JOURNAL OF COMPUTATIONAL PHYSICS J14, !

46--159 ( 1994)

A Level Set Approach for Computing Solutions to

Incompressible Two-Phase Flow
MARK SUSSMAN,* PETER SMEREKA, t AND STANLEY OSHER I

Depart1ne11t of Marhen1atics, University of California, Los Angeles, California 90024-1555

Received July 6, 1993~ revised March I, 1994

(Lipschitz continuous) function, denoted as .p, which

A level set approach for computing solutions to incompressible two· eliminates the problems that conventional difference
phase flow is presented. The interface between the two fluids is schemes incur. This formulation also eliminates the problem
considered to be sharp and is described as the zero level set of a smooth of adding/subtracting points to a moving grid and it
function. We use a second-order projection method which implements
a second-order upwinded procedure for differencing the convection automatically takes care of merging and breaking of the
terms. A new treatment of the level set method allows us to include interface. Furthermore, the level set formulation generalizes
large density and viscosity ratios as well as surface tension. We consider well to three dimensions. The actual front location never has
to he computed. Instead, the front is embedded as a
thn rnotion of air bubbles in wator and falling water drops in air.
•C' 1994 Academic P1ess, Inc.
particular level set in a fixed domain PDE.
An application of the level set formulation was used in
[23] for compressible fluid flow. Examples from [23]
I. INTRODUCTION
include Kelvin-Helmholtz instability and Rayleigh-Taylor
A numerical method is developed for computing the instability for helium and air. The density ratio was about
motion of incompressible two-phase flow. We will consider 29 to 4 and both gases were treated as perfect gases. These
immiscible fluids where steep gradients in density and investigators found it was best to initialize <fi as the signed
viscosity exist across the interface. Instead of explicitly distance from the front, thus eliminating steep gradients
tracking the interface, we intend to implicitly "capture" the from <fi. A second-order non-conservative ENO scheme was
interface using a level set approach. The interface will be used for solving the equation for ¢.
identified as the zero level set of a smooth function. Instead of helium and air, we shall consider water and air.
As mentioned in [29], conventional conservative The density ratio of water to air is about IOOO to I and the
schemes will incur excessive numerical diffusion which will equation of state for water is not that of a perfect gas. One
destroy the sharpness of the front. High order conservative can approximate the flow of water in the compressible
schemes can produce numerical oscillations around framework if one replaces pressure P with P + B and y with
the front. The approach presented in [29] was to track the N. B and N are constants derived from the modified Tait
velocity using an Eulerian grid while explicitly tracking the equation (see, for example, [9]). For compressible flow
interface using a grid that moves through the stationary there would be a restriction put on the size of the time step
grid. The investigators cited good results, but the algorithm since the sound speed of water is about five times that of air
for tracking the front seems hard to implement. Complica- at sea level. In order to avoid this restriction, we will solve
tions occur when one needs to add or subtract points to the the problems involving water and air using the equations for
n1oving grid. "fhcse problcn1s arc :1111plific<l when solving a incon1prcssiblc now. This is a very good approximation as
three <li111cnsional problcrn. long as the nuid velocities arc rnuch srnallcr than the speed
In [27), a level set fonnulation for n1oving interfaces was of sound.
introduced. The level set function is typically a smooth Incompressible flow algorithms that have been used to
track the interface of air/water problems include vortex
methods [ l ], boundary integral methods [2], volume of
•Research supported by ARO DAA L03·01-GOl62.
t Research supported by an NSF postgraduate fellowship.
fluid methods [12, 15], front tracking methods [29], and
f Research supported in part by DARPA/ONR-NOOOJ4-92~J-1890, projection methods [ 5, I l, 18, 17]. We will use a projection
ARO DAA L03-91-GB-t62. and NSF DMS-91-03t04. method similar to [ 5]. If one combines the level set techni-

0021-9991/94 16.00 t46

Copyright© 1994 by Academic Press, Inc.
All rights or reproduction in any form reserved.
 LEVEL SET APPROACH IN TWO-PHASE FLOW 147

ques of [27] with a projection method (see [33]), one can

avoid having to explicitly track the interface (hence the term u,= -(u·V)u+g.
"front capturing"). Furthermore, one can use higher order
upwinded methods for differencing the non-linear convec-
tive terms. These methods provide robust treatment of
+!
p
(-vp + ..!..Re v. (2µD) + .!.B KJ(d)n)
convective terms at high Reynolds numbers [28, 32, 30, 14]. "'Lu - Vp/p. (6)
In this paper we shall also combine a level set technique
with a projection method. Our approach allows for large The key parameters are p .!p °" dimensionless density inside
density ratios (about I 000 to I), surface tension, and jumps the bubble; µ./µ 0 , dimensionless viscosity inside the bubble;
in viscosity, while retaining a high order of accuracy. The
Re= (2R) 312 Jg p 0 /µ 0 Reynolds number; and B =
important feature of our method is that we maintain our
4p 0 gR 2/a, Bond number. The dimensionless density and
level set function as a distance function for all time, without
viscosity outside the bubble are equal to I. We have g.
reconstructing the interface. This prevents the interface from
represent a unit gravitational force.
ever changing thickness.

2.2. Projection
2. DESCRIPTION OF ALGORITHM
From (6), we have u, =Lu -Vp/p. Let V =Lu. As noted
2.1. Equations of Motion in [ 4, ll ], if the initial value problem for Eq. (6) is well
posed, then there exists a unique decomposition (Hodge
In our study we shall consider the fluid motion for rising decomposition), where V = Vd + VI/! and Vd is divergence
air bubbles in water and falling water drops in air. We shall free. As in [ 5 ], we define a density weighted inner product
denote the density and viscosity inside the bubble (or drop) such that we can decompose V into Vdand Vl/!/p. Given our
by p 6 andµ,, respectively, and for the continuous phase by density weighted norm, we have Vd J_ Vl/J/p. Given a vector
p 0 andµc- The equations of motion are given by the incom-
pressible Navier-Stokes equations
=
V, we define our projection operator as P p(V) Vd· Since
the Hodge decomposition is unique and u, is divergence
free, we have u, = P /Lu). So, (6) and (2) are reduced to
1
u,+(u·V)u=F+-(-Vp+V ·(2µD)+ox/i(d)n) (I)
p u,=Pp(Lu). (7)
V·u=O, (2)
In order to compute the projection, we take the curl of both
sides of the equation p V = p Vd + VI/! to obtain
where u=(u, v) is the fluid velocity, p=p(x, t) is the fluid
density, µ = µ( x, t) is the fluid viscosity, D is the viscous
V x (p V) = V x (p VJ)·
stress tensor, and Fis a body force. The surface tension term
is considered to be a force concentrated on the interface. We
denote a as the surface tension, K as the curvature of the Given any divergence free vector Vd• there exists a stream
front, d as the normal distance to the front, J as the Dirac function 'I' such that Vd = V x 'I'. Furthermore, in two
delta function, and n as the unit outward normal vector at dimensions we have 'I'= (0, 0, '!'). The above equation can
the front. For immiscible liquids the density and viscosity now be written as
are constant on particle paths, therefore
-V(p V'l')=V x(pV). (8)
p,+(u·V)p=O (3)
We consider problems obeying the free-slip condition (5);
µ,+(u·V)µ=O. (4) hence '!' = 0 on the boundary.

We will assume solid wall boundaries with the free-slip

2.3. Level Set Description
condition
Since p and µ change sharply at the front, conventional
u ·0=0, (5) finite difference schemes will incur excessive numerical diffu-
sion when solving ( 3) and ( 4 ). Instead, we shall use the level
where n is the normal vector at the boundary. set technique to "capture" the interface as in [23, 33]. Our
The initial radius of the bubble (or drop) is denoted as R level set function is denoted as <fi and it is taken positive
and the only body force we consider is gravity denoted as g. outside the bubble and negative inside the bubble. There-
After the non-dimensionalization of (I) we have fore, the bubble interface is the zero level set of <fi. We
148 SUSSMAN, SMEREKA, AND OSHER

shall initialize </! to be the signed normal distance from the The above equation for p effectively gives the interface a
interface. Consider the following equation: finite thickness IX; IX is constant for all time while IX decreases
at a rate of O(h ). A justification for spreading the interface
</i,+(u·V)</i=O.
can be found on page 29 of [ 29]. Implicit in the above
formula is that q, is a distance function. This important point
This equation will move the zero level of</! exactly as the
actual bubble interface moves. Since </! is a smooth function, will be discussed in Section 2.5 below.
unlike p or µ, the above equation is more easily solved The surface tension force is represented by
numerically. Therefore Eq. (3), (4), and (7) can be written
as
u, = P ,(Lu) (9)
</!,= -(u-V)</J (10) As pointed out in [16, 8], the surface tension force can be
cast in the level set formulation and smoothed using
if </!>0
p=
{'
PdPc
(p.+pJ/(2pJ
if </!<0
if </!=0
(I I) (l/B) Ko(d)n = (l/B) K(</i) o(</i) V</i,

where the curvature is

if </!>0
µ=
{'
µb/µc
(µ. + µJ/(2µJ
if </!<0
if </i=O.
(12) K(</i)=V c;:I). (14)

If we maintain q, as a distance function, as we shall do, then

We solve the above system in the domain we may numerically approximate o( </!) by a mollified delta
function, o.(</i), smoothed in similar fashion as in [25]:
Q = { ( x, y) I 0 .:; x .:; 7R, O.:; y.:; 7R},
where u obeys the free slip condition (5) on iJQ.
o.(</i) ={~(I + cos(ir</i/1X) )/1X if It/ii< IX,
otherwise.
It should be noted that while </! is initially a distance
function it will not remain so. Furthermore, solutions of</! We denote IX as the prescribed ''thickness" of the interface. In
can develop a jump at the interface when interfaces merge. our computations we use oc = ~ L1x.
1
Below we shall present a novel way ofreinitializing </!so that
it remains a distance function. 2.4.1. Inclusion of Surface Tension into the Projection Step
2,4, Smoothing When surface tension is active, special care must be taken
when computing the right-hand side of (8 ). The contribu-
Special care must be taken when resolving the discon- tion of surface tension to V x (pV) is
tinuity in the equation for density at </! = 0 and when
computing the delta function that appears in the surface
tension term. -BI ({K(</J) o(</J)</iy)x-(K(</i) o(</J)</i,)y)). (15)
If we use ( 11) for determining p( ¢ ), the solution of the
elliptic system ( 8) will yield unwanted instabilities at the
We write o(</J) as iJH(</J)/iJ</i. It follows that we can reduce
interface; especially for large density, ratios. A method was
(15) to
proposed in [20] for solving (8) with a discontinuous p.
Unfortunately, this method would require the solution of a (16)
non-symmetric matrix and the method is difficult to imple-
ment. In order to prevent instabilities, we decided to smooth
pat the interface. We have the following equation for p(</J): The equation for H is

jj=(p.+pJ/(2pc)
2
Jp=(Pc-P•)/(2pJ

I if </!>IX
H.(</i) = 2
if </J< -IX

p(</i) = P•IPc if </J< -IX ( 13)

{ .!.2 (!+.!. sin(n</i/1X)) otherwise.
p + Jp sin(ir</i/(21X)) otherwise. IX 7!
 LEVEL SET APPROACH IN TWO-PHASE FLOW 149

a t"' 3.50 b t= 3.50 (a) l- 3.50 (b) t- 3.50

0 0
CONTOUR FROM -.5 TO .5 BY .25 CONTOUR FROM - 5 TO .5 BY .25
FIG. 3. Without surface tension, large drop should deform as it hits
FIG. 1. Level sets of a large water drop; Re= 10.0, Bd =inf., density
the base; Re = l 0.0, Bd =inf., density 1/1000, grid 50 x 100: (a) reinit; (b)
1/1000, grid 50 x JOO: (a) reinit; (b) no reinit.
no reinit.

By writing the surface tension contribution in our

drops) merge, a steep gradient in </> will arise between the
"Heaviside" formulation, we eliminate the numerical
bubbles. Also, a drop moving at constant speed will cause a
instabilities that occur when differentiating a delta function.
steep gradient to form in the distance function after a finite
amount of time. Maintaining 1> as a distance function is
2.5. Keeping cl> a Distance Function
essential for providing the interface with a width fixed in
While Eq. ( !O) will move the level set 1> = 0 at the correct time. Computation of surface tension is difficult to compute
velocity, </> will no longer be a distance function (i.e., near a steep gradient in the distance function. The values
[V</>l ¥ 1 ). </>can become irregular after some period of time for p( ~ ), especially for large density ratios, will be greatly
(see Fig. 1 and 2). For example, when two bubbles (or distorted if [V</>l is far from one. In Fig. 3, we see that

0
0
a l= 8.60 b l'"- 8.60 (a) l• 6.BO (b) \."' B.BO
CONTOUR FROM - .5 TO .5 BY 25 CON10UR ffiOM - 5 TO -5 8"f 25
F1G. 4. With large surface tension, bubble should reach e!Jipsoidal
FIG. 2. Level sets of rising bubble; Re=5.0, Bd=0.4, density 40/l, steady state; Re= 5.0, Bd = 0.4, density 40/1, grid 64 x 128: (a) reinit; (b)
grid 64 x 128: (a) reinit; (b) no reinit. no reinit.
150 SUSSMAN, SMEREKA, AND OSHER

" ~ \___)
\ / \ .14 - - - - - - - - - - - - - - - - - - - -

111 .811

0.53

t=12.0D t=12.8D t=l3.8D t=H-.40

'
0 0 0 0.23

-0.1118
32x64
64x12B

t:o 8.80 l= 9.60 l=>I0.40 t.,11.20 -0. 38 L--L--'----L--'----~-'----~-'----~~

0.0111 1 .8111 3.6111 5.40 7.20

FIG. 5. Evolution of rising bubble with large surface tension. Bubble time
reaches a steady shape and velocity; Re = 5.0, Bd = 0.4, density 40/l; grid
64 x 128. FIG. 7. Convergence test for bubble rising with large surface tension:
velocity steady state, Re= 5.0, Bd = 0.4, density 40/1.

without reinitialization, the drop loses 41 % of its mass at

time t = 3.5 while mass is conserved with reinitialization. In minor axis attain a steady state and the bubble mass is
Fig. 4, we compare results at time t = 8.8 for a gas bubble conserved (see Fig. 5, 6, 7, and 8).
rising with large surface tension and viscosity. Without Conventional routines for reinitializing a distance func-
reinitialization, the bubble loses 19 % of its mass and the tion have to explicitly find the contour r/! = 0 and reset r/! at
velocity does not attain a steady state. Furthermore, the all points close to the front. This takes O(n 3 ) operations (see
bubble shape does not attain the characteristic steady state [ 10]) and can distort the front (e.g., mass loss, corners)
of an ellipsoid. With reinitialization, the bubble velocity and depending on how one reconstructs the shape of the contour

3.70 - - - - - - - - - - - - - - - - - - - - 2.11 ~-------------------

32x64
3.40 2 .02 64x12B

•
N
·;;;
-
.,
3.22 1 . 94

~
00
•• •
"• "a
2.Q8 "' "
§
1 .85

2.74 1 -76
32x64-
64x128

. 1 .66
2.50
111.00 3.0111 0.00 q .0111 12.00 15.1110 e.00 1 . 80 3.6e 5.4e 7.20 9.00
time time

FIG. 6. Convergence test for bubble rising with large surface tension: FIG. 8. Convergence test for bubble rising with large surface tension:
mass conservation, Re= 5.0, Bd = 0.4, density 40/1. minor axis steady state, Re= 5.0, Bd = 0.4, density 40/1.
 LEVEL SET APPROACH IN TWO-PHASE FLOW 151

<P = 0. Through our experiments, we have found that one for one time step with p( <P) given by ( 13 ), µ( <P) given by ( 12 ),
needs to reinitial\ize <P after every time step in order to keep and the surface tension force given by (16 ). Denote the
the solution accurate. Thus any distortions from constantly updated </! by <P'" + 1m, and the updated u by u<H 1'.
reinitializing <P are amplified. Step 3. Construct a new distance function by solving
An iteration method for reinitializing <P was introduced by
[26]. Given a region Q+ with <P>O on Q+ and </J=O on with efi(x,O)=.P'"+ 112 '(x)
8Q +, evolve the equation ¢, = 1 -1/V</JI until <P reaches a
steady state. If <P is already close to a distance function, then to steady state. We denote the steady state solution by
one should not have to evolve too fadn time. tj>(n+ I)_
Unfortunately, one still has to prescribe boundary condi-
Step 4. We have now advanced one time step. The zero
tions on aQ+ which entails explicitly finding the interface.
level set of <t><n+ I) gives the new interface position and <t><n+ 11
We can eliminate the problem of finding the interface.
is a distance function. Repeat Steps 2 and 3.
Consider the following function </J 0 (x) whose zero level set is
the air-liquid interface; ¢ 0 ( x) need not be a distance func-
tion, however. We shall construct a function, </J(x), with the
properties that its zero level set is the same as ¢ 0( x) and that 3. DISCRETIZATION
<P is the signed normal distance to the interface. This is
A staggered mesh will be used for the velocity and the
achieved by solving the following problem to steady state
distance function. With h as the mesh size, we define
</J,=S(</Jol0-J¢;+¢;l
x,_ 1 = <U + ~Jh, u+ Dhl
(17)
</J(x, 0) = </J 0 (x), ( 18)
ui, 1 =u(xi,J)
where S is the sign function. For numerical purposes it is ¢,. 1 = efi(x;.)
useful to smooth the sign function; we do this as
i=O···M-1

S,.,,o-
(A. ) - <Po j=O···.N-1.
~ (19)
v''Po+s-
For a square box, we have Mh = 7R and Nh = 7R. For a
Equation ( 17) has the property that <P remains unchanged rectangular box, we have Mh =SR and Nh = !OR.
at the interface; therefore the zero level set of ¢ 0 and <P are
the same. Away from the interface <P will converge to 3.1. Discretization in Time
IV <PI = I. Therefore, it will converge to the actual distance. We will use a second-order Adams-Bashforth method for
The above algorithm completely avoids finding the interface evolving the equation in time (see [.J8]),
and it proves to be efficient to implement numerically. In
our computations, one iteration per time step was usually
enough for meeting our convergence criterion. In [26],
existence and uniqueness proofs are provided for the where k is the time step. A similar formula as above is used
problem: in discretizing the equation for efi.

IV</J(x)I = .<(x) in Q+
(20) 3.2. Convection Terms
</J(x) = 0 on aQ+.
We will use a second-order ENO method for the
In our case .<(x) =I. approximation of the convective terms. For u divergence
free, we have
2.6. Summary
(u · V)efi = (u<filx + (vefi)y (22)
We can now summarize our algorithm.
(u·V)u=fx+gy, (23)
Step I. Initialize </J( x, 0) to be signed normal distance to
the front. where
Step 2. Solve
(24)
</J,+u·V</J=O

581/114/1-11
152 SUSSMAN, SMEREKA, AND OSHER

For a conservative scheme, we approximate fx using We now have

f;_ 112 ,1 . Unfortunately, there are unwanted
ft+ 112 ,1 -
if uL<;;OanduR;;.o
oscillations since (23) is not always numerically accurate.
We will use an algorithm similar to that ofEq. (2.9) in [5]. if uM<;;O and uR<;;O
For the equation for ¢, we have if uM;;.O and uL;;.O.

(uefi)x + (vt/J)y 3.3, Viscous and Curvature Terms

"' ( ( u¢ ); + 11,, 1 - ( u¢) ;- 11z.1 We approximate the components of the viscous stress
tensor D using central differencing:
+ (vt/! L. 1+ 112 - ( v¢ );, 1 -112)/(2h)
(ux)1+ 1/2,J+ 1/2 ~ (u;+ l,J + Ui+ 1,J+ I - U;,j- ui,J+ I )/(2h)
~ (u,+ 112.1+u.--112.1)(¢,+112.1-tP;-11i.1)/(2h)
(uyL+ 1/2,J+ 1;2 ~ (u;+ t.J+ i - ui+ l,J + ui,J+ 1 - ui,J)/(2h ).
+ (V;.1 + 1/2 + V;,1- !/,)( t/J ;,l + 1/2 - t/J ;,1- 1p)/(2h)
+ (t/J;+ 1/2.1+t/J;-1/2.)(U;+1/2.1- U;- l/2.)/(2h) Similar equations are used for vx and Vy.
The divergence of the stress tensor is computed as
+ (q, ;,1+1/2 + q, 1.1- 112)( v ;,1 + 1/2 - v ;,1- 112)/(2h) follows:
((µDrn•nJxl;.l
For smooth data, we have r/J; + 1;2. 1 +</>i-112.J ~ 4> i,J + 112 +
rfiu_ 112 . Since u is numerically divergence free, we have ~ ((µDm·'.!L+ 1/2,j+ 1/2 + (µDm• )i+
11
l/2,j-1/2

u;+ 1; 2 , 1 -u;_ 112 •1 ';::;', -(vi,J+t;2 -v;, 1 _ 112 ). We find the -(µDrn·n); _ 1/2.1+1/2 - (µDrn·n); _ 1/2.1- l/2)/(2h)
following approximation
((µDm·n)y);.l
(ut/J)x + (vt/J)y ~ ( (µDm· );+ 1/2,j+
11
l/2 - (µDm• 11 );+ 1/2,j-1/2
"' ( U; + 1/2.1 + U; _ 1/2,)( </!; + 1/2.1 - </!;- 112)/(2h) +(µDrn·n); _ 1/12.1+1/2 - (µDrn·n); _ l/2.1-1/2)/(2h)
+ (v ;,1 + 1/2 + v ;,1- 112)( ¢ ;,1+1/2 - ¢ ;,1- 112)/(2h ). with

Similarly we have
We use the free-slip condition (5) to determine the
discretization at the left boundary. For example,
~ (u; + 112:1 + U;-112.1)(u; + I/2,J- U;-1;2.1)/(2h)
(ux)_ 112.1+ 112 = 2(uo.1 + Uo. 1+ i)/(2h)
+ (v i,J + 1;2 + v i,J- 112)( U;,J + 112 - ui,J- 1/2)/(2h)
µ-1/2,1+1/2 = (µ0,l + µ0.1+ I )/2.
(/2)x + (g,)y Similar equations are used for the other boundaries. The
"'(u; + 112.1+U;-112.)(v;+112. 1 - V;-1;2)/(2h) curvature is discretized in the same fashion as the discretiza-
tion of the divergence of the viscous stress tensor.
+ (v;,1+1/2 + v1.1- 112)( v ;,1+1/2 - v;,1-112)/(2h ).
3.4, Discretization of the Projection
For computing U;+ 112 , 1 (similarly for u 1,1 + 112 , 4' 1+ 112 , 1 , ••• ),
Given V=Lun, we decompose V into the form Vd+
we use a second-order ENO scheme (see [28, 23]):
Vi/J/p, where Vdis divergence free and define P pn(Lun) o= Vd
Define
and Vpn =Vi/I. Following [5 ], in order to define the discrete
approximation of the projection, we must first define
if lal <;; lbl
m(a, b) = {~ otherwise.
discrete divergence and gradient operators and a discrete
p-weighted inner product.
For divergence we have
Let
(V · VL+112.1+112

UL= ui,j + !m(U;+ l,j- ui,j• U;,;- U1-1.1) ~ (DU)i+ 112.1+ 1;2
uR=U·+1
l
1
, ).--, m(U+2
l , ].-U·+1.
I , ) ' U·+1
I , ].-u 1,)
. .) =(ui+ l,J+ I - ui,J+ I + Ui+ l,j- U;,1)/(2h)
+(Vi+ 1,j+ I - V;+ l,j +Vi,}+ 1 - V;,j)j(2h ).
 LEVEL SET APPROACH IN TWO-PHASE FLOW 153

For the gradient we have (PCG) algorithm using an incomplete Cholesky factoriza-
tion as a preconditioner. The initial data for the PCG
( V<P )1.1 "°' (G<P )1.1 = (( G x <P )1.1, ( Gy <P )1.) algorithm is a linear combination of the results of previous
( G x<P);,J =(<P; + 112.1+ 112 - <Pi-112.1+ 112
time steps.

+ <PI+ 1/2,j- 1/2 - <PI- 1/2,j- 11il/(2h) 3.5. The Reinitialization of¢
( G y<P)1.1 =(<Pi+ 112.1+ 112 - <Pi+ 112.1- 112 In this section, we describe how to numerically evolve
+ <PI - 1/2,j+ 1/2 - <PI- l/2,j-1/2)/(2h) (17) to steady state. We can write (17) in the form

For the inner product we have </!, + w · V</J = S(</! 0 ), (28)

M-1 N-1 where

(V1, V,)P= L L (V1.u· V2.ulPu w = S(</J 0 )(V¢/IV¢1 ).
i=O j=O

Note. The divergence operator and <Pare defined at the Equation (28) is a nonlinear hyperbolic equation whose
cell comers X;+ 112.i+ I/2• where i = -1 · · · M - I and j = characteristics are given by w. The vector w is a unit normal
-1 .. ·N-l. always pointing outward from the zero level set (</! = 0 ).
With the above definitions for D and G, the discrete One possible discretization of ( 17) is as follows. We define
operators are skew adjoint (G= -DT; see [5]). Using our
definitions of G, D, and (.,. )p, discretely divergence free a= D; </1 1.1 = ( ¢ 1.j- <Pi-1.)/h
vector fields with zero normal components are orthogonal
to discrete vector fields of the form G<P/p. Consequently, we
b =n: </11.j = (¢1+ 1.1- ¢1.)/h
can uniquely decompose any discrete vector field into c =D; ¢ 1.1 = (</!1,j- </!1.1-1 )/h
U + G<P/p, where DU= 0.
d =D; ¢1.1 = (</11.j+ 1'1.;l/h
I -
In two dimensions, a divergence free vector can be written
as the curl of a vector 'I'= (0, 0, 'I') (see pp. 77-78 of [3] ). and
Define a discrete function 'I' at the cell corners x 1 + 112 .1+ 112 .
Since our divergence free vector field has zero normal
component at the boundary, 'I'= 0 at the boundary ( i = -1,
jJ¢~ 1 + e2
S,( ¢ )1•1 = </! 1

M - 1 orj= -1, N - 1 ). Define G.LW as the discrete curl of Jmax( (a+ ) 2 , (b-) 2 ) + max((c+ )2 , (d-) 2 ) - I
W; e.g., G.L 'I'= ( GY 'I', -G, '!'). Then we have if ¢7,j>O
G(<fi),. 1 = Jmax((a-) 2, (b+) 2 ) + max((c-) 2, (d+ )2 )-1
(25)
if ¢7,j<O
GJ_(p GJ_ 'I')= GJ_(p V) (26)
0 otherwise,
-G,(p(Gx 'l'))-Gy(p(Gy '!')) = G,(pV2 )-G/pV 1 ). (27)
where the + superscript denotes the positive part and the -
When surface tension is active, we modify the right-hand superscript denotes the negative part. Equation ( 17) is then
side of (27) using our Heaviside formulation (16). The updated using
difference formula for (27) is
(29)
- [ Pi,j tpi- 1/2,j- 1/2 + p i,j +I 'P; _ 1/2.J + 3/2
It is shown in [26) that (29) is a consistent, monotone
+Pi+ 1,j 'P;+ 3/2,j- l/2 +Pi+ 1,j+ I 'P;+ 3j2,J+ 3/2
scheme of (20), which is known to converge to the unique
- (Pi,)+ Pi+ I,}+ Pt,)+ I +Pi+ l,j+ I) tpi+ l/2,j+ 1;2] viscosity solution of (20). If q, 0 > 0 and G( ¢ 0 ) < 0, then [26]
proved that ¢ /' implies G( ¢) /'. An extension of this can
= G,(p V2 ) - G,(pV i)
easily be deduced for q, 0 < 0. If ¢ 0 < 0 and G(¢ 0 ) < 0, then
q, > implies G( ¢) /'. So, if G( q, 0 ) "'0 and if Lit is sufficiently
-9l (GXK(ljJ) GyH - GyK(ljJ) GXH), small, then G(</JN) /' 0 as N - oo. This follows from the fact
that as long as G( </JN) < 0, then </JN"' q,N + 1 for q,N ';3' 0 and
where V is V with the surface tension terms excluded. q,N ';i' </IN+ I for q,N"' 0.
Once 'P is known, we can set U G_i_ 'P. The matrix = While (29) has the advantage of being a monotone
system is solved using a preconditioned conjugate gradient scheme, it has the disadvantage of being only first order. We
154 SUSSMAN, SMEREKA, AND OSHER

observed a significant improvement in our results by

approximating the derivatives of¢ by higher order schemes.
In our tests, we used a second-order ENO scheme upwinded
in the proper direction (see [ 28] ). As mentioned previously,
we typically needed only one iteration per time step in order
to meet our convergence criterion. We have also done tests
A A 0
(a) t= 4.40 (b) t= 4.40 (c) t"' 4.40
where¢ was initialized as + I outside of a unit circle and - I
inside of a unit circle and our iteration converged success- FIG. 9. Convergence test for bubble rising with medium range
fully. Reynolds number and small surface tension; Re= 100.0, Bd = 200.0,
The stopping criterion for the iteration is density 1000/1; (a) 36 x 36; (b) 72 x 72; (c) 144 x 144.

number and small surface tension. The second problem is a

bubble rising with low Reynolds number and high surface
tension. We let the second bubble rise to a steady state.
where M =number of grid points where 1¢~1 I <a. In our For the first test, we have B= 200.0, Re= 100.0, p,/Pb =
experiments we had At= h/IO and we used s = h in the 1000.0, and µJµb = 100.0. Grids of 36 x 36, 72 x 72, and
expression for Se. 144 x 144 are used. The number of time steps for each of the
respective grids is 734, 1467, and 2934. We compare results
3.5.1. Computation of At up to time t = 4.4 (see Fig. 9 and 10). We compute the
The time step must obey the CFL conditions due to the relative error between successive grids as
convective terms. Restrictions due to stiff source terms (e.g., 4.4
gravity and surface tension), and due to viscous terms must Eh.2h= I l/.-/2.I·
also be satisfied (see (6.3) of [23] and (61) or [8]): t=O.O

At,= j(p, + p.)B/8n h 312 Table I contains values for the position and velocity of the
center of the bubble. Table I also contains relative errors for
At,= mJn (i~ (p(Re)h 2/µ)) minor axis size and area. Figure 11 compares the position of
the center of the bubble.

At,=mJn (i~ 1 )

~ ~
~
0 0 0 0 0 0
4. ANALYSIS OF RESULTS
l= 5.20 t= 5.60 t= 6.00

Our experiments simulate the flow of air bubbles and

water drops. We will assume that the flow is two-dimen-
sional and symmetric about the axis x = 0. The key
parameters are density ratio Pc/Pb, viscosity ratio µc/µb,
Bond number B = 4p, gR 2/a, and Reynolds number Re=
(2R) 3i 2 ~ p,/µ" For the bubble problems R = 2.5 cm and
n l= 4.00
0
t= 4.40
~
·D (',,•

l= 4.80
for the drop problems R = 0.125 cm. Using the tables
in [ 3], we have Pa"= 1.226 x 10- 3 g/cm', Pw'"' =
l.OOOg/cm', µ,"=l.78xl0- 4 g/(cms), µw.,"=l.137x

~
10- 2 g/(cm s), g= 980 cm/s 2 , and a= 72.8 dynes/cm. These
physical constants are used in determining the above
parameters. c?D (())
t= 2.80 l= 3.20 t= 3.60
4.1. Convergence Study
FIG. 10. Evolution of rising bubble with medium range Reynolds
We consider two rising bubble problems. The first number and small surface tension; Re= 100. Bd = 200, density 1000/I, grid
problem is a bubble rising with medium range Reynolds 144x 144.
 LEVEL SET APPROACH IN TWO-PHASE FLOW 155

TABLE I
Convergence Study: t = 4.4

E1i.21i E21i,41i Order

Position O.oJ 0.05 2.3

Velocity 0.02 O.oJ 1.8
Area O.G2 0.16 3.0
Minor axis 0.04 0.12 1.6

(a)

For the second test, we have B = 0.40, Re= 5.0, FIG. 12. Surface tension effects on bubble rising with medium range
Reynolds number; Re=lDO, density 1000/1, grid 72x72: (a) Bond
pJph = 40.0, and µJµ 6 = 500. Grids of 16 x 32,32 x 64, and number 200.0; (b) Bond number 25.0.
64 x 128 are used (see Fig. 5). The bubble reaches a steady
where its oblate shape is the expected result (see [29],
p. 31 ). The area, velocity, and minor axis are each plotted
with respect to time (see Fig. 6, 7, and 8). 4.2.2. Effects of Viscous Terms

We compare the results of Fig. 13 (Re= 10.0), Fig. 10

4.2. Analysis of Air Bubble Problem
(Re= 100.0), and Fig. 14 (Re= 1000.0). The change in
shape due to low Reynolds number is similar to that found
4.2.1. Effects of Surface Tension
in the experiments of [ 7] (see Fig. 3 of [ 7] ). The results for
We study the effects of adding surface tension to the rising high Reynolds number flow compares well to the results in
bubble problem where Re = 100.0. In Fig. 12, we compare [2]. In [2], the authors computed the solution until pinch
the results with Bond number of 200.0 (a) to that of off (t = 4.0) of an inviscid gas bubble. They used both the
Bd = 25.0 (b ). The change in shape is similar to that found boundary integral method and the point vortex method and
in [ 21] (experiments using the boundary integral method), obtained almost identical results using those methods. The
where the bubble skirts become thinner and the indentation fact that our results using an Eulerian grid are also very
at the bottom flattens out. close to the above Lagrangian schemes validates our code
for high Reynolds number and for steep density ratios.
3.13 ~~-~-~-~-~-,----~~-~--,,

2.50
A A A
t= 4.BO t= 5.20 l= 5.60
1 .88

1 .25
A A A
t= 3.60 t= -4.00 t= 4.40
0.63
--36x36
---72x72
- - - - 144xi44
0 .00 l-0-:,_
0.00
_L_~-~-~-L-~~~-~-_J

0.88 1 . 76
time
2. 64 3.52 4. 40 GJ c=:J A
t= 2.40 t= 2.80 l= 3.20

FIG. 11. Convergence test for bubble rising with medium range
Reynolds number and small surface tension: position, Re= 100.0, FIG. 13. Evolution ofrising bubble with low Reynolds number and no
Bd = 200.0, density 1000/1. surface tension; Re= 10.0, Bd =inf, density 1000/1, grid 72 x 72.
156 SUSSMAN, SMEREKA, AND OSHER

c=::i
'.o o:
~

·o o· 0
~

0 n n n
t= 5.20 t= 5.60 t= 6.00 l= 6.00 t= 6.60 t= 7.20

~
t= 4.00
0
t= 4.40
0

·0
~
0
~

t= 4.80
fi
t= 4.20
n t= 4.80
n l= 5.40

cru t"' 2.60

(f1)
t= 3.20
~
t= 3.60
w l= 2.40
~
t= 3.00
~
t= 3.60

FIG. 14. Evolution of rising bubble with high Reynolds number and FIG. 16. Evolution of rising bubble with slight density ratio and
low surface tension; Re= 1000, Bd = 200, density 1000/1, grid 140 x 140. medium range Reynolds number; Re= 100, Bd =inf, density 1.01/1, grid
140 x 140.

n n
1. 78 ~-~-~-~-~-~-~~-~-~-~-~

~
~ d - - - density 1000/1
1 . 31 - - - density 5/1
t= 5.20 t"' 5.60 t= 6.00 ----density 1.1/1

~
t= 1.00
~
t= 4.-'l-O
n t= 4.80
' 83

0. 36

-0. 12

cr=u (fl) ~
t= 2.60 t= 3.20 t= 3.60 -0.5'1 ~-~-~-~-~-~-~~-~-~-~-~

0.00 1 .2b 2.52 3.78 5.04 6.30

time
FIG. 15. Evolution of rising bubble with low density ratio and
medium range Reynolds number; Re= 100, Bd =inf, density 5/1, grid FIG. 17. Effects of density ratio on bubble rise velocity; Re= 100,
72 x 72. Bd =inf, grid 50 x 100.
 LEVEL SET APPROACH IN TWO-PHASE FLOW 157

l= 3.50 l= 4.00 l= 4.50 l= 5.00

1:'1.03

0.01

0 ~ I I

-0.01 t= 1.50 t= 2.00 l= 2.50 l= 3.00

-0.03 0 0 0 0
-0.05
0.00 1 .83 3. 67 5.50 7.33 9. 17
lime

FIG. 18. Amplitudinal oscillations of a two-dimensional drop driven

FIG. 20. Evolution of a water drop with surface tension. Drop remains
by surface tension forces. Amplitude decays exponentially due to viscosity;
circular as it hits the base; Re= 10, Bd = 1/800, density 1/1000, grid
Re~ 20, Bd ~ 1/2, density 1/100, h ~ 0.07.
50 x 100.

l= 3.50 l= 4.00 l= 4.50 l= 5.00

5.50

3.50

0 ~
t= 1.50 l= 2.00 t"' 2.50 t= 3.00

1 .50

0 0 0 -0.50

0
-2.50
0.00 0.70 1 . 40 2. 10 2.80 3.50
time
FIG. 19. Evolution of a large water drop (no surface tension). Drop
defonns as it hits the base; Re= IO, Bd =inf, density 1/1000, grid 50 x 100. FIG. 21. Position of water drop with surface tension versus time;
Re~ 10, Bd ~ 1/800, density 1/1000, grid 50 x 100.
158 SUSSMAN, SMEREKA, AND OSHER

density-ratio bubble and 5.6 for the low density ratio

bubble. We also compare the above results to that of
Fig. 16. For slight density ratios, the skirts of the bubble
begin to roll up (see [ l] ). In Fig. 17, we plot the velocity of
the bubble for different density ratios.

4.3. Analysis of Water Drop Problems

We consider three problems in 2D flow. The first problem
is a stationary drop that is driven by surface tension. The
second problem is a drop that is allowed to fall from rest
and hit the bottom of our closed box. The last problem
involves two drops which are propelled towards each other
and allowed to collide.
For the first problem, we compute the frequency of
oscillations due to surface tension driven flow (see Fig. 18 ).
We compare our results to that which is predicted by [ 19],
p. 472, (Eq. (8)). We have Re=20.0, B=0.5, pJp.=0.01,
µJµb = 0.01, a= 1.0, A= 0.06, s = 2, and Ax= 0.07. "A" is
the amplitude of the initial perturbation, "s" is the distur-
bance type, and "a" is the base radius of the drop. Our
FIG. 22. Evolution of two water drops colliding with each other.
computed period of 1.92 agrees closely with the result
Combined drop experiences surface tension driven oscillations; Re= 20,
Bd = 2.0, density 1/14, grid 44 x 44. predicted by the formula 2n/a ( 1.81 ), where a 2 = 6/B.
For the second problem, we study the flow of a falling
water drop. We refer the reader to Figs. 19 and 20. When the
4.2.3. Effects of Different Density Ratios required surface tension is added, we obtain the expected
circular shape. We compare our results (R=0.125 cm, B=
We compare the results of Fig. 10 (pJp. = 1000.0) to that
0.00125, Re= 10.0) to that of [22] (R = 0.140 cm). Our
of Fig. 15 (pJp. = 5.0). The shape of the bubbles are
results are similar to those of [ 22]. The added surface
similar, except that the skirts on the low density-ratio
tension prevents the drop from "flattening out" near the
bubble are allowed to grow longer than those on the high
lower boundary. Figure 21 displays the position of the drop
density-ratio bubble. The pinch-off time is 4.5 for the high
versus time. The average acceleration was 0.95 (5 % error).
For the last problem, we study the head-on collision of
two water drops in the absence of graviational forces (see
[ 24, 13] ). We refer the reader to Fig. 22. The two drops are
each accelerated at each other with a body force of0.5. After
6.50 time t = 2.0, when the drops are traveling with a non-dimen-
sional velocity of about 1.0, the force is turned off. Mass is
conserved throughout the collision (see Fig. 23 ). After the
collision, the combined drop undergoes oscillation due to
6.00
the surface tension forces (Re= 20.0, Bd = 2.0 ). As with
•
0 bubble pinch-off above, our level set formulation enables us
•" to merge the two drops without any extra programming .
5.50

5. CONCLUSIONS

5.00 In summary, we have designed a formally second-order

accurate algorithm for tracking the interface of two incom-
pressible fluids. The interface remains sharp ( p J p, = 1000)
4.50 ~~-~-~-~-~-~-~~-~-~
without ever having to explicitly find the front. As in [29],
0.00 3.00 6.00 9.00 12.00 15.00 the front is given a finite thickness 0( h) which does not
time change in time; hence there is no added numerical diffusion.
FIG. 23. Conservation of mass for colliding drop problem; Re= 20, Since we solve a PDE fort/> (ensured to be a smooth distance
Bd = 2.0, density 1/14, grid 44 x 44. function for all time) instead of for p, relatively coarse grids
 LEVEL SET APPROACH IN TWO-PHASE FLOW 159

can be used. The algorithm is easy to code since the initial 9. T. J. Chen and C. H. Cooke, On the Riemann Problem for Liquid or
Eulerian grid remains the same throughout simulation. Gas/Liquid Media, Dept. of Mathematics and Statistics, Old Dominion
There is no extra code needed for handling merging, break- University, 1991 (unpublished).
up, dilation, or contraction of the interface. Since the algo- JO. D. L. Chopp, Ph.D. thesis, Lawrence Berkeley Laboratory and
Department of Mathematics, University of California Berkeley, 1991
rithm does not have to explicitly find the interface, the code (unpublished).
can be easily generalized to three dimensions. Furthermore, 11. A. J. Chorin, Math. Comput. 22, 745 (1968).
surface tension is incorporated as a body force term which 12. E. Fatemi, J. Compur. Phys. 108, 209 ( 1993).
is easy to compute. Because of the special treatment of the I 3. J. Fukai, Z. Zhao, D. Poulikakos, C. M. Megaridis, and 0. Miyatake,
convective terms, the algorithm can accurately handle high Phys. Fluids A 5 ( 11 ), 2588 (1993).
Reynolds number flow. In the future, we would like to 14. S. K. Godunov, Mar, Sb. 47, 271 (1959) [Russian]; USJPRS Transl.
generalize the algorithm to handle axisymmetric flow, for 7226 (1960).
computations involving spherical bubbles as well as cylin- 15. C. W. Hirt and B. D. Nichols, J. Comput. Phys. 39, 201 ( 1981 ).
drical bubbles. Furthermore, we would like to simulate fully 16. Y. C. Chang, T. Y. Hou, B. Merriman. and S. Osher, preprint, 1994.
three-dimensional problems involving many bubbles and 17. J. Van Kan, SIAM J. Sci. Statist. Comput. 1, 870 ( 1986).
drops which can interact with each other. Finally, we would 18. J. Kim and P. Moin, J. Compur. Phys. 59, 308 (1985).
like to simulate flow involving several ( > 2) immiscible 19. H. Lamb, Hydrodynamics (Dover, New York, 1945).

fluids using the work of [ 6]. 20. R. J. LeVeque and Z. Li, Technical Report 92-12, Dept of Appl. Math.,
University of Washington, 1991 (unpublished).
21. T. S. Lundgren and N. N. Mansour, J. Fluid Mech. 224, 177 (1991).
ACKNOWLEDGMENT 22. J.E. McDonald, Sd Am. 19Q, 18, 64 (1954).
23. W. Mulder, S. Osher, and J. A. Sethian, J. Compur. Phys. 100, 209
We thank J. M. Morel for bringing [ 26] to our attention. (1992).
24. M. R. Nobari, Y. J. Jan, and G. Tryggvason, NASA Technical
Memorandum 106394, 1993 (unpublished).
REFERENCES 25. C. S. Peskin, J. Comput. Phys. 25, 220 ( 1977).
26. E. Rouy and A. Tourin, SIAM J. Numer. An'al. 29 (3), 867 (1992).
I. C.R. Anderson, J. Comput. Phys. 61, 3 (1985). 27. S. Osher and J. A. Sethian, J. Comput. Phys. 19 (I), 12 (1988).
2. G. R. Baker and D. W. Moore, Phys. Fluids A 1 {9), (1989). 28. C. W. Shu and S. Osher, J. Comput, Phys. 83, 32 ( 1989).
3. G. K., Batchelor, An Introduction to Fluid Dynamics, Cambridge Univ. 29. S. 0. Unverdi and G. Tryggvason, J. Comput. Phys. 100, 25 (1992).
Press, Cambridge, UK, 1967.
30. B. Van Leer, "Multidimensional Explicit Difference Schemes for
4. J.B. Bell, P. Colella, and H. M. Glaz, J. Comput. Phys. 85. 257 (1989). Hyperbolic Conservation Laws," in Computing Methods in Applied
5. J.B. Bell and D. L. Marcus, J. Comput. Phys. 101. 334 ( 1992). Sciences and Engineering, Vol. VI, p. 493 (Elsevier Science,
6. J. Bence, UCLA, Ph.D. thesis, June 1993 (unpublished). Amsterdam, 1984 ).
7. D. Bhaga and M. E. Weber, J. Fluid Mech. 105, 61 (1981). 31. J. K. Walters and J. F. Davidson, J. Fluid Mech. 12, 408 ( 1962).
8. J. U. Brackbill, D. B. Kothe, and C. Zemach, J. Comput. Phys. 100, 335 32. P. Woodward and P. Colella, J. Comput. Phys. 54, 115 ( 1984 ).
(1992). 33. J. Zhu and J. Sethian, J. Comput. Phys. 102, 128 (1992).