Advances in Water Resources Vol. 22, No. 5, pp.

461–478, 1999
䉷 1999 Elsevier Science Ltd
Printed in Great Britain. All rights reserved
PII: S 0 3 0 9 - 1 7 0 8 ( 9 8 ) 0 0 0 2 2 - 0 0309-1708/99/$ - see front matter

Practical implementation of the fractional flow

approach to multi-phase flow simulation
Philip Binning a,* & Michael A. Celia b
a
Department of Civil, Surveying and Environmental Engineering, The University of Newcastle, Callaghan, N.S.W. 2308, Australia
b
Department of Civil Engineering and Operations Research, Princeton University, Princeton, NJ 08544, USA

(Received 13 August 1997; revised 24 June 1998; accepted 30 June 1998)

Fractional flow formulations of the multi-phase flow equations exhibit several

attractive attributes for numerical simulations. The governing equations are a
saturation equation having an advection diffusion form, for which characteristic
methods are suited, and a global pressure equation whose form is elliptic. The
fractional flow approach to the governing equations is compared with other
approaches and the implication of equation form for numerical methods discussed.
The fractional flow equations are solved with a modified method of characteristics for
the saturation equation and a finite element method for the pressure equation. An
iterative algorithm for determination of the general boundary conditions is
implemented. Comparisons are made with a numerical method based on the two-
pressure formulation of the governing equations. While the fractional flow approach is
attractive for model problems, the performance of numerical methods based on these
equations is relatively poor when the method is applied to general boundary
conditions. We expect similar difficulties with the fractional flow approach for more
general problems involving heterogenous material properties and multiple spatial
dimensions. 䉷 1999 Elsevier Science Ltd. All rights reserved.

Keywords: multi-phase fluid flow, numerical methods, fractional flow, MMOC.

1 INTRODUCTION concomitant introduction of ancillary functions such as the

fractional flow function. The latter approach derives almost
Numerical simulation of multi-phase flow in complex exclusively from the petroleum literature. In 1973, Morel-
porous media remains one of the outstanding difficulties Seytoux1 wrote a classic paper in which he drew together
in the field of computational hydrology. The reasons for much of the previous research and explored unifying themes
this difficulty include the highly nonlinear nature of the between the disciplines. He showed that the flow of air and
coupled partial differential equations that govern the water in unsaturated soils can be viewed as a multi-phase
system, and the lack of reliable constitutive data for these system, and showed how the experience of petroleum
problems. This difficulty has led many researchers to engineering could aid in understanding of multi-phase
explore alternative forms of the governing equations, and flow problems in hydrology. Since then many developments
to seek specialized numerical algorithms that can improve have occurred, particularly in the area of numerical methods
the computational performance of a simulator. for the solution of the multi-phase flow equations.
Historically, there have been two main approaches to Numerical methods are very sensitive to the choice of
modeling multi-phase flow, arising in the disciplines of form of the governing equation. Morel-Seytoux1 showed
hydrology and petroleum engineering. The first is based that there are several ways to write the governing equations
on individual balance equations for each of the fluids, of fluid flow, and that each method offered its own insights
while the second involves manipulation and combination into the solution. More recently, Ewing2 has also examined
of those balance equations into modified forms, with recent developments in the choice of equation form for
multi-phase flow. In the light of the new and continuing
*Corresponding author. E-mail: [email protected] developments in numerical methods for the solution of the
461
462 P. B. Binning, M. A. Celia

multi-phase flow equations, it is worthwhile revisiting the dimensions with variability in the intrinsic permeability of
question of the form of the governing equations and explor- the medium. Hansen et al.23 considered systematic treat-
ing the implications of this equation form for a numerical ment of the gravity terms in the fractional flow formulation.
method based on it. The problem of boundary condition implementation has
Two approaches to writing the governing equations will also been considered recently by Chen et al.24.
be examined here: the two-pressure approach; and the frac- The pressure equation has been solved by a variety of
tional flow approach. The two-pressure approach to the methods. In one dimension, for incompressible flow the
governing equations has been widely used in the hydrologic total velocity is a constant in space that depends only on
literature. In this approach, the governing equations are the time varying boundary conditions, and so the pressure
written in terms of the pressures in each of the two phases equation may be solved analytically1. In higher dimensions,
through a straightforward substitution of Darcy’s equation numerical methods must be employed. Ewing25 reviews
into the mass balance equations for each phase. This numerical techniques for solution of the pressure equation
approach has been adopted by a number of authors includ- and demonstrates the importance of accurate determination
ing: Pinder and Abriola3 who employed a finite difference of velocities. Variations in material properties, particularly
approximation of the governing equations to describe non- the permeability, can cause sharp changes in pressures.
aqueous phase liquid flow in the saturated zone; Sleep and Large errors can result if these pressures are differentiated
Sykes4, and Kaluarachchi and Parker5, who used a finite- to obtain the necessary velocities. Ewing and
element solver with a Newton–Raphson scheme to linearize Heinemann26,27 developed a mixed method for solution of
the equations; Celia and Binning6–8 who employed a finite the pressure equation, with the aim of accurate determi-
element discretization with fully implicit time stepping and nation of velocities. They showed that mixed methods
Picard iteration to solve the air and water flow equations; were far superior to conventional approaches when applied
and Schrefler and Xiaoyong9, who solved the consolidation to standard five-spot problems. Mixed methods are the
problem in the unsaturated zone with a finite element state-of-the-art at the current time.
discretization of the governing equations. Other approaches to the numerical solution of the frac-
The fractional flow approach originated in the petroleum tional flow equations include the work of Guarnaccia and
engineering literature, and employs the saturation of one of Pinder,28 who employed the fractional flow formulation
the phases and a pressure as the independent variables. The with the sequential solution method of Spillette et al.29 to
fractional flow approach treats the multi-phase flow decouple the equations and a collocation method to dis-
problem as a total fluid flow of a single mixed fluid, and cretize the equations to solve the problem of NAPL
then describes the individual phases as fractions of the total migration in the water and gas phases.
flow. This approach leads to two equations: the pressure While the fractional flow and two-pressure approaches
equation; and the saturation equation. The pressure equation have been studied by various researchers, other methods
is an elliptic equation that is solved for the pressures and the have also been developed to solve multi-phase flow
total flux. The saturation equation is written in advection– equations. One of the more important variants on these
diffusion form with a hyperbolic characteristic that methods is the pressure–saturation approach. Since the
describes the speed of an infiltrating front. The advective saturation is a function of the capillary pressure or differ-
term is nonlinear and usually leads to shock formations. Its ence between phase pressures, it is possible to reformulate
general behavior may be exploited to design a Lagrangian the governing equations in terms of a saturation and one of
numerical algorithm that allows for relatively large time steps the phase pressures. The second phase pressure can then be
by projecting the solution forward along the characteristics. removed from the equations by expressing it in terms of the
In the absence of capillarity and gravity, the saturation saturation and the pressure in the other phase. The attraction
equation may be solved analytically; this approach dates to of this approach is that it is very well suited to problems
Buckley and Leverett10. Interest in the saturation equation where there may be phase disappearance. If the saturation in
and its analytic solution continues today. For example, one of the phases is zero, the pressure in that phase is poorly
Sander et al.11 have found an analytic solution for a defined. A standard two-pressure formulation will not be
restricted class of functional forms. Numerical solutions able to handle this problem. The pressure saturation
have been considered more recently. Morel-Seytoux and approach has been employed by Faust,30 Kueper and
Billica12,13, and Wangen14 used finite difference methods, Frind,31 Moridis and Reddell,32 and Pruess.33 Forsyth34
and Chen et al.15 employed finite elements and mixed and Unger et al.35 have developed compositional simulators
methods to solve the fractional flow equation for one- to solve two- and three-phase multi-component systems,
dimensional systems with incompressible fluids, while a based on the P–S formulation and finite-volume approxima-
general characteristic-based approach was presented by tions and nonlinear flux limiters. Abriola and Rathfelder36
Douglas16 and Espedal and Ewing17. The computational implement and review both the two-pressure and the
work of Dahle et al.18,19, Celia and Binning20, and Langlo pressure–saturation formulations, examining the mass bal-
and Espedal21 demonstrated the computational benefits of ance properties of each formulation. They show that, if the
the characteristic-based method, with the paper by Langlo coefficients and initial conditions are properly treated,
and Espedal22 presenting simulation results in two accurate solutions can be obtained with both formulations.
 Multi-phase flow simulation 463

Since the governing equations for multi-phase flow are question of general boundary conditions, consider the impli-
highly nonlinear, numerical techniques must in general be cations of general material heterogeneities, speculate on the
employed for their solution. There are many choices of potential difficulties with multiple infiltration and drainage
numerical method, with temporal discretization being a fronts, and consider some general questions related to fluid
topic of great interest in the literature. In hydrologic appli- compressibility and gravity terms.
cations the predominant approach has been to employ fully
implicit schemes. Haverkamp et al.37 compared temporal
discretizations of Richards’ equation, and showed that 2 GOVERNING EQUATIONS
because of the stability restrictions on explicit solvers,
implicit solvers provide solutions five to ten times faster In this presentation, we consider balance and constitutive
than explicit solvers. The better performance of the implicit equations applicable to each fluid phase; we will not con-
solvers occurs, despite the need for iterations to deal with sider modeling of individual components or processes of
the nonlinearities. Multi-phase flow simulators in hydrology interphase transfer. For each fluid phase a, the mass balance
have similar behavior to that observed by Haverkamp, and equation may be written as
so fully-implicit time stepping is the dominant approach in ⳵(ra JSa )
hydrology and has been adopted by many authors36,38. þ ⵜ·(ra qa ) ¼ 0 (1)
⳵t
The petroleum engineering literature contrasts with the
where r a is the density of fluid a, J is porosity, S a is fluid
hydrology literature with the predominant approach being
saturation, and q a is the volumetric flux vector for phase a.
explicit or semi-implicit. The most common approach
For a two-fluid-phase system, eqn (1) is written for each
employed38 is the IMPES scheme (implicit pressure,
fluid phase, and is augmented by constitutive relationships
explicit saturation). This scheme has been favored because
that relate saturations to capillary pressure (eqn (2)),
it decouples the pressure and saturation equations, thus
specify a closure condition on saturations (eqn (3)), and
reducing the computational effort required for their solution.
relate volumetric fluxes to pressure gradients via a multi-
The IMPES scheme has been used by Sleep and Sykes39
phase version of the Darcy equation (eqn (4)) which
who used it to simulate the behavior of an air–water system.
includes relative permeabilities (eqn (5)),
Sleep and Sykes compared the efficiency of an IMPES
scheme with a fully implicit scheme. They showed that Sa ¼ Sa (pc ) ¼ Sa (pnw ¹ pw ) (2)
the IMPES was more restricted in time step size than the
implicit scheme. The difference in time step size made up Sw þ Snw ¼ 1 (3)
for any speed the IMPES gained by decoupling the
kkra
equations. However, the IMPES scheme requires less qa ¼ ¹ ·(ⵜpa ¹ ra g) (4)
memory, which may be a consideration for larger problems. ma
A new development in hydrology is the use of an adaptive
kra ¼ kra (Sa ) (5)
implicit scheme, which was first employed in the petroleum
literature. A fully implicit formulation is used in regions of where p c ⬅ p w ¹ p n w is the capillary pressure, p a denotes
sharp changes of independent variables and the explicit the individual phase pressures with w and nw denoting the
scheme used elsewhere. An example is the work of wetting and non-wetting phases respectively, k is the intrin-
Reeves and Abriola.40 sic permeability tensor, m a is the viscosity in phase a, g is
In this paper, the equations that describe two-phase fluid the gravitational vector (the vertical coordinate is oriented
flow will be reviewed. The equations will be presented in a positive down), and k ra is the relative permeability to phase
two-pressure, or standard mass balance, form that is com- a which is a nonlinear function of saturation.
monly used by hydrologists, as well as a modified form that If fluid or matrix compressibility is considered, then
we will refer to as the fractional flow or global pressure– appropriate compressibility coefficients also need to be
saturation form that has been developed by petroleum defined. For the purposes of this paper all fluids, as well
engineers. Computations will be used to demonstrate the as the solid matrix, will be considered to be incompressible.
differences in performance between numerical approxima- Given this set of equations, boundary and initial condi-
tions of the different equation forms, and the implications tions must be supplied to complete the mathematical
for the future multi-phase modeling efforts discussed. The description. These are usually given as known pressures,
paper will not discuss the various choices of temporal dis- saturations or fluxes in each of the fluid phases. Many dif-
cretization, as these are examined in other works. A fully ferent combinations of these boundary conditions occur in
implicit time discretization will be adopted for all numerical practical problems. An important criteria for acceptance of a
schemes, and the focus of the paper will be on the form of numerical method is that it must be able to solve the govern-
the governing equation and the implications of that form on ing equations for the wide variety of possible boundary
the performance of the numerical method. conditions.
The paper will also consider questions related to practical The governing equations, eqns (1)–(5) are a set of
implementation of the fractional flow approach to multi- coupled, nonlinear partial differential equations. The basic
phase flow simulation. In particular, we will revisit the equations can be mathematically manipulated into several
464 P. B. Binning, M. A. Celia

alternate forms with various choices of primary dependent where P c(S c) ¼ 0, and f w is the fractional flow function,
variables. The choice of equation form and primary solution defined by
variables have considerable implications for the numerical krw =mw
method used to solve the equations. Two possible formula- f w (Sw ) ⬅ (10)
(krw =mw ) þ (krnw =mnw )
tions will be considered here and the benefits of each
discussed. In the two-pressure formulation, the pressures Given these definitions, the total velocity may be related to
in each phase are chosen as independent variables and the the total pressure.41 This expression for the total velocity
governing equations are parabolic in nature. The other may then be substituted into eqn (8) to obtain the nonlinear
formulation employed is the fractional flow approach elliptic partial differential equation that governs the
where a saturation and a ‘global pressure’ are chosen as evolution of the total pressure P,
independent variables. In this case, the governing equation ⵜ·L·ⵜP ¹ ⵜ·(LG) ¼ 0 (11)
for saturation is parabolic (hyperbolic in the case of zero
capillarity) and the equation for pressure elliptic. where
 
krw krnw
2.1 Two-pressure approach L(Sw ) ⬅ k þ
mw mnw

The conventional approach to solving the governing is the total mobility coefficient, and
equations, eqns (1)–(5) in hydrology involves a straight- G(Sw ) ⬅ (fw rw þ (1 ¹ fw )rnw )g
forward substitution of the Darcy equation (eqn (4)) into
is a nonlinear function of saturation that accounts for
the mass balance equation (eqn (1)) to eliminate the fluxes
gravitational effects. Eqn (11) forms the first of the two
from the equations. The primary unknowns are chosen to be
governing equations in the fractional flow approach.
the two phase pressures, p w and p n w. The saturation in the
The second equation used in the fractional-flow equation
non-wetting phase is eliminated using eqn (3).
  set is derived from manipulation of eqn (1) with a ¼ w, such
⳵Sw kkrw that the primary dependent variable is wetting-phase satura-
J ¹ ⵜ· ·(ⵜpw ¹ rw g) ¼ 0 (6)
⳵t mw tion S w. Because eqn (1) contains an individual phase flux
  that is dependent on an individual phase pressure, an
⳵Sw kkrnw equation is found relating the individual phase flux, q w to
¹J ¹ ⵜ· ·(ⵜpnw ¹ rnw g) ¼ 0 (7)
⳵t mnw the total flux
The equations are coupled and nonlinear, and are closed by qw ¼ F w (Sw , P) ¼ f w (Sw )
specification of the pressure–saturation relation (eqn (2))   
kkrnw dpc
and relative permeability relation (eqn (5)). ⫻ qT þ ⵜS þ (rw ¹ rnw )g ð12Þ
mnw dSw w
2.2 Fractional flow approach In eqn (12), F w is the fractional flow function incorporating
both gravity and capillarity. In one dimension, substitution
The fractional flow approach is again based on a set of two of eqn (12) into eqn (1) with a ¼ w, and assuming incompres-
nonlinear equations, each of which has its origin in eqns sible flow, results in the following saturation equation.1,41
(1)–(5), but with a different choice of dependent variables.  
⳵Sw ⳵Sw ⳵ ⳵Sw
Definition of these variables requires extensive manipula- J þ Fw ⬘(Sw ) ¹ D(Sw ) ¼0 (13)
tion and rearrangement of the two-phase equations. For ⳵t ⳵z ⳵z ⳵z
simplicity of presentation, consider the case of constant where
fluid densities and constant porosity. Then the individual  
kkrnw
phase mass-balance equations (eqn (1)) sum to form Fw (Sw ) ¼ fw qT þ (r ¹ rnw )g (14)
mnw w
ⵜ·(qw þ qnw ) ¼ ⵜ·qT ¼ 0 (8)
and
where q T, called the total flux, is the sum of the phase- kkrnw dpc
volumetric fluxes. In one dimension this equation has the D(Sw ) ¼ ¹ fw (15)
mnw dSw
particularly simple solution that the total flux is a constant
in space determined by the boundary conditions. A new The coefficient D(S w) is the capillary diffusion term, while
independent variable is now introduced, called the global F w is the fractional flow, containing f w as well as the gravity
pressure, with the aim of expressing the total flux solely in drive, and is defined so that q w ¼ F w(S w,P) in the absence of
terms of the gradient of the global pressure and not in terms capillary forces. The definition F w degenerates into q Tf w in
of the individual phase pressures. Following Chavent and the absence of gravity. Note that the definition of F w in eqn
Jaffré,41 a global pressure is defined by (14) differs slightly from that of Chavent and Jaffré41 and
ZSw   Morel Seytoux,1 whose definition is divided by the total
1 1 dpc
P ¼ (pw þ pa ) ¹ fw ¹ dS (9) flux q T. The equation form presented here follows the
2 Sc 2 dSw w approach of Hansen et al.,23 who included the multiplicand
 Multi-phase flow simulation 465

q T in F w, in order that it remained well defined in the case

of zero total flux.
When eqns (11) and (13) are supplemented with appro-
priate boundary conditions, they may be solved for the
unknowns P and S w. Implementation of boundary con-
ditions is discussed below. The equations are coupled
through the dependence of the nonlinear coefficients L
and G in the pressure equation on the saturation S w and
through the dependence of the total velocity q T appearing
in the saturation equation on the global pressure P. Note that
the definition of global pressure means that the pressure and
saturation equations are only loosely coupled. Other defini-
tions of pressure and saturation will increase the coupling
between the equations. The other advantage of the choice of
global pressure and saturation as independent variables is
that these variables are smoother than the individual phase
pressures, and so the performance of numerical methods
based on these equations should in principle be better than
solvers employing phase pressures as the independent
variables.2
In an oil–water system, with typical relative permeability
functions given by18,42
krw ¼ S3w (16)

krnw ¼ (1 ¹ Sw )3 (17)
the fractional flow function has characteristic S-shape, in
the absence of gravity and assuming equal fluid viscosities.
Fig. 1 shows both the relative permeability functions and
the fractional flow function f. This S-shaped fractional flow
function leads to solutions that combine a shock front with
a rarefaction wave when gravity is absent; this is the
classical Buckley–Leverett solution. When capillary diffu-
sion is considered, the front exhibits some spreading, but
this is often relatively minor. Fig. 2 shows the Buckley–
Leverett solution, together with the full solution of the
saturation equation with permeabilities given by eqns (16)
and (17) and D(S w) ¼ 0.04␧S w(1 ¹ S w) with ␧ ¼ 0.01 (See
Dahle18 for details of these functional forms). In either
case, the frontal speed may be estimated from the classic
theory of hyperbolic equations.
For an air–water system the fractional flow and capillary
diffusion functions are quite different from the oil–water
case, because of the much higher viscosity difference
between fluids. For air and water m w ¼ 0.001 N s m ¹2 and
m a ¼ 1.83 ⫻ 10 ¹5 N s m ¹2 at 20⬚C, leading to a mobility
ratio of 54.6, as opposed to the oil–water system above,
which was modeled with a mobility ratio of 1. As pointed
out by Morel Seytoux,1 this viscosity contrast causes both
curves to shift significantly toward the higher water
saturations, greatly diminishing the rarefaction wave and
expanding the shock range. As an example, the relative
permeability functional forms determined experimentally
by Touma and Vauclin43 are used to derive the fractional Fig. 1. Functional forms of Young42 and Dahle et al.18 for an oil–
water system. (a) Relative permeability curves. (b) Fractional flow
flow and capillary diffusion functions. The appropriate function. The slope of the Welge49 tangent line drawn gives the
functional forms, determined experimentally by speed of shock fronts formed in the solution. (c) Approximate
Touma and Vauclin, are as follows: the capillary capillary diffusion function with ␧ ¼ 0.01.
466 P. B. Binning, M. A. Celia

Fig. 2. Solution of the saturation eqn 13) using the oil–water functional relationships shown in Fig. 1. The MMOC solution of the full
saturation equation is shown as a line with nodal positions marked. The Buckley–Leverett solution is shown as a solid line for comparison.
The Buckley–Leverett front speed is 4/3.

pressure–saturation relationship, obtained by fitting a van second and third solve the fractional flow equations
Genuchten44 form to the data, is (eqns (11) and (13)) using a finite difference and
Sws ¹ Swr characteristics formulation. All numerical results are
Sw ¼ þ Swr presented here for one-dimensional problems. The
1
 n 1¹ two-pressure solution has been generalized to higher
p n
1þ a c dimensions.45 The numerical solutions of the fractional
row g flow equations (eqns (11) and (13)) have not been extended
where S ws ¼ 0.843; S wr ¼ 0.0716; a ¼ 0.044 cm ¹1; n ¼ 2.2 to higher dimensions using the methodology presented
and r ow is the density of water at standard temperature and here, although several other authors have presented
pressure. For the hydraulic conductivities, Touma and solutions to the saturation equation (eqn (13)) in
Vauclin fitted a polynomial to experimental observations two-dimensions.18,22,46,47
with the results:
3.1 Solution of two-pressure equations
A
kw ¼ w (JSw )Bw
Kws The two-pressure equation formulation was employed by
Celia and Binning6 in their numerical simulation of two-
Aa
kra ¼   phase air and water flow in the unsaturated zone. Their
pc B a method is based on a finite element discretization in
Aa þ
row g space, a backward Euler approximation in time, and a so-
with A w ¼ 18 130 cm h ¹1; B w ¼ 6.07; the porosity J ¼ 0.37; called modified Picard linearization. With superscripts n and
the saturated conductivity of water K w s ¼ 15.40 cm h ¹1; m denoting time level and iteration level, respectively, the
A a ¼ 3.86 ⫻ 10 ¹5; and the saturated conductivity of air time-discretized equation takes the following form (see
is K a s ¼ 2800 cm h ¹1. The air and water functional forms Celia and Binning for details):
are shown in Fig. 3. The figure is plotted using the normal-
ized water saturation S ¼ SSwsw ¹¹SSwrwr . The water relative perme- n þ 1, m
dSw dpan þ 1, m ¹ dpwn þ 1, m
ability does not reach k r w ¼ 1 because the maximum J:
saturation S w s is less than 1 during infiltration. The figure dpc Dtn þ 1
also shows the fractional flow function and the capillary  n þ 1, m 
⳵ kkrw ⳵dpnwþ 1, m
diffusion function. ¹
⳵z mw ⳵z
 n þ 1, m  n þ 1, m 
⳵ kkrw ⳵pw
¼ ¹ rw g
3 NUMERICAL METHODS ⳵z mw ⳵z

Three numerical methods are presented. The first solves Swn þ 1, m ¹ Snw
¹J ¼ Rwn þ 1, m ð18Þ
eqns (6) and (7) and is called a two-pressure solution. The Dtn þ 1
 Multi-phase flow simulation 467

Fig. 3. Functional forms of Touma and Vauclin43 for air and water in a coarse sand. (a) Relative permeabilities. (b) Pressure saturation
curve. (c) Capillary diffusion function. (d) Fractional flow function f. (e) Fractional flow function F for the case q T ¼ 0. The dashed line is
discussed in Section 3.2.3.
468 P. B. Binning, M. A. Celia

n þ 1, m first is an Eulerian approach due to Morel-Seytoux and

dS dpan þ 1, m ¹ dpwn þ 1, m
¹ J: w Billica,12,13 who employed a simple finite difference
dpc Dtn þ 1 scheme. The second is an Eulerian–Lagrangian approach
 n þ 1, m  that solves the saturation equation using a modified
⳵ kkra ⳵dpan þ 1, m
¹ method of characteristics (MMOC).
⳵z ma ⳵z
 n þ 1, m  n þ 1, m 
n þ 1, m
Sw n
¹ Sw ⳵ kkra ⳵pa 3.2.2 Finite difference solution of the saturation equation
¼J þ ¹ ra g
Dtn þ 1 ⳵z ma ⳵z The finite difference scheme developed by Morel-Seytoux
and Billica employs a centered-in-space discretization and a
¼ Ran þ 1, m ð19Þ Euler backward time stepping, with a Picard iteration
scheme for the nonlinear coefficients. For evenly spaced
where Dt nþ1 ¼ t nþ1 ¹ t n is the time step size, nodes and a constant (in space) total velocity, the scheme
n þ 1, m
kra ¼ kra (Snwþ 1, m ), and dpan þ 1, m ¼ pan þ 1, m þ 1 ¹ pan þ 1, m . can be written
These equations are linearized with unknowns dpwn þ 1, m
and dpan þ 1, m . The right hand side Ran þ 1, m is a function of " #
DtDjn¹þ1=2
1, m
the known pressures at the new time level and old iteration ¹ Sjn¹þ11, m þ 1
(n þ 1,m) and old time (n) levels. At each iteration step the Dz2
equations are solved and the dpan þ 1, m used to update the " n þ 1, m
!#
pressures. Iteration proceeds until the residual Ran þ 1, m Dt Dj ¹ 1=2 Djnþþ1=2
1, m
þ Jþ þ Sjn þ 1, m þ 1
given by the right hand side of each equation reaches a Dz Dz Dz
user specified error tolerance. The residual can be seen to " #
be a measure of the degree to which the numerical solution DtDjnþþ1=2
1, m
¹ Sjnþþ11, m þ 1 ¼ JSnj
satisfies the governing equation. Dz2
To complete the discretization, a spatial approximation is
Dt n þ 1, m
required. A lumped Galerkin finite element discretization is þ (F ¹ Fjnþþ1=2
1, m
) ð20Þ
chosen using piecewise linear basis functions. Lumping of Dz j ¹ 1=2
the mass matrix is useful in controlling oscillations that
arise through the finite-element discretization of the time where j is an index over space, n denotes the time level, and
derivative. The numerical equations are solved simul- m the iteration level.
taneously for the air and water pressures, with the air and
water equations at each node being written alternately in the 3.2.3 MMOC solution of the saturation equation
matrix formation. This leads to a five-banded block While the finite difference approximation (eqn (20)) has
symmetric matrix. certain benefits, especially for one-dimensional problems,
it ignores the fact that the hyperbolic part of the saturation
3.2 Numerical solution of the fractional flow equations equation may be solved by numerical methods that take
advantage of the secondary (hyperbolic) characteristics of
The fractional flow approach comprises two equations—the eqn (13). The total flux is constant in one dimension and
pressure equation (eqn (11)) and the saturation equation smooth in higher dimensions, and so an infiltrating front of
(eqn (13)). These equations present a greater numerical the wetting fluid will move at a fairly uniform velocity with
challenge than those of the two-pressure approach. The a fairly constant shape.
two-pressure approach involves two equations with a simi- The saturation equation (eqn (13)) can be solved by a
lar mathematical character, and so the same numerical characteristic-based approximation. One such method,
method can be used for both equations leading to a simple based on the MMOC with localized grid refinement, was
numerical technique. In contrast, the fractional flow outlined and implemented by Celia and Binning for the case
approach involves two equations with completely different of known q T and simple boundary conditions. Other similar
characters, and so separate numerical techniques must be approaches include the recent Eulerian–Lagrangian loca-
devised for each equation. lized adjoint method (ELLAM) approach of Dahle et al.,48
as well as earlier work of Dahle,18 Espedal and Ewing,17 and
3.2.1 Saturation equation others cited earlier in this paper. The procedure used herein
One of the main attractions of the fractional flow approach is an extension of the MMOC approach outlined in Celia
is the form of the nonlinear advection term in eqn (13). The and Binning.20 The key enhancements to the original
family of characteristics associated with the hyperbolic part algorithm of Espedal and Ewing17 include the changed defi-
of these equations provide the physically-correct speed for nition of F w so that it remains well defined in the case of zero
an infiltrating front. Numerical methods may be developed total flux, the definition of operator splitting for growing
to take advantages of this information, thereby leading to infiltration fronts, and the development of an algorithm to
potentially improved numerical performance. Two methods handle general boundary conditions.
of solution of the saturation equation are presented here. The The numerical method solves the saturation equation
 Multi-phase flow simulation 469

(eqn (13)), which we will rewrite: The operator splitting discussed above is for the case of
  the wetting fluid infiltrating into non-wetting fluid. Under
DSw ⳵ ⳵Sw
J ¹ D(Sw ) ¼0 (21) drainage, and combined drainage–infiltration events, the
Dt ⳵z ⳵z
definition needs to be modified. The discussion of these
where modifications is beyond the scope of this work. Questions
DSw ⳵Sw Fw ⬘(Sw ) ⳵Sw also arise in heterogeneous materials where the character-
¼ þ (22) istics cross the boundaries between two materials having
Dt ⳵t J ⳵z
different fractional flow function definitions. It is not clear
is the material derivative of saturation. The hyperbolic part how this would be handled with the current method. While
of the equation DSDt ¼ 0 is nonlinear and produces solutions
w
the work of Langlo and Espedal22 addresses the issue of
with shocks, although these shocks (or fronts) move with heterogeneity, it avoids this problem by using hetero-
well-defined speed. This frontal speed is calculated using geneous materials where only the absolute permeability
the convex envelope of the fractional flow function through varies, with the nonlinear functional forms remaining the
the following algorithm. The saturation equation is split same.
into two equations Numerically, eqns (23) and (24) are solved separately
18,50
. The characteristic eqn (23) is solved using a character-
DSw ⳵Sw F̄⬘(Sw ) ⳵Sw istic algorithm, and the diffusion correction eqn (24) is
¼ þ (23)
Dt ⳵t J ⳵z solved using a spatially optimal Petrov–Galerkin finite
  element method.
DSw ⳵(b(Sw )Sw ) ⳵ ⳵Sw The first step in the numerical algorithm is the numerical
J þ ¹ D(Sw ) ¼0 (24)
Dt ⳵z ⳵z ⳵z determination of the operator splitting. This is done by find-
ing Stw and Sbw . Eqn (26) is solved using the method of
where the new functions F̄⬘ and b are defined in order that bisection and the result compared to the maximum satura-
the equation DSDt ¼ 0 gives the physical frontal speed. Eqn
w
tion in the domain to define Stw as above. Sbw is defined to be
(23) can be termed a ‘characteristic equation’ and eqn (24) the minimum saturation in the domain.
a ‘diffusion correction equation’. The correct definitions for Localized grid refinement is very attractive with a
F̄(Sw ) and b(S w) are found by examining the Buckley– Lagrangian approach to the governing equations. The
Leverett solution to the equation. This leads to the saturation equation can be solved analytically for frontal
following definition: speed, and the speed used a priori to refine the grid at the
F(Sw ) ¼ F̄(Sw ) þ b(Sw )Sw projected location of the infiltrating front at the new time
step. The frontal location is defined by examining the
where
2 gradient of saturation. In regions where the grid has been
Fw (Sw ) Stw ⱕ Sw ⱕ 1 refined but the gradient of the solution is no longer large, the
6
F̄(Sw ) ¼ 4 Fw (Stw ) ¹ Fw (Sbw ) (25) mesh refinement is removed.
Sw ⱕ Stw The characteristic equation (eqn (23)) is solved by back-
Stw ¹ Sbw tracking along characteristics from the new time level to
n
and b(S w) is defined in the appropriate manner. Note that find the characteristic solution S̄ 17. Because the velocity
the above splitting is based on the assumption that F w(S w) is of the front is a function of the saturation, an iterative
concave for S w ⱖ Stw . scheme must be employed to find the correct solution.
The simulations presented in this paper are all one- Note that a major source of error in the numerical solution
dimensional. Splitting in the higher-dimensional case is through the interpolation of the saturation to grid
including gravity has been considered by Hansen47 and locations inbetween nodes.
Hansen et al.23 The function F̄(Sw ) for oil and water, as The material derivative in the diffusion correction
used by Dahle,18 is shown in Fig. 1b, as the concave equation is then approximated using the characteristic
envelope of the fractional flow function. solution and the diffusion correction equation is written as
While the shock is never fully developed for problems n
Swn þ 1 ¹ S̄w ⳵
with non-zero capillary diffusion, the algorithm is designed J þ (b(Swn þ 1 )Swn þ 1 )
to assume that the infiltrating front is fully developed, so Dt ⳵z
 n þ 1
that the Buckley–Leverett saturation is used for Stw . The ⳵ n þ 1 ⳵Sw
Buckley–Leverett saturation is found using the tangent ¹ D(Sw ) ¼0 ð27Þ
⳵z ⳵z
law of Welge49 which solves the equation
F(Sw ) dFw The diffusion correction is discretized using quadratic
¹ ¼0 (26) Petrov–Galerkin finite elements, and linearized using a
Sw dSw
standard Picard iteration scheme. The Picard iteration
for the Buckley–Leverett saturation S BL. If the maximum scheme lags the nonlinear coefficients in the diffusion cor-
saturation in the domain is less than the Buckley–Leverett rection equation (eqn (27)) by one iteration, thus linearizing
saturation then Stw is reduced to this lower value. the equation. The Petrov–Galerkin method divides the
470 P. B. Binning, M. A. Celia

domain into E elements Q e ¼ [z e,z e þ1]. Piecewise linear variable). Boundary conditions may also be specified in
basis functions, v i, and spatially ‘optimal’ test functions, terms of the total flux q T (or gradients of P). The total-
w i are used, with sufficient upstream weighting added to flux-type boundary condition is often used in simulations
provide enough artificial diffusion at the solution front so of petroleum reservoirs. In petroleum reservoirs the total
that non-unique solutions do not form.50 The test functions flux can conveniently be specified as zero, as the domain
are weighted ‘upstream’ in relation to the b(Swn þ 1, m ) term, of simulation is often enclosed by no-flow boundaries (for
which is an advection term in the opposite direction to the example in a five-spot problem).
flow. Fig. 4 shows a fully upstream weighted test function In hydrology, complex boundary conditions involving
for a front that is moving from left to right (since b is an combinations of individual fluid fluxes or pressures often
advection term that sharpens the front it is in the opposite must be specified. These types of boundary conditions at
direction to the frontal movement). Within the Petrov– the land surface are difficult to implement using the
Galerkin procedure, the temporal derivatives are lumped fractional flow form of the equations, because of the non-
using the L1 scheme of Milly,51 and the grouped term physical nature of the global pressure. For example, for
b(S)S is expanded using the finite element representation. infiltration problems, the boundary conditions at the land
The resultant algebraic system is tridiagonal in structure, surface are typically specified as a given air pressure and
and is solved using the Thomas algorithm. See Binning45 a water flux. From these two values, neither P nor S w can be
for details. calculated.
In one-dimensional problems it is possible to derive
3.3 Solution of pressure equation analytically an expression for total flux1. Because we wish
to develop algorithms that are applicable to higher-
Compared to the saturation equation the pressure equation is dimensional problems, where analytical expressions are
very well behaved. The pressure equation is a highly non- not available, we will not employ the analytical expressions
linear elliptic equation. The pressure, however, is relatively in the one-dimensional problem. Rather, a more general
smooth and the numerical approximation is straightforward. technique will be developed.
Since all simulations presented here are for a homogeneous An iterative technique is required for implementation of
medium, a simple Galerkin finite element method with general boundary conditions. Consider the case where the
piecewise linear basis functions is used instead of the flux of wetting phase q w and pressure of the non-wetting
mixed methods of Ewing and Heinemann.27 phase p nw are known at the top boundary and the saturation
and global pressure are specified at the lower boundary. In
3.4 Treatment of boundary conditions this case the saturation and global pressure within the
domain, boundary conditions at the top end of the column
If the fractional flow approach is applied to general multi- and the total velocity are unknown. An iterative solution for
dimensional, incompressible fluid flow problems, the P, S w and q T in the interior and P and S w at the boundary is
appropriate equations are the pressure equation (eqn (11)), developed. The procedure can be described as follows:
and the saturation equation (eqn (13)). This set of equations
is solved for the water saturation, S w, and the global pressure, 1. Begin the solution by estimating the total flux by
P. The boundary conditions for these equations can be using the known wetting phase boundary flux q T ¼
specified either in terms of the saturation of water or in q w.
terms of the global pressure (which is a non-physical 2. Next estimate S top, the saturation at the top boundary
by dividing the given value of q w by the saturated
hydraulic conductivity. Use this ratio as a guess for
k rw, and determine the value of S top implied by
that value of k rw. Mathematically, solve
krow g=mw ¹ krw (Stop ) ¼ 0 for the saturation S top at
qw

the top boundary.

3. Given S top at the boundary solve the saturation
eqn (13) using MMOC.
4. Given the solution S w, calculate the capillary pressure
at the top boundary, then with the given p a at the
boundary, calculate p w. Now with p a, p w and S top
known at the boundary, calculate the total pressure,
P top, at that boundary using eqn (9). The integral in
eqn (9) must be evaluated with care, because of the
highly nonlinear nature of the integrand. In particular
the derivative of capillary pressure is infinite at both
S w ¼ S w s and S w ¼ S c ¼ S w r. The extended midpoint
Fig. 4. Basis and test funtions for the MMOC. rule with Romberg’s method is used 52 to construct a
 Multi-phase flow simulation 471

ZSw  
1 dpc
Fig. 5. Integral ¹ f w ¹ dS as a function of saturation in the definition of global pressure using the functional forms of Touma
Sc 2 dSw w and Vauclin43 for the case of air and water phases.

lookup table for the integral in eqn (9). The integral particular for the combination of a known flux in the
from the definition of global pressure is shown in water phase and a known pressure in the air phase. The
Fig. 5. algorithm is not directly applicable to other combinations
5. Given P top at the boundary, solve the pressure eqn of phase fluxes/pressures at the boundary. Different
(11). algorithms based on similar iterative techniques will need
6. Update q T from the pressure solution, and then find to be developed for these combinations. The requirement for
q w at the top boundary using eqn (12). different iterative algorithms for each boundary type will
7. Update S w at the top boundary by matching the not add to the computational time required to solve the
calculated q w at the top boundary with the known fractional flow equations. However, the complexity of the
boundary flux. The algorithm has the following steps: computer code will be greatly increased in fractional flow
codes that are developed to satisfy generalized boundary
(a) In the first iteration on the boundary conditions
conditions. This contrasts with the two-pressure approach,
estimate DS top by employing a discrete version of the
where different boundary conditions can be accommodated
mass balance eqn (1) in the boundary element
relatively easily.
qw2 ¹ qw1 Dt
DStop ¼
Dz J 3.5 Mass-balance
where q w 1, q w 2 are the fluxes at the first and second
A mass-balance calculation provides necessary (but not suf-
nodes of the domain respectively. These fluxes are
ficient) verification of the validity of the numerical solution.
estimated using the current saturation solution and
The mass-balance for the two-pressure solution has been
eqn (12).
described in Celia and Binning.6 They showed that mass-
(b) In the first and subsequent iterations, compare the balance errors for this numerical scheme are small. In
calculated flux q w 1 with the desired boundary flux. If contrast, a MMOC discretization of the fractional flow
the calculated flux is less than the desired boundary equations is prone to mass-balance errors. These arise
flux, increment the saturation at the top boundary, through a poor treatment of the boundary conditions in the
S top, by DS top. Otherwise, decrement S top by DS top. In
each iteration decrease the size of DS top by dividing by
2. This iterative approach is similar to the method of
bisection and will converge provided the initial esti-
mate of DS top is large enough.
8. Re-solve the saturation equation by MMOC and
repeat the loop (steps 3–8) until the solution
converges.
Fig. 6 shows a simplified schematic of the algorithm. The
Fig. 6. Sequence of computation for solution of the pressure–
algorithm follows closely those proposed by Chen et al.24 saturation equations with general boundary conditions. The
and Vassilev (A. Vassilev, personal communication, 1996). numbers in the diagram refer to sequence of steps explained in
The boundary condition algorithm described above is the text.
472 P. B. Binning, M. A. Celia

MMOC. The mass-balance errors inherent in the MMOC comparison is made of the two-pressure approach, and the
have been noted by Healy and Russell.53 Recently, Dahle finite difference and MMOC solutions of the saturation
et al.48 have shown how an ELLAM can be used to address equation. A second problem is then solved where general
the mass-balance errors inherent in the MMOC discretiza- hydrologic boundary conditions are applied, so the global
tion of the saturation equation. The ELLAM was devised by pressure equation must also be solved in the fractional flow
Celia et al.54 to provide a rigorous framework for character- approach. For both problems, material properties are taken
istic methods and an improved treatment of boundary from the measurements of Touma and Vauclin,43 as
conditions. presented above.
The finite difference algorithm of Morel-Seytoux and
Billica is based on the conservative form of the governing 4.1 Example 1
saturation equation (eqn (13)), and as with any implicit
centered finite difference approximation of the conservative In petroleum engineering applications it is natural to specify
form of the advection–dispersion equation, it is perfectly boundary conditions in terms of the total flux and the satura-
mass conservative. tion of a single fluid. The first example considers a one-
Mass-balance errors have been calculated for all three dimensional problem in hydrology where it is possible to
methods by comparing the mass change in the domain specify such boundary conditions. For one-dimensional,
calculated from the saturations with the boundary fluxes, incompressible flow the pressure equation is degenerate
either specified as a boundary condition or calculated and so the total flux is constant in space and determined
from the solution. The results are reported for a number of by the boundary conditions. If the total flux is specified at
simulations below. the boundary then it is not necessary to solve the pressure
equation. The first example considers infiltration of water
into an initially dry soil with a uniform normalized satura-
4 RESULTS tion of 0.1646. The normalized water saturation at the soil
surface is specified to be 0.9159, and at the bottom of the
Several problems relevant to applications in hydrology are soil column it is 0.1646. The bottom of the column is sealed
presented. They consider multi-phase infiltration of water to both the air and water phases, so that the total flux is zero.
into a porous medium initially filled with air and water. First Note that the boundary conditions at the bottom of the
a simulation is presented where saturations are specified as column, q w ¼ q a ¼ 0 and S w ¼ 0.1646, are over-specified
the boundary conditions and the total flux is known, so it (normally the saturation at the bottom of the column would
is only necessary to solve the saturation equation. A not be given) to avoid the need for an iterative procedure to

Fig. 7. Solution of a problem with constant saturation boundary conditions using three different numerical methods. The MMOC is shown
with nodes marked. The finite difference solution and the two-pressure solution are shown as solid lines and are indistinguishable from each
other. The computational effort involved in the MMOC solution is about one tenth that of the other techniques (see Table 1). There is,
however, some loss of accuracy in the colution.
 Multi-phase flow simulation 473

determine the bottom boundary conditions. For water to with the MMOC are due to the iteration scheme in the
infiltrate into the column it is necessary for the air to escape solution of the ‘diffusion correction’ equation with a larger
through the soil surface. This is possible since the soil number of iterations required as the strength of the non-
surface is not saturated. linearities increases. The single step integration of the
The problem is solved using three numerical methods. hyperbolic part of the equation will also become more
The first is the two-pressure solution. The others are based inaccurate with choice of a larger time step size. In contrast,
on the fractional flow approach, and use the MMOC and a the time step choices for the finite difference and two-
finite difference approximation to solve the saturation pressure solutions are restricted by a Courant number
equation. The two-pressure solution employs a spatial criterion, which for explicit schemes can be given by35
discretization with elements of size Dz ¼ 1 cm and variable- !
sized time steps. The total simulated time is 100 min, and 1 Fw (Stw ) ¹ Fw (Sbw ) Dt
Cr ¼ ⬍1
the initial time step size is Dt ¼ 2 s. Initially it is necessary J Stw ¹ Sbw Dx
to employ a small time step as the pressure changes are steep
at the infiltrating front and so the nonlinearities are strong. For the implicit schemes employed here the Courant
However, as the water infiltrates into the column the infil- number criterion is less strict,55,56 but is still a good
trating front smooths, so the nonlinearities weaken and guide for choice of time step.
larger time steps can be taken. In this case the time step
can be increased from its initial size of 2 s, up to 116 s at 4.2 Example 2
the end of the simulation. In each iteration the error
tolerance on the residual (See eqns (18) and (19)) is set to In hydrology, boundary conditions are frequently specified
be 1 ⫻ 10 ¹3. in each phase separately. A common problem specifies a
The finite difference solution employs the same spatial flux in the water phase and a pressure in the air phase. In
grid as the two-pressure solution, and constant sized time this case the coupled pressure and saturation equations must
steps of size Dt ¼ 200.0 s. The time step size is chosen to be be solved simultaneously. For the present example, a
as large as possible, and is constrained by the need to ensure column of soil is considered with a normalized initial
that the nonlinear iteration scheme is convergent. The error saturation of 0.1655. The boundary conditions for the
tolerance in the nonlinear iteration scheme on saturation is problem are that the water flux is fixed at 8.3 cm h ¹1 and
0.001. the air pressure equal to 0 cm water at the soil surface. The
The MMOC solution employs a coarse spatial grid with air pressure is set to be 0.1204 cm and the water pressure
Dz ¼ 10.0 cm and localized grid refinement with the same ¹99.8796 cm at the bottom of the soil column. The air
grid size as used in the two-pressure solution. The grid is boundary conditions are chosen to be those for a static
refined whenever the saturation change between neigh- equilibrium in the air phase and the bottom boundary con-
boring nodes exceeds 0.05 (see the solution in Fig. 7 for a dition on the water phase is chosen to match the initial water
typical grid). A constant sized time step of 600 s is used and content of the column. The column is filled with a coarse
the error tolerance on the nonlinear iteration schemes is set sand, having the properties given in Fig. 3. The problem was
to be 0.001 on the saturation. modeled by Celia and Binning6 using the two-pressure
A comparison of the solutions obtained by the three approach. The problem is solved here using both the
methods is given in Fig. 7. Table 1 gives a comparison of fractional flow approach and the two-pressure solution and
the computational effort required by the three methods. The comparisons are made between the methods.
results from the three simulations suggest a significant The fractional flow approach employs the MMOC to
difference in computational efficiency. The two-pressure solve the saturation equation and the finite element method
solution was the most inefficient scheme, the finite to solve the pressure equation. The iterative methodology
difference solution of the saturation equation marginally described previously is used to determine the boundary con-
better, and the MMOC solution of the saturation equation ditions. A coarse grid of 11 nodes with a spacing of 10 cm is
the most efficient solution technique. used. The grid is locally refined, with each element being
The efficiency of the MMOC is due to the lack of restric- broken into 10 smaller elements whenever the normalized
tion on time step size. The only restrictions of time step size saturation change between neighboring nodes exceeds 0.05.

Table 1. Comparison of the numerical efficiency of techniques for solving the two-phase flow problem in a case with fixed saturation
boundaries
Two-pressure solution MMOC Finite difference
Computational time (s) (Sun SPARCstation 5) 43.1 60.0 34.0
Number of time steps 96 10 30
Courant number 0.055–1.06 5.52 1.84
Iterations/time step 3–10 14–2 6–13
Mass balance error 0% 3.5% 0%
474 P. B. Binning, M. A. Celia

This value was chosen as it provides a good balance Fig. 8 shows that the infiltrating water has a fairly
between computational efficiency and accuracy. Ten time uniform frontal speed and a constant shape. The uniform
steps of size 10 min are employed in the solution. Error frontal speed is given by the Buckley–Leverett equation.
tolerances on each of the iterative processes in the code The infiltrating front retains its shape as the capillary diffu-
must also be specified. These are the iterative characteristic sion is exactly balanced by the self-sharpening character of
solver and the iterative ‘diffusion correction’ in the solution the hyperbolic part of the equation. These observations are
of the saturation equation, and the iteration criteria for con- the key to the computational efficiency of the MMOC solu-
vergence of the boundary conditions. In all three cases an tion. Note also that Fig. 3c shows that the capillary diffusion
error tolerance of 1 ⫻ 10 ¹3 was specified on the normalized acts only over high ranges of saturation, as observed in
saturation. Fig. 8. Analysis of the fractional flow function and the
The two-pressure solver uses a grid with the same resolu- capillary diffusion function give a good a priori idea of
tion as the refined grid of the fractional flow solver, i.e. 101 the form of the solution to the equations.
nodes of spacing 1 cm. A time step size of 6 s is initially Fig. 9 shows the global pressure corresponding to the
applied, and the time step is increased by a factor of 1.05 at saturation solution shown in Fig. 8. The global pressure is
later times when the number of nonlinear iterations falls a mathematical convenience and so no physical interpreta-
below 10. Using this time step acceleration scheme the tion can be made of this figure. The figure does show one
time step size could be increased from the initial 6 s to problem with the fractional flow approach. The localized
101 s at the end of the simulation. The convergence criteria grid refinement is based on gradients of the normalized
on the solution is 1 ⫻ 10 ¹3 on the residual. saturation and not on the global pressure. As can be seen
The two solutions are compared in Figs 8 and 9, and the from Fig. 9 the global pressure gradient is quite large in
computational cost of each solution is shown in Table 2. regions remote from the saturation front. In particular, the
Fig. 8 shows saturation as a function of depth at various time global pressure variation at the boundary is quite large.
intervals. The distribution of nodes used in the fractional However, the coarse grid is used at the boundary after the
flow solution is also shown in the figure. As can be seen infiltration front has moved into deeper parts of the column.
from the figure, the fractional flow solver has used a refined This leads to quite large errors in the determination of the
grid in the neighborhood of the steep infiltration front. The boundary conditions. A modified grid refinement algorithm
two-pressure solution is perfectly mass conservative. In could be employed where both the pressure and saturation
contrast the fractional flow solution shows a 4% mass gradients are used to determine the degree of refinement.
balance error, with the infiltrating front lagging behind the However, such an approach would lead to large portions of
correct location. the grid requiring refinement, reducing the computational

Fig. 8. Saturation at various times in an initially uniform dry soil with a fixed water flux at the oil surface (depth ¼ 0 cm). The figure shows
a comparison between the fractional flow solution and the two-pressure solution.
 Multi-phase flow simulation 475

Fig. 9. Global pressure obtained using the two-pressure and fractional flow approaches for the problem illustrated in Fig. 8.

advantages of localized grid refinement for the example 5 DISCUSSION

being considered here.
The two-pressure solution is more accurate than the frac- Numerical solutions based on the fractional flow form of the
tional flow solution, with large errors in the fractional flow governing equations appear potentially attractive, because
solution due to discretization errors leading to a poor treat- of the relatively simple form of the pressure equation
ment of the boundary conditions. The computational effort (eqn (11)), and the hyperbolic character of the saturation
involved in obtaining these solutions is given in Table 2. equation (eqn (13)). Numerical algorithms can be designed
The two-pressure solution requires about twice as much to take advantage of these properties, with the aim of gain-
computational time as the fractional flow solution. The frac- ing significant improvements in computational efficiency.
tional flow solution required 31.85 s on a Sun workstation. The results presented here show that for problems in one
Of this time the majority was spent in solving the highly dimension where the saturation equation alone can be
nonlinear saturation solution (29.9 s), with only a small solved, the fractional flow approach is far more efficient
amount of time required to solve the much better behaved than equivalent solvers employing the two-pressure
pressure equation (1.0 s). approach. However, for generalized boundary conditions,
The simulation shown here has been chosen to illustrate such as those frequently employed in hydrology, the full
the deficiencies of the fractional flow solution. As the spatial pressure–saturation solution of the fractional flow equations
and temporal discretizations are refined the solutions con- must be coupled with an iterative scheme to find the
verge. To obtain a solution with the fractional flow approach correct boundary conditions. In this case there was little
that is of similar accuracy to the two-pressure solution or no computational advantage gained by the fractional flow
requires a uniform fine grid. Localized grid refinement approach.
based only on the saturation solution does not give satis- The fractional flow approach, while offering some
factory results. If the fractional flow solution is required to computational attractions, is also far more complex to
obtain a high accuracy solution the advantages in computa- understand and to implement in computer codes. The two-
tional efficiency of the method over the two-pressure pressure code requires approximately 690 lines of
approach disappear. generously commented code, whereas the fractional flow

Table 2. Comparison of computational effort for the fractional flow and two-pressure solutions for the problem illustrated in Fig. 8

Two-pressure solution Finite difference

Computational time (s) (Sun SPARCstation 5) 48.9 31.85
Time steps 105 10
Iterations/time step 5–10 5–19 diffusion, 3–11 boundary
Mass balance error 0% 4%
476 P. B. Binning, M. A. Celia

code requires 2080 lines. The largest portion of the frac- implications for the back-tracking step of the characteristic
tional flow code is the MMOC solution of the saturation solution and the treatment of the ‘anti-diffusion’ correction
equation. Given that no computer code is ever bug free, that is associated with the fractional flow splitting. Because
the significantly larger amount of computer code required numerical solutions are quite sensitive to the treatment of
for the fractional flow approach should be a serious the anti-diffusion, additional study appears to be needed to
consideration when choosing the solution methodology. incorporate general material heterogeneities. A similar
The results presented here are for one dimension. It is situation arises when multiple infiltration and drainage
possible that in higher dimensions the fractional flow fronts occur, as might be expected for simulation of inter-
approach will become more competitive. For example, it mittent rainfall events. In this case, different fractional flow
was shown in Binning45 that the time step acceleration splittings are required for the different fronts, even when
that can be used very successfully in one dimension with hysteresis is ignored.
the two-pressure approach, gives little advantage in higher One of the most appealing features of the fractional flow
dimensions. In contrast the fractional flow approach has formulation is the physical insight offered by the form of the
been demonstrated to be equally effective with large time saturation equation, as recognized by pioneers in this area
steps for petroleum problems in both one and higher dimen- including Buckley and Leverett10 and Morel-Seytoux.1
sions.18 Until the time when the higher-dimensional version While numerical methods based on the fractional flow
of the fractional flow approach with general boundary formulation of the governing equations are very attractive
conditions has been completed, no final conclusions on for simple model problems, their extension to practical
the efficiency of the method can be made. problems remains to be demonstrated. The results presented
There are a number of drawbacks to the fractional flow here for general boundary conditions suggest that potential
formulation. These include the non-physical nature of the gains in numerical performance due to the equation form
total pressure term; the associated complications in imple- may be offset by the additional complications of the
menting boundary conditions whose specification is method. Much work needs to be done before numerical
associated naturally with individual phases instead of methods based on the fractional flow approach will be
combinations of phase information; the complications intro- able to compete with methods having more general applic-
duced by the nature of the gravity terms; the difficulties in ability, such as those of Celia and Binning6 and Forsyth and
dealing with multiple infiltration and drainage fronts; the coworkers.34,35
problem of including compressibility; and the complications
in the characteristic solution when material heterogeneity is
introduced. ACKNOWLEDGEMENTS
Implementation of general boundary conditions requires
different iteration schemes for each different pair of phase- This work was supported in part by the U.S. Department
specific boundary conditions, one example of which is given of Energy under Grant DE-FG02-95ER25267. We wish to
above. This causes both additional coding and additional acknowledge the contributions of Richard Ewing, Magne
computations which largely offset the gains made by Espedal, and Helge Dahle to this work. We would also
developing numerical methods that exploit the form of the like to thank the Journal Editor and the anonymous
governing equations. Note that for one-dimensional reviewers for their comments on the paper.
problems, the analytical techniques of Morel-Seytoux and
Billica12,13 can be used to avoid solving the full pressure
equation. However, that approach does not generalize easily REFERENCES
to multiple dimensions, while the approach presented herein
does. The gravity issue has been addressed effectively by 1. Morel-Seytoux, H.J. Two-Phase flows in porous media.
Hansen et al.,23 although its implementation in multiple Advances in Hydroscience, 1973, 9, 119–202.
dimensions implies some additional coding. 2. Ewing, R.E., Three-phase flow formulations. In Computa-
tional Methods in Water Resources XI, Vol. 1, ed. A.A.
The two outstanding issues that appear to be most Aldama, J. Aparicio, C.A. Brebbia, W.G. Gray, I. Herrera
significant at this time are the treatment of material hetero- and G.F. Pinder. Computational Mechanics Publications,
geneities in which the fractional flow function varies Southampton, 1996.
spatially, and the treatment of multiple infiltration and 3. Pinder, G.F. and Abriola, L.M. On the simulation of non-
aqueous phase organic compounds in the subsurface. Water
drainage fronts. Earlier work, by for example Langlo and
Resources Research, 1986, 22, 109S–119S.
Espedal,21 incorporated heterogeneity in the intrinsic 4. Sleep, B.E. and Sykes, J.F. Modeling the transport of volatile
permeability only, with the relative permeabilities (and organics in variably saturated media. Water Resources
therefore the fractional flow functions) remaining spatially Research, 1989, 25, 81–92.
constant. The more difficult case of heterogeneous 5. Kaluarachchi, J.J. and Parker, J.C. An efficient finite
elements method for modeling multiphase flow. Water
fractional flow functions requires that the splitting technique
Resources Research, 1989, 25, 43–54.
used to determine frontal speeds and the associated 6. Celia, M.A. and Binning, P. Two-phase unsaturated flow: one
numerical characteristic curves must account for underlying dimensional simulation and air phase velocities. Water
variability within a time step. This appears to have Resources Research, 1992, 28, 2819–2828.
 Multi-phase flow simulation 477

7. Binning, P. and Celia, M.A., Coupled air-water flow and 24. Chen, Z., Ewing, R.E. and Espedal, M., Multiphase flow
contaminant transport in the unsaturated zone: A numerical simulation with various boundary conditions. In Computa-
model. In Shallow Groundwater Systems, ed. P. Dillon and I. tional Methods in Water Resources X, Vol. 2, ed. A.
Simmers, IAH Int. Contributions to Hydrogeology 18. A.A. Peters, G. Wittum, B. Herrling, U. Meissner, C.A. Brebbia,
Balkema, 1998. W.G. Gray and G.F. Pinder, Kluwer Academic Publishers,
8. Binning, P., Celia, M.A. and Johnson, J.C., Two-phase flow Dordrecht, 1994.
and contaminant transport in unsaturated soils with applica- 25. Ewing, R.E. Simulation of multiphase flows in porous media.
tion to low-level radioactive waste disposal. Nuclear Transport in Porous Media, 1991, 6, 479–499.
Regulatory Commission, Washington, DC, NUREG/CR- 26. Ewing, R.E. and Heinemann, R.F., Incorporation of mixed
6114 Vol. 2, 1996. finite elements methods in compositional simulation for
9. Schrefler, B.A. and Xiaoyong, Z. A fully coupled model for reduction of numerical dispersion. Reservoir Simulation Sym-
water flow and air flow in deformable porous media. Water posium, San Francisco, CA, 1983, Soc. Petrol. Eng., Dallas,
Resources Research, 1993, 29, 155–167. TX, SPE 12267, 1983.
10. Buckley, S.E. and Leverett, M.C. Mechanism of Fluid 27. Ewing, R.E. and Heinemann, R.F. Mixed finite elements
Displacement in Sands. Trans. AIME, 1942, 146, 107–116. approximations of phase velocities in compositional reservoir
11. Sander, G.C., Norbury, J. and Weeks, S.W. An exact solution simulation. Computer Methods in Applied Mechanics and
to the nonlinear diffusion–convection equation for two-phase Engineering, 1984, 47, 161–175.
flow. Quarterly Journal of Mechanics and Applied 28. Guarnaccia, J.F. and Pinder, G.F., NAPL: A mathematical
Mathematics, 1993, 46, 709–727. model for the study of NAPL contamination in granular
12. Morel-Seytoux, H.J. and Billica, J.A. A two-phase numerical soils, equation development and simulator documentation.
model for prediction of infiltration: Applications to a semi- The University of Vermont, RCGRD #95-22, 1997.
infinite column. Water Resources Research, 1985, 21, 607– 29. Spillette, A.G., Hillestad, J.G. and Stone, H.L., A highly
615. stable sequential solution approach to reservoir simulation.
13. Morel-Seytoux, H.J. and Billica, J.A. A two-phase numerical Soc. Pet. Eng. 48th Ann. Meet., Las Vegas, NV, SPE Paper
model for prediction of infiltration: Case of an impervious no. 4542, 1973.
bottom. Water Resources Research, 1985, 21, 1389–1396. 30. Faust, C.R. Transport of immiscible fluids within and below
14. Wangen, M. Vertical migration of hydrocarbons modelled the unsaturated zone. Water Resources Research, 1985, 21,
with fractional flow theory. Geophysical Journal 587–596.
International, 1993, 115, 109–131. 31. Kueper, B.H. and Frind, E.O. Two-phase flow in hetero-
15. Chen, Z., Espedal, M. and Ewing, R.E. Continuous-time geneous porous media, 1. Model development.. Water
finite element analysis of multiphase flow in groundwater Resources Research, 1991, 27, 1049–1057.
hydrology. Applications of Mathematics, 1995, 40, 203–226. 32. Moridis, G.J. and Reddell, D.L. Secondary water recovery by
16. Douglas, J. Jr Finite difference method for two-phase air injection, 1. The concept and the mathematical and
incompressible flow in porous media. SIAM Journal of numerical model. Water Resources Research, 1991, 27,
Numerical Analysis, 1983, 20, 681–696. 2337–2352.
17. Espedal, M.S. and Ewing, R.E. Characteristic Petrov– 33. Pruess, K., TOUGH user’s guide. U.S. Nuclear Regulatory
Galerkin subdomain methods for two-phase immiscible Comission, Washington, DC, CR-4645, 1987.
flow. Computer Methods in Applied Mechanics and 34. Forsyth, P.A. A positivity preserving method for simulation
Engineering, 1987, 64, 113–135. of steam injection for NAPL site remediation. Advances in
18. Dahle, H.K., Espedal, M.S., Ewing, R.E. and Saevareid, O., Water Resources, 1993, 16, 351–370.
Characteristic adaptive subdomain methods for reservoir 35. Unger, A.J.A., Forsyth, P.A. and Sudicky, E.A. Variable
flow problems. Numerical Methods for Partial Differential spatial and temporal weighting schemes for use in multi-
Equations, 1990, 6, 279–309. phase compositional problems. Advances in Water
19. Dahle, H.K., Espedal, M.S. and Saevereid, O. Characteristic, Resources, 1996, 19, 1–27.
local grid refinement techniques for reservoir flow problems. 36. Abriola, L.M. and Rathfelder, K. Mass balance errors in
International Journal for Numerical Methods in Engineering, modelling two-phase immiscible flows: causes and remedies.
1992, 34, 1051–1069. Advances in Water Resources, 1993, 16, 223–239.
20. Celia, M.A. and Binning, P., Multiphase models of unsatu- 37. Haverkamp, R. et al. A Comparison of Numerical Simulation
rated flow: approaches to the governing equations and Models for One-Dimensional Infiltration. Soil Sci. Soc. Am.
numerical methods. In Proceedings IX International J., 1977, 41, 285–294.
Conference Computational Methods Water Resourses Vol 38. Russell, T.F., Modeling of multiphase multicontaminant
2: Mathematical Modeling in Water Resources, ed. transport in the subsurface. Reviews of Geophysics,
T.F. Russell et al. Elsevier Applied Science, London, 1992. Supplement, 1995, 1035–1047.
21. Langlo, P. and Espedal, M.S. Macrodispersion for two phase, 39. Sleep, B.E. and Sykes, J.F. Compositional simulation of
immiscible flow in porous media. Advances in Water groundwater contamination by organic compounds, 1.
Resources, 1994, 17, 297–316. Model development and verification. Water Resources
22. Langlo, P. and Espedal, M.S., Heterogeneous reservoir Research, 1993, 29, 1697–1708.
models, two-phase immiscible flow in 2-D. In Computational 40. Reeves, H.W. and Abriola, L.M. An iterative compositional
methods in Water Resources IX, Vol. 2: Numerical Methods model for subsurface multiphase flow. Journal of
in Water Resources, ed. T.F. Russell, R.E. Ewing, C.A. Contaminant Hydrology, 1994, 15, 249–276.
Brebbia, W.G. Gray, and G.F. Pinder, Elsevier Applied 41. Chavent, G. and Jaffré, J., Mathematical Models and Finite
Science, London, 1992. Elements for Reservior Simulation. North Holland,
23. Hansen, R., Espedal, M.S. and Nygaard, O., An operator Amsterdam, 1986.
splitting technique for two-phase immiscible flow dominated 42. Young, L.C. A study of spatial approximations for simulating
by gravity and capillary forces. In IX International Con- fluid displacements in petroleum reserviors. Computer
ference on Computational Methods in Water Resources, Methods in Applied Mechanics and Engineering, 1984, 47,
Vol. 1, ed. T.F. Russell, R.E. Ewing, C.A. Brebbia, W.G. 3–46.
Gray, and G.F. Pinder, Elsevier Applied Science, London, 1992. 43. Touma, J. and Vauclin, M. Experimental and numerical
478 P. B. Binning, M. A. Celia

analysis of two-phase infiltration in a partially saturated soil. 50. Dahle, H.K., Adaptive characteristic operator splitting tech-
Transport in Porous Media, 1986, 1, 27–55. niques for convective-dominated diffusion problems in one
44. van Genuchten, M.T. A closed form equation for predicting and two space dimensions. Thesis, University of Bergen,
the hydraulic conductivity in soils. Soil Sci. Soc. Am. J., Department of Applied Mathematics, 1988.
1980, 44, 892–898. 51. Milly, P.C.D. A mass-conservative proceedure for time step-
45. Binning, P., Modeling unsaturated zone flow and contami- ping in models of unsaturated flow. Advances in Water
nant transport in the air and water phases, Ph.D. Thesis, Resources, 1985, 8, 32–36.
Department of Civil Engineering and Operations Research, 52. Press, W.H., Flannery, B.P., Teukolsky, S.A. and Vetterling,
Princeton University, 1994. W.T. Numerical Recipes in C. Cambridge University Press,
46. Stothoff, S.A. and Dougherty, D.E., Implementing a three- Cambridge, 1988.
dimensional two-phase flow model on the connection 53. Healy, R.W. and Russell, T.F. A finite volume Eulerian–
machine. in Computational Methods in Water Resources Lagrangian localized adjoint method for the solution of the
IX, Vol. 1, ed. T.F. Russell, R.E. Ewing, C.A. Brebbia, advection–diffusion equation. Water Resources Research,
W.G. Gray and G.F. Pinder, Elsevier Applied Science, 1993, 29, 2399–2413.
London, 1992. 54. Celia, M.A. et al. An Eulerian–Lagrangian localized adjoint
47. Hansen, R., On the numerical solution of nonlinear reservoir method for the advection–diffusion equation. Advances in
flow models with gravity. Dr. Scient. Thesis, Department of Water Resources, 1990, 13, 187–206.
Mathematics, University of Bergen, 1993. 55. Daus, A.D., Frind, E.O. and Sudicky, E.A. Comparative error
48. Dahle, H.K., Ewing, R.E. and Russell, T.F. Eulerian– analysis in finite element formulations of the advection–
Lagrangian localized adjoint methods for a nonlinear dispersion equation. Advances in Water Resources, 1985, 8,
advection–diffusion equation. Computer Methods in Applied 86–95.
Mechanics and Engineering, 1995, 122, 223–250. 56. Noorishad, J. et al. A perspective on the numerical solution
49. Welge, H.J. A simplified method for computing oil recovery of convection dominated transport problems: a price to pay
by gas or water drive. Petroleum Transactions AIME, 1952, for the easy way out. Water Resources Research, 1992, 28,
195, 91–98. 551–564