Direct Numerical Simulation
Direct Numerical Simulation
1. INTRODUCTION
The numerical simulation of flows with interfaces and free-surface flows is a
vast topic, with applications to domains as varied as environment, geophysics,
engineering, and fundamental physics.
In engineering, as well as in other disciplines, the study of liquid-gas inter-
faces is important in combustion problems with liquid and gas reagents. The
formation of droplet clouds or sprays that subsequently burn in combustion
chambers originates in interfacial instabilities, such as the Kelvin-Helmholtz
instability. What can numerical simulations do to improve our understanding of
these phenomena? The limitations of numerical techniques make it impossible
to consider more than a few droplets or bubbles. They also force us to stay at
low Reynolds or Weber numbers, which prevent us from finding a direct solu-
tion to the breakup problem. However, these methods are potentially important.
First, the continuous improvement of computational power (or, what amounts
to the same, the drop in megaflop price) continuously extends the range of af-
fordable problems. Second, and more importantly, the phenomena we consider
often happen on scales of space and time where experimental visualization is
difficult or impossible. In such cases, numerical simulation may be a useful
prod to the intuition of the physicist, the engineer, or the mathematician.
A typical example of interfacial flow is the collision between two liquid
droplets. Finding the flow involves the study not only of hydrodynamic fields
in the air and water phases but also of the air-water interface. This latter part
567
0066-4189/99/0115-0567$08.00
P1: ARS/mbg P2: ARS/vks QC: ARS
November 16, 1998 19:16 Annual Reviews AR075-16
is more difficult than the former because the interface is subject to a number of
relevant physical phenomena, at scales much smaller than the typical sizes of
the droplets. For instance, a major difficulty is caused by the change in interface
topology that occurs when the colliding droplets coalesce. This complicates
the physics and sharpens the requirements that a numerical method must satisfy
in order to resolve the motion in a satisfactory way.
These difficulties are dealt with in a variety of ways, depending on the physi-
cal modeling and the numerical methods. First comes a choice of modeling the
interface as either a thin or a thick region of space. In engineering problems,
such as combustion studies, the interface may be considered a discontinuity
Annu. Rev. Fluid Mech. 1999.31:567-603. Downloaded from www.annualreviews.org
novice practitioner.
by University of Massachusetts - Amherst on 09/24/12. For personal use only.
The normal velocity is related to the fluid velocity field u in various ways.
In the simplest case, without phase change, one assumes continuity of fluid
velocity
u1 = u2 . (1)
(We note in general xi , the limiting values of a variable x when the interface
is approached from phase i.) This may be written in jump notation, i.e. the
notation [x] = x2 − x1 :
[u] S = 0. (2)
Annu. Rev. Fluid Mech. 1999.31:567-603. Downloaded from www.annualreviews.org
The interface velocity is then the normal velocity, the same on both sides of the
by University of Massachusetts - Amherst on 09/24/12. For personal use only.
interface:
V = u1 · n = u2 · n. (3)
In the case of evaporation, or more generally phase change, there may be a mass
flow q from phase 1 to phase 2. This mass flow is connected to the velocities on
both sides of the interface in the following way. Consider the frame of reference
where the interface is at rest. The normal velocities in that frame of reference
are u 0 = u · n − V . The mass flow is the amount of mass that goes from phase 1
to phase 2, and it must be the same on both sides of the interface by conservation
of mass (the Rankine-Hugoniot condition):
ρ1 u 01 = ρ2 u 02 = q. (4)
Then, in the general frame of reference,
ρ1 (u1 · n − V ) = ρ2 (u2 · n − V ) = q. (5)
Thus, if there is no phase change, i.e. q = 0, one recovers Equation 3, which
now appears clearly to be a consequence of mass conservation. The continuity
of tangential velocities [ut = u − (u · n)n], on the other hand, is a physical
assumption akin to the assumption that the slip velocity on a solid wall vanishes.
Then interface motion may be described in weak form by
∂t χ + V n · ∇χ = 0, (6)
which in the no-phase-change case amounts to
∂t χ + u · ∇χ = 0. (7)
(Notice that ∇χ = nδ S .) Here we use a weak formulation of the partial dif-
ferential equation (PDE), since derivatives of the discontinuous function χ
are singular. In this weak formulation, the PDE is interpreted by the space
integrals of Equation 7. In other words, the function cannot be differentiated,
but integrals of Equation 7 are well-defined: They express volume evolution
and correspond to interface motion with velocity V.
P1: ARS/mbg P2: ARS/vks QC: ARS
November 16, 1998 19:16 Annual Reviews AR075-16
∇ · u = 0. (8)
ρ(∂t u + u · ∇u) = −∇ p + 2∇ · µD + 2σ κδ S n,
by University of Massachusetts - Amherst on 09/24/12. For personal use only.
(9)
where ρ is the density and µ the dynamical viscosity. These are two phys-
ical properties of the bulk phases that may have discontinuities across phase
boundaries. The rate-of-strain tensor is D:
µ ¶
1 ∂u j ∂u i
Di j = + . (10)
2 ∂ xi ∂x j
In these equations, the term 2σ κδ S n represents capillary forces, with a surface
tension σ and n is again the unit normal to the interface. The mean curvature
is κ = (1/r1 + 1/r2 )/2. It may also be expressed as
1
κ = − ∇ S · n, (11)
2
where ∇ S is the gradient operator restricted to the surface S.
T = −σ (1 − n ⊗ n)δ S , (12)
2σ κnδ S = −∇ · T, (13)
This condition may be split into a normal and tangential stress condition
[− p + 2µn · D · n] S = 2σ κ, (16)
[µt(k) · D · n] S = 0, (17)
where the vectors tk may be any set of d − 1 independent tangent (to S) vectors,
and d is the dimension of space.
Away from the interface, Equation 9 takes the usual form
∂t u + u · ∇u = −∇( p/ρ) + ν∇ 2 u, (18)
with ν = µ/ρ. All the above formulations are equivalent.
2.5 Problems with Reconnecting Interfaces
When two interfaces reconnect and change topology, as in Figures 3 and 4, the
above macroscopic model becomes even more singular. Indeed, immediately
after reconnection the surface is no longer a smooth surface but has cusp-like
singularities.
The curvature κ is then ill-defined and arbitrarily large in the instants imme-
diately following reconnection. Another difficulty arises in the case represented
in Figure 4. In that case it is reasonable to assume that the velocity field is hy-
perbolic near the reconnection point. The interfaces then approach each other
exponentially, and their separation d(t) varies as d0 exp(−λt), where λ is the
largest eigenvalue of the rate-of-strain tensor D of the local velocity field. It is
obvious that the separation d(t) rapidly reaches scales much smaller than the
scales on which microscopic molecular forces act to attract or repel interfaces
(de Gennes 1985). The exact nature of these forces depends on the fluids and
on the presence of surfactants. As a rule of thumb, they must be considered
prevalent on scales of 10–40 nm.
This difficulty is less important, from the numerical point of view at least,
in the case of the filamentary reconnection (Figure 3). Indeed, in that case the
P1: ARS/mbg P2: ARS/vks QC: ARS
November 16, 1998 19:16 Annual Reviews AR075-16
Figure 3 Interface reconnection through filament breakup. This kind of reconnection is accel-
erated by capillary tension. The Navier-Stokes equation predicts it will happen in finite time.
Figure 4 Interface reconnection through sheet breakup. This reconnection may occur in a hyper-
bolic region of the flow, in two-dimensional as well as in three-dimensional flow. The computation
of the final phase of reconnection requires the estimation of microscopic forces.
and
t(k) · D · n| S = 0. (20)
It may be noted that the last condition, which results from the nullity of tangen-
tial stresses, is purely kinematic: It does not involve material properties at all.
For this reason it has interesting consequences for the value of the vorticity on
a free surface (Batchelor 1967, Longuet-Higgins 1998).
then exists, in which the interface is both a free boundary between phases and
a vortex sheet (Kane 1994). The boundary integral may at first view be solved
more simply than the full Navier-Stokes or Euler equations. In a cubic domain
of size L and with a spatial grid size h, one needs N = L/ h points along
each dimension. The discretization of the full two-dimensional problem would
require N 2 grid points. If the interface is also of length |S| ' O(L), then the
boundary integral may be discretized with O(N ) points, which is much less than
what is needed for the full Euler equation. However, if there are many droplets
or bubbles, or if the interface is very convoluted, then |S| À L. However, for
many problems the boundary-integral method offers a substantial savings of
Annu. Rev. Fluid Mech. 1999.31:567-603. Downloaded from www.annualreviews.org
dynamics ... and seems to be hardly known in other areas of fluid dynamics” is,
however, still quite accurate.
5. FIXED-GRID METHODS
In fixed-grid methods the grid used to solve the Navier-Stokes equation is
entirely or quasi-entirely fixed. The grid may be quasi-entirely fixed when, for
P1: ARS/mbg P2: ARS/vks QC: ARS
November 16, 1998 19:16 Annual Reviews AR075-16
example, it is adaptive just for those cells cut by the interface, as in Figure 7
(Glimm et al 1986, 1987).
5.1 Marker Methods
In marker methods, tracers or marker particles are used in the algorithm to
locate the phases. Interfacial or surface-marker methods use marker particles
only on the interfaces. Volume-marker methods have marker particles in the
whole domain.
For two-phase flow, surface markers are more accurate than volume markers
because they track exactly the location of the interface. However, when there
are more than two phases, it may become difficult to handle the complexity of
triple lines (lines in three-dimensional space where three phases meet) and other
effects associated with the presence of several phases. Volume markers afford
a simple way of dealing with the problem. Volume markers become distorted
as time goes by, just as does the Lagrangian grid of Figure 6.
P1: ARS/mbg P2: ARS/vks QC: ARS
November 16, 1998 19:16 Annual Reviews AR075-16
Figure 6 In a grid that follows the fluid motion, a considerable distortion may appear even if
the interface has undergone relatively mild deformation. [Redrawn after Figure 5 of McHyman
(1984).]
Figure 7 An example of local adaptation on an otherwise fixed grid. (Left) The undeformed finite
element triangular grid; (Right) the thick line representing the interface is made up of sides of the
new triangles.
P1: ARS/mbg P2: ARS/vks QC: ARS
November 16, 1998 19:16 Annual Reviews AR075-16
Figure 8 An example of the use of surface markers on an underlying square grid. The use of
surface markers allows maintenance of the thin layer shown in the figure. In other types of fixed-grid
methods, this layer may break.
Figure 9 Another example of the use of surface markers: A vortical structure entrains the interface
in a spiraling motion. Surface markers allow the capture of details of interface motion on scales
much smaller than the grid spacing. However, as explained in the text, some small-scale detail of
the hydrodynamics may be lost when scales much smaller than the grid are created.
P1: ARS/mbg P2: ARS/vks QC: ARS
November 16, 1998 19:16 Annual Reviews AR075-16
the same viscosity and density and there is no surface tension, the interface is
“transparent” to the fluid: The fluid, in a way, does not know that there is an
interface. The phases are only distinguished by their “color” as an ideal passive
scalar with zero diffusivity. In that case it makes sense to track details smaller
than the grid.
On the other hand, if there is, for instance, a difference in density, there
will be scales in the velocity and pressure fields much smaller than the grid,
and these scales will not be well resolved using the uniform square grid of
Figure 9.
Surface-marker methods have been used extensively by the Glimm group
Annu. Rev. Fluid Mech. 1999.31:567-603. Downloaded from www.annualreviews.org
(Glimm et al 1986, 1987) and by the Tryggvason group (Tryggvason & Unverdi
by University of Massachusetts - Amherst on 09/24/12. For personal use only.
1990; Unverdi & Tryggvason 1992a,b). Marker methods have other advantages,
and an important one is the high degree of accuracy that may be achieved by
representing the interface through high-order interpolation polynomials. This
accuracy may improve the accuracy of surface tension calculations (Popinet &
Zaleski 1998a,b).
5.2 Volume of Fluid Methods
As stated in the introduction, we consider in some detail VOF methods relative
to two-dimensional incompressible flows, without mass transfer. Extension to
compressible fluids can be found (Norman & Winkler 1986, Miller & Puckett
1996, Puckett & Saltzman 1992, Saurel & Abgrall 1998). Fluxes are defined
on cell faces of a square mesh with constant grid spacing h = 1x = 1y. The
generalization to three-dimensional domains and to nonuniform rectangular
and parallelepipedic grids can be done with some extra conceptual effort, but
it would unnecessarily complicate this presentation (Popinet et al 1997).
A discrete analog of the characteristic function χ used in Equation 7 is the
scalar field Ci j , known as volume fraction or color function (it actually should
be area fraction in two dimensions, but the standard appellation is volume
fraction). The color function Ci j represents the portion of the area of the cell
(i, j) filled with phase 1:
ZZ
Ci j h ≈
2
χ (x, y) d x d y.
(i, j)
We have 0 < C < 1 in cells cut by the interface S and C = 0 or 1 away from it.
An example of a color function corresponding to a circle arc is shown in
Figure 10. In an incompressible flow, mass conservation is equivalent to con-
servation of volume and hence of the characteristic function χ . There is a vast
literature of numerical methods for conservation laws. However, an explicit
account needs to be taken of the special nature of the problem, which is entirely
concentrated on the interface S. Moreover, particular care has to be given to the
P1: ARS/mbg P2: ARS/vks QC: ARS
November 16, 1998 19:16 Annual Reviews AR075-16
Figure 10 The exact volume-of-fluid color function for a smooth circular arc over a square grid.
constraint 0 < C < 1, because numerical errors in the estimation of the fluxes
can lead to values of C outside the physical range of validity.
VOF methods have been known for several decades and have gone through a
continuous process of improvement. Their use and effectiveness are widespread,
for several reasons:
4. The scheme is local in the sense that only the C values of the neighboring
cells are needed to update the C value in the cell (i, j). For this reason, it
is relatively simple to implement these algorithms in parallel, in particular
within the framework of domain decomposition techniques.
the volume fraction C in that cell and in the neighboring ones. The simplest
types of VOF-methods are the simple line interface calculation (SLIC) of Noh
& Woodward (1976) or the SOLA-VOF algorithm of Hirt & Nichols (1981).
These are first order, O(h), in the accuracy of the reconstruction of the interface.
Typically, the reconstructed interface is made up of a sequence of segments
aligned with the grid, as shown in Figure 11a. The reconstruction is relatively
crude, and its advection, even with simple velocity fields such as translations or
solid body rotations, generates a large amount of flotsam. A few improvements
have been made by several authors (Chorin 1980, Lafaurie et al 1994), but these
features still remain.
Annu. Rev. Fluid Mech. 1999.31:567-603. Downloaded from www.annualreviews.org
More accurate VOF techniques attempt to fit the interface through piecewise
by University of Massachusetts - Amherst on 09/24/12. For personal use only.
linear segments. This was already implemented in some of the earliest VOF
methods (DeBar 1974). The more accurate techniques are now known as the
piecewise linear interface construction (PLIC) method (Ashgriz & Poo 1991,
Li 1995, Parker & Youngs 1992, Rider & Kothe 1995, Pilliod & Puckett 1997,
Rudman 1997, 1998, Rider & Kothe 1998, Gueyffier et al 1998).
To understand the reconstruction problem we have represented in each cell
of Figure 12a possible distribution of the color function on the square grid.
In Figure 12b, it is harder to determine the location of the interface, but a
robust reconstruction method should be able to handle this case too without any
ambiguity.
One of the critical simplifying features of VOF/PLIC algorithms is that one
does not attempt to reconstruct the interface as a chain of joined segments (a
continuous chain of segments) but rather as a discontinuous chain with however
asymptotically small discontinuities.
Whenever the curvature is small (i.e. the radius of curvature is large with
respect to the grid size) the method will be accurate. However, it is robust
in the sense that it does not have a catastrophic behavior when the curvature
increases or even becomes infinite, for instance at reconnection. Specifically,
Figure 11 Two attempts at interface reconstruction. (a) First-order or simple line interface cal-
culation; (b) second-order or piecewise linear interface construction.
P1: ARS/mbg P2: ARS/vks QC: ARS
November 16, 1998 19:16 Annual Reviews AR075-16
Figure 12 Representing the color function as shaded squares of size proportional to the fractional
by University of Massachusetts - Amherst on 09/24/12. For personal use only.
volumes may help to explain the reconstruction problem. (a) A well-behaved distribution of the
color function, corresponding to a smooth circular arc. (b) A more confused distribution.
This estimation of the normal is sometimes called the Parker and Youngs ap-
proach (PY). In two dimensions, the discrete gradient ∇ h C is constructed from
the volume fraction values of a 3 × 3 block of cells centered at (i, j).
An alternate approach is a least-square method (Puckett 1991). The same
block of cells is considered, but now the interface is approximated by a straight
line in the whole block, with the constraint that the volume fraction of the
central cell is always the true value C. By changing the slope of the line, one
can minimize the discrete error between the true value C and the value C̃ given
by the linear approximation. For example, the discrete error E in L 2 norm at
cell (i, j) is given by the expression
Annu. Rev. Fluid Mech. 1999.31:567-603. Downloaded from www.annualreviews.org
à 1 ! 12
by University of Massachusetts - Amherst on 09/24/12. For personal use only.
X
E(m) = (C̃ i+k, j+l (m) − Ci+k, j+l )2
. (25)
k,l=−1
More recently, the ELVIRA approach was proposed (Pilliod & Puckett 1997).
In ELVIRA the slope is chosen among six “candidates” given by the backward,
central, and forward differences of the column sums of the volume fractions
done along the x and the y directions, respectively. The candidate that yields
the lowest value of the error in Equation 25 wins.
The ELVIRA approach reconstructs exactly (with machine precision) a
straight line interface. This method is thus truly second order, O(h 2 ), unlike PY,
which is intermediate between first and second order (Pilliod & Puckett 1997).
Numerical experiments show, however, that at low-to-medium resolution it does
not always give better results than PY, but as the resolution increases, ELVIRA
gives the best results, as should be expected. This typically occurs when there
are more than 20 grid points along the bubble diameter.
In the second part of the reconstruction, a linear interface that divides the
computational cell into two parts containing the proper area of each fluid must
be found. In general, the “forward” problem of finding the area within a square
on each side of a given linear interface is more straightforward than the “inverse”
problem of obtaining the equation for the linear interface, given the fraction
of area contained on each side and the normal direction. Both are needed in
the reconstruction and propagation steps of PLIC. We achieve this by deriving
an explicit expression that relates the “cut” area to a parameter α, which fully
defines the straight line (Gueyffier et al 1998).
In two dimensions, the problem can be stated as follows. Given a square cell
of side h in the (x, y) plane and a straight line (such as EH in Figure 13) with
normal vector m, find the area of the region below the line that also lies within
the square cell. This corresponds to the area ABFGD in Figure 13. To obtain an
expression for this area, let us suppose that the components m x and m y of the
normal m are both positive (this can always be arranged by a simple coordinate
P1: ARS/mbg P2: ARS/vks QC: ARS
November 16, 1998 19:16 Annual Reviews AR075-16
This corresponds to the second term within the square brackets on the right-
hand side of Formula 27, which also contains the Heaviside step function
H (α − m x h), since the area of the triangle BEF is only subtracted if E is
to the right of B. Similarly, the third term within the square brackets in the
formula subtracts the area of the triangle DGH when α > m y h and point H
lies above point D. The single Formula 27 thus provides the area of the region
below the straight line (Equation 26), which lies in the original square of side
h for all possible cases. The area is a continuous, one-to-one, monotonically
increasing function of α. It ranges from zero, when α = 0, to h 2 , when α
reaches its maximum value of (m x + m y )h. There are two critical values of α,
Annu. Rev. Fluid Mech. 1999.31:567-603. Downloaded from www.annualreviews.org
Formula 27, at which the function changes form. This occurs when the straight
line (Equation 26) passes through the corners B and D of the square, i.e. when
α = m x h or α = m y h.
In practice, not only does one need the “forward” Formula 27 between the
cut area and the parameter α, but the method also requires the “inverse” problem
of determining the α that corresponds to a given cut area and normal direction
in a computational cell. There are a number of ways to achieve this. One can
simply use a standard root-finding approach to find the particular value of α at
which the cut area has the desired value, for instance, an iterative method may be
found in Rider & Kothe (1998). Another option is as follows: Corresponding to
each critical value of α for which the interface passes through one of the corners
of the square, there exists a critical value of the cut area. In between any two
critical values, the function on the right hand side of Formula 27 is a known
polynomial in α whose roots can be evaluated analytically. Thus, to resolve the
inverse problem, we first identify which two critical values bound it on either
side and then obtain the root of the correct polynomial in α in that range.
Figure 14 A simple scheme for the split computation of the fluxes. As in the upwind differencing
scheme, the whole block of fractional volume in a band of width uτ is transferred from the upwind
cell to the downwind cell. No account, however, is taken of the change of shape of the interface
during the time step.
Figure 15 A Lagrangian propagation step. The advection of the reconstructed segment is com-
puted, defining a new segment that straddles two cells. From that new segment, fluxes are calculated.
P1: ARS/mbg P2: ARS/vks QC: ARS
November 16, 1998 19:16 Annual Reviews AR075-16
area φ 0 of the fluid contained at the beginning of the step in the control cell
that remains there. The updated volume fraction in each cell after the fractional
step along the x direction is then given by
£ ¤
Ci,(x)j = φi,−j + φi,0 j + φi,+j . (28)
Equation 16 from one element to the next, or as a boundary condition for the
pressure shown in Equation 19.
On fixed grids, there are two options:
1. Smoothing the capillary force: Surface tension is implemented in a simple,
albeit approximate way, by distributing it over neighboring grid points.
2. Accurate finite volume balance of surface tension forces as realized in
Popinet & Zaleski 1998a: This option has been used only for two-dimension-
al calculations because of the complexity of the geometric considerations
involved.
Annu. Rev. Fluid Mech. 1999.31:567-603. Downloaded from www.annualreviews.org
by University of Massachusetts - Amherst on 09/24/12. For personal use only.
It is possible to connect the use of a filtered surface tension with the idea of a
thickened interface (Anderson et al 1998). A thick interface occurs naturally in
the Van der Waals-Cahn-Hilliard theory of interfaces. In this theory, matter is
considered at an intermediate scale between the classical continuum mechanics
scale at which interfaces are sharp and the molecular scale where matter is
discontinuous. At this scale, continuous fields of density and velocity still exist
and interfaces have a finite thickness. An example of a dynamical equation
with thick interfaces for a one-species, liquid-vapor flow as well as numerical
simulations of that model may be found in Nadiga & Zaleski (1996). For two
incompressible phases, the Van der Waals-Cahn-Hilliard framework was used
Annu. Rev. Fluid Mech. 1999.31:567-603. Downloaded from www.annualreviews.org
Seppecher (1996).
6.2 Momentum Conservation
The smoothing method makes it possible to find an approximation of the sin-
gular tensor defined in Equation 12,
T = σ (I − n ⊗ n)||∇ C̃||, (32)
where the unit normal n is
∇ χ̃
n= . (33)
||∇ χ̃||
The momentum-conserving formulation is a finite difference approximation
of Equation 32. In practice, a 4- or 6-point approximation was used for the
computation of the normals.
As in the method used by Brackbill et al (1992), this approximation con-
verges in theory toward the true normal as H → δ S . When the kernel H² is
narrowed one should keep m f large in order to have h ¿ ². However, numer-
ical experiences have shown that the best results are obtained when m f = 1
or 2. Larger values of m f increase the spurious currents.
6.3 Other Methods for Surface Tension
The direct measurement of normal vectors and their derivatives is possible when
the interface is tracked by markers or when it is an element boundary.
We have tested a method of surface markers together with the MAC finite
difference method (Popinet & Zaleski 1998a). The surface stresses were again
computed and integrated over the boundary of a finite volume for the x and
y momentum. When carefully implemented, this method makes it possible to
considerably reduce the spurious currents.
6.4 Capillary Waves
The simulation of capillary waves is an important test for the interface simula-
tion methods in general. It has been performed in detail by Fyfe et al (1988) for
P1: ARS/mbg P2: ARS/vks QC: ARS
November 16, 1998 19:16 Annual Reviews AR075-16
their free Lagrangian method. Their computations were carefully analyzed for
possible errors coming from finite domain size, nonlinear effects, and viscous
effects. High levels of accuracy were obtained for waves on planar interfaces,
but the accuracy was less for capillary oscillations of cylinders (two-dimensional
flows around circular patches). For the conservative method described above,
tests were presented by Gueyffier et al (1998). For comparisons with the level-
set method see Sussman et al (1994) and Sussman & Smereka (1997). For all
these fixed-grid methods the accuracy is of the order of 1%. Marker meth-
ods with smoothed surface tension also give comparable results (Nobari et al
1996). However, somewhat more accuracy, of the order of 0.5%, may be ob-
Annu. Rev. Fluid Mech. 1999.31:567-603. Downloaded from www.annualreviews.org
with viscosity. This means that either the capillary number Ca = µU/σ is small
or that the Laplace number La = σρ R/µ2 is large. Numerical experiments
show that at small Ca, but moderate La, no catastrophic instability occurs,
and that the only troublesome effects are the spurious currents. However, at
large La, parasite currents may grow and become large enough to destroy the
interface. This may be easily explained by the connection between La and
Reynolds numbers of spurious currents given by Equation 34,
La = 100 Re p , (35)
Annu. Rev. Fluid Mech. 1999.31:567-603. Downloaded from www.annualreviews.org
∼106 . This is approximately the value for a 1-cm droplet or bubble. Thus, this
droplet size represents the borderline for current-day simulations.
For larger droplets and Reynolds numbers, one needs to compare surface
tension and inertia. At small Weber numbers, We = U 2 ρ R/σ , capillary ten-
sion dominates inertial phenomena. Since this happens for smaller R, it may
explain why several articles and private communications report difficulties with
smaller particles.
have been carried out for this physical situation using the VOF method by
by University of Massachusetts - Amherst on 09/24/12. For personal use only.
Gueyffier et al (1998) and the results show an excellent agreement with the
pictures of Peregrine et al (1990).
Figure 16 Ratio between sedimentation velocity of an infinite array of drops and Stokes velocity
for a single drop.
Figure 17 Droplet formation in a 2D sheared layer. The parameters are ρ L /ρG = 10, ReG =
7500, and W eG = 150. The liquid phase is the grey area. The lines are iso-vorticity levels.
Figure 18 Droplet formation at higher capillary and Weber numbers. The intermediate steps are
not shown. The level lines of the vortices are omitted in b to avoid overcrowding the plot.
P1: ARS/mbg P2: ARS/vks QC: ARS
November 16, 1998 19:16 Annual Reviews AR075-16
Figure 19 Breakup of the interface through a secondary instability of the sheet. ρ L /ρG =
10, Re = 1000, and W e = 300.
authors have attempted to solve the full Navier-Stokes equations. The dissipa-
tion of surface waves was studied by Yang & Tryggvason (1998). Wave breaking
and the mixing of the fluid from the plunging jet was studied by Abadie & Calt-
agirone (1998). Breaking waves with plunging jets and several splash-ups were
simulated by Chen et al (1998).
Applications of surface marker methods to biology are a promising area of
research. One example in this direction is Agresar et al (1998). Boiling flow is
also a new application of surface-tracking methods (Juric & Tryggvason 1998).
Finally, bubbly flow with a large number of bubbles has been intensively
studied both in two and three dimensions by several groups (see Esmaeeli &
Tryggvason 1996, 1998a,b; Tomiyama 1998).
P1: ARS/mbg P2: ARS/vks QC: ARS
November 16, 1998 19:16 Annual Reviews AR075-16
8. CONCLUSION
In our opinion, the main progress witnessed over the past 10 years is the advent
of many three-dimensional calculations. Two-dimensional calculations have
also matured, producing either highly accurate results or calculations over im-
Annu. Rev. Fluid Mech. 1999.31:567-603. Downloaded from www.annualreviews.org
The one new method that appeared is level sets. It is seducing by its sim-
plicity but has not yet produced the wide range of results, especially in three-
dimensions, achieved by older methods. The VOF method has been consider-
ably improved by the systematic use in the scientific literature of the higher
order piecewise linear interface calculation method.
Important calculations have not yet been done. For instance, there is no
three-dimensional calculation of a falling drop or rising bubble in the difficult
air/water conditions and in the nontrivial regimes where axial symmetry does
not hold. Many complex flows, such as atomizing jets, splashing droplets, or
breaking waves, have been studied only in a preliminary way. Finally, for some
flows, quantitative results are difficult to get, in particular, agreement between
simulations and linear theory for shear flow instabilities.
For engineering applications, new methods need to be developed that could
adapt to complex geometries, such as combustors and nozzles, to cite just two
of a whole world of examples.
It is our opinion that these difficulties should soon be resolved. This would
allow these methods to connect directly with scientific problems and engineer-
ing applications in multiphase flow. From there the simulation of interfaces
could move on to incorporate more complex physics, such as models of inter-
molecular forces, complex surface rheology, and the transport and effects of
tensio-active molecules.
Literature Cited
Abadie S, Caltagirone J. 1998. Breaking waves. and adhesion of circulating cells. J. Comput.
Paris: C. R. Acad. Sci. In press Phys. 143:346–80
Agresar G, Linderman J, Tryggvason G, Pow- Aleinov I, Puckett EG. 1995. Computing sur-
ell K. 1998. An adaptive, Cartesian Front- face tension with high-order kernels. In Proc.
tracking method for the motion, deformation 6th Int. Symp. Comput. Fluid Dyn. ed. K
P1: ARS/mbg P2: ARS/vks QC: ARS
November 16, 1998 19:16 Annual Reviews AR075-16
Oshima, pp. 6–13. Lake Tahoe, CA. May axisymmetric free-surface flow. Phys. Rev.
be obtained from the author’s web page at Lett. 71:3458–61
http://math.math.ucdavis.edu/∼egp/ Eggers J. 1997. Nonlinear dynamics and break-
Anderson DM, McFadden GB, Wheeler AA. up of free-surface flows. Rev. Mod. Phys.
1998. Diffuse-interface methods in fluid me- 69:865–929
chanics. Annu. Rev. Fluid Mech. 30 Eggers J, Dupont TD. 1994. Drop forma-
Ashgriz N, Poo JY. 1991. FLAIR: Flux line- tion in a one-dimensional approximation of
segment model for advection and interface the Navier-Stokes equation. J. Fluid Mech.
reconstruction. J. Comput. Phys. 92:449–68 262:205–21
Baker GR, Meiron DI, Orszag SA. 1982. Gener- Ervin E, Tryggvason G. 1997. The rise of bub-
alized vortex methods for free surface flows bles in a vertical shear flow. ASME J. Fluid
problems. J. Fluid Mech. 123:477 Eng. 119:443–49
Baker GR, Meiron DI, Orszag SA. 1984. Esmaeeli A, Tryggvason G. 1996. An inverse
Boundary integral methods for axisymmetric energy cascade in two-dimensional, low Rey-
Annu. Rev. Fluid Mech. 1999.31:567-603. Downloaded from www.annualreviews.org
and three dimensional Rayleigh-Taylor insta- nolds number bubbly flows. J. Fluid Mech.
bility problems. Physica D 12:19–31 314:315–30
by University of Massachusetts - Amherst on 09/24/12. For personal use only.
Batchelor GK. 1967. An Introduction to Fluid Esmaeeli A, Tryggvason G. 1998a. Direct nu-
Dynamics. Cambridge: Cambridge Univ. merical simulations of bubbly flows. Part I.
Press Low Reynolds number arrays. Submitted for
Bell J, Dawson C, Shubin G. 1990. An un- publication
split, higher order Godunov method for scalar Esmaeeli A, Tryggvason G. 1998b. Direct nu-
conservation laws in multiple dimensions. J. merical simulations of bubbly flows. Part II.
Comput. Phys. 74:171–200 Moderate Reynolds number arrays. Submit-
Benzi R, Succi S, Vergassola M. 1992. The lat- ted for publication
tice Boltzmann equation: theory and appli- Floryan JM, Rasmussen H. 1989. Numeri-
cations. Phys. Rep. 222:145–97 cal methods for viscous flow with moving
Boulton-Stone JM, Blake JR. 1993. Gas bub- boundaries. Appl. Mech. Rev. 42:323–40
bles bursting at a free surface. J. Fluid Mech. Fritts MJ. Cowley W, Trease HE, eds. 1985. The
254:437–66 Free Lagrange Method. Lect. Notes Phys.
Brackbill J, Kothe DB, Zemach C. 1992. A con- Vol. 238. New York: Springer-Verlag
tinuum method for modeling surface tension. Fukai J, Shiiba Y, Yamamoto T, Miyatake O,
J. Comput. Phys. 100:335–54 Poulikakos D, et al. 1995. Wetting effects on
Chen G, Kharif C, Li J, Zaleski S. 1998. the spreading of a liquid droplets colliding
Two-dimensional Navier-Stokes simulation with a flat surface: experiment and model-
of breaking waves. Submitted for publication ing. Phys. Fluids 7:236–47
Chorin AJ. 1980. Flame advection and propa- Fukai J, Zhao Z, Poulikakos D, Megaridis CM,
gation algorithms. J. Comput. Phys. 35:1–11 Miyatake O. 1993. Modeling of the deforma-
Collela P. 1990. Multidimensional upwind tion of a liquid droplet impinging on a flat
methods for hyperbolic conservation laws. J. surface. Phys. Fluids A 5:2588–99
Comput. Phys. 87:171–200 Fyfe DE, Oran ES, Fritts MJ. 1988. Surface ten-
Coward AV, Renardy YY, Renardy M, Richards sion and viscosity with Lagrangian hydro-
JR. 1997. Temporal evolution of periodic dis- dynamics on a triangular mesh. J. Comput.
turbances in two-layer Couette flow. J. Com- Phys. 76:349–84
put. Phys. 132:346–61 Glimm J, Bryan OM, Menikoff R, Sharp D.
Cuenot B, Magnaudet J, Spennato B. 1997. 1986. Front tracking applied to Rayleigh-
The effects of slightly soluble surfactants on Taylor instability. SIAM J. Sci. Stat. Comput.
the flow around a spherical bubble. J. Fluid 7:230–51
Mech. 339:25–53 Glimm J, McBryan O, Menikoff R, Sharp DH.
Dandy DS, Leal LG. 1989. Buoyancy-driven 1987. Front tracking applied to Rayleigh-
motion of a deformable drop through a qui- Taylor instability. SIAM J. Sci. Stat. Comput.
escent liquid at intermediate Reynolds num- 7:230–51
bers. J. Fluid Mech. 208:161–92 Gueyffier D, Nadim A, Li J, Scardovelli R, Za-
DeBar R. 1974. Fundamentals of the KRAKEN leski S. 1998. Volume of fluid interface track-
code. Tech. rep. UCIR-760. Lawrence Liver- ing with smoothed surface stress methods for
more Nat. Lab. three-dimensional flows. Submitted for pub-
de Gennes PG. 1985. Wetting: statics and dy- lication
namics. Rev. Mod. Phys. 57:827–63 Gunstensen AK. 1992. Lattice-Boltzmann stud-
Duraiswami R, Prosperetti A. 1992. Orthogo- ies of multiphase flow through porous media.
nal mapping in two dimensions. J. Comput. PhD thesis. MIT, Cambridge, MA
Phys. 98:254–68 Harlow F. 1964. The Particle-in-Cell Comput-
Eggers J. 1993. Universal pinching of 3D ing Method for Fluid Dynamics, In Methods
P1: ARS/mbg P2: ARS/vks QC: ARS
November 16, 1998 19:16 Annual Reviews AR075-16
of Computational Physics, Vol. 3, pp. 313– dynamics. Annu. Rev. Astron. Astrophys. 30:
43. New York: Academic 543–74
Hirt CW, Amsden AA, Cook JL. 1974. An arbi- Monaghan J. 1994. Simulating free surface
trary Lagrangian Eulerian computing method flows with SPH. J. Comput. Phys. 110:399–
for all speeds. J. Comput. Phys. 14:227–53 406
Hirt CW, Nichols BD. 1981. Volume of fluid Monaghan J, Bicknell P, Humble R. 1994. Vol-
(VOF) method for the dynamics of free boun- canoes, tsunamis and the demise of the Mi-
daries. J. Comput. Phys. 39:201–25 noans. Physica D 77:217–28
Hirt CW, Nichols BD. 1988. Flow-3D users Nadiga BT, Zaleski S. 1996. Investigations of
manual. Tech. Rep. Flow Sciences, Inc. a two-phase fluid model. Eur. J. Mech. B 15:
Hoyt JW, Taylor J. 1977. Waves on water jets. 885–96
J. Fluids Mech. 83:119–27 Nichols BD, Hirt CW, Hotchkiss RS. 1980.
Jacqmin D. 1996. An energy approach to the SOLA-VOF: A solution algorithm for tran-
continuum surface tension method. AIAA sient fluid flow with multiple free boundaries.
Annu. Rev. Fluid Mech. 1999.31:567-603. Downloaded from www.annualreviews.org
CONTENTS
Linear and Nonlinear Models of Aniosotropic Turbulence, Claude
1
Cambon, Julian F. Scott
Transport by Coherent Barotropic Vortices, Antonello Provenzale 55