Professional Documents
Culture Documents
5-03
-02-03
Procedures and Guidelines Page 1 of 20
Guideline
7.5-03 CFD
Table of Contents
This guidelines describe recommendations For the applications covered by the present
and common practices for ship CFD applica- guidelines, the flow is mostly turbulent and the
tions in a general way. Separate specific guide- Reynolds numbers (Re) are of the order of 106
lines cover the CFD process for other applica- in model scale and 109 in full scale,
tions in detail. ρVL
Re = (0)
µ
In this guideline the CFD process is divided
into three steps: pre-processing, computation, where V is the ship speed, L the length be-
and post-processing. The pre-processing step in- tween the perpendiculars, ρ and µ are the
volves the definition of the geometry, the com- density and viscosity of the fluid, respectively.
putational domain, and the computational grid.
The choice of the governing equations to be The current guidelines are intended for ap-
solved, the most suitable numerical techniques plications with high Reynolds numbers, typical
to be used to solve them and the real computa- of ship hydrodynamics.
tional work are performed in the computational
For free surface flow simulations, where the
step. The post-processing step is the visualisa-
free surface between air and water is solved, the
tion, the analysis, and the verification and vali-
Froude numbers (Fr) are typically of the order
dation of the results.
of 0.1 and rarely above 1,
All these steps may be significantly different
V
for different naval hydrodynamic problems, Fr = (0)
gL
such as hull resistance, propulsion, manoeuvra-
bility, sea keeping, sail aerodynamics, etc. How- where g is the acceleration of gravity.
ever, the present document aims at providing
good practice guidelines, which can be applied Both air and water are typically assumed to
to most ship hydrodynamic applications. ITTC be incompressible, allowing uncoupling mo-
specific guidelines are available for ship re- mentum and continuity equations from the en-
sistance (7.5-03-02-04), self-propulsion (7.5- ergy equation. Therefore, velocity and pressure
fields can be computed independently from the
ITTC – Recommended 7.5-03
-02-03
Procedures and Guidelines Page 4 of 20
temperature field. This assumption remains Indicative tolerances for the geometry are
valid for Mach numbers (Ma) smaller than about ±10−5 L , where L is the ship length for the hull
0.3, for Re ~ 106 . This tolerance may need to be re-
duced when smaller scale appendages, such as
V
Ma = (0) rudder or propellers, are to be included in the
a
flow computations or when the flow computa-
tion requires smaller grid cells than the indica-
where a is the speed of sound.
tive CAD tolerance. Conversely, as this toler-
Finally, these guidelines are written assum- ance is based on the distance to the wall of the
ing that the solver is grid-based (either struc- first grid point (see §2.3.6) if wall functions are
tured or unstructured) and based on the Reyn- used the geometry tolerance can be increased
olds Averaged Navier-Stokes equations appropriately.
(RANSE), as found in most commercial and ac-
Due care and attention are required to re-
ademic CFD packages. Use of mesh-free meth-
solve geometry features such as trailing edges
ods is not covered in this guidelines. For a com-
that may be less than an order of magnitude
plete overview on the computational techniques,
larger than the geometry tolerance. Geometry
the reader is referred for example to Hirsch
features that are smaller than the geometric tol-
(1989 and 1990) and Ferziger & Peric (2002).
erance should be removed from the geometry
before it is exported. It is recommended that any
2. PRE-PROCESSING additional surfaces, curves or any other geomet-
rical entities that are produced within the pre-
2.1 Geometry processor software, be generated with the same
tolerance of the original geometry.
It is common practice to define the geometry
in a CAD (Computer Aided Design) software, The origin and orientation of the reference
which allows for a high control of the surfaces. coordinate system should be chosen with care
The geometry can be exported from the CAD because the equations of motion will be solved
software and imported into a CFD pre-processor with respect to this coordinate system and so
software (grid generator) with a range of for- will be the computation of forces, moments, etc.
mats. One format compatible with most com- Therefore, if the aim of the simulation is the es-
mercial software is IGES. The accuracy of the timation of a force (for example the resistance),
geometry should be checked to ensure that the it is recommended to have one coordinate axis
imported surface definitions are reasonably aligned with that force, this to minimize the pos-
smooth and connect within a given tolerance. sibility of mistakes and to have the opportunity
ITTC – Recommended 7.5-03
-02-03
Procedures and Guidelines Page 5 of 20
to monitor directly the quantity of interest, if an experimental data is known, the boundary con-
iterative procedure (in time, for example) is in- dition should be placed sufficiently far from any
volved. On the other hand, if body motion is in- solid wall in order to assume an undisturbed far
volved (as in the case of manoeuvring or sea field velocity. When significant lift force is gen-
keeping computations) the best choice of the erated with respect to the far field velocity di-
frame of reference could be driven by the re- rection, computing the correct deflection of the
quirement for the rigid body motion solver. upstream streamlines is critical to the correct
Moreover, in order to minimise chances of er- computation of the lift force. If L is the length of
rors, it is recommended that the same coordinate the lifting surface, the inlet face should be posi-
system is used in both the CAD software and the tioned at least 10L upstream of the lifting sur-
pre-processor software. face. Conversely, when the lift is negligible and
the body is streamlined, the inlet face can be po-
2.2 Computational Domain and Boundary sitioned as close as one length upstream the
Conditions body.
In ship hydrodynamics, most of the prob- Different boundary conditions can be used
lems involve external flows, i.e. the fluid is not on the outflow face, though the most robust op-
contained into physical boundaries. Therefore, tion is a Neumann condition (zero gradient) for
the computational domain, which represents the velocity and pressure. The distance between the
volume of fluid to be modelled, must be closed outflow face and the body mainly depends on
by additional surfaces that do not represent any the generation of lift and, for free surface com-
physical boundary. In the more general case, the putations, on the reflection of gravity waves.
tested object is completely surrounded by fluid. When large lift force is generated, the outlet face
In this case a prismatic computational domain should be at least 20L downstream the lifting
can be considered. However, different geome- surface, while for other cases the outlet face
tries could better suit different grid topologies. should be at least L downstream the geometry.
Therefore, the shape of the computational do-
The other non-physical boundaries of the do-
main should be dictated by the grid strategy in
main should also be placed at least L away from
place (see §2.3.1).
the geometry. For towing tank walls symmetry,
The computational domain should include imposed velocity or slip conditions can be used,
inflow and an outflow surfaces. A Dirichlet con- with symmetry the most commonly used.
dition, i.e. a known velocity field, should be im-
posed on the inflow boundary. When no specific
ITTC – Recommended 7.5-03
-02-03
Procedures and Guidelines Page 6 of 20
2.3.3 Hybrid Grids and Other Techniques occur at the interfaces or overlapping regions if
the interpolation scheme is unable to resolve
Hybrid grids can use both structured and un- correctly the change in grid. This can be allevi-
structured grids in different parts of the domain. ated by ensuring that the local change in the grid
Non-conformal grids are those where the points across the non-conformal region is minimized
or cells on a surface do not match. Non-confor- with similar grid resolution and spacing used for
mal grids are often used at the boundaries be- both grid regions. Non-conformal grids can also
tween blocks discretized with different grid den- use different cell types to assist the grid genera-
sities, or on moving (sliding) grids. These types tion process, for example a grid for a detailed
of grids are often required for wake flow analy- rudder with skeg and end plates can be produced
sis where it is important to capture small flow using a local prism/tetrahedral grid which is em-
features. bedded inside a multiblock grid for the hull.
The flow solution algorithm must use special Overset grids are formed by blocks that
coding in order to interpolate between the non- overlap in space. These grids are particularly
connected grids. For some methods this interpo- convenient when the different blocks experience
lation scheme may be defined by interfaces, a relative motion and therefore the overlapping
where the interpolation method is defined across changes with time. For instance, overset grids
grid boundary surfaces, for other methods the can be used for ship propulsion where one grid
interpolation scheme is defined by overlaps, is used for the hull and another is used for the
where the interpolation is defined across local propeller.
grid volumes. For both of these types of interpo-
lation schemes the formal order of accuracy is 2.3.4 Cell Types
likely to be reduced, especially when there are
large differences between the grid resolutions Quadrilateral (2D, 4-sided) and hexahedral
and topologies of donor and receptor grids. (3D, 6-sided) cells are supported by almost all
However, this type of approach can be used to CFD codes. Topological attributes of these cells
considerably simplify the grid generation pro- – presence of opposing faces, and relative loca-
cess so that locally better quality grids can be tions of cell centres and face centres – have been
produced around the various geometry compo- found to be beneficial to spatial accuracy of nu-
nents and assembled together to form a com- merical solutions. In structured grid-based solv-
plete grid. ers, the presence of stencils (e.g., i, j, k coordi-
nates in the computational domain) readily ac-
Non-conformal grids should be used with commodates high-order discretization schemes
care because undesirable wave reflections can (e.g., fifth-order convection scheme) that can
ITTC – Recommended 7.5-03
-02-03
Procedures and Guidelines Page 8 of 20
enhance spatial accuracy. They are also efficient Which cell types to use for a given problem
in terms of usage of cells, since they can be clus- really depends on many factors such as the
tered and/or stretched as needed to economically solver (what cell types can the solver support),
increase the resolution in particular regions objective of the computation (do fine details of
(such as boundary layers, or region where large the flow need to be resolved?) and hardware re-
velocity gradients occur). The main shortcom- sources (are computational resources available
ing of these structured grid cells is that it is often to run cases involving large grids?). For rela-
very hard to generate high-quality grids for tively simple geometries, consider using a high-
complex geometries. quality hexahedral grid. However, if body-mo-
tion (e.g. free sinkage and trim) is involved, con-
Unstructured grids give more flexibility in sider using overset grids if your CFD solver in-
the choice of cell types, simplifying grid gener- cludes this capability.
ation for complex geometries. The majority of
unstructured CFD solvers allow use of arbitrary For complex geometries (e.g. a fully-ap-
polyhedral cells such as quadrilaterals and trian- pended ship) for which a high-quality structured
gles in 2D; and hexahedra, tetrahedra, wedges, grid is difficult to generate, consider using an
pyramids and prisms in 3D to name a few, and unstructured grid (preferably a hybrid unstruc-
combination of all them (hybrid grids). In a typ- tured grid) or overlapping grids.
ical unstructured grid, the boundary layer is dis-
Avoid using tetrahedral cells for boundary
cretized using a prism grid grown out of a trian-
layers, near the free surface, and in the regions
gular grid on the wall, and a tetrahedral grid is
where high resolution is required. The use of
used elsewhere away from wall. Compared to
hexahedral or prismatic cells would result in
typical structured grids, gridding time can be
higher accuracy and better convergence rates.
dramatically reduced with unstructured grids.
Spatial accuracy for unstructured grid cells such
2.3.5 Grid Resolution
as triangles, tetrahedra, and pyramids can be
lower than that for quadrilateral and hexahedral The number of grid points, and therefore the
cells. Unstructured grids usually require a larger resolution of the grid, should be decided based
number of computational points than structured on the chosen turbulence model and on the tem-
grids in order to achieve a comparable accuracy. poral and spatial scales of the flow features of
Furthermore, spatial accuracy of the majority of interest. In the following some practical guide-
unstructured grid finite volume solvers is lim- lines are provided but interested readers can find
ited to second order. in Pope (2000) and Sagaut (2006) more rigorous
discussion based on the physics of the resolved
ITTC – Recommended 7.5-03
-02-03
Procedures and Guidelines Page 9 of 20
time and statistical averages of the flow field, structure can enable more efficient use of com-
and the modelled turbulent fluctuations. puting resources but at the expense of increased
grid generation time and complexity.
Whenever possible, use hyperbolic grid gen-
erators to guarantee as much as possible an or- In free surface flow simulations, when the
thogonal grid near the wall. Grids orthogonal to free surface is modelled, always use orthogonal
the domain boundaries, where the boundary grids to resolve it if possible. Use no less than
conditions are imposed, are recommended. For 40 grid points to resolve the shortest wave
some boundary conditions, such as symmetry length when a second-order spatial accuracy is
and wall conditions, this orthogonality condition used, though this condition may be extremely
may be mandatory for some solvers. expensive and unnecessary for low Froude num-
ber flows. The same resolution can be achieved
Provide refinement where flow features of with different spatial accuracies varying linearly
interest are expected, in accordance with the size the spatial resolution. For instance, for a fourth-
of the feature to be simulated. Where the flow order spatial resolution, using 20 grid points is
features of interest are not known before-hand, equivalent to using 40 grid points with a second-
it may be necessary to use an iterative process to order spatial resolution.
establish the existence of the key flow features
for a given geometry. This requires obtaining 2.3.6 Near-wall Region
and examining an initial flow solution, and sub-
sequent grid refinement in the regions of interest. The number of grid points within the bound-
This process should be carried out until some ary layer is determined by the level of accuracy
measure of grid refinement is satisfied to ensure required, the turbulence model chosen, and
that flow features are sufficiently resolved. whether wall functions are used.
If overset grids are used, check that overlap Guidelines are provided in terms of the non-
is sufficient for the number of fringes that are dimensional wall distance of the first point from
needed in the code and the order of accuracy. If
the wall, y + . We define
motions are computed, this requirement holds
for all time steps.
Friction velocity:
Ensure sufficient resolution where high cur-
vature geometry is present, especially around τw
u* = (0)
leading and trailing edges. An appropriate grid ρ
ITTC – Recommended 7.5-03
-02-03
Procedures and Guidelines Page 10 of 20
A near-wall turbulence model resolving the to comply with wall function requirements or to
laminar sub-layer needs at least 3 points inside resolve the boundary layer), the distance y of
the first point from the wall can be computed as
it, which using y + = 2 results in an expansion
follows:
ratio of 1.5, the largest acceptable, but a maxi-
y+ L
mum value closer to 1.2 is recommended. In y = (0)
C
ReL f
most cases a y + ≤ 1 will be used with expansion 2
0.455
Cf = (0) 2.3.7 Grid Quality
ln ( ReL )
2.58
It should be noted that Eq. (12), though As mentioned in the Introduction, these
widely used and accepted, was developed for guidelines assume that the incompressible Na-
typical hull forms in existence before 1957 and vier-Stokes equations for Newtonian fluids are
thus it may not produce good results for modern solved. The Reynolds number range is such that
and unconventional hull forms. laminar to turbulent transition occurs and turbu-
lent flow must be modelled. In the present guide
ITTC – Recommended 7.5-03
-02-03
Procedures and Guidelines Page 12 of 20
it is assumed that the turbulence is modelled Detached Eddy Simulation (DES) is a hybrid
with a RANS approach, where the instantaneous method that reduces the required computational
velocity is split into the sum of its statistical av- effort by solving the (unsteady) RANS equa-
erage and a turbulent fluctuation which is mod- tions in the near-wall region and applying LES
elled by the turbulence model. It has to be high- in the rest of the domain.
lighted that, when dealing with unsteady flow
fields, it is assumed that a clear separation be- 3.1.2 RANS Models
tween the time scales of the mean motion and
Turbulence modelling has been an important
time scales of turbulent motion exists. The fun-
research topic over the past decades. A large
damental assumption in unsteady RANS com-
number of models have been proposed, tested
putations is that the averaging time is bigger
and applied, but no universal model has been de-
than turbulent time scales (which are averaged)
veloped. Thus one is forced to choose the best
but much smaller than mean flow time scales
model available for each specific application.
(which are solved).
The majority of turbulence models are based on
It is worth noting more computationally de- the so-called Boussinesq hypothesis, which de-
manding approaches are becoming more and fines a turbulent or eddy viscosity (as opposed
more common in recent years. Direct Numerical to the molecular viscosity) to account for the ef-
Simulation (DNS) solves the Navier-Stokes fect the turbulence motion has on the mean flow.
equations to the resolution of the smallest turbu- For a detailed description of turbulence models,
lent scales but needs grids and time resolution see for example Wilcox (2006).
which are not yet achievable for high Reynolds
Zero-equation, or algebraic models express
numbers, such as those typical of ship hydrody-
the eddy viscosity in terms of the mean flow var-
namic applications.
iables and mean flow gradients without solving
Large Eddy Simulation (LES) does not aver- any additional equations. They are hardly ever
age the Navier-Stokes equations in time, but fil- used in ship hydrodynamic applications.
ters them in space. This results in transient com-
One-equation models solve one additional
putations on extremely dense grids as they aim
equation (i.e. in addition to the momentum and
to resolve the larger turbulence scales and mod-
mass conservation equations) for the eddy vis-
elling the scales smaller than the grid resolution.
cosity. Regularly used in ship hydrodynamics
LES is used currently for highly demanding
are models by Menter and by Spalart-Allmaras.
flows where the transient nature of the turbu-
These models are sometimes extended with a
lence needs to be resolved to smaller scales, like
in the case of acoustic noise.
ITTC – Recommended 7.5-03
-02-03
Procedures and Guidelines Page 13 of 20
typically at zero pressure gradients and it is are used, then different functions for Δu + are
well-known that these analytical expressions be-
used depending on the range of ks and the types
come less valid, or even invalid, with increasing
adverse pressure gradients. Though corrections of roughness, such as sand grains or spheres. In-
exist for wall functions in the presence of pres- terested readers are invited to consult Schlicht-
sure gradients, these make the wall functions ing and Gersten (2000). No general guidelines
less stable. Thus it is a trade-off between accu- can yet be given for wall-resolved boundary lay-
racy and computational effort. Wall functions ers. However, the roughness profile is typically
should be avoided if possible, and used with care not resolved, the blockage effect of the rough-
when necessary. ness is taken into account using a modified wall
+
= y + + k s / 2 , and therefore the wall-
unit, ymod
3.3 Surface Roughness
The flow in the boundary layer is not af- adjacent cell should be higher than ks .
fected by the surface roughness when the phys-
the thickness of the laminar sublayer. In partic- In the majority of industrial CFD codes, dif-
ular, the effect of the surface roughness is negli- fusion terms in the governing equations are dis-
cretized using a second-order (central differenc-
gible when ks is lower than approximately
ing) scheme by default. Thus, spatial accuracy is
2ν / u * , in this case the surface is said hydrody- largely determined by discretization scheme
namically smooth. For hydrodynamically rough used for convection terms.
surfaces, the logarithmic Law of the Wall is
The first-order upwind (FOU) scheme, of-
modified adding a term, Δu + , which is a func-
fered in many commercial CFD codes often as a
tion of ks . Unfortunately this function depends default scheme, is very stable. However, it intro-
duces large numerical (artificial) diffusion - that
on the type of roughness and no universal for-
is why it is so stable. Its use can only be justified
mulation is known.
when the physical diffusion (turbulent + molec-
In most ship CFD applications, surfaces are ular) is much higher than the introduced numer-
assumed to be hydraulically smooth. Otherwise ical diffusion. This is true within the boundary
the effect of the roughness can be modelled ei- layer, allowing use of hybrid schemes that use
ther modifying the wall functions or the turbu- lower order near the wall and higher order far-
lence boundary conditions. When wall functions ther out.
ITTC – Recommended 7.5-03
-02-03
Procedures and Guidelines Page 15 of 20
The majority of high-order convection dis- numerical simulation (DNS) for its low-dissipa-
cretization schemes in popular use today for- tion that is critical to accurately resolve turbu-
mally have a second-order of accuracy with an lent structures. However, CD scheme is inher-
upwind bias. All these second-order upwind ently unstable, raising convergence issues for
(SOU) schemes differ from one another in terms cases involving fine grids and small effective
of the flux limiter used to suppress unphysical viscosity (large local Reynolds number). One
oscillations in the solutions. Still higher-order should consider using a stabilized form of cen-
schemes such as fifth-order scheme exist; how- tral differencing.
ever, not all CFD solvers offer such higher-order
schemes. Moreover, it has to be taken into ac- 3.5 Free Surface
count that higher order schemes are, in general,
There are two major categories in free sur-
less robust than first or second order accurate
face models. For interface fitting approaches the
schemes. Therefore, their use should be consid-
numerical grid is conformed to the free surface
ered with special care. These schemes can be
shape. Interface capturing approaches define the
used in regions of high-quality grid (near or-
free surface as an iso-surface of a marker func-
thogonal, low aspect ratio, low expansion ratio),
tion and thus the grid does not have to conform
like in Cartesian overset refinements, while
to the free surface.
lower-order schemes can be used in others re-
gions. Higher-order schemes are standard in im- In general, the interface fitting approach is
mersed boundary methods in Cartesian grids. more accurate and efficient than the capturing
approach, since free surface boundary condi-
The SOU scheme is both reasonably accu-
tions can be applied in the exact free surface lo-
rate and robust, and for that reason is an indus-
cation. Therefore the interface fitting model
trial workhorse for convection discretization in
may be selected whenever possible. However, it
body-fitted approaches. The SOU scheme is
should be noted that a deformation/re-gridding
therefore recommended for all convection-dif-
procedure is essential in order to keep the grid-
fusion transport equations.
lines following the deformation of the free sur-
First order accurate schemes in time may be face. This may cause severe distortion of grid-
used only if a steady-state is sought. lines even though the initial grid conforming to
the undisturbed free surface has a good quality.
Although outside the scope of the present
guidelines, it is worth mentioning that second- Difficulties in grid generation and/or re-
order central differencing (CD) scheme is often gridding can be avoided with interface capturing
used in large eddy simulation (LES) and direct approaches. Also, for large deformation of free
ITTC – Recommended 7.5-03
-02-03
Procedures and Guidelines Page 16 of 20
surface, such as overturning or breaking waves, particle of fluid within a time interval equal to
interface capturing methods should be used the time step:
since surface-tracking methods cannot resolve
u Δt
the change in surface topology. Since the cap- < Cmax (0)
turing methods demand finer grid resolution Δx
near the interface, grid generation requires more where |𝑢𝑢�| is the local velocity, Δt is the time
attention. The choice of level-set method or vol- step, and Δx is the linear cell size in the flow
ume-of-fluid method in for capturing approach
has little impact in the final solutions. Although direction. In explicit solvers, where Cmax < 1 ,
the numerical methods are different, well imple- the CFL condition is more demanding than in
mented flow solvers provide similar results for
both models. implicit solvers, where Cmax can be larger than
cases fewer points per core can be used if the due to the convergence should always be esti-
CFD code scales well to obtain faster turna- mated with a verification procedure.
round, or more points per core if processor
Convergence should be achieved for both
memory allows it and turnaround time is not
steady-state and transient computations. In the
critical.
case of implicit transient computations, conver-
To maximize performance, try to distribute gence must be evaluated at each time step.
the load evenly between nodes. For instance, if
The level of convergence should be assessed
running in a Linux cluster with 2 quad-core pro-
by the history of residual variations for the mass
cessors (8 cores) per node, and your case needs
and momentum equations. Residuals indicate
10 cores, you can distribute your load in two
how far the present approximate solution is
nodes using 8-2 or 5-5 configurations. The sec-
away from perfect conservation of mass and
ond balances the load per node better since it
momentum. Thus the residual for a discretized
provides more memory bandwidth per process.
equation is defined as the L1-Norm of the imbal-
Modern workstations with shared memory ance between the left and the right hand side of
are available with up to 64 processors, though that equation over all the computational points.
much larger specialized systems are produced. Usually the residual is scaled by a reference
High-performance clusters are typically cheaper value. Sometimes the L2-Norm or L∞-Norm are
per processor for large systems (thousands of used to define the residual.
cores) but use distributed memory. Shared-
CFD users do not need to worry about the
memory systems allow all processors access all
definition of residuals, as they are often pre-de-
memory, resulting in easier programming and
fined by code developers. Instead, attention
better scalability of most applications. On the
should be paid to the selection of convergence
other hand, distributed memory systems provide
criteria. The recommended criterion is the drop
massive number of processors for very large
of scaled residuals by at least three orders of
computations.
magnitude off their initial values. However it is
acknowledged that due to complexity of some
3.8 Convergence
problems or oscillatory convergence this crite-
A number of convergence criteria should be rion may not be satisfied. In such cases, the con-
defined and examined in order to ensure reliable vergence of forces and moments, and the con-
convergence of solution. While different levels vergence of the unknowns in key regions of the
of convergence are acceptable, the uncertainty flow field (e.g. at propeller plane for nominal
wake calculations) can be monitored.
ITTC – Recommended 7.5-03
-02-03
Procedures and Guidelines Page 18 of 20
The verification and validation of numerical The estimate of the uncertainty must be
simulations is an essential tool for the interpre- based on a verification and validation proce-
tation of the results and the identifications of dures but it can be assumed that the uncertainty
those aspects of the simulation that can be im- does not change significantly with small
proved. changes of the boundary and initial conditions.
Therefore, once the uncertainty has been esti-
The numerical uncertainty is due to the nu- mated for one simulation, this can often be used
merical error introduced by the method used to as a reasonable estimate of the uncertainty of
solve the chosen set of equations with the given similar simulations. For example, if a parameter
boundary and initial conditions, while the mod- study on the hull geometry is performed, verifi-
elling uncertainty is associated to the modelling cation can be performed for only one or few val-
error, which is the difference between the exact ues of such parameter. Care should be taken on
solution of this set of equations and the reality. these extrapolations that no significant changes
The numerical uncertainty can be estimated per- on the flow field are observed between similar
forming a set of numerical simulations, while cases.
ITTC – Recommended 7.5-03
-02-03
Procedures and Guidelines Page 19 of 20
It must be noted that the difference between Pope, S. (2000) Turbulent Flows, Cambridge
the uncertainties of different quantities com- University Press.
puted with the same simulation can be higher
Sagaut, P. (2006) Large Eddy Simulation for In-
than one order of magnitude. For instance, the
compressible Flows, an Introduction,
uncertainty of the ship resistance is expected to
Springer.
be significantly lower than the uncertainty of the
local flow speed in one point of the domain. Schlichting, H., and Gersten, K. (2000) Bound-
Also, the higher the order of the derivatives of ary-layer theory, Springer.
the primitive unknowns, the higher the uncer-
tainty expected. For instance, the uncertainty on Wilcox, D.C. (2006) Turbulence Modeling for
the velocity is expected to be lower than the un- CFD (Third Edition), DCW Industries.
certainty on the vorticity. Therefore, the uncer-
Larsson, L., Stern F., Vissonneau, M. (2013)
tainty must be estimated for every quantitative
Numerical Ship Hydrodynamics: An assess-
result gathered from the simulation.
ment of the Gothenburg 2010 Workshop,
Springer.
5. REFERENCES
ITTC Recommended Procedures and Guide- ITTC Recommended Procedures and Guide-
lines. Uncertainty analysis in CFD, Guide- lines. Practical Guidelines for Ship Re-
lines for RANS codes, No. 7.5-03-01-02 sistance CFD, No. 7.5-03-02-04 (2014).
(1999).
ITTC Recommended Procedures and Guide-
ITTC Recommended Procedures and Guide- lines. Practical Guidelines for Ship Self-Pro-
lines. CFD User’s Guide, No. 7.5-03-01-03 pulsion CFD No. 7.5-03-03-01 (2014).
(1999).
ITTC Recommended Procedures and Guide-
ITTC Recommended Procedures and Guide- lines. Practical Guidelines for RANS Calcu-
lines. CFD Verification, No. 7.5-03-01-04 lation of Nominal Wakes, No. 7.5-03-03-02
(1999). (2014).