Sei sulla pagina 1di 26

A numerical method for the simulation of steady and

unsteady cavitating ows


Yiannis Ventikos
a,
*, George Tzabiras
b
a
Environmental Hydraulics and Water Resources Group, School of Civil and Environmental Engineering, Georgia
Institute of Technology, Atlanta, GA, 30332-0355, USA
b
Department of Naval Architecture and Marine Engineering, National Technical University of Athens, Athens, P.O.
Box 64070, GR-15710, Greece
Received 25 September 1997; received in revised form 07 May 1998; accepted 24 July 1998
Abstract
A numerical method for the prediction of cavitating ows around lifting bodies is presented. The
algorithm employs a one-uid NavierStokes-enthalpy solver that can handle variable uid properties,
along with properly formulated watervapor mixture state laws, in order to account for the two-phase
ow of water and vapor and the transition from one phase to the other. The method allows for the
simulation of steady and unsteady, two-dimensional, cavitating ows for low and moderate Reynolds
numbers. Employment of the method for various test cases shows good agreement with available
experimental measurements and observations and demonstrates interesting phenomena related to
cavitation and its interaction with viscous mechanisms. Simulation of Reynolds number dependent
phenomena, like cavitation inception and total section drag force, are only predicted qualitatively in the
present study. However, the method shows promise for reliable quantitive predictions. # 1999 Elsevier
Science Ltd. All rights reserved.
1. Introduction
Cavitation is a phenomenon that plays a major role in surface sea-going vessel design and
operation, as well as in hydraulic equipment. Propellers, hydrofoil ships, hydraulic turbines and
pumps may suer from its consequences in many ways. All these devices encompass lifting
surfaces and a liquid working medium, water in our case. When cavitation occurs
Computers & Fluids 29 (2000) 6388
0045-7930/00/$ - see front matter # 1999 Elsevier Science Ltd. All rights reserved.
PII: S0045- 7930( 98) 00061- 9
www.elsevier.com/locate/compfluid
* Corresponding author. Tel.: +1-404-894-2721; fax: +1-404-894-2677.
E-mail address: yvent@ce.gatech.edu (Y. Ventikos)
unexpectedly, severe decrease in performance is observed and in many cases, the machine itself
suers from damages. Systems that operate for long time intervals without maintenance, like a
marine propeller, may be completely destroyed by such damages.
The systematic study of cavitation is relatively new in engineering. The literature provided by
Knapp et al. [1] and Young [2] are considered reference texts nowadays. They treat many
aspects of cavitation, from hydrodynamic cavitation (which is of interest to us) to
sonoluminescence and can be very useful for an overview of the phenomenon.
Several researchers have investigated cavitation numerically. We could categorize such
numerical eorts under two general thrusts: the rst employs boundary element-type
methodologies while for the second, treatment of the whole eld, i.e. full spatial discretization,
is attempted. Along the rst direction, Doctors [3] dealt with the linearized potential ow over
a two-dimensional supercavitating hydrofoil. The solver he developed was based on the
distribution of Kelvin sources and vortices along the mean line of the prole. The cavity
surface was considered a free surface and the foil was assumed to be supercavitating, in order
to avoid the, inherent, problem of nding a suitable shape for the closure of the cavity inherent
in such methodologies, Uhlman [4] employed a similar approach; however, he did not linearize
the problem. The distribution of the boundary elements was of a more complex nature here
and, upon convergence, satisfaction of the boundary conditions both on the foil and on the
surface cavity was achieved. A heuristic model for the cavity closure was adopted. Lee et al. [5]
also have treated the problem in an analogous manner. They estimated the shape of the cavity
along with the strengths of the distributed sources and dipoles iteratively, solving for the
satisfaction of the exact (free surface-like) cavity boundary conditions. Kinnas and Fine [6] and
Kinnas et al. [7,8] have treated the partially cavitating two-dimensional and three-dimensional
problems using nonlinear theory and a low-order potential-based boundary element method.
The cavity shape was determined from two independent boundary value problems: for the rst,
the cavity length is specied and the cavitation number, s =
p
o
p
v
(
ru
o)
a2
, (where u
o
and p
o
are the
undisturbed velocity and pressure at innity, r is the density of the liquid and p
v
is the critical
vapor pressure for the temperature of the liquid) is unknown while for the second the opposite
holds. The same team has introduced either empirical or boundary layer relations for the
estimation of the inuence of viscosity and has developed design and optimization tools based
on the aforementioned methodologies. All the works referenced so far have been based on the
boundary element method and on potential ow theory. They have the advantage of being
computationally ecient and, provided that the heuristic cavity closure model is ecient and
the viscous eects are either negligible or readily accountable with simplied techniques, they
give rather accurate results.
The second broad line of research involves spatial discretization of the full computational
domain and not just the boundary. Delanoy [9] solved again for the inviscid case using a
SIMPLE-like pressure correction solver along with a simple constitutive relation linking the
state of the uid (in this case the density at every point in the computational domain) with the
pressure. Kubota et al. [10] went one step further and included viscosity in their method. They
employed a NavierStokes solver in what they named a Bubble Two-phase Flow model. The
uid was considered a compressible continuum with widely varying density. A time marching
nite dierence numerical scheme was used for the numerical solution. The growth, variation
of size and collapse of the bubbles was given by a modied Rayleigh equation solution for
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 64
every point in the ow eld and every time step. Chen and Heister [11] employed a Marker-
and-Cell NavierStokes solver and used a combination of simple constant and linear
relationships (choice based on location in the ow eld) to link density with pressure. All the
research eorts referenced so far can be characterized as void fraction methods, the term
indicating that the vapor phase is considered completely passive, not contributing in the
momentum, continuity and energy equations. In other words, no consideration is taken for the
ow characteristics within vapor cavities and bubbles. Ventikos and Tzabiras [12] worked
along the same broad direction, only they introduced a set of pragmatic state laws for the
behavior of the vapor phase. Although their work is of the same general strategy with the
latter three, there is a signicant dierence: the vapor phase is not treated as passive, thus the
method is not a void fraction method. Instead, the actual watervapor state laws are
incorporated in this model and the vapor phase contributions are fully accounted for in the
mass and momentum balance. Moreover, physical properties of the one-uid mixture
(viscosity, thermal conductivity and, of course, density) are computed without neglecting the
contribution of the gaseous phase. It is an extension of this last research eort that we are
presenting in this paper. The latter four works correspond, as mentioned before, to eld
solutions instead of boundary solutions that are common in the study of cavitation. They
suer from being rather expensive on the computer but, especially the viscous approaches, are
very promising because they can simulate cavityviscosity interaction eects and may lead to
advanced design tools.
In this work, we are presenting our attempt to develop a numerical method that is based on
a viscous solver and accounts in a realistic manner for the water to vapor transition. This work
also contains a selection of results from the simulations performed concerning various, steady
and unsteady, cavitating ow cases. The focus here is mainly to examine the behavior and
validity of the developed numerical method. However, some interesting observations are
presented concerning the unsteady case where cavityvortex interactions take place.
2. Governing equations and solution procedure
We shall describe the partial dierential equations that govern the ow conguration under
consideration and the numerical method developed for their solution. We shall deal with the
motion of innite uid around a still hydrofoil, with steady boundary conditions.
2.1. The transport equations
As we have seen in Section 1, numerous approaches for the solution of cavitating ows have
appeared in the literature. We have decided to treat the phenomenon of cavitation by
considering the uid existing in two phases, the liquid phase (water) and the gaseous phase
(vapor). These two phases are supposed to co-exist in the ow eld and to transform from one
to the other, depending on the local conditions at every instance in time. This approach, as
indicated in Section 1, is becoming more and more popular recently because of its potential to
include all the physics of the cavitating ow in the solution. However, it carries a possible
drawback, that it is a continuum approach with the ambition to simulate a phenomenon
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 65
exclusively discontinuous like cavitation, where the phase change boundary is so abrupt that it
practically forms a free surface. However, as we shall see in the sequel, the results obtained
from the employment of this method are quite satisfactory.
Following the aforementioned considerations, two dierent approaches concerning the
expression of the dierential equations have appeared so far in the literature: the
interpenetrating continua (or two uids) method and the variable properties (or one uid, or
mixture) method. As the name implies, the rst approach [13,14] assumes that both phases co-
exist at every point in the ow eld in varying proportions and each phase is governed by its
own set of dierential equations. This assumption introduces quantities that account for the
trans-phase exchanges of mass, momentum, energy etc. To clarify this approach, we can write
the transport equations for one direction with x
i,j
coordinates and corresponding u
i,j
velocity
components in the following form:
. u
1
momentum equation (water phase):
df
w
r
w
u
1w
dt

_
d
dx
1
_
f
w
r
w
u
2
1w
_

d
dx
2
_
f
w
r
w
u
1w
u
2w
_
_
= f
w
d
p
dx
1

_
d
dx
1
(f
w
s
11
)
d
dx
2
(f
w
s
12
)
_
C
f(
u
1v
u
1w)
m
vw
u
1w
(1)
. u
1
momentum equation (vapor phase):
df
v
r
v
u
1v
dt

_
d
dx
1
_
f
v
r
v
u
2
1v
_

d
dx
2
_
f
v
r
v
u
1v
u
2v
_
_
= f
v
d
p
dx
1

_
d
dx
1
(f
v
s
11
)
d
dx
2
(f
v
s
12
)
_
C
f(
u
1w
u
1v)
m
vw
u
1w
(2)
In the above equations, r
w,v
are the densities of water and vapor respectively, f
w,v
are the
volume fractions (f
w
f
v
= 1), p is the shared pressure of the two phases, s
ij
are the viscous
stresses and C
f
is the inter-phase slip coecient. This viscous friction between the two phases is
a quantity not only generally unknown but also not easily measurable in experiments. This
fact, along with the total number of partial dierential equations to be solved, which is almost
twice that of the one-uid approach, to be described in the sequel, present some of the inherent
disadvantages of this technique. The system closes by proper equations of state. We should
note that, to our best knowledge, the interpenetrating continua method has not been used so
far in the study of cavitation.
On the contrary, the variable properties method treats the working medium as a mixture of
two uids behaving as one with variable properties from point-to-point, depending on the
relative ratio of the two phases. Thus, one set of dierential equations expresses the uid
motion and state laws. The system closes again by proper constitutive relations, only in this
case the exact thermodynamic properties of the mixture have to be dened instead of its
fractions. This is the main drawback of this approach: we have to know the physical properties
(density, viscosity, etc.) for a mixture of two uids, for every possible ratio. However, this
weakness does not apply when our working medium is water, because the watervapor mixture
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 66
properties (or equivalently, the relations that provide them) are well-known and widely used in
physics and engineering.
We have decided to follow the second approach in the present study (i.e., the variable
properties method) and we shall illustrate many of its aspects by analyzing it further. Three
main reasons led us to this choice:
. The physics of the second approach are more straightforward.
. There are reliable sources for the watervapor mixture physical properties like the UK
Steam Tables volume [15].
. The computer cost is substantially lower in comparison to the interpenetrating continua
approach, because there are half the dierential equations that need to be discretized and
solved.
The dierential equations which govern the ow eld according to the variable properties
method, expressed in Cartesian tensor notation and in conservative form, can be written as
[16]:
. Momentum
d
dt
(
ru
i )

d
dx
j
(
ru
j
u
i )
=
dp
dx
i

d
dx
j
_
m
_
e
ij

1
3
e
ii
d
ij
__
(3)
. Continuity
dr
dt

d
dx
i
(
ru
i )
= 0 (4)
. Stagnation enthalpy of mixture
d
dt
(rh)
d
dx
j
_
ru
j
h
_
=
d
dx
j
_
K
dT
dx
j
_

d
dx
j
u
i
m
_
e
ij

1
3
e
ii
d
ij
_
(5)
The rate of strain vector is dened as
e
ij
=
du
i
dx
j

du
j
dx
i
(6)
and the enthalpy
h = h
ref

_
T
T
ref
c
p
dT
1
2
u
i
u
j
(7)
When we compare these equations with that of the interpenetrating continua approach (Eqs.
(1) and (2)), it is easy to identify the much simpler form of the terms, as well as the lack of
inter-phase coecients, a fact that reduces the number of necessary experimental constants for
our system signicantly. In the above equations, t is the time, x
i,j
are the independent
coordinates of the coordinate system, u
i,j
correspond to the velocity vectors in the two
directions of the coordinate system, r is the density, p is the pressure, m is the molecular
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 67
viscosity, d
ij
is the Kronecker delta, T is the absolute temperature, K is the thermal
conductivity and c
p
is the specic heat of the uid at constant pressure. The temporal
derivatives are dropped in all steady cases. The dilation term
d
dx
j
(
1
3
e
ii
d
ij
) (which corresponds
to the viscous forces exerted to the uid while undergoing change of volume due to hydrostatic
pressure) has been retained in the formulation of the discrete equations. The properties of the
uid at various locations are computed via the employment of the watervapor tables, in the
form of
r =
1
V
sp
(h,p)
, m = m(h,p)
T = T(h,p), K = K(h,p), c
p
= c
p
(h,p) (8)
where V
sp
stands for the specic volume of the uid.
It should be noted here that since the ratio of the density of the uid and gas phase is
approximately O(10
3
), a big cavity can be lled with a very small mass of vaporized liquid.
Also, this sudden phase change corresponds to a very small change of temperature of the
liquid, the temperature that is directly linked with the latent heat of this small mass of
vaporized liquid. These two facts are of importance to our analysis, because they constitute the
driving terms of the continuity and energy equations: the essence of the dierence of this
approach and the void fraction methods. The small liquid mass decit that the appearance of
the cavity creates may be negligible compared to the total mass of the system, but it is very
important to the size of the cavity that this mass decit lls when turned to vapor. The size of
this cavity subsequently inuences and modies the characteristics of the liquid-ow eld and
this strong coupling, originating from the small vaporized liquid mass, renders the accurate
and exact treatment of the vapor phase important. Moreover, vapor ows within the cavities
and the characteristics of this ow inuence the shape of the boundary in a similar sense as
described above.
2.2. Initial and boundary conditions
Initial conditions and boundary conditions on all boundaries are required for the solution of
the described system of equations. Various kinds of initial conditions were tested, but we
concluded in using uniform conditions of undisturbed velocity, pressure, enthalpy and density
throughout the ow eld.
The nature of the equations dictates the application of proper boundary conditions on all
boundaries. We apply the no-slip, no ux boundary condition for the velocity on the surface of
the airfoil and constant velocity on boundary I, Fig. 1. Neumann conditions are applied for
the enthalpy on all boundaries and an adiabatic condition is applied for this quantity on the
airfoil surface.
On boundary II (Fig. 1) the outow boundary, a Neumann condition for the velocities
suces for the steady case. However, for the unsteady case, when large scale organized vortical
structures exit the computational domain through this boundary, velocity derivative conditions
cause blockage eects that alter the solution in an unallowable manner (or, even lead the
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 68
solution to divergence). One solution to this problem is to use extremely long computational
domains, which may extend 100150 times the length of the solid body downstream. Such a
strategy requires a lot of computer memory and CPU time to account for the extra grid points.
To avoid this, more advanced exit boundary conditions were used, i.e. wave type boundary
conditions. In brief, following the analysis presented in [17,18], we consider that momentum
behaves like a transported wave quantity, thus yielding an anisotropic propagation wave
equation on boundary II (Fig. 1). In our 2D case, this gives
d
2
u
dt
2
c
2
x
d
2
u
dx
2
c
2
y
d
2
u
dy
2
= 0 (9)
where c
x
and c
y
are the characteristic velocities of the wave propagation in the two directions.
After some manipulation [17], the outow boundary condition may be written as
d
dt
(
ru
1)

d
dx
1
(
ru
1
u
1)
=
d
dx
2
m
_
du
1
dx
2
_
(10)
This equation resembles strongly the NavierStokes equation for the direction parallel to the
ow, vanishing the pressure gradient and the streamwise viscous stresses and when employed
as an outow boundary condition gave signicantly better results. The use of this boundary
condition has permitted the utilization of reasonably long computational domains, only 1020
chord lengths downstream, resulting in great computer time economy. Moreover, the behavior
of the vortical structures leaving the computational domain was quite satisfactory, without any
deformation or dissipation of the eddies and no pressure feedback eects.
2.3. The discretized equations
For the solution of the equations of the ow eld, a C-type orthogonal curvilinear
coordinate system is adopted (Fig. 1). This type of grid geometry is particularly benecial for
pressure-driven and stagnation-oriented ow phenomena due to its good leading edge
Fig. 1. Computational domain and notation.
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 69
resolution features [19]. In Fig. 2, the coordinate system along with the denition of the
velocity vectors is presented. The velocity vectors used in the equations of motion always refer
to the local coordinate system, i.e. the velocity components are aligned with the grid lines. For
instance, the u
1
momentum equation would read [20]:
d
dt
(
ru
1)

1
h
1
h
2
_
d
dx
1
_
rh
2
u
2
1
_

d
dx
2
(rh
1
u
1
u
2
)
_
=
1
h
1
dp
dx
1

1
h
1
h
2
_
d
dx
1
(h
2
s
11
)
d
dx
2
(h
2
s
12
)
_
K
12
s
12
2K
21
s
22
(11)
In this equation, K
ij
correspond to the curvatures of the orthogonal curvilinear system and h
ij
are the metrics.
Following the control volume approach [21, 22], the transport equations are integrated over
nite volumes dened by the grid cells. The volume integral for each cell and each term of
every equation is then transformed to integrals on the four grid lines dening the cell by
employment of the divergence theorem. After some algebra and assuming linear distribution of
the variables within each control volume, this formulation leads to sets of algebraic equations
of the form:
j
P
j
n1
P
Dt
DV
P
A
P
j
P
= A
N
j
N
A
S
j
S
A
E
j
E
A
W
j
W
S
j
(12)
which obviously can be written as
Fig. 2. Coordinate system and control volumes.
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 70
_
DV
P
Dt
A
P
_
j
P
= A
N
j
N
A
S
j
S
A
E
j
E
A
W
j
W

_
S
j

DV
P
Dt
F
n1
P
_
(13)
Similarly, we can employ second order of accuracy dierencing in time, leading to an
expression such as,
_
3DV
P
2Dt
A
P
_
j
P
= A
N
j
N
A
S
j
S
A
E
j
E
A
W
j
W

_
S
j

DV
P
2Dt
_
4F
n1
P
F
n2
P
_
_
(14)
Dt is the time step employed, DV
P
is the volume of the corresponding cell, F
n1
is the solution
of the previous time step and F
n2
is the solution of the one before the previous time step. As
can be deducted from the above formulations, the numerical scheme is fully implicit. This
means that for each stage of computation, only current time-step estimates of the variables are
used and the previous stage (or stages in the case of higher order of temporal accuracy)
solutions appear only in the source terms. This requires solution, for every variable, of a N
N system of equations (where N = n
x
n
y
is the total number of cells in the computational
domain) and full convergence for each time step, before proceeding to the next. In these
relations we have dropped the superscript n dening the present time-step, for convenience. N,
S, E, W denote the north, south, east and west neighbors of the central point P, Fig. 2.
A
N,S,E,W
include the combined eect of convective and diusive terms and A
P
= (SA
i
) S
P
where S
P
is the part of the source term that is proportional to F
P
(i = N,S,E,W). For the
steady case, the temporal term is dropped and the method of under-relaxation is adopted to
stabilize convergence [20]. As one can see from Fig. 2, a staggered grid arrangement is
implemented. The velocities correspond to transposed control volumes, whereas pressure
correction and enthalpy are dened in the center of the control volumes.
The numerical method was tested with central dierencing and with a hybrid scheme. This
hybrid scheme [23] consists of a mechanism selecting between central dierencing or rst-order
upwind, depending on the value of the local cell Pechlet number (Pc rulam) representing the
ratio of the strengths of convection and diusion at every cell. For the grid resolution nally
employed (see Section 3.1), it was found that central dierencing, accompanied by small under-
relaxation factors (for the steady case) or small time-steps (for the unsteady case) converged
and produced wiggle-free solutions (for the Reynolds number 2000 computed) and was used
throughout this study.
2.4. Pressure correction
The source terms S
F
of the algebraic systems of the momentum equations carry pressure
gradients that are not known a-priori. A variant of the PISO approach [24] is adopted, to
calculate the pressure eld simultaneously with the transport equations, at every time step.
According to this, the satisfaction of continuity leads to a pressure correction equation, derived
from the general approximation:
A
P
du
P
=

A
i
du
i
D
u
Dp
/
(15)
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 71
where du is the velocity change and Dp
/
is the corresponding pressure gradient change. If we
combine this expression with the integrated continuity equation, we get
A
P
p
/
P
=

A
i
p
/
i
DF (16)
where p
/
P
is the pressure correction for node P, and D is the velocity divergence in the
corresponding cell. A two-step process is followed in the standard PISO approach: rst, F is
considered zero (which reduces the algorithm to standard SIMPLE form [25]) and then F is
considered a linear combination of all the summation terms of the right-hand side of Eq. (15)
and the divergence is taken to be zero.
In this study, following [26], we accept that for ne grids, the pressure derivatives do not
vary signicantly over neighboring cells. Then, we can rewrite the rst approximation as
du
e
I
_
Du
e

i
f
i
_
A
/
i
Du
i
_
e
_
_
p
/
p
p
/
E
_
= Du
(1)
e
_
p
/
P
p
/
E
_
(17)
say for one of the velocity corrections, on the east-side face of the pressure control volume
(Fig. 2). In this relation, f
i
are geometrical weighting factors and the terms under summation of
the original relation have been truncated in the calculation of A
/
i
Du
i
. The coecient Du
e
includes the eect of the neighboring grid cells. We could extend this by summing all the grid
points' contributions by a repetitive procedure, but the inuence of cells other than the
neighboring ones decays rapidly, yielding a third additive step impractical. This approach is
benecial for two reasons: rst, only one solution of the algebraic system is required,
comparing to the standard PISO approach where two successive solutions are necessary and
second, a signicant enhancement in the convergence rate was observed.
2.5. Density correction and constitutive relations
In order to consider the strong eects of the two-phase nature of the ow, the algorithm
was modied to incorporate the dependence of density on the pressure correction. This is
important in our case, because cavitation is an extremely abrupt phenomenon where
minimal variation of pressure changes the density of the uid mixture from values of
O(10
3
) (water) to values of almost O(10
1
) (vapor). For this reason, following [27], a
supplementary correction r
/
= kp
/
is introduced in the pressure correction equation, where
r
/
is the resulting density dierence from a p
/
pressure correction and k is a constant.
Moreover, upwinding was applied for density in the pressure correction equation, by
modifying accordingly the procedure described in [27]. The value of the constant k should
(at least in principle) be close to the slope of the pressure-specic volume curve of the
uid; however, because when convergence is achieved p
/
vanishes (it represents pressure
correction and not the pressure itself, therefore as the system approaches converged states
this quantity diminishes gradually), the importance of this correction also vanishes. Thus, k
is picked in a suitable manner to increase the convergence speed.
Several mechanisms for the choice of k were tested. We have tried uniform k throughout the
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 72
ow eld with values ranging from 0.1 to 1000. This approach resulted in poor convergence
and, moreover, we could not come to a nal decision on an optimum value of k since the
convergence behavior changed drastically from the rst to the last iterations and this did not
happen in a systematic manner, considering the various values of k. The next approach we
tried was to give k non-zero values only in vapor-phase regions. This failed completely, leading
to fast divergence. We think that this is because the watervapor interface boundary is
extremely steep (as it should be and as we shall see in the sequel) and from iteration to
iteration the cells around that inter-phase boundary change from water-lled cells to vapor-
lled cells very rapidly. Therefore, it is not benecial for the numerical solution to have a
number of cells being computed with non-zero k for iteration | and with zero k for iteration
{ 1 (and vice versa).
The approach that did work was slightly more complicated. We have assumed so far
that the variation of r
/
with p
/
is practically linear. This does not actually hold, especially
in the sharp phase change regime (which is very abrupt, as mentioned before, resembling
strongly a free surface) since an extremely small variation in pressure changes the density
by more than three orders of magnitude. This was incorporated in the choice of k by
selecting k = 1 for the water phase and k = 1 CS
F
[ S
F
[ for the vapor phase, where C is
a normalizing positive coecient picked empirically for every cavitation number tested and
S
F
the mass residual for each cell. The linking of k with the mass residual of every cell is
the key here. Inter-phase boundary cells may keep one face in water and the other in
vapor (although this is an extreme case, the phase change boundary is relatively coarser,
i.e. usually 34 cells thick). Thus we have the density as a multiplier in the momentum and
mass transfer terms, varying by four orders of magnitude. This fact can prevent
convergence from being achieved and by strengthening k locally there, we have improved
the solution process signicantly. Note that works of similar nature [9] mention that this
great discrepancy of density in adjacent cells causes severe convergence problems and can
even leave remaining residuals in the iterative procedure thus preventing full convergence of
the solution. This problem has been tackled in this work by the technique described above.
Moreover, when we wanted to compute very small cavitation number cases, we had to
gradually reduce the reference pressure, thus computing (however, not to full convergence)
a few intermediate steps and using each one as the initial solution for the next step in
order to avoid divergence. This technique led to the achievement of convergence for very
dicult cases (cavitation number-wise), see Section 3.
At this point we are ready to discuss the way the physical properties of the watervapor
mixture are estimated from the computed values of the enthalpy and the pressure. In the
Appendix of the UK Steam Tables [15], one can nd the so-called `derived functions', derived
directly by partial dierentiation of the canonical or characteristic function g = g(p,T) (in case
p and T are chosen as independent variables), where g is the specic free enthalpy (Gibbs
function). Of course, similar results can be obtained for any other pair of thermodynamic
variables. All the possible states of the uid are divided into six categories or sub-regions, and
formulas are given for various physical quantities within each sub-region. The regimes of
interest to us are the sub-cooled region (where we have only water) and the mixture area, along
with the saturation line that separates the two. The quantities of interest to us are specic
volume (which when inverted gives the density), viscosity, thermal conductivity and specic
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 73
heat at constant pressure. Since the relations that provide these data are rather standard, are
occasionally very long and moreover they carry a huge amount of constants (double precision
constants one should note), it would be redundant to present them in detail here. Two things
must be kept in mind though:
. For almost all quantities in question, dierent relations are provided for the two sub-regions
of interest. Therefore, we must have two relations for each physical quantity and proper
controls that route the algorithm to the appropriate formula for every computational cell
and time step.
. All the quantities are not given as direct expressions of the pressure and the enthalpy, but
usually as q
i,j
= f(p
i,j
,T
i,j
), where q stands for any of the quantities mentioned before, p for
the pressure, T for the temperature and i,j indicating all the grid cells in the ow domain.
Therefore, we had to employ small iterative procedures to get the results in question, using a
function of the form h
i,j
= f(p
i,j
,T
i,j
) [28] employing a combination of two of the formulas
given in the Steam Tables [15]. Both these issues led to a signicant increase in necessary
computational time. It should also be pointed out that the watervapor charts used in this
study correspond to one particular condition as far as water purity is concerned. Since
cavitation inception and the hysteresis mechanisms that are present depend strongly on the
quality of water and in particular the amount of cavitation nuclei present, it is obvious that
all results obtained correspond to that particular water quality situation. The numerical
method is capable, in principle, to accommodate any similar watersteam chart (or any
uidvapor chart for that matter) available.
A comment on our choice to solve the pressureenergy equations for the determination of
state is appropriate at this point. As mentioned before, in order to ll the volume of a
cavity, only a very small quantity of water needs to be vaporized. Thus, one can say that
it is not necessary to solve the continuity equation for the vapor phase since its
contribution to the mixture mass balance is negligible. A similar argument would yield the
energy equation useless for both phases since the variations of temperature in the water
phase are also almost negligible. Such an approach may be computationally more ecient,
but treats the vapor phase as completely passive. However, as mentioned before, this is not
the case, since the vapor phase itself ows inside the cavity and the bubbles and the small
amounts of vaporized liquid dictate partly the size of the cavities. Especially in the case of
the unsteady ow, the evolution of many of the most intriguing phenomena depends
strongly on the vapor phase ow characteristics. Throughout the presentation of the
fundamental aspects of this method, we have implied that thermodynamic and
hydrodynamic equilibrium prevail in the ow eld. These assumptions are not necessarily
true [29] and, in particular, there is strong evidence that thermodynamic equilibrium is not
the case at the rear part of the cavitation bubble where collapse of the bubbles take place.
However, we should note that although lack of equilibrium might be the case in the
subgrid scale region, the present method deals with volume and time interval averages and,
in this respect, reacts in an `averaging' manner. Therefore, since neither our time scale
resolution, nor our spatial discretization are near the range where non-equilibrium
phenomena might dominate, we believe that our equilibrium assumptions are sucient for
the requirement of accurate engineering predictions.
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 74
2.6. Computational aspects
The numerical procedure can be outlined in a owchart-like manner:
1. initialize all variables
2. solve u
1
and u
2
momentum equations, with the available pressure and density
3. solve the pressure correction equation
4. correct the pressure, velocities and density according to the modied PISO manner
5. solve the enthalpy equation
6. employ the steamvapor formulas to update the physical properties at every cell, using the
computed values of pressure and enthalpy
7. go to 2 and iterate from 2 to 7 until the dimensionless total residuals of all four algebraic
systems fall under a prescribed level. If a steady computation is performed, exit
8. proceed to the next time step and iterate from 2 to 8 until the desired time interval is
simulated.
We should mention that the resulting systems of linear equations for each equation and each
time step are solved iteratively using a GaussSeidel, alternate direction, banded matrix solver
[16].
The resulting numerical code was developed on SUN and SGI workstations and the nal
runs were executed on R8K SGI workstations and servers. Typical steady runs required 1000
3000 iterations for convergence (45 orders of magnitude reduction of the residuals and
stationarity of C
p
and C
f
distributions on the airfoil). For the unsteady runs, 37 iterations
were required within each time step for convergence and approximately 20,000100,000 time
step histories constituted one unsteady run. Each steady-state run required 210 hours of
execution time on the aforementioned hardware, whereas each unsteady simulation had to run
for approximately 2 weeks.
In the sequel, we shall present results from two kinds of simulations: steady cavitating ows
and unsteady cavitating ows. Although cavitation is an essentially unsteady phenomenon,
there are several occasions when we can consider the characteristics of the ow domain xed in
time and obtain a solution that is valid in a large scale time-averaged sense. From an
engineering point of view, this approach has the benet of allowing faster computations and
results can be readily comparable with experimental ones, for example, the pressure
distribution on the surface of the prole.
3. Results and discussion
As mentioned earlier, it was not feasible (and out of the scope of this study) to simulate
ows at technically interesting Reynolds numbers. Since direct simulation of such Reynolds
numbers is still out of the question, even for simple ow congurations and single-phase ows,
such an attempt would necessitate the introduction of a proper turbulence model.
Unfortunately, there is no widely acceptable turbulence model that can handle the uncertainties
of either unsteady single-phase ows, or cavitating two-phase ows. This being the case, we
actually had a choice of picking a Reynolds number on terms other than similarity. Not
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 75
desiring to oversimplify the test cases by an extremely small Reynolds number, or
overcomplexing them by choosing a Reynolds number at the limit of our ner grid resolution
capability, we decided to solve for Reynolds number 2000. We believe that it is a good choice,
since it is quite satisfactorily resolved by the grid and several interesting viscous phenomena
are present, as we shall see in the sequel. It should be made clear at this point that since the
method at its current state of development cannot achieve Reynolds number similarity, it
should be anticipated that certain aspects of cavitating ow predictions will yield poor or no
quantitative information. For example, it is known that cavitation inception depends on the
Reynolds number [30]. Although the unsteady solver can simulate this phenomenon of
cavitation initiation, the cavitation number where it occurs should be treated only as
qualitative evidence.
3.1. Grid dependence, time-step sensitivity and numerical accuracy
We have conducted a large number of test runs with various grid sizes, time steps, choice of
hybrid scheme or central dierencing and choice of rst or second order of accuracy for time
integration. Regarding the temporal integration, we have found that for relatively big time
steps (Dt = 0X05, which was the biggest time step we have tested), rst and second order of
accuracy for the temporal integration gave substantially dierent results. However, as we
reduced the time step, the two tended to overlap and when we reached time steps of about
Dt = 0X0004, there was no observable dierence between the two, in terms of mean values or
Strouhal numbers for the various phenomena. Therefore, we believe that keeping second order
of accuracy for the temporal integration along with a xed time step of Dt = 0X0004 places the
computations on the safe side of accuracy as far as this aspect is concerned. The time step is
always normalized with the velocity at innity and the chord length.
With respect to grid resolution, we have tested grids ranging from 100 70 nodes to 800
400 nodes. For the coarser grids, we have had problems achieving convergent solutions using
the central dierencing scheme. The solutions obtained by the hybrid scheme had a strong
tendency to underestimate viscous vortical structures, especially in the unsteady ow case.
Moreover, the cavity shape had a step-like outline, because of the coarseness of the grid cells
adjacent to the inter-phase boundary. Grid resolutions of 500 300 and ner seemed to
converge to the same solution (within 2% maximum dierence) as far as pressure coecient
and skin friction coecient distributions were concerned, and the shape of the cavity was
exactly identical. Thus, a grid of 600 400 was used throughout the rest of this study.
For the aforementioned nal choice of time step and grid resolution, test runs employing
either the hybrid scheme or pure central dierencing converged and produced wiggle-free
solutions. Since it has been observed that the hybrid scheme has a tendency to underestimate
secondary vortical structures, central dierencing was used for the rest of this study. It must be
noted that for the steady-state test case, central dierencing imposed further restrictions with
respect to the choice of under-relaxation factors. Moreover, upwinding for the density in the
coecients of the pressure correction equation [27] is, as mentioned before, an inherent part of
the algorithm.
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 76
3.2. Application of the solver to the steady cavitation problem
The rst test case we shall present is that of a NACA 0012 airfoil at 88 angle of attack and a
cavitation number, s =
p
o
p
v
(ru
o)a2
equal to 1.25 (where u
o
and p
o
are the undisturbed velocity
and pressure at innity, r is the density of the liquid and p
v
is the critical vapor pressure for
the temperature of the liquid). The iso-density contour lines for this test case appear in Fig. 3.
The rst thing one can note is that the density contour lines are very closely packed together,
indicating a sudden pass from one phase to the other, in compliance with experimental
observations. Actually, the variation from water to vapor at the front part of the cavity takes
place within three cells. Moreover, the density contour lines expand signicantly at the closing
end of the cavity. Experiments show that in this region the inter-phase boundary is not so
sharply dened, including small bubbles etc. The size of these minor formations would be
smaller than the size of the cells of our grid, so the model cannot simulate them at these
discretizations. However, the method adapts the solution in an average manner by making the
inter-phase boundary broader. According to experiments and potential ow calculations, one
might expect the cavity to be somewhat longer for this cavitation number. However, the low
Reynolds number of the simulations allows the development of a recirculation zone carrying
mild pressures that terminate the cavity prematurely. In Fig. 4, the pressure coecient
distribution is presented for the same computation, and also for the non-cavitating ow case.
One can easily observe the at area at the top front part of the airfoil corresponding to the
cavity. This area is responsible for a loss of lift of about 22% in this case, a fact indicating the
signicance of studying cavitation for lifting devices, where achievement of high lift is crucial.
The reader might note that the pressure drop in this gure (and in almost all of the subsequent
ones) is of a smaller magnitude than usually observed. This is again a Reynolds number eect,
observed also in other viscous computations of similar nature [10]. It is probably even more
pronounced in the present work because of the intensity of the viscous eects at the rear part
of the airfoil.
The second case we studied is that of a NACA 0009 prole at 2.58 angle of attack and
cavitation number 0.9. The rather low cavitation number made convergence of this test case
quite tedious: very low under-relaxation factors had to be employed to avoid divergence and
this led to a large number of iterations (almost 7500) necessary for convergence. In Fig. 5, the
comparison of the suction side C
p
distribution with experimental results [31] is given. We can
say that the region of the cavity, both length-wise and pressure distribution-wise, is predicted
Fig. 3. Iso-density contours, s = 1X25, a = 88, NACA 0012 airfoil.
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 77
Fig. 4. The C
p
distributions, s = 1X25 and non-cavitating case, a = 88, NACA 0012 airfoil.
Fig. 5. The C
p
distributions, simulation and experiment, s = 0X9, a = 2X58, Re = 2X5 10
6
, suction side, NACA 0009
airfoil.
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 78
rather well. In fact, as we move downstream along the airfoil surface, there is a growing
discrepancy between the experiment and the computations, since the experiment was carried
out at Reynolds number 2X5 10
6
while the computation at 2 10
3
. This results in milder
pressures due to the much thicker boundary layer appearing in the computations. The pressure
drop after the cavity is resolved more accurately, due to the small angle of attack that reduces
the viscous eects compared to the previous case. Apart from the inuence of viscosity, this
could also be attributed to the shape of the pressure distribution of the non-cavitating ow:
higher angles of attack correspond to higher negative pressure peaks near the leading edge
while smaller angles would result in pressures much closer to zero, right after the cavity
pressure plateau. This comparison with experimental results demonstrates well the method's
strong points as well as its limitations at the current state of its development.
We shall present now the results obtained from two tests performed: the rst concerns the
behavior of the ow eld when the angle of the airfoil changes systematically for a constant
cavitation number, and the second examines the ow eld as the cavitation number changes
systematically for a constant angle of attack. A large number of computer runs were executed
for these tests [18], but only a selection of the results is presented here.
For the rst test, we chose s = 0X95 and angles of attack of 28, 48, 68 and 88. The results
concerning the iso-density contours appear in Fig. 6. The length of the cavity grows with the
increase of the angle of attack, as expected, apart from the case of 88 angle of attack where
viscous eects result in a slight reduction of the cavity length.
A better perspective is given by the C
p
distributions appearing in Fig. 7. In general, there is
a growth of the area where C
p
is at and nearly constant in the cavity region. However, the
growth rate is signicantly reduced for the 68 and the 88 angle of attack, probably due to the
Fig. 6. Iso-density contours, s = 0X95, a = 28, a = 48, a = 68, a = 88, NACA 0012 airfoil.
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 79
development and growth of a recirculation zone as discussed earlier. An overall picture of the
phenomenon is given in Fig. 8 where the lift-angle of attack curve is presented. The co-
existence of viscous separation and cavity inuences on the pressure distribution on the prole
yields this non-monotonical behavior.
The second test examines various cavitation numbers at a xed angle of attack. We chose
a = 48 and cavitation numbers equal to 1.05, 0.95, 0.90 and 0.80. Unfortunately, we could not
lower the cavitation numbers any further in our computations, because cases below 0.80 were
almost impossible to converge to a steady state, the solution uctuated broadly for extremely
long computer times. We believe that this is because the resulting cavity lengths are very long
and either shed bubbles or fall deep within the recirculation zone. We can deduce that both
these mechanisms co-exist and interact, requiring very high discretizations and usually some
Fig. 8. Lift vs. angle of attack, NACA 0012 airfoil.
Fig. 7. The C
p
distributions, s = 0X95, a = 28, a = 48, a = 68, a = 88, suction side, NACA 0012 airfoil.
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 80
form of stabilization procedure in order to achieve convergence. Very small under-relaxation
factors were employed for the s = 0X8 case (about 0.03 for all variables but the pressure, for
which the under-relaxation was set to 0.1). The results concerning the iso-density contours
appear in Fig. 9. The length of the cavity grows with the decrease of the cavitation number, as
expected.
A clearer view is given by the C
p
distributions appearing in Fig. 10. Again, there is a growth
of the area where C
p
is at and nearly constant in the cavity region. A global picture of the
phenomenon is given in Fig. 11 where the lift-cavitation number curve is presented. The shape
of the curve follows the general trend of experimental results [32], although the overall curve is
shifted, due to the Reynolds number eect (again, three orders of magnitude dierence exist
between experiment and computation).
As an overall conclusion from this section, we can say that the method works rather well
and gives reasonable results while it suers from scaling limitations, especially when large
angles of attack are concerned. Moreover, it is rather expensive on the computer, especially for
small cavitation numbers.
3.3. Application of the solver to the unsteady cavitation problem
In this section, we shall present the results concerning a simulation of the formation, growth,
shedding and collapse of bubbles from the suction side of a NACA 0012 airfoil at 88 angle of
Fig. 9. Iso-density contours, s = 1X05, s = 0X95, s = 0X90, s = 0X80, a = 48, NACA 0012 airfoil.
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 81
Fig. 11. Lift vs. cavitation number, NACA 0012 airfoil.
Fig. 10. The C
p
distributions, s = 1X05, s = 0X95, s = 0X90, s = 0X80, a = 48, NACA 0012 airfoil.
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 82
attack, Reynolds number 2000 and cavitation numbers s = 1X3 and s = 1X07. We nd that the
evolution of the phenomenon is very well illustrated schematically in [2], an evolution similarly
depicted in Fig. 12. We can see the existence of a main cavity and the formation of a bubble at
its edge, being swept downstream by the mean ow. An indication of the formation of a re-
entrant jet is also sketched. The re-entrant jet is a stream of water that is observed
experimentally to be injected inside the cavity from its rear lower part.
Apart from the aforementioned Reynolds number limitation, another restriction applies in
the unsteady case: cavitation bubble collapse is an extremely abrupt phenomenon and the
temporal resolution of our simulation is limited by the time step used. For the study that
follows we have employed a time step Dt = 0X0004 seconds. Numerical reasons (loss of
accuracy and excessive roundo error in the temporal term
3DV
P
2Dt
A
P
, where the rst part
becomes disproportionally large and the inuence of A
P
is lost) forbade the use of even smaller
time steps. It is speculated that we could reduce the time step even further with the use of
much ner grids. However, the combination of ne spatial discretization (the grid used was the
same as in the steady case, 600 400), along with the necessity of acquisition of hundreds of
thousands of time steps for the proper illustration of the phenomena inquired, made further
renement of the grid impossible. This in turn prohibited the use of smaller time steps and as a
consequence it was impossible to capture clearly the phenomenon of cavity collapse. We have
managed to do that only in one case, where the bubbles traveled a signicant fraction of the
chord before they vanished. The gures to follow reect these eorts and are to be studied in
parallel with the scenario described in [2] and Fig. 12.
The simulation was conducted as follows: rst the cavitation number was set to a big (non-
cavitating) value and the ow was left to develop past any transients. After 30 seconds of
simulated time, the cavitation number was reduced gradually to 1.3 (by proper adjustment of
Fig. 12. Formation and collapse of bubbles.
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 83
the pressure at innity). After simulating 40 seconds of cavitation number 1.3, s was lowered
gradually to 1.07 and another 30 seconds of simulated time was spent there. Then, the
cavitation number was again gradually raised to a non-cavitating condition. This gradual
progression of the parameter was done for two reasons: rst, to ensure convergence of this
very dicult ow case, and secondly to record the cavitation inception and desinence numbers
[1,2,12]. These numbers indicate the values of s where cavitation appears and disappears,
respectively. The two values are not the same, due to a well established [1,2] hysteresis
mechanism. The values recorded were 1X5220X005 and 1X5820X005, respectively. Experiments
have shown the great inuence of water quality on the aforementioned numbers. Degassed
water is not only much more resistant to tension [1,2], but the appearance of cavitation in the
experimental rig is much more scattered and the hysteresis is bigger. We could not test this
eect consistently, since we only have one set of watervapor charts at hand (that corresponds
to natural water). We believe that for a phenomenon like cavitation, where random factors
initiate the process, the aforementioned scattering is the rule and not something exceptional.
The ow eld at times 40.3 and 40.6 seconds and cavitation number 1.3 can be seen in Fig.
13. We are at the stage where the main cavity forms and grows and, at the same time, parts of
it are swept downstream. The traveling bubbles correspond to the case mentioned before i.e.
they are yet to reach the unsteady recirculation region [33,34] which starts at about 35% of the
chord's length. The comment that must be repeated here is that the iso-density contours are
very densely packed.
The next plot in Fig. 13 depicts time instant 42 seconds. The main cavity has grown in size
and the swept bubble has moved downstream. After this time step, the bubble vanished, thus it
is obvious that we are just a few fractions of a second before collapse. We must say here that
since we actually treat both phases as compressible, the method could, in principle, sense the
Fig. 13. Iso-density contours. T = 40X3, T = 40X6, T = 42X0, T = 45X1, s = 1X3, NACA 0012 airfoil.
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 84
pressure wave that is generated by the collapse of a bubble and propagated by density
variations (much like a near-spherical shock wave) throughout the uid. However, the velocity
of this propagating wave would be equal to the speed of sound in water, a value that is too big
to be resolved by the employed temporal discretization. This brings up again the issue of the
time step, that is, very small time steps would be necessary to simulate such shock waves, time
steps that are out of our capabilities for the time being. In the last time-step depicted in Fig.
13, the main cavity is fully formed and signicantly bigger in volume than the previous time
steps.
The time sequence presented in Fig. 14 demonstrates the formation and disappearance of the
re-entrant jet. We can see how the rear part of the cavity deforms, how the water column
moves inside the vapor phase, at rst continuously and then as water droplets, how the drops
swirl there and move even deeper in the cavity, and nally how they dissolve and practically
vanish. This cycle has been left to repeat itself a few times during the simulation. In fact, each
cycle is not exactly the same as the previous ones, but is similar in the manner described in
[33,34]. One cannot but notice the complex formations that appear in these gures, especially
apparent in time step 48.5 seconds. The curling of the jet is the outcome of a complex
interaction of the jet with a localized vortex appearing in that location [18,33,34]. This vortical
structure alters the form of the re-entrant jet in the way appearing in these plots, actually
stalling a little its full penetration in the cavity, something that happens a little latter, as seen
in the plot depicting time step 48.5 seconds
In Fig. 15, the cavitation number is lowered to 1.07. For this value, the size of the cavity is
of greater extend and this leads to the situation mentioned before, where the cavity ends well
within the strong recirculation zone starting somewhere in the middle of the airfoil. We can
observe the fast alternating motion of the closing boundary of the cavity and the beginning of
Fig. 14. Iso-density contours. T = 45X7, T = 46X8, T = 47X1, T = 47X9, T = 48X5, T = 49, s = 1X3, a = 88, NACA
0012 airfoil.
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 85
the formation of the detaching bubble, however, which is not actually observed to travel
downstream.
A general observation one can also make from these contour plots is the occasional zig-zag
formation of the iso-density lines at the top front part of the cavity boundary. This can be
explained if we refer to [33,34], pointing out the existence of a KelvinHelmholtz instability at
that location. This is a shear layer type instability, positioned between the high velocity free
ow regime above the airfoil and the decelerated uid in the thick boundary layer that
develops near the surface of the body. This instability manifests its existence with irregular
shaped pressure and vorticity contours and, for higher Reynolds numbers, with distinct shear
layer vortices. Fourier analysis of pressure or velocity signals at that location verify the
existence of a high frequency oscillation [33,34] in accordance with the aforementioned
visualizations of computed ow elds.
4. Conclusions
We have developed a numerical method for the prediction of steady and unsteady cavitating
ows. The algorithm combines a robust one-uid variable-properties NavierStokes-enthalpy
equation solver along with a realistic set of state relations for the watervapor mixture. The
results obtained displayed reasonable cavity shapes with sharp phase-change boundaries.
Moreover, computations for steady-state cavitating ows compare satisfactorily with available
experiments. For the unsteady cavitation case, interesting phenomena involving the interaction
of viscous mechanisms with cavitation have been predicted in good qualitative agreement with
Fig. 15. Iso-density contours. T = 90X4, T = 91X2, T = 92X1, T = 92X8, s = 1X07, a = 88, NACA 0012 airfoil.
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 86
available experimental observations. The test cases computed demonstrate well both the
complexity and richness of the phenomena that take place, as well as the capability of the
method to predict such phenomena. The method can be readily extended to three dimensions
and providing that there is availability of computer power, it can be used eectively to draw
conclusions on complicated cavitation mechanisms, like re-entrant jets. Due to the Reynolds
number discrepancies between our laminar computations and technically interesting (turbulent)
ows, several Reynolds number dependent phenomena are only qualitatively predicted.
Employment of proper turbulence closures for the prediction of technically interesting, high
Reynolds number ows is the obvious next step for future research in this direction.
Acknowledgements
The authors wish to thank Prof. C. Frangopoulos of the Department of Naval Architecture
and Marine Engineering, National Technical University of Athens, for kindly providing a
comprehensive set of modules for the estimation of the watervapor properties, Dr. M. Braza
of the Institut de Mecanique des Fluides de Toulouse for helpful discussions and Prof. F.
Sotiropoulos of the School of Civil and Environmental Engineering, Georgia Institute of
Technology for his insightful comments.
References
[1] Knapp RJ, Daily JW, Hammit FG. Cavitation. New York: McGraw-Hill Company, 1970.
[2] Young FR. Cavitation. New York: McGraw-Hill Book Company, 1989.
[3] Doctors LJ. Eects of a nite Froude number on a supercavitating hydrofoil. J Ship Res 1986;30:111.
[4] Uhlman JS. The surface singularity method applied to partially cavitating hydrofoils. J Ship Res 1987;31:107
24.
[5] Lee C-S, Kim Y-G, Lee J-T. A potential-based panel method for the analysis of a two-dimensional super- or
partially-cavitating hydrofoil. J Ship Res 1992;36:16881.
[6] Kinnas SA, Fine NE. A numerical non-linear analysis of the ow around two- and three-dimensional partially
cavitating hydrofoils. J Fluid Mech 1993;254:15181.
[7] Kinnas SA, Mishima S, Villeneuve R. A viscous/inviscid analysis method for cavitating ows on two and three
Dimensions. In: 20th Symposium of Naval Hydrodynamics, Santa Barbara, USA. 1994.
[8] Kinnas SA, Mishima S, Savineau C. Application of optimization techniques to the design of cavitating hydro-
foils and wings. In: CAV'95 International Conference, Deauville, France. 1994.
[9] Delannoy Y. Modelisation d'ecoulements instationnaires et cavitants. Thesis. Institut National Polytechnique
de Grenoble, 1989.
[10] Kubota A, Kato H, Yamaguchi H. A new modeling of cavitating ows: a numerical study of unsteady cavita-
tion on a hydrofoil section. J Fluid Mech 1992;240:5996.
[11] Chen Y, Heister S. Two-phase modeling of cavitating ows. Computers and Fluids 1995;24(7):799809.
[12] Ventikos Y, Tzabiras G. A numerical study of the steady and unsteady cavitation phenomenon around hydro-
foils. In: CAV'95 International Conference, Deauville, France. 1995.
[13] Spalding DB. Numerical computation of multi-phase ows. V.K.I. for Fluid Dynamics. Lecture Series, 1981.
[14] Markatos NC. Modeling of two-phase transient ow and combustion of granular propellants. Int J Multiphase
Flow 1986;12:91333.
[15] U.K. Steam Tables in SI Units 1970. U.K. Committee on the Properties of Steam. London: Edward Arnold,
1970.
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 87
[16] Antonopoulos K. Prediction of ow and heat transfer in rod bundles. Thesis. Imperial College, London, 1979.
[17] Jin G, Braza M. A nonreecting outlet boundary condition for incompressible unsteady NavierStokes calcu-
lations. J Comput Phys 1993;107:23953.
[18] Ventikos Y. Numerical investigation of unsteady, cavitating and non-cavitating ows around hydrofoils.
Doctorate Thesis. National Technical University of Athens, 1996.
[19] Tzabiras G, Garofallidis D, Ventikos Y. On the numerical solution of the Reynolds equations around airfoils
by various orthogonal curvilinear co-ordinate systems. In: Second National Conference on Mechanics,
E.E.Th.E.M.-I.U.T.A.M., Athens, Greece. 1989.
[20] Tzabiras G, Dimas A, Loukakis T. A numerical method for the calculation of incompressible, steady, separated
ows around airfoils. Int J of Num Methods in Fluids 1986;6(11):789809.
[21] Patankar SV. Numerical heat transfer and uid ow. Washington, DC: Hemisphere, 1981.
[22] Gosmann AD, Lai KYM. Finite dierence and other approximations for the transport and NavierStokes
equations. In: IAHR Symposium on Rened Modelling of Flows. 1982.
[23] Spalding DB. A novel nite-dierence formulation for dierential expressions involving both rst and second
derivatives. Int J of Num Methods in Eng 1972;4:5519.
[24] Issa RJ. Numerical methods for two- and three-dimensional viscous ows. V.K.I. Lecture Series, 19811985.
[25] Patankar SV, Spalding DB. A calculation procedure for heat, mass and momentum transfer in three dimen-
sional parabolic ows. Int J Heat and Mass Transfer 1972;15:1787806.
[26] Tzabiras G. A numerical investigation of the Reynolds scale eect on the resistance of bodies of revolution.
Ship Technol Res 1992;39:2844.
[27] Karki KC, Patankar SV. Pressure based calculation procedure for viscous ows at all speeds in arbitrary con-
gurations. AIAA J 1989;27(9):116774.
[28] Frangopoulos C. private communication, 1992.
[29] Chen Y, Heister S. Modeling hydrodynamic nonequilibrium in cavitating ows. J Fluids Eng 1996;118:1728.
[30] Farrel KJ, Billet ML. A correlation of leakage vortex cavitation in axial-ow pumps. J Fluids Eng
1994;116(3):5517.
[31] Avellan F, Dupont P, Ryhming I. Generation mechanism and dynamics of cavitation vortices downstream of a
xed leading edge cavity. In: 17th O.N.R. 1988.
[32] Silberman E. Experimental studies of supercavitating ow about simple two-dimensional bodies in a jet. J Fluid
Mech 1959;5:33754.
[33] Ventikos Y, Tzabiras G, Braza M. The eect of viscous dissipation on the organized structures in the wake
past an airfoil in transition to turbulence. In: Ninth Symposium on Turbulent Shear Flows, Kyoto, Japan.
1993.
[34] Ventikos Y, Tzabiras G, Braza M. Identication of aperiodicity of a NavierStokes solution around an airfoil
with dynamical systems theory tools. In: Fifth International Symposium on Rened Flow Modeling and
Turbulence Measurements, Editions E.N.P.C., Paris, France. 1993.
Y. Ventikos, G. Tzabiras / Computers & Fluids 29 (2000) 6388 88

Potrebbero piacerti anche