1

2 A review on the turbulence modelling strategy for ship hydrodynamic simulations

4 Blanca Pena and Luofeng Huang*

5 Department of Mechanical Engineering, University College of London, United Kingdom

7 Abstract: Ship operations are accompanied by turbulent regimes that play a significant role in the
8 hydrodynamic characteristics of a flow. With the ongoing development of computational technologies, it is
9 now feasible to numerically simulate turbulent ship flows with a high degree of detail. Turbulent
10 simulations, however, tend to be computationally expensive and require a trade off between computational
11 costs and fidelity. Whilst a range of turbulence modelling strategies is available in Computational Fluid
12 Dynamics, there is a lack of up-to-date recommendations on their suitability for different ship flow
13 simulation scenarios. Addressing this gap, the present work reviews the state-of-the-art of turbulence
14 modelling for ship hydrodynamic applications. As a result, this paper introduces the most known turbulence
15 modelling approaches used in ship hydrodynamics, followed by a thorough discussion of their
16 applicabilities and limitations. Furthermore, this paper provides recommendations for the selection of
17 turbulence modelling strategies versus various ship simulation scenarios, such as resistance prediction, ship
18 flow modelling, self-propulsion, and cavitation analyses. It is expected that the present paper will provide
19 decision-making support by helping CFD users minimise the time spent on trial and error, as well as
20 providing valuable insights to promote the advancement of turbulence modelling.

21 Keywords: Ship, Turbulence, Computational Fluid Dynamics, Reynolds-Average Navier-Stokes, Large-

22 Eddy Simulation, Detached Eddy Simulation.

23

24

25

26

Corresponding author: [email protected]

1
27 1. Introduction

28 Computational fluid dynamics (CFD) is one of the most important approaches used in ship design. From
29 integral ship resistance to local structural response, CFD has been proven to be a legitimate prediction
30 technique. Compared to lab experiments, CFD is much cheaper and can provide much more information
31 about the flow that cannot be captured during experiments. Hereby, it has been a widespread trend in the
32 shipping industry, academia and international societies that designate CFD as an essential skill (ASME,
33 2009; ITTC, 2014a).

34 The comprehensive capabilities of CFD provide abundant options on numerical set-ups, bringing many
35 questions on how to make choices. For example, the spatial and temporal discretisation resolutions, known
36 as mesh density and timestep size respectively, play a crucial role in computational accuracies (Jasak et al.,
37 2007). Generally speaking, a higher spatial and temporal resolution can yield more accurate results while
38 the demanded computational recourses to solve a simulation will increase accordingly. Due to this reason,
39 sensitivity tests on mesh/timestep size are required to secure optimal performance and a balance between
40 computational costs and fidelity.

41 Another compromise is also faced regarding the options of turbulence modelling schemes, which is to a
42 lesser extent covered by existing CFD guidelines. Direct Numerical Simulation (DNS) can accurately
43 replicate turbulent flows, but this requires solving the Navier-Stokes equations at the Kolmogorov micro-
44 scale, which is extremely demanding and makes DNS inaccessible to most CFD users. As a result,
45 assumption-based turbulence modelling is commonly used. Such assumptions have categorised
46 corresponding turbulence modelling strategies into several groups, known as Reynolds-Average Navier-
47 Stokes equations (RANS), Large-Eddy Simulation (LES) and their combination (Hybrid). These schemes
48 equip different capabilities, and one can require thousandfold computational resources than another
49 (Tezdogan et al., 2015; Liefvendahl and Fureby, 2017). Therefore, it is of great importance to justify that
50 an appropriate turbulence modelling scheme has been applied.

51 First of all, it is essential to use a turbulence modelling strategy when simulating ship flows, rather than just
52 adopting a laminar assumption, because the kinematic energy dissipation within turbulent flows, if non-
53 negligible, has to be taken into consideration. Khojasteh et al. (2020) compared shipping waters predicted
54 by laminar modelling, turbulence modelling and experiments. They demonstrated that laminar modelling
55 could cause a significant deviation in the fluid shape as well as the calculated velocity and force. They
56 showed that such deviation could be avoided when applying RANS turbulence modelling. Moreover,
57 researchers have proven that appropriately choosing a turbulence modelling strategy can provide accurate
58 predictions in ship resistance (Zha et al., 2014a), motions (Cha and Wan, 2015) and wake (Shen et al.,

2
59 2002), as well as replicating complex ship flows such as propeller cavitation (Watanabe et al., 2003) and
60 vortex shedding (Arslan et al., 2016). However, some turbulence schemes may cause important phenomena
61 to be wholly neglected. For example, the time-averaging process of RANS negates the flow unsteadiness;
62 thus, the cavitation observed in experiments could not be appropriately modelled when using RANS, as
63 illustrated by Bensow (2011). Another example of Arslan et al. (2016) shows that small-scale fluid vortexes
64 that are essential for predicting local structural response were captured by LES but not by RANS.

65 In such a context, appropriately choosing a turbulence modelling strategy becomes crucial when simulating
66 ship hydrodynamic flows. Ideally, CFD users should count on a clear and up-to-date recommendation of
67 suitable turbulence modelling schemes for different scenarios. The guidance is expected to avoid
68 unnecessary testing time and improve simulation accuracy. To serve such a purpose, this work intends to
69 provide a study on the applicability of available turbulence schemes while focusing on the applicability in
70 model-scale and full-scale ship hydrodynamics. Being specific and critical, this work aims to provide handy
71 insights on how existing turbulence modelling schemes can simulate ship hydrodynamic behaviours,
72 focusing on their capability, limitation, computational cost and accuracy.

73 The present paper starts by briefly introducing the theories of different turbulence modelling schemes,
74 linking to a review of how those schemes have been applied to various ship-flow interaction scenarios.
75 Next, each scheme's capability and limitation are discussed and insightful explanations are given.
76 Subsequently, turbulence modelling selection recommendations are provided based on a combination of
77 well simulating the case and saving the computational cost, with respect to particular scenarios such as
78 resistance prediction, ship flow modelling, self-propulsion, and cavitation. Finally, the paper discusses the
79 main challenges faced by contemporary turbulence modelling methods and offers suggestions for future
80 work.

81

82 2. Turbulence Modelling Approaches

83 2.1 Direct Numerical Simulation (DNS)

84 DNS principle is based on a direct resolution of the Navier-Stokes (N-S) equations representing the most
85 complete mathematical description of the continuum flow of a fluid. Therefore, DNS is recognised to be
86 the most expensive method to simulate turbulence flows. The N-S mass and continuum conservation
87 equations are given below in the most general way (Peric and Ferziger, 2002).

88

3
 89 ∇⋅𝐯=0 (1)

∂(ρ𝐯)
90 ∂t
+ ∇ ⋅ (ρ𝐯𝐯) = −∇p + ∇ ⋅ τ + ρg (2)

91

92 where 𝐯 is velocity vector, P is pressure, ρ is the density and 𝛕 = μ(∇𝐯 + ∇𝐯 T ) is the viscous stress, in
93 which μ is the dynamic viscosity.

94 Any DNS simulation requires to solve all essential scales of motion contained in the flow up to an order of
95 magnitude of the Kolmogorov micro-scale (ηk ) (Kolmogorov, 1941):

96

ν3
97 ηk = ( ε )1/4 (3)

98

99 where ε represents the dissipation rate and ν the molecular viscosity. In general, Kolmogorov micro-scales
100 refer to the smallest turbulent vortex scales in a turbulent flow at which the viscosity dominates and the
101 turbulent kinetic energy is dissipated into heat. DNS is the most accurate tool available for modelling
102 viscous flow; however, the prohibitive computational cost of DNS has hindered it from solving most
103 industrial problems.

104

105 2.2 Reynolds-Averaged Navier–Stokes (RANS)

106 The prohibitive cost of DNS motivated the development of more affordable numerical strategies to account
107 for the turbulence effects. One of the most common methods is the Reynolds-Average Navier-Stokes
108 equations (RANS), in which instantaneous turbulent velocity is decomposed into its time-averaged and
109 fluctuating quantities, by modifying the N-S equations into:

110

111 ∇⋅𝐯=0 (4)

∂(ρ𝐯)
112 + ∇ ⋅ (ρ𝐯𝐯) = −∇p + ∇ ⋅ (τ − ρ𝐯 ′ 𝐯 ′ ) + ρ𝐠 (5)
∂t

113
4
114 where 𝐯 is the time-averaged velocity and 𝐯 ′ is the fluctuating one, ρ is the fluid density, 𝑝 denotes the
115 time-averaged pressure, 𝜏 = µ[∇v+ (∇v)T] is the viscous stress term, µ is the dynamic viscosity and g is the
116 gravitational acceleration.

117 As additional unknowns have been introduced into equation (5), some approximations and additional
118 models are required to solve the new appeared term. To model this term, two approaches are typically used:

119 a) Eddy viscosity models. This approach relates the turbulent stresses appearing in the RANS
120 equations to the gradients of time-averaged velocity (Peric and Ferziger, 2002). Most eddy viscosity
121 models are based on the Boussinesq approximation. This approach relates the Reynolds Stresses to
122 the mean velocity gradients. The most known turbulent models in this group are the Spalart-
123 Allmaras (SA), the k − ε family and the k − ω family; they have several branches specialising in
124 different problems and adopt specific empirical coefficients based on data fitting for a wide range
125 of turbulent flows (McComb, 1990).
126 b) An alternative approach to solve the Reynolds stresses is by using Reynolds Stress Models (RSM).
127 An RSM model directly calculates all the specific Reynolds stress tensor components by solving
128 their governing transport equations together with an equation for the dissipation rate (Peric and
129 Ferziger, 2002). This turbulence model can predict complex flows than eddy viscosity models
130 because the transport equations for the Reynolds stresses naturally account for the effects of
131 turbulence anisotropy and streamline curvature. However, it is essential to remember that this
132 model belongs to the RANS category, averaging pressure and velocities, leading to inaccurate
133 prediction of the unsteady vorticity fields.

134

135 2.3 Large Eddy Simulation (LES)

136 Kolmogorov (1941) proposed to describe the turbulent energy spectrum and its eddies into a range of scales.
137 LES's principal idea is to reduce the computational cost of DNS by filtering the Navier–Stokes equations
138 and modelling the smallest turbulent scales in the flow (Smagorinsky, 1963). LES's main principle is to
139 approach the modelling of turbulence by considering that the large vortical structures created by the
140 geometry contain most of the energy within the bulk flow. Thus, LES resolves turbulent structures in space
141 everywhere in the flow domain down to a grid limit, while it uses a Sub-Grid Scale (SGS) model to simulate
142 the effect of small eddies (smaller than filter size) on the mean flow.

5
143 The difference between DNS and LES approaches are summarised in Figure 1. DNS resolves all turbulent
144 scales down to the Kolmogorov micro-scale from the N-S equations, whereas LES only resolves the largest
145 turbulent scales in the flow.

146

147

148 Figure 1: LES and DNS approach representation (Ferziger and Perić, 2002).

149

150 The low-pass filtering operation which is used in LES methods may be applied to spatial and temporal
151 fields of a flow variable. In one-dimensional notation the filtered velocity (Leonard, 1975) is defined by:

152

153 𝑢̅𝑖 (𝑥) = ∫ 𝐺(𝑥, 𝑥 ′ )𝑢𝑖 (𝑥 ′ )𝑑𝑥′ (6)

154

155 where G(x, x') represents the filter kernel which has a length scale associated with it, ∆. Eddies with a size
156 larger than ∆ represent large eddies while those smaller than ∆ represent eddies that need to be modelled.
157 The Navier-Stokes equations can then be filtered to obtain the LES governing equations. More details can
158 be found in (Ferziger and Perić, 2002).

6
159 For those scales smaller than the filter, RANS' Reynolds stress term in the governing equations is modelled
160 using Sub-Grid Scale (SGS) models in LES. It is well known that in Newton's law of viscosity for
161 incompressible flow:

162

163 τNewtonian = 2νS = ν(∇v + ∇v T ) (7)

164

165 where S is the rate of deformation of fluid elements. It has been found that the turbulent stresses increase
166 as the mean rate of deformation increases. The Boussinesq hypothesis proposed that the Reynolds stress in
167 RANS is proportional to the mean rates of deformation. In the SGS model, this theory is interpreted as SGS
168 stresses are proportional to the instantaneous rates of deformation, i.e.

169

1 T 1
170 τ = −2νt S + tr(τ)𝐈 = −νt (∇𝐔 + ∇𝐔 ) + tr(τ)𝐈 (8)
3 3

171

172 where νt is the SGS eddy viscosity. On dimensional grounds, it is assumable that νt can be expressed as a
173 product of an SGS velocity scale, ϑ, and an SGS length scale, L, as

174

175 νt = CϑL (9)

176

177 where C is a dimensionless constant. Therefore, the turbulence model based on Eddy Viscosity theory is to
178 find appropriate equations for ϑ and L by either algebraic relations or transport equations and then use them
179 to obtain νt thus closing the filtered N-S equations; examples can be found in (Pope, 2001).

180 The main disadvantage of using pure LES for high Reynolds number flows is the requirement of very fine
181 grids, particularly in the near-wall regions of the flow domain and very small time steps (Hanjalić and
182 Launder, 2009). Therefore, it is anticipated that using LES for full-scale ship problems is not practical due
183 to the high computational requirements.

184

7
185 2.4 Hybrid (RANS+LES)

186 Hybrid methods or RANS+LES methods are based on the principle of reproducing a RANS in regions not
187 subjected to flow separation and an LES in regions with significant flow separation, respectively reflected
188 as coarse and dense meshes, as presented by Guilmineau et al. (2018) in Figure 2. Hybrid methods were
189 initially conceived for improving the numerical prediction of complex flows encountered in the aviation
190 industry (Spalart et al., 1997). An example of this is the simulation of high–Reynolds flows subject to
191 significant flow separation.

192 The concept of a hybrid method can be considered remarkably useful in ship hydrodynamics. For example,
193 some problems require accurately modelling the ship's wake using a more detailed calculation of large eddy
194 turbulent interactions, which can be done using LES. By contrast, regions of less complex physics can be
195 calculated satisfactorily using RANS models. Such a hybrid approach is typically useful during propellers'
196 design and performance evaluation and energy-saving devices or cavitation assessments.

197 Hybrid methods are typically classified into two categories: zonal and non-zonal methods. On the one hand,
198 hybrid zonal methods rely on a RANS model and a subgrid-scale model. This approach is applied in
199 different domains separated by a sharp or dynamic interface. By contrast, non-zonal methods assume that
200 the governing set of equations smoothly converts from a RANS behaviour to an LES behaviour, based on
201 criteria updated during the computation. Zonal hybrid methods have been successfully validated to resolve
202 ship hydrodynamic problems, being Detached Eddy Simulation (DES) the most popular.

203 In general, DES is computationally cheaper than LES while able to simulate detailed turbulent vortex
204 structures. Figure 3 shows the results from the flow's simulation at low Reynolds numbers for a sphere
205 using different turbulence modelling strategies (Constantinescu et al., 2003). The figure reveals that RANS
206 models cannot model the vortex shedding phenomena with details obtained during the LES and DES
207 computations. In this case, all models were meshed using the same number of elements; however, DES was
208 considerably faster than LES.

209

8
210

211 Figure 2: Hybrid mesh of Guilmineau et al. (2018): panel (a) shows general mesh in line with the
212 requirement of RANS, panel (b) shows LES level mesh applied to the region where significant flow
213 separation is expected.

214

215

216 Figure 3: Turbulence modelling strategies comparison made by Constantinescu et al. (2003).

217
9
218 2.4.1 Detached Eddy Simulation (DES)

219 Compared with RANS and LES, DES is a new turbulence strategy that has been developed in recent years.
220 DES allows the near-wall region to be treated by RANS, and regions of significant flow recirculation are
221 modelled using an LES approach. The interface between RANS and LES is called the 'grey region'. The
222 first DES (also known as the DES97) relied upon the isotropic Spalart-Allmaras (SA) model. The DES97
223 creates a hybrid method by introducing the cell size dependency into the SA turbulence modelling
224 approach's destruction term. By these means, the detached eddies distant from the wall are resolved using
225 LES, whereas the attached vortices are modelled using RANS. For this reason, the DES97 method was
226 entitled the Detached-Eddy Simulation (Spalart et al., 1997).

227 The DES length scale was introduced to replace dw in all terms of the SA model equations as follows:

228

229 LDES97 = min (dw ; LLES ), LLES = CDES Δ, Δ = max (Δi , Δj , Δk ) (10)

230

231 Where dw < LLES, the DES length scale is equal to the RANS SA model length scale. By contrast, where
232 LLES < dw, the length scale LLES is employed. Being CDES a constant and Δ corresponds to the grid size
233 formulated as the maximum cell length in each index direction. This formulation is proposed assuming that
234 no knowledge of the local turbulent structure is available in advance. The smallest resolvable isotropic
235 eddies would scale with the coarsest grid cell dimension. Therefore, the DES97 formulation can
236 successfully relate a near-wall region to RANS and fine mesh regions far from the wall to LES. However,
237 this method reported several fundamental issues as follows.

238 The first issue is called 'The Grey Area'. The grey issue was known by Spalart (2009) and it refers to the
239 zone where the DES model switches from RANS to LES mode (dw ≈ CDES Δ). This region was reported to
240 be problematic unless the separation is abrupt and fixed by the geometry (Spalart, 2009). As the DES97
241 stipulates, the attached boundary layer should be handled entirely by RANS mode, where the turbulent
242 kinetic energy is fully modelled. By contrast, the region containing flow separation should be handled by
243 LES, fully resolving the turbulent kinetic energy. Therefore, a region between the two modes may contain
244 little resolved turbulence even though the computation is in LES mode. The grey area issues are well known
245 by DES users and tend to be more noticeable in flows with thin recirculation regions (Spalart, 2009). Also,
246 problems can be encountered if the grid is too fine for the RANS region or too coarse for the LES region

10
247 which can lead to too little turbulence is modelled in the RANS region and too little turbulence is resolved
248 in the LES region, This phenomenon is called 'Modelled Stress Depletion' (MSD) (Spalart et al., 2006).

249 The second issue is called the 'Incursion of LES Mode Inside the Boundary Layer'. The DES97 was based
250 on the assumption that the near-wall tangential grid spacing exceeds the boundary layer thickness and only
251 outside the boundary layer limits dw should be equal to CDES Δ. Nevertheless, the issues associated with this
252 condition's violation were anticipated in the original DES97 publication (Spalart, 2009). Menter and Kuntz
253 (2004) reported the former issue as part of their investigations using the DES97 model for an airfoil with
254 flow separation near the trailing edge. In general, their analysis predicted a separation point further upstream
255 than the location found by using a pure SA model. This phenomenon was named 'Grid Induced Separation'
256 (GIS). Overall, this issue affects the RANS Reynolds stresses that reduce the RANS mode's skin friction
257 calculation in certain situations. Therefore, the grid resolution's sensitivity that gives rise to the appearance
258 of MSD and GIS is identified as the most significant deficiencies of the original DES97.

259 The third issue is called the 'Log-layer Mismatch' (LLM). As reported by Nikitin et al. (2000), a consistent
260 pathology was found in DES where the modelled wall-shear stress deviates from the true one by
261 approximately 15%.

262

263 2.4.2 Delayed Detached Eddy Simulations (DDES) and Improved Delayed Detached Eddy Simulation
264 (IDDES)

265 The mathematical issues outlined in the previous section have been tackled by improvements in the original
266 DES97 turbulence modelling strategy. The most known improved versions of DES97 are the DDES and
267 the IDDES.

268 DDES can correct the MSD issue detected on the original DES97, which typically affects the RANS
269 Reynolds stresses calculations (Spalart et al., 2006). The DDES new capability detects boundary layers and
270 extends the RANS mode further, even if the cell size corresponds to an LES model. This model has also
271 demonstrated the ability to handle the GIS issue successfully. Hence, the method was proposed to replace
272 the DES97 even though it could not handle the log-layer mismatch yet. The IDDES is the latest and more
273 ambitious turbulence modelling strategy developed by Shur et al. (2008), which showed the potential to
274 solve the LLM in addition to the MSD issue. However, it is essential to emphasize that IDDES, is still
275 susceptible to mesh definition as much as DDES. An IDDES-based model may still present MSD if the
276 mesh is inappropriately defined.

11
277 A comparison between experiments and various turbulence modelling strategies is shown in Figure 4, where
278 the vortex structure generated during the interaction of flow with a blunt body is presented. It can be seen
279 that the IDDES method can accurately replicate the flow structure, while RANS approaches show a large
280 deviation from the experimental observation.

281

282

283 Figure 4: Flow behind a blunt body: a comparison between RANS (a) and (b), DES (c), IDDES (d) and
284 experiments (e). (Guilmineau et al. 2018)

285

12
286 2.5. Turbulence modelling and y+

287 The boundary layer of ships is commonly modelled by mesh layers built upon the hull surface in which the
288 near-wall cell is very thin and the thickness of each contiguous layer gradually increases outwards so that
289 the thickness of the outmost layer is close to the size of cells outside of the boundary layer mesh, as shown
290 in Figure 5.

291 The y+, as shown in Equation 11 (ITTC, 2014a), is the dimensionless quantity for the distance from the wall
292 up to the centre of the first grid cell.

293
ℎ
294 𝑦+ = 𝐿 × 𝑅𝑒√0.0375/(log10 𝑅𝑒 − 2)2 (11)
𝑝𝑝

295

296 where h is the thickness of the innermost cell layer, Lpp is the ship length between perpendiculars and Re
297 is Reynolds number.

298 To calculate the near-hull fluid physics in CFD, y+ requires to be between 0 and 5 (most times y+=1) which
299 allows for the resolution of the boundary layer viscous sub-layer. This is translated in a very dense near-
300 wall region which tends to result in computationally expensive simulations. On the other hand, Wall
301 Functions (WF) can get around the unaffordable computational cost by modelling the inner part of the near-
302 wall region (viscous and buffer sub-layers. When using WF, y+ tends to reach values significantly higher
303 than 1 (in the order of 70-100 and even more). More details of WF can be seen in (Pope, 2001).

304

305

306 Figure 5: Illustration of near-hull mesh setup: mesh layers were built between the hull geometry (upper
307 right) and uniform domain mesh (lower left).

308

13
309 In general, it has been well proved that RANS associated with WF allows accurate prediction of ship
310 resistance when Y+ is up to 100, as also recommended by (ITTC, 2014a). However, except resistance
311 prediction, a Y+ of 100 is not sufficiently accurate for other applications, such as propulsions (y+ < 5, (Sun
312 et al., 2020)) and hydroacoustics (y+ < 1, (Smith and Ventikos, 2021)). Meanwhile, other turbulence
313 schemes rather than RANS should be used in those applications, for which, a specific review will be given
314 in Section 3.

315

316 3. Ship hydrodynamic applications

317 This section presents the state of the art of turbulence modelling strategies and their applicability for the
318 prediction of ship resistance, boundary layer, propeller propulsion, cavitation, and structural vibration,
319 alongside comments on their capability and limitations.

320

321 3.1 Resistance

322 DNS, the most complex and detailed strategy for modelling the flow, has been successfully applied to
323 simulate free-surface turbulent wake behind ships in model scale (Shen et al., 2002). The authors tested
324 three 50-cm towed ships with different beams to draft rations during the experiments and at very low Froude
325 number Fr = 0.04. The hull forms were relatively simple, resembling a cylindrical shape as the main interest
326 was to understand the underlying physics of turbulent wake at a Kolmogorov micro-scale and their
327 dependence on basic hull geometric parameters. In general, the results from the DNS simulations (Figure
328 6) are satisfactorily comparative to the experimental results. However, it should be emphasised that the
329 simulations were conducted at very low Reynolds numbers, far from real towing tank experiments.

14
330

331 Figure 6: Instantaneous contours of spanwise vorticity ωz of ship wake at the free surface, obtained using
332 (i) experimental PIV technique and (ii) the DNS simulations. The upper and lower panels show two
333 different hull geometries. (Shen et al., 2002)

334

335 DNS simulations for a realistic hull form geometry in full/model scale have not been reported yet. In
336 general, simulations with much higher Reynolds numbers would present a significantly more turbulent
337 flow. As a consequence, the Kolmogorov micro-scale ηk would reduce dramatically. For example,
338 considering a full-scale DNS simulation of a typical cargo ship with a length L = 150 m at a Reynolds
339 number Re = 109, for such a problem, it is expected that the Kolmogorov micro-scale will be in the order
340 of at least 2 x 10-19 m. Therefore, modelling all the turbulent scales up to the Kolmogorov micro-scale would
341 require an incredibly dense mesh with cell sizes in the ship wall's vicinity measuring at least 2 x 10-19 m.
342 This numerical order of magnitude clearly shows that DNS implies a huge numerical task and remains
343 unlikely to achieve in practice.

344 RANS has been successfully applied to predict ship resistance and the results agree well with test data
345 (ITTC, 2017). Remarkable examples are validations against the model-scale KRISO Container Ship (KCS),
346 designed by the Maritime and Ocean Engineering Research Institute of South Korea, the model-scale Japan
347 Bulk Carrier (JBC), designed by the National Maritime Research Institute, Yokohama National University
348 and Ship Building Research Centre of Japan, and the full-scale Ragel, a real general cargo ship. In addition,
349 there have been a range of workshops reporting work on representative tests and comparing numerical
350 approaches from different participants during blind tests which were validated with experimental data.
351 Some examples are the Gothenburg (Larsson et al., 2013) and the 2015 Tokyo (Larsson et al., 2015)

15
352 workshops in ship hydrodynamics. These efforts have facilitated best practices of ship simulations that
353 provide insights into the applicabilities of different numerical strategies (Eça et al., 2009).

354 According to the ITTC recommendation for resistance predictions (ITTC, 2014b), a y+ value of 30~100 is
355 in most cases sufficient when employing WF RANS. This conclusion was given as part of the 2015 Tokyo
356 Workshop in Ship Hydrodynamics (Hino et al., 2020), where participants demonstrated that WF could
357 achieve more accurate resistance prediction than directly resolving the wall region. Therefore, it is seen that
358 WF is computationally cheaper than wall-resolved methods while it can still provide comparable resistance
359 prediction. This phenomenon is related to the fact that ship resistance is an integral force value, not because
360 WF obtained better fluid details than wall-resolved methods.

361 The computational cost of WF RANS is mostly affordable for standard computers nowadays. The ITTC
362 suggested a value of y+ 30~100 which needs several million elements of mesh for a full-scale container
363 ship (Tezdogan et al., 2015, 2016), and the required elements will be much fewer for model-scale; Terziev
364 et al. (2019) conducted comprehensive model scale resistance computations on the KCS geometry and
365 demonstrated that RANS models could accurately obtain ship resistance when the total cell number is 1~3
366 million.

367 RANS eddy-viscosity models: SST k − ω and k-ε are the most common turbulence modelling strategies
368 used to simulate viscous marine flows (Terziev et al., 2019). In terms of the numerical setup, RANS eddy-
369 viscosity models shelter the mesh near ship walls from the extreme density of the Kolmogorov micro-scale.
370 In general, the SST k − ω has been demonstrated to be a robust turbulence modelling strategy due to its
371 capability to model adverse pressure gradients and flow separation (Paterson et al., 2003). A negative
372 pressure gradient means the static pressure increases in the flow direction, which can happen when a water
373 flow encounters a hull, especially around the stern region (ITTC, 2014b).

374 Model scale investigations using RANS have reached a certain maturity and confidence (Wackers et al.,
375 2011). Abundant comparisons between RANS-CFD and model-test results reported good agreement,
376 during open-water resistance, propulsion, seakeeping, and manoeuvring simulations. Zha et al. (2014a)
377 demonstrated the SST k − ω model can be applied to accurately predict the resistance of large
378 container/tanker ships and Zha et al. (2014b) shown that the same approach also works well for high-speed
379 hulls. Yet, accurate results could also be obtained by the k − ε method, as proved by Dashtimanesh et al.
380 (2020).

381 Wackers et al. (2011) presented that the wake of an advancing ship can be modelled in a high-fidelity
382 manner; Figure 7 and Figure 8 respectively show the validation of overall wave pattern and free-surface

16
383 line along the hull. The authors demonstrated that the selection of various RANS models does not make a
384 significant difference in this particular application. Since ships' wave pattern is more of a qualitative
385 criterion and subject to certain uncertainties, subtle disagreements between different RANS models are
386 reasonable. Shen and Wan (2013) presented comparisons between RANS (the SST k-ω) simulations and
387 the experiment of a ship advancing in head waves, demonstrating good accuracy for the prediction of ship
388 motions and added resistance. The SST k-ω method has also successfully replicated measurements during
389 CFD self- propulsion tests (Winden et al., 2014) and ship performance in confined waterways (Huang et
390 al., 2020, 2021a, 2021b). In addition, Huang et al. (2021c) successfully reproduced the fierce water-entering
391 process of a freefall lifeboat by the SST k-ω model, demonstrating that the calculated impact forces agree
392 well with full-scale measurements. These examples clearly show that RANS models, particularly the SST
393 k-ω model, can provide accurate predictions of ship resistance, motions and wave patterns.

394

395

396 Figure 7: Comparison of computed (top) and measured (bottom) wave pattern for the Series 60 ship
397 (Wackers et al., 2011).

398

17
399
400 Figure 8: Validation of wave profile for the standard Wigley hull advancing in calm water. (Wackers et
401 al., 2011)

402

403 In shallow water CFD simulations, Terziev et al. (2019) compared results from different eddy viscosity
404 models with experimental results to calculate ship resistance. They demonstrated that the results from
405 different RANS schemes are at the same accuracy level, as shown in Figure 9.

406

407

408 Figure 9: Resistance deviation of different RANS eddy viscosity models. (Terziev et al., 2019)

409

18
410 In full scale, turbulence modelling strategies validations for the simulation of ship viscous flows have barely
411 been conducted. In general, it is known that Reynolds number differences between full-scale and model-
412 scale simulations significantly affect the boundary layer structure. This leads to considerable discrepancies
413 between ship hydrodynamics in model scale and full scale. ITTC provided a guideline on correlating ship
414 resistance from model-scale to full-scale by correcting the viscous resistance (ITTC, 2008). However,
415 detailed physical features would not be able to be revealed using such a correlation approach. It is worthy
416 to mention that the hull roughness is particularly influential for full-scale modelling, because higher skin
417 friction results in a higher roughness Reynolds number, which makes the scale effect significant. The
418 validation study of Song et al. (2021) shows different levels of roughness setups in CFD will result in
419 different boundary layer behaviours as well as resistance deviations against experiments, where the
420 modelling of roughness in CFD is achieved by modifying the WF model. The detailed equations for
421 roughness modelling can be referred to (Song et al., 2021).

422 Unsteady flow features have been successfully modelled by using the Reynolds Stress Model (RSM).
423 Sotiropoulos and Patel (1995) compared the stern and wake flow predicted by an eddy viscosity model and
424 RSM. The comparisons show that the RSM accurately predicts most of the experimentally observed flow
425 features in the stern and near-wake regions, whereas the two-equation model predicts only the overall
426 qualitative trends. In particular, solutions with the RSM accurately display the origin of the stern vortex.
427 Furthermore, Zou et al. (2010) shown the excellent capability of an EARSM model (Explicit Algebraic
428 Reynolds Stress Model) (Deng et al., 2005) on simulating manoeuvring, by which they presented a tanker
429 performing turns at different speeds; and Mucha et al. (2016) reported that both the EARSM and the SST
430 k-ω RANS models can accurately predict ship hydrodynamics in a canal scenario. Although RSM has
431 shown the potential of obtaining better accuracies than eddy-viscosity based RANS models on flow feature
432 predictions, it has a much higher computational requirement. Thus, RSM usage has not been popular in
433 ship design, since the much cheaper eddy-viscosity-based RANS can already offer a relatively accurate
434 prediction on the integral performance of ships, e.g. resistance or motions.

435 IDDES ship resistance calculations found in the literature were performed by Kornev et al. in model scale
436 (2011). The authors compared the ship resistance components' prediction using an IDDES based
437 computational model vs different RANS turbulence approaches. Their investigation showed that the IDDES
438 set-up could predict the total resistance with just a 1% deviation from the experimental results measured
439 during the towing tank calculation (Kornev et al., 2011). By contrast, all RANS models deviated from the
440 experimental outputs with a margin between 3% to 7.5%.

19
441 Another set of model scale ship resistance calculations using the IDDES formulation with a y+ of less than
442 one were performed by Kornev et al. (2018). This time, they tested the Japan Bulk Carrier (JBC) benchmark
443 test case presented during the 2015 Tokyo workshop in Ship Hydrodynamics (Hino et al., 2020). They
444 compared the ship resistance components' prediction when using an IDDES based computational model vs
445 RANS calculations. Their investigation showed that the IDDES set-up could predict the total resistance
446 with a 4% deviation from the experimental results measured during the towing tank calculation (Kornev
447 and Abbas, 2018). The RANS set-up, however, deviated by just 2.3% from the experiments.

448 In addition, Kornev et al. (2018) compared RANS and IDDES on predicting friction and pressure resistance
449 components, CF and CP, and the results were unexpected. The RANS setup CF was compared with calculated
450 CF using the well-known ITTC-57 friction line (ITTC, 1957), showing a 2% deviation between the CFD
451 output and the empirical friction line. By contrast, the IDDES computational setup deviated by 15% from
452 the ITTC57 empirical formulation. Kornev et al. (2019) attributed this deviation to the difficulty to
453 reproduce the turbulence transport through the interface between the RANS and LES zones. They believe
454 that the LES area approaches the wall on certain occasions, resulting in an underestimation of the vortex
455 generation in the near-wall region. This phenomenon results in a Grid Induced Septation, which
456 overpredicts the pressure resistance. Nevertheless, the ITTC friction line was designed for ship hull
457 geometries built before 1957. Therefore, considering the demonstrated accuracy of the IDDES for
458 calculating the total resistance, a deviation in CF with the 1957 experimental line could be impractical when
459 analysing a modern hull form.

460 Kornev et al. (2018) results may not be taken as conclusive. It has been observed that the RANS setup
461 showed grid convergence between the mesh refinements. However, for the IDDES case, it can be seen that
462 setup convergence was not fully achieved, and perhaps one further refinement would have been needed.
463 Moreover, RANS simulations demonstrated to provide substantially accurate results when compared to sea
464 trials. Therefore, it seems that there is not much room for further improvement by IDDES.

465 LES for ship resistance modelling has been shown to be prohibitive as well as unnecessary. Liefvendahl
466 and Fureby (2017) estimated the cell number of LES required to simulate a whole hull are around 8 x 10 8
467 in model-scale and around 6.7 x 1013 in full-scale. Simulations may be therefore achievable in model-scale
468 (Nishikawa et al., 2012) but still unaffordable in full-scale even though High-Performance Computing
469 (HPC) facilities are available. To date, there is no publication found on simulating a whole full-scale ship
470 using LES, partially because it is too expensive, and also because RANS has been widely proved accurate
471 enough to provide an estimation of integral forces such as resistance or thrust.

472

20
473 3.2 Boundary layer and propeller wake

474 Boundary layer (local flow) and wake studies tend to be significantly more complex than resistance
475 investigations. The Tokyo 2015 Workshop in Ship Hydrodynamics (Larsson et al., 2015) closely examined
476 the accuracy of different turbulence modelling strategies and different wall treatments for different
477 geometries. When applying RANS turbulence approaches to high block coefficient ships such as the JBC
478 in model scale, where significant flow separation is expected, it was found that using low values of y+
479 provides a more accurate boundary layer structure than high y+. This is displayed in Figure 10, which
480 compares experimental measurements of the non-dimensional velocity (left subfigure) against CFD
481 computations using wall modelled RANS (high y+, middle subfigure) and wall-resolved RANS (low y+,
482 right subfigure). The comparison shows that only low y+ functions were able to calculate the 0.4 non-
483 dimensional velocity isolines. Based on the comparison wall-resolved RANS performs better than wall-
484 modelled RANS when flow separation is important.

485

486
487 Figure 10: Velocity contour comparison for the model-scale JBC geometry between experiments (left),
488 high y+ (middle) and low y+ (right). (Larsson et al., 2015)

489

490 Kornev et al. (2018), who participated in the 2015 Tokyo Workshop, studied the velocity contours on the
491 aft region for the JBC hull with a high block coefficient (Cb > 0.8) in model scale. The authors compared a
492 RANS and an IDDES with experimental results at a Re = 7.46 x 106 as shown in Figure 11. The figure
493 shows the results from the towing tank tests, the IDDES approach's results, and the computations using the
494 RANS SST k-ω. In general, it can be seen that the IDDES approach was able to capture the bilge vortex

21
495 with a high degree of precision. Therefore, the experiments corroborated the enhanced IDDES' capability
496 to accurately predict the vorticity and vortex shedding structures emanated from the hull. The authors also
497 remarked that IDDES capability is particularly noticeable if Cb > 0.8. By contrast, the RANS model
498 significantly deviated from the experimental results and calculated a much weaker bilge vortex, which
499 indicates that the unsteady turbulent behaviours can be eliminated by the averaging-process of RANS,
500 alongside corresponding structural vibrations.

501

502

503 Figure 11: Study of the instantaneous velocity contours on the aft end of a model scale ship by (a) model
504 scale experiments, (b) IDDES and (c) RANS. (Kornev et al, 2018)

505

506 DES models are known to overperform RANS when separated flows are expected. A typical case is seen
507 when computing geometries both slender and bluff bodies at a static drift angle (Xing et al., 2012) and
508 (Bhushan et al., 2011). However, for low-blocky ships with minimal separated flows, the literature reveals
509 some limitations in DES models. For example, the 2010 Gothenburg Workshop in Ship Hydrodynamics
510 (Larsson et al., 2013) reported that DES models suffer from modelling issues such as Modelled Stress
511 Depletion (MSD) which manifest as a difficulty to reproduce the turbulence through the interface between
512 RANS and LES zones in the flow attached regions (Spalart et al., 1997). MSD is seen in cases when the
513 LES length scale drops below the RANS length scale close to the wall region, triggering the numerical
514 model into LES mode close to the wall. In such a scenario it is possible that the desired LES region is not

22
515 refined enough and cause an error. This issue may significantly affect the performance of the model for
516 non-separated flows, which are more typical on low block coefficient geometries. As corroboration of this
517 theory the 2010 Gothenburg Workshop (Larsson et al., 2013) and 2015 Tokyo Workshop (Hino et al., 2020)
518 both confirmed that RANS models are able to provide a good representation of the wake and accurate self-
519 propulsion results for low blocky ships, such as KCS and DTC.

520 Viitanen et al. (2018) conducted open water model scale marine propeller DDES simulations that were
521 compared with experimental results. Propeller global forces, local flow phenomena, and cavitation patterns
522 were compared to cavitation tunnel tests results. As shown in Figure 12, the results from their investigation
523 suggested that the DDES numerical approach accurately predicted the flow pattern. The authors emphasised
524 that the propeller's DDES open water characteristics underpredicted the thrust by 2-8%. By contrast, the
525 torque was accurately predicted with minimal deviations of just 1%. The author attributed the deviations to
526 possible confinement effects due to the geometry of the cavitation tunnel, which was simplified during the
527 simulations. This study also found a good agreement between the numerically simulated propeller wake
528 patterns and experimental LDV measurements.

529

530

531 Figure 12: Effective wake for open water marine propeller (a) experimental results and (b) DDES
532 (Viitanen et al., 2018)

533

534 The enhanced capability of the IDDES has also been noted in full-scale. Pena et al. (2020a; 2020b)
535 compared nominal wakes (Figure 13) and vorticity fields on the aft end of the ship in full-scale. Two RANS
536 turbulence modelling approaches suitable for the simulation of ship flows have also been assessed: the k-ε

23
537 model and the k-ω SST model. The results from their investigations revealed significant deviations from
538 the calculation of the wakes in full-scale when using different turbulence modelling strategies. Both RANS
539 setups demonstrated a fair agreement. Differences were observed in the region closer to the propeller centre.
540 More substantial velocity retardation and a higher degree of three-dimension of the wakefield can be seen
541 for the SST model. On the other hand, the hybrid IDDES model showed close to zero velocity fluid on the
542 region under the shaft.

543

544

545 Figure 13: Nominal wake prediction using different turbulence modelling strategies Pena et al. (2020a;
546 2020b)

547

548 The work of Pena et al. (2020a; 2020b) confirmed that RANS numerical simulations in full scale resulted
549 in a less accurate prediction of the velocity fields within the ship boundary layer and wake. For the first
550 time, they attributed numerical issues linked to choosing an improper turbulence modelling strategy to the
551 discrepancies between sea trials and numerical simulations of energy-saving devices performance (ITTC,
552 2014b). They recommended using an IDDES-based numerical approach to investigate the wake and
553 boundary layer of full-scale ships with a particular focus on the design of propellers, rudders or energy-
554 saving devices.

555 LES has been applied to predict high-vortex flows generated with a ship. Yang et al. (2008) simulated a
556 whole ship in model scale using Wall Resolved LES (WRLES), as shown in Figure 14. Their results did

24
557 show that WRLES produced high-fidelity boundary layer structures, as shown in Figure 15. However, they
558 reported that the simulations required a cell number of around 268 million even in model scale, which is
559 hardly affordable in the industry. On the other hand, Liefvendahl and Fureby (2017) conducted model scale
560 simulations of the JBC using Wall Modelled LES (WMLES) and WRLES approaches in model scale with
561 meshes of 19 million and 800 million elements respectively. From Liefvendahl et al. experience, it is
562 possible to remark that the WRLES approach is feasible if only one simulation is required, while ship design
563 usually needs to run extensive simulations considering a combination of different inputs (e.g. operation
564 condition, ship hull variant). A later study from Liefvendahl and Johansson (2021) compared again the two
565 approaches and its applicability in the marine industry for the prediction of model scale ship bulk carrier.
566 Their investigation showed that WMLES is feasible for model scale ship hydrodynamics and with
567 substantial mesh savings when compared to wall-resolving LES. Therefore, WMLES could be the solution
568 to make LES accessible for full ship simulation.

569

570

571 Figure 14: LES simulation of vortex structures associated with an advancing ship. (Yang et al., 2008)

572

25
573
574 Figure 15: Flow pattern comparison near the bow. (Yang et al., 2008)

575

576 Alternatively, LES can be applied to localised ship design. In this manner, high-density meshes are just
577 needed to be applied to a part of a hull where turbulent behaviours are of great significance. For example,
578 Arslan et al. (2016) put only two-dimensional sections of a hull into LES simulations, thus investigating
579 the motions of a ship subjected to lateral flows, i.e. cross currents. They provided parallel CFD and model-
580 test comparisons to show that considerable vortex structures are generated near the ship section and can
581 bring about instability to manoeuvring, with LES's excellent capability illustrated by Figure 16. Moreover,
582 they reported RANS failed in this case, as the small-scale vortexes are roughly left out and the vortex centre
583 is inaccurately predicted, see Figure 17. Thus, it remains doubtful on applying RANS to predict the vortex
584 structures associated with the interactions of flow with a certain ship part. More examples of comparing
585 LES and RANS in localised ship design will be presented in Section 3.4.

586

587

26
588

589 Figure 16: LES (left) and experimental (right) views of vortex structures occurring when a lateral section
590 of a hull is subjected to a cross current. (Arslan et al., 2016)

591

592 Figure 17: LES (left) and RANS (right) views of vortex structures occurring when a lateral section of a
593 hull is subjected to a cross current. (Arslan et al., 2016)

27
594 3.3 Self-Propulsion

595 Self-propulsion simulations in full scale revealed that RANS could provide accurate propeller thrust and
596 torque predictions using the sliding mesh approach (Ponkratov and Zegos, 2015; Pena et al., 2019), or the
597 actuator disc approach (Jasak et al., 2019; Bakica et al., 2020). Both approaches have demonstrated provide
598 results in line with the Lloyds Register First Full-scale Ship Hydrodynamics Workshop (Lloyds Register,
599 2016). Jasak et al. (2019) demonstrated that using the SST k − ω RANS scheme could achieve agreement
600 with experiments with only 6 million cells, as shown in Figure 18, while it should be noted that Jasak et al.
601 did not discretise the propeller geometry which results in less computational mesh requirements. It should
602 also be noted that they used very sophisticated local mesh refinements in their in-house code, shown in
603 Figure 19, which may be hard to replicate by others. This indicates local mesh refinements can be an
604 important skill when considering the compromise between turbulence modelling strategies and
605 computational costs. Nevertheless, the results are still encouraging as even 10~20 million cells would be
606 affordable with HPC that have been popularised nowadays.

607

608

609 Figure 18: Mesh sensitivity study for predicting self-propulsion thrust, alongside a comparison with full-
610 scale measurement data. (Jasak et al., 2019)

611

28
612

613 Figure 19: Mesh layout of a self-propulsion ship in full scale, in which sophisticated local mesh
614 refinements are applied. (Jasak et al., 2019)

615

616 Using DES, model scale self-propulsion simulations in rough waters were conducted by Carrica et al.
617 (2011) based on a fully discretised propeller with dynamic overset grids approach. They compared results
618 between CFD computations and experiments and showed deviations in calculating the thrust and torque
619 coefficients KT and KQ of 6% and 5%, respectively. The deviances with the experimental results are likely
620 to be minimised by implementing an IDDES numerical model to avoid the DES97 numerical issues.

621 Using IDDES, Kornev et al. (2018) numerically and experimentally compared model-scale calm water self-
622 propulsion tests using the JBC benchmark test case presented during the 2015 Tokyo workshop in Ship
623 Hydrodynamics (Hino et al., 2020). During the analysis the authors used the Arbitrary Mesh Interface
624 (AMI) module provided in OpenFOAM to model the interface between static (hull) and rotating (propeller)
625 grids with a y+ of 2.9 at the propeller and 15 on the hull. As a result, the authors used dense meshes
626 exceeding 10 million elements, which requires high computational resources. The results from this
627 investigation revealed that KT and KQ were predicted with a degree of accuracy of 7% and 2% respectively.
628 They also used RANS to do the same validation exercise and found that the RANS set-up showed a higher
629 degree of accuracy with KT and KQ deviances of 1 and 5% respectively.

29
630 Pena et al. (2020a; 2020b) compared the torque coefficients obtained at the self-propulsion point, using
631 three different turbulence approaches, against the experimental values measured on the sea trials (Figure
632 20). This work was based on a ship geometry with a Cb = 0.77 and sea trial measurements that enabled the
633 numerical model's validation. The authors remarked that three numerical models produced consistent
634 results compared with the experimental data, with an 8% difference for the k-ε and a 5% difference for the
635 SST. The IDDES was the most accurate of the three set-ups, producing results within an excellent 2%
636 difference from the sea trials measurements.

637

638

639 Figure 20: KQ for the three turbulence modelling strategies and the experimental measurements at three
640 shaft speeds. (Pena et al., 2020a; 2020b)

641

642 Also, a recent trend for self-propulsion modelling recommends the implementation of wall-roughness
643 functions which seem to yield more accurate results than the ones obtained when using a smooth wall non-
644 slip condition (Mikkelsen and Walther, 2020) and which may account for biofouling, the paint roughness,
645 welding seams or hull plate deformations on the hull (Pena et al., 2020b). Implementing roughness in WF
646 is expected to increase the boundary layer thickness and nominal wake disturbances as shown by Song et
647 al. (2019), which may impact the calculated propeller performance. This new approach seems promising,
648 however, multiple studies presented in the literature have shown accuracy during self-propulsion tests
649 without the necessity of the implementation of wall-roughness in full-scale (Ponkratov and Zegos, 2015;

30
650 Ponkratov, 2017; Jasak et al., 2019; Pena et al., 2020b; Bakica et al., 2020). Therefore, further investigations
651 are required to clarify the role of wall roughness for self-propulsion modelling.

652 The reviewed literature shows that it is not yet well-established which turbulence modelling strategy yields
653 the most accurate results versus which propeller modelling approach (e.g. a virtual disk or a physical disk).
654 Different approaches do not show an obvious difference in terms of macro-parameters calculation such as
655 torque and thrust coefficients. For example, Carrica et al., (2011) using a physical-disk + DES approach
656 obtained a higher level of KT and KQ than those obtained by Jasak et al. (2019) using a virtual-disk + RANS.
657 It is therefore suggested to use a virtual disk approach to predict macro-parameters due to its relatively low
658 computational requirements. By contrast, if local flows need to be studied as part of the self-propulsion
659 analysis, then a fully discretised propeller approach is recommended. In this case, high computational
660 resources (very dense meshes) will be required. If during the self-propulsion tests flow separation is
661 expected and the unsteady loads at the propeller need to be studied, then may be recommended to implement
662 a DES approach. However, the simplest cases with no flow separation could still be modelled by a RANS
663 approach.

664

665 3.4 Cavitation

666 Cavitation is a phenomenon in which rapid changes of pressure in a liquid lead to the formation of small
667 vapour-filled cavities in places where the pressure is relatively low, commonly occurring with the operation
668 of propeller and hydrofoil. RANS has shown capable of generating gas-phase out of water-phase during
669 propeller cavitation simulations (Watanabe et al., 2003; Zhang et al., 2006), this means the RANS equations
670 have no problem on accurately obtaining pressure solutions and capture such low pressure and phase
671 changing. It has been shown in the literature that RANS has been successfully applied to predict the risk of
672 cavitation occurring. For instance, Lu et al. (2012) applied RANS to investigate the cavitating flow on a
673 marine propeller. Upon comparison against experiments, they found that RANS has captured the occurrence
674 of the root cavitation, despite that the evolution of cavitation was only partially simulated. For the above
675 reason, RANS is still reckoned as the most productive approach for cavitation modelling.

676 RANS may be applied as an indicator of whether or not a cavitation phenomenon should appear, while a
677 higher-order approach (Hybrid, LES) is still required to study the changes in ship characteristics due to the
678 presence of cavitation. RANS was found to neglect bubble phenomena that are common in cavitation, as
679 shown in Figure 21. This can be certainly attributed to the time-averaging processing of RANS. This was
680 confirmed by Bensow (2011), who provided a comparison between LES, DDES and RANS on modelling
681 the cavitation vapour generated within the wake of a foil as shown in Figure 22. The figure confirms that
31
682 RANS fails in simulating the detailed vapour, while LES and DDES are both capable of such modelling,
683 with LES presenting more detailed small-scale vapours than DES.

684

685

686

687 Figure 21: Cavity bubbles generated by rotating blades: a comparison between RANS and experimental.
688 (Watanabe et al., 2003)

689

32
690

691

692

693 Figure 22: Simulation of cavitation on the Delft Twist11 foil: a comparison between LES (left column),
694 DDES (middle column), and RANS (right column). (Bensow, 2011)

695

33
696 Viitanen et al. (2018) used DDES to simulate the cavitation patterns generated by propellers and provided
697 a comparison against experiments. Their results are presented in Figure 23 in which cavitation is seen to be
698 exceptionally replicated. The superiority of the DDES hybrid approach is demonstrated for cavitation
699 assessments.

700

701

702 Figure 23: Tip and hub vortex structures during (a) experiments and (b) DDES computations. (Viitanen et
703 al., 2018)

704

705 Cavitation is an important factor to cause erosions on ship propeller blades, as it can cause severe damages,
706 erosions and noise. Therefore, accurately modelling propeller erosion demonstrates that it is of great
707 importance that the selected turbulent modelling strategy captures vapour bubbles. This fact remarks on the
708 unsuitability of RANS for erosion predictions. Ponkratov (2015) conducted erosion studies on a full-scale
709 self-propelled ship using DES. The results from their numerical assessments were compared to borescopes
710 observations and underwater propeller inspections. The author reported that CFD underpredicted the
711 development of the tip vortex. This phenomenon was attributed to an under the mesh definition, which was
712 not considered fine enough. However, the numerical model was able to predict cavitation bubbles collapse
713 on the blade surfaces accurately. Therefore, the authors considered their DES model able to accurately
714 predict the propeller erosion areas as confirmed by comparison with underwater propeller inspection results.

34
715 Usta et al. (2017) studied ship propellers erosion caused by cavitation using a DES approach and provided
716 a comparison with experiments showing promising capabilities of DES on predicting the erosion damage.

717 On the other hand, a more accurate way to represent erosion could be by investigating localised turbulent
718 flows using an LES approach. Mahesh et al. (2015) performed LES simulations on ships' three important
719 phenomena: crash-back, cavitation, and hydro-acoustics. Following validation against experiments, the
720 authors demonstrated that the turbulent behaviours within the phenomena can be accurately captured by
721 LES. They also showed RANS are defeated by LES in those tests.

722 Crashback is an off-design condition in which the marine vessel moves in the forward direction, whereas
723 the propeller rotates in the reverse direction, yielding negative thrust. Flow around the propeller in
724 crashback is characterised by large-scale unsteadiness and flow separation. High amplitude off-axis forces
725 and moments are produced by this unsteadiness, which is transmitted to the body, adversely affecting its
726 manoeuvrability and causing damage and reduction in the performance of propeller blades (Mahesh et al.,
727 2015). Jang and Mahesh (2013) presented the forces of crashback on a ship propeller in the frequency
728 domain, as in Figure 24, alongside validation showing the superb capability of LES on reproducing the
729 unsteady process, especially the low-frequency forces due to flow separation. Wang et al. (2009) assessed
730 the noise caused by a foil, also showing LES can accurately predict the unsteady flow at a frequency-domain
731 level, as shown in Figure 25.

732 Moreover, Liefvendahl (2010) demonstrated that LES can provide the essential fluid instability within
733 propeller flows, important for noise estimate, and such information would be ignored by RANS' averaging
734 nature. More examples of LES simulations can be found as assessing unsteady force during crash-back
735 (Verma, Jang and Mahesh, 2012; Jang and Mahesh, 2013), cavitation (Bensow and Bark, 2010; Dittakavi
736 et al., 2010; Gnanaskandan and Mahesh, 2015; Lu et al., 2014), and noise (Wang and Moin, 2000).

737 Following the discussion about, LES has been widely applied to localise ship design. It proves capable of
738 capturing desired information, and those simulations are achievable with a cell number at a level of 107 to
739 mesh a specific ship component (Liefvendahl et al., 2010). However, considering multiple components in
740 a larger scale will significantly increase the computational cost and lead LES to be an unaffordable
741 approach, limiting the applicability of standalone LES, because the interactions between multiple
742 components could be essential (Benites-Munoz et al., 2020). In this context, the hybrid method, that allows
743 the simulation of multiple components/zones with different fidelities, is feasible to model a larger ship part
744 using local refinements at key locations.

745

35
746

747 Figure 24: Validation of thrust coefficient (left) and side-force coefficient (right) in the frequency domain.
748 Solid line: LES; dash line: experimental. (Jang and Mahesh, 2013)

749

750

751 Figure 25: Frequency spectra of pressure fluctuations on the suction side. Solid line: LES; dash line:
752 experimental. (Wang et al., 2009)

753

754 4. Discussion

755 RANS, as illustrated by multiple sources above, neglects the unsteady flow features. This lies in the
756 averaging nature of its mathematical formulation, as shown in Equations (4) and (5). However, RANS
757 proved to provide sufficient accuracy in resistance predictions wake (thus wave-making resistance) and
758 vessels' motions. This essentially means RANS can accurately predict the total hydrodynamic loads of a
759 ship. As expressed in Equation (12), the hydrodynamic load on a ship is integrated of pressure and viscous
760 force on its surface mesh.

761
36
 (12)
𝐅𝐡 = ∫ (− p 𝐧 + τ ∙ 𝐧) dS

762

763 The hydrodynamic force then subsequently governs the motions of a ship, if the ship is not fixed. This can
764 be considered as the combination of translation and rotation, which was solved with the rigid-body motion
765 equations in the body-fixed system based on the mass centre of the ship G-x'y'z', as expressed in Equation
766 (13) and (14) (Huang and Thomas, 2019). This is to say, if pressure and viscous terms are solved correctly
767 as in the RANS equations, plus an appropriate dynamic mesh is applied, they should integrate into a correct
768 hydrodynamic force, e.g. resistance, despite it is an averaged value for each timestep. This, on the other
769 hand, corroborated why mesh convergence is so important.

770

⃗⃗⃗⃗G
dV (13)
𝐅=m
dt

dω
⃗⃗⃗⃗⃗G (14)
𝐓 = [J] ∙ + ⃗⃗⃗⃗⃗
ωG × ([J] ∙ ⃗⃗⃗⃗⃗
ωG )
dt

771

772 where F and T are the total force and torque on the ship, induced by the gravity, the hydrodynamic load
773 from the surrounding fluid Fh; m and [J] are the mass and inertia moment tensor respectively, and VG and
774 ωG are the translational and rotational velocity vectors of the ship respectively.

775 In line with these features, previous studies have demonstrated that RANS models can accurately obtain
776 ship resistance and motions in various situations. This conclusion has been supported by extensive model
777 tests. Among mainstream RANS options, the k − ω SST scheme is the most favoured and validated one.
778 For full scale, the derivation of formulae from model tests usually needs to apply the ITTC extrapolation
779 procedure since it is impossible to ensure Froude and Reynolds numbers are both equal between full scale
780 and model scale, in which, the former governs gravity/inertia (waves) forces and the latter dictates viscous
781 forces. This procedure divides the total ship resistance into a wave component and a friction component.
782 Scaling based on a consistent Froude number is practical in model tests, which scales the wave component
783 correctly while bringing about certain errors within the friction component due to a changed Reynolds
784 number. The latter may be corrected using the ITTC correlation method (ITTC, 2008).

37
785 Directly simulating ship performance in full scale is applicable, as it does not conflict with the CFD theories.
786 Full-scale simulations will make the ITTC extrapolation procedure unnecessary, thus simplifying the
787 procedure of resistance prediction. Moreover, it can solve the boundary layer issues associated with a
788 changed Reynolds number and remain the correct scale's fluid behaviours. However, due to a lack of full-
789 scale sea trial data, this has not been fully validated and will need years of a process. Preliminary results
790 have shown to be promising (Lloyds Register, 2016), which gives confidence in RANS. Considering the
791 mathematic roots, model-scale validations and full-scale validations, this review proposes that RANS can
792 provide sufficiently accurate resistance predictions and there is not much room for improvement from
793 applying more complex turbulence modelling strategies, i.e. RANS is the fastest and cheapest approach to
794 predict ship resistance.

795 On the other hand, as the statistically time-averaging process within RANS ignores certain fluid
796 instabilities, important physics might be oversimplified in certain scenarios. This caused RANS to be
797 proved incapable of applications where fluid instabilities are significant, such as structural vibration and
798 noise predictions. Meanwhile, the relatively coarse mesh of RANS has been shown to introduce
799 inaccuracies to small vortex structures, important within propeller flows. These incapabilities have limited
800 the application of RANS to assessing integrated ship performance. Thus, RANS suitability depends on the
801 problem of investigation, e.g. integrate or local, steady or unsteady; and based on its certain inapplicability,
802 there is a need to use higher-order turbulent strategies to simulate such problems. It has been a trend that
803 RSM models can remedy the rough flow features predicted by eddy-viscosity based RANS; however, RSM
804 has to date much less applied to ship hydrodynamic simulations. DNS, on the other hand, could allow the
805 modelling of more complex flows. Yet, DNS has no practicality for the simulation of ship viscous flows in
806 the near future. Moreover, the numerical order of magnitude of mesh and timestep requirements makes
807 DNS a huge numerical task and remains unlikely to be achieved in practice. Thus DNS is more suitable for
808 relatively pure fluid dynamic analyses. Also, as a cheaper solution than DNS, LES has proved to provide
809 sufficient accuracy in complex ship design applications.

810 LES allows the modelling of complex flow features that RANS inherently neglects. It was shown to be very
811 accurate in predicting detailed ship flow behaviours, able to capture the instabilities ignored by RANS.
812 Thus, LES could be successfully applied to the assessment of structural vibration, noise, and erosion.
813 Multiple publications even used frequency spectrums to demonstrate the excellence of LES. However, the
814 main disadvantage of using LES is the requirement of very fine grids in the near-wall regions of the flow
815 domain and a minimal timestep size. Consequently, pure LES applicability for the simulation of an entire
816 hull is deemed far from reality in the industry. The literature has demonstrated that LES is currently just
817 applied to localised simulations limited to a specific component's scale, such as a propeller. Thus, LES has

38
818 still been deemed an unaffordable approach for scenarios where interactions between ship components are
819 important.

820 Hybrid turbulence modelling strategies based on the combination of RANS and LES techniques are
821 becoming increasingly popular in ship hydromechanics. They have the inherent ability to predict well
822 highly unsteady and separated flows while their mesh requirements are lower than LES. Also, they tend to
823 present fewer convergence issues than RSM and be less sensitive to inlet boundary conditions. One of the
824 main applications where hybrid DES approaches have shown clear superiority is for the prediction of local
825 flows, ship boundary layers and wakes when flow separation is expected. Therefore, it is recommended to
826 use a hybrid approach for simulating flows for propellers, rudder loading, energy-saving devices, and any
827 other applications where boundary layer and wake flows need to be assessed in detail.

828 However, there are two important challenges for hybrid DES approaches. First, the transition between
829 RANS and LES reckons to be the main obstacle of DES. One of the consequences is that there is a high
830 risk from the grid point of view that the simulation can act in RANS mode in the targeted LES region. In
831 the same line, the supposed LES zone might end up being treated as RANS. This phenomenon happens if
832 no special care is taken while meshing, which might hinder hybrid models' popularisation. Even with the
833 most advanced DES approach, the IDDES still shows incapable of addressing physical abnormalities
834 present in the formulation. Therefore, it is clear that more computational developments are expected to
835 avoid the grey area issues present in contemporary IDDES approaches. Secondly, more advanced DES
836 models need still to be validated for the simulation of ship viscous flows. Other DES approaches have
837 already demonstrated their superiority for the simulation of viscous flows in industrial applications. For
838 example, ZDES (Zonal Detached Eddy Simulations) allows the CFD user to define the RANS and LES
839 zones, but this approach has not been validated for ship viscous flow applications yet. Therefore, a CFD
840 user intending to implement hybrid approaches should always check that the targeted RANS/LES zones are
841 actually treated as RANS/LES in the CFD software before accepting the results obtained from the run.

842 An alternative approach, WMLES may offer significant advantages when compared to wall-resolving LES.
843 As demonstrated by Liefvendahl and Johansson (2021), WMLES is excellent for modelling detailed ship
844 hydrodynamics and meanwhile offers a substantial mesh saving when compared to a pure wall-resolving
845 LES approach. However, WMLES techniques also present challenges and may provide numerical
846 inaccuracies such as the log-layer mismatch. According to Larsson et al. (2016), in WMLES, the height of
847 the wall-modelled region should be carefully chosen as this parameter directly impact the mesh resolution
848 and modelled results; An inappropriate selection may cause the wall shear stress to be under/overpredicted,

39
849 influencing the modelling of flow details such as separation. Besides, WMLES simulations are expected to
850 have higher computational costs than DES simulations.

851 In general, DES and WMLES (both of them model the near-wall region) are still under development and
852 being studied for marine applications. They are more expensive than RANS, although cheaper than Wall-
853 Resolved LES. It is expected that the errors and inaccuracies of these models may be corrected in a near
854 future. An enhanced practice on local mesh refinement may also speed up this popularisation (Jasak et al.,
855 2019). In addition, developments in data-driven CFD (Pena et al., 2020c; Pena and Huang, 2021) could
856 make the DES set-up task semiautomatic. This would make high-order turbulence modelling approaches
857 more accessible for the ship hydrodynamic CFD community.

858 In the present discussion, decision-making mainly involves making decisions considering the compromise
859 between turbulent fidelity and computational costs. It is worth noting that such difference in computational
860 costs lies in mesh requirement, specifically, the precise level of the modelling boundary layer, reflecting
861 the Y+ value. Thus, it has been a consensus that RANS simulations can provide accurate results with a Y+
862 of 100, while that shall be 1 for LES (ITTC, 2014a; Arslan et al., 2016), and the Hybrid approach has
863 different requirements in the RANS region and the LES region respectively. This determines the mesh
864 density around the hull geometry. As numerical divergence will occur when the expansion ratio between
865 mesh layers is too high (a ratio of less than 1.25 is recommended), the cell thickness can only gradually
866 increase layer-by-layer from the geometry surface towards the whole computational domain. Thus, the
867 limitation of boundary layer geometry will cost the computational cells to change as a whole. On the other
868 hand, denser mesh shall be matched by a smaller timestep. This relationship lies in the definition of Courant
869 number (Co):

870

uΔt
871 Co = (15)
Δx

872

873 where Δt is the timestep size, u/Δx is its normal velocity divided by the distance between the cell centre
874 and the centre of the neighbour cell. To avoid numerical convergency, the Co value should be restricted for
875 different applications, such as less than 1 for relative mild flows and less than 0.3 breaking type flows
876 (Huang et al., 2019, 2021c). Therefore, the turbulence modelling strategy selection will dictate y+ then link
877 to overall mesh density and timestep choices, thus enormously influencing the computational costs.

878 Thereafter, the recommendations of the present work will be made based on two principles:

40
879 (a) the purpose of the specific ship-design task can be done, and the accuracy should be acceptable.

880 If principle (a) is satisfied:

881 (b) using as cheap a turbulence modelling strategy as possible, here meaning the least mesh requirement.

882 Based on this paper's review and discussion, the recommendation of turbulence strategies for various ship
883 hydrodynamic simulations is given in Table 1.

884

885

886

887

888

889

890

891

892

893

894

895

896

897

898

899

900

901

41
902

903 Table 1: Applicability summary and recommendations of turbulence modelling strategies.

904

Recommending
Maturity and accuracy (deviation against
Applications turbulence Mesh estimation Y+
measurements)
strategy

Model scale: 1-5 million (Zha et al., Mature technology,

Ship 2014a; Dashtimanesh et al., 2020)
resistance/motio RANS 30-100 less than 10% deviation
ns Full scale: 5-10 million (Tezdogan et
al., 2015)

Model scale: 10-25 million ( Kornev et

Boundary layer al., 2018) Early Technology,
DES 0-5
flow Full-scale: ~50 million (Pena et al., has shown great accuracy
2019)

Self-propulsion
Model scale 5-10 million (Tokyo 2015 Mature technology,
(only predict RANS + Workshop)
integrated thrust virtual/actuator 30-100 less than 10% deviation
disc approach Full-scale: 5-20 million (Jasak et al.,
and torque
2019)
coefficients)

Self-propulsion Model scale: 5-10 million (Kornev and

(for local flow RANS with fully Abbas, 2018) Mature Technology
analysis with discretised 30-100
insignificant propeller Full scale: 5-20 million (Sun et al, Less than 10% deviation
flow separation) 2020)

Self-propulsion Model scale 5-10 million ( Kornev and

DES with fully Abbas, 2018) Early Technology
(for local flows discretised 0-5
with significant Full-scale: ~50 million has shown great accuracy
propeller
flow separation) ( Pena et al., 2019)

Mature technology,
Cavitation
~2 million Can predict whether or not a cavitation
(only predict RANS 30-100 phenomenon should appear, but cannot
(Lu et al., 2012)
occurance) predict the changes in ship characteristics due
to the presence of cavitation

Cavitation Mature technology,

20 - 100 million
(full flow LES 0-5 while the development of DES can
(Mahesh et al., 2015)
behaviours) potentially reduce the computational cost

Component 5-20 million Mature technology,

vibration/noise/ LES (Wang et al., 2009; Jang and Mahesh, 0-5 very accurate and validated in a frequency-
damage 2013) domain level

905

906
42
907 5. Conclusions

908 This paper has analysed primary turbulence modelling strategies’ capability, limitation, computational cost,
909 and accuracy for various ship hydrodynamic applications, providing a turbulence modelling selection
910 recommendation based on the findings. The work has the potential to serve as an up-to-date guideline for
911 ship hydrodynamic simulations, as well as facilitating a wider CFD community in terms of turbulence
912 modelling.

913 In summary, RANS shows excellent capabilities in predicting the integral performance of a ship such as
914 forces and moments in both model scale and full scale. However, RANS omits certain details of the flow,
915 which typically causes inaccuracies when assessing ship performance related to detailed flow velocity and
916 vorticity. The detailed flow features can be effectively modelled using LES, but this method is only cost-
917 effective to study a small-scale problem that does not cover a large part of a ship (e.g. studying propeller
918 cavitation), as larger-scale LES simulations are still prohibitively expensive. On the other hand, the
919 WMLES approach and the hybrid DES approach are shown promising to model high-fidelity flows whilst
920 keeping the computational cost affordable, but both approaches are harder to use than RANS or LES
921 because they remain certain numerical challenges that are very sensitive to setups. With experience being
922 gained, WMLES and DES are expected to develop into mature alternatives to RANS and LES.

923

924 References

925 Arslan, T., Pettersen, B., Andersson, H.I., 2016. Large-eddy simulation of cross-flow around ship sections.
926 Journal of Marine Science and Technology 21, 552–566.
927 ASME, 2009. Verification & Validation in Computational Fluid Dynamics & Heat Transfer [WWW
928 Document]. URL https://www.asme.org/codes-standards/find-codes-standards/v-v-20-standard-
929 verification-validation-computational-fluid-dynamics-heat-transfer (accessed 2.21.20).
930 Bakica, A., Vladimir, N., Gatin, I., Jasak, H., 2020. CFD simulation of loadings on circular duct in calm
931 water and waves. Ships and Offshore Structures 1–13.
932 Benites-Munoz, D., Huang, L., Anderlini, E., Marín-Lopez, J.R., Thomas, G., 2020. Hydrodynamic
933 Modelling of An Oscillating Wave Surge Converter Including Power Take-Off. Journal of Marine
934 Science and Engineering 8, 771.
935 Bensow, R.E., 2011. Simulation of the unsteady cavitation on the Delft Twist11 foil using RANS, DES and
936 LES, in: Second International Symposium on Marine Propulsors, Hamburg, Germany.
937 Bensow, R.E., Bark, G., 2010. Implicit LES predictions of the cavitating flow on a propeller. Journal of
938 Fluids Engineering.
939 Bhushan, S., Carrica, P., Yang, J., Stern, F., 2011. Scalability and validation study for large scale surface
940 combatant computations using CFDShip-Iowa. The International Journal of High Performance
941 Computing Applications 25, 466–487.

43
942 Carrica, P.M., Fu, H., Stern, F., 2011. Computations of self-propulsion free to sink and trim and of motions
943 in head waves of the KRISO Container Ship (KCS) model. Applied Ocean Research 33, 309–320.
944 https://doi.org/10.1016/j.apor.2011.07.003
945 Cha, R., Wan, D., 2015. Numerical investigation of motion response of two model ships in regular waves.
946 Procedia engineering 116, 20–31.
947 Constantinescu, G., Chapelet, M., Squires, K., 2003. Turbulence modeling applied to flow over a sphere.
948 AIAA Journal 41, 1733–1742. https://doi.org/10.2514/2.7291
949 Dashtimanesh, A., Tavakoli, S., Kohansal, A., Khosravani, R., Ghassemzadeh, A., 2020. Numerical study
950 on a heeled one-stepped boat moving forward in planing regime. Applied Ocean Research 96,
951 102057.
952 Deng, G.B., Queutey, P., Visonneau, M., 2005. Three-dimensional flow computation with Reynolds stress
953 and algebraic stress models, in: Engineering Turbulence Modelling and Experiments 6. Elsevier,
954 pp. 389–398.
955 Dittakavi, N., Chunekar, A., Frankel, S., 2010. Large eddy simulation of turbulent-cavitation interactions
956 in a venturi nozzle. Journal of Fluids Engineering.
957 ATKINS and VIRTUE. Best Practice Guidelines for the application of Computational Fluid Dynamics in
958 Marine Hydrodynamics. Tech. rep., VIRTUE-The Virtuel Tank Utility in Europe.
959 Ferziger, J.H., Perić, M., Street, R.L., 2002. Computational methods for fluid dynamics. Springer.
960 Gnanaskandan, A., Mahesh, K., 2015. A numerical method to simulate turbulent cavitating flows.
961 International Journal of Multiphase Flow.
962 Guilmineau, E., Deng, G.B., Leroyer, A., Queutey, P., Visonneau, M., Wackers, J., 2018. Assessment of
963 hybrid RANS-LES formulations for flow simulation around the Ahmed body. Computers & Fluids
964 176, 302–319.
965 Hanjalić, K., Launder, B., 2009. Modelling turbulence in engineering and the environment: Second-moment
966 routes to closure. https://doi.org/10.1017/CBO9781139013314
967 Hino, T., Stern, F., Larsson, L., Visonneau, M., Hirata, N., Kim, J., 2020. Numerical Ship Hydrodynamics:
968 An Assessment of the Tokyo 2015 Workshop. Springer Nature.
969 Huang, L., Li, M., Romu, T., Dolatshah, A., Thomas, G., 2021a. Simulation of a ship operating in an open-
970 water ice channel. Ships and Offshore Structures 16, 353–362.
971 https://doi.org/10.1080/17445302.2020.1729595
972 Huang, L., Li, Z., Ryan, C., Ringsberg, J.W., Pena, B., Li, M., Ding, L., Thomas, G., 2021b. Ship resistance
973 when operating in floating ice floes: Derivation, validation, and application of an empirical
974 equation. Marine Structures 79, 103057.
975 Huang, L., Ren, K., Li, M., Tuković, Ž., Cardiff, P., Thomas, G., 2019. Fluid-structure interaction of a large
976 ice sheet in waves. Ocean Engineering 182, 102–111.
977 Huang, L., Tavakoli, S., Li, M., Dolatshah, A., Pena, B., Ding, B., Dashtimanesh, A., 2021c. CFD analyses
978 on the water entry process of a freefall lifeboat. Ocean Engineering 232, 109115.
979 Huang, L., Thomas, G., 2019. Simulation of Wave Interaction With a Circular Ice Floe. Journal of Offshore
980 Mechanics and Arctic Engineering 141, 041302.
981 Huang, L., Tuhkuri, J., Igrec, B., Li, M., Stagonas, D., Toffoli, A., Cardiff, P., Thomas, G., 2020. Ship
982 resistance when operating in floating ice floes: A combined CFD&DEM approach. Marine
983 Structures 74, 102817.
984 ITTC, 2017. Benchmark Database for CFD Validation for Resistance and Propulsion. Recommended
985 Procedures and Guidelines.
44
 986 ITTC, 2014a. Guidelines: Practical Guidelines for Ship CFD Applications. ITTC Report.
987 ITTC, 2014b. 27th Propulsion Committee Proceedings. ITTC Report.
988 ITTC, 2008. 1978 ITTC Performance Prediction Method. Recommended Procedures and Guidelines.
989 ITTC, 1957. ITTC Proceedings of the 8th International Conference, in: 8th ITTC Conference.
990 Jang, H., Mahesh, K., 2013. Large eddy simulation of flow around a reverse rotating propeller. Journal of
991 Fluid Mechanics.
992 Jasak, H., Jemcov, A., Tukovic, Z., 2007. OpenFOAM: A C++ library for complex physics simulations, in:
993 International Workshop on Coupled Methods in Numerical Dynamics. IUC Dubrovnik, Croatia,
994 pp. 1–20.
995 Jasak, H., Vukčević, V., Gatin, I., Lalović, I., 2019. CFD validation and grid sensitivity studies of full scale
996 ship self propulsion. International Journal of Naval Architecture and Ocean Engineering 11, 33–
997 43.
998 Khojasteh, D., Tavakoli, S., Dashtimanesh, A., Dolatshah, A., Huang, L., Glamore, W., Sadat-Noori, M.,
999 Iglesias, G., 2020. Numerical analysis of shipping water impacting a step structure. Ocean
1000 Engineering 209, 107517.
1001 Kolmogorov, A.N., 1941. The Local Structure of Turbulence in Incompressible Viscous Fluid for Very
1002 Large Reynolds Numbers. Proceedings of the Royal Society A: Mathematical, Physical and
1003 Engineering Sciences 434, 9–13. https://doi.org/10.1098/rspa.1991.0075
1004 Kornev, N., Abbas, N., 2018. Vorticity structures and turbulence in the wake of full block ships. Journal of
1005 Marine Science and Technology 23, 567–579.
1006 Kornev, N., Shevchuk, I., Abbas, N., Anschau, P., Samarbakhsh, S., 2019. Potential and limitations of scale
1007 resolved simulations for ship hydrodynamics applications. Ship Technology Research 66, 83–96.
1008 https://doi.org/10.1080/09377255.2019.1574965
1009 Kornev, N., Taranov, A., Shchukin, E., Kleinsorge, L., 2011. Development of hybrid URANS-LES
1010 methods for flow simulation in the ship stern area. Ocean Engineering 38, 1831–1838.
1011 https://doi.org/10.1016/j.oceaneng.2011.09.024
1012 Larsson, J., Kawai, S., Bodart, J., Bermejo-Moreno, I., 2016. Large eddy simulation with modeled wall-
1013 stress: recent progress and future directions. Mechanical Engineering Reviews 3, 15–00418.
1014 Larsson, L., Stern, F., Visonneau, M., 2013. Numerical ship hydrodynamics: an assessment of the
1015 Gothenburg 2010 workshop. Springer.
1016 Larsson, L., Stern, F., Visonneau, M., Hino, T., Hirata, N., Kim, J., 2015. Proceedings, Tokyo 2015
1017 Workshop on CFD in Ship Hydrodynamics.
1018 Leonard, A., 1975. Energy cascade in large-eddy simulations of turbulent fluid flows, in: Advances in
1019 Geophysics. Elsevier, pp. 237–248.
1020 Liefvendahl, M., 2010. Investigation of propeller wake instability using LES. Ship Technology Research
1021 57, 100–106.
1022 Liefvendahl, M., Alin, N., Chapuis, M., Fureby, C., Svennberg, U., Troëng, C., 2010. Ship and propulsor
1023 hydrodynamics, in: V European Conference on Comput. Fluid Dynamics.
1024 Liefvendahl, M., Fureby, C., 2017. Grid requirements for LES of ship hydrodynamics in model and full
1025 scale. Ocean Engineering 143, 259–268. https://doi.org/10.1016/j.oceaneng.2017.07.055
1026 Liefvendahl, M., Johansson, M., 2021. Wall-modeled LES for ship hydrodynamics in model scale. Journal
1027 of Ship Research 65, 41–54.
1028 Lloyds Register, 2016. Workshop on Ship Scale Hydrodynamic Computer Simulation Proceedings.

45
1029 Lu, N.X., Bensow, R.E., Bark, G., 2014. Large eddy simulation of cavitation development on highly skewed
1030 propellers. Journal of Marine Science and Technology 19, 197–214.
1031 Lu, N.X., Svennberg, U., Bark, G., Bensow, R., 2012. Numerical simulations of the cavitating flow on a
1032 marine propeller, in: 8th International Symposium on Cavitation.
1033 Mahesh, K., Kumar, P., Gnanaskandan, A., Nitzkorski, Z., 2015. LES applied to ship research. Journal of
1034 Ship Research 59, 238–245.
1035 McComb, W.D., 1990. The physics of fluid turbulence. Chemical physics.
1036 Menter, F.R., Kuntz, M., 2004. Adaptation of Eddy-Viscosity Turbulence Models to Unsteady Separated
1037 Flow Behind Vehicles 339–352. https://doi.org/10.1007/978-3-540-44419-0_30
1038 Mikkelsen, H., Walther, J.H., 2020. Effect of roughness in full-scale validation of a CFD model of self-
1039 propelled ships. Applied Ocean Research 99, 102162.
1040 Mucha, P., Deng, G., Gourlay, T., Moctar, O., 2016. Validation studies on numerical prediction of ship
1041 squat and resistance in shallow water, in: 4th MASHCON-International Conference on Ship
1042 Manoeuvring in Shallow and Confined Water with Special Focus on Ship Bottom Interaction. pp.
1043 122–133.
1044 Nikitin, N. V., Nicoud, F., Wasistho, B., Squires, K.D., Spalart, P.R., 2000. An approach to wall modeling
1045 in large-eddy simulations. Physics of Fluids 12, 1629–1632. https://doi.org/10.1063/1.870414
1046 Nishikawa, T., Yamade, Y., Sakuma, M., Kato, C., 2012. Application of fully-resolved large eddy
1047 simulation to KVLCC2. Journal of the Japan Society of Naval Architects and Ocean Engineers 16,
1048 1–9.
1049 Paterson, E.G., Wilson, R. V., Stern, F., 2003. General-purpose parallel unsteady rans ship hydrodynamics
1050 code: Cfdshipiowa., IIHR Report 432. Iowa Institute for Hydraulic Research. The University of
1051 Iowa, Iowa, USA.
1052 Pena, B., Huang, L., 2021. Wave-GAN: A deep learning approach for the prediction of nonlinear regular
1053 wave loads and run-up on a fixed cylinder. Coastal Engineering 103902.
1054 Pena, B., Muk-Pavic, E., Fitzsimmons, P., 2020a. Detailed analysis of the flow within the boundary layer
1055 and wake of a full-scale ship. Ocean Engineering 218, 108022.
1056 Pena, B., Muk-Pavic, E., Ponkratov, D., 2019. Achieving a high accuracy numerical simulations of the flow
1057 around a full scale ship, in: International Conference on Offshore Mechanics and Arctic
1058 Engineering. American Society of Mechanical Engineers, p. V07AT06A058.
1059 Pena, B., Muk-Pavic, E., Thomas, G., Fitzsimmons, P., 2020b. An approach for the accurate investigation
1060 of full-scale ship boundary layers and wakes. Ocean Engineering 214, 107854.
1061 Pena, B., Huang, L., Ahlgren, F., 2020c. A Review on Applications of Machine Learning in Shipping
1062 Sustainability, in: SNAME Maritime Convention. The Society of Naval Architects and Marine
1063 Engineers.
1064 Ponkratov, D., 2017. Proceedings: 2016 Workshop on Ship Scale Hydrodynamic Computer Simulations.
1065 Lloyd’s Register, Southampton, United Kingdom.
1066 Ponkratov, D., Zegos, C., 2015. Validation of ship scale CFD self-propulsion simulation by the direct
1067 comparison with sea trials results, in: Proceedings of the Fourth International Symposium on
1068 Marine Propulsors.
1069 Pope, S.B., 2001. Turbulent flows. IOP Publishing.
1070 Shen, L., Zhang, C., Yue, D.K.P., 2002. Free-surface turbulent wake behind towed ship models:
1071 Experimental measurements, stability analyses and direct numerical simulations. Journal of Fluid
1072 Mechanics 469, 89–120. https://doi.org/10.1017/S0022112002001684
46
1073 Shen, Z., Wan, D., 2013. RANS computations of added resistance and motions of a ship in head waves.
1074 International Journal of Offshore and Polar Engineering 23, 264–271.
1075 Shur, M.L., Spalart, P.R., Strelets, M.K., Travin, A.K., 2008. A hybrid RANS-LES approach with delayed-
1076 DES and wall-modelled LES capabilities. International Journal of Heat and Fluid Flow 29, 1638–
1077 1649. https://doi.org/10.1016/j.ijheatfluidflow.2008.07.001
1078 Smagorinsky, J., 1963. General circulation experiments with the primitive equations: I. The basic
1079 experiment. Monthly weather review 91, 99–164.
1080 Smith, T.A., Ventikos, Y., 2021. Wing-tip vortex dynamics at moderate Reynolds numbers. Physics of
1081 Fluids 33, 035111.
1082 Song, S., Demirel, Y.K., Atlar, M., 2019. An investigation into the effect of biofouling on the ship
1083 hydrodynamic characteristics using CFD. Ocean Engineering 175, 122–137.
1084 Song, S., Demirel, Y.K., Muscat-Fenech, C.D.M., Sant, T., Villa, D., Tezdogan, T., Incecik, A., 2021.
1085 Investigating the effect of heterogeneous hull roughness on ship resistance using CFD. Journal of
1086 Marine Science and Engineering 9, 202.
1087 Sotiropoulos, F., Patel, V.C., 1995. Application of Reynolds-stress transport models to stern and wake
1088 flows. Journal of Ship Research 39, 263–283.
1089 Spalart, P.R., 2009. Detached-Eddy Simulation. Annual Review of Fluid Mechanics 41, 181–202.
1090 https://doi.org/10.1146/annurev.fluid.010908.165130
1091 Spalart, P.R., Deck, S., Shur, M.L., Squires, K.D., Strelets, M.K., Travin, A., 2006. A new version of
1092 detached-eddy simulation, resistant to ambiguous grid densities. Theoretical and Computational
1093 Fluid Dynamics 20, 181–195. https://doi.org/10.1007/s00162-006-0015-0
1094 Spalart, P.R., Jou, W.H., Strelets, M.K., Allmaras, S.R., 1997. Comments on the feasibility of LES for
1095 wings and on a hybrid RANS/LES approach. Advances in DNS/LES 1, 4–8.
1096 Sun, W., Hu, Q., Hu, S., Su, J., Xu, J., Wei, J., Huang, G., 2020. Numerical Analysis of Full-Scale Ship
1097 Self-Propulsion Performance with Direct Comparison to Statistical Sea Trail Results. Journal of
1098 Marine Science and Engineering 8, 24.
1099 Terziev, M., Tezdogan, T., Incecik, A., 2019. Application of eddy-viscosity turbulence models to problems
1100 in ship hydrodynamics. Ships and Offshore Structures 1–24.
1101 Tezdogan, T., Demirel, Y.K., Kellett, P., Khorasanchi, M., Incecik, A., Turan, O., 2015. Full-scale unsteady
1102 RANS CFD simulations of ship behaviour and performance in head seas due to slow steaming.
1103 Ocean Engineering 97, 186–206. https://doi.org/10.1016/j.oceaneng.2015.01.011
1104 Tezdogan, T., Incecik, A., Turan, O., 2016. Full-scale unsteady RANS simulations of vertical ship motions
1105 in shallow water. Ocean Engineering 123, 131–145.
1106 Usta, O., Aktas, B., Maasch, M., Turan, O., Atlar, M., Korkut, E., 2017. A study on the numerical prediction
1107 of cavitation erosion for propellers. Fifth International Symposium on Marine Propulsion smp’17.
1108 Verma, A., Jang, H., Mahesh, K., 2012. The effect of an upstream hull on a propeller in reverse rotation.
1109 Journal of Fluid Mechanics.
1110 Viitanen, V.M., Hynninen, A., Sipilä, T., Siikonen, T., 2018. DDES of wetted and cavitating marine
1111 propeller for CHA underwater noise assessment. Journal of Marine Science and Engineering 6.
1112 https://doi.org/10.3390/jmse6020056
1113 Wackers, J., Koren, B., Raven, H.C., Van der Ploeg, A., Starke, A.R., Deng, G.B., Queutey, P., Visonneau,
1114 M., Hino, T., Ohashi, K., 2011. Free-surface viscous flow solution methods for ship
1115 hydrodynamics. Archives of Computational Methods in Engineering 18, 1–41.

47
1116 Wang, M., Moin, P., 2000. Computation of trailing-edge flow and noise using large-eddy simulation. AIAA
1117 Journal.
1118 Wang, M., Moreau, S., Iaccarino, G., Roger, M., 2009. Les prediction of wall-pressure fluctuations and
1119 noise of a low-speed airfoil. International Journal of Aeroacoustics.
1120 Watanabe, T., Kawamura, T., Takekoshi, Y., Maeda, M., Rhee, S.H., 2003. Simulation of steady and
1121 unsteady cavitation on a marine propeller using a RANS CFD code, in: Proceedings of The Fifth
1122 International Symposium on Cavitation (Cav. Citeseer.
1123 Winden, B., Turnock, S.R., Hudson, D.A., 2014. Self-propulsion modelling of the KCS container ship using
1124 an open source framework.
1125 Xing, T., Bhushan, S., Stern, F., 2012. Unsteady vortical flow and turbulent structures for a tanker hull form
1126 at large drift angles. Ocean Eng 55, 23–43.
1127 Yang, J., Bhushan, S., Suh, J., Wang, Z., Koo, B., Sakamoto, N., Xing, T., Stern, F., 2008. Large-eddy
1128 simulation of ship flows with wall-layer models on Cartesian grids, in: Proceedings of 27th
1129 Symposium on Naval Hydrodynamics. Seoul, Korea.
1130 Zha, R, Ye, H., Shen, Z., Wan, D., 2014. Numerical study of viscous wave-making resistance of ship
1131 navigation in still water. Journal of Marine Science and Application 13, 158–166.
1132 Zha, R, Ye, H., Shen, Z., Wan, D., 2014. Numerical computations of resistance of high speed catamaran in
1133 calm water. Journal of Hydrodynamics 26, 930–938.
1134 Zhang, Z., Liu, H., Zhu, S., Zhao, F., 2006. Application of CFD in ship engineering design practice and
1135 ship hydrodynamics. Journal of Hydrodynamics 18, 308–315.
1136 Zou, L., Larsson, L., Orych, M., 2010. Verification and validation of CFD predictions for a manoeuvring
1137 tanker. Journal of Hydrodynamics, Ser. B 22, 438–445. https://doi.org/10.1016/S1001-
1138 6058(09)60233-X
1139

48