JOURNAL OF COMPUTATIONAL PHYSICS lo&2537 (1992)

A Front-Tracking Method for Viscous, Incompressible,

Multi-fluid Flows
SALIH OZEN UNVERDI AND G&TAR TRYGGVASON
Department of Mechanical Engineering and Applied Mechanics, The University of Michigan, Ann Arbor, Michigan 48109

Received July 5, 1990; revised May 10, 1991

Basically, two strategies exist to deal with flows con-

A method to simulate unsteady multi-fluid flows in which a sharp
taining sharp fronts. In front-capturing, the basic approach
interface or a front separates incompressible fluids of different density is to use a finite volume discretization of the whole domain,
and viscosity is described. The flow field is discretized by a conservative while modifying the numerical approximation to minimize
finite difference approximation on a stationary grid, and the interface is numerical difficulties. This generally involves using a high
explicitly represented by a separate, unstructured grid that moves
(usually second) order scheme and adding artificial
through the stationary grid. Since the interface deforms continuously,
it is necessary to restructure its grid as the calculations proceed. In addi- viscosity around the front to diffuse it slightly to avoid
tion to keeping the density and viscosity stratification sharp, the tracked oscillations. The fundamental idea goes back to von
interface provides a natural way to include surface tension effects. Both Neumann and Richtmayer. Since then shock-capturing
two- and three-dimensional, full numerical simulations of bubble has become considerably more sophisticated because of
motion are presented. 0 1992 Academic Press. Inc
the introduction of schemes that explicitly enforce
monotonicity through a nonlinear step and at the sametime
maintain high order “almost everywhere.” These methods
1. INTRODUCTION (see, e.g., Boris [7] for a review) generally serve well for
shocks but work lesswell for contact discontinuity. Further-
Accurate simulation of fluid flow with a sharp front more, relatively fine grids are still needed,thus limiting most
presents a problem of considerable difficulty which has unsteady flow simulations to two dimensions.
challenged inventors and users of numerical methods since An alternative to front-capturing is tracking, where addi-
the beginning of large-scale computational work. Fronts or tional computational elements are introduced explicitly to
interfaces occur commonly in fluid mechanics problems. keep track of the front. This generally reduces by a
Shocks in compressible flows, vortex sheets in inviscid considerable amount the resolution needed to keep the
flows, and boundaries between immiscible fluids are some front sharp, and eliminates numerical diffusion altogether.
of the better known examples. In principle, all of these However, tracking is best suited for well-defined fronts that
problems can be handled numerically by working with the are easily identifiable in the initial conditions.
governing equations in integral form, which lead to conser- Tracking methods can be divided further into volume-
vative finite difference schemesand thus avoid any assump- tracking and front-tracking. In volume-tracking methods
tions about the differentiability of the solution. In practice, a marker, used to reconstruct the interface, is advected
however, straightforward application of a conservative with the flow. In the well-known marker-and-cell (MAC)
schemeis severely limited when a sharp interface is present. method of Harlow and Welch [25], marker particles were
If the numerical method is of a low order, excessivenumeri- inserted to identify the spatial region occupied by a single
cal diffusion will quickly destroy the sharpness of the front; fluid with a free surface. Daly [15] relined this method to
a higher order scheme will lead to numerical oscillations allow it to deal with two fluid problems. Other extensions
around the front that may couple into other parts of the and refinements can be found in Daly [16] and Amsden
solution in an undesirable way. Here, we present a method and Harlow [3], for example. Hirt and Nichols [26]
for incompressible flows of two or more immiscible fluids abandoned the use of marker particles and instead used a
where the interface is explicitly tracked to keep the density marker function advected by the flow. At each time step
and viscosity fields sharp and to allow the inclusion of sur- they reconstructed the front from the marker function by
face tension forces at the interface. The method has been generating either a horizontal or a vertical interface within
implemented for both two- and three-dimensional flows. each interface cell. This interface was then advected with
25 0021-9991/92 $5.00
Copyright 0 1992 by Academic Press, Inc.
All rights of reproduction in any form reserved.
26 UNVERDI AND TRYGGVASON

its normal velocity to update the marker function. Later Several specialized methods for flows that can be treated
developments include the simple line interface calculations either as inviscid, or that are so viscous that inertial force5
(SLIC) of Noh and Woodward [33] and refinements by may be neglected (Stokes flows), are properly classified as
Chorin [lo] and others. front-tracking methods. In these cases the governing equa-
Volume-tracking methods enjoy considerable popularity tions can be reformulated as an integral equation over the
because they are reasonably accurate and relatively simple moving boundaries and interfaces so the only grid is the
to implement. For higher accuracy at the cost of con- front grid. These methods are generally referred to as
siderably more complexity, it is necessary to use front- boundary integral, or boundary element methods, but vor-
tracking, where the interface itself is described by additional tex methods for generalized vortex sheets also belong to this
computational elements. The basic idea is discussed by class. Although it is usually highly desirable to reformulate
Richtmyer and Morton [39], but its primary implementa- a problem as an integral equation, occasionally it is more
tion has been through the work of J. Glimm and coworkers efficient to retain an Eulerian grid for part of the calculation,
(see, e.g., Glimm, Grove, Lindquist, McBryan, and The well-known vortex-in-cell (VIC) method is such a
Tryggvason [24]). They represent the moving front by a hybrid, where the vorticity is carried by distinct Lagrangian
connected set of points, which forms a moving internal particles, but a stationary Eulerian grid is used to find the
boundary. To calculate the evolution inside the fluid in the velocities. In its original form the VIC method was
vicinity of the interface, an irregular grid is constructed and developed for homogeneous fluids, but extensions to weakly
a special finite difference stencil is used on these irregular stratified flows were devised by Meng and Thomson [30]
grids. Several problems have been addressed by this and by Tryggvason [47] for arbitrary stratification.
method, notably compressible flows of stratified fluids (e.g., In this article we present a front-tracking method for
Chern, Glimm, McBryan, Plohr, and Yaniv [S]) and the incompressible, viscous, multi-fluid flows. While the inter-
motion of saturation fronts in porous medium (e.g., Daripa, face is explicitly tracked, it is not kept completely sharp but
Glimm, Lindquist, Maesumi, and McBryan [ 171). For is rather given a finite thickness of the order of the mesh size
aerodynamics calculations with another front-tracking to provide stability and smoothness. This thickness remains
method see the review by Moretti [31]. A front-tracking constant for all time (no “numerical diffusion”) but
technique based on a somewhat different idea than those decreases with liner resolution of the stationary grid. Since
described above has been developed by Peskin and the fluids are incompressible, the interface simply moves
collaborators (see e.g., Peskin, [35]; Fauci and Peskin, with the fluid velocity, which is interpolated from the grid.
[19]; Fogelson and Peskin, [20]). In this work, which is The method incorporates some of the features of volume-
limited to incompressible and homogeneous fluids, the tracking and shock-capturing schemes in that the original
interface is again represented by a connected set of particles underlying grid is retained through the simulations and no
which carry forces, either imposed externally or adjusted in restructuring is needed to align the grid with the interface.
such a way as to achieve a specific velocity at the interface. However, the interface is explicitly tracked, as in the front-
In this case, however, the fixed grid is kept unchanged, tracking method of Glimm and coworkers (e.g., Glimm et
even near the interface, and the interface force is simply al. [23]), and like methods based on moving grids, such as
distributed onto the fixed grid. the one discussed in Fritts and Boris [21] and Fyfe, Oran,
The major drawback of direct front-tracking is its com- and Fritts [22]. When the method is described in detail, it
plexity. In addition to the obvious question of how the inter- becomes clear that it is actually most related to the VTC
face grid interacts with the stationary grid, and vice versa, it method and Peskin’s method. Its major differences are that
is generally necessary to restructure the interface grid the tracked interfaces carry the jump (or gradient) in
dynamically as the calculations proceed. Computational properties across the interface and that, at each time step,
points must be added in regions where the grid stretches, the property field is reconstructed by solving a Poisson
and usually it is desirable to eliminate points from regions equation (using a fast solver). The primary advantage of
of compression. Additional complications, to be discussed this approach is that interfaces can interact in a rather
later, arise in three dimensions. Another major problem for natural way, since the gradients simply’add or cancel as the
front-tracking results from the interaction of a front with grid distribution is constructed from the information carried
another front (or another part of the same front). Generally, by the tracked front. This interaction-automatically taken
the computational procedure does not recognize more than care of in our method-is considered one of the great dif-
one front in each cell of the stationary grid, and therefore ficulties of front-tracking methods and one of the reasons
double interfaces have to be merged into one interface that volume-tracking methods have been favored in the past
or eliminated. A merging algorithm that has been used (see, e.g., Hirt and Nichols, [26]; Hyman, [27]; Oran and
with some success in two-dimensional simulations is Boris, [34]; Floryan and Rasmussen, [IS]).
described by Glimm, Grove, Lindquist, McBryan, and The rest of the paper is arranged as follows: A short over-
Tryggvason [ 241. view of previous numerical calculations of bubble motion
 VISCOUS, MULTI-FLUID FLOWS 21

is given in Section 2. Section 3 discusses the numerical For bubbles rising due to buoyancy, inertial effects are
method. Since the finite difference method is a conventional often of great importance, and bubble deformation can be
one, we only describe it briefly and focus instead on the considerable. Ryskin and Lea1 [4&42] examined the
front-tracking part. In Section 4 we present several sample steady-state shape of a number of rising axisymmetric
calculations of bubbles rising in a gravity field, both for a bubbles using a finite difference technique and a boundary-
two- and a three-dimensional flow. Our conclusions appear fitted coordinate system. They also applied their numerical
in Section 5. A short account of some of the work reported technique to an investigation of the steady-state shape of
here was presented at the Annual Meeting of the Division of bubbles in an axisymmetric strain flow. More recently
Fluid Dynamics of the American Physical Society at Ames, Dandy and Lea1 [ 143 extended this method to the two-fluid
CA, November 19-21, 1989, and at the regional meeting of problem, where the bubble is assumedto have finite density
the American Mathematical Society in Manhattan, KS, and viscosity. Unsteady numerical calculations of the initial
March 16-17, 1990.Calculations of the evolution of a three- deformation of two-dimensional, inviscid, bubbles include
dimensional, weakly stratified Rayleigh-Taylor instability those done by Anderson [2] in the case where the density
with a preliminary version of our front-tracking scheme ratio is very small and Baker and Moore [4] in the caseof
appeared in Tryggvason and Unverdi [48]. a bubble of zero density. Other unsteady, two-dimensional
flow calculations around deformable drops have been
2. BACKGROUND presented by Fyfe, Oran, and Fritts [22], who used a
moving triangular grid. Calculations of inviscid, fully three-
In order to demonstrate its capabilities, we have used our dimensional bubbles, for weak stratification and excluding
method to simulate the motion of bubbles. To put this work surfacetension, were presented by Brecht and Ferrante [6].
into proper perspective, a short overview of previous bubble The above numerical studies, as well as a large number
simulations follows. Several investigators have applied of experimental investigations, have led to a reasonably
boundary integral techniques to predict the deformation of good understanding of the motion of a single buoyant
drops and bubbles at zero Reynolds number (Stokes flow) bubble. A summary and graphic representation of the
in a strain field. Youngren and Acrivos [SO] considered a various rise modes can be found in Clift, Grace, and Weber
gas bubble in viscous extensional flow, and Rallison [37] [13] and Churchill [12], for example. Bubbles can be
considered the time-dependent deformation of a drop with characterized by the Morton number, M= gpz/p,o’, and
a viscosity equal to the surrounding fluid in the only fully the Eotvos number, Eo = p,gdz/a (also called the Bond
three-dimensional calculation so far. More recent work number), as well as by their density, pb/pO, and viscosity,
includes Stone and Leal’s [44] study of the breakup of pb/p,, ratios. Here the subscript o refers to the outer fluid,
extended drops, and investigations of the deformation of an b to the bubble fluid, p is the viscosity, p the density, c the
initially non-spherical drop in the Stokes limit by Koh and surface tension, g the gravity acceleration, and d, the “effec-
Lea1 [28] and Pozrikidis [36]. The last study shows that tive” diameter of the bubble. The Morton number involves
the spherically shaped solution of Hadamart (see,e.g.,Clift, the fluid properties only, and the Eotvos number is the non-
Grace, and Weber [ 133) is indeed very stable, while the dimensional size of the bubble. The rate of rise is expressed
relaxation of a non-spherical drop toward a spherical shape as a Reynolds number in terms of the rise velocity,
includes some rather remarkable processes,including the Re = pUdJp, where U is the velocity. Often the Weber num-
shedding of the fluid in the drop and, in other cases, ber, We = pU2dJo, is also used to characterize the bubble.
ingestion of the outer fluid. For recent experiments on the Obviously Eo and M are the parameters that one actually
breakup of drops in simple flow fields, see Stone, Bentley, controls. Re and We are found only after the terminal
and Lea1 [43]. A thorough review of the deformation of velocity has been determined. For all M, sufficiently small
small drops and bubbles in viscous flows and a list of other bubbles remain spherical and rise in a steady-state way.
contributions has been given by Rallison [38] and Acrivos Likewise, large enough bubbles become a “spherical cap”
[ 11. Other computations using the Stokes flow approxima- and rise in a steady-state fashion. In the transition zone
tion include, for example, a study of the axisymmetric from small to large bubbles, the motion depends on the
motion of a viscous drop toward a fluid interface for a range Morton number of the liquid. Bubbles in high M fluids
of capillary numbers and viscosity ratios by Chi and Lea1 become ellipsoidal (or “dimpled ellipsoidal-cap” for even
[9], simulation of the formation of two-dimensional drops larger M) before turning into a spherical cap shape, but
due to a Rayleigh-Taylor instability in a thin viscous film for continue to have a well-defined steady-state motion. For
fluids of equal viscosities by Yiantsios and Higgins [Sl 1, low M liquids, like water, while the bubbles also become
and an investigation of the motion of drops through circular ellipsoidal, their motion is unsteady, following a zigzag or
tubes by Martinez and Udell [29]. Computations of the helical path. One they become a spherical cap and resume
breakup of contaminated drops in a strain field are reported steady-statemotion, their wake is usually turbulent for high
by Stone and Lea1 [45]. Re. While ellipsoidal bubbles are usually spherical on top,
28 UNVERDI AND TRYGGVASON

and flatter on the bottom, Ryskin and Lea1 [41], in their

numerical study, found bubbles with a relatively flat top and
more spherical back, for low viscosity outer fluid.

3. NUMERICAL METHOD

The equations governing the motion of unsteady, viscous,

incompressible, immiscible two-fluid systems are the
Navier-Stokes equations. In conservative form these are

= -Vp+pg+V.(2pD)+a~n&x-x~~). (1)

Here u is the velocity field; p and p are the discontinuous

density and viscosity fields, respectively. D is the rate
FIG. 1. The computational setup. The computational domain is
of deformation tensor, whose components are D,= discretized by a regular, three-dimensional, Cartesian, stationary grid.
&(ui,j + u~,~),0 is the surface tension coefficient, K is the The front is represented by a separate unstructured, two-dimensional,
curvature, and n is a normal to the interface separating the triangular grid. The computations are done in a box with rigid (full slip)
two fluids. 6(x-x1) is a delta function that is zero top and bottom and full slip or periodic boundaries in the horizontal
everywhere except at the interface, where x = x’; g is the direction.
gravity force.
These are supplemented by the incompressibility condi- equation for the pressure. This equation is solved by a mul-
tion tigrid method (MGD described in Wesseling [49]) in two-
dimensions, while a simple, two-color, successive over-
v.u=o (2)
relaxation schemeis used for the three-dimensional runs.
The major novelty of the method described here is the
and equations of state for the density and the viscosity. Here way p and p are updated. Since these variables are both
the density and viscosity fields of a material particle remain discontinuous across the interface, we may expect either
constant, so the equations of state are simply excessivenumerical diffusion or problems with oscillations
around the jump if no special treatment is used at the
front. To avoid these problems we introduce an additional
s+u.vp=o, ap
t+u.vp=o. (3) computational element-the interface grid-that explicitly
marks the position of the interface. An indicator funcion,
Z(x), that is 1 inside the bubble and 0 in the outer fluid, is
Inside each fluid, p and p are constants.
then constructed from the known position of the interface.
The above equations are solved in a rectangular two- and
Since p and p are constant within each fluid, this indicator
three-dimensional domain with a finite difference method,
function allows us to evaluate the proper values of these
conventional in every way except that the advection of the
variables at each grid point by
density and the viscosity fields is achieved by explicit
tracking of the interface between the two fluids, instead of
solving Eq. (3) directly. The spatial differentiation is P(X) = PO+ (PI3- PO)Z(x)
(4)
calculated by second-order finite differences on a staggered P(X) = PO+ (Pb- PO)Z(x).
Eulerian grid, and the time advection is either by a simple,
explicit Euler method, or, for some of the two-dimensional To avoid introducing disturbances of length scale equal to
calculations, by a second order Adams-Bashforth method. the mesh by having the properties jump abruptly from one
When the Euler time integration method is used, the finite grid point to the next, the interface is not kept completely
difference approximation is exactly the same as in the MAC sharp but given a small thickness of the order of the mesh
method. The boundary conditions on our computational size. In this transition zone the fluid properties change
domain are either periodic or full slip in the horizontal smoothly from the value on one side of the interface to the
directions and a rigid, stress-free top and bottom. The value on the other side. This artificial interface thickness is
general setup of our problem is sketched in Fig. 1. a function of the mesh size used, only, and does not change
Combining the incompressibility condition and the during the calculations. Therefore no numerical diffusion is
momentum equations results in a nonseparable, elliptic present. This finite thickness also serves to position the
 VISCOUS, MULTI-FLUID FLOWS 29

interface more accurately on the grid. If the indicator func- between two grid points. Suppose that another part of the
tion were to take only the values 0 and 1, then it is only same interface returns between these same two gridpoints.
possible to state that the interface lies somewhere between Then a grid point that has been identified to be on the 0 side
grid points with different values. When the interface has a of an interface in one pass will be on the 1 side of the same
finite thickness, larger than the mesh size, it becomes interface when it returns. This is obviously unacceptable. An
possible to specify its position exactly. We therefore have to alternative to this local approach is a more “global” con-
live with the “uncertainty principle” that, when a jump struction, where the whole interface is examined for every
is represented on a stationary, discrete grid, it is not grid point where the properties are needed. An example of
possible to specify a precise location and zero thickness this approach would be to construct the indicator function,
simultaneously. in two-dimension and for a closed contour, by
A justification for spreading the sharp interface onto the
grid may be found in considering the definition of a 6 func-
tion as the limit of a smooth function when a parameter
defining the smoothness of the function goes to zero. Here,
just as finite difference approximations to the derivative are If g(r) = 1, this expression is the winding number of the con-
obtained by stopping the limiting process at some finite tour with respect to the grid point and is thus either 1 for
interval (h), the 6 function is approximated by taking the points inside the contour or 0 for points outside. The func-
smoothness parameter equal to some small, but finite, tion g(r) is selectedin such a way as to make the transition
value, instead of to zero. Approximating a 6 function by a smooth (e.g., g(r) = 1 - ~~r2’sZ,where 6 is a “smoothness”
smooth grid distribution is not a new approach. Our parameter). The advantage of this global approach is that
inspiration comes from two well-known methods where interacting interfaces are accounted for, but the drawback is
such approximation is used. The first is the so-called VIC that it is expensive, even if only grid points in the vicinity of
method for simulations of vortex sheet problems, where an the interface are considered.
infinitely thin vortex sheet is replaced by a finite thickness To take advantage of the desirable properties of the
vortex layer on a stationary grid for the velocity calcula- global property construction approach, but to defeat the
tions. The other is the Navier-Stokes calculation of fluid high cost of the straightforward implementation, we have
motion in the heart by Peskin and collaborators, where the developed a fast global property construction method. The
boundaries of the heart are approximated by infinitely thin, basic elements of this method are as follows:
moving walls in a fluid-filled domain and the force induced
by the heart muscles is distributed on to the grid. Here we (1) The jump in the indicator function carried by the
work also with a step-function, in addition to a b-function, interface is spread to the grid points nearest to the interface.
but similar considerations apply. This generates a grid-gradient field which is zero, except
In cases where the tracked interface carries a quantity near the interface, and has a finite thickness. The spreading
that vanishes everywhere except at the interface (vorticity in of the jump onto the grid is done in such a way that the
the VIC method and forces in Peskin’s calculations), we volume integral of the gradient (or the jump) is conserved.
simply proceed to solve the governing equations after this So, if G(x) is the gradient of the indicator function
quantity has been distributed to the stationary grid. evaluated at a stationary grid point x, and D is a “distribu-
However, when the tracked interface marks the change or tion function” that determines what fraction of the interface
jump in some fluid property, it is necessary to change the quantity should go to each grid point, then
grid values of this property every time the interface moves.
Several possibilities exist for accomplishing this update. The G(x) = c D(x -xc’)) n(l) Ad’).
simplest is a “local” approach, where one identifies the side
of the interface, where the grid points next to it appear, and
Here n(‘) is the unit normal vector to an interface element of
one assigns properties accordingly. This, of course, only
area ds”’ (or 1’me segment in two-dimensions) whose
updates points next to the interface, but since the properties
centroid is at x(I).
are constant within each fluid, they can only change if the
interface has recently passedby. Since the interface motion (2) By numerical differentiation, using second-order
must be limited to less than a mesh size in each time step centered differences, we find the divergence of the gradient
for stability and accuracy reasons, an algorithm may be field (V . G), thus calculating the Laplacian of the indicator
constructed in such a way that all points where a change function. This is again zero, except near the interfaces.
might have taken place are visited. (3) To find the indicator function everywhere we use a
However, a problem with this method occurs when two fast Poisson solver to solve
interfaces or different segmentsof the same interface merge
or lie close to each other. Consider an interface passing V21=V.G (7)
30 UNVERDI AND TRYGGVASON

with the proper boundary conditions. Here the right-hand gradient was distributed to and the same weight functions.
side is calculated in (1) and (2). The Laplacian is The velocity at interface point 1is, therefore,
approximated by the standard second-order, finite
difference approximation. UC’)= 1 qxi- .(‘)) u,, (11)
This approach leads to a field quantity Z(x), which is con-
stant within each fluid, but has a finite thickness transition where the sum is over the points on the stationary grid in
zone around the interface and therefore approximates a the vicinity of the interface point. The interface is then
two-dimensional step-function. Obviously the method advected by integrating
extends easily to problems where there are more than two
fluids present.
In the calculations reported here we have used the dx”‘= uu,,
dt (121
distribution function

D(x - .(‘I) =
(4h)-’
fi ( 1 +cos;
i= I
(x;-xxj’))
>
,
Discrete points on the interface are advected with the
flow and the interface itself is formed by connecting these
if ~x~-x;‘)/ <2h, i= 1, cc; points. In two dimensions, two points are connected by a
straight line segment, but in three dimensions, three points
otherwise, (8) are connected by a triangular element. The data structure is
therefore the same as is commonly used in finite element
introduced by Peskin [35]. Here h is the Eulerian mesh modeling, which makes the use of graphics tools developed
width and c(= 2, 3 (two- and three-dimensions, respec- to visualize such models convenient. (The three-dimen-
tively). This function has several desirable properties, as dis- sional figures in this paper were generated by the Applica-
cussed by Peskin [35]. We have also experimented with a tion Visualization System software on a Stellar work-
simpler area-weighting distribution as frequently used in station.)
VIC calculations (see,e.g., Christiansen [ 111). This gives a As the interface deforms, some parts of it are depleted of
sharper interface, but not quite as smooth a transition. The computational points, and other parts may becomecrowded
area weighting distribution is obviously much simpler than with points. To maintain an adequate resolution, we must
(8) but is known to introduce small-scale anisotropy (e.g., add computational elements as the interface stretches
Tryggvason [47]). We have not yet conducted a detailed and (although often not necessary) remove elements that
investigation of under what circumstances such anisotropy become very small. In two dimensions, where the front is
can be tolerated, but it is likely that when surface tension is simply a line, these modifications are a relatively simple
important small scale smoothness is desirable. matter. The distance between two markers is kept within a
In those calculations where surface tension is included we lower bound and an upper bound by adding and deleting
find the magnitude of the surface tension force from the points. When the interface is a surface embedded in a three-
curvature of the interface: dimensional flow, this aspect literally takes on a whole new
dimension. Our regridding procedure consists of three basic
steps: (1) node addition (for elements that become too
large); (2) node deletion (when elements become small);
Here 0 is the surface tension coefficient, and K(‘) is the and (3) reconnection or restructuring (to eliminate bad
curvature of the interface at the centroid of element (I). “aspect ratios, ” i.e., elements that have a small area but a
This force is then distributed onto the grid in the same way large perimeter). Restructuring is sometimes also neededfor
as the gradient, thus constructing a grid-force field three-dimensional surfaces to improve their smoothness.
Figure 2 depicts these three basic regridding operations
F(x) = 1 D(x - xc’)) f (‘). schematically. (For a discussion of these problems for two-
(10) dimensional simulations on a moving triangular grid see,
e.g., Fritts and Boris [21].)
This ensures that the total force is conserved when going For the calculation of interface curvature we have
from the interface to the grid. The distributed surface implemented somewhat different methods for the two-
tension force gives a body-force-like term in the equations dimensional and the three-dimensional calculations. In two
of motion, and this term is included in calculation of the dimensions we lit a local, cubic, Hermite polynomial to four
velocity field. Its divergence also appears in the source term points on the interface and find the curvature by differentia-
for the pressure equation. tion with respect to the arc length. The surface tension force
To interpolate the velocities of the interface points from is found by evaluating the curvature at the middle of the
the grid velocities we use the same gridpoints that the element. To calculate the mean curvature of an interface
 VISCOUS, MULTI-FLUID FLOWS 31

the interface grid is restructured to allow long time simula-

tions.
A number of bubbles for various Eo and M are shown in
Fig. 3. The figure demonstrates the (essentially) steady
states, but in each case the full unsteady problem is solved.
The physical times are equal for each row, but the middle
and bottom row are at twice the time of the top row. These
(b’ @ r) @ calculations are all done on a 65 by 129 grid, and the com-
putational box has rigid, full slip boundaries. Therefore,
considerable boundary effects may be expected. Here,
p,,/pb = 40, and the Eiitvijs number, the Morton number,
and the viscosity ratio for each frame are noted in the
caption. For small Eo (i.e., large surface tension or small
bubble), the bubbles remain nearly circular, but for higher
Eo the bubbles attain a steady state with a somewhat flat or
FIG. 2. The basic operations in dynamic restructuring of the three-
dimensional interface grid: (a) addition of a point by bisection of the
longest side of a large element; (b) point deletion to eliminate small

0 0
elements and remove points; (c) reconnection to eliminate elements with
large perimeter and small area.

in three dimensions we use a method described by Todd

and McLeod [46]. At each point on the interface grid we
identify its neighbors. By fitting circles through the point of
interest and neighboring points, two at a time, we obtain a
set of tangent vectors and curvatures. The normal is found
by averaging the vector product of the various tangent
vectors. This information is then used to estimate the
coefficients in the so-called Dupin indicatrix by a least
I (a) (b) (d)

squares method. These coefficients give the mean curvature,

in, directly. Multiplying the average of the mean curvature
vector at the three vertices of each triangular element by the a 0
surface tension coefficient and the area of the element, CJAA,
gives the resultant surface tension force on the element. Dis-
tribution of this force onto the grid by the three-dimensional
distribution function completes the interaction of the
interface and the Eulerian grid. In addition to calculations
of surface tension forces, the curvature is needed when (e) (f) (h)
interface points are added.

4. COMPUTATIONAL EXAMPLES

A preliminary version of the method described above has

been used to study the Rayleigh-Taylor instability of two
fluids of similar density, where the Boussinesq approxima-
tion is applicable. The Boussinesq approximation leads to a
separable pressure equation and hence simplifies the
calculations considerably. These calculations, reported in (i) Cj) Ik)
1 (I)
Tryggvason and Unverdi [48], included neither surface
FIG. 3. The steady-state rise of a rising bubble for various Eo and M.
tension nor dynamic restructuring of the interface mesh. Here po/pb=40. Top row, Eo= 1; M= IO-‘; 10e6; 10e5; lo-“;
Here, we present calculations of buoyant bubbles, both in p0/pb=88; 156; 277; 493. Middle row, Eo= 10; M= 10m4;10m3; IO-*;
two and three dimensions where the fluids can have 10-i; pO/pb = 88; 156; 277; 493. Bottom row, Eo = 104; M= 10-l; 1; 10;
arbitrary density difference, surface tension is included, and 102;po/pb=85; 151;269;479.
32 UNVERDI AND TRYCGVASON

the skirts pulled off the bubble with the separated flow
increase continuously in length. At the highest Eotvos
number (bottom row) the position of the bubble is relatively
independent of the Morton number, but for the lowest Eo a
strong dependency is observable. This is as we would expect.
Low Eo corresponds to small bubbles where viscous drag
is important; for larger bubbles (and Eo) the drag is
mostly pressure drag and the velocity is therefore relatively
independent of the viscosity.
To examine the flow field around the bubbles in more
(a) (b) (d) detail, we plot the stream lines in Fig. 4 in a frame of
referencemoving with the bubble for the bubbles in Figs. 3a
through h. For high Morton numbers (and therefore low
Reynolds numbers), the flow around the bubble is nearly a
Stokes flow. No wake is visible behind the bubble, and the
interior of the bubble consists of a single vortex pair. As the
Morton number decreases,the Reynolds number increases
and a wake appears behind the bubble. For small Eo, where
the bubbles remain nearly circular, the wake is relatively
small, but as the bubble deforms more, a larger wake
appears. To show both the vortex behind the bubble as well
(e) (f) (g) (h) as the one inside of it, the stream line values are not equi-
FIG. 4. The stream lines for the top and the middle frames in Fig. 3.
spaced. The secondary vortex inside the bubbles in (e) and
(f) is actually considerably weaker than the primary vortex,
which, in turn, is weaker than the recirculating region
elliptic shape. Even higher Eo results in bubbles with a behind the bubble. Even though the bubble itself has
nearly flat back and semicircular front. For low Eo and M reached an approximate steady-state shape, the wake is
(low viscosity but high surface tension or small bubble, i.e., actually still growing in most of the casesshown here.
the top left frame), this latter shape is actually inverted with To illustrate the disturbance the bubble causes on the
the top being slightly flatter than the back. This agreeswith pressure field, in Fig. 5 we have plotted a three-dimensional
Ryskin and Leal’s [41] prediction for steady-state axisym- view of the pressure field for two cases.Here Eo = 1.0, and
metric bubbles. For the highest Eo and lowest M (lower left- in(a)M=10P8andin(b)M=10-5.Thepressurerisesdue
hand corner), the solution shown is not truly a steady state; to hydrostatic forces toward the bottom of the box (away

(a) (b)
FIG. 5. The pressurefield for EO = 1: (a) M = 10-s; (b) A4 = 10-s. The pressure increases toward the bottom ofthe domain d.ue to hydrostatic
 VISCOUS, MULTI-FLUID FLOWS 33

from the viewer). The surface tension forces on the surface Notice in particular that the position of the bubble is in
of the bubble keep the pressure inside it nearly constant and good agreement in both cases.
considerably higher than the ambient pressure. The main In addition to a poorer resolution of the flow field itself,
difference is that the bubble in the less viscous fluid (a) the coarse grid also makes the artificial interface thickness
causes considerably larger changes in the pressure field of larger. In Fig. 8 we show the indicator function for the run
the outer flow. in Fig. 6 as reconstructed from the tracked front. The top
The bubbles in Fig. 3 are essentially fully converged solu- frame is the coarse grid solution, and the bottom frame is
tions, and, to demonstrate the convergence, we compare the from the liner grid. In both cases this function is uniformly
steady-state solution for Eo = 10 and M = 0.1, as calculated 0 outside the bubble and 1 inside. Furthermore, in both
on two different grids, in Fig. 6. The left frame is calculated cases the transition from one value to the other takes about
on a 33 by 65 grid, and the right one is on a 65 by 129 grid three meshes. Obviously, that corresponds to a thicker layer
(as in Fig. 3). The interface itself is shown in (a) and the on the coarse grid.
stream function, in (b). Obviously, the interfaces are We have compared our calculations with the results
virtually identical. The stream functions are also close, but presented by Clift, Grace, and Weber [ 131 (see also Bhage
since the solution is in primitive variables and the stream and Weber [S]), who chart the various shape regimes for a
function is obtained by processing the velocity field rising bubble. In spite of the considerable differences
(differentiate to find the vorticity and solving a Poisson between our setup and the one the chart is based on (i.e.,
equation), we expect slightly worse agreement. The resolu- fully three-dimensional or axisymmetric situations and
tion requirement increases as the viscosity decreases, and faraway boundaries), there is a remarkably good qualitative
Fig. 7 shows the interface for Eo = 10 and A4 = 0.001, and agreement. The shapes also correspond qualitatively to the
same grids as in Fig. 6. Here, the coarse grid interface axisymmetric steady-state calculations of Ryskin and Lea1
deviates slightly from the line grid solution at the rim of the [41]. The limitations of our calculations are primarily
bubble where curvature is highest. Nevertheless, the overall resolution at high Eo and low M. Furthermore, as with any
shape is in close agreement with the liner grid solution. explicit method, simulations of very low Reynolds number
flows are subject to strict stability limitations on the size of
the time step and are therefore expensive. At zero Reynolds
number, however, boundary integral solutions of the Stokes
equations are applicable.
To demonstrate how our method deals with interacting

ii 0 interfaces, Fig. 9a shows the collision of two bubbles of

equal size. Here Eo = 10, M= lo-*, pO/pb =40 and
pO/pLb= 27. Initially both bubbles start out as circles near
the bottom of the box. As the bubbles rise, the lower bubble
is engulfed into the top bubble, and eventually both bubbles
move together with relatively little change in shape. The
indicator function for this simulation is shown in Fig. 9b,
(a)
where it is clear that the densities merge smoothly into a
uniform field as the bubbles come close enough. The double
interface remains in the flow and actually maintains a slight

(b)
(a) (b)
FIG. 6. Resolution test for Eo= 10, M= 10-l: (a) bubble shape:
(b) stream function. The coarse grid (left frame) is 33 by 65 meshes; the FIG. 7. Resolution test for Eo = 10, M= lo-‘. The coarse grid (a) is
tine grid (right frame) is 65 by 129 meshes. 33 by 65 meshes; the tine grid (b) is 65 by 129 meshes.
34 UNVERDI AND TRYGGVASON

(a)

(b)

FIG. 8. The indica .tor function (scaled density: 1 inside the bubble and 0 outside) for the run in Fig. 7: (a) 33 by 65 grid; (b) 65 by 129 grid

pressure difference between the bubbles. Notice that there is

a small accumulation of fluid in the middle of the double
interface and that this shows up in the indicator field. At the
point where the interfaces have come completely together
the indicator function shows no trace of the interface.
To test the method on fully three-dimensional problems
we have made a number of calculations for both single 0
bubbles as well as two interacting bubbles. The initial defor-
mation and the subsequent steady-state rise of two fully
three-dimensional bubbles are shown in Fig. 10. The
0
viewing angle lies slightly below the box, and the bubble is
shown at four different times. In (a) Eo = 10 and M = lo-‘,
pO/p,, =40, ,LL~/,u~ = 156, and in (b) Eo= lo', M= lo*,
p0/pb=40, pL,/pLb= 49. Initially the bubble deforms to a
somewhat ellipsoidal shape in (a) and a dimpled ellipsoid in
(b). It then rises with constant velocity and shape. The
stationary grid in these calculations was 32 by 32 meshes in 0
the horizontal direction, and 64 meshes long in the vertical
direction. The interface grid, shown in Fig. 11 for the last
time in Fig. IOa, moves with the fluid and is thus constantly
0 (bl
changing, even though the bubble may have reached a
FIG. 9. Collision of two bubbles of equal size. Here Eo = 10,
steady-state shape. M= 10m2, pa/p,, =40, and /.~~/fi~ =27. The grid is 65 by 129 meshes:
In Fig. 12 the merging of two bubbles of equal size and (a) bubble shapes; (b) the indicator function. The contours levels are 0.167,
small surface tension is simulated. The grid is the same as in 0.333,0.500,0.667,0.833.
 VISCOUS, MULTI-FLUID FLOWS 35

FIG. 11. The interface grid for a rising bubble corresponding to the
last time in Fig. 10a.

(a) W upward suction by the top bubble forms it into a somewhat

cylindrical shape, eventually leaving a thin tail behind as
FIG. 10. The initial deformation and rise of a three-dimensional most of the lower bubble merges with the top one. Although
bubble. The bubble is shown at four different times. The simulation is on
a 32 x 32 x 64 grid. The initial bubble diameter is 0.4 times the width of the
we have calculated the process for a longer time, the lower
computational box: (a) Eo = 10 and M = 10-3, pO/pb = 40, po/pLb= 156; part of the bottom bubble is clearly underresolved in the last
(b) Eo = 10’ and M= lo*, po/pb =40, p,,/pb = 49. frame in Fig. 12, and the physical relevance of the results
may therefore be doubtful beyond this stage. For low sur-
face tension, as was the case here, the merging process is
Fig.lO, and Eo=50, M= 1, po/p,,==20, p,/pb=26. The relatively fast compared with the rise velocity, and the
bubbles are initially spherical and are put relatively close bubbles move only about one bubble diameter during the
together, slightly off the centerline of the box. As the bubbles coalescence process, which is in qualitative agreement with
rise they deform considerably. The bottom of the top bubble de Nevers and Wu [32], for example. For larger surface
folds upward and deforms the lower bubble into a “pear” tension the coalescence process takes longer as is seen in the
shape, pointing toward the top one. The top bubble con- two-dimensional calculation in Fig. 9.
tinues to deform into a hemispherical shell with most of the Not ali bubble interactions involve merging bubbles.
fluid contained in the rim. The lower bubble is drawn into In Fig. 13 the motion of a large and a small bubble is
this shell. Initially, the back of the lower bubble deforms in simulated. Here M = 10P3, pO/pt, = 20, pLo/pb= 28, Eo = 10,
a similar way as the back of the top one does, but the and Eo = 2.5 for the big and small bubble, respectively, and

FIG. 12. Non-axisymmetric merging of two bubbles on 32 by 32 by 64 grid. Here Eo = SO,M = I, p,/p, = 20, and p, /,ub = 26
36 UNVCWI ANIJ I KYWVASUN

FIG. 13. Collision of a large bubble with a smaller one. Here M = 10m3,p,/pb = 20, and pO/pb = 28. For the bigger bubble Eo = 10, and Eo = 2.5
for the smaller one. The grid is 32 by 32 by 96 meshes.

the grid is 32 by 32 by 96 meshes. The surface tension is the method interfaces well with standard Poisson solvers.
relatively high, so the large bubble takes on an ellipsoidal Several test cases suggest that the method is both robust and
shape and the small bubble remains nearly spherical. The computationally efficient.
simulations are initiated with the bubbles relatively close The major limitation of the method is, of course, resolu-
and the smaller bubble lying above the larger one. As the tion requirements at large Eo and low M. As with all other
bubbles rise, the larger bubble pushes the smaller to the numerical schemes, this is particularly limiting for three-
side, passes by, and finally leaves the smaller bubble behind dimensional simulations. Another limitation, more specific
to rise more slowly. The interaction process is now slower to this method, is the treatment of interfaces that interact. If
than in the low surface tension case, and the large bubble two interfaces (or different segments of the same interface)
rises about three diameters from the first frame to the last. some sufftciently close together so that they influence the
same grid point, the property profile gradients simply cancel
5. CONCLUSIONS each other. The resulting double interface will, however,
remain in the flow, even though the net effect may by zero.
A numerical method for the simulation of unsteady, For simulations where surface tension is so small that the
incompressible, multi-fluid flows has been presented. The pressure difference across such double interface is of minor
Navier-Stokes equations are solved on a stationary finite significance, the calculations can proceed without any
difference grid, and the sharp interface separating the fluids special treatment. When the double interfaces are of
is explicitly tracked by an additional moving interface grid. dynamic significance, such as in the generation of froth
This prevents both numerical diffusion and oscillations and when bubbles arrive at a free surface, or when a thin film
allows surface tension to be incorporated in a natural way. between adjacent bubbles ruptures, either a subgrid model
The original Eulerian grid is retained throughout the for the film behavior or a mechanism to decide on the
simulation. Since a conservative finite difference scheme is rupture time and the subsequent changes in interface
used, no special difference formulas are needed near the topology have to be added. The proper inclusion of rupture
interface, nor is any restructuring required to align the grid of thin films necessitates additional physical modeling.
with the interface. The use of a regular fixed mesh makes Generally rupture is highly dependent upon the properties
coding the method relatively simple, and, more importantly, of the fluids involved (compare air bubbles in clean and
makes most of the method easily vectorizable. Furthermore, soapy water), and additional material properties are
 VISCOUS, MULTI-FLUID FLOWS 31

needed. To accomplish the actual implementation of such a 12. S. W. Churchill, “Viscous Flows. The Practical Use of Theory”
rupture model a topological change algorithm (see, for (Butterworths, London, 1988).
example, Glimm et al. [24] ) would be necessary. The ability 13. R. Clift, J. R. Grace, and M. E. Weber, “Bubbles, Drops, and Particles”
(Academic Press, New York/London, 1978).
of the present method to deal with interacting interfaces by
allowing them to form passive structures when suitable is 14. D. S. Dandy and L. G. Leal, J. Fluid Mech. 208, 161 (1989).
15. B. J. Daly, Phys. Fluids 10, 297 (1967).
comparable to the treatment in volume-tracking methods.
16. B. J. Daly, J. Comput. Phys. 4, 97 (1969).
(In its present form the method actually has a built-in model
17. P. Daripa, J. Glimm, B. Lindquist, M. Maesumi, and 0. McBrian, “On
for thin films; it allows no rupture to take place and sets the
the Simulation of Heterogeneous Petroleum Reservoirs,” in Numerical
pressure jump across the film equal to twice the jump across Simulation in Oil Recovery, edited by M. Wheeler (Springer Verlag,
a single interface with the same curvature). We note that, New York, 1988).
while a volume-tracking method can also account for the 18. J. M. Floryan and H. Rasmussen, Appl. Mech. Rev. 42, 323 (1989).
interaction of two interfaces, modeling of thin films and 19. L. J. Fauci and C. S. Peskin, J. Comput. Phys. 77, 80 (1988).
rupture is completely inaccessible to volume-tracking, 20. A. L. Fogelson and C. S. Peskin, J. Comput. Phys. 79, 50 (1988).
since a double interface will always disappear whether it is 21. M. J. Fritts and J. P. Boris, J. Compuf. Phys. 31, 173 (1979).
physical or not. 22. D. E. Fyfe, E. S. Oran, and M. J. Fritts, J. Comput. Phys. 76, 349 (1988).
Although our method has been tested extensively on two- 23. J. Glimm, 0. McBryan, R. Menikoff, and D. H. Sharp, SIAM J. Sci.
dimensional problems, the major purpose of developing Stat. Comput. 7, 230 (1987).
it was to simulate fully three dimensional situations. The 24. J. Glimm, J. Grove, B. Lindquist, 0. McBryan, and G. Tryggvason,
test cases presented here suggest that larger simulations SIAM J. Sci. Stat. Comput. 9,61 (1988).
involving several bubbles are entirely feasible, although 25. F. H. Harlow and J. E. Welch, Phys. Fluids 8, 2182 (1965).
such simulations will, naturally, always be limited to some- 26. C. W. Hirt and B. D. Nichols, J. Comput. Phys. 39, 201 (1981).
what modest-sized systems. Our experience suggests that, 27. J. M. Hyman, Physica D 12, 396 (1984).
for high viscosity and high surface tension, a meaningful 28. C. J. Koh and L. G. Leal, Phys. Fluids A 1, 1309 ( 1989).
resolution of a single bubble can be achieved on a grid as 29. M. J. Martinez and K. S. Udell, J. Fluid Mech. 210, 565 (1990).
small as 16’ (our three-dimensional bubbles had a diameter 30. J. C. S. Meng and J. A. L. Thomson, J. Fluid Mech. 84,433 (1978).
of about 16 meshes). Larger simulations, for example, on a 31. G. Moretti, Annu. Rev. Fluid Mech. 19, 313 (1987).
643 or a 1283 grid, could include several bubbles. 32. N. de Nevers and J.-L. Wu, AIChE J. 17, 182 (1971).
33. W. F. Noh and P. Woodward, Lecture Notes in Physics, Vol. 59
(Springer-Verlag, New York/Berlin, 1976) p. 330.
ACKNOWLEDGMENTS 34. E. S. Oran and J. P. Boris, Numerical Simulafions of Reuclive Fiow
(Elsevier, Amsterdam/New York, 1987).
We wish to acknowledge the help of C. Churchill at the University of
35. C. S. Peskin, J. Comput. Phys. 25, 220 (1977).
Michigan Visualization Laboratory with the graphics, and P. Wesseling at
the University of Delft who sent us his multigrid code (MGD). This work 36. C. Pozrikidis, J. Fluid Mech. 210, 1 (1990).
was supported by NSF Grant No. MSM87-07646. The calculations were 37. J. M. Rallison, J. Fluid Mech. 109, 465 (1981).
done on the computers at the San Diego Supercomputer Center which is 38. J. M. Rallison, Annu. Rev. Fluid Mech. 16, 45 (1984).
sponsored by the NSF. 39. R. D. Richtmyer and K. W. Morton, Bfference Methods for Initial-
Value Problems (Interscience, New York, 1967).
REFERENCES 40. G. Ryskin and L. G. Leal, J. Fluid Mech. 148, 1 (1984).
41. G. Ryskin and L. G. Leal, J. Fluid Mech. 148, 19 (1984).
I. A. Acrivos, “The Deformation and Break-up of Single Drops in Shear 42. G. Ryskin and L. G. Leal, J. Fluid Mech. 148, 37 (1984).
Fields,” in Physicochemical Hydrodynamics, Inlerfacial Phenomena,
43. H. A. Stone, B. J. Bentley, and L. G. Leal, J. Fluid Mech. 173, 131
edited by M. G. Velarde (Plenum, New York, 1987) p. 1.
(1986).
2. C. R. Anderson, J. Comput. Phys. 61,417 (1985).
44. H. A. Stone and L. G. Leal, J. Fluid Mech. 198, 399 (1989).
3. A. A. Amsden and F. H. Harlow, J. Comput. Phys. 6, 322 (1970).
45. H. A. Stone and L. G. Leal, J. Fluid Mech. 220, 161 (1990).
4. G. R. Baker and D. W. Moore, Phys. Fluids A 1, 1451 (1989).
46. P. H. Todd and R. J. Y. McLeod, Cornput. Aided Design 18, 33 (1986).
5. D. Bhaga and M. E. Weber, J. Fluid Mech. 105,61 (1981).
A-l G. Tryggvason, J. Comput. Phys. 75, 253 (1988).
7,.
6. S. H. Brecht and J. R. Ferrante, Phys. Fluids A 1, 1166 (1989).
48. G. Tryggvason and S. 0. Unverdi, Phys. Fluids A 2, No. 5, 656 (1990).
7. J. P. Boris, Annu. Rev. Fluid Mech. 21, 345 (1989).
49. P. Wesseling, “A Robust and Efficient Multigrid Method,” in Mull&id
8. I.-L. Chern, J. Glimm, 0. McBryan, B. Plohr, and S. Yaniv, J. Comput.
Methods, edited by W. Hackbusch and U. Trottenberg (Proceedings,
Phys. 62, 83 (1986).
KolnPorz, 1981) Lect. Notes in Math., Vol. 960 (Springer-Verlag,
9. B. K. Chi and L. G. Leal, J. Fluid Mech. 201, 123 (1989). Berlin, 1982), p. 614.
IO. A. J. Chorin, J. Compuf. Phys. 35, 1 (1980). 50. G. K. Youngren and A. Acrivos, J. Fluid Mech. 76, 433 (1976).
Il. J. P. Christiansen, J. Comput. Phys. 13, 363 (1973). 51. S. G. Yiantsios and B. G. Higgins, Phys. F/uids A 1, 1484 (1989).