You are on page 1of 24

Advances in Water Resources 145 (2020) 103716

Contents lists available at ScienceDirect

Advances in Water Resources


journal homepage: www.elsevier.com/locate/advwatres

Upscaled equations for two-phase flow in highly heterogeneous porous


media: Varying permeability and porosity
Tufan Ghosh a,c,∗, Carina Bringedal b,c, Rainer Helmig c, G.P. Raja Sekhar a
a
Department of Mathematics, Indian Institute of Technology Kharagpur, Kharagpur 721302, India
b
Stuttgart Center for Simulation Science (SimTech), University of Stuttgart, Pfaffenwaldring 5a, Stuttgart 70569, Germany
c
Institute for Modelling Hydraulic and Environmental Systems, University of Stuttgart, Pfaffenwaldring 61, Stuttgart 70569, Germany

a r t i c l e i n f o a b s t r a c t

Keywords: In this article we consider a two-phase flow model in a highly heterogeneous porous column. The porous column
Homogenization consists of homogeneous blocks, where the porosity and permeability vary from one block to the other. The
Upscaling flow direction is perpendicular to the layering of the porous column, and hence can be approximated by one-
Heterogeneous medium
dimensional model equations. The periodic change in porosity and absolute permeability enforce the fluid to
Two-phase flow
be trapped at the interface between the blocks, leading to a highly varying saturation. In order to capture the
effective behavior, upscaled equations for the average saturation are derived via homogenization. This technique
relies on a notion of periodicity and allows averaging over any number of blocks that may have any internal
distributions of the rock parameters. Moreover, the present article also derives effective equations for randomly
distributed layers of different porosity and absolute permeability. Numerical experiments are performed which
show good agreement between the averaged solutions of the original micro-scale equations and the solutions of
the upscaled equations, also in the case of randomly distributed layers. In particular these numerical experiments
show how the internal distribution of the permeability and porosity affect the effective behavior of the flow.

1. Introduction to flow from the injection to the production well. In this case, during
water-drive, oil phase trapping is encountered at the interface between
Multi-phase flow through highly heterogeneous porous medium is high and low permeability regions and flow becomes stagnant. This
encountered in various natural systems; for example, CO2 sequestration again causes reduction in the recovery rate. Fluid flow inside these cross-
in subsurface flow, oil/natural gas transport inside petroleum reservoirs, laminated heterogenous structures are studied by Kortekaas (1985) and
groundwater networks, etc. Generally these natural systems are spatially by Duijn et al. (1995). In a similar framework Dale et al. (1997) studied
heterogeneous and range from small-scale to large-scale formations. A a steady two-phase flow in a heterogeneous medium.
typical heterogeneity stems from the composition of two homogeneous A very common challenge in petroleum reservoir models is the effect
materials separated by a sharp interface, where this transition is non- of small-scale heterogeneities which significantly impact the large-scale
smooth. Understanding the effect of such a sharp interface on the move- behavior of flow through the reservoir formation. Specifically, the reser-
ment of these phases is an important topic and has drawn the attention of voir characteristics like porosity, intrinsic permeability, relative perme-
the research community in the last few decades. Kueper et al. (1989) ex- ability, etc., vary at a microscopic scale compared to the field scale.
perimentally studied the effect of the heterogeneities on two-phase flow In general these micro-scale variations largely control the macro-scale
through porous structures. van Lingen (1998) carried out a laboratory flow behavior. In order to capture these micro-scale variations one has
experiment on a column of porous layers with periodically varying per- to perform detailed simulations which are very expensive with respect
meability. It may be noted that in petroleum recovery process, rock to time and other resources. Hence, one should look for upscaling these
heterogeneities lead to counter-productive results. For example, if the oscillatory quantities and obtain a macro model or an intermediate-scale
permeability of the reservoir formation is high from the injection well model that is more amenable to perform simulations. Thus it is neces-
to the production well, a large amount of oil is bypassed during the sary to derive effective equations and constitutive relationships, which
water-drive due to the preferential flow paths. This leads to a substan- can further capture the influence of micro-scale heterogeneities and ef-
tial reduction in the recovery rate. On the other hand let us consider the fectively transfer them to the large-scale modeling approach. One may
case when the rock inhomogeneity is perpendicular or cross-laminated


Corresponding author at: Department of Mathematics, Indian Institute of Technology Kharagpur, Kharagpur 721302, India.
E-mail addresses: tufan.ghosh@iws.uni-stuttgart.de, tufu0321@iitkgp.ac.in (T. Ghosh), carina.bringedal@iws.uni-stuttgart.de (C. Bringedal),
rainer.helmig@iws.uni-stuttgart.de (R. Helmig), rajas@iitkgp.ac.in (G.P. Raja Sekhar).

https://doi.org/10.1016/j.advwatres.2020.103716
Received 27 December 2019; Received in revised form 31 July 2020; Accepted 3 August 2020
Available online 5 August 2020
0309-1708/© 2020 Elsevier Ltd. All rights reserved.
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

refer to Wen and Gómez-Hernández (1996) for a detailed overview of ear degenerate parabolic diffusion-convection equation for the wetting
the upscaling techniques for absolute permeabilities in heterogeneous phase saturation is developed. Considering a rapidly oscillating poros-
media. A comparative investigation between various dynamic upscal- ity function and absolute permeability tensor, a nonlinear homogenized
ing methods for two-phase flow based on fluid potentials is presented problem is obtained.
in Darman et al. (2002). It is noticed that a large variation in the porosities, permeabilities
A wide range of literature is available on the upscaling of multi- and the capillary pressure curves significantly control the rate and extent
phase fluid flow through porous media. Volume averaging techniques of the oil trapping or the storage of supercritical CO2 . Moreover, analyz-
and homogenization or two-scale asymptotic expansion techniques are ing the impact of block type heterogeneities on the oil trapping in the up-
popular upscaling methods for multi-phase flow through porous media. scaling context is an important area of research in the field of petroleum
Whitaker (1986) used volume averaging techniques and developed a extraction process. Accounting for the influence of the capillary and vis-
macroscopic model for Stokes flow of two immiscible fluids through cous forces, Virnovsky et al. (2004) presented an upscaled model for
a medium consisting of a rigid solid skeleton. Whitaker (1994) and steady-state two-phase flow in a heterogeneous medium with periodic
Lasseux et al. (1996) improved this process by transforming the closure structures. They computed rate-dependent relative permeabilities and
problem to Stokes’-like equations that can be solved to determine the effective capillary pressure for the above mentioned setup. Adopting
permeability tensors. Later, Lasseux et al. (2008) derived a macroscopic two-scale asymptotic expansion methods, van Duijn et al. (2002) de-
model through volume averaging techniques for inertial two-phase rived the effective flow equations for a layered media with flow orthog-
fluid flow through homogeneous porous media. In the context of two- onal to the strata. They also considered a medium consisting of multi-
phase flow through heterogeneous permeable structures, Quintard and ple layers and with random properties. Hence, upscaled equations are
Whitaker (1988) derived large-scale transport equations using volume derived for periodically as well as randomly layered media, where the
averaging. They also developed a closure scheme which can theoret- random nature is treated as a stationary ergodic structure. van Duijn
ically determine the large-scale permeability tensor along with the et al. (2002) assumed continuity of the first-order term in the asymptotic
large-scale capillary pressure. Quintard and Whitaker (1990a) used expansion of the oil saturation. This has lead to enhanced diffusivity
the method of volume averaging to capture the complexities of two- effect in the effective equations (van Duijn et al., 2007). Consequently,
phase flow in heterogeneous porous media and derived a macroscopic assuming continuity of the first-order terms in the asymptotic expansion
form of the momentum equations for two immiscible fluids, along with of the capillary pressure, van Duijn et al. (2007) derived the effective
the large-scale permeability tensor, and a dynamic, large-scale capil- quantities which are the weighted harmonic mean of the micro-scale
lary pressure. Here the corresponding closure problem was expressed properties and these are different from the ones derived by van Duijn
in terms of integro-differential equations for each of the phases in- et al. (2002). The differences in the effective quantities are mainly in
volving two permeability tensors. Continuing their work, Quintard and terms of the weights. In this article we adopt a similar approach as in
Whitaker (1990b) presented numerical experiments in a two-region van Duijn et al. (2007).
model of a stratified system based on local volume-averaged equations Schweizer (2008) presented a convergence proof of the homogeniza-
and compared these with the theoretical results obtained from large- tion procedure for the setting used in van Duijn et al. (2002, 2007). More
scale averaging (Quintard and Whitaker, 1988; 1990a). precisely, Schweizer (2008) proved the weak convergence of the multi-
In general, volume-averaging techniques provide the form of the scale saturation and flux towards their effective counterparts. However,
upscaled equations, but in order to determine the coefficients it without hinting to the explicit form, only the structure of the effec-
must be supplemented with physical arguments and/or associated tive quantities are provided for the upscaled models. Schweizer and
data (Berryman, 2005). Through assuming periodicity, explicit ex- Pop (2007) have shown that the effective models derived in van Duijn
pressions for the coefficients in the upscaled equations can be ob- et al. (2007) and Schweizer (2008) are equivalent. Using the concept of
tained through homogenization. One may refer Davit et al. (2013) for weighted solutions, Henning et al. (2013) derived a rigorous homog-
a comparative investigation between volume averaging with closure enization result for degenerate two-phase flow in a porous medium.
and formal periodic asymptotics based on multi-scale expansions. In Szymkiewicz et al. (2011) proposed a macroscopic model for two-phase
this context, a noteworthy upscaling method of considerable inter- flow accounting the trapping effects in a heterogeneous porous medium
est is the homogenization using two-scale asymptotic expansion meth- composed of disconnected inclusions embedded in the background ma-
ods. We refer to Hornung (1997), Cioranescu and Donato (1999) and terial. Later Szymkiewicz et al. (2012) extended the previous study of
Auriault et al. (2009) for more details on the homogenization technique. Szymkiewicz et al. (2011) with two specific objectives: (i) pointing out
Lewandowska et al. (2004) and Neuweiler and Cirpka (2005) used a ho- the limitations of the Richard’s equation for porous media with discon-
mogenization approach to derive upscaled quantities for unsaturated nected coarse inclusions; (ii) accounting for the effect of apparent air en-
porous media flow. Mikelić (2000) presented a detailed note on ho- try pressure due to heterogeneous structures, deriving an upscaled form
mogenization theory and its application to infiltration through porous of the modified Richard’s equation. In the context of unsaturated porous
structures. Several researchers (Beliaev, 2003; Bourgeat and Hidani, media flow, Szymkiewicz et al. (2014) have shown that the performance
1995; Yeh, 2006; Zijl and Trykozko, 2002) have relied on this homoge- of the Richard’s equation can be improved by introducing appropriately
nization approach and analyzed incompressible, immiscible two-phase defined effective permeability and capillary pressure functions, repre-
flow through heterogeneous porous media. Through homogenization, senting large-scale behavior of the heterogeneous medium.
Bourgeat and Panfilov (1998) constructed a macro-scale model for two- Bertsch et al. (2003) analyzed the existence, uniqueness and the
phase flow in a dual porosity medium. Moreover, they determined a regularity properties along with the matching conditions at the in-
capillary relaxation time. terface between the two types of rock with different permeabilities.
The existing literature indicates that upscaled models have been A generic model for capillary entrapment together with the conver-
exploited to address multi-phase flow through layered porous media. gence study of a numerical scheme of industrial interest is presented
Mouche et al. (2010) developed such a model for the vertical mi- by Enchery et al. (2006). Using an appropriate regularization and time
gration of a CO2 plume through a vertical column filled with a pe- discretization, Buzzi et al. (2009) showed the existence for degener-
riodically layered porous medium. Since the configuration is vertical, ate two-phase flow equations in one space dimension. For a simpli-
both capillarity and buoyancy are taken into account and semi-explicit fied model of two-phase flow, the influence of discontinuous capillary
upscaled flux functions are proposed. Transport of immiscible com- forces at the interfaces between several rock-types is investigated by
pressible fluids such as water-gas through porous media is discussed Cances (2010). Ahmadi et al. (2010) relied on numerical simulations to
by Amaziane et al. (2010). There a weak formulation of a nonlinear analyze the generalized Darcy-Forchheimer model for two-phase flow in
parabolic equation for the non-wetting phase pressure and a nonlin- a heterogeneous porous medium. Ern et al. (2010) presented a sequen-
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

tial discontinuous Galerkin method to approximate the trapping effects The governing equations are the mass conservation:
for two-phase flow through heterogeneous porous media with capillary 𝜕𝑆𝛼 𝜕𝑞
pressure discontinuity. Recently, van Duijn et al. (2016) studied a two- 𝜙(𝑥) + 𝛼 =0 (with 𝛼 = 𝑤, 𝑛𝑤), (2.1)
𝜕𝑡 𝜕𝑥
phase flow in heterogeneous porous media with dynamic capillarity ef-
and the multi-phase extended Darcy law (with x-axis along the flow
fects. The medium that they considered was composed of two adjacent
direction):
homogeneous blocks with a sharp interface, and they consequently de-
rived the coupling conditions at the sharp transition face. It is noticed 𝑘𝑟𝛼 (𝑆𝛼 ) 𝜕𝑝𝛼
𝑞𝛼 = −𝑘(𝑥) . (2.2)
that in most of the existing literature heterogeneity is mainly due to dif- 𝜇𝛼 𝜕𝑥
ferent permeabilities in the adjacent layers. Only a few researchers have The system is closed with the constraints on phase saturation and cap-
treated heterogeneities due to both porosity and absolute permeability. illary pressure as:
Moreover, the majority of the available studies in upscaling deals with
a single discontinuity in the unit cell. Models allowing multiple types of 𝑆𝑤 + 𝑆𝑛𝑤 = 1, (2.3)
heterogeneities and hence averaging over several discontinuities in the
unit cell expected to better describe the average behavior in a highly 𝑝𝑛𝑤 − 𝑝𝑤 = 𝑝𝑐 (𝑥, 𝑆𝑤 ). (2.4)
heterogeneous reservoir on a fairly coarse grid. We address such mod-
els in the current work. Here 𝜙, k, S, 𝜇, kr , p and q denote the porosity, the (intrinsic) ab-
Here, in this paper, we generalize the problem considered by van solute permeability, the saturation - mapped to the standard interval
Duijn et al. (2002, 2007), which deals with the trapping phenomena [0,1], the viscosity, the relative permeability, the phase pressure and the
at the micro-scale. We adopt homogenization techniques and two-scale specific discharge, respectively, with subscripts for each of the phases.
asymptotic expansion methods to derive the effective equations and The capillary pressure is described by the widely used Leverett model
the corresponding constitutive relations. We consider only the capillary (Leverett, 1941):
limit case; i.e., capillary number is of order one in the dimensionless √
form. Physically we are interested in a situation when the capillary ef- 𝜙(𝑥)
𝑝𝑐 (𝑥, 𝑆𝑤 ) = 𝜎 𝐽 (𝑆𝑤 ), (2.5)
fects due to surface tension are balanced with the viscous effects. The 𝑘(𝑥)
present study can not be thought of just as a simple generalization of van where 𝜎 and J are respectively the interfacial tension between the two
Duijn et al. (2002, 2007), as the current work derives effective equations fluids and the Leverett function. It is assumed that J is strictly decreas-
in two different configurations: (i) cell containing a single discontinuity ing, going to infinity as Sw ↘0, and that J(1) > 0. Physically this indicates
and (ii) cell containing multiple discontinuities. We allow porosity and that the pressure difference in the two phases should exceed the capil-
intrinsic permeability to vary on the micro-scale. We demonstrate the √
𝜙(𝑥)
lary entry pressure given by 𝑝𝑐 (𝑥, 1) = 𝜎 𝑘(𝑥)
𝐽 (1), before the oil phase
convergence through numerical examples. Further, this study considers
a variety of orientations for distribution of the porosity and permeabil- can penetrate into a fully water saturated core. It may be noted that
ity layers. In addition, this article derives upscaled equations for ran- the entry pressure is inversely proportional to the square root of the ab-
domly distributed layers of different porosity and absolute permeabil- solute permeability and directly related to the square root of porosity,
ity. This study also briefly presents upscaled quantities for any number and therefore varies between the layers. Thus the entry pressure will be
of discontinuities within a unit cell. Although we consider upscaling of discontinuous at the interface separating the two layers.
a one-dimensional model, understanding the complex behavior for this The mass conservation Eq. (2.1) together with the saturation con-
setting is important for extending to two or three dimensions, which will straint (2.3) implies that the total specific discharge (Darcy’s velocity)
𝜕𝑞
be taken up in future. 𝑞 ∶= 𝑞𝑤 + 𝑞𝑛𝑤 is constant in space, i.e., 𝜕𝑥 = 0. For the sake of simplicity
In Section 2 we present the mathematical model for two-phase flow we assume that the total specific discharge is constant with time as well.
through heterogeneous porous media. Section 3 deals with the deriva- Hence one can choose q > 0 as a prescribed quantity, i.e. the flow takes
tion of the upscaled equations for highly heterogeneous porous struc- place from left to right.
tures. Effective quantities for a single discontinuity within a cell can Following van Duijn et al. (2002, 2007), we set 𝑢 = 𝑆𝑛𝑤 and con-
be found in Section 3.2, and extended results for multiple discontinu- sequently 𝑆𝑤 = 1 − 𝑢. This eventually prompts us to redefine the func-
ities within a cell are shown in Section 3.3. Section 4 deals with the tional form of 𝑘𝑟𝛼 with 𝛼 = 𝑤, 𝑛𝑤, pc and J in terms of u. Now J be-
derivation of effective equations for randomly distributed blocks. A brief comes strictly increasing with u and approaches to infinity as u↗1, and
discussion on the numerical discretization of the micro-scale and the J(0) > 0.
upscaled equations are in Section 5.1 and 5.2 respectively. Various nu- Taking into account the constant nature of the total Darcy velocity
merical test cases are presented in Section 5.3. With the final remarks (specific discharge) one can combine Eqs. (2.1), (2.2) and conditions
in Section 6, we conclude this article. (2.3), (2.4) into a single transport equation for the oil saturation. Con-
sequently, the saturation equation for the oil phase in terms of the new
2. Mathematical formulation variable u can be written as:
[ ]
𝜕𝑢 𝜕 𝑘𝑟𝑛𝑤 𝑘(𝑥) 𝑘𝑟𝑛𝑤 𝑘𝑟𝑤 𝜕𝑝𝑐
𝜙(𝑥) + 𝑞− = 0. (2.6)
This section deals with the mathematical formulation of a specific 𝜕𝑡 𝜕𝑥 𝑘𝑟𝑛𝑤 + 𝑀𝑘𝑟𝑤 𝜇𝑤 𝑘𝑟𝑛𝑤 + 𝑀𝑘𝑟𝑤 𝜕𝑥
two-phase flow model. The model is defined on a highly heterogeneous
Note that in the Leverett model of capillary pressure (Eq. (2.5)) an ex-
porous structure. The highly heterogeneous porous medium consists of
plicit relation between the capillary pressure, porosity and absolute per-
homogeneous layers with different porosity and absolute permeability.
meability of the medium is prescribed. This expresses the saturation
These layers are perpendicular to the flow direction, leading to a one-
Eq. (2.6) precisely in terms of these discontinuous quantities of inter-
dimensional flow and a medium with different coarse and fine layers
est i.e., porosity and absolute permeability, which in turn enable us to
(refer Fig. 1). Here we assume that the porosity 𝜙(x) as well as absolute
effectively capture the impact of these quantities in the formal deriva-
(intrinsic) permeability k(x) are varying between the layers. Note that
tion of the effective equations.
the porosity and intrinsic permeability are assumed to be constant in-
Introducing the non-dimensional variables as:
side each layer. Hence here we analyze the one-dimensional flow of two
𝑥 𝑡𝑞 𝑘
incompressible and immiscible phases through a medium with homoge- 𝑥̂ ∶= , 𝑡̂ ∶= , 𝑘̂ ∶= , (2.7)
neous layers of any porosity and intrinsic permeability zones. Through- 𝐿 𝐿 𝐾
out this investigation water is being considered as the wetting fluid and where L is a macroscopic characteristic length scale and K is the charac-
oil as the non-wetting fluid. teristic absolute permeability. For notational convenience we drop the
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

Fig. 1. Periodically varying porous medium with different porosity and permeability layers. Large scale and small scale are highlighted and associated with the
length scales L and l, respectively.

hats in the rest of this study. In this way we bring Eq. (2.6) to a dimen- media equation with regularity away from the trap location. In addition,
sionless form as: they also established that the solution is bounded. In the present arti-
𝜕𝑢 𝜕𝐹 cle, we assume similar properties for the oil saturation in the case of
𝜙(𝑥) + = 0, (2.8)
𝜕𝑡 𝜕𝑥 arbitrarily distributed multiple traps. In particular, 0 ≤ u ≤ 1.
We mention that Bourgeat and Panfilov (1998) investigates a similar
with
problem as considered here, although using a phase pressure-saturation
𝜕
𝐹 = 𝑓 (𝑢) − 𝑁𝑐 𝑘(𝑥)𝑘𝑟𝑤 (𝑢)𝑓 (𝑢) 𝑝 (𝑥, 𝑢). (2.9) formulation for the two-phase flow and not the fractional flow formula-
𝜕𝑥 𝑐
tion as applied here. Where we apply continuity of flux and capillary
Here ( )
pressure if 𝑢(𝑥𝑖+ , 𝑡) > 0 , they apply continuity of the fluxes in each
𝑘𝑟𝑛𝑤 phase and of the pressure in each phase (and hence also of the capil-
𝑓 (𝑢) = , (2.10)
𝑘𝑟𝑛𝑤 + 𝑀𝑘𝑟𝑤 lary pressure) at the discontinuity. In this present article, we investigate
the trapping phenomena in highly heterogeneous media. In this context,
is the fractional flow function, and
it may be noted that the capillary entry pressure has a significant impact

𝜙(𝑥) on the trapping phenomena (van Duijn et al., 2007). Hence to incorpo-
𝑝𝑐 (𝑥, 𝑢) = 𝐽 (𝑢), (2.11) rate the effect of the capillary entry pressure in this current setting, it is
𝑘(𝑥)
necessary to assume that the capillary pressure is continuous only when
is the dimensionless capillary pressure. The capillary number Nc and the oil phase is present on the either side of the discontinuity. Otherwise, a
viscosity ratio M are defined as follows jump in the capillary pressure is noticed at the interface. Moreover, in
√ their non-dimensionalization they open for the permeability ratio and
𝜎 𝐾 𝜇
𝑁𝑐 = , 𝑀= 𝑜. (2.12) the ratio of characteristic capillary pressure values to also depend on
𝜇𝑤 𝑞𝐿 𝜇𝑤 the scale separation 𝜖 = 𝐿𝑙 , which is not considered here.
One must note that based on the specific application, the value of the
capillary number Nc may vary considerably. The jumps in the poros- 3. Homogenization technique for periodic layers
ity 𝜙 and intrinsic permeability k produce discontinuities in the cap-
illary pressure. Hence across the interfaces, the spatial derivatives in Here in this section we derive effective/upscaled two-phase flow
Eqs. (2.8) and (2.9) need a special attention. To deal with this complex- equations for periodically distributed layers corresponding to the model
ity, Duijn et al. (1995), Bertsch et al. (2003) have suggested to consider equations introduced in the earlier section. The present study includes
Eq. (2.8) only in the homogeneous layers with constant 𝜙 and k, and to the differences in entry pressure corresponding to the different homo-
enforce matching conditions across the discontinuities. Specifically, we geneous layers, which is eventually the main reason for trapping at the
assume that the discontinuity (interface) is located at 𝑥 = 𝑥𝑖 . For given micro-scale.
𝜙 and k, the matching conditions read, for all t > 0,

[𝐹 (𝑡)]𝑥𝑖 = 0, (2.13) 3.1. Homogenization procedure

In order to derive an effective/upscaled model we make use of the


𝑢(𝑥𝑖+ , 𝑡)[𝑝𝑐 (𝑡)]𝑥𝑖 = 0, [𝑝𝑐 (𝑡)]𝑥𝑖 ≥ 0, (2.14)
homogenization technique (Hornung, 1997). We refer to the model
where [𝐹 (𝑡)]𝑥𝑖 = 𝐹 (𝑥𝑖+ , 𝑡) − 𝐹 (𝑥𝑖− , 𝑡) is the jump in the flux at the in- equations as described in Section 2. The highly heterogeneous nature
terface xi and [𝑝𝑐 (𝑡)]𝑥𝑖 likewise. Condition (2.13) expresses the flux of the medium considered stems from the varying porosity (𝜙j ) and ab-
continuity. From condition (2.14), the capillary pressure is continuous solute permeability (kj ), as depicted in Fig. 1. Each of the homogeneous
( )
[𝑝𝑐 (𝑡)]𝑥𝑖 = 0 only if the oil phase is present on the both sides of the layers is of thickness l (micro-scale reference length), which is very small
( ) compared to the macro-scale reference length L. Naturally this leads to a
interface 𝑢(𝑥𝑖+ , 𝑡) > 0 . On the other hand, the capillary pressure is dis-
( ) small expansion parameter 𝜖 = 𝐿𝑙 << 1. One may note that the trapping
continuous across the interface [𝑝𝑐 (𝑡)]𝑥𝑖 ≠ 0 if the oil phase disappears effects are prominent in all the transitions from a high permeability to
for x > xi i.e., 𝑢(𝑥𝑖+ , 𝑡) = 0. To complete the model, suitable initial and a low permeability region.
boundary conditions are also needed. As the upscaling procedure is in- One may note the presence of two dimensionless numbers, M and
dependent of these, they will be specified in Section 5. It may be noted Nc in the dimensionless Eq. (2.8), through the flux (2.9). van Duijn
that Bertsch et al. (2003) studied a simplified version of the problem et al. (2002) analyzed the problem separately for the two specific cases
considered in this article with only one trap location. They have shown of small and moderate capillary number Nc , while keeping the viscosity
the existence and uniqueness of a solution satisfying the standard porous ratio 𝑀 = 1. In their investigation convergence results for balance case
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

Fig. 2. Periodic distribution of porosity and


absolute permeability in a cell with single dis-
continuity.

i.e., 𝑁𝑐 = (𝜖) are shown using numerical examples. Whereas, these nu- (a) If 𝑘1𝑖 > 𝑘2𝑖 then
merical evidences are really difficult to attain in the case of the capil- √ √
lary limit i.e., 𝑁𝑐 = (1), until the capillary effects are dominant. This √ 1 √ 2
√𝜙 √
√ 𝑖 𝐽 (𝑢∗ ) = √ 𝜙𝑖 𝐽 (0),
restriction is due to the micro-scale assumption of continuity of the (𝜖) 2 𝑖 +1
(3.4a)
𝑘1𝑖 𝑘2𝑖
term in the asymptotic expansion of the oil saturation. Later with a
different micro-scale assumption concerning the continuity of the (𝜖)
(b) If 𝑘1𝑖 < 𝑘2𝑖 then
term in the asymptotic expansion of the capillary pressure, van Duijn
√ √
et al. (2007) investigated the capillary limit case and consequently de- √ 1 √ 2
√𝜙 √
rived the macro-scale equations which are also suitable for flow regimes √ 𝑖 𝐽 (0) = √ 𝜙𝑖 𝐽 (𝑢∗ ). (3.4b)
2𝑖+1
that are not dominated by capillary effects. Here we follow a similar ap- 𝑘𝑖
1 𝑘2𝑖
proach as described in van Duijn et al. (2007). In this present investiga-
tion we look only at the capillary limit case and derive the corresponding At 𝑥 = (2𝑖 + 2)𝜖 we impose a similar matching condition as that of (3.3)
effective equations for a highly heterogeneous porous structure. depending on the criterion, either 𝑘2𝑖 > 𝑘1𝑖+1 or 𝑘2𝑖 < 𝑘1𝑖+1 as:
One can now proceed with the derivation of the upscaled (effective) (a) For 𝑘2𝑖 > 𝑘1𝑖+1 :
equations for 𝜖↘0, using the well known two-scale or homogenization
technique. Here we assume that all the quantities depend on the two spa- if 𝑢((2𝑖 + 2)𝜖 − ) < 𝑢∗2𝑖+2 , then 𝑢((2𝑖 + 2)𝜖 + ) = 0,
√ √
tial variables, x - the macro-scale or slow variable, and 𝑦 = 𝑥𝜖 - the micro- 𝜙1𝑖+1
𝜙2𝑖
scale or fast variable. Based on these two variables, one may expand all if 𝑢((2𝑖 + 2)𝜖 − ) ≥ 𝑢∗2𝑖+2 , then 𝐽 ( 𝑢 ((2 𝑖 + 2) 𝜖 − ))=
1 𝐽 (𝑢((2𝑖 + 2)𝜖 )).
+
2 𝑘𝑖 𝑘𝑖+1
the quantities involved in terms of 𝜖 and equate the similar order terms
in 𝜖. Here we proceed with the derivation of the effective/upscaled equa- (3.5a)
tions with two sub-cases, namely, (i) single discontinuity within a cell,
(b) For 𝑘2𝑖 < 𝑘1𝑖+1 :
and (ii) multiple discontinuities within a cell. Details are presented in
the forthcoming subsections. √ √
𝜙1𝑖+1 𝜙2𝑖
3.2. Single discontinuity within a cell if 𝑢((2𝑖 + 2)𝜖 + ) ≥ 𝑢∗2𝑖+2 , then 𝐽 (𝑢((2𝑖 + 2)𝜖 + ))= 𝐽 (𝑢((2𝑖 + 2)𝜖 − )),
𝑘1𝑖+1 𝑘2𝑖

if 𝑢((2𝑖 + 2)𝜖 + ) < 𝑢∗2𝑖+2 , then 𝑢((2𝑖 + 2)𝜖 − ) = 0,


For notational convenience we assume that the periodic structures
{ } (3.5b)
with traps are located at the points 𝑥𝑖 = 𝜖𝑦𝑖 = 𝜖𝑖 ∶ 𝑖 ∈ ℤ (refer Fig. 2).
𝜖 𝜖
The corresponding permeability k (x) and the porosity 𝜙 (x) are defined where 𝑢∗2𝑖+2 is the threshold (entry point) saturation for fluid to flow
by 𝑘𝜖 (𝑥) = 𝑘(𝑥∕𝜖) and 𝜙𝜖 (𝑥) = 𝜙(𝑥∕𝜖), where through the interface 𝑥 = (2𝑖 + 2)𝜖. This is similarly defined as 𝑢∗2𝑖+1 in
{ 1 ( )
𝑘𝑖 , on 𝑦2𝑖 , 𝑦2𝑖+1 , Eq. (3.4).
𝑘= ( ) (3.1)
𝑘𝑖 , on 𝑦2𝑖+1 , 𝑦2𝑖+2 ,
2 We now replace 𝜙 and k by 𝜙𝜖 and k𝜖 in (2.8) and (2.9). Now

{ 1 ( ) Eqs. (2.8) and (2.9) hold in the domain Σ𝜖 = ℝ ⧵ 𝑇𝜖 , where 𝑇𝜖 = 𝜖 𝑖∈ℤ 𝑖.
𝜙𝑖 , on 𝑦2𝑖 , 𝑦2𝑖+1 , 𝜖
𝜙= ( ) (3.2) Let u be a solution of (2.8) satisfying the matching conditions (3.3) and
𝜙𝑖 , on 𝑦2𝑖+1 , 𝑦2𝑖+2 .
2
(3.5). We consider the following two-scale asymptotic expansion:
We distinguish two kinds of matching conditions: one going from 𝑘1𝑖 to
𝑢𝜖 (𝑥, 𝑡) = 𝑢0 (𝑥, 𝑦, 𝑡) + 𝜖𝑢1 (𝑥, 𝑦, 𝑡) + 𝜖 2 𝑢2 (𝑥, 𝑦, 𝑡) + ⋯ , (3.6)
𝑘2𝑖 and the other from 𝑘2𝑖 to 𝑘1𝑖+1 .
At 𝑥 = (2𝑖 + 1)𝜖 we impose the following: where the functions uj are periodic in 𝑦 = 𝑥∕𝜖, the fast scale. This as-
(a) For 𝑘1𝑖 > 𝑘2𝑖 : sumption is the counterpart of the stationarity condition in a random
medium for the two-scale asymptotic expansion method of homogeniza-
if 𝑢((2𝑖 + 1)𝜖 − ) < 𝑢∗2𝑖+1 , then 𝑢((2𝑖 + 1)𝜖 + ) = 0,
√ √ tion. Since the microstructure is periodic here, the property is stronger
𝜙1𝑖 𝜙2𝑖 i.e, each uj is y-periodic (Auriault et al., 2009).
if 𝑢((2𝑖 + 1)𝜖 − ) ≥ 𝑢∗2𝑖+1 , then 𝐽 ( 𝑢 ((2 𝑖 + 1) 𝜖 − ))=
2 𝐽 (𝑢((2𝑖 + 1)𝜖 )).
+
1 𝑘𝑖 𝑘𝑖
It is noted that the continuity of the first order term in the asymp-
(3.3a) totic expansion of oil saturation Eq. (3.6) at the interface between two
(b) For 𝑘1𝑖 < 𝑘2𝑖 : homogeneous layers used in van Duijn et al. (2002) leads to enhanced
diffusivity in the resulting upscaled equation (van Duijn et al., 2007).
√ √ Moreover, this also produces discontinuities of order 𝜖 in the capillary
𝜙2𝑖 𝜙1𝑖
if 𝑢((2𝑖 + 1)𝜖 + ) ≥ 𝑢∗2𝑖+1 , then 𝐽 (𝑢((2𝑖 + 1)𝜖 + ))= 1 𝐽 (𝑢((2𝑖 + 1)𝜖 )),

pressure. However, although these are at the micro-scale, they result in
𝑘2𝑖 𝑘𝑖
if 𝑢((2𝑖 + 1)𝜖 + ) < 𝑢∗2𝑖+1 , then 𝑢((2𝑖 + 1)𝜖 − ) = 0, a significant effect at the macro-scale (van Duijn et al., 2007). These
observations prompts us to adopt the approach proposed in van Duijn
(3.3b)
et al. (2007) i.e., the continuity of first order terms in the expansion of
where 𝑢((2𝑖 + 1)𝜖 − ) = lim𝑥→(2𝑖+1)𝜖 − 𝑢(𝑥), 𝑢((2𝑖 + 1)𝜖 + ) = lim𝑥→(2𝑖+1)𝜖 + 𝑢(𝑥), the capillary pressure. The averaged solution of the micro-scale model
and 𝑢∗2𝑖+1 is the threshold (entry point) saturation for fluid to flow is compared with the effective solutions developed in van Duijn et al.
through the interface 𝑥 = (2𝑖 + 1)𝜖, which has to be exceeded before the (2002, 2007). The averaged solution is found to be better represented
non-wetting/oil phase can enter the fine material (low permeability re- by the effective solution developed in van Duijn et al. (2007). To pro-
gion) from the coarse material (high permeability region). This thresh- ceed with the derivation of the upscaled model, one may notice that the
old saturation is uniquely defined as: capillary pressure can be expanded asymptotically as:
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716


𝜙(𝑥) which holds for every 𝑥, 𝑦 ∈ ℝ and for all t > 0. This is due to the conti-
𝜖
𝑝𝑐 (𝑢 ) = 𝐽 (𝑢𝜖 ) nuity of the flux. Moreover we also have the following:
𝑘(𝑥)
{ }
√ ( ) 𝜕𝑝0𝑐
𝜕𝐹 0 𝜕𝐹 1 𝜕
𝜙(𝑥) 𝜕𝐽 (𝑢) || 𝜕 2 𝐽 (𝑢) || (𝜖 ) ∶ 0 =
−1
+ = −𝜖 𝑁𝑐 𝑘𝜆(𝑢 )
−1 0
= 𝐽 (𝑢0 ) + Δ𝑢 + | (Δ𝑢 ) 2
+ ⋯ , (3.7) 𝜕𝑥 𝜕𝑦 𝜕𝑥 𝜕𝑦
𝑘(𝑥) 𝜕𝑢 ||𝑢0 𝜕𝑢2 ||𝑢0 [ { ( ) }]
𝜕 𝜕𝑝𝑐
1 𝜕𝑝0𝑐 𝜕𝑝0𝑐
with Δ𝑢 = 𝜖𝑢1 + 𝜖 2 𝑢2 + ⋯. This becomes + 𝑓 (𝑢 ) − 𝑁𝑐 𝑘 𝜆(𝑢 )
0 0
+ + 𝜆 (𝑢 )𝑢
′ 0 1
,
𝜕𝑦 𝜕𝑦 𝜕𝑥 𝜕𝑦

𝜖 𝜙(𝑥) ( 0 ) (3.19)
𝑝𝑐 (𝑢 ) = 𝐽 (𝑢 ) + 𝜖𝑢1 𝐽 ′ (𝑢0 ) + (𝜖 2 ) . (3.8)
𝑘(𝑥)

The first two terms in the pc -expansion are 𝜕𝑢0 𝜕𝐹 1 𝜕𝐹 2 𝜕𝑢0


√ (𝜖 0 ) ∶ 0 = 𝜙(𝑦) + + = 𝜙(𝑦)
𝜕𝑡 𝜕𝑥 𝜕𝑦 𝜕𝑡
𝜙(𝑥) [ { ( ) }]
𝑝𝑐 =
0
𝐽 (𝑢0 ), (3.9) 𝜕𝑝𝑐 1 𝜕𝑝0 𝜕𝑝0
𝑘(𝑥) 𝜕
+ 𝑓 (𝑢0 ) − 𝑁𝑐 𝑘 𝜆(𝑢0 ) + 𝑐 + 𝜆′ (𝑢0 )𝑢1 𝑐
√ 𝜕𝑥 𝜕𝑦 𝜕𝑥 𝜕𝑦
𝜙(𝑥) 1 ′ 0 { [ ( )
𝑝𝑐 =
1
𝑢 𝐽 (𝑢 ). (3.10) 𝜕 𝜕𝑝𝑐2 𝜕𝑝 1
𝑘(𝑥) + 𝑓 ′ (𝑢0 )𝑢1 − 𝑁𝑐 𝑘 𝜆(𝑢0 ) + 𝑐
𝜕𝑦 𝜕𝑦 𝜕𝑥
It may be noted that the expansion (3.7) is admissible only when J(u) ( ) ( ( 1 )2 ) ]}
and its higher order partial derivatives exist and are continuous at u0 . 𝜕𝑝𝑐1 𝜕𝑝 0
𝑢 𝜕𝑝0𝑐
+𝜆′ (𝑢0 )𝑢1 + 𝑐 + 𝜆′′ (𝑢0 ) + 𝜆′ (𝑢0 )𝑢2 .
Substituting these expansions into (2.8) and equating the terms of the 𝜕𝑦 𝜕𝑥 2 𝜕𝑦
same order of 𝜖 gives equations for u0 , u1 , … at different orders of 𝜖. (3.20)
The details follow. We define the effective oil saturation as
𝑦 We now look for 𝑦−periodic solutions of (3.18) subject to the matching
∫𝑦 2𝑖+2 𝜙(𝑦)𝑢0 (𝑥, 𝑦, 𝑡)𝑑𝑦
𝑈 (𝑥, 𝑡) = 2𝑖
, (3.11) conditions (3.3) and (3.5), with x and t as given parameters. Our aim is
𝑦
∫𝑦 2𝑖+2 𝜙(𝑦)𝑑𝑦 now to prove that 𝐹 0 = 0. Let F0 < 0. Define
2𝑖

which is the weak limit of u𝜖 , and where u0 (x, y, t) is the leading order 𝑤(𝑦) ∶= 𝐽 (𝑢0 (𝑦)), 𝜆(𝑤) ∶= 𝑘𝑟𝑤 (𝐽 −1 (𝑤))𝑓 (𝐽 −1 (𝑤)) (3.21)
approximation to u𝜖 in some norm. Schweizer (2008) proved the con- and
vergence of the homogenization procedure for nonlinear flow problems 𝑤
in inhomogeneous geometries, which is similar to the setting considered Λ(𝑤) = 𝜆(𝑠)𝑑𝑠. (3.22)
∫𝐽 (0)
in this present article.
Capillary limit: 𝑁𝑐 = (1). Introducing the oil flux as Note that the last function is strictly increasing and from Eq. (3.18), we
have
𝜕𝑝𝜖𝑐 √ 𝜕𝑤
𝐹 𝜖 = 𝑓 (𝑢𝜖 ) − 𝑁𝑐 𝑘𝜖 (𝑥)𝜆(𝑢𝜖 ) , (3.12) 𝜆(𝑤) 𝜙𝑘 =−
𝐹0
=∶ 𝐹 > 0. (3.23)
𝜕𝑥 𝜕𝑦 𝑁𝑐
where
√ Hence, for 𝑦2𝑖 < 𝑦 < 𝑦2𝑖+1 ,
𝜙𝜖 (𝑥) 𝐹
𝜆(𝑢𝜖 ) = 𝑘𝑟𝑤 (𝑢𝜖 )𝑓 (𝑢𝜖 ) and 𝑝𝜖𝑐 = 𝐽 (𝑢𝜖 ), (3.13) Λ(𝑤(𝑦2𝑖+1− )) − Λ(𝑤(𝑦2𝑖+ )) = √ , (3.24)
𝑘𝜖 (𝑥)
𝜙1𝑖 𝑘1𝑖
Eq. (2.8) becomes giving
𝜖 𝜕𝑢𝜖 𝜕𝐹 𝜖 𝜖 𝐹 𝜖
𝜙 (𝑥) + =0 in Σ × (0, ∞). (3.14) 𝑤(𝑦2𝑖+1− ) ≥ 𝑤(𝑦2𝑖+ ) + √ , (3.25)
𝜕𝑡 𝜕𝑥 ‖𝜆‖∞
𝜙𝑖 𝑘𝑖
1 1
We now apply expansions (3.6) and (3.8) to F𝜖 , which gives
𝜕𝑝0 where
{ ‖𝜆‖∞ denotes } the standard sup-norm defined by ‖𝜆‖∞ =
𝐹 𝜖 = −𝜖 −1 𝑁𝑐 𝑘𝜆(𝑢0 ) 𝑐 + 𝑓 (𝑢0 )
𝜕𝑦 sup |𝜆(𝑢)| ∶ 0 < 𝑢 < 1 . Similarly, for 𝑦2𝑖+1 < 𝑦 < 𝑦2𝑖+2 ,
{ ( ) }
𝜕𝑝1𝑐 𝜕𝑝0 𝜕𝑝0
−𝑁𝑐 𝑘 𝜆(𝑢0 ) + 𝑐 + 𝜆′ (𝑢0 )𝑢1 𝑐 𝐹 𝜖
𝜕𝑦 𝜕𝑥 𝜕𝑦 𝑤(𝑦2𝑖+2− ) ≥ 𝑤(𝑦2𝑖+1+ ) + √ , (3.26)
‖𝜆‖∞
{ [ ( ) ( ) 𝜙2𝑖 𝑘2𝑖
𝜕𝑝2𝑐 𝜕𝑝1 𝜕𝑝1𝑐 𝜕𝑝0
+ 𝜖 𝑓 ′ (𝑢0 )𝑢1 − 𝑁𝑐 𝑘 𝜆(𝑢0 ) + 𝑐 + 𝜆′ (𝑢0 )𝑢1 + 𝑐 where y2i , 𝑦2𝑖+1 and 𝑦2𝑖+2 are the location of the interfaces at the micro-
𝜕𝑦 𝜕𝑥 𝜕𝑦 𝜕𝑥
( ) ]} scale.
( 1 )2
𝑢 𝜕𝑝𝑐 0 Let us first consider the case 𝑘1𝑖 > 𝑘2𝑖 . Assuming 𝑤(𝑦2𝑖+1− ) ≤ 𝐽 (𝑢∗2𝑖+1 )
+ 𝜆′′ (𝑢0 ) + 𝜆′ (𝑢0 )𝑢2 + (𝜖 2 ). (3.15) and using the strict monotonicity of w, we get 𝑤(𝑦2𝑖+ ) < 𝐽 (𝑢∗2𝑖+1 ), giving
2 𝜕𝑦
𝑤(𝑦2𝑖− ) = 𝐽 (0). This eventually contradicts the monotonic behaviour of
Equivalently we define w in (𝑦2𝑖−1 , 𝑦2𝑖 ). Now assuming 𝑤(𝑦2𝑖+1− ) > 𝐽 (𝑢∗2𝑖+1 ) and the continuity

𝐹 𝜖 ∶= 𝜖 −1 𝐹 0 + 𝐹 1 + 𝜖𝐹 2 + (𝜖 2 ). (3.16) of capillary pressure implies 𝑤(𝑦2𝑖+1+ ) =
𝜙1𝑖 𝑘2𝑖
2 1 𝑤(𝑦2𝑖+1− ). Consequently
𝜙𝑖 𝑘𝑖
Using the expansion (3.15) in the Eq. (3.14) leads to the following equa- from the relations (3.25) and (3.26), we have
tions √
( ) √ 1 2
√𝜙 𝑘 𝐹 𝜖
𝜕𝐹 0 𝜕 𝜕𝑝0𝑐 𝑤(𝑦2𝑖+2− ) ≥ √ 𝑖 𝑖 𝑤(𝑦2𝑖+1− ) + √
(𝜖 ) ∶ 0 =
−2
= −𝑁𝑐 𝑘𝜆(𝑢 )
0
; (3.17) 𝜙2𝑖 𝑘1𝑖 ‖𝜆‖∞
𝜕𝑦 𝜕𝑦 𝜕𝑦 𝜙2 𝑘2 𝑖 𝑖
√ √
now incorporating the continuity of F𝜖 , we have √ 1 2 ⎛ ⎞
√𝜙 𝑘 𝐹 𝜖 ⎜ 𝑘2𝑖 1 ⎟
≥ √ 𝑖 𝑖
𝑤(𝑦2𝑖 + ) + √ + √ ⎟.
‖𝜆‖∞ ⎜⎜ 𝑘1
𝜕𝑝0𝑐 (3.27)
−𝑁𝑐 𝑘𝜆(𝑢0 ) = 𝐹 0 = 𝐹 0 (𝑥, 𝑡), (3.18) 𝜙2𝑖 𝑘1𝑖 𝜙2𝑖 𝑖 𝑘2𝑖 ⎟
𝜕𝑦 ⎝ ⎠
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

Repetition of this reasoning backwards in i shows that w will drop below Finally we use the 𝜖 0 -Eq. (3.20). Since F2 is continuous in the fast scale,
𝐽 (𝑢∗2𝑖+1 ) at the right half of a certain transition face, which again leads we find the averaged oil saturation from Eq. (3.11) as
to a contradiction. Hence F0 ≥ 0. Similar arguments imply F0 ≤ 0, and 𝑦
∫𝑦 2𝑖+2 𝜙(𝑦)𝑢0 (𝑥, 𝑦, 𝑡)𝑑𝑦 𝜙1 𝐶𝑖 + 𝜙2𝑖 𝐶 𝑖
hence 𝐹 0 = 0. 𝑈 (𝑥, 𝑡) = 2𝑖 𝑦 = 𝑖 , (3.35)
∫𝑦 2𝑖+2
𝜙(𝑦)𝑑𝑦 𝜙1𝑖 + 𝜙2𝑖
Now consider the case 𝑘1𝑖 < 𝑘2𝑖 . Note that in this analysis, flow direc- 2𝑖

tion is from left to right. Here in this case flow is from a low permeable and the effective equation has the following form
medium to a relatively high permeable medium; i.e., capillary pressure ( 1 )
𝜙𝑖 + 𝜙2𝑖 𝜕𝑈 [ }]
𝜕 𝜕 {
in the right portion of the transition face is lower than the same in the +  (𝑈 ) − 𝑁𝑐 Λ(𝑈 )  (𝑈 ) = 0. (3.36)
2 𝜕𝑡 𝜕𝑥 𝜕𝑥 𝑐
left side, if the non-wetting phase is present on both sides. Whereas in
this scenario zero non-wetting phase saturation in the right part even- The effective saturation is expressed in terms of Ci and 𝐶 𝑖 . For 𝐶 𝑖 > 0,
tually implies zero non-wetting phase saturation in the left part as well. we end up with the relation given in (3.28). This gives
Hence in this case it is trivial that 𝐹 0 = 0. This conclusion allows us to ⎧ √ √
⎛√ √
√ 𝜙1 √ 𝑘2 ⎞⎫
solve Eq. (3.18) with the matching conditions (3.3) and (3.5). 1 ⎪ 1 2 −1 ⎜√ 𝑖 √ 𝑖 ⎟⎪
𝑈 = 𝐺(𝐶𝑖 ) = 𝜙 𝐶
⎨ 𝑖 𝑖 + 𝜙𝑖 𝐽 𝐽 ( 𝐶 𝑖 ⎬.
) (3.37)
It may be noted that at the interface between the two homogeneous 𝜙1𝑖 + 𝜙2𝑖 ⎪ ⎜ 𝑘1 𝜙2 ⎟⎪
⎩ ⎝ 𝑖 𝑖 ⎠⎭
layers, fluxes are continuous at all scales. In addition to this, the capil-
lary pressure condition is also imposed at the interface, and the period- ( )
𝜙1𝑖 𝑢∗2𝑖+1
icity is assumed at the end points of the cells. These are satisfied only if Notice that G defined above maps (0,1) onto , 1 . As J is strictly
𝜙1𝑖 +𝜙2𝑖
u0 is the lowest term in the asymptotic expansion (3.6), and is constant increasing, the function G shares the same property and can therefore
at the micro-scale in the homogeneous layers. In particular, Ci and 𝐶 𝑖 be inverted. Hence one can write Ci in terms of U as
are related by:
𝐶𝑖 = 𝐺−1 (𝑈 ). (3.38)
√ √
⎛√ √
√ 𝜙1 √ 𝑘2 ⎞
−1 ⎜√ 𝑖 √ 𝑖 Inserting the above into (3.28) one can obtain an equivalent expression
𝐶𝑖 = 𝐽 𝐽 (𝐶𝑖 )⎟, (3.28)
⎜ 𝑘1 𝜙2𝑖 ⎟ for 𝐶 𝑖 . With the above expressions one can identify the upscaled frac-
⎝ 𝑖 ⎠
tional flow as
where Ci and 𝐶 𝑖 are the values of u0 at the micro-scale in the left and ⎧ 𝜙1 𝑢∗
⎪0, for 0 ≤ 𝑈 ≤ 𝑖1 2𝑖+1 ,
right part of the interface located at 𝑦 = 2𝑖 + 1 respectively. ⎪ 𝜙 + 𝜙2
⎪( )/( ) 𝑖 𝑖
 (𝑈 ) = ⎨ 𝜙1 𝑢∗
Now whenever 𝐶 𝑖 > 0, we find 1 1
⎪ 𝑘1 𝑘 (𝐶 ) + 𝑘2 𝑘 (𝐶 )
1
+ 2 1 𝑖 2𝑖+1
, for 1 2 < 𝑈 < 1,
𝑘1 𝑘𝑟𝑤 (𝐶𝑖 )𝑓 (𝐶𝑖 ) 𝑘 𝑘𝑟𝑤 (𝐶𝑖 )𝑓 (𝐶𝑖 ) 𝜙 +𝜙
{ ⎪ 𝑖 𝑟𝑤 𝑖 𝑖 𝑟𝑤 𝑖 𝑖 𝑖 𝑖 𝑖
𝐶𝑖 > 𝑢∗2𝑖+1 , for 𝑦2𝑖 < 𝑦 < 𝑦2𝑖+1 , ⎪1, for 𝑈 = 1.

𝑢0 (𝑦) = (3.29)
𝐶𝑖, for 𝑦2𝑖+1 < 𝑦 < 𝑦2𝑖+2 , (3.39)

if Ci exceeds the threshold saturation 𝑢∗2𝑖+1 and the non-wetting phase This function can be interpreted as the weighted harmonic mean of
can infiltrate into the immediate fine material, or the micro-scale fractional flow. The weight functions, 1∕𝑘1𝑖 𝑘𝑟𝑤 (𝐶𝑖 ) and
{ 1∕𝑘2𝑖 𝑘𝑟𝑤 (𝐶𝑖 ), depend on the water relative permeability, which in turn
𝐶𝑖 ≤ 𝑢∗2𝑖+1 , for 𝑦2𝑖 < 𝑦 < 𝑦2𝑖+1 ,
𝑢0 (𝑦) = (3.30) depends on Ci and 𝐶𝑖 .
0, for 𝑦2𝑖+1 < 𝑦 < 𝑦2𝑖+2 ,
The upscaled Λ term is the harmonic average of the terms 𝑘1𝑖 𝜆(𝐶𝑖 )
if there is no non-wetting phase flow from the high into the low perme- and 𝑘2𝑖 𝜆(𝐶𝑖 ),
able medium. Note that in the first case the capillary pressure is contin- ⎧ 𝜙1𝑖 𝑢∗2𝑖+1
uous at the transition face. ⎪0 , for 0 ≤ 𝑈 ≤ ,
⎪ 𝜙1𝑖 + 𝜙2𝑖
Now consider the 𝜖 −1 -Eq. (3.19). Since 𝐹 0 = 0 and the flux is contin- ⎪ /( )
Λ(𝑈 ) = ⎨ 𝜙1𝑖 𝑢∗2𝑖+1 (3.40)
uous, we obtain 𝐹 1 = 𝐹 1 (𝑥, 𝑡). Using (3.29) and (3.30), the local form of 1 1
, for < 𝑈 < 1,
⎪2 +
F1 is ⎪ 𝑘𝑖 𝜆(𝐶𝑖 ) 𝑘2𝑖 𝜆(𝐶𝑖 )
1 𝜙1𝑖 + 𝜙2𝑖
( ) ⎪0 , for 𝑈 = 1.
𝜕𝑝1𝑐 𝜕𝑝0𝑐 (𝐶𝑖 ) ⎩
𝐹 = 𝑓 (𝐶𝑖 ) − 𝑁𝑐 𝑘𝑖 𝜆(𝐶𝑖 )
1 1
+ , (3.31)
𝜕𝑦 𝜕𝑥 To express the upscaled capillary pressure we use (3.28). This means
that, whenever oil is present on both sides of the interface, the first term
for 𝑦2𝑖 < 𝑦 < 𝑦2𝑖+1 , and in the expansion (3.8) is constant in the entire micro-scale cell. Hence
( ) we get
⎧ 𝜕𝑝1𝑐 𝜕𝑝0𝑐 (𝐶 𝑖 ) √ √
⎪ 𝑓 ( 𝐶 ) − 𝑁 𝑘 2 𝜆(𝐶 ) + , for 𝐶𝑖 > 𝑢∗2𝑖+1 , √ 1 √ 2
𝐹1 = ⎨ 𝑖 𝑐 𝑖 𝑖
𝜕𝑦 𝜕𝑥 (3.32) √𝜙 √𝜙
𝑐 (𝑈 ) = √ 𝑖
𝐽 (𝐶𝑖 ) = √ 𝑖 𝐽 (𝐶 𝑖 ). (3.41)
⎪0 , for 𝐶𝑖 ≤ 𝑢∗2𝑖+1 ,
⎩ 𝑘𝑖
1 𝑘2𝑖

for 𝑦2𝑖+1 < 𝑦 < 𝑦2𝑖+2 . Clearly one has to consider only the nontrivial case To determine the effective diffusivity  we use (3.38), (3.40) and (3.41),
𝐶𝑖 > 𝑢∗2𝑖+1 . From (3.31) and (3.32) we obtain which implies

√ 1
√ 𝜙 𝐽 ′ (𝐺−1 (𝑈 ))
⎧ 𝑓 (𝐶𝑖 ) − 𝐹 1 𝜕𝑝0𝑐 (𝐶𝑖 ) (𝑈 ) = Λ(𝑈 )𝑐′ (𝑈 ), where 𝑐′ (𝑈 ) = √ 𝑖 . (3.42)
⎪ − ∶= 𝐵1 (𝑥, 𝑡), for 𝑦2𝑖 < 𝑦 < 𝑦2𝑖+1 , 𝑘1𝑖 𝐺′ (𝐺−1 (𝑈 ))
𝜕𝑝1𝑐 ⎪ 𝑁 𝑘1 𝜆(𝐶𝑖 ) 𝜕𝑥
=⎨ 𝑐 𝑖 (3.33)
𝜕𝑦 ⎪ 𝑓 (𝐶𝑖 ) − 𝐹 − 𝜕𝑝𝑐 (𝐶𝑖 ) ∶= 𝐵 (𝑥, 𝑡), for 𝑦
1 0
3.3. Multiple discontinuities within a cell
⎪ 𝑁 𝑘2 𝜆(𝐶 ) 2 2𝑖+1 < 𝑦 < 𝑦2𝑖+2 .
𝜕𝑥
⎩ 𝑐 𝑖 𝑖
Here we consider the upscaling over several discontinuities (refer
It is noted that after the integration, 𝐵1 + 𝐵2 = 0. Thus one can solve for Fig. 3). Although this procedure can be done for any n ≥ 2 sub-layers,
F1 and obtain for convenience we will present here considering 𝑛 = 4 (results for
1 1 any n are given in Appendix A). For notational convenience we as-
+ 𝜕𝑝0𝑐 (𝐶𝑖 ) 𝜕𝑝0𝑐 (𝐶𝑖 )
𝑘1𝑖 𝑘𝑟𝑤 (𝐶𝑖 ) 𝑘2𝑖 𝑘𝑟𝑤 (𝐶𝑖 ) + sume that the periodic structures with traps are located at the points
𝐹 =
1
− 𝑁𝑐 𝜕𝑥 𝜕𝑥
. (3.34) { }
1 1 1 1 𝑥𝑖 = 𝜖𝑦𝑖 = 𝜖𝑖 ∶ 𝑖 ∈ ℤ . The corresponding permeability k𝜖 (x) and the
+ +
𝑘1𝑖 𝑘𝑟𝑤 (𝐶𝑖 )𝑓 (𝐶𝑖 ) 𝑘2𝑖 𝑘𝑟𝑤 (𝐶𝑖 )𝑓 (𝐶𝑖 ) 𝑘1𝑖 𝜆(𝐶𝑖 ) 𝑘2𝑖 𝜆(𝐶𝑖 ) porosity 𝜙𝜖 (x) are defined by 𝑘𝜖 (𝑥) = 𝑘(𝑥∕𝜖) and 𝜙𝜖 (𝑥) = 𝜙(𝑥∕𝜖), where
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

Fig. 3. Periodic distribution of porosity and absolute permeability in a cell with multiple discontinuities.

fourth compartment/block of the cell i.e., 𝑦4𝑖+3 < 𝑦 < 𝑦4𝑖+4 , then
⎧ 1 ( )
⎪𝑘𝑖 ,on (𝑦4𝑖 , 𝑦4𝑖+1 , )
⎪𝑘2 ,on 𝑦4𝑖+1 , 𝑦4𝑖+2 , ⎧ 1
⎪𝐶𝑖 > 𝑢4𝑖+1 , for 𝑦4𝑖 < 𝑦 < 𝑦4𝑖+1 ,

𝑘 = ⎨ 𝑖3 ( ) (3.43)
⎪𝑘𝑖 ,on (𝑦4𝑖+2 , 𝑦4𝑖+3 ), ⎪𝐶 > 𝑢∗4𝑖+2 , for 𝑦4𝑖+1 < 𝑦 < 𝑦4𝑖+2 ,
2
𝑢0 (𝑦) = ⎨ 𝑖3 (3.47)
⎪𝑘𝑖 ,on 𝑦4𝑖+3 , 𝑦4𝑖+4 ,
4
⎪𝐶𝑖 ≤ 𝑢4𝑖+3 , for 𝑦4𝑖+2 < 𝑦 < 𝑦4𝑖+3 ,


⎪0, for 𝑦4𝑖+3 < 𝑦 < 𝑦4𝑖+4 .
⎧ 1 ( ) ⎩
⎪𝜙𝑖 ,on (𝑦4𝑖 , 𝑦4𝑖+1 , )
⎪𝜙 ,on 𝑦4𝑖+1 , 𝑦4𝑖+2 ,
2
Further, if 𝐶𝑖1 exceeds the threshold saturation 𝑢∗4𝑖+1 but 𝐶𝑖2 stays below
𝜙 = ⎨ 𝑖3 ( ) (3.44)
⎪𝜙𝑖 ,on (𝑦4𝑖+2 , 𝑦4𝑖+3 ), or equal to 𝑢∗4𝑖+2 and the non-wetting phase can not infiltrate into the
⎪𝜙𝑖 ,on 𝑦4𝑖+3 , 𝑦4𝑖+4 .
4
next fine material at the third compartment of the cell i.e., 𝑦4𝑖+2 < 𝑦 <

𝑦4𝑖+3 and consequently in the fourth compartment, then

We distinguish four kinds of matching conditions: one going from 𝑘1𝑖 to ⎧ 1


⎪𝐶𝑖 > 𝑢4𝑖+1 , for 𝑦4𝑖 < 𝑦 < 𝑦4𝑖+1 ,

𝑘2𝑖 , a second from 𝑘2𝑖 to 𝑘3𝑖 , another from 𝑘3𝑖 to 𝑘4𝑖 and finally from 𝑘4𝑖
to 𝑘1𝑖+1 . Note that we impose matching conditions similar to Eqs. (3.3) ⎪𝐶 ≤ 𝑢∗4𝑖+2 ,
2 for 𝑦4𝑖+1 < 𝑦 < 𝑦4𝑖+2 ,
𝑢0 (𝑦) = ⎨ 𝑖 (3.48)
and (3.5) at the locations 𝑥 = (4𝑖 + 1)𝜖, 𝑥 = (4𝑖 + 2)𝜖, 𝑥 = (4𝑖 + 3)𝜖 and ⎪0, for 𝑦4𝑖+2 < 𝑦 < 𝑦4𝑖+3 ,
𝑥 = (4𝑖 + 4)𝜖. It may be noted that the unit cell construction used in ⎪0, for 𝑦4𝑖+3 < 𝑦 < 𝑦4𝑖+4 .

Fig. 3 is referred to as the original or non-shifted approach in the rest of
this article. From Eq. (3.23) with similar arguments as before one can Finally, if 𝐶𝑖1 stays below or equal to 𝑢∗4𝑖+1 and the non-wetting phase
prove that 𝐹 0 = 0. This conclusion allows us to solve Eq. (3.18) with can not infiltrate into the next fine material at the second compartment
the matching conditions similar to Eqs. (3.3) and (3.5). In the case of a of the cell i.e., 𝑦4𝑖+1 < 𝑦 < 𝑦4𝑖+2 and also the consequent compartments,
single transition within a cell, fluid can pass through the cell whenever then we have
it exceeds the threshold saturation at that transition face. However, for
multiple transitions within a cell, fluid can pass through the cell only if it ⎧ 1
⎪𝐶𝑖 ≤ 𝑢4𝑖+1 , for 𝑦4𝑖 < 𝑦 < 𝑦4𝑖+1 ,

exceeds the maximum threshold saturation of the transition faces. Oth- ⎪0, for 𝑦4𝑖+1 < 𝑦 < 𝑦4𝑖+2 ,
erwise, fluid flow may cease at any of the transition faces having higher 𝑢0 (𝑦) = ⎨ (3.49)
threshold saturation. In this context, we find different possibilities as ⎪0, for 𝑦4𝑖+2 < 𝑦 < 𝑦4𝑖+3 ,
⎪0, for 𝑦4𝑖+3 < 𝑦 < 𝑦4𝑖+4 .
follows ⎩

⎧ 1 Note that in the first case the capillary pressure is continuous at all the
⎪𝐶𝑖 > 𝑢4𝑖+1 , for 𝑦4𝑖 < 𝑦 < 𝑦4𝑖+1 ,

transition faces, in the second case the capillary pressure is continuous
⎪𝐶 2 > 𝑢∗ , for 𝑦4𝑖+1 < 𝑦 < 𝑦4𝑖+2 , at the first two transition faces, further in the third case the capillary
𝑢0 (𝑦) = ⎨ 𝑖3 4𝑖+2 (3.45)
𝐶
⎪ 𝑖 > 𝑢 ∗
4𝑖+3
, for 𝑦4𝑖+2 < 𝑦 < 𝑦4𝑖+3 , pressure is only continuous at the first transition face, whereas in the
⎪𝐶𝑖4 , for 𝑦4𝑖+3 < 𝑦 < 𝑦4𝑖+4 , fourth case the capillary pressure is discontinuous at all the transition

faces.
where Here in this subsection, we will only point out the differences be-
tween the single discontinuity case and the multiple discontinuities case
√ √ for the effective equations. For all the necessary steps in the derivation
⎛√ √
√ 𝜙1 √ 𝑘2 ⎞
−1 ⎜√
𝑖 √ 𝑖
𝐶𝑖2 =𝐽 𝐽 (𝐶𝑖1 )⎟, (3.46a) we refer to the previous subsection and also to van Duijn et al. (2007).
⎜ 𝑘1 𝜙𝑖2 ⎟ For multiple transitions within a cell, we define the averaged oil satu-
⎝ 𝑖 ⎠
√ √ ration as
√ √
⎛√ 𝜙1 √ 𝑘3 ⎞
𝐶𝑖3 = 𝐽 −1 ⎜√ 𝑖 √ 𝑖 𝐽 (𝐶𝑖1 )⎟,
𝑦
(3.46b) ∫𝑦 4𝑖+4 𝜙(𝑦)𝑢0 (𝑥, 𝑦, 𝑡)𝑑𝑦 𝜙1𝑖 𝐶𝑖1 + 𝜙2𝑖 𝐶𝑖2 + 𝜙3𝑖 𝐶𝑖3 + 𝜙4𝑖 𝐶𝑖4
⎜ 𝑘1 𝜙𝑖3 ⎟ 𝑈 (𝑥, 𝑡) = 4𝑖
= , (3.50)
⎝ 𝑖 ⎠ 𝑦
∫𝑦 4𝑖+4 𝜙(𝑦)𝑑𝑦 𝜙1𝑖 + 𝜙2𝑖 + 𝜙3𝑖 + 𝜙4𝑖
√ √ 4𝑖
⎛√ √ 4
√𝜙 √ 𝑘
1 ⎞
𝐶𝑖4 = 𝐽 −1 ⎜√ 𝑖 √ 𝑖 𝐽 (𝐶𝑖1 )⎟, (3.46c) and the corresponding effective equation is given by:
⎜ 𝑘1 𝜙4𝑖 ⎟
⎝ 𝑖 ⎠
∑4
𝑗=1 𝜙𝑗𝑖 𝜕𝑈 𝜕
[
𝜕 { }]
+  (𝑈 ) − 𝑁𝑐 Λ(𝑈 )  (𝑈 ) = 0. (3.51)
if 𝐶𝑖1 , 𝐶𝑖2 , 𝐶𝑖3
exceeds the threshold saturations 𝑢∗4𝑖+1 , 𝑢∗4𝑖+2 ,
re- 𝑢∗4𝑖+3 4 𝜕𝑡 𝜕𝑥 𝜕𝑥 𝑐
spectively and the non-wetting phase can infiltrate into the next fine
material. On the other hand, if 𝐶𝑖1 , 𝐶𝑖2 exceeds the threshold saturations The effective saturation is expressed in terms of 𝐶𝑖𝑗 , 𝑗 = 1, 2, 3, 4.
𝑢∗4𝑖+1 , 𝑢∗4𝑖+2 respectively but 𝐶𝑖3 stays below or equal to 𝑢∗4𝑖+3 and the Fluid can pass through the cell only if 𝐶𝑖𝑗 > 0 ∀𝑗. Hence we consider
non-wetting phase can not infiltrate into the next fine material at the 𝐶𝑖𝑗 > 0 to derive the upscaled quantities. Next, for 𝐶𝑖𝑗 > 0, 𝑗 = 2, 3, 4, we
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

end up with the relations given in (3.46). This gives entire micro-scale cell. Hence we get
√ √ √ √
√ 1 √ 2 √ 3 √ 4
⎧ √ √ √𝜙 √𝜙 √𝜙 √𝜙
⎪ 1 1
⎛√ √
√ 𝜙1 √ 𝑘2 ⎞ 𝑐 (𝑈 ) = √ 1 𝐽 (𝐶𝑖1 ) = √ 2 𝐽 (𝐶𝑖2 ) = √ 3 𝐽 (𝐶𝑖3 ) = √ 4𝑖 𝐽 (𝐶𝑖4 ).
𝑖 𝑖 𝑖
(3.56)
1 2 −1 ⎜√ 𝑖 √ 𝑖 1 ⎟ 𝑘𝑖 𝑘𝑖 𝑘𝑖 𝑘𝑖
𝑈 = 𝐺(𝐶𝑖1 )
= ∑ ⎨ 𝜙 𝑖 𝐶 𝑖 + 𝜙 𝑖 𝐽 𝐽 ( 𝐶 𝑖 )
4 𝑗 ⎜ 𝑘1 𝜙2𝑖 ⎟
𝑗=1 𝜙𝑖 ⎪ ⎝ 𝑖 ⎠
⎩ To determine the effective diffusivity  we use (3.53), (3.55) and (3.56),
√ √ √ √ which implies
⎛√ √
√ 𝜙1 √ 𝑘3 ⎞ ⎛√ √
√ 𝜙1 √ 𝑘4 ⎞⎫ √
√ 1
3 −1 ⎜√ 𝑖 √ 𝑖 1 ⎟ 4 −1 ⎜√ 𝑖 √ 𝑖 ⎪ √ 𝜙 𝐽 ′ (𝐺−1 (𝑈 ))
+ 𝜙𝑖 𝐽 𝐽 (𝐶𝑖 ) + 𝜙𝑖 𝐽 𝐽 (𝐶𝑖1 )⎟⎬. (3.52)
⎜ 𝑘1 𝜙 3 ⎟ ⎜ 𝑘1 𝜙4𝑖 ⎟⎪ (𝑈 ) = Λ(𝑈 )𝑐′ (𝑈 ), where 𝑐 (𝑈 ) = √ 1𝑖 ′ −1

. (3.57)
⎝ 𝑖 𝑖 ⎠ ⎝ 𝑖 ⎠⎭ 𝑘𝑖 𝐺 (𝐺 (𝑈 ))
( ) It may be noted that when the permeability in 𝑘4𝑖 (last compart-
𝜙1𝑖 𝑢∗4𝑖+1
Notice that G defined above maps (0, 1) onto ∑4 𝑗 , 1 . As J is strictly ment/block of ith cell) is much larger than 𝑘1𝑖+1 (first compartment/block
𝑗=1 𝜙𝑖
increasing, the function G shares the same property and can therefore of (𝑖 + 1)th cell), the choice of the above cell as in Fig. 3 may not be a
be inverted. Hence one can write 𝐶𝑖1 in terms of U as good periodic structure for the upscaling procedure, since the threshold
saturation value at the micro-scale location 𝑦4𝑖+4 controls significantly
𝐶𝑖1 = 𝐺−1 (𝑈 ). (3.53) the flow transitions from the ith to (𝑖 + 1)th cell. In this context, we con-
sider a shifted periodic structure (refer Fig. 4), for a medium with a
Inserting the above into (3.46) one can obtain equivalent expressions for staircase upwards type porosity and intrinsic permeability distribution
𝐶𝑖𝑗 , 𝑗 = 2, 3, 4. Note that fluid flow can be ceased at any of the transition (refer Fig. 5). Please note that the upscaled quantities can be derived in a
faces unless it exceeds the corresponding threshold saturation. Hence it similar way as earlier. Here in the following we only hint the deviations
can pass through any of the transitions including the entire cell. Consid- in the local reconstruction of the flow quantities.
ering this and with the above expressions one can identify the upscaled For notational convenience we assume that the periodic structures
{ }
fractional flow as with traps are located at the points 𝑥𝑖 = 𝜖𝑦𝑖 = 𝜖𝑖 ∶ 𝑖 ∈ ℤ . The cor-
responding permeability k𝜖 (x) and the porosity 𝜙𝜖 (x) are defined by
⎧ 𝜙1𝑖 𝑢∗4𝑖+1
⎪0,

for 0 ≤ 𝑈 ≤ ∑4 𝑗
, 𝑘𝜖 (𝑥) = 𝑘(𝑥∕𝜖) and 𝜙𝜖 (𝑥) = 𝜙(𝑥∕𝜖), where
𝑗=1 𝜙𝑖
⎪( ) /( ) ( )
⎪ ∑2 1 ∑2 1
,
𝜙1𝑖 𝑢∗4𝑖+1 ⎧𝑘1 , on 𝑦4𝑖+1∕2 , 𝑦4𝑖+1 ,
⎪ 𝑖
⎪ for ∑4

𝑗=1 𝑗
𝑘 𝑘
𝑖 𝑟𝑤 ( 𝐶 𝑗
𝑖 )
𝑗=1 𝑗
𝑘 𝑘
𝑖 𝑟𝑤 ( 𝐶 𝑗
𝑖 ) 𝑓 ( 𝐶 𝑗
𝑖 ) 𝑗=1 𝜙𝑖
𝑗
( )
⎪ 𝜙1𝑖 𝑢∗4𝑖+1 + 𝜙2𝑖 𝑢∗4𝑖+2 ⎪𝑘2𝑖 , on 𝑦4𝑖+1 , 𝑦4𝑖+2 ,


<𝑈 ≤ ∑4 𝑗
, ⎪ ( )
⎪( ) /( ) 𝑗=1 𝜙𝑖 𝑘 = ⎨𝑘3𝑖 , on 𝑦4𝑖+2 , 𝑦4𝑖+3 , (3.58)
 ( 𝑈 ) = ⎨ ∑3 1 ∑3 1 𝜙1𝑖 𝑢∗4𝑖+1 + 𝜙2𝑖 𝑢∗4𝑖+2
⎪ 4 ( )
⎪𝑘𝑖 , on 𝑦4𝑖+3 , 𝑦4𝑖+4 ,
⎪ 𝑗=1 𝑗 𝑗 𝑗=1 𝑗 𝑗 𝑗
, for ∑4 𝑗
𝑘 𝑘 ( 𝐶 ) 𝑘 𝑘 ( 𝐶 ) 𝑓 ( 𝐶 ) 𝑗=1 𝜙𝑖
⎪𝑘5 = 𝑘1 , on (𝑦 )
⎪ 𝑖 𝑟𝑤 𝑖 𝑖 𝑟𝑤 𝑖 𝑖
⎪ 𝜙1𝑖 𝑢∗4𝑖+1 + 𝜙2𝑖 𝑢∗4𝑖+2 + 𝜙3𝑖 𝑢∗4𝑖+3
⎪ <𝑈 ≤ ∑4 , ⎩ 𝑖 𝑖 4𝑖+4 , 𝑦4𝑖+9∕2 ,
⎪ 𝑗
𝑗=1 𝜙𝑖
⎪( ) /( ) ( )
𝜙𝑖 𝑢4𝑖+1 + 𝜙𝑖 𝑢4𝑖+2 + 𝜙𝑖 𝑢4𝑖+3 ⎧𝜙1 ∕2,
1 ∗ 2 ∗ 3 ∗
⎪ ∑4 1 ∑4 1 on 𝑦4𝑖+1∕2 , 𝑦4𝑖+1 ,
⎪ , for < 𝑈 < 1,

𝑗=1 𝑗 𝑗
𝑘𝑖 𝑘𝑟𝑤 (𝐶𝑖 )
𝑗=1 𝑗 𝑗
𝑘𝑖 𝑘𝑟𝑤 (𝐶𝑖 )𝑓 (𝐶𝑖 ) 𝑗 ∑4
𝑗=1 𝜙𝑖
𝑗
⎪ 𝑖 ( )
⎪1,
⎩ for 𝑈 = 1. ⎪𝜙2𝑖 , on 𝑦4𝑖+1 , 𝑦4𝑖+2 ,
⎪ ( )
(3.54) 𝜙 = ⎨𝜙3𝑖 , on 𝑦4𝑖+2 , 𝑦4𝑖+3 , (3.59)
⎪ 4 ( )
𝜙
⎪ 𝑖 , on 𝑦 , 𝑦
4𝑖+3 4𝑖+4 ,
⎪𝜙5 ∕2 = 𝜙1 ∕2, on (𝑦 )
This function can be interpreted as the weighted harmonic mean of
the micro-scale fractional flow. The weight functions, 1∕𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 ), 𝑗 = ⎩ 𝑖 𝑖 4𝑖+4 , 𝑦4𝑖+9∕2 .
1, 2, 3, 4 depend on the water relative permeability, which eventually, Note that the porosities in the first and last interval are weighted with
depend on 𝐶𝑖𝑗 , 𝑗 = 1, 2, 3, 4. Note that the harmonic means in (3.54) are 1/2 to account for these blocks having half width. We distinguish four
only over the layers for which fluid could enter. This is to illustrate the kinds of matching conditions: one going from 𝑘1𝑖 to 𝑘2𝑖 , other from 𝑘2𝑖 to
potential effect that fluid could enter partly into a macroscopic grid cell 𝑘3𝑖 , other from 𝑘3𝑖 to 𝑘4𝑖 and another from 𝑘4𝑖 to 𝑘1𝑖+1 . Note that we impose
with several layers, but not completely through. For example, if the sat- matching conditions similar to Eq. (3.3) at the locations 𝑥 = (4𝑖 + 1)𝜖,
uration is such that fluid can enter the first two blocks but is too low to 𝑥 = (4𝑖 + 2)𝜖, 𝑥 = (4𝑖 + 3)𝜖 and 𝑥 = (4𝑖 + 4)𝜖. From Eq. (3.20) with similar
proceed to the third, one would be in the second case of (3.54). Note arguments as before one can prove that 𝐹 0 = 0. This conclusion allows
however that it would not be possible for the flow to pass through the us to solve Eq. (3.17) with the matching conditions. Note that in this
entire cell in this case. shifted approach the unit cell has four transition faces as opposed to the
The upscaled Λ term is the harmonic average of the terms original approach. Now through similar arguments as in the case of the
𝑘𝑗𝑖 𝜆(𝐶𝑖𝑗 ), 𝑗 = 1, 2, 3, 4, original approach we find
⎧𝐶𝑖 > 𝑢4𝑖+1 , for 𝑦4𝑖+1∕2 < 𝑦 < 𝑦4𝑖+1 ,
1 ∗
⎧ 𝜙1𝑖 𝑢∗4𝑖+1
⎪0, for 0 ≤ 𝑈 ≤ ∑ 𝑗
, ⎪𝐶 2 > 𝑢 ∗ , for 𝑦4𝑖+1 < 𝑦 < 𝑦4𝑖+2 ,

⎪ /(
4
𝑗=1 𝜙𝑖 ⎪ 𝑖 4𝑖+2

)
𝜙𝑖 𝑢4𝑖+1
1 ∗ 𝜙1𝑖 𝑢∗4𝑖+1 + 𝜙2𝑖 𝑢∗4𝑖+2 𝑢0 (𝑦) = ⎨𝐶𝑖3 > 𝑢∗4𝑖+3 , for 𝑦4𝑖+2 < 𝑦 < 𝑦4𝑖+3 , (3.60)
∑2 1
⎪2 𝑗=1 𝑗 𝑗
, for ∑ 𝑗
<𝑈 ≤ ∑4 𝑗
, ⎪𝐶 4 > 𝑢 ∗ , for 𝑦4𝑖+3 < 𝑦 < 𝑦4𝑖+4 ,
⎪ 𝑘𝑖 𝜆(𝐶𝑖 ) 4
⎪ 𝑖5 4𝑖+4
⎪ /( ) 𝑗=1 𝜙𝑖 𝑗=1 𝜙𝑖
Λ(𝑈 ) = ⎨ ∑3 𝜙𝑖 𝑢4𝑖+1 + 𝜙𝑖 𝑢4𝑖+2
1 ∗ 2 ∗ 𝜙𝑖 𝑢4𝑖+1 + 𝜙2𝑖 𝑢∗4𝑖+2 + 𝜙3𝑖 𝑢∗4𝑖+3
1 ∗ ⎩𝐶 𝑖 , for 𝑦4𝑖+4 < 𝑦 < 𝑦4𝑖+9∕2 ,
1
⎪3 𝑗=1 𝑗 𝑗
, for ∑4 𝑗
<𝑈 ≤ ∑4 𝑗
,
⎪ 𝑘𝑖 𝜆(𝐶𝑖 ) 𝑗=1 𝜙𝑖 𝑗=1 𝜙𝑖 where
⎪ /( )
√ √
𝜙𝑖 𝑢4𝑖+1 + 𝜙𝑖 𝑢4𝑖+2 + 𝜙𝑖 𝑢4𝑖+3
1 ∗ 2 ∗ 3 ∗
⎛√ √
⎪ ∑4
⎪4 𝑗=1 𝑗
1
𝑗
, for ∑4 𝑗
< 𝑈 < 1, √ 𝜙1 ∕2 √ 𝑘2 ⎞
⎪ 𝑘𝑖 𝜆(𝐶𝑖 ) 𝑗=1 𝜙𝑖 −1 ⎜√ 𝑖 √ 𝑖 𝐽 (𝐶 1 )⎟,
⎪1, 𝐶𝑖2 =𝐽 𝑖 ⎟ (3.61a)
⎩ for 𝑈 = 1. ⎜ 𝑘1𝑖 𝜙2𝑖
⎝ ⎠
(3.55) √ √
⎛√ √
√ 𝜙1 ∕2 √ 𝑘3 ⎞
Also here we present the upscaled Λ for the case of partial intrusion (sec- 𝐶𝑖3 = 𝐽 −1 ⎜√ 𝑖 √ 𝑖 𝐽 (𝐶𝑖1 )⎟, (3.61b)
⎜ 𝑘𝑖1 𝜙𝑖3 ⎟
ond and third case), but we note that there would not be any diffusion ⎝ ⎠
√ √
across the macro-scale block. To express the upscaled capillary pressure ⎛√ √ 4
√ 𝜙 ∕2 √ 𝑘
1 ⎞
we use (3.46). This means that, whenever oil is present on both sides of 𝐶𝑖4 = 𝐽 −1 ⎜√ 𝑖 √ 𝑖 𝐽 (𝐶𝑖1 )⎟, (3.61c)
all the interfaces, the first term in the expansion (3.8) is constant in the ⎜ 𝑘1𝑖 𝜙4𝑖 ⎟
⎝ ⎠
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

Fig. 4. Periodic distribution of porosity and absolute permeability in a cell with multiple discontinuities for shifted approach.

Fig. 5. Permeability and porosity in a staircase upwards distribution on the small scale.
⎛ 𝜙1𝑖 𝑢∗ ⎞
(√ √ ) (√ √ ) Notice that G defined above maps (0, 1) onto ⎜ ∑24 4𝑖+1𝑗 , 1⎟. With similar
𝜙1 ∕2 𝑘5 𝜙1𝑖 ∕2 𝑘1𝑖 ⎜ 𝑗=1 𝜙𝑖 ⎟
𝐶𝑖5 = 𝐽 −1 𝑖
𝑘1
𝑖
𝜙5 ∕2
𝐽 (𝐶𝑖1 ) = 𝐽 −1 𝑘1𝑖 𝜙1𝑖 ∕2
𝐽 (𝐶𝑖1 ) ⎝ ⎠
( ) 𝑖 1 𝑖 (3.61d)
= 𝐽 𝐽 (𝐶𝑖 ) = 𝐶𝑖 ,
−1 1 arguments as earlier one can write 𝐶𝑖1 in terms of U as in Eq. (3.53).
Inserting the above into (3.61) one can obtain equivalent expressions for
if 𝐶𝑖1 , 𝐶𝑖2 , 𝐶𝑖3 , 𝐶𝑖4 exceeds the threshold saturations 𝐶𝑖𝑗 , 𝑗 = 2, 3, 4. With the above expressions one can identify the effective
𝑢∗4𝑖+1 , 𝑢∗4𝑖+2 , 𝑢∗4𝑖+3 , 𝑢∗4𝑖+4 respectively and the non-wetting phase fractional flow as
can infiltrate into the next fine material. We also have other possibil-
ities analogous to the original approach given in Eqs. (3.47)–(3.49).
⎧ 𝜙1𝑖
However to compute the effective quantities we require 𝐶𝑖𝑗 > 0 ∀𝑗, ⎪0 , for 0 ≤ 𝑈 ≤ ∑2 4
𝑢∗4𝑖+1
,

𝜙𝑗𝑖
hence these alternatives are not relevant and to avoid repetitions we ⎪ 𝑗=1
⎪( ) /( ) 𝜙1𝑖
𝑢 ∗ 𝜙1𝑖 ∗
𝑢4𝑖+1 + 𝜙2𝑖 𝑢∗4𝑖+2
skip these expressions. ⎪ ∑2 1 ∑2 1
, for ∑2 4
4 𝑖 +1
< 𝑈 ≤ 2 ∑4 ,
⎪ 𝑗=1
𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 )
𝑗=1
𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 )𝑓 (𝐶𝑖𝑗 ) 𝑗 𝑗
Now in the case of the shifted approach, for multiple transitions ⎪( 𝑗=1 𝜙𝑖 𝑗=1 𝜙𝑖
⎪ ) /( ) 𝜙1𝑖 ∗
∑ ∑3 𝑢4𝑖+1 + 𝜙𝑖 𝑢4𝑖+2
2 ∗
within a cell, we define the averaged oil saturation as ⎪ 3 1 1
, for 2 ∑4
⎪ 𝑗=1
𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 )
𝑗=1
𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 )𝑓 (𝐶𝑖𝑗 ) 𝑗
𝑦 ⎪ 𝑗=1 𝜙𝑖
𝜙1𝑖 𝜙5𝑖
∫𝑦 4𝑖+9∕2 𝜙(𝑦)𝑢0 (𝑥, 𝑦, 𝑡)𝑑𝑦 𝐶𝑖1 + 𝜙2𝑖 𝐶𝑖2 + 𝜙3𝑖 𝐶𝑖3 + 𝜙4𝑖 𝐶𝑖4 + 𝐶𝑖5 ⎪ 𝜙1𝑖 ∗
𝑢 + 𝜙2𝑖 𝑢∗4𝑖+2 + 𝜙3𝑖 𝑢∗4𝑖+3
4𝑖+1∕2 2 2 ⎪ <𝑈 ≤ 2 4𝑖+1
,
𝑈 (𝑥, 𝑡) = 𝑦 = , ⎪ ∑4 𝑗
𝑗=1 𝜙𝑖
∫𝑦 4𝑖+9∕2 𝜙(𝑦)𝑑𝑦 𝜙1𝑖
+𝜙2𝑖 +𝜙3𝑖 +𝜙4𝑖 +
𝜙5𝑖  (𝑈 ) = ⎨
⎪( ) /( ) 𝜙1𝑖 ∗
4𝑖+1∕2 𝑢4𝑖+1 + 𝜙𝑖 𝑢4𝑖+2 + 𝜙𝑖 𝑢4𝑖+3
2 ∗ 3 ∗
2 2 ⎪ ∑4 1 ∑4 1
, for 2
⎪ 𝑗=1 𝑗=1 ∑4
(3.62) 𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 ) 𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 )𝑓 (𝐶𝑖𝑗 ) 𝑗=1 𝜙𝑖
𝑗

𝜙1𝑖 ∗
⎪ 𝑢 + 𝜙2𝑖 𝑢∗4𝑖+2 + 𝜙3𝑖 𝑢∗4𝑖+3 + 𝜙4𝑖 𝑢∗4𝑖+4
⎪ <𝑈 < 2 4𝑖+1
,
and the corresponding effective equation is given by Eq. (3.51). The ⎪ ∑4 𝑗
𝑗=1 𝜙𝑖
effective saturation is expressed in terms of 𝐶𝑖𝑗 , 𝑗 = 1, 2, 3, 4. For 𝐶𝑖𝑗 > ⎪(
⎪ ∑
) /( ) 𝜙1𝑖 ∗
𝑢 + 𝜙𝑖 𝑢4𝑖+2 + 𝜙3𝑖 𝑢∗4𝑖+3 + 𝜙4𝑖 𝑢∗4𝑖+4
2 ∗
1 ∑5 1 2 4𝑖+1
0, 𝑗 = 2, 3, 4, we end up with the relations given in (3.61). This gives ⎪

5
𝑗=1
𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 )
𝑗=1
𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 )𝑓 (𝐶𝑖𝑗 )
, for ∑4 𝑗
⎪ 𝑗=1 𝜙𝑖
⎧ √ √ < 𝑈 < 1,
⎛√ √ ⎪
√ 𝜙1 ∕2 √ 𝑘2 ⎞ ⎪
1 ⎪ 1 1 2 −1 ⎜√ 𝑖 √ 𝑖 𝐽 (𝐶 1 )⎟ ⎩1 , for 𝑈 = 1.
𝑈 = 𝐺(𝐶𝑖 ) = ∑4
1
𝜙 𝐶 + 𝜙𝑖 𝐽
𝑗⎨ 𝑖 𝑖 𝑖
⎜ 𝑘𝑖1
𝜙𝑖 2 ⎟
𝑗=1 𝜙𝑖 ⎪ ⎝ ⎠
(3.64)

√ √ √ √
⎛√ √
√ 𝜙1 ∕2 √ 𝑘3 ⎞ ⎛√ √
√ 𝜙1 ∕2 √ 𝑘4 ⎞⎫
⎪ This function can be interpreted as the weighted harmonic mean of
+ 𝜙3𝑖 𝐽 −1 ⎜√ 𝑖 1 √ 𝑖3 𝐽 (𝐶𝑖1 )⎟ + 𝜙4𝑖 𝐽 −1 ⎜√ 𝑖 1 √ 𝑖4 𝐽 (𝐶𝑖1 )⎟⎬.
⎜ 𝑘𝑖 𝜙𝑖 ⎟ ⎜ 𝑘𝑖 𝜙𝑖 ⎟⎪ the micro-scale fractional flow. The weight functions, 1∕𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 ), 𝑗 =
⎝ ⎠ ⎝ ⎠⎭
1, 2, 3, 4, 5 are depending on the water relative permeability, which even-
(3.63) tually, in turn are depending on 𝐶𝑖𝑗 , 𝑗 = 1, 2, 3, 4, 5.
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

Fig. 6. Random distribution of blocks.

The effective Λ term is the harmonic average of the terms ergodic structure is characterized by a probability space (Ω, 𝜇), along
𝑘𝑗𝑖 𝜆(𝐶𝑖𝑗 ), 𝑗 = 1, 2, 3, 4, 5, with an ergodic dynamical system 𝑇 (𝑥), 𝑥 ∈ ℝ (Papanicolaou, 1995; van
Duijn et al., 2002). Now for a 𝜇-measurable subset  ⊂ Ω, we introduce
⎧ 𝜙1

𝑖 𝑢∗ 𝑃 = 𝑃 (𝜔) ⊂ ℝ by
2 4𝑖+1
⎪0, for 0 ≤ 𝑈 ≤ ∑
4 𝑗
,
⎪ 𝑗=1 𝜙𝑖 𝑃 (𝜔) = {𝑥 ∈ ℝ ∶ 𝑇 (𝑥)𝜔 ∈  }, (4.1)
⎪ /( ) 𝜙1 𝜙1
⎪ 𝑖 𝑢∗ 𝑖 𝑢∗ + 𝜙2𝑖 𝑢∗4𝑖+2
⎪2 ∑2 2 4𝑖+1 2 4𝑖+1
⎪ 𝑗=1 𝑗
1
𝑗
, for ∑
4 𝑗
< 𝑈 ≤ ∑4 𝑗
, which is a random stationary set. In this case we assume that P(𝜔) is of
⎪ 𝑘𝑖 𝜆(𝐶𝑖 ) 𝑗=1 𝜙𝑖 𝑗=1 𝜙𝑖
⎪ /(
the form
) 𝜙 1
𝑖 𝑢∗
𝜙 1
𝑖 𝑢∗
⎪ ∑3 1 2 4𝑖+1
+ 𝜙2𝑖 𝑢∗4𝑖+2 2 4𝑖+1
+ 𝜙2𝑖 𝑢∗4𝑖+2 + 𝜙3𝑖 𝑢∗4𝑖+3 ( )
⎪3 𝑗=1 , for ∑4 <𝑈 ≤ ∑4 , 𝑃 (𝜔) = ∪𝑖∈ℤ 𝑦2𝑖 , 𝑦2𝑖+1 , (4.2)
⎪ 𝑗 𝑗 𝑗 𝑗
𝑘𝑖 𝜆(𝐶𝑖 ) 𝜙
𝑗=1 𝑖 𝑗=1 𝜙𝑖

Λ(𝑈 ) = ⎨ /( ) 𝜙1
⎪ ∑4 1
𝑖 𝑢∗
2 4𝑖+1
+ 𝜙2𝑖 𝑢∗4𝑖+2 + 𝜙3𝑖 𝑢∗4𝑖+3 in which 𝑦𝑖 ∈ ℝ are strictly increasing random variables with respect
⎪4 , for ∑4
𝑗=1 𝑗
𝑘𝑖 𝜆(𝐶𝑖 )
𝑗 𝑗 to i. According to the ergodic theorem of Birkhoff (1931) there exists a
⎪ 𝑗=1 𝜙𝑖
⎪ 𝜙1 density function of P, given by
⎪ 𝑖 𝑢∗ + 𝜙𝑖 𝑢∗4𝑖+2 + 𝜙3𝑖 𝑢∗4𝑖+3 + 𝜙4𝑖 𝑢∗4𝑖+4
2
⎪ 2 4𝑖+1
<𝑈 < ∑4 , 𝑏

⎪ 𝑗=1 𝜙𝑖
𝑗
1 ∑
⎪ /( ) 𝜙 1 𝜓 ∶= 𝜇(𝑃 ) = lim ∣ 𝑦2𝑖+1 (𝜔) − 𝑦2𝑖 (𝜔) ∣, (4.3)
⎪ ∑5 1
𝑖 𝑢∗
2 4𝑖+1
+ 𝜙2𝑖 𝑢∗4𝑖+2 + 𝜙3𝑖 𝑢∗4𝑖+3 + 𝜙4𝑖 𝑢∗4𝑖+4 𝑎,𝑏→∞ 𝑦2𝑏 − 𝑦−2𝑎 −𝑎
⎪5 𝑗=1 𝑗 𝑗
, for ∑4 𝑗
< 𝑈 < 1,
⎪ 𝑘𝑖 𝜆(𝐶𝑖 ) 𝑗=1 𝜙𝑖
⎪ for nearly all 𝜔 ∈ Ω, satisfying 0 ≤ 𝜓 ≤ 1. The corresponding random
⎩1, for 𝑈 = 1.
porosity and absolute permeability is given by
(3.65)
{ 1
𝜙𝑖 , if 𝑥𝑖 ∈ 𝑃
To express the effective capillary pressure we use (3.61). This means 𝜙(𝑥𝑖 , 𝜔) = 𝜙(𝑇 (𝑥𝑖 )𝜔) = (4.4)
𝜙2𝑖 , if 𝑥𝑖 ∈ ℝ∖𝑃
that, whenever oil is present on both sides of all the interfaces, the first
{ 1
term in the expansion (3.8) is constant in the entire micro-scale cell. 𝑘𝑖 , if 𝑥𝑖 ∈ 𝑃
𝑘(𝑥𝑖 , 𝜔) = 𝑘(𝑇 (𝑥𝑖 )𝜔) = (4.5)
Hence we get 𝑘2𝑖 , if 𝑥𝑖 ∈ ℝ∖𝑃
√ √ √ √
√ 1 √ 2 √ 3 √ 4 Note that porosity 𝜙 and absolute permeability k are stationary random
√ 𝜙 ∕2 √𝜙 √𝜙 √𝜙
𝑐 (𝑈 ) = √ 𝑖
𝐽 (𝐶𝑖 ) =
1 √ 𝑖
𝐽 (𝐶𝑖 ) =
2 √ 𝑖
𝐽 (𝐶𝑖 ) = √ 4𝑖 𝐽 (𝐶𝑖4 )
3
variables. Now
𝑘1𝑖 𝑘2𝑖 𝑘3𝑖 𝑘𝑖
√ 𝜙𝜖 (𝑥, 𝜔) = 𝜙(𝑇 (𝑥∕𝜖)𝜔) and 𝑘𝜖 (𝑥, 𝜔) = 𝑘(𝑇 (𝑥∕𝜖)𝜔). (4.6)
√ 5
√ 𝜙 ∕2
= √ 𝑖 5 𝐽 (𝐶𝑖5 ). (3.66) Consequently, = 𝑢∗𝑖 𝑢∗𝑖 (𝜔)
through the relation (3.4). This being stochas-
𝑘𝑖 tic, we use the two-scale asymptotic expansion for the saturation, cap-
To determine the effective diffusivity  we use (3.53), (3.65) and (3.66), illary pressure and flux function given by relation (3.6), (3.7) and
which implies (3.16) respectively. Note that here Fj are stationary ergodic random
√ fields. Using these expansions in Eq. (3.14), we get
√ 1
√ 𝜙 ∕2 𝐽 ′ (𝐺−1 (𝑈 ))
(𝑈 ) = Λ(𝑈 )𝑐′ (𝑈 ), where 𝑐′ (𝑈 ) =√ 𝑖1 . (3.67) 𝜕𝐹 0
=0 ⟹ 𝐹 0 = 𝐹 0 (𝑥, 𝑡), (4.7)
𝑘𝑖 𝐺′ (𝐺−1 (𝑈 )) 𝜕𝑦
while the random variable 𝑝0𝑐 satisfies the Eq. (3.18). Now for a given
4. Effective equations for randomly distributed blocks
realization of 𝜔 one needs to reconsider this Eq. (3.18). Following anal-
ogous steps as van Duijn et al. (2002) and as in the periodic layer case,
In this section, we derive effective equations for a random distribu-
it follows that 𝐹 0 = 0. Now for 𝑖 ∈ ℤ,
tion of blocks (refer to Fig. 6) in the capillary limit. Here we present
the derivation for the cell with a single discontinuity and for multiple ⎧𝐶 ( 𝜔 ) > 𝑢 ∗ ( 𝜔 ) , for 𝑦2𝑖 (𝜔) < 𝑦 < 𝑦2𝑖+1 (𝜔),
discontinuities we refer to Appendix B. The derivation steps are simi- ⎪ 𝑖 2𝑖+1(√ √ )
⎪ 𝜙 1 𝑘 2
lar to the approach explained by van Duijn et al. (2002), and are thus 𝑢0 (𝑦, 𝜔) = ⎨𝐶 𝑖 (𝜔) = 𝐽 −1 𝑖 𝑖
𝐽 (𝐶𝑖 (𝜔)) , for 𝑦2𝑖+1 (𝜔)
𝑘1𝑖 𝜙2𝑖
only explained briefly. However, here we assume the continuity of the ⎪
first order term in the asymptotic expansion of the capillary pressure ⎪ < 𝑦 < 𝑦2𝑖+2 (𝜔),

instead of the oil saturation that is used in van Duijn et al. (2002). (4.8)
We give more attention to the differences between the approach of
van Duijn et al. (2002) and what is followed here in this study. We if Ci (𝜔) exceeds the threshold saturation 𝑢∗2𝑖+1 (𝜔) and the non-wetting
consider a stationary ergodic geometrical structure instead of a peri- phase can infiltrate into the immediate fine material, or
odic distribution of blocks. A stationary ergodic structure is a stochastic {
𝐶𝑖 (𝜔) ≤ 𝑢∗2𝑖+1 (𝜔), for 𝑦2𝑖 (𝜔) < 𝑦 < 𝑦2𝑖+1 (𝜔),
structure that exhibits both stationarity and ergodicity. Essentially, this 𝑢0 (𝑦, 𝜔) = (4.9)
0, for 𝑦2𝑖+1 (𝜔) < 𝑦 < 𝑦2𝑖+2 (𝜔),
implies that the random structure will not change its statistical proper-
ties with time and that its statistical properties can be deduced from if there is no non-wetting phase flow from high to low permeable
a single, sufficiently large random sample of the process (Daley and medium. Note that in the first case capillary pressure is continuous at
Vere-Jones, 1988; Papanicolaou, 1995). It is assumed that the stationary the transition face.
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

Now consider the (𝜖 −1 )-Eq. (3.19). As 𝐹 0 = 0, the ergodic nature of One may observe that this function is again the weighted harmonic
F1 suggests that 𝐹 1 = 𝐹 1 (𝑥, 𝑡) and given by mean of the micro-scale fractional flow function, as in the periodic
( ) layer case. Here we mention that the difference in the effective frac-
𝜕𝑝1𝑐 𝜕𝑝0
𝐹 1 = 𝑓 (𝑢0 ) − 𝑁𝑐 𝑘𝑖 (𝜔)𝜆(𝑢0 ) + 𝑐 . (4.10) tional flow in van Duijn et al. (2002) compared to the present study
𝜕𝑦 𝜕𝑥
is in terms of the weights. A straightforward mathematical calculation
Let 𝐶𝑖 (𝜔) ≤ 𝑢∗2𝑖+1 (𝜔), then 𝐹 1 = 0 on (𝑦2𝑖 (𝜔), 𝑦2𝑖+1 (𝜔)) implies 𝐹 1 = shows that the weights in van Duijn et al. (2002) are 𝜓∕𝑘1𝑖 𝑘𝑟𝑤 (𝐶𝑖 )𝐽 ′ (𝐶𝑖 )
0 ∀𝑥 ∈ ℝ and𝑡 > 0. Next if 𝐶𝑖 (𝜔) > 𝑢∗2𝑖+1 (𝜔), then and (1 − 𝜓)∕𝑘2𝑖 𝑘𝑟𝑤 (𝐶𝑖 )𝐽 ′ (𝐶𝑖 ). We refer to Appendix B for the functional
form of the effective fractional flow in case of multiple discontinuities.
⎧ 𝑓 (𝐶𝑖 (𝜔)) − 𝐹 1 𝜕𝑝0 (𝐶𝑖 (𝜔)) In addition, the effective/upscaled Λ term is also the harmonic average
⎪ − 𝑐 ∶= 𝐵1 (𝜔), for 𝑦2𝑖 (𝜔) < 𝑦 < 𝑦2𝑖+1 (𝜔),
𝜕𝑝1𝑐 ⎪ 𝑁 𝑘 (𝜔)𝜆(𝐶𝑖 (𝜔))
1 𝜕𝑥 of the micro-scale quantities, hence given by
=⎨ 𝑐 𝑖
𝜕𝑦 𝜕𝑝 (𝐶𝑖 (𝜔))
⎪ 𝑓 (𝐶𝑖 (𝜔)) − 𝐹
1 0
− 𝑐 ∶= 𝐵2 (𝜔), for 𝑦2𝑖+1 (𝜔) < 𝑦 < 𝑦2𝑖+2 (𝜔). ⎧0 , 𝜓𝜙1𝑖 𝑢∗2𝑖+1
⎪ 𝑁 𝑘2 (𝜔)𝜆(𝐶 (𝜔)) 𝜕𝑥 for 0 ≤ 𝑈 ≤ ,
⎩ 𝑐 𝑖 𝑖 ⎪ / 𝜓 𝜙1𝑖 +(1−𝜓 )𝜙2𝑖
⎪ ( )
(4.11) Λ(𝑈 ) = ⎨1 𝜓
+ 1−𝜓
, for
𝜓𝜙1𝑖 𝑢∗2𝑖+1
< 𝑈 < 1, (4.19)
⎪ 𝑘1𝑖 𝜆(𝐶𝑖 ) 𝑘2𝑖 𝜆(𝐶𝑖 ) 𝜓 𝜙1𝑖 +(1−𝜓 )𝜙2𝑖
It may be noted that van Duijn et al. (2002) has the local representa- ⎪
𝜕𝑝1 ⎩0 , for 𝑈 = 1.
tion for 𝜕𝑢
1
𝜕𝑦
, however in this article we have the same for 𝜕𝑦𝑐 . This is
due to the differences in the micro-scale assumption between van Duijn For the effective Λ term in case of multiple discontinuities refer to
𝜕𝑝1𝑐 Appendix B. With similar arguments as in the periodic case, one can
et al. (2002) and the present study. Since is the local representation 𝜕𝑦 also define the effective capillary pressure and the effective diffusivity
of a stationary random variable with zero mean, we get
following the Eqs. (3.41) and (3.42) respectively.
𝜒1 𝐵 1 ( 𝜔 ) + 𝜒2 𝐵 2 ( 𝜔 ) = 0 , (4.12) Note that for a single discontinuity in a cell the periodic case corre-
where 𝜒 is the characteristic function. Now Eq. (4.12) gives sponds to 𝜓 = 12 , hence for each realization 𝜔 we end up with an effec-
tive equation similar to the periodic case (Eq. (3.36)). In addition, one
𝑓 (𝐶𝑖 (𝜔)) − 𝐹 1 𝑓 (𝐶𝑖 (𝜔)) − 𝐹 1 𝜕𝑝0𝑐 (𝐶𝑖 (𝜔)) may observe that for 𝜓 = 12 , the average oil saturation relation (4.15) re-
𝜓 + (1 − 𝜓) =𝜓
𝑁𝑐 𝑘1𝑖 (𝜔)𝜆(𝐶𝑖 (𝜔)) 𝑁𝑐 𝑘2𝑖 (𝜔)𝜆(𝐶𝑖 (𝜔)) 𝜕𝑥 duces to the relation (3.35), which is the average oil saturation for the
periodic case.
𝜕𝑝0𝑐 (𝐶𝑖 (𝜔))
+ (1 − 𝜓) . (4.13)
𝜕𝑥 5. Comparison between micro-scale and upscaled model
Eq. (4.13) can be solved for F1 and we obtain (dropping 𝜔-dependency
for convenience) We here compare the original micro-scale Eq. (2.8) and the upscaled
𝜓 1−𝜓 𝜕𝑝0𝑐 (𝐶𝑖 ) 𝜕𝑝0𝑐 (𝐶𝑖 ) equation for periodic layers (3.36) or (3.51), with special emphasis on
𝑘1𝑖 𝑘𝑟𝑤 (𝐶𝑖 )
+ 𝜓 + (1 − 𝜓)
𝑘2𝑖 𝑘𝑟𝑤 (𝐶𝑖 ) 𝜕𝑥 𝜕𝑥
𝐹1 = − 𝑁𝑐 . (4.14) showing the differences with respect to the upscaled equation developed
𝜓 1−𝜓 𝜓 1−𝜓
+ + in van Duijn et al. (2007). We first describe the numerical implemen-
𝑘1𝑖 𝑘𝑟𝑤 (𝐶𝑖 )𝑓 (𝐶𝑖 ) 𝑘2𝑖 𝑘𝑟𝑤 (𝐶𝑖 )𝑓 (𝐶𝑖 ) 𝑘1𝑖 𝜆(𝐶𝑖 ) 𝑘2𝑖 𝜆(𝐶𝑖 )
tation of the equations before moving to the numerical examples. For
Finally we use the (𝜖 0 )-Eq. (3.20). Using the ergodic nature of F2 , we all the numerical examples we let 𝑁𝑐 = 1 and consider the following
find the averaged oil saturation from Eq. (3.11) as Brooks-Corey type relative permeabilities and Leverett type capillary
𝑦 pressure:
∫𝑦 2𝑖+2 𝜙(𝑦)𝑢0 (𝑥, 𝑦, 𝑡)𝑑𝑦 𝜓𝜙1𝑖 𝐶𝑖 + (1 − 𝜓)𝜙2𝑖 𝐶 𝑖
2𝑖
𝑈 (𝑥, 𝑡) = 𝑦 = , (4.15)
∫𝑦 2𝑖+2 𝜙(𝑦)𝑑𝑦 𝜓𝜙1𝑖 + (1 − 𝜓)𝜙2𝑖 𝑘𝑟𝑤 (𝑢) = (1 − 𝑢)2 ; 𝑘𝑟𝑛𝑤 = 𝑢2 ; 𝐽 (𝑢) = (1 − 𝑢)−1∕2 . (5.1)
2𝑖

and the effective equation has the following form In van Duijn et al. (2007, 2002) the porosity was always constant, and
( 1 ) 𝜕𝑈 [ }]
𝜕 𝜕 { our model reduces to the model found in van Duijn et al. (2007) when
𝜓𝜙𝑖 + (1 − 𝜓)𝜙2𝑖 +  (𝑈 ) − 𝑁𝑐 Λ(𝑈 ) 𝑐 (𝑈 ) = 0. (4.16) the porosity is set as constant and averaging over two blocks. Hence,
𝜕𝑡 𝜕𝑥 𝜕𝑥
the model of van Duijn et al. (2007) can be seen as a special case of the
The effective saturation can be expressed in terms of Ci (𝜔) and 𝐶 𝑖 (𝜔).
model derived in the current paper. When allowing porosity to vary with
For 𝐶 𝑖 (𝜔) > 0, we have
the permeability, we will in most of the numerical examples assume a
1 Kozeny-Carman like relation between the (non-dimensional) permeabil-
𝑈 = 𝐺(𝐶𝑖 (𝜔)) =
𝜓𝜙1𝑖 + (1 − 𝜓)𝜙2𝑖 ity and porosity; that is,
⎧ √ √
⎛√ √
√ 𝜙1 √ 𝑘2 ⎞⎫ 𝑘(𝜙) =
𝜙3
,
⎪ 1 2 −1 ⎜√ 𝑖 √ 𝑖 ⎟⎪
(5.2)
× ⎨𝜓𝜙𝑖 𝐶𝑖 (𝜔) + (1 − 𝜓)𝜙𝑖 𝐽 𝐽 𝐶 𝜔 . (1 − 𝜙)2 𝛾
⎟⎬
( 𝑖 ( )) (4.17)
⎪ ⎜ 𝑘 1
𝜙2

⎩ ⎝ 𝑖 𝑖 ⎠⎭ with a material constant 𝛾 = 2.5. However, note that the considered
( ) models and the upscaling procedure in the previous sections do not rely
𝜓𝜙1𝑖 𝑢∗2𝑖+1
Note that G defined above maps (0,1) onto , 1 . We use on such a relation between porosity and permeability, and is mainly
𝜓 𝜙1𝑖 +(1−𝜓 )𝜙2𝑖
similar arguments to write Ci (𝜔) in terms of U as in Eq. (3.38). With applied for convenience.
the above expressions one can identify the effective/upscaled fractional For all the simulations we consider the initial condition
flow as
𝑢(𝑡 = 0) = 0.9999 and 𝑈 (𝑡 = 0) = 0.9999 (5.3)
𝜓𝜙1 𝑢∗
⎧0 , for 0 ≤ 𝑈 ≤ 𝜓 𝜙1 +(1− 𝑖 2𝑖+1
, for the micro-scale and the upscaled equations, respectively. The reason
𝜓 )𝜙2𝑖
⎪( )/ 𝑖

⎪ 𝜓
for not using an initial value of 1 is to avoid the singularity of the cap-
⎪ 1 + 2 1−𝜓 illary pressure (5.1). The corresponding boundary conditions over the
 (𝑈 ) = ⎨( 𝑘𝑖 𝑘𝑟𝑤 (𝐶𝑖 ) 𝑘𝑖 𝑘𝑟𝑤 (𝐶𝑖 ) )
⎪ domain x ∈ (0, 2) are
𝜓 1−𝜓 𝜓𝜙1𝑖 𝑢∗
⎪ 𝑘1 𝑘 (𝐶 )𝑓 (𝐶 ) + 𝑘2 𝑘 (𝐶 )𝑓 (𝐶 ) , for 𝜓 𝜙1 +(1−2𝑖𝜓+1)𝜙2 < 𝑈 < 1, 𝜕𝑢
⎪ 𝑖 𝑟𝑤 𝑖 𝑖 𝑖 𝑟𝑤 𝑖 𝑖 𝑖 𝑖
𝑢(𝑥 = 0) = 0 and (𝑥 = 2) = 0, (5.4)
⎩1 , for 𝑈 = 1. 𝜕𝑥
(4.18) and correspondingly for the upscaled saturation U.
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

5.1. Numerical discretization of the micro-scale equation van Duijn et al. (2002), we separate between transitions from low to
high or from high to low permeability to calculate 𝑢𝑘𝑗± and to modify
Implementing the original Eqs. (2.8) is a non-trivial task due to the the fluxes. Following the capillary pressure conditions (3.3) we have;
discontinuities appearing at the transitions between the permeability (a) If 𝑘𝑗 > 𝑘𝑗+1 then
layers. The main steps follow the approach developed and applied by
van Duijn et al. (2002, 2007), hence we explain these only in brief. if 𝑢𝑘𝑗− < 𝑢∗𝑗 , then 𝑢𝑘𝑗+ = 0,
√ √
We give more attention to the differences between the numerical ap- 𝜙𝑗 𝜙𝑗+1 (5.8a)
if 𝑢𝑘𝑗− ≥ 𝑢∗𝑗 , then 𝑘
𝐽 ( 𝑢 𝑘 )=
𝑗− 𝑘
𝐽 (𝑢𝑘𝑗+ ).
proach of van Duijn et al. (2002) and the current one. The micro-scale 𝑗 𝑗+1

Eq. (2.8) will be discretized with a cell-centered two-point finite vol-


(b) If 𝑘𝑗 < 𝑘𝑗+1 then
ume scheme using upwinding for the spatial discretization and forward
√ √
Euler for the time discretization. Such an approach has proven success- 𝜙𝑗+1 𝜙𝑗
ful in previous applications (as by van Duijn et al. (2002, 2007)) and if 𝑢𝑘𝑗+ ≥ 𝑢∗𝑗 , then 𝑘𝑗+1
𝐽 (𝑢𝑘𝑗+ ) = 𝑘𝑗
𝐽 (𝑢𝑘𝑗− ),
(5.8b)
is analyzed by Evje and Karlsen (2000). Dealing with the fluxes at the if 𝑢𝑘𝑗+ < 𝑢∗𝑗 , then 𝑢𝑘𝑗− = 0,
transitions between the layers is of high importance to correctly capture
the trapping and to get an accurate numerical approximation. There- where 𝑢∗𝑗 is the threshold saturation associated with the edge 𝑥𝑗+1∕2 ,
fore, we adopt a similar strategy here as applied by van Duijn et al. defined similarly as (3.4).
𝑘
The flux 𝐹𝑗+1∕2 is modified using values only from the left or the
(2002, 2007). However, similar considerations to incorporate capillary
pressure conditions and entry pressure effects are also taken into ac- right side of the edge 𝑥𝑗+1∕2 ; that is,
count by Ern et al. (2010), Henning et al. (2013), Duijn et al. (1995),
√ 𝛽(𝑢𝑘𝑗− ) − 𝛽(𝑢𝑘𝑗 )
van Duijn et al. (2016). Discretizations with such modified flux cal- 𝑘
𝐹𝑗+1∕2 = 𝑓 (𝑢𝑘𝑗 ) − 𝑁𝑐 𝑘𝑗 𝜙𝑗 , (5.9a)
culations are also investigated analytically by Brenner et al. (2013), Δ𝑥∕2
Enchery et al. (2006), Henning et al. (2015), Jaffré (1995). We men-
√ 𝛽(𝑢𝑘𝑗+1 ) − 𝛽(𝑢𝑘𝑗+ )
𝑘
tion that although the choice of explicit time discretization gives a time 𝐹𝑗+1∕2 = 𝑓 (𝑢𝑘𝑗+ ) − 𝑁𝑐 𝑘𝑗+1 𝜙𝑗+1 . (5.9b)
step restriction, the explicit time discretization also makes each time Δ𝑥∕2
step much cheaper as fewer nonlinear problems have to be solved. From the assumptions on 𝑢𝑘𝑗± , these two fluxes have to be equal. Since
The domain x ∈ (0, 2) is divided into uniform grid cells, where we use
𝑢𝑘𝑗is known, we reformulate (5.9) together with (5.8) as equations for
more grid cells than number of discontinuities, and where grid cell j is
the interval 𝑥 ∈ (𝑥𝑗−1∕2 , 𝑥𝑗+1∕2 ) = (𝑗, 𝑗 + 1)Δ𝑥, and Δx is the grid size. The determining 𝑢𝑘𝑗± . We are searching for the zero of the function
cell center xj is hence located in (𝑗 + 1∕2)Δ𝑥. For the time discretization √ √
Δ𝑥
we use forward Euler with a constant time step Δt. The flow is always 𝑔(𝑢𝑘𝑗− , 𝑢𝑘𝑗+ ) = 𝑓 (𝑢𝑘𝑗+ ) + 𝑁𝑐 𝑘𝑗+1 𝜙𝑗+1 𝛽(𝑢𝑘𝑗+ ) + 𝑁𝑐 𝑘𝑗 𝜙𝑗 𝛽(𝑢𝑘𝑗− ) − 𝑇 , (5.10)
2
from left to right and we apply upwinding for the advective flux. We let
where T is given by
𝑢𝑘𝑗 denote the discrete unknown in grid cell j at time step k. The grid
√ √
size is such that a discontinuity in permeability/porosity always occurs Δ𝑥
𝑇 = 𝑓 (𝑢𝑘𝑗 ) + 𝑁𝑐 𝑘𝑗+1 𝜙𝑗+1 𝛽(𝑢𝑘𝑗+1 ) + 𝑁𝑐 𝑘𝑗 𝜙𝑗 𝛽(𝑢𝑘𝑗 ). (5.11)
on a grid edge. We first describe the discretization away from the dis- 2
continuities before explaining the procedure for correctly incorporating Note that 𝑔(𝑢𝑘𝑗− , 𝑢𝑘𝑗+ ) is a strictly increasing function in both argu-
the discontinuities on the transitions. Discontinuities are found in both
ments. We assume that a CFL condition is fulfilled such that 𝑇 ≤ Δ𝑥 +
permeability k(x) and in porosity 𝜙(x), but these are assumed to occur √ √ 2
at the same locations. Hence, for an interior unknown 𝑢𝑘𝑗 +1 away from 𝑁𝑐 ( 𝑘𝑗 𝜙𝑗 + 𝑘𝑗+1 𝜙𝑗+1 )𝛽(1).
a discontinuity in permeability and porosity, we have To incorporate the capillary pressure conditions, we separate be-
tween four cases corresponding to (5.8). We emphasize that van Duijn
Δ𝑡
𝑢𝑘𝑗 +1 = 𝑢𝑘𝑗 − (𝐹 𝑘 𝑘
− 𝐹𝑗−1∕2 ), (5.5) et al. (2002) describes the same cases with respect to shifts in perme-
Δ𝑥𝜙𝑗 𝑗+1∕2
ability, but uses different bounds on T than considered in the follow-

𝑘
where 𝐹𝑗±1∕2 is the approximation of the flux function (2.9) on the ing. First, assume 𝑘𝑗 > 𝑘𝑗+1 and that 𝑇 < 𝑁𝑐 𝜙𝑗 𝑘𝑗 𝛽(𝑢∗𝑗 ). In this case the
edges xj ± 1/2 . Similarly as van Duijn et al. (2002) and Evje and function 𝑔(𝑢𝑘𝑗− , 0) has a unique zero in the interval 𝑢𝑘𝑗− ∈ [0, 𝑢∗𝑗 ). It hence
Karlsen (2000) we apply a Kirchhoff transformation to the capillary follows that 𝑢𝑘𝑗+ = 0 and the first pressure condition in (5.8a) is fulfilled.

pressure term in (2.9). We define Secondly, if 𝑘𝑗 > 𝑘𝑗+1 and 𝑁𝑐 𝜙𝑗 𝑘𝑗 𝛽(𝑢∗𝑗 ) ≤ 𝑇 , the only possibility for
𝑢 𝑔(𝑢𝑘𝑗− , 𝑢𝑘𝑗+ ) to reach zero is for a non-zero 𝑢𝑘𝑗+ and 𝑢𝑘𝑗− ≥ 𝑢∗𝑗 , and we are
𝛽(𝑢) = 𝑘𝑟𝑤 (𝑣)𝑓 (𝑣)𝐽 ′ (𝑣)𝑑𝑣, (5.6)
∫0 in the second case of (5.8a)√ . We hence search for the unique solution
( 𝜙𝑘 )
𝑘
𝑢𝑗− ∈ [𝑢𝑗 , 1) and 𝑢𝑗+ = 𝐽
∗ 𝑘 −1 𝑗 𝑗+1
𝐽 (𝑢𝑘𝑗− ) > 0 fulfilling 𝑔(𝑢𝑘𝑗− , 𝑢𝑘𝑗+ ) = 0.
where 𝛽(u) is now a strictly increasing function in u and smooth for our 𝑘𝑗 𝜙𝑗+1
choices of krw (u), krnw (u) and J(u). The flux function (2.9) can then be √
Similarly, when 𝑘𝑗 < 𝑘𝑗+1 and 𝑇 < Δ𝑥 𝑓 (𝑢∗𝑗 ) + 𝑁𝑐 𝜙𝑗+1 𝑘𝑗+1 𝛽(𝑢∗𝑗 ), the
approximated as 2
function 𝑔(0, 𝑢𝑗+ ) has a unique zero in the interval 𝑢𝑘𝑗+ ∈ [0, 𝑢∗𝑗 ). In
𝑘
√ 𝛽(𝑢𝑘𝑗+1 ) − 𝛽(𝑢𝑘𝑗 ) this case 𝑢𝑘𝑗− = 0, fulfilling the second pressure condition in (5.8b).
𝑘
𝐹𝑗+1∕2 = 𝑓 (𝑢𝑘𝑗 ) − 𝑁𝑐 𝜙𝑗 𝑘𝑗 , (5.7) √
Δ𝑥 Finally, if 𝑘𝑗 < 𝑘𝑗+1 and Δ𝑥 2
𝑓 (𝑢∗𝑗 ) + 𝑁𝑐 𝜙𝑗+1 𝑘𝑗+1 𝛽(𝑢∗𝑗 ) ≤ 𝑇 , the func-
where 𝜙j and kj are the (constant) porosity and permeability found in tion 𝑔(𝑢𝑘𝑗− , 𝑢𝑘𝑗+ ) has a unique zero fulfilling 𝑢𝑘𝑗+ ∈ [𝑢∗𝑗 , 1) and 𝑢𝑘𝑗− =
grid cell j. (√ 𝜙 𝑘 )
𝐽 −1 𝑗+1 𝑗
𝐽 ( 𝑢 𝑘 ) > 0.
When there is a discontinuity in porosity and permeability on the 𝑘 𝜙 𝑗+1 𝑗 𝑗+
cell edge 𝑥𝑗+1∕2 , we have to modify the corresponding approximation Although van Duijn et al. (2002) considered the same cases as here
𝑘
of the flux 𝐹𝑗+1∕2 . To this aim we use a similar approach as van Duijn (with constant 𝜙) with the corresponding reasoning, they use the oppo-
et al. (2002), but include also the effect of the varying porosity. In this site cases for T giving unique solutions for when 𝑘𝑗 > 𝑘𝑗+1 and 𝑘𝑗 < 𝑘𝑗+1 .
case the permeability and porosity values in grid cells j and 𝑗 + 1 are The bounds given by van Duijn et al. (2002) cannot yield unique solu-
different, and are denoted by kj , 𝑘𝑗+1 and 𝜙j , 𝜙𝑗+1 . To determine the tions for 𝑢𝑘𝑗− and 𝑢𝑘𝑗+ under their conditions. It seems that they simply
correct flux on 𝑥𝑗+1∕2 , we apply the capillary pressure conditions as de- switched the cases 𝑘𝑗 > 𝑘𝑗+1 and 𝑘𝑗 < 𝑘𝑗+1 while describing the bounds.
scribed in (3.3). We introduce dummy variables 𝑢𝑘𝑗− and 𝑢𝑘𝑗+ representing We could only reproduce their simulation results when using the cases
the saturation just left and right of the edge 𝑥𝑗+1∕2 , respectively. As in for T as described here.
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

These four cases cover the conditions in (5.8) and enable us to deter- average is weighted with respect to the layer widths in the random case.
mine the modified fluxes (5.9) on edges 𝑥𝑗+1∕2 where the permeability Note that for (3.36) and (3.51) the grid size is constant. The upscaled
and porosity is discontinuous. Any of the two expressions in (5.9) can be flux  is approximated by
applied in (5.5) as they are equal. The values for 𝑢𝑘𝑗± have to be updated 𝑘 ) −  (𝑈 𝑘 )
𝑐 (𝑈𝑗+1 𝑐 𝑗
for every time step, but depend only on information from the previous 𝑘𝑗+1∕2 =  (𝑈𝑗𝑘 ) − 𝑁𝑐 Λ𝑘𝑗+1∕2 , (5.14)
(known) time step. Δ𝑥𝑗+1∕2
To ensure that the saturation remains within the physical bounds 0
where Λ𝑘𝑗+1∕2 is the (width-weighted) harmonic average of the Λ(𝑈𝑗+1
𝑘 )
and (up to) 1, we get a time step restriction in (5.5). Following argu-
ments by Evje and Karlsen (2000) we get for (5.5) that and of Λ(𝑈𝑗𝑘 ), where Λ(U) is given by (3.40) for 𝑛 = 2, by (3.55) or
(3.65) for 𝑛 = 4, and by (4.19) for the randomly distributed blocks. The
( 𝑁𝑐 √ )
Δ𝑡 upscaled fractional flow function  (𝑈 ) is given by (3.39) for 𝑛 = 2, by
max |𝑓 ′ (𝑢)| + 2 max 𝜙(𝑥)𝑘(𝑥) max |𝛽(𝑢)| ≤ min 𝜙(𝑥)
Δ𝑥 𝑢∈[0,1) Δ𝑥 𝑥∈[0,2] 𝑢∈[0,1) 𝑥∈[0,2] (3.54) or (3.64) for 𝑛 = 4, and by (4.18) for the random blocks. The
(5.12) averaged grid size Δ𝑥𝑗+1∕2 = (Δ𝑥𝑗 + Δ𝑥𝑗+1 )∕2 is used in the case for
the randomly distributed blocks (4.16), but is constant equal to Δx for
guarantees a solution within these bounds as long the initial condition (3.36) and (3.51). To evaluate  (𝑈𝑗𝑘 ) and Λ(𝑈𝑗𝑘 ), the local saturations
fulfills these. For a domain with many discontinuities in the permeability have to be reconstructed for each layer found within grid cell j as ex-
and porosity a very small grid size Δx will be needed, which will also plained in Sections 3.2, 3.3 and 4. The discontinuities between the layers
pose severe restrictions on the time step Δt. For example, for Δ𝑥 = 1∕512, are incorporated into the upscaled equations, hence they do not require
𝑁𝑐 = 1, 𝑀 = 1, 𝜙 varying between 0.5 and 0.66, and maximum k being any special attention except to ensure that the local reconstruction re-
1, we get Δ𝑡 ≤ 9 ⋅ 10−6 . The simulations in Section 5.3 are made with a mains within [0,1). Hence, (5.13) is applied for the entire domain.
time step fulfilling (5.12), hence the saturations are expected to remain Similarly as with the discretization for the micro-scale equations, we
within [0,1) as proven by Pop and Yong (2002). Although we use the get a bound on the time step size in (5.13) to ensure solutions remaining
initial condition (5.3), which is below 1, we could in theory have that between 0 and (up to) 1. Here we get
the saturation in the first time steps increases to saturation values very
( 𝑁 )
Δ𝑡
close to 1 near the layer transitions, which would lead to numerical max | ′ (𝑈 )| + 2 𝑐 max |Λ(𝑈 )𝑐 (𝑈 )| ≤ min 𝜙(𝑥). (5.15)
difficulties. This was however not the case for the examples considered Δ𝑥 𝑈 ∈[0,1) Δ𝑥 𝑈 ∈[0,1) 𝑥∈[0,2]

in the current study, where only a small increase related to the machine In the case of a non-uniform grid, the above bound should be fulfilled
precision was encountered. with respect to the smallest appearing grid size. As the upscaled equa-
tions are in general solved on a much coarser mesh, this constraint will
5.2. Numerical discretization of the upscaled equations not be a severe one. For the numerical examples in the following section,
the upscaled equations are solved with the same time step as applied for
The upscaled Eqs. (3.36), (3.51) and (4.16) are also discretized us- the corresponding micro-scale equations, which always fulfilled the up-
ing a cell-centered finite volume scheme. Although the upscaled solu- scaled time step constraint.
tion represents the case where the number of layers have approached
infinity, we will generally use the same number of grid points for the 5.3. Numerical examples
upscaled solution as there are layers divided by the number of layers we
average over. In this case the upscaled solution has the same number of We first assess the behavior of the micro-scale and upscaled equa-
grid points as the averaged micro-scale solution, giving a fair compar- tions with respect to an increasing number of repeated layers and as-
ison between the two solutions. Hence, if considering a case with e.g. sess the effect of averaging over several layers as well as the num-
16 layers, with every second low and high permeabilities and porosi- ber of grid points used. Further, we compare our model to van Duijn
ties, we would only apply 8 upscaled grid cells when considering the et al. (2007) and show the differences that appear when letting the
formulation from Section 3.2. porosity vary, before considering the effect of large permeability (and
For (3.36) and (3.51) a uniform grid is applied. Each grid cell j is porosity) difference and viscosity ratio. We then assess some interesting
the interval 𝑥 ∈ (𝑥𝑗−1∕2 , 𝑥𝑗+1∕2 ) = (𝑗, 𝑗 + 1)Δ𝑥, where Δx is the grid size, behavior when letting n > 2, and finally we address a case with randomly
but where the grid size is now determined by the number of permeabil- distributed blocks. The original (non-shifted) approach is applied for the
ity and porosity layers and of how many are averaged over. The cell upscaled equations unless otherwise is stated. Unless otherwise stated,
center xj is located in (𝑗 + 1∕2)Δ𝑥. If there are N layers in total and we we apply for the micro-scale simulations 𝑁𝑥 = 8 grid points within each
average over n layers, where N/n is an integer, the discretization of the layer (and hence Nx · N grid points in total for N layers), while for the
upscaled model hence has N/n grid cells and Δ𝑥 = 2𝑛∕𝑁. But Δx can upscaled equations we apply N/n grid points.
also be chosen differently, and in this case there is no connection be-
tween the number of grid points used for the upscaled solution and the
5.3.1. Behavior with respect to number of grid points and averaging layers
actual number of layers.
We let N denote the total number of layers between x ∈ [0, 2]. The
For the randomly distributed blocks, the upscaled Eq. (4.16) is dis-
permeability values are oscillating between 0.2 and 0.4, with corre-
cretized on a non-uniform grid that reflects the widths of the underlying
sponding porosity values of 0.5 and 0.57. We let the viscosity ratio
random layers that are averaged over. Here each grid cell j is the inter-
be 𝑀 = 1. Increasing N for the micro-scale equations corresponds to
val 𝑥 ∈ (𝑥𝑗−1∕2 , 𝑥𝑗+1∕2 ) where Δ𝑥𝑗 = 𝑥𝑗+1∕2 − 𝑥𝑗−1∕2 and the cell center xj
layers becoming thinner and the medium acting more homogeneous.
is in 𝑥𝑗 = (𝑥𝑗+1∕2 + 𝑥𝑗−1∕2 )∕2.
Fig. 7 shows the average behavior of the micro-scale equations (black
The flow is always from left to right, and we apply upwinding. For
lines), where the solution has been averaged over the two internal lay-
the time discretization we use forward Euler with a constant time step
ers. The micro-scale simulations have here used a discretization with
Δt. Letting 𝑈𝑗𝑘 denote the discrete unknown in grid cell j at time step k,
𝑁𝑥 = 8 grid points in each layer. The average is a porosity-weighted
the upscaled Eqs. (3.36), (3.51) or (4.16) is
average to coincide with the averaging applied in the upscaling proce-
Δ𝑡 dure. In Fig. 7 we also show the convergence of the upscaled solution
𝑈𝑗𝑘+1 = 𝑈𝑗𝑘 − (𝑘𝑗+1∕2 − 𝑘𝑗−1∕2 ), (5.13)
Δ𝑥𝑗 𝜙𝑗 with respect to refining the grid. As mentioned in Section 5.2, the num-
ber of grid points for the upscaled equations can, but does not have to,
where 𝑘𝑗+1∕2 is the approximation of the flux found in (3.36), (3.51) or represent the number of layers. We here use N to denote the number of
(4.16), and 𝜙𝑗 is the average of the n porosities found in grid cell j. This layers also for the upscaled equations; since we always average over two
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

Fig. 7. Porosity-weighted average of the solutions of the micro-scale equations (black lines with markers) and solutions of the upscaled equations (purple lines with
markers) at non-dimensional time 𝑡 = 0.4. The right figure shows a zoomed-in view of the full solution seen to the left. (For interpretation of the references to colour
in this figure legend, the reader is referred to the web version of this article.)

layers this means N/2 grid points have been used. This way, the aver- and 𝑛 = 8 layers, using N/n grid points. The results are seen in Fig. 8.
aged and upscaled solutions with same N in Fig. 7 use the same number Note that when solving the 𝑛 = 8-upscaled equations, only 16 macro-
of grid points. The purple lines are the results of the upscaled equa- scale grid points are used. This gives a very cheap, and fairly accurate
tions where we always average over two transitions; hence 𝑛 = 2. The discretization, although the shape of the front is not captured so accu-
upscaled solution shows a small overestimate of the saturation-values rately due to a low amount of grid points. However, from Fig. 8 it is
near the inlet as also observed in van Duijn et al. (2007) for constant clear that all the upscaled approaches do a good job in approximating
porosities. This overestimate is however decreasing with increasing N. the average behavior of the micro-scale equations. As in Fig. 7 we see a
Hence, Fig. 7 shows how the increasing number of layers support a con- slight overestimate of the upscaled solutions compared to the averaged
verging average behavior, while increasing number of grid points for micro-scale solutions. However, this overestimate is less when only av-
the upscaled equations also support convergence, and that these two eraging over 𝑛 = 2.
approaches converge towards the same averaged/upscaled solution. In Finally, we assess the grid convergence of the micro-scale equations.
the following we will always use N/n grid points for the upscaled equa- We here fix 𝑁 = 96 layers and let Nx be the number of grid points
tions. One can of course use a larger amount of grid points, but this used within each layer. Fig. 9 shows the micro-scale solutions using
choice give a more fair comparison with the averaged solution as they 𝑁𝑥 = 4, 8, 16 grid points within each layer. The zoomed-in view in the
then have the same number of grid points. right part of Fig. 9 shows that 𝑁𝑥 = 4 is deviating slightly from the other
We now fix 𝑁 = 128, and assess the effect of averaging or upscaling two near the block transitions, while 𝑁𝑥 = 8 and 𝑁𝑥 = 16 give very sim-
over two or more internal layers. The setting is the same as before, with ilar results at the micro scale. The averaged solutions are visually iden-
permeability values oscillating between 0.2 and 0.4, porosity between tical for all three cases, with the averaged 𝑁𝑥 = 4-solution lying slightly
0.5 and 0.57, and viscosity ratio is 𝑀 = 1. The original micro-scale equa- above the other two. Calculating the L2 -norm of the difference between
tions are solved for 𝑁 = 128 and with 𝑁𝑥 = 8 grid points within each the two coarser solutions with the 𝑁𝑥 = 16-solution, gives 6.6 ⋅ 10−3 and
layer, and the micro-scale solution is averaged, using porosity-weighted 2.6 ⋅ 10−3 for 𝑁𝑥 = 4 for the averaged and micro-scale solution respec-
averages, over 𝑛 = 2, 𝑛 = 4 and 𝑛 = 8 layers. Also, the upscaled equa- tively, and 2.2 ⋅ 10−3 and 9.1 ⋅ 10−4 for 𝑁𝑥 = 8. The small difference be-
tions are correspondingly solved for when averaging over 𝑛 = 2, 𝑛 = 4 tween 𝑁𝑥 = 8 and 𝑁𝑥 = 16, both for the micro-scale solution and the

Fig. 8. The micro-scale solution (dashed line) together with the porosity-weighted average of the micro-scale solution over 𝑛 = 2, 4, 8 (black lines with markers) at
non-dimensional time 𝑡 = 0.4. The purple lines with markers are the solutions of the upscaled equations, averaged over 𝑛 = 2, 4, 8. The right figure shows a zoomed-in
view of the full solution seen to the left. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

Fig. 9. Micro-scale solutions (lines) together with their porosity-weighted averages (lines with markers) at time 𝑡 = 0.4 using either 𝑁𝑥 = 4 (solid lines), 𝑁𝑥 = 8
(dashed lides) or 𝑁𝑥 = 16 (dashed-dotted lines) in each block. The right figure shows a zoomed-in view of the full solution seen to the left. Note that the three
averaged lines are on top of each other and difficult to separate even with markers.

Fig. 10. Left: Average of the micro-scale solutions (black lines with markers) and solutions of the upscaled equations (purple lines with markers) for either constant or
varying porosity. Right: Zoomed-in view of the micro-scale solutions (black dashed and dashed-dotted lines) together with their average and the upscaled solutions.
The plots are at non-dimensional time 𝑡 = 0.4 for the varying porosity and 𝑡 = 0.75 for constant porosity, which corresponds to the same time due to different
non-dimensionalizations. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

averaged solution, indicates that the gain of increasing the number of of permeability 𝑘 = 0.2 and 𝑘 = 1, while porosity correspondingly os-
grid points per layer is marginal. We will hence use 𝑁𝑥 = 8 grid points cillates between 𝜙 = 0.5 and 𝜙 = 0.66. Fig. 11 shows the upscaled and
for each layer in all following numerical examples. averaged solution after 𝑡 = 0.1, 𝑡 = 0.3 and 𝑡 = 0.5 using 𝑁 = 128 and
𝑛 = 2, both for the averaged and upscaled solutions. Comparing to the
5.3.2. Constant versus varying porosity solutions found with a smaller permeability and porosity difference in
We now consider 𝑁 = 128 and let the permeability vary between Figs. 8 and 10, we see that the averaged and upscaled solutions approach
𝑘 = 0.2 and 𝑘 = 0.4. To compare with the model considered by van Duijn a higher plateau level, as expected. Also, the micro-scale oscillations are
et al. (2007) we either let the porosity be constant, or vary between much larger where substantially larger saturation values are found in the
𝜙 = 0.5 and 𝜙 = 0.57 with the fine and coarse layers. Fig. 10 shows the high-porosity and permeability layers.
average of the micro-scale equations and the upscaled equations (𝑛 = 2) We consider now a larger permeability ratio, and also explore a
for each of the two cases, together with a zoomed-in view of also the different relation between porosity and permeability than the Kozeny-
original micro-scale solution. Since van Duijn et al. (2007) has incor- Carman relation (5.2). The repeating layers have a permeability of
porated the porosity into the non-dimensionalization, we calculate a 𝑘 = 0.1 and 𝑘 = 1, while porosity either follows (5.2) giving 𝜙 = 0.43 and
corresponding non-dimensional time for the varying porosity based on 𝜙 = 0.66 in the layers, or we consider porosities of 𝜙 = 0.2 and 𝜙 = 0.8
the average porosity. The effect of varying porosity mainly affects the in the layers, to at the same time show how a larger porosity ratio af-
slope of the saturation front. Also, the varying porosity gives a different fects the behavior. The averaged and upscaled results at time 𝑡 = 0.25
internal distribution of the saturation between the fine and the coarse using 𝑁 = 128 and 𝑛 = 2 are shown in Fig. 12, where the results marked
layers. KC corresponds to the porosity ratio following (5.2), while the results
marked large have the larger porosity ratio. Compared to Fig. 11 we
5.3.3. Effect of strong discontinuity between layers see that the plateau values of trapped oil near the inlet is even larger
If increasing the permeability and porosity ratio between the repeat- for both cases. Although the averaged and upscaled behavior of the KC
ing layers, the effect is similar as observed in van Duijn et al. (2007), and large cases are quite similar, with only a minor difference in the
where the front becomes more steep. We here apply repeating layers
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

Fig. 11. Left: Average of the micro-scale solutions (black lines with circles) and the upscaled solutions (purple lines with triangles) for non-dimensional times 𝑡 = 0.1,
𝑡 = 0.3, 𝑡 = 0.5 when there is a permeability ratio of 5 between the layers, and a porosity ratio following (5.2). Right: Zoomed-in view of the micro-scale solution
(black dashed line) at 𝑡 = 0.3 together with the average and the upscaled solution. (For interpretation of the references to colour in this figure legend, the reader is
referred to the web version of this article.)

Fig. 12. Left: Average of the micro-scale solutions (black lines with markers) and the upscaled solution (purple lines with markers) for non-dimensional time 𝑡 = 0.25
when there is a permeability ratio of 10, and porosity ratio either follows (5.2) (solid lines with circles) or have a large porosity ratio (dotted lines with stars). Right:
Zoomed-in view of the micro-scale solution (dashed line for the KC case, dashed-dotted for large porosity ratio) together with the average and the upscaled solutions.
(For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

front speed and the internal distribution of the micro-scale solution is increasing viscosity ratio allows more trapping of fluid near the inlet, in
quite different. The difference in front speed can be connected to the both the high-permeability and low-permeability layers.
average porosity being different. Note that the averaged and upscaled
solutions are based on porosity-weighted averages. In the case of a large 5.3.5. Four repeated layers
porosity difference between the layers, the larger saturations in the high- We here design two test cases where four layers are repeated
porosity layers have a large impact on the averaged solution. The in- throughout the domain. The four layers have increasingly finer/coarser
creased porosity ratio between the layers, makes the jump in the micro- permeabilities and porosities, and attain the values
scale saturation between the layers smaller, which was also observed in 𝑘1𝑖 = 0.2; 𝑘2𝑖 = 0.34; 𝑘3𝑖 = 0.58; 𝑘4𝑖 = 0.99,
Fig. 10.
𝜙1𝑖 = 0.5; 𝜙2𝑖 = 0.55; 𝜙3𝑖 = 0.61; 𝜙4𝑖 = 0.66.

5.3.4. Effect of large viscosity ratio These values are periodically repeated through the domain, and we con-
We also consider the viscosity ratio between the fluids being large, sider two cases: One where the internal distribution is a Staircase up-
namely 𝑀 = 5 and 𝑀 = 10. Going back to the 𝑁 = 128 permeability wards with stepwise increasing values, as in the above equation, and
and porosity layers of 𝑘 = 0.2 and 𝑘 = 0.4, and 𝜙 = 0.5 and 𝜙 = 0.57 as the other where the internal distribution is a Staircase downwards where
earlier, and average over 𝑛 = 2, we obtain the averaged and upscaled the permeabilities and porosities are stepwise decreasing, oppositely as
solutions found in Fig. 13. The solutions are shown at the times 𝑡 = 0.1, the above equation. These two distributions represent the extreme cases
𝑡 = 0.25 and 𝑡 = 0.4, which approximately corresponds to the times of that can be expected when there are four repeated values. This is due to
𝑡 = 0.2, 𝑡 = 0.5 and 𝑡 = 0.8 considered in van Duijn et al. (2007). Also the fact that both include a step between the highest and lowest porosity
when including varying porosity, we see that the larger viscosity ratios and permeability values. In particular for the Staircase upwards case, the
lead to a more smooth front as capillary forces become more dominant step from the last high-permeability layer to the first low-permeability
and the front speed increases. This is particularly more visible for 𝑀 = layer poses a severe restriction on the flow. For this case the shifted up-
10. The zoomed-in view in the right part of Fig. 13 also shows how the scaling approach is expected to better represent the average behavior.
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

Fig. 13. Left: Average of the micro-scale solutions (black lines with circles, respectively stars) and the upscaled solutions (purple lines with circles, respectively stars)
for 𝑀 = 5, respectively 𝑀 = 10 for times 𝑡 = 0.1, 𝑡 = 0.25, 𝑡 = 0.4 (lines reoccurring from left to right). Right: Zoomed-in view of the micro-scale solutions (black,
dashed line for 𝑀 = 5, black, dashed-dotted line for 𝑀 = 10) at 𝑡 = 0.25 together with the average and the upscaled solutions. (For interpretation of the references
to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 14. Effective fractional flow  (𝑈 ) from (3.54) (left) and effective diffusivity (𝑈 ) from (3.57) (right) for the two cases Staircase upwards and Staircase downwards
using the original approach. The effective properties are equal for U > 0.45.

The Staircase downwards case has several minor steps that all limit the and they are identical for U > 0.45. Compared to the fractional flow and
flow, while the large step from the last low-permeability layer to the effective diffusivity found in e.g. van Duijn et al. (2007), we see a non-
first high-permeability layer does not pose a restriction. Here both the monotone behavior for low values of U in the effective quantities in
upscaling approaches are expected to behave reasonably well. the staircase downwards case. Recall that the effective fractional flow
Although the permeability and porosity values are the same in these (3.54) and (3.64), and diffusivity (3.57) and (3.67) are defined as the
two cases, the effective fractional flow functions and effective diffusivity harmonic averages over the layers where fluid could enter. This non-
are generally not the same when averaging over 𝑛 = 4. The reason is that monotone behavior is due to the fluid in this case would theoretically be
the local reconstructions of the saturation values (using (3.52) or (3.63)) able to enter the first high-permeability block (or two first blocks) in the
are different. However, if applying the original multi-block averaging cell, but not pass through the cell completely as the last block represents
described in Section 3.3, the effective diffusivity (3.57) and fractional the largest restriction on the flow. Only the staircase downwards case
flow (3.54) functions are the same for values larger than 𝑈 = 0.45, and includes such artifacts as the internal transitions pose restrictions on the
they would both approach zero as U↘0.45. This is due to only the inter- flow, while for the staircase upwards case the flow could infiltrate the
nal transitions being properly accounted for. Especially for the staircase entire layer when being able to enter the first low-permeability layer.
upwards case, this is a crude approach as the transition from the last, However, since we initially start with a large saturation which grad-
high-permeability block into the first low-permeability block in the next ually decreases, the effective saturation U will not reach values below
cell, represents a large restriction on the flow. 𝑈 = 0.45 or 𝑈 = 0.48, as the effective quantities do not allow any fluid
In Figs. 14 and 15, the effective fractional flow (3.54) and (3.64), and movement then. Values below 𝑈 = 0.45 or 𝑈 = 0.48 correspond to at
diffusivity (3.57) and (3.67) using the original and shifted approach for least one block (those with lowest porosity and permeability) having
the staircase upwards and downwards cases are shown. As mentioned zero saturation. This means that the non-monotone behavior seen in
earlier, the original approach gives curves that are similar to each other, the effective flux and diffusivity functions is not relevant as we will
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

Fig. 15. Effective fractional flow  (𝑈 ) from (3.64) (left) and effective diffusivity (𝑈 ) from (3.67) (right) for the two cases Staircase upwards and Staircase downwards
using the shifted approach.

Fig. 16. Micro-scale (black, dashed line), averaged (black, solid line with circles) and upscaled (purple lines with markers) solutions of the 𝑛 = 4 periodically
distributed layers named staircase upwards (left) and staircase downwards (right) at non-dimensional time 𝑡 = 0.4. (For interpretation of the references to colour in
this figure legend, the reader is referred to the web version of this article.)

not be able to attain effective saturation values below 0.45 or 0.48 upscaled solutions in Fig. 16, the internal distribution of the layers is
for the staircase downwards case. We also see that for the shifted ap- not irrelevant. In particular for the shifted approach, the lowest pos-
proach, the possible upscaled saturation values are only above 𝑈 = 0.48 sible value of upscaled saturation is 𝑈 = 0.48 for the staircase down-
and 𝑈 = 0.31 for the staircase downwards and staircase upwards cases, wards case, while the lowest for the staircase upwards case is 𝑈 = 0.31,
respectively. as seen in Fig. 15. For the original approach the lowest upscaled sat-
Fig. 16 shows the micro-scale solutions of the two cases together uration is 𝑈 = 0.45 in both the cases (Fig. 14). This is reflected also
with the averaged and upscaled solutions, using 𝑛 = 4. The upscaled so- in Fig. 16 where these plateau values are approached near the inlet
lutions using both the original and shifted approaches are shown. Note in the upscaled solutions. The averaged saturation of the staircase up-
that the solutions for staircase upwards and staircase downwards using wards case is approaching a plateau value similar to as predicted by the
the original approach are the same as they share the same effective frac- shifted upscaling approach, while the averaged saturation of the stair-
tional flow and effective diffusivity in the relevant range of U. It is also case downwards case approaches a plateau value slightly below both
clear from Fig. 16 that the shifted upscaling approach represents, as ex- upscaling approaches, but closer to the original approach. The stair-
pected, a much better approximation of the averaged behavior of the case downwards case is hence not as limited in its averaged transport
staircase upwards case compared to the original approach. The shifted as predicted by the upscaled equations. The amount of trapping in the
approach leads to an overestimation of the plateau level near the inlet staircase downwards case is overestimated by both approaches. A pos-
in the staircase downwards case. Hence, for the staircase downwards ex- sible way to improve the upscaling approach for such a transport would
ample, the original approach approximates the average behavior slightly be to better incorporate the behavior between the low-permeability and
better, but both the original and shifted approach lead to fairly good the next high-permeability block between the cells. The non-monotone
representations of the average behavior in this case. behavior discussed earlier, where fluid would propagate partly into a
As seen in the curves for the effective fractional flow and effective new cell but do not cross completely, as seen in Figs. 14 and 15, in-
diffusivity in Figs. 14 and 15 and in the corresponding averaged and dicates that the staircase downwards case should indeed be more mo-
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

Fig. 17. Left: Random realization of micro-scale block width. Each block has a width being an integer multiplied with Δ𝑥 = 1∕512. The sum of the block widths is
2. Right: The corresponding macro-scale grid size for each macro-scale grid cell.

Fig. 18. The micro-scale solution (dashed line) together with the average (black line with circles) and solution of the upscaled equations (purple line with triangles).
The right figure shows a zoomed-in view of the full solution seen to the left. (For interpretation of the references to colour in this figure legend, the reader is referred
to the web version of this article.)

bile than currently predicted by the upscaled approaches. However, while the upper could be set to any number larger than 4. The last block
the upscaled quantities below 𝑈 = 0.45 or 𝑈 = 0.48 are now never has been adjusted in size such that the domain is x ∈ (0, 2). The applied
relevant. realization resulted in 124 blocks, where the width of each block can be
seen in Fig. 17. Here we also show the macro-scale grid size for each of
5.3.6. Randomly distributed blocks the 62 macro-scale grid cells.
Finally we investigate the case of randomly distributed blocks as The micro-scale solution together with the averaged and upscaled
described in Section 4. We consider two different layers with porosity solutions at time 𝑡 = 0.4 are shown in Fig. 18. The micro-scale solution
𝜙 = 0.5 and 𝜙 = 0.57 and permeability 𝑘 = 0.2 and 𝑘 = 0.4. The width shows a qualitative similar behavior as e.g. in Fig. 9, although with a
of each layer is randomly chosen. For simplicity we keep the grid on clear visible effect from the varying layer width. The upscaled and aver-
the micro-scale uniform and the random width is chosen to be a multi- aged solutions are not behaving monotone anymore, which is due to the
ple of the chosen grid size. This ensures that the same discretization as averages being weighted with the layer width as well as the porosity.
described in Section 5.1 can be directly applied. Hence, when a high-porosity (and high-permeability) layer is wide, the
To ease the comparison with the averaged and upscaled solutions, average over the two blocks will also be larger. The averaged and up-
the upscaled solution is chosen to have a grid such that each macro- scaled solutions are however still behaving very similar to each other,
scale grid cell corresponds to two micro-scale blocks. Since the blocks showing that the upscaled solution can still give a very good represen-
have a random width, the upscaled Eq. (4.16), with corresponding flux tation of the averaged behavior also in the case of randomly distributed
function (4.18) and diffusivity (4.19), is discretized on a non-uniform blocks.
grid, as described in Section 5.2. Note that the averaged and upscaled
saturations represent the averages weighted with respect to porosity and 6. Final remarks
layer width, as seen in (4.15).
In order to realize a random distribution of the blocks, we choose We have derived an upscaled approach for two-phase flow through
Δ𝑥 = 1∕512 and pick for each micro-scale block a random number from highly heterogeneous media, where both permeability and porosity can
a uniform distribution of integers between 4 and 12. The lower limit is vary substantially. The upscaling procedure applies to periodically or
set to 4 to ensure the presence of at least four grid points in each layer, randomly distributed heterogeneities, and where these heterogeneities
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

can have any internal distributions compared to earlier approaches. Ac- for the original approach. In the case of multiple transitions within a
counting for porosity differences in the model gives new conditions for cell (refer to Fig. A.19), one may define the average oil saturation as
the capillary pressure determining the overall flow, and also affects the 𝑦 ∑𝑛 𝑗 𝑗
∫𝑦 𝑛(𝑖+1) 𝜙(𝑦)𝑢0 (𝑥, 𝑦, 𝑡)𝑑𝑦 𝑗=1 𝜙𝑖 𝐶𝑖
internal distribution of the fluid and hence the local trapping of the fluid.
𝑈 (𝑥, 𝑡) = 𝑛𝑖 𝑦 = ∑𝑛 𝑗
, (A.1)
The inclusion of several blocks that are periodically repeated with ∫𝑦 𝑛(𝑖+1) 𝜙(𝑦)𝑑𝑦 𝑗=1 𝜙𝑖
𝑛𝑖
different internal distributions in the block properties, opens for the
consideration of local trapping in different ways. In particular we are and the corresponding effective equation is given by
able to observe behavior with respect to trapping that would not have ∑𝑛 𝑗
𝑗=1 𝜙𝑖 𝜕𝑈
[ }]
been visible for when only averaging over two periodically repeated 𝜕 𝜕 {
+  (𝑈 ) − 𝑁𝑐 Λ(𝑈 )  (𝑈 ) = 0. (A.2)
layers or when considering a constant porosity. The upscaled fractional 𝑛 𝜕𝑡 𝜕𝑥 𝜕𝑥 𝑐
flow function and effective diffusion depend both on the permeability The effective saturation is expressed in terms of 𝐶𝑖𝑗 , 𝑗 = 1, 2, … , 𝑛. For
and porosity values of the internal blocks, but also of their internal dis- 𝐶𝑖𝑗 > 0, 𝑗 = 2, 3, … , 𝑛, we get
tribution as the ordering of high-permeability versus low-permeability
√ √
(and likewise, porosity) blocks affects the effective flow and diffusivity ⎛√ √
√ 𝜙1 √ 𝑘𝑗 ⎞
behavior. Especially the location of the lowest permeability and poros- 𝐶𝑖𝑗 = 𝐽 −1 ⎜√ 𝑖 √ 𝑖𝑗 𝐽 (𝐶𝑖1 )⎟, 𝑗 = 2, 3, … 𝑛. (A.3)
⎜ 𝑘 1
𝜙𝑖 ⎟
ity values compared to its neighbors is important for the overall flow. ⎝ 𝑖 ⎠
To better account for the extreme case where the flow is limited by a
transition from a high- to a low-permeability block, a shifted upscaling This leads to
⎧ √ √
⎞⎫
approach was found to better represent the average behavior in such
𝑛 ⎛√ 𝜙1 √
√ 𝑗
scenarios. The upscaling has been generalized to be valid for averaging ⎪ 1 1 ∑ 𝑗 −1 ⎜√
1 √ 𝑖 √ 𝑘𝑖 𝐽 (𝐶 1 )⎟⎪.
𝑈= 𝐺(𝐶𝑖1 )
= ∑𝑛 𝜙
𝑗⎨ 𝑖 𝑖
𝐶 + 𝜙𝑖 𝐽 𝑖 ⎟⎬ (A.4)
⎜ 𝑘1 𝜙𝑗𝑖
𝑗=1 𝜙𝑖 ⎪
over any n number of periodally repeated layers, but also to random ⎪

𝑗=2 ⎝ 𝑖 ⎠⎭
layers where both porosity and permeability are randomly distributed.
The upscaled solution for the randomly distributed blocks was found to ( 1 ∗ )
𝜙 𝑢
still give a very good representation of the average behavior, showing Notice that G defined above maps (0, 1) onto ∑𝑖𝑛 𝑛𝑖+1𝑗 , 1 . As J is strictly
𝑗=1 𝜙𝑖
the applicability of the upscaling procedure also in more general cases. increasing, the function G shares the same property and can therefore be
It should however be stressed that the considered models are applicable inverted. Hence one can write 𝐶𝑖1 in terms of U as in (3.53). Considering
for moderate capillary numbers. the above expressions one can identify the effective/upscaled fractional
The upscaling approach we have presented here for one-dimensional flow as
porous media represents a necessary basis for later extensions to two
or three spatial dimensions. With the current development, we have ⎧ 𝜙1𝑖 𝑢∗𝑛𝑖+1
⎪0 , for 0 ≤ 𝑈 ≤ ∑𝑛 ,
𝑗
derived a novel approach that allows the upscaling of trapping effects ⎪ 𝑗=1 𝜙𝑖
⎪( )/( ) ∑2 𝑗 ∗
𝜙𝑖 𝑢𝑛𝑖+1 𝑗=1 𝜙𝑖 𝑢𝑛𝑖+𝑗
1 ∗
based on differences in the fractional flow and capillary effects through ⎪ ∑2 1 ∑2 1
⎪ 𝑗=1 𝑗=1
, for ∑𝑛 < 𝑈 ≤ ∑𝑛 ,
⎪ 𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 ) 𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 )𝑓 (𝐶𝑖𝑗 ) 𝜙 𝑗
𝑗=1 𝜙𝑖
𝑗
varying material properties. ⎪( )/( ) ∑𝑗=1
2
𝑖
𝑗 ∗ ∑3 𝑗 ∗
𝜙 𝑢 𝜙 𝑢
⎪ ∑3 1 ∑3 1
, for ∑𝑛
𝑗=1 𝑖 𝑛𝑖+𝑗
< 𝑈 ≤ ∑𝑛
𝑗=1 𝑖 𝑛𝑖+𝑗
,
⎪ 𝑗=1
𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 )
𝑗=1
𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 )𝑓 (𝐶𝑖𝑗 ) 𝜙𝑗 𝜙𝑗
Declaration of Competing Interest ⎪ 𝑗=1 𝑖 𝑗=1 𝑖
 (𝑈 ) = ⎨ ⋮
⎪( )/( ) ∑𝑚−1 𝑗 ∗ ∑𝑚
⎪ ∑𝑚 1 ∑𝑚 1 𝑗=1
𝜙𝑖 𝑢𝑛𝑖+𝑗 𝜙𝑗 𝑢∗
𝑗=1 𝑖 𝑛𝑖+𝑗
The authors declare that they have no known competing financial ⎪ 𝑗=1 𝑗=1
, for ∑𝑛 < 𝑈 ≤ ∑𝑛 ,
⎪ 𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 ) 𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 )𝑓 (𝐶𝑖𝑗 ) 𝑗=1 𝑖
𝜙 𝑗
𝜙𝑗
𝑗=1 𝑖
interests or personal relationships that could have appeared to influence ⎪ ⋮
⎪( )/( ) ∑𝑛−1 𝑗 ∗
the work reported in this paper. ⎪ ∑𝑛 1 ∑𝑛 1 𝑗=1 𝜙𝑖 𝑢𝑛𝑖+𝑗
⎪ 𝑗=1 𝑗=1
, for ∑𝑛 < 𝑈 < 1,
⎪ 𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 ) 𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 )𝑓 (𝐶𝑖𝑗 ) 𝑗=1 𝜙𝑖
𝑗

⎪1 , for 𝑈 = 1.
CRediT authorship contribution statement ⎩
(A.5)
Tufan Ghosh: Conceptualization, Methodology, Formal analysis,
Writing - original draft, Writing - review & editing, Visualization. Carina This function can be interpreted as the weighted harmonic mean of
Bringedal: Conceptualization, Formal analysis, Validation, Software, the micro-scale fractional flow. The weight functions, 1∕𝑘𝑗𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 ), 𝑗 =
Writing - original draft, Writing - review & editing, Visualization. Rainer 1, 2, … , 𝑛 depend on the water relative permeability, which eventually,
Helmig: Conceptualization, Writing - review & editing, Supervision, depends on 𝐶𝑖𝑗 , 𝑗 = 1, 2, … , 𝑛.
Funding acquisition. G.P. Raja Sekhar: Formal analysis, Writing - re- The effective/upscaled Λ term is the harmonic average of the terms
𝑘𝑗𝑖 𝜆(𝐶𝑖𝑗 ), 𝑗 = 1, 2, … , 𝑛,
view & editing, Supervision.

⎧ 𝜙1𝑖 𝑢∗𝑛𝑖+1
Acknowledgement ⎪0 , for 0 ≤ 𝑈 ≤ ∑𝑛 ,
⎪ 𝑗
⎪ /( 𝑗=1 𝜙𝑖
) ∑
We thank the anonymous referees for useful suggestions which im- ⎪ ∑2 1 𝜙1𝑖 𝑢∗𝑛𝑖+1
2
𝜙𝑗 𝑢 ∗
𝑗=1 𝑖 𝑛𝑖+𝑗
proved the quality of the manuscript. One of the authors (TG) acknowl- ⎪ 2 𝑗=1 𝑗 , for ∑𝑛 < 𝑈 ≤ ∑𝑛 ,
⎪ 𝑘𝑖 𝜆(𝐶𝑖𝑗 ) 𝑗=1 𝜙𝑖
𝑗
𝑗=1 𝜙𝑖
𝑗
edges the support from the institute in the form of institute research ⎪ /( ) ∑2 𝑗 ∗ ∑3 𝑗 ∗
fellowship (Grant no: IIT/Acad/PGS&R/F.II/2/15/MA/90J09) of Indian ⎪ ∑3 1 𝑗=1 𝜙𝑖 𝑢𝑛𝑖+𝑗 𝑗=1 𝜙𝑖 𝑢𝑛𝑖+𝑗
⎪ 3 𝑗=1 𝑗 , for ∑𝑛 < 𝑈 ≤ ∑ 𝑛
,
Institute of Technology Kharagpur. We thank the Deutsche Forschungs- ⎪ 𝑘𝑖 𝜆(𝐶𝑖𝑗 ) 𝑗=1 𝜙𝑖
𝑗
𝑗=1 𝜙𝑖
𝑗

gemeinschaft (DFG, German Research Foundation) for supporting this Λ(𝑈 ) = ⎨ ⋮


⎪ /( ) ∑𝑚−1 𝑗 ∗ ∑𝑚 𝑗 ∗
work by funding SFB 1313, Project Number 327154368. The authors are ⎪ ∑𝑚 1 𝑗=1 𝜙𝑖 𝑢𝑛𝑖+𝑗 𝑗=1 𝜙𝑖 𝑢𝑛𝑖+𝑗
greatly thankful to Professor C.J. van Duijn for his valuable suggestions ⎪𝑚 𝑗=1 𝑗 , for ∑𝑛 < 𝑈 ≤ ∑𝑛 ,
⎪ 𝑘𝑖 𝜆(𝐶𝑖𝑗 ) 𝑗=1 𝜙𝑖
𝑗
𝑗=1 𝜙𝑖
𝑗

and critical remarks on this manuscript. ⎪ ⋮


⎪ /( ) ∑𝑛−1 𝑗 ∗
⎪ ∑𝑛 1 𝜙𝑢
𝑗=1 𝑖 𝑛𝑖+𝑗
Appendix A. Upscaled equations for n sub-layers within a cell ⎪ 𝑛 𝑗=1 𝑗
, for ∑𝑛 < 𝑈 < 1,
⎪ 𝑘𝑖 𝜆(𝐶𝑖𝑗 ) 𝑗
𝑗=1 𝜙𝑖
⎪1 , for 𝑈 = 1.
Here we present briefly the generalization of the results presented in ⎩
Section 3.3. In following we demonstrate the effective flow quantities (A.6)
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

Fig. A19. Schematic of a cell consisting of periodically distributed multiple sub-layers.

Fig. B20. Schematic of a cell consisting of ran-


domly distributed multiple sub-layers.

To express the effective/upscaled capillary pressure, we adopt a similar


relation as in (A.3). This means that, whenever oil is present on either ⎧𝑘 1 , if 𝑥𝑖 ∈ 𝑃1 ,
⎪ 𝑖2
side of all the interfaces, the first term in the expansion (3.8) is constant ⎪𝑘 𝑖 , if 𝑥𝑖 ∈ 𝑃2 ,
in the entire micro-scale cell. Hence we obtain ⎪ ⋮
√ √ √ 𝑘(𝑥𝑖 , 𝜔) = 𝑘(𝑇𝑗 (𝑥𝑖 )𝜔) = ⎨ 𝑛−1 (B.5)
√ 1
√𝜙
√ 2
√𝜙 𝜙𝑛𝑖 ⎪𝑘 𝑖 , if 𝑥𝑖 ∈ 𝑃𝑛−1(, )
𝑐 (𝑈 ) = √ 𝑖
𝐽 (𝐶𝑖 ) =
1 √ 𝑖
𝐽 (𝐶𝑖 ) = ⋯ =
2
𝐽 (𝐶𝑖𝑛 ). (A.7) ⎪ 𝑛 ∑𝑛−1
𝑘1 𝑘2 𝑘𝑛𝑖 ⎪𝑘 𝑖 , if𝑥𝑖 ∈ ℝ∖ 𝑗=1 𝑃𝑗 .
𝑖 𝑖 ⎩
To determine the effective diffusivity  we use (3.53), (A.5) and (A.6),
which implies Note that these are stationary random variables. Now

√ 1
√ 𝜙 𝐽 ′ (𝐺−1 (𝑈 ))
(𝑈 ) = Λ(𝑈 )𝑐 (𝑈 ), where 𝑐 (𝑈 ) = √ 𝑖
′ ′
. (A.8) 𝜙𝜖 (𝑥, 𝜔) = 𝜙(𝑇 (𝑥∕𝜖)𝜔) and 𝑘𝜖 (𝑥, 𝜔) = 𝑘(𝑇 (𝑥∕𝜖)𝜔). (B.6)
𝑘1𝑖 𝐺′ (𝐺−1 (𝑈 ))
Similarly one can obtain the effective flow quantities for the shifted Following similar steps as in Section 4, one can derive the effective equa-
approach in a cell with n sub-layers. tions for n randomly distributed sub-layers in a cell. When there are ran-
domly distributed, multiple transitions within a cell, we can define the
Appendix B. Upscaled equations for n randomly distributed average oil saturation as
sub-layers within a cell
∑𝑛−1 ( ∑ −1 ) 𝑛 𝑛
𝑦
∫𝑦 𝑛(𝑖+1) 𝜙(𝑦)𝑢0 (𝑥, 𝑦, 𝑡)𝑑𝑦 𝑗=1𝜓𝑗 𝜙𝑗𝑖 𝐶𝑖𝑗 + 1 − 𝑛𝑗=1 𝜓𝑗 𝜙𝑖 𝐶𝑖
Here we present briefly the generalization of the results presented 𝑛𝑖
𝑈 (𝑥, 𝑡) = 𝑦 = ∑𝑛−1 ( ∑ ) ,
in Section 4. We assume that the cell contains n randomly distributed ∫𝑦 𝑛(𝑖+1) 𝜙(𝑦)𝑑𝑦 𝑗 𝑛−1 𝑛
𝑛𝑖 𝑗=1 𝜓𝑗 𝜙𝑖 + 1 − 𝑗=1 𝜓𝑗 𝜙𝑖
sub-layers (see Fig. B.20). Again we consider a stationary ergodic ge-
ometrical structure, which is characterized by a probability space (Ω, (B.7)
𝜇), along with ergodic dynamical systems 𝑇𝑗 (𝑥), 𝑗 = 1, 2, … 𝑛 − 1, 𝑥 ∈ ℝ.
Now for a 𝜇-measurable subset 𝑗 ⊂ Ω, 𝑗 = 1, 2, … 𝑛 − 1, we introduce and the corresponding effective equation is given by
𝑃𝑗 = 𝑃𝑗 (𝜔) ⊂ ℝ by
{ } (𝑛−1 ( ) )
∑ 𝑛−1
∑ [ }]
𝑃𝑗 (𝜔) = 𝑥 ∈ ℝ ∶ 𝑇𝑗 (𝑥)𝜔 ∈ 𝑗 , (B.1) 𝜕𝑈 𝜕 𝜕 {
𝜓𝑗 𝜙𝑗𝑖 + 1− 𝜓𝑗 𝜙𝑛𝑖 +  (𝑈 ) − 𝑁𝑐 Λ(𝑈 ) 𝑐 (𝑈 ) = 0.
𝜕𝑡 𝜕𝑥 𝜕𝑥
which is a random stationary set. Here we assume that Pj (𝜔) is of the 𝑗=1 𝑗=1
form (B.8)
( )
𝑃𝑗 (𝜔) = ∪𝑖∈ℤ 𝑦𝑛𝑖+𝑗−1 , 𝑦𝑛𝑖+𝑗 , 𝑗 = 1, 2, … 𝑛 − 1, (B.2)
The effective saturation is expressed in terms of 𝐶𝑖𝑗 , 𝑗 = 1, 2, … , 𝑛. For
in which 𝑦𝑖 ∈ ℝ are strictly increasing random variables with respect to 𝐶𝑖𝑗 > 0, 𝑗 = 2, 3, … , 𝑛, we get a similar relation as in (A.3). This leads
i. Using the ergodic theorem by Birkhoff (1931), we get that there exists to
a density function of Pj , given by
1
∑𝑏 𝑈 = 𝐺(𝐶𝑖1 ) = ∑ ( ∑ −1 ) 𝑛
1 𝑛−1 𝑗
𝜓𝑗 ∶= 𝜇(𝑃𝑗 ) = lim ∣ 𝑦𝑛𝑖+𝑗 (𝜔) − 𝑦𝑛𝑖+𝑗−1 (𝜔) ∣, (B.3) 𝜓 𝜙
𝑗=1 𝑗 𝑖 + 1 − 𝑛𝑗=1 𝜓𝑗 𝜙𝑖
𝑎,𝑏→∞ 𝑦𝑛𝑏+𝑗−1 − 𝑦−𝑛𝑎+𝑗−1 −𝑎
⎧ √ √
for nearly all 𝜔 ∈ Ω, satisfying 0 ≤ 𝜓 j ≤ 1. The corresponding random 𝑛−1
∑ ⎛√ √
√ 1√ 𝑗 ⎞
⎪ 𝑗 −1 ⎜√ 𝜙𝑖 √ 𝑘𝑖 1 ⎟
porosity and absolute permeability are given by × ⎨𝜓1 𝜙𝑖 𝐶𝑖 +
1 1
𝜓𝑗 𝜙𝑖 𝐽 𝐽 ( 𝐶 𝑖 )
⎪ ⎜ 𝑘1 𝜙𝑗𝑖 ⎟

𝑗=2 ⎝ 𝑖 ⎠
⎧𝜙1 , if 𝑥𝑖 ∈ 𝑃1 ,
⎪ 𝑖2 ( ) √ √
⎪𝜙𝑖 , if 𝑥𝑖 ∈ 𝑃2 , 𝑛−1
∑ ⎛√
√ 𝜙1 𝑘𝑛 ⎞⎫
⎪ 𝑛 −1 ⎜√ 𝑖 𝑖 1 ⎟⎪
⋮ + 1− 𝜓𝑗 𝜙𝑖 𝐽 𝐽 ( 𝐶 𝑖 ⎟⎬.
) (B.9)
𝜙(𝑥𝑖 , 𝜔) = 𝜙(𝑇𝑗 (𝑥𝑖 )𝜔) = ⎨ 𝑛−1 (B.4) ⎜ 𝑘1 𝜙𝑛𝑖
⎪𝜙𝑖 , if 𝑥𝑖 ∈ 𝑃𝑛−1(, )
𝑗=1 ⎝ 𝑖 ⎠⎪⎭
⎪ 𝑛 ∑𝑛−1
⎪𝜙𝑖 , if 𝑥𝑖 ∈ ℝ∖ 𝑗=1 𝑃𝑗 ,

T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

( )
𝜓1 𝜙1𝑖 𝑢∗𝑛𝑖+1
Notice that G defined above maps (0, 1) onto ∑𝑛−1 ( ∑ ) ,1 . Since J is strictly increasing, the function G shares the same property and
𝑗 𝑛−1 𝑛
𝑗=1 𝜓𝑗 𝜙𝑖 + 1− 𝑗=1 𝜓𝑗 𝜙𝑖

can therefore be inverted. Hence one can write 𝐶𝑖1 in terms of U as in (3.53).
Considering the above expressions one can identify the effective/upscaled fractional flow as

⎧ 𝜓1 𝜙1𝑖 𝑢∗𝑛𝑖+1
⎪0 , for 0 ≤ 𝑈 ≤ ∑
𝑛−1 𝑗
( ∑𝑛−1 ) 𝑛 ,
⎪ 𝑗=1 𝜓𝑗 𝜙𝑖 + 1 − 𝑗=1 𝜓𝑗 𝜙𝑖
⎪( )/( ) ∑2 𝑗 ∗
𝜓 𝜙 𝑗=1 𝜓𝑗 𝜙𝑖 𝑢𝑛𝑖+𝑗
1 𝑢∗
⎪ ∑2 𝜓𝑗 ∑2 𝜓 𝑗 1 𝑖 𝑛𝑖+1
⎪ 𝑗=1 𝑗 𝑗=1 𝑗 , for ∑𝑛−1 ( ∑ −1 ) < 𝑈 ≤ ∑𝑛−1 ( ∑ −1 ) 𝑛 ,
⎪ 𝑘𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 ) 𝑘𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 )𝑓 (𝐶𝑖𝑗 ) 𝜓𝑗 𝜙𝑗𝑖 + 1 − 𝑛𝑗=1 𝜓𝑗 𝜙𝑛𝑖 𝜓𝑗 𝜙𝑗𝑖 + 1 − 𝑛𝑗=1 𝜓𝑗 𝜙𝑖
⎪( 𝑗=1 𝑗=1
)/( ) ∑2 𝑗 ∗ ∑3 𝑗 ∗
⎪ ∑ 𝜓𝑗 ∑3 𝜓𝑗 𝑗=1 𝜓𝑗 𝜙𝑖 𝑢𝑛𝑖+𝑗 𝑗=1 𝜓𝑗 𝜙𝑖 𝑢𝑛𝑖+𝑗
⎪ 3
, for ∑
⎪ 𝑗=1 𝑗
𝑘𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 )
𝑗=1 𝑗
𝑘𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 )𝑓 (𝐶𝑖𝑗 ) 𝑛−1 𝑗
( ∑𝑛−1 ) 𝑛 < 𝑈 ≤ ∑𝑛−1 𝑗
( ∑𝑛−1 ) 𝑛 ,
⎪ 𝑗=1 𝜓𝑗 𝜙𝑖 + 1 − 𝑗=1 𝜓𝑗 𝜙𝑖 𝑗=1 𝜓𝑗 𝜙𝑖 + 1 − 𝑗=1 𝜓𝑗 𝜙𝑖
⎪ ⋮


 ( 𝑈 ) = ⎨( )/( ) ∑𝑚−1 𝑗 ∗ ∑𝑚 𝑗 ∗ (B.10)
∑𝑚 𝜓𝑗 ∑𝑚 𝜓𝑗 𝑗=1 𝜓𝑗 𝜙𝑖 𝑢𝑛𝑖+𝑗 𝑗=1 𝜓𝑗 𝜙𝑖 𝑢𝑛𝑖+𝑗
⎪ , < 𝑈 ≤
⎪ 𝑗=1 𝑗
𝑘𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 )
𝑗=1 𝑗
𝑘𝑖 𝑘𝑟𝑤 (𝐶𝑖𝑗 )𝑓 (𝐶𝑖𝑗 )
for ∑𝑛−1 𝑗
( ∑𝑛−1 )
𝑛 ∑𝑛−1 𝑗
( ∑𝑛−1 ) 𝑛 ,
⎪ 𝑗=1 𝜓 𝑗 𝜙𝑖 + 1 − 𝑗=1 𝜓 𝑗 𝜙 𝑖 𝑗=1 𝜓 𝑗 𝜙𝑖 + 1 − 𝑗=1 𝜓𝑗 𝜙𝑖
⎪ ⋮

⎪ ∑ −1
⎪ ∑𝑛−1 𝜓𝑗 1 − 𝑛𝑗=1 𝜓𝑗
⎪ + ∑𝑛−1
𝑗=1 𝑗 𝑗 𝑛 𝑘 (𝐶 𝑛 ) 𝑗 ∗
⎪ 𝑘𝑖 𝑘𝑟𝑤 (𝐶𝑖 ) 𝑘 𝑖 𝑟𝑤 𝑖 𝑗=1 𝜓𝑗 𝜙𝑖 𝑢𝑛𝑖+𝑗
⎪ ,
⎪ ∑ for ∑ ( ∑𝑛−1 ) 𝑛 < 𝑈 < 1,
𝜓𝑗 1 − 𝑛𝑗=1
−1
𝜓𝑗 𝑛−1 𝑗
⎪ ∑𝑛−1 𝑗=1 𝜓𝑗 𝜙𝑖 + 1 − 𝑗=1 𝜓𝑗 𝜙𝑖
+
⎪ 𝑗=1 𝑘𝑗 𝑘 (𝐶 𝑗 )𝑓 (𝐶 𝑗 ) 𝑘𝑗 𝑘 (𝐶 𝑛 )𝑓 (𝐶 𝑛 )
⎪ 𝑖 𝑟𝑤 𝑖 𝑖 𝑖 𝑟𝑤 𝑖 𝑖
⎩1 , for 𝑈 = 1.

One may observe that this function is again the weighted harmonic mean of the micro-scale fractional flow function, as in periodic layer case. In
addition, the effective/upscaled Λ term is also the harmonic average of the micro-scale quantities, and given by

⎧ 𝜓1 𝜙1𝑖 𝑢∗𝑛𝑖+1
⎪0 , for 0 ≤ 𝑈 ≤ ∑
⎪ 𝑛−1 𝑗
( ∑𝑛−1 ) 𝑛 ,
⎪ 𝑗=1 𝜓𝑗 𝜙𝑖 + 1 − 𝑗=1 𝜓𝑗 𝜙𝑖
⎪( /( ) ∑2 𝑗 ∗
⎪ ∑2 ) ∑2 𝜓𝑗 𝜓1 𝜙1𝑖 𝑢∗𝑛𝑖+1 𝑗=1 𝜓𝑗 𝜙𝑖 𝑢𝑛𝑖+𝑗
⎪ 𝑗=1 𝜓𝑗 𝑗=1 𝑗 , for ∑ ( ∑𝑛−1 ) < 𝑈 ≤ ∑𝑛−1 ( ∑𝑛−1 ) 𝑛 ,
𝑘𝑖 𝜆(𝐶𝑖𝑗 ) 𝑛−1 𝑗 𝑛 𝑗
⎪ 𝑗=1 𝜓𝑗 𝜙𝑖 + 1 − 𝑗=1 𝜓𝑗 𝜙𝑖 𝑗=1 𝜓𝑗 𝜙𝑖 + 1 − 𝑗=1 𝜓𝑗 𝜙𝑖
⎪ /( ) ∑2 𝑗 ∗ ∑3 𝑗 ∗
⎪(∑3 ) ∑3 𝜓𝑗 𝑗=1 𝜓𝑗 𝜙𝑖 𝑢𝑛𝑖+𝑗 𝑗=1 𝜓𝑗 𝜙𝑖 𝑢𝑛𝑖+𝑗
⎪ 𝜓
𝑗=1 𝑗 𝑗=1 𝑗 , for ∑ ( ∑𝑛−1 ) 𝑛 < 𝑈 ≤ ∑𝑛−1 ( ∑𝑛−1 ) 𝑛 ,
⎪ 𝑘𝑖 𝜆(𝐶𝑖𝑗 ) 𝑛−1 𝑗 𝑗
𝑗=1 𝜓𝑗 𝜙𝑖 + 1 − 𝑗=1 𝜓𝑗 𝜙𝑖 𝑗=1 𝜓𝑗 𝜙𝑖 + 1 − 𝑗=1 𝜓𝑗 𝜙𝑖

⎪ ⋮
Λ(𝑈 ) = ⎨ (B.11)
⎪ /( ) ∑𝑚−1 𝑗 ∗ ∑𝑚 𝑗 ∗
⎪(∑𝑚 ) ∑𝑚 𝜓𝑗 𝑗=1 𝜓𝑗 𝜙𝑖 𝑢𝑛𝑖+𝑗 𝑗=1 𝜓𝑗 𝜙𝑖 𝑢𝑛𝑖+𝑗
⎪ 𝜓
𝑗=1 𝑗 𝑗=1 𝑗 , for ∑ ( ∑𝑛−1 ) < 𝑈 ≤ ∑𝑛−1 ( ∑𝑛−1 ) 𝑛 ,
⎪ 𝑘𝑖 𝜆(𝐶𝑖𝑗 ) 𝑛−1 𝑗 𝑛 𝑗
𝑗=1 𝜓𝑗 𝜙𝑖 + 1 − 𝑗=1 𝜓𝑗 𝜙𝑖 𝑗=1 𝜓𝑗 𝜙𝑖 + 1 − 𝑗=1 𝜓𝑗 𝜙𝑖

⎪ ⋮

⎪ / ∑ −1 ⎞ ∑𝑛−1
⎪ ⎛
∑ −1 𝜓𝑗 1 − 𝑛𝑗=1 𝜓𝑗 𝜓𝑗 𝜙𝑗𝑖 𝑢∗𝑛𝑖+𝑗
𝑗=1
⎪1 ⎜ 𝑛𝑗=1 ⎟,
⎪ ⎜ 𝑗 𝑗
𝑘𝑖 𝜆(𝐶𝑖 )
+ 𝑛 𝑛
𝑘𝑖 𝜆(𝐶𝑖 ) ⎟
for ∑
𝑛−1 𝑗
( ∑𝑛−1 ) 𝑛 < 𝑈 < 1,
⎪ ⎝ ⎠ 𝑗=1 𝜓𝑗 𝜙𝑖 + 1 − 𝑗=1 𝜓𝑗 𝜙𝑖
⎪1 , for 𝑈 = 1.

With similar arguments as in the periodic case, one can also define the effective capillary pressure and the effective diffusivity following the
Eqs. (A.7) and (A.8) respectively.
Note that for n sub-layers in a cell, the periodic case corresponds to 𝜓𝑗 = 1𝑛 for all 𝑗 = 1, 2, … , 𝑛 − 1; hence for each realization 𝜔 we end up with
an effective equation similar to the periodic case (Eq. (A.2)). In addition, one may observe that for 𝜓𝑗 = 1𝑛 for all 𝑗 = 1, 2, … , 𝑛 − 1, the average oil
saturation relation (B.7) reduces to the relation (A.1), which is the average oil saturation for the periodic case. Moreover, similar observation also
holds for other upscaled quantities.
T. Ghosh, C. Bringedal and R. Helmig et al. Advances in Water Resources 145 (2020) 103716

References Lasseux, D., Quintard, M., Whitaker, S., 1996. Determination of permeability tensors for
two-phase flow in homogeneous porous media: theory. Transp. Porous Media 24 (2),
Ahmadi, A., Arani, A.A.A., Lasseux, D., 2010. Numerical simulation of two-phase inertial 107–137.
flow in heterogeneous porous media. Transp. Porous Media 84 (1), 177–200. Leverett, M.C., 1941. Capillary behavior in porous solids. Trans. AIME Petr. Eng. Div. 142,
Amaziane, B., Antontsev, S., Pankratov, L., Piatnitski, A., 2010. Homogenization of im- 152–169.
miscible compressible two-phase flow in porous media: application to gas migration Lewandowska, J., Szymkiewicz, A., Burzyński, K., Vauclin, M., 2004. Modeling of unsatu-
in a nuclear waste repository. Multiscale Model. Simul. 8 (5), 2023–2047. rated water flow in double porosity soils by the homogenization approach. Adv. Water
Auriault, J., Boutin, C., Geindreau, C., 2009. Homogenization of Coupled Phenomena in Resour. 27 (3), 321–345.
Heterogenous Media. John Wiley & Sons, Inc., New Jersey, USA. Mikelić, A., 2000. Homogenization theory and applications to filtration through porous
Beliaev, A., 2003. Homogenization of two-phase flows in porous media with hysteresis in media. In: Fasano, A. (Ed.), Filtration in Porous Media and Industrial Application. In:
the capillary relation. Eur. J. Appl. Math. 14, 61–84. Lecture Notes in Math. 1734. Springer-Verlag, Berlin, pp. 127–214.
Berryman, J.G., 2005. Comparison of upscaling methods in poroelasticity and its general- Mouche, E., Hayek, M., Mügler, C., 2010. Upscaling of CO2 vertical migration through
izations. J. Eng. Mech. 131 (9), 928–936. a periodic layered porous medium: the capillary-free and capillary-dominant cases.
Bertsch, M., Dal Passo, R., van Duijn, C.J., 2003. Analysis of oil trapping in porous media Adv. Water Resour. 33, 1164–1175.
flow. SIAM J. Math. Anal. 35, 245–267. Neuweiler, I., Cirpka, O.A., 2005. Homogenization of richards equation in permeability
Birkhoff, G.D., 1931. Proof of the ergodic theorem. Proc. Nat. Acad. Sci. USA 17, 656–660. fields with different connectivities. Water Resour. Res. 41, W02009.
Bourgeat, A., Hidani, A., 1995. Effective model of two-phase flow in a porous medium Papanicolaou, G.C., 1995. Diffusion in random media. In: Keller, J.B., Mc Laughlin, D.W.,
made of different rock types. Appl. Anal. 58, 1–29. Papanicolaou, G.C. (Eds.), Surveys in Applied Mathematics, Vol. I. Plenum Press, New
Bourgeat, A., Panfilov, M., 1998. Effective two-phase flow through highly heterogeneous York, pp. 205–254.
porous media: capillary nonequilibrium effects. Comput. Geosci. 2, 191–215. Pop, I.S., Yong, W.A., 2002. A numerical approach to degenerate parabolic equations.
Brenner, K., Cances, C., Hilhorst, D., 2013. Finite volume approximations for an immisci- Numer. Math. 92, 357–381.
ble two-phase flow in porous media with discontinuous capillary pressure. Comput. Quintard, M., Whitaker, S., 1988. Two-phase flow in heterogeneous porous media: the
Geosci. 17, 573–597. method of large-scale averaging. Transp. Porous Media 3 (4), 357–413.
Buzzi, F., Lenzinger, M., Schweizer, B., 2009. Interface conditions for degenerate Quintard, M., Whitaker, S., 1990. Two-phase flow in heterogeneous porous media I: the in-
two-phase flow equations in one space dimension. Analysis 29, 299–316. fluence of large spatial and temporal gradients. Transp. Porous Media 5 (4), 341–379.
Cances, C., 2010. On the effects of discontinuous capillarities for immiscible two phase Quintard, M., Whitaker, S., 1990. Two-phase flow in heterogeneous porous media II: nu-
flows in porous media made of several rock-types. Netw. Heterog. Media 5, 635–647. merical experiments for flow perpendicular to a stratified system. Transp. Porous Me-
Cioranescu, D., Donato, P., 1999. An Introduction to Homogenization. Oxford University dia 5 (5), 429–472.
Press Inc., New York, USA. Schweizer, B., 2008. Homogenization of degenerate two-phase flow equations with oil
Dale, M., Ekrann, S., Mykkeltveit, J., Virnovsky, G., 1997. Effective relative permeabili- trapping. SIAM J. Math. Anal. 39, 1740–1763.
ties and capillary pressure for one-dimensional heterogeneous media. Transp. Porous Schweizer, B., Pop, I.S., 2007. On the homogenization of the Buckley-Leverett equa-
Media 26, 229–260. tion including trapping effects at the micro scale. Proc. Appl. Math. Mech. 7,
Daley, D.J., Vere-Jones, D., 1988. An Introduction to the Theory of Point Processes. 1041403–1041404.
Springer-Verlag„ New York. Szymkiewicz, A., Helmig, R., Kuhnke, H., 2011. Two-phase flow in heterogeneous porous
Darman, N.H., Pickup, G.E., Sorbie, K.S., 2002. A comparison of two-phase flow dynamic media with non-wetting phase trapping. Transp. Porous Media 86, 27–47.
upscaling methods based on fluid potentials. Comp. Geosci. 6, 5–27. Szymkiewicz, A., Helmig, R., Neuweiler, I., 2012. Upscaling unsaturated flow in binary
Davit, Y., Bell, C.G., Byrne, H.M., Chapman, L.A.C., Kimpton, L.S., Lang, G.E., porous media with air entry pressure effects. Water Resour. Res. 48, W04522.
Leonard, K.H.L., Oliver, J.M., Pearson, N.C., Shipley, R.J., Waters, S.L., Whiteley, J.P., Szymkiewicz, A., Helmig, R., Neuweiler, I., 2014. Influence of heterogeneous air en-
Wood, B.D., Quintard, M., 2013. Homogenization via formal multiscale asymptotics try pressure on large scale unsaturated flow in porous media. Acta Geophys. 62,
and volume averaging: how do the two techniques compare? Adv. Water Resour. 62, 1179–1191.
178–206. van Duijn, C.J., Cao, X., Pop, I.S., 2016. Two-phase flow in porous media: dynamic capil-
Duijn, C.J., Molenaar, J., de Neef, M.J., 1995. The effect of capillary forces on immiscible larity and heterogeneous media. Transp. Porous Media 114, 283–308.
two-phase flow in heterogeneous porous media. Transp. Porous Media 21, 71–93. van Duijn, C.J., Eichel, H., Helmig, R., Pop, I.S., 2007. Effective equations for two-phase
Enchery, G., Eymard, R., Michel, A., 2006. Numerical approximation of a two-phase flow flow in porous media: the effect of trapping on the microscale. Transp. Porous Media
problem in a porous medium with discontinuous capillary forces. SIAM J. Numer. 69, 411–428.
Anal. 43, 2402–2422. van Duijn, C.J., Mikelić, A., Pop, I.S., 2002. Effective equations for two-phase flow with
Ern, A., Mozolevski, I., Schuh, L., 2010. Discontinuous Galerkin approximation of trapping on the microscale. SIAM J. Appl. Math. 62 (5), 1531–1568.
two-phase flows in heterogeneous porous media with discontinuous capillary pres- van Lingen, P., 1998. Quantification and Reduction of Capillary entrapment in Cross-Lam-
sures. Comput. Methods Appl. Mech. Eng. 199, 1491–1501. inated Oil Reservoirs. Sub-Faculty of Applied Earth Sciences, Delft University of Tech-
Evje, S., Karlsen, K.H., 2000. Monotone difference approximations of BV solutions to de- nology Ph.d. thesis.
generate convection-diffusion equations. SIAM J. Numeri. Anal. 37, 1838–1860. Virnovsky, G., Friis, H., Lohne, A., 2004. A steady-state upscaling approach for immiscible
Henning, P., Ohlberger, M., Schweizer, B., 2013. Homogenization of the degenerate two-phase flow. Transp. Porous Med. 54, 167–192.
two-phase flow equations. Math. Models Methods Appl. Sci. 23 (12), 2323–2352. Wen, X.H., Gómez-Hernández, J.J., 1996. Upscaling hydraulic conductivities in heteroge-
Henning, P., Ohlberger, M., Schweizer, B., 2015. Adaptive heterogeneous multiscale meth- neous media: an overview. J. Hydrol. 183, ix–xxxii.
ods for immiscible two-phase flow in porous media. Comput. Geosci. 19, 99–114. Whitaker, S., 1986. Flow in porous media II: the governing equations for immiscible,
Hornung, U., 1997. Homogenization and porous media. Interdisciplinary Applied Mathe- two-phase flow. Transp. Porous Media 1 (2), 105–125.
matics, Vol. 6. Springer-Verlag, New York, USA. Whitaker, S., 1994. The closure problem for two-phase flow in homogeneous porous me-
Jaffré, J., 1995. Flux calculation at the interface between two rock types for two-phase dia. Chem. Eng. Sci. 49 (5), 765–780.
flow in porous media. Transp. Porous Media 21, 195–207. Yeh, L.M., 2006. Homogenization of two-phase flow in fractured media. Math. Models
Kortekaas, T.F.M., 1985. Water/oil displacement characteristics in crossbedded reservoir Methods Appl. Sci. 16, 1627–1651.
zones. Soc. Pet. Eng. J. 25 (6), 917–926. Zijl, W., Trykozko, A., 2002. Numerical homogenization of two-phase flow in porous me-
Kueper, B.H., Abbott, W., Farquhar, G., 1989. Experimental observations of multiphase dia. Comput. Geosci. 6, 49–71.
flow in heterogeneous porous media. J. Contam. Hydrol. 5, 83–95.
Lasseux, D., Ahmadi, A., Arani, A.A.A., 2008. Two-phase inertial flow in homogeneous
porous media: a theoretical derivation of a macroscopic model. Transp. Porous Media
75 (3), 371–400.

You might also like