Comparison between the diffuse interface and volume of fluid methods for

simulating two-phase flows

Shahab Mirjalilia,∗, Christopher B. Iveyb , Ali Mania

a Department of Mechanical Engineering, Stanford University, Stanford, CA 94305, USA
b Cascade Technologies, Inc., Palo Alto, CA 94303, USA
arXiv:1803.07245v1 [physics.flu-dyn] 20 Mar 2018

Abstract
A wide variety of interface capturing methods have been introduced for simulating two-phase flows through-
out the years. However, there is a noticeable dearth of literature focusing on objective comparisons between
these methods, especially when they are coupled to the momentum equation and applied in physically
relevant regimes. In this article, we compare two techniques for simulating two-phase flows that possess at-
tractive qualities, but belong to the two distinct classes of diffuse interface (DI) and volume of fluid (VOF)
methods. Both of these methods allow for mass-conserving schemes that can naturally capture large interfa-
cial topology changes omnipresent in realistic two phase flows. The DI solver used in this work is based on a
conservative and bounded phase field method, developed recently. Similar to level set methods, this diffuse
interface method takes advantage of the smoothness of the phase field in computing curvature and surface
tension forces. Geometric VOF methods track the fractional tagged volume in a cell. The specific geometric
VOF scheme used here is a discretely conservative and bounded implementation that uses geometric algo-
rithms for unsplit advection and interface reconstruction, while employing height functions for normal and
curvature calculation. We present a quantitative comparison of these methods on Cartesian meshes in terms
of their accuracy, convergence rate, and computational cost using canonical two-dimensional (2D) two-phase
test cases: a very dense drop moving through a quiescent gas, the Rayleigh-Taylor instability, an equilibrium
static drop, an oscillating drop and the damped surface wave. We further compare these methods in their
ability to resolve thin films by simulating the impact of a water drop on a deep water pool. Using results of
these studies, we suggest qualitative guidelines for selection of schemes for two-phase flow calculations.
Keywords: diffuse interface, volume of fluid, two-phase flow

1. Introduction

In general, numerical methods for simulating immiscible two phase flows can be split into front tracking
and front capturing methods [1]. For example, in the realm of front tracking methods, Unverdi and Tryg-
gvason [2] coupled the Navier-Stokes equations with the use of marker points for following the immiscible
fluid interfaces. In these methods, marker points are advected and marker functions are reconstructed from
their location. We refer the reader to a more comprehensive overview of such methods provided in [3].
These methods are not easy to adopt for capturing many realistic flows where severe topological changes
can occur at the interfaces. The most popular methods among the front capturing schemes, which require
no additional treatment to capture topological changes, are categorized into volume of fluid methods, level
set methods and phase field (diffuse interface) methods. The interested reader is referred to [1] for a detailed
review of interface capturing methods.

∗ Corresponding author: Phone: 650-561-5951

Email addresses: [email protected] (Shahab Mirjalili), [email protected] (Christopher B. Ivey),
[email protected] (Ali Mani)

Preprint submitted to Journal of Computational Physics March 14, 2018

 Volume of fluid (VOF) methods are the oldest and most commonly used class of two-phase CFD methods
to this day [4, 5, 6]. These methods can be categorized based on their advection and reconstruction routines
[7]. While algebraic VOF methods are simple and do not require tedious tracking of intersections, geometric
VOF methods have become more widespread due to their higher accuracy. The piecewise-linear interface
calculation (PLIC) VOF schemes have become particularly popular [8] because they can leverage geometric
and analytic routines for the advection and reconstruction steps. This allows for the development of efficient
toolboxes for performing these operations in an accurate and conservative manner. The advection operation
for PLIC-VOF can be either directionally split or unsplit. Directionally split advection routines are simpler
to implement, but incur extra cost due to their strict time-step requirement and their need for separate
advection and reconstruction operations for each dimension. Moreover, split advection schemes are not
suitable for unstructured meshes, which are critical for realistic two-phase systems. Unsplit methods on the
other hand, require only one advection and reconstruction step per time-step [9, 10]. It was shown in [11]
that unsplit VOF advection schemes are more accurate than their split counterparts when handling sharp
interfacial features. In addition, unsplit methods are more suitable for unstructured meshes, evidenced by
the development of accurate unsplit advection schemes for arbitrary non-convex polyhedral meshes [12, 13].
On the downside, unsplit schemes require expensive computation of complex flux polyhedra in interfacial
cells. This can potentially lead to difficulties with parallel load balancing [14]. The computations during
the reconstruction step of PLIC-VOF methods are mostly spent on accurate estimation of normal vectors
after volume fraction advection is completed. The discontinuous nature of the volume fraction field presents
a challenge for calculation of the normal vectors. Thus, several methods for calculation of normal vectors
(and interface reconstruction based on them) have been proposed. While many schemes have emerged
[15, 16, 11, 17, 18], these methods either do not offer second order accuracy or achieve that through iterative
costly geometric operations. A well-known non-iterative method for estimating the normal vectors is the
height-function (HF) technique [19, 20]. This method has received significant attention recently. In [21], a
framework was presented for extending the HF method to 3D unstructured non-convex polyhedral meshes
with application to PLIC-VOF. Their proposed method, which uses embedded height-functions (EHF)
calculates interface normals with second order accuracy, with much less cost than other existing second
order accurate methods.
Level set methods were introduced by [22] and further developed by many other authors. The idea
of using level set methods for capturing interfaces in two-phase flows was introduced by [23]. Level set
methods use the signed distance function to define an interface location. In a subsequent interface advection
process, the level set field diverges from the signed distance function, potentially deteriorating the accuracy
of the representation. To maintain accuracy, the values of the level set function are periodically reinitialized
to represent signed distances but with minimal intrusion to the physical interface location. Due to the
smoothness of the field, standard numerical schemes can be applied in the advection step of these methods
and they can naturally handle topological changes. Moreover, since the level set function (a signed distance
function) is a smoothly varying field, it can be conveniently used to obtain the normal and curvature values
for the calculation of surface tension forces. While level set methods continue to be popular for two-phase
flow simulations, inherent lack of mass conservation can be problematic for simulations performed on coarse
meshes, two-phase flows with small features, and simulations with long time horizons. Improvements by [24]
and [25] have reduced the mass loss by introducing a refined auxilliary mesh and combining with a PLIC-
VOF method, respectively. Nevertheless, these methods introduce complexities and extra costs that the
original level set methods did not have. Another solution was given by [26] and [27], where they presented
the conservative level set approach. This method conserves the mass of each phase and preserves most of
the desirable properties of the original level set method. Converse to VOF methods and traditional level set
methods, conservative level set (CLS) is a diffuse interface method where the interface acquires a hyperbolic
tangent profile in the interface-normal direction. Advection of the interface distorts this profile shape and
reinitialization is required to correct the deviations and restore the desired shape of the interface after any
advection step.
Phase field methods are a class of interface capturing schemes that have recently become popular for
simulating immiscible two-phase flows [28, 29, 30, 31, 32, 33, 34]. These methods which are also often known
as diffuse interface methods do not require any reinitialization step, unlike CLS methods. In other words, one
2
only needs to time-integrate the phase field and momentum equations at each time-step of these methods.
Typically, the Allen-Cahn or Cahn-Hilliard [35] equations are used in phase field methods. These methods
admit discrete energy laws that guarantee boundedness of total energy in the domain as explained in [33] and
[36]. While the Allen-Cahn method is not particularly suitable for two-phase flow simulations due to a lack
of mass conservation, the Cahn-Hilliard equation is mass-conserving. However, the fourth-order derivative
in the Cahn-Hilliard phase field equation incurs numerical difficulties. Furthermore, even at the PDE level,
the Cahn-Hilliard equation produces shifts in the value of the phase function far from the interface, making
it rather unsuitable for large density ratios [37]. Recent works have therefore sought to combine Allen-Cahn
and Cahn-Hilliard equations in new phase field models that possess the advantages of both equations but
avoid their issues [31, 38]. The nonconservative phase field model proposed in [31], and the CLS method
inspired the conservative second order phase field method of [38]. Their method conserved the total mass
similar to CahnHilliard equation, while being simpler to implement. However, in spite of a dispersion-
relation-preserving upwind scheme developed for the convection-diffusion equation, the discretized phase
field variable would not remain bounded between the values for the two pure phases. Numerical overshoots
and undeershoots were handled via unphysical mass-clipping and mass-redistribution algorithms. In [39],
a space-time discretization for the phase field equation of [38] was introduced for staggered grids. It was
analytically and numerically shown that with appropriate selection of the free parameters in the phase field
equation, time-integration of the phase field would always guarantee bounded values for the phase field. The
resulting phase field (diffuse interface) method avoids diffusive upwinding, unphysical mass redistribution
and interface reinitialization.
Before specifying the two interface capturing methods we have chosen to compare, an overview of surface
tension force implementations is due. As described in [40], surface tension forces can be either implemented
as stresses or body forces, respectively known as integral and volumetric formulations. While integral
formulations conserve the momentum of the system automatically, no version of this type of formulation is
yet to be found that allows for the discrete balance of pressure gradients and surface tension forces. This
imbalance leads to the unwelcome creation of spurious currents. Volumetric formulations on the other hand,
are amenable to balanced force methods that eliminate spurious currents if curvature is computed exactly
[18]. Hence, volumetric formulations are adopted for modeling capillary forces in most two-phase solvers.
For all such methods, calculation of normal vectors and subsequently curvature values is an important step
that can significantly affect accuracy of solutions. Curvature of an interface can be defined (and computed)
as the rate that the normal vector turns along the interface. As a result, accurate calculation of the
normal vectors is required for accurate prediction of curvatures and surface tension forces. Two of the most
popular methods for implementation of surface tension forces in VOF solvers are the convolved VOF [41]
and reconstructed-distance function [41, 42] methods. Such methods however do not converge under mesh
refinement. Alternatively, height functions can be employed for computing curvatures [41]. The height
function (HF) method achieves second order convergence on uniform Cartesian meshes [43, 44, 45]. In [20],
a second order HF technique was presented for quad and octree discretizations. For general 3D unstructured
non-convex polyhedral meshes, the embedded height function (EHF) technique by [21] provided the first
convergent method for computing curvatures on unstructured grids. Curvature evaluation is much more
straight-forward in level set and phase field methods. In such methods, a smooth field is available from which
the computation of normal vectors and curvature is possible via spatial differentiation. After computing
normal vectors and curvatures the surface tension force is often implemented as a continuum surface force
[42] or by way of ghost fluid method [46].
In this work, we compare the accuracy and computational cost of two recently developed VOF and DI
two-phase solvers using a variety of canonical problems. In particular, our tests assess the performance of
the fully coupled two phase flow solvers, where at each time-step in addition to the interface transport step,
the coupled momentum equation is solved. The first solver uses a finite volume PLIC-VOF method with
unsplit advection [13]. This approach is second-order accurate and discretely conservative and bounded on
general non-convex polyhedral meshes. For surface tension force calculation in this method, the embedded
height function (EHF) of [21] is used. The second solver uses a finite difference phase field method presented
in [39]. The phase transport equation is coupled to Navier-Stokes and the continuum surface force (CSF)
method is implemented for surface tension forces. More details on these solvers is provided in Section 2.
3
 Although many numerical recipes have been developed for handling realistic two-phase flow simulations,
there is a lack of understanding with regard to the relative cost-accuracy trade-off of all these methods.
Comparisons of different components of two-phase flow solvers are common. For example, interface capturing
schemes have been compared via standard advection tests, with prescribed velocity fields [11, 47, 48, 49].
Moreover, surface tension implementations have been compared in articles such as [41] and [50]. However,
studies like [51] and [52], where the relative performance of two full two-phase solvers are compared, are
scarce. In [51], the number of test cases were limited and the selected methods were quite similar. In [52],
by using a common Navier-Stokes solver, 3D tests were used to compare the performance of four different
interface capturing schemes with similar curvature calculation methods utilizing level sets. In their study,
however, all four schemes tested were VOF methods, and because of the introduction of level set all schemes
involved some mass conservation errors. The selected DI and VOF methods considered in the present study
are both mass-conservative interface capturing schemes. Furthermore, they are on opposite ends of the
spectrum in terms of cost and implementation complications.
Although both of these methods are bounded, mass conserving front-capturing schemes that are capable
of automatically representing large interfacial changes, some clear differences between the two approaches
should be emphasized. The VOF method is a sharp interface approach. On the other hand, the DI technique
thickens a realistic phase transition zone from a few nanometers to few cells width. Figure 1 depicts the
representation of a drop with diameter 0.5 in a 1 × 1 domain using the DI and VOF methods at two
resolutions. It is clear that the interface location is approximately captured in a cell using discontinuous
planes in the VOF method, while the DI method smears the interface over a few cells. Figure 1 gives the
impression that VOF is more accurate than DI. Indeed, we will show that at the same resolution, VOF is
more accurate than DI. On the other hand, VOF is much more expensive than DI. In this work the central
theme is to clarify that for different types of problems which of the two methods is more accurate at equal
cost levels.
In what follows, we first introduce the two specific DI and VOF numerical schemes chosen for comparison
in Section 2. Section 3 compares the accuracy and cost of the these numerical methods on several canonical
two-phase problems. From the results of these studies, we make recommendations for using these methods
to simulate two-phase flows. Section 4 concludes the article with a summary of our findings. Appendix A
presents classical curvature and advection tests for the DI method. For details on the VOF method, we refer
the interested reader to [53].

2. Numerical Methods

2.1. Volume of Fluid Method

Piecewise-linear interface calculation (PLIC) VOF methods describe the interface by a series of discon-
nected planes, each oriented by a unit normal, n̂, and positioned by a constant, C, such that

~n · ~x + C = 0, (1)

where ~n points outward with respect to the reference phase, and C enforces the volume fraction in the
cell [54]. PLIC-VOF methods have the potential to provide discrete mass conservation, which is important
to many two-phase engineering applications including those involving large density ratios, such as breaking
waves, and those involving wide range of length scales associated with secondary breakups to disperse phases.
In our adopted methodology, the equations governing the PLIC-VOF scheme are as follows. The volume
fraction, which is advected geometrically using the non-intersecting flux polyhedron advection (NIFPA)
scheme described in [55] is governed by:
∂f
+ ∇ · (f~u) = 0 (2)
∂t
The momentum equation is given by
∂ρ~u


+ ∇ · (ρ~u ⊗ ~u) = −∇P + ∇ · µ ∇~u + ∇~u> + ρ~g + T~ , (3)
∂t
4
 0.8 0.9
0.7 0.8

0.6 0.7
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2 0.2
0.1 0.1

(a) DI 10 × 10 (b) DI 20 × 20

(c) VOF 10 × 10 (d) VOF 20 × 20

Figure 1: A drop with R = 0.25 placed in a 1 × 1 domain for DI and VOF at two different resolutions

5
where the surface tension force per volume is

T~ = σκ~nδΣ , (4)

in which δΣ is a numerical Dirac delta function. The convective flux in Equation 3 is computed in a
geometric manner consistent with the geometric operator used for volume fraction advection. A balanced
force discretization of surface tension forces is used to allow for pressure gradients to balance surface tension
forces at equilibrium and reduce spurious currents. Curvature values and the normal vector are computed
using embedded height functions (EHF) and utilized in computing the surface tension force per volume, T~ ,
which is implemented as a continuum-surface-force (CSF) as shown in Equation 4. In addition, density and
viscosity fields are coupled to Equation 2 via

ρ = (ρl − ρg )f + ρg , (5)

and
µ = (µl − µg )f + µg . (6)
The flow is incompressible,
∇ · ~u = 0, (7)
and an adaptation of the fractional-step time integration method of [56] was used for second order time
integration and enforcement of Equation 7. It is noteworthy that this novel discretization of the one-fluid
formulation discretely conserves mass, and in the absence of viscous and capillary effects, momentum.

2.2. Diffuse Interface Model

Similar to [26], [27] and [38], the phase field chosen for our diffuse interface method has a hyperbolic
tangent shape given by:
1 s
φ = [1 + tanh( )], (8)
2 2
where s is the signed distance from the interface and  is a smallness parameter, setting the interface
thickness. Moreover, φ represents the phase field. Note that the hyperbolic tangent interface profile chosen
here is the same as the equilibrium profile given by the Cahn-Hilliard equations [35, 33]. However, as
shown by [38], the same equilibrium solution can be captured by a second order evolution equation. For an
incompressible flow, this equation is given by
  
∂φ ∇φ
+ ∇ · (~uφ) = ∇ · γ ∇φ − φ (1 − φ) . (9)
∂t |∇φ|
We note that the conservative form of this equation naturally ameliorates the problem of mass loss in
two-phase flows as compared to level set methods. In other words, the total amount of φ is conserved
globally. The two flux terms on the right hand side are numerical contrivances responsible for preserving the
shape of the interface. The second term sharpens the interface while the first term opposes this action by
diffusing the phase field, giving a transition region of fixed width. Both terms are multiplied by a common
factor related to the maximum velocity in the domain to ensure the interface is held together even in the
harshest strain conditions [38],[27], [39]. Both of these constants are set based on findings in [39] such that
the discrete solution to the phase field remains bounded and conservative throughout the simulation while
taking advantage of second order central finite differences. For simulations presented in this work, we adopt
 = ∆x and γ = |~u|max where |~u|max is the magnitude of the maximum velocity in the domain.
In a two-phase flow calculation, Equation 9 must be solved coupled with the Navier-Stokes equation
to obtain velocity fields. Surface tension force is included in the Navier-Stokes equations in the form of a
continuum surface force [42, 26]. For this, curvature values must be calculated accurately by taking the
divergence of the normal vector field. Moreover, the density and viscosity fields are linearly related to the
phase field. Thus for an incompressible flow

∇ · ~u = 0, (10)
6
and the Navier-Stokes equations can be written as

∂~u 1
+ ∇ · (~u ⊗ ~u) = −∇P + ∇ · [µ(∇~u + ∇T ~u)] + ρ~g + σκ∇φ , (11)
∂t ρ

where the normal vector (~n) is simply

∇φ
~n = , (12)
|∇φ|
curvature (κ) is computed as
κ = ∇ · ~n, (13)
density is linear with respect to φ
ρ = (ρl − ρg )φ + ρg , (14)
and similarly for viscosity we have
µ = (µl − µg )φ + µg . (15)
ρl , ρg and µl , µg are the densities and viscosities of the two fluids, respectively. Additional implementation
details include:
• A staggered grid is used where P ,φ,ρ and µ are stored at cell centers. Velocity values are stored
at the center of their corresponding cell faces. This is a standard approach that results in desirable
conservation properties when central difference schemes are used [57].

• We use second order central differences in space to have mass and momentum conservation (in the
absence of surface tension forces) in addition to boundedness of φ. As explained in [39] by choosing
the appropriate RHS parameters( and γ) for Equation 9, desirable conservation properties can be
retained while upwinding schemes or complicated methods for the advection of the phase field are
avoided.

• A second order three stage Runge-Kutta stepping scheme adopted from [58] is used for time stepping.
This time stepping scheme only requires one Poisson solve per time step.
• The convective term in the momentum equation is in divergence form.
• Since there are no overshoots or undershoots in the phase field equation, no mass-redistribution is
required.

2.3. Convergence of Diffuse Interface Methods

When performing convergence studies on diffuse interface methods, the behavior of two sources of error
should be taken into consideration [33, 39]. First, there are discretization errors between the numerical
solution and exact solution of Equations 9,10,11,13,14,15 which we will refer to as E∆ . To be clear, this
error is because of numerical truncations in approximating the continuous diffuse equations. Second, errors
are incurred by approximating a sharp interface with a smoother transitional zone which we denote as E .
In order to make a diffuse interface simulation converge to sharp interface dynamics one needs to reduce the
thickness of the interface () and the mesh spacing (∆x) during convergence studies. Moreover, as explained
in [33] and [39], optimal convergence is obtained when both errors scale the same way. The dependence of
E∆ and E on , γ and ∆x is however problem dependent. In [39], these errors were studied for the very
simple problem of advection of a 1D droplet with a prescribed velocity field. It was found that for that
problem, optimal convergence was obtained when /∆x was kept fixed while the mesh was refined, resulting
in a fixed number of mesh points across the interface for different resolutions. Unless explicitly mentioned,
we will perform mesh refinement in this manner for the canonical test cases presented herein.

7
2.4. Cost Comparison
An objective and accurate comparison of different numerical methods is a delicate task. There are many
challenges in comparing the cost of two codes that consist of multiple subroutines and employ different
discretizations and frameworks. To ensure that a fair comparison between the two solvers is performed we
have taken the following steps:
• For all the canonical problems (all cases except the droplet impact problem), the simulations were
performed on one single processor on the same computing machine.
• While the cost of the DI solver depends almost entirely on the size of the problem (not the density ratio
or dynamics), the routines employed in the VOF simulations loop over the interfacial cells, making
the run times problem dependent and mainly a function of the number of interfacial cells. Therefore,
for each specific problem, we exactly match the input parameters for the two solvers (e.g. densities,
surface tension, viscosities and initial conditions), and document the cost for the specific physical
problem. This is in contrast to reporting cost based on scaling relations from generic simulations.
• In this work, by cost we are referring to total run time for performing a simulation. In some figures,
cost per time-step, which is the run time for one time-step is compared for the different methods.
• In all of the cost figures the relative cost of simulating each case with respect to the coarsest DI
calculation (cheapest simulation) is plotted.
• For the droplet impact problem, we use the same clusters to perform the parallel simulations. Overall
cost is presented in terms of relative CPU-Hours of computation.
• Generally speaking, the computational cost of an unstructured flow solver is larger than that of an
equivalent structured flow solver because the memory layouts of structured flow solvers can be lever-
aged by the machine (e.g. reduce cache misses). However, the arithmetic intensity of the geometric
VOF scheme of Ivey and Moin [13] exceeds that of the momentum solve and curvature computation
(the embedded height function scheme of Ivey and Moin [13] reduces to the traditional height function
method on uniform hexahedral meshes). The geometric advection method is Lagrangian in nature
and has similar costs on both structured and unstructured hexahedral meshes. Therefore, any com-
putational savings provided by a structured memory layout is expected to be negligible compared to
the geometric VOF method. As such, the unstructured nature of the VOF solver does not invalidate
findings described herein.
• For Cartesian grids, the computational cost of performing VOF simulations in this work is dominated
by the computational geometric operations in the advection step. These operations are performed for
interfacial cells and computational cost scales with the number of such cells. On the other hand, the
cost of the diffuse interface method is dominated by the Poisson solve required to find pressure, and
we observe this cost to scale with the total number of cells in the domain. The relative cost of the
two methods is problem dependent. For an N × N two-dimensional domain, if the interface remains
local and limited to a narrow region of the domain, the cost of performing one time-step of VOF and
DI simulations can be approximated as CV OF N and CDI N 2 respectively, where CV OF and CDI are
constants and CV OF  CDI . If the interfacial zones are distributed throughout the domain, like in
droplet laden turbulence, bubbly flows or turbulent breaking waves, the cost of one time-step of VOF
would then be approximated by C 0 V OF N 2 where again C 0 V OF  CDI . This has two implications.
First, for cases where the interface is limited to a specific portion of the domain, the relative cost of
VOF compared to DI shrinks as the mesh is refined. Second, for cases where the interface is spread
around the whole domain, diffuse interface is significantly cheaper regardless of the resolution.
• Although the Poisson solvers used in the two codes are not identical, for the same number of unknowns,
the two Poisson solvers give similar run times. Hence, it is safe to assume that the cost comparisons
reported in this work are independent of the specific choice of Poisson solvers. It is worth noting

8
 however, that typically solving the Poisson system subsumes the majority of the cost of the DI, while
only taking a small portion (problem-dependent but generally < 25%) of the cost of the VOF solver
[53].

3. Results

In this section we consider multiple two-dimensional (2D) canonical problems in which interface transport
is coupled with the Navier-Stokes equations. For each test case, we present results from DI and VOF
simulations. Many numerical tests that are decoupled from the Navier-Stokes equations are presented
in [55, 13] and [21] to verify the accuracy of the numerical methods used for the advection and normal
vector/curvature calculation steps of the VOF method chosen for this comparison. Furthermore, the two-
phase discretization has been verified using canonical two-phase flows [59]. Similar verification tests can
be found in the appendix for the DI method under study herein and in [39]. We have excluded these
tests from our comparison section since we intend to make conclusions for realistic applications where the
momentum equation and two-phase solver are coupled and computational cost may be a factor which needs
to be considered.
We present simulation results from several test cases that evaluate different aspects of the two solvers we
have chosen to compare. Section 3.1 uses the problem of advecting a very dense drop to assess the solvers’
robustness and ability to handle large density ratios. Next, we proceed to simulations of a more complicated
flow, with Rayleigh-Taylor (RT) instability in Section 3.2 providing a measure for comparing the accuracy
of interface and momentum transport via the two solvers. This case is unique in the sense that interfacial
regions are not confined in space and grow in time, allowing us to study how that affects the cost levels of
each solver. Surface tension is zero in the RT simulations, decoupling the accuracy of momentum transport
from surface tension force implementation. Moreover, this case has similarities with turbulent two-phase
flows as we will explain in Section 3.2. In Section 3.3, spurious currents are reported. We compare the
accuracy of the surface tension force implementations via examination of these artificial currents, which
are especially detrimental in simulations of surface tension dominant flows. In Sections 3.4 and 3.5, the
numerical interplay of inertia and surface tension forces is studied in the context of droplet oscillations
and damping surface waves, respectively. Finally, the realistic problem of droplet-pool impact is studied
in Section 3.6. This test involves interactions between inertia, surface tension forces and lubrication flow.
Moreover, the complex topologies that arise in these simulations provide another opportunity to compare
the competence of the solvers.

3.1. High Density Ratio Advecting Drop

This test case is adopted from [60], in which a drop of fluid 1 with diameter D = 0.2 is advected in
the x-direction with initial velocity U = 1 in a periodic 1×1 box of initially stationary fluid 2. The spatial
resolution is chosen to be 128×128 (corresponding to about 25 mesh points across the drop diameter) and
no surface tension or viscous forces are present. We consider very high density ratios up to ρ1 /ρ2 = 107
to test the robustness of the solvers. There is no exact boundary for the drop in the DI simulations, and
the density is very high within the transition zone. In this case, the velocity in the domain is initialized
using a hyperbolic tangent profile where the U = 0.5 contour lies on points where the density is not very
large, say 103 ρ2 . The drop is examined once after traveling a distance equal to one nominal diameter and
again after traveling a distance of five nominal diameters (see Figure 2). A CFL number of 0.25 is chosen
for these simulations. Theoretically, due to the high density ratio the drop should not ”feel” the presence
of the surrounding phase and should not deform. The diffuse interface is found to be robust for simulations
performed at density ratios up to ρ1 /ρ2 = 107 . For the VOF method, the simulations are stable for density
ratios up to ρ1 /ρ2 = 105 . For density ratios above that, we resort to lower order advection schemes to
keep the simulation running stably. Nevertheless, compared to the DI results, the profile shapes obtained
from geometric VOF seem to be closer to the ideal drop shape. This can be attributed to the consistent
geometric scheme used for computing convective fluxes in the VOF solver, as explained in [53]. A similar
correction can be implemented for the DI momentum transport but is not within the scope of this work.

9
 (a) DI, advection distance of 1D (b) DI, advection distance of 5D

(c) VOF, advection distance of 1D (d) VOF, advection distance of 5D

Figure 2: Interface profiles for the dense drop advection problem at density ratio of 107 on a 128 × 128 grid with around 25
cells across the drop diameter, using DI and VOF with lower order advection operators

The main conclusion from this test case is that both chosen two-phase solvers are very robust in the face of
large density ratios.

3.2. Rayleigh-Taylor Instability

As we proceed to gradually add more physical effects into the canonical test cases, we study the classical
problem of the Rayleigh-Taylor (RT) instability similar to [24]. In this study, no capillary effects exist, so the
curvature calculation and surface tension force implementation do not affect the results of this section. There
is a dense fluid with ρ1 = 1.225 and µ1 = 0.00313 sitting on a light fluid with ρ2 = 0.1694 and µ2 = 0.00313.
The interface between the two phases is at y = 0 in a [0, 1] × [−2, 2] box with slip boundary conditions on all
sides. Gravity is set to g = 9.81 and the interface is initially perturbed with a cosine wave with amplitude
0.05. As the instability grows, the heavier fluid penetrates into the lighter fluid and structures such as the
ones shown in Figure 3 emerge. Four different resolutions(64 × 256,128 × 512,256 × 1024,512 × 2048) were
examined for either of the two-phase methods and Figure 3 shows the interface location at four different
times(0.6, 0.7, 0.8, 0.9) long after the linear growth regime of the instability has passed. The VOF simulations
show convergence between interfacial profiles at 256 × 1024 and 512 × 2048, thus the results from 512 × 2048
VOF are chosen as reference solutions and plotted in black in the background of Figure 3. On top of the
background reference solution in Figure 3, at different resolutions(increasing from left to right) and different
times(increasing from top to bottom), interfacial profiles from DI and VOF are plotted on the left (in blue)
and right (in red) side respectively. It is clear from this figure that at the same numerical resolution, VOF
outperforms DI in capturing the interface location. Apparently however, DI can reproduce similar quality as
of VoF when it uses twice the resolution. Note that surface tension is zero in these simulations, so curvature
and surface tension force accuracy do not affect the results of this comparison. In other words, this test case
assesses the accuracy of phase and momentum advection for DI and VOF, where the coupling is due to the
dependence of density and viscosity on the phase (Equations 14 and 15 for DI and 5 and 6 for VOF).
Figure 4a compares the cost of performing the Rayleigh-Taylor simulations on different resolutions using
DI and VOF. While DI is found to be an order of magnitude cheaper than VOF at all resolutions, the
cost grows as Nx 2 for both methods. The cost of DI scales with the number of mesh points in the domain
which is expected for a phase field model. The cost of the VOF simulations, however, are mostly due to
the advection and surface tension force calculations which scale with the number of interfacial cells. It is
apparent from Figure 3 that the number of interfacial cells is growing in time. Therefore, the cost of the
VOF calculations increases accordingly as the simulation progresses in time, as shown in Figure 4b. The

10
 t = 0.6 [s] t = 0.6 [s] t = 0.6 [s] t = 0.6 [s]
0.5 0.5 0.5 0.5

0 0 0 0
y [m]

y [m]

y [m]

y [m]
-0.5 -0.5 -0.5 -0.5

-1 -1 -1 -1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x [m] x [m] x [m] x [m]
t = 0.7 [s] t = 0.7 [s] t = 0.7 [s] t = 0.7 [s]
0.5 0.5 0.5 0.5

0 0 0 0
y [m]

y [m]

y [m]

y [m]
-0.5 -0.5 -0.5 -0.5

-1 -1 -1 -1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x [m] x [m] x [m] x [m]
t = 0.8 [s] t = 0.8 [s] t = 0.8 [s] t = 0.8 [s]
0.5 0.5 0.5 0.5

0 0 0 0
y [m]

y [m]

y [m]

y [m]
-0.5 -0.5 -0.5 -0.5

-1 -1 -1 -1

-1.5 -1.5 -1.5 -1.5

0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1
x [m] x [m] x [m] x [m]
t = 0.9 [s] t = 0.9 [s] t = 0.9 [s] t = 0.9 [s]

0.6 0.6 0.6 0.6

0.4 0.4 0.4 0.4

0.2 0.2 0.2 0.2

0 0 0 0

-0.2 -0.2 -0.2 -0.2

y [m]

y [m]

y [m]

y [m]

-0.4 -0.4 -0.4 -0.4

-0.6 -0.6 -0.6 -0.6

-0.8 -0.8 -0.8 -0.8

-1 -1 -1 -1

-1.2 -1.2 -1.2 -1.2

-1.4 -1.4 -1.4 -1.4

0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1

x [m] x [m] x [m] x [m]

Figure 3: Rayleigh-Taylor instability. Resolution is increased from left (64 × 256) to right (512 × 2048), while physical time
increases from top (t = 0.6) to bottom (t = 0.9). For each panel, blue dashed lines show the interface location predicted by
DI on the left and red dashed lines show the interface location from VOF simulations on the right. The reference solution at
different times is also plotted in solid black.

11
 5
10 3
4.5

4
relative cost

10 2
3.5

cost
3
10 1 2.5
DI
VOF 2
Nx 2
10 0 1.5
10 2 10 3 0 0.2 0.4 0.6 0.8 1
Nx t[s]

(a) (b)

Figure 4: Relative cost of performing Rayleigh-Taylor simulations with DI and VOF compared at different resolutions (a), and
cost of performing one time-step of the 64 × 256 VOF simulation plotted against physical time (b)

overall effect is that VOF cost scales with Nx 2 , which is different from most other simulations we will go
through in the remainder of the paper. As mentioned before, we see from Figure 3 that the accuracy of
the VOF simulation is comparable to the DI simulation with twice the resolution. On the other hand, from
Figure 4a it can be seen that the cost associated with each VOF simulation is comparable to the cost of DI
performed at four times that resolution. As a result, we can conclude that at higher numerical resolution but
equal cost levels, DI yields more accurate results for this test case. Similar to the RT instability, turbulent
two-phase flows also cause interfacial regions to fill the space, resulting in the cost of VOF to scale with the
total number of cells. Based on the results in this section, it is reasonable to conclude that for turbulent
flows, a higher resolution DI can achieve similar accuracy to a VOF calculation but with less cost.

3.3. Spurious Currents in a Stationary Drop

In this test case we have a drop with diameter D = 0.4 in a 1 × 1 box with slip boundary conditions.
The initially stationary drop and surrounding fluid have equal densities and viscosities. The dimensionless
parameter characterizing this problem is the Laplace number, La = σρD/µ2 . In order to vary La, the
fluid densities are varied for different tests while everything else including the surface tension (σ = 1), and
viscosity(µ=0.1) are kept constant.
If curvatures were computed exactly, the selected DI and VOF solvers would both generate no spurious
currents, as they allow for the discrete balance between pressure gradients and surface tension forces, re-
sulting in a stagnant condition as the exact solution to this problem. Therefore, any sustained flow in the
numerical solution is due to errors in estimation of curvature from discretized field data, when the surface
tension force uses a balanced scheme [61], as done in both methods. We will examine the magnitude of the
largest velocity in the domain for La = 12000 at non dimensional time t∗ = σt/(µD) = 250, which is a
large enough time for the surface tension and viscous forces to have physically balanced each other and√for
spurious currents to be fully developed. To obtain optimal convergence using DI, we set  = ∆x2/3 / 3 32
instead of  = ∆x which is used for all other test cases in this work. A grid size of 128×128 is used for the
cases shown in Figure 5, where the spurious flow field induced by the DI and VOF methods are compared.
For the diffuse interface method(Figure 5a), the spurious currents are minimal at locations of the interface
that are either aligned with the grid, or angled at ∼ 45 degrees. In addition, the streamlines close to the
interface are mostly in the tangential direction and are mostly due to 8 vortex pairs inside and outside of
the drop. On the other hand, for VOF(Figure 5b), the flow seems to be due to 8 vortices which are centered
on the interface, resulting in maximal influx at angles that are multiples of 45 degrees. This is an interesting
observation regarding the flow patterns of spurious currents. Based on our examinations, this subtle contrast
in flow patterns seems to be generalizable to all sharp interface (like VOF) and diffuse interface (like phase
field) methods, when surface tension force is represented using a CSF form.
12
 y 0.6

0.4

0.2
(a) DI (b) VOF

Figure 5: Zoom-in on spurious currents in and around a stationary drop after equilibrium for the diffuse interface and VOF
simulations performed on 128×128 grids

0.4 0.6 0.8

0 x
0 0.2 0.4 0.6 0.8 1
10 -4
x

10 -5

10 -6

10 -7
10 1 10 2

Figure 6: Ca = umax µ/σ associated with spurious flow for the stationary drop problem at different resolutions

In Figure 6 we have compared the magnitude of Ca = umax µ/σ for the two methods at equilibrium
time for different resolutions. It is apparent that while the two methods perform similarly at the coarser
levels (∼ 6 mesh points across the drop) at higher resolutions VOF produces about an order of magnitude
smaller artificial currents compared to the diffuse interface. Moreover, VOF seems to have about second
order convergence, while DI has close to first order convergence rate with mesh refinement. These results
are in accordance with the predictions of [61], who showed that the magnitude of the maximum spurious
current in the domain should have the same convergence rate as the curvature evaluation as one refines the
mesh.
We have also compared the accuracy of the two methods against cost for these simulations. In Figure 7a,
Ca = umax µ/σ is plotted versus cost of performing one time-step using either method. However, due to its
finer resolution at a comparable cost, DI requires finer time-step and thus it is more reasonable to compare
the total cost for simulations performed over a fixed physical time (t∗ = 250). This comparison is plotted in
Figure 7b. However, it is also important to consider cost per time-step because in a realistic simulation, the
time-step restriction may indeed be imposed by stiffer physics or processes that are happening elsewhere in
the domain.

13
 DI DI
10 -4 VOF 10 -4
VOF
Ca = umax µ/σ

Ca = umax µ/σ
10 -5 10 -5

10 -6 10 -6

0 1 2
10 10 10 10 0 10 1 10 2 10 3 10 4
relative cost/time-step relative cost

(a) (b)

Figure 7: Ca = umax µ/σ in the stationary drop problem at equilibrium plotted against the cost per time step (a) and total
cost (b) for DI and VOF

From Figures 7a and 7b, it is apparent that if one desires very high accuracy with minimal tolerance for
spurious currents, VOF would be the method that provides lower Ca at a low cost. Nevertheless, in most
applications spurious current is an issue at low resolutions (e.g. poorly resolved drops in a liquid atomization
process [62] or droplet-laden turbulence [63]), where a refined DI is the better option.
We have also provided comparisons of cost of performing each simulation at different resolutions in
Figures 8a and 8b. In Figure 8a the relative cost of performing one time step of either method is plotted.
Diffuse interface is consistently one to two orders of magnitude cheaper than VOF at the same resolution.
An important distinction between the cost of VOF and DI is highlighted by the scalings of the cost of the two
methods. As we previously explained, most of the cost associated with VOF is due to geometric operations
for advection and curvature calculation which are performed at cells containing the interface. The number
of cells containing the interface scales with N , and as a result the whole cost of computation per time-step
for VOF scales with N α , where 1 < α < 2. On the other hand, as explained before, DI is blind to the
intricacies of the interface and the cost of all operations (including the sparse matrix Poisson solve step)
scales with the number of mesh points, N 2 . The total cost of reaching t∗ = 250 is shown in figure 8b. Since
the capillary time-step is used here the cost of each time-step is multiplied by a factor of N 3/2 , thus while
the ratio of the costs is preserved from Figure 8a to Figure 8b, the total cost evidently scales with N β and
N 3.5 for VOF and DI respectively, where 2.5 < β < 3.5.

3.4. Oscillating drop

In this simulation, a fluid 1 drop of radius R0 = 2 is perturbed with a mode n = 2 perturbation and
amplitude A0 = 0.01R0 in a box of size 20×20 filled with fluid 2. Slip boundary conditions are applied
on all walls. The physical parameters of the problem are given by σ = 1, ρ1 = 1, ρ2 = 0.01, µ1 = 1 and
µ2 = 0.0001. The initial perturbation activates the surface tension forces that causes the drop to oscillate.
During these oscillations, the extra energy goes back and forth between surface energy and kinetic energy.
The theoretical oscillation frequency for columns in the linear regime was found by [64] to be

n(n2 − 1)σ
ω2 = . (16)
(ρ1 + ρ2 )R0 3

In our simulations we can measure the period of oscillations, Tcalc , and compare it against the theoretical
prediction. Specifically the error of calculated period is reported as

ωTcalc
ET = −1 . (17)
2π

14
 10 4

10 2
10 2

10 0 10 0

10 1 10 2 10 1 10 2

(a) (b)

Figure 8: Comparison of cost per time-step (a) and total cost (b) of DI and VOF for the spurious current problem

The error in the calculated period of oscillations from the DI and VOF methods can be seen in Figure 9. By
examining this error we assess the accuracy of surface tension force calculations in the two different methods.
Similar to previous cases, VOF outperforms DI at the same resolution for this test case. When comparing
the period error values from these methods against cost, the conclusion is quite different. This comparison
is illustrated in Figure 10. The cost per time-step and total cost are used in this figure to demonstrate that
the two methods are performing at comparable levels for this test case.
We provide the relative cost per time-step and relative total cost of DI and VOF in Figure 11. The cost
is again around an order of magnitude less for DI. The logic used to interpret Figures 8a and 8b applies here
as well. Hence, cost per time-step of DI and VOF scale with N 2 and N α respectively, where 1 < α < 2. This
naturally leads to the total cost scalings of N 3.5 and N β for DI and VOF respectively, where 2.5 < β < 3.5.

3.5. Standing Wave

As the final canonical test case, we look at interfacial oscillations for a standing wave. This problem
assesses the accuracy of the two methods at capturing the oscillatory interplay between inertia and surface
tension in the presence of viscous dissipation. The mechanism underlying the oscillations is similar to the
oscillating drop problem studies in Section 3.4. Initially a single small amplitude wave with wavelength
λ = 2π is placed between two immiscible fluids in a [0, 2π] × [0, 2π] domain. The initial perturbation
amplitude is A0 = 0.01λ and y0 = π. The initial conditions for the wave position are given by
hwave (x, t = 0) = y − y0 + A0 cos(x − ∆x/2). (18)
Boundary conditions are periodic in the x direction and slip boundary conditions hold for the top and
bottom walls. If the two fluids have equal kinematic viscosity ν, then the amplitude of the wave will be
given by([65]):
X4
4(1 − 4β)ν 2 √ zi ω02 A0 √
Aex = A 0 erf c νt + ( 2 )exp[(zi2 − ν)t]erf c(zi t), (19)
8(1 − 4β)ν 2 + ω02 Z
i=1 i
zi − ν

where zi are the roots of

√
z 4 − 4β νz 3 + 2(1 − 6β)νz 2 + 4(1 − 3β)ν 3/2 z + (1 − 4β)ν 2 + ω02 = 0. (20)
For the test case simulated, σ = 2 ρ1 = ρ2 = 1, ν = 0.064720863 and the time step was dt = 0.003.
The amplitude of the standing wave measured at x = ∆x/2 versus time using DI and VOF is plotted for
15
 10 0

10 -1

10 -2

10 -3

10 1 10 2

Figure 9: Period error in the oscillating drop problem for different methods and resolutions

10 -1 DI 10 -1
VOF
ET = |(Tcalc /Texact − 1)|

10 -2 10 -2

10 0 10 1 10 2 10 0 10 2
relative cost/time-step
(a) (b)

Figure 10: Period Error in the oscillating drop problem for different methods plotted against computational cost per time-step
(a) and total computational cost (b)

16
 10 3

10 4
10 2

10 1 10 2

10 0
10 0

10 1 10 2 10 1 10 2

(a) (b)

Figure 11: Comparison of cost per time-step (a) and total cost (b) of DI and VOF for the oscillating drop period calculation
problem

different resolutions alongside the exact solution in Figures 12a and 12b respectively. It is clear that the
sharp interface calculations from VOF are more accurate at capturing the correct location of the interface.
Using the theoretical amplitude evolution equation (Equation 19) a better comparison of the performance
of the two methods can be obtained from examining the root-mean-squared (r.m.s.) error of the simulation
predictions. This comparison is given in Figure 13a . While VOF errors for this test case are about an
order of magnitude smaller than DI for the same resolution, both methods show first order convergence. It
is important to note that the initial amplitude of the wave is small and comparable to the thickness of the
interface for the DI simulations. This will lead to errors of the order of the thickness of the interface ()
for the location of the contour φ = 0.5, explaining the first order trend in the errors. Moreover, since our
measure for comparison in this test case is the evolution of the small amplitude standing wave, we believe
that part of the failure of DI relative to VOF is because of the need for secondary interpolation for finding
the exact location of the interface, a step unnecessary for sharp interface models such as VOF.
Once again, we compare the error of the methods against computational cost. Figure 13b shows that
VOF is more cost-efficient when very high accuracy in determining the location of the interface is sought.
On the other hand, in lieu of a low resolution VOF, DI with higher resolution provides higher accuracy while
maintaining the same cost.
In Figure 14, we compare the relative cost of performing the simulations using DI and VOF at different
resolutions. Diffuse interface seems to be between one and two orders of magnitude cheaper at the same
resolution. However for this topologically simple problem , since the number of interfacial cells clearly scales
with N , and the majority of the work for the VOF method is done on these cells, the cost seems to be
proportional to N , as opposed to DI which computes the evolution of φ in all cells in the domain, resulting
in a cost (for Poisson and the rest of computations) scaling with N 2 .

3.6. Droplet Impact

We briefly compare the performance of DI and VOF on a complex problem relevant to air entrainment
during droplet-pool impact events. When a water droplet with a O(1mm) diameter impacts a deep pool of
water with normal velocity of O(1m/s), a thin gas film between the drop and pool cushions the impact. This
air layer becomes entrapped between the two liquid bodies after they contact at a distance from the center,
and then retracts to leave a large bubble behind. Figure 15a shows a schematic showing the thin film with
respect to the drop and pool. While it is clear from the experimental images in [66] that the stages after
the thin film is entrapped require 3D simulations, the events leading to the formation of the thin gas film
can be well approximated with 2D simulations [67]. Moreover, although in these experiments undulations
17
 10 -3 10 -3
10
8
6
5
4
2
0 0
-2
-4
-5
-6

5 10 15 20
0 5 10 15 20

(a) DI
(b) VOF

Figure 12: Amplitude of the wave at x = ∆x/2 from DI and VOF simulations are plotted against time for different resolutions
alongside the theoretical prediction. In both figures, the theoretical curve is plotted in solid black, while the numerical results
from 16 × 16, 32 × 32, 64 × 64 and 128 × 128 simulations are plotted using dotted, dashed-dotted, dashed and solid blue lines
respectively.

10 0 DI
VOF
10 -1
Erms

10 -2
Erms

DI -2
10
VOF
N −1
10 -4 N −2

10 1 10 2 10 0 10 1 10 2
N relative cost

(a) (b)

Figure 13: Standing wave r.m.s error values for different methods plotted against resolution (a), and standing wave r.m.s error
values for different methods plotted against computational cost (b)

18
 relative cost

10 2

DI
VOF
N1
N2

10 0
10 1 10 2
N
Figure 14: Cost comparison of DI and VOF for the standing wave problem

appear on the edge of the retracting thin film, these azimuthal instabilities do not lead to pinch-off of smaller
bubbles. As such a 2D simulation of the full problem is also expected to result in one large bubble after the
thin film has retracted.
Boundary element methods (BEM) coupled with lubrication flow in the gas are usually employed for
simulating this phenomenon [67, 68]. Such methods are preferred since they do not require resolving the
thickness of the very thin gas film. Nevertheless, in this section we compare the performance of VOF and DI
at capturing the outcome of the drop-pool impact problem using highly resolved 2D two-phase simulations.

We simulate a water-air system with impact velocity U = 1m/s and droplet diameter D = 3.4mm on
stretched meshes concentrated around the region of impact. Due to the resolution requirements of this
problem, we perform these simulations utilizing parallel computing. In Figure 15b, the thin gas film is
shown just prior to the contact of the liquid bodies, depicting how the simulations converge to a solution
that is also close to the predictions of a BEM simulation similar to [67]. The mean film thickness of the film
is also in agreement with experimental measurements by [69]. After the drop and pool contact however,
some smaller bubbles are observed in the simulations. These bubbles are unphysical, as their size shrinks
almost proportional with the mesh, and also as the experiments show, only one large bubble remains after
retraction of the thin gas film. These numerical bubbles can be seen in Figure 15c, which shows a zoom-in
on the contact zone of a 6144 × 6144 DI simulation. The diameter of the large bubble and the numerical
bubbles generated from VOF and DI are sorted and plotted in Figure 16. Indeed, the size of these artificial
bubbles shrink linearly with mesh refinement as shown in Figure 16.
Through rigorous comparisons we find that both methods converge on the geometry of the film and the
diameter of the large physical bubble resulting from it’s retraction, but VOF requires less resolution as one
would expect. Figure 16 shows how results from a DI simulation on a 6144 × 6144 stretched mesh capture
features with the same length scales as a 2048 × 2048 stretched mesh VOF simulation of this problem.
Curiously however, the VOF simulation (with 3 times mesh points in both directions) requires about 5
times more computational cost than the DI simulation. It’s worth a reminder that in the VOF method
19
 (a) (b)

(c)

Figure 15: Schematic of how a thin gas film is entrapped during the drop-pool impact event (a). Entrapped thin film at contact
time from DI simulations at three different resolutions and a BEM simulation for a D = 3.4mm water drop impacting a deep
pool with U = 1m/s(b). A zoomed-in snapshot of the interface and numerical air bubbles around the region where the drop
and pool first contact for a 6144 × 6144 DI simulation of a D = 3.4mm water drop impacting the pool with U = 1m/s (c)

20
 DI2048 × 2048
DI3072 × 3072
DI4096 × 4096
DI6144 × 6144
V OF 4096 × 4096
Bubble Diameter(m)

V OF 2048 × 2048
10 -5 V OF 1024 × 1024
V OF 512 × 512

10 -6

2 4 6 8 10 12
Bubble Index

Figure 16: Sorted bubble diameters for VOF and DI, demonstrating convergence on the size of the large bubble while numerical
bubbles shrink with mesh refinement

studied here, the operations deployed to find the volume fractions and their normal vector are performed
in regions containing the interface. As a result, when a highly stretched mesh is used to concentrate most
mesh points around the interface, the cost almost scales with the number of total mesh points. Moreover,
in such large parallel computations, load balancing is much more difficult for VOF compared to DI which is
load-balanced in nature.

4. Conclusions

In this paper we compared two fundamentally different yet promising numerical methods from the class
of VOF and DI methods. While both of these methods have attracted plenty of attention from researchers,
an unbiased comparison of these methods’ capabilities was warranted. Several test cases were chosen to
assess the relative performance and computational cost of these methods at solving the interface evolution
equation coupled to the Navier-Stokes equation.
It was observed that for most of the test cases VOF was more accurate than DI at the same resolution. In
terms of convergence behavior, both methods were convergent on all test cases with comparable convergence
rates. However, the computational cost of DI was typically an order of magnitude less than VOF at the same
resolution. We generally observed that while maintaining accuracy, one could save a factor of about 2-3 in
resolution by switching from DI to VOF. However, more often than not, the cost of these two comparably
accurate simulations would be similar or even cheaper in the case of DI.
For most test cases, DI with more resolution was found to be a better option than coarse or under-
resolved VOF simulations. At high resolutions VOF yielded higher accuracy at lower cost compared to DI.
This was especially found to be true for problems where the interfacial cells were confined in space. For
such flows, if the space has λ dimensions and N cells per direction, the cost of VOF and DI scale with
N λ−1 and N λ respectively, giving VOF an advantage at higher N values. We verily observed this for some
of our small canonical problems; however, large realistic problems requiring parallel computing incur load
balancing issues for the geometric VOF, while posing no extra challenge to the scalable DI. Alternatively if
the interface is not confined to a particular region of space, as is the case for most high Re two-phase flows,
high resolution would be needed at all points in space. For a flow in λ dimensions, this results in the VOF

21
 0.4
0.4
0.35 0.35
0.8 0.8
0.3 0.3

0.6 0.25 0.6 0.25

y
0.2
y

0.2
0.4 0.4 0.15
0.15

0.1 0.1
0.2 0.2
0.05 0.05

0 0
0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8
x x

(a) Exact pressure (b) DI pressure

Figure 17: Pressure field from diffuse interface compared to the theoretical value of pressure for a sharp interface drop. In
addition, the interface location is plotted in black in part (b)

cost to scale with N λ , similar to DI. As such, compared to VOF, the low computational cost of DI would
allow for cheaper DI calculations at higher accuracy and cost levels for these problems. Based on these
discussions and our results, it seems likely that in both types of flows, the accuracy of DI accompanied by
its scalability and low cost allow for DI to outperform VOF in many practical simulations.

Appendix A Diffuse Interface Tests

A.1 Curvature Accuracy Test

From Equations 11 and 13 we know that the surface tension force takes the form of σκ∇φ, where
curvature is obtained from the normal vector using κ = ∇ · ~n. Calculating curvature error is different
between diffuse interface methods and sharp interface methods. While in sharp interface methods curvature
calculation is only undertaken at the interface, in diffuse interface methods curvature must be accurately
computed in a transition layer between the two phases. Strictly speaking, the interface exists everywhere in
the domain; thus, a global indicator would be more suitable to study the surface tension force calculation
accuracy. In the absence of the inertial and viscous terms in Equation 11, pressure gradients must balance
the surface tension forces. Hence, the value of the pressure drop across an interface with known curvature is
a practical indicator of the accuracy of surface tension forces. Specifically, we will utilize a test case where
a circle of radius R = 0.25 is placed in the center of a 1 × 1 box and the pressure computed from DI will
be compared to the theoretical pressure field(Figure 17). Note that theoretically, the pressure jump at the
drop interface(corresponding to φ = 0.5) is given by ∆P = σκ/R. If we neglect the error in solving the
linear Poisson system for pressure, the error in pressure is solely due to the errors committed during surface
tension force calculation in DI. On a Cartesian mesh, this error can be calculated via
Ny
Nx X
X σ
EP = P (i, j) − HS(R − r(i, j)) ∆x∆y, (21)
i=1 j=1
R

where HS(x) is the heavy side step function, and r is the distance of a point from the center of the domain.
In Figure 18, the pressure error, EP is plotted against resolution. These calculations were performed by
selecting  = ∆x = ∆y in all simulations. Clearly, the error in the pressure calculation decreases linearly
with mesh refinement.

22
 DI
N −1
N −2
10 -2
EP

10 -3

10 -4

10 1 10 2
ND
L
Figure 18: Convergence study on the accuracy of pressure/surface tension force calculation

A.2 Advection Test

A standard test for two-phase flow advection schemes is the 2D drop in shear flow test case. A column
of radius R0 = 0.15 and center (0.5; 0.75) is placed inside a unit sized box. The velocity field is given by
the stream function,
1 πt
Ψ(x, t) = sin2 (πx)sin2 (πy)cos( ) (22)
π T
, and T = 8. The velocity field specified for this test case dictates the initial circle to be retained at the
end of simulation time. Hence, an insightful check is to compare the final captured interface with the exact
solution. For these test cases, a fixed time step of dt = 0.5∆x was chosen. In Figure 19, the state of the
drop at t = T /2 can be seen for different resolution simulations. As Figure 19 shows, the drop gets stretched
in the vortex field and artificial drops appear at low resolutions. In Figure 20, the drops interfacial profiles
at the end of the simulation(t=T), which should analytically be the same as initial conditions is plotted
against the initial drop for different resolutions.
If we define the final error in the shape of the drop to be defined by:
Ny
Nx X
X
Eshape = |HS(φ(i, j, t = T ) − 0.5) − HS(φ(i, j, t = 0) − 0.5)| ∆x∆y (23)
i=1 j=1

Then, Figure 21 shows how errors are reduced with approximately second order convergence rate.

5. Acknowledgements

This work was supported by Office of Naval Research (Grant No. 119675) and NASA (Grant No. 127881)
.

23
 0.8 0.8

0.6 0.6
y

y
0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
x x
(a) 128 × 128 (b) 256 × 256

0.8 0.8

0.6 0.6
y

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
x x
(c) 512 × 512 (d) 1024 × 1024

Figure 19: Interfacial profiles at t = T /2 for drop in deformation field. Increasing resolution from left to right and top to
bottom results in an intact ligament which does not break into drops

24
 0.8 0.8

0.6 0.6
y

y
0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
x x
(a) 128 × 128 (b) 256 × 256

0.8 0.8

0.6 0.6
y

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
x x
(c) 512 × 512 (d) 1024 × 1024

Figure 20: Interfacial profiles at t = T for drop in deformation field plotted against initial drop. The final solution and initial
drop should lay on top of each other with a perfect two-phase flow numerical method.

25
 10 -2

10 -3

10 -4
10 2

Figure 21: Shape error for the deforming drop test case versus resolution

6. References
[1] S. Mirjalili, S. Jain, M. Dodd, Interface-capturing methods for two-phase flows: An overview and recent developments,
Center for Turbulence Research Annual Research Briefs (2017) 117–135.
[2] S. O. Unverdi, G. Tryggvason, A front-tracking method for viscous, incompressible, multi-fluid flows, Journal of compu-
tational physics 100 (1) (1992) 25–37.
[3] G. Tryggvason, B. Bunner, A. Esmaeeli, D. Juric, N. Al-Rawahi, W. Tauber, J. Han, S. Nas, Y.-J. Jan, A front-tracking
method for the computations of multiphase flow, Journal of Computational Physics 169 (2001) 708–759.
[4] R. Debar, Fundamentals of the KRAKEN code, Technical Report UCIR-760, Lawrence Livermore National Laboratory,
CA.
[5] B. D. Nichols, C. W. Hirt, Methods for calculating multi-dimensional, transient free surface flows past bodies, Tech. Rep.
LA-UR-75-1932, Los Alamos National Laboratory (1975).
[6] W. F. Noh, P. Woodward, SLIC (simple line interface calculation), in: A. I. van de Vooren, P. J. Zandbergen (Eds.),
Proceedings of the Fifth International Conference on Numerical Methods in Fluid Dynamics June 28 – July 2, Twente
University, Enschede, Vol. 59 of Lecture Notes in Physics, Springer, Berlin, Heidelberg, 1976, pp. 330–340.
[7] R. Scardovelli, S. Zaleski, Direct numerical simulation of free-surface and interfacial flow, Annual review of fluid mechanics
31 (1) (1999) 567–603.
[8] G. Tryggvason, R. Scardovelli, S. Zaleski, Direct numerical simulations of gas–liquid multiphase flows, Cambridge Univer-
sity Press, 2011.
[9] J. Lpez, J. Hernndez, P. Gmez, F. Faura, A volume of fluid method based on multidimensional advection and spline
interface reconstruction, Journal of Computational Physics 195 (2) (2004) 718 – 742. doi:https://doi.org/10.1016/j.
jcp.2003.10.030.
URL http://www.sciencedirect.com/science/article/pii/S0021999103005801
[10] M. Owkes, O. Desjardins, A computational framework for conservative, three-dimensional, unsplit, geometric transport
with application to the volume-of-fluid (vof) method 270 (2014) 587–612.
[11] J. E. P. Jr., E. G. Puckett, Second-order accurate volume-of-fluid algorithms for tracking material interfaces, Journal of
Computational Physics 199 (2) (2004) 465 – 502.
[12] L. Jofre, O. Lehmkuhl, J. Castro, A. Oliva, A 3-D Volume-of-Fluid advection method based on cell-vertex velocities for
unstructured meshes 94 (2014) 14–29.
[13] C. Ivey, P. Moin, Conservative and bounded volume-of-fluid advection on unstructured grids, Submitted to the J. of
Comput. Phys.

26
[14] L. Jofre, R. Borrell, O. Lehmkuhl, A. Oliva, Parallel load balancing strategy for Volume-of-Fluid methods on 3-D un-
structured meshes 282 (2015) 269–288.
[15] B. Parker, D. Youngs, Two and three dimensional Eulerian simulation of fluid flow with material interfaces, Atomic
Weapons Establishment, 1992.
[16] E. Aulisa, S. Manservisi, R. Scardovelli, S. Zaleski, Interface reconstruction with least-squares fit and split advection in
three-dimensional cartesian geometry, Journal of Computational Physics 225 (2) (2007) 2301 – 2319.
[17] R. Scardovelli, S. Zaleski, Interface reconstruction with least-square fit and split eulerianlagrangian advection, International
Journal for Numerical Methods in Fluids 41 (3) (2003) 251–274.
[18] M. M. Francois, S. J. Cummins, E. D. Dendy, D. B. Kothe, J. M. Sicilian, M. W. Williams, A balanced-force algorithm
for continuous and sharp interfacial surface tension models within a volume tracking framework, J. Comput. Phys. 213 (1)
(2006) 141–173. doi:http://dx.doi.org/10.1016/j.jcp.2005.08.004.
[19] M. Sussman, M. Ohta, A stable and efficient method for treating surface tension in incompressible two-phase flow, SIAM
Journal on Scientific Computing 31 (4) (2009) 2447–2471.
[20] S. Popinet, An accurate adaptive solver for surface-tension-driven interfacial flows, Journal of Computational Physics
228 (16) (2009) 5838–5866.
[21] C. Ivey, P. Moin, Accurate interface normal and curvature estimates on three-dimensional unstructured non-convex poly-
hedral meshes, Journal of Computational Physics 300 (2015) 365–386.
[22] S. Osher, J. A. Sethian, Fronts propagating with curvature-dependent speed: Algorithms based on hamilton-jacobi for-
mulations, Journal of Computational Physics 79 (1) (1988) 12 – 49.
[23] M. Sussman, P. Smereka, S. Osher, A level set approach for computing solutions to incompressible two-phase flow, Journal
of Computational Physics 114 (1) (1994) 146 – 159. doi:https://doi.org/10.1006/jcph.1994.1155.
URL http://www.sciencedirect.com/science/article/pii/S0021999184711557
[24] M. Herrmann, A balanced force refined level set grid method for two-phase flows on unstructured flow solver grids,
J. Comput. Phys. 227 (2008) 2674–2706.
[25] M. Sussman, A second order coupled level set and volume-of-fluid method for computing growth and collapse of vapor
bubbles, Journal of Computational Physics 187 (1) (2003) 110 – 136.
[26] E. Olsson, G. Kreiss, A conservative level set method for two phase flow, J. Comput. Phys. 210 (2005) 225.
[27] E. Olsson, G. Kreiss, S. Zahedi, A conservative level set method for two phase flow ii, J. Comput. Phys. 225 (2007) 785.
[28] D. Anderson, G. McFadden, A. Wheeler, Diffuse-interface methods in fluid mechanics, Annual Review of Fluid Mechanics
30 (1998) 139–165.
[29] V. Badalassi, H. Ceniceros, S. Banerjee, Computation of multiphase systems with phase field models, Journal of Compu-
tational Physics 190 (2) (2003) 371–397.
[30] J. Kim, A continuous surface tension force formulation for diffuse-interface models, Journal of Computational Physics
204 (2) (2005) 784–804.
[31] Y. Sun, C. Beckermann, Sharp interface tracking using the phase-field equation, Journal of Computational Physics 220 (2)
(2007) 626 – 653.
[32] H. Ding, P. D. Spelt, C. Shu, Diffuse interface model for incompressible two-phase flows with large density ratios, Journal
of Computational Physics 226 (2) (2007) 2078 – 2095.
[33] D. Jacqmin, Calculation of two-phase navierstokes flows using phase-field modeling, Journal of Computational Physics
155 (1) (1999) 96 – 127.
[34] J. Kim, S. Lee, Y. Choi, A conservative allencahn equation with a spacetime dependent lagrange multiplier, International
Journal of Engineering Science 84 (2014) 11 – 17. doi:https://doi.org/10.1016/j.ijengsci.2014.06.004.
URL http://www.sciencedirect.com/science/article/pii/S0020722514001311
[35] J. W. Cahn, J. E. Hilliard, Free energy of a nonuniform system. i. interfacial free energy, The Journal of Chemical Physics
28 (2) (1958) 258–267.
[36] J. Shen, X. Yang, A phase-field model and its numerical approximation for two-phase incompressible flows with different
densities and viscosities, SIAM Journal on Scientific Computing 32 (3) (2010) 1159–1179.
[37] P. Yue, C. Zhou, J. J. Feng, Spontaneous shrinkage of drops and mass conservation in phase-field simulations, Journal of
Computational Physics 223 (1) (2007) 1 – 9.
[38] P. H. Chiu, Y. T. Lin, A conservative phase field method for solving incompressible two-phase flows, J. Comput. Phys.
230 (2011) 185.
[39] S. Mirjalili, C. B. Ivey, A. Mani, On the boundedness properties of a conservative diffuse interface method for two-phase
flowsarXiv:arXiv:1803.01262.
[40] S. Popinet, Numerical models of surface tension, Annual Review of Fluid Mechanics 50 (1) (2018) 49–75. doi:10.1146/
annurev-fluid-122316-045034.
[41] S. J. Cummins, M. M. Francois, D. B. Kothe, Estimating curvature from volume fractions, Computers and Structures
83 (67) (2005) 425 – 434, frontier of Multi-Phase Flow Analysis and Fluid-StructureFrontier of Multi-Phase Flow Analysis
and Fluid-Structure.
[42] J. Brackbill, D. Kothe, C. Zemach, A continuum method for modeling surface tension, Journal of Computational Physics
100 (2) (1992) 335 – 354.
[43] G. Bornia, A. Cervone, S. Manservisi, R. Scardovelli, S. Zaleski, On the properties and limitations of the height function
method in two-dimensional cartesian geometry, Journal of Computational Physics 230 (4) (2011) 851 – 862.
[44] J. López, C. Zanzi, P. Gómez, R. Zamora, F. Faura, J. Hernández, An improved height function technique for computing
interface curvature from volume fractions, Computer Methods in Applied Mechanics and Engineering 198 (33-36) (2009)
2555–2564.

27
[45] J. López, J. Hernández, On reducing interface curvature computation errors in the height function technique, Journal of
Computational Physics 229 (13) (2010) 4855–4868.
[46] R. P. Fedkiw, T. Aslam, B. Merriman, S. Osher, A non-oscillatory eulerian approach to interfaces in multimaterial flows
(the ghost fluid method), Journal of computational physics 152 (2) (1999) 457–492.
[47] K. Yokoi, Efficient implementation of thinc scheme: a simple and practical smoothed vof algorithm, Journal of Computa-
tional Physics 226 (2) (2007) 1985–2002.
[48] S. S. Deshpande, L. Anumolu, M. F. Trujillo, Evaluating the performance of the two-phase flow solver interfoam, Com-
putational science & discovery 5 (1) (2012) 014016.
[49] W. Rider, D. Kothe, Stretching and tearing interface tracking methods, in: 12th Computational Fluid Dynamics Confer-
ence, 1995, p. 1717.
[50] D. Gerlach, G. Tomar, G. Biswas, F. Durst, Comparison of volume-of-fluid methods for surface tension-dominant two-phase
flows, International Journal of Heat and Mass Transfer 49 (3-4) (2006) 740–754.
[51] C. Bilger, M. Aboukhedr, K. Vogiatzaki, R. Cant, Evaluation of two-phase flow solvers using level set and volume of fluid
methods, Journal of Computational Physics 345 (2017) 665 – 686. doi:https://doi.org/10.1016/j.jcp.2017.05.044.
URL http://www.sciencedirect.com/science/article/pii/S0021999117304278
[52] W. Aniszewski, T. Ménard, M. Marek, Volume of fluid (vof) type advection methods in two-phase flow: a comparative
study, Computers & Fluids 97 (2014) 52–73.
[53] C. B. Ivey, Geometric volume-of-fluid framework for simulating two-phase flows on unstructured meshes, Ph.D. thesis,
Stanford University (2017).
[54] D. L. Youngs, Time-dependent multi-material flow with large fluid distortion, in: K. W. Morton, M. J. Baines (Eds.),
Numerical Methods for Fluid Dynamics, Academic Press, New York City, NY, 1982, pp. 273–285.
[55] C. Ivey, P. Moin, Conservative and bounded volume-of-fluid advection on unstructured grids, Center for Turbulence
Research Annual Research Briefs (2016) 163–179.
[56] F. Ham, G. Iaccarino, Energy conservation in collocated discretization schemes on unstructured meshes, Annual Research
Briefs, Center for Turbulence Research, Stanford University (2004) 3–14.
[57] Y. Morinishi, T. S. Lund, O. V. Vasilyev, P. Moin, Fully conservative higher order finite difference schemes for incom-
pressible flow, Journal of computational physics 143 (1) (1998) 90–124.
[58] H. Le, P. Moin, An improvement of fractional step methods for the incompressible navier-stokes equations, Journal of
Computational Physics 92 (2) (1991) 369 – 379.
[59] C. Ivey, P. Moin, S. Bose, D. Kim, Geometric volume-of-fluid framework for the simulation of two-phase flows on unstruc-
tured grids (2016) 163–179.
[60] M. Raessi, H. Pitsch, Consistent mass and momentum transport for simulating incompressible flows with large density
ratios using the level set method, Computers and Fluids 63 (2012) 70–81.
[61] M. Magnini, B. Pulvirenti, J. Thome, Characterization of the velocity fields generated by flow initialization in the cfd
simulation of multiphase flows, Applied Mathematical Modelling 40 (1516) (2016) 6811 – 6830.
[62] M. Gorokhovski, M. Herrmann, Modeling primary atomization, Annual Review of Fluid Mechanics 40 (1) (2008) 343–366.
arXiv:https://doi.org/10.1146/annurev.fluid.40.111406.102200, doi:10.1146/annurev.fluid.40.111406.102200.
URL https://doi.org/10.1146/annurev.fluid.40.111406.102200
[63] M. S. Dodd, A. Ferrante, On the interaction of taylor length scale size droplets and isotropic turbulence, Journal of Fluid
Mechanics 806 (2016) 356412. doi:10.1017/jfm.2016.550.
[64] H. Lamb, Hydrodynamics, Cambridge university press, 1932.
[65] A. Prosperetti, Motion of two superposed viscous fluids, Phys. Fluids 24 (7) (1981) 1217–1223.
[66] S. Thoroddsen, T. Etoh, K. Takehara, Air entrapment under an impacting drop, Journal of Fluid Mechanics 478 (2003)
125–134.
[67] P. D. Hicks, R. Purvis, Air cushioning in droplet impacts with liquid layers and other droplets, Physics of Fluids 23 (6)
(2011) 062104.
[68] M. H. Hendrix, W. Bouwhuis, D. van der Meer, D. Lohse, J. H. Snoeijer, Universal mechanism for air entrainment during
liquid impact, Journal of fluid mechanics 789 (2016) 708–725.
[69] S. Thoroddsen, T. Etoh, K. Takehara, N. Ootsuka, Y. Hatsuki, The air bubble entrapped under a drop impacting on a
solid surface, Journal of Fluid Mechanics 545 (2005) 203–212.

28