase and
double body modes is shown in Figure 3.4 (λ=31.599).

42
 Figure 3.4. Residual time-history (“Water” refers to the volume fraction
residual). Depicted: λ=31.599

In all cases, following the end of the transient oscillatory motion, residuals did
not exhibit signs of reducing further. This is because the flow had fully
developed at this stage. This is not problematic per se, because ship CFD –
especially towed, calm water predictions – are not characterised by large ship
motions or deformations of the free surface. Thus, while it is desirable to
achieve a magnitude of the residuals that is as low as possible, a small
reduction does not imply a ‘bad’ solution by itself.

According to ITTC (2014), even if the recommended three order of magnitude

residual reduction cannot be achieved, integral values can be used to assess
convergence of the solution, in particular, forces and moments acting on the
hull (this is primarily in view of the complexity associated with the numerical
simulation of ship flows). This was done earlier in this section in the case of
iterative errors in resistance for λ=31.599. The following section presents the
numerical verification study, which expands on the expected uncertainties
due to spatial and temporal discretisation.

43
 3.4.3 Verification study

It is inevitable to induce errors when temporally or spatially discretising the

Navier-Stokes equations. Since the governing equations cannot be modelled
continuously, it is assumed that errors decay rapidly as Δt and Δx decrease. In
other words, the continuous equations should match the discretised versions
as Δt, Δx→0.

The current method of estimating uncertainty in CFD simulations is based on

expanding the error as a power series with integer powers of Δt or Δx (Xing
and Stern, 2010), introduced by Richardson (1911). The four types of
conditions which govern whether a solution is convergent or divergent as Δt
and Δx are refined can be summarised as follows:

1. Monotonic convergence: 0 < R < 1

2. Oscillatory convergence: R < 0; |R| < 1
3. Monotonic divergence: R > 1
4. Oscillatory divergence: R < 0; |R| > 1

For conditions 3 and 4, neither error nor uncertainty can be estimated. Here, R
is the convergence ratio, defined in Eq. (3.5):

𝑅𝑘 = 𝜀21 ⁄𝜀32 (3.5)

where 𝜀21 is the difference between the medium (𝑓2 ) and fine (𝑓1 ) solutions,
while 𝜀32 is the difference between the coarse (𝑓3 ) and medium (𝑓2 ) solutions.
These (𝑓1,2,3) are obtained by systematically coarsening (by using the
refinement ratio r = √2, recommended by the ITTC, (2008)) the respective input
parameter – time-step or mesh (Stern et al., 2006). In other words, the base size
is multiplied by √2 while maintaining the smallest time step. The resulting cell
counts are 1,674,346 and 880,876 for the medium and coarse mesh study for
λ=75, respectively. The same procedure is applied to the time step, which is
lessened by the same factor on the finest grid to isolate errors due to changes
in temporal discretisation.

Since r is maintained constant, the order-of-accuracy (p) takes the form of:

𝑝𝑘 = ln(𝜀𝑘32 ⁄𝜀𝑘21 )⁄ln (𝑟) (3.6)

21
Thus, one arrives at the extrapolated value (𝑓𝑒𝑥𝑡 ), according to Celik et al.
(2008):
21
𝑓𝑒𝑥𝑡 = (𝑟 𝑝 × 𝑓1 − 𝑓2 )/(𝑟 𝑝 − 1) (3.7)

44
Next, the approximate relative error, defined in Eq. (3.8), and extrapolated
relative error, defined in Eq. (3.9), can be estimated.

𝑒𝑎21 = |(𝑓1 − 𝑓2 )/𝑓1 | (3.8)

21 12 12 |
𝑒𝑒𝑥𝑡 = |(𝑓𝑒𝑥𝑡 − 𝑓1 )/𝑓𝑒𝑥𝑡 (3.9)

Table 3.3. Grid convergence for trim and total resistance coefficient
Trim at CoG (with monotonic 𝐶𝑇 (with monotonic
convergence) convergence)
r √2 √2
𝑓1 0.1642 4.2908×10-3
𝑓2 0.1653 4.2794×10-3
𝑓3 0.1693 4.2454×10-3
R 0.2963 0.3600
p 3.5102 2.9480
21 0.1637 4.2972×10-3
𝑓𝑒𝑥𝑡
𝜀𝑎21 (%) 0.6914 0.2673
21
𝜀𝑒𝑥𝑡 (%) 0.3010 0.1486
21
𝐺𝐶𝐼𝑓𝑖𝑛𝑒 (%)0.3638 0.1879

Finally, the grid convergence index (GCI) can be calculated, shown in Eq.
(3.10). This marks the end of the error estimation, mentioned earlier.
21
𝐺𝐶𝐼𝑓𝑖𝑛𝑒 = 1.25𝜀𝑎21 ⁄(𝑟 𝑝 − 1) (3.10)

Table 3.4. Time step convergence for trim and total resistance coefficient
Trim at CoG (with monotonic 𝐶𝑇 (with monotonic
convergence) convergence)
r √2 √2
φ1 0.1642 4.2908×10-3
φ2 0.1612 4.3002×10-3
φ3 0.1566 4.3315×10-3
R 0.6716 0.2983
p 1.1483 3.4904
21 0.1703 4.2868×10-3
𝑓𝑒𝑥𝑡
𝜀𝑎21 (%) 1.8436 0.2180
21 3.6346 0.0927
𝜀𝑒𝑥𝑡 (%)
21
𝐺𝐶𝐼𝑓𝑖𝑛𝑒 (%) 4.7147 0.1158

45
Table 3.3 and Table 3.4 collectively show the numerical uncertainty is bound
within acceptable limits (arbitrarily set as 5%). The adopted numerical set-up
is more sensitive to changes due to the time step than due to cell number
variations in the case of trim, as shown in Table 3.3 and Table 3.4. The results
suggest the opposite is true for the computed total resistance coefficients. The
suspected cause for this is the turbulence model: even in the coarse mesh
condition, the y+≲1 condition is not deviated from, where the k-ε turbulence
model is known to perform well. It is not difficult to see that the y+ condition
has been maintained because the number of wall-normal cells, as well as the
thickness of the prism layer are independent of the base size, although the cell
aspect ratio changes dramatically as the mesh is coarsened. On the other hand,
trim is tied to the dynamic behaviour of the ship, where the lessening of Δt has
greater potential for impact. This, coupled with the uncertainty described
previously regarding the effect turbulence modelling has on sinkage and trim
is thought to be the root cause of the elevated 𝐺𝐶𝐼 indices for trim.

3.5 Results and discussion

In this section, the obtained results are presented and compared against
experimental values where possible. At the onset of this chapter, the
parameters of interest were defined as the total resistance and its constituent
components. Although prevalence will be given to these, running trim and
sinkage are also considered.

3.5.1 Error evaluation

Experimental data were collected for three different scale factors. For λ=75,
Shivachev et al. (2017) performed both tests, and CFD computations on the
KCS without appendages. Simonsen et al. (2013) modelled the rudder, but not
the propeller of the KCS in λ=52.667. Kim et al. (2001) performed experiments
on the KCS without the rudder or propeller, and did not allow the ship to trim
or sink. While the aim is to present results that are as realistic as possible, it is
not possible to satisfy all three experimental set-ups. The decision made for
the purposes of this thesis is that the ship should be modelled with the rudder,
since it has the potential to modify substantially the flow field near the stern.
On the other hand, the propeller was omitted so as to avoid deviating too
much from the experiments, the majority of which omitted the propeller.
Finally, as explained in the previous section, the DFBI method is used to model
sinkage and trim.

46
Table 3.5 shows the numerical and experimental total resistance coefficients.
The error in λ=31.599 is based on the results of Kim et al. (2001). In all model
scales, the error is bounded within an acceptable limit (3% deviation). An
easily observed trend is that 𝐶𝑇 values are always under predicted, albeit
slightly, by the adopted numerical set-up.

In the cases of sinkage and trim, the numerical calculations show a greater
scatter. In λ=75, the trim was predicted within 1.3% of the experimental value
(Table 3.6), which the present CFD set-up suggests remain largely unchanged
as the ship is scaled. However, the experimental results show a greater trim in
λ=52.667. The opposite seems to be suggested by the limited sample points for
sinkage, which has been non-dimensionalised by ship length, shown in Table
3.7. Here, the results predict a reduction in sinkage with scale. Despite the
aforementioned errors in trim and sinkage, the results presented suggest the
principle of dimensional similarity holds to a greater degree than experiments
do.

The errors reported could stem from a variety of sources. In model-scale

experiments, turbulence stimulators are mounted near the bow of the ship.
Thus, the numerical and experimental flows around the hull in each scale
would have exhibited different turbulence characteristics. Furthermore, some
variability is expected in the approach to turbulence stimulators between
testing facilities. While their effect cannot be accounted for in sinkage and trim,
the induced parasitic drag is subtracted. As reported in Larsson et al. (2014),
turbulence modelling has a pronounced effect on sinkage, and one could,
therefore, speculate that it could also have a similar effect on the trim.
Factoring in the small uncertainty reported in each experimental study, it can
be said that the resistance values have been predicted very well.

Although an acceptable level of error was established in all model-scale case

studies examined, it would be wrong to generalise this to full-scale. Indeed,
without full-scale measurements one should be wary of making such a claim.
The choice of the adopting the KCS provided the possibility of comparing
against experiments at each model-scale λ. However, it also prohibits us from
attempting to carry this forward to full-scale since no real, full-scale equivalent
of the KCS exists. Therefore, the full-scale wave resistance coefficients and
form factors for the highest Reynolds number should be considered as an
estimation rather than a concrete prediction. Furthermore, it is not practical to
achieve y+ ≲1 in full-scale due to the prohibitively large number of cells doing
so would entail. Hence, the numerical set-up, where it can confidently be

47
claimed that the results are accurate, strictly speaking, is not identical to the
implemented full-scale set-up. Keeping this in mind, there are sufficient
grounds and indeed verifiable data samples in model-scale to justify every
conclusion drawn in the following sub-sections.

Table 3.5. Numerical and experimental total resistance coefficients, Fr = 0.26

Scale Software Experime Error
Method or source Re Numerical
(λ) package ntal (%)
4.32×10-3 2.041
Shivachev et al.
4.3×10 -3 2.494
(2017), no appendages
75 4.909×106 Star-CCM+ 4.23×10 -3 4.41×10-3 4.023
The current CFD:
4.291×10-3 2.703
multiphase
CFDSHIP-
4.07×10-3 5.568
Simonsen et al. (2013) IOWA
52.667 8.342×106 4.42×10-3 4.31×10-3 -2.552
The current CFD:
4.232×10-3 1.818
multiphase
The current CFD:
31.599 1.794×107 Star-CCM+ 3.51×10-3 3.557×10-3 1.312
multiphase
Tezdogan et al. (2015) 2.295×10-3 - -
The current CFD: 1 3.188×10 9
2.294×10-3 - -
multiphase

Table 3.6. Numerical and experimental trim, Fr = 0.26

Scale Software Numerical Experimental Error
Method or source Re
(λ) package (deg) (deg) (%)
Shivachev et al. 0.198 -22.222
(2017), no 0.198 -22.222
appendages 75 4.909×106 Star-CCM+ 0.195 0.162 -20.370
The current CFD:
0.164 -1.330
multiphase
Simonsen et al. CFDSHIP-
0.178 3.940
(2013) IOWA
Simonsen et al. 52.66
8.342×106 0.18 0.185 2.860
(2013) 7
The current CFD:
0.165 10.708
multiphase
Star-CCM+
The current CFD: 31.59
1.794×107 0.163 - -
multiphase 9
The current CFD:
1 3.188×109 0.153 - -
multiphase

48
Table 3.7. Numerical and experimental sinkage/length, Fr = 0.26
Software Error
Method or source Scale (λ) Re Numerical Experimental
package (%)
Shivachev et al. -1.957×10-3 14.286
(2017), no -1.950×10-3 14.571
appendages 75 4.909×10 6 Star-CCM+ -1.924×10-3 -2.283×10 -3
15.714
The current CFD:
-1.890×10-3 17.218
multiphase
Simonsen et al. CFDSHIP-
-2×10-3 4.762
(2013) IOWA
Simonsen et al.
52.667 8.342×106 -2.706×10-3 -2.100×10-3 -9.524
(2013)
The current CFD: -1.897 ×10-
-9.667
multiphase 3
Star-CCM+
The current CFD:
31.599 1.794×107 -1.907×10-3 - -
multiphase
The current CFD:
1 3.188×109 -1.902×10-3 - -
multiphase

3.5.2 Resistance decomposition

Here, the form factor approach is adopted, because it contains one parameter
more than the Froude’s approach on which scale effects can be examined: the
form factor itself. As explained in earlier, by making use of the double body
approach, there is no free surface (𝐶𝑊 =0) which modifies the form factor
equation to yield (1+k) = 𝐶𝑇 /𝐶𝐹 . In this form, (1+k) can be calculated directly
using CFD (under the double body assumption), whether scaled
geometrically, or via a modification to the value of dynamic viscosity (μ).

An initial estimate of the form factor during the early design stage can be as
valuable as a precise calculation down the line. Therefore, three empirical
relations to predict (1+k) have been adopted from Molland et al. (2017).
Alongside these, the corrections of García-Gómez (2000) and Min and Kang
(2010) are applied.

3.5.2.1 Wave resistance

While Reynolds number effects on the form factor are practically uncontested,
scale effects on wave resistance are largely unexamined. Doctors et al. (2007)
suggested a ‘wave resistance form factor’ (1+kw), but did not provide a
recommendation regarding its value. Doctors (2007) and Lazauskas (2009)
defined this as |kw| < 1, and kw < 0: a fundamentally different concept from the
‘traditional’ form factor. Unfortunately, a sufficiently large dataset for the KCS
to establish a value for kw using a regression procedure is not available. Even if

49
that were the case, many hull forms must be examined under the same criteria
to achieve a meaningful estimation applicable in general.

Figure 3.5. Wave resistance coefficients.

Figure 3.5 shows the predicted wave resistance coefficients for each scale
according to the three different methods using CFD (calculated as the
difference in total resistance of the multiphase and double body simulations).
Here, the Slender-body prediction, calculated using Bentley’s Maxsurf
Resistance software, is represented by a flat line because it is invariant with a
change in Reynolds number. This figure presents the first indication that it
may not be possible to achieve a smoothly varying curve for which general
predictions can be made when large changes are applied to the Re. While all
methods agree the general direction of the curve in every examined scale, there
are disagreements between them at every examined Re. It is worth mentioning
that the first point of the viscous scaled values, i.e. λ=75, is the same as the
translated double body value.

50
 Figure 3.6. Wave cuts at y/L=0.1, 0.2, 0.3 for all examined scale factors. The
ship is located at 0 < x/L <1, whereas the flow is in the negative x direction.

Figure 3.7. Relative locations of the wave cuts. Depicted: λ=75.

The change in wave resistance coefficient implies a geometrically non-similar

wave pattern between scale factors. Here, these are not examined, as

51
identifying differences is difficult and subjective. Instead, to reveal differences,
wave cuts are used, shown in Figure 3.6, at three different locations next to the
ship. The relative positions of the wave cuts are shown in Figure 3.7. In all λ>1
examined, the curves are sufficiently close to make them largely
indistinguishable from each other. In full-scale, the differences become more
pronounced. This is the suspected source of differences in wave resistance
coefficient. It is interesting that changes are magnified as one moved further
away from the ship centreline, i.e. into the fully developed wake. In other
words, changes to the wave field with scale require time to propagate some
distance from the wave maker before differences become apparent. Therefore,
the geometric scaling of the ship’s bulbous bow, whose primary function is to
modify the wave field is responsible for the observed differences in the wake.
Indeed, the laminar portion of the boundary layer at the bulbous bow is
appreciably stronger in model scale than that at full scale, inevitably creating
discrepancies (Hochkirch and Mallol, 2000). On the other hand, the stern,
being the second main source of wave-making, together with the interaction
from the bow wave cause the dominant features observed between -2.5 < x/L
< 0 in Figure 3.6.

Albeit small, errors in the predicted total resistance coefficient and their
influence on the wave resistance should not be disregarded. Nevertheless,
once decomposed, it is reasonable to expect that these are distributed, to some
extent, among the constituent components of the total resistance. Thus, it could
be argued that the predicted error, spread over the frictional resistance, wave
resistance, and form factor, is insignificant (in the 3D extrapolation case).
Therefore, scale effects on wave resistance have been proven, conclusively
rendering it a function of both the Reynolds number.

In reality, the boundary layer’s interaction with the wave field, and therefore
wave resistance, is expected to resemble an iterative process, where the former
modifies the latter and vice versa. Although this is largely ignored, several
studies have documented such effects, for example Stern (1985). The effects of
free surface flow effects on flat plates were examined in Longo et al. (1998),
who reported changes in Reynolds stresses and mean velocity components.
Marquardt (2009) stated that wave effects on the boundary layer are
proportional to wave steepness and may influence the boundary layer up to
half a wavelength (in the -z direction). Suh et al. (2011) also identified
deformation of vortical structures as a consequence of the presence of a phase
interphase.

52
 3.5.2.2 Frictional resistance

To quantify free surface effects on the frictional component of resistance,

Figure 3.8 contains all CFD simulations performed as part of this chapter as
well as the friction lines used for extrapolation. Although not frequently used
in the literature, (in full-scale) the Telfer (1927) line passes just below the CFD
predictions, while the Grigson (1999) method slightly overpredicts the
multiphase result. The method proposed by Gadd (1967), on the other hand,
passes through the predicted frictional resistance coefficients. The ITTC’57 line
also provides a good approximation, underpredicting the values slightly.

The results of Figure 3.8 reveal that in the low Re range (high λ), free surface
effects are much more pronounced than in full-scale. There are several reasons
for this. Primarily, the boundary layer affected by the free surface grows much
more slowly than the skin friction of a fully submerged body. The latter
increases with the wetted area (λ2), while the former changes approximately
with the reciprocal of the 7th root of the Reynolds number. Clearly, the
boundary layer length will be equal to the length along the waterline, whereas
the vertical distance at which it is disturbed can be estimated by the
wavelength of the Kelvin wake. Methods to calculate this vary in complexity
and robustness. Such a computation is not attempted, instead leaving it as a
piece of future work. Such a study could seek to establish a relationship
expressing the exact volume of the disturbed boundary layer by the generated
waves. Simultaneously, the contribution of air resistance to the total friction
was calculated as less than 0.3% of in all cases using the multiphase
simulations.

Naturally, at low Froude numbers, the free surface is hardly disturbed (which
also forms the basis for calculating the form factor), and double body
approximations can provide good predictions (Landweber and Patel, 1979).
Since the relative importance of free surface effects on the boundary layer
decay rapidly with scale (similar to the displacement thickness), the
aforementioned approximation can be used successfully at full-scale without
significantly compromising the accuracy of the desired solution. That is,
provided one is equipped with a tool capable of estimating the wave resistance
separately with sufficient fidelity directly at full-scale.

Mitchell’s integral is used for many theoretical predictions of wave resistance,

and forms the basis for the slender body method used in Figure 3.5. Its
accuracy in full scale, however, is not as well documental as in model scale.

53
Gotman (2002) asserted that if the ship has a convex transom, the wave
resistance is over predicted. The opposite was suggested to be the case for
concave transoms, while if the waterline is straight, or near straight, the values
agree well with experiments in the range Fr < 0.29. The results presented
earlier suggest that the above is only true in λ=31.599 (the KCS transom is
convex). Thus, while confidence in numerical tools is largely lacking, it should
also be questionable to use a single wave resistance coefficient for all scales.

It is a well-reported fact that the Schoenherr line matches almost exactly

experimental values for the skin friction of plates (Bertram, 2012). The reason
why a deviation is observed here between the friction lines and CFD results
was partially examined by Magionesi and Di Mascio (2016). They focused their
attention on flows over ship bulbous bows, because, while a flat plate is
characterised by a 0 pressure gradient (Peltier and Hambric, 2007), this is not
the case for large parts of the ship’s wetted area. Bulbous bows in particular
exhibit pressure gradients, the boundary layer is three dimensional, and they
are subject to free surface effects, which cannot be ignored (Ciappi and
Magionesi, 2005). Ishihara et al. (2015) stated that turbulent boundary layers
not only influence the large-scale wake behind a structure, but also the location
of separation. In the case of a ship, the boundary layer at the bulbous bow,
being different from that of a flat plate, causes changes to cascade astern, and
likely into the wake itself.

54
 Figure 3.8. Frictional resistance coefficients

One final point to consider here is the Froude number. Having used a single
Fr (0.26) in all simulations allowed the isolation of the effects of Re. Froude
number effects on wave resistance are well documented. What is lacking in
particular is its effect on the form factor. To resolve this, a larger scale study is
required, featuring both numerical and experimental geosim analyses from a
single institution to eliminate variability across testing facilities. This is
suggested as a piece of future work.

3.5.2.3 Form factor

It is now appropriate to introduce the calculated form factors, shown in Figure

3.9, against Reynolds number. Surprisingly, the (1+k) values do not exhibit a
well-defined trend. Instead, some scatter is evident as the Reynolds number
increases. A similar trend was discovered by Min and Kang, (2010), whose
empirical correction depends solely on the Re. García-Gómez (2000) on the
other hand, devised a formulae based on the scale factor. Both of these are
shown in Figure 3.9 with the dashed, and dotted lines, respectively. What is
most evident here is their change with scale factor or Reynolds number.

55
Indeed, the error between the empirically corrected, translated double body
form factor and multiphase (1+k) value found is approximately 8.6%. In the
level double body case, this is even higher, reaching 10.9%. It should be noted
that the empirical form factors are placed for a λ=31.599. The abovementioned
corrections are applied to them in order to demonstrate how they perform
compared to the full-scale CFD results.

Figure 3.9. Calculated form factors

56
 3.5.2.4 Extrapolation to full-scale

The presence of the scatter in values described previously implies there is a

high degree of variability in the possible extrapolated value of the total
resistance coefficient in full-scale. For example, two testing facilities using
different scale factor to determine the wave resistance would find inconsistent
𝐶𝑇 values. To illustrate this, Figure 3.10 was compiled using all combinations
of wave resistance coefficients (CFD and slender body), form factors (CFD and
empirical, with and without the corrections of Min and Kang (2010) and
García-Gómez (2000)), yielding 126 unique predictions for each friction line.
Clearly, it is not possible to show the corresponding category for each point.
Instead, the average of each friction line, global average, and the multiphase
CFD prediction for 𝐶𝑇 are shown in Figure 3.10. Now, it is imperative to
highlight the importance of reliable predictions for each component of the total
resistance. For instance, although the Telfer (1927) line showed excellent
predictions for the frictional resistance, its use may be questioned depending
on the values chosen for the form factor and wave resistance. Of course, this is
also the case for all other methods used. Figure 3.10 suggests a ‘band’ of
possible values for each friction line exists. Therefore, a large degree of
uncertainty can be expected depending on the adopted methodology to
calculating each component of 𝐶𝑇 .

57
 Figure 3.10. Extrapolated total resistance coefficient based on the results
obtained in this chapter

The assertion that a small error in the total resistance coefficient is decomposed
into even smaller errors over each constituent component worked
advantageously earlier. However, if the process is reversed, the effect is
magnified to yield a high degree of uncertainty. Thus, the importance of high-
fidelity methods cannot be overstated. One possible problem is that virtually
all of the available literature treats these problems (the evaluation of the total’s
constituent components) separately. For example, one may determine a
method to calculate the exact frictional resistance for a specific Reynolds
number. Indeed, the available methods perform very well, as shown in this
chapter. However, if free surface effects are not accounted for in the
extrapolation procedure from a Reynolds number in the region of 106, to one
near 5×109, the results will not resemble reality. This would influence all
parameters present in the decomposition of the total resistance coefficient and
would likely be amplified when carried to full scale. An analogous argument
can be made for the estimation of wave resistance, although there is
uncertainty in its estimation due to the degree of complexity associated with
ship wave making.

A more fundamental problem arises from the ambiguity resulting from the
amplification of errors discussed above: it becomes difficult to justify

58
validating the estimation of each constituent component of the total resistance
when calculated separately. Perhaps the best example in this context is the
prediction of wave resistance by potential flow methods. These are typically
validated by subtracting the viscous resistance (1+k)𝐶𝐹 from the
experimentally obtained total, for instance, Tuck and Lazauskas (2008). Now
𝐶𝐹 is usually obtained via one of the friction lines in Figure 3.8 (typically the
ITTC57 line), while (1+k) – experimentally using the procedure stated in
earlier. As demonstrated, both of these are susceptible to scale effects and
therefore contain a certain amount of error when carried to full-scale.
Consequently, the use of wave resistance estimations using potential flow
methods in extrapolation procedures should be approached with caution, not
least because they fail to model scale effects. Referring to validation versus
experiments, the problem is not confined to the use of methods based on
Michell's (1898) integral. Instead, wave cut methods (Janson and Spinney,
2004) and panel methods (Newman, 1992), both of which are deemed reliable
and robust, must suffer from the above issue. To circumvent this obstacle,
wave probes can be used during tank testing (Kim et al., 2001; Townsin, 1971,
1968; Troesch and Beck, 1974).

Alternatively, a recently emerging method is to capture high quality optical

images of the free surface around the ship (Gomit et al., 2014). Applying this
method, Caplier et al. (2016) estimated the energy contained in a ship’s wake
by using its spectrogram, defined as a heat map used to visualise the time-
dependent height of the water surface in terms of a frequency spectrum
(Pethiyagoda et al., 2018, 2017). Here, only half of the obstacle has been
addressed. The remainder (scale effects) seems insurmountable with present
methods, especially with the apparent scatter of 𝐶𝑊 observed in Figure 3.5.

The use of CFD methods undoubtedly increases understanding of the

underlying phenomena and their relative importance. However, there is still
a fundamental source of uncertainty present in their use: the statistical
modelling of turbulence, usually referred to as turbulence modelling. In this
chapter it was shown that it is the most probable source of errors in sinkage
and trim, whose effect cannot be neglected. One could argue along the lines of
Moore (1965) with regards to the availability of computational power, and
claim that Direct Numerical Simulation (DNS) will solve the problems in this
respect if one allows sufficient time to increase the available computational
power. However, even today, the use of DNS is limited to a narrow range of
Reynolds numbers, which is far from sufficient even for model-scale

59
computations (Beck and Reed, 2001). Simultaneously, the field of wall function
derivation is very active (Kiš and Herwig, 2012). Even if one chooses to accept
the use of turbulence modelling, Pereira et al. (2017) found that different
models are best suited for hull resistance prediction and propeller dynamics.

3.6 Summary and conclusion

In line with the research aims and objectives, this chapter focused on scale
effects on ship bare hull resistance in unrestricted waters. Several hypotheses
were tested using a commercially available RANS solver. Emphasis was
placed on challenging the widely used assumption of geometric similarity in
ship wave patterns with scale, and thus a scale invariant wave resistance,
which was shown to be untrue, confirming the results of Raven et al. (2008).
Scale effects on all examined parameters were shown to be magnified with
large changes in the Reynolds number. This was accomplished by maintaining
the Froude number constant throughout all adopted case-studies (Fr = 0.26).

It was demonstrated that the relative importance of free surface effects of

frictional resistance decay rapidly with an increase in Reynolds number, but
can substantially influence extrapolated results if not properly accounted for.
Several sources for this have been identified, including boundary layer
thickness and flow separation. These are also the likely cause of non-
geometrically similar wave patterns observed.

An alternative path worth exploring is to perform all computations directly in

full-scale. While the resources required to perform a full-scale computation
seem to be limited, experience suggests that time improves computational
availability. It could then be argued that it is only a matter of time before it
becomes commonplace to routinely perform full-scale CFD simulations. One
of the main issues academia must seek to resolve is that of turbulence
modelling, which is a very active field of research where strides of progress
are being made.

In performing numerical simulations, the intention was to maintain the

highest possible degree of similarity with the real-life physics of the problem.
However, incorporating propeller effects was omitted. The complex, non-
uniform flow, generated by the presence of a rotating propeller is responsible
for stark changes in the overall pressure and velocity fields near the ship’s
stern. While by ignoring such effects, representing the overall problem may
have been deviated from, it is thought that isolating the bare hull resistance is
a worthwhile endeavour. Indeed, to the best of the author’s knowledge, the

60
work presented herein is the first to examine scale effects on the wave
resistance and free surface effects on the frictional resistance using CFD.

61
 APPLICATION OF EDDY-
VISCOSITY TURBULENCE MODELS
TO PROBLEMS IN SHIP
HYDRODYNAMICS

This chapter examines different approaches to close the Reynolds averaged

form of the Navier-Stokes equations. This is identified as the most widely used
method to model flows in marine hydrodynamics. A comparative assessment
is performed on two families of two-equation eddy viscosity turbulence
models and a one-equation closure. The time per iteration characteristics of
each model is also recorded to determine which model is most economical. A
statistical approach is proposed to rank the turbulence closures.

4.1 Introduction

Since the beginning of the 19th century, the fields of science and engineering
have been underpinned by the concept of determinism. Its principle was first
articulated by Pierre-Simon Laplace (van Strien, 2014) to express the idea that
every event is causally determined by previous occurrences. Therefore, if one
possesses sufficient information about these causes, it is possible to predict the
state of a system at any point in the future. This idea likely feels natural to
engineers in particular.

Mathematically, determinism is merely an expression that a unique solution

of the Navier-Stokes equations exists. In fact, this has been shown to be the
case in two dimensions, but for three-dimensional cases it holds for finite times

62
only (Lesieur, 2008). In other words, it is not always possible to obtain
sufficient knowledge of antecedent events in order to determine the exact state
of the system at some future time. In practice, the engineer is rarely, if ever,
interested in the level of detail referred to above. Hence, the use of statistical
modelling of physical processes is justified as long as an accurate mean, or
integral, value can be obtained.

The question then shifts from the ability to capture the physical interaction of
every molecule of the fluid, to its overall properties, such as velocity and
pressure. In other words, the fluid is treated as a continuum. The difficulty
here is associated with the complexity of turbulent phenomena. Although
these can be observed in everyday life, modelling exact statistical averages of
fluctuations inherent in turbulence has proven impossible (Durbin and
Pettersson Reif, 2011). Thus, every time the numerical solution to a problem
involving turbulent properties is invoked, it contains a degree of empiricism.
This empiricism is required due to the averaging process, whether temporally
or spatially, applied to the Navier-Stokes equations. Upon performing what is
known as Reynolds averaging, additional terms are introduced as a
consequence of nonlinearities in the governing equations. Perhaps
surprisingly, the degree of empiricism (the number of empirically defined
coefficients, or lack thereof) is not always an indication of accuracy.

The addition of new unknowns to the governing equations means that the
Reynolds Averaged Navier-Stokes (RANS) equations no longer form a closed
set. In simple terms, there are now more unknowns than there are equations,
which is why turbulence modelling is also referred to as providing closure.
The additional (partial differential) equations introduced define the type of
turbulence model. Hence, zero, one, or two-equation closures will be referred
to. Now, the topic of investigation can safely be shifted to identifying which
turbulence model is optimal for the particular case it is applied to. The optimal
solution must exhibit several characteristics:

1. The solution must lie within some predefined measure of accuracy.

2. The effort required to obtain the solution must not be excessive.
3. There should be no simpler method to achieve the same, or reasonably
similar result.
4. The method used should represent the underlying physical phenomena
correctly.

63
In most cases, one can gain an idea of the expected level of accuracy depending
on the method used. For example, it is well-known that linear potential flow
theories do not perform well in extreme cases, where nonlinear phenomena
dominate. However, the above (or any other) classification of turbulence
models, cannot be used to establish such a hierarchy. Indeed, one-equation
models can, and do perform better than two equation models in certain cases.
This is true because the derivation of most of the currently available models
are not rooted in the physics of the problem per se (Argyropoulos and
Markatos, 2015). Simply put, some of the empirical constants contained within
the turbulence models are determined by fitting data to simple experimental
observations. There is no telling what their effect would be if they were
applied to a different problem, as is often the case. Moreover, the mathematical
formulation of the equations used to provide closure are rarely related to the
physics of the problem either (Markatos, 1986).

This chapter is addressed to practitioners as well as researchers of RANS-

based ship hydrodynamic analysis, with the goal of reducing some of the
ambiguity inherent in the field. The aim is to analyse the issue of optimal
turbulence model selection as applied to problems pertaining to the realm of
ship hydrodynamics. Complex problems in the field are sought to test the
turbulence models robustly. To elaborate, it would be counterproductive to
select a simple, two-dimensional case since the aim is not to improve on the
selected models. The parameters of interest are ship bare hull resistance,
sinkage, and trim, due to the difficulties associated in with their prediction.
Furthermore, Larsson et al., (2014) reported a scatter in sinkage and trim
results when examining submissions to the Gothenburg numerical ship
hydrodynamics workshop, justifying the above selections. Deliberately
accentuating the effect each of these has on the remaining parameters as well
as their overall importance implies the selection of a shallow water case study.

In shallow water, the sinkage, trim, and resistance are influenced by the
adverse pressure gradient acting longitudinally on the hull, generated as a
result of the hull’s proximity with the seabed. Additionally, in terms of
turbulence, pressure gradients present an extra layer of complexity,
superimposed of the already difficult problems of predicting highly three-
dimensional boundary layers inherent in ship underwater shapes. The hull
form selected is the benchmark KCS container ship, while the analysis is
performed using Star-CCM+, version 13.01.011.

64
The remainder of this chapter will continue with an overview and
identification of selected turbulence models. Then, the case studies and
numerical implementation are presented in Section 4.3 and 4.4, respectively.
Results and their analysis follows form these in Section 4.5, while a summary
and conclusions are given in Section 4.6.

4.2 Background

This chapter is motivated partially by the findings of Larsson et al., (2014), who
observed a scatter in sinkage and trim predictions, as mentioned previously,
and partially by recent work undertaken within Chapter 3. Sinkage and trim
proved more difficult to simulate accurately than resistance in CFD (Terziev
et al., 2019a). In deep waters, the combined effect of sinkage and trim is known
to be small on resistance and in some cases may even be neglected (Ponkratov,
2016). However, their magnitude and influence are accentuated by the
proximity of the seabed in shallow water cases. This not only represents a case
where the underlying physics are not well understood, but also, an everyday
situation all ships must cope with. More importantly, groundings, where the
combined effect of sinkage and trim (termed ship squat) play a central role,
are consistently in the top five incident categories of EMSA (European
Maritime Safety Agency, 2018, 2017, 2016, 2015) since records began in 2011.
In fact, the annual figures published by EMSA suggest more than half of all
incidents occur in shallow water, where ship squat is always a contributing
factor. Thus, the accurate prediction of ship squat is of critical importance.

While ship squat is a phenomenon requiring further research, turbulence

modelling is a ‘known unknown’, which contributes to the present levels of
ambiguity inherent in shallow water predictions. Greater levels of attention
are typically attributed to the prediction of resistance. However, there is less
ambiguity in its predictions. This has led to rule of thumb approaches even in
CFD by international bodies, such as the International Towing Tank
Conference (ITTC, 2014). On the other hand, sinkage and trim have proved
more evasive. The results from the present research will enable generalisations
to be carried into deep water cases for enhanced predictions, even when
running trim and sinkage are of little importance. The remainder of this
section will proceed by examining the turbulence models.

65
 4.2.1 Turbulence

Attempts to bring turbulence under mathematical scrutiny begin with

Boussinesq’s concept of eddy-viscosity (Wilcox, 2006), which he developed
from ideas expressed by Saint-Venant (Darrigol, 2017). This makes it possible
to express the Reynolds stress tensor, which is the additional term introduced
in the process of Reynolds averaging, as a function of mean flow properties
(such as the eddy-viscosity). Here, Reynolds averaging refers to the process of
averaging the decomposed total velocity (mean and fluctuating parts) (Durbin
and Pettersson Reif, 2011). The eddy-viscosity is a property of the flow, i.e. a
function of space and time, unlike the molecular viscosity, which is an intrinsic
physical constant (Bailly and Comte-Bellot, 2004).

The first turbulence model was introduced by Prandtl (1925) using the mixing
length concept to compute the eddy-viscosity. This model did not contain any
Partial Differential Equations (PDEs), and is therefore known as zero-equation
or algebraic. In practice, an n-equation model refers to the number of
additional PDEs introduced to close the Reynolds Averaged Navier-Stokes
(RANS) equations (Wilcox, 2006). Twenty years passed before the next
conceptual leap was made by Prandtl, when he modelled a PDE to express the
turbulent kinetic energy, k, thereby creating the first one-equation turbulence
model. Fundamentally, this allowed the local flow properties to be dependent
on antecedent events. Then, van Driest (1956) devised a viscous damping
modification to the mixing length model, which has been applied to virtually
all algebraic closures since (Wilcox, 2006). A noteworthy example of a zero-
equation model includes that devised by Baldwin and Lomax (1978).

Since the first one-equation model was born, an explosion of competing

turbulence closures has occurred. Here, prevalence is given to those, used in
the assessment. Thus, attention will be confined to the main eddy-viscosity
models and their significant variants as implemented in the commercial RANS
solver, Star-CCM+, version 13.01.011. The fundamental purpose of this class of
closures is to predict an eddy-viscosity. Since this was first introduced by
Boussinesq, it has since become the underlying hypothesis used to derive the
vast majority of turbulence models.

Although the Reynolds Stress turbulence (RST) model boasts as the most
physically sound closure strategy (Sarkar and Lakshmanan, 1991), there are
problems related to computational stability in its implementation (Parneix et
al., 1998). Furthermore, for the present class of problems (ship

66
hydrodynamics), eddy-viscosity models dominate the literature, despite the
fact that Rotta (1951) derived the first RST model decades prior to the first
practical application of eddy-viscosity-type models. The prevalence of eddy-
viscosity models in the field is illustrated at the end of this section.

Broadly speaking, the problem of closure is solved via two main classes of
turbulence models. These are the k-ε, k-ω models and their variants. Here, k is
the turbulent kinetic energy, ε is the dissipation rate, and ω is the dissipation
frequency. Before each class is examined sequentially, a qualitative description
of a one equation turbulence model is given. It should be noted that the
underlying mathematical relations, expressing all turbulence models utilised
in this thesis can be found in Appendix C.

4.2.1.1 One equation Spalart-Allmaras turbulence closure

A popular alternative to two-equation turbulence models is presented in one-

equation closures. The most successful of these is the formulation proposed by
Spalart and Allmaras (1992). Breaking precedent, this closure uses a diffusivity
equation in place of the turbulent kinetic energy to establish an eddy-viscosity.
All prior work on models of this type had used the turbulent kinetic energy
instead (Baldwin and Timothy, 1990; Johnson and King, 1984). The original
formulation presented in Spalart and Allmaras (1992) was specifically
designed for unstructured codes. Due to its ability to model separated flows
and its popularity in the field of aerodynamics (Siemens, 2018), this model is
incorporated.

4.2.1.2 The k-ε model and its variants

The k-ε model’s foundations were laid by Jones and Launder (1972). It is
interesting to note that it the k-ε model does not owe its popularity to
robustness or accuracy, but to the fact that it was the first two equation model
applied in practice (Durbin and Pettersson Reif, 2011). That is, even though
Kolmogorov's (1942) k-ω model had been described decades earlier. In Star-
CCM+, the original coefficients proposed have been replaced with those
suggested by Launder and Sharma (1974) in the case of the standard
formulation. However, the standard k-ε model cannot be applied on low
Reynolds number (LRN) type cases due to the occurrence of a singularity near
solid boundaries. Here, low Reynolds number refers to turbulent 𝑅𝑒𝑇 , which
is a function of wall distance (𝑅𝑒𝑇 = √𝑘𝑑/𝑣, where k is the turbulent kinetic
energy, d is the distance from the wall to the nearest cell, and 𝑣 is the viscosity,

67
as opposed to 𝑅𝑒 = 𝑈𝐿/𝑣, where 𝑈 is the speed, and L represents the length of
the ship). Thus, when y+ ~ 1, models capable of coping with LRN must be used.
On the other hand, if y+ > 30, high Reynolds number (HRN) models are
required.

Jones and Launder (1972) originally formulated the k-ε model for HRN-type
cases. However, in such a scenario, the use of wall functions is required, which
are not compatible with complex flows (Pettersson Reif et al., 2009). Therefore,
researchers use either a LRN or realizable versions, discussed below.
According to Durbin (1996), two equation turbulence models have a tendency
to overpredict turbulent kinetic energy in stagnation points. Realizability
refers to the imposition of a suitable constraint to the model to alleviate this
shortcoming. Additionally, Star-CCM+ offers the two-layer method, where a
PDE is solved for k near the wall, but ε is algebraically prescribed based on the
wall distance (Chen and Patel, 1988). The wall proximity indicator is then used
to patch the two solutions onto each other, as suggested by Jongen (1998). On
the other hand, the length scale and turbulent viscosity ratio, used to calculate
ε, are adopted from Wolfshtein (1969) in the two layer approach.

The realizable, two layer k-ε model, hereafter referred to as “k-ε 2l” for brevity,
has been extensively used for practical applications (Cakici et al., 2017; Liu et
al., 2018; Ozdemir et al., 2016; Song et al., 2019; Terziev et al., 2018; Tezdogan
et al., 2016a, 2016b, to name but a few), making it an ideal candidate for this
chapter’s purposes. However, there are a large number of variants of the k-ε
model put forth by researchers, each claiming superiority. The motivation
behind the plethora of alternatives lies with the shortcomings on the model.
While its inability to model near-wall effects in the original form can be
circumvented easily, the implicitly assumed linear stress-strain relationship
poses further problems: linearity may be violated in strongly 3D boundary
layers (Durbin and Pettersson Reif, 2011). This is certainly the case for ship
boundary layers, as asserted by Magionesi and Di Mascio, (2016). Despite
these shortcomings, the k-ε 2l has enjoyed considerable success and
widespread implementation for industrial applications. It is therefore
interesting to compare the performance of some k-ε variants available in Star-
CCM+. Alternatives include proposals by Shih et al. (1995), which cannot be
incorporated since it is not offered by the RANS solver.

Abe et al. (1994) proposed a modification of the standard k-ε model, henceforth
referred to as ‘AKN’ (after the authors of the paper – Abe, Kondoh, Nagano),
claiming their variant can predict flow reattachment better than the original.

68
As such, their model is of the LRN-type. The authors also altered the
coefficients of the original model in their derivation. According to Siemens
(2018), the AKN model is a suitable choice for flows over complex geometries,
which is undoubtedly representative of a ship underwater shape. For these
reasons, this model is selected for the present examination. Additionally, flow
reattachment following separation is a crucially important parameter in the
prediction of ship resistance, trim and sinkage.

A widely discussed closure in turbulence research is the k-ε-v2-f model,

hereafter referred to as ‘v2-f ’ (Durbin, 2017a), where v2 is the wall-normal
stress component and f is the elliptic relaxation parameter. This model solves
two additional equations in order to predict the eddy-viscosity. Its origins can
be traced to work done by Durbin (1996, 1995, 1993, 1991), whereas some
recent extensions and modifications include Pettersson Reif et al. (2009),
Pettersson Reif, (2006), Laurence et al. (2004), Davidson et al. (2003), and Lien
et al. (1998). While this model cannot be classified as a two-equation model
for obvious reasons, it retains its larger classification (eddy-viscosity-type
model). The benefits of utilising the v2-f model include its ability to handle wall
effects, simultaneously accounting for flow nonlocality. Thus, it can be applied
directly to LRN-type meshes, without any modifications, or the need to patch
regions of different solutions as was the case earlier. Siemens (2018) stated that
selecting this model “is known to capture the near-wall turbulence effects
more accurately, which is crucial for the accurate prediction of heat transfer,
skin friction and flow separation.” Thus, it is included in the investigation
within the k-ε-class models. The fundamental advantage in implementing the
v2-f model is related to its ability to account for anisotropy of turbulence
stresses, which are not addressed in the variants examined so far (Gorji et al.,
2014).

The final two models incorporated within the present category are elliptic
blending closures, henceforth referred to as ‘EB’. Both of the variants
integrated in this chapter solve for the turbulent kinetic energy, the dissipation
rate, the normalised reduced wall-normal component, and the elliptic
blending factor (α, rather than f). This is thought to be beneficial due to its
inclusion of nonlocal, and pressure echo effects. The latter, observed near
impermeable boundaries, are solved for via the method of images. The wall-
normal component is defined as the ratio of v2 and k, and therefore requires
the solution of v2 in a similar manner to the v2-f model. The concept of elliptic
relaxation was first introduced for the RST-class models by Durbin (1993b). As

69
such, it mandated the solution of six equations to provide closure, which was
later simplified to a single equation (α). Later, Manceau and Hanjalić (2002)
further simplified the EB model rendering it more industry friendly, which is
the form utilised by Star-CCM+. The elliptic blending model’s conception can
be traced to being a consequence of the development of the v2-f model.

Up to this point, linear eddy-viscosity models have been the focus. Although,
the v2-f has been extended to incorporate nonlinearities, Pettersson Reif (2006)
found its linear variant to perform equally well, which is the version employed
here. On the other hand, non-linear eddy-viscosity models, specifically of the
elliptic blending-type have recently emerged and gained popularity. The main
contributors to their development other than Manceau and Hanjalić (2002)
include Manceau (2015) and Manceau et al. (2001), however, the conceptual
origins of 2nd order closures can be traced to Chou (1945). In the present work,
the standard and lag EB (referred to as ‘Lag-EB’) are adopted. The former,
according to Billard and Laurence (2012), circumvents the k-ε 2l models near-
wall issues, while it exhibits stable performance for practical applications.

The alterative, the Lag-EB model of Revell et al. (2005), was devised to address
a fundamental shortcoming of almost all eddy-viscosity models. Namely, their
imposition of a linear relationship between the stress tensor and rate-of-strain
tensor, which follows directly from the Boussinesq hypothesis. Revell et al.
(2005) implemented a phase lag, or angle between these two parameters to
express their misalignment. In Star-CCM+, the implementation of Lag-EB
follows the modifications suggested by Billard and Laurence (2012) for
stability reasons (Siemens, 2018).

This marks the end of the description of the five k-ε variants used. The next
class of turbulence models replace the dissipation rate, ε, with ω – the
dissipation frequency.

4.2.1.3 The k-ω model and its variants

The second most widely used strategy to close the governing equations
involves the adoption of a k-ω type turbulence model. Here, only a brief
overview of the variants used is given. Wilcox (2006) can be consulted for a
comprehensive overview of the subject. It is worth noting that without having
prior knowledge of the Kolmogorov (1942) k-ω model, Saffman (1970)
formulated a k-ω model to close the RANS equations. The latter was applicable
down to the wall, immediately showing distinct advantages over the k-ε
model.

70
The modern version of the k-ω variant, here referred to as the standard ‘k-ω
Wilcox’ model, can be traced to Wilcox (1988), although earlier work provided
its basis (Saffman and Wilcox, 1974), it was later revised several times by the
same author. Most notably, by Wilcox (2008, 2006), the most recent of which is
incorporated in Star-CCM+. Siemens (2018) advise that the model coefficients
of Wilcox (2008) have not been extensively verified for flows, other than two-
dimensional parabolic cases. Therefore, these corrections are included as an
optional add-on. Since the objective of the present work is to test well-
established closures, these have not been made use of.

The k-ω Wilcox model has faced primarily the criticism of free-stream
sensitivity (Cazalbou et al., 1994). This refers to the fact that turbulent statistics
must vanish at some distance from an impermeable wall. In the free-stream,
their influence on the irrotational region (the region not affected by viscosity)
is thought to decay as (y+)-4 (Phillips, 1955) which is not always the case when
employing the k-ω closure. This problem was addressed by Wilcox (2008) and
is thought to have been resolved in the latest version of his model. Menter
(1994) transformed the transport of ε from the k-ε model to resolve the free-
stream sensitivity by a variable substitution in the k-ω Wilcox model. The
formulation Menter (1994) arrived at is similar to the standard k-ω Wilcox
model, but contains an additional term expressing the cross-diffusion of k and
ω.

In other words, the form derived by Menter (1994), referred to as the ‘k-ω SST
model’ (where SST stands for Shear Stress Transport), could potentially
provide identical results to the k-ε model. By incorporating a blending
parameter, expressed as a function of wall distance, the k-ω SST model is
thought to preserve the advantages inherited by both main closures used
today (k-ε and k-ω). This is also one of the models gaining popularity in the
field of ship hydrodynamics (Deng et al., 2015; Farkas et al., 2018; Haase et al.,
2016b; van Wijngaarden, 2005; Wang et al., 2018; Wnęk et al., 2018). As such,
it is important that it is incorporated it within the set of case studies.

The final aspect of turbulence research considered in this sub-section is

associated with the concept of transition. This term is used to describe the
change of a boundary layer from laminar to turbulent (Saric et al., 2002). This
change is of practical importance because a turbulent boundary layer imparts
a greater skin friction to the surface from which it is shed, than when in
laminar state. Thus, the onset of turbulence, and its possible delay, is of great
importance, especially considering that skin friction is one of the main

71
contributing components of ship resistance. According to Durbin (2017b), the
primary modes in which transition occurs are:

1. Natural
2. Bypass (as a response to external disturbances)
3. Separation-induced

Transition models are required because most turbulence closures are by

definition applicable to the fully turbulent part of a flow (HRN-type models).
A related example is the difference and range of applicability of the Blasius
(1908) and any other friction lines, for instance, the well-known Schoenherr
(1932) line. The AKN model and the standard k-ε model, for example, contain
empirical damping functions, which allow them to exhibit transition
properties. However, this is a consequence of solution bifurcation, resulting
from the use of damping functions rather than physical insight into the
phenomenology of the process itself (Lesieur, 2008). Their use has also been
widely criticised (Patel et al., 1984), in part because the only similarity between
different empirical damping functions is the presence of exponentials in their
derivation (Durbin, 1991; Shih, 1990). Some notable examples include Chien
(2008) and Launder and Sharma (1974), to name but a few.

The RANS solver employed, Star-CCM+, offers a variety of methods of

modelling transition. The first, and arguably more difficult to implement (the
turbulence suppression model) requires prior knowledge of the location of
transition. Thus, it is not applicable in many cases. An alternative solution was
provided by Abu-Ghannam and Shaw (1980), which is based on the concept
of intermittency. This parameter can range from 0 to 1, describing the time the
flow is in a turbulent state (in percentage values). The reason this is not
implemented is related to its requirement of nonlocal quantities, which cannot
be evaluated at cell centres. A local alternative was suggested by Langtry
(2006). Since the formulation was originally incomplete, its adaptation in Star-
CCM+ requires user programming and definition of additional functions.
Recognising that it is highly unlikely for the average user to accomplish this,
the simpler ‘γ’ transition model has been opted for. Selecting the γ transition
model activates the solution of an additional intermittency equation (γ) and
couples its solution with the k-ω SST model via the kinetic energy – k (Menter
et al., 2015). Non-local effects are handled via an approximation of the three-
dimensionality of boundary layer by using the concept of helicity, which is a
measure of the cross-flow within a turbulent boundary layer (Müller and
Herbst, 2014).

72
 4.2.1.4 Potential alternatives

As has been illustrated thus far, research and development of turbulence

models has been rather chaotic, featuring small steps and large conceptual
leaps. For instance, while elliptic blending-type models are thought to be
superior, the development of one, and two-equation closures has continued
since the former’s inception. One way to characterise the field would be via
additions of small theoretically and empirically acquired pieces of information
to existing models. Indeed, this sort of ‘tinkering’ is a defining feature of the
field of turbulence research (Roache, 2016).

Initially, model constants were calibrated against experiments. When Direct

Numerical Simulation (DNS) data became available, the validation procedure
shifted. Nowadays, the aim is to match DNS data instead of experiments.
Regardless, due to the complexity and limitations of the DNS approach, the
possible benchmarks are limited to simple two, or three-dimensional cases at
best (a classical example is the backward facing step (Darrigol and Turner,
2006)). Thus, the coefficients selected for each model have been calibrated to
perform best for these specific cases. It is therefore not possible to determine
the precision, accuracy, or even tendency to over or underpredict values of
any one turbulence model when applied elsewhere. The matter is worsened
by the previously illustrated fact that all turbulence models contain a large
degree of empiricism (arrived at by a process of tinkering), compounded by
the simple truth that the mathematical formulations used do not represent
real, physical processes.

The degree of empiricism is a reflection of the complexity of the processes

occurring in turbulent flows. If analytical solutions were possible, turbulence
models would be exact, and empirical coefficients would not be required
(Durbin and Pettersson Reif, 2011). Contrary to expectation, empiricism is not
necessarily a drawback. The problem addressed here is the inability to
establish a hierarchy in the modelling of turbulence in any meaningful way
other than complexity. For this reason, research of the type of Gorji et al. (2014),
and El-Behery and Hamed (2011) have emerged attempting to provide specific
recommendations with regards to the use of different turbulence closures in a
particular context.

Turbulence models described thus far model statistically almost the entirety
of the turbulent kinetic energy spectrum. This expresses a concept similar to
wave spectra (St Denis and Pierson, 1953), where components of different

73
wavelength coexist simultaneously (turbulent eddies replacing waves in the
present context). Richardson (1922) and Kolmogorov (1941) are chiefly
attributed with the conceptualisation of the turbulent energy spectrum and its
cascade into smaller scales. It is this cascading process that gives rise to
nonlinear interactions, and thus inhibits its description. Researchers have
devised a way to statistically model part of the spectrum via the turbulence
models described above, leaving the remainder to be captured in the
simulation. This is known as Large (or Very Large) Eddy Simulation (LES or
VLES, alternatively, Detached Eddy Simulation - DES). Alongside the twofold
increase of the order of magnitude in computational power requirements,
Kornev et al. (2019) reported several difficulties when applying the LES
technique to ship hydrodynamics problems. The main issue was related to the
region used to patch the statistically averaged and resolved solution together.
In light of this, the adoption of LES will not be considered. Studies, detailing
the use of hybrid RANS-LES techniques in marine hydrodynamics include the
work of Bhushan et al. (2013), Carrica et al. (2010), Posa et al. (2019), and can
be consulted for further information.

Before the case study description is given, it is worth examining which

turbulence models tend to be preferred in marine hydrodynamics. Here, the
term ‘marine hydrodynamics’ is used in reference to research encompassing
all aspects of ship and offshore research that includes the simulation of fluid
flow using a RANS solver. Additionally, investigations centred on propeller
performance are incorporated.

The compiled dataset presents a sample of research work done during the past
two decades using RANS solvers in the related field. The sample, collected
from two of the biggest academic publishers (ScienceDirect, and Taylor and
Francis), is split into categories corresponding to the turbulence model used in
each work. No differentiation is made between different subtypes of models
(e.g. realizable vs standard k-ε). Figure 4.1 presents the breakdown of the
turbulence models among the same categories and provides a graphical
representation of their cumulative use over the past roughly two decades. The
results shown in Figure 4.1 demonstrate that the k-ω SST model has gained
increasing popularity in the past decade, and that virtually all work done is
concentrated in two categories. The purpose of this chapter will then serve to
introduce some of the newer turbulence models (v2-f, EB, and Lag-EB) to the
field of marine hydrodynamics.

74
 Figure 4.1. Graphical representation of the use of turbulence models in
marine hydrodynamics2.

4.3 Case-studies

An integral part of testing the performance of any turbulence model or

numerical set-up involves comparing results against experimental data.
Having established shallow water ship hydrodynamics will be focused upon
due to the innate complexity of the field, restricts the choices available. In a
recent study by Elsherbiny et al. (2019), the behaviour and performance of the
well-known benchmark KCS container ship were tested. Although focus was
placed on the new Suez Canal in the abovementioned work, a rectangular
canal case study was investigated as well, and is adopted here. Doing so will
allow the reduction in cell numbers because simulating the thin layer of water
near the intersection of the sloping seabed and sides (as was the case in
Tezdogan et al. (2016b)) is bypassed. Thus, the computational mesh can be
refined in the vicinity of the ship, and gradually coarsened in every direction

2 The full list of references used to construct Figure 4.1 can be found in Appendix B

75
without compromising the solution. Indeed, doing so is a mainstay in the
practice of computational hydrodynamics.

The case-studies examined by Elsherbiny et al. (2019) allow us to vary the

speed of the ship and test the turbulence models while keeping all other
parameters constant. However, one could argue that in shallow water ship
hydrodynamics, a change in the depth can have an effect, comparable to the
change in speed. Limiting the case-studies in such a way would prevent the
results from being generalised to a wider pool of problems in the field. For this
reason, two case studies adopted from Mucha and el Moctar (2014) are also
replicated numerically. In their work, the authors used the same ship as in
Elsherbiny et al. (2019), with a different draught (T=10m, rather than T=10.8m
full-scale equivalent) and scale factor (λ=40, rather than λ=75 in Elsherbiny et
al. (2019)). For the purposes of the present research, two ratios of depth and
draught have been selected for a single ship speed. This translates into a
different Depth Froude number (𝐹ℎ =𝑈/ √𝑔 × ℎ, where 𝑈 is the speed, 𝑔 is the
gravitational acceleration and h is the depth), although the speed has been
maintained constant. In shallow water, waves are nondispersive and can
attain a single propagation speed defined by the denominator of the depth
Froude number.

The KCS principal characteristics are summarised in Table 4.1, while the case
studies are graphically depicted in Figure 4.2. The conditions, against which
the turbulence models are assessed are given in Table 4.2. All nine turbulence
models are tested against each of the four case studies delineated jointly in
Table 4.1 and Table 4.2, yielding a total of 36 simulations.

Table 4.1. Principal characteristics of the KCS

Elsherbiny et al. (2019) Mucha and el Moctar (2014)
Quantity
Full-scale Model-scale (1:75) Full-scale Model-scale (1:40)
L [m] 230 3.067 230 5.75
B [m] 32.2 0.429 32.2 0.805
T [m] 10.8 0.144 10 0.25
CB [-] 0.651 0.651 0.64 0.64
S [m ]2
9530 1.694 8992 5.62

Table 4.2. Case-studies description

76
 h/ h Umodel Uship
ID Reference λ 𝐹ℎ [-] Re [-]
T [m] [m/s] [kn]
0.30 1.8297×10
1 Elsherbin 0.534 9.0
7 3 6
y et al. 2.2 0.32
5 3.4759×10
2 (2019) 0.57 6
1.005 16.9

Mucha 0.32 4.7116×10

3 1.3 0.37 8.1
and el 4 5 6
0.73
Moctar 0 4.7116×10
4 1.6 0.4 0.41 8.1
(2014) 6

Figure 4.2. Graphical depiction of the adopted case-studies (not drawn to

scale): (a) refers to the study of Elsherbiny et al. (2019), (b) and (c) refer to
Mucha and el Moctar (2014).

4.4 Numerical implementation

The RANS solver used, Star-CCM+, employs the Finite Volume Method (FVM)
to model the problem at hand by using the integral form of the governing
equations and by discretising the computational domain into a finite number

77
of adjoining cells. Within the framework of Star-CCM+, pressure and
continuity are linked via a predictor-corrector scheme.

To simulate turbulent properties within the fluid, the closures described in

Section 4.2.1 are applied. These are summarised in Table 4.3, where Roman
numerals are used to assign each case with a number.

Table 4.3. Summary of the tested turbulence models.

ID Class Identifier
I AKN
II v2- f
III k-ε EB
IV Lag-EB
V k-ε 2l
VI k-ω γ
VII k-ω k-ω SST
VIII k-ω Wilcox
IX One-equation Spalart-Allmaras

4.4.1 Physics modelling

The motion of the fluid is modelled via a flat wave by the Volume of fluid
(VOF) method (Hirt and Nichols, 1981). Alternatives include the Level-Set
method. However, Starke et al. (2017) showed that some features of the bow
wave generated by the ship could not be captured using the Level-Set method.
In some deep-water cases, the presence of the free surface may be omitted.
However, in shallow water flows, the energy imparted onto the waves due to
the disturbance caused by the ship has a higher relative importance than in
deep waters. Thus, it is not admissible to neglect the presence of the water
surface. A third approach was proposed by Carrica et al. (2007), where a single
phase level set method is introduced. This eliminated the need to account for
the air-filled part of the domain, reducing cell numbers. Unfortunately, it has
not been incorporated within the solver used. Therefore, it cannot be applied
for the here-examined conditions.

The phase interface is captured by the RANS solver by assigning a scalar value
between 0 and 1 to every cell in the domain. A value of 1 implies the cell is
filled with water, conversely, 0 means the cell is in the air-filled part of the
domain. A value of 0.5 indicates a cell is half full of air and water, thereby
defining the location of the free surface. The adoption of the VOF method

78
requires that both immiscible fluids account for large parts of the domain,
while their contact area should be relatively small (Siemens, 2018). To prevent
reflections from the outlet, a VOF damping zone, equal to 1.25L is
implemented in all cases.

A critical feature of the simulation is the type of convection scheme, which was
selected as second order upwind. This choice is made because reducing the
order (to 1st order) compromises the accuracy, whereas increasing the order
(to 3rd order) can potentially lead to problems related to stability (Siemens,
2018). This is applied to all turbulence models in all case-studies. The
segregated flow solver is used solve the Navier-Stokes equations in an
uncoupled manner with the aid of a SIMPLE (Semi-Implicit Method for
Pressure Linked Equations) algorithm.

To model the motions of the ship, the Dynamic Fluid-Body Interaction (DFBI)
module offered by Star-CCM+ is used. The ship is allowed to sink and trim
only, which occur as a result of pressure (normal) and shear (tangential) forces
acting on the hull. To avoid a large initial shock, resulting when the simulation
is initiated, the ship is constrained during the first 5 seconds in all simulations.
Implementing this constraint on the motions of the ship reduces the transient
oscillatory motion typically observed in the early stages of the simulation.
Since the type of problem is pseudo-steady, it is desirable to reduce the time
taken to achieve a steady state. The time taken for the solver to complete a
single iteration is also reduced during these first 5 seconds, since an update to
the position of the ship is not required at the end of the inner loop, which is
set to comprise of 10 iterations. It is important to note that the free surface is
updated at every iteration, while sinkage, trim, and resistance are recorded
and updated once the inner iterations loop has competed.

4.4.2 Time step selection

At the end of the inner loop of iterations, the time step must be updated. This
is selected in all cases as Δt = 0.0035L/𝑈, following Terziev et al., (2018), and
Tezdogan et al. (2016b), where the efficacy of this method of setting Δt has
been demonstrated. This value is calculated for each case study, but is not
changed across turbulence models. An advantage of using this method of
setting the time step is that a change in the mesh is not necessary when the
speed is altered. Thus, for the four case-studies, only 3 meshes are required.
Cases 1 and 2 can be used with the same computational mesh (whose
properties are examined in the following sub-section), while the time step is

79
altered according to the formula above. Case study 4 requires a slight increase
in cell numbers due to the different depth, when compared to case 3 (see
section 4.4.4, Table 4.4). All other characteristics, including the time step for
cases 3 and 4 is the same because the speed is not changed. Alternative
approaches to setting the time step include the ITTC's (2011) Δt=0.005 ~
0.01L/𝑈. Throughout all examined cases, the temporal discretisation is set as
first order, and the solution is allowed to develop for a minimum of 150
seconds of physical time to ensure convergence of the results.

4.4.3 Computational domain

The computational domain has been structured according to the

recommendations of ITTC (2011) and Siemens (2018). Namely, the inlet
boundary in resistance computations should be located between one and two
ship lengths from the ship bow, while the outlet – between two and three ship
lengths downstream. The velocity inlet is placed 1.5 ship lengths upstream
from the forward perpendicular, while the pressure outlet – 2.5 ship lengths
downstream from the aft perpendicular. The domain width and depth are
dictated by the case studies, which are summarised in Table 4.2. The side
boundary is set as a wall, while the domain bottom is a velocity inlet. The latter
choice is made because the cells in the domain are linked to the ship’s centre
of gravity (CoG), which is set as the origin of the local coordinate system. As
the CoG translates and rotates, some cells cross the domain bottom. To ensure
that the same amount of water flows past the ship, the domain bottom is
allowed to introduce the lost amount of water back into the domain. This is
done adding the same VOF flat wave speed to the domain bottom in the
negative x – direction. In other words, the bottom acts as a reflective boundary,
without developing a boundary layer as a result of the relative motion
between it and the fluid, while simultaneously ensuring the h/T condition is
maintained.

To alleviate the computational cost, a symmetry boundary is imposed,

coincident with the ship and canal centreline. Finally, the domain top is
located 1.25 ship lengths from the undisturbed water level, where a velocity
inlet condition is imposed. The domain is graphically shown in Figure 4.3.

80
 Figure 4.3. Boundary conditions and dimensions for all computational
domains.

4.4.4 Mesh generation

The mesh was generated within the facilities of Star-CCM+, which allow the
user to make full use of the software’s automatic operations. The static, region-
based mesh comprises mainly of hexahedral cells with minimal skewness,
generated via the trimmed cell mesher. According to the findings of Jones and
Clarke (2010), tetrahedral cells can compromise the accuracy of the solution.
Concentric local volumetric refinements are imposed in the vicinity of the hull,
ensuring the accurate representation of flow phenomena. Furthermore, the
location where the free surface is expected to deform has been systematically
refined.

a b

Figure 4.4. Sample y+ value distribution along the wetted hull: (a) refers to
the study of Elsherbiny et al. (2019), (b) refers to Mucha and el Moctar (2014)
– case 2.

The prism layer mesher is used to capture the ship boundary layer by creating
orthogonal prismatic cells. Resolving near-wall flow accurately is of critical

81
importance, which is why the y+ values are monitored, ensuring the average
values remain below 1 in all cases (Eca et al., 2018). The distribution of y+ values
along the wetted part of the hull are shown in Figure 4.4 for cases 1 and 3.

While the vast majority of turbulence models predict the same y+ values, the
AKN model (bright green) deviates significantly. In the next section, it will be
shown that such behaviour leads to a compromised accuracy. For case 1 (of
Elsherbiny et al. (2019)) the large scale factor (λ=75) allows an increased
resolution to be imposed without the need to resort to an excessively large
number of cells. The second case is not as straightforward, which is why the y+
values are allowed to reach 0.9, although the average in all cases is smaller
than 1, despite the two-fold increase in cell numbers (Table 4.4). The resulting
computational mesh, mirrored using the central symmetry for illustration
purposes, is shown in Figure 4.5, where the local volumetric refinements are
clearly visible, including the Kelvin wedge. The properties of the
computational domain for all case-studies are shown in Table 4.4.

Strictly speaking, the imposition of a symmetry plane has the potential of

invalidating the results presented herein. However, Mucha and el Moctar
(2014) demonstrated a negligible difference resulting from the presence of a
symmetry plane. For this reason, it is deemed safe to proceed directly with the
computational mesh, as described previously. Finally, to enable a reliable
comparison of the results, free of unregistered unknowns, all parameters
described in this section have been kept identical across the case-studies.

Figure 4.5. 3-D view of the computational mesh. Depicted: case 1 (mirrored
using the central symmetry plane).

82
Table 4.4. Computational domain cell properties.
Case study Number of cells Number of faces Number of
number vertices
1 and 2 1,446,076 4,281,940 1,491,537
3 2,995,685 8,868,880 3,087,140
4 3,032,015 8,984,576 3,125,549

4.5 Results and discussion

In this section, the results are presented and compared, both against each other
and against the relevant experiments. The first sub-section is dedicated to the
numerical uncertainties and error estimations against experiments. Presenting
these first will later allow a fuller discussion of the results.

4.5.1 Comparison against experimental data

The first step taken in this section is to present the results in tabular form
alongside the experimentally obtained values (EFD – Experimental Fluid
Dynamics) and calculated error (E). This step is used to validate the solution.
The error is defined as E=(EFD-CFD)/EFD×100, and is shown in Table 4.5. The
values, highlighted in red were not used in the computation of averages
shown in the same table. This is done in an attempt to allow the global average
(shown in bold) of the error to represent the values in a meaningful way. If the
error values for trim were taken into account, the global average error would
be much higher.

The first conclusion drawn from the errors shown in Table 4.5 suggest that
resistance can be predicted by almost all turbulence models effectively. This is
true especially considering the experimental force measurement uncertainty,
which is up to 2.20% in Elsherbiny et al. (2019). As referred to previously,
sinkage and trim present an additional layer of complexity. The numerical set-
up has not predicted values within acceptable margins for one row (case 3), in
the case of sinkage and three (cases 1, 3, and 4) in the case of trim (discussed
further in the following sections). This is the case for two reasons, which
include numerical and experimental uncertainty.

83
Table 4.5. Numerical results and error calculations
Description
ID Type AKN v2- f EB Lag-EB k-ε 2l k-ω γ k-ω SST k-ω Wilcox SP-All Average
λ Fh h/T
EFD 5.775×10-3 -
0.303 1 CFD 4.401×10-3 5.559×10-3 5.559×10-3 5.614×10-3 5.439×10-3 5.298×10 -3 5.340×10 -3 5.505×10 -3 5.596×10 -3 5.489×10-3
Resistance coefficient [-]

E 23.79% 3.74% 3.75% 2.79% 5.83% 8.26% 7.53% 4.67% 3.10% 4.96%
75 2.2
EFD 5.224×10-3 -
0.570 2 CFD 4.087×10-3 5.212×10-3 5.199×10-3 5.247×10-3 5.096×10-3 5.034×10-3 5.014×10-3 5.138×10-3 5.239×10-3 5.147×10-3
E 21.77% 0.24% 0.49% -0.44% 2.46% 3.64% 4.02% 1.65% -0.27% 1.47%
EFD 2.784×10-3 -
0.370 1.3 3 CFD 1.120×10-3 2.802×10-3 2.668×10-3 2.869×10-3 2.851×10-3 2.587×10-3 2.559×10-3 2.661×10-3 2.842×10-3 2.730×10-3
E 59.76% -0.64% 4.15% -3.06% -2.42% 7.07% 8.09% 4.42% -2.07% 1.94%
40
EFD 4.74×10-3 -
0.410 1.6 4 CFD 1.80×10-3 5.02×10-3 4.67×10-3 4.99×10-3 4.97×10-3 4.49×10-3 4.45×10-3 4.66×10-3 4.98×10-3 4.781×10-3
E 62.02% -5.86% 1.50% -5.27% -4.79% 5.28% 6.24% 1.89% -5.01% -0.75%
EFD -2.09×10-3 -
0.303 1 CFD -2.09×10-3 -2.00×10-3 -2.00×10-3 -2.01×10-3 -2.01×10-3 -2.02×10-3 -2.03×10-3 -2.02×10-3 -2.01×10-3 -2.015×10-3
E -0.07% 4.10% 4.18% 3.58% 3.54% 3.23% 2.97% 3.09% 3.65% 3.14%
75 2.2
EFD -8.85×10-3 -
Sinkage [m]

0.570 2 CFD -8.48×10-3 -8.18×10-3 -8.20×10-3 -8.22×10-3 -8.19×10-3 -8.25×10-3 -8.28×10-3 -8.25×10-3 -8.18×10-3 -8.221×10-3
E 4.15% 7.52% 7.29% 7.09% 7.41% 6.76% 6.45% 6.78% 7.57% 7.11%
EFD -6.16×10-3 -
0.37 1.3 3 CFD -7.73×10 -3 -7.60×10 -3 -7.54×10 -3 -7.56×10 -3 -7.69×10-3 -7.61×10-3 -7.61×10-3 -7.60×10-3 -7.51×10-3 -7.591×10-3
E -25.46% -23.35% -22.47% -22.72% -24.83% -23.59% -23.59% -23.34% -22.00% -23.24%
40
EFD -6.16×10-3 -
0.41 1.6 4 CFD -6.01×10-3 -5.83×10-3 -5.85×10-3 -5.83×10-3 -5.88×10-3 -5.86×10-3 -5.89×10-3 -5.88×10-3 -5.77×10-3 -5.848×10-3
E 2.49% 5.28% 5.11% 5.37% 4.58% 4.85% 4.37% 4.60% 6.38% 5.07%
Table continued on the next page
84
 EFD -2.2×10-2 -
0.303 1 CFD -3.79×10-2 -3.56×10-2 -3.52×10-2 -3.58×10-2 -3.55×10-2 -3.58×10-2 -3.54×10-2 -3.59×10-2 -3.64×10-2 -3.568×10-2
E -69.15% -58.78% -56.92% -59.82% -58.38% -59.67% -57.84% -60.35% -62.29% -60.36%
75 2.2
EFD -1.03×10-1 -
0.570 2 CFD -1.07×10-1 -9.95×10-2 -9.75×10-2 -9.96×10-2 -9.95×10-2 -1.02×10-1 -1.01×10-1 -1.02×10-1 -1.03×10-1 -1.005×10-1
Trim [°]

E -4.49% 3.22% 5.20% 3.12% 3.21% 1.27% 1.82% 0.67% -0.62% 1.49%
EFD -2.24×10-2 -
0.37 1.3 3 CFD 3.97×10-4 -5.39×10-3 -6.01×10-3 -5.67×10-3 -4.81×10-3 -5.44×10-3 -4.97×10-3 -4.99×10-3 -6.23×10-3 -5.438×10-3
E 101.77% 75.93% 73.16% 74.69% 78.49% 75.70% 77.77% 77.73% 72.19% 75.71%
40
EFD -2.24×10-2 -
0.41 1.6 4 CFD -9.10×10-4 -5.37×10-3 -5.48×10-3 -4.86×10-3 -4.98×10-3 -5.42×10-3 -5.03×10-3 -4.88×10-3 -6.54×10-3 -5.319×10-3
E 95.93% 76.03% 75.53% 78.27% 77.74% 75.78% 77.55% 78.19% 70.79% 76.24%
Average 24.57% 2.20% 3.96% 1.65% 2.48% 5.05% 5.19% 3.47% 1.59% 3.13%

85
Particularly in the case of trim for case-studies 3 and 4, no turbulence model
has adequately predicted the experimentally measured results. This can stem
from a variety of sources. One interesting observation is that in Mucha and el
Moctar (2014) the reported Longitudinal Centre of Gravity (LCG) is different
from what is typically reported elsewhere. In the numerical work done here,
the LCG is located as prescribed in the relevant reference. However, this serves
to highlight a problem in the results: the two pieces of research found a
disagreement in a metric as fundamental as the LCG. In the presence of such
uncertainties, it is imperative that an inter-facility test is conducted and
assessed via a Youden (1972) style technique.

4.5.2 Numerical verification

The numerical verification adopted herein follows widely used procedures in

the ship hydrodynamics community (Xing and Stern, 2010). Specifically, the
Richardson Extrapolation (RE) procedure (Richardson, 1911). This consists of
expressing the error as an expanded power series with integer powers of grid
spacing (Δx) or time step (Δt) as a finite sum. If one assumes the solutions lie
within the asymptotic range, it is admissible to take only the dominant term
into account, leading to the so-called grid triplet study. Following Xing and
Stern (2010), the Grid Convergence Index (GCI) is used, devised by Roache
(1998). This method can be used to establish the uncertainty due to grid
spacing and time step errors, as demonstrated in Terziev et al. (2018),
Tezdogan et al. (2016b). This step is a crucially important part of any numerical
solution, and its omission in computational science and engineering is not
admissible (Roache et al., 1986). On the other hand, the iterative errors are
calculated as being virtually zero in all cases following the procedure of Roy
and Blottner (2001).

The GCI procedure begins by calculating the convergence ratio (R), defined as
the ratio between 𝜀21 = (𝑓2 – 𝑓1 ) and 𝜀32 = (𝑓3 – 𝑓2 ). Here, 𝑓𝑖 refers to the
solution obtained via the respective input parameter (mesh or time-step) using
the ith solution. The solutions (i) are obtained by systematically coarsening each
parameter by a factor of √2 (also known as the refinement ratio – r), as
recommended by ITTC (2008). Thus, in the case of grid dependence, the base
size is magnified by the above factor, whereas in the case of time dependence,
the time step is lessened by r. This procedure yields a total of four extra
solutions, used to define four possible scenarios for R for time and grid
dependence (Stern et al., 2006):

86
 1. Monotonic convergence, if 0>R >1
2. Oscillatory convergence, if R <1 and |R| < 1
3. Monotonic divergence, if R > 1
4. Neither error nor uncertainty can be evaluated

Once the type of convergence is known, the order of accuracy (Celik et al.,
2008), p is calculated as shown in Eq. (4.1). A monotonic convergence is
assumed in all equations presented herein.

𝑝 = ln(𝜀23 ⁄𝜀21 )⁄ln (𝑟) (4.1)

The next step is to find the extrapolated solution – Eq. (4.2), relative error – Eq.
(4.3), and extrapolated error – Eq. (4.4):
21
𝑓𝑒𝑥𝑡 = (𝑟 𝑝 × 𝑓1 − 𝑓2 )/(𝑟 𝑝 − 1) (4.2)

𝜀𝑎21 = |(𝑓1 − 𝑓2 )/𝑓1 | (4.3)

21 12 12
𝜀𝑒𝑥𝑡 = |(𝑓𝑒𝑥𝑡 − 𝑓1 )/𝑓𝑒𝑥𝑡 | (4.4)

Finally, the grid convergence index can be calculated as shown in Eq. (4.5):
21
𝐺𝐶𝐼𝑓𝑖𝑛𝑒 = 1.25𝜀𝑎21 ⁄(𝑟 𝑝 − 1) (4.5)

To obtain the solutions in the grid convergence study, the smallest time step
is used while coarsening the grid. On the other hand, to obtain the time step
convergence study results, the finest grid is used while lessening the time step.
Throughout the numerical verification procedure, surface mesh characteristics
were kept constant to maintain the accurate representation of the ship
geometry. The convergence studies are summarised in Table 4.6. The results
suggest that simulations are subject to greater uncertainty stemming from the
grid, rather than the time step. Although a GCI of 12.47% (for sinkage) may
seem high, Elsherbiny et al. (2019) report a 90% uncertainty for sinkage and
trim. In other words, the simulations exhibit uncertainty, with an order of
magnitude smaller than the experiment. The reason why this is the case is
related to experimental measurement equipment. The displacements used to
calculate sinkage and trim are too small to be reliably measured. Even in the
case of resistance, the numerical uncertainty (1.94% and 0.01%) is smaller than
the experimental uncertainty – 2.2%. This also explains the observed errors in
Table 4.5. While the resistance has been predicted reasonably well in all cases,
sinkage and trim are harder to capture, both experimentally and numerically.

87
Table 4.6. Convergence study for sinkage, trim, and resistance for case 1,
λ=75, h/T=2.2, 𝐹ℎ =0.303, EB model.
Grid dependence Time dependence
Paramete
Sinkage Resistanc Sinkage Resistanc
r Trim [°] Trim [°]
[m] e [N] [m] e [N]
r √2 √2 √2 √2 √2 √2
𝑓1 -2.002×10-3 -3.516×10-2 1.252 -2.002×10-3 -3.516×10-2 1.252

𝑓2 -2.181×10-3 -3.402×10-2 1.269 -2.000×10-3 -3.510×10-2 1.252

𝑓3 -2.655×10-3 -3.209×10-2 1.316 -1.946×10-3 -3.498×10-2 1.251

R 0.380 0.591 0.360 0.367 0.575 0.111

p 2.792 1.519 2.900 9.536 1.599 6.334

12
𝑓𝑒𝑥𝑡 -1.822×10-3 -3.630×10-2 1.235 -2.004×10-3 -3.523×10-2 1.252

𝜀𝑎21 8.99% 3.25% 1.37% 0.10% 0.18% 0.01%

21
𝜀𝑒𝑥𝑡 9.88% 3.14% 1.39% 0.10% 0.18% 0.00%
21
𝐺𝐶𝐼𝑓𝑖𝑛𝑒 12.47% 7.56% 1.94% 0.19% 0.38% 0.01%

Finally, the modelling errors induced by a change in turbulence models are

known to differ depending on the closure selected (Pereira et al., 2017).
Therefore, the results presented in Table 4.6 are not uniform across all case-
studies. However, having compared the solution against experimental data, it
is not though necessary to verify the numerical solution for each turbulence
model. The number of additional solutions required per case study (four)
implies an unreasonably high number (140) of additional simulations
necessary to bind the numerical error completely, as would be desirable. In
consequence, due to the extreme computational expense associated with the
procedure, this is left as a piece of future work.

4.5.3 Results comparison

Having established that resistance is predicted reasonably well by all

turbulence models (with the exception of AKN), it is worthwhile examining its
constituent components in more detail. In CFD assessments the resistance of a
body subject to fluid flow is defined as the sum of tangential (shear) and
normal (pressure) components. Each of these is non-dimensionalised by
division by 0.5ρU2S, where ρ=997.561 kg/m3 is the water density, and S is the
wetted surface area (CF and CP, respectively). Figure 4.6-Figure 4.9 show the
distribution of resistance among the two categories (shear and pressure) as
well the experimental values and error calculated for each turbulence model.

88
In the figures, the errors calculated by the AKN model have not been included
in order to preserve the y-scale of the error within reasonable values for the
remaining turbulence closures.

4.5.3.1 Pressure resistance

The results (Figure 4.6-Figure 4.9) suggest that all turbulence models predict
the pressure component of resistance well, with little disagreement between
the different approaches to closing the governing equations. This may seem a
trivial observation, since the pressure resistance is not strongly coupled with
viscous flow behaviour near impermeable walls. This has been widely used as
a justification for the coupling of RANS and potential flow solvers (Tahara and
Stern, 1994). However, this observation has wider implications: it suggests the
wave resistance, which is part of pressure resistance, is computed with good
accuracy regardless of the turbulence model. Wave resistance has been a
particularly important and difficult metric to accurately capture in the past for
several reasons. These include the complexity of the flow surrounding a ship.
In some cases, the deep-water wave resistance problem is thought to be
addressed reasonably well by potential flow. However, the vast majority of
theories break down for shallow waters. A good example illustrating this is
Michell's (1898) integral. As shown by Tuck and Lazauskas (2008), Beck (1977),
Beck et al., (1975), Tuck (1967, 1966), Michell's (1898) integral can be extended
to a large variety of shallow water problems, including deep water wave
resistance, but not to the prediction of shallow water wave resistance.

Figure 4.6. Resistance coefficient comparison for case 1

89
Figure 4.7. Resistance coefficient comparison for case 2.

Figure 4.8. Resistance coefficient comparison for case 3.

90
 Figure 4.9. Resistance coefficient comparison for case 4.

Traditionally, ship resistance is extrapolated from model to full-scale

following a towing tank experiment (Molland et al., 2017). Once the total
resistance has been obtained, it is decomposed into a frictional component, a
component due to wave making, and a viscous component. The assumption
is that a ship’s form factor and wave resistance components remain constant
with scale, while friction varies as prescribed by the friction line of choice
(Grigson, 1999; ITTC, 1999; Katsui et al., 2005). The findings presented herein
suggest the wave resistance (if assumed constant with scale) can be established
with sufficient accuracy regardless of the turbulence model employed at a
high scale factor, where the computational expense is small (refer to Table 4.4
for cell numbers). Finally, the well-established trend of increase in pressure
resistance with decreasing depth is observed in Figure 4.8 and Figure 4.9.

91
 Figure 4.10. Free surface profile along the hull for all case-studies.

One way to confirm that wave resistance is predicted well is to examine

whether any differences are present in the location of the waterline along the
hull. In this respect, in Figure 4.10, all wave-hull profiles are shown. The first
observation made here is the agreement established between all turbulence
models. The AKN closure seems to deviate slightly in cases 1 and 3. In case 2,
the k-ω SST model diverges from the other turbulence models near the stern
of the ship. Furthermore, it exhibits a slight oscillatory behaviour in case 4, not
predicted by other closures. Indeed Table 4.5 can be consulted to confirm the
presence of a systematic deviation in the predicted resistance values obtained
via the k-ω SST model. This is surprising considering its popularity among the
ship hydrodynamics research community. The reasons for this are explored
later in this chapter.

92
 4.5.3.2 Friction resistance

The difference between the smallest and largest skin friction coefficient
prediction, excluding AKN, is 2.7×10-3 for cases 1 and 2, and 2.9×10-3 for cases
3 and 4. This an interesting observation, because it serves to highlight that
depending on the turbulence model, vastly different results can be obtained in
terms of friction. Naturally, this has a strong influence on the total resistance,
since friction is one of the main contributing components. Keeping in mind the
y+ distribution shown in Figure 4.4 (Section 4.3.1), it seems unlikely that
predictions made by different turbulence models will collapse into a single
value were the grid to be refined further. The skin friction coefficients are
graphically compared to established friction lines in Figure 4.11. With the
exception of the two variants based on the SST model (k-ω SST, and k-ω γ), the
closures predict a significantly higher skin friction than any friction line would
suggest. This can be used to highlight the shortcomings of form factor and
wave resistance extrapolation techniques, since a reliable frictional resistance
is integral to the procedure (Terziev et al., 2019a). Confidence in the
predictions can be established due to their systematic predictions, as shown in
Figure 4.11, in terms of their relative location on the plot. To elaborate, k-ω
SST, and k-ω γ models are consistently the lowest predictions, whereas k-ε Lag-
EB – the highest.

There are several likely sources of error in the solutions presented herein
associated with the RANS technique. Stern et al. (2006) identify these errors as
turbulence modelling (which is being here), artificial compressibility (also a
part of some turbulence models), domain size (dictated by the experimental
facilities), and round off errors. It is generally accepted that the latter are
negligible. In Section 4.5.2 the numerical uncertainty was bound for an
example test case. This allows for the turbulence model to be identified as the
dominant contributor to all errors found.

In terms of 𝐶𝐹 , Fukagata et al. (2002) derived a relationship between skin

friction and the Reynolds stress distribution across a surface. Their results
confirm experimental observations suggesting that the Reynolds stress within
80 wall units (y+ <80) accounts for the vast majority of skin friction (up to 90%).
This region is viscosity-dominated and coincides with the location where
turbulence modelling can have the greatest impact. Indeed, one of the aims of
turbulence modelling is to predict a Reynolds stress. The three-dimensionality
of ship boundary layers is also a critically important fact, since closures are

93
calibrated for two dimensions in almost all cases . In fact, the skin friction is
known to consist of pressure development, laminar contribution, spatial
development, and Reynold stress components (Stroh, 2016). Shallow waters
amplify the three-dimensionality of boundary layer, which was one of the
motivating factors cited in the adoption of the case-studies (Zeng et al., 2018).

Figure 4.11. Skin friction comparison. Shapes connected via dotted lines
correspond to cases 1 and 2. Solitary coloured shapes correspond to case 3,
while the black shapes – case 4.

Finally, the results shown in Figure 4.11 suggest that a change in water depth
can have influence skin friction substantially. The mechanism by which this
occurs is not well understood. Zeng et al. (2018) presented a modification of
the ITTC friction line for shallow waters (ITTC, 2017a). Their derivation is
based on flat plates, and is only applicable to the flat region of a ship’s bottom,
whereas the friction line proposed by Katsui et al. (2005) is used for the
remainder of the wetted area. There are obvious problems with this, mainly
due to the fact that the KCS was allowed to sink and trim, rendering the
boundary layer on the flat bottom highly three-dimensional, unlike that of a
flat plate. The flow is also forced past the sides of the ship in virtually all
shallow water cases, which avoids the violation of the Bernoulli principle.
Figure 4.12 depicts the boundary layer and wake generated by the KCS in case
4. The top half of the figure was generated to represent 90% of the free stream

94
velocity (𝑈90%=0.657 m/s)., while the bottom half – 99% of the free-stream
velocity (𝑈99%=0.7227 m/s). Figure 4.12 shows the boundary layer is highly
three-dimensional, especially near the bow and stern, where the 99% free-
stream iso-surface is in contact with the seabed over a significant area. In the
figure, the dynamic pressure is used to colour the surfaces. This is done to
highlight the differences in pressure, acting longitudinally and transversely
due to the proximity of the seabed. Finally, the waves resulting from the
disturbance, caused by the ship also register in the plot.

Figure 4.12. Boundary layer and wake (top view), depicted: case 4. Top: iso-
surfaces for 0.9×𝑈. Bottom: iso-surfaces for 0.99×𝑈.

4.5.3.3 An attempt to identify the optimum turbulence model

The problem of consistency can be addressed via a modified Youden (1972)

plot. Although the original purpose of the technique proposed by the
abovementioned author was to establish experimental biases, it can also be
used to provide an indication of accuracy and consistency. The required
modification is that instead of dimensional quantities, plotted in the x and y
axes, the error is used, calculated in Section 5.1.1, Table 4.5.

The modified technique begins by plotting the error for each parameter in an
x-y plane. As shown in Figure 4.13, the y-axis is chosen to represent error in
resistance as calculated for case-studies 2 and 4, whereas the x-axis: the error
in case-studies 1 and 3. Thus, the empty shapes in Figure 4.13 represent the
points with coordinates [x, y] = [E1, E2] where the subscripts refer to case study
number, and E is the error. The filled shapes correspond to points with
coordinates [E3, E4]. For example, the empty blue circle (v2 – f model) in Figure
4.13 has an x – coordinate equal to the total resistance error for case 1 (3.74%),

95
and a y – coordinate equal to the total resistance error for case 2 (0.24%).
Similarly, the filled blue circle has x – coordinates equal to the 𝐶𝑇 error for case
3, and y – coordinate equal to 𝐶𝑇 error for case 4.

As described by Youden (1972), the next step is to construct the straight lines,
representing the median x and y values. Depending on the location of each
point within the plot, a systematic bias, or lack thereof can be established. For
instance, a point lying in the first quadrant (the standard convention is
adopted), established via the intersection of the median values, suggests a
systematic overprediction. Using the intersection of the median x and y values,
a 45-degree diagonal is drawn. This diagonal can be useful in giving a measure
of consistency each turbulence model provides in its predictions. In other
words, points close to the diagonal perform in a similar manner in all four
case-studies, i.e. systematically. An extra layer of information can be extracted
due to the fact that the error was made use of in Figure 4.13.

By requiring that the error is the metric against which the points are assessed,
it is possible to simultaneously take into account the closures’ performance
against the relevant experiment. An additional consequence of the above
choice is that the experimental uncertainty can be drawn directly onto the plot.
For case-studies 1 and 2, this is 2.20%. For cases 3 and 4, the uncertainty is not
reported, therefore, the same magnitude as in cases 1 and 4 is assumed. Armed
with this information, a circle, centred at [0, 0], with a radius of 2.20 can be
drawn, giving a graphical interpretation of the error and biases in all four cases
(the red circle in Figure 4.13). If all experimental uncertainties were known, an
ellipse would be required, as is the case with modified versions of the plot
(Velázquez and Asuero, 2017).

The mathematics behind different turbulence models cannot provide any

information that could forecast whether the values will under or overpredict
the experiment prior to running the simulation. If all turbulence models were
made equal, the values would be scattered evenly in each quadrant around
the median x-y intersection. This is true because all models are calibrated
against a similar set of experiments, and therefore should be equally robust.
Therefore, displacement along the diagonal can be interpreted as bias, whereas
offset from the diagonal is an indication of erratic behaviour.

96
 Figure 4.13. Bivariate error plot for resistance.

A high deviation from the circle (representing the experimental uncertainty)

indicates a turbulence model’s inability to cope well with the case it is applied
to. Thus, the proximity of k-ω Wilcox, and k-ε EB (and its Lag extension)
models, coupled with their small distance from the diagonal, indicates they
perform well, consistently, and reliably. By contrast, the k-ε 2l model shows
erratic behaviour, over and underpredicting results by a large margin
depending on the case study. Thus, the results suggest that applying this
closure, the user cannot hope to predict with any confidence what the outcome
of a simulation will be. If either of the well-performing models mentioned
above were to be used, one could expect to achieve results with a small,
positive error. That is, provided the numerical set-up is reasonably similar to
that presented herein.

Unfortunately, Figure 4.13 takes into account only resistance. Thus, no

information regarding the turbulence models’ behaviour in sinkage and trim
can be deduced from it. For this reason, Figure 4.14 was constructed using
case-studies 1, 2 and 4. On the x-axis, the error in sinkage is plotted, while on
the y-axis – the corresponding case study’s error in resistance. As before, the
empty shapes represent case 1, the filled shapes – case 2, and the black shapes

97
– case 4. Case 3 is omitted due to the large errors in sinkage, which would push
the x scale too far for any meaningful discussion.

Figure 4.14. Bivariate error plot: sinkage vs. resistance.

A consequence of the arrangement of the points is that the diagonal,

constructed from the intersection of the median x and y values has a negative
slope. Since different experimental uncertainties are reported in Elsherbiny et
al. (2019) for each axis, the shape constructed in Figure 4.14 is a rectangle with
dimensions 3.47×2.2, matching the uncertainties in sinkage and resistance,
respectively. As before, the lack of reported data for case-studies 3 and 4
means a uniform cross-experimental uncertainty is assumed.

All turbulence models seem to follow certain tendencies in Figure 4.14, where
k-ε type models tend to fall beneath the diagonal, while k-ω based models are
located above it. A notable exception to this is the k-ε 2l model, which has two
points above, and one below the diagonal. Figure 4.14 happens to be
constructed so that points below the diagonal are closer to the limits of the
rectangle, and therefore closer to the experimental uncertainty bounds.

98
It is important to keep in mind that the ‘distance to the diagonal’ metric is
purely artificial. However, since it depends on the overall scatter of the
predictions, it serves as a good indication of erratic behaviour. As mentioned
previously, if all turbulence models were equally robust, they would all lie
within a small circle, centred at the median x and y values. Large deviations,
prevalent to the k-ε-class models, point towards unreliable predictions. This is
not in reference to the accuracy itself. Indeed, in some cases their predictions
fall within the experimental uncertainty (e.g. v2 – f, EB, and LagEB models in
resistance for case 2). Instead, reliable turbulence models are those that
preform similarly across all case-studies they are applied to.

For Figure 4.14, the set of criteria used to establish a hierarchy of performance
follow from the assessment performed for Figure 4.13. Namely, the quadrant
in which each point lies, as well as the distance to the diagonal, and origin are
taken into account. These metrics, in addition to their overall Root Mean
Square (RMS) values have been constructed graphically in Figure 4.15, where
only the results depicted in Figure 4.14 are used. Here, the k-ω Wilcox model
exhibits the desired characteristics to greater extent than the remaining
candidates. Firstly, the distance to the diagonal is small throughout, relative
to other cases. This is an indication the turbulence model is not an outlier in
the scatter of predictions. Secondly, the distance to the origin is reasonable.
Finally, two out of three points predicted via this model fall within the same
(2nd) quadrant, indicating a degree of consistency.

99
 Figure 4.15. Distances in the Error-Error plane and RMS error for case-
studies 1-3.

The k-ε EB model performs marginally worse, with all three points within
different quadrants. Other strong candidates include the v2 – f model, whose
predictions tend to lie close to those obtained via the k-ε EB closure. Perhaps
surprisingly, the results suggest that the two models accounting for over 80%
of all research done in the field of ship hydrodynamics should be ruled out
(based on the dataset established earlier). Namely, the k-ε 2l and k-ω SST
models (see Figure 4.1). On the other hand, the SPAL model performs almost
as well as the more complex, recent closures, such as k-ε EB and Lag-EB
models. In part, the findings of Eça et al. (2018) have been echoed, who found
that for the k-ω SST model, a y+≈0.1 is required to achieve the same level of
accuracy as other turbulence models. This requirement restricts the use of the
k-ω SST closure, because of the difficulties related to the constraint.
Specifically, the cell numbers would increase substantially, as well as the their
aspect ratio, which compromises the solution’s stability and convergence
properties (Eca et al., 2015; ITTC, 2011).

100
 4.5.3.4 Quantification of the computational resource

The final metric, against which turbulence models are evaluated is time per
iteration. Although the solver’s time per iteration has been recorded
throughout the entire duration of the simulation, only values achieved during
the first 5 seconds are reported. This is done because once the ship is allowed
to translate and rotate, the time history is contaminated with the numerical
algorithms responsible for doing this. Additionally, the auto save function of
Star-CCM+ was employed to routinely save the simulation (every 5 seconds of
physical time), which registers as a spike in the time per iteration monitor.
Therefore, global averages would not be a suitable metric.

Figure 4.16. Time per iteration for case 1.

All simulations were performed using the high performance computer

facilities at the University of Strathclyde, Archie-WeSt. The facility features 2
Intel Xeon Gold 6138 20 core 2.0GHz CPUs, with 40 cores per node, and 192
GB RAM allocated per node. To run the simulations described herein, one
node (with 40 cores) was used per case per turbulence model (i.e. per
simulation). The resulting time per iteration is shown graphically in Figure
4.16 for case 1. It should be stressed that all results presented in Figure 4.16 are

101
highly dependent on the solver employed. The relationship between different
closures would most likely be maintained if carried forward to a case using
Star-CCM+ featuring different cell numbers. However, scalability is not
uniform across all RANS solvers, as discussed by Robertson et al. (2015).

The relationship between core numbers and efficiency in obtaining the

solution is known to be inverse in nature (Axtmann and Rist, 2016). Therefore,
as the cell numbers increase, the difference in computational time will likely
decrease. For this reason, case 1, with the smallest number of cells, is chosen
for this assessment. In this respect, the development of algorithms tasked with
optimising parallelisation is an active field of research, although its
relationship with DNS is stronger than RANS approaches (Cifani et al., 2018;
Kooij et al., 2018; van der Poel et al., 2015).

As expected, the SPAL model provides the smallest computational time

requirements, and is used to establish a proportional increase metric for all
other models. The k-ω Wilcox model provides the smallest time after the SPAL
closure, closely followed by k-ω SST. All k-ε variants exhibit a higher time per
iteration due to the additional near-wall treatments or extra equations
introduced. For instance, the main source of differences between k-ω Wilcox
and k-ε 2l in terms of time is the realizability requirements for the latter. On
the other hand, the imposition of a γ transition is shown to increase
substantially the computational requirements, while providing little, if any
improvements in terms of accuracy, consistency or reliability of the solution.
Furthermore, introducing additional equations into the turbulence closure
implicitly destabilises the solution. Specifically, more ways in which the
numerical calculation could fail are introduced. This is sometimes thought of
as numerical stiffness of the solution, and was the reason given for employing
a 2nd order convection scheme, rather than the more accurate 3rd order option.
In this respect, the practitioner must have a sense of the accuracy, consistency,
and computational resource required. Improving the latter is normally
associated with a detrimental effect in the former two. The results suggest this
trade-off may not always be necessary. That is, provided a suitable numerical
set-up is employed.

One final aspect requiring discussion is associated with the solver itself.
Research has shown that it is dangerous to generalise outcomes achieved in
one software to another. Specifically, the same mesh and set-up can lead to
different solutions depending on the solver employed. This problem was
highlighted by Iaccarino (2002), who found significant deviations across

102
RANS software. In consequence, the hierarchy established in this thesis
would, in all likelihood differ if repeated using a different solver. This is true
because software providers do not disclose all information related to the
algorithms and sub-routines used. It is therefore not possible to predict what
the effect of replicating the work reported in this chapter elsewhere would
have.

4.6 Summary and conclusion

This chapter has focused on reducing the uncertainty in turbulence model

selection for a class of problems in ship hydrodynamics. Nine turbulence
models were identified as potential candidates for the assessment based on a
literature survey. Emphasis was placed on devising a hierarchy depending on
performance, since it is not possible to use any other meaningful metric. The
adopted case-studies, attempting to replicate a set of shallow water
experiments, were specifically selected to challenge the turbulence models.

The validity of the implemented numerical set-up was demonstrated by

validating the solutions against experimental data and by estimating the
numerical uncertainty for an example case study. The results indicate that
pressure resistance and its constituent components are predicted well by all
turbulence models. This finding has wider implications in the extrapolation of
ship resistance from model to full-scale, as it implies wave resistance is largely
independent of turbulence modelling. Friction resistance was shown as the
main source of errors. There are several identified areas contributing to the
observed discrepancies, other than those due to the numerical set-up, which
are bound within acceptable levels. While turbulence models tend to be
calibrated against two dimensional cases, a strong three-dimensional
boundary layer is observed in the investigated cases.

The observed difference between the highest and smallest frictional resistance
prediction was calculated to be nearly constant across all case-studies. This
indicates that a strong turbulence modelling sensitivity should be prevalent
for full-scale computations, as forecasted by Duvigneau et al. (2003). The
difficulties in creating sufficiently fine meshes to resolve the boundary layer
at full-scale is also a suspected contributor to the here-described uncertainty.
Specifically, the work presented herein can be said to have established a
hierarchy for similar numerical set-ups. If the y+ values are varied significantly,
the ranking establishing could well be different.

103
An attempt to rank turbulence models with respect to their accuracy and
consistency was performed, which identified the k-ω Wilcox and elliptic
blending models as exhibiting the best performance. The assessment was
carried out via a modified Youden bivariate plot, where the error in relation
to the experimental values was plotted on both axes. The performance criteria
included are thought to be capable of providing an initial estimate of the
outcome prior to completing the simulation. Coupling these findings with the
time per iteration characteristics of each turbulence model suggests the
optimal choice is the k-ω Wilcox turbulence model for bare hull ship resistance
computations.

104
 A NUMERICAL ASSESSMENT OF
THE SCALE EFFECTS OF A SHIP
ADVANCING THROUGH
RESTRICTED WATERS

This chapter uses the findings of the previous two chapters of this thesis to
assess scale effects in restricted waters. A geosim series analysis is performed
on a containership advancing through a canal. Multiphase and double body
simulations are performed as part of the assessment. Viscous scaling is also
explored for the adopted case study. Numerical estimates of the scale effects
are shown to conform to theoretical analysis of the impact of viscosity and
vorticity on the flow around the ship.

5.1 Introduction

Historically, naval architecture has primarily relied on Experimental Fluid

Dynamics (EFD) due to the lack of consistently reliable theoretical predictions
in the field. The advent of analytical and computational methods has not done
enough to encourage naval architects to adopt theoretical predictions in their
toolkit. Even where this has been the case, computational work usually takes
a secondary place. While experimental work has its distinct advantages, the
tendency of overreliance on the EFD predictions has some major drawbacks.

Experiments are expensive, especially shallow water cases (Jiang, 2001), they
require time, as well as facilities with adequate equipment. Even if all of these
requirements are satisfied, one can run into the assumptions of the

105
extrapolation procedures used to determine the full-scale parameters.
Specifically, scale effects have been documented in every component of ship
resistance (García-Gómez, 2000; Kouh et al., 2009; Raven et al., 2008).

The case of shallow water presents an additional layer of complexity, because

scale effects are expected to be greater than in unrestricted waters (Tuck, 1978).
Here, it is important to distinguish between inland ships, which spend their
entire operational lives in restricted waters, and seagoing ships. Shallow water
studies merit investigation because even seagoing ships enter shallow waters
multiple times each voyage. It is precisely in these cases that a significant
proportion of accidents occur according to EMSA (European Maritime Safety
Agency, 2019, 2018, 2017, 2016, 2015). However, this does not represent the full
picture. Inland transportation will play a major role if carbon dioxide
emissions due to transportation are to be reduced. This has led to policies
aimed at encouraging the use of navigable rivers and canals (Caris et al., 2014;
European Commission, 2018; Mihic et al., 2011).

To facilitate the transition to safer operations in shallow waters, the underlying

hydrodynamic phenomena must be better understood. An action taken by the
ship in deep water can have counter-intuitive consequences in shallow water
(Tuck, 1978). These consequences are caused by the hydrodynamic interaction
between the ship’s hull and the surrounding bathymetry. Effects include a
reduction in under keel clearance which translates into a grounding hazard.
Additionally, the resistance is known to increase, and the manoeuvrability
characteristics are compromised (Fujino, 1976; Millward, 1996). Faced with the
above challenges, many analysis methods are either inapplicable, or perform
poorly in shallow water.

The primary goal of this chapter is to examine scale effects of the total
resistance and its constituent components in confined water. It is important to
mention that there have been reports of scale effects in sinkage between
model- and full-scale measurements in shallow water (Dand, 1967; Duffy,
2008; Ferguson, 1977; Shevchuk et al., 2019; K. Song et al., 2019). The fact that
external parameters, such as wind and waves are impossible to control, as well
as the difficulties one faces in full-scale measurements might preclude the
identification of the specific root these scale effects experimentally. It may not
always be possible to ascertain whether a true scale effect is observed, or if the
apparent differences are due to uncontrolled parameters, such as surface
roughness, bathymetry irregularity, etc. The adopted case-studies therefore
neglect the effects of sinkage ad trim.

106
The lack of experimental data at different scale factors (i.e. a geosim series in a
controlled, laboratory environment) for the same ship in confined water
motivates a purely numerical study in all but the smallest scale factor, where
data is available. The geosim analysis is applied on the well-known KCS hull
form, with conditions replicated from recent experimental work, reported in
Elsherbiny et al. (2019). To reveal scale effects, double body and multiphase
simulations are performed. In the present context, double body simulations
refer to the modelling approach where the free surface has been replaced by a
symmetry plane. This has the effect of eliminating the wave resistance
component from the total resistance. The novelty the work presented in this
chapter is expressed in the approaches used to determine the parameters of
interest, as well as the adopted case study.

The remainder of this chapter proceeds with a brief description of the

methodology in Section 5.2, which also contains the ship geometry and case-
studies. Section 5.3 contains the results and relevant discussion, whereas a
summary and conclusion is given in Section 5.5.

5.2 Methodology

This section is split into two major parts. In the first part, the overall procedure
and case studies are presented. The second section contains a description of
the numerical set-up used, together with details regarding its implementation.

5.2.1 Approach to the problem at hand

The approach to the problem adopted herein is to perform a numerical

simulation in a single scale factor, where experimental data is available. In
particular, the work of Elsherbiny et al. (2019) was selected. Verification is
performed to demonstrate the efficacy of the numerical set-up for this case.
Specifically, the well-known KCS ship, without appendages was used, in a
depth to draught ratio of 2.2 and a depth Froude number (𝐹ℎ ) of 0.303.
Although different speeds are also available, as a result of the experiment,
𝐹ℎ =0.303 was selected as it guarantees a reasonable speed when full-scale is
reached (approximately 9 knots). This is chosen to increase the practical
relevance of the presented numerical simulations.

The choice of the next, higher scale factor (λ) is trivial in the absence of
experimental data. For this reason, it was decided to divide λ by 2, followed
by a full-scale simulation to arrive at the scale factors and ship properties
shown in Table 5.1. In Table 5.1, the field labelled as “Dynamic viscosity” is

107
used to reproduce the approach of Haase et al. (2016). In the aforementioned
study, the authors devised a procedure whereby a modification of the value
of viscosity, a ship may satisfy both Reynolds and Froude similarity
simultaneously. The value highlighted in bold is the default used in all
simulations, whereas the fields corresponding to the remaining scale factors
contain the value used to push the Reynolds number its corresponding value
for each λ. This is performed while the linear dimensions are maintained the
same. The approach allows one to use a single grid for scale effects
assessments.

Table 5.1. Case-studies

Quantity Symbol Value Unit
Scale Factor λ 75 37.5 1 -
Length L 3.067 6.133 230 m
Beam B 0.429 0.859 32.2 m
Draught T 0.144 0.288 10.8 m
Depth D 0.253 0.507 19 m
Water depth h 0.317 0.634 23.760 m
Block coefficient CB 0.651 0.651 0.651 -
Longitudinal Centre
LCG 1.488 2.976 111.593 m
of Gravity
Wetted area S 1.694 6.777 9530 m2
Speed U 0.535 0.756 4.630 m/s
Reynolds number Re 1.840×106 5.205×106 1.195×109 -
Dynamic viscosity μ 8.8871×10-4 3.1421×10-4 1.3683×10-6 Pa-s
Depth Froude number Fh 0.303 -

In Chapter 3, the same approach alongside linear scaling and demonstrated

that the approach provides results that are close to those obtained by a
traditional double body simulation (Terziev et al., 2019a). Sezen and Cakici
(2019) performed a similar study, but arrived at the opposite conclusion. This
is the case for several reasons. Firstly, in Chapter 3, solely double body
simulations were performed using the approach referred to as “viscous
scaling”. This eliminates the issue of viscous effects on the free surface,
resulting from the change in the physical properties of the fluid surrounding
the ship. However, in the evaluation the performance of the viscous scaling
approach, the authors (Sezen and Cakici, 2019) assumed that the residuary
coefficient must remain constant. Moreover, the initial methodology of Haase
et al. (2016a) validates a grid in model scale, then repeats the simulation with
a change in the viscous properties of the fluid. No change in mesh was
originally envisioned. Although it is true that the y+ values cannot remain the

108
same between the different case-studies, a change in mesh characteristics
voids the first step, in which validation is performed. In addition, the
approach’s appeal is expressed in the fact that low cell numbers can be used
to perform a full-scale simulation. Reconstructing the mesh and matching the
y+ values negates this appeal as it corresponds to a drastic increase in cell
numbers.

Once the viscously scaled simulation has run its course, the results are
multiplied by the ratio of scale factors to the third power, thus correcting the
discrepancy in linear dimensions.

5.2.1 Numerical implementation

The placement of the inlet and outlet boundaries follows the recommendations
of ITTC (2011) and is shown in Figure 5.1. The domain top is placed at 1.25×L
from the undisturbed water surface, where a velocity inlet condition is
imposed. The domain bottom is set to match the experimental condition of
h/T=2.2 in all scales, specified as a velocity inlet. Such a boundary condition
guarantees that there will be no relative motion between the fluid and the
seabed. A velocity inlet may also be preferable due to the fact that open
boundaries have a stabilising effect on the numerical solution. In any case, the
use of velocity inlets to represent the domain bottom has been validated in
recent studies (Elsherbiny et al., 2020). The side boundary is also positioned
following the experiment, at 2.3m from the ship centreline in λ=75, and is
scaled accordingly. The accompanying mesh for the full-scale multiphase
simulation is shown in Figure 5.2, whereas Table 5.2 contains the resulting cell
numbers for all simulations. It should be noted that the multiphase simulation
for λ=75 corresponds to the cell numbers used for viscous scaling. To ensure
that the longitudinal extent of the computational domain does not impact
detrimentally the solution, the domain was extended by one ship length on
either side of the ship and the simulation repeated. This revealed no
discernible change in the results. The near-wall mesh is set to maintain an
average of y+<1 in all model-scale computations, whereas its average value in
full-scale is approximately 300.

The latter value is thought to provide sufficiently accurate results for full-scale
simulations (Peric, 2019).

The simulations which contain both air and water and represent the
experimental set-up as accurately as possible are labelled as multiphase. Here,

109
the interphase between the two mediums is modelled by the Volume of Fluid
(VOF) method (Hirt and Nichols, 1981). This is a standard approach adopted
in the vast majority of marine CFD where the resolution of the free surface is
important. The VOF method is also used in Star-CCM+ to model air and water
currents and therefore the ship’s speed. This is done via the concept of a flat
wave, and is set appropriately for each scale, as shown in Table 5.1. The
velocities specified at the inlet boundary, while the outlet is required to
maintain the hydrostatic pressure.

Figure 5.1. Domain characteristics and boundary conditions

Figure 5.2. Full-scale mesh generated in Star-CCM+

Table 5.2. Cell numbers for all simulations.

Scale factor 1 37.5 75
Multiphase 26644375 7938801 4046168
Double body 14339889 2387454 1050032

To enable the assessment of scale effects, an estimate of the wave resistance

and form factor is necessary. In CFD, these can be extracted by performing
what is known as a double body simulation. In essence, this is equivalent to
replacing the free surface with a symmetry plane. Thus, wave resistance is no
longer a component of the total, as shown in Eq. (5.1a) for multiphase regime,
and Eq. (5.1b) for double body regime:

CT =CF×(1+k)+CW (5.1a)

CT = CF×(1+k) (5.1b)

In Eq. (5.1a) and Eq. (5.1b), all resistance parameters are shown in non-
dimensional form, achieved by division by 0.5×S×𝑈2×ρ, where S is the wetted

110
area in m2, 𝑈 is the ship velocity in m/s, and ρ is the water density (997.561
kg/m3). To obtain the wave resistance, one must simply subtract the total
resistance in double body mode from the multiphase condition, while the form
factor (1+k) is obtained by division of CTdb by CFdb (Eq. (5.1a) and Eq. (5.1b):
CTdb/CFdb) (Molland et al., 2017), where the m subscript refers to multiphase
solutions, while db indicates double body. It is important to note that CFD
predicts ship resistance (CT) as the sum of normal (pressure resistance CP –
which contains 3D effects (viscous pressure) as well was wave resistance (CW)),
and tangential (frictional resistance CF) components. It should be noted that in
this chapter, flat plate friction lines are not used to determine the form factor
as is typically done according to the ITTC (1999). This done in favour of the
frictional resistance coefficient obtained from CFD because this matric has
been shown to be highly sensitive to ship underwater form as well as depth,
rendering the usually used friction lines inapplicable to shallow waters. In
other words, the approach of Zeng et al. (2019) is followed.

The standard k-ω model (Wilcox, 2006) is adopted, as implemented in Star-

CCM+ version 13.06.012. The stability and consistency of the k-ω model for the
class of problems examined here was demonstrated in Chapter 4 (Elsherbiny
et al., 2020). Moreover, it proved the least computationally expensive two-
equation turbulence model. The k-ω model showed an increase in solution
time of 8% compared to a one-equation turbulence model, whereas the k-ε
model increased the wall time by approximately 16%. As a result of the high
relative importance of turbulent properties, the convective term is set to 2nd
order.

The temporal evolution of the solution is resolved via a first order implicit
unsteady scheme, with a time step (Δt) equal to Δt = 0.0035×L/𝑈. This has been
demonstrated to be a good choice in several works, and is adopted (Tezdogan
et al., 2016a). However, it is important to state that any discretisation of the
temporal term of the Navier-Stokes equations will inevitably result in some
numerical error. These are explored in the following section. The remaining
physics, modelled by the incompressible RANS equations, are solved for
numerically via the segregated flow solver offered in Star-CCM+. More details
can be accessed in Siemens (2018).

111
 5.3 Results and discussion

5.3.1 Verification study

As mentioned previously, the first step in the procedure is to determine the

uncertainties of the numerical set-up of λ=75. The predicted multiphase total
resistance coefficient (5.123×10-3) shows reasonable agreement with the
experimental value (5.505×10-3), underpredicting the result by -6.85%. This is
thought to be sufficiently accurate, especially considering that in the
experiment, the ship was allowed to sink and trim, whereas during the
numerical simulation it was kept fixed. The verification study is presented in
Table 5.3 for spatial and Table 5.4 for temporal discretisation, respectively. It
should be noted that the relevant equations and relationships used in the
production of Table 5.3 and Table 5.4 are omitted. Instead, the reader is
referred to the report by the ITTC (2017b). To compute the numerical
uncertainty, the Grid Convergence Index (GCI) is used, which is typically
treated as the standardised approach to reporting numerical uncertainties.

Table 5.3 and Table 5.4 contain the grid and time-step studies for the numerical
simulations in both physics regimes (multiphase and double body). The
results indicate that the largest uncertainty can be expected from the
multiphase RANS simulation (3.348%). In terms of temporal dependence, the
simulations do not show significant errors. According to Table 5.3 and Table
5.4, the numerical simulations (regardless of physics approach) are more
sensitive to grid refinement than they are to a change in the time step.

The choice of refinement ratio is of critical importance in verification studies

(Phillips, 2012). This is used as a multiplicative factor to the grid size or time
step to coarsen the grid. The choice of √2 is chosen in line with the
recommendations of the ITTC (2008). In general, the refinement ratio should
be chosen to attain a value between 1.1 and 2, as suggested by ASME
(American Society of Mechanical Engineers, 2009).

The achieved grid numbers were as follows. The double body cell numbers for
the medium and coarse solution numbered 679,472 and 480,040, respectively.
Similarly, the multiphase cell numbers were 2,384,829 and 1,395,411. In the
process of coarsening the mesh for the verification study, the properties of the
mesh used in defining the surface of the ship have been maintained identical.
This is done to preserve an accurate representation of the ship geometry in the

112
process of determining the numerical uncertainty. Such an approach was
adopted by Tezdogan et al. (2016b, 2015) and is followed here as well.

The wave resistance coefficient’s numerical uncertainty characteristics are also

shown in Table 5.3 and Table 5.4 for spatial and temporal discretisation,
respectively. Here, 𝐶𝑊 is calculated as the difference of the multiphase and
double body resistance values at each refinement level. In the case of time
dependence, the double body simulations exhibit a smaller variation in
resistance characteristics than the multiphase results. Therefore, 𝐶𝑊 is
predicted to exhibit an oscillatory behaviour. For this case, the modified
relationships, as given in the recent work of Song et al. (2019) are used to
predict the uncertainty, since they can cope with oscillations in the data.

Table 5.3. Spatial discretisation-induced numerical uncertainty (for λ=75).

The wave resistance coefficient listed in this table was arrived at by
subtracting the double body resistance from the multiphase resistance.
Parameter Multiphase resistance Double body resistance 𝐶𝑤
Refinement ratio √2 √2 √2
Fine 5.123×10-3 4.752×10-3 0.371×10-3
Medium 5.607×10-3 4.745×10-3 0.862×10-3
Coarse 6.877×10-3 4.720×10-3 2.157×10-3
Convergence Monotonic Monotonic Monotonic
Order of accuracy 2.792 3.377 2.798
GCI (%) 3.348 0.020 0.014

Table 5.4. Temporal discretisation-induced numerical uncertainty (for λ=75).

The wave resistance coefficient listed in this table was arrived at by
subtracting the double body resistance from the multiphase resistance.
Parameter Multiphase resistance Double body resistance 𝐶𝑤
Refinement ratio √2 √2 √2
Fine 5.123×10-3 4.752×10-3 0.371×10-3
Medium 5.215×10-3 4.793×10-3 0.422×10-3
Coarse 5.293×10-3 5.010×10-3 0.289×10-3
Convergence Monotonic Monotonic Oscillatory
Order of accuracy 2.472 4.79 2.8930
GCI (%) 1.109 0.016 0.0012

The uncertainty estimation technique also requires that other sources of error
are small. These include round-off error and iterative error (Ferziger and Peric,
2002). The former is thought negligible in most cases, whereas the latter can
have a significant impact. Iterative errors are assessed via the procedure of
Roy and Blottner (2006). The results suggest that the smallest iterative errors
are found in the case of double body simulations are negligible. On the other

113
hand, the RANS multiphase simulation demonstrated an iterative error of circa
0.08%, which is considered sufficiently small. To ensure that the solution has
converged, the resistance time-history is monitored alongside the residuals.
The former are allowed to decrease by at least three orders of magnitude
before the solution is stopped.

5.3.2 Numerical results

In this section, the computed skin friction data are shown for each scale factor
according to the three different methods in Figure 5.3 along some established
friction lines. Here, it is evident that the viscous scaling procedure may be used
with good accuracy to determine the frictional resistance coefficient. This
conclusion may be drawn from the fact that the difference between the linearly
scaled multiphase predictions and their viscously scaled counterparts are not
substantial. These seem to increase as the Reynolds number approaches its
full-scale value, where the viscously scaled simulation predicts the skin
friction within 0.1% of the double body result.

Figure 5.3. Skin friction coefficients calculated at each scale and established
friction lines, used to demonstrate the relative difference between the
shallow water line of Zeng et al. (2019) compared to other predictions.

114
Discrepancies between double body and multiphase results may stem from a
variety of sources. These include the small changes of the wetted surface area
resulting from the deformation of the free surface. Such effects have been
neglected in the present study. Alternatively, research has shown that vortex
shedding is modified as a result of the presence of a free surface (Suh et al.,
2011). Moreover, such an influence has been documented experimentally by
Dand (1967). The same researcher also predicted co-dependence of wave and
frictional resistance of a flat plate. Thus, the changes observed in the frictional
resistance coefficients are not strictly a manifestation of numerical
assumptions.

Now, it is important to put the findings presented in Figure 5.3 in context and
compare the data with other research conducted recently. For this purpose,
the friction line, specifically designed for the KCS in shallow water by Zeng et
al. (2019) is included alongside the remaining friction lines. One may draw an
immediate conclusion that the frictional coefficient is predicted with high
accuracy in both model scale factors examined. Indeed, the line of Zeng et al.
(2019) outperforms any of the remaining lines in the field. Naturally, this is
solely due to the shallow water effect, which is not accounted for in the
derivation of any other friction line. However, the full-scale results derived
from the present study indicate a problematic trend.

In reaching full-scale Reynolds numbers, the friction line of Zeng et al. (2019)
exhibits too great a slope. Thus, the frictional resistance coefficients do not
agree well with the data found here. Simultaneously, lines with milder slopes,
specifically that of Grigson (1999) and Gadd (1967) are closer to the full-scale
data. This suggests that at full-scale, the frictional resistance coefficient may be
affected by the depth restriction to a lesser extent (especially considering the
fact that Gadd's (1967) line was also found to be in good agreement with deep
water predictions in Chapter 3). The information presented here also points
towards the fact that lateral restrictions might not impact the ship resistance
significantly in terms of frictional resistance. That is at least at the restriction
level posed by the adopted case study. However, one may expect that upon
reaching significantly more restricted waters, such as narrow canals, the bank
effect would be noticeable in the frictional resistance coefficient. In summary,
the friction line of Zeng et al. (2019) is shown to perform well in model scale.
To determine if the observed discrepancy in full-scale is due to the numerical
set-up adopted here requires further research.

115
For instance, it may be the case that in full-scale, the effect of the bank is greater
than in model scale. To prove or disprove this, analysis is required for different
widths, although an attempt at quantifying such an influence is made later in
this section. Such assessments do not seem popular in the literature due to the
fact that the water depth has a greater bearing on the parameters of interest.
One final aspect of the solution that one should consider is the highly specific
nature of the friction coefficients and associated line devised by Zeng et al.
(2019). The solution included in Figure 5.3 was generated specifically for the
KCS. Indeed, within their work, Zeng et al. (2019) produced lines for two other
hull forms. Unfortunately, generalisations to other ships are not possible due
to the highly specific nature of the flow in shallow water, which depends
heavily on the ship form. This also points to the fact that each underwater
shape influences the frictional resistance even in deep waters. Thus, the use of
friction lines universally might not be the best approach.

To further support the argument laid out previously, that a free surface
modifies the boundary layer, Figure 5.4 depicts the numerical boundary layer
extents in the smallest and largest scale factors. Typically, the extent of the
boundary layer is taken as the location where the velocity near a body reaches
99% of its free stream value. In the present case, it was found that such a
condition does not lead to a single line, rather, to a small area where the flow
attains practically the same speed. For illustration purposes, the boundary
layer definition has been slightly altered to 90% of the free stream velocity.
This is sampled at four locations, namely, at the forward perpendicular,
amidships, at 0.25×L and at the aft perpendicular.

Even after restricting the definition of the boundary layer, it is apparent that
amidships the flow velocity near the free surface exhibits several z/L points
with the same velocity for a single y/L position. However, it is not thought
necessary to restrict the boundary layer definition further as this may impact
the resulting data detrimentally. Specifically, the difference in distribution of
velocities within the boundary layers of different speeds will reduce as one
approaches the solid boundary, where the flow is stationary with respect to
the body.

116
 Figure 5.4. Predicted boundary layer thickness at different scales.

As asserted earlier, all subplots within Figure 5.4 confirm that at full-scale the
boundary layer is thinner. However, the reduction in thickness at the aft
perpendicular is seen as the largest. The well-known keel vortex is prevented
from forming in the present case study likely due to the proximity of the
seabed. This causes the flow to accelerate as the water is passes beneath the
ship. Nevertheless, amidships in model scale, connotations of an increasing
boundary layer thickness are observed. This phenomenon is predicted by both
the free surface and double body method in λ=75. However, the full-scale
results exhibit an even weaker vortex (Zeng et al., 2019), this specific feature
being hardly discernible in both multiphase and double body simulations for
λ=1. It should be noted that in their recent work, Song et al. (2019) obtained
similar results in terms of boundary layer thickness variations.

In terms of viscous scaling, it is evident that the method performs adequately.

To elaborate, the boundary layer seems to follow the full-scale prediction
closely. It is also important to note that in model-scale, the free-surface effect
is visible at the forward perpendicular, amidships and at the aft perpendicular,
where the boundary layer broadens as it approaches z/L=0. The same locations

117
are characterised by the absence of the viscously scaled method’s boundary
layer, in agreement with the full-scale data.

The fact that the viscously scaled predictions model similar behaviour near the
free surface is encouraging. However, there is an apparent difference between
the frictional resistance coefficients predicted by this method and the linearly
scaled simulations. This may stem from a difference in the wetted area, which
has been assumed constant (in non-dimensional form) throughout all cases.
To further elucidate the potential influence of such an effect, Figure 5.5
contains the free surface elevations for the largest and smallest scale factors on
the ship hull. The intermediate scale has been omitted to allow a clearer
depiction of the generated results.

Figure 5.5. Comparison of the wave elevation on the ship hull.

For consistency, all dimensions have been normalised by the ship length in
Figure 5.5. Here, the result label with a “vs” subscript indicates the viscously
scaled result in full-scale. Figure 5.5 shows that better agreement with the full-
scale result is achieved near the stern of the ship via the viscously scaled
model, rather than λ=75. Therefore, viscous effects are of lesser consequence in
full-scale. This may be deduced by considering the fact that the viscously
scaled simulation features a value of viscosity, that is significantly lower than
one would normally observe (Table 5.1 may be consulted for the values).

To provide supporting evidence for the observed phenomena, the reader is

directed to the work of Brard (1970) and Tatinclaux (1970), who demonstrated
that the action of viscosity, vorticity and turbulence are expected to have an
impact on the flow properties. The abovementioned authors studied the effects
of viscous, vortical flows on the wave resistance of a ship. Their findings
include that a viscous contribution may be identified as part of the wave
resistance of a ship.

118
At this stage, it is worthwhile exploring the reason why potential flow theories
do not account for wave elevation changes and the related consequences.
According to Brard (1970), vortical and turbulent effects act on the ship in a
manner proportional to 1/(𝑅𝑒 × 𝐹ℎ2 )1/3 in terms of local waves and
1⁄(𝑅𝑒 × 𝐹ℎ4 ) in terms of far field waves. Unfortunately, the analysis presented
in Brard (1970) is for deep, unrestricted waters. The relative magnitude of the
aforementioned terms is shown graphically in Figure 5.6. However, one may
reasonably expect the above effects to be of greater significance in restricted
shallow waters. Thus, no logical contradiction is expected when carrying the
above relationships to the present analysis. Since the depth Froude number
has been maintained constant, it is not though necessary to examine the
specific relationships as a function of this particular parameter. Instead, Figure
5.6 depicts the relative contribution of each wave component (far-field and
near-field) with increasing Reynolds number.

Figure 5.6. Effect of viscosity and turbulence on near and far field waves
according to Brard (1970).

Figure 5.6 suggests the assertion that viscous effects are of lesser consequence
at full-scale in Figure 5.5 is justified according to the mathematical analysis of
Brard (1970). Although the present case studies are restricted to a single speed

119
for which no far field waves are present, it is worthwhile to comment on their
potential effect. If a ship propagates at a speed where far field waves are
generated, regardless of the water depth and/or restriction, the effect in the far
field waves is expected to be greater than that in the near field disturbance.
This follows because although the region where viscous effects dominate has
become smaller, it has not completely disappeared. Thus, significant
proportions of the near field disturbance will be generated and will lie within
this region. Conversely, the far field waves will be impacted by a smaller wake,
as demonstrated in Figure 5.4.

As a certain Reynolds number (~107) is passed, the relative difference between

the model and full-scale waves decreases rapidly. This may be confirmed by
examination of Figure 5.6, where it is apparent that the slope of the far field
effect is nearly zero for Reynolds numbers past 108. A small effect may be
expected because the majority of changes in the wake occur in the region
Reynolds numbers in the region of 106 – 107. Coincidently, this is the region
where all model tests are performed due to size limitations. Thus, it may be an
inescapable fact that such effects cannot be negated completely by adopting a
model with greater linear dimensions. This is also augmented by the fact that
as one enters the lower range of Reynolds numbers, both curves increase in
magnitude rapidly.

The effects demonstrated in Figure 5.6 are typically omitted from potential
flow theories, even when a nonlinear vortical flow is sought. This is the case
because of the small relative magnitude both the near and far field
disturbances exhibit, as well as their nonlinear nature. Therefore, an analysis
where terms to, say, second order are sought would justifiably not take these
terms into account (Brard, 1970; Tatinclaux, 1970).

The next step is to examine the predicted wave resistance. This is shown in
Figure 5.7, using the aforementioned methods (viscous and viscous scaling).
The viscously scaled wave resistance coefficient is estimated by subtracting
the total resistance as scaled (viscously) and the double body resistance at the
specific scale factor.

Figure 5.7 clearly shows that the methods agree in terms of trend – an overall
reduction as the scale factor approaches full-scale is observed. Not
surprisingly, the predictions follow a pattern closely resembling that of Figure
5.6, characterised by a sharp decline in the low Reynolds number range,
followed by a mild slope. As demonstrated previously, the smallest linear

120
dimensions coincide with those where the viscous effect is expected to be
highest. Therefore, the difference between the wave resistance observed at the
two adjacent model scale factors is justified. The source of the persistent
discrepancy between the methods is likely related to the assumptions in terms
of viscosity and double body approximation, although one would expect this
to decline further if the Reynolds number were to be increased.

Figure 5.7. Predicted wave resistance coefficients.

Prior to providing further justification of the results shown thus far, the final
set of data is presented. Specifically, Figure 5.8 depicts the predicted form
factors. The overall trend observed in the figure is that of reduction in (1+k)
with higher Reynolds numbers. The best-behaved curve is that calculated via
the multiphase method. Indeed, the experimental work of Elsherbiny et al.
(2019) suggested the form factor should be in the region of 1.16. The double
body prediction seems to resemble this to a lesser extent, and as the scale factor
is increased, the data do not decrease monotonically as is the case with the
multiphase results.

However, it is not possible to asses scale effects in (1+k) in the absence of

experimental data for each scale factor. Moreover, the multiphase method is
not characterised by an increase for λ=37.5. This points to the fact that the

121
double body simulation at λ=37.5 may be inaccurate rather than the
multiphase one. The change in form factor may be justified by referring to the
recent work of Zeng et al. (2019). In the aforementioned work, the authors
derived a similar shape for (1+k). The authors also defined a new relationship
for the form factor of the KCS sailing in shallow waters, which is employed in
Figure 5.8.

Figure 5.8. Predicted form factors.

The relationship defining the solid line in Figure 5.8 due to Zeng et al. (2019)
does not take into account the lateral confinement, which is the suspected
cause of the observed difference. However, the agreement between the
multiphase data and the approximation of Zeng et al. (2019) is seen to be good.
Thus, the results confirm the efficacy of the method determined in the
previously mentioned reference. Unfortunately, the curve fitting approach
used to derive the relationship must be performed anew for each ship. This is
because no method is known to determine such an equation for any ship
without the presence of data to fit. To prove that the lateral confinement’s
effect is not as significant as the depth restriction, the velocity distribution
along a line in the x –z plane at the aft perpendicular is used. An example of
this for [x/L, z/L]=[0, 0.5T] is shown in Figure 5.9, where the velocity has been

122
normalised by the free-stream velocity. In the present context, this is defined
as the flow velocity specified at the inlet. The specific location is chosen in line
with the significant difference in boundary layer thickness observed in Figure
5.4.

Figure 5.9. Velocity distribution along a line at the aft perpendicular, z=0.5T.
Depicted: multiphase simulations, λ=75 and λ=1.

It is apparent in Figure 5.9 that the flow is accelerated in a different manner in

the two scale factors. More importantly, the flow speed achieved near the tank
wall (which is set as a slip wall and therefore does not impact the flow velocity
as a no-slip wall would) is higher than the free-stream velocity. In λ=75, the
flow velocity is 2.6% higher, whereas in λ=1 – 1.9% higher than that specified
at the inlet. In other words, a net difference of 0.7%. Although one may argue
that this is not a significant difference, its impact is nonetheless of some, albeit
small importance to the ship, particularly on the frictional resistance. The
manner in which the different scale factors achieve their maximum flow
speeds near the wall is different from one another. Thus, the results from can
be used to signify that in full-scale, side wall effects are of (slightly) smaller
influence than in model scale factors.

123
Importantly, Figure 5.9 suggests that flow properties do not scale linearly in
highly restricted waterways. Had this been the case, no difference would be
present in the curves shown in Figure 5.9. Therefore, forming a geosim series
in such conditions is not as straightforward as simply scaling the tank
dimensions. In practice, tanks equipped with false bottoms could be used
effectively in this respect. However, if one is to accept the results associated
with Figure 5.9, then the tank dimensions should scale non-linearly. The
manner in which this should occur is not known at present, but correction
methods similar to Raven (2019) could be considered as a starting point. The
issue with such corrections is that they are inherently designed to remove side
wall effects, whereas one might wish to maintain this influence when
designing, say, a canal boat, or river cruise ship.

The results presented in Figure 5.9 should also be considered in conjunction

with the boundary layer thickness assessment carried out in Figure 5.4. These
jointly suggest that the influence of the water depth scales non-linearly as well
as the width. Thus, rendering the possibility of forming a geosim series in
shallow water of infinite width equally complex. In fact, the greater proximity
of the seabed amplifies the influence of the boundary. Effects of this kind were
used as a justification for at the onset. It could be considered that these have
been proven to a sufficient extent.

The collapsing difference between an infinitely wide water case-study, where

the velocity ratio would reach unity, and the canal case explains the
discrepancies observed between the present CFD method and the data of Zeng
et al. (2019). Specifically, in the low Reynolds number range, the relative
difference between the predicted (1+k) and the method of Zeng et al. (2019) is
larger than at high Reynolds numbers. This observation fits neatly with the
presented data.

5.4 Summary and conclusion

This chapter presented a numerical assessment of scale effects of a ship

advancing through a canal. To assess the scale effects, a geosim series was
formed and evaluated at three different scale factors. The numerical methods
used comprise RANS-based multiphase and doubly body simulations. These
enabled the assessment of the form factor and wave resistance.

Comparison with recently developed equations describing the frictional

resistance revealed excellent agreement with the present CFD set-up in model-
scale. For high Reynolds numbers, the large slope of the curve terminated at

124
too low values according to the present CFD method. The aforementioned
equations were developed for an infinitely wide shallow water case. Results
quantified the influence of the particular canal as small. The flow being
accelerated by less than 3% in locations near the wall as a result of the reduced
clearance in model scale.

Scale effects on the accelerated velocity were demonstrated in the case of flow
near the ship and canal walls. This amounted to 0.7% difference between
model and full-scale. In terms of boundary layer, the CFD set-up captured the
well-known decrease in thickness. A measurement of the velocity profiles at
the aft perpendicular also suggested that the wake volume as a fraction of the
ship’s displacement is also significantly reduced from model to full-scale.

The predicted form factor showed good agreement with recently established
relationships for the KCS. This parameter, along with the wave resistance
exhibits a monotonic decline until full-scale. This was confirmed by invoking
well-known mathematical analysis which suggests the influence of vorticity
and turbulence on the ship decay rapidly with increasing Reynolds number.

125
 A POSTERIORI ERROR AND
UNCERTAINTY ESTIMATION IN
COMPUTATIONAL SHIP
HYDRODYNAMICS

This chapter presents the application of local and global numerical uncertainty
estimators and provides a framework to interpret solution data and what it
suggests with regards the computational grid’s performance in view of
demonstrating a small numerical error. The analysis is performed on a set of
full-scale simulation because high Reynolds number flows are difficult and
expensive to measure experimentally. Therefore, it is frequently the case that
one lacks validation data for full-scale flows. Such conditions are where
numerical error estimation has the highest value in simulation-based design.

6.1 Introduction

The year 2014 saw the publication of NASA’s study on the future directions of
Computational Fluid Dynamics (CFD) (Slotnick et al., 2014). Although their
focus was predominantly on aerospace applications, several aspects overlap
significantly with the field of computational ship hydrodynamics. This
prompted Hawkes et al. (2018) to address one of the concerns raised by
Slotnick et al. (2014), specifically, scalability problems of CFD simulations with
increasing cell numbers.

A second key bottleneck, identified by Slotnick et al. (2014), is associated with

solution uncertainty and robustness. This is directly related to the confidence

126
levels one can attribute to a numerically derived solution. Such metrics are of
critical importance in simulation-based design. Since computational power
increases exponentially, it is reasonable to anticipate that Reynolds averaged,
or other Navier-Stokes-based techniques will eventually become the norm in
most forms of engineering analysis. Ship hydrodynamics is by no means an
exception to this statement.

One issue that remains challenging is that of grid generation. At present,

constructing a mesh, onto which the solution of the governing Partial
Differential Equations (PDEs) is to be obtained is as much an art as it is a
science. This is true especially in cases where the requirements in terms of
accuracy are high, as is often the case in ship hydrodynamics. Specifically, with
ever tightening EEDI-related (Energy Efficiency Design Index) regulations,
margins of error and uncertainty are slim. It is therefore imperative that once
the practitioner has invested in performing what is currently the state-of-the-
art analysis method in ship hydrodynamics, he/she can obtain an adequate
margin of error associated with the solution. In other words, it is critical to
determine the confidence one can place in the numerical solution, and
discount specific results if necessary.

Lack of robust procedures to address confidence levels in numerical data that

do not require user intervention (i.e. are automatically generated) is one of the
key issues hindering in the large scale adoption of simulation-based design
(Slotnick et al., 2014). This is particularly true of ship hydrodynamics which
inherently features turbulent flows and fluid interfaces. These require
knowledge of the simulation-specific phenomena, such as wakes and free
surface deformations, where the computational mesh must be refined to
capture the underlying physics well. Therefore, a key issue is related to mesh
definition, which is almost always the greatest source of error (Eca and
Hoekstra, 2006). Although grid sensitivity studies are frequently performed in
an academic context, their outcomes are not always understood according to
Salas (2006). In industrial applications on the other hand, the analyst may
simply consider a single mesh due to the perception that error estimation is
difficult and time consuming (Freitas, 2002).

This chapter will describe in detail, and demonstrate the use of a method
developed by Cadafalch et al. (2002), and supplemented by Phillips and Roy
(2017). This method utilises well-known procedures to predict the numerical
error in the domain, simultaneously equipping the analyst with a measure of
how well the grid has performed in reducing the numerical error. Full-scale

127
simulations are sought in line with the fact that high Reynolds number flows
are difficult to experimentally measure, and are where numerical work has the
highest potential impact in simulation-based design (Stern et al., 2001).
Specifically, the well-known KCS was simulated in shallow water, replicating
the experiments of Elsherbiny et al. (2019) in a rectangular canal at full-scale.
Use is made of the commercially available RANS solver, STAR-CCM+, version
13.06.011 to perform the present analysis.

The remainder of this chapter will continue by introducing the necessary

background in section 6.2, followed by a description of the case study in
section 6.3. The numerical set-up, and convergence properties of the solution
are examined in Section 6.4. Finally, concluding remarks are given in section
6.5.

6.2 Background

The accuracy of CFD methods is of great interest in academia for several

reasons. When compared to analytical methods, computational approaches
have a greater breadth of application and are capable of addressing
significantly more complex physical phenomena (Oberkampf and Blottner,
1998). The main advantage of analytical methods is expressed in the fact that
they are built on sound, reproducible and traceable mathematical arguments.
However, this is simultaneously their major drawback. Many processes of
immense practical importance, such as the turbulent motion of a fluid, cannot
be described analytically (Durbin and Pettersson Reif, 2011). If this were
possible, turbulence closures would be exact, and the field of turbulence
modelling would not exist in its current form.

On the other hand, by avoiding the issues relating to analytical modelling of

physical phenomena, the CFD method runs into a separate set of problems.
These mainly relate to the equivalence between the continuum form of the
governing PDEs and their discrete approximations. Lilek and Perić (1995)
separate the errors into three categories:

1. Modelling errors, which can be thought of as the difference between

the exact solution of the equations describing the fluid flow (i.e.
satisfying the conservative laws), and the actual flow. For laminar
flows, the Navier-Stokes equations are sufficiently accurate, but to
account for cases where turbulence is important, additional models are
required (Jasak, 1996). These errors are separate from the numerical
errors discussed here (Oberkampf and Blottner, 1998).

128
 2. Iterative errors, which arise as a result of the nonlinearity of the
governing equations. The iterative fashion in which these are solved
introduces errors, also known as iteration convergence errors (Eca et al.,
2013).
3. Discretisation errors, which stem from the mapping of the continuum
PDE and their related auxiliary models (for example, the turbulence
model) into algebraic equations. Formally, this ‘replacement’ of
equations rests upon Lax’s Equivalence Theorem (LET) (Lax and
Richtmyer, 1956). In essence, LET contains two statements regarding
the discrete approximation of the PDEs. Firstly, the numerical
procedure must be consistent, i.e. as the discretisation length (or mesh
length – h) approaches 0, the error must vanish. Secondly, the numerical
method must be stable, i.e. if the one were to allow the solver to run
indefinitely, the solution must remain bounded (Morton and Mayers,
2005). It should be noted that the LET is valid for linear initial value
problems. As such, it is applied to the linearised form of the mapped
governing equations.

Round-off errors are sometimes included in the above list. They can be traced
to the finite digit storage on digital computers, but their influence is thought
negligible (Roy, 2005). For this reason, their influence will not be discussed
further. As mentioned above, modelling errors are not the primary focus and
are therefore discounted. Iterative errors are thought of as a simpler problem.
In any case, their approximation is important because numerical error and
uncertainty estimators have a tendency of magnifying solutions with large
iterative errors, or incomplete iterative convergence (Larsson et al., 2014). For
the adopted case-studies, these have been estimated in the region of 10-4% by
using the procedure of Roy and Blottner (2006). Instead, the primary goal of is
to apply to RANS-based ship hydrodynamics the discretisation error
estimation procedures of local quantities and demonstrate their use. This is
because the discretisation error category can account for more than 90% of all
error (Xu et al., 2019). The use and application of discretisation error estimators
is discussed in detail in the following sub-section.

6.2.1 Discretisation error

As stated previously, discretisation errors stem from the mapping of the

continuum PDEs onto discrete locations. This is done by splitting the solution
domain into a finite number of solution nodes. There are two methods to

129
combat discretisation errors. The usual approach is to increase the density of
space or time intervals, which the region of interest is subdivided into
(although spatial discretisation is of interest here). This is typically referred to
as spatial or temporal refinement, but cannot guarantee the solution will be
improved per se. The above is exemplified in the findings of Larsson et al.
(2014). In the aforementioned work on numerical hydrodynamics, solutions
obtained with less than one million cells were compared solutions produced
by 20+ million grids from different participants. The graphical summary
provided in Larsson et al. (2014) indicates that participants with low cell
numbers often perform better than those with large cell counts. In other words,
the arrangement, or properties of the grid are equally important in reducing
errors (Salas, 2006). This is an issue that will be examined in the present work.

The alternative is to shift the focus from the subdivisions, explained above,
towards equation discretisation. In the Finite Volume Method (FVM), which
is used by Star-CCM+, the equation discretisation employed is 2nd order. A
second order discretisation approach is typical of RANS solvers, and implies
that according to Lax’s equivalence theorem, the error must reduce with the
square of the grid size for asymptotic grids (Roy, 2005). In other words, the
formal order of accuracy of the simulation is two (𝑝𝑓 =2) (Roache and Knupp,
1993). Therefore, the second method to reduce the discretisation error would
be to increase the formal order of accuracy, such as the 4th order accurate
method, devised by Lilek and Perić (1995). In practice, the choices in this
respect are limited to 𝑝𝑓 =2. That is, unless the user is re-coding the entire RANS
solver. Consequently, the focus of this chapter will be on discretisation errors,
stemming from grid density. The first method devised for this purpose is the
Grid Convergence Index (GCI), devised by (Roache, 1998).

The GCI method predicts an uncertainty, which is used to bracket the

numerically calculated solution. The true solution is then expected to lie within
this bracket 95 out of 100 times, i.e. with 95% confidence (Roache, 1997). The
GCI method is based on Richardson's (1911) work, who devised a method to
estimate a solution of fourth order accuracy based on second order finite
differencing. Later, the approach, widely known as Richardson Extrapolation
(RE), was improved upon by Richardson (1927). RE was originally formulated
to use a refinement ratio (r) of two (𝑟 = 2), which translates into grid doubling.
However, it can be used with any factor larger than 1.1. That is, as long as the
produced solutions are both sufficiently different, and asymptotic, to enable

130
the validity of the method. This is in view of avoiding interference of other
numerical errors (Roache, 1998).

The error of a numerical solution can be defined via a Taylor series expansion
in the form shown in Eq. (6.1):
∞
𝜀 = 𝑓𝑒𝑥 − 𝑓1 = ∑ 𝛼ℎ𝑝 = 𝛼𝑝𝑓 ℎ𝑝𝑓 + 𝐻𝑖𝑔ℎ𝑒𝑟 𝑜𝑟𝑑𝑒𝑟 𝑡𝑒𝑟𝑚𝑠 (6.1)
𝑝=𝑝𝑓

where 𝑓𝑒𝑥 is the exact solution, 𝑓1 is the solution of obtained on a grid with
characteristic size h, and α is a constant. In the present context, the error is
defined as the difference between the exact solution and the solution obtained
with a grid spacing h. In an attempt to reasonably approximate the error,
higher order terms may be neglected. Thus, reducing the form of Eq. (6.1) into:

𝜀 ≈ 𝛼𝑝 𝑓 ℎ 𝑝 𝑓 (6.2)

The omission of higher order terms requires that their combined effect is much
smaller than the error derived at the formal order of accuracy. The
implications of this are that the grid must be asymptotic.

The reason why Richardson Extrapolation is not used per se is that it provides
a 50% confidence level, whereas GCI boasts of 95%, as stated earlier. This is
achieved by magnifying the predicted numerical error by a Factor of Safety
(FS). The value assigned to FS is clearly of high importance. In the GCI method,
this is set as 3 when two grids are available, and 1.25 when three grids are used
(Phillips, 2012).

Several methods have emerged in the recent past, many based on the GCI that
attempt to modify the prediction of the FS. Some have made the FS a function
of the observed order of accuracy, shown in Eq. (6.3) (Celik et al., 2008):
𝑓 −𝑓
𝑝 = ln (𝑓3−𝑓2)⁄ln (𝑟) (6.3)
2 1

where r is the refinement ratio.

The fundamental notion is that if the observed and formal orders of accuracy
are close, the grids are asymptotic. Otherwise, the GCI uncertainty diminishes
rapidly with growing p, rendering unrealistically low numerical errors.
Roache (1998) limited the observed order of accuracy to prevent this from
occurring. Clearly, Eq. (6.3) can only be used when the argument of the natural
logarithm in the numerator is positive and larger than 1. In other words, when
the solutions exhibit monotonic convergence.

131
Unfortunately, the exact solution (𝑓𝑒𝑥𝑡 ) is known rarely, if ever, in fluid flows
of practical interest. Using the order of convergence, one may re-write the
error as a function of p:

𝜀(𝑝) = (𝑓2 − 𝑓1 )⁄(𝑟 𝑝 − 1) (6.4)

The uncertainty, (U) is then simply the absolute error, as a function of the order
of convergence, multiplied by the Factor of Safety:

𝑈𝐺𝐶𝐼 = 𝐹𝑆 × |𝜀(𝑝)|/(𝑟 𝑝 − 1) (6.5)

Clearly, the Factor of Safety is characterised by a step change (1.25 to 3) within

the framework of the GCI method. Stern et al. (2001) implemented a correction
factor (CF) approach to resolve this. In essence, their approach consists of
introducing a correction factor, used to account for higher order effects. This
is used to magnify the error predicted via Eq. (6.4). In the above work, the
correction factor is introduced as a metric describing the distance from the
asymptotic range, where Eq. (6.4) is thought to be inadequate. Thus, the error
according to Stern et al. (2001) takes the form of:

𝜀𝐶𝐹 (𝑝) = 𝐶 × (𝑓2 − 𝑓1 )⁄(𝑟 𝑝 − 1) (6.6)

where C is the correction factor, predicted as shown in Eq. (6.7):

𝐶 = (𝑟 𝑝 − 1)⁄(𝑟 𝑝𝑓 − 1) (6.7)

Finally, the uncertainty can be estimated by:

[9.6(1 − 𝐶)2 + 1.1] × 𝜀𝐶𝐹 𝑤ℎ𝑒𝑛 |1 − 𝐶| < 0.125
𝑈𝐶𝐹 = { (6.8)
[2|1 − 𝐶| + 1] × 𝜀𝐶𝐹 𝑤ℎ𝑒𝑛 |1 − 𝐶| ≥ 0.125

Alternatively, the Factor for Safety approach of Xing and Stern (2010) uses the
normalised order of convergence (P), defined as the ratio of the observed and
theoretical order of convergence (𝑃 = 𝑝/𝑝𝑓 ). Based on the value, attained by P,
the uncertainty can be estimated as shown in Eq. (6.9):
(16.4𝑃 − 14.8) × |𝜀(𝑝)| 𝑤ℎ𝑒𝑛 𝑃 > 1
𝑈𝑅𝐸𝐹𝑆 = { (6.9)
(2.45 − 0.85𝑃) × |𝜀(𝑝)| 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

All methods described thus far rely, in one form or another, on the observed
order of accuracy. However, approximating an error does not require any
knowledge of p. Indeed, p is little more than the power, to which the grid (h)
is raised to in the dominant term of the Taylor series expansion, as shown in
Eq. (6.1). Following such an approach, Roy (2008) split the problem into first
and second order components. Thus, the fine solution (𝑓1 ) can be expressed in

132
terms of the extrapolated solution and a set of error terms as demonstrated in
Eq. (6.10):

𝑓1 = 𝑓𝑒𝑥𝑎𝑐𝑡 + 𝑔1 ℎ1 + 𝑔2 ℎ12 + 𝐻𝑖𝑔ℎ𝑒𝑟 𝑜𝑟𝑑𝑒𝑟 𝑡𝑒𝑟𝑚𝑠 (6.10)

In this approach, the constants 𝑔1 and 𝑔2 are defined as shown in Eq. (6.11)
and Eq. (6.12), respectively:

𝑔1 = (𝑟 2 𝜀21 − 𝑟𝜀32 )/(𝑟 × (𝑟 − 1)2 ) (6.11)

𝑔2 = (𝜀32 − 𝑟𝜀21 )/(𝑟 × (𝑟 + 1)(𝑟 − 1)2 ) (6.12)

where 𝜀21 = 𝑓2 − 𝑓1 , and 𝜀32 = 𝑓3 − 𝑓2 . Here, it is also useful to mention the

convergence ratio (R), defined as the ratio 𝜀32 /𝜀21 , which is also the inverse of
the argument in the numerator of Eq. (3). Armed with these parameters, the
extrapolated solution may be estimated by:

𝑓𝑒𝑥𝑎𝑐𝑡 = 𝑓1 + (𝜀32 − (𝑟 2 + 𝑟 − 1) × 𝜀21 )⁄[(𝑟 + 1)(𝑟 − 1)2 ] (6.13)

Roy (2008) defined the above procedure solely as an error estimator and did
not provide a Factor of Safety, which can be used to predict an uncertainty.
Therefore, one can make no claim with regards to the confidence level the
method provides. To establish the range of applicability of the method, it must
be applied to cases where exact solutions are known. This would allow the
introduction of a Factor of Safety, which adequately accounts for the
conservatism of the scheme.

The reason why this method is included in the present assessment is twofold.
Firstly, the behaviour of the error can be examined with grid refinement,
separating linear and quadratic terms. This permits one to determine whether
the asymptotic range is approached from above, below, or from different
directions by each term in the expansion (Eq. 6.10). The implications of such
an analysis are important, because one may gauge whether the solution
exhibits linear or quadratic convergence depending on which constant (𝑔1 , 𝑔2 )
dominates in each range (coarse, medium, fine grids). The extracted
information can then be used to assess whether further mesh refinement is
warranted. This decision would rest on the approximated rate at which the
error reduces. In other words, one may attempt to justify using a specific grid
based on its convergence properties with greater ease than is possible with
other methods.

For example, if a particular parameter exhibits second order convergence, the

CFD practitioner may choose to refine the grid further. This could be a good

133
choice since even a small reduction in the grid size can be expected to lead to
a palpable reduction in the predicted error. Such tools are especially powerful
when considering cases where experimental data are not available.

In second place, the spatial error, defined in Eq. (6.13), has third order
accuracy, which is independent of the observed order of accuracy (Roy, 2008).
Therefore, the method does not rely on the positivity of 1/R, or the related
condition R< 1, in Eq. (6.3) and will produce an error estimate regardless of
the parameters fed into Eqs. (6.10-6.13). In practice, provided one has a greater
number of solutions, a system of equations can be constructed in the form of
Eq. (6.10). The first n terms would allow the calculation of the n-1st constant
(𝑔𝑛−1 ). Thus, when using three solutions, as is the case in this for the present
set of simulations, the linear and quadratic terms (1st and 2nd order) are
computed. Since the theoretical order of accuracy of the solver is 𝑝𝑓 = 2, there
is little merit in seeking solutions of order 3 or higher.

It should be noted that a disagreement between the observed and theoretical

order of accuracy does not necessarily imply inconsistency (Thomas and
Langley, 2008). This is the case because Richardson Extrapolation was devised
for structured grids. In fact, Diskin and Thomas (2010) point out that in
unstructured grids, an order of accuracy near unity, or higher than two is
frequently observed. They further state that this does not contradict the Lax
and Richtmyer (1956) equivalency theorem, because the LET requirements are
sufficient, but not strictly necessary to demonstrate the validity of a numerical
scheme.

6.2.2 Local error and uncertainty

Having described the error and uncertainty estimation techniques, it is

prudent to put into context the contribution of this chapter to the wider
literature. Here, the local error and uncertainty estimator, originally devised
by Cadafalch et al. (2002) is introduced. They began by defining Richardson
and oscillatory nodes as follows:

• Richardson nodes: (𝑓3 − 𝑓2 ) × (𝑓2 − 𝑓1 ) > 0

• Oscillatory nodes: (𝑓3 − 𝑓2 ) × (𝑓2 − 𝑓1 ) < 0

Additionally, converged nodes can be computed, where the above product lies
below some predefined measure of accuracy. For the purposes of this chapter,
no such limit is defined. Therefore, it is thought prudent to proceed along the

134
lines of Phillips and Roy (2012) and discount converged nodes as a possibility
in the numerical work.

Cadafalch et al. (2002) then ensue to use the local observed order of accuracy
to produce a global average. Following a similar rationale, Phillips and Roy
(2017) provide an analogous framework, which is employed here. The first
step is a small modification to Eq. (6.3), shown in Eq. (6.14). Specifically, the
absolute value of the quotient in the numerator is taken. This ensures that
oscillatory nodes are also taken into account. Such a change has been proposed
previously in, for example, Celik and Karatekin (1997) and Celik et al. (2008).
In other words, the change breaks no precedents.
𝑓 −𝑓
𝑝̂ = ln (|𝑓3−𝑓2|)⁄ln (𝑟) (6.14)
2 1

The resulting order of accuracy at each location of interest is not used directly.
Instead, Phillips and Roy (2017) define the global deviation from the formal
order of accuracy (𝑝𝑓 = 2), shown in Eq. (6.15):
𝑁
1
𝛥𝑝 = min [𝑁 ∑ min(|𝑝𝑓 − 𝑝̂𝑖 |, 4𝑝𝑓 ) , 0.95𝑝𝑓 ] (6.15)
𝑖=1

Eq. (6.15) can be interpreted as the mean local deviation of the observed order
of accuracy from the theoretical order of accuracy. The maximum deviation is
restricted to 4𝑝𝑓 in an attempt to avoid skewing the average (Phillips and Roy,
2017). In the process of derivation of 𝛥𝑝, Phillips and Roy (2017) considered
multiplicative factors of 𝑝𝑓 of 2, 4, 6, and 8, but the choice for this particular
parameter was shown to be of little consequence. To prevent values close to
zero, the maximum distance of the observed order of accuracy from the formal
order is limited to 95% of 𝑝𝑓 . Having obtained 𝛥𝑝, one can progress to
calculating the global distance from the formal order of accuracy (𝑝∗ ), shown
in Eq. (6.16). This enables the calculation of the Factor of Safety (FS), as a
function of 𝑝∗ , shown in Eq. (6.17).

𝑝∗ = 𝑝𝑓 − 𝛥𝑝 (6.16)
8
𝑝∗
𝐹𝑆(𝑝∗ ) = [𝐹0 − (𝐹0 − 𝐹1 ) (𝑝 ) ] (6.17)
𝑓

where 𝐹0 =3, and 𝐹1 =1.1. Eq. (6.17) can be used to construct a smoothly varying
FS with distance from the asymptotic range.

Finally, the uncertainty can be estimated as shown in Eq. (6.18):

135
 𝑓 −𝑓
𝑈𝑛𝑐𝑒𝑟𝑡𝑎𝑖𝑛𝑡𝑦 = 𝐹𝑆(𝑝∗ ) |𝑟2𝑝∗ −11 | (6.18)

In the process of any RE-based method, a solution is extrapolated. This can be

thought of as the solution, which one may expect to obtain on a grid with a
cell size of 0. In other words, it approximates the analytical solution. This is
done by virtue of three systematically refined or coarsened grids. Emphasis is
placed on solutions achieved with uniform refinement ratios, although it is
possible to vary this particular parameter. For the relevant background on this,
the reader is referred to ITTC (2002) and Roache (1997). The use of an
extrapolated solution as the final outcome of a simulation is inadmissible
because it is not possible to prove that such a solution satisfies the conservative
laws (Roy, 2008).

The methods presented thus far can be classified as a posteriori, because they
allow the estimation of uncertainty after the simulation has run its course.
However, there are a priori methods, which are, in theory, capable of
performing the same function as soon as the grid has been generated. Since no
such method is available in the RANS solver used, this course of action has not
been explored further. Alternatives also include intrusive methods. For
example, an error transport equation may be produced, which generates an
estimate of the numerical accuracy while the simulation is running. In
recognition of the fact that in practice, the average RANS practitioner will not
be able to achieve this, such methods have not been incorporated. A full
classification and discussion on the above methods can be found in Jasak
(1996), Phillips (2014, 2012), Roache (1997) and Roy (2005).

6.3 Case-studies and numerical set-up

For the purposes of this chapter, a set of full-scale numerical simulations are
used. These form a full-scale replica of the experimentally investigated case-
studies of Elsherbiny et al. (2019). The refinement ratio 21/3 was selected to
avoid a sudden explosion in cell numbers. In other words, making this
selection allowed the adopted simulations to have a greater number of cells
while avoiding a sudden increase in cell numbers beyond which the
simulations would become difficult to handle. The present simulations feature
cell numbers ranging from approximately 10 million to 32 million. As
explained previously, these cell numbers are achieved by a systematic,
consecutive spatial refinement.

136
 6.3.1 Spatial characteristics of the computational domain

The dimensions of the computational domain satisfy two requirements.

Firstly, the width and water depth are set to represent the experimental
campaign of Elsherbiny et al. (2019), as shown in Figure 6.1. To reduce the
required cell numbers, a symmetry condition is imposed, coincident with the
ship and canal centreline. The width of the domain was originally 2.3m in
λ=75, whereas the depth (h) was 2.2 times the ship draught (ℎ⁄𝑇 = 2.2).
Naturally, these conditions are maintained to reproduce the experiment as
closely as possible in full-scale. To accomplish this, the side boundary is set as
a slip wall, while the domain bottom is prescribed as a velocity inlet.

Figure 6.1. Domain dimensions and boundary conditions

The remaining boundaries are placed following widely used

recommendations regarding resistance predictions in ship CFD. Specifically,
the inlet boundary is positioned 1.5 ship lengths upstream of the forward
perpendicular, whereas the outlet is placed 2.5 ship lengths downstream of the
aft perpendicular (ITTC, 2014; Siemens, 2018). The domain top is set as an inlet,
positioned 1.25 ship lengths from the undisturbed waterline. For reference, the
ship and domain dimensions are shown in Table 6.1.

All simulations were performed in the commercially available RANS solver,

Star-CCM+, version 13.06.011. As mentioned previously, the software uses the
Finite Volume Method (FVM) to discretise the governing PDEs onto a finite
number of predominantly hexahedral cells with minimal cell skewness. The
latter is defined as the angle between the face area vector (face normal) and
the vector connecting the centroids of two adjacent cells. The reason why this
parameter is important lies in the fact that diffusion terms contain the dot
product of the two vectors referred to above (Siemens, 2018). Therefore, when
these are perpendicular (i.e. the mesh is non-orthogonal), the diffusive terms

137
are seemingly divided by zero, rendering severe convergence problems. The
mesh must be as orthogonal as possible to avoid such a scenario.

Table 6.1. KCS and computational domain principal characteristics in full-

scale
Quantity Symbol Value Unit
Scale Factor λ 1 -
Length L 230 m
Beam B 32.2 m
Draught T 10.8 m
Depth D 19 m
Water depth h 23.760 m
Block coefficient CB 0.651 -
Longitudinal Centre
LCG 111.593 m
of Gravity
Wetted area S 9530 m2
Speed 𝑈 4.630 m/s
Reynolds number Re 1.195×10 9
-
Depth Froude 0.303
𝐹ℎ -
number
Water 2.2
h/T -
depth/Draught
Domain half width w 172.5 m

In the free stream, there is little preventing the mesh from near perfect
orthogonality. However, near the ship surface, the mesh must conform to the
inherent curvature of the ship’s underwater shape. This property,
superimposed on the large aspect ratio cells, required to adequately capture
boundary layer physics can create significant problems. The optimum aspect
ratio of a grid was examined by Salas (2006), who concluded that it is possible
to reduce the numerical error based solely on grid aspect ratio. Meshes with
different aspect ratios are not manufactured, since this is too computationally
expensive, considering the 32 million cells in the finest grid. Instead, the
numerical error, as predicted for the finest solution will be examined against
the aspect ratio of the cell, where the solution has been sampled from.

Two distinct metrics will be assessed. The first is the free surface deformation,
while the second – skin friction, acting on the hull. Since the ship was allowed
to sink and trim via the Dynamic Fluid-Body Interaction module, offered by
Star-CCM+, one cannot take free surface samples from the immediate vicinity

138
of the ship hull. In the case of skin friction, the ship centre of gravity is used as
the coordinate system origin, thus, automatically correcting for any
differences resulting from the ship’s squat. The sinkage and trim, computed
from the three grids are shown in Table 6.2, while the resulting grids are
shown in Figure 6.2. The free surface and skin friction are sampled at a
simulation time of 500s to ensure all transient effects have decayed. Table 6.2
also contains the temporal uncertainty estimates, which where arrived at by
magnifying the time-step by a factor of 20.5 to produce the medium (𝑓2 ) and
coarse (𝑓3 ) solutions.

Table 6.2. Computed integral quantities and related uncertainties

Spatial uncertainty Temporal uncertainty
Parameter Sinkage [m] Trim [°] Sinkage [m] Trim [°]
𝑓1 -0.186 -0.0752 -0.186 -0.0752
𝑓2 -0.186 -0.0753 -0.186 -0.0753
𝑓3 -0.187 -0.0756 -0.187 -0.0754
𝑝 [-] 4.097 1.9643 2.450 0.7896
GCI [%] 0.0220 0.1852 0.42 0.877
Convergence mode Monotonic Monotonic Monotonic Monotonic

Figure 6.2. Resulting mesh

The mesh, depicted in Figure 6.2, was designed to capture the expected free
surface disturbance by concentric volumetric refinements. This is done in the
vicinity of the free surface, as well in the proximity of the ship hull. The top
view of the mesh, shown in Figure 6.2 represents half of the computational
grid. This is constructed within the automatic facilities Star-CCM+. The prism
layer mesh is utilised to manufacture the near-wall mesh, which is
accomplished by generating an exploded sub-surface, using the ship as input.
Within the region between the sub-surface and the ship hull, a fine mesh can
be imposed in order to capture boundary layer physics.

139
 6.3.2 Physics modelling

Typically, the near-wall performance of a simulation is characterised via the

dimensionless distance from the wall, the y+ function. In model-scale
computations, the desirable y+ values are smaller than 1 (y+<1). In full-scale
cases, setting the y+ values below one is a significant challenge, involving very
high aspect ratio cells. These are not a desirable feature in a numerical
simulations, because they destabilise the convergence properties of the
solution by introducing numerical stiffness (Deng et al., 2004; Eca et al., 2015;
ITTC, 2011). It is also important to consider the transition of cell size. As the
flow moves between two neighbouring cells, partial reflections may occur if
there are large changes in the properties of the mesh (Siemens, 2018).
Additionally, mesh coarsening introduces diffusion (Perić and Abdel-
Maksoud, 2016), which is not desirable in locations featuring near-wall
influences. Fortunately, in full-scale applications, flow separation is less likely,
and of less consequence. The question of whether employing wall functions
amounts to correct flow features in the wake of a full-scale ship is yet to be
determined. However, this requires experimental measurements in the wake
of a full-scale ship, as well as the modelling of a spinning propeller, which is
not within the scope of this thesis. The use of wall functions is therefore
assumed justifiable (mean y+≈200 in the present case).

The near-wall characteristics are governed by the distribution of velocity

within the boundary layer. Naturally, this is dependent on the method,
selected to close the Navier-Stokes equations. Such a model is required due to
the fact that in their Reynolds averaged form, the governing equations do not
form a closed set. For the purposes of this thesis, use is made of the standard
k-ω model. This choice is made following the findings made in Chapter 4,
where the k-ω model was demonstrated to perform consistently and reliably.
Moreover, the time per iteration was shown to be the smallest of all two-
equation turbulence models assessed. Coupling these findings with the fact
that the k-ω model can be applied seamlessly for any y+ value without any
major modifications make it a good choice for the present assessment.

The importance of turbulence modelling in full-scale ship hydrodynamics has

been discussed in Deng et al. (2004), who found that the non-linear Reynolds
Stress Transport (RST) model exhibits best performance. However, the RST
model suffers from several problems, not encountered in two-equation
turbulence models. In particular, a greater number of equations is required to

140
predict the Reynolds stress. Thus, the computational time is expected to be
several times that achieved with the standard k – ω model. That is, albeit the
solution of turbulence equations scales better than momentum equations
(Hawkes et al., 2018). Although the RST model boasts of modelling turbulent
physics more robustly than two-equation models, there are problems in its
practical implementation (Parneix et al., 1998). The above suggest a conclusion
along the lines of that made by Eca and Hoekstra (2001). Namely, two-
equation turbulence models remain a good option in full-scale ship
hydrodynamics. Moreover, the discrepancies between turbulence models are
thought to be less significant with an increase in Reynolds number (Eca and
Hoekstra, 2001), although more recent work suggests this may not be the case
(Visonneau, 2005).

To guarantee that convective terms are represented as accurately as possible,

a second order upwind scheme is adopted. This is done in view of the fact that
upwind-biased schemes are the only available variety that can guarantee
boundedness of the solution (Jasak, 1996). Moreover, adopting a higher (3rd
order) scheme can destabilise the solution because it introduces significant
numerical diffusion (Vanka, 1987). The use of a second order accurate method
is in line with the findings of Andrun et al. (2018). The aforementioned authors
recommended the use of at least 2nd order methods, although the results
suggested little impact on free surface modelling. In the above reference, the
authors also recommended the use of a least-squares (LSQ) approach to
discretising the gradients. For all simulations performed here, the hybrid
Gauss-LSQ method is used following the recommendations of software
developer (Siemens, 2018). The definition of such a method is required because
in addition to variable values, variable gradients must be computed at cell
centres. These are used chiefly for the estimation of diffusion and convection
properties.

The segregated flow method is used to solve the integral conservation laws in
a sequential manner. This requires the solution of the three velocity
components and pressure iteratively, one after the other. Velocity and
pressure are coupled via a SIMPLE algorithm. Although the solver can cope
with weakly compressible flows, the incompressible form of the governing
equations is employed.

141
 6.3.3 Time step selection and temporal discretisation

The temporal term in the Navier-Stokes equations is discretised via a first

order method. This is done based the fact that the present class of problems
falls within the ‘pseudo-steady’ category. To elaborate, the physical problem
should not depend on transient terms per se. However, in practice, the RANS
method requires the definition of temporal discretisation. Even though this
may introduce an undue reliance of the computed results onto a metric they
are theoretically independent of (Jasak, 1996). Time is advanced every 15
iterations by 0.0035L/𝑈 in the present simulations (note that this is smaller than
the ITTC (2014) recommendation of 0.01~ 0.05L/V). This choice for the time-
step has been widely used and has been proven to provide sufficiently
accurate results for similar cases (Terziev et al., 2019b, 2018; Tezdogan et al.,
2016a, 2015; Zhang et al., 2018).

The reason why 15 inner iterations are used relates to residual reduction.
Residuals indicate the degree to which the discretised equation is satisfied in
each cell. Star-CCM+ uses the root mean-square value of the absolute error of
all cells in the domain to provide a single metric (Siemens, 2018). Typically,
one would seek residuals that are as low as possible, say in the range 10-6 ~ 10-
8. In the present case, the specific dissipation rate was found to be satisfied to

the greatest extent, achieving values in the range of 10-8. The remaining
residuals, relating to the turbulent kinetic energy, x, y, z momentum equations
and continuity achieved values in the order of 10-4. The simulations were
initially run with 10 inner iterations, but this was found to be insufficient to
reduce the residuals to the desired range. This range is defined based on the
recommendations of the ITTC (2014). Specifically, the residuals should
decrease by 2 – 3 orders of magnitude.

6.4 Results and discussion

The first step in this section is to present the input data required. As mentioned
previously, two different metrics are examined: free surface elevation and skin
friction. The former is sampled 25×104 times, while the latter consists of 1.5×104
points on the ship hull.

The abovementioned points of the free surface are spaced uniformly in the x –
y plane on the port side of the ship and are depicted in Figure 6.3. Here, all
spatial dimensions are normalised by ship length for consistency. The final
sub-plot, shown in Figure 6.3 contains the convergence ratio (R), whose extents

142
are limited to the range of interest, namely ±1. This is done due to the fact that
if a point lands outside this range, the solution is divergent. From the figure it
is immediately apparent that human visual perception of the differences with
mesh refinement is by no means adequate to detect the discrepancies,
highlighted by the convergence ratio.

Figure 6.3. Fine, medium, coarse solutions, and convergence ratio of the free
surface
It is important to mention that the use of the refinement ratio of 21/3 means it is
not possible to sample the free surface at cell centres of each grid, where the
computations are performed. In cases where this is possible, the coarse grid
determines the sampling locations, whereas the medium and fine grid are
sampled every rth cell. In other words, the grids are nested. Instead, in the work
presented herein, the RANS solver interpolates the free surface at the
requested locations automatically. Such interpolation, and multiple sampling

143
from a single cell in the coarse mesh may introduce additional errors.
However, since the user is normally interested in an error estimator that is
slightly leaning towards conservativeness, this is thought as an acceptable
price to pay. In any case, an attempt to demonstrate the effect of sample
density will be demonstrated later.

In terms of skin friction, the ship hull is sampled at ten equally spaced lines,
i.e. buttock planes. In the context of ship hydrodynamics, the underwater
shape is typically the main contributor to resistance. Therefore, only the
submerged underwater area is normally sampled for skin friction. However,
if this were to be done, it would cause a different number of samples to be
acquired from each grid. This is the case because even a small variation in the
water surface near the hull may have a significant effect on the number of
samples. For this reason, the entirety of the ship is taken. To eliminate the
different sinkage and trim effect, the coordinate system is altered in each case
to match the orientation and position of the centre of gravity of the ship. This
allows all differences in position, orientation, and water elevation to be
accounted for.

Moreover, a change in the free surface elevation in any single grid would cause
a large numerical error and consequently uncertainty in the later stages. In
other words, employing the here-described technique of sampling the ship
hull for points, one avoids all potential problems, simultaneously accounting
for disparities in the water surface elevation on the hull which are not
addressed in the free surface sampling described previously.

The resulting points on the hull are shown in Figure 6.4, along with the
convergence ratios. Points lying on the above water part of the ship naturally
exhibit a much smaller skin friction than the submerged points, as expected.
The former are found at the bottom of the plot when examining the
distribution of skin friction in the longitudinal direction. It is also important to
mention that all coordinates have been normalised by ship length, as was the
case for the free surface.

In the present context, skin friction is defined via the wall shear stress (τw). As
before, the distribution of the convergence ratio in space reveals a complex
pattern. It is clearly discernible that points near the parallel midbody may be
responsible for significant deviations based on the value of R. This is an
unexpected observation, since the curvature of the hull in this particular

144
location is milder than it is near the more complex structures of the underwater
shape, specifically, the bulb and the stern.

Figure 6.4. Skin friction nodes from the ship hull

6.4.1 Local classification

The next step is to determine which nodes can be classed as type ‘Richardson
nodes’, and which are oscillatory. For illustrative purposes, this will be
coupled with the local observed order of accuracy, as defined by Eq. (6.14).
This is shown in Figure 6.5, where the Richardson nodes are predicted to
account for approximately 43.8% of all nodes, i.e. circa 11×104 points in the case
of the free surface. The remainder exhibit non-monotonic convergence or
divergence. It is important to note that the local observed order of accuracy
has been limited to 20. In any case, few points exceed this value. Figure 6.5
demonstrates that the computed free surface is separated into regions. The
boundaries of these regions are characterised by a high observed order of
accuracy.

The patches of Richardson or oscillatory nodes are clearly evident in locations,

next to the ship hull, as well as up to half a ship length downstream. These
locations are coincidently characterised by significant increases in mesh

145
density (refer to Figure 6.2). The clearly distinguishable patches coincide with
the wedge, prescribed to capture the potential effect of the Kelvin wake. In the
present simulations, the low depth Froude number (0.303), coupled with the
lateral restrictions allow for near-field waves only, as shown in Figure 6.3.

Figure 6.5. Observed local order of accuracy on the free surface nodes.

The observed local order of accuracy based on the skin friction distribution on
the ship hull is shown in Figure 6.6. Here, the maximum p has been maintained
as pmax=20 to retain consistency across the different parameters. However, as
was previously observed, few points approach this limiting value.

The pattern of patches of similar behaviour seems to be largely maintained in

Figure 6.6. To elaborate, it is apparent that nodes tend to switch from
Richardson to oscillatory behaviour in patches along the ship hull, although
this is not as easily observed as was the case in Figure 6.5. Many of the nodes
previously highlighted as potentially problematic – near the parallel midbody
of the hull – are shown to be characterised as of Richardson type. In the case
of skin friction, fewer than half of all sampled nodes exhibit oscillatory
behaviour. This shift, with respect the free surface highlights that different

146
parameters may require separate consideration in examinations of the type
presented herein.

Figure 6.6. Local observed order of accuracy on the ship hull

6.4.2 Error analysis and decomposition

It is now prudent to examine the behaviour of the error constants on the free
surface. These are calculated following the proposition of Roy (2008), and are
shown in Eq. (6.11) and Eq. (6.12) for the linear and quadratic components,
respectively. As stated in section 6.2.1, there are two possibilities for the
computed error constants. The 1st and 2nd order components either amplify or
attenuate the overall error once summed. The first case can occur regardless of
the sign of each constant, provided they approach the asymptotic range from
the same side. The contribution of these nodes in the case of the free surface is
demonstrated in Figure 6.7. Here, it is evident that the vast majority of error
contributions are a consequence of linear and quadratic terms of opposite sign.
Those with identical sign, i.e. approaching the asymptotic range from the top,
or bottom are coloured in red, and account for a small fraction of the overall
(less than 1% of all samples). Clearly, they can only exist in the narrow range
where the two constants cross over to a different quadrant in Figure 6.7. Thus,

147
leaving the vast majority of nodes to be classed as having opposite linear and
quadratic signs.

Therefore, all points, located near the extremities of Figure 6.7 contribute little
to the overall error, and consequently uncertainty. In the same figure, the
distribution of the error constants with respect to each other is also included.
This is done to demonstrate the fact that most predictions are located near the
centre of the plot. In other words, one should not expect excessive errors
arising from misalignment of the linear and quadratic components of the error.

Figure 6.7. Free surface error constants.

In the case of skin friction distribution, the error constants magnitude is not
similar to that of Figure 6.7, where the free surface error constants are given.
The skin friction error constants are depicted in Figure 6.8, which replicates
what was previously shown in the case of the free surface. Clearly, the number
of samples in Figure 6.8 is significantly smaller than in Figure 6.7. Although
this makes observations on the overall behaviour slightly more difficult,
several general trends can be identified.

148
In Figure 6.8, the range of both error constants is several orders of magnitude
higher than was the case for the free surface points. This is a direct
consequence of the fact that the latter were normalised by ship length prior
being decomposed into 1st and 2nd order contributions. Therefore, the range of
the axes should not be taken as an indicator of a problem in the assessment per
se. Secondly, as was the case in Figure 6.7, the error is distributed along the
diagonal of the plot in Figure 6.8. This observation, coupled with the small
number of points where the error is reinforced in the process of summation,
has several implications.

Figure 6.8. Skin friction error constants.

Primarily, the alignment of points along the aforementioned diagonal of

Figure 6.8 suggests that most points will exhibit a small error, resulting from
the cancellation of linear and quadratic components. Additionally, the error

149
amplifying points are constrained within a small wedge near the origin,
implying a small numerical error can be expected to stem from these samples.

In the following section, the spatial distribution of the error is assessed for both
parameters. For consistency, the free surface is examined first.

6.4.2.1 Spatial distribution of the error

The spatial distribution of the error constants (in the x – y plane) is depicted in
Figure 6.9 for the free surface. In the first sub-plot of the figure, the locations
where the linear and quadratic components reinforce the overall error have
been removed. With the exception of a line of samples near the outlet, the
nodes where the error is amplified exhibit no discernible pattern. Thus, it may
be postulated that such samples are a random occurrence, or are associated
with a type of numerical error that is not known at present. In any event, the
number of such points on the free surface is sufficiently small to be discounted.

The remaining sub-plots (2nd – 4th), shown in Figure 6.9 highlight the absolute
values of the error constants. These allow one to pinpoint the specific locations,
contributing to increased numerical error. Although any error estimation
technique could have been used, Roy's (2008) breakdown is preferable because
it allows a more detailed analysis. Moreover, since the same input is used for
all uncertainty estimators, the predicted locations will remain constant,
regardless of the method, as will be demonstrated later. The differences will
arise in terms of magnitude, rather than relative error distribution in the
domain.

Figure 6.9 can also be examined from a different viewpoint. There are two
distinctive features that are shown in all three possible error contributions (1st,
2nd, and 3rd order accurate solutions). Specifically, the locations where the
Kelvin wedge approaches the side wall exhibit elevated levels of numerical
error. It may be suggested that this is due to abrupt changes in mesh density.
However, no such pattern can be seen over the remainder of the boundaries
of the Kelvin wedge, whose arrangement can be consulted in Figure 6.2.
Therefore, it may be speculated that the elevated levels of error are attributable
to a combination of the change in mesh density, coupled with the proximity of
a solid wall. In fact, such an observation is supported by the presence of nearly
concentric semi-circles in the immediate vicinity of the ship hull (i.e. between
0 < x/L < 1).

150
 Figure 6.9. Spatial distribution of the linear and quadratic terms of the error
in the free surface, computed via Eq. (6.11) and Eq. (6.12).

However, the argument that an abrupt mesh density change in proximity of a

no-slip wall may not be sufficient to describe the entirety of the ‘error field’
shown in Figure 6.9. In other words, the arcs of elevated error, located entirely
within the Kelvin wedge do not fall within the above category. That is unless
the interpretation of extended influence of the solid boundary onto the flow is
adopted. To elaborate, it is possible that the hull’s direct influence as a
numerical error generator is greater than that of the side wall. This issue may
require further study to accurately pinpoint the cause of these arcs.

151
The same argument in terms of method used to assess the error applies to the
distribution of the skin friction on the ship hull, whose decomposed error
distribution is depicted in Figure 6.10. As before, the first step is to examine
the relative locations where the error is augmented versus those where it is
attenuated. In the present case, 3.11% of all samples fall within the former
category – a significant increase when compared to the free surface. As was
the case for the fee surface, no discernible pattern can be identified in the
spatial distribution of these points at present.

In Figure 6.10, the error constants are limited to a maximum magnitude of 60

to avoid skewing the axes too much because few points approach the limit.
This also allows the identification of the points with elevated values with ease:
almost the entire bulbous bow. On the other hand, locations near the parallel
midbody show low levels of error. This is also the case for locations above the
waterline. Such a remark may be interpreted as an indication that surface
curvature creates elevated errors. To examine this claim further, the skewness
angle of the first near-wall cell, along with its aspect ratio are shown in Figure
6.11.

Although significant skewness angles and aspect ratios are observed in Figure
6.11, no correlation can be identified between either of the above mesh
properties and the observed error distribution. At the onset, it was stated that
an attempt is to be made to correlate the abovementioned characteristics of the
numerical simulation. This was done in view of the fact that it has been
demonstrated to be possible for simple cases (Salas, 2006). This may not have
been achieved because a single mesh set-up was used and refined to produce
the three solutions.

Therefore, as a piece of future work, the aspect ratio of near-wall cells, could
be varied in a systematic manner to produce recommendations in this respect.
A problem worth considering in the course of such a study would be whether
the y+ values on the hull are targeted at the same value. In the event where they
are, the change in aspect ratio must stem from a lengthwise reduction, i.e. grid
refinement. Conversely, if the y+ values are not maintained constant across
cases, the grid can be coarsened to achieve different case-studies. In the present
case, the first near-wall cell’s spanwise and lengthwise dimensions are a
function of the base size. The base size represents the largest cell in the domain,
and is found near the outlet and inlet. It is important to maintain the same
aspect ratio across generated grids, because otherwise the procedure of
Richardson extrapolation is invalid (Salas and Atkins, 2009).

152
Figure 6.9 and Figure 6.10 reveal two important aspects of the numerical
simulation. Firstly, the linear contribution to the error is dominant in both
figures. This observation, coupled with the fact that the majority of samples
do not exhibit a behaviour of error amplification when summed suggests good
performance of the numerical simulation. Such a conclusion can be drawn
because the dominant error behaviour is known to be of 1st order as the grid is
refined past a certain point, and of second order in coarser grids. This is
exemplified in the following sub-section.

Figure 6.10. Spatial distribution of the linear and quadratic terms of the error
in the skin friction, computed via Eq. (6.11) and Eq. (6.12).

153
 Figure 6.11. Cell aspect ratio and skewness angle on ship hull: fine solution

6.4.3 Error behaviour considerations

For this sub-section, samples from the free surface are considered only. This is
done because of the greater number of available points. Regardless, the
discussions and conclusions reached in what follows are equally valid for both
parameters.

Figure 6.12. Example of error reinforcement and attenuation

154
The error behaviour, shown in Figure 6.12 is assembled by randomly taking
two samples from each error category depicted in Figure 6.7. Specifically, a
node where the 1st and 2nd order normalised error approach the asymptotic
range from different sides, and a node where they approach the asymptotic
range from the same side are extracted. These are classified with ‘-‘ and ‘+’ in
Figure 6.12, respectively. To simplify the figure, the logarithm of each term is
used. This choice is made to allow the error line’s slope to depict the order to
which the term is raised. In other words, the line with slope one represents the
first term of the Taylor series expansion shown in Eq. (6.10).

As referred to previously, Figure 6.12 depicts the presence of a hiatus when

the first and second order error contributions intersect. This only occurs in
cases where the asymptotic range is approached from opposite sides. Such a
scenario is, as established in Figure 6.7, representative of the vast majority of
all samples of the free surface and skin friction. Points near the hiatus are to be
avoided, because of the possibility of spurious behaviour upon employing a
grid in the proximity of the cusp. This is because Richardson extrapolation
assumes smooth error derivatives, which is not the case in error attenuating
scenarios near the cusp (Celik et al., 2005).

It is worthwhile to examine what would happen if, by chance, a solution were

to be produced at precisely the grid size corresponding to the hiatus. In a
typical Richardson extrapolation, for sufficiently fine grids the dominant error
is raised to the lowest power, i.e. p. Thus, setting the order of accuracy as the
lowest, dominant component of the Taylor series expansion and neglecting
higher order terms is justifiable (Salas and Atkins, 2009). However, as h is
increased, multiple errors of different behaviour begin competing in the
Taylor series expansion. Thus, setting p as any one value and neglecting higher
order terms is not correct in the vicinity of the hiatus.

The reason why Figure 6.12 is incorporated is to demonstrate a fact concerning

the error’s behaviour with grid refinement. Specifically, second order
convergence, or more precisely, error reduction, can only be expected for
coarse grids. This is independent of whether the error is reinforced or
attenuated by its separate contributions.

On the other hand, once the hiatus (if one is present) has been overcome, the
cumulative error exhibits first order convergence. This follows as a
consequence of the second order term’s contribution decreasing at a much
faster rate than the remainder of the error. Therefore, if one observes a global

155
error with convergence properties that are close to linear, it can be claimed that
the simulation is characterised by acceptable numerical error. It is important
to note that no claim of whether a particular grid is asymptotic or not is made
based on the above criterion. Instead, it is suggested that little improvement,
i.e. reduction in numerical error can be achieved with grid refinement.

To elaborate, depending on the case, it may be justifiable to reject further grid

refinement if the predicted numerical error will not decrease noticeably. Such
an argument could be made if the error is shown to be located in the range
where linear convergence dominates. On the other hand, if the specific grid
produces errors in the quadratically convergent region, further grid
refinement may be warranted. Further supporting evidence to this
observation can be extracted from several works on numerical uncertainty
estimation. In particular, the least-squares (LSQ) estimator of Eca and
Hoekstra (2009, 2006). The LSQ method was designed to treat solutions with
0.5 < 𝑝 < 2.1 as asymptotic.

Another aspect of Figure 6.12 worth bearing in mind relates to the apparent
order of convergence with grid refinement. It has already been stated that for
fine grids, p is close to unity, while for coarse grids, it attains a value of in the
vicinity of two. However, it is important to consider the manner in which such
a transition occurs. In the case where the error is supplemented by the two
components of the Taylor series, the error smoothly transitions from 1st to 2nd
order. This can be easily deduced by an examination of the slope of the dashed
blue line in Figure 6.12. There are no discontinuities, and its slope varies from
1 in the fine mesh region to 2 in the coarse mesh region. On the other hand,
when the error is attenuated, the slope of the solid blue line exhibits a different
behaviour. Although it too is of 1st order in the fine mesh range, the transition
to 2nd order occurs though −∞ and +∞. In the course of this transition, the
order of convergence attains all values except those between 1 and 2 (Salas and
Atkins, 2009).

Combining the findings of Figure 6.9 and Figure 6.10 with those of Figure 6.12
suggests that the linear error has a much greater magnitude than its quadratic
counterpart for the computed free surface and skin friction. Therefore, the
locations where the majority of the error is expected to be found in the log-log
plane of Figure 6.12 is to the left of the hiatus. That is, for most points, whose
error approaches the asymptotic range from opposite directions. The error
supplementing cases also follow a similar trend, albeit without the presence of
a hiatus, which may produce spurious solutions.

156
In the context of asymptotic behaviour of the grid, there is one final angle from
which the solution may be considered. As explained previously, the above
method decomposes the error into first and second order components.
Naturally, these are not dependent on the local observed order of accuracy,
which is estimated via Eq. (6.14). This is true because 𝑝 (nor 𝑝̂ ) is not used in
Roy's (2008) method. However, the order of accuracy can be interpreted as a
measure of the convergence mode. That is to say, if 𝑝 contains a real part only,
then the solutions supplied into the equation converge monotonically. This is
analogous to maintaining the convergence ratio between 0 and 1.

To exemplify the behaviour of the error constants versus each component of a

typical Richardson extrapolation, Figure 6.13 depicts the behaviour of the
error constants against the convergence ratio, whereas Figure 6.14 – the error
constants against the observed order of accuracy. Figure 6.13 is characterised
by distinct asymptotes, located at a different R value for each error constant.
The prediction of where this location will lie depends exclusively on the
refinement ratio. Specifically, no solutions can exist outside the boundaries
defined by Eq. (6.20):

Figure 6.13. Error constants vs. convergence ratio

1⁄𝑟 𝑓𝑜𝑟 1𝑠𝑡 𝑜𝑟𝑑𝑒𝑟

Asymptote locations { 1⁄𝑟 2 𝑓𝑜𝑟 2𝑛𝑑 𝑜𝑟𝑑𝑒𝑟 (6.20)
1⁄(𝑟 2 + 𝑟 − 1) 𝑓𝑜𝑟 1𝑠𝑡 + 2𝑛𝑑 𝑜𝑟𝑑𝑒𝑟

157
Having employed 𝑟 = 21/3 , the asymptotes are located at 𝑅 = 0.7937, 0.6300,
and 0.5413 for the linear, quadratic terms, and their sum, respectively. Since
one would ideally seek to obtain solutions that exhibit solely monotonic
convergence, the x – axis has been limited between 0 and 1 for R. Having
established that the prediction of the relative location of the asymptotes is
trivial, one could employ this information before conducting the numerical
simulations. The choice of r could be dictated so as to maximise the available
area in Figure 6.13. In the present case, the choice of convergence ratio was
governed exclusively by the rapid increase in cell numbers. Had a larger
convergence ratio been selected, the cell numbers would have increased
beyond a manageable size, whereas the idea was to allow for a large number
of sequentially refined sets of simulations to be achieved.

Figure 6.14 demonstrates that although Roy's (2008) method does not
explicitly depend on the order of accuracy, discernible patterns can be
identified by combining the two. The error constant’s behaviour on
monotonically convergent samples shows clear trends, whereas the non-
monotonic nodes do not. The former are characterised by a sharp cusp. The
linear component shows that this cusp is located at p=2, the quadratic at p=1,
while their sum: p=2.654. Figure 6.14 also confirms that the error due to the 2nd
order term attains a smaller magnitude over the entire range examined. The
1st order term also grows at a greater rate than the quadratic term as the
distance from the cusp is increased. The relative location of where this occurs
is again governed exclusively by the refinement ratio. To demonstrate this, one
may simply substitute the terms of Eq. (6.20) into Eq. (6.14). In the non-
monotonic case, all points seem to have a tendency of increased error when
compared to the monotonic samples. Therefore, despite the fact that the
method does not take into account the convergence ratio, nor the order of
accuracy per se, clear relationships can be established based on these
components and the error constants.

158
 Figure 6.14. Error constants vs. local observed local of accuracy

6.4.4 Local uncertainty and extrapolated solutions

In this sub-section, the uncertainty estimates are presented for both the free
surface and the skin friction data. A discussion is also given on the choice of
the Factor of Safety, as well as the best choice for the order of convergence.

6.4.4.1 Factor of Safety

The first step is to determine whether the suggested FS of 1.25 as part of the
GCI procedure agrees with the procedure of Phillips and Roy (2017) for
estimating the FS. This is presented for both local parameters in Figure 6.15.

The results reveal that the skin friction data is asymptotic based on the
criterion of Eq. (6.17), which exhibit a Factor of Safety of 1.2532 with a distance
from the asymptotic range of 0.0209. The reader is reminded that in the present
context, the distance from the asymptotic range metric is defined as the
deviation from the formal order of accuracy (𝑝𝑓 = 2). Therefore, the skin
friction data is highly asymptotic, suggesting that further refinement is of little

159
use in the present case. This is an encouraging finding considering the high
aspect ratios of the near-wall cells and their skewness angles, shown in Figure
6.11.

Figure 6.15. Computed Factors of Safety using Eq. (6.17).

The free surface on the other hand is highly non-asymptotic. It may seem that
the point depicting its relative location on the Δp range is not at its further
possible limit. However, Eq. (6.17) limits the distance from the asymptotic
range to 95% of 𝑝𝑓 , i.e. a maximum of 1.9, where the free surface Factor of
Safety lies. If this limit were not imposed, the attained value along the Δp axis
is predicted as 3.0571.

The remainder of the factors of safety and modified orders of convergence are
not shown because they depend on simple manipulations of 𝑝 and 𝑝𝑓 . Instead,
prevalence is given on the uncertainty itself in what follows.

6.4.4.2 Uncertainty spatial distribution

In this sub-section, the uncertainties for both sampled parameters are given.
This is done according to all methods described in Section 2.1.

In Figure 6.16, the uncertainty of the free surface based on the method of
Phillips and Roy (2017) is presented with 𝑝 = 2 in the denominator of Eq.
(6.18). This is done following the recommendation of the above authors. The
uncertainty shown in second place, i.e. that due to Roy (2008) is computed

160
with a variable factor of safety, and the error provided by the error constants,
shown in Eq. (6.11) and Eq. (6.12) multiplied by the fine grid spacing and the
FS shown in Figure 6.15. This is then divided by 𝑟 𝑝𝑓 − 1 in line with other error
estimators.

Figure 6.16. Local uncertainty spatial distribution on the free surface

according to all methods.

The remainder of uncertainty estimators of Figure 6.16 follow as explained in

Section 2. In other words, no modifications have been attempted on these. This

161
is done to avoid compromising the confidence level of each prediction. The
main conclusion, drawn from Figure 6.16 is that the first two methods, due to
Phillips and Roy (2017) and due to Roy (2008) exhibit similar levels of
convergence, in line with other methods with a longer history of
implementation. This is a promising outcome, demonstrating that their use is
likely capable of providing good results. Coupling this with the in-depth
assessment possible based on the abovementioned methods makes them a
desirable approach to error and uncertainty estimation.

As postulated earlier, the locations of heightened uncertainty do not change

across methods. However, the magnitude of the uncertainty is highly sensitive
to the method employed. For instance, the GCI method predicts heightened
uncertainty near y/L between -2 and -1.5. Although the decomposed linear and
quadratic components also exhibit increased uncertainty in the vicinity of this
region, error cancellation seems to dampen the uncertainty magnitude.
Therefore, a greater level of detail can be extracted by the decomposed form
of the error. Moreover, the uncertainty is highest in this type of error estimator.
However, it should be employed with caution, because as stated previously,
its confidence level has yet to be established. Based on the generated data the
level of conservativeness predicted for the mixed order method is circa 51.64%.

Conservativeness is defined as the percentage of points, where 𝑓𝑖 − 𝑈 <

𝑓𝑒𝑥𝑡 < 𝑓𝑖 + 𝑈 is satisfied. In other words, points where the predicted
extrapolated solution lies within the uncertainty band. Since the mixed order
method of Roy (2008) is designed for specific orders of magnitude, it is not
capable of providing a sufficiently large band of uncertainty to accommodate
an extrapolated solution with an observed order of accuracy far from 1 st and
2nd powers of the grid spacing. Alternatively, the extrapolation itself may not
be a good estimate if the observed order of accuracy is far from the pre-defined
powers of the grid spacing. On the other hand, the method of (Phillips and
Roy, 2017) provided a coverage of 73.60%.

A confidence interval below 95% is to be expected, since nearly half of all

points were not classified as ‘Richardson nodes’, which is a precondition for
the procedure itself. It is likely that the Factor of Safety’s value (FS =3) is in part
the reason why 73.60% coverage is achieved, rather than a confidence interval
equal to the fraction of Richardson nodes. If the present exercise were to be
repeated on a structured solver, one may expect that a greater confidence level
would be achieved. Unstructured solvers tend to deteriorate the convergence

162
properties of the solution (Thomas and Langley, 2008). Due to the small
number of skin friction samples, confidence bounds are not predicted for τw.

The skin friction uncertainty is shown in Figure 6.17 according to the same
methods as done for the free surface. In this case, the limits of the uncertainty
have been increased for two methods: Roy's (2008) and the FSRE method. This
is done to accommodate the larger predictions provided by these estimators.
The degree to which the grid is asymptotic, demonstrated in the previous
section (Figure 6.15), has rendered the uncertainty prediction due to Phillips
and Roy (2017) the smallest. Such behaviour is justified, because the near-wall
grid was shown to be highly asymptotic of all methods. The difference
between the abovementioned technique and the GCI method is that the former
uses the theoretical order of accuracy. The GCI method on the other hand
employs the locally observed order, which is in large parts of the domain
larger than two. This can be consulted in Figure 6.14.

One final aspect of Figure 6.17 worth considering relates to the distinction of
wet and dry areas of the hull. In all methods, the above distinction is clearly
visible, with submerged points tending towards greater uncertainty. To
provide a more intuitive representation of the uncertainty, Figure 6.18 is
constructed to show the fine solution along with the uncertainty computed by
each method. In the plot it is more evident which locations along the x – axis
are culpable for greater uncertainty levels. It is also manifest that points
located near the origin of the abscissa contribute virtually nothing to the
overall levels of uncertainty. As stated previously, these points represent the
dry parts of the ship hull.

The representation used in Figure 6.18 highlights the existence of several

locations where certain uncertainty estimators predict a significantly greater
level of uncertainty than others. For example, the GCI method shows that few
points in the stern area of the ship are characterised by large errors. On the
other hand, Roy's (2008) method predicts a greater number of points behaving
in this manner. This is suspected to be due to the interaction of error constants,
used by the latter method. Since the GCI method does not distinguish between
different orders of convergence in this respect, the decomposition approach
can unveil greater detail. At this stage, no comment in terms of robustness is
possible, due to the previously examined reasons.

163
Figure 6.17. Local uncertainty spatial distribution skin friction on the ship
hull according to all methods

164
 Figure 6.18. Skin friction with error bars.

6.4.4.3 Extrapolated solutions

This section examines the final aspect of the numerical verification procedure.
Namely, the extrapolated solutions. There are two such solutions for each
sampled point in each parameter (free surface and skin friction). These are the
GCI and Roy's (2008) extrapolated solutions. A comparison between the two
enables the analyst to confirm the locations of heightened error. Such locations
will register in the extrapolated solution as large deviations from the fine
solution. Furthermore, upon comparing extrapolated solutions achieved by
different methods, it is possible to show which method can cope with non-
monotonic input better. This is the case, because such a method would allow
for oscillatory samples, providing an extrapolated solution that resembles the
input to a greater extent.

For consistency, the free surface is examined first in Figure 6.19, where the two
extrapolated solutions are accompanied by the ratio of the fine and respective
extrapolated solution. This is done to better highlight the differences between

165
the separate solutions, and their interaction with the fine solution. The figure
serves to demonstrate that both methods predict a discontinuous line
emanating from the aft perpendicular and progressing towards the centre of
the domain. However, they disagree on its shape and extent. The same is true
for a line in the forward part of the hull.

Figure 6.19. Extrapolated free surfaces

166
Figure 6.19 also reveals patches of differences forward of the ship, the origin
of which can only be speculated at. The relative discrepancies between the two
methods also highlight that the GCI method predicts to a lesser extent the
semi-circular arcs in the proximity of the ship. The elevated levels of error at
the side wall are also shown to be of different magnitude. More importantly,
the region prescribed to provide the VOF damping seems to be a major source
of disagreement between the two methods. This observation indicates that a
wide range of numerical parameters simultaneously influence, and therefore
compete for the dominant contribution in the error and uncertainty estimate.
Splitting these from the grid-induced errors and bifurcating them as separate
components may therefore not be the best approach in non-asymptotic grids.

The next step is to examine how the two extrapolated solutions compare in
terms of skin friction. This is depicted in Figure 6.20 in the same order as was
previously done for the free surface. Here, the GCI method is shown to
provide solutions that are practically indistinguishable from the fine solution
itself. This can be verified by consulting the plot representing the ratio of the
GCI solution and fine solution. Almost all points in this category have attained
a value of either one, or are sufficiently close to one. The mixed order method
on the other hand does not resemble the fine solution as closely. This causes
differences in the predicted solution by the aforementioned approach to be
evident when considering its ratio with the fine solution.

Upon comparing the ratio of the two extrapolated solutions in terms of skin
friction, further differences emerge. These are visible primarily near the
extremities of the ship, suggesting that areas of high curvature may cause
disagreement between the two methods. It is likely that the cause of this
observation once again stems from the interaction of linear and quadratic
components. Figure 6.10 points towards the existence of large errors in both
the linear and quadratic terms of the error in this region. Even a small error
misalignment will create heightened levels of uncertainty, which may be
responsible for the observed levels of disagreement in Figure 6.20.

In summary, this section demonstrated that each error and uncertainty

estimator has a distinct tendency when it comes to predictions. The
formulation of each distinct approach governs which regions of the sampled
domain are highlighted with an increased concentration of numerical
problems. In many cases, different aspects of the numerical solution may, and
do interact in unpredictable ways. This may be the error itself, exhibiting
discontinuities in terms of value as well as order, or other aspects, such as free

167
surface modelling, convection properties or separate unexamined effects.
These include sharpening factors in the VoF scheme, convection order, as well
as the grid resolution. On the other hand, the interplay of cell aspect ratio,
skewness angle, y+ value, turbulence model, and convection scheme order
generate a distinct problem. In many cases, this interplay may be fortuitously
favourable, but unknown, as is probably the case in terms of skin friction. On
the other hand, the level of complexity of all components influencing the free
surface make it impossible to accurately pinpoint the specific cause-effect
relationship, leading to the observed error and uncertainty. However, by
attempting to isolate separate aspects of the numerical solution, such as grid
density, it is possible to provide recommendations relating to best practices.

Figure 6.20. Extrapolated skin friction distribution on the hull.

168
It was demonstrated that in full-scale applications, a y+ value below one is not
a necessary precondition to classifying the near-wall grid as asymptotic. On
the other hand, a highly dense grid, encompassing the area where the free
surface deforms was shown to be insufficient to brand the sampled points as
generated by an asymptotic grid. This also highlights the fact that depending
on the grid arrangement and numerical set-up, vastly different behaviours in
terms of error can be achieved in separate aspects of the simulation. It should
be borne in mind that an asymptotic solution does not necessarily translate
into a low validation error. Therefore, the next step would be to compare the
solution, generated by a study similar to that presented herein, with
experimental data. Recently emerging methods to capture free surface
deformations experimentally show potential in this respect (Caplier et al.,
2016).

6.4.5 Effect on sampling density

The present assessment used 25×104 points to sample the free surface. It is not
reasonable to expect such level of detail be used in industrial applications. This
is the case for two reasons. Primarily, because exporting 25×104 points requires
both time and care. But also, one would ideally like to know what is the
minimal sample density, required to produce a reliable result. This section will
attempt to answer the above question by taking a different number of samples,
randomly from the available points.

The present sampling study is performed in steps. Firstly, statistical measures

are used to determine the error due to sample size, assuming the exact number
of Richardson nodes is known. Then, the available free surface is randomly
sampled repeatedly at different levels of density.

The error due to sample size (E) is estimated as shown in Eq. (6.21) for a 95%
confidence level, which is depicted graphically in Figure 6.21.

𝐸 = 1.96 × √𝑚 × (𝑚 − 1)⁄𝑛 (6.21)

where m is the currently estimated percentage of Richardson nodes (43.9696%

of all free surface points), and n is the total number of available points. This
procedure implicitly assumes that the above fraction of the total represents the
true number of Richardson nodes in the domain. Such an assertion should be
made with caution, because there is no manner to determine the exact number
of nodes in each category. Moreover, the classification of nodes may change if
the grid is refined further

169
 Figure 6.21. Error induced by sample size as predicted by Eq. (6.21).

For the reasons explained above, the free surface is also sampled using
random, uniformly distributed points 25×104 times. This is performed at the
density depicted in the title of each sub-plot of Figure 6.22, i.e. by taking [50%,
25%, 12.5%, 6.25%, 3%, 1%, 0.1%, 0.01%] of all samples. The results from this
are presented in Figure 6.23. Here, the Factor of safety and fraction of
Richardson nodes are monitored at each sampling iteration. Then, the
standard deviation and mean values are used to construct the plot. No
individual sample revealed a FS value other than 3, which is why no standard
deviation of the Factor of Safety is included. The only varying metric was
determined as the % of Richardson nodes. The sample independent solution
is also estimated in Figure 6.23 using the GCI procedure to extrapolate a point,
based on the final three solutions. Since these exhibit oscillatory convergence,
the absolute value modification is employed in the estimation of the order of
convergence. The sample-independent solution is shown as a filled circle at
the end of the abscissa.

170
Figure 6.22. Sampling of the free surface. Each sampling, featuring less than 100% of all points is repeated 25×104 times.

171
 Figure 6.23. Influence of sample size.

The results presented in this Section suggest that 104 samples should be
sufficient to estimate the Factor of Safety and fraction of Richardson nodes
with sufficient confidence. The first part of the assessment performed here
suggests that such a sampling density would induce errors smaller than 1%.
The second assessment agrees with this conclusion, showing a standard
deviation below 1% in the region of 104 samples. Since both types of
assessments show similar behaviour, it is concluded that the aforementioned
number of samples in full-scale is sufficient to accurately estimate the
parameters of interest.

6.5 Conclusion and recommendations for future work

The importance of verification of CFD work has gained increased importance

and attention, with many influential academic journals and institutes revising
their editorial policies to reject any papers not featuring such assessments
(American Institute of Aeronautics and Astronautics (AIAA), 1994; Celik et al.,
2008; Roache et al., 1986). This fact, coupled with the increased reliance on
simulation-based design in all fields of engineering suggests that numerical
verification will only increase in importance. It is therefore paramount that
new methods are assimilated within all fields of engineering soon after their
inception.

172
This chapter presented the application of non-intrusive a posteriori methods to
assess the numerical performance of a set of CFD simulations. The main goal
was to introduce some well-known local methods of numerical verification to
the field of ship hydrodynamics. To increase the practical applicability of the
work presented herein, a set of full-scale simulations of the KCS advancing in
a canal were adopted. Then, the numerical verification procedures were
applied on local parameters, specifically, the free surface elevation and the
non-dimensional shear stress, i.e. the skin friction.

The obtained results were presented in terms of several parameters of interest,

including the decomposed error, uncertainty as well as the extrapolated
solutions. In the case of the free surface, the assessment revealed semi-circular
arcs of heightened numerical uncertainty emanating from the hull. The
numerical damping beach implemented aft of the ship was also identified as a
definite source of uncertainty. More importantly, it was demonstrated that the
numerical uncertainty predictions in the damping region were in
disagreement between the examined methods. Thus, suggesting that
numerous aspects of the numerical simulation likely compete for the
dominant error contribution. The grid, containing the free surface was also
demonstrated to be far from asymptotic. This is likely due to interaction of
different numerical parameters involving the definition of the free surface,
which may affect the observed order of convergence.

In terms of skin friction, the assessment revealed that the near-wall grid is
highly asymptotic. Regardless of the proximity of the generated grid to the
asymptotic range, several uncertainty estimators were tested. These revealed
significant disagreements in the magnitude of the uncertainty depending on
the adopted approach. Based on the above findings it can be stated that the
uncertainty is expected to be highly dependent on the parameter used in its
prediction. Different parts of the computational domain are governed by
distinct numerical schemes and approaches, thereby increasing the complexity
of the problem at hand significantly.

The viscous-dominated part of the domain, located near impermeable

boundaries, can be refined sufficiently to bring it in line with the asymptotic
range. That is even though this particular part of the computational domain
features uncertainties stemming from the choice of turbulence model, y+
strategy, as well as a phase interphase (to name but a few aspects of the
solution). On the other hand, the free surface grid resolution, containing the
majority of the cells in the domain was not near the asymptotic range. The

173
distinction of the grid’s performance with respect to the asymptotic range is
important, because verification procedures implicitly contain the above
requirement. Therefore, the determined uncertainty is a function of grid
performance in this respect.

174
 VIRTUAL REPLICA OF A TOWING
TANK EXPERIMENT TO
DETERMINE THE KELVIN HALF-
ANGLE OF A SHIP IN RESTRICTED
WATER

This chapter presents a virtual towing tank that is consistent with physical
towing facilities. This is achieved by removal of all open boundaries within
the computational domain. The adopted case studies represent long, narrow
canals. The resistance of the ship in the New Suez Canal and a rectangular
canal are compared with experimental data. The generated wavefield,
including the predicted Kelvin half-angle are validated by means of a Fourier
approach. The assessment suggests it may be possible to validate numerical
free surfaces even in the absence of experimental data.

7.1 Introduction

Computational Fluid Dynamics (CFD) has become widely accepted as a useful

tool to predict flow around a ship. This is facilitated by the increase in available
computational power, which has allowed practitioners to re-create the flow
around a vessel even on a standard computer. Thus, the number of cells, or
more generally, the computational effort required to perform a numerical
simulation in model-scale is not thought to be prohibitive for practical
applications.

175
Regardless of the advances in every field of numerical modelling, CFD is not
yet considered a replacement of model-scale experimentation. This is because
it is not possible to guarantee that a particular numerical model will perform
with the same level of accuracy across all possible case studies. For example,
new energy-saving devices, or novel underwater shapes may require research
into the best applicable modelling approaches. Additionally, the consequences
of implementing modelling assumptions may not be fully understood.
Specifically, although in model-scale computations, a significant portion of the
ship hull is covered by a laminar layer, most turbulence models assume the
flow is fully turbulent. Yet, results with accuracy of a few percentage error can
be found in the open literature (Bašić et al., 2017; Bechthold and Kastens, 2020;
Farkas et al., 2018; Larsson et al., 2014; Razgallah et al., 2018; Simonsen et al.,
2013; Toxopeus, 2013).

There are also different aspects of the problem of modelling ship flows that
can be validated with different levels of confidence. For instance, the resistance
of a ship can be measured accurately. However, velocities in the wake of the
ship, or free surface elevations require complex and expensive equipment.
Thus, in the course of validating a numerical result, researchers typically
analyse the error in observed integral quantities (resistance, motions, etc.), but
tend to assume that other flow features are also accurately modelled as a
consequence. Although this may be the true in many cases, an approach to
validate aspects of the flow around a ship, such as the generated wave field, is
necessary. Ideally, such a method would not rely on expensive equipment, nor
complex mathematics, in other words, it should be accessible. It is important
to mention that some experimental campaigns report on a wide range of
features of the flow around the ship, for instance, the flow properties in the
wake (Longo and Stern, 2002; Tahara et al., 2002b).

The research presented in this chapter is motivated primarily by the manner

in which the problem of ship resistance is typically solved. That is, the
principle of Galilean relativity is invoked (also called frame invariance; further
information can be found in Kundu et al. (2012)). Namely, the water is flowing
over a stationary ship (in the direction of the incoming flow). This assumption
has several consequences. Those particularly important to the naval architect
are:

176
 1. Levels of inlet turbulence.
This can have an impact on the overall properties of the flow (Wang et
al., 2015) and may require calibration in some cases. For example,
according to Lopes et al. (2017), the onset of transition from laminar to
turbulent boundary layer is strongly dependent on the level of free-
stream turbulence. Some two-equation models, such as the SST k – ω
model (which is widely used in marine hydrodynamics), are known to
predict excessive decay of free-stream turbulence, which may affect the
results. More recently, Lopes et al. (2019) examined the same topic.
According to them, even if one were to employ a more advanced eddy-
viscosity model, capable of accounting for transition, the location of
where laminar-turbulent transition occurs is highly dependent on the
level of inlet turbulence.
2. Wave reflections and their damping.
In cases where the Volume of Fluid (VoF) method is used (Wackers et
al., 2011), a damping length is often prescribed. That is, a length over
which all waves are damped, extending from the boundary it is applied
to in the normal direction. Setting an inappropriate damping length can
have severe consequences to the predicted parameters (Perić and
Abdel-Maksoud, 2016).
3. The temporal dependency of free surface flows.
The simulation of free surface flows via CFD cannot be solved using
steady-state solvers (except in rare academic cases), because they
require that properties are convected through the domain (Wackers et
al., 2011). Theoretically, in the frame of reference of the ship, the flow –
once converged – is steady (provided no separation and wave breaking
occur). Therefore, ship resistance is frequently classed as a pseudo-
steady problem. In reality, towing takes place over time, and is a
fundamentally unsteady process. Here, the presence of turbulence,
which is by definition time-dependent (Durbin and Pettersson Reif,
2011), should also be kept in mind.

A second aspect, inspiring this Chapter partly stems from point (2) above.
Although these may be of less interest to the naval architect, they carry their
own importance, nonetheless. Specifically, the destruction of ship waves,
regardless of whether or not damping is prescribed. Once a ship-generated
wave reaches the outlet, it is irreversibly destroyed, and the information it
carries – lost. In shallow and restricted waters, ship waves are of great
importance because they cause bank erosion, and may even lead to destruction

177
of coastal features/infrastructure (Grue, 2017; Sorensen, 1997). In extreme
cases, they may even be the cause of loss of life, as stated by Soomere (2007).
Therefore, the accurate modelling of ship waves and their interactions with
riverbeds, or canal sides is important. Soomere (2007) also advances a criticism
of ship-induced flow predictions. Namely, that the flow is only described at a
distance of few ship lengths.

Clearly, ship waves are both of practical and research interest. Therefore, the
validation of numerical ship-generated waves is of high importance. In this
respect, the work of Caplier et al. (2016), Fourdrinoy et al. (2019), and Gomit
et al. (2014) is important to mention. The authors of the aforementioned
references systematically developed and implemented a technique to capture
and analyse ship-generated waves from a model experiment. Of interest to the
present research is the fact that in their studies, the authors proved the
dispersion relation in deep and shallow water and demonstrated its validity
for ships experimentally. Since the developed technique relies primarily on
spectral representation of the wave field, it is thought prudent to attempt its
application to numerically generated free surface disturbance caused by a
ship. It is expected that, if applied correctly, it is possible to validate a
numerical wave field simply by means of processing a virtual free surface,
which would be undoubtedly of practical use. Such a method has the potential
to change how numerical solutions of surface piercing bodies are treated.

The present chapter will attempt to apply the aforementioned spectral

technique on a different type of numerical towing tank. Instead of relying on
Galilean relativity, the present paper will present a numerical replica of a
towing tank, where the ship advances over a stationary fluid. This is achieved
via the overset domain method, where the ship is encased in what is essentially
a moving box. To perform the numerical simulations, the commercial Reynold
Averaged Navier-Stokes (RANS) solver, Star-CCM+ version 13.06, is used. The
specific case studies adopted in this chapter are selected to maximise the
practical relevance of the present work. Specifically, the New Suez Canal is
replicated, alongside a standard rectangular canal, which were investigated
experimentally by Elsherbiny et al. (2019). The KRISO container ship (KCS)
with a scale factor of 1:75, following available experimental data is used for all
simulations.

The aim of this chapter is primarily to demonstrate that it is possible to create

a virtual towing tank where the ship is towed using the overset method, i.e. a
virtual towing tank that does not rely on Galilean relativity. The generated

178
wave field will then be used to estimate the Kelvin half-angle for an example
case. The adopted approach also allows one to split the near- and far-field
wave systems, which is used on the fully non-linear disturbance, generated by
the KCS at a variety of speeds in two different canals.

This chapter is organised as follows. Section 7.2 contains a description of the

adopted case studies, while section 7.3 explains the adopted methodology,
which is split into the two techniques used in this chapter. Namely, the
computational set-up and the spectral representation techniques. Section 7.4 is
dedicated to results and their discussion, whereas section 7.5 contains
conclusions and summary.

7.2 Case studies

As mentioned in the previous section, the adopted case studies are taken from
the experimental work of Elsherbiny et al. (2019). The rationale behind this
choice relates to the particular objective of this chapter. To elaborate, shallow
water studies are a natural choice for the examination of ship-generated
waves. This is because they present several features, absent in deep water ship-
generated waves. Shallow water waves are nonlinear, and their Kelvin half-
angle is speed-dependent (Tunaley, 2014; Yang et al., 2011). This is illustrated
in Figure 7.1, which is constructed via Havelock's (1908) linear method. Here,
the Kelvin wake angle increases from its deep-water value of ≈19.47° to 90°.
The theory predicts that at a depth Froude number (𝐹ℎ = 𝑈/√𝑔ℎ, where 𝑈 is
the ship speed in m/s, 𝑔 is the gravitational acceleration, and h is the water
depth) of one, 𝐹ℎ = 1, the ship-generated waves will travel at the same speed
as the disturbance, indicating the Kelvin wedge fills the entire half-plane
aft of the disturbance, i.e. at a half-angle of 𝜃 = 90°. The relationships derived
by Havelock (1908) are omitted in the present work, as they are available in
the open literature.

179
Figure 7.1. Kelvin half-angle of ship-generated waves in shallow waters as a
function of the depth Froude number (Havelock, 1908).

Ship-generated waves are also of greater concern in restricted areas than in

deep waters, because they may affect the surrounding environment
detrimentally. In navigational fairways, bank erosion is of particular concern,
which has led to authorities restricting the speed with which vessels are legally
allowed to operate (Suez Canal Authority, 2019). Such a restriction
simultaneously guards against groundings.

The case studies adopted herein are chosen to reflect the aforementioned
points. In this respect, the recent work of Elsherbiny et al. (2019b) is used as a
benchmark. From their experimentally investigated cases, two different canal
cross sections are selected: The New Suez Canal, and a standard rectangular
canal. These are graphically depicted in Figure 7.2.

180
 Figure 7.2. Graphical depiction of the cross-section of the selected case-
studies. Top: New Suez Canal, bottom: rectangular canal.

The ship used also follows from the experimental campaign of Elsherbiny et
al. (2019b). Namely, the KCS hull form is used, scaled by a factor of 1:75. This
translated into a depth-to-draught ratio of 2.2, based on the ship’s design
draught. In order to ensure that a well-defined Kelvin wake is simulated, the
selected depth Froude numbers are towards the high end of the
experimentally available conditions. The ship’s particulars are given in Table
7.1, whereas test matrix alongside the predicted Kelvin half-angles (via
Havelock's (1908) method) are described in Table 7.2. It should be noted that
Havelock's (1908) method was originally devised for point sources, and is
therefore not expected to be perfectly accurate for non-linear three-
dimensional surface piercing bodies. Nonetheless, it is a useful starting point.
Additionally, turbulence, viscosity and vorticity may influence ship-generated
waves, particularly in the near-field (Lee and Lee, 2019).

The relatively high depth Froude numbers ensure the numerically generated
wave field will be discernible. The spectral method used also performs best at
high speeds, where the near and far-field disturbances generated by the ship
are well visible. This can be seen by consulting the results of Caplier et al.
(2016).

7.3 Methodology

This section is presented in two major parts. These reflect the methodologies
used within this chapter. The first section presents the numerical set-up, which
is followed by an explanation of the spectral method in the second sub-section.

181
Table 7.1. Ship characteristics.
Quantity Symbol Value Unit

Scale Factor λ 75 -

Length L 3.067 m

Beam B 0.429 m

Draught T 0.144 m

Depth D 0.253 m

Water depth h 0.32 m

Block coefficient CB 0.651 -

Longitudinal Centre of Gravity LCG 1.488 m

Wetted area S 1.694 m2

Table 7.2. Test matrix and resultant Kelvin half-angles according to

Havelock's (1908) method.
Kelvin half-
Case Depth-to- Depth Froude Ship speed angle (°)
Canal
No draught (h/T) number (𝐹ℎ ) (m/s) (Havelock,
1908)

Rectangular 1 0.57 1.01 19.52

canal 2 0.77 1.364 21.58
2.2
New Suez 3 0.47 0.815 19.47
Canal 4 0.57 1.01 19.52

7.3.1 Numerical aspects

The solver, Star-CCM+ version 13.06, employs the finite volume method to
model the flow, which uses the integral form of the incompressible RANS
equations and divides the computational domain into a finite number of
adjoining cells. Continuity and momentum are linked via a predictor-corrector
approach. Further details pertaining to the implementation and algorithms
used can be accessed in Siemens (2018) and Ferziger and Peric (2002)

182
To account for turbulence within the fluid, the k–ω model due to (Wilcox, 2008)
is used. This choice is made following the findings of Chapter 4, which showed
the particular model to be stable and provide the fastest solution time of all
two-equation variants. Benefits of using the k–ω model include its seamless
application to low y+ type meshes (y+ < 1). This is a desirable feature, because it
avoids the use of wall functions, or any other bifurcations of the solution, as is
the case with the k–ε model (Siemens, 2018). Although wall functions can
predict the forces acting on a body with good accuracy, they may introduce
errors in the modelling of hydrodynamic properties in the wake of a ship. For
instance, they are unable to account for flow separation (Pettersson Reif et al.,
2009). To facilitate a good representation of turbulent properties, a second
order convection scheme is applied throughout all simulations.

To characterise the fluid interphase, the Volume of Fluid (VoF) method is used
(Hirt and Nichols, 1981). Moreover, Star-CCM+ offers a High Resolution
Interphase Capturing (HRIC) scheme to enhance the definition of the free
surface, which is applied to all numerical simulations (Muzaferija and Peric,
1999). Vertical ship movement, i.e. sinkage and trim, are not accounted for to
reduce the complexity of the simulations. Instead, the ship’s position in the x
– z plane is adjusted prior to initiating the simulation (by manually changing
the ship’s position). This is done in an attempt to reduce the discrepancy
between the experimental results, and those derived herein. However, some
difference is expected to persist since the experimental data, reported by
Elsherbiny et al. (2019b) was determined for a free to sink and trim KCS model.

7.3.1.1 Computational domain

As stated previously, frame invariance is not used. Instead, the ship is given
the corresponding velocity, which can be consulted in Table 7.2 for each canal.
To model the motion of the ship along the canal, the overset domain approach
is used. Thus, the ship is towed in the virtual environment over a static fluid.
This has two main consequences. Firstly, the computational domain can no
longer conform to the recommendations of the ITTC (2014) relating to the
positioning and dimensions of the computational boundaries. Instead, an
attempt is made to replicate the towing tank used for the experimental work,
used as a benchmark. Specifically, the Kelvin Hydrodynamics laboratory at
the University of Strathclyde. Naturally, the width and depth of the
computational domain must satisfy the test cases (given in Table 7.2). On the
other hand, the length of the computational domain is set as 60 m long. The

183
dimensions are kept the same across case studies (pertaining to the overset
domain and the length of the tank). These are shown in Figure 7.3. The height
of the static domain is set as 1.23 ship lengths from the undisturbed water
surface in all cases to eliminate any possible effects stemming from the height
of the domain.

Figure 7.3. Length of the computational domain and dimensions of the

overset domain.

The dimensions of the overset domain, which are maintained identical across
case studies are also shown in Figure 7.3. It should be noted that for
visualisation purposes, the figures have been mirrored about the central plane.
Other than the boundary, coincidental with the canal and ship centrelines,
where a symmetry condition is imposed, all other boundaries within the
background domain are no-slip walls. This is in line with the goal of designing
a more realistic representation of a towing tank. Specifically, so-called ‘open
boundaries’ do not exist in reality (Oberkampf and Blottner, 1998). Examples
of such boundaries include velocity inlets and pressure outlets. Although it is
easier to define the conditions at such boundaries mathematically, they are a
definite source of modelling error as discussed earlier.

The manner in which the computational domain is constructed allows the

removal of wave damping. Moreover, the definition of turbulent properties on
boundaries (such as levels of inlet turbulence) of the fluid is not necessary
since there are no inlets nor outlets present. However, it should be noted that
the initial conditions in terms of turbulence in the fluid follow the
recommendations of the software developers, namely that a turbulent

184
viscosity ratio of 10 should be used. This decays rapidly and is close to zero in
the non-disturbed region of the domain at the end of the acceleration phase.

7.3.1.2 Computational mesh

The computational mesh is generated entirely within the automatic facilities

of Star-CCM+. As stated earlier, the near-wall mesh is generated so that y+ < 1
over the wetted area of the ship. This is achieved via the prism layer mesher,
offered by Star-CCM+. The choice of background and overset mesh is of
critical importance. This must be done in a way that enables the solver to
adequately capture flow properties as they transition from the background
into the overset mesh. The cell distribution of each domain is depicted
graphically in Figure 7.4, whereas Figure 7.5 depicts the y+ distribution on the
hull at a physical time of 40s for the 𝐹ℎ =0.77 case.

Figure 7.4. 3D depiction of the computational mesh.

The arrangement of the mesh does not vary across case studies; the total cell
numbers for each canal are shown in Table 7.3. The circa 8 million cell
difference between the two adopted canals is a direct result of the smaller
wetted volume occupied by the Suez Canal.

Figure 7.4 also depicts the manner in which the mesh coarsens as the distance
from the waterline is increased. This gradual coarsening is implemented to
reduce the overall number of computational cells.

185
 Figure 7.5. The y+ distribution of the ship hull at 𝐹ℎ =0.77, sampled at 40s
physical time.

Table 7.3. Cell numbers for both canals.

Canal Background cells Overset cells Total

Rectangular canal 29,312,452 766,402 30,078,854

Suez Canal 21,496,179 766,402 22,262,581

7.3.1.3 Time-step selection

Time-step selection is of high importance in CFD. In this respect, the Courant–

Friedrichs–Lewy (CFL) number may be used as an assessment criterion. The
CFL number is defined as the product of the flow speed and time-step, divided
by the mesh size (Ferziger and Peric, 2002). As a fluid parcel propagates
through a mesh, one would ideally aim to capture its properties at each cell.
This is satisfied when CFL<1. Since the mesh is kept identical for all cases, the
highest speed can be used to assess the CFL condition. Moreover, a CFL
condition onto the background domain is not a meaningful metric, since the

186
majority of the fluid is static. Instead, the CFL number within the overset box
is monitored throughout the duration of the simulation.

Typically, when solid body motion is present, the time-step requirements are
relatively low. Here, a trial with a time-step of 0.0035L/𝑈, where L is the ship
length and 𝑈 is the ship speed in m/s was used. The results indicated good
agreement with experimental data, as will be demonstrated in the following
section. For this reason, the time-step is set at 0.0035L/𝑈 for all simulations. In
all RANS solution, there is a trade-off between computational resource
consumption and turn-around time. For the highest speed, the average CFL
within the overset domain did not exceed 0.7, which is considered adequate
for a first order temporal discretisation scheme. It should also be born in mind
that if the time-step is too low, severe numerical noise may be noticed in the
solution time-history (Yuan, 2019).

7.3.1.4 Time-history of the solution

An example time-history of the resistance of the ship is given in Figure 7.6.

The figure is characterised by two distinct regions. Firstly, the ship is
accelerated linearly up to the target velocity. This is done by linking the time-
step and velocity. The specific approach adopted requires the ship speed to
reach its steady value at the end of the first 1000 time-steps. In other words,
the ship’s velocity is increased by 𝑈𝑡𝑎𝑟𝑔𝑒𝑡 /1000 each time-step. Thus, after an
initial oscillatory behaviour, the resistance time-history exhibits oscillatory
convergence towards its steady-state value. The oscillations are liked with the
reflections of waves from the side walls, which also impact the observed
resistance (Yuan et al., 2018). All final values, reported in are obtained by
averaging over one period of oscillation of the resistance curve. The specific
point where averaging is performed is chosen to be sufficiently far from the
acceleration phase to eliminate its effect.

Figure 7.6 is characterised by a sharp cusp at the end of the acceleration phase.
This stems from the approach adopted to towing the ship through the domain.
Once the velocity of the ship has reached its final value, the simulation is
stopped, and the towing speed altered. It is thought that this change induces
a shock on the ship, which causes the observed cusp. This is likely not the
observed in physical towing. Nevertheless, the results, used for the present
analysis are taken a considerable time after this shock occurs. Therefore, the
results are not thought contaminated by this cusp.

187
 Figure 7.6. Example time-history of the solution (depicted: 𝐹ℎ =0.57).

7.3.1.5 Verification

This section contains the numerical verification of the case study in the
rectangular canal, 𝐹ℎ =0.57. The numerical uncertainties are estimated via the
Grid Convergence Index (GCI) method. This is the standard way to report
numerical uncertainties in ship CFD (ITTC, 2002). It is assumed that the
remaining cases are characterised by similar levels of numerical uncertainty.

The GCI method requires three systematically coarsened solutions for the
same case. The recommendations of the ITTC (2002) are followed. Specifically,
the refinement ratio, 𝑟 = √2 is adopted. The refinement ratio is used to coarsen
and lessen the grid size and time step, respectively. The GCI method assumes
that all three solutions are close to the asymptotic range, and are sufficiently
different, which may be difficult to achieve in practice. The proximity to the
asymptotic range is typically characterised by the convergence ratio, p, which
is shown in Eq. (7.1).

𝑝 = ln(𝜀32 ⁄𝜀21 ) / ln 𝑟 (7.1)

where 𝜀32 =𝑓3 – 𝑓2 , and 𝜀21 = 𝑓2 − 𝑓1 . Here, 𝑓𝑖 represents the ith solution,
generated by a systematic coarsening/lessening of the input parameter (grid
or time-step). If 𝑝 = 2, then the grid or time-step can be deemed asymptotic
(Roy, 2005). The convergence properties, however, can be determined in a
different manner, which also carries information on what type of
convergence/divergence is achieved with refinement. This is known as the
convergence ratio, R=𝜀21 /𝜀32 (Stern et al., 2006). Based on the value of R, the
following may be interpreted:

188
 • Monotonic convergence is observed if 0 < 𝑅 < 1
• Oscillatory convergence is observed if 𝑅 < 0
• Divergence is observed if 𝑅 > 1

The GCI method is only applicable in case (1). Next, an error estimate (𝜀21 ) is
defined as (Celik et al., 2008):

𝜀21 =(𝑓1 − 𝑓2 )/𝑓1 (7.2)

Once the error is known, the numerical uncertainty can be calculated as shown
in Eq. (7.3) (Roache, 1997):

GCI=1.25 𝜀21 /(𝑟𝑘𝑝 − 1) (7.3)

where k represents the kth input variable (grid or time-step). The factor 1.25 in
the numerator of the expression, defining the numerical uncertainty
represents a Factor of Safety. This has been devised to ensure that the true
solution lies within the bracket provided by the GCI with 95% confidence. The
results from the convergence study can be seen in Table 7.4

The successive grid coarsening resulted in 10,955,825 and 4,155,326 cells for
the medium and coarse solutions, respectively. In terms of spatial dependence,
the solution exhibited rapid ranges with reduction in cell numbers. This
‘superconvergence’ can be deduced by examining the order of accuracy, p.
While in the case of grid coarsening it is approximately 9, when the time-step
is lessened, the solution changes according to 𝑝𝑡𝑖𝑚𝑒 =0.463. According to Eca
and Hoekstra (2009), orders of accuracy between in the range 0.5 and 2 can still
be treated as asymptotic.

Eca and Hoekstra (2009) devised a procedure based on a least-squares fit to

estimate the numerical uncertainty. Their method is not employed here,
because it requires a minimum of four solutions. Further coarsening of the
computational mesh resulted in divergent behaviour in the simulation. The
consequence of the time-step exhibiting the above order of accuracy means
that the GCI method predicts large numerical uncertainties. That is, even
though the overall change between the coarse and fine solution is less than 3%
of the fine solution’s value.

It should be noted that the time-step as kept at the smallest value while
coarsening the grid. Conversely, the finest grid was maintained throughout
the temporal convergence analysis. To ensure that the ratio of overset cell to
background cell dimension is kept constant, both domains were coarsened

189
simultaneously. It is assumed that all examined cases will exhibit similar levels
of spatial and temporal dependence. For this reason, the above procedure is
not repeated.

Table 7.4. Grid and time independence (rectangular canal, 𝐹ℎ =0.57). EFD
result: 4.5047 N.
Parameter Mesh Time-step Units

r √2 √2 -

f1 4.325 (29,312,452 cells) 4.325 N

f2 4.356 (10,955,825 cells) 4.381 N

f3 5.045 (4,155,326 cells) 4.446 N

R 0.044 0.852 -

p 9.005 0.463 -

GCI (%) 0.6704 19.609 -

7.3.2 Spectral representation of the wave field

In this section, the method devised by Caplier et al. (2016) and Gomit et al.
(2014) is briefly examined. This method has previously been used to determine
a ship’s speed via satellite imagery (Arnold-Bos et al., 2007; Wu, 1992). The
essence of the approach is to process an available water surface in Fourier
space. To achieve this, a 2D Fourier transform is used to represent the
disturbance in the spectral space (𝑘𝑥 , 𝑘𝑦 ). These arise by defining the angular
wavenumber, k, as a vector containing x and y components, 𝑘 = √𝑘𝑥2 + 𝑘𝑦2 . The
spatial equivalent of these components are used to calculate the extents of the
spectrum. The x-direction length of the entire water surface (𝐿𝑥 ) becomes
𝑘𝑥,𝑚𝑎𝑥 , similarly, using the extent in the y-direction (𝐿𝑦 ) one obtains 𝑘𝑦,𝑚𝑎𝑥 , as
shown in Eq. (7.4). Likewise, the resolution of the water surface in real space
dictates the steps in Fourier space 𝛥𝑘𝑥 and 𝛥𝑘𝑦 in the x and y directions,
respectively (Eq. (7.5)).

𝑘𝑥,𝑚𝑎𝑥 = 2𝜋/𝐿𝑥 , 𝑘𝑦,𝑚𝑎𝑥 = 2𝜋/𝐿𝑦 (7.4)

1 2𝜋 1 2𝜋
Δ𝑘𝑥 = 2 × Δ𝑋 , Δ𝑘𝑦 = 2 × Δ𝑌 (7.5)

where ΔX and ΔY are the resolutions in the x and y directions, respectively.

190
If surface tension is ignored, the dispersion relation in shallow water may be
expressed via the angular frequency (ω) of the waves as shown in Eq. (7.6).
This is justified, because surface tension becomes important only in waves,
characterised by wavelengths smaller than 7cm. Alternatively, the travelling
disturbance should propagate with a speed higher than 0.23 m/s (Lighthill,
1990).

𝜔(𝑘) = ±√𝑔𝑘 tanh 𝑘ℎ (7.6)

A moving ship will cause the waves to be Doppler-shifted, and setting 𝜔′ =𝜔,
and 𝑘 ′ = 𝑘, the resulting dispersion relation becomes (Caplier et al., 2016):

𝜔′ (𝑘) = ±√𝑔𝑘 tanh 𝑘ℎ − 𝑈𝑘𝑥 (7.7)

To obtain the locus of the dispersion relation, Eq. (7.7) is solved for 𝜔′ (𝑘) = 0
(Carusotto and Rousseaux, 2013), which yields:

𝑈 2 𝑘𝑥2 − 𝑔√𝑘𝑥2 + 𝑘𝑦2 tanh(ℎ√𝑘𝑥2 + 𝑘𝑦2 ) = 0 (7.8)

Eq. (7.8) is symmetrical with respect to both axes (Crapper, 1964). This is
demonstrated in Figure 7.7, which includes to computed loci (this is used to
represent the solutions of Eq. (7.8)) for 𝐹ℎ =0.57, 0.77, according to the adopted
cases. Alongside these, the critical depth Froude number is depicted to
demonstrate the effect of ship speed on the dispersion relation in shallow
water. It should be noted that, the dispersion relation is speed independent in
deep water (Caplier et al., 2016). The arms of the loci always begin at
𝑘𝑦,𝑥 /(𝑔/𝑈 2 ) =1 in deep water. This is also the cut-off wavenumber. In shallow
waters on the other hand, the cut-off wavenumber varies with speed. This can
be seen by consulting Figure 7.7, specifically, where the loci cross the abscissa.
Here, the case for 𝐹ℎ =0.47 is not shown, as it is practically impossible to
distinguish it from the 𝐹ℎ =0.57 case. Deep and shallow water cases are
essentially identical when 𝑘ℎ ≫ 1.

The cut-off wavenumber separates the near-field disturbance from the far-
field waves, generated by the ship (Caplier et al., 2016). Thus, useful analysis
with applications to loads on coastal structures may be performed by
removing the near-field disturbance, which does not propagate away from the
ship. To determine the cut-off wavenumber (𝑘𝑥𝑐 ) in shallow water, Caplier et
al. (2016) solved Eq. (7.9).

𝑈 2 𝑘𝑥𝑐 − 𝑔𝑘𝑥𝑐 tanh(ℎ𝑘𝑥𝑐 ) = 0 (7.9)

191
Figure 7.7. Solutions to Eq. (7.8). Figure depicts the examined loci, alongside
the critical depth Froude number (𝐹ℎ =1) to demonstrate the effect of speed.

Finally, the Kelvin half-angle may be determined by computing tan 𝜃 =

−1
(𝑑𝑘𝑦 ⁄𝑑𝑘𝑥 ) at the inflection point. According to Nakos and Sclavounos
(1989), numerical errors will manifest near cut-off wavenumbers. This can be
deduced by examining Figure 7.7. Even a small deviation in the intersection
between the locus and the abscissa will lead to large errors. The specific
example given demonstrates the relatively low distance between the
intersection point of 𝐹ℎ =0.57 and 0.77. On the other hand, as the speed is
increased further, the locus approaches the origin. At the critical depth Froude
number, the locus will transition into crossing the origin and progressing into
a quadrant characterised by opposite signs of the abscissa and ordinate.

The manner in which a numerical free surface generates this pattern is in terms
of maxima of the spectrum in Fourier space. This can be extracted and
compared to the theoretical prediction, provided by Eq. (7.8). (Caplier et al.,
2016) demonstrated that the relationship holds well despite its neglect of non-
linear and three-dimensional terms. Thus, if one can prove that a numerically
generated free surface (once processed to the spectral domain) provides

192
maxima, near the locus, then the free surface can be considered validated. This
is explored in the following section. The analysis method can be summarised
as follows:

1. Using the required input in terms of ship geometry, speed and

underwater topography, solve the governing equations and obtain a
numerical free surface.
2. Compute the locus using Eq. (7.8).
3. Predict the location of the cut-off wave number using Eq. (7.9).
4. Export the numerical free surface and transform it using a Fast Fourier
Transform.
5. Filter the matrix, resulting from step 4 into two parts:
a. The far field wave components are given by locations of the
matrix, obtained during step 4 which lie in the region |𝑘𝑥 | > |𝑘𝑥𝑐 |
(high pass filter).
b. The near field wave components are given by locations of the
matrix obtained during step 4 which lie in the region |𝑘𝑥 | < |𝑘𝑥𝑐 |
(low pass filter).
6. Extract the coordinates (𝑘𝑥 , 𝑘𝑦 ) of the maxima of each column of the
matrix obtained in step 5.a.
7. Fit a curve through the points obtained in step 6 and compare with the
locus, computed in step 2.
8. Compute the inflection point of the curve fit obtained in step 7 with the
locus’ inflection point (obtained in step 2) by using tan 𝜃 =
−1
(𝑑𝑘𝑦 ⁄𝑑𝑘𝑥 ) , computed at the inflection point.
9. Represent each component of the free surface (near and far field waves)
by using an Inverse fast Fourier Transform on the components 5.a and
5.b.

7.4 Results and discussion

The presentation and analysis of the results are split into two major parts. The
first relates to the computed ship resistance coefficients, while the second sub-
section relates to the spectral analysis of ship waves.

7.4.1 Ship resistance

In this sub-section, the resistance, obtained numerically is briefly discussed

and compared to the experimental data. To begin with, the total resistance
coefficients are presented in Figure 7 for all cases. The experimental data of

193
Elsherbiny et al. (2019b) is included alongside each numerical result to enable
comparison. Figure 7.8 clearly indicates the numerical prediction has a well-
defined tendency to slightly underpredict the experimental data. Here, the
subscripts refer to depth Froude number. As one might expect, the ship’s
resistance at higher speeds becomes more challenging to predict by CFD. This
is evident, especially in the resistance characteristics at the highest depth
Froude number for each case.

The overall agreement between the experimental and numerical data is

encouraging. This is the first sign that the constructed towing tank is capable
of providing good predictions for the resistance of a ship. Possible sources of
discrepancy are suspected to stem from the fact that the numerical approach
did not model sinkage and trim. In shallow waters, their combined effect,
termed ship squat, is attributed greater relative importance than in deep
waters. Thus, the imperfect modelling adopted may have been partly the cause
of the observed levels of discrepancy between the experimental and numerical
results. Moreover, as the ship speed is increased, the difference also grows.
This matches the pattern observed in the sinkage and trim curves for a ship,
both theoretically (Tuck, 1967, 1966), as well as numerically (Jachowski, 2008).

In practice, the cell numbers used tend towards being prohibitively high. As
stated earlier, a virtual towing tank, where the principle of Galilean invariance
has been utilised will consist of no more than 2-3 million cells. Such
simulations can be performed within a few days on a standard computer.
Thus, the adopted approach of virtually towing the ship may not become
widespread soon. Nevertheless, the additional information that may be
extracted from a case such as this can be useful. An example of this is given in
the following sub-section.

194
 Figure 7.8. Comparison of total resistance coefficients for all cases (R
indicates the rectangular canal case, whereas S – the Suez Canal). Subscripts
refer to depth Froude number.

7.4.2 Spectral analysis of the numerical free surface

The highest speed is examined to begin with, simulated in the virtual towing
tank, 𝐹ℎ =0.77 in the rectangular canal. The numerical free surface is depicted
in Figure 7.9. Here, the far field waves, generated by the ship are clearly
visible. Due to the lateral restriction, waves have reflected approximately 2.5
ship lengths aft of the ship. It should be noted that Figure 7.9, and other figures
henceforth, are reflected around the central symmetry plane to enable a better
visualisation.

195
 Figure 7.9. Generated wave field in the rectangular canal at 𝐹ℎ =0.77.

At this point, it is useful to attempt to determine the wave angle. According to

Havelock's (1908) method, the Kelvin half-angle is 𝜃=21.58°. Figure 7.10
depicts an attempt at solving this problem by projecting a (dashed) straight
line form the forward perpendicular to the sides (and its reflection) at an angle
of 21.58°. The line ‘lands’ at a wave trough on the canal wall – clearly, this is
not the correct approach. The solid line in Figure 7.10 represents the same
process, but beginning from the nearest peak downstream at the wall, and
projecting in both directions. The line intersects the ship approximately ¼L
from the forward perpendicular. Then, the broken line is initiated at the
highest peak at the wall, where wave reflection occurs. This intersects the ship
approximately ¾L from the FP. Finally, the dotted line shows the same
process. It originates at the point [min(x), max(y)], representing the point
where the aft wave system is generated.

Figure 7.10. Generated wave field in the rectangular canal at 𝐹ℎ =0.77 and
corresponding half angle according to Havelock (1908). Dashed line
originates at FP, solid line originates at the nearest downstream peak, where
the dashed line is reflected. Broken line originates at the highest wave
elevation on the wall, dotted line originates at the ship coordinates
representing the point where the aft wave system is generated.

Clearly, neither line in Figure 7.10 accounts for the wave angle well. This is not
surprising since the method used to estimate the half angle is linear and
devised for a point source. In this case, the spectral representation may be used
to approach the problem. As explained earlier, the first step is to calculate the
Fourier transform of the wave field. This is performed in MATLAB, which
uses grayscale images. For this reason, the images used henceforth to represent
the free surface will be shown in the grayscale format, used to perform the
analysis. By doing so, other researchers may cross-reference results obtained
herein by analysing the provided images.

196
Figure 7.11. Processing of the wave field. Depicted: 𝐹ℎ =0.77 in the rectangular
canal. Top: the raw image – real space extents are 32m in the stream wise and
4.6m in the span wise directions. Bottom left: the Fourier representation of
the wave field. Bottom right: detected maxima (red points) and fit (dashed
line), superimposed onto the theoretical relationship (solid line), Eq. (7.8).

Figure 7.11 depicts the adopted method of analysis. Specifically, the free
surface is first represented in Fourier space. Then, for each column of the
matrix defining the Fourier transform, a maximum is identified. The 𝑘𝑥 , 𝑘𝑦
components of these maxima are then compared to the theoretical relationship
provided by Eq. (7.8). A polynomial fit is constructed from these points to
demonstrate the accuracy of large and small 𝑘𝑥 on the fit. In the present case,
it is apparent that as one progresses in the 𝑘𝑥 range to higher values,
agreement deteriorates quickly. According to Nakos and Sclavounos (1989),
insufficient grid resolution will be manifest as numerical dispersion in the 𝑘𝑥 ,
𝑘𝑦 plane being curved towards high values of 𝑘𝑥 , eventually forming closed
curves. Since this is not what is observed in the present study, it may be
concluded that the numerical wave field is represented with sufficient grid
resolution. Nonlinear phenomena will be revealed in the appearance of
additional branches in the spectrum. In each quadrant, a branch, emanating
from the origin and propagating linearly to the edge of the plot, where it
reflects, is observed. Moreover, smaller branches of the dispersion relation are

197
observed, with origins at higher kx values, indicating nonlinearity (Fourdrinoy
et al., 2019). These lead to deformation of the wake in the real space.

Now, it is important to deduce the origin of the apparent disagreement in the

high 𝑘𝑥 region, as well as its effect on the predicted Kelvin half-angle. One
source of disagreement inevitably stems from the fact that the dispersion
relation used for comparison is linear (Whitham, 2011). Ship waves,
particularly in shallow water are nonlinear. Even in deep waters, Ma et al.
(2018) found significant nonlinear influence on the dynamic pressure of the
KCS. On the other hand, the experimental data presented in Caplier et al.
(2016) suggest that the theory approximates real ship waves very well for the
Wigley hull. This may not be a fair comparison, because nonlinear effects for
the Wigley hull are known to be small (Chen et al., 2016; Ma et al., 2018; Wu et
al., 2019). In other words, using the parabolic hull plays to the method’s
strengths.

The neglect of nonlinear terms is chiefly manifest in the near-field disturbance,

close to the ship. Although a modification to the Kelvin half-angle may be
produced as a consequence of modified pressure in the near field, the
magnitude of such an effect is not known. Interference between of transverse
and divergent waves, generated at the ship’s bow may be one cause in the
observed disagreement (Noblesse et al., 2014). Such an effect, coupled with
nonlinearity exhibited by the KCS and influence of viscosity are thought to be
the dominant sources of discrepancy.

There is one more aspect of the solution one should consider carefully. This
relates to the curve fit used to approximate the numerical dispersion for higher
𝑘𝑥 values than maxima were detected for. In shallow waters, the arms of the
locus are typically not well developed (Caplier et al., 2016). Thus, it is difficult
to extract sufficiently many points to perform the analysis. For this reason, the
only fair assessment recommended is within the range where maxima have
been detected from the Fourier transform. The range, 𝑘𝑥 /(𝑔/𝑈 2 ) ∈ [1, 2.5], is
used to perform all subsequent analysis. This includes part of the fit over
which no maxima have been detected to illustrate the effect of limited data
samples.

198
 Figure 7.12. Derivatives 𝑑𝑘𝑦 /𝑑𝑘𝑥 for deep water, shallow water, and
numerical shallow water cases.

Figure 7.12 contains the derivatives 𝑑𝑘𝑦 /𝑑𝑘𝑥 for deep and shallow water based
on the dispersion relation, alongside the numerically generated fit from CFD.
This is shown because upon evaluating 𝑑𝑘𝑦 /𝑑𝑘𝑥 at the inflection point, the
Kelvin half-angle can be obtained. Moreover, Nakos and Sclavounos (1990)
recommend the examination of these derivatives to highlight differences
between numerical and theoretical dispersion relations. Figure 7.12 also
includes the area under each curve for reference. Clearly, assessing solely the
area under each curve is not a good approach to determine an apparent
disagreement, or error, which is -3.901% in this case. This is the case because
different parts of the kx range over which the derivative is shown may
attenuate or reinforce the total favourably. On the other hand, the RMS (Root-
Mean-Square) of the difference between the theoretical and numerical curve is
predicted as 0.456. The effect of this on the predicted half-angle is illustrated
in Figure 7.13, where the consequences of the previously examined differences
are highlighted.

199
 Figure 7.13. Predicted and theoretical half-angles.

The net effect of the difference between the fit and theoretical curves is
translated into a difference of approximately 2.6° in the predicted half-angle.
This is not considered as a substantial discrepancy. However, the location,
where inflection occurs is significantly different between the two sets of data.
This occurs at 𝑘𝑥,𝑡ℎ𝑒𝑜𝑟𝑦 /(𝑔/𝑈 2 )=1.06 according to the theoretical relationship,
whereas CFD predicts this at 𝑘𝑥,𝐶𝐹𝐷 /(𝑔/𝑈 2 ) ≈1.46. The identification of this
half-angle does not help in visualising the wake better. Plotting a line with
origins at the bow with an angle of 24.1° causes an intersection with the wall
earlier than what is shown in Figure 7.10.

There are several aspects of this technique that should be improved. Firstly,
the range, over which it is acceptable to find maxima of the Fourier transform
should be defined. The only way to accomplish this is via an extensive
experimental campaign. If such an interval is known, then it may be possible
to define a metric expressing the degree to which waves are correctly
modelled. It may also be possible to link specific parts of the computational
free surface with increased error in the representation of ship waves. The only
manner in which this can be achieved is via experimental work, which should
demonstrate the validity of the assumptions as they apply to waves generated
by a ship, rather than point sources. Specifically, ships causing highly
nonlinear flows, and thereby – waves.

200
 Figure 7.14. Splitting of near and far field components via manipulations of
the spectrum (cut-off wave number 𝑘𝑥𝑐 = 4.7885). Top: original free surface, (a)
indicates the far field component, whereas (b) indicates the near-field
disturbance and their corresponding Fourier representations. Longitudinal
extent: 32 m.

This section proceeds with the next aspect of the solution, which one may
obtain via the spectral representation. Specifically, splitting the near field from
the far field components. This is illustrated for the rectangular case study at
𝐹ℎ =0.77 in Figure 7.14, where the cut-off wave number is 𝑘𝑥𝑐 = 4.7885. Here,
the shape of the ship leaves a small effect onto the corresponding near and far
field systems because the outline of the vessel forms part of the free surface

201
itself. The intensity of the spectrum depends on the input, and can be changed
based on the brightness of the supplied input. The range of the spectrum is
therefore not shown.

The near field disturbance is not confined in the immediate vicinity of the ship
in Figure 7.14, contrary to expectations. Instead, it is shed from the ship
downstream, with its influence being clearly visible near the domain walls.
This representation also allows the detection of the far field waves, as well as
their reflection form the side walls with ease. It is apparent that the wave
system is convex with respect to the ship centreline. Once reflected, this is not
as clearly visible. The full spectrum for this case can be consulted in Figure
7.11.

Figure 7.15. Computed free surface in the rectangular canal, 𝐹ℎ =0.57. (a), far
field (b), and near field (c) representations in the real and spectral space
(𝑘𝑥𝑐 =9.5796). Longitudinal extent: 16.5 m.

Figure 7.15 shows the spectral decomposition process as applied to the

rectangular canal for 𝐹ℎ =0.57. Here, the arms of the spectrum, previously used
to extract maxima and compare with the theoretical relationship are not clearly
formed. This is consistent with findings of other researchers (Caplier et al.,
2016). Specifically, the higher the depth Froude number is, the more clearly the
arms of the spectrum are formed. The physical origins of this relate to the

202
relatively small far field disturbance generated by the ship in the examined
speed range. Simultaneously, speeds corresponding to 𝐹ℎ ≥ 1 are impractical,
which is why they have not been investigated. Experimental data in terms of
resistance is also not available for the adopted case studies at the
aforementioned speeds. In any case, several features of the spectrum can be
observed. Firstly, the low-intensity arms, observed in Figure 7.11 and Figure
7.14, extending into the far field are reproduced. However, in Figure 7.14, the
arms are not reflected from the boundaries of the plot, instead, they exhibit
periodic structures, which vanish near the limits.

Figure 7.16. Computed free surface in the Suez Canal, 𝐹ℎ =0.57. (a), far field
(b), and near field (c) representations in the real and spectral space
(𝑘𝑥𝑐 =9.5796). Longitudinal extent: 13 m.

It is now appropriate to shift the focus onto the Suez Canal and the spectral
representation of the wave field obtained. As before, the higher speed
(𝐹ℎ =0.57) is examined first, shown in Figure 7.16. An immediately apparent
difference relates to the structure of the wave field. Specifically, the slopped
canal banks have caused a rundown of the water surface. Since the theoretical
relationship used to plot the solid line in Figure 7.16a (top right) can only
account for a single depth, it is not seen to represent the Fourier representation
of the numerical wave field well. Interestingly, the spectrum contains maxima,
arranged in semi-circular arcs. An interpretation of this is not attempted at
present, instead, leaving this for a more theoretical piece of work. Such a

203
research would need to determine the form of the dispersion relation in non-
constant water depths. The cut-off wave number, used to produce Figure 7.15
and Figure 7.16 is same (𝑘𝑥𝑐 =9.5796 m-1), for this reason. As was the case for
𝐹ℎ =0.77, the near field disturbance is trapped near the canal walls, and over a
great distance downstream.

Figure 7.17. Computed free surface in the Suez Canal, 𝐹ℎ =0.47. (a), far field
(b), and near field (c) representations in the real and spectral space (𝑘𝑥𝑐 =
14.1833). Longitudinal extent: 20 m.

The lower speed investigated in the Suez Canal, and the Fourier
representation of its wave field is depicted in Figure 7.17. As expected, the
disturbance generated by the ship at 𝐹ℎ =0.47 is significantly smaller than that
produced at 𝐹ℎ =0.57 (Figure 7.16). The spectrum exhibits a similar structure
to what was previously observed for 𝐹ℎ =0.57. For both results shown in
Figure 7.16 and Figure 7.17, the near field hydrodynamic response has caused
bright parts of the spectrum, which are periodically broken. These correspond
to what is identified as a near field wave by the method, trapped at the lateral
extents of the tank. This is primarily the case due to the relative size of these
disturbances. Namely, their wavelength is of the order of magnitude of the
ship itself. Whether this classification itself is correct probably requires further
research. However, their effect onto the Fourier representation is clearly

204
visible, especially in Figure 7.16, where a high intensity patch can be seen
undulating along the ordinate.

The lower speed investigated in the Suez Canal, and the Fourier
representation of its wave field is depicted in Figure 7.18. As expected, the
disturbance generated by the ship at 𝐹ℎ =0.47 is significantly smaller than that
produced at 𝐹ℎ =0.57 (Figure 7.17). The spectrum exhibits a similar structure to
what was previously observed for 𝐹ℎ =0.57. For both results shown in Figure
7.16 and Figure 7.18, the near field hydrodynamic response has caused bright
parts of the spectrum, which are periodically broken. These correspond to
what is identified as a near field wave by the method, trapped at the lateral
extents of the tank. This is primarily the case due to the relative size of these
disturbances. Namely, their wavelength is of the order of magnitude of the
ship itself. Whether this classification itself is correct probably requires further
research. However, their effect onto the Fourier representation is clearly
visible, especially in Figure 7.17, where a high intensity patch can be seen
undulating along the ordinate.

Figure 7.18. Computed free surface in the Suez Canal, 𝐹ℎ =0.47. (a), far field
(b), and near field (c) representations in the real and spectral space (𝑘𝑥𝑐 =
14.1833). Longitudinal extent: 20 m.

205
 7.5 Summary and conclusion

This chapter presented a towing tank which does not reply on the principle of
Galilean relativity. This was accomplished via the overset method, which was
used to actually ‘tow’ the KCS model in a virtual environment. The main
benefits of adopting such an approach were identified in terms of the reduced
number of assumptions needed to perform the analysis. Specifically, all
boundaries (except those prescribed as symmetries or overset boundaries) are
no-slip walls, which is more physically consistent than ‘traditional‘ virtual
towing tanks. Since the ship advances over a static fluid, the approach
presented in this thesis does not require the definition of inlet turbulent
properties, as is usually the case. Thereby, removing one major source of
modelling error and uncertainty.

The adopted case studies numerically replicated recently published results in

a rectangular canal and in the New Suez Canal (Elsherbiny et al., 2019b). The
computed resistance of the ship was compared to the experimentally obtained
values. Good agreement was found, although some discrepancy persisted in
the highest speeds examined. The source of the difference between the
experimental and numerical results is primarily attributed to the fixed sinkage
and trim used in the virtual towing tank. The study was supplemented by a
method to decompose the wave field and determine the Kelvin wake angle. In
terms of the former, it was discovered that near field disturbances propagate
outwards towards the canal sides and are shed by the ship downstream. Their
effect persisted over a significant distance. This is of practical interest, because
near field disturbances are typically linked with strong pressure variations.
Thus, information extracted via the spectral decomposition method may be
used to assess the optimum slope and positioning of canal sides to avoid
excessive forces, linked with bank erosion.

On the other hand, it was shown that in a narrow canal, it is difficult to identify
the boundaries of the Kelvin wake. Values for the half-angle computed via
linear point-source methods were compared with those obtained by CFD. The
effects of nonlinearity and interference of wave systems shed by the bow and
stern were identified potential sources of discrepancy. However, the
numerical (𝜃𝐶𝐹𝐷 =24.1°) and theoretical (𝜃𝑇ℎ𝑒𝑜𝑟𝑦 =22.756°) Kelvin wake half-
angles for 𝐹ℎ =0.77 in the rectangular canal were found to compare reasonably
well. The inflection point, which governs the value of the half-angle however
was found to be in some disagreement. According to the theory, this should

206
occur at approximately 𝑘𝑥 =1.06, whereas CFD suggests the inflection point is
located at 𝑘𝑥 = 1.46. The potential sources of this discrepancy likely pertain to
limitations in terms of mesh size, time step in CFD, but also, and likely more
importantly, the theoretical assumptions.

207
 MODELLING THE
HYDRODYNAMIC EFFECT OF
ABRUPT WATER DEPTH CHANGES
ON A SHIP TRAVELLING IN
RESTRICTED WATERS USING CFD

This chapter examines the hydrodynamics of ships advancing past depth

discontinuities in restricted waters. Numerical simulations are performed
using two speeds and four different levels of water depth reduction. To gauge
the effects of transcritical flows, the case studies are specifically designed to
result in such conditions. The assessment suggests a strong influence in the
relative change of parameters of interest depending on the initial speed.

8.1 Introduction

Although there is a significant body of literature devoted to the study of ship

hydrodynamics in confined waters, several open research questions remain.
Contemporary interest in the field is driven by the fact that according to
EMSA (European Maritime Safety Agency, 2019, 2018, 2017, 2016, 2015), a
large proportion of all ship incidents occur in restricted waters. Although
human factors are predominantly thought to be the root cause of this, counter-
intuitive ship behaviour can occur in shallow waters, magnifying the overall
risk (Tuck, 1978). Therefore, understanding the hydrodynamic phenomena
occurring in shallow, and more generally, restricted waters, is of practical
relevance.

208
In some cases, the ship operator may not be aware of the consequences caused
by their operation in shallow water. An excellent example of this is described
and analysed at length by Grue (2017), who investigated waves caused by
ships sailing past abrupt depth transitions. He demonstrated that long waves
can be generated at the depth transition, which the author termed “mini-
tsunamis”. The resulting waves were shown to cause substantial damage to
coastal infrastructure several kilometres from their inception point. An even
more extreme case is reported to have caused the loss of life (Soomere, 2007).

The present chapter takes inspiration form Grue's (2017) work and seeks to
explore the associated effects further, using numerical methods. A survey of
the literature revealed that studies model similar problems using potential-
flow-based methods. These generally fall within three categories including
slender body methods (Gourlay, 2003; Plotkin, 1977, 1976; Tuck, 1967), Green
function-based methods (Yang et al., 2001; Yuan, 2014; Yuan et al., 2018; Yuan
and Incecik, 2016), methods based on the Boussinesq approach (Dam et al.,
2008; David et al., 2017; Grue, 2017; Jiang et al., 2002; Torsvik et al., 2006; Wu
and Wu, 1982), methods based on the Korteweg-de Vries equation (Cole, 1987;
Hur, 2019; Katsis and Akylas, 1987, 1984), and methods based on the
Kadomtsev-Petviashvili equation (Beji, 2018; Mathew and Akylas, 1990;
Sharma, 1995).

Most methods mentioned previously can be thought of as long wave theories.

The long wave family of theories can be arrived at by applying a combination
of assumptions and appropriate boundary conditions to the Euler equations,
which are known to model ship waves with adequate accuracy. However, as
is often the case, there is some disparity between different approaches as
illustrated by Torsvik (2009) in terms of dispersive properties exhibited by
generated waves. Neglecting viscosity may not be a valid assumption, based
on the findings of recent numerical and experimental studies, which observed
the formation of a boundary layer on the seabed in very shallow conditions
(Böttner et al., 2020; Shevchuk et al., 2016).

By contrast, studies have shown that the fully nonlinear Reynolds averaged
Navier-Stokes (RANS) equations can model the present class of problems well
(Bechthold and Kastens, 2020; Elsherbiny et al., 2020; Shevchuk et al., 2016;
Terziev et al., 2018; Tezdogan et al., 2016a). It is therefore prudent to attempt
to construct a fully nonlinear viscous towing tank with a varying bathymetry.
The specific object of this chapter is thus to simulate the hydrodynamic effects
caused by a ship passing over a step change in the water depth using CFD. To

209
the best of the author’s knowledge, such towing tanks have only been
constructed via the use of the family of long wave theories, described
previously.

According to Jiang et al. (2002), the unsteadiness and three-dimensionality of

the problem to be investigated herein precludes the use of many methods.
Even the applicable methods rely on the assumption of inviscid flow, which
may not hold for near-critical speeds or very shallow waters. Therefore,
provided one can cope with the computational effort, it is desirable to
investigate the effects of depth changes via a RANS method.

This chapter presents an attempt at modelling the above scenario using the
commercial RANS solver Star-CCM+, version 14.06. As a starting point, the
experimentally investigated rectangular canal of Elsherbiny et al. (2019) is
used. In their study, the authors of the aforementioned work performed an
experimental investigation into the hydrodynamics of the KCS containership
in a rectangular canal and the Suez Canal. Their findings indicate a strong
dependence of the canal cross-sectional area on all examined parameters.
Their study featured a depth-to-draught ratio of 2.2 and a width of 4.6m for
the rectangular canal.

To examine phenomena other methods may not be well-suited for, the chosen
speed range for the present work is trans-critical and is applied to the
rectangular canal of Elsherbiny et al. (2019). Specifically, two subcritical (𝐹ℎ <
1) depth Froude numbers are chosen to begin with. As the ship advances
through the domain, it encounters a step decrease in the water depth,
rendering a higher value of 𝐹ℎ . To model the longitudinal motion of the ship
along the canal, the overset domain approach is utilised, which follows from
the preceding chapter. To simplify the numerical simulations, the ship is not
allowed to move in any direction, other than along the canal’s length.
Therefore, ship squat is not modelled in this chapter.

The specific cases examined here are detailed in the following section. Section
8.3 presents the numerical implementation, which also contains estimates for
the numerical uncertainty. Then, the generated results and their analysis are
given in Section 8.4. Finally, Section 8.5 contains a summary and conclusion.

8.2 Case studies

This section is devoted to an overview of selected case studies. To begin with,

justification is given in terms of the selected conditions.

210
As stated previously, transitions past the critical depth Froude number are
sought. This is to demonstrate that RANS solvers are well-equipped to handle
such problems, that present significant difficulties for several theoretical
methods as reported in the literature. For example, the slender body theory
has a singularity at 𝐹ℎ = 1. Several studies have devised approaches to handle
the behaviour of the theory around this depth Froude number, but none has
become widely used (Alam and Mei, 2008; Gourlay and Tuck, 2001; Lea and
Feldman, 1972; Miles, 1986; Tuck, 1967, 1966). Therefore, the decision to
simulate conditions where the critical speed is met and exceeded was taken.
Alongside these, subcritical conditions are also modelled.

To achieve a high depth Froude number, a relatively high ship speed is

required. Although it may be the case that few vessels would operate under
such conditions, Grue (2017) reported fast ferries travelling at 𝐹ℎ = 0.7 prior to
the depth transition. Therefore, the high-speed choice is deemed both
acceptable and practically relevant for ship operations. Since ship waves are
of greater concern in restricted waterways due to bank erosion, such as rivers
and canals, a corresponding case study is sought.

As mentioned earlier, the work of Elsherbiny et al. (2019) is used as a starting

point for three reasons. Firstly, they investigated a canal case, which matches
the requirements set out previously. Secondly, the groundwork in
constructing and validating the wavefield of this case study was laid out in
chapter 7. Finally, replicating the towing conditions allows other researchers
to compare resistance and wavefield data against the work presented herein.
This could be done by other researchers using numerical methods.
Alternatively, an experimental version of the case studies presented here
should also be carried out.

Initial speeds are selected based on the highest available 𝐹ℎ explored by

Elsherbiny et al. (2019), namely, 𝐹ℎ𝑖 = 0.77. Since one of the objectives is to
model a critical case, the water depth restricted so that when the ship crosses
the step at a constant speed, the resulting depth Froude number, 𝐹ℎ𝑠 = 1.
Henceforth, the subscripts i and s will be used to denote the initial condition,
and the condition past the step, respectively. Naturally the same ship as used
in the work of Elsherbiny et al. (2019) is utilised, namely the KCS, whose
principal characteristics are shown in Table 8.1.

211
Table 8.1. KCS principal characteristics (in model scale).
Quantity Symbol Value Unit
Scale factor λ 75 -
Length L 3.067 m
Beam B 0.429 m
Draught T 0.144 m
Block coefficient CB 0.651 -
Longitudinal Centre of Gravity LCG 1.488 m
Wetted area S 1.694 m2

Making the above choice while maintaining a constant velocity requires the
depth to change from ℎ𝑖 /𝑇=2.2 to ℎ𝑠 /𝑇 ≈ 1.3 (note that the width is
maintained constant w=4.6m). Further cases to gauge the sensitivity of the flow
to water depth are specified at three equal intervals between the two
extremities in terms of the depth Froude number. Moreover, to include
additional supercritical cases, a second initial depth Froude number is selected
as 𝐹ℎ𝑖 = 0.9 and investigated for the same depths. The resulting test matrix is
given in Table 8.2. Further reductions in the water depth are not implemented
to avoid numerical problems in the implementation of the overset domain
approach, used to tow the ship. A schematic drawing of the different steps
investigated can be seen in Figure 8.1.

Table 8.2. Text matrix

No U (m/s) 𝐹ℎ𝑖 ℎ𝑖 (m) ℎ𝑖 /T 𝐹ℎ𝑠 ℎ𝑠 (m) ℎ𝑠 /T ℎ𝑖 /ℎ𝑠
1 1 0.190 1.304 1.687
2 0.943 0.214 1.468 1.500
1.364 0.77
3 0.885 0.242 1.665 1.321
4 0.826 0.277 1.905 1.155
0.32 2.2
5 1.169 0.190 1.304 1.687
6 1.102 0.214 1.468 1.500
1.595 0.9
7 1.034 0.242 1.665 1.321
8 0.967 0.277 1.905 1.155

Figure 8.1. Schematic drawing of the step changes in water depth. Not drawn
to scale.

212
 8.3 Numerical implementation

This section is devoted to the numerical implementation, with subsections

assigned to different aspects of the simulations.

8.3.1 The numerical environment

As stated in the earlier, the commercial solver, Star-CCM+, version 14.06 is

used. The solver is based on the Finite Volume Method (FVM). To avoid
deviating from the core topic of this chapter, specific details in terms of
algorithms and subroutines are not discussed. Instead, the reader is referred
to the user manual (Siemens, 2018) and texts containing detailed information
of the numerics used (Ferziger and Peric, 2002).

There are two main aspects of the solution in the present CFD modelling
requiring particular attention. The first relates to the definition of the free
surface. The disturbance caused by the ship may be significant and will
influence the solution substantially, therefore, an accurate representation is
necessary. The definition of the water surface is modelled via the Volume of
Fluid (VoF) method (Hirt and Nichols, 1981), with the High Resolution
Interphase Capturing (HRIC) scheme to enhance its sharpness (Muzaferija and
Peric, 1999, 1997). The grid on which the problem at hand is discretised is
discussed in the following subsections.

The second aspect of the numerical implementation that is of importance

relates to turbulence. The k-ω model of Wilcox (2008) is used, which has been
utilised to obtain consistently good predictions in terms of computational
resources and resistance predictions, as demonstrated in recent studies (Eca
and Hoekstra, 2008; Elsherbiny et al., 2020). The two-equation eddy-viscosity
turbulence closure is also selected due to its seamless application to all types
of meshes. This is an advantage because the model does not require
modifications independent on whether a wall function is used or not.
Although a low y+ mesh is constructed on the ship hull, as will be
demonstrated at a later stage, this is not the case for the canal sides and bottom.
Therefore, wall functions are used at all domain boundaries. Finally, to ensure
a good representation of turbulent properties, all simulations are run with a
second order accurate convection scheme.

To model the ship’s longitudinal motion along the domain, the overset domain
approach is used. In essence, this results in the creation of a box, enveloping
the hull. To re-create the ship’s motion in the x direction, the overset domain

213
is assigned the ship’s velocity, which appears in Table 8.2. Adopting the
overset domain allows the efficient modelling of the problem at hand. The
alternative would involve re-meshing at each time-step, which would cause
considerable increase in the computational load.

8.3.2 Computational domain and boundary conditions

The computational domain dimensions and arrangement are depicted in

Figure 8.2. The location labelled as “Step” represents where the water depth
transition is located. The values used can be consulted in Table 8.2 and Figure
8.1.

Figure 8.2. Depiction of the computational domain (depicted: ℎ𝑖 /ℎ𝑠 =1.687).

The main dimensions of both the background and overset domains are
unaltered across case studies. The only change stems from the difference in the
water depth after the midpoint of background domain. This is also set as the
global origin of the coordinate system to simplify the representation of the
results. To further simplify the results, the ship is modelled with an even keel,
and ship squat is not accounted for. The reason behind this relates to the
restricted water depth. If the overset domain were to collide with the
background domain, the simulation could fail or cause unreasonable results.
Since a shock is expected as the ship transits past the step, ship squat is not
modelled to avoid the aforementioned effects.

The modelling of the KCS with an even keel results in substantially different
resistance values to those recorded in the article by Elsherbiny et al. (2019),
precluding the possibility of a validation study. However, using an identical
set-up, the resistance and wavefield have been validated in Chapter 4 and
Chapter 7 (Elsherbiny et al., 2019a; Terziev et al., 2020, 2019b; Tezdogan et al.,
2016b, 2015).

The manner in which the numerical towing tank is constructed allows the
removal of all open boundary conditions. This carries positive and negative
impacts on the solution simultaneously. Specifically, the modelling

214
assumptions related to inlets and outlets, damping lengths and clearances
between an open boundary and the ship, inlet turbulence, etc. are no longer of
consequence because they no longer exist in the CFD simulations. Therefore,
the modelling assumptions and sources of uncertainty are significantly
reduced. On the other hand, from a mathematical point of view, open
boundaries are easier to implement. Their removal may destabilise the
solution in some cases. This is particularly the case when performing grid
refinement studies – if the grid is too coarse, the simulation diverges during
the early stages of the solution.

The numerical implementation of the domain requires in three types of

boundary conditions. Symmetry planes are instituted in the overset and
background domains, coincident with the centreline to reduce the
computational effort. The overset domain requires the appropriate boundaries
to imposed on the moving box, encasing the ship. All other boundaries are set
as no-slip walls, as would be the case in a physical towing tank. Therefore, the
numerical tank is physically consistent with real towing facilities.

8.3.3 Computational mesh

The computational mesh onto which the RANS equations are discretised is
generated within the automatic facilities of the software package utilised to
run the analysis. To ensure a good representation of the water surface, the
mesh used in Chapter 7 is duplicated, due to the fact that those results are
already validated. The prism layer mesher, offered by Star-CCM+ is used to
create near-wall cells at the ship hull, with the average y+ not exceeding 0.8 for
the highest speed examined. This is used to construct the near-wall cells on the
ship hull, responsible for accounting for the high velocity gradients within the
boundary layer. On the other hand, the y+ values on the side walls and bottom
are allowed to exceed 1, resulting in the use of wall functions on these
boundaries. A close-up of the generated grid on the undisturbed free surface
prior to initiating the simulation is shown in Figure 8.3.

215
 Figure 8.3. Close-up of the computational mesh on the free surface.

Table 8.3 contains the numbers of cells, generated for each case. It should be
noted that these do not vary with 𝐹ℎ𝑖 . Therefore, both 𝐹ℎ𝑖 = 0.77, 0.9 are
simulated with the same numbers of cells for their corresponding cases.

Table 8.3. Cell numbers for all four depth transitions

Cases ℎ𝑖 /ℎ𝑠 Number of cells
1 and 5 1.687 25,248,501
2 and 6 1.500 25,632,314
3 and 7 1.321 26,392,544
4 and 8 1.155 26,776,892

8.3.4 Time-step selection

Making an adequate choice for the time-step (Δ𝑡) in unsteady simulations is of

critical importance. If the Δ𝑡 value is too large, the numerical solution may
become unstable, or give unrealistic results. This is to be balanced with
computational time, i.e. it is not practical for a simulation to run over
unnecessary long periods of time while consuming computational resources.
Numerical noise may also manifest itself in the solution if the time-step is not
chosen correctly. Based on the results of Chapter 7, Δ𝑡 =0.0035𝐿/𝑈 is chosen
with a first order discretisation scheme. To ensure an adequate representation
of the physics, the Courant number is monitored throughout the simulation,
with a stopping criterion imposed to end the simulation if the Courant number
equals or exceeds unity within the overset domain.

The overset domain is chosen instead of the background to represent the

Courant number criteria because in the former, the majority of the flow is
accelerated. Had this been applied to the background domain, where the
majority of the fluid is static, the results in terms of the Courant number would

216
be misleadingly low. The generated grid succeeded in preserving a Courant
number, CFL>1.

8.3.5 Numerical verification

This subsection contains estimates of the numerical uncertainty, induced by

the discretisation of the RANS equations in time and space. The approach used
follows the recent work of Bechthold and Kastens (2020), who followed the
guide of Celik et al. (2008). This begins with the definition of a refinement
factor, r. According to ASME (American Society of Mechanical Engineers,
2009), acceptable values of r range between 1.1 and 1.5. The value of √2 is
adopted for the refinement ratio for the present set of simulations. This is
applied as a multiplicative factor to the mesh and time-step, which are
magnified successively, creating a medium (i=2) and coarse (i=3) solution for
each metric (mesh and time-step). To simplify the analysis, the refinement
ratio is kept constant, i.e. 𝑟21 = 𝑟32 = √2. Nevertheless, the relationships used
to perform the analysis with non-uniform r are adopted to enable comparison
of results with other studies.

Once the medium and coarse solutions have been obtained, the observed order
of accuracy can be determined as shown in Eq. (8.1):
|ln|𝜀32 ⁄𝜀21 |+𝑞(𝑝)|
𝑝= , (8.1)
ln(𝑟21 )

with
𝑝
𝑟 −𝑠
𝑞(𝑝) = ln (𝑟21
𝑝 ), (8.2)
−𝑠 32

and
𝜀
𝑠 = 𝑠𝑔𝑛 (𝜀32 ), (8.3)
21

where 𝜀32 = 𝑓3 − 𝑓2 , and 𝜀21 = 𝑓2 − 𝑓1 , with 𝑓𝑖 denoting the ith solution. For a
constant refinement case, the function 𝑞(𝑝) = 0. In the case of mesh
independence, the medium and coarse solution featured 8,684,955 and
3,167,970 cells, respectively.

The next step is to estimate the uncertainty, denoted GCI (Grid Convergence
Index), after Roache (1998), shown in Eq. (8.4):
𝑓1 −𝑓2 𝑝
𝐺𝐶𝐼 = 1.25 × | |⁄(𝑟21 − 1) (8.4)
𝑓1

217
In the present case, the resistance of the ship before and after the step change
in depth is used in the assessment. Therefore, two different estimates of the
uncertainty are obtained for the mesh and two for the time-step. The specific
case to which this is applied is case 1, 𝐹ℎ𝑖 = 0.77, 𝐹ℎ𝑠 = 1, as shown in Table
8.4.

Table 8.4. Numerical uncertainty study results (results are given for the ship
resistance).
Before step (𝐹ℎ𝑖 =0.77) After step (𝐹ℎ𝑠 =1)
Mesh Time-step Mesh Time-step
Fine (N) 19.74 19.74 64.43 64.43
Medium (N) 23.09 23.05 53.40 65.76
Coarse (N) 23.85 23.04 47.64 65.76
GCI 6.20% 0.03% 10.94% 0.01%
𝑈𝑐 0.06% 0.11%

In Table 8.4, the GCI value before and after the step is reported. Then, the
combined uncertainty, 𝑈𝑐 is estimated for each case as shown in Eq. (8.5):

2 2
𝑈𝑐 = √𝐺𝐶𝐼𝑚𝑒𝑠ℎ + 𝐺𝐶𝐼𝑡𝑖𝑚𝑒 (8.5)

It is apparent from the results of Table 8.4 that the results of the study are more
sensitive to variations in the mesh than they are to variations in the time-step.
Although the uncertainty exceeds 10% for the mesh in the critical region, this
is considered a tolerable level of uncertainty considering the challenging case
to which the analysis is applied to. The results from this subsection can be
interpreted as follows. The mesh requirements for critical depth Froude
number cases are considerably higher than those for subcritical cases. This
serves to highlight that the examination of critical speeds is not a trivial
problem even for RANS solvers. However, results with the given uncertainty
can be obtained, whereas many potential flow-based methods predict
singularities at 𝐹ℎ = 1.

Finally, it should be noted that while coarsening the grid, the finest time-step
was maintained. The mesh was magnified by the same factor (𝑟 = √2) in both
the overset and background domains to preserve the transitional ratio
between the two. Conversely, the temporal dependence study was carried out
on the finest mesh only.

218
 8.4 Results and discussion

This section presents the obtained results, their analysis and discussion. To
begin with, the resistance characteristics of the ship are given. For
convenience, the percentage increase in resistance as the water depth changes
is presented. This is thought to be a more suitable way to enable other
researchers to compare different hull forms subjected to similar conditions.
The initial resistance recorded for the model-scale ship at 𝐹ℎ𝑖 =0.77 was 19.74N,
while at 𝐹ℎ𝑖 =0.9 the value was 55.891 N.

Figure 8.4. Resistance increase resulting from the depth change.

Figure 8.4. demonstrates that the initial velocity is critically important for the
relative increase in resistance. To elaborate, the cases where the initial speed is
𝐹ℎ𝑖 = 0.77 exhibit several times the increase in resistance when compared to
the cases with initial speed corresponding to 𝐹ℎ𝑖 = 0.9. More importantly, no
jumps or sharp cusps are observed, as predicted by linear potential flow
theory. The results in Figure 8.4 also point towards the possibility to minimise
the total drag by small variations in the ship speed. For example, case 3
(𝐹ℎ𝑠 =0.885) is considerably more favourable than case 1 (𝐹ℎ𝑠 =1).

For both initial depth Froude numbers, the resistance does not peak at 𝐹ℎ = 1.
This is in line with experimental data for a family of hulls given in Benham et
al. (2020) and Benham et al. (2019), where the authors predicted that wave
resistance peaks well before the critical depth Froude number. The reason why
ship resistance is thought to exhibit a peak near 𝐹ℎ = 1 is strictly related to the
sustained generation of waves, which is widely considered an independent

219
component of the total resistance. Linear potential flow theories place the peak
at the critical depth Froude number; however, nonlinear phenomena are
known to deform the wave resistance curve’s peak towards the lower 𝐹ℎ range.
A similar effect can be observed in other experimental data, in terms of the
Kelvin wake angle (𝜃). For example, Johnson's (1957) experiments showed that
the peak in 𝜃 can occur at around 𝐹ℎ,𝜃=𝑚𝑎𝑥 =0.9. Therefore, the trend exhibited
by both curves in Figure 8.4 matches expectations.

In the cases where 𝐹ℎ𝑠 ≥ 1.1, one might expect to observe a reduction in the
resistance. This is a well-known phenomenon and has recently been
demonstrated by Benham et al. (2020, 2019). There are two possible
explanations as to why this is not observed in the presented results, which
suggest the resistance increases in each case. One way to look at the data
would be to suggest that the expected decrease in resistance is too narrow over
the depth Froude number range. In such a scenario, the dip would be observed
if further simulations were carried out between 1.09 < 𝐹ℎ𝑠 < 1.15. Evidence to
suggest that this may be the case can be found in Benham et al. (2019). The
decrease in resistance for their hull forms is narrow. It is also worth noting that
in their study, a family of simple hull shapes were examined. Similar hulls, for
example, the Wigley hull, are known to produce a predominantly linear flow
field (Chen et al., 2016).

Figure 8.5. Wavefield for case 1, 𝐹ℎ𝑖 =0.77 and 𝐹ℎ𝑠 =1. The solution time and
increment interval unit at which the free surface is shown is based on the end

220
 of the acceleration phase (shown in the first tile). The dashed line indicates
the position of the step change in water depth.

A second way of interpreting these results may be in terms of the wavefield.

When the depth Froude number past the step exceeds or is equal to unity, any
solitons generated by the ship may not be shed sufficiently quickly, to the
forward part of the tank. Therefore, the ship may be trapped sailing against a
wave. One way to determine whether this is the case would be to examine the
wave field. This is shown in Figure 8.6 and Figure 8.6 for cases 1 and 7,
respectively (𝐹ℎ𝑖 = 0.77 with 𝐹ℎ𝑠 = 1, and 𝐹ℎ𝑖 = 0.9 with 𝐹ℎ𝑠 = 1.03,
respectively).

Figure 8.6. Wavefield for case 7, 𝐹ℎ𝑖 =0.9 and 𝐹ℎ𝑠 = 1.03. The solution time
and increment interval unit at which the free surface is shown is based on the
end of the acceleration phase (shown in the first tile). The dashed line
indicates the position of the step change in water depth.

The wave elevation carried in front of the ship is longer for 𝐹ℎ𝑖 =0.9 and
𝐹ℎ𝑠 =1.03 (case 7). Thus, it is conceivable that this is could be the root cause for
the elevated resistance. The second main difference between Figure 8.5 and
Figure 8.6 relates to the generated soliton. While the soliton is clearly visible
in Figure 8.5, it has not detached from the bow wave elevation in Figure 8.6.
This is the case because the ship speed is close to the wave speed (𝐹ℎ𝑖 =0.9),

221
therefore, the soliton would require a much longer domain to be properly
shed. Solitons are also known to be essentially two-dimensional (Gourlay,
2001). To check whether the wavefield is 2D on the step, the time-history of
the wave elevation on the step is recorded and shown in Figure 8.7. The free
surface is monitored at four points next to the ship to provide a picture of the
generated disturbance with distance in the y direction. Since the towing tank
is symmetrical about the centreline, the probes are not mirrored. For
comparison, the time-history of the same probes shown in Figure 8.7 for
𝐹ℎ𝑖 =0.77, are shown for 𝐹ℎ𝑠 =0.9 in Figure 8.8. These demonstrate that the
wavefield in the latter case is 2D to a much greater extent and that the change
in step height (ℎ𝑠 ) has a smaller relative influence on the deformations of the
free surface.

Figure 8.7. Wave probes at the step for 𝐹ℎ𝑖 =0.77.

222
 Figure 8.8. Wave probes at the step for 𝐹ℎ𝑖 =0.9.
Figure 8.7 indicates that the wave field is only uniform along the y axis prior
to the ship’s interaction with the step. This can be seen by referring to the wave
elevation between 20 and 30 seconds of physical time in Figure 8.7 and Figure
8.8. The interactions differ with 𝐹ℎ𝑠 due to the physics of wave reflection and
transition from a submerged step. Since the wave speed in the deeper region
is higher than that of the shallower region, past the step, the wave profile must
transform upon transiting from one depth to the other. There have been many
studies into how this occurs. The first such work is thought to be that of Lamb
(1932), who derived an expression for the ratio of transmitted and incident
waves. His assumption of zero vertical velocity at the step seemed
inappropriate to Bartholomeusz (1958), who presented a more in-depth study.
However, the end result was identical to Lamb's (1932). Later, both of these
studies were put under question by Newman (1965), who also ended up with
Lamb's (1932) formulation for very shallow water cases.

The contribution of Newman (1965) however was expressed in the fact that he
obtained an expression for an infinitely deep incident wave transforming into
a shallow region. He provided a physical interpretation as to why a
transmitted wave asymptotically tends to a wave height 𝜁𝑠 =2× 𝜁𝑖 (where 𝜁 is
the wave elevation, whereas the subscripts maintain their earlier designation)

223
as ℎ𝑠 → 0. Newman's (1965) interpretation is that as the shallower region’s
depth vanishes, two phenomena occur. Firstly, the entire incident wave’s
amplitude is reflected, which is physically consistent. Secondly, that as the
wave transits to much shallower regions, the energy transmitted into the
region of depth ℎ𝑠 reduces at a rate proportional to the that depth (ℎ𝑠 ), causing
the transmitted component to be twice the incident wave’s height. Newman
(1965) then presents experimental results, which show that the theory is
consistent, although some scatter in the tank data is observed around the
theory. To check whether the physical phenomena occur in agreement with
the aforementioned studies, the wave elevation is recorded along the entire
tank at different times.

To simplify the discussion of the results, the obtained wavecuts are split into
different phases of the simulation as follows (Figure 8.6 and Figure 8.6 can also
be consulted in this respect):

I. End of acceleration – this occurs when the ship has reached its target
speed. The wavefield at this stage is not yet fully developed and differs
from its pseudo-steady state in several important ways. Discussion of
these can be found in Doctors (1975) and Day et al. (2009).
II. Subcritical wavefield development – this phase of the simulation
contains the time required by the wavefield to approach its steady state.
This process occurs in all towing tanks, whether virtual or physical. In
numerical tanks where the ship’s position in the x direction is
maintained constant, this phase is equivalent to the time allowed for
convergence.
III. Prior to the step – at this stage, the ship begins to interact with the step.
Initially, this is indirectly via the bow wave, which is partially
compressed by the additional blockage.
IV. Transiting the step – this phase occurs while the step is located under
the ship itself.
V. After the step – this phase begins as soon as the stern of the ship has
cleared the step. Interactions between the depth transition and the step
do not cease here. Instead, the accelerated fluid aft of the ship, interacts
with the step continuously for a considerable time. This effect is
subsequently demonstrated.
VI. Critical wavefield development – once the ship has cleared the step
and advanced about one ship length along the canal, the wavefield

224
 corresponding to the depth ℎ𝑠 has begun developing. This can be
thought of in similar terms as explained in II, i.e. a convergence stage.

The development of the wavefield is split into the above stages and given in
Figure 8.9 and Figure 8.10 for case 1 (𝐹ℎ𝑖 =0.77, 𝐹ℎ𝑠 =1) along y/w=0.1 and y/w
=0.2. In these figures, the maximum and minimum wave elevation for each
phase are recorded and marked. Evidently, for y/w =0.2, shown in Figure 8.10,
the disturbance caused by the ship decays in the y direction, allowing the
soliton to assume the maximum value in phases II, III, and IV. For y/w =0.1
(Figure 8.9), this is only the case in phase II.

An interesting property, observed in Figure 8.9 and Figure 8.10 relates to the
difference of the wave field aft of the ship in phases III and IV (prior and in
transit of the step) when compared to phase V (after the step). The oscillatory
pattern observed in the earlier phases, corresponding to the Kelvin wake (refer
to Figure 8.6 for a top view) is transformed as the ship enters the region of
depth ℎ𝑠 . The oscillatory pattern are replaced by a substantial depression,
following the ship, as evident in the final stage, given in Figure 8.9 and Figure
8.10 along each wavecut. The length and height of the wave, trapped at the
ship’s stern is seen to decrease substantially, while the bow wave is
considerably stronger in both respects. This observation partly explains the
increase in resistance, shown in Figure 8.4. Namely, the ship carries with it a
greater volume of water at its bow.

225
 Figure 8.9. Wavecut 1 (y/w=0.1) evolution for case 𝐹ℎ𝑖 =0.77, 𝐹ℎ𝑠 =1, made
dimensionless by the initial depth ℎ𝑖 =0.32. Maxima and minima are marked
with green and red points, respectively. The outline of the ship is scaled
down in the vertical direction by a factor of 8 to enable a visualisation of the
ship’s position.

In investigating the wavefield, it is important to keep in mind that the RANS

solver models the flow in a fully nonlinear manner. In Chapter 7, it was
demonstrated that the present set-up models the dispersive properties of the
Kelvin wake in good agreement with the linear dispersion relation. In this
chapter the opportunity to compare the evolution of the numerical wavefield
with linear potential flow approximations of wave transmission past different
steps, is used to provide a form of validation.

As stated earlier, Lamb (1932), Bartholomeusz (1958), and Newman (1965) all
arrived at the same relationship describing the transmission coefficient,
expressed as shown in Eq. (8.6):
2√ℎ𝑖
𝑇𝑅 = (8.6)
√ℎ𝑖 +√ℎ𝑠

226
where TR is the ratio of transmitted and incident wave height. This relationship
follows directly from the wave speed in each shallow water region, whose
linear form is √𝑔ℎ, with 𝑔 being the gravitational acceleration. Lamb (1932)
showed that this relationship can be arrived at simply by imposing continuity
and equivalence of the two waves (transmitted and reflected component) at
the point directly above the step. All possible values of TR are shown in Figure
8.11 alongside the numerical predictions for cases 1-4 using this method. Here,
the theoretical predictions are marked along each line representing the
possible coefficient values to enable a better visualisation of the numerical
results and their deviation from the theory.

Figure 8.10. Wavecut 2 (y/w=0.2) evolution for case 𝐹ℎ𝑖 =0.77, 𝐹ℎ𝑠 =1, made
dimensionless by the initial depth ℎ𝑖 =0.32. Maxima and minima are marked
with green and red points, respectively. The outline of the ship is scaled
down in the vertical direction by a factor of 8 to enable a visualisation of the
ship’s position.

227
 Figure 8.11. Transition and reflection coefficients for cases 1-4 (𝐹ℎ𝑖 =0.77).

Figure 8.11 also contains the reflected coefficient values for the entire range of
depth ratios. In this case, no numerical predictions are given because a
reflected wave elevation is not observed. The suspected cause of this is the
ship’s interaction with the wavefield. To elaborate, any reflected wave will be
disturbed almost as soon as it is created by the passing of the ship. On the other
hand, the numerical results for TR (transmission coefficient) show good
agreement with the theoretical line. In fact, the present datapoints are less
scattered than the experimental results shown in Newman (1965). It should be
noted that in the aforementioned work, the author used the infinitely deep
initial region theory to construct his line. Nevertheless, he demonstrated that
the experimental data are scattered around the line, providing a form of
validation for the present wavefield.

The reason why several closely positioned datapoints are shown in Figure 8.11
relates to the manner in which the wavefield is sampled. Specifically, Figure
8.9 and Figure 8.10 show that one has a range of choices when it comes to
taking the incident wave height and transmitted wave height. Therefore, it is
thought important to demonstrate that this particular choice is of little
importance on the positioning of the calculated transmission coefficients. For
this reason, only one datapoint is given for cases II ~ IV. To further investigate
the significance of the location over which the soliton is taken, the wavefield

228
maximum and minimum along the wavecuts given in Figure 8.9 and Figure
8.10 are shown in Figure 8.12.

Figure 8.12. Analysis of the wavefield for 𝐹ℎ𝑖 =0.77, 𝐹ℎ𝑠 =1 along wavecuts
y/w=0.1 and y/w=0.2

Here, a simple relation was used to predict the transmitted component of the
soliton, which was shown to agree well with the numerical results. However,
it also important to state that following the work of Newman (1965), Lamb
(1932), and Bartholomeusz (1958), considerable developments have been
made in the field of predicting the behaviour of waves at depth transitions.
Some recent studies include Bender and Dean, (2003), and discussions thereof
(Bender and Dean, 2005; Liu and Lin, 2005), which can give the reader a much
more in-depth appreciation of the complexities encountered in the related
field. Many methods, such as that of Marshall and Naghdi (1990) use the
wavenumber, k, in each region to arrive at an expression for TR.

229
To estimate the wavenumber, Lee and Lee (2019) give several methods. Guo
(2002), and Newman (1990) can be consulted as well. In this chapter, the
approaches of Hunt (1979) and Havelock (1908) are used. These techniques
agree well for the examined range (𝐹ℎ𝑖 =0.77), as shown in Figure 8.13.

Figure 8.13. Predicted wavenumber (𝑘𝑖 ) for each depth Froude number

Clearly, one runs into problems using these methods when predicting the
wavenumber in the region past the step if 𝐹ℎ𝑠 =1. In this respect, Marshall and
Naghdi (1990) proposed the following relationships between the two regions’
wavenumbers:

𝑘𝑖 tanh(𝑘𝑖 ℎ𝑖 ) = 𝑘𝑠1 tanh (𝑘𝑠1 ℎ𝑠 ) (8.7)

ℎ𝑖 𝑘𝑖2
𝑘𝑠2 = √ℎ 2 2 2 (8.8)
𝑠 (1+𝑘𝑖 ℎ𝑖 ⁄3−𝑘𝑖 ℎ𝑖 ℎ𝑠 ⁄3)

𝑘𝑠3 = 𝑘𝑖 √ℎ𝑖 ⁄ℎ𝑠 (8.9)

Then, the reflected and transmitted and coefficients are given in Eq. (8.10) and
Eq. (8.11), respectively:
𝑘 −𝑘
𝑅𝑅 = 𝑘 𝑠 +𝑘𝑖 (8.10)
2 𝑖

230
 2𝑘𝑠
𝑇𝑅 = 𝑘 (8.11)
𝑠 +𝑘𝑖

In Eq. (8.7) – Eq. (8.9), the subscripts 1, 2, 3 are used to differentiate the
wavenumber predictions. To examine the predictions graphically, Figure 8.14
was constructed showing the relationships between 𝑘𝑠1−3 and the
wavenumber, as predicted by Hunt's (1979) method.

Figure 8.14. Wavenumber predictions for the region past the step. Depicted:
case 1, 𝐹ℎ𝑖 =0.77 and 𝐹ℎ𝑠 =1.

As 𝑘𝑠2 increases past a value of 100, it essentially ceases to grow. Fortunately,

the values of interest are far from this boundary. The predicted wave numbers
using the aforementioned methods are summarised in Table 8.5.

231
Table 8.5. Summary of wavenumbers for cases 1 ~ 4 (𝐹ℎ𝑖 =0.77).
Method Description Wavenumber value
Havelock (1908) 7.591
Hunt (1979) Initial region 7.367
Relative difference -3.04%
𝐹ℎ𝑠 =1 7.974
𝐹ℎ𝑠 =0.943 7.777
Eq. (8.7)
𝐹ℎ𝑠 =0.885 7.608
𝐹ℎ𝑠 =0.826 7.470
𝐹ℎ𝑖 =0.77 𝐹ℎ𝑠 =1 7.224
𝐹ℎ𝑠 =0.943 7.093
Eq. (8.8)
𝐹ℎ𝑠 =0.885 7.031
𝐹ℎ𝑠 =0.826 7.085
𝐹ℎ𝑠 =1 5.672
𝐹ℎ𝑠 =0.943 6.018
Eq. (8.9)
𝐹ℎ𝑠 =0.885 6.409
𝐹ℎ𝑠 =0.826 6.855

Figure 8.15. Transmission and reflection coefficients based on Marshall and

Naghdi's (1990) method (𝐹ℎ𝑖 =0.77).

232
Since Marshall and Naghdi (1990) derived Eq. (8.7) without additional
assumptions, it is used to construct transmission coefficients shown in Figure
8.15 for cases 1~4 (𝐹ℎ𝑖 =0.77). Here, the range has been retained from Figure
8.11 for consistency. Figure 8.15 demonstrates that the much simpler approach
given in Figure 8.11 can be used to quickly estimate the values, with essentially
the same accuracy in the transmission coefficients. This is valid due to the
particular case studies selected. Had a deeper water region been chosen, the
approach of Marshall and Naghdi (1990) would have been recommended.

At this stage it is also worthwhile to mention that the decrease in height of the
soliton with distance, given in Figure 8.12 is also physically sound. As stated
earlier, the height of the wave in front of the ship was shown to decay. This
was given as a justification to the question why, there are several datapoints
for 𝐹ℎ𝑠 =1 in terms of transmission coefficients. The dissipation of the wave is
due to a combination of viscous action in the fluid and friction at the edges
and bottom of the tank. Here, the reader is reminded that all boundary
conditions, with the exception of the symmetry plane and overset box are no-
slip walls. Therefore, the dissipation of the wave is an expected outcome.

The dissipation in the present context can be approximated as shown in Figure

8.16. Here, Eq. (8.12) and Eq. (8.13) are used to construct the dissipation and
amplitude-change parts of the plot, as reported in Lamb (1932). Further
discussion of these equations can be found in the relevant literature (Denner
et al., 2017; Dorn, 1966; Hunt, 1964; Keulegan, 1948; Liang and Chen, 2019).

2𝑘 𝜈 𝑘𝑤+sinh (2𝑘ℎ)
𝐷= √2𝜔 2𝑘ℎ+sinh (2𝑘ℎ) (8.12)
𝑏

𝛿 = 𝜁𝑒 −𝐷𝑥 (8.13)

where 𝐷 is the dissipation, 𝜔2 = 𝑔𝑘 tanh(𝑘ℎ), 𝜈 is the kinematic viscosity, 𝛿 is

the amplitude of the wave having travelled x metres, and 𝜁 is the elevation of
the initial wave. It should be noted that 𝑘𝑠1 , as given in Table 8.5 and Eq. (8.7)
are used throughout for consistency.

233
 Figure 8.16. Viscous dissipation on a unit wave with different dispersive
properties travelling a unit distance.

In the present set of simulations, it is not immediately obvious where the

origins of the soliton lie. It is therefore difficult to determine an exact damped
amplitude. Thus, tracing the soliton’s decay continuously and comparing it
with the analytical solution is not attempted. Instead, the value of the soliton
after it has transferred onto the step is used. Then, the damping relations given
in Eq. (8.12) and Eq. (8.13) are applied to arrive at a decrease in magnitude by
2.923%. Similarly, the numerical result is that the soliton reduced in magnitude
by 2.994% between phases III (prior to the step) to VI (critical wavefield
development). Note that the soliton has already cleared the step at phase III.
This agreement is excellent and indicates that the present solution has
captured the physics of wave propagation and transmission well, as indicated
by the results given here and in Figure 8.11 and Figure 8.15.

The numerical modelling of the wavefield, including the friction on the side
walls and bottom was shown to agree well with analytical solutions.
Therefore, the use of wall functions at the boundaries of the tank is shown to
provide sufficient accuracy. This result also suggests that numerical diffusion,
incurred by the grid density is minimal for the soliton. To check whether the
solitary wave propagates with a speed, consistent with theoretical predictions,
the formulation given in Ertekin et al. (1986), Eq. (8.14) is used.

234
𝐶 = √𝑔ℎ × (1 + 𝜁/ℎ) (8.14)

where 𝐶 is the solitary wave speed in a region of depth h. The formulation is

used for case 1 to obtain 𝐶 = 1.44 m/s, which compares well with the
numerically observed 1.46 m/s, exhibiting a relative difference of 1.71%. In
this calculation, the average 𝜁 value is used form the final two measurements,
given in Figure 8.9 and Figure 8.12.

Another aspect of the generated data considered here relates to the velocity
field produced by the soliton, and the ship and its wave system as they interact
with the step. As was the case earlier, focus is placed on the critical transition
case (case 1, 𝐹ℎ𝑖 =0.77 and 𝐹ℎ𝑠 =1). To examine the velocity field, the
proportion of the domain beneath the undisturbed free surface is plotted at
various times in Figure 8.17. The symmetry plane is used as a reference in this
case throughout. Longitudinally, one ship length before and after the step is
included in the plots.

Figure 8.17. Generated velocity field for case 1, 𝐹ℎ𝑖 =0.77 and 𝐹ℎ𝑠 =1 as the
ship and soliton interact with the step.

235
The first part of Figure 8.17 shows the process of the soliton as it has cleared
the step. Here, the step causes a region of elevated velocity magnitude to
persist even after the soliton has propagated a full ship length past the depth
reduction. This would likely persist for a long time had the ship not disturbed
the flow field. The most surprising aspect of the solution is that a significant
proportion of the fluid maintains its velocity a significant time after the ship
has cleared the step. The flow field in the final part of Figure 8.17 is in the
direction of the ship, with a vortex persisting at the step’s edge.

The observations made in Figure 8.17 are in agreement with recent numerical
and experimental work, which demonstrated that a boundary layer will form
on the seabed in very shallow conditions (Böttner et al., 2020; Shevchuk et al.,
2016). In this Chapter it is also demonstrated that this effect persists in the final
part of Figure 8.17. However, as mentioned earlier, the domain does not
feature any inlets or outlets. Therefore, any fluid, accelerated in the ship’s
direction must return to equalise the pressure and water elevation behind the
ship. Ship-generated waves would also contribute to this, but they diffuse by
a combination of sides/bottom friction, numerical and viscous dissipation.
Indeed, the smaller the waves, the greater the action of numerical dissipation.

To investigate the mechanism by which the fluid returns after being

accelerated in the direction of the ship, the velocity field at the step is taken at
the end of the simulation (i.e. once the ship’s bow is about 1 m from the end
boundary) and shown in Figure 8.18. Although the magnitude of velocity is
given in the contour of the plot, the x-direction velocity is also shown as a
vector field. This reveals that fluid is returning at a palpable rate to the step.
However, once the step is cleared, the velocity diminishes rapidly. Such an
effect may explain the elevated time-history of the free surface in Figure 8.7
and Figure 8.8. Specifically, these indicate that considerable volumes of water
have been swept along with the ship and are subsequently returning to
equilibrium.

Figure 8.18. Example velocity field near the step at the end of the simulation
for case 1: 𝐹ℎ𝑖 =0.77, 𝐹ℎ𝑠 =1.

236
A final consideration is given to the wavecuts shown in Figure 8.9 and Figure
8.10. Although difficult to spot, in phase VI (critical wavefield development),
a wave trough can be seen propagating in the direction, opposite that of the
ship’s. This is not a purely numerical phenomenon. Grue's (2017) theory and
observations in a Norwegian fjord demonstrated that this is a key feature of
depth transitions. In the case depicted in Figure 8.9 and Figure 8.10, the
depression is of approximately –40cm height when converted to full-scale.
This consequence of depth transitions is clearly an important part of the
physics and has the potential to cause severe infrastructure damage.

8.5 Conclusion and summary

This chapter examined the impact of a step change in the water depth on ship
performance using the commercial RANS solver Star-CCM+. The adopted case
studies reflected a narrow canal, along the length of which, the water depth
changes abruptly. This scenario was modelled to reflect recent work
performed experimentally (Elsherbiny et al., 2019c) over a constant canal
cross-section, and subsequently validated in Chapter 7. Four depth reductions
were modelled with two constant speeds. These were deliberately selected to
provide transcritical depth Froude numbers, with one case targeted
specifically at the critical speed.

The results were reported in terms of percentage increase of the base resistance
encountered by the ship prior to, and after the step. These indicated that a
resistance increase of up to approximately 226% may occur if the transition
results in subcritical depth Froude numbers approaching the critical speed. On
the other hand, it was shown that when the ship has a high initial speed, close
to the sub-supercritical boundary, the resistance increase is considerably
milder. Results indicated in this case a change of less than 50% for the model
scale ship.

In cases where the initial speed is near 𝐹ℎ =1, it was shown that the wavefield
is largely two-dimensional and unaffected by the height of the step. On the
other hand, when the speed is lower, the interaction of the ship’s wavefield
with the depth transition is highly sensitive to the step height. For this reason,
the transition of waves past shallow water depth discontinuities were
examined in some detail. Two theories with varying complexity were
employed, both of which agreed well with the numerical results. This
suggested that the shallow water celerity approximation √𝑔ℎ holds well even

237
when the critical depth Froude number is reached in water depths of up to
ℎ⁄𝑇 =2.2.

Simultaneously, the damping of the soliton was shown to be in excellent

agreement with analytical relations for viscous dissipation of waves in canals.
The theoretical result for damping in the amplitude of the soliton was 2.923%,
whereas the numerically obtained decrease in amplitude was 2.994%. This
suggests that the present RANS approach can model the dissipation of solitons
with high accuracy. It was also shown that a boundary layer is formed on the
canal bottom, which persisted long after the ship has passed through. The
accelerated fluid also requires considerable time to return to its quiescent state.

In terms of numerical uncertainty, it was observed that the requirements for

the grid are considerably higher. The numerical uncertainty analysis was
performed for both regions, i.e. before and after the step. This revealed that as
the depth Froude number decreased, the requirements in terms of mesh
increased noticeably, resulting in an elevated uncertainty. On the other hand,
both regions showed relatively low temporal dependency.

238
 CONCLUSIONS AND FUTURE
WORK

9.1 Introduction

This chapter summarises the work performed in this thesis. The completion
of the aims and objectives is assessed based on the contents of each chapter.
Future work is then suggested based on the findings of the thesis.

9.2 Conclusions

This thesis presented examinations on a variety of subjects within the field of

ship hydrodynamics. Each chapter presented an assessment, designed to
allow a better understanding of the underlying phenomena. These were
defined at the onset of this thesis in terms of research aims and objectives, and
were met as follows.

The first objective, listed in Chapter 1 was:

✓ To perform a thorough review of the literature on computational ship

hydrodynamics and identify open research questions.

This objective was met in Chapter 2, which presented a literature review on a

variety of issues within the field of ship hydrodynamics. It included scale
effects in unrestricted and restricted waters. The survey revealed several key
characteristics of investigations in the area. For instance, assumptions, such as
that of the double body flow are frequently applied, as well as the high
popularity of linear and potential flow methods is noteworthy. Until recently,
this was justified with the scarcity of computational resources. However,
double body simulations of ship hydrodynamics can also be beneficial in

239
revealing underlying physics. For example, the interaction between frictional
and wave resistance, as pointed out in Chapter 3.

The following objectives, listed in Chapter 1 were:

✓ To perform a geosim analysis in deep and shallow waters.

✓ To investigate the scale effects on ship resistance and identify the sources of
these scale effects.

These objectives were addressed in two parts. Firstly, a geosim analysis was
performed in deep, unrestricted waters, where the physics of the problem are
simpler (Chapter 3). Then, the analysis was performed for the shallow water
equivalent (Chapter 5). Both of these assessments revealed an interaction
between different components of ship resistance, specifically, wave and
frictional resistance. Moreover, the existence of scale effects was demonstrated
on the wave resistance and form factor. Therefore, the choice of the scale factor
was shown to be one of the important aspects determining the value of the
extrapolated full-scale resistance of a ship. It was also demonstrated that one
should account for interactions between the components of ship resistance,
and that neglecting these can have a palpable influence on the extrapolated
value.

The sources of scale effects were shown to be rooted in the complexity of the
problem. The disparity between the Reynolds number of the model and full-
scale ships are chiefly responsible as causing scale effects. Specifically, the
action of viscosity and vorticity, which largely defy analytical description,
were identified as the responsible factors. This is due to their non-linear
variations with the Reynolds number, which changes several orders of
magnitude between the experimental and full-scale conditions.

The scale effects on the wave field were analysed from two points of view.
Firstly, the observed numerical wave profiles were shown to vary with scale.
In the deep-water case, this was done for the far-field waves, which are
assumed geometrically similar and inviscid. In the case of shallow water, the
wave elevation on the hull was examined due to the fact that the adopted case
produced near-field waves only. Both examinations revealed a change in the
wave elevation with variations in Reynolds number. The second point of view,
from which the wave field was examined was to seek the origin of the
observed scale effect. In both cases, this was identified as the boundary layer,
and its viscous and vortical action on the surrounding fluid.

240
The next objectives, listed in Chapter 1 were:

✓ To identify the best turbulence model to close the RANS equations, providing
consistent predictions over a range of parameters, at a small computational
cost.

The approach to addressing these objectives was designed to maximise their

practical relevance. To achieve this, Chapter 4 presented a turbulence
dependence study in restricted waters. Firstly, an overview was given to
familiarise the reader with the fundamental concepts, strengths, and
outstanding problems of the field of turbulence modelling. These were placed
in context by reference to workshops on numerical ship hydrodynamics and
current trends in the field. Then, four case studies in restricted waters were
adopted to demonstrate the relative effect of turbulence closure on sinkage,
trim, and resistance, since all of these parameters are magnified in confined
waters.

A consequence of the complexity of the problem of turbulence modelling was

found to preclude a straightforward ranking of the models. For this reason, a
statistical approach, originally devised to determine consistency in
experimental facilities was implemented with a small variation in the
methodology. This revealed that it is possible to obtain both qualitative and
quantitative characteristics of the turbulence models in terms of consistency
and accuracy across a range of different case studies. The assessment
suggested that the standard k-ω model provides the best compromise between
accuracy and consistency. Moreover, it was demonstrated that the k-ω model
requires the least computational resources of all two-equation eddy-viscosity
turbulence closures examined. Thus, the k-ω model was identified as a good
choice for studies in ship hydrodynamics.

The following objectives were listed in Chapter 1 as:

✓ To demonstrate the application of numerical uncertainty estimators on

different parameters and their use in full-scale ship hydrodynamics.
✓ To determine the performance of computational grids in terms of proximity to
the asymptotic range.

To maximise the practical relevance of the work and meet the above objectives,
a full-scale set of simulations was adopted in Chapter 6, where the objectives
were addressed. This choice was made in line with the fact that high Reynolds
number flows are difficult and expensive to measure, making them scarce in

241
the open literature and allowing little room for validation. In such cases,
numerical uncertainty estimators are frequently the only manner in which one
can assess the performance of a simulation. Therefore, an in-depth study into
the application of numerical uncertainty estimators, both local and global, was
performed. To facilitate the interpretation of data, generated as part of the
uncertainty study, a variety of methods were adopted. The resulting data was
shown to be able to aid the analyst in making decisions regarding the
performance of the simulation in terms of proximity to the asymptotic range.

The importance of choice of a parameter for the assessment was demonstrated

by sampling the free surface around the ship and the skin friction it
experiences. The assessment showed that despite the use of wall functions in
the latter case, the grid behaved in close proximity to the asymptotic range,
and was therefore associated with a relatively low numerical uncertainty. On
the other hand, the free surface was highly non-asymptotic, rendering the
uncertainty which was observed to be found higher. These contradicting
outcomes within the same simulation demonstrated that it is of high
importance to assess different parameters to determine the performance of a
numerical grid.

The next objectives, given in Chapter 1 read as follows:

✓ To determine the Kelvin half-angle of a ship in restricted waters based on a

numerical free surface analysis method.

These objectives were met within Chapter 7, where a virtual towing tank was
constructed, featuring no open boundaries (inlets or outlets), thereby being
consistent with physical experiments. In this towing tank, Galilean relativity
was not invoked to assign a velocity to the fluid to flow past a static hull.
Instead the ship propagates over a static fluid by use of the overset domain
method. It was demonstrated that via this approach, it is possible to achieve
reasonable accuracy in terms of ship resistance in model scale.

On the other hand, the wave field was analysed via a Fourier representation
of the numerical free surface. Despite the anticipated nonlinearities, resulting
from the solver and hullform, good agreement was established between a
theoretical prediction of the wave field in the tank and the numerical
observation. The numerically modelled Kelvin half-angle was shown to differ
in magnitude by a reasonably small amount when compared with the
theoretical approach (less than 2° deviation).

242
The final objective, listed in Chapter 1 was:

✓ To predict the impact of an abrupt change in the water depth on the resistance
of a ship and its wavefield in shallow waters.

To achieve this objective, Chapter 8 presented a numerical model of the

hydrodynamic effect of abrupt water depth changes on a ship travelling in
restricted waters using CFD. The virtual approach to tow the ship within the
computational tank, developed in Chapter 7, was applied. The resistance was
shown to vary substantially depending on the chosen depth Froude numbers,
both initial and those resulting after the step.

Significant interactions between the wave field and submerged depth

discontinuity were observed in the wave elevation time-history. The wave
field was analysed using two methods. Firstly, transmission coefficients were
calculated using two theories. One assuming the limit of very shallow water
has been achieved, and one taking into account the linear dispersion relation
for intermediate depths. Both theories showed agreement amongst
themselves, and with the numerical data. Secondly, the viscous dissipation,
resulting due to the friction with the side walls and inherent properties of the
fluid were analysed. These showed excellent agreement with an analytical
approximation.

9.3 Discussion

This thesis relied almost exclusively on the Computational Fluid Dynamics

approach to analysing ship hydrodynamics in deep and shallow waters.
Nevertheless, several theoretical methods were included to either augment the
discussions, provide room for comparison, or offer a type of validation. The
author of this thesis believes that it is justifiable to expect that CFD will become
the norm in hydrodynamic analysis and simulation-based design. However,
the utility of theories, relying on a number of well-posed assumptions should
not be discounted.

It was demonstrated in Chapter 2 that scientific papers concerning ship CFD

analysis have been increasing exponentially in recent years (Figure 2.3). Open
research questions in the field have persisted, presenting both researchers and
practitioners with several highly relevant unresolved issues. In Chapter 2 the
unresolved issues were identified as the resistance extrapolation procedures,
the approaches to modelling turbulence, shallow water effects on ship
performance, particularly as they relate to full-scale studies, the sensitivity of

243
the solution to grid density and topography, the representation of ship waves,
and the modelling of depth transitions in shallow waters. This thesis has
showed that each of these problems can be solved with the CFD approach.
Moreover, it was shown that the conceptual problems plaguing simpler
methods are largely overcome in this way. However, useful information can
be obtained from potential flow approaches, which can increase the
confidence in numerical approaches.

Specifically, Chapter 5 showed that theoretical analysis predicts the correct

behaviour of scale effects. Chapter 7 demonstrated that a fully nonlinear
towing tank produces a wavefield in agreement with linear methods, whereas
Chapter 8 validated the transmission coefficients and viscous dissipation on
ship-induced disturbances.

Alongside the successes of numerical approaches in modelling a variety of

phenomena, their limitations should also be borne in mind. These were
examined in Chapter 4, where turbulence dependence was reviewed. On the
other hand, Chapter 6 presented a detailed application and interpretation of
numerical uncertainty, induced as a result of mapping the governing
equations onto a discrete grid. Therefore, there are many aspects in which the
work presented in this thesis could be extended. These are examined in the
following section.

9.4 Future work

Without a doubt, the field of ship hydrodynamics presents many

opportunities of both theoretically engaging and practically relevant work.
Some specific examples for further research are given in this section.

It is hoped that the work presented in Chapter 3 and Chapter 5 will create an
interest in wave resistance changes with scale, a topic largely ignored in the
study of ship resistance. A more scientific extrapolation procedure is required
to replace the experience-based approach most facilities seem to have adopted.
While the ITTC’s procedure has been widely criticised for years, little progress
has been made towards its improvement, largely stagnating new
developments in this direction. The problems associated with scale effects do
not necessarily come from the decomposition and extrapolation procedure
itself, rather the assumptions imposed as a part of it. Were these to be relaxed
or re-examined with modern computational tools, progress could be achieved.

244
In this thesis, the presence of a rotating propeller was not modelled. This is
justifiable due to the scarcity such of studies performed even for un-appended
hulls where the complexity is much smaller compared to modelling a hull with
its appendages (propeller). Therefore, the studies presented in this thesis
should be used as a stepping stone to lead to research where self-propulsion
is considered.

A different aspect of ship hydrodynamics that is currently popular amongst

researchers is the estimation of roughness and fouling effects on ship hulls.
Although several studies have already been performed in this area, the
combined effect of shallow waters and fouling is yet unexamined. A related
study could also examine the effect of roughness on the canal walls or seabed
on ship performance. As demonstrated in Chapter 8, a boundary layer forms
on the ship and domain bottom surfaces. If both of these feature a degree of
fouling, it is justifiable to expect different performance characteristics.

Fouling and surface roughness should also be examined under different

turbulence models and their wider effect on ship performance. Ideally, such a
work would model the propeller’s action. The assessment of numerical
uncertainty should be extended to propellers. Since the wave field would be
fundamentally time-dependent, one could focus on the pressure pulses on the
ship hull, the friction experienced by the propeller blades, or the position and
volume of sheets of cavitation.

Finally, it would be of practical interest to determine which acceleration profile

results in the smallest unsteadiness and therefore quickest convergence of the
resistance curve. In both Chapter 7 and Chapter 8, a constant acceleration was
used to ramp up the velocity until the target speed is achieved. Determining a
manner of achieving the target speed with the smallest possible soliton
magnitude would be undoubtedly of interest to the towing tank community.

245
 REFERENCES

Abe, K., Kondoh, T., Nagano, Y., 1994. A new turbulence model for predicting
fluid flow and heat transfer in separating and reattaching flows-l. Flow
field calculations. Int. J. Heat Mass Transf. 37, 139–151.
https://doi.org/10.7566/JPSJ.84.054710

Abu-Ghannam, B.J., Shaw, R., 1980. Natural Transition of Boundary Layers -

The Effects of Turbulence, Pressure Gradient, and Flow History. J. Mech.
Eng. Sci. 22, 213–228.
https://doi.org/https://doi.org/10.1243/JMES_JOUR_1980_022_043_02

Alam, M.R., Mei, C.C., 2008. Ships advancing near the critical speed in a
shallow channel with a randomly uneven bed. J. Fluid Mech. 616, 397–
417. https://doi.org/10.1017/S0022112008004035

American Institute of Aeronautics and Astronautics (AIAA), 1994. Editorial

policy statement on numerical accuracy and experimental uncertainty.
AIAA J. 32, 3. https://doi.org/10.2514/3.48281

Andrun, M., Blagojević, B., Bašić, J., 2018. The influence of numerical
parameters in the finite-volume method on the Wigley hull resistance.
Proc. Inst. Mech. Eng. Part M J. Eng. Marit. Environ.
https://doi.org/10.1177/1475090218812956

Argyropoulos, C.D., Markatos, N.C., 2015. Recent advances on the numerical

modelling of turbulent flows. Appl. Math. Model. 39, 693–732.
https://doi.org/10.1016/j.apm.2014.07.001

Arnold-Bos, A., Martin, A., Khenchaf, A., 2007. Obtaining a ship’s speed and
direction from its Kelvin wake spectrum using stochastic matched
filtering, in: International Geoscience and Remote Sensing Symposium
(IGARSS). pp. 1106–1109. https://doi.org/10.1109/IGARSS.2007.4422995

246
ASME (American Society of Mechanical Engineers), 2009. Standard for
Verification and Validation in Computational Fluid Dynamics and Heat
Transfer - ASME V&V 20-2009, ASME International.

Atencio, B.N., Chernoray, V., 2019. A resolved RANS CFD approach for drag
characterization of antifouling paints. Ocean Eng. 171, 519–532.
https://doi.org/10.1016/j.oceaneng.2018.11.022

Axtmann, G., Rist, U., 2016. Scalability of OpenFOAM with Large

EddySimulations and DNS on High-PerformanceSystems, in: Nagel W.,
Kröner D., Resch M. (Eds) High Performance Computing in Science and
Engineering ´16. Springer, Cham. https://doi.org/10.1109/CAHPC.2005.27

Bailly, C., Comte-Bellot, G., 2004. Turbulence, Springer. Springer, Cham.

https://doi.org/https://doi.org/10.1007/978-3-319-16160-0

Bakica, A., Gatin, I., Vukčević, V., Jasak, H., Vladimir, N., 2019. Accurate
assessment of ship-propulsion characteristics using CFD. Ocean Eng. 175,
149–162. https://doi.org/10.1016/j.oceaneng.2018.12.043

Baldwin, B.S., Lomax, H., 1978. Thin Layer Approximation and Algebraic
Model for Separated Turbulent Flows, In 16th aerospace sciences meeting.

Baldwin, B.S., Timothy, B., 1990. A one-equation turbulence transport model

for high Reynolds number wall-bounded flows, National Aeronautics
and Space Administration. https://doi.org/10.2514/6.1991-610

Banks, J.B., Phillips, B., Bull, W., Turnock, S.R., 2010. RANS Simulations of the
Multiphase Flow Around the KCS Hullform, in: Gothenburg 2010
Workshop on Numerical Ship Hydrodynamics. pp. 601–606.

Bartholomeusz, B.Y.E.F., 1958. Reflexion of long waves at a step. Math. Proc.

Cambridge Philos. Soc. 54, 106–118.

Bašić, J., Degiuli, N., Dejhalla, R., 2017. Total resistance prediction of an intact
and damaged tanker with flooded tanks in calm water. Ocean Eng. 130,
83–91. https://doi.org/10.1016/j.oceaneng.2016.11.034

Bechthold, J., Kastens, M., 2020. Robustness and quality of squat predictions
in extreme shallow water conditions based on RANS-calculations. Ocean
Eng. 197, 106780. https://doi.org/10.1016/j.oceaneng.2019.106780

Beck, R.F., 1977. Forces and Moments on a Ship Moving in a Shallow Channel.
J. Sh. Res. 21, 107–119.

Beck, R.F., Newman, J.N., Tuck, E.O., 1975. Hydrodynamic forces on ships in
dredged channels. J. Sh. Res. 19, 166–171.

247
Beck, R.F., Reed, A.M., 2001. Modern Computational Methods for Ships in a
Seaway. SNAME Trans. 109, 1–55.

Begovic, E., Day, A.H., Incecik, A., Mancini, S., Pizzirusso, D., 2015. Roll
damping assessment of intact and damaged ship by CFD and EFD
methods. Proc. 12th Int. Conf. Stab. Ships Ocean Veh. 501–512.

Beji, S., 2018. Kadomtsev-Petviashvili type equation for entire range of relative
water depths.
https://doi.org/10.1080/21664250.2018.1436241org/10.1080/21664250.2018.
1436241

Bellafiore, D., Zaggia, L., Broglia, R., Ferrarin, C., Barbariol, F., Zaghi, S.,
Lorenzetti, G., Manfè, G., De Pascalis, F., Benetazzo, A., 2018. Modeling
ship-induced waves in shallow water systems: The Venice experiment.
Ocean Eng. 155, 227–239. https://doi.org/10.1016/j.oceaneng.2018.02.039

Bender, C.J., Dean, R.G., 2005. Reply to discussion of “Wave transformation by

two-dimensional bathymetric anomalies with sloped transitions” [Coast.
Eng. 50 (2003) 61-84]. Coast. Eng. 52, 201.
https://doi.org/10.1016/j.coastaleng.2004.11.003

Bender, C.J., Dean, R.G., 2003. Wave transformation by two-dimensional

bathymetric anomalies with sloped transitions. Coast. Eng. 50, 61–84.
https://doi.org/10.1016/j.coastaleng.2003.08.002

Benham, G.P., Bendimerad, R., Benzaquen, M., Clanet, C., 2020. Hysteretic
wave drag in shallow water. Phys. Rev. Fluids 5, 64803.
https://doi.org/10.1103/physrevfluids.5.064803

Benham, G.P., Boucher, J.P., Labbé, R., Benzaquen, M., Clanet, C., 2019. Wave
drag on asymmetric bodies. J. Fluid Mech. 878, 147–168.
https://doi.org/10.1017/jfm.2019.638

Bertram, V., 2012. Resistance and Propulsion, Practical Ship Hydrodynamics.

https://doi.org/10.1016/B978-0-08-097150-6.10003-X

Bhushan, S., Alam, M.F., Walters, D.K., 2013. Evaluation of hybrid RANS/LES
models for prediction of flow around surface combatant and Suboff
geometries. Computers and Fluids 88, 834–849.
https://doi.org/10.1016/j.compfluid.2013.07.020

Bhushan, S., Xing, T., Carrica, P., Stern, F., 2009. Model- and Full-Scale URANS
Simulations of Athena Resistance, Powering, Seakeeping, and 5415
Maneuvering. J. Sh. Res. 53, 179–198.

Bhushan, S., Xing, T., Carrica, P., Stern, F., Des, U., 2007. Model- and Full-Scale

248
 URANS / DES Simulations for Athena R / V Resistance , Powering , and
Motions, in: 9th International Conference on Numerical Ship
Hydrodynamics Ann Arbor, Michigan. pp. 5–8.

Billard, F., Laurence, D., 2012. A robust k-ε-v2̄/k elliptic blending turbulence
model applied to near-wall, separated and buoyant flows. Int. J. Heat
Fluid Flow 33, 45–58. https://doi.org/10.1016/j.ijheatfluidflow.2011.11.003

Blasius, H., 1908. The boundary layers in fluids with little friction. National
Advisory Committee for Aeronautics. Technical Memorandum 1256.

Böttner, C.U., Anschau, P., Shevchuk, I., 2020. Analysis of the flow conditions
between the bottoms of the ship and of the waterway. Ocean Eng. 199.
https://doi.org/10.1016/j.oceaneng.2020.107012

Bourne, J., 2000. Louisiana’s Vanishing Wetlands: Going, Going ... Science (80-
. ). 289, 1860 LP – 1863. https://doi.org/10.1126/science.289.5486.1860

Brard, R., 1970. Viscosity, Wake, and Ship Waves. J. Sh. Res. 14, 1–34.

Bugalski, T., 2007. An overview of the selected results of the European Union
Project EFFORT. Arch. Civ. Mech. Eng. 7, 55–67.
https://doi.org/10.1016/S1644-9665(12)60013-2

Cadafalch, J., Pérez-Segarra, C.D., Cònsul, R., Oliva, A., 2002. Verification of
finite volume computations on steady-state fluid flow and heat transfer.
J. Fluids Eng. Trans. ASME 124, 11–21. https://doi.org/10.1115/1.1436092

Cakici, F., Sukas, O.F., Kinaci, O.K., Alkan, A.D., 2017. Prediction of the
vertical motions of the dtmb 5415 ship using different numerical
approaches. Brodogradnja 68, 29–44. https://doi.org/10.21278/brod68203

Caplier, C., Rousseaux, G., Calluaud, D., David, L., 2016. Energy distribution
in shallow water ship wakes from a spectral analysis of the wave field.
Phys. Fluids 28. https://doi.org/10.1063/1.4964923

Caris, A., Limbourg, S., Macharis, C., van Lier, T., Cools, M., 2014. Integration
of inland waterway transport in the intermodal supply chain: A
taxonomy of research challenges. J. Transp. Geogr. 41, 126–136.
https://doi.org/10.1016/j.jtrangeo.2014.08.022

Carlson, J.A., Jaffe, A., Wiles, A., 2014. The Millenium Prize Problems,
American Mathematical Society.

Carrica, P.M., Huang, J., Noack, R., Kaushik, D., Smith, B., Stern, F., 2010.
Large-scale DES computations of the forward speed diffraction and pitch
and heave problems for a surface combatant. Computers and Fluids 39,

249
 1095–1111. https://doi.org/10.1016/j.compfluid.2010.02.002

Carrica, P.M., Wilson, R. V., Noack, R.W., Stern, F., 2007. Ship motions using
single-phase level set with dynamic overset grids. Computers and Fluids
36, 1415–1433. https://doi.org/10.1016/j.compfluid.2007.01.007

Carusotto, I., Rousseaux, G., 2013. The Cerenkov effect revisited: From
swimming ducks to zero modes in gravitational analogues, in: Analogue
Gravity Phenomenology. Cham, pp. 109–144.

Castiglione, T., Stern, F., Bova, S., Kandasamy, M., 2011. Numerical
investigation of the seakeeping behavior of a catamaran advancing in
regular head waves. Ocean Eng. 38, 1806–1822.
https://doi.org/10.1016/j.oceaneng.2011.09.003

Castro, A.M., Carrica, P.M., Stern, F., 2011. Full scale self-propulsion
computations using discretized propeller for the KRISO container ship
KCS. Computers and Fluids 51, 35–47.
https://doi.org/10.1016/j.compfluid.2011.07.005

Cazalbou, J.B., Spalart, P.R., Bradshaw, P., 1994. On the behavior of two-
equation models at the edge of a turbulent region. Phys. Fluids 6, 1797–
1804. https://doi.org/10.1063/1.868241

Celik, I.B., Ghia, U., Roache, P.J., Freitas, C.., 2008. Procedure for Estimation
and Reporting of Uncertainty Due to Discretization in CFD Applications.
J. Fluids Eng. 130, 078001. https://doi.org/10.1115/1.2960953

Celik, I.B., Karatekin, O., 1997. Numerical Experiments on Application of

Richardson Extrapolation With Nonuniform Grids. J. Fluids Eng. 119,
584–590. https://doi.org/10.1115/1.2819284

Celik, I.B., Li, J., Hu, G., Shaffer, C., 2005. Limitations of Richardson
Extrapolation and Some Possible Remedies. J. Fluids Eng. 127, 795.
https://doi.org/10.1115/1.1949646

Chen, H.C., Patel, V.C., 1988. Near-wall turbulence models for complex flows
including separation. AIAA J. 26, 641–648. https://doi.org/10.2514/3.9948

Chen, X.-N., Grounarz, A., List, S., 2001. Flow Around Ships Sailing in Shallow
Water—Experimental and Numerical Results, in: Twenty-Second
Symposium on Naval Hydrodynamics Office of Naval Research Bassin
d’Essais Des CarenesNational Research Council. pp. 968–982.
https://doi.org/10.17226/9771

Chen, X., Zhu, R., Ma, C., Fan, J., 2016. Computations of linear and nonlinear
ship waves by higher-order boundary element method. Ocean Eng. 114,

250
 142–153. https://doi.org/10.1016/j.oceaneng.2016.01.016

Chien, K.Y., 1982. Predictions of Channel and Boundary-Layer Flows with a

Low-Reynolds-Number Turbulence Model. AIAA J. 20, 33–38.
https://doi.org/10.2514/3.51043

Choi, J.E., Kim, J.H., Lee, H.G., Choi, B.J., Lee, D.H., 2009. Computational
predictions of ship-speed performance. J. Mar. Sci. Technol. 14, 322–333.
https://doi.org/10.1007/s00773-009-0047-4

Choi, J.E., Min, K.S., Kim, J.H., Lee, S.B., Seo, H.W., 2010. Resistance and
propulsion characteristics of various commercial ships based on CFD
results. Ocean Eng. 37, 549–566.
https://doi.org/10.1016/j.oceaneng.2010.02.007

Chou, P.Y., 1945. On velocity correlations and the solutions of the equations of
turbulent fluctuation. Q. Appl. Math. 3, 38–54.
https://doi.org/10.1090/qam/11999

Chun, H.H., Park, I.R., Lee, S.K., 2001. Analysis of Turbulence Free-Surface
Flow around Hulls in Shallow Water Channel by a Level-set Method, in:
Twenty-Second Symposium on Naval Hydrodynamics Office of Naval
Research Bassin d’Essais Des CarenesNational Research Council. pp. 941–
956. https://doi.org/10.17226/9771

Ciappi, E., Magionesi, F., 2005. Characteristics of the turbulent boundary layer
pressure spectra for high-speed vessels. J. Fluids Struct. 21, 321–333.
https://doi.org/10.1016/j.jfluidstructs.2005.07.006

Cifani, P., Kuerten, J.G.M., Geurts, B.J., 2018. Highly scalable DNS solver for
turbulent bubble-laden channel flow. Computers and Fluids 172, 67–83.
https://doi.org/10.1016/j.compfluid.2018.06.008

Cole, S.L., 1987. Transient waves produced by a moving pressure distribution.

Q. Appl. Math. 45, 51–58. https://doi.org/10.1090/qam/885167

Crapper, G.D., 1964. Surface waves generated by a travelling pressure point.

Proc. R. Soc. London. Ser. A. Math. Phys. Sci. 282, 547–558.
https://doi.org/10.1098/rspa.1964.0250

Dam, K.T., Tanimoto, K., Fatimah, E., 2008. Investigation of ship waves in a
narrow channel. J. Mar. Sci. Technol. 13, 223–230.
https://doi.org/10.1007/s00773-008-0005-6

Dand, W.I., 1967. The wavemaking resistance of ships: Vertical force and form
resistance of a hull at uniform velocity. PhD Thesis. University of
Glasgow.

251
Darrigol, O., 2017. Joseph Boussinesq’s legacy in fluid mechanics. Comptes
Rendus - Mec. 345, 427–445. https://doi.org/10.1016/j.crme.2017.05.008

Darrigol, O., 2003. The spirited horse, the engineer, and the mathematician:
Water waves in nineteenth-century hydrodynamics. Arch. Hist. Exact Sci.
58, 21–95. https://doi.org/10.1007/s00407-003-0070-5

Darrigol, O., Frisch, U., 2008. From Newton’s mechanics to Euler’s equations.
Phys. D Nonlinear Phenom. 237, 1855–1869.
https://doi.org/10.1016/j.physd.2007.08.003

Darrigol, O., Turner, J.S., 2006. Worlds of flow: A history of hydrodynamics

from the bernoullis to prandtl, Physics Today.
https://doi.org/10.1063/1.2349735

Date, C.J., Turnock, S., 1999. A study into the techniques needed to accurately
predict a skin friction using RANS solvers with validation against
Froude’s historical flat place experimental data, Ship science report No.
114.

David, C.G., Roeber, V., Goseberg, N., Schlurmann, T., 2017. Generation and
propagation of ship-borne waves - Solutions from a Boussinesq-type
model. Coast. Eng. 127, 170–187.
https://doi.org/10.1016/j.coastaleng.2017.07.001

Davidson, L., Nielsen, P. V, Sveningsson, A., 2003. Modifications of the V2F

Model for Computing the Flow in a 3D Wall Jet. Turbul. Heat Mass Transf.
4, 577–584.

Day, A.H., Clelland, D., Doctors, L.J., 2009. Unsteady finite-depth effects
during resistance tests on a ship model in a towing tank. J. Mar. Sci.
Technol. 14, 387–397. https://doi.org/10.1007/s00773-009-0057-2

Demirel, Y.K., Khorasanchi, M., Turan, O., Incecik, A., Schultz, M.P., 2014. A
CFD model for the frictional resistance prediction of antifouling coatings.
Ocean Eng. 89, 21–31. https://doi.org/10.1016/j.oceaneng.2014.07.017

Demirel, Y.K., Turan, O., Incecik, A., 2017. Predicting the effect of biofouling
on ship resistance using CFD. Appl. Ocean Res. 62, 100–118.
https://doi.org/10.1016/j.apor.2016.12.003

Deng, G., Duvigneau, R., Queutey, P., Visonneau, M., 2004. Assessment of
turbulence models for ship flow at full scale, in: Comp. Mech., WCCM VI.
Beijing, China.

Deng, G., Nantes, E.C. De, Guilmineau, E., Queutey, P., Visonneau, M., 2014.
Simulation of Container Ship in Shallow Water at Model Scale and Full

252
 Scale, in: Proceedings of the 3rd National CFD Workshop for Ship and
Offshore Engineering.

Deng, G.B., Queutey, P., Visonneau, M., 2010. RANS prediction of the
KVLCC2 tanker in head waves. J. Hydrodyn. 22, 476–481.
https://doi.org/10.1016/S1001-6058(09)60239-0

Deng, R., Li, C., Huang, D., Zhou, G., 2015. The Effect of Trimming and
Sinkage on the Trimaran Resistance Calculation, Procedia Engineering.
Elsevier B.V. https://doi.org/10.1016/j.proeng.2015.11.199

Denner, F., Paré, G., Zaleski, S., 2017. Dispersion and viscous attenuation of
capillary waves with finite amplitude. Eur. Phys. J. Spec. Top. 226, 1229–
1238. https://doi.org/10.1140/epjst/e2016-60199-2

Dhinesh, G., Murali, K., Subramanian, V.A., 2010. Estimation of hull-propeller

interaction of a self-propelling model hull using a RANSE solver. Ships
Offshore Struct. 5, 125–139. https://doi.org/10.1080/17445300903231109

Dias, F., 2014. Ship waves and Kelvin. J. Fluid Mech. 746, 1–4.
https://doi.org/10.1017/jfm.2014.69

Diskin, B., Thomas, J.L., 2010. Notes on accuracy of finite-volume

discretization schemes on irregular grids. Appl. Numer. Math. 60, 224–
226. https://doi.org/10.1016/j.apnum.2009.12.001

Doctors, L.J., 2007. A Numerical Study of the Resistance of Transom-Stern

Monohulls. Sh. Technol. Res. 54, 134–144.
https://doi.org/10.1179/str.2007.54.3.005

Doctors, L.J., 1975. The experimental wave resistance of an accelerating two-

dimensional pressure distribution. J. Fluid Mech. 72, 513–527.
https://doi.org/10.1017/S0022112075003114

Doctors, L.J., Macfarlane, G.J., Young, R., 2007. A study of transom-stern

ventilation. Int. Shipbuild. Prog. 54, 145–163.

Dorn, W.G.V., 1966. Boundary dissipation of oscillatory waves. J. Fluid Mech.

24, 769–779. https://doi.org/10.1017/S0022112066000995

Duffy, J., 2008. Modelling of Ship-Bank Interaction and Ship Squat for Ship-
Handling Simulation by.

Durbin, P.A., 2017a. Some Recent Developments in Turbulence Closure

Modeling. Annu. Rev. Fluid Mech. 50, 77–103.
https://doi.org/10.1146/annurev-fluid-122316-045020

Durbin, P.A., 2017b. Perspectives on the Phenomenology and Modeling of

253
 Boundary Layer Transition. Flow, Turbul. Combust. 99, 1–23.
https://doi.org/10.1007/s10494-017-9819-9

Durbin, P.A., 1996. On the k-3 stagnation point anomaly. Int. J. Heat Fluid
Flow 17, 89–90. https://doi.org/10.1016/0142-727X(95)00073-Y

Durbin, P.A., 1995. Separated flow computations with the k-epsilon-v-squared

model. AIAA J. 33, 659–664. https://doi.org/10.2514/3.12628

Durbin, P.A., 1993a. Application of a near-wall turbulence model to boundary

layers and heat transfer. Int. J. Heat Fluid Flow 14, 316–323.
https://doi.org/10.1016/0142-727X(93)90004-7

Durbin, P.A., 1993b. A Reynolds stress model for near-wall turbulence. J. Fluid
Mech. 249, 465–498. https://doi.org/10.1017/S0022112093001259

Durbin, P.A., 1991. Near-wall turbulence closure modeling without “damping

functions.” Theor. Comput. Fluid Dyn. 3, 1–13.
https://doi.org/10.1007/BF00271513

Durbin, P.A., Pettersson Reif, B.A., 2011. Statistical theory and modelling for
turbulent flow, Second Edi. ed. Wiley.

Duvigneau, R., Visonneau, M., Deng, G.B., 2003. On the role played by
turbulence closures in hull shape optimization at model and full scale. J.
Mar. Sci. Technol. 8, 11–25. https://doi.org/10.1007/s10773-003-0153-8

Eca, L., Hoekstra, M., 2009. Evaluation of numerical error estimation based on
grid refinement studies with the method of the manufactured solutions.
Computers and Fluids 38, 1580–1591.
https://doi.org/10.1016/j.compfluid.2009.01.003

Eca, L., Hoekstra, M., 2008. The numerical friction line. J. Mar. Sci. Technol. 13,
328–345. https://doi.org/10.1007/s00773-008-0018-1

Eca, L., Hoekstra, M., 2006. Discretization Uncertainty Estimation based on a

Least Squares version of the Grid Convergence Index. Proc. Second Work.
CFD Uncertain. Anal. 1–27.

Eca, L., Hoekstra, M., 2001. Numerical Prediction of Scale Effects in Ship Stern
Flows with Eddy-Viscosity Turbulence Models, in: Twenty-Second
Symposium on Naval Hydrodynamics Office of Naval Research Bassin
d’Essais Des CarenesNational Research Council. pp. 553–568.
https://doi.org/10.17226/9771

Eca, L., Hoekstra, M., Beja Pedro, J.F., Falcao de Campos, J.A.C., 2013. On the
characterization of grid density in grid refinement studies for

254
 discretization error estimation. Int. J. Numer. Methods Fluids 72, 119–134.
https://doi.org/https://doi.org/10.1002/fld.3737

Eca, L., Pereira, F.S., Vaz, G., 2018. Viscous flow simulations at high Reynolds
numbers without wall functions: Is y+≃1 enough for the near-wall cells?
Computers and Fluids 170, 157–175.
https://doi.org/10.1016/j.compfluid.2018.04.035

Eca, L., Saraiva, G., Vaz, G., Abreu, H., 2015. The Pros and Cons of Wall
Functions. OMAE2015, St.Johns, Newfoundland, Canada 0–11.
https://doi.org/10.1115/OMAE2015-41518

El-Behery, S.M., Hamed, M.H., 2011. A comparative study of turbulence

models performance for separating flow in a planar asymmetric diffuser.
Computers and Fluids 44, 248–257.
https://doi.org/10.1016/j.compfluid.2011.01.009

el Moctar, O., Lantermann, U., Mucha, P., Höpken, J., Schellin, T.E., 2015.
RANS-Based Simulated Ship Maneuvering Accounting for Hull-
Propulsor-Engine Interaction. Sh. Technol. Res. 61, 142–161.
https://doi.org/10.1179/str.2014.61.3.003

Elsherbiny, K., Terziev, M., Tezdogan, T., Incecik, A., Kotb, M., 2020.
Numerical and experimental study on hydrodynamic performance of
ships advancing through different canals. Ocean Eng. 195.
https://doi.org/10.1016/j.oceaneng.2019.106696

Elsherbiny, K., Terziev, M., Tezdogan, T., Incecik, A., Kotb, M., 2019a.
Numerical and experimental study on hydrodynamic performance of
ships advancing through different canals. Ocean Eng. 106696.
https://doi.org/10.1016/j.oceaneng.2019.106696

Elsherbiny, K., Tezdogan, T., Kotb, M., Incecik, A., Day, A.H., 2019b. An
experimental investigation of the trim effect on the behaviour of a
containership in shallow water, in: Proceedings of the ASME 2019 38th
International Conference on Ocean, Offshore and Arctic Engineering
OMAE2019, Glasgow, UK.

Elsherbiny, K., Tezdogan, T., Kotb, M., Incecik, A., Day, S., 2019c.
Experimental analysis of the squat of ships advancing through the New
Suez Canal. Ocean Eng. 178, 331–344.
https://doi.org/10.1016/j.oceaneng.2019.02.078

EMSA (European Maritime Safety Agency), 2019. Annual Overview of Marine

Casualties and Incidents 2019.

255
EMSA (European Maritime Safety Agency), 2018. Annual Overview of Marine
Casualties and Incidents 2018.

EMSA (European Maritime Safety Agency), 2017. Annual overview of marine

casualties and incidents 2017.

EMSA (European Maritime Safety Agency), 2016. Annual overview of marine

casualties and incidents 2016.

EMSA (European Maritime Safety Agency), 2015. Annual overview of marine

casualties and incidents 2011-2015.

Ertekin, R.C., Webster, W.C., Wehausen, J. V., 1986. Waves caused by a moving
disturbance in a shallow channel of finite width. J. Fluid Mech. 169, 275–
292. https://doi.org/10.1017/S0022112086000630

European Commission, 2018. Inland waterways [WWW Document]. Mobil.

Transp. URL https://ec.europa.eu/transport/modes/inland_en (accessed
5.17.19).

Farkas, A., Degiuli, N., Martić, I., 2018. Assessment of hydrodynamic

characteristics of a full-scale ship at different draughts. Ocean Eng. 156,
135–152. https://doi.org/10.1016/j.oceaneng.2018.03.002

Farkas, A., Degiuli, N., Martić, I., 2017. Numerical investigation into the
interaction of resistance components for a series 60 catamaran. Ocean Eng.
146, 151–169. https://doi.org/10.1016/j.oceaneng.2017.09.043

Ferguson, A.M., 1977. Factors affecting the components of ship resistance. PhD
Thesis. University o Glasgow, Department of Naval Architecture.
University of Glasgow.

Ferziger, J.H., Peric, M., 2002. Computational Methods for Fluid Dynamics,
Springer. https://doi.org/10.1016/S0898-1221(03)90046-0

Fourdrinoy, J., Caplier, C., Devaux, Y., Rousseaux, G., et. al., 2019. The naval
battle of actium and the myth of the ship-holder : the effect of bathymetry,
in: 5th MASHCON, Ghent University, May 2019.

Fox, R., McDonald, A., Prichard, P., 2015. Introduction to fluid mechanics, 6th
ed, Hoboken, NJ. Wiley.

Freitas, C.J., 2002. The issue of numerical uncertainty. Appl. Math. Model. 26,
237–248. https://doi.org/10.1016/S0307-904X(01)00058-0

Froude, W., 1874. Report to the lords commissioners of the admiralty on

experiments for the determination of the frictional resistance of water on
a surface, under various conditions, performed at Chelston cross, under

256
 the authority of their lordships.

Fujino, M., 1976. Maneuverability in restricted waters, state of the art. National
Science Foundation Report No. 184. Department of Naval Architecture
and Marine Engineering, College of Engineering, University of Michigan,
Ann Arbor, Michigan. Ann Arbor, Michigan.

Fukagata, K., Iwamoto, K., Kasagi, N., 2002. Contribution of Reynolds stress
distribution to the skin friction in wall-bounded flows. Phys. Fluids 14,
L73–L76. https://doi.org/10.1063/1.1516779

Fureby, C., Toxopeus, S.L., Johansson, M., Tormalm, M., Petterson, K., 2016. A
computational study of the flow around the KVLCC2 model hull at
straight ahead conditions and at drift. Ocean Eng. 118, 1–16.
https://doi.org/10.1016/j.oceaneng.2016.03.029

Gadd, G.E., 1967. A new turbulent friction formulation based on a reappraisal

of Hughes’ results. Trans. RINA 109, 109–511.

Gaggero, S., Villa, D., Viviani, M., 2015. The Kriso container ship (KCS) test
case: An open source overview. Comput. Methods Mar. Eng. VI, Mar.
2015 735–749.

García-Gómez, A., 2000. On the form factor scale effect. Ocean Eng. 27, 97–109.
https://doi.org/10.1016/S0029-8018(98)00042-0

Golbraikh, E., Eidelman, A., Soloviev, A., 2013. On helical behavior of

turbulence in the ship wake. J. Hydrodyn. 25, 83–90.
https://doi.org/10.1016/S1001-6058(13)60341-8

Gomit, G., Rousseaux, G., Chatellier, L., Calluaud, D., David, L., 2014. Spectral
analysis of ship waves in deep water from accurate measurements of the
free surface elevation by optical methods. Phys. Fluids 26.
https://doi.org/10.1063/1.4902415

Gorji, S., Seddighi, M., Ariyaratne, C., Vardy, A.E., O’Donoghue, T., Pokrajac,
D., He, S., 2014. A comparative study of turbulence models in a transient
channel flow. Computers and Fluids 89, 111–123.
https://doi.org/10.1016/j.compfluid.2013.10.037

Gotman, A., 2007. A history of ship resistance evaluation. J. Ocean Technol. 2,

74–96.

Gotman, A., 2002. Study Of Michell’s Integral And Influence Of Viscosity And
Ship Hull Form On Wave Resistance. Ocean. Eng. Int. 6, 74–115.

Gourlay, T., 2014. Duisburg Test Case Containership Squat Prediction using

257
 ShallowFlow Software, in: , PreSquat - Numerische Vorhersagen von
Dynamischem Squat in Begrenzten Gewässern, Bericht F005/2014. pp. 1–
6.

Gourlay, T., 2006. Flow beneath a ship at small underkeel clearance. J. Sh. Res.
50, 250–258.

Gourlay, T., 2003. Ship Squat In Water of Varying Depth. Int. J. Marit. Eng. 145,
12. https://doi.org/10.3940/rina.ijme.2003.a1.1031

Gourlay, T., 2001. The supercritical bore produced by a high-speed ship in a

channel. J. Fluid Mech. 434, 399–409.
https://doi.org/10.1017/S002211200100372X

Gourlay, T., Tuck, E.O., 2001. The maximum sinkage of a ship. J. Sh. Res. 45,
50–58.

Grigson, C.W.B., 1999. A planar friction algorithm and its use in analysing hull
resistance. Trans. RINA.

Grue, J., 2017. Ship generated mini-tsunamis. J. Fluid Mech. 816, 142–166.
https://doi.org/10.1017/jfm.2017.67

Guo, B.J., Deng, G.B., Steen, S., 2013. Verification and validation of numerical
calculation of ship resistance and flow field of a large tanker. Ships
Offshore Struct. 8, 3–14. https://doi.org/10.1080/17445302.2012.669263

Guo, C., Zhang, Q., Shen, Y., 2015. A non-geometrically similar model for
predicting the wake field of full-scale ships. J. Mar. Sci. Appl. 14, 225–233.
https://doi.org/10.1007/s11804-015-1316-8

Guo, J., 2002. Simple and explicit solution of wave dispersion equation. Coast.
Eng. 45, 71–74. https://doi.org/10.1016/S0378-3839(02)00039-X

Haase, M., Davidson, G., Binns, J., Thomas, G., Bose, N., 2016a. Full-scale
resistance prediction in finite waters: A study using computational fluid
dynamics simulations, model test experiments and sea trial
measurements. Proc. Inst. Mech. Eng. Part M J. Eng. Marit. Environ. 231,
316–328. https://doi.org/10.1177/1475090216642467

Haase, M., Zurcher, K., Davidson, G., Binns, J.R., Thomas, G., Bose, N., 2016b.
Novel CFD-based full-scale resistance prediction for large medium-speed
catamarans. Ocean Eng. 111, 198–208.
https://doi.org/10.1016/j.oceaneng.2015.10.018

Hai-Long, S., Obwogi, E.O., Yu-Min, S., 2016. Scale effects for rudder bulb and
rudder thrust fin on propulsive efficiency based on computational fluid

258
 dynamics. Ocean Eng. 117, 199–209.
https://doi.org/10.1016/j.oceaneng.2016.03.046

Havelock, T., 1908. The Propagation of Groups of Waves in Dispersive Media,

with Aplication to Waves on Water produced by a Travelling Disturbance
422–451. https://doi.org/10.1098/rspa.1933.0074

Hawkes, J., Vaz, G., Phillips, A.B., Cox, S.J., Turnock, S.R., 2018. On the strong
scalability of maritime CFD. J. Mar. Sci. Technol. 23, 81–93.
https://doi.org/10.1007/s00773-017-0457-7

Hirt, C.. W., Nichols, B.. D., 1981. Volume of fluid (VOF) method for the
dynamics of free boundaries. J. Comput. Phys. 39, 201–225.
https://doi.org/10.1016/0021-9991(81)90145-5

Hochkirch, K., Mallol, B., 2000. On the Importance of Full-Scale CFD

Simulations for Ships. Numeca 85–95.

Hoekstra, M., Eca, L., Windt, J., Raven, H.C., 2000. Viscous flow calculations
for KVLCC2 and KCS models using the PARNASSOS code, in: Gotenburg
2000: A Workshop on Numerical Ship Hydrodynamics.

Hofmann, H., Lorke, A., Peeters, F., 2008. The relative importance of wind and
ship waves in the littoral zone of a large lake. Linnol. Ocean. 53, 368–380.
https://doi.org/10.1029/2010WR010012

Hou, Y. hang, Li, Y. jia, Liang, X., 2019. Mixed aleatory/epistemic uncertainty
analysis and optimization for minimum EEDI hull form design. Ocean
Eng. 172, 308–315. https://doi.org/10.1016/j.oceaneng.2018.12.003

Hughes, G., 1954. Friction and form resistance in turbulent flow and a
proposed formulation for use in model and ship correlation. Trans. Inst.
Nav. Arch. 96.

Hunt, J.N., 1979. Direct solution of wave dispersion equation. J. Waterw. Port,
Coastal, Ocean Eng. 106, 457–459.

Hunt, J.N., 1964. The viscous damping of gravity waves in shallow water. La
Houille Blanche 685–691. https://doi.org/10.1051/lhb/1964038

Hur, V.M., 2019. Shallow water models with constant vorticity. Eur. J. Mech.
B/Fluids 73, 170–179. https://doi.org/10.1016/j.euromechflu.2017.06.001

Iaccarino, G., 2002. Predictions of a Turbulent Separated Flow Using

Commercial CFD Codes. J. Fluids Eng. 123, 819.
https://doi.org/10.1115/1.1400749

IMO, 2011. Annex 19: Resolution MEPC.203(62).

259
Inui, T., 1954. Wave-Making Resistance in Shallow Sea and in Restricted
Water, with Special Reference to its Discontinuities. J. Zosen Kiokai 76, 1–
10.

Inui, T., 1936a. On Deformation, Wave Patterns and Resoncance Phenomenon

of Water Surface due to a Moving Disturbance. II. Proc. Physico-
Mathematical Soc. Japan 18.

Inui, T., 1936b. On Deformation, Wave Patterns and Resoncance Phenomenon

of Water Surface due to a Moving Disturbance. I. Proc. Physico-
Mathematical Soc. Japan 18.

Irkal, M.A.R., Nallayarasu, S., Bhattacharyya, S.K., 2019. Numerical prediction

of roll damping of ships with and without bilge keel. Ocean Eng. 179, 226–
245. https://doi.org/10.1016/j.oceaneng.2019.03.027

Ishihara, T., Ogasawara, H., Hunt, J.C.R., 2015. Analysis of conditional

statistics obtained near the turbulent/non-turbulent interface of turbulent
boundary layers. J. Fluids Struct. 53, 50–57.
https://doi.org/10.1016/j.jfluidstructs.2014.10.008

ITTC, 2017a. Recommended Procedures 1978 ITTC Performance Prediction

Method, 4th revision, 7.5 – 02 03 – 01.4. 28th Int Towing Tank Conf.

ITTC, 2017b. ITTC-Recommended Procedures and Guidelines Uncertainty

Analysis, Instrument Cali-bration ITTC Quality System Manual
Recommended Procedures and Guidelines Procedure.

ITTC, 2014. ITTC – Recommended Procedures and Guidelines - Practical

guidelines for ship CFD applications. 7.5-03-02-03 (Revision 01). ITTC –
Recomm. Proced. Guidel. 19.

ITTC, 2011. Recommended Procedures and Guidelines: Practical Guidelines

for Ship CFD. 26th Int. Towing Tank Conf.

ITTC, 2008. Uncertainty Analysis in CFD Verification and Validation

Methodology and Procedures. 25th ITTC 2008, Resist. Comm. 12.

ITTC, 2002. ITTC – Recommended Procedures and Guidelines Uncertainty

Analysis , Example for Resistance Test.

ITTC, 1999. 1978 ITTC Performance Prediction Method 7.5 - 02 - 03 - 01.4.

Jachowski, J., 2008. Assessment of ship squat in shallow water using CFD.
Arch. Civ. Mech. Eng. 8, 27–36. https://doi.org/10.1016/S1644-
9665(12)60264-7

Jagadeesh, P., Murali, K., 2010. Rans Predictions of Free Surface Effects on

260
 Axisymmetric Underwater Body. Eng. Appl. Comutational Fluid Mech.
4, 301–313. https://doi.org/10.1080/19942060.2010.11015318

Janson, C.-E., Spinney, D., 2004. A Comparison of Four Wave Cut Analysis
Methods for Wave Resistance Prediction. Sh. Technol. Res. 51, 173–184.
https://doi.org/10.1179/str.2004.51.4.004

Jasak, H., 1996. Error analysis and estimation for finite volume method with
applications to fluid flow. PhD Thesis. Imperial College of Science,
Technology and Medicine.

Jasak, H., Vukčević, V., Gatin, I., Lalović, I., 2018. CFD validation and grid
sensitivity studies of full scale ship self propulsion. Int. J. Nav. Archit.
Ocean Eng. https://doi.org/10.1016/j.ijnaoe.2017.12.004

Jiang, T., 2001. A New Method For Resistance and Propulsion Prediction of
Ship Performance in Shallow Water, in: Proceedings of the Eighth
International Symposium on Practical Design of Ships and Other Floating
Structures 16 – 21 September 2001 Shanghai, Chaina. pp. 509–515.

Jiang, T., Henn, R., Sharma, S.D., 2002. Wash waves generated by ships
moving on fairways of varying topography. 24th Symp. Nav. Hydrodyn.
2, 8–13.

Johnson, D.A., King, L.S., 1984. A mathematically simple turbulence closure

model for attached and separated turbulent boundary layers. AIAA J. 23,
1684–1692. https://doi.org/10.2514/3.9152

Johnson, J.W., 1957. Ship Waves in Navigation Channels. Coast. Eng. Proc. 1,
40. https://doi.org/10.9753/icce.v6.40

Jones, D.A., Clarke, D.B., 2010. Fluent Code Simulation of Flow around a
Naval Hull: the DTMB 5415. Def. Sci. Technol. Organ. Victoria (Australia),
Marit. platforms Div.

Jones, W.P., Launder, B.E., 1972. The Prediction of Laminarization with a Two-
Equation Model of Turbulence. Int. J. Heat Mass Transf. 15, 301–314.
https://doi.org/10.1016/0017-9310(72)90076-2

Jongen, T., 1998. Simulation and Modeling of Turbulent Incompressible Flows.

Kaidi, S., Smaoui, H., Sergent, P., 2017. Numerical estimation of bank-
propeller-hull interaction effect on ship manoeuvring using CFD method.
J. Hydrodyn. 29, 154–167. https://doi.org/10.1016/S1001-6058(16)60727-8

Katsis, C., Akylas, T.R., 1987. On the excitation of long nonlinear water waves
by a moving pressure distribution. part 2. three dimensional effects. J.

261
 Fluid Mech. 177, 49–65. https://doi.org/10.1017/S0022112087000855

Katsis, C., Akylas, T.R., 1984. On the excitation of long nonlinear water waves
by a moving pressure distribution. J. Fluid Mech. 177, 49–65.
https://doi.org/10.1017/S0022112087000855

Katsui, T., Asai, H., Himeno, Y., Tahara, Y., 2005. The proposal of a new
friction line, in: Fifth Osaka Colloquium on Advanced CFD Applications
to Ship Flow and Hull Form Design, Osaka, Japan.

Kellett, P., Turan, O., Incecik, A., 2013. A study of numerical ship underwater
noise prediction. Ocean Eng. 66, 113–120.
https://doi.org/10.1016/j.oceaneng.2013.04.006

Keulegan, G.H., 1948. Gradual damping of solitary waves. J. Res. Natl. Bur.
Stand. (1934). 40, 487. https://doi.org/10.6028/jres.040.041

Kianejad, S.S., Enshaei, H., Duffy, J., Ansarifard, N., 2019. Prediction of a ship
roll added mass moment of inertia using numerical simulation. Ocean
Eng. 173, 77–89. https://doi.org/10.1016/j.oceaneng.2018.12.049

Kim, M., Hizir, O., Turan, O., Day, S., Incecik, A., Day, A.H., 2017a. Estimation
of added resistance and ship speed loss in a seaway. Ocean Eng. 141, 1–
12. https://doi.org/10.1016/j.oceaneng.2017.06.051

Kim, M., Hizir, O., Turan, O., Incecik, A., 2017b. Numerical studies on added
resistance and motions of KVLCC2 in head seas for various ship speeds.
Ocean Eng. 140, 466–476. https://doi.org/10.1016/j.oceaneng.2017.06.019

Kim, W.J., Van, S.H., Kim, D.H., 2001. Measurement of flows around modern
commercial ship models. Exp. Fluids 31, 567–578.
https://doi.org/10.1007/s003480100332

Kim, Y.C., Kim, K.S., Kim, J., Kim, Y., Park, I.R., Jang, Y.H., 2017. Analysis of
added resistance and seakeeping responses in head sea conditions for
low-speed full ships using URANS approach. Int. J. Nav. Archit. Ocean
Eng. 9, 641–654. https://doi.org/10.1016/j.ijnaoe.2017.03.001

Kinaci, O.K., Gokce, M.K., 2015. A computational hydrodynamic analysis of

duisburg test case with free surface and propeller. Brodogradnja 66, 23–
38.

Kinaci, O.K., Sukas, O.F., Bal, S., 2016. Prediction of wave resistance by a
Reynolds-averaged Navier-Stokes equation-based computational fluid
dynamics approach. Proc. Inst. Mech. Eng. Part M J. Eng. Marit. Environ.
230, 531–548. https://doi.org/10.1177/1475090215599180

262
Kiš, P., Herwig, H., 2012. The near wall physics and wall functions for
turbulent natural convection. Int. J. Heat Mass Transf. 55, 2625–2635.
https://doi.org/10.1016/j.ijheatmasstransfer.2011.12.031

Kok, Z., Duffy, J., Chai, S., Jin, Y., Javanmardi, M., 2020. Numerical
investigation of scale effect in self-propelled container ship squat. Appl.
Ocean Res. 99. https://doi.org/10.1016/j.apor.2020.102143

Kok, Z., Jin, Y., Chai, S., Denehy, S., Duffy, J., 2018. URANS prediction of
berthed ship–passing ship interactions. Ships Offshore Struct. 13, 561–574.
https://doi.org/10.1080/17445302.2018.1429136

Kolmogorov, A.N., 1942. Equations of turbulent motion of an incompressible

fluid. Izv. Acad. Sci. USSR Phys 6, 56–58.

Kolmogorov, A.N., 1941. The local structure of turbulence in incompressible

viscous fluid for very large Reynolds numbers. Cr Acad. Sci. URSS 30,
301–305.

Kooij, G.L., Botchev, M.A., Frederix, E.M.A., Geurts, B.J., Horn, S., Lohse, D.,
van der Poel, E.P., Shishkina, O., Stevens, R.J.A.M., Verzicco, R., 2018.
Comparison of computational codes for direct numerical simulations of
turbulent Rayleigh–Bénard convection. Computers and Fluids 166, 1–8.
https://doi.org/10.1016/j.compfluid.2018.01.010

Korkmaz, K.B., Werner, S., Bensow, R.E., 2019. Numerical friction lines for
CFD based form factor determination, in: VIII International Conference
on Computational Methods in Marine Engineering MARINE 2019.

Kornev, N., Abbas, N., 2018. Vorticity structures and turbulence in the wake
of full block ships. J. Mar. Sci. Technol. 23, 567–579.
https://doi.org/10.1007/s00773-017-0493-3

Kornev, N., Shevchuk, I., Abbas, N., Anschau, P., Samarbakhsh, S., 2019.
Potential and limitations of scale resolved simulations for ship
hydrodynamics applications. Sh. Technol. Res. 7255.
https://doi.org/10.1080/09377255.2019.1574965

Kouh, J.S., Chen, Y.J., Chau, S.W., 2009. Numerical study on scale effect of form
factor. Ocean Eng. 36, 403–413.
https://doi.org/10.1016/j.oceaneng.2009.01.011

Kundu, P., Cohen, I.M., Dowling, D.R., 2012. Fluid Mechanics. Elsevier.

Lam, W.H., Hamill, G.A., Robinson, D.J., 2013. Initial wash profiles from a ship
propeller using CFD method. Ocean Eng. 72, 257–266.
https://doi.org/10.1016/j.oceaneng.2013.07.010

263
Lamb, H., 1932. Hydrodynamics. Cambridge Univ. Press 6th revise, 262–264.
https://doi.org/10.1017/CBO9781107415324.004

Landweber, L., Patel, V.C., 1979. Ship Boundary Layers. Annu. Rev. Fluid
Mech. 11, 173–205. https://doi.org/10.1146/annurev.fl.11.010179.001133

Langtry, R.B., 2006. A Correlation-Based Transition Model using Local

Variables for Unstructured Parallelized CFD codes. Univ. Stuttgart.
https://doi.org/10.1115/1.2184352

Larsson, L., Stern, F., Visonneau, M., 2014. Numerical Ship Hydrodynamics:
An assessment of the Gothenburg 2010 Workshop. Springer.
https://doi.org/10.1007/978-94-007-7189-5

Launder, B.E., Sharma, B.I., 1974. Application of the energy-dissipation model

of turbulence to the calculation of flow near a spinning disc. Lett. Heat
Mass Transf. 1, 131–137. https://doi.org/10.1016/0094-4548(74)90150-7

Laurence, D., Uribe, J.C., Utyuzhnikov, S.V., 2004. A robust formulation of the
v2-f model and its applications. Int. J. Numer. Methods Fluids 00, 1–6.

Lax, P.D., Richtmyer, R.D., 1956. Survey of the stability of linear finite
difference equations. Commun. Pure Appl. Math. 9, 267–293.
https://doi.org/10.1002/cpa.3160090206

Lazauskas, L. V, 2009. Resistance, Wave-Making and Wave-Decay of Thin

Ships, with Emphasis on the Effects of Viscosity.

Lea, G.K., Feldman, J.P., 1972. Transcritical flow past slender ships, in: 9th
Symposium on Naval Hydrodynamics. Washington DC, pp. 1527–1542.

Lee, B.W., Lee, C., 2019. Equation for ship wave crests in the entire range of
water depths. Coast. Eng. 153, 103542.
https://doi.org/10.1016/j.coastaleng.2019.103542

Lee, S.J., Kim, H.R., Kim, W.J., Van, S.H., 2003. Wind tunnel tests on flow
characteristics of the KRISO 3,600 TEU containership and 300K VLCC
double-deck ship models. J. Sh. Res. 47, 24–38.

Lee, Y.G., Ha, Y.J., Lee, S.H., Kim, S.H., 2018. A study on the estimation
method of the form factor for a full-scale ship. Brodogradnja 69, 71–87.
https://doi.org/10.21278/brod69105

Lesieur, M., 2008. Turbulence in fluids, 4th ed, Physics of Fluids. Springer.
https://doi.org/10.1063/1.1711203

Li, T., 2003. Computation of turbulent free-surface flows around modern

ships. Int. J. Numer. Methods Fluids 43, 407–430.

264
 https://doi.org/10.1002/fld.580

Liang, H., Chen, X., 2019. Viscous effects on the fundamental solution to ship
waves. J. Fluid Mech. 879, 744–774. https://doi.org/10.1017/jfm.2019.698

Liefvendahl, M., Fureby, C., 2017. Grid requirements for LES of ship
hydrodynamics in model and full scale. Ocean Eng. 143, 259–268.
https://doi.org/10.1016/j.oceaneng.2017.07.055

Lien, F.S., Kalitzin, G., Durbin, P.A., 1998. RANS modeling for compressible
and transitional flows. Cent. Turbul. Res. Summer Progr. 1998 1–20.

Lighthill, M.J., 1990. Waves in Fluids.

Lilek, Ž., Peric, M., 1995. A fourth-order finite volume method with colocated
variable arrangement. Computers and Fluids 24, 239–252.
https://doi.org/10.1016/0045-7930(94)00030-3

Linde, F., Ouahsine, A., Huybrechts, N., Sergent, P., 2016. Three-Dimensional
Numerical Simulation of Ship Resistance in Restricted Waterways: Effect
of Ship Sinkage and Channel Restriction. J. Waterw. Port, Coastal, Ocean
Eng. 143, 06016003. https://doi.org/10.1061/(asce)ww.1943-5460.0000353

Liu, H.W., Lin, P.Z., 2005. Discussion of “Wave transformation by two-

dimensional bathymetric anomalies with sloped transitions” [Coast. Eng.
50 (2003) 61-84]. Coast. Eng. 52, 197–200.
https://doi.org/10.1016/j.coastaleng.2004.11.002

Liu, W., Demirel, Y.K., Djatmiko, E.B., Nugroho, S., Tezdogan, T., Kurt, R.E.,
Supomo, H., Baihaqi, I., Yuan, Z.M., Incecik, A., 2018. Bilge Keel Design
for the Traditional Fishing Boats of Indonesia’s East Java. Int. J. Nav.
Archit. ure Ocean Eng.

Liu, Y., Zou, L., Zou, Z.J., 2017. Computational fluid dynamics prediction of
hydrodynamic forces on a manoeuvring ship including effects of dynamic
sinkage and trim. Proc. Inst. Mech. Eng. Part M J. Eng. Marit. Environ.
https://doi.org/10.1177/1475090217734685

Longo, J., Huang, H.P., Stern, F., 1998. Solid/free-surface juncture boundary
layer and wake. Exp. Fluids 25, 283–297.
https://doi.org/10.1007/s003480050232

Longo, J., Stern, F., 2002. Effects of drift angle on model ship flow. Exp. Fluids
32, 558–569. https://doi.org/10.1007/s00348-001-0397-0

Lopes, R., Eca, L., Vaz, G., 2019. On the Numerical Behavior of Rans-based
Transition Models. J. Fluids Eng. 142, 1–14.

265
 https://doi.org/10.1115/1.4045576

Lopes, R., Eca, L., Vaz, G., 2017. On the Decay of Free-stream Turbulence
Predicted by Two-equation Eddy-viscosity Models. NuTTS-2017,
Wageningen, the Netherlands 1–6.

Ma, C., Zhu, Y., He, J., Zhang, C., Wan, D.C., Yang, C., Noblesse, F., 2018.
Nonlinear corrections of linear potential-flow theory of ship waves. Eur.
J. Mech. B/Fluids 67, 1–14.
https://doi.org/10.1016/j.euromechflu.2017.07.006

Ma, S.J., Zhou, M.G., Zou, Z.J., 2013. Hydrodynamic interaction among hull,
rudder and bank for a ship sailing along a bank in restricted waters. J.
Hydrodyn. 25, 809–817. https://doi.org/10.1016/S1001-6058(13)60428-X

Maasch, M., Mizzi, K., Atlar, M., Fitzsimmons, P., Turan, O., 2019. A generic
wake analysis tool and its application to the Japan Bulk Carrier test case.
Ocean Eng. 171, 575–589. https://doi.org/10.1016/j.oceaneng.2018.12.030

Magionesi, F., Di Mascio, A., 2016. Investigation and modelling of the

turbulent wall pressure fluctuations on the bulbous bow of a ship. J.
Fluids Struct. 67, 219–240.
https://doi.org/10.1016/j.jfluidstructs.2016.09.008

Majidian, H., Azarsina, F., 2019. Numerical simulation of container ship in

oblique winds to develop a wind resistance model based on statistical
data. J. Int. Marit. Safety, Environ. Aff. Shipp. 2, 67–88.
https://doi.org/10.1080/25725084.2018.1564471

Maki, K.J., Broglia, R., Doctors, L.J., Di Mascio, A., 2013. Numerical
investigation of the components of calm-water resistance of a surface-
effect ship. Ocean Eng. 72, 375–385.
https://doi.org/10.1016/j.oceaneng.2013.07.022

Manceau, R., 2015. Recent progress in the development of the Elliptic Blending
Reynolds-stress model. Int. J. Heat Fluid Flow 51, 195–220.
https://doi.org/10.1016/j.ijheatfluidflow.2014.09.002

Manceau, R., Hanjalić, K., 2002. Elliptic blending model: A new near-wall
Reynolds-stress turbulence closure. Phys. Fluids 14, 744–754.
https://doi.org/10.1063/1.1432693

Manceau, R., Wang, M., Laurence, D., 2001. Inhomogeneity and anisotrophy
effects on the redistribution term in Reynolds-averaged Navier-Stoke
modelling. J. Fluid Mech. 438, 307–338.
https://doi.org/10.1017/S0022112001004451

266
Mancini, S., Begovic, E., Day, A.H., Incecik, A., 2018. Verification and
validation of numerical modelling of DTMB 5415 roll decay. Ocean Eng.
162, 209–223. https://doi.org/10.1016/j.oceaneng.2018.05.031

Markatos, N.C., 1986. The mathematical modelling of turbulent flows. Appl.

Math. Model. 10, 190–220. https://doi.org/https://doi.org/10.1016/0307-
904X(86)90045-4

Marquardt, M.W., 2009. Effects of waves and the free surface on a surface-
piercing flat-plate turbulent boundary layer and wake. University of
Iowa.

Marshall, J.S., Naghdi, P.M., 1990. Wave reflection and transmission by steps
and rectangular obstacles in channels of finite depth. Theor. Comput.
Fluid Dyn. 1, 287–301. https://doi.org/10.1007/BF00271583

Mathew, J., Akylas, T.R., 1990. On three-dimensional long water waves in a

channel with sloping sidewalls. J. Fluid Mech. 215, 289–307.
https://doi.org/10.1017/S0022112090002658

Menter, F.R., 1994. Two-equation eddy-viscosity turbulence models for

engineering applications. AIAA J. 32, 1598–1605.
https://doi.org/10.2514/3.12149

Menter, F.R., Smirnov, P.E., Liu, T., Avancha, R., 2015. A One-Equation Local
Correlation-Based Transition Model. Flow, Turbul. Combust. 95, 583–619.
https://doi.org/10.1007/s10494-015-9622-4

Michell, J.H., 1898. The wave-resistance of a ship. London, Edinburgh, Dublin

Philos. Mag. J. Sci. 45, 106–123.

Mihic, S., Golusin, M., Mihajlovic, M., 2011. Policy and promotion of
sustainable inland waterway transport in Europe - Danube River. Renew.
Sustain. Energy Rev. 15, 1801–1809.
https://doi.org/10.1016/j.rser.2010.11.033

Miles, J.W., 1986. Stationary, transcritical channel flow. J. Fluid Mech. 162, 489–
499. https://doi.org/10.1017/S0022112086002136

Millward, A., 1996. A review of the prediction of squat in shallow water. J.

Navig. 77–88. https://doi.org/10.1017/S0373463300013126

Milne-Thomson, L.M., 1962. Theoretical hydrodynamics, MacMillan & Co

LTD.

Min, K.-S., Kang, S.-H., 2010. Study on the form factor and full-scale ship
resistance. J. Mar. Sci. Technol. 15, 108–118.

267
 https://doi.org/10.1007/s00773-009-0077-y

Molland, A.F., Turnock, S.R., Hudson, D.A., 2017. Model-Ship Extrapolation,

in: Ship Resistance and Propulsion: Practical Estimation of Ship
Propulsive Power. Cambridge University Press, Cambridge, pp. 70–85.
https://doi.org/10.1017/9781316494196.006

Moore, G.E., 1965. Cramming more components onto integrated circuits, in:
Proceedings Of The IEEE. pp. 82–85. https://doi.org/10.1109/N-
SSC.2006.4785860

Morton, K.W., Mayers, D.F., 2005. Numerical Solution of Partial Differential

Equations: An Introduction. Cambridge University Press, Cambridge.
https://doi.org/10.1017/CBO9780511812248

Mucha, P., 2017. On Simulation-based Ship Maneuvering Prediction in Deep

and Shallow Water.

Mucha, P., Deng, G., Gourlay, T., Moctar, O. el, 2016. Validation studies on
numerical prediction of ship squat and resistance in shallow water. Proc.
4th MASHCON 83–92. https://doi.org/10.18451/978-3-939230-38-0

Mucha, P., el Moctar, B., 2014. Numerical Prediction of Resistance and Squat
for a Containership in Shallow Water.

Mucha, P., El Moctar, O., Böttner, C.U., 2014. Technical note: PreSquat -
Workshop on numerical prediction of ship Squat in restricted waters. Sh.
Technol. Res. 61, 162–165. https://doi.org/10.1179/str.2014.61.3.004

Müller, C., Herbst, F., 2014. Modelling of crossflow-induced transition based

on local variables, in: 6th European Conference on Computational Fluid
Dynamics (ECFD). Barcelona, Spain, pp. 20–25.

Muzaferija, S., Peric, M., 1999. Computation of free-surface ows using

interface- tracking and interface-capturing methods, in: Chap. 2 in O.
Mahrenholtz and M. Markiewicz (Eds.), Nonlinear Water Wave
Interaction, Computational Mechanics Publications, WIT Press,
Southampton.

Muzaferija, S., Peric, M., 1997. Computation of free-surface flows using the
finite-volume method and moving grids. Numer. Heat Transf. Part B
Fundam. 32, 369–384. https://doi.org/10.1080/10407799708915014

Nakos, D.E., Sclavounos, P.D., 1990. On steady and unsteady ship wave
patterns. J. Fluid Mech. 215, 263–288.
https://doi.org/10.1017/S0022112090002646

268
Newman, J.N., 1992. Panel methods in marine hydrodynamics. 11th Australas.
Fluid Mech. Conf.

Newman, J.N., 1990. Numerical solutions of the water-wave dispersion

relation. Appl. Ocean Res. 12, 14–18. https://doi.org/10.1016/S0141-
1187(05)80013-6

Newman, J.N., 1970. Recent research on ship waves, in: Proceedings of the 8th
Symposium on Naval Hydrodynamics, August 24-28. Rome, Italy, pp.
519–545.

Newman, J.N., 1965. Propagation of water waves over an infinite step. J. Fluid
Mech. 23, 399–415. https://doi.org/10.1017/S0022112065001453

Niklas, K., Pruszko, H., 2019a. Full-scale CFD simulations for the
determination of ship resistance as a rational, alternative method to
towing tank experiments. Ocean Eng. 190, 106435.
https://doi.org/10.1016/j.oceaneng.2019.106435

Niklas, K., Pruszko, H., 2019b. Full scale CFD seakeeping simulations for case
study ship redesigned from V-shaped bulbous bow to X-bow hull form.
Appl. Ocean Res. 89, 188–201. https://doi.org/10.1016/j.apor.2019.05.011

Noblesse, F., He, J., Zhu, Y., Hong, L., Zhang, C., Zhu, R., 2014. Why can ship
wakes appear narrower than Kelvin’s angle ? Eur. J. Mech. B/Fluids 46,
164–171. https://doi.org/10.1016/j.euromechflu.2014.03.012

Nwogu, O., 1993. Alternative form of Boussinesq equations for nearshore

wave propagation. J. Waterw. Port, Coast. Ocean Eng. 119, 618–638.
https://doi.org/10.1061/(ASCE)0733-950X(1993)119:6(618)

Oberkampf, W.L., Blottner, F.G., 1998. Issues in computational fluid dynamics

code verification and validation. AIAA J. 36, 687–695.
https://doi.org/10.2514/2.456

Oggiano, L., Pierella, F., Nygaard, T.A., De Vaal, J., Arens, E., 2017.
Reproduction of steep long crested irregular waves with CFD using the
VOF method, in: Energy Procedia.
https://doi.org/10.1016/j.egypro.2017.10.351

Oh, K.J., Kang, S.H., 1992. Full scale Reynolds number effects for the viscous
flow around the ship stern. Comput. Mech. 9, 85–94.
https://doi.org/10.1007/BF00370064

Ohashi, K., Kobayashi, H., Hino, T., 2018. Numerical simulation of the free-
running of a ship using the propeller model and dynamic overset grid
method. Sh. Technol. Res. 65, 153–162.

269
 https://doi.org/10.1080/09377255.2018.1482610

Ozdemir, Y.H., Barlas, B., 2017. Numerical study of ship motions and added
resistance in regular incident waves of KVLCC2 model. Int. J. Nav. Archit.
Ocean Eng. 9, 149–159. https://doi.org/10.1016/j.ijnaoe.2016.09.001

Ozdemir, Y.H., Cosgun, T., Dogrul, A., Barlas, B., 2016. A numerical
application to predict the resistance and wave pattern of KRISO
containership. Brodogradnja 67, 47–65.

Pacuraru, F., Domnisoru, L., 2017. Numerical investigation of shallow water

effect on a barge ship resistance. IOP Conf. Ser. Mater. Sci. Eng. 227.
https://doi.org/10.1088/1757-899X/227/1/012088

Parneix, S., Durbin, P.A., Behnia, M., 1998. Computation of 3-D turbulent
boundary layers using the V2F model. Flow, Turbul. Combust. 60, 19–46.
https://doi.org/10.1023/A:1009986925097

Patel, V.C., Rodi, W., Scheuerer, G., 1984. Turbulence models for near-wall and
low Reynolds number flows - A review. AIAA J. 23, 1308–1319.
https://doi.org/10.2514/3.9086

Patel, V.C., Sarda, O.P., 1990. Mean-flow and turbulence measurements in the
boundary layer and and wake of a ship double model. Exp. Fluids 8, 319–
335. https://doi.org/https://doi.org/10.1007/BF00217197

Peltier, L.J., Hambric, S.A., 2007. Estimating turbulent-boundary-layer wall-

pressure spectra from CFD RANS solutions. J. Fluids Struct. 23, 920–937.
https://doi.org/10.1016/j.jfluidstructs.2007.01.003

Pena, B., Muk-Pavic, E., Ponkratov, D., 2019. Achieving a high accuracy
numerical simulations of the flow around a full scale ship, in: Proceedings
of the International Conference on Offshore Mechanics and Arctic
Engineering - OMAE. pp. 1–10. https://doi.org/10.1115/OMAE2019-95769

Peregrine, D.H., 1967. Long waves on a beach. J. Fluid Mech. 27, 815–827.
https://doi.org/10.1017/S0022112067002605

Pereira, F.S., Eca, L., Vaz, G., 2017. Verification and Validation exercises for the
flow around the KVLCC2 tanker at model and full-scale Reynolds
numbers. Ocean Eng. 129, 133–148.
https://doi.org/10.1016/j.oceaneng.2016.11.005

Peric, M., 2019. White paper: Full-scale simulation for marine design. Siemens
White Pap.

Perić, R., Abdel-Maksoud, M., 2016. Reliable damping of free-surface waves in

270
 numerical simulations. Sh. Technol. Res. 63, 1–13.
https://doi.org/10.1080/09377255.2015.1119921

Pethiyagoda, R., McCue, S.W., Moroney, T.J., 2017. Spectrograms of ship

wakes: Identifying linear and nonlinear wave signals. J. Fluid Mech. 811,
189–209. https://doi.org/10.1017/jfm.2016.753

Pethiyagoda, R., McCue, S.W., Moroney, T.J., 2014. What is the apparent angle
of a Kelvin ship wave pattern? J. Fluid Mech. 758, 468–485.
https://doi.org/10.1017/jfm.2014.530

Pethiyagoda, R., Moroney, T.J., Macfarlane, G.J., Binns, J.R., McCue, S.W.,
2018. Time-frequency analysis of ship wave patterns in shallow water:
modelling and experiments. Ocean Eng. 158, 123–131.
https://doi.org/10.1016/j.oceaneng.2018.01.108

Pettersson Reif, B.A., 2006. Towards a nonlinear eddy-viscosity model based

on elliptic relaxation. Flow, Turbul. Combust. 76, 241–256.
https://doi.org/10.1007/s10494-006-9013-y

Pettersson Reif, B.A., Mortensen, M., Langer, C.A., 2009. Towards sensitizing
the nonlinear v 2-f model to turbulence structures. Flow, Turbul.
Combust. 83, 185–203. https://doi.org/10.1007/s10494-008-9194-7

Phillips, O.M., 1955. The irrotational motion outside a free turbulent

boundary. Math. Proc. Cambridge Philos. Soc. 51, 220–229.
https://doi.org/10.1017/S0305004100030073

Phillips, T., 2014. Residual-based Discretization Error Estimation for

Computational Fluid Dynamics 148.

Phillips, T., 2012. Extrapolation-based Discretization Error and Uncertainty

Estimation in Computational Fluid Dynamics.

Phillips, T., Roy, C.J., 2017. A New Extrapolation-Based Uncertainty Estimator

for Computational Fluid Dynamics. J. Verif. Valid. Uncertain. Quantif. 1,
041006. https://doi.org/10.1115/1.4035666

Phillips, T., Roy, C.J., 2012. Evaluation of Extrapolation-Based Discretization

Error and Uncertainty Estimators, AIAA 2011-215. 49th AIAA Aerosp.
Sci. Meet. Incl. New Horizons Forum Aerosp. Expo. Orlando, F, 1–18.
https://doi.org/10.2514/6.2011-215

Piomelli, U., Balaras, E., 2002. Wall Layer Models for Large Eddy Simulations.
Annu. Rev. Fluid Mech. 34, 349–374.
https://doi.org/10.1146/annurev.fluid.34.082901.144919

271
Plotkin, A., 1977. Slender-ship shallow-water flow past a slowly varying
bottom. J. Eng. Math. 11, 289–297. https://doi.org/10.1007/BF01537089

Plotkin, A., 1976. The flow due to a slender ship moving over a wavy wall in
shallow water. J. Eng. Math. 10, 207–218.
https://doi.org/10.1007/BF01535383

Ponkratov, D., 2016. Lloyd’s Register workshop on ship scale hydrodynamics,

in: Ponkratov, D. (Ed.), 2016 Workshop on Ship Scale Hydrodynamic
Computer Simulation. p. 2016. https://doi.org/10.1002/ejoc.201200111

Posa, A., Broglia, R., Felli, M., Falchi, M., Balaras, E., 2019. Characterization of
the wake of a submarine propeller via Large-Eddy simulation. Computers
and Fluids 184, 138–152. https://doi.org/10.1016/j.compfluid.2019.03.011

Prakash, S.M.N., Chandra, B., 2013. Numerical Estimation of Shallow Water

Resistance of a River-Sea Ship using CFD. Int. J. Comput. Appl. 71, 33–40.
https://doi.org/https://doi.org/10.5120/12357-8670

Prandtl, L., 1925. Report on the investigation of developed turbulence,

Translation of “ Bericht über Untersuchungen zur ausgebildeten
Turbulenz.” Zeitschrift für angewandte Mathematik und Mechanik, vol.
5, no. 2, April 1925.

Quérard, A., Temarel, P., Turnock, S.R., 2008. Influence of viscous effects on
the hydrodynamics of ships-like sections undergoing symmetric and anti-
symmetric motions, using RANS, in: Proceedings of the ASME 27th
International Confer Ence on Offshore Mechanics and Arctic Engineering
OMAE2008, Estoril, Portugal. pp. 1–10.
https://doi.org/10.1115/OMAE2008-57330

Queutey, P., Visonneau, M., 2015. Free-Surface Capturing RANSE Simulations

for a Ship at Steady Drift. Sh. Technol. Res. 51, 106–122.
https://doi.org/10.1179/str.2004.51.3.003

Rabaud, M., Moisy, F., 2013. Ship wakes: Kelvin or mach angle? Phys. Rev.
Lett. 110, 1–5. https://doi.org/10.1103/PhysRevLett.110.214503

Raven, H.C., 2019. Shallow-water effects in ship model testing and at full scale.
Ocean Eng. 189, 106343. https://doi.org/10.1016/j.oceaneng.2019.106343

Raven, H.C., van der Ploeg, A., Starke, A.R., Eca, L., 2008. Towards a CFD-
based prediction of ship performance—progress in predicting full-scale
resistance and scale effects. Int. J. Marit. Eng. 150.

Razgallah, I., Kaidi, S., Smaoui, H., Sergent, P., 2018. The impact of free surface
modelling on hydrodynamic forces for ship navigating in inland

272
 waterways: water depth, drift angle, and ship speed effect. J. Mar. Sci.
Technol. 1–22. https://doi.org/10.1007/s00773-018-0566-y

Revell, A.J., Benhamadouche, S., Craft, T., Laurence, D., Yaqobi, K., 2005. A
Stress-Strain Lag Eddy Viscosity Model for Unsteady Mean Flow. Eng.
Turbul. Model. Exp. 6 27, 117–126. https://doi.org/10.1016/B978-
008044544-1/50010-8

Richardson, L.F., 1927. Deferred approach to the limit. Trans. R. Soc. London,
Ser. A 226, 299–361.

Richardson, L.F., 1922. Weather prediction by numerical process, 1st ed.

Cambridge University Press.

Richardson, L.F., 1911. The Approximate Arithmetical Solution by Finite

Differences of Physical Problems involving Differential Equations, with
an Application to the Stresses in a Masonry Dam. Philos. Trans. th R. Soc.
London,Containing Pap. a Math. Phys. Character 210, 307–357.

Roache, P.J., 2016. Verification and Validation in Fluids Engineering: Some

Current Issues. J. Fluids Eng. 138, 101205.
https://doi.org/10.1115/1.4033979

Roache, P. J., 1998. Validation and Verification in computational science and

engineering. Hermosa Albuquerque, NM.

Roache, Patrick J., 1998. Verification of codes and calculations. AIAA J. 36, 696–
702. https://doi.org/10.2514/2.457

Roache, P.J., 1997. Quantification of Uncertainty in Computational Fluid

Dynamics. Annu. Rev. Fluid Mech. 29, 123–160.
https://doi.org/10.1146/annurev.fluid.29.1.123

Roache, P.J., Ghia, K.N., White, F., 1986. Editorial Policy Statement on the
Control of Numerical Accuracy. J. Fluids Eng. 108, 2.

Roache, P.J., Knupp, P.M., 1993. Completed richardson extrapolation.

Commun. Numer. Methods Eng. 9.

Robertson, E., Choudhury, V., Bhushan, S., Walters, D.K., 2015. Validation of
OpenFOAM numerical methods and turbulence models for
incompressible bluff body flows. Computers and Fluids 123, 122–145.
https://doi.org/10.1016/j.compfluid.2015.09.010

Rotta, J., 1951. Statistical Theory of Inhomogeneous Turbulence. NASA Tech.

Doc.

Rotteveel, E., Hekkenberg, R.G., 2015. The Influence of Shallow Water and

273
 Hull Form Variations on Inland Ship Resistance 2, 220–236.

Roy, C.J., 2008. Grid Convergence Error Analysis for Mixed-Order Numerical
Schemes. AIAA J. 41, 595–604. https://doi.org/10.2514/2.2013

Roy, C.J., 2005. Review of code and solution verification procedures for
computational simulation. J. Comput. Phys. 205, 131–156.
https://doi.org/10.1016/j.jcp.2004.10.036

Roy, C.J., Blottner, F.G., 2006. Review and Assessment of Turbulence Models
for Hypersonic Flows: 2D/Asymmetric Cases. 44th AIAA Aerosp. Sci.
Meet. Exhib. https://doi.org/10.2514/6.2006-713

Roy, C.J., Blottner, F.G., 2001. Assessment of One- and Two-Equation

Turbulence Models for Hypersonic Transitional Flows. J. Spacecr. Rockets
38, 699–710. https://doi.org/10.2514/2.3755

Rozman, S., 2009. Wake pattern of a boat. (Doctoral Diss. Ljubljana, Slov. Univ.
Ljubliana).

Saaty, L.T., Bram, J., 1964. Nonlinear mathematics, Vol 12. ed. McGraw-Hill.

Saffman, P.G., 1970. A Model for Inhomogeneous Turbulent Flow. Proc. R. Soc.
A Math. Phys. Eng. Sci. 317, 417–433.
https://doi.org/10.1098/rspa.1970.0125

Saffman, P.G., Wilcox, D.C., 1974. Turbulence-Model Predictions for

Turbulent Boundary Layers. AIAA J. 12, 541–546.
https://doi.org/10.2514/3.49282

Salas, M.D., 2006. Some observations on grid convergence. Computers and

Fluids 35, 688–692. https://doi.org/10.1016/j.compfluid.2006.01.003

Salas, M.D., Atkins, H.L., 2009. Problems associated with grid convergence of
functionals. Comput. Fluid Dyn. 2008 38, 309–314.
https://doi.org/10.1016/j.compfluid.2008.01.015

Salim, M., Cheah, S.C., 2009. Wall y + strategy for dealing with wall-bounded
turbulent flows. Int. MultiConference Eng. Comput. Sci. II, 1–6.
https://doi.org/10.1.1.149.722

Saric, W.S., Reed, H.L., White, E.B., 2002. Stability and Transition of three -
Dimensional Boundary Layers. Annu. Rev. Fluid Mech. 35, 413–440.
https://doi.org/10.1146/annurev.fluid.35.101101.161045

Sarkar, S., Lakshmanan, B., 1991. Application of a Reynolds stress turbulence

model to the compressible shear layer. AIAA J. 29, 743–749.
https://doi.org/10.2514/3.10649

274
Saydam, A.Z., Taylan, M., 2018. Evaluation of wind loads on ships by CFD
analysis. Ocean Eng. 158, 54–63.
https://doi.org/10.1016/j.oceaneng.2018.03.071

Schlichting, H., 1979. Boundary-Layer Theory, 7th ed. McGraw-Hill.

https://doi.org/10.1007/978-3-662-52919-5

Schoellhamer, D.H., 1996. Anthropogenic sediment resuspension mechanisms

in a shallow microtidal estuary. Estuar. Coast. Shelf Sci. 43, 533–548.
https://doi.org/10.1006/ecss.1996.0086

Schoenherr, K., 1932. Resistance of flat surfaces moving through a fluid. Trans.
Soc. Nav. Arch. Mar. Eng. 40, 279–313.

Schweighofer, J., 2004. Numerical Investigation of the Turbulent Free-Surface

Flow around the Series 60 Ship at Model- and Full-Scale Ship Reynolds
Numbers. Yearb. Soc. Nav. Archit. Ger. STG Summer Meet. Pol. 1–10.

Schweighofer, J., Regnstr, B., 2005. Viscous-Flow Computations of Two

Existing Vessels At Model- and Full-Scale Ship Reynolds Numbers - a
Study Carried Out Within the European Union Project , Effort ., in:
International Conference on Computational Methods in Marine
Engineering. pp. 1–11.

Sezen, S., Cakici, F., 2019. Numerical Prediction of Total Resistance Using Full
Similarity Technique. China Ocean Eng. 33, 493–502.
https://doi.org/10.1007/s13344-019-0047-z

Sharma, S.D., 1995. A Slender Ship Moving at a Near-Critical Speed in a

Shallow Channel. J. Fluid Mech. 291, 263–285.
https://doi.org/10.1017/S0022112095002692

Shenoi, R.R., Krishnankutty, P., Panneer Selvam, R., 2016. Study of

manoeuvrability of container ship by static and dynamic simulations
using a RANSE-based solver. Ships Offshore Struct. 11, 316–334.
https://doi.org/10.1080/17445302.2014.987439

Shevchuk, I., Bottner, C.-U., Kornev, N., 2019. Numerical investigation of scale
effects on squat in shallow water, in: 5th MASHCON, Ostend, Belgium.
pp. 410–422.

Shevchuk, I., Böttner, C.U., Kornev, N., 2016. Numerical Analysis of the Flow
in the Gap Between the Ship Hull and the Fairway Bottom in Extremely
Shallow Water. Proc. 4th Int. Conf. Sh. Manoeuvring Shallow Confin.
Water (MASHCON), 23 - 25 May 2016, Hamburg, Ger. 0, 37–42.
https://doi.org/10.18451/978-3-939230-38-0

275
Shevchuk, I., Kornev, N., 2017. Study of unsteady hydrodynamic effects in the
stern area of river cruisers in shallow water. 7th Int. Conf. Comput.
Methods Mar. Eng. Mar. 2017 2017-May, 440–448.
https://doi.org/10.1080/09377255.2017.1349599

Shi, A., Wu, M., Yang, B., Wang, X., Wang, Z., 2012. Resistance calculation and
motions simulation for free surface ship based on CFD. Procedia Eng. 31,
68–74. https://doi.org/10.1016/j.proeng.2012.01.992

Shih, T., 1990. An improved k-epsilon model for near-wall turbulence and
comparison with direct numerical simulation. NASA STI/Recon Tech.
Rep. N 0–21.

Shih, T.H., Liou, W.W., Shabbir, A., Yang, Z., Zhu, J., 1995. A new k-ε eddy
viscosity model for high reynolds number turbulent flows. Computers
and Fluids 24, 227–238. https://doi.org/10.1016/0045-7930(94)00032-T

Shivachev, E., Khorasanchi, M., Day, A.H., 2017. Trim influence on KRISO
container ship (KCS); an experimental and numerical study, in:
Proceedings of the ASME 2017 36th International Conference on Ocean,
Offshore and Arctic Engineering. pp. 1–7.
https://doi.org/10.1115/OMAE2017-61860

Siemens, 2018. Star-CCM+ User Guide version 13.04.

Sigmund, S., Moctar, O., 2018. Numerical and experimental investigation of

added resistance of different ship types in short and long waves. Ocean
Eng. 147, 51–67. https://doi.org/10.1016/j.oceaneng.2017.10.010

Simonsen, C.D., Otzen, J.F., Joncquez, S., Stern, F., 2013. EFD and CFD for KCS
heaving and pitching in regular head waves. J. Mar. Sci. Technol. 18, 435–
459. https://doi.org/10.1007/s00773-013-0219-0

Slotnick, J., Khodadoust, A., Alonso, J., Darmofal, D., Gropp, W., Lurie, E.,
Mavriplis, D., 2014. CFD Vision 2030 Study: A Path to Revolutionary
Computational Aerosciences, NASA/CR–2014-218178.
https://doi.org/10.1017/CBO9781107415324.004

Song, K., Guo, C., Wang, C., Sun, C., Li, P., Zhong, R., 2019. Experimental and
numerical study on the scale effect of stern flap on ship resistance and
flow field. Ships Offshore Struct. 0, 1–17.
https://doi.org/10.1080/17445302.2019.1697091

Song, S., Demirel, Y.K., Atlar, M., 2019. An investigation into the effect of
biofouling on the ship hydrodynamic characteristics using CFD. Ocean
Eng. 175, 122–137. https://doi.org/10.1016/j.oceaneng.2019.01.056

276
Soomere, T., 2007. Nonlinear components of ship wake waves. Appl. Mech.
Rev. 60, 120–138. https://doi.org/10.1115/1.2730847

Soomere, Tarmo, 2009. Long ship waves in shallow water bodies, in: Quak, E.,
Soomere, T. (Eds.), Applied Wave Mathematics: Selected Topics in Solids,
Fluids, and Mathematical Methods. pp. 193–228.
https://doi.org/10.1007/978-3-642-00585-5_12

Sorensen, R.M., 1997. Prediction of Vessel-Generated Waves with Reference to

Vessels Common to the Upper Mississippi River System, US Army Corps
of Engineers, Waterways Experiment Station, ENV Report 4.

Spalart, P., Allmaras, S., 1992. A one-equation turbulence model for

aerodynamic flows. 30th Aerosp. Sci. Meet. Exhib.

Srividya, K., Thandaveswara, B.S., 2005. CFD analysis for sea-chest design.
ISH J. Hydraul. Eng. 11, 58–72.
https://doi.org/10.1080/09715010.2005.10514781

St Denis, M., Pierson, W.J., 1953. On the Motions of Ships in Confused Seas.
Trans. Soc. Nav. Archit. Mar. Eng. 61, 280–357.

Starke, A.R., Drakopoulos, K., Toxopeus, S.L., Turnock, S.R., 2017. RANS-
based full-scale power predictions for a general cargo vessel, and
comparison with sea-trial results. 7th Int. Conf. Comput. Methods Mar.
Eng. Mar. 2017 2017-May, 353–364.

Stern, F., 1985. Effects of Waves on the Boundary Layer of a Surface-Piercing

Body.

Stern, F., Wilson, R., Shao, J., 2006. Quantitative V&V of CFD simulations and
certification of CFD codes. Int. J. Numer. Methods Fluids 50, 1335–1355.
https://doi.org/10.1002/fld.1090

Stern, F., Wilson, R. V., Coleman, H.W., Paterson, E.G., 2001. Comprehensive
approach to verification and validation of CFD simulations—Part 1:
Methodology and procedures. J. Fluids Eng. Trans. ASME 123, 793–802.
https://doi.org/10.1115/1.1412235

Stern, F., Yang, J., Wang, Z., Sadat-Hosseini, H., Mousaviraad, M., 2013.
Computational ship hydrodynamics: Nowadays and way forward. Int.
Shipbuild. Prog. 60, 3–105. https://doi.org/10.3233/ISP-130090

Stroh, A., 2016. Control of Spatially Developing TurbulentBoundary Layers

for Skin Friction DragReduction, in: Nagel W., Kröner D., Resch M. (Eds)
High Performance Computing in Science and Engineering ´16.
https://doi.org/10.1109/CAHPC.2005.27

277
Suez Canal Authority, 2019. Suez Canal Rules of Navigation.
https://doi.org/https://www.suezcanal.gov.eg/English/Navigation/Pages
/RulesOfNavigation.aspx

Suez Canal Authority, 2018. Suez Canal Traffic Statistics: Annual Report 2018.

Suh, J., Yang, J., Stern, F., 2011. The effect of air-water interface on the vortex
shedding from a vertical circular cylinder. J. Fluids Struct. 27, 1–22.
https://doi.org/10.1016/j.jfluidstructs.2010.09.001

Sun, W., Hu, Q., Hu, S., Su, J., Xu, J., Wei, J., 2020. Numerical Analysis of Full-
Scale Ship Self-Propulsion Performance with Direct Comparison to
Statistical Sea Trail Results. J. Mar. Sci. Eng. 8, 1–22.

Tahara, Y., Katsui, T., Himeno, Y., 2002a. Computation of Ship Viscous Flow
at Full Scale Reynolds Number. J. Soc. Nav. Archit. Japan 92, 89–101.

Tahara, Y., Longo, J., Stern, F., 2002b. Comparison of CFD and EFD for the
Series 60 C B = 0.6 in steady drift motion, J Mar Sci Technol.

Tahara, Y., Stern, F., 1994. Validation of an interactive approach for calculating
ship boundary layers and wakes for nonzero froude number. Computers
and Fluids 23, 785–816. https://doi.org/10.1016/0045-7930(94)90066-3

Tatinclaux, B.J., 1970. Effect of a Rotational Wake on the Wavemaking

Resistance of an Ogive. J. Sh. Res. 14, 84–99.

Telfer, E. V., 1927. Ship resistance similarity. Trans. R. Inst. Nav. Archit. 69,
174–190.

Terziev, M., Tezdogan, T., Incecik, A., 2019a. A geosim analysis of ship
resistance decomposition and scale effects with the aid of CFD. Appl.
Ocean Res. 92. https://doi.org/10.1016/j.apor.2019.101930

Terziev, M., Tezdogan, T., Incecik, A., 2019b. Application of eddy-viscosity

turbulence models to problems in ship hydrodynamics. Ships Offshore
Struct. 1–24. https://doi.org/10.1080/17445302.2019.1661625

Terziev, M., Tezdogan, T., Oguz, E., Gourlay, T., Demirel, Y.K., Incecik, A.,
2018. Numerical investigation of the behaviour and performance of ships
advancing through restricted shallow waters. J. Fluids Struct. 76, 185–215.
https://doi.org/10.1016/j.jfluidstructs.2017.10.003

Terziev, M., Zhao, G., Tezdogan, T., Yuan, Z., Incecik, A., 2020. Virtual Replica
of a Towing Tank Experiment to Determine the Kelvin Half-Angle of a
Ship in Restricted Water. J. Mar. Sci. Appl. 8, 1–24.
https://doi.org/10.3390/jmse8040258

278
Tezdogan, T., Demirel, Y.K., Kellett, P., Khorasanchi, M., Incecik, A., Turan,
O., 2015. Full-scale unsteady RANS CFD simulations of ship behaviour
and performance in head seas due to slow steaming. Ocean Eng. 97, 186–
206. https://doi.org/10.1016/j.oceaneng.2015.01.011

Tezdogan, T., Incecik, A., Turan, O., 2016a. A numerical investigation of the
squat and resistance of ships advancing through a canal using CFD. J.
Mar. Sci. Technol. 21, 86–101. https://doi.org/10.1007/s00773-015-0334-1

Tezdogan, T., Incecik, A., Turan, O., 2016b. Full-scale unsteady RANS
simulations of vertical ship motions in shallow water. Ocean Eng. 123,
131–145. https://doi.org/10.1016/j.oceaneng.2016.06.047

Tezdogan, T., Incecik, A., Turan, O., Kellett, P., 2016c. Assessing the Impact of
a Slow Steaming Approach on Reducing the Fuel Consumption of a
Containership Advancing in Head Seas. Transp. Res. Procedia 14, 1659–
1668. https://doi.org/10.1016/j.trpro.2016.05.131

Thomas, J.L., Langley, N., 2008. Toward Verification of Unstructured-Grid

Solvers. AIAA J. 46. https://doi.org/10.2514/1.36655

Thomson, W., 1887. On ship waves, in: Proceedings of the Institution of

Mechanical Engineers. pp. 409–434.
https://doi.org/https://doi.org/10.1243/PIME_PROC_1887_038_028_02

Tin Htwe, N.T., Hino, T., Suzuki, K., 2015. Computation of Free Surface Flows
around Box-Shaped Ships by an Unstructured Navier-Stokes Solver. Sh.
Technol. Res. 60, 104–117. https://doi.org/10.1179/str.2013.60.3.001

Toki, N., 2008. Investigation on Correlation Lines through the Analyses of

Geosim Model Test Results. J. Japan Soc. Nav. Archit. Ocean Eng. 8, 71–
79.

Torsvik, T., 2009. Modelling of ship waves from high-speed vessels, in:
Applied Wave Mathematics: Selected Topics in Solids, Fluids, and
Mathematical Methods. Springer Berlin Heidelberg, pp. 229–263.

Torsvik, T., Dysthe, K., Pedersen, G., 2006. Influence of variable Froude
number on waves generated by ships in shallow water. Phys. Fluids 18.
https://doi.org/10.1063/1.2212988

Torsvik, T., Pedersen, G., Dysthe, K., 2009. Waves Generated by a Pressure
Disturbance Moving in a Channel with a Variable Cross-Sectional
Topography. J. Waterw. Port Coast. Ocean Eng. 135, 120–123.
https://doi.org/Doi 10.1061/(Asce)0733-950x(2009)135:3(120)

Townsin, R.L., 1971. The viscous drag of a “Victory” model—Results from

279
 wake and wave pattern measurements. Trans. RINA 113, 307–321.

Townsin, R.L., 1968. Viscous drag from a wake survey. Measurements in the
wake of a ‘Lucy Ashton’ mode. Trans. RINA 110.

Toxopeus, S.L., 2013. Viscous-flow calculations for KVLCC2 in deep and

shallow water. Comput. Methods Appl. Sci. 29, 151–169.
https://doi.org/10.1007/978-94-007-6143-8_9

Toxopeus, S.L., Simonsen, C.D., Guilmineau, E., Visonneau, M., Xing, T., Stern,
F., 2013. Investigation of water depth and basin wall effects on KVLCC2
in manoeuvring motion using viscous-flow calculations. J. Mar. Sci.
Technol. 18, 471–496. https://doi.org/10.1007/s00773-013-0221-6

Troesch, A., Beck, R.F., 1974. Experiments on ship motions in shallow water.

Tropea, C., Yarin, A.L., Foss, J.F., 2007. Springer handbook of experimental
fluid mechanics, Springer Berlin, Heidelberg. https://doi.org/10.1007/978-
3-540-30299-5

Tuck, E.O., 1978. Hydrodynamic Problems of Ships in Restricted Waters.

Annu. Rev. Fluid Mech. 10, 33–46.

Tuck, E.O., 1967. Sinkage and Trim in Shallow Water of Finite Width.
Schiffstechnik 14, 92–94.

Tuck, E.O., 1966. Shallow-Water Flows Past Slender Bodies. J. Fluid Mech. 26,
81–95. https://doi.org/10.1017/S0022112066001101

Tuck, E.O., Lazauskas, L. V, 2008. Drag on a ship and Michell’s integral, in:
22nd Int. Congress of Theoretical and Applied Mechanics, Adelaide,
South Australia, August.

Tuck, E.O., Taylor, J.P., 1970. Shallow wave problems in ship hydrodynamics,
in: 8th Symposium Naval Hydrodynamics. pp. 627–659.

Tunaley, J.K.E., 2014. Ship Wakes in Shallow Waters. LRDC Rep. 6–9.

UNFCCC, 1998. Kyoto Protocol To the United Nations Framework Kyoto

Protocol To the United Nations Framework, United Nations.
https://doi.org/10.1111/1467-9388.00150

van der Poel, E.P., Ostilla-Mónico, R., Donners, J., Verzicco, R., 2015. A pencil
distributed finite difference code for strongly turbulent wall-bounded
flows. Computers and Fluids 116, 10–16.
https://doi.org/10.1016/j.compfluid.2015.04.007

van Driest, E.R., 1956. On Skin Friction and Heat Transfer Near the Stagnation

280
 Point, NACA Report, AL-2267.

van Mannen, J.D., van Oossanem, P., 1988. Principles of Naval Architecture.
The Society of Naval Architects and Marine Engineers, Jersey.

Van, S.H., Ahn, H., Lee, Y.Y., Kim, C., Hwang, S., Kin, J., Kim, S.K., Park, I.R.,
2011. Resistance Characteristics And Form Factor Evaluation For Geosim
Models Of KVLCC2 And KCS, in: Advanced Model Measurement
Technology for EU Maritime Industry - AMT’11. pp. 282–293.

van Strien, M., 2014. On the origins and foundations of laplacian determinism.
Stud. Hist. Philos. Sci. Part A 45, 24–31.
https://doi.org/10.1016/j.shpsa.2013.12.003

van Wijngaarden, E., 2005. Recent developments in predicting propeller-

induced hull pressure pulses. Proc. 1st Int. Sh. Noise Vib. Conf. 1–8.

Vanka, S.P., 1987. Second-order upwind differencing in a recirculating flow.

AIAA J. 25, 1435–1441. https://doi.org/10.2514/3.9801

Velázquez, M.J.N., Asuero, A.G., 2017. Youden Two-Sample Method, in:

Quality Control and Assurance-An Ancient Greek Term Re-Mastered.
IntechOpen.

Visonneau, M., 2005. A Step Towards the Numerical Simulation of Viscous

Flows Around Ships at Full-Scale. Mar. CFD.

Wackers, J., Koren, B., Raven, H.C., van der Ploeg, A., Starke, A.R., Deng, G.B.,
Queutey, P., Visonneau, M., Hino, T., Ohashi, K., 2011. Free-Surface
Viscous Flow Solution Methods for Ship Hydrodynamics. Arch. Comput.
Methods Eng. 18, 1–41. https://doi.org/10.1007/s11831-011-9059-4

Wang, J., Yu, H., Zhang, Y., Xiong, X., 2016. CFD-based method of determining
form factor k for different ship types and different drafts. J. Mar. Sci. Appl.
15, 236–241. https://doi.org/10.1007/s11804-016-1372-8

Wang, J., Zou, L., Wan, D.C., 2018. Numerical simulations of zigzag maneuver
of free running ship in waves by RANS-Overset grid method. Ocean Eng.
162, 55–79. https://doi.org/10.1016/j.oceaneng.2018.05.021

Wang, J., Zou, L., Wan, D.C., 2017. CFD simulations of free running ship under
course keeping control. Ocean Eng. 141, 450–464.
https://doi.org/10.1016/j.oceaneng.2017.06.052

Wang, Z.Z., Xiong, Y., Shi, L.P., Liu, Z.H., 2015. A numerical flat plate friction
line and its application. J. Hydrodyn. 27, 383–393.
https://doi.org/10.1016/S1001-6058(15)60496-6

281
Web of Science, 2020. Publication statistics [WWW Document]. Clarivate Anal.
URL
https://wcs.webofknowledge.com/RA/analyze.do?product=WOS&SID=F
6qCjOD5enYzEEUuVEv&field=PY_PublicationYear_PublicationYear_en
&yearSort=true (accessed 4.26.20).

Weymouth, G.D., Wilson, R., Stern, F., 2005. Rans computational fluid
dynamics predictions of pitch and heave ship motions in head seas. J. Sh.
Res. 49, 80–97.

White, F., 2010. Fluid Mechanics. McGraw-Hill,New York 862.

https://doi.org/10.1111/j.1549-8719.2009.00016.x.Mechanobiology

White, F., 2006. Viscous fluid flow, 3rd ed. McGraw-Hill.

Whitham, G.B., 2011. Linear and nonlinear waves.

Wilcox, D.C., 2008. Formulation of the k-w Turbulence Model Revisited. AIAA
J. 46, 2823–2838. https://doi.org/10.2514/1.36541

Wilcox, D.C., 2006. Turbulence modeling for CFD, 3rd ed, Transportation
Research Record. DCW Industries.
https://doi.org/10.1016/j.aqpro.2013.07.003

Wilcox, D.C., 1988. Reassessment of the scale-determining equation for

advanced turbulence models. AIAA J. 26, 1299–1310.
https://doi.org/10.2514/3.10041

Wilson, R. V., Carrica, P.M., Stern, F., 2006. URANS simulations for a high-
speed transom stern ship with breaking waves. Int. J. Comut. Fluid Dyn.
20, 105–125. https://doi.org/10.1080/10618560600780916

Witherden, F.D., Jameson, A., 2017. Future directions of computational fluid

dynamics, in: 23rd AIAA Computational Fluid Dynamics Conference,
2017. pp. 1–16.

Wnęk, A.D., Sutulo, S., Soares, C.G., 2018. CFD Analysis of Ship-to-Ship
Hydrodynamic Interaction. J. Mar. Sci. Appl. 17, 21–37.
https://doi.org/https://doi.org/10.1007/s11804-018-0010-z

Wolfshtein, M., 1969. The velocity and temperature distribution in one-

dimensional flow with turbulence augmentation and pressure gradient.
Int. J. Heat Mass Transf. 12, 301–318. https://doi.org/10.1016/0017-
9310(69)90012-X

Wolter, C., Arlinghaus, R., 2003. Navigation impacts on freshwater fish

assemblages: The ecological relevance of swimming performance. Rev.

282
 Fish Biol. Fish. 13, 63–89. https://doi.org/10.1023/A:1026350223459

Wortley, S., 2013. CFD Analysis of Container Ship Sinkage, Trim and
Resistance.

Wu, C.S., Zhou, D.C., Gao, L., Miao, Q.M., 2011. CFD computation of ship
motions and added resistance for a high speed trimaran in regular head
waves. Int. J. Nav. Archit. Ocean Eng. 3, 105–110.
https://doi.org/10.3744/JNAOE.2011.3.1.105

Wu, D.-M., Wu, T.Y., 1982. Three-Dimensional Nonlinear Long Waves Due to
Moving Surface Pressure, in: 14th Symposium on Naval Hydrodynamics,
Ann Arbor.

Wu, H., Wu, J., He, J., Zhu, R., Yang, C.J., Noblesse, F., 2019. Wave profile along
a ship hull, short farfield waves, and broad inner Kelvin wake sans
divergent waves. Phys. Fluids 31. https://doi.org/10.1063/1.5088531

Wu, Z., 1992. On the Estimation of a Moving Ship’s Velocity and Hull
Geometry Information from its Wave Spectra.

Xing-Kaeding, Y., Jensen, G., 2015. Simulation of Ship Motions during

Maneuvers. Sh. Technol. Res. 53, 159–182.
https://doi.org/10.1179/str.2006.53.4.003

Xing-Kaeding, Y., Jensen, G., Hadžić, I., Peric, M., 2015. Simulation of Flow-
Induced Ship Motions in Waves using a RANSE Method. Sh. Technol.
Res. 51, 56–68. https://doi.org/10.1179/str.2004.51.2.002

Xing, T., Stern, F., 2010. Factors of Safety for Richardson Extrapolation. J.
Fluids Eng. 132, 061403. https://doi.org/10.1115/1.4001771

Xu, X., Li, H.A.O., Lin, Y., 2019. Mesh – Order Independence in CFD
Simulation. IEEE Access 7, 119069–119081.
https://doi.org/10.1109/ACCESS.2019.2937450

Yang, Q., Faltinsen, O.M., Zhao, R., 2001. Wash and Wave Resistance of Ships
in Finite Water Depth. Pract. Des. Ships Other Float. Struct. 475–483.
https://doi.org/10.1016/b978-008043950-1/50060-0

Yang, R., Shugan, I. V, Fang, M., 2011. Kelvin ship wake in the wind and waves
fiend and on the finite sea depth. Environment 27, 71–77.
https://doi.org/10.1115/OMAE2011-49872

Youden, W.J., 1972. Graphical diagnosis of inter-laboratory test results. J. Qual.

Technol. 4.
https://doi.org/https://doi.org/10.1080/00224065.1972.11980509

283
Yuan, Z.M., 2019. Ship Hydrodynamics in Confined Waterways. J. Sh. Res. 63,
1–14. https://doi.org/http://dx.doi.org/10.5957/JOSR.04170020

Yuan, Z.M., 2014. Hydrodynamic interaction between ships travelling or

stationary in shallow waters. Universtiy of Strathclyde.

Yuan, Z.M., Incecik, A., 2016. Investigation of side wall and ship model
interaction, in: 2016 International Conference on Maritime Technology.
pp. 1–10.

Yuan, Z.M., Zhang, X., Ji, C.Y., Jia, L., Wang, H., Incecik, A., 2018. Side wall
effects on ship model testing in a towing tank. Ocean Eng. 147, 447–457.
https://doi.org/10.1016/j.oceaneng.2017.10.042

Zeng, Q., Hekkenberg, R., Thill, C., 2019. On the viscous resistance of ships
sailing in shallow water. Ocean Eng. 190, 106434.
https://doi.org/10.1016/j.oceaneng.2019.106434

Zeng, Q., Thill, C., Hekkenberg, R., Rotteveel, E., 2018. A modification of the
ITTC57 correlation line for shallow water. J. Mar. Sci. Technol. 0, 0.
https://doi.org/10.1007/s00773-018-0578-7

Zhang, S., Tezdogan, T., Zhang, B., Xu, L., Lai, Y., 2018. Hull form optimisation
in waves based on CFD technique. Ships Offshore Struct. 13, 149–164.
https://doi.org/10.1080/17445302.2017.1347231

Zhang, S., Zhang, B., Tezdogan, T., Xu, L., Lai, Y., 2017. Research on bulbous
bow optimization based on the improved PSO algorithm. China Ocean
Eng. 31, 487–494. https://doi.org/10.1007/s13344-017-0055-9

Zhang, Z., 2010. Verification and validation for RANS simulation of KCS
container ship without/with propeller. J. Hydrodyn. 22, 932–938.
https://doi.org/10.1016/S1001-6058(10)60055-8

Zou, L., Zou, Z., Liu, Y., 2019. CFD-based predictions of hydrodynamic forces
in ship-tug boat interactions. Ships Offshore Struct. 0, 1–11.
https://doi.org/10.1080/17445302.2019.1589963

284
 PUBLICATIONS

The following papers have been either published, or submitted for

publication. Please note that not all papers are included in this thesis.

Journal articles

1. Terziev, M., Tezdogan, T. and Incecik, A., 2019. A geosim analysis of

ship resistance decomposition and scale effects with the aid of CFD.
Applied Ocean Research, 92, p.101930.
2. Terziev, M., Tezdogan, T. and Incecik, A., 2019. Application of eddy-
viscosity turbulence models to problems in ship hydrodynamics. Ships
and Offshore Structures, pp.1-24.
3. Elsherbiny, K., Terziev, M., Tezdogan, T., Incecik, A. and Kotb, M.,
2020. Numerical and experimental study on hydrodynamic
performance of ships advancing through different canals. Ocean
Engineering, 195, p.106696.
4. Terziev, M., Tezdogan, T. and Incecik, A., 2020. A numerical
assessment of the scale effects of a ship advancing through restricted
waters (submitted to Ships and Offshore Structures).
5. Terziev, M., Tezdogan, T., Incecik, A., 2020. A posteriori error and
uncertainty estimation in computational ship hydrodynamics. Ocean
Eng. 208, 107434. https://doi.org/10.1016/j.oceaneng.2020.107434
6. c) Terziev, M., Zhao, G., Tezdogan, T., Yuan, Z. and Incecik, A.,
2020. Virtual replica of a towing tank experiment to determine the
Kelvin half-angle of a ship in restricted water. Journal of Marine Science
and Engineering, 8(4), p.258.
7. Terziev, M., Tezdogan, T. and Incecik, A., 2020. Modelling the
hydrodynamic effect of abrupt water depth changes on a ship travelling

285
 in restricted waters using CFD (submitted to Ships and Offshore
Structures).

Conference papers

1. Terziev, M., Tezdogan, T. and Incecik, A., 2019, February. Influence of

mixed flows on ship hydrodynamics in dredged channels. In ASME
2019 38th International Conference on Ocean, Offshore and Arctic
Engineering, Glasgow, UK.
2. Terziev, M., Elsherbiny, K., Tezdogan, T. and Incecik, A., 2020, March.
Experimental and numerical study of an obliquely towed ship model in
confined waters. In 39th International Conference on Ocean, Offshore
& Arctic Engineering. (In press).

286
 APPENDIX A

The data obtained from Web of Science (2020) reads as follows:

Publication year Records % of total

2020 – as of 26.04.2020 34 2.56%
2019 206 15.50%
2018 168 12.64%
2017 154 11.59%
2016 130 9.78%
2015 107 8.05%
2014 89 6.70%
2013 58 4.36%
2012 70 5.27%
2011 50 3.76%
2010 38 2.86%
2009 36 2.71%
2008 34 2.56%
2007 27 2.03%
2006 28 2.11%
2005 26 1.96%
2004 5 0.38%
2003 17 1.28%
2002 15 1.13%
2001 10 0.75%
2000 5 0.38%

287
 1999 3 0.23%
1998 7 0.53%
1997 3 0.23%
1996 5 0.38%
1995 2 0.15%
1992 2 0.15%
Total 1329

The fit used to extrapolate to the year 2030 with 𝑅 2 = 0.9908 was determined using
a curve fit within MATLAB as shown in Eq. (A.1):

𝑓(𝑥) = 𝑎𝑒 𝑏𝑥 + 𝑐𝑒 𝑑𝑥 (A.1)

where, x represents the year. The fit was performed excluding data point for
year 2020.

Coefficients of Eq. (A.1):

𝑎 = 8.718 × 10−149
𝑏 = 0.1715
𝑐 = 0
𝑑 = −0.1013

288
 APPENDIX B

The data, used to construct Figure 4.1 is listed in the table below (all citations
are listed in the reference list):

Turbulence Percentage of
Reference Count
model total
Spalart- (Duvigneau et al., 2003; Eca et al., 2018; Eca 7 6.31%
Allmaras and Hoekstra, 2001; Hoekstra et al., 2000; Maki
et al., 2013; Pereira et al., 2017; Tin Htwe et al.,
2015)
k-ω Wilcox (Eca and Hoekstra, 2001; Hoekstra et al., 2000; 6 5.41%
Lam et al., 2013; Pereira et al., 2017;
Weymouth et al., 2005; Wortley, 2013)
k-ω SST (Atencio and Chernoray, 2019; Banks et al., 50 45.05%
2010; Bhushan et al., 2009, 2007; Castiglione
et al., 2011; Castro et al., 2011; Demirel et al.,
2017, 2014; Deng et al., 2014, 2010; Duvigneau
et al., 2003; Eca et al., 2018; Eca and Hoekstra,
2001; el Moctar et al., 2015; Farkas et al., 2018;
Gaggero et al., 2015; Guo et al., 2013, 2015;
Haase et al., 2016a; Hoekstra et al., 2000; Irkal
et al., 2019; Jasak et al., 2018; Kaidi et al.,
2017; Kianejad et al., 2019; Kok et al., 2018;
Liu et al., 2017; Maasch et al., 2019; Majidian
and Azarsina, 2019; Mucha et al., 2016;
Oggiano et al., 2017; Ohashi et al., 2018;
Pacuraru and Domnisoru, 2017; Pereira et al.,
2017; Prakash and Chandra, 2013; Queutey and
Visonneau, 2015; Razgallah et al., 2018;
Saydam and Taylan, 2018; Schweighofer and
Regnstr, 2005; Shenoi et al., 2016; Shi et al.,
2012; Sigmund and Moctar, 2018; Simonsen et
al., 2013; S. Song et al., 2019; Toxopeus et al.,

289
 2013; Visonneau, 2005; Wang et al., 2017,
2016; Wilson et al., 2006; Zhang, 2010; Zou et
al., 2019)
k-ε (Bakica et al., 2019; Begovic et al., 2015; 41 36.94%
Bellafiore et al., 2018; Chen et al., 2001; Choi
et al., 2009; Chun et al., 2001; Dhinesh et al.,
2010; Eca and Hoekstra, 2001; Farkas et al.,
2017; Hai-Long et al., 2016; Hoekstra et al.,
2000; Jagadeesh and Murali, 2010; Kellett et
al., 2013; M. Kim et al., 2017b, 2017a; Y. C.
Kim et al., 2017; Kinaci et al., 2016; Kinaci and
Gokce, 2015; Kouh et al., 2009; Lee et al.,
2018; Linde et al., 2016; Liu et al., 2018; Ma et
al., 2013; Mancini et al., 2018; Oh and Kang,
1992; Ozdemir and Barlas, 2017;
Schweighofer, 2004; Shivachev et al., 2017;
Srividya and Thandaveswara, 2005; Tahara et
al., 2002a; Terziev et al., 2018; Tezdogan et al.,
2016c, 2016b, 2016a, 2015; Van et al., 2011;
Visonneau, 2005; Wang et al., 2016; Wu et al.,
2011; Xing-Kaeding et al., 2015; Xing-Kaeding
and Jensen, 2015; Zhang et al., 2018, 2017)
AKN (Jagadeesh and Murali, 2010; Pereira et al., 2 1.80%
2017)
RST (Choi et al., 2010; Duvigneau et al., 2003; Lee 5 4.50%
et al., 2018; Visonneau, 2005; Wang et al.,
2016)

290
 APPENDIX C

This Appendix details the equations used by in this thesis. Sources for the
information presented herein concerning governing equations is drawn from
Ferziger and Peric (2002) and Durbin and Pettersson Reif (2011). These
references are omitted henceforth.

Firstly, the governing equations, describing viscous incompressible flow can

be written as follows:
1
𝜕𝑡 𝑢̃ + 𝑢̃𝑖 𝜕𝑗 𝑢̃𝑖 = − 𝜌 𝜕𝑖 𝑝̃ + 𝜈∇2 𝑢̃𝑖 (C.1)

𝜕𝑖 𝑢̃𝑖 = 0 (C.2)

Eq. (C.1) is the conservation of momentum, whereas Eq. (C.2) is the

conservation of mass. In the aforementioned equations, the notation 𝜕𝑖 was
used to denote the partial derivative 𝜕/𝜕𝑥𝑖 , whereas the subscript 𝑖 gives the
respective component of the vector. We may decompose the total velocity
𝑢̃(𝑥, 𝑡) into a mean and fluctuating component 𝑢̃(𝑥, 𝑡) = 𝑈(𝑥, 𝑡) + 𝑢(𝑥, 𝑡), and
𝑈 = 𝑢̅̃ . Here, a bar above a variable indicates averaging.

Utilising the abovementioned decomposition, Eq. (C.1) and Eq. (C.2) become:
1
𝜕𝑡 (𝑈𝑖 + 𝑢𝑖 ) + (𝑈𝑖 + 𝑢𝑖 )𝜕𝑗 (𝑈𝑖 + 𝑢𝑖 ) = − 𝜌 𝜕𝑖 (𝑃 + 𝑝) + 𝜈∇2 (𝑈𝑖 + 𝑢𝑖 ) (C.3)

𝜕𝑖 (𝑈𝑖 + 𝑢𝑖 ) = 0 (C.4)

In cases, such as those presented throughout the thesis, where the fluctuating
component is not modelled, one may average Eq. (C.3) and Eq. (C.4), which
become:
1
𝜕𝑡 𝑈𝑖 + 𝑈𝑖 𝜕𝑗 𝑈𝑖 = − 𝜌 𝜕𝑖 𝑃 + 𝜈∇2 𝑈𝑖 − 𝜕𝑗 ̅̅̅̅̅
𝑢𝑗 𝑢𝑖 (C.5)

291
𝜕𝑖 𝑈𝑖 = 0 (C.6)

Eq. (C.5) and Eq. (C.6) are the Reynolds averaged Navier-Stokes equations.
The final term in Eq. (C.5) is the derivative of the Reynolds stress tensor ̅̅̅̅̅,
𝑢𝑗 𝑢𝑖
given in Eq. (C.7):
𝑢1 𝑢1
̅̅̅̅̅̅ 𝑢1 𝑢2
̅̅̅̅̅̅ 𝑢1 𝑢3
̅̅̅̅̅̅
𝑢𝑗 𝑢𝑖 = (̅̅̅̅̅̅
̅̅̅̅̅ 𝑢2 𝑢1 ̅̅̅̅̅̅
𝑢 2 𝑢2 𝑢
̅̅̅̅̅̅
2 𝑢3 ) (C.7)
𝑢3 𝑢1
̅̅̅̅̅̅ 𝑢3 𝑢2
̅̅̅̅̅̅ 𝑢
̅̅̅̅̅̅
3 𝑢3

The resultant set of 4 equations contain 10 unknowns, which is why turbulence

modelling is frequently referred to as providing closure, i.e. additional
equations, allowing the system to form a closet set. The unknowns, present in
the above equations are 𝑈𝑖 , 𝑃, and ̅̅̅̅̅
𝑢𝑖 𝑢𝑗 with 𝑖, 𝑗 = 1, 2, 3. The Reynolds stress
term is modelled an eddy-viscosity concept, which translates into:

−𝑢
̅̅̅̅̅
𝑖 𝑢𝑗 ≈ 𝑣𝑇 [𝜕𝑗 𝑈𝑖 + 𝜕𝑖 𝑈𝑗 ] (C.8)

It is the purpose of all eddy-viscosity turbulence models to predict the eddy-

viscosity, 𝑣𝑇 . This may be achieved by a variety of methods.

C.1 Spalart-Allmaras model

The one-equation Spalart and Allmaras (1992) closure models the eddy-
viscosity via Eq. (C.9):

𝑣𝑇 = 𝜌𝑓𝑣 (𝑣̃/𝑣) (C.9)

where 𝑣̃ is the modified diffusivity term, and 𝑓𝑣 is the damping function,

shown in Eq. (C.10)
(𝑣̃ ⁄𝑣)3
𝑓𝑣 (𝑣̃/𝑣) = (𝑣̃⁄𝑣)3+7.13 (C.10)

Finally, the modified diffusivity is modelled via a transport equation in the

form of:
1
𝜕𝑡 𝑣̃ + 𝑈 ∙ ∇𝑣̃ = 𝒫𝑣 − 𝜀𝑣 + 𝜎 [∇((𝑣 + 𝑣̃)∇𝑣̃) + 𝑐𝑏2 |∇𝑣̃|2 ] (C.11)
𝑣

Where according to Spalart and Allmaras (1992), 𝑐𝑏2 = 0.622, 𝜎𝑣 = 2/3, while
𝒫 is the production term, modelled using Eq. (C.12), where 𝑐𝑏1 = 0.1355, and
𝑆 being the magnitude of the mean vorticity:

𝒫𝑣 = 𝑐𝑏1 𝑆𝑣̃ (C.12)

292
 C.2 The 𝒌 − 𝜺 family of turbulence models

The 𝑘 − 𝜀 family of models predicts the eddy viscosity by using the turbulent
kinetic energy (𝑘) and dissipation rate (𝜀):

𝑣𝑇 = 𝐶𝜇 𝑓𝜇 𝑘 2 /𝜀 (C.11)

where the standard value of 𝐶𝜇 is a constant, and 𝑓𝜇 is a damping function.

These parameters are summarised in . The mean flow properties are the
expressed via:
2
−𝑢
̅̅̅̅̅
𝑗 𝑢𝑖 = 2𝑣𝑇 𝑆𝑖𝑗 − 3 𝑘𝛿𝑖𝑗 (C.12)

with 𝑆𝑖𝑗 being the mean rate of strain tensor, 𝑆𝑖𝑗 = 0.5(𝜕𝑖 𝑈𝑗 + 𝜕𝑈𝑖 ).

The transport equations for the turbulent kinetic energy and dissipation rate
are:

𝜕𝑡 (𝑘) + 𝑈𝑗 𝜕𝑗 𝑘 = 𝒫 − 𝜀 + 𝜕𝑗 ((𝑣 + 𝑣𝑇 /𝜎𝑘 )𝜕𝑗 𝑘) (C.13)

𝐶𝜀1 𝒫−𝐶𝜀2 𝜀 𝑣
𝜕𝑡 𝜀 + 𝑈𝑗 𝜕𝑗 𝜀 = + 𝜕𝑗 ((𝑣 + 𝜎𝑇 ) 𝜕𝑗 𝜀) (C.14)
𝑇 𝜀

In the above equations, 𝒫 = 2𝑣𝑇 |𝑆|2 , and 𝑇 = 𝑘/𝜀 is the turbulent time scale.
Eq. (C.14) is constructed by dimensional analogy, and is an assumed form. For
this reason, Eq. (C.14) contains empirical constants 𝐶𝜀1 , 𝐶𝜀2 , 𝜎𝜀 , while 𝜎𝑘 is
usually taken as unity. The respective coefficient values for the models, used
in this thesis (the AKN model and the realizable 𝑘 − 𝜀 models are given in
Table 0.1.

Table 0.1. Model coefficients for AKN and 𝑘 − 𝜀 closures.

Coefficient AKN Realizable 𝑘 − 𝜀
𝐶𝜀1 1.5 max(0.43, 𝜂/(5 + 𝜂)), with 𝜂 = 𝑆𝑘/𝜀
𝐶𝜀2 1.9 1.9
𝜎𝜀 1.4 1.2
𝜎𝑘 1.4 1

The (Lag) Elliptic blending and v2 – f models share their origins and are
presented jointly.

Firstly, the turbulent eddy viscosity of the v2 – f model is predicted via:

𝜀𝑣̅̅̅2̅
̅̅̅2 + 𝑈𝑗 𝜕𝑗 ̅̅̅
𝜕𝑡 𝑣 𝑣 2 + 𝑘 = 𝑘𝑓 + 𝜕𝑘 [𝑣𝑇 𝜕𝑘 ̅̅̅
𝑣 2 ] + 𝑣∇2 ̅̅̅
𝑣2 (C.15)

293
 𝑐2 𝒫 𝑐1 ̅̅̅
𝑣 2̅ 2
𝐿2 ∇2 𝑓 − 𝑓 = − + ( 𝑘 − 3) (C.16)
𝑘 𝑇

̅̅̅2 𝑇.
with 𝑐2 = 0.3, and 𝑐1=0.4. The eddy viscosity is predicted as 𝑣𝑇 = 𝐶𝜇 𝑣

The Elliptic blending model solves the transport equations for the turbulent
kinetic energy (Eq. (C.13)) and dissipation rate (Eq. (C.14)), alongside a
reduced wall-normal stress component 𝜙, and the elliptic blending factor 𝛼.

∇(𝐿2 ∇𝛼) = 𝛼 − 1 (C.17)

𝜇 𝜇
𝜕𝑡 (𝜌𝜙) + ∇(𝜌𝜙𝑈) = ∇ [(2 + 𝜎 𝑡 ) ∇𝜙] + 𝒫𝜙 + 𝑆𝜙 (C.18)
𝜙

where 𝒫𝜙 is the production term, 𝜎𝜙 = 1, and

𝑘3 𝑣3
𝐿 = 𝐶𝐿 √ 𝜀2 + 𝐶𝜂2 √ 𝜀 (C.19)

with 𝐶𝐿 = 0.164, 𝐶𝜂 = 75.

The 𝒌 − 𝝎 family of turbulence models

Wilcox (2006) represented the transport equation for the turbulent kinetic
energy as:

𝑣
𝜕𝑡 𝑘 + 𝑈𝑗 𝜕𝑗 𝑘 = 2𝑣𝑇 |𝑆|2 − 𝐶𝜇 𝑘𝜔 + 𝜕𝑗 ((𝑣 + 𝜎𝑇 ) 𝜕𝑗 𝑘) (C.20)
𝑘

and the dissipation frequency as:

𝑣
𝜕𝑡 𝜔 + 𝑈𝑗 𝜕𝑗 𝜔 = 2𝐶𝜔1 |𝑆|2 − 𝐶𝜔2 𝜔2 + 𝜕𝑗 ((𝑣 + 𝜎𝑇 ) 𝜕𝑗 𝜔) (C.21)
𝜔

where 𝐶𝜔1 = 5/9, 𝐶𝜔2 = 3/40, 𝜎𝜔 = 𝜎𝑘 = 2, and 𝐶𝜇 = 0.09. The eddy-viscosity

is predicted as 𝑣𝑇 = 𝑘/𝜔.

Menter (1994) proposed the Shear Stress Transport turbulence closure. He

noted certain properties of solutions, obtained with the 𝑘 − 𝜔 model, for
example, overprediction of shear stress. Menter's (1994) solution was to
introduce a bound on the stress-intensity ratio |𝑢𝑣
̅̅̅̅|/𝑘 = 𝑎1 . Then, by using the
identity −𝑢𝑣
̅̅̅̅ = 𝑘𝜕𝑘 𝑈/𝜔, he arrived at:
𝒫 ̅̅̅̅ 2 𝜔
𝑢𝑣 1 ̅̅̅̅ 2
𝑢𝑣
= = 𝐶 |𝑘| (C.22)
𝜀 𝑘𝜀 𝜇

294
Menter (1994) also proposed bounds on 𝑣𝑇 in the form of:
𝑘 √𝐶𝜇 𝑘
𝑣𝑇 = min [𝜔 , |2𝜴|
] (C.23)

where |𝛀| is the mean flow rotation tensor. Next, a limiting function (𝐹1 ) is
introduced:

𝐹1 = tanh(𝑎𝑟𝑔14 ) (C.24)

with
√𝑘 500𝑣 2𝑘𝜔
𝑎𝑟𝑔1 = min [𝑚𝑎𝑥 (𝐶 , 𝜔𝑦 2 ) , 𝑦 2 𝑚𝑎𝑥(𝛻𝑘∙𝛻𝜔,10−20 )]
𝜇 𝜔𝑦
(C.25)

Finally, the transport equation for the turbulent dissipation rate is:

𝐶𝜀1 𝒫−𝐶𝜀2 𝜀 𝑣
𝜕𝑡 𝜀 + 𝑈𝑗 𝜕𝑗 𝜀 = + 𝜕𝑗 ((𝑣 + 𝜎𝑇 ) 𝜕𝑗 𝜀) + 𝐹1 𝑆𝜔 (C.26)
𝑇 𝜀

which is identical to Eq. (C.14), with the exception of the final term on the
right-hand side. Here, 𝑆𝜔 is found by:
2 𝑣 |∇𝑘|2 ∇𝑘∙∇𝜀
𝑆𝜔 = 𝑇 (𝑣 + 𝜎𝑇 ) [ − ] (C.27)
𝜔 𝑘 𝜀

The model constants are also interpolated as follows:

𝐶𝜀 1 = 1 + (1 − 𝐹1 )0.44 + 𝐹1 𝐶𝜔1 (C.28)

𝐶𝜀2 = 1 + (1 − 𝐹1 )0.92 + 𝐹1 𝐶𝜔2 /𝐶𝜇 (C.29)

The 𝛾 (intermittency) transition model adds a transport equation to be solved

alongside those stated earlier for the SST model:
𝑑𝛾 𝜇𝑇
(𝜌𝛾) + ∇(𝜌𝛾𝑈) = ∇∇ [(𝜇 + ) ∇𝛾] + 𝑃𝛾 − 𝐸𝛾 (C.30)
𝑑𝑡 𝜎𝑓

where:

𝑃𝛾 = 100𝜌𝑆𝛾(1 − 𝛾)𝐹𝑜𝑛𝑠𝑒𝑡 (C.31)

𝐹𝑜𝑛𝑠𝑒𝑡 = max (𝐹𝑜𝑛𝑠𝑒𝑡2 − 𝐹𝑜𝑛𝑠𝑒𝑡3 , 0) (C.32)

in which

𝐹𝑜𝑛𝑠𝑒𝑡2 = min (𝐹𝑜𝑛𝑠𝑒𝑡1 , 2) (C.33)

𝑅𝑒 3
𝐹𝑜𝑛𝑠𝑒𝑡3 = max [1 − ( 3.5𝑡 ) , 0] (C.34)

295
 𝑅𝑒𝑣
𝐹𝑜𝑛𝑠𝑒𝑡1 = 𝐶 (C.35)
𝑜𝑛𝑠𝑒𝑡1 𝑅𝑒𝜃𝑐

where:

• 𝐶𝑜𝑛𝑠𝑒𝑡1 = 2.2
• 𝜎𝑓 = 1
• 𝑅𝑒𝑡 is the turbulent Reynolds number, 𝑅𝑒𝑡 = 𝑘 2 ⁄(𝑣𝜀) = 𝑘⁄(𝑣𝜔)
• 𝑅𝑒𝑣 is the strain-rate Reynolds number 𝑅𝑒𝑣 = 𝑑 2 𝑆⁄𝑣 , with 𝑑 being the
distance from the wall
• 𝑅𝑒𝜃𝑐 is the correlation for the critical Reynolds number and is defined
as follows:
𝑅𝑒𝜃𝑐 = 𝐶𝑇𝑈1 + 𝐶𝑇𝑈2 exp [−𝐶𝑇𝑈3 𝑇𝑢𝐿 𝐹𝑃𝐺 (𝜆𝜃𝐿 )] (C.36)
where 𝐶𝑇𝑈1 = 100, 𝐶𝑇𝑈2 = 1000, 𝐶𝑇𝑈3 = 1, and
100√2𝑘⁄3
𝑇𝑢𝐿 = min ( , 100) (C.37)
𝜔𝑑
10−3 𝑑2
𝜆𝜃𝐿 = min [𝑚𝑎𝑥 (−7.57 × 𝛻(𝑛𝑈) × 𝑛 + 0.0128, −1) , 1] (C.38)
𝑣
max[min(1 + 14.68𝜆𝜃 , 1.5) , 0] , 𝜆𝜃𝐿 ≥ 0
𝐹𝑃𝐺 (𝜆𝜃𝐿 ) = { (C.39)
max[min(1 − 7.34𝜆𝜃 , 3) , 0] , 𝜆𝜃 < 0

The production term 𝐸𝛾 = 𝐶𝑎2 𝜌𝑊𝛾𝐹𝑡𝑢𝑟𝑏 (𝑐𝑒2 𝛾 − 1),

where 𝐹𝑡𝑢𝑟𝑏 = exp[−(𝑅𝑒𝑡 ⁄2)4 ]. 𝑊 is the modulus of the mean vorticity tensor,
while 𝐶𝑎2 = 0.06, and 𝐶𝑒2 = 50.

296