i An update to this article is included at the end

Journal of Computational Physics 225 (2007) 840–873

www.elsevier.com/locate/jcp

A robust and accurate approach to computing compressible

multiphase flow: Stratified flow model and AUSM+-up scheme
a,* b
Chih-Hao Chang , Meng-Sing Liou
a
Center for Risk Studies and Safety, University of California, Santa Barbara, Goleta, CA 93117, USA
b
Propulsion Systems Division, NASA Glenn Research Center, Cleveland, OH 44135, USA

Received 25 February 2006; received in revised form 21 December 2006; accepted 4 January 2007
Available online 19 January 2007

Abstract

In this paper, we propose a new approach to compute compressible multifluid equations. Firstly, a single-pressure com-
pressible multifluid model based on the stratified flow model is proposed. The stratified flow model, which defines different
fluids in separated regions, is shown to be amenable to the finite volume method. We can apply the conservation law to
each subregion and obtain a set of balance equations1. Secondly, the AUSM+ scheme, which is originally designed for the
compressible gas flow, is extended to solve compressible liquid flows. By introducing additional dissipation terms into the
numerical flux, the new scheme, called AUSM+-up, can be applied to both liquid and gas flows. Thirdly, the contribution
to the numerical flux due to interactions between different phases is taken into account and solved by the exact Riemann
solver. We will show that the proposed approach yields an accurate and robust method for computing compressible mul-
tiphase flows involving discontinuities, such as shock waves and fluid interfaces. Several one-dimensional test problems are
used to demonstrate the capability of our method, including the Ransom’s water faucet problem and the air–water shock
tube problem. Finally, several two dimensional problems will show the capability to capture enormous details and com-
plicated wave patterns in flows having large disparities in the fluid density and velocities, such as interactions between
water shock wave and air bubble, between air shock wave and water column(s), and underwater explosion.
 2007 Elsevier Inc. All rights reserved.

PACS: 83.85.Pt; 47.11.+j; 47.55.Kf

Keywords: Multiphase flow; Multifluid model; AUSM+ scheme; Stratified flow method

1. Introduction

Seeking an accurate method to simulate compressible multifluid and multiphase flows has been an impor-
tant research topic for many engineering applications. Examples of compressible multifluid flows include the

*
Corresponding author.
E-mail addresses: [email protected] (C.-H. Chang), [email protected] (M.-S. Liou).
1
However, conservative form is lost in these balance equations when considering each individual phase; in fact, the interactions that
exist simultaneously in both phases manifest themselves as nonconservative terms.

0021-9991/$ - see front matter  2007 Elsevier Inc. All rights reserved.
doi:10.1016/j.jcp.2007.01.007
 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873 841

cooling system in the conventional nuclear reactor, the fuel transport system in which the fuel and gas are
transported simultaneously, and the generation, deformation and collapse of cavities around the underwater
propeller or within the high pressure injector. The flow phenomena involved in these systems are very com-
plicated, entailing reliable mathematical description (modeling) and numerical methods.
In this paper, we will concentrate on the numerical algorithm for compressible multifluid equations in
which the fluids are assumed inter-penetrating, non-homogeneous and non-equilibrium; that is, each fluid
has its own velocity and temperature fields at the same location, but all fluids share the same pressure. As
shown by Buyevich [1] and Ishii [2], one can use two sets of Euler/Navier–Stoke equations to describe the
motion of gas and liquid phase fluids respectively, as given in the following:
oðai qi Þ
þ r  ðai qi~vi Þ ¼ S q ;
ot
oðai qi~vi Þ
þ r  ðai qi~ vi Þ þ rðai pÞ ¼ prai þ ~
vi~ Sv; ð1Þ
ot
oðai qi Ei Þ oai
þ r  ðai qi H i~
vi Þ ¼ p þ Se;
ot ot
where the subscript ‘‘i’’=‘‘g’’ or ‘‘l’’, representing gas or liquid phase fluid respectively. ai is the void fraction of
fluid ‘‘i’’, and must satisfy the constraint, ag þ al ¼ 1. The RHS of Eq. (1) represents the interactions that cou-
ple both fluid motions together. The ‘‘S’’ terms on the right hand side represent a group of terms that arise
from interfacial physics, viscous effects, phase change, body forces, etc. They are expressed in differential (first
or higher derivatives) or non-differential form. The system of equations is quite general and has been widely
used to describe multiphase flows in a variety of applications. However, the inviscid limit of the multifluid
model has been known to be problematic, giving rise to instability, loss of accuracy, and non-convergence
in numerical solution. It is primarily attributable to the fact that the system can become non-hyperbolic
and ill-posed. Hence, the root of the problems points to the first derivative terms that differ from that for
the single fluid equations (which are known to be hyperbolic), namely, p$ai and p oaoti . Additionally, these
terms are not in conservative form. It is not clear that a discontinuity in the sense of weak solution still remains
valid. Some comprehensive reviews of the compressible multifluid models can be found in [3,4].
It has been recognized that non-hyperbolicity is a major reason for causing numerical instability. Thus, a
great deal of efforts have been focused on how to improve the hyperbolicity of the system and make the mul-
tifluid equations well-posed; clearly altering first derivative terms with some physical basis is necessary, such as
interfacial pressure correction [5], virtual mass [6], or separate pressures [3,7]. However, we will show that non-
hyperbolicity is not the only reason for causing numerical instability. It is as important to properly handle the
discretization and define numerical fluxes that includes all relevant interactions terms between the same and
different phases in order to obtain a stable and accurate numerical solution.
To highlight the importance of a numerical method for solving a multifluid model, we consider the case of
an 1D moving fluid interface with constant velocity and pressure. Since all convection fluxes cancel out with
each other, Eq. (1) reduces to a simple equation that the numerical solution must satisfy the so called pressure
non-disturbing condition [7] given as follows:
rðai pÞ ¼ prai : ð2Þ
The LHS of the above equation is in conservative form and similarly exists in the single fluid equations, hence
it is rather clear what to do about its discretization. However, it is not so obvious how to discretize the RHS
term, $ai, which is in nonconservative form, because it will have to be compatible with how the pressure flux
aip is evaluated in order to satisfy the above equation. For example, a central differencing of $ai is unlikely to
be compatible with an upwind differencing of $(aip). On the other hand if the discretization of the LHS re-
quires analytical form of the eigenstructure of Eq. (1), which unfortunately is not available in general, then
it will be difficult to come up with a compatible RHS.
While Eq. (1) is usually derived from an averaging procedure [1,2], Stewart and Wendroff [3] gave another
approach for derivation based on the stratified flow model. Adopting the concept of the stratified flow model
has several advantages. Firstly, it gives a clear view of the mathematical representation of physics involved in
multifluid flow. Secondly, we find that it provides a clue as to the construction of numerical fluxes. It has led us
842 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873

to recognize various types of interactions, not only that occurring within a cell between different phases, but
also that at the cell boundaries, as illustrated in Fig. 1. The former is the in-cell interaction that gives rise to the
nonconservative term in Eq. (2), marked by A. The latter, marked by B in Fig. 1b, is a natural consequence of
the finite volume method and is consistent with the stratified model; it is however not observed previously in
the literature. We refer these interactions due to different phases collectively to as inter-phasic terms. We first
explore the stratified flow model in our earlier paper [8] and further refine the numerical procedure in [9] by
including the Riemann solver and applications to various problems. Independently, the idea of recognizing the
presence of different phases in each sub-volume appears in [10]. Their method is further employed for different
studies [11,12] in the multifluid framework.
To our knowledge, all numerical methods for multifluid flows are extended from the ones for single fluid.
The extension is not necessarily straightforward; in fact, difficulties arise because we need be concerned with
additional issues. These include: (1) disparities in fluid velocities and properties, (2) non-hyperbolicity of the
partial differential equations, (3) terms in nonconservative form, (4) surface tension force, etc. The manifesta-
tion of difficulties can be in stability, accuracy (e.g., unwanted oscillations and smearing), or uniqueness of
solution. Moreover, even if the multifluid model is rendered hyperbolic, its eigensystem is still too complicated
to be put in an analytical form, hence making it difficult to use the characteristic-based approximate Riemann
solvers such as the Roe’s scheme or the Osher’s scheme. On the other hand, the simplicity of the flux vector
splitting scheme [13] or the AUSM-family schemes [14] makes them an attractive alternative for the current
multifluid model.
The AUSM+ scheme proposed by Liou [15] is known to be accurate and robust for compressible gas flows,
especially for its ability in capturing shock and contact discontinuities. It can be easily extended to multispe-
cies equations, and it can also handle flows of very low Mach number with the help of a pre-conditioning
matrix and the numerical speed of sound [16]. While being successful for computing gas flows, the AUSM+
scheme is found to yield oscillatory solutions for liquid fluid, for which the equation of state, such as the stiff-
ened gas model [17] or the Tait’s model, is stiff. To overcome this problem due to stiffness, new diffusion terms
based on the pressure and velocity fields are introduced to the AUSM+ scheme [8,18]. The modified scheme is
essentially a variation of the AUSM+-up scheme of Liou [14,19]. These diffusion terms are used to enhance the
coupling between the pressure and velocity fields. We will show that they can effectively suppress numerical
oscillations behind the pressure waves.
The rest of the paper is organized as follows. Section 2 shows the details of our method. We will discuss the
stratified flow model, the discretization, the AUSM+-up scheme, exact Riemann solver for gas–liquid inter-

Gas
Gas

A
A
Liquid Liquid

Gas A Gas
B
A
Liquid
Liquid

Fig. 1. Illustration of the inter-phasic terms between different phases.

 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873 843

face, the interfacial pressure correction, the updating procedure, and the extension to multidimensional flows.
Section 3 presents results of several 1D and 2D problems. Finally, concluding remarks are given in Section 4.

2. Numerical method

In this section, we shall present the conservation laws based on the stratified flow model, its discrete coun-
terpart and the numerical method. We begin by defining the control volume of each phase and identifying
interfaces between the same and different fluids on the control surface. Then, we shall introduce the
AUSM+-up scheme and the exact Riemann solver we use to calculate the numerical flux. In addition, we will
discuss how the interfacial pressure correction term, which is included to make the system hyperbolic, will be
incorporated in our framework. Finally, the updating of variables requires additional attention.

2.1. The stratified flow model

By adopting the concept of stratified model [3]2, we find several advantages. It gives a clear physical inter-
pretation of the interfacial terms manifesting the interactions between phases and it allows an easy procedure
to balance the pressure terms, Eq. (2). We also show that the inviscid limit of Eq. (1) can be derived by the the
stratified flow model. As illustrated in Fig. 2, two different fluids, delineated by the void fraction function ai,
are conceptually considered to occupy in two separate regions. Within each region, a single fluid (phase) is
defined, with its properties and flow variables clearly identifiable. Hence, it is straightforward to apply the con-
servation laws to each fluid, say fluid ‘‘i’’, in partial control volume Vi of the total volume V,
Z I
o
q dV i þ ðqi~ vi Þ  ~
n dS i ¼ 0;
ot V i i Si
Z I I
o
vi dV i þ ðqi~
qi~ vi Þ  ~
vi~ n dS i þ n dS i ¼ 0;
p~ ð3Þ
ot V i Si Si
Z Z I
o oai
q Ei dV i þ p dV þ ðqi H i~ vi Þ  ~
n dS i ¼ 0;
ot V i i V ot Si

where V i ¼ ai V and Si is the surface area enclosing Vi. It is noted that the above integral equations appear in
conservative form, as in the case of single phase flow. When expressed in differential form, it can be shown that
in the inviscid limit and ignoring body and surface forces, the above equations are equivalent to Eq. (1) if ai is
continuous because V i ¼ ai V is varying with space and time.
To close this system with equal numbers of unknown variables and equations, it is necessary to supple-
ment with equations of state and the constraint ag þ al ¼ 1. In this paper, we assume that the gas and
liquid phases are respectively described by the ideal gas and stiffened gas equations of state. The latter
is given in Appendix A.
For clarity, we will first develop the numerical method for the 1D system and then extend it for the mul-
tidimensional system. In what follows, we will consider the discretization of Eq. (3) for a stationary (Eulerian)
cell located in the closed interval ½x  Mx=2; x þ Mx=2. Making appropriate substitutions of the partial volume
with the total volume, the surface integral now is broken into two parts: (1) the usual flux integral evaluated at
the cell boundaries x  Mx=2, and (2) the additional one arising at the in-cell interface (volume-fraction line).
Since there is not a net flow crossing the in-cell interface, there is no convective flux and pressure is the only
force exerted there. Thus, the final equations become
0 1 0 1 0 1 0 1
Z xþMx2 ai qi Z xþMx2 ai qi ui Z xþMx2 0 Z xþMx2 0
o B C oB 2 C B0C B oai C
a q
@ i i iAu dx þ a ðq
@ i i i u þ pÞ A dx þ p @ A dx  p @ ox A dx ¼ 0: ð4Þ
ot xMx2 xMx
2
ox xMx
2 oai xMx
2
ai qi E i ai qi ui H i ot 0

2
Here we do not require that the fluid be physically stratified; we use the concept only to facilitate the derivation of the multifluid model
in an average sense, and more importantly the discretization of numerical fluxes.
844 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873

(1-αi)
Fluid 2

αi
Fluid 1

Fig. 2. Illustration of the stratified flow.

The second term is similar to the typical flux for the single phase flow, except each phase now contributes a
portion corresponding to its volume fraction ai. Additional terms reflect interactions due to the presence of the
other fluid. The third term is the work done to phase ‘‘i’’ due to its volume change and the last term represents
the force exerted at the in-cell interface.

2.2. The discretization of the stratified model

A general discretized form of the above equations may be expressed as

0 1 0 1 0 1
ai qi 0 0
Mx B C Mx B C B C
dt @ ai qi ui A þ dx ðai f i Þj þ pj @ 0 A  pj @ dx ai A ¼ 0; ð5Þ
Mt Mt
ai qi Ei j dt ai j 0 j

where f is the numerical flux having an identical expression as for single phase flow,
T
f ¼ ðqu; qu2 þ p; quH Þ : ð6Þ
nþ1 n
The time and spatial difference operators are: dt ðÞ ¼ ðÞ  ðÞ and dx ðÞj ¼ ðÞjþ1=2  ðÞj1=2 . Note that the
integrals in Eq. (5) are accurate up to at least second order if one assumes that p and a are smooth within each
cell and the pressure pj is taken to be the cell-averaged value.
With the finite volume method, the flow variables are described by a piecewise function within the cell and
may be discontinuous at the cell boundaries. Using the reconstructed volume-fraction function and following
the framework of the stratified model, the flow configuration, which now contains an internal (within each cell)
structure separating both fluids in the cell, is depicted in Fig. 3, with gas in ðacghÞ and liquid in ðcdef Þ. That is,
one can express a discontinuous change at the interface between phases via the volume fraction function
within each cell. Three types of interfaces between the same and different fluids can be recognized at a cell
boundary, as seen in Fig. 3: the gas–gas ðab and ghÞ, liquid–liquid ðcd and ef Þ, and gas–liquid interfaces
ðbc; cg and fgÞ. Naturally, the numerical flux at the cell interface now comprises contributions from these
interfaces. Thus, Eq. (5) can be expanded as follows:
0 1 0 1 0 1
ai q i 0 0
Mx B C Mx B C B H C
dt @ ai qi ui A þ dx ð#ii f ii þ #i0 i f i0 i þ #ii0 f ii0 Þ þ pj @ 0 A  pj @ dx ai A ¼ 0: ð7Þ
Mt Mt
ai E i j dt ai j 0

d e
Liquid
ϑl-l Liquid f
Liquid
g
c
ϑl-g Gas
Gas
b
Gas ϑg-g
a h

j-1/2 j j+1/2

Fig. 3. Illustration of the one dimensional stratified flow.

 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873 845

The subscript i 0 denotes the phase being different from i, e.g., i0 ¼ l when i ¼ g. It is understood that the
letter ‘‘j’’ is reserved for mesh indexing. Also, the numerical fluxes on the interfaces between the same
fluids are denoted by fi–i and the fluxes between different fluids denoted by f ii0 and f i0 i . The functions,
#ii , #ii0 and #i0 i , denote the effective lengthes of the interfaces at the cell boundary and they are defined
as
#g–g ¼ minððag ÞL ; ðag ÞR Þ; ð8Þ
#l–l ¼ minððal ÞL ; ðal ÞR Þ; ð9Þ
#g–l ¼ maxð0; Dag Þ ¼ maxð0; Dal Þ; ð10Þ
#l–g ¼ maxð0; Dal Þ ¼ maxð0; Dag Þ; ð11Þ
where
DðÞ ¼ ðÞR  ðÞL ; ð12Þ
where subscripts ‘‘L’’ and ‘‘R’’ refer to the ‘‘left’’ and ‘‘right’’ sides of the cell interface.
Then, it is easy to show the following identities:
#g–g þ #l–l þ #g–l þ #l–g ¼ 1; ð13Þ
#g–g þ #g–l ¼ ðag ÞL ; ð14Þ
#l–l þ #l–g ¼ ðal ÞL ; ð15Þ
and
#g–l  #l–g ¼ 0: ð16Þ
The notion of effective lengthes is new. Hence, further clarification is useful. For example, #g–l will represent
the effective length of the gas–liquid interface with gas fluid on its left side (hence the left subscript) and liquid
on the right side. The definition for #l–g is similar, but with liquid on the left and gas on the right.
For the pressure force at the in-cell interface, the last term in Eq. (5), a new spatial difference operator is
warranted to reflect the fact that it is only applied to the portion inside the cell.
dH
x ðÞ ¼ ðÞjþ1=2;L  ðÞj1=2;R : ð17Þ
Noting that the difference operator in Eq. (12) is referred to the same cell interface, while dH
x is referred to two
different cell interfaces.

2.3. The AUSM+-up method

The AUSM-family schemes have been developed over the years for single phase flows; they have been
shown to possess several desirable properties which lead to a robust and accurate method for solving flows
at all speeds, see [15,19,20] for details. Its algorithm is rather simple, yet its accuracy rivals that often associ-
ated with typical flux difference schemes, such as the Roe scheme and the Osher scheme. A unique feature of
the AUSM schemes is that it can be readily applied to equations with complicated eigensystem, because it is
not explicitly based on the structure of eigensystem. In this paper we shall extend the single phase version of
the AUSM methods, specifically the one given in [19], to the multifluid system and demonstrate that the new
method, AUSM+-up, can provide accurate and stable results. In what follows, we will discuss the algorithm in
details.
For a single fluid, the AUSM+ scheme writes the interface flux fi–i as a sum of convection flux f cii and pres-
sure flux f pii :
f ii ¼ f cii þ f pii ¼ m_ ii Wi;L=R þ pii ; ð18Þ
T
in which the convected variables W ¼ ð1; u; H Þ are evaluated by a simple upwind definition. That is, depend-
ing on the sign of the interface Mach number Mi–i (whose definition is to be defined later), we define, by using
the notation ‘‘L/R’’,
846 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873

WL if M ii P 0;
Wi;L=R ¼ ð19Þ
WR otherwise:
T
The definition of mass flux m_ ii and pressure flux pii ¼ ð0; pii ; 0Þ is central in the AUSM-family methods;
hence, extra attention is required for the multifluid model, which will be described in detail below. First, the
mass flux is further expressed as
m_ ii ¼ aii M ii qi;L=R ; ð20Þ
with
M ii ¼ Mþ 
ð4Þ ðM i;L Þ þ Mð4Þ ðM i;R Þ: ð21Þ
The Mach number at the ‘‘left’’ and ‘‘right’’ states are defined as
ui;L=R
M i;L=R ¼ ; ð22Þ
aii
where ai–i is known as ‘‘numerical’’ speed of sound [16], whose specification is interesting because certain
numerically advantageous features can be obtained by using different considerations. In other words, its spec-
ification is not unique, but it is nevertheless related to and on the order of the physical speed of sound. Based
on our previous study of single phase flows [15], it is advantageous for calculating a moving discontinuity to
use a common speed of sound at the cell interface. Here, we adopt the same idea by defining a common value
for multifluid flows.
For the multifluid model, various choices for defining the speed of sound have appeared in the literature.
We employ the numerical speed of sound amix derived for stratified (separated) flow, which is given by Wallis
[21], also used in Hancox [22] and Toumi [23], based on a weighted harmonic mean of two quantities,
!
1 al ag al ag
2
þ ¼ 2
þ : ð23Þ
amix ql qg ql al qg a2g

It is noted that a2mix lies between a2g and a2l and has the following limiting values: (1) a2mix ¼ Oða2g Þ for
ag=l ¼ Oð1Þ, since ql  qg , and (2) a2mix ! Oða2l=g Þ when ag=l ! 0, that is, when one fluid disappears, the speed
of sound of the other fluid naturally assumes. This formula is different from the speed of sound appropriate for
the homogeneous flow where only one set of equations is needed and its value can fall much below ag and al,
especially when ag ¼ 0:5. It is numerically preferable that the numerical speed of sound falls between the range
of both individual fields in order to maintain smoothness of solution. By virtue of the above consideration, we
now use Eq. (23) for the common speed of sound to define the Mach number for both fluids, i.e.,
aii ¼ amix ; i ¼ l or g: ð24Þ
The polynomial functions in Eq. (21) are given as
( 

Mð1Þ if jMj P 1;
Mð4Þ ðMÞ ¼  
ð25Þ
Mð2Þ ½1  2Mð2Þ  otherwise:
where
1
M
ð1Þ ¼ ðM  jMjÞ; ð26Þ
2
 1 2
Mð2Þ ¼  ðM  1Þ : ð27Þ
4
Next, the interface pressure pi–i is defined as
pii ¼ Pþ 
ð5Þ ðM i;L Þp L þ Pð5Þ ðM i;R Þp R : ð28Þ
Notice that we do not assign the phase index ‘‘i’’ to the pressure pL or pR because we assume a single pressure
common to both phases in the two-fluid model under consideration. Similar to the Mach number polynomials,
we also have the corresponding ones for pressure, but using the same set of basis polynomials,
 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873 847
( 1
M
M
ð1Þ if jMj P 1;
P
ð5Þ ðMÞ ¼ ð29Þ
M
ð2Þ ½2 M  3MM
ð2Þ  otherwise:
Our previous work [8] shows that when the equations of state for water is used, both the AUSM+ and
AUSM+-a [14] schemes generate numerical oscillations near discontinuities. To suppress these oscillations,
we can add some properly-scaled dissipation terms to the AUSM+ scheme to enhance the coupling between
the mass flux (or convection terms) and pressure. We respectively introduce into the convection flux a dissi-
pation term Dp based on pressure difference and into the pressure flux a dissipation term Du based on velocity
difference. The use of the dissipation terms based on velocity and pressure are originally proposed to solve low
speed single phase flows [18–20]; they can effectively enhance the stability and convergence. Here, we rewrite
the mass flow rate m_ ii and the interface pressure pi–i in the following forms:
m_ ii ¼ Eq: ð20Þ þ Dp ii ; ð30Þ
pii ¼ Eq: ð28Þ þ Duii : ð31Þ
When dealing with the compressible liquid fluid, we adopt the same idea and make appropriate scaling adjust-
ments to the dissipation terms. Formal derivation of the scaling of the dissipation terms can be found in [19].
Specifically, we write Dp and Du as:
DM i maxð1  M 2i ; 0ÞðpL  pR Þ
Dp ii ¼ jp ; ð32Þ

ai
Duii ¼ ju Pþ 
ð5Þ ðM i ÞPð5Þ ðM i Þ ai ðui;L  ui;R Þ;
qi  ð33Þ
where
DM i ¼ Mþ þ  
ð4Þ ðM i;L Þ  Mð1Þ ðM i;L Þ  Mð4Þ ðM i;R Þ þ Mð1Þ ðM i;R Þ:

The parameters a i and M i here are obtained with the arithmetic mean of the ‘‘L’’ and ‘‘R’’ states, i.e.
i ¼ ðai;L þ ai;R Þ=2. In this paper, the AUSM+-up scheme is uniformly applied to gas–gas and liquid–liquid
a
interface. And the coefficients jp ¼ ju ¼ 1 are used for all the test problems, unless they are noted otherwise,
e.g. jp ¼ ju ¼ 0.

2.4. Exact Riemann solver for gas–liquid interface

As noted above, we use AUSM+-up scheme to calculate the numerical flux between the same fluid, f g–g and
f l–l . What remains to be done is the evaluation of numerical flux between different fluids, f g–l and f l–g . To our
knowledge, no exact Riemann solver is known in the literature for numerical flux of a general two-fluid model
of a real fluid. Fortunately, because the stiffened gas EOS is only slightly different from the ideal gas EOS, an
exact form of the Riemann solution is possible, requiring a small numerical iteration for solving pressure. It
involves exactly the same procedure as in the ideal gas case. Since there is still an algebraic difference from that
of ideal gas, it is instructive to show the derivation of the exact Riemann solution. Its details are given in
Appendix B for the interested reader3.
Since the exact Riemann solver is still more expensive to compute than other approximate solvers, we shall
only use the exact Riemann solver in the vicinity of the fluid interface. More specifically, we use the exact Rie-
mann solver to calculate f ii0 in Eq. (7) when #ii0 > 5, where  is used to define the minimum value of ai and,
generally, we have  ¼ 106  108 in our computation. Otherwise, we just use the simple upwind scheme to
calculate the convection flux and use the central difference scheme to calculate the pressure flux.
Referring to Appendix B, the fluid velocity in the transition region uw can be calculated by Eq. (B.3) or
(B.5). The numerical fluxes across the inter-phasic interface depend on the relative position of fluids and
the direction of the cell boundary. Considering the situation where ðag ÞL > ðag ÞR , hence #g–l > 0 and
#l–g ¼ 0. Then for the left cell (j – 1), the numerical flux on the cell boundary is

3
One of the reviewers points out that the exact Riemann solver was given for the stiffened gas in [24], which we were not aware of. Since
this work is written in Russian and no English translation of it is available, we include its derivation in Appendix B for completeness.
848 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873

8 8 9 9
>
> maxð0; uH ÞðqH
g ÞL > 0 > >
>
>
> >
> >
> >
>
>
> H H
maxð0; u Þðqg ÞL u H >
> pH >
> >
>
>
> >
> >
> >
>
>
< H H H H < 0 >
> = >
=
maxð0; u Þððqg ÞL ðEg ÞL þ p Þ
f g–l ¼ þ : ð34Þ
>
> minð0; uH ÞðqH ÞR >
> >
> 0 >
>
>
> l >
> >
> >
>
>
> >
> >
> 0 >
>
>
> minð0; uH ÞðqH l ÞR u
H >
> >
> >
>
>
: >
; : ;
H H H H
minð0; u Þððql ÞR ðEl ÞR þ p Þ 0

And for the right cell (j), the numerical flux is

8 9 8 9
>
> maxð0; uH ÞðqH g ÞL >
> > 0 >
>
> >
> >
> >
>
>
> H
maxð0; u Þðqg ÞL u H H >
> >
> 0 >
>
>
> >
> >
> >
>
>
< H H H
> >
H = < 0 >
=
maxð0; u Þððqg ÞL ðEg ÞL þ p Þ
f g–l ¼ þ ; ð35Þ
>
> minð0; uH ÞðqH ÞR >
> > > 0 >
>
>
> l >
> >
> H>>
>
> >
> > >p >>
>
> minð0; u H
Þðq H
Þ u H >
> >
> >
>
:
l R >
; : > ;
H H H
minð0; u Þððq Þ ðE Þ þ p Þ H 0
l R l R

where the first and second terms in the RHS of above equations represent the convection flux and pressure flux
respectively. It should be noted that the pressure in the momentum equation is applied to the gas fluid in the
left cell and liquid fluid in the right cell. This is the inter-phasic interaction across the cell boundary, which
makes our method different from others.
We remark that recognizing the use of flux contributions between unlike phases, f g–l and f l–g , is essential in
the present formulation. However, the numerical representation of these fluxes is open to different ideas, as in
the development of single fluid numerical fluxes; using the exact Riemann solver appears to be a safe route
since we did not know anything better than it. As will be discussed later in Section 3.4, the substitute of
our previous rough approximation [8] with the current one has yielded improvements in accuracy and robust-
ness. On the other hand, there are disadvantages with using the exact Riemann solver, especially in the case of
general fluids having a complex equation of state, thus preventing from obtaining an analytical equation to
solve for pw. It is likely that one can develop other robust and accurate ways of dealing with the fluxes between
mixed fluids. A research in this direction will be very desirable since it is general and the one for single fluid is
just its special case.

2.5. Summary of the numerical fluxes for the multifluid model

In summary, we can give the numerical fluxes in Eq. (7) in a general form. For simplicity, we shall inter-
changeably use the following subscripts: 1=2 j þ 1=2, and 1=2 j  1=2. Then f i;1=2 denotes the numer-
ical fluxes of fluid i at the cell boundaries j  1=2 of cell j. Substituting the component fluxes developed above,
we have the numerical flux in Eq. (7) expanded as
f i;1=2 ¼ ½#ii f ii þ #ii0 f ii0 þ #i0 i f i0 i 1=2
¼ #ii;1=2 ½ðaii M ii qi;L=R þ Dp ii ÞWi;L=R þ ð0; pii ; 0ÞT 1=2 þ #ii0 ;1=2 ½maxð0; uH ÞqH H
i Wi
T H
þ ð0; pH ; 0Þ 1=2 þ #i0 i;1=2 minð0; uH ÞðqH
i Wi Þ1=2 ; ð36Þ
where
pii;1=2 ¼ ½Pþ 
ð5Þ ðM i;L Þp i;L þ Pð5Þ ðM i;R Þp i;R þ Duii 1=2 :

And the numerical flux f i;1=2 for cell j is

T H
f i;1=2 ¼ #ii;1=2 ½ðaii M ii qi;L=R þ Dp ii ÞWi;L=R þ ð0; pii ; 0Þ 1=2 þ #ii0 ;1=2 maxð0; uH ÞðqH
i Wi Þ1=2
H T
þ #i0 i;1=2 ½minð0; uH ÞqH H
i Wi þ ð0; p ; 0Þ 1=2 : ð37Þ
 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873 849

A benefit of using the stratified flow method is that it automatically meets the pressure non-disturbing
condition described in Introduction. To show this, we consider the case of a moving contact discontinuity
across which velocity and pressure are continuous, i.e., pL ¼ pR ¼ p and uL ¼ uR ¼ u. Hence it is easy to
show
p1=2;L=R ¼ pii;1=2 ¼ pH
1=2 ¼ p; ð38Þ

and
u1=2;L=R ¼ ðai M i Þ1=2 ¼ uH
1=2 ¼ u: ð39Þ
First, referring to Eqs. (7), (36) and (37), we collect all the pressure terms in the momentum equation and de-
note it as Cm:
Cm ¼ ½#ii pii þ #ii0 pH 1=2  ½#ii pii þ #i0 i pH 1=2  pj dH
x ai : ð40Þ
Since
½#ii þ #ii0 1=2 ¼ ðai Þ1=2;L ;
½#ii þ #i0 i 1=2 ¼ ðai Þ1=2;R : ð41Þ
dH
x ðÞ ¼ ðÞ1=2;L  ðÞ1=2;R :

Then, Eq. (40) simplifies to

Cm ¼ p½#ii þ #ii0 1=2  p½#ii þ #ii0 1=2  pdH
x ai ¼ 0: ð42Þ
Similarly in the energy equation, we can also collect all the pressure terms and denote it by CE:
4x
CE ¼ p dt aj þ ½#ii ai M i pL=R þ #ii0 maxð0; uH ÞpH þ #i0 i minð0; uH ÞpH 1=2
4t
 ½#ii ai M i pL=R þ #ii0 maxð0; uH ÞpH þ #i0 i minð0; uH ÞpH 1=2 : ð43Þ

It is further reduced to
4x
CE ¼ p dt aj þp½#ii uþ#ii0 maxð0;uÞþ#i0 i minð0;uÞ1=2 p½#ii uþ#ii0 maxð0;uÞþ#i0 i minð0;uÞ1=2 : ð44Þ
4t
If u P 0, we get
 
4x
CE ¼ p dt aj þ u½ðai Þ1=2;L  ðai Þ1=2;L  ¼ 0; ð45Þ
4t

by virtue of the discrete form of the transport equation of volume fraction function. For uj < 0, we also have
 
4x
CE ¼ p dt aj þ u½ðai Þ1=2;R  ðai Þ1=2;R  ¼ 0: ð46Þ
4t

Consequently, all the terms involved with pressure are summed up to zero, the system is left with convective
fluxes only, which are convected with the constant speed u. Thus, the pressure non-disturbing condition is met
and the stratified flow model can capture a stationary contact discontinuity exactly.

2.6. Multi-dimensional extension

To extend the above 1D method to higher dimension, we first write the general version of Eq. (4) as
0 1 0 1 0 1
Z ai qi I Z 0 Z 0
o B C o B C B C
@ ai qi~ vi A dV þ ðai f i Þ dS þ p@ 0 A dV  p@ rai A dV ¼ 0: ð47Þ
ot V ot V V
ai q i E i ai 0
850 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873

After discretization performed as before, the above equation becomes

0 1 0 1 0 1
ai qi 0 0
Vj B C X
m
VjB C Xm
B ~C
dt @ ai qi~vi A þ ðai f i Þl S l þ pj @ 0 A  pj @ ai;L n A S l ¼ 0; ð48Þ
Mt Mt
ai qi Ei j l¼1 dt a i j l¼1
0 l

with
ai f i ¼ ð#ii f ii þ #i0 i f i0 i þ #ii0 f ii0 Þ: ð49Þ

Here m is the number of cell faces enclosing cell j. All of the cell faces are defined to be directed outward from
the cell and ai;L is the reconstructed void function located just at the inner side of cell faces.

2.7. Interfacial pressure correction term

It is well known that the system of inviscid multifluid equations, Eq. (1), is non-hyperbolic. To make the
system hyperbolic, thus well-posed, an interfacial pressure correction term in the form of first spatial deriva-
tive can be added in order to alter the characteristics of the system. Hence, the original multifluid model of Eq.
(1), enhanced with the pressure correction, pint, now appears as
oðai qi Þ
þ r  ðai qi~vi Þ ¼ 0;
ot
oðai qi~vi Þ
þ r  ðai qi~ vi Þ þ rðai pÞ ¼ ðp  pint Þrai ;
vi~ ð50Þ
ot
oðai qi Ei Þ oai
þ r  ðai qi H i~vi Þ ¼ ðp  pint Þ :
ot ot
The pressure correction term we adopt is the one first proposed by Stuhmiller [5],
ag qg al ql 2
pint ¼ r j~
vg ~
vl j : ð51Þ
ag ql þ al qg
where r is a parameter and the system becomes hyperbolic if
r P 1; ð52Þ
under the low relative Mach number condition of j~ vg ~vl j ag . Unfortunately this condition is shown re-
cently in [25] to be insufficient to guarantee hyperbolicity if the relative Mach number is finite, it requires a
much larger value than unity, depending on the flow conditions and ai. The effect of r on the solution will
be addressed in detail in the Ransom’s faucet problem in Section 3.2.
Here we adopt the idea of pressure correction into our framework of stratified fluid and introduce pressure
correction pint into Eq. (48). As described above, the stratified fluid model suggests two types of inter-phasic
fluxes – one is within the same cell and the other between neighboring cells. Recognizing this difference makes
it quite clear as to how the interfacial pressure correction term should be implemented in the discretized equa-
tion, specifically with respect to the second, third, and fourth terms in Eq. (48). That is to say, only those pres-
sure fluxes that envisage interactions between different fluids need be implemented with the pressure
correction. Hence, Eq. (48) becomes
0 1
ai qi
C X
m
Vj B
vi A þ
dt @ ai qi~ ½#ii f ii þ #ii0 ðf ii0  pint int
ii0 Þ þ #i i ðf i i  pi0 i Þl S l
0 0
Dt l¼1
ai E i
0 1 0 1
0 0
VjB C XBm
C
þ ðpj  pint j Þ @ 0 A  ðpj  pint j Þ @ ai;L~ n A S l ¼ 0: ð53Þ
Dt l¼1
dt ai 0 l
 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873 851

The interfacial pressure correction is seen to appear in two places, respectively as pint
j (which is simply Eq. (51)
evaluated at ‘‘j’’) for the interface within the cell and pint
ii 0 between the cells, which is given as follows.

pint int
n; 0ÞT ;
ii0 ¼ p ii0 ð0;~ ð54Þ

where
^ ai0 ðqi Þj ðqi0 Þj0
ai ^ 2
pint
ii0 ¼ r jð~
vi Þj  ð~
vi 0 Þ j0 j ; ð55Þ
ai ðqi0 Þj0 þ ^
^ ai0 ðqi Þj
and
1
ai ¼ ððai Þj þ ðai Þj0 Þ ¼ 1  ^
^ ai0 : ð56Þ
2
In the above equations, we use the subscript j 0 to denote the cell that neighbors cell j through the cell
boundary l.
The introduction of the interfacial pressure correction pint is to improve the stability of the system, and it
should not be too large to change the physics of the flow. In most of our test problems, pint is much smaller
than the local pressure pj in the flow. However, under some extreme conditions it is possible that pint is larger
than the surrounding pressures, thereby resulting in a significant departure from the original problem. Hence,
an additional limit on the pressure correction terms is imposed.
pint int
j ¼ minðp j ; ep p j Þ; ð57Þ
and
pint int
ii0 ¼ minðpii0 ; ep minðp j ; pj0 ÞÞ: ð58Þ
2 3
In our practice, we usually set ep ¼ 10  10 , which has been found sufficient to keep the simulation well-
behaved4. Several runs were conducted to investigate the effects on the solution due to ep for some computa-
tionally-challenging problems, e.g., shock-droplet problem in Section 3.5; no noticeable effects were found for
it varying from 0.001 to 0.1.

2.8. Time integration method and updating of variables

A multi-step Runge–Kutta method is used for advancing the solution in time and we will omit its descrip-
tion since it is rather standard. However, unlike the single fluid ideal gas flow, the updating of variables at the
new time level for multifluid equations gives rise to a new problem requiring additional care. Below describes
the procedure used in this study. For two dimensional flow, the time discretization of Eq. (53) may be written
as:
0 1 0 1 0 1
Wi asi qsi ani qni
BM C B asi qsi usi C B ani qni uni C Dt
B iC B C B C
B C¼B s s s C ¼ B n n n C  x Ri ; ð59Þ
@ Ni A @ ai qi vi A @ ai qi vi A V
Ei asi qsi Esi þ ðpn  pint s
j Þai ani qni Eni þ ðpn  pint n
j Þai

where Ri is the summation of the second and fourth terms in Eq. (53). The superscripts n and s are the indices
for the present and next time steps respectively, and x is the parameter used for each sub-step of the Runge–
Kutta method. To calculate the primitive variables at the next time step, the EOS of fluids must be applied to

4
The physical meaning of the interfacial pressure is still debated every now and then and the correctness of the form, aside from
rudimentary arguments, is far from resolved. Hence, it is argued here that this additive term must be small compared to the surrounding
pressure (which is physical) so that the physical problem at hand is not significantly altered and yet it is sufficiently large to maintain the
solution well behaved (a manifestation of hyperbolicity).
852 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873

the above equation to update the pressure via an algebraic equation: a quadratic equation in the case of using
the stiffened gas model,
ðps Þ2 þ ðA þ B þ DÞps  AC þ BD ¼ 0; ð60Þ
where
A ¼ ðcg  1ÞðEg  Hg Þ;
B ¼ ðcl  1ÞðEl  Hl  pn Þ þ cl p1 ;
C ¼ cl p1 þ ðcl  1Þðpn  pint
j Þ;

D ¼ ðcg  1Þðpn  pint

j Þ;

and
M2i þ N2i
Hi ¼ :
2Wi
The pressure ps is the positive root of the above equation. Then the other primitive variables can be derived
easily,
Eg  Hg
asg ¼ ps ; asl ¼ 1  asg ;
cg 1
þ ðpn  pint
j Þ

Mi Ni 1 1 s 2 s 2
usi ¼ ; vsi ¼ ; esi ¼ ðEi  asi ðpn  pint
j ÞÞ  ððui Þ þ ðvi Þ Þ;
Wi Wi Wi 2
ðcg  1Þesg c es ðps þ p1 Þ
T sg ¼ ; T sl ¼ l l s :
Rg ðC p Þl ðp þ cl p1 Þ
Notice that the coefficients in Eq. (60) differ enormously, at least on the order of p1 of the liquid. The ratio
between the largest and smallest coefficient might be as large as 1018 in some of our test cases, resulting in a
large numerical error in solving ps and asg , even when double precision is used in the calculation. Hence, an
additional Newton iteration method is used to improve the accuracy. From the energy equation of both fluids
in Eq. (59), ps and asg are solved from the following equations simultaneously.
!
ps int
F g ¼ ags
þ ðp  pj Þ þ Hsg  Esg ¼ 0;
n
ð61Þ
cg  1
 s 
p þ cl p 1
F l ¼ ð1  asg Þ þ ðpn  pint
j Þ þ Hsl  Esl ¼ 0: ð62Þ
cl  1
Here Hsl=g and Esl=g are the known quantities from the current updates of conservative variables as given in Eq.
(59). To obtain ðasg ; ps Þ, the above two equations are solved simultaneously via the Newton iteration method
with the analytical solution of Eq. (60) as the initial guess. Generally, we can drive the functions, F g and F l ,
from O(103) to O(105) within 20 iterations. We emphasize that this additional step is required to ensure the
decoding is precise, free of roundoff errors; it is performed at each time step before continuing on to the new
time step. It should not be confused with the pressure relaxation used in [10] because this iterative procedure is
merely an attempt to get numerically consistent solutions to Eqs. (61) and (62) before continuing on to the
next time step. More discussion of this procedure is given in [8].
Another concern when solving Eq. (53) is that the numerical error of variables may be amplified when one
of the phase is disappearing (or similarly appearing in full), i.e., ag or al ! 0. In this case, although the amount
of the fluid is near void and its contributions to the flow field is essentially negligible, the calculation however
can still become unstable if any variable associated to this fluid diverges. To control the numerical error
(numerical instability) when one phase of fluid vanishes, we adopt the idea suggested in Paillère et al. [26],
in which the fluids assume equilibrium immediately by mixing (blending) fluid states. For example, if asg , which
is obtained after the time integration, approaches zero and falls in the range min 6 asg 6 max , then we re-set the
velocity and temperature fields of the disappearing phase by blending,
 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873 853

400 1.2

ug , εmax=1.0e-3
350
ug , εmax=1.0e-6 1
300 ul , εmax=1.0e-3
ul , εmax=1.0e-6 0.8
250
ug , u l

200 0.6

αg
150 0.4

100
αg
0.2
50

0
0
2 4 6 8 10
X
Fig. 4. Effect of max on the solution of air–water shock tube problem (see Section 3.3); min ¼ 107 .

ðqsg Þadjust ¼ Gðng Þqsg þ ð1  Gðng ÞÞqsl ; q ¼ u; T ð63Þ

The function GðnÞ; n 2 ½0; 1; is a smooth cubic polynomial interpolation with the properties: Gð0Þ ¼ 0,
Gð1Þ ¼ 1 and G0 ð0Þ ¼ G0 ð1Þ ¼ 0. One easily finds the choice,
GðnÞ ¼ n2 ð2n  3Þ;
where the normalized parameter is defined as
a  min
n¼ :
max  min
In these equations, min > 0 is a very small value defined by the lowest limit of ai ; i ¼ g; l, e.g.,
min ¼ minðag ; 1  ag Þ in the problem or a pre-set value (106  107 in this paper), whichever is smaller; we
chose max ¼ 10min  100max in our calculations.
We present in Fig. 4 the effect on the solution of the air–water shock tube (see Section 3.3 for details) by
varying the value of max in a wide range, 103 6 max 6 106 , while min is set to be the same as ðag Þmin ¼ 107 .
In the figure, the solutions using max ¼ 103 and 106 are indistinguishable (maximum difference in jug  ul j is
0.05 m/s). It confirms that the solutions are convergent and the present method is robust and accurate to han-
dle the situation of phase disappearing/re-appearing.

3. Results and discussion

The 4-stage Runge–Kutta method is used for the time integration. The Osher–Chakravarthy TVD scheme
[27] is chosen to provide third-order spatial accuracy for all results presented in this section, unless noted
otherwise.
Several one and two dimensional test problems are used to study the capability of our method. We apply
the new method described above to simulate several 1D multiphase problem, such as the Ransom’s faucet
problem and air–water shock tube problems. Then, two dimensional problems will be presented, including
the shock–droplet interactions, underwater explosion, and underwater shock–bubble interactions.
For ease of referencing, we list in Table 1 the parameters relevant to the computation of each problem in
this paper. The interfacial pressure constant is set to 2 throughout, except in the Ransom’s faucet problem
where a range of values are tested to show its effect on the solution. The CFL number is set to 0.5 for all cases
except in the case of shock and water column(s) interaction for which a smaller value of 0.3 is used to maintain
854 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873

Table 1
Parameters used in computed problems
Problem r CFL (jp)g (ju)g (jp)l (ju)l ep
Moving contact 2 0.5 0/1 0/1 0/1 0/1 102
Ransom’s Faucet 2 0.5 0/1 0/1 0/1 0/1 102
Air–water shock tube 2 0.5 1 1 1 1 102
Water–air shock tube 2 0.5 1 1 1 1 102
Shock and water column 2 0.3 0 0 1 1 102
Shock and bubble 2 0.5 0 0 1 1 102
Underwater explosion 2 0.5 1 1 1 1 102
See Eq. (51) for r, (32) and (33) for ðjp ; ju Þ, and (57) for ep. For the cases of the first two rows, 0/1 denotes that both values were used,
resulting in no visible differences in the solutions.

stability. The diffusion terms Dp ðjp Þ and Du ðju Þ in Eqs. (32) and (33), used in the mass and pressure fluxes, are
used in most cases. For the moving discontinuity problem and faucet problem, we do not use these diffusion
terms since the pressure is basically constant throughout the computation domain. We also turn off these
terms in the gas fluxes in the shock–water column and shock–bubble problems.
Unless stated otherwise, all variables are presented in the SI units, e.g., u in [m/s], T in [K], and p in Pa.

1.1
1.0
0.9
0.8
AUSM+-up
0.7 Analytical
0.6
αg

0.5
0.4
0.3
0.2
0.1
0.0
-0.1
0 2 4 6 8 10
X
110000.0

107500.0
AUSM+-up
Analytical
105000.0

102500.0

100000.0
p

97500.0

95000.0

92500.0

90000.0
0 2 4 6 8 10
X
Fig. 5. Void fraction and pressure profiles of a moving discontinuity problem at time t = 0.03 s, 200 cells.
 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873 855

3.1. The moving discontinuity problem

We first applied our method to the moving contact discontinuity problem as a basic test. To accurately cap-
ture the moving discontinuity a numerical scheme is required to keep the positivity of the void fraction func-
tion and the uniformity of the pressure field as the flow continues. We set the initially condition as
ðp; ag ; ui ; T i ÞL ¼ ð105 Pa; 1:0  ; 100:0 m=s; 300:0 KÞ;
ðp; ag ; ui ; T i ÞR ¼ ð105 Pa; ; 100:0 m=s; 300:0 KÞ;
where  ¼ 1:0 108 . A mesh of 200 cells was used for the computation. For this problem and the Ransom’s
faucet problem, we use jp ¼ ju ¼ 0, as there is no sharp pressure gradient in the flow.
The result at time t = 0.03 s is presented in Fig. 5. The proposed method is able to capture the contact dis-
continuity accurately without inflicting any disturbance in the pressure field, confirming that the pressure non-
disturbance condition is satisfied and is consistent with the analysis given in Section 2.5.

3.2. Ransom’s faucet problem

Referring to Fig. 6, the faucet problem introduced by Ransom [28] consists of air and a water jet confined in
a channel, with the flow configurations at three time instants depicted. In the beginning, both the water col-
umn and the surrounding air are moving at a constant speed of 10 m/s. Then the gravity force is applied to the
fluid, causing the water column to accelerate and get narrower so that the mass conservation is met. This is
accompanied with a void wave moving toward the outlet. Finally, the flow becomes steady when its wavefront
moves out of the computational domain. An analytical solution can be derived by assuming that the water is
incompressible and the effect of air negligible.
We set up the inflow boundary condition as: the void fraction of air ag ¼ 0:2, the velocities of both air and
water ug ¼ ul ¼ 10 m=s, the temperatures T g ¼ T l ¼ 300 K, and the pressure is extrapolated from the interior
point. At the outflow boundary, we set the pressure p ¼ 105 Pa, and all the other primitive variables are
extrapolated from inside. A mesh of 500 cells was used for the computation.
Fig. 7 shows the time evolution of the void fraction profile. The computed result compares well with the
analytical solution and there are no oscillations in the result. It should be noted that most of the published
results show a severe overshoot in void fraction profile behind the advancing wavefront. It is noted that

10 m/s 10 m/s 10 m/s

Water

Initial condition Transition state Steadystate

Fig. 6. Illustration of Ransom’s water faucet problem.

856 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873

any numerical scheme that does not satisfy the pressure non-disturbing condition will generate instability in
the region where rai 6¼ 0, i.e., in the vicinity of fluid interface. This test demonstrates that the present method
automatically satisfies the pressure non-disturbance condition at the discrete level and is able to enhance the
stability of simulation.
Fig. 8 demonstrates how the interfacial pressure correction pint influences the simulation result. Void frac-
tion profiles calculated by different r are shown. When r ¼ 0, the multifluid model is non-hyperbolic and ill-
posed mathematically. Although the wavefront is sharply captured, there is still a minor glitch in the profile.

0.5
Analytical,t=0.5sec
A t= 0.1 sec
0.45
B t= 0.2 sec
C t= 0.3 sec
0.4 D t= 0.4 sec
E t= 0.5 sec
E
0.35
g

D
0.3
C

0.25 B
A

0.2

0.15
0 2 4 6 8
x
Fig. 7. Time evolution of the void fraction profile in the water faucet problem.

Analytical
0.45 A σ =0
B σ =2.0 A
C σ =5.0
0.4 D σ =10.0

0.35
g

0.3 0.46

0.45

0.25 0.44
D C B A
0.43
B C D
0.2 5.2 5.4 5.6 5.8 6 6.2

0 2 4 6 8
x
Fig. 8. The void fraction profiles based on different r in the water faucet problem at time t = 0.5 s.
 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873 857

The use of the interface pressure correction pint (r P 1) can effectively stabilize the simulation. As shown in the
figure, the void fraction profile is smeared as r increases, indicating that additional numerical dissipation is
introduced, especially in front of the wave. In the rest of our test problems, we will simply set r ¼ 2:0.

3.3. Air-to-water shock tube problem

We consider the air–water shock tube problem with nearly pure air and water separated by a diaphragm. A
large pressure ratio of 104 is imposed across the diaphragm. First, we set the initial condition as follows:
ðp; ag ; ui ; T i ÞL ¼ ð109 Pa; 1  ; 0 m=s; 308:15 KÞ;
ðp; ag ; ui ; T i ÞR ¼ ð105 Pa; ; 0 m=s; 308:15 KÞ:
where  ¼ 1:0 107 . Two different meshes of 500 and 5000 cells are used for the computation. The fine mesh
solution is used as a reference. The results are presented in Fig. 9.
In this case, the pressure in the air side is much higher than that in the water side. A very strong shock is
transmitting into the water, and a rarefaction wave is travelling back into the air. As the sound speed in water
is much faster than in air, the speed of shock is much higher than the rarefaction wave. The shock wave, rare-
faction wave and the fluid interface are sharply resolved.

1
1E+09 N= 500
N = 5000
0.8 N= 500 8E+08
N = 5000
0.6
6E+08
αg

0.4 4E+08

0.2 2E+08

0 0

0 2 4 6 8 10 0 2 4 6 8 10
x x
300

N= 500 340
N= 500
250 N = 5000 N = 5000

320
200
300
Tavg
uavg

150
280

100
260

50
240

0 220
0 2 4 6 8 10 0 2 4 6 8 10
x x

Fig. 9. State profiles for the air–water shock tube problem. The initial condition is ðp; ag ; ui ; T i ÞL ¼ ð1:0 109 ; 1  ; 0; 308:15Þ,
ðp; ag ; ui ; T i ÞR ¼ ð1:0 105 ; ; 0; 308:15Þ,  ¼ 1:0 107 .
858 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873

5E+08

κp= κu=0 .1
4E+08
κp= κu=0 .5
κp= κu=1 .0
3E+08 κp= κu=3 .0

2E+08

1E+08

8.8 8.9 9 9.1 9.2 9.3

X
Fig. 10. Effects of diffusion parameters ju and jp on the solution of the previous air–water shock tube problem.

Next, we use this case to study the influence of diffusion parameters used in the AUSM+-up, namely ju and
jp, on the solution. Fig. 10 shows that a small amount of diffusions, ju ¼ jp ¼ 0:1, are not sufficient to remove
oscillations and a monotonic profile is obtained by setting their values beyond 0.5. The values of ju ¼ jp ¼ 1:0
appear to be a good balance.

3.4. Water-to-air shock tube problem

Next we change the initial condition so that the shock wave propagates from water to air. We set
ðp; ag ; ui ; T i ÞL ¼ ð1:0 107 Pa; ; 0 m=s; 308:15 KÞ;
ðp; ag ; ui ; T i ÞR ¼ ð5:0 106 Pa; 1  ; 0 m=s; 308:15 KÞ:
The results are shown in Fig. 11. It is interesting to note that due to the differences in density and compress-
ibility, it is easy for the shock wave to transmit from air into water as shown in the previous example. On the
other hand, it is difficult for the shock to travel from water into air in the sense that a shock wave with only a
small pressure ratio is transmitted into air, and a very strong rarefaction is generated from the fluid interface
and travelling back into the water. We find that our current method is accurate, providing an improvement in
accuracy over our previous method [8] in which a rough scheme was used to represent the flux between unlike
phases, as shown in detail by the enlarged insert in the figure. It is also more robust with the current method
for a relatively high pressure ratio between water and air. However, it becomes less robust when we increase
the pressure ratio to 1000, for example in the case of the following initial condition:
ðp; ag ; ui ; T i ÞL ¼ ð1:0 108 Pa; ; 0 m=s; 308:15 KÞ;
ðp; ag ; ui ; T i ÞR ¼ ð1:0 105 Pa; 1  ; 0 m=s; 308:15 KÞ:
As mentioned in Section 2.8, the numerical error in the limit as ai ! 0 is amplified significantly with the large
pressure ratio across the rarefaction wave. To avoid divergence in the solution, we need to increase the min-
imum value of ai from  ¼ 107 to  ¼ 105 . It is noted that the elevation of the amount of the diminishing
phase does not result in a significant impact on the overall behavior of the solution since the magnitude of
 ¼ 105 is still negligibly small.
We present the results in Fig. 12. The strength of the shock wave in air is much smaller than the rarefaction
wave in water; the pressure ratio across the rarefaction wave is about 790 times bigger than the pressure ratio
 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873 859

1 1E+07
Chang & Liou (2003)
N= 500
0.8 9E+06 N = 5000

5.6E+06
0.6 8E+06
5.4E+06
αg

p
0.4 7E+06 5.2E+06

5E+06
0.2 6E+06
0.8 1 1.2 1.4 1.6

0 5E+06

0 2 4 6 8 10 0 2 4 6 8 10
x x

301
3.5

2.5 300.5
T avg
u avg

1.5 300

0.5
299.5

0 2 4 6 8 10 0 2 4 6 8 10
x x

Fig. 11. State profiles for the water–air shock tube problem. The initial condition is ðp; ag ; ui ; T i ÞL ¼ ð1:0 107 ; ; 0; 308:15Þ,
ðp; ag ; ui ; T i ÞR ¼ ð5:0 106 ; 1  ; 0; 308:15Þ, and  ¼ 1:0 107 .

across the shock wave in air. Since the shock is too weak to be recognized in the pressure profile, we simply use
log scale for the display. It is clear that solutions based on both meshes agree very well.
From the above 1D tests, we may conclude that the AUSM+-up scheme is able to provide accurate and
convergent solutions under conditions with large disparities in density, pressure and void fraction, in which
the shock and rarefaction waves can be captured well. In what follows, we shall demonstrate its capability
for several 2D problems, complicated in flow patterns and rich in flow physics.

3.5. Shock and water–column interaction problem

Here we study phenomena caused by a planar shock wave impinging upon a cylindrical water column. The
configuration is shown in Fig. 13. Initially we have a 2D water column with diameter of 3.5 mm located at the
origin and an air shock wave set up at x = 2.0 mm. The initial condition is given as: before the shock wave,
p ¼ 1:0 105 Pa; ui ¼ vi ¼ 0 m=s; T i ¼ 346:98 K;
and the states behind the shock,
p ¼ 1:0333 106 Pa; ui ¼ 831:48 m=s; vi ¼ 0 m=s; T i ¼ 929:57 K:
860 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873

8
1 10
N= 500
N = 5000
0.8
N= 500
107
N = 5000
0.6
αg

p
0.4
106

0.2

5
0 10

0 2 4 6 8 1 0 2 4 6 8 1
x x
80
N= 500
N= 500
N = 5000 320
N = 5000
60

310
T avg
u avg

40

300

20

290

0 2 4 6 8 1 0 2 4 6 8 1
x x

Fig. 12. State profiles for the water–air shock tube problem. The initial condition is ðp; ag ; ui ; T i ÞL ¼ ð1:0 108 ; ; 0; 308:15Þ,
ðp; ag ; ui ; T i ÞR ¼ ð1:0 105 ; 1  ; 0; 308:15Þ, and  ¼ 1:0 105 .

Incoming shock

Water (0.0mm,0.0mm)

Air r=1.75mm Air

2.0 mm

Fig. 13. Illustration of a shock–water column interaction problem.

 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873 861

This yields a shock wave moving at Mach 3 in the air. A mesh of about 670,000 cells is used, covering the
domain [0.03 m,0.04 m] · [0.02 m, 0.02 m]. Fig. 14 shows the time evolution of the simulation result with
two different pictures shown side by side for each time frame. One shows the pressure contours together with
contour lines of ag ¼ 0:1; 0:5 and 0.9 to identify the fluid interface. The other gives contours of density gradi-
ent 5qavg (or numerical Schlieren photo). Because of the large difference in compressibility between air and
water, we modify the scale of density gradient to accentuate the details of the flow structure. Here, the function
ð1 þ a2l Þ logðj5qj þ 1Þ is chosen. We find this modification is able to describe the pressure wave within the
water column well.
As in the previous 1D air–water shock tube problem, when the incident shock hits the water column,
part of the shock wave is transmitted into the water and part of it is reflected. Since the compressibility
of water is much smaller than the air, the shock wave within the water column moves much faster than
the shock wave in the air, as evident in (Fig. 14a). When the penetrated shock wave hits the curved interface
in the rear of the water cylinder, most of it is reflected as a rarefaction wave, just the opposite of what hap-
pens in the front part (or similar to the case of water–air shock tube). The curved interface would act as an
‘‘optical’’ mirror and the reflected rarefaction waves are focused in the rear area, as shown in Fig. 14b and
c. The rarefaction waves will be reinforced, eventually creating a low pressure region. As the pressure
approaches zero, the air begins to expand and creates a cavitation-like structure in the rear of the water
column, even though the air takes only a very small volume fraction in our initial setting (ag ¼ 106 within
the water5). The cavity-like structure can be clearly seen in Fig. 14c–e6. While the pressure in the cavitation
region continues to drop dramatically (with p ! 10:0 Pa in this case), its area is also shrinking smaller and
smaller because the surrounding pressure is higher. Eventually the area collapses to a singular point and a
shock is generated and propagates outward – a cylindrical shock inside the water column, as seen in
Fig. 14f.
In Fig. 14d to f, the incident shock of air glances over the water column and meets in the wake region. It is
interesting to note that the foot of the glancing shock continuingly transmits pressure waves into the water
column, clearly identifiable with the pressure contours. As the incoming air speed is high and the water speed
within the column is close to zero, a strong shear layer (contact discontinuity with a large difference in tan-
gential velocities in this case) is formed in the forward interface. As a result, the interface is getting thicker
and thicker, and part of the water appears to be dragged away by the air flow7.
As we have understood earlier in Section 3.3, the pressure wave in the air transmits into the water easily,
but much harder from water to air, due to the mismatch in acoustic impedance. In this case, the pressure
wave, once transmitted into the water column, is essentially confined, propagating back and forth, within
the water. On the other hand, the pressure field outside of the water column is basically not affected by
the waves within water column, with only a very small fraction of the pressure waves being transmitted
across the interface. Because the strength of the transmitted waves from the water is very weak, e.g. pressure
ratio of 1.013, compared to 10.333 of the incident shock wave, it is too faint to see in the pressure contours.
By increasing the image contrast, the shock wave is now visible, as marked by ‘‘A’’ in Fig. 15 in the
enlarged view, although the impinging shock is many times more clear. Nevertheless, it demonstrates that
our method is capable of resolving well even a faint shock wave, among complicated wave interactions and
much stronger waves.
In Fig. 16, we compare the calculated result with the experiment data obtained by Theofanous et al. [29]
where a Tributyl Phosphate (TBP) droplet is subject to a Mach 3.0 incident shock wave. Since the material
properties (density, viscosity and surface tension coefficient) of TBP is similar to that of water and we are una-

5
This in fact is not far from reality since water generally contains a slight amount of air.
6
It is noted that the stiffened gas EOS is not valid at low pressure; in fact, it allows a negative pressure of enormous value, which is
clearly not physical. In the real fluid, the water should evaporate when its pressure drops below the vapor pressure. The so-generated vapor
would fill the space and prevent the pressure from dropping without limit. Hence, an EOS having the ability to describe phase transition is
called for in this situation.
7
We believe that some of these are the artifacts attributable to numerical dissipations which act more overtly to the contact
discontinuity. However, these numerical dissipations covertly compensate for the physical diffusion omitted in our calculation, thus
making the qualitative comparison with the reality justifiable.
862 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873

Fig. 14. Time evolution of a shock–water column interaction problem. Incoming shock speed is M ¼ 3:0. Left: contours of pressure and
void fraction function; Right: contours of function ð1 þ a2l Þ logðj5qavg j þ 1Þ (Schlieren photo).

ware of other more appropriate EOS, we simply use the water properties for the purpose of a macroscale com-
parison. Since the Weber number for the TBP droplet is large ðWe ¼ 30;000Þ, we can safely neglect surface
tension. We also ignore the viscous effect which is deemed important in the thin region next to the interface.
Nevertheless, the comparison given in Fig. 16 yields an astonishing similarity between the computed and
experimental results. It is observed that the shape of droplet is not only changed by the compression force
imposed by the incoming air flows, but also equally importantly by the low pressure region created by the
reflected rarefaction wave in the rear. As a result, the rear interface is pulled inward. making the droplet rear
face flattened. The flattened interface is captured well by the method. This comparison implies that at least at
the initial stage the dominant phenomena are basically governed by nonlinear and inviscid mechanisms.
 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873 863

Fig. 14 (continued)

Next, we apply our method to a more complicated problem in which two water columns are considered.
The geometry of the initial setting is presented in Fig. 17, with an initial condition somewhat different from
that of previous case, with the incident shock moving at Mach 6. Hence, we set up the flow parameters for
the states before the shock as
p ¼ 1:0 105 Pa; ui ¼ vi ¼ 0 m=s; T i ¼ 346:98 K;
while those behind the shock as
p ¼ 4:18375 106 Pa; ui ¼ 1:818957 103 m=s; vi ¼ 0 m=s; T i ¼ 2755:48 K:
864 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873

Fig. 15. Enlarged view of numerical Schlieren, corresponding to Fig. 14d at t ¼ 4:0 ls, showing the very weak transmitted shock behind
the water column and the rarefied region in the rear.

Fig. 16. Schlieren photos for shock-droplet interaction problem. Left: TBP droplet subjected to M ¼ 3:0 flow, We ¼ 30; 000 (Theofanous
et al. [29]); Right: numerical result by water droplet with M ¼ 3:0 and We ¼ 1.

Incoming shock
Air

Water (0.0 mm, 0.0 mm)

Air
r1 =3 .2 mm
(7.0mm,-4.0mm)
Water

4.0 mm
r2 =2 .5 mm

Fig. 17. Illustration of the shock–water column interaction problem. Two water columns case.
 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873 865

The numerical Schlieren photos of the simulation result are shown in Fig. 18. For the first (left) water column,
we find in Fig. 18a that the shock wave travels faster outside of the water column than inside, because the
sound speed in water is about 4.4 times that in air while the shock Mach number is 6. At the interface facing
the incoming flow, the width of the fluid interface increases quickly as shown in Fig. 18b. A complicated flow
structure is developed at the interface when the air drags the water away from it, see Fig. 18c–f.
The result shows that after the incoming shock has passed the first water column, the lower part of the
shock hits on the second water column (Fig. 18b). The shock is reflected and forms a second bow shock in
front of the second column, which in turn transmits into the first water column and modulates the flow field
initiated previously by the first incident shock, resulting in asymmetrical profiles, as seen in Fig. 18c–f. The

Fig. 18. Time evolution of a shock interacting with two water columns. Contours of function logðj5qavg j þ 1Þ (numerical Schlieren photo)
is presented.
866 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873

lower and upper parts of the initial shock finally met, see Fig. 18c, in the wake of the first column, creating
asymmetric patterns due to the presence of the second column. A very complicated fluid structure is gen-
erated in the wake region after the shock passes and the complexities continue to grow. Vortices, shocks,
expansion and compression waves, contact lines, etc., all can be clearly seen in the numerical Schlieren
photo.
Interestingly, a ripple-like structure is seen, apparently bouncing back and forth between the primary bow
shock and the water column.
As in the single column case, a sequence of rarefaction waves make the pressure decrease in the rear region
of both columns. Because the incoming shock wave is stronger in this problem than the single column one, the

Fig. 19. Pressure and void fraction contours for the shock–bubble interaction problem.
 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873 867

reflected rarefaction wave can drive the pressure to a negative value, hence preventing the calculation of air
phase from going forward.

3.6. Shock and air-bubble interaction problem

The interaction of underwater shock wave with air bubble is also studied. The initial condition is the same
as that used by Hankin [30]. An air bubble with diameter of 6.0 mm is immersed in the water with its center at
the origin. The incoming shock wave initially located at x ¼ 4:0 mm is moving at Mach number of 1.509,
which is obtained by the following initial conditions. Before the shock wave,
p ¼ 1:01325 105 Pa; ui ¼ vi ¼ 0 m=s; T i ¼ 292:98 K;
and behind of the shock wave,
p ¼ 1:6 109 Pa; ui ¼ 661:81 m=s; vi ¼ 0 m=s; T i ¼ 595:14 K:
The simulation is performed on a mesh of about 600,000 cells. The time evolution of the simulation result is
presented in Fig. 19, showing the contours of pressure and void fraction ag. We observe that, after the water
shock wave hits the bubble, a strong rarefaction wave is reflected back into water from the fluid interface and a
relatively weak shock is transmitted into the air ðt ¼ 1:2–2:4 lsÞ. This is manifested by little variation inside

Fig. 20. Numerical Schlieren photo (logðj5qavg j þ 1Þ) of shock–bubble interaction problem.
868 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873

the air bubble, instead the air bubble begins deforming as soon as it is hit by the impinging shock. Again, be-
cause of the weak shock wave within the air bubble (no more than 0.1% of strength of the incident shock), it is
difficult to identify it in the plot of pressure contours, but more visible in the numerical Schlieren as evident in
Fig. 20a.
A water jet generated by the rarefaction wave is formed at the centerline and continuously pushes the bub-
ble into a crescent shape. The speed of the jet can be more than 2000 m/s. As a result, the bubble is finally
forced to break up into two separate bubbles and the water jet will collide with the still water behind the bub-
ble at about t ¼ 3:6l s (Fig. 19c), thus generating a shock wave which continues to expand radially, as seen in
Fig. 19d–f.
As time progresses, the separated air bubbles are compressed into a smaller volume due to the high pres-
sure imposed on it by the surrounding fluids. As the size of the bubble decreases to the local mesh size, the
solution becomes under-resolved. Then, additional mesh must be refined or mesh adaptation would be
necessary.
Details of the flow are vividly revealed in the enlarged view, shown in Fig. 20b. It consists of several clearly
identifiable shock waves: (1) the shock wave transmits across the rear interface from the bubble and enters the
water region behind bubble (as indicated by A), (2) the shock that is reflected from the rear and still remains
within the bubble (indicated by B), (3) the incident shock, which has passed over the bubble (indicated by C),
and (4) the shock waves generated by the high speed water jet (denoted by D and E). All have been well cap-
tured, clearly demonstrating the capability of our method, with no serious smearing found in the result.

3.7. 2D underwater explosion problem

As another example, an underwater explosion problem is computed. This problem is also difficult because it
entails a large difference in fluid properties, very strong shock/rarefaction wave, and two fluid interfaces – one
being the water free surface and another under the water. Referring to Fig. 21, the initial condition is set as
follows: the depth of the water is 1.5 m and a cylindrical core of detonation product is located 1.0 m below the
water surface, located at y ¼ 0. The diameter of the core is 0.2 m. The fluid states within the core are set as
ag ¼ 1  ; p ¼ 1:0 109 Pa; ui ¼ vi ¼ 0 m=s; T i ¼ 2000:0 K;
and the fluid states outside the core are

 if y < 0;
ag ¼
1   otherwise;
p ¼ 1:0 105 Pa; ui ¼ vi ¼ 0 m=s; T i ¼ 300:0 K;
where  ¼ 1:0 106 . The computation is performed on a mesh of 760 530 quadrilaterals. Initially, this
problem is set up as a 1D cylindrical shock tube problem, in which the high pressure gas of the explosion prod-
uct in the core is separated from the water by the fluid interface.
We show the time evolution of the simulation result in Fig. 21, in the form of numerical Schlieren photo
(contours of 5qavg ). It is noted that 5qavg is very large on the fluid interface (about O(107) in this case).
We need to modify its scale in order to make the flow structure appear more clearly; here, we use the function
ð1 þ 0:2a5g Þ logðj5qavg j þ 1Þ. We also show the contour of ag ¼ 0:5 with a white dash line to mark the location
of the fluid interface.
Referring to Fig. 21a, a shock wave is transmitted from the explosion product into the liquid water sur-
rounding the core and propagates radially outward from the core. At the same time, a rarefaction wave is
propagating from the core boundary into the center of the core, which can be observed as a pressure drop
in the vicinity of the core region by a dark strip inside the core in Fig. 21b. When the shock hits the solid wall
at the bottom, it is reflected, generating a high pressure region behind the shock, as seen also in Fig. 21b.
In Fig. 21c, the shock wave hits the upper water surface and pushes it upward. While a small fraction of the
shock wave is transmitted into the air, a strong rarefaction wave is reflected from the water surface, propagat-
ing downward back to the water and generating a low pressure region near the water surface. Similarly, when
the reflected shock from the bottom wall hits the explosion core, a small part of the shock transmits into the
 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873 869

core, which can be clearly seen in the Schlieren, and a rarefaction wave is bounced back from the core inter-
face, producing a low pressure region beneath the core. Since the rarefaction wave is strong, the pressure will
drop much more after it has hit the bottom wall and reflected as shown in Fig. 21d.
In Fig. 21e and f, the wave interactions has become progressively more complicated both inside and outside
the core. It is noticed that the core has gradually expanded, but essentially retained its shape, because the flow
is driven by high pressure in the core, while the shock and air-bubble interactions given in Section 3.6 are
unsymmetrically initiated from outside of the air bubble. Again, a faint shock wave transmitted above the
water surface is distinguishable and the surface is now deformed noticeably in Fig. 21e and f.

Fig. 21. Time evolution of the underwater explosion problem. Left: Numerical Schlieren photos for underwater explosion problem. The
fluid interfaces (ag ¼ 0:5) are outlined by white dashed lines.
870 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873

4. Conclusion

In this paper, we have proposed a new compressible multiphase multifluid method that makes use of the
stratified flow concept. As a result, it is natural to conclude that the numerical fluxes are comprised of two
types: interaction of the same fluids, f g–g and f l–l , and interactions of different fluids, f g–l and f l–g . This formu-
lation lends itself to satisfy the pressure non-disturbing condition, which is found to be advantageous for cap-
turing discontinuities, such as the fluid interface and shock wave. We extended the recent AUSM+-up method
to solve the multifluid equations, where the liquid phase is described by the stiffened gas model. The new
method has been applied to a wide variety of one and two dimensional problems. The computed results indi-
cate that our method is accurate and robust, notably for problems with shock-interface interactions, yielding
amazingly fine details about complicated interaction phenomena inside the air or water phases, suggesting that
the method does not overly smear out even the details of small strength. For a more realistic description of the
flow under low pressure, a more accurate description for the liquid phase is needed, together with the capa-
bility of handling phase transition.

Acknowledgements

The authors thank the reviewers for their valuable suggestion and comments that have contributed to the
improvement of this paper. They are grateful to NASA Glenn Research Center, Cleveland, Ohio, for support-
ing the research reported herein.

Appendix A. The stiffened gas model

In this paper, the gas and liquid fluids are governed by the perfect gas model and stiffened gas model,
respectively. The stiffened gas model proposed by Harlow and Amsden [17] is expressed by
cl  1
pl ¼ ql C pl T l  p1 ; ðA:1Þ
cl
Cp p
el ¼ l T l þ 1 ; ðA:2Þ
cl ql
where the parameters ðcl ; C pl ; p1 Þ are constants, depending to the material we use. Hence,the speed of sound is
simply
 1=2
cl ðpl þ p1 Þ
al ¼ : ðA:3Þ
ql
It is clear that the perfect gas model becomes a special case of the stiffened gas model, by taking p1 ¼ 0 and a
different set of values for cl and C pl . Here, the parameters used for water are determined to meet the following
criteria. When T l ¼ 293:15 K and p ¼ 1:0132 105 Pa, we set the following properties based on the following
data:
ql ¼ 998:23kg=m3 ;

oel  Cal J
C vl ¼ ¼ 1000 ¼ 4190:0 ; ðA:4Þ
oT l v kg K kg K
al ¼ 1482 m=s:
Then, we have
J
cl ¼ 1:932; C pl ¼ 8095:08 ; p1 ¼ 1:1645 109 Pa:
kg K
All the simulations presented in this paper will be based on these parameters. It is remarked that this set of
parameters are different from those used by others. In principle, the compressibility factor (ratio of specific
heat) cl for water should be close to 1.0. The cl (=1.932) we use is much closer to 1.0 then 4.4 used by Saurel
 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873 871

and Abgrall [31] and 2.8 used by Paillère et al. [26]. The parameters we choose also give correct density, sound
speed and heat capacity in standard atmosphere condition.
The stiffened gas model can provide a reasonable approximation to a fluid in high pressure. Noting that the
speed of sound in liquid is substantially higher than that in gas and this fact is accommodated in the equation
of state (EOS) by an enormous value of p1. As a result of having a large speed of sound, liquid flows usually
fall in the low Mach number flow regime and the coupling between the pressure and density fields becomes
weak. That is, the variation in density is insignificant even when a very large pressure gradient is imposed
on the flow, making the fluid essentially incompressible. On the other hand, small changes in density field
can result in huge changes in pressure, making numerical solutions prone to oscillations, due to the presence
of a large p1 in the EOS. In mathematical terms, the gas–liquid system is said to be stiff.

Appendix B. Exact Riemann solution for the stiffened gas

In what follows we present the exact Riemann solution for the fluid described by the stiffened-gas equation
of state. The ideal-gas Riemann solution then becomes a special subset by letting p1 ¼ 0 and using the cor-
responding value for c. The general procedure for constructing the exact Riemann solution is standard. e.g.,
available in the book by Toro [32]. For completeness, we give the specific formulas for the stiffened gas EOS as
described in Appendix A.
We consider a one dimensional Riemann problem, as shown in Fig. 22. The computation domain is divided
into four regions by the waves propagating across the domain. The wave associated with the eigenvalues
(u þ a) and (u  a) can be either a shock wave or a rarefaction wave, and the wave associated with the eigen-
value u is a contact discontinuity. Then we can assign the flow properties in the four regions respectively as
QL ; QH H
L ; QR and QR. The QL and QR are the initial states of the flow on the left and right respectively. The
goal is to determine the middle states QH H
L ; QR , which are separated by the contact discontinuity, across which
we know the following relations.
pH H H
L ¼ pR ¼ p ; ðB:1Þ
uH
L ¼ uH
R
H
¼u : ðB:2Þ
These ‘‘w’’ states will be related to the ‘‘L’’ and ‘‘R’’ states through the ‘‘u  a’’ and ‘‘u + a’’ waves respec-
tively. Across the ‘‘u – a’’ wave, we can write the velocity uw as a function of QL and pw
uH ¼ uL  fL ðpH ; QL Þ; ðB:3Þ
with
8 h i1=2
>
> ðp H
 p Þ 2
if pH P pL ;
< L qL ððcL þ1Þ½pH þðp1 ÞL þðcL 1Þ½pL þðp1 ÞL Þ
fL ðpH ; QL Þ ¼   ðB:4Þ
>
> 2aL
cL 1
: cL 1
ðn L Þ 2cL
 1 otherwise:

And across the ‘‘u þ a’’ wave, we have a similar relation among uw, QR and pw, expressed as
uH ¼ uR þ fR ðpH ; QR Þ; ðB:5Þ

t
u-a u
u+a

Fig. 22. Illustration of the Riemann problem.

872 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873

with
8 h i1=2
>
>
< ðpH  pR Þ qR ððcR þ1Þ½pH þðp1 ÞR2þðcR 1Þ½pR þðp1 ÞR Þ if pH P pR ;
fR ðpH ; QR Þ ¼   ðB:6Þ
>
> 2aR
cR 1
: cR 1 ðnR Þ 2cR
1 otherwise;

and
pH þ ðp1 Þk
nk ¼ ; ðB:7Þ
pk þ ðp1 Þk
where the subscript k can be either L or R. From Eqs. (B.3) and (B.5), we get a nonlinear algebraic equation
for pw as a function of known states QL and QR,
f ðpH ; QL ; QR Þ ¼ fL ðpH ; QL Þ þ fR ðpH ; QR Þ þ uR  uL ¼ 0: ðB:8Þ
We use the Newton iteration method to solve for pw. Then we obtain uw by substituting pw into Eqs. (B.3) or
(B.5). Finally, the densities qH H
L and qR can be calculated from the following equation.
8 
< q hL nL þ1 if pH P pL ;
L hL þnL
qH
L ¼ ðB:9Þ
: 1;
qL ðnL ÞcL otherwise;
8 
< q hR nR þ1 if pH P pR ;
R hR þnR
qH
R ¼ ðB:10Þ
: 1;
qR ðnR ÞcR otherwise;
with
ck þ 1
hk ¼ : ðB:11Þ
ck  1
Note that the above formulas are for the subsonic case; the ones for supersonic are obtained by simple
upwinding, e.g., pH ¼ pL and uH ¼ uL if uL  aL > 0 and uR  aR > 0.

References

[1] Y.A. Buyevich, Statistical hydrodynamics for dispersed system, physical background and general equations, J. Fluid Mech. 49 (1971)
489–507.
[2] M. Ishii, Thermo-fluid dynamic theory of two-phase flow, Eyrolles, Paris, 1975.
[3] H.B. Stewart, B. Wendroff, Two-phase flow: models and methods, J. Comput. Phys. 56 (1984) 363–409.
[4] I. Toumi, A. Kumbaro, H. Paillère, Approximate Riemann solvers and flux vector splitting schemes for two-phase flow, Lecture series
1999–03, von Karman Institute for Fluid Dynamics, 1999.
[5] J. Stuhmiller, The influence of interfacial pressure forces on the character of two-phase flow model equations, Int. J. Multiphase Flow
3 (1977) 551–560.
[6] D. Drew, L. Cheng, J.R.T. Lahey, The analysis of virtual mass effects in two-phase flow, Int. J. Multiphase Flow 5 (1979) 233–242.
[7] R. Saurel, R. Abgrall, A multiphase Godunov method for compressible multifluid and multiphase flows, J. Comput. Phys. 150 (1999)
425–467.
[8] C.-H. Chang, M.-S. Liou, A new approach to the simulation of compressible multifluid flows with AUSM+ scheme, AIAA paper 03–
4107 (2003).
[9] C.-H. Chang, M.-S. Liou, Simulation of multifluid multiphase flows with AUSM+-up scheme, in: Third Internation Conference of
Computational Fluid Dynamics, Toronto, Canada, 2004.
[10] R. Abgrall, R. Saurel, Discrete equations for physical and numerical compressible multiphase mixtures, J. Comput. Phys. 186 (2003)
361–396.
[11] R. Saurel, S. Gavrilyuk, F. Renaud, A multiphase model with internal degrees of freedom: application to shock–bubble interaction, J.
Fluid Mech. 495 (2003) 283–321.
[12] O. LeMétayer, J. Massoni, R. Saurel, Modeling evaporation fronts with reactive Riemann solvers, J. Comput. Phys. 205 (2005) 567–
610.
[13] B. Van Leer, Flux vector splitting for the Euler equations, in: Proceedings of the 8th International Conference on Numerical Methods
in Fluid Dynamics, Springer Verlag, Berlin, 1982.
 C.-H. Chang, M.-S. Liou / Journal of Computational Physics 225 (2007) 840–873 873

[14] M.-S. Liou, Ten years in the making – AUSM-family, AIAA paper 2001–2521, in: 15th Computational Fluid Dynamics Conference
Proceedings, 2001.
[15] M.-S. Liou, A sequel to AUSM: AUSM+, J. Comput. Phys. 129 (1996) 364–382.
[16] M.-S. Liou, J.R. Edwards, Numerical speed of sound and its application to schemes for all speeds, AIAA paper 99–3268 (1999).
[17] F. Harlow, A. Amsden, Fluid dynamics, Technical Report LA-4700, Las Alamos National Laboratory, 1971.
[18] M.-S. Liou, A further development of the AUSM+ scheme toward robust accurate solution for all speed, AIAA paper 2003–4116, in:
16th Computational Fluid Dynamics Conference Proceedings, 2003.
[19] M.-S. Liou, A sequel to AUSM, part II: AUSM+-up for all speeds, J. Comput. Phys. 214 (2006) 137–170.
[20] J.R. Edwards, M.-S. Liou, Low-diffusion flux-splitting methods for flows at all speeds, AIAA Journal 36 (9) (1998) 1610–1617.
[21] G. Wallis, One-dimensional two-phase flow, McGraw-Hill, New York, 1964.
[22] W. Hancox, R. Ferch, W. Liu, R. Nieman, One-dimensional models for transient gas–liquid flows inducts, Int. J. Multiphase Flow 6
(1980) 25–40.
[23] I. Toumi, An upwind numerical method for two-fluid two-phase flow models, Nuclear Sci. Eng. 123 (1996) 147–168.
[24] S.K. Godunov, Numerical solution of multidimensional gas dynamics problems, Nauka, Moscow, 1976.
[25] M.-S. Liou, L. Nguyen, C.-H. Chang, S. Sushchikh, R. Nourgaliev, T. Theofanous, Hyperbolicity, discontinuities, and numerics of
two-fluid models, Technical report, Springer, in: 4th International Conference on Computational Fluid Dynamics, 2006.
[26] H. Paillère, C. Corre, J.R.G. Cascales, On the extension of the AUSM+ scheme to compressible two-fluid models, Comput. Fluids 32
(2003) 891–916.
[27] S.R. Chakravarthy, S. Osher, Computing with high-resolution upwind schemes for hyperbolic equation, Lect. Appl. Math. 22 (1985)
57–86.
[28] V.H. Ransom, Numerical benchmark tests, in: G.F. Hewitt, J.M. Delhaye, N. Zuber (Eds.), Multiphase Science and Technology, vol.
3, Hemishpere Publishing Corporation, 1987.
[29] T. Theofanous, G. Li, T. Dinh, Aerobreakup in rarefied supersonic gas flows, J. Fluids Eng. 126 (2004) 516–527.
[30] R.K.S. Hankin, The Euler equations for multiphase compressible flow in conservation form: simulation of shock–bubble interactions,
J. Comput. Phys. 172 (2) (2001) 808–826.
[31] R. Saurel, R. Abgrall, A simple method for compressible multifluid flows, SIAM J. Sci. Comput. 21 (3) (1999) 1115–1145.
[32] E.F. Toro, Riemann solvers and numerical methods for fluid dynamics: a practical introduction, Springer-Verlag Telos, 1997.
 Update
Journal of Computational Physics
Volume 227, Issue 10, 1 May 2008, Page 5360

DOI: https://doi.org/10.1016/j.jcp.2008.01.041
 Available online at www.sciencedirect.com

Journal of Computational Physics 227 (2008) 5360

www.elsevier.com/locate/jcp

Erratum

Erratum to ‘‘A robust and accurate approach to

computing compressible multiphase flow: Stratified
flow model and AUSM+-up scheme”
[J. Comput. Phys. 225 (2007) 840–873]
Chih-Hao Chang a,*, Meng-Sing Liou b
a
Center for Risk Studies and Safety, University of California, Santa Barbara, 6740 Cortona Dr., Goleta, CA 93117, USA
b
Aeropropulsion Division, NASA Glenn Research Center, Cleveland, OH 44135, USA

Available online 6 February 2008

In the equation after (A.4) on page 870, we incorrectly copied the values of parameters; they should be
given as:
J
cl ¼ 1:9276; C pl ¼ 8076:6 ; p1 ¼ 1:1373  109 Pa:
kg K
These are the values used in all the results presented in the referenced paper.

DOI of original article: 10.1016/j.jcp.2007.01.007.

*
Corresponding author. Tel.: +1 805 252 3425.
E-mail addresses: [email protected] (C.-H. Chang), [email protected] (M.-S. Liou).

0021-9991/$ - see front matter Ó 2008 Elsevier Inc. All rights reserved.
doi:10.1016/j.jcp.2008.01.041