You are on page 1of 8

Engineering Applications of Computational Fluid Mechanics Vol. 1, No. 1, pp.

49-56 (2007)

FREE-SURFACE SIMULATIONS USING AN INTERFACE-TRACKING


FINITE-VOLUME METHOD WITH 3D MESH MOVEMENT
J. J. R. Williams
Department of Engineering, Queen Mary, University of London, U.K.
E-mail: J.J.R. Williams@qmul.ac.uk
ABSTRACT: In the numerical solution of a free-surface fluid flow problem, an additional conservation equation
(called the space conservation law) has to be solved along with the mass, momentum and energy conservation equations.
In this paper a surface adaptive curvilinear finite volume method for solving free-surface flows is presented. The
computational mesh defines hexahedra that are used as the control volumes in a finite-volume method in which the mesh
moves to adapt to the free-surface. Dependent variables are located at the centre of each volume. Code verification is
carried out on a number of inviscid, laminar and turbulent flow cases and the results show good agreement with
experimental data and theoretical solutions.

1. INTRODUCTION
In many problems of engineering interest, the
boundary of the computational domain can change
with time. For instance, for flow within the cylinder
of an internal combustion engine, the boundary
motion is prescribed for a given engine speed. For
other problems, such as liquid wave propagation
and open channel flows, the motion of the surface is
determined by the fluid flow itself. Many flows in
geophysics and industrial engineering have a gasliquid interface (free-surface). The free-surface
alters the turbulence structure near the surface
significantly. In particular, the kinetic energy of
vertical velocity fluctuations is redistributed to
horizontal motions. Computations of free-surface
flows have now been carried out by a number of
workers [Demirdzic & Peric (1988), Farmer et al
(1994), Kordulla (1983), Muzaferija & Peric (1998),
Th, et al (1994), Thomas, Leslie & Williams (1995)
and Meselhe & Sotiropoulos (2000)] with, in
general, results showing good agreement with
experiments and theoretical solutions. Also, many
methods have been used to track the free-surface.
These can, however, be classified into two major
groups in a manner similar to the techniques used to
follow shocks: interface-capturing and interfacetracking methods. For the interface capturing
method the surface is determined from the volume
of liquid within the computational cell (usually
Cartesian) and an additional equation has to be
solved for this fraction (or concentration). With the
interface-tracking method however a boundaryfitted grid is used which is re-adjusted each time the
free- surface is moved (usually at each time step).

The free-surface may undergo arbitrarily large


deformations but is not able to model breaking
waves.
The purpose of this study is to develop an energy
conservative finite volume interface-tracking
method to model turbulent flow using either large
eddy or direct numerical simulation in which the
grid moves with the motion of the free-surface.
Results will be presented for a number of inviscid
and laminar flow studies that were used to verify
the correct working of the code and also for large
eddy simulation (LES) of turbulent open channel
flow.
Thomas, Leslie & Williams (1995) reported on a
method that was developed to simulate turbulent
flow in an open-channel. This technique used splitmerged cells at the free-surface and worked within
a Cartesian framework. The code, at the time, was
unique in its ability to model free-surface turbulent
flows but had the drawbacks of not being able to
model wave slopes greater than the cell aspect
ratios (hence necessitating the need for a large
number of cells in the horizontal directions) and
geometries that were not composed of rectangular
elements. One of the major objectives of this work
was therefore to develop a more robust method of
simulating free-surface wave phenomena within a
curvilinear framework. As a consequence, a new
three dimensional curvilinear free-surface code has
been developed that can handle relatively large
deformations of the surface. The adopted scheme is
based on a combination of a three stage RungeKutta (RK3) for the explicit velocity advancement
followed by a backward Euler Volume of Fluid
(VOF) step to update the cell vertices. The work

Received: 1 Nov. 2006; Revised: 23 Dec. 2006; Accepted: 3 Jan. 2007


49

Engineering Applications of Computational Fluid Mechanics Vol. 1, No. 1 (2007)

reported here shows these techniques applied to 3D


free surface problems that have allowed for the
inclusion of a sub-grid stress model. With this
extension it will be possible to carry out LargeEddy Simulations of free-surface flows in complex
geometries, around surface piercing structures (such
as bridge piers), and around floating bodies. The
technique is simple, conservative, accurate and
stable and it also does not require curviliear transformations of the grid as used by other workers
such as Hodges and Street (1999) and Komari et al
(1993).

control volume moves, the space conservation law


(SCL) which is expressed by the following relation
between the rate of change of control volume and
its surface velocity, has to be satisfied as following:
d
d + W ndS = 0 ,
(5)
s
dt

This is subtracted from equation (1) to give:

U ndS = 0,
s

In this section we briefly describe the finite volume


method (FVM) for solving the three-dimensional
flow of an incompressible viscous fluid with
moving control volume surfaces and start with the
conservation equations for mass and momentum in
their integral form:

d + (U W ) ndS = 0,

d
dt

ui d + ui (U W ) ndS

P
=
d +
x
i

g i d + ij n j dS

(6)

Equation (2) is solved in conservative form for


hexahedral control volumes using a RK3 scheme
and the pressure correction technique; the pressure
being determined by the solution of equation (6).
Free-surface cells are treated in a similar manner to
interior cells except that there is no flux across the
free-surface and the tangential stress is zero.
Surface tension effects are ignored and the pressure
in free-surface cells is determined by interpolation
between P0 and the cell below. The solution domain
has to be subdivided into a finite number of CVs
which can, in principle, be of any shape and are
defined by edges which form the grid. These grids
are initially formed by either directly specifying
their co-ordinates or by solving functional
relationships. The elevation of the free-surface is
given in terms of a single-valued height function h:
z = h( x , y ,t )
(7)
Surface movement takes place along the corner
grids that can stretch and contract with turbulent
and/or wave motion. The movement is determined
by summing the flows around columns centred at
cell vertices. These flows are determined as a
weighted average at the beginning of the time
step ( = 0 - Forward-Euler), at the end ( = 1 Backward-Euler) or at some interval in between.
For 1 an iterative procedure is adopted. The
corner vertices move in proportion to the surface
movement in order to keep the cell sizes well
distributed. In order that the grid movements are
available for the momentum equations, the verticies
are moved and all (except the surface ones) are
fixed using the velocities at the beginning of the
time step.
Using the above control volume approach,
discretised versions of the Navier-Stokes equations
are established at each nodal point. For control
volumes that are adjacent to the domain boundaries,
the general discretised equations are modified to
incorporate the appropriate boundary conditions
(wall, inflow, outflow and periodicity). The

2. GOVERNING EQUATIONS

d
dt

(1)

(2)

Here, is the fluid density, is the control


volume (CV) bounded by a closed surface S, U is
the fluid velocity vector whose Cartesian
components are ui (i = 1, 2, 3), W is the velocity of
the control volume surface, t is time, P is the
pressure, gi is the body force in the direction of the
Cartesian coordinate xi, n is the unit vector normal
are the
to S and directed outwards, and
components of the viscous stress tensor defined as:
u u j

(3)
ij = i +
x j xi

where = m + e being the molecular and


Smagorinsky eddy viscosity of the fluid
respectively. The pressure P is split into a
hydrostatic and kinematic component, Pk, as
follows:
P = { P0 g [ z t ( x , y ) z ]} + Pk
(4)
Where P0 is the surface pressure and zt (x,y) is the
vertical distance z to the free-surface. When the
50

Engineering Applications of Computational Fluid Mechanics Vol. 1, No. 1 (2007)

resulting system is then solved explicity (for the


momentum equations) and implicity (for the surface
advancement equations). At the end of the RK3
steps and before the surface is advanced, it is
necessary to impose mass continuity on the system
and this is acheived by the projection method in
which the mass flux into each cell is considered and
then solving iteratively for the pressure Pk.

is related through Equations (13 - 15) to the change


in surface elevation by
x( i , j ,k ) = ( i , j ,k )xt ( i , j ,k )
(14)
y( i , j ,k ) = ( i , j ,k )y t ( i , j ,k )
(15)
z( i , j ,k ) = ( i , j ,k )z t ( i , j ,k )
(16)
where it has been assumed that xb, yb, zb do not
change with time. For the problems solved in this
paper, however, it is sufficient for the mesh to move
only vertically. In this case, in Eq. (15) increases
monotonically with k from 0.0 at zb to 1.0 at zt and
==0. This kind of grid distribution is an
application of the coordination system of Phillips
(1957), which is a special case of the spine method
of Kistler and Scriven (1984).

3. NUMERICAL METHOD

The code has been developed mainly at Queen


Mary, University of London. It has a collocated grid
in which the pressure and all three velocities are
located at the centre of hexahedra control volumes,
the boundaries of which can initially be set to
follow a curvilinear arrangement and, for equally
sized control volumes, the special discretisation is
second order. The code can simulate flow between
solid walls or be periodic in one or more horizontal
directions.

5. NUMERICAL RESULTS

Here we report on the application of the code to a


number of tests that were chosen to test its validity.
Except for the LES tests, a neutrally stable
Forward-Euler
free-surface
advancement/
Backward-Euler velocity advancement scheme was
used. The choice of resolution for the tests was
based on both the individual problems
requirements and experience.

4. COMPUTATIONAL MESH

The free-surface is approximated by the piecewiselinear grid, as shown in 2D in Fig. 1. The time
evolution of the free-surface is therefore defined by
the motion of the corner points on the surface and,
to keep the internal mesh well distributed, it is
important that it also moves with the surface. This
internal mesh motion can be obtained by solving a
set of equations or can be algebraically related to
the surface motion and either method could be used
in the code. The relation between internal and
surface nodes are given by the following
equations:x( i , j ,k ) x b ( j ,k )
= ( i , j ,k ),
(8)
x t ( j , k ) x b ( j ,k )
y ( i , j , k ) y b ( i ,k )
= ( i , j ,k ),
(9)
y t ( i ,k ) y b ( i ,k )
z( i , j , k ) z b ( i , j )
= ( i , j ,k ),
(10)
zt ( i , j ) zb ( i , j )
where xb, yb, zb and xt, yt, zt denote the elevations of
the corners on the bottom and top boundaries.
Over the time step t = tn - tn-1, the change in
elevation of a corner point:
x( i , j ,k ) = x n ( i , j ,k ) x n 1 ( i , j ,k )
(11)

y( i , j ,k ) = y n ( i , j ,k ) y n 1 ( i , j ,k )
z( i , j ,k ) = z ( i , j ,k ) z
n

n 1

( i , j ,k )

Fig. 1. Computational grid in x and z direction. Solid


circles denote control volume centers

5.1
Wave tests
5.1.1 Tank sloshing
The transient capability of the code was tested by
predicting the evolution of a wave with an initial
shape given by 1.0 + 0.01cos((1 + x)) as it sloshed
in a tank. The length of the tank and mean depth
was = 1.0, the amplitude a = 0.01, g = 9.8 and the
viscosity and initial velocities were zero. Wall

(12)
(13)

51

Engineering Applications of Computational Fluid Mechanics Vol. 1, No. 1 (2007)

The viscosity, , and acceleration due to gravity, g,


were set to 0.02 and 2500 respectively to give
Reynolds (umaxH/) and Froude (umax/(gH)) numbers
of 1250 and 0.5. The code was able to reproduce a
parabolic velocity profile and, with an evenly
spaced resolution of 16 cubed, the computed
surface velocity was 24.78 compared with a
theoretical value of 25.0. Periodic boundary
conditions were used in both horizontal directions
and a no-slip linear/viscous boundary condition was
applied at the bed.

boundary conditions were used throughout (apart


from the free-surface). As the flow was inviscid, it
is expected that the sloshing should repeat itself
without significant decay. For a relatively course
grid where x = y = z = 1/8 and the time step t
= 0.005, the amplitude decayed by only 1.5% over
one wave period. The time dependence of the
amplitude of the fluid on the left and right walls of
the tank (denoted by a solid and a dashed line
respectively) are shown in Fig. 2.

5.2.2 Airy wave


The energy decay of a viscous Airy wave of
steepness 0.05 was investigated for varying degrees
of spatial and time resolution. The wavelength, box
length and fluid depth were each given a value of
1.0 and g was 9.81. This resulted in a wave celerity
of 1.25 with a period of 0.8. Periodic boundary
conditions were used in the horizontal directions
and, for the inviscid runs, a free-slip velocity
condition on the bed. Whilst, for the viscous runs, a
no-slip linear/viscous boundary condition was used
on the bed. For x = y = z = 1/16 and the time
step t = 0.0064 the wave Courant number was
0.32 and the changes in potential (PE), kinetic (KE)
and total energy (TE) over one wave period are
shown in Table 1. Also shown are results for a
value of viscosity of =0.001 together with the
numerical predictions of Thomas, Leslie &
Williams (1995) who used a Cartesian code with
split-merged surface cells. For comparison purposes,
it is useful to know that the theoretical total energy
decay is calculated to be 11.9% for =0.001.

Fig. 2. Sloshing in a tank after release of a wave:


elevation on the left (solid line) and right (broken line)
walls for a zero-viscosity fluid and free-slip condition on
the walls

5.1.2 Gaussian hill


In order to test the three-dimensional characteristics
of the code the decay of a small amplitude inviscid
radially symmetric wave was modelled. The wave
had an initial Gaussian profile following the
relationship 0.7 + 0.01exp((r r0)2/10), where r2 =
2
2
x y
x2 + y2 and r02 = n + n and periodic
2 2
boundary conditions were used in both horizontal
directions. Results show the wave to be perfectly
symmetrical about its centre and for the peak to be
at an approximate radial distance of 17 cells. This
compares favourably with a theoretical value of
16.9 cells (determined by the wave velocity) and
with the approximate value of 18 cells obtained
from the numerical predictions of Thomas, Leslie &
Williams (1995).

Table 1. Energy comparisons for the Airy wave

5.2.3 Trapezoidal channel


Further testing was then carried out by simulating
viscous flow in a trapezoidal channel. The channel
had dimensions of length = 6, depth = 1, base width
= 2 and the side slopes were 45 degrees. and g
were set to 0.02 and 2500 respectively to give

5.2
Viscous flow tests
5.2.1 Couette flow
The code was first run for a channel of depth H = 1,
length = 6 and width = 4 and periodic boundary
conditions were used in both horizontal directions.
52

Engineering Applications of Computational Fluid Mechanics Vol. 1, No. 1 (2007)

Reynolds (umaxH/) and Froude (umax/(gH) numbers


of 1250 and 0.5. The number of grids used in the
streamwise, spanwise and vertical directions was 64,
32 and 32 respectively with the grids in the
streamwise and vertical directions being evenly
spaced. However, in the spanwise direction, the 32
cells were evenly spaced both across the bottom and
top of the trapezium and the vertical grids were
fanned out between. A periodic boundary condition
was used in the streamwise direction and a no-slip
linear/viscous boundary condition along the walls
and bed. Fig. 3 shows the computed steady state
streamwise velocity contours, which appear to be
very satisfactorily. A steady state maximum
velocity of 21.26 was obtained which compares
with a theoretical value of 25.0 for an infinitely
wide channel. A similar problem was run by
Thomas, Leslie & Williams (1995) with the same
box size and resolution but with the sides treated as
a series of steps in the Cartesian grid. The
maximum velocity obtained from this simulation
was 22.22. This result, together with the above
mentioned velocity contour plots, indicated that the
code was working satisfactorily.

both horizontal directions. A Smagorinsky subgrid


model was used with a constant set to 0.1 and with
Van Driest damping (1956) near the bed. Cartesian
coordinates were used and the grid resolution (in
wall
units)
was
x+=32.0,
y+=21.38,
+
+
z (min)=2.67 and z (max)=9.30 for the
streamwise, spanwise and vertical directions
repectively. Thus regular spacing was used in the
horizontal directions but the vertical spacing was
such that the first grid above the bed was at the
resolution of 1/64 and the others were stretched
gradually. A second order time discretisation
scheme was used.
The mean velocity profile u+=um/ur normalized by
the wall shear velocity is shown in fig. 4. Within the
laminar sublayer z+ < 6 the profiles converge onto
the linear law u+ = z+, and beyond z+ > 30 show
logarithmic behaviour. The computed profile
closely follows the log-law u+ = k-1 ln (z+) + 5.29
(where the von Karman constant is determined as k
= 0.41) and the experimental data of Nezu & Rodi
(1986). There is, however, some evidence of a wake
region near the free surface associated with the
boundary condition du+ / dz = 0 at z+ = Re+.

Fig. 4. Mean velocity profiles: -, computed 64x64x64


mesh and subgrid model; , Nezu & Rodi (1986),
experimental data; - - -, wall laws
Fig. 3. Velocity profiles in the trapezoidal channel flow

The turbulence intensities normalized by the bed


shear velocity are shown in fig. 5. The comparisons
of the computed turbulence intensities are made
against the experimental measurements of Nishino
& Kasagi (1989) and Komori et al (1982).

5.3 Turbulent flow test


Preliminary results are given of a turbulent flow
large eddy simulation that has been carried out for
which experimental data is available. The Reynolds
number, Re+ (defined as urd/), was equal to 171
and the Froude number (umax/gd) was 0.55. The
channel has sizes of depth, d =1, length = 6, width =
2 and periodic boundary conditions were used in

53

Engineering Applications of Computational Fluid Mechanics Vol. 1, No. 1 (2007)

resolution used in this simulation is 40 x 16 x 21.


Fig. 6, provides a detailed schematic of the dune
geometry used, lengths H and L are 20mm and
6.6H, respectively. The overall length was 400mm
and the width of the computation was set equal to L.
Again a Smagorinsky subgrid model was used with
a constant set to 0.1 together with Van Driest
damping (1956) near the bed.

Fig. 5. Turbulent intensities: -, urms/ur; - - -, rms/ur; ,


Komori et al. (1982), computed Re+=176; , Nishino &
Kasagi (1989) experimental Re+=180
Fig. 6. Dune geometry

There is very good agreement between the


computed and experimental urms near the bed.
However, away from the bed, the computed results
are slightly lower than those measured. The
agreement between the computed and measured vrms
values is reasonably good but the computed are,
again, slightly lower than those measured. There is
excellent agreement amongst the wrms computed
profiles with the experimental measurements except
very near the surface. Here there is a slight upturn
in the computed velocity which is attributed to
resolution effects as a result of the vertical grid
stretching.
Near the surface, the agreement between measured
and computed values is not as good as near the bed.
Although, the measurements of vrms and wrms of
Komori et al (1982) are also not consistent with
those of Nishino & Kasagi (1989) where they
overlap in the middle depth. It is possible that the
differences may be due the different physical
boundary conditions in place, i.e. channel centerline
and free-slip surface. The effect of the free surface
is to increase the intensity of the horizontal
components of turbulence whilst reducing the
vertical component in a thin layer close to the
surface.

In the experiments of Balachandar et al. (2002) data


was taken for flow across multiple dune beds,
across 17th and 18th dune out of 22, to ensure fully
developed flow. Computationally this is achieved
by enforcing periodic boundaries in both the
streamwise and spanwise domain. The simulation
was allowed to run for 46 large eddy turnover times
(letots) to allow for the flow to become fully
developed and then statistics were taken over a
further 16 letots. A letot for this work is defined as:
L
T=
(17)
U*
where L and U* are the characteristic length and
shear velocity respectively.
The experimental Reynolds and Froude numbers
were 5.8 x 104 and 0.39 respectively, giving a
computational maximum streamwise velocity of
40.8.
In order to capture the near-wall large velocity
gradients, the cells in the vertical direction were
gradually stretched from z+(min)= z+(average)/2
using a hyperbolic-tan function. The resolution in
wall units was x+=107.6, y+=88.8 and
z+(average)=67.6.
Figs. 7 and 8 compare the mean streamwise
velocity and turbulence intensity at various
locations across the dune to measured experimental
data. Both mean and intensity data computed agree
well with available data, however the intensities
tend to show a larger discrepancy. Differences
can be attributed to the relatively large
resolution used in the simulation and to the
difficulty in accurately measuring turbulent flows.

5.4 Turbulent dune flow


Large eddy simulation of flow over a twodimensional dune has been computed and compared
with available of Balachandar et al. (2002). The
geometry and flow conditions are computed from
Yue, Lin and Patel (2005), however a coarser
resolution has been used. Yue, Lin and Patel used a
resolution of 80 x 64 x 84 (x,y,z), the current
54

Engineering Applications of Computational Fluid Mechanics Vol. 1, No. 1 (2007)

Experimental [Balachandar et al. (2002)], Computational


Fig. 7. Mean velocity profiles at various locations

Experimental [Balachandar et al. (2002)], Computational


Fig. 8. RMS plots of streamwise velocity normalised by the free-stream velocity

55

Engineering Applications of Computational Fluid Mechanics Vol. 1, No. 1 (2007)

6. CONCLUSIONS AND DISCUSSION

This paper has reported on a method of calculating


free-surface flows that utilizes boundary fitted grids.
These grids may have a curvilinear arrangement and
are free to move with the surface motion. A number
of invisid and viscous flow tests have been carried
out that show that the code is both accurate and
robust. Results have also been given of two largeeddy simulations of open channel flow and these
have been compared with previous simulations at a
similar Reynolds number and with the experimental
measurements. The agreement with published
numerical and experimental data is generally very
good and provides confidence in the codes accuracy
although it has not been tested for very steep waves.
The code has been developed with the long term aim
of simulating turbulent flow around surface piercing
structures, floating and submersible craft and work is
continuing along these lines.

9.
10.

11.

12.
13.

REFERENCES

1. Balachandar R, Polate C, Hyun KY, Lin, C-L,


Yue W, Patel VC (2002). LDV, PIV and LES
investigation of flow over a fixed dune.
Proceedings of the symposium held at Monte
Verit: Sedimentation and Sediment Transport,
kluwer academic, 171-178.
2. Demirdzic I, Peric M (1988). Space Conservation
Law in Finite Volume Calculations of Fluid Flow.
Int. J. Numer. Meth. Fluids 8:1037-1-50.
3. Eckelmann H (1974). The structure of the viscous
sublayer and the adjacent wall region in a
turbulent channel flow. J. Fluid Mech. 65:439-.
4. Farmer J, Martinelli L, Jameson A (1994). Fast
Multigrid Method for Solving Incompressible
Hydrodynamic Problems with Free Surfaces.
AIAA J. 32(6), June 1994.
5. Hodges BR, Street RL (1999). On Simulation of
Turbulent Nonlinear Free-Surface Flows. J.
Comp. Phys. 151:425-457.
6. Homori S, Nagaosa R, Murakami Y (1993).
Direct Numerical Simulation of Threedimensional Open-channel Flow with Zero-shear
Gas-Liquid Interface. Phys. Fluids A 5(1).
7. Kistler SF, Scriven LE (1984). Coating Flow
Theory by Finite Element and Asymptotic
Analysis of the Navier Stokes System. Int. J.
Numer. Meth. Fluids 4:207-229.
8. Komori S, Ueda H, Ogino F, Mizushina T (1982).
Turbulence structure and transport mechanism at
56

14.
15.

16.
17.

18.
19.

the free surface in an open channel flow. Int. J.


Heat Mass Trans. 26:513-521.
Kordulla W (1983). Efficient Computational of
Volume in Flow Predictions. AIAA J. 21.
Meselhe E, Sotiropoulos F (2000). ThreeDimensional Numerical Model with Deformable
Free-Surface for Open-Channels. IAHR J.
Hydraulic Research 38(2).
Muzaferija S, Peric M (1998). Computational of
Free-surface Flows Using Interface-tracking and
interface capturing Methods. Nonlinear Water
Wave Interaction, O. Mahrenholtz, and M.
Markiewicz, eds., Chap. 2, Computational
Mechanics Publications, Southampton.
Nezu I, Rodi W (1986). Open channel flow
measurements with a laser doppler anemometer.
J. Hyd. Engng., ASCE, 112-5:335-354.
Nishino K, Kasagi N (1989). Turbulence
statistics measurement in a two-dimensional
channel flow using a three-dimensional particle
tracking velocimeter. Seventh Symposium on
Turbulent Shear Flows. Stanford University,
2:1063-1075.
Phillips NA (1957). A Coordinate System
Heaving Some Special Advantages for Numerical
Forecasting. J. Meteorol. 14:184-185.
Th JL, Raithby GD, Stubley GD (1994).
Surface-Adaptive Finite-Volume Method For
Solving free Surface Flows. Numer. Heat
Transfer, Part B, 26:367-380.
Thomas TG, Leslie DC, Williams JJR (1995).
Free Surface Simulations Using a Conservative
3D Code. J. Comp. Phys. 116:52-68.
Thomas PD, Lombard CK (1979). Geometric
Conservation Law and Its Application to Flow
Computations on Moving Grids. AIAA J. 17(10),
1979.
Van Driest ER (1956). On turbulent flow near a
wall. J. Aero. Sci. 23:1007-1011.
Yue W, Lin C-L, Patel VC (2005). Large eddy
simulation of turbulent open-channel flow with
free surface simulated by level set method.
Physics of Fluids 17(2), Art. 025108.

You might also like