INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN FLUIDS

Int. J. Numer. Meth. Fluids 2009; 60:149–175

Published online 6 August 2008 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/fld.1877

A dual-time central difference finite volume scheme for interface

capturing on unstructured meshes

H. Gough, Ann L. Gaitonde∗, † and D. P. Jones

Department of Aerospace Engineering, University of Bristol, University Walk, Bristol BS8 1TR, U.K.

SUMMARY
This paper describes a central difference scheme for the prediction of flows with an interface. The interface
is captured rather than tracked and the key to the current approach is a correction to the hydrostatic
pressure. The correction enables the scheme to evaluate pressures at cell faces in a consistent manner so
that the source term in the equations is correctly balanced at the interface and on non-equispaced meshes.
This prevents the development of large errors in the solution, which can lead to the divergence of the
numerical scheme. The current approach allows interface flows to be calculated by a simple modification
of existing central difference codes. Results for a number of test cases are presented, with comparisons
made with both experimental data and other numerical solutions. Copyright q 2008 John Wiley & Sons,
Ltd.

Received 4 March 2008; Revised 29 May 2008; Accepted 2 June 2008

KEY WORDS: interface capturing; dual-time; central difference; unsteady; finite volume; CFD

1. INTRODUCTION

Numerical simulations of flows with a moving interface separating two immiscible fluids are
complex. The numerical simulation of two fluid flows, particularly water and air flows, is central
to many different types of applications, including the flow through rivers and aqueducts, and is
crucial in the design and construction of naval hydrodynamics, examples of which include ships,
submerged bodies, oil platforms and buoys. Various numerical techniques have been developed
to predict the motion of an interface between two fluids. The main approaches that have been
developed can be characterized as surface fitting, surface tracking and surface capturing [1].
Surface-fitting methods involve adapting the topology of the computational mesh to match the

∗ Correspondence to: Ann L. Gaitonde, Department of Aerospace Engineering, University of Bristol, University Walk,
Bristol BS8 1TR, U.K.
†
E-mail: [email protected]

Contract/grant sponsor: EPSRC

Contract/grant sponsor: BAe Systems

Copyright q 2008 John Wiley & Sons, Ltd.

150 H. GOUGH, A. L. GAITONDE AND D. P. JONES

interface position [2–5]. The use of surface fitting maintains a high interface resolution; however,
such methods are limited to surfaces that do not undergo large deformations. Large deformations in
the interface would lead to significant distortion of the computational mesh, which degrade solution
convergence and quality. Previous results [6] have shown that even simple interface geometries
can produce large mesh deformations. Another disadvantage of surface-fitting methods is the
need to recompute the metrics of the computational mesh after each mesh adaption, which can
be computationally costly depending upon the implementation used. Both surface-tracking and
surface-capturing schemes use meshes that do not move to conform to the free surface; hence, the
grid-related issues of surface-fitting schemes are avoided. Surface-tracking methods explicitly track
the interface location through marker points or functions. The reconstruction of the free surface
location, which is required for such methods, can be complex, especially in three dimensions. The
surface-capturing approach treats the interface as a contact discontinuity in the density field that
is captured by the numerical scheme along with the velocity and pressure fields. This approach is
analogous to shock capturing in single-phase compressible flow. The need for complex surface-
tracking and reconstruction procedures is eliminated [7–9] and complicated interface motion can
be predicted. The use of such schemes on arbitrarily unstructured meshes is no more complicated
than on its structured counterpart. Furthermore, this approach can readily handle situations of fluid
break-up, entrapment, reconnection and also situations in which the interface develops as part of
the flow solution provided the computational mesh resolution is sufficiently fine for the case in
question. Existing schemes of this type utilize an upwind flux formulation based on the Riemann
problem at each cell face [1, 10–12]. A central difference formulation with added dissipation
has not previously been reported for interface capturing and is described in this paper. This
represents a direct extension of the widely used scheme developed for compressible flow [13, 14]
and previously extended to single-phase incompressible flows via an artificial compressibility
approach [15].

2. GOVERNING EQUATIONS FOR INTERFACE CAPTURE

The fundamental assumptions of the approach to interface capturing used in this paper are that each
of the fluids is assumed to be incompressible and isothermal, so that a solution can be obtained
for the pressure and velocity without regard to energy transport.
In a single-phase flow of an incompressible isothermal fluid, the non-dimensional form of the
governing equations of motion is
∇ ·U = 0 (1)
*U 1
+∇ ·(UU) = −∇ p + ∇ ·[∇U+(∇U)T ]+S (2)
*t Re
where S contains any source terms, the Reynolds number Re =U0 L 0 /0 and U0 , L 0 and 0 are
reference values of speed, length and kinematic viscosity, respectively. However, for a two-phase
flow, the density will change in space and time at and near the interface. To allow for this situation,
the density  is reintroduced into the incompressible momentum equation. In addition, a separate
equation for the density is solved to capture the interface position. The density is assumed to
remain constant along particle paths with the resulting transport equation written in conservative
form using the incompressible continuity equation [16]. The new system of equations is given in

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
 A DUAL-TIME CENTRAL DIFFERENCE FINITE VOLUME SCHEME 151

non-dimensional form by
∇ ·U = 0 (3)
*U 
+∇ ·(UU) = −∇ p + ∇ ·[∇U+(∇U)T ]+S (4)
*t Re
*
+∇ ·(U) = 0 (5)
*t
where the reference density for non-dimensionalization is taken to be that of the denser fluid. The
source term is given by
ĝ
S= +f (6)
Fr2
where the two terms on the right-hand side represent √ the buoyancy forces due to gravity and surface
tension, respectively. The Froude number Fr =U0 / |g|L 0 results from non-dimensionalization and
represents the ratio of the fluid flow speed to the speed of small-amplitude surface waves [17]. The
surface tension term is obtained using the continuum surface force (CSF) model, which represents
the surface tension effects as a continuous volumetric force acting in the direct proximity of the
interface, see [16, 18]. The stress tensor and continuity equations remain the same as for a single-
phase incompressible isothermal fluid, since the fluids are still both incompressible and isothermal.
An expression for the mixture viscosity is required and is given by
 = +(1−)r (7)
where r is the viscosity ratio and the volume fraction  is given by
−r
= (8)
1−r
and r is the density ratio of the fluids. The density of the heavier fluid has been used as the
reference density; hence, the non-dimensional fluid density for the heavier fluid is equal to one and
the non-dimensional density of the lighter fluid equals the density ratio r . Similarly the reference
viscosity is that of the heavier fluid; hence, the non-dimensional fluid viscosity for the heavier
fluid equals one and the non-dimensional viscosity of the lighter fluid is given by r , the viscosity
ratio of the fluids.
As a fixed mesh finite volume scheme is to be used, the equations are cast into an integral
form as
  
*
I0 W dV + F·dS = S dV (9)
*t V *V V
These are solved using a dual-time-stepping approach, which introduces an additional fictitious
derivative with respect to a pseudo-time to the equations:
   
* *
Cb W dV +I0 W dV + F·dS = S dV (10)
* V *t V *V V
A solution of this modified system that is steady state in pseudo-time is a time-accurate solution
of the original system of equations (9).

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
152 H. GOUGH, A. L. GAITONDE AND D. P. JONES

The vector W is given by

W = ( p, u, v, w, )T (11)
The flux vector can be decomposed into convective and diffusive components:
F = FC −FD (12)
where
FC = (U, uU+ p î, vU+ p ĵ, wU+ p k̂, U)T
       T (13)
 *U  *U  *U
F = 0,
D
+∇u , +∇v , +∇w , 0
Re *x Re *y Re *z
The matrices I0 and Cb are given by
⎛ ⎞ ⎛ ⎞
0 0 0 0 0 1/2 0 0 0 0
⎜ ⎟ ⎜ ⎟
⎜ .. .. .. .. ⎟ ⎜ .. .. .. .. ⎟
⎜ . 1 . . .⎟ ⎜ . 1 . . .⎟
⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟
⎜ .. ⎟ , ⎜ .. .. ⎟
I0 = ⎜ ... ..
. 1
..
. .⎟ Cb = ⎜ ..
.
..
. 1 . .⎟ (14)
⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟
⎜ .. ⎟ ⎜ .. ⎟
⎜· · · ··· ··· 1 .⎟ ⎜ ··· ··· ··· 1 .⎟
⎝ ⎠ ⎝ ⎠
··· ··· ··· ··· 1 ··· ··· ··· ··· 1
where  is the artificial compressibility parameter [19] used to obtain a hyperbolic system of
equations and the source vector is given by
       T
ĝx ĝ y ĝz
S = 0  2 + f x î,  2 + f  y ĵ,  2 + f z k̂, 0 (15)
Fr Fr Fr

3. CENTRAL DIFFERENCE SCHEME

Equations (10) are solved using a standard central difference scheme [14] together with a new
method to modify the hydrostatic pressure on cell faces. If the standard scheme is used without
the pressure modification, then the source term will not be correctly balanced in the equations.
This can lead to divergence of the numerical scheme.

3.1. Spatial discretization

A finite volume discretization of Equation (10) is formulated by integrating over an arbitrary fixed
cell, P, of an unstructured mesh
dW P dW P
V P  +V P I0 +C P = S P V P (16)
d dt
where W P is the cell-average value of the flow variables located at the cell-centroid, S P is the
source term evaluated using cell-average values, V P is the cell area and C P is the approximation

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
 A DUAL-TIME CENTRAL DIFFERENCE FINITE VOLUME SCHEME 153

to the flux integral. In evaluating the flux terms, the values of the flow variables, W, together
with velocity and temperature gradients on the cell face are needed. The face values of the flow
variables are approximated by taking the average of the flow variables of the cell-centres bounding
the face [20]:
W f = 12 (W L +W R ) (17)
and the gradients are first obtained at cell-centres using Green’s theorem before being averaged
as above to give face values. The hydrostatic component of pressure is modified on the face to
ensure that the source term is correctly balanced, see Section 3.4. In order to prevent odd–even
decoupling, numerical dissipation, either scalar or matrix, is explicitly added to the solution as
detailed in Section 3.5.
The convective, viscous, numerical dissipative fluxes and the source term can all be combined
into a single residual R∗P so that Equation (16) can now be written in semi-discrete form as follows:
dQ P dQ P
V P Cb +V P I0 +R∗P = 0 (18)
d dt
3.2. Time integration
An implicit real-time algorithm is obtained by approximating Equation (18) at real-time level n+1
and putting the real-time derivative into discrete form using a second-order backwards difference
formula [21]:
dQn+1 3Qn+1 n−1
P −4Q P +Q P
n
V P Cb P
+V P I0 +R∗ n+1
P =0 (19)
d 2t
By combining the discrete real-time discretization with the residual term, R∗P n+1 , Equation (19)
can be written in a more compact form:
dQn+1
V P Cb P
+Rn+1
P =0 (20)
d
The required time-accurate solution is found by marching (20) to steady state in pseudo-time at
each real time step using an explicit four stage Runge–Kutta scheme [14, 15]. Local pseudo-time-
stepping and multigrid are used for convergence acceleration.

3.3. Local pseudo-time-stepping

The local pseudo-time-step in cell P is calculated from both convective and diffusive components:
1 1 1
= + (21)
 P CP DP
where the convective time step, for a fixed mesh, is given by
2V P Co
CP =  (22)
f (| f ||S f |)

where Co is the Courant number, f is the cell spectral radius in the face normal direction and
S f is the cell face area. This differs from the structured mesh approach to the local time step [22],

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
154 H. GOUGH, A. L. GAITONDE AND D. P. JONES

which uses a spectral radius scaled in each of the coordinate directions. However, this information
is not available on a fully unstructured mesh. The diffusive component of the local time step is
given by

Co(V P )2
DP =  (23)
2k/Re( f |S f |2 )

where k is an adjustable constant that usually takes a value of 2. For unsteady calculations where
the real time step, t, is small compared with the local pseudo-time-step, as calculated above,
it has been shown that the discretization scheme can become unstable [13]. To overcome this a
further limit is applied to the pseudo-time-step [13] to prevent it becoming too large in comparison
with the real time step. For the current scheme this leads to
 P = min( P , 23 t) (24)

3.4. Hydrostatic pressure discretization

Central difference schemes calculate the value of a variable at a cell face, for surface flux evalua-
tions, using a simple average of the cell values bounding the face. When this is applied to obtain
the face pressure, both the kinematic and hydrostatic pressure components will be averaged. The
source term on the right-hand side of Equation (16) should be balanced by the hydrostatic pressure
flux calculated using the central difference scheme. Away from the interface, on an equispaced
computational mesh, the hydrostatic pressure flux will correctly balance the gravitational source
term. This is not true when the central difference scheme is applied using a computational mesh
that is not equispaced. The situation is further complicated at the interface itself where the steep
change in the density profile can introduce large errors into the computation especially for low
Froude number problems where the source term becomes large and correspondingly the hydrostatic
pressure gradient increases. An upwind scheme uses gradient extrapolation to calculate the flow
variables at a cell face. Therefore, in regions away from the interface the correct hydrostatic pres-
sure field will be calculated at the cell faces even on a mesh that is not equispaced. In the vicinity
of the interface, however, even for an upwind scheme the problem of correctly balancing the
hydrostatic pressure field with the source term is a known issue [23]. If the position of the interface
were known, then it would be possible to average the kinematic pressure component and directly
include the correct hydrostatic pressure. However, this is not possible with interface capturing
schemes since the interface position is never explicitly known. Therefore, a novel approach has
been developed to calculate the hydrostatic pressure consistently at the cell faces for a central
difference scheme without the position of the interface being known. The new scheme produces
values for pressure at cell faces throughout the computational domain, including the interface itself,
which are consistent and balance the source term correctly. The hydrostatic pressure on the face
is first approximated by the cell-centred scheme as a simple average:
( p H )L +( p H )R
( pH ) f = (25)
2
Figure 1 shows the situation for a mesh that is not equispaced and for the purposes of drawing the
figure, gravity is assumed to act in the negative z direction. Two cells with different densities are
considered bounding a face f . Then the cell-centre hydrostatic pressure values in Equation (25),
which are set correctly at the start of the computation, are assumed to remain correct for all

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
 A DUAL-TIME CENTRAL DIFFERENCE FINITE VOLUME SCHEME 155

Figure 1. Hydrostatic pressure gradients at a cell face—shown for case when

cell left is at a greater depth than cell right.

time. They can then be rewritten in terms of the correct face hydrostatic pressure ( p corr
H ) f and the
hydrostatic gradient in each of the cells adjacent to the face, which are given by
L ĝ
∇ pH
L
=
Fr2
R ĝ
∇ pH
R
=
Fr2
Equation (25) becomes

H ) f +∇ p H ·(xL −x f )+( p H ) f +∇ p H ·(xR −x f )

( p corr L corr R
( pH ) f =
2
 
L (xL −x f )+R (xR −x f )
= ( p H ) f + ĝ·
corr
2Fr2
where x f is the position vector of the cell face centre and xL and xR are the position vectors of
the cell-centres of the left and right cells bordering the face.
On rearranging, this gives
( p corr
H ) f = ( p H ) f −
(26)
where the error term
is given by
 
 (xL −x f )+R (xR −x f )

= ĝ· L (27)
2Fr2
Using the corrected value for the hydrostatic pressure, the cell face value of the total pressure, p,
can be obtained from
pL + pR
( p corr ) f = −
(28)
2
Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
156 H. GOUGH, A. L. GAITONDE AND D. P. JONES

Note that if the cell densities and the cell dimensions in the direction of the gravity field are equal
L = R
ĝ·(x f −xL ) = −ĝ·(x f −xR )
then the error term will be zero. This approach for the hydrostatic pressure discretization ensures
that the field is correctly reproduced throughout the flow domain including at the interface. A
different hydrostatic gradient correction approach has been recently developed to address this issue
for an upwind Roe flux scheme by Qian et al. [23].

3.5. Numerical dissipation

Numerical dissipation is added to central difference schemes to prevent numerical oscillations in the
solution. Both scalar [14] and matrix [24] dissipation models have been implemented in this study.

3.5.1. Scalar dissipation model, JST. The scalar dissipation model used in this research is a
modified version of that first described by Jameson et al. [14] for compressible flows. The original
dissipation model consisted of a fourth-order term that stabilizes the central difference scheme
without smearing the solution and a second-order term that is activated in regions of strong
gradients; typically for compressible flows this would be regions of ‘shock’ where the pressure
gradient changes rapidly. For simple incompressible flows no regions of large gradients equivalent
to shock waves in compressible flows exist; therefore, only the fourth-order dissipation would be
necessary. However, for a flow with an interface, regions of large gradients will exist and it is
necessary to use both the second- and fourth-order dissipation terms in the current scheme. The
dissipation ‘shock detector’ or switch for compressible flow is based on pressure. In this work the
switch has the same form as for compressible flow, but the detector is in this case based on density.
The scalar dissipation model is easy to formulate and quick to compute. However, each dissipative
flux equation is scaled by the same factor, the spectral radius of the Jacobian matrix. This can
degrade the numerical results particularly for flows that include regions of flow within boundary
layers and for flows that contain large solution gradients where the second-order dissipation is
necessary, e.g. when the flow features an interface. This issue is partly addressed by the matrix
dissipation model.

3.5.2. Matrix dissipation model. The matrix dissipation model [24] allows different magnitudes of
dissipation for each of the discrete equations. Matrix dissipation was developed by considering the
differences between upwind and central difference schemes and formulating the matrix dissipation
to mimic a first-order upwind scheme in the vicinity of ‘shocks’ or large solution gradients. The
vector form of the dissipative flux is, in this case, scaled by the absolute value of the Jacobian
matrix. The switch is modified compared with the scalar case. Some implementations of matrix
dissipation have used a switch based on the Van Leer flux limiter that produces a total variation
diminishing (TVD) scheme [25, 26]. However, in this study a weaker form of the switching function
has been used as this gives improved convergence [16, 27] compared with the Van Leer limiter.

3.5.3. Numerical dissipation correction. The numerical dissipation schemes are based on the
differences of the flow variables in W. The pressure variable, p, contains the combined kinematic
pk and hydrostatic p H components. If differencing of p itself is used it will introduce unnec-
essary dissipation. For example, consider a container filled with water at rest, then the correct

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
 A DUAL-TIME CENTRAL DIFFERENCE FINITE VOLUME SCHEME 157

Figure 2. Set-up for Rayleigh–Taylor instability in dimensional variables.

steady solution will have a varying pressure due to the hydrostatic pressure gradient. This solution
should not have any dissipation introduced into the continuity equation as the variations in the
pressure field are a correct feature of the flow. Thus, the dissipation should only use the differ-
ences of the kinematic pressure. The correct differences in the kinematic pressure can be made
using a similar approach to that in Section 3.4. The difference in pressure at the face will be
given by
 p f = ( p H ) f +( pk ) f (29)
where  p f = pR − pL and similarly for  p H and  pk . For the numerical dissipation scheme the
difference of the kinematic pressure at the face is required. This can be obtained by rearranging
Equation (29) as
( pk ) f =  p f −( p H ) f (30)
and noting that if
( p H ) f = ( p H )R −( p H )L

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
158 H. GOUGH, A. L. GAITONDE AND D. P. JONES

Figure 3. Contours for the Rayleigh–Taylor instability matrix scheme: (a) t = 0.8; (b) t = 1.6; (c) t = 2.4;
(d) t = 3.2; (e) t = 4.0; and (f ) t = 4.8.

then expressing ( p H )R and ( p H )L in terms of the correct hydrostatic pressure on the face
gives

ĝ·[L (x f −xL )−R ·(x f −xR )]

( p H ) f = (31)
Fr2
Therefore, the final required difference term for the numerical dissipation will be given by the
following expression:

ĝ·[L (x f −xL )−R ·(x f −xR )]

( pk ) f = ( pR − pL )− (32)
Fr2

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
 A DUAL-TIME CENTRAL DIFFERENCE FINITE VOLUME SCHEME 159

Figure 4. Contours for the Rayleigh–Taylor instability JST scheme: (a) t = 0.8; (b) t = 1.6; (c) t = 2.4;
(d) t = 3.2; (e) t = 4.0; and (f ) t = 4.8.

The expression for kinematic pressure differences given by Equation (32) is used in the numerical
dissipation given for the continuity equation.

4. BOUNDARY CONDITIONS AND INITIALIZATION

The boundary conditions are implemented using ‘ghost’ or ‘halo’ cells at the boundary.

4.1. Impermeable no-slip walls

The density is extrapolated from the interior value to the wall using a zero gradient and the pressure
is extrapolated with a fixed gradient. The velocity of the fluid at the wall is set equal to that of
the wall.

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
160 H. GOUGH, A. L. GAITONDE AND D. P. JONES

Figure 5. Contours for the Rayleigh–Taylor instability upwind scheme: (a) t = 0.8; (b) t = 1.6; (c) t = 2.4;
(d) t = 3.2; (e) t = 4.0; and (f ) t = 4.8.

4.2. Inlet/outlet
The method of characteristics has been used at inlet/outlet boundaries. The characteristic variables
of the interface capturing equations are extrapolated either from the flow interior or from the
exterior values depending upon the sign of the corresponding eigenvalue. In order to avoid wave
reflection at the outflow, wave damping is used [28–31] to dissipate all the wave energy of the
outgoing waves. In the surface-tracking approach the damping in the outflow can be achieved
by damping the surface elevation. For the surface-capturing approach this is not directly possible
because the interface location is not known; hence, a different approach is needed. One possibility
would be to use a spatially static viscosity ramp [32, 33] such that the fluid thickens near the
outflow, which would have the desired effect of dissipating the energy of the outgoing waves.

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
 A DUAL-TIME CENTRAL DIFFERENCE FINITE VOLUME SCHEME 161

Figure 6. Contours for the Rayleigh–Taylor instability from Kelecy and Pletcher [1].

However, for calculations with explicit pseudo-time-stepping,‡ the increase in viscosity yields a
considerable limitation on the allowable local time step. This viscous limitation on the time step
becomes rapidly more restrictive than the convective limitation. Consequently, the viscous time
step restriction deteriorates the overall efficiency of the numerical method. A second approach that
has been utilized here is to damp the fluid motion in the direction of the gravity vector [31] using
a damping factor that is switched on in a damping zone around the outflow boundary.

‡
The maximum local time step depends on both inviscid and viscous contributions.

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
162 H. GOUGH, A. L. GAITONDE AND D. P. JONES

1
Upwind
0.95 MATRIX
JST
0.9
0.85

Density Profile
0.8
0.75
0.7
0.65
0.6
0.55
0.5
0 0.2 0.4 0.6 0.8 1
x

Figure 7. Comparison of density profiles at z = −0.3, t = 4 for Rayleigh–Taylor instability.

Figure 8. Cut location at z = −0.3, t = 4 for Rayleigh–Taylor instability density profile comparison.

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
 A DUAL-TIME CENTRAL DIFFERENCE FINITE VOLUME SCHEME 163

0.24
Chandrasekhar
0.22 JST

non-dimensional growth rate

MATRIX
Upwind
0.2 Daly

0.18
0.16
0.14
0.12
0.1
0.08
0.06
0 50 100 150 200 250 300
Reynolds Number

Figure 9. Non-dimensional growth rate for Rayleigh–Taylor instability at Re = 39, 72 and 176.

4.3. Initialization
The initialization of the flow variables is of particular importance when dealing with the interface
capturing scheme. In particular, the pressure field has to be initialized to the correct hydrostatic
distribution in order to satisfy the continuity and momentum equations. The initial interface location
is known at the start of the simulation; hence, the hydrostatic pressure distribution can be initialized
to the correct values easily. The correct initialization is dependent on the particular test case and
is explained in the description of the results for each test case.

5. RESULTS

For the new interface capturing scheme, results have been produced for various test cases in order
to verify the accuracy of the predicted results, in particular the interface position. The interface is
represented by the contour of the average value of the densities of the two fluids, which is given
in non-dimensional variables by
1+r
ave = (33)
2
For the purpose of comparison, results are also presented using an upwind scheme implemented in
the same computational environment [23, 34, 35]. Note that in all cases the value of the artificial
compressibility parameter is taken to be 1 and the CFL number is 1. The environment in which this
scheme was implemented allows for any general unstructured mesh. In this study hybrid meshes
are used with a background Cartesian mesh with enriched pockets around features of interest such
as solid and free surface [36].

5.1. Rayleigh–Taylor instability

The Rayleigh–Taylor instability is a well-known fluid dynamics problem [37, 38] for which an
analytical solution can be produced for the initial stages using linear perturbation analysis [39].
The problem consists of a heavy fluid sitting on top of a lighter fluid, with the interface subject

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
164 H. GOUGH, A. L. GAITONDE AND D. P. JONES

Figure 10. Set-up for the broken dam problem in dimensional variables.

5
MATRIX
JST
4.5 Upwind
Martin & Moyce

3.5
surge front position

2.5

1.5

1
0 0.5 1 1.5 2 2.5 3 3.5
t*

Figure 11. Surge front position versus non-dimensional time for the coarse mesh.

to a perturbation. Buoyancy forces cause the amplitude of the perturbation to grow with time.
Previous computational studies have investigated the problem [1, 10, 11, 40] because it provides
an interesting test case for interface schemes due to the mixing of the two fluids. The specific test
case considered here has a density ratio r = 0.5 and the kinematic viscosities of the two fluids
are set to be equal so that r = 0.5. A single wavelength perturbation is introduced at the fluid
interface using the following non-dimensional velocity field [40]:

 sin(x)e−|z| , z>0
u=
−|z|
− sin(x)e , z<0

w = − cos(x)e−|z|

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
 A DUAL-TIME CENTRAL DIFFERENCE FINITE VOLUME SCHEME 165

1
MATRIX
JST
0.9 Upwind
Martin & Moyce

0.8

water column height

0.7

0.6

0.5

0.4

0.3
0 0.5 1 1.5 2 2.5 3 3.5
t*

Figure 12. Dam break column height versus non-dimensional time for the coarse mesh.

where  is a constant that governs the amplitude of the initial displacement. For this test case
the value was set to  = 0.25, as data are available for comparison [11]. Figure 2 shows the
initial set-up. The computational mesh has dimensions of 30×90. The reference length for non-
dimensionalization L 0 is taken to be the width of the computational domain and the depth of each
fluid is H = 1.5L 0 . The Reynolds number was set to Re = 283. In keeping with other studies, the
boundary conditions applied are no-slip walls on the top and bottom surfaces of the computational
domain and symmetry conditions on the sides of the domain.
Figures 3 and 4 show the results obtained from the central difference scheme with the new
hydrostatic pressure correction using matrix and JST dissipation, respectively. Figure 5 shows results
obtained using a standard upwind scheme implemented in the same computational environment as
the central difference scheme by the authors. The plots show density contours around ave = 0.75,
with five contour lines plotted in the range 0.70.8. It can be seen from the figures that the
initial velocity perturbation causes the fluid to rise on the left-hand side and sink on the right-hand
side. This motion causes a mixing of the two fluids resulting in the familiar mushroom shape.
Figure 6 shows the results obtained using an upwind scheme by Kelecy and Pletcher [1] for
comparison. There are some small differences in the shape of the interface between both central
difference scheme results and the upwind results of Kelecy and Pletcher. However, the central
difference results compare well with the results from the upwind scheme implemented by the
authors. This suggests that the difference from the results of Kelecy and Pletcher is mainly due to
differences in the computational environment rather than the use of central differencing in place
of upwinding.

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
166 H. GOUGH, A. L. GAITONDE AND D. P. JONES

5
MATRIX
JST
4.5
Upwind
Martin & Moyce
4

3.5
surge front position

2.5

1.5

1
0 0.5 1 1.5 2 2.5 3 3.5
t*

Figure 13. Surge front position versus non-dimensional time for the fine mesh.

In order to further study the density spread, Figure 7 displays the density profile at the position
z = −0.3 for time t = 4. The location of the slice is more clearly depicted in Figure 8, which shows
the location of the cut plane. At this time the mushroom shape of the Rayleigh–Taylor instability
is starting to form. Figure 7 shows that the cut profile: passes from the lighter fluid to the heavier
one at x = 0.28; then transforms back to the lighter fluid at x = 0.42; is influenced by the proximity
of the region of heavier fluid at x = 0.58, which causes a small secondary spike; and passes into
the heavier region x = 0.81. The profiles for all the schemes have a similar shape, but there are
differences particularly at the peaks and troughs between all three schemes. For this test case the
matrix scheme took 40% more computational time than the JST scheme and the upwind scheme
took approximately 50% more computational time than the JST scheme.
Chandrasekhar [39] derived an analytical expression for the linear growth rate n ∗ of the Rayleigh–
Taylor instability as a function of the Reynolds number. In order to verify that the current interface
capturing methodology can accurately predict the growth rates, several cases were run at different
Reynolds numbers. The amplitude of the displacement was then used to calculate the linear growth
rate as outlined in [41]. The amplitude of the displacement is taken to be the average of the absolute
value of the displacement of the left and right of the instability. After the stable modes of the
disturbance in the initial transient have decayed away, the natural logarithm of the amplitude of the
displacement becomes linear with time. The slope of the linear region gives the growth rate n ∗ of
the corresponding Reynolds number. The slope is calculated based on the time interval t = 0.2–0.5.
The linear growth rate for the Reynolds numbers 39, 72 and 176 has been computed such that a
comparison with other research can be made. The predicted growth rate values in Figure 9 show

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
 A DUAL-TIME CENTRAL DIFFERENCE FINITE VOLUME SCHEME 167

1
MATRIX
JST
Upwind
0.9
Martin & Moyce

0.8

water column height

0.7

0.6

0.5

0.4

0.3
0 0.5 1 1.5 2 2.5 3 3.5
t*

Figure 14. Dam break column height versus non-dimensional time for the fine mesh.

good agreement with the analytical curve based on the linear analysis of Chandrasekhar and also
with the results obtained in another numerical study by Daly [40, 41].

6. BROKEN DAM

The broken dam or water column collapse has been extensively investigated by various researchers
both numerically [42–45] and also experimentally [46, 47]. It is therefore often used as a benchmark
test case. The initial configuration and computational domain is depicted in Figure 10 at t = 0
when the column of water is confined to the left of a tank. For t>0 the restraint is removed
and the water is allowed to collapse under the influence of gravity. The reference length used in
non-dimensionalization is the height of the water column
√ L 0 , the reference density is taken to be
that of the water and the reference velocity U0 = |g|L 0 . The density ratio r = 0.001 and the
viscosity ratio r = 0.01. The Reynolds number is then given by
3/2
0 |g|1/2 L 0
Re = (34)
0
The pressure distribution is initialized to a hydrostatic distribution relative to the top of the
water column. The computational domain of this simulation represents a closed container with
impermeable walls. This set-up differs from the experimental one that utilized an open container,

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
168 H. GOUGH, A. L. GAITONDE AND D. P. JONES

Figure 15. Contours of  for the breaking dam problem: (a) t = 0.0; (b) t = 0.6; (c) t = 1.2;
(d) t = 1.8; (e) t = 2.4; and (f ) t = 3.0.

but is the usual set-up for numerical simulations. All of the computational boundaries are set to
be slip walls for consistency with other numerical studies [43, 44].
For the current test case L 0 = 0.05715 m, Re = 43 000 and ave = 0.5005. Two different compu-
tational meshes were constructed for the simulation: a coarse mesh of dimensions 80×20 and a
fine mesh of dimensions 160×40.
Figure 11 plots the location of the surge front against the non-dimensional time for the coarse
computational mesh and is compared with experimental data from Martin and Moyce [46]. The
matrix and JST solutions are closer to the experimental surge front position than the upwind results,
which has over-predicted the surge front position. Near the end of the simulation, t>2.50, the JST
dissipation starts to under-predict the surge front position. The column height is plotted against
the non-dimensional time for the coarse mesh in Figure 12. The graph shows that there are some
differences between all the schemes and the experimental data. Solutions at this mesh density can

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
 A DUAL-TIME CENTRAL DIFFERENCE FINITE VOLUME SCHEME 169

Figure 16. Vector plots for the breaking dam problem: (a) t = 0.6; (b) t = 1.2;
(c) t = 1.8; (d) t = 2.4; and (e) t = 3.0.

seem to show an exaggerated level of agreement with the experimental data. In fact, these solutions
were not fully converged with respect to the mesh size and obtaining a converged solution led to
results further from the experimental data.
Figure 13 shows the surge front position on the fine computational mesh. The results of the
matrix dissipation more closely match those obtained using the upwind formulation. This is in
keeping with the way the matrix dissipation scheme has been constructed to closely mimic an
upwind formulation. However, the central differences schemes and the upwind scheme all over-
predict the surge front location; that is, the simulated surge front moves faster than the experimental
one. This phenomenon has been reported in other research [43, 44, 47] and is attributed to the
difficulty in determining the exact location of the leading edge. A thin layer, similar to a jet,
shoots along the bottom of the tank, which is difficult to capture, followed by the bulk of the
flow. Furthermore, the discrepancy between the simulated and experimental results might also
be due to the removal of the barrier holding the liquid in the left-hand corner at the start of
the experimental simulation. In the numerical experiment this barrier is removed instantaneously,
whereas in the physical world the removal of the barrier would have a time associated with it.

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
170 H. GOUGH, A. L. GAITONDE AND D. P. JONES

The predicted results for the column height on the fine mesh, Figure 14, agree well with the
experimental data.
Figure 15 shows density contours and Figure 16 shows vector plots calculated using matrix
dissipation on the fine computational mesh. In Figure 15, 10 density contours are plotted within

Figure 17. Pressure contours for the breaking dam problem: (a) t = 0.0; (b) t = 1.2; and (c) t = 2.4.

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
 A DUAL-TIME CENTRAL DIFFERENCE FINITE VOLUME SCHEME 171

the range 0.30.6. The interface has been captured with a small spread. For this test case and
the others reviewed in this research, the spread of the standard deviation of the interface was over
approximately 10 cells and this was consistent as the computational mesh was refined. Figure 16
shows an interesting feature of the dam break, which is the vortex that develops just above the
interface and remains there for the duration of the simulation as a result of the forces upon the air
next to the water.
Figure 17 shows the pressure contours obtained at times t = 0.0, 1.2 and 2.4 for the matrix
dissipation. At time t = 2.4 contours arise in the lower right-hand corner. These are due to the
fact that, consistent with other numerical studies, the computational domain is a closed container.
The thin layer that crosses the domain ahead of the bulk of the flow hits the side of the container
giving rise to the pressure disturbance.

7. THREE-DIMENSIONAL WATER COLUMN COLLAPSE

The preceding test cases have been simulated as two-dimensional by creating a thin slice of the flow
to create a three-dimensional mesh with appropriate boundary conditions on the symmetry planes.
In order to test the current implementation in a fully three-dimensional regime, a challenging test
case is considered here. A series of three-dimensional dam break experiments was conducted by
Martin and Moyce [46]. These consisted of a cylinder of water at t = 0.0, which was subsequently
allowed to collapse under the influence of gravity. The experimental data measured the axial surge
front position as a function of time. The initial configuration is axisymmetric and the initial radial
slice is the same as shown in Figure 10. The reference length was taken to be the base radius
of the initial water column and the reference density was that of water, resulting in a Reynolds
number of Re = 43 000 and density and viscosity ratios of r = 0.001, r = 0.01, respectively. All
of the boundaries were set to be viscous walls. The computational mesh used was of dimensions
120×120×40 in the radial, circumferential and vertical directions, respectively.
Figure 18 shows the predicted surge front position plotted against time for JST, matrix and
upwind schemes together with the results obtained from Martin and Moyce. This shows that the

5
MATRIX
4.5 JST
Upwind
axial surge front position

4 Martin & Moyce

3.5

2.5

1.5

1
0 0.5 1 1.5 2 2.5 3 3.5
t

Figure 18. Axial surge front position for three-dimensional dam break.

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
172 H. GOUGH, A. L. GAITONDE AND D. P. JONES

JST scheme slightly under-predicts the surge front position. The upwind and matrix schemes give
similar results until the latter stages when the upwind scheme over-predicts the surge front position
The difference between the JST and matrix numerical dissipation schemes is more pronounced
for this three-dimensional test case because of the three-dimensional motion of the liquid, which
creates diffusion in both the normal and stream-wise directions relative to the interface. Each
equation in the JST dissipation is scaled by the spectral radius, whereas the matrix dissipation
scales the respective equations corresponding to the characteristic wave speed of the flow. The
slight oscillations in the JST curve due to the increased diffusion of the scheme makes it more
difficult to identify the location of the surge front. Figure 19 shows the iso-surfaces of the average
density ave computed using the matrix dissipation scheme. The progression of the water column
collapse can be seen.

Figure 19. Iso-contours of ave for the three-dimensional breaking dam problem: (a) t = 0.0; (b) t = 0.6;
(c) t = 1.2; (d) t = 1.8; (e) t = 2.4; and (f ) t = 3.0.

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
 A DUAL-TIME CENTRAL DIFFERENCE FINITE VOLUME SCHEME 173

8. CONCLUSIONS

This paper has described a central difference scheme for the interface capturing scheme. Although
unstructured meshes have been used in this study, the methodology could equally be applied to
structured mesh codes. Results for a number of test cases have been simulated using the JST and
matrix dissipation. Upwind results have also been calculated for comparison purposes. For these
test cases the new hydrostatic pressure discretization ensures that the source term is balanced by
the hydrostatic pressure flux term. The results for the internal flows using either JST and matrix
scheme show that this formulation can accurately capture the main physical characteristics of the
flow for the test cases considered, but at a lower computational cost compared with the upwind
scheme. The JST scheme has been shown to introduce a larger amount of interfacial spread but still
performs well. This work shows that sufficiently small mesh spacing is required within the region
of the interface for these flows to be accurately simulated with the central difference schemes.
Insufficient mesh spacing can cause diffusion in the interface that degrades the results. With a mesh
spacing O(10−3 ), the current implementation has been shown to predict the flow with reasonable
accuracy. From the results of the interface capturing test cases, the standard deviation of the
interface was found to be approximately 10 cells on average. Further, mesh spacing reduction
would improve the predicted results but this would create very large computational meshes and
perhaps another approach, e.g. mesh adaption, is needed in order to produce an implementation
that can be used as an efficient design tool.

ACKNOWLEDGEMENTS
This work was funded by EPSRC and BAe Systems. This support is gratefully acknowledged.

REFERENCES
1. Kelecy FJ, Pletcher RH. The development of a free surface capturing approach for multidimensional free surface
flows in closed containers. Journal of Computational Physics 1997; 138:939–980.
2. Ferziger JH, Perić M. Computational Methods for Fluid Dynamics. Springer: Berlin, 1996.
3. Takizawa A, Koshizuka S, Kondo S. Generalization of physical component boundary fitted coordinate (PCBFC)
method for the analysis of free-surface flow. International Journal for Numerical Methods in Fluids 1992;
15:1213–1237.
4. Li YS, Zhan JM. An efficient three-dimensional semi-implicit finite-element scheme for simulation of free surface
flows. International Journal for Numerical Methods in Fluids 1993; 5(8):1904–1913.
5. Farmer J, Martinelli L, Jameson A. Fast multigrid method for solving incompressible hydrodynamic problems
with free surfaces. AIAA Journal 1994; 32:1175–1182.
6. Ramaswamy B, Kawahara M. Lagrangian finite element analysis applied to viscous free surface fluid flow.
International Journal for Numerical Methods in Fluids 1987; 7(9):953–984.
7. Hirt CW, Nichols BD. Volume of fluid VOF method for the dynamics of free boundaries. Journal of Computational
Physics 1981; 39:201–225.
8. Youngs DL. Time-dependent multi-material flow with large fluid distortion. Numerical Methods for Fluid
Dynamics. Academic Press: London, 1982; 273–285.
9. Unverdi SO, Tryggvason G. A front tracking method for viscous incompressible flows. Journal of Computational
Physics 1992; 100:25–37.
10. Pan D, Chang C-H. The capturing of free surfaces in incompressible multi-fluid flows. International Journal for
Numerical Methods in Fluids 2000; 33:203–222.
11. Zhao Y, Hui Tan H, Zhang B. A high-resolution characteristics based dual time-stepping VOF method for free
surface flow simulation on unstructured grids. Journal of Computational Physics 2002; 183:233–272.

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
174 H. GOUGH, A. L. GAITONDE AND D. P. JONES

12. Qian L, Causon M, Ingram DM, Mingham CG. Cartesian cut cell two-fluid solver for hydraulic flow problems.
Journal of Hydraulic Engineering 2003; 129(9):688–696.
13. Arnone A, Liou M-S, Povinelli LA. Multigrid time-accurate integration of the Navier–Stokes equations. AIAA
93-3361, 1993.
14. Jameson A, Schmidt W, Turkel E. Numerical solution of the Euler equations by finite volume methods using
Runge–Kutta time stepping schemes. AIAA Paper 81-1259, 1981.
15. Gaitonde AL. A dual-time method for two-dimensional unsteady incompressible flow calculations. International
Journal for Numerical Methods in Engineering 1998; 41:1153–1166.
16. Gough HJ. Incompressible interface capturing methods for application to free surfaces and cavitation. Ph.D.
Thesis, University of Bristol, 2007.
17. Acheson DJ. Elementary Fluid Dynamics. Oxford Applied Mathematics and Computing Science Series. Oxford
University Press: Oxford, 2003.
18. Brackbill JU, Kothe DB, Zemach C. A continuum method for modelling surface tension. Journal of Computational
Physics 1992; 100:335–354.
19. Chorin AJ. A numerical method for solving incompressible viscous flow. Journal of Computational Physics 1967;
2:12–26.
20. Jameson A, Mavriplis D. Finite volume solution of the two-dimensional Euler equations on a regular triangular
mesh. AIAA Paper 85-0435, 1985.
21. Jameson A. Time dependent calculations using multigrid, with applications to unsteady flows past airfoils and
wings. AIAA Paper 91-1596, 1991.
22. Tejera-Cuesta P. A parallel dual-time multigrid method for incompressible flows over deforming geometries.
Ph.D. Thesis, University of Bristol, 2000.
23. Qian L, Causon DM, Mingham CG, Ingram DM. A free-surface capturing method for two fluid flows with
moving bodies. Proceedings of the Royal Society A 2006; 462:21–42.
24. Swanson RC, Turkel E. On central-difference and upwind schemes. Journal of Computational Physics 1992;
101:292–306.
25. Sweby PK. High resolution schemes using flux limiters for hyperbolic conservation laws. SIAM Journal on
Numerical Analysis 1984; 21:995–1011.
26. Jorgensson P, Turkel E. Central difference TVD schemes for time dependent and steady state problems. Journal
of Computational Physics 1993; 107:297–308.
27. Swanson RC, Radespiel R, Turkel E. On some numerical dissipation schemes. AIAA 97-1945, 1993.
28. Hino T. An unstructured grid method for incompressible viscous flows with a free surface. AIAA-1997-0862,
1997.
29. Löhner R, Yang C, Oñate E, Idelssohn S. An unstructured grid-based, parallel free surface solver. AIAA-1997-1830,
1997.
30. Hino T. A finite-volume method with unstructured grid for free surface flow simulations. Proceedings of 6th
International Symposium on Numerical Ship Hydrodynamics, Iowa City, 1993.
31. Park J-C, Kim M-H, Miyata H. Fully non-linear free surface simulations by a 3D viscous numerical wave tank.
International Journal for Numerical Methods in Fluids 1999; 29:685–703.
32. Wasistho B, Geurts BJ, Kuerten GM. Simulation techniques for spatially evolving instabilities in compressible
flow over a flat plat. Computers and Fluids 1997; 26(7):713–739.
33. Liu S, Liu C. Fourth order finite difference and multigrid methods for modelling instabilities in flat plate boundary
layers. Computers and Fluids 1994; 23:955–982.
34. Roe PL. Approximate Riemann solvers, parameter vectors, and difference schemes. Journal of Computational
Physics 1981; 43:357.
35. Swanson RC, Turkel E. On central difference and upwind schemes. Technical Report 182061, ICASE, 1990.
36. Leatham M, Stokes S, Shaw JA, Cooper J, Appa J, Blaylock TA. Automatic mesh generation for rapid-response
Navier–Stokes calculations. AIAA2000-2247, 2000.
37. Rayleigh L. Investigation of the character of the equilibrium of an incompressible heavy fluid of variable density.
Scientific Papers 1900; 200.
38. Taylor GI. The instability of liquid surfaces when accelerated in a direction perpendicular to their planes.
Proceedings of the Royal Society of London, Series A 1950; 201:192.
39. Chandrasekhar S. Hydrodynamic and Hydromagnetic Stability, Chapter X. Oxford University Press: Oxford,
1961; 428–447.
40. Daly BJ. Dynamics of liquids in moving containers. Physics of Fluids 1967; 10:297.

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld
 A DUAL-TIME CENTRAL DIFFERENCE FINITE VOLUME SCHEME 175

41. Daly JB. Numerical study of two fluid Rayleigh–Taylor instability. Physics of Fluids 1967; 10(2):297–307.
42. Walhorn E, Kölke A, Hübner B, Dinker D. Fluid–structure coupling within a monolithic model involving free
surface flows. Computers and Structures 2005; 83:2100–2111.
43. Greaves D. Simulation of viscous water column collapse using adapting hierarchical grids. International Journal
for Numerical Methods in Fluids 2005; 50(6):693–711.
44. Ubbink O. Numerical prediction of two fluid systems with sharp interfaces. Ph.D. Thesis, Imperial College,
University of London, 1997.
45. Andrillon Y, Bertrand A. A 2D+T VOF fully coupled formulation for the calculation of breaking free-surface
flow. Journal of Marine Science and Technology 2004; 8:159–168.
46. Martin JC, Moyce WJ. An experimental study of the collapse of liquid columns on a rigid horizontal plane.
Philosophical Transactions of the Royal Society of London, Series A: Mathematical, Physical and Engineering
Sciences 1952; 244:312–324.
47. Koshizuka S, Tamako H, Oka Y. A particle method for incompressible viscous flow with fluid fragmentation.
Computational Fluid Mechanics Journal 1995; 4:29–46.

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2009; 60:149–175
DOI: 10.1002/fld