Upscaling of Geocellular Models for

Reservoir Flow Simulation:

A Review of Recent Progress
Louis J. Durlofsky
Department of Petroleum Engineering, Stanford University, Stanford, CA
ChevronTexaco Exploration & Production Technology Company,
San Ramon, CA
Paper presented at 7
th
International Forum on Reservoir Simulation
B uhl/Baden-Baden, Germany, June 23-27, 2003
Abstract
A variety of approaches for the upscaling of detailed geocellular models for ow simu-
lation are reviewed. Local, extended local, global and quasi global techniques for the
calculation of equivalent permeability and transmissibility are discussed. Special-
ized procedures for permeability upscaling in the vicinity of wells are also described.
Flow-based grid generation techniques, in which a curvilinear grid is introduced to
resolve the eects of permeability connectivity, are described along with procedures
for computing upscaled properties for such grids. Coarse scale simulation results
generated using many of these upscaling techniques are presented. These results il-
lustrate the capabilities of existing upscaling procedures and demonstrate the levels
of accuracy attainable using the various approaches.
1 Introduction
Typical reservoir simulators can handle on the order of 10
5
10
6
simulation cells.
The exact number will vary considerably depending on the type of simulation to be
performed (dead oil, black oil, compositional) and the available computer hardware.
For many types of models, production version simulations with over 10
6
cells are still
not performed routinely. Geological characterizations, by contrast, typically contain
on the order of 10
7
10
8
cells. These models, which are referred to as ne grid models,
geostatistical models or simply geocellular models, represent geological variation on
very ne scales vertically, though their areal resolution is still relatively coarse. For
example, a typical geostatistical model might contain layering of thickness 1 ft or
less, though cell sizes in the areal direction might be about 50 - 100 ft. Thus, ne
1
grid geological descriptions can be expected to grow further, so the need for reliable
upscaling techniques will continue.
Another issue of considerable importance is the need for the assessment of risk
and uncertainty in reservoir performance. Nearly every aspect of the reservoir char-
acterization contains some degree of uncertainty, so predictions are necessarily of
a statistical character. The uncertainty in reservoir performance can be gauged by
simulating a number of dierent geological realizations or scenarios. Thousands of
such runs may be required to cover the range of parameter variation. It is not
computationally feasible (or desirable) to perform these simulations on the ne grid
model. Signicantly upscaled models are required if a full assessment of project risk
and uncertainty is to be accomplished. These models should ideally be even coarser
than the O(10
5
10
6
) grid block models referred to above. In addition, if thousands
of coarse models are to be simulated, the upscaling must be highly automated.
Dierent upscaling procedures are appropriate in dierent situations. The ideal
procedure to use on a particular problem depends on the simulation question being
addressed, the production mechanism, and the level of detail that can be accommo-
dated in the coarse model. For a problem involving primary production with only
oil being produced, the coarse model should correctly capture the eects of near-
well heterogeneity as well as the general large scale ow response of the reservoir.
For scenarios involving displacement of oil by water or gas, it may be important to
accurately capture the eects of key ow paths between injection and production
wells. This may require the use of specialized gridding procedures.
Following the description of the governing equations below, we classify upscaling
techniques in terms of the types of parameters that are upscaled (single or two-
phase ow parameter upscaling). For moderate degrees of coarsening, single-phase
upscaling techniques often provide acceptable results, particularly when used in con-
junction with specialized (ow-based) coarse grids. At higher degrees of coarsening,
some type of relative permeability upscaling is also generally required. In order to
maintain focus, this review will cover only techniques for single-phase upscaling and
for the generation of coarse grids. The calculation of upscaled relative permeabil-
ities is an important topic that has been addressed in several recent reviews and
comparison studies; see for example Christie [16], Barker and Thibeau [8], Barker
and Dupouy [7] and Darman et al. [18]. Recent reviews on upscaling with a focus
closer to that of this review include those by Wen and G omez-Hernandez [77], Re-
nard and de Marsily [67] and Farmer [32]. Our emphasis here will be on methods
implemented for structured grids, though many of the methods discussed could be
readily adapted to unstructured grids.
This review proceeds as follows. In 2 we present the equations governing single
and two-phase ow on the ne scale and then describe some ways in which upscaling
procedures can be classied. We present the coarse scale pressure equation and in-
troduce the upscaled permeability tensor k

in 3. Analytical upscaling procedures

2
are then briey considered. In 4 we discuss numerical methods for the local cal-
culation of k

. We consider various boundary conditions for these calculations and

describe dierent post-processing procedures for computing k

from the local solu-

tions. Next, we describe extended local upscaling techniques (e.g., the use of border
regions) and the determination of k

for irregular coarse grid control volumes. In

5 we discuss a number of techniques based on transmissibility upscaling. These
include global procedures, a local-global technique, and near-well upscaling. Flow-
based grid generation procedures are considered in 6. We briey discuss approaches
for assessing the quality of the upscaled model in 7. In 8, numerical results for
many of the procedures described in previous sections are presented. Results using
local, extended local and local-global techniques are included, as are results for up-
scaling to irregular coarse cells. We also display results using near-well upscaling
and ow-based gridding. We conclude in 9 with a summary and a discussion of
potential future directions for research in the area of upscaling.
2 Fine scale equations and classications of
upscaling procedures
We now consider the equations describing single and two-phase ow on the ne
scale. These equations are strictly appropriate for use with a fully resolved or ne
scale geocellular model; i.e., one in which no upscaling has been performed.
2.1 Single-phase ow equations
The equation governing single-phase ow in the absence of gravity is formed by
combining Darcys law
u =
1

k p, (1)
with a statement of mass conservation

t
() + (u) + m = 0. (2)
The resulting equation, referred to as the pressure equation, is given by:

t
()
_

k p
_
+ m = 0. (3)
In Eqs. (1)-(3), u is the Darcy velocity, is viscosity, k is the (symmetric positive
denite) permeability tensor, p is pressure, t is time, is porosity, is density and
3
m is the source/sink term (positive for production) expressed as a mass ow rate
per unit volume.
Upscaling procedures are generally formulated based on a simplied form of
Eq. (3). Specically, if we assume that the uid and rock are incompressible (i.e.,
does not vary in space or time and /t = 0), we obtain a simplied pressure
equation:

_
1

k p
_
= q, (4)
where q = m/ is the volumetric source term. The equations above can be modied
to account for the eects of gravity by replacing p with the potential = p gz,
where g is the gravitational acceleration and the z axis is assumed to point vertically
downwards (in the direction of gravity), as is the convention for most reservoir
simulators.
2.2 Two-phase ow equations
The equations describing two-phase ow on the ne scale can again be formed by
combining Darcys law with a statement of mass conservation. In this case, in the
absence of gravity, Darcys law can be expressed as:
u
j
=
k
rj

j
k p
j
, (5)
where the subscript j refers to the phase (j = w for water and j = o for oil) and k
rj
is the relative permeability to phase j. Mass conservation is given by:

t
(
j
S
j
) + (
j
u
j
) + m
j
= 0, (6)
where S
j
is the saturation (volume fraction) of phase j. If we assume that /t = 0,
that
j
does not vary in time or space, and that capillary pressure (p
c
) is negligible;
i.e., p
c
(S
w
) = p
o
p
w
= 0, we obtain:
u
t
= q
t
, (7)
where q
j
= m
j
/
j
and the total volumetric source term is q
t
= q
w
+ q
o
. The total
Darcy velocity u
t
is given by:
u
t
= u
w
+u
o
= k
_
k
rw

w
+
k
ro

o
_
p. (8)
The water velocity u
w
can now be expressed as u
w
= f(S
w
)u
t
where f(S
w
) is the
usual Buckley-Leverett fractional ow function. Inserting this form for u
w
in Eq. (6),
we can write the water saturation equation as:

S
w
t
+ [u
t
f(S
w
)] = q
w
. (9)
4
The corresponding pressure equation can be formed by introducing the expression
for u
t
(Eq. (8) into Eq. (7)):
(k
t
(S
w
) p) = q
t
, (10)
where we have introduced the total mobility
t
, dened via:

t
=
w
+
o
=
k
rw

w
+
k
ro

o
. (11)
The pressure and saturation equations describe the ow of two immiscible uids.
The pressure equation (10) is very similar to the single-phase pressure equation
(4) except the
t
term replaces 1/. The single-phase limit is recovered if the two
phases have identical properties and do not interfere. For such a system, k
rw
= S
w
,
k
ro
= S
o
= 1 S
w
, and
w
=
o
, giving
t
= 1/
w
= 1/
o
and f(S
w
) = S
w
(in
Eq. (9)). In this case, the system is of unit mobility ratio.
More general models for two-phase ow include the eects of compressibility,
gravity and capillary pressure. The representation of the system in terms of pressure
and saturation equations is still possible; see Peaceman [60] and Aziz and Settari [5]
for details.
2.3 Classications of upscaling procedures
There are several dierent ways in which upscaling techniques can be classied.
Here we will classify the various methods in terms of the types of parameters that
are upscaled (single or two-phase ow parameters) and the way in which these
parameters are computed (e.g., using local or global calculations). We rst consider
classication by the types of parameters that are upscaled.
For single-phase ow involving a single component, the only parameters to be
upscaled are porosity and the absolute permeability or transmissibility (transmis-
sibility appears in the discretized ow equations). In the more general case of
two-phase ow, the absolute permeability (or transmissibility) and porosity as well
as the relative permeability can be upscaled. However, in many cases it is possible
to develop reasonably accurate coarse scale models for two-phase ow with only the
absolute permeability (or transmissibility) and porosity upscaled, particularly when
accurate upscaling is used in conjunction with ow-based grid generation. In mod-
els of this type, the geocellular scale relative permeabilities are used directly on the
coarse scale. Thus, even for two (or three) phase ow systems, we can still generate
coarse scale models with only absolute permeability or transmissibility and porosity
upscaled in some cases. We refer to this type of approach as single-phase parameter
upscaling or single-phase upscaling, with the understanding that it can be used
for both single and two-phase (or multiphase) ow problems. Such approaches can
5
be further classied as permeability or transmissibility upscaling procedures. We
will in fact distinguish between permeability and transmissibility upscaling in 4
and 5 below.
In other cases, the two-phase ow parameters (e.g., k
rj
and p
c
) are also upscaled.
We refer to this type of approach as a two-phase parameter upscaling or simply
two-phase upscaling. Another way to view this classication is in terms of the
governing pressure and saturation equations. In single-phase upscaling, the pres-
sure equation is modied but the saturation equation appears essentially the same
(though is upscaled). In two-phase upscaling, by contrast, parameters in both
equations are modied.
The second type of classication is related to the way in which upscaled param-
eters (single or two-phase) are computed. In all cases the intent of the upscaling
procedure is to replace the ne model with a coarse model. In a purely local pro-
cedure, coarse scale parameters are computed by considering only the ne scale
region corresponding to the target coarse block. No additional ne scale informa-
tion is included in the upscaling calculation. In a global upscaling technique, the
entire ne scale model is simulated for the calculation of the coarse scale parameters.
The assumption here is that the coarse scale parameters will be applicable to other
(related) ow scenarios.
There are several important variants of the purely local and global approaches.
Of interest to us will be extended local and quasi global upscaling techniques. In
extended local procedures, coarse scale parameters are computed by considering the
region corresponding to the target coarse block plus a ne scale border region or
skin around this region. Coarse scale quantities are generally computed by averaging
the ne scale solution only over the region corresponding to the target coarse block.
In quasi global methods, global ow data informs the upscaling technique but this
information is only approximate. For example, in the case of a quasi global two-
phase parameter upscaling, the global ow eld might be estimated from a single
solution of the single-phase pressure equation rather than by a more computationally
expensive transient solution of the two-phase ow equations. We will discuss local,
extended local, global and quasi global upscaling techniques later in this review.
2.4 Multiscale methods
Another important class of upscaling procedures are multiscale nite element and
nite volume approaches. Within the context of reservoir simulation, for exam-
ple, Hou and Wu [42] and Arbogast and Bryant [3] developed nite element based
approaches, while Jenny et al. [43] developed an approach based on a full tensor
ux-continuous nite volume procedure. These methods have the benet that the
ne scale permeability information enters into the global solution in a systematic
way. To date, these procedures have mainly been applied to the pressure equation;
6
transport calculations have generally been performed by reconstructing the ne scale
velocity eld. This reconstruction, which can also be accomplished using other pro-
cedures (see, e.g., [66, 37, 34, 14]), entails the use of the ne grid for the solution of
the saturation equation.
Dual grid approaches (e.g., [66, 37, 34]) are also related to multiscale methods.
In these procedures, dierent grids are dened and reconstruction techniques are
applied to determine ne scale variables from the global coarse scale solution. These
methods can provide accurate results in many cases, though they, like multiscale
procedures, require that the ne scale permeability information be stored and used
during the global solution.
3 Coarse scale pressure equation and
analytical upscaling methods
We now consider the upscaling of absolute permeability. As indicated above, even
for many two-phase ow problems, particularly when the overall level of coarsening
is not excessive (e.g., 1-2 order of magnitude reduction in the total number of cells
for a typical problem), this has been found to be a reasonable approach. Some
theoretical justication for neglecting the upscaling of relative permeability, in the
context of moderately coarsened models and ow-based grids, is oered by Durlofsky
[23]. As the degree of coarsening becomes very high, however, upscaled two-phase
ow functions will generally be required.
3.1 Fine and coarse scale pressure equations
For now we consider steady, single-phase incompressible ow with no source terms
(Eq. (4) with q = 0). We introduce a conceptual two-scale model for permeability;
i.e., permeability varies on two distinct scales referred to as x and y. The x scale
is a slow scale, meaning that variations in x are relatively gradual. The y scale, by
contrast, is a fast scale, and captures the ne scale variation of permeability. With
this k(x, y) representation, the dimensionless pressure equation is written as follows:
(k(x, y) p) = 0. (12)
Eq. (12) is the pressure equation with the permeability eld fully resolved. Homoge-
nization procedures, as applied by Bourgeat [10] and Saez et al. [69], among others,
allow Eq. (12) to be replaced with an analogous equation in which variations need
only be resolved on the x (slow) scale. The pressure equation in this case can be
written as:
(k

(x) p
c
) = 0, (13)
7
where k

is referred to as the eective permeability tensor and p

c
is the coarse scale
pressure. Note that k

is dened on the scale of x; y-scale variations, which exist

in k, have been homogenized or averaged. This means that, in solving Eq. (13), we
need not resolve eects on the scale of y, which leads to signicant computational
savings.
The upscaling of k(x, y) to k

(x), and the replacement of Eq. (12) by Eq. (13), is

mathematically valid only in certain circumstances [10, 69]. Specically, the region
over which k

is computed must be large relative to the fast (y) scale of variation

(e.g., the correlation length of the heterogeneity eld). In addition, Eq. (13) is
technically not applicable near boundaries or sources. However, even though there
is often not a clear theoretical justication for replacing Eq. (12) by Eq. (13), there
is a large body of numerical evidence justifying this procedure. The coarse scale
permeability in these more general cases is properly referred to as the equivalent
grid block permeability tensor rather than an eective permeability. See Durlofsky
[22] for more discussion of this point. A wide variety of techniques for computing
these grid block permeabilities is available, as we shall discuss.
3.2 Power averaging procedures
The simplest procedures for computing grid block permeabilities are power averaging
procedures, introduced by Deutsch [19] (see also discussion in Wen and G omez-
Hernandez [77]). These approaches do not require any numerical solutions so they
are very ecient computationally. The basic approach entails computing upscaled
permeability components, here designated k

i
, via:
k

i
=
_
1
V
b
_
V
b
[k
i
(y)]

i
dV
_
1/
i
, (14)
where V
b
is the coarse block bulk volume and both the ne scale permeability k(y)
and the upscaled permeability k

are considered to be diagonal tensors, with i

designating a diagonal component. The power averaging exponent
i
can vary with
direction i. This type of averaging procedure can be readily applied to coarse grid
cells of any shape, so it is suitable for use with irregular grids.
The power averaging exponent
i
is constrained to lie between 1 and 1. The
extremes correspond to layered systems: for ow parallel to the layers = 1 (arith-
metic average), while for ow perpendicular to the layers = 1 (harmonic av-
erage). The geometric mean corresponds to 0 (in this case k

i
is replaced by
ln(k
i
)). Power averaging can also be applied using a combination of two dierent
values of . For example, for a structured, approximately layered system in the
x z coordinate system, k

x
might be computed by harmonically averaging along
each layer in x and then arithmetically averaging these layer averages. The use of
this type of procedure provides power averaging approaches with a higher degree of
8
applicability. Power averaging exponents may be determined in practice by tuning
against numerical upscaling results. The assumption then is that the same
i
can
be used for models with similar permeability distributions.
The fact that
i
can vary with direction leads to the general observation that
the upscaled permeability can be anisotropic even when the underlying permeability
eld k(y) is everywhere isotropic. For the power averaging approach described here,
the upscaled permeability is still a diagonal tensor. However, for the more general
numerical methods described below, the upscaled permeability is typically a full
tensor quantity.
Other analytical procedures for permeability upscaling include the renormaliza-
tion approach of King [48] and the full tensor averaging technique of Kasap and Lake
[45]. Both of these approaches are very ecient and have been shown to perform
well for some classes of problems. Like power averaging techniques, however, they
lack the generality of the numerical procedures described below.
3.3 Upscaling of porosity
Porosity on the coarse scale, designated

, is computed such that pore volume is

exactly conserved between the ne and coarse scales. Specically,

is computed
via:

=
1
V
b
_
V
b
(y) dV, (15)
where V
b
designates bulk volume.
4 Numerical procedures to compute k

The more robust and accurate procedures for computing k

require the solution of

the ne scale pressure equation over the target coarse block. As discussed below,
in some cases it is benecial to use an extended local approach in order to include
the eects of neighboring regions in the calculation of k

. Global and quasi global

procedures can also be applied for this calculation. We now consider a variety of
approaches, beginning with the simplest purely local techniques.
Dierent upscaling techniques force agreement in dierent quantities. For ex-
ample, in recent work Zijl and Trykozko [81] describe approaches for computing
upscaled permeabilities using pressure-ux averaging, pressure-dissipation averag-
ing, or ux-dissipation averaging. The relative advantages and disadvantages of the
various approaches are also discussed. When periodic boundary conditions are ap-
plied (discussed below), the upscaled permeability is the same in all cases. In the
discussion here we do not consider all of these approaches for upscaling but focus
9
instead on the use of pressure and velocity (or ux) averaging. It is possible that
alternate procedures may provide improved results in some cases.
In all of the procedures described below, nite volume (often referred to as nite
dierence) methods are applied for the numerical solutions unless otherwise speci-
ed. If the ne scale permeability is a diagonal tensor and the grid is orthogonal,
standard procedures (using harmonic averages of the appropriate permeability com-
ponent) can be used. In this case the usual ve point nite dierence stencil is
obtained (in two dimensions). In cases in which permeability is a full tensor and/or
the grid is nonorthogonal, we apply the multi-point ux techniques as described in
[2, 30, 52].
4.1 Fixed pressure boundary conditions
In computing equivalent grid block permeabilities, we solve Eq. (12) over the ne
scale region corresponding to the target coarse block. A signicant issue in any
local or extended local upscaling technique is the choice of boundary conditions to
be imposed. Because the actual conditions imposed on the region during the course
of a ow simulation are not known a priori and will in general vary, there is always
some ambiguity in specifying the boundary conditions in the upscaling procedure.
There is additionally some freedom in how the upscaled k

is computed from the

local ne grid solution. We now discuss these issues in turn.
Upscaling methodologies will rst be discussed within the context of a structured,
rectangular grid. The extension of these approaches to more general control volumes
will then be described. Consider the rectangular domain illustrated in Fig. 1. This
domain corresponds to a single coarse grid block. The ne grid permeability eld
(on the y-scale) is not shown. Our intent is to solve Eq. (12) over this domain
and then use this solution to compute the equivalent grid block permeability tensor.
We rst discuss the boundary conditions for this problem and then discuss how to
compute k

from the ne grid solution.

The simplest and in many ways the most intuitive boundary conditions for this
problem might be a constant pressure - no ow boundary specication. For the
system shown in Fig. 1, these boundary conditions require us to solve Eq. (12)
twice. In the rst solution we set
p(0, y
2
) = 1, (16a)
p(L
1
, y
2
) = 0, (16b)
u(y
1
, 0) n
1
= u(y
1
, L
2
) n
2
= 0, (16c)
while in the second solution the pressure dierence is specied to be in the y
2
direction. From these two solutions we can compute total ow rates through the
10
Figure 1: Schematic of rectangular local solution domain.
faces of the region; i.e., from the rst solution:
q
1
=
_
L
2
0
u(L
1
, y
2
) n
4
dy
2
. (17)
Similarly, q
2
can be computed from the second solution. Then, using an averaged
form of Darcys law, k

1
can be computed via:
k

1
=
q
1
L
1
L
2
p
, (18)
where p = 1 from Eqs. (16). The quantity k

2
is computed in a similar manner.
In many cases, the k

1
and k

2
computed in this way provide reasonably accurate
coarse block permeabilities. However, the boundary conditions and method for
computing k

from the local ne grid solution in this case preclude the calculation
of the cross terms of k

. These terms can be signicant in cases where the grid is

not locally K-orthogonal (by K-orthogonality we mean that the grid geometry and
permeability can be represented in terms of two-point uxes in the nite volume
discretization). Therefore, procedures for computing a full tensor k

are required
for some cases.
One approach for generating full tensor coarse block permeabilities is to post-
process the ne grid solution in a manner dierent to that used in Eqs. (17) and
11
(18). Because the no ow boundary conditions do not permit the calculation of
a full tensor k

via integration over boundaries, we instead compute volume aver-

aged velocities and pressure gradients over the entire ow domain (e.g., [68, 80]).
Specically, we compute
< u >
j
=
1
V
_
V
u
j
dV, (19a)
< p >
j
=
1
V
_
V
(p)
j
dV, (19b)
where j = 1, 2 indicates the ow solution (i.e., j = 1 denotes the solution with the
pressure dierence in the y
1
direction; j = 2 the solution with the pressure dierence
in the y
2
direction). Because both < u > and < p > have two components (for a
two dimensional problem), and because we solve two ow problems, four components
of k

can be calculated from these two ow solutions. Specically, we can write:

< u >
1
1
= (k

11
< p >
1
1
+k

12
< p >
1
2
), (20a)
< u >
1
2
= (k

21
< p >
1
1
+k

22
< p >
1
2
), (20b)
< u >
2
1
= (k

11
< p >
2
1
+k

12
< p >
2
2
), (20c)
< u >
2
2
= (k

21
< p >
2
1
+k

22
< p >
2
2
), (20d)
where the subscript on < u > and < p > designates the vector component and
the superscript the ow problem. This set of equations can be rearranged to give a
matrix equation:
_
_
_
_
< p >
1
1
< p >
1
2
0 0
0 0 < p >
1
1
< p >
1
2
< p >
2
1
< p >
2
2
0 0
0 0 < p >
2
1
< p >
2
2
_
_
_
_
_
_
_
_
k

11
k

12
k

21
k

22
_
_
_
_
=
_
_
_
_
< u >
1
1
< u >
1
2
< u >
2
1
< u >
2
2
_
_
_
_
, (21)
which can now be solved to determine the components of k

.
With the boundary conditions described by Eqs. (16), the k

computed via
Eq. (21) will not in general be symmetric. Various procedures can be applied to
enforce symmetry; the simplest approach is simply to set each of the cross terms
equal to (k

12
+ k

21
)/2. A better approach to ensure symmetry is to solve a least
square problem rather than Eq. (21). In this case we enforce symmetry by adding
an equation of the form k

12
k

21
= 0. The resulting set of equations is now given
by:
_
_
_
_
_
_
< p >
1
1
< p >
1
2
0 0
0 0 < p >
1
1
< p >
1
2
< p >
2
1
< p >
2
2
0 0
0 0 < p >
2
1
< p >
2
2
0 1 1 0
_
_
_
_
_
_
_
_
_
_
k

11
k

12
k

21
k

22
_
_
_
_
=
_
_
_
_
_
_
< u >
1
1
< u >
1
2
< u >
2
1
< u >
2
2
0
_
_
_
_
_
_
. (22)
12
The last equation can be rescaled if necessary so these matrix elements are of the
same magnitude as the other terms.
In addition to symmetry, we also require that k

be positive denite (i.e., have

positive eigenvalues). In two dimensions this requires that k

11
> 0, k

22
> 0 and
k

11
k

22
> (k

12
)
2
. This requirement is generally satised by the methods described
here. In the few cases when this is not satised, k

can be recomputed using outlet

averaging (in which case k

12
is undetermined) or using periodic boundary conditions
(described below).
The pressure - no ow boundary conditions just discussed are not the most gen-
eral boundary conditions that can be used. Another alternative is to use boundary
conditions that specify a linear pressure variation along the sides parallel to the
direction of the pressure gradient (e.g., King et al. [46, 47]). Then, rather than the
boundary conditions of Eqs. (16), the boundary conditions for the rst ow problem
are:
p(0, y
2
) = 1, (23a)
p(L
1
, y
2
) = 0, (23b)
p(y
1
, 0) = p(y
1
, L
2
) = 1 y
1
/L
1
. (23c)
Unlike the pressure - no ow specication, these boundary conditions do al-
low for ow out of the domain in a direction transverse to the pressure dierence.
Thus, a full tensor k

can be computed from the integrated ow rates through the

boundaries. For example, from the solution of Eq. (12) subject to Eqs. (23), we can
compute
q
1
1
=
_
L
2
0
u
1
(L
1
, y
2
) n
4
dy
2
, (24a)
q
1
2
=
_
L
1
0
u
1
(y
1
, L
2
) n
2
dy
1
. (24b)
Then,
k

11
=
q
1
1
L
1
L
2
p
, (25a)
k

21
=
q
1
2
p
, (25b)
where again p = 1 in this case. The other two components of k

can be computed
from the second solution (pressure dierence in y
2
direction). The tensor computed
in this manner will again not be symmetric in general.
The solution using linear pressure boundary conditions can also be used to com-
pute k

within the volume averaging context described via Eqs. (19) to (22). The k

13
thus computed will in general dier from the k

computed via Eqs. (24) and (25).

It is not clear from previous work which of these approaches is the more accurate
it is likely that the method of choice will be case dependent.
4.2 Periodic boundary conditions
We next describe the use of periodic boundary conditions for the calculation of k

.
This boundary specication eliminates some of the ambiguity of the other methods
in that it provides the same result for either method of post-processing of the ne
grid solution. Periodic boundary conditions have been applied and analyzed by a
number of investigators including Durlofsky [22], Boe [9], Pickup et al. [63] and Wen
et al. [76, 75]. These boundary conditions again require that two local ne scale
problems be solved. The specic form of the boundary conditions derives from the
assumption that the system is replicated periodically in space and that the global
pressure (on the scale of x) can be approximated as p = p
0
+ G (x x
0
), where
G = G
1
i
1
+ G
2
i
2
is a constant vector. These are both reasonable approximations
in many cases, though inaccuracy can result when the assumption of periodicity
disrupts larger scale permeability connectivity.
With reference to Fig. 1, periodic boundary conditions can be specied via:
p(y
1
, 0) = p(y
1
, L
2
) G
2
L
2
on D
1
and D
2
, (26a)
p(0, y
2
) = p(L
1
, y
2
) G
1
L
1
on D
3
and D
4
, (26b)
u(y
1
, 0) n
1
= u(y
1
, L
2
) n
2
on D
1
and D
2
, (26c)
u(0, y
2
) n
3
= u(L
1
, y
2
) n
4
on D
3
and D
4
. (26d)
By selecting two linearly independent pairs of (G
1
, G
2
), and by solving Eq. (12)
subject to Eqs. (26), we can compute k

using either Eqs. (19) to (21) or Eqs. (24)

and (25), with identical results. In practice, in the rst problem we set G
1
= 1 and
G
2
= 0 and in the second problem G
1
= 0 and G
2
= 1.
Periodic boundary conditions have several useful features. They guarantee that
the resulting k

will be symmetric and positive-denite. Thus, no post-processing

of the result is necessary to ensure that these two criteria are met. In a study
comparing the use of several dierent boundary conditions for the calculation of
k

, Pickup et al. [63] demonstrated that periodic boundary conditions performed

reliably in the example problems considered. However, they did demonstrate that
in many cases the dierences between the various methods were slight.
4.3 Extended local procedures: use of border regions
It has been observed by a number of authors that improved accuracy in k

can be
achieved if a larger local problem is solved (see, e.g., G omez-Hernandez and Journel
14
[36]; Holden and Lia [39]; Hou and Wu [42]; Wu et al. [80]; Wen et al. [76, 74, 75]).
By including neighboring regions in the calculation of k

for a particular coarse

block, the eects of large scale permeability connectivity (or lack of connectivity)
can be better captured, particularly when the permeability eld contains features
that are not oriented with the grid. A border region, containing the ne scale
permeability eld corresponding to a single ring of coarse blocks around the target
coarse block, is shown in Fig. 2. In the gure, the ner lines represent the ne grid,
the heavier lines the coarse grid, and the shaded block is the target cell for which
k

is to be computed. Any of the boundary conditions discussed above can now be

applied on the expanded domain shown in the gure.
Figure 2: Fine and coarse grids with border region (target cell in center).
Because we now wish to compute k

for only a portion of the ne scale domain over

which Eq. (12) is solved (the target coarse block), we apply the volume averaging
procedure described above rather than the integration over boundaries. Specically,
using Eqs. (19), we compute < u > and < p > over the shaded region in Fig. 2
and then apply Eq. (21) or (22) to form k

. We note that, even if periodic boundary

conditions are applied, the k

computed using Eq. (21) will not in general be sym-

metric. This is because the symmetry provided by periodic boundary conditions for
purely local upscaling is lost when border regions are applied. Symmetry can be
approximately recovered through use of the least square technique (Eq. (22)).
It is reasonable to expect that the eect of the boundary specication on the
computed k

will be less when border regions are used and that results using various
boundary conditions will tend to converge. In limited tests, some of which we will
describe below, this did in fact appear to be the case [75]. We note additionally
that, again in limited tests, the use of one bordering ring surrounding the target
15
cell appears to suce. Relatively little improvement was observed when a two-ring
region was applied [76, 75].
4.4 Irregular coarse grid control volumes
Up to this point, we have considered upscaling procedures applicable for structured
grids comprised of rectangular blocks. The procedures described can readily be
extended to three dimensions, so they are equally applicable to brick-shaped control
volumes. We now discuss the generalization of these approaches to irregular control
volumes. Two basic procedures will be described, though many other variants are
of course possible.
We assume for the present discussion that the underlying geocellular grid is
structured and is comprised of rectangular cells. In general, the irregular coarse
scale cell will not be aligned with respect to the ne grid. By this we mean that
the vertices of the coarse cell will not coincide with the corner points of the ne
geocellular grid. This case is illustrated in Fig. 3. Nine coarse blocks are shown in
the gure, as is the underlying ne grid. The central coarse cell is the target coarse
cell; i.e., the one for which we wish to compute k

. Techniques for upscaling with

general quadrilaterals when the ne and coarse grids are aligned were developed by
Edwards [27] and Eek-Jensen et al. [31].
Figure 3: Non-aligned ne and coarse grids with border region.
The border region approach described above lends itself quite naturally to this
calculation. We dene the local solution domain as the rectangular region containing
the coarse cell corner points plus a number of additional ne grid cells corresponding
16
to the dimensions of the target coarse cell. For example, for the case where the coarse
cell corner points fall within a 6 4 rectangle of ne cells (as in Fig. 3), the local
problem will be expanded by 6 cells in the y
1
and +y
1
directions and by 4 cells in
the y
2
and +y
2
directions. In this case, the local solution will contain 18 12 ne
cells (the entire domain shown in Fig. 3).
In the case of a triangular or polygonal coarse scale control volume, the local
solution domain could be dened to include all ne scale blocks whose centroids
fall within a prescribed distance of the centroid of the target cell. This distance
could be specied in terms of a typical linear dimension of the target cell. Thus,
regardless of the shape of the coarse cell, a rectangular domain containing ne scale
cells corresponding to the neighbors of the target cell can be established.
We then solve Eq. (12) subject to boundary conditions to determine the ne
scale solution over the domain of Fig. 3. Following the solution of the local ne
scale problem, we compute < u > and < p > over the coarse cell region and then
apply Eq. (21) or (22) to form k

. Some care is needed in computing these averages

since the ne and coarse grids are not aligned.
The simplest approach is to include in the averaging only those ne cells whose
centers fall within the boundaries of the coarse cell. This approach is illustrated
in Fig. 3, where the ne blocks indicated by shading (the blocks falling within the
coarse cell) are the cells contributing to the calculation of < u > and < p >.
Another approach, which might oer better accuracy in some cases, would be to
apply an area weighting using all ne scale blocks that have some portion of their
area within the coarse cell:
< u >
j
=
1
A
t

l
u
j
l
A
l
, (27)
where u
j
l
is the velocity for ne scale block l as computed from problem j, A
l
is the
area of this ne scale block that falls within the target coarse cell, and A
t
is the
total coarse cell area. Applying an analogous expression for < p >, k

can then
be computed from Eq. (21) or (22).
4.5 Conforming scale up method
In a number of test cases, the procedure described above was shown to provide
coarse scale ow results of reasonable accuracy (Wen et al. [74]). However, it is
possible that an improved treatment of the coarse cell geometry may be required
for some types of problems. This can be accomplished through use of a triangle
based nite element method for the solution of the local ne grid problem, as we
now describe.
The local ne grid meshing of the coarse quadrilateral cell is formed in this case
17
via a triangulation as illustrated in Fig. 4. The underlying ne grid is also shown in
the gure. In this case no border region is applied - the solution domain is simply
the coarse cell control volume - though the method can be readily generalized to
include border regions. The triangulation is accomplished here and in the examples
below using the public domain software qhull (Barber et al. [6]). The triangulation
is performed such that all internal permeability discontinuities (as dened by the
underlying geocellular model) and the coarse cell boundaries are resolved, subject
to constraints on the sizes and shapes of the triangular elements. Fine grid per-
meabilities can then be inserted directly from the underlying description into the
corresponding grid of triangular elements. The overall approach is referred to as the
conforming scale up method because it ensures that the region over which k

is
computed conforms exactly to the coarse quadrilateral cell.
Figure 4: Triangulated coarse scale quadrilateral cell (from [38]).
Once the triangulation is performed, Eq. (12) can be solved over the local ne
scale region. This can be accomplished using a mixed or non-conforming nite
element method. These methods are appropriate for this calculation because they
treat permeability as a cell (rather than a nodal) quantity and because they are
well suited to problems containing highly discontinuous permeability elds. He et
al. [38] presented an approach using the lowest order non-conforming nite element
method, which has been shown by Marini [54] to be identical to the lowest order
mixed nite element method for the solution of Eq. (12).
Any type of boundary condition can be applied to solve Eq. (12) over the tri-
18
angulated domain. If periodic boundary conditions are applied, there are some
geometric issues that must be addressed, such as the treatment of unequal distri-
butions of element edges on opposite sides of the coarse cell. This can be handled
consistently by viewing the coarse cell boundaries in transform space (see [38] for
a detailed discussion of this issue). In addition, the pressure dierences between
elements on opposite sides of the coarse cell must be modied from that specied
in Eqs. (26). This is because corresponding points are not, in this case, separated
by a distance that is the same at all locations. This, coupled with the fact that
the periodicity condition derives from the assumption that the coarse scale pressure
varies as p = p
0
+G (xx
0
), provides the following conditions for G
1
= 0, G
2
= 0:
p
1
= p
2
G
1
y
12
1
on D
1
and D
2
, (28a)
p
3
= p
4
G
1
y
34
1
on D
3
and D
4
, (28b)
where p
1
and p
2
designate pressure at corresponding points on sides 1 and 2 (see
Fig. 1) and y
12
1
designates the dierence in y
1
between those two points (and
similarly for p
3
, p
4
and y
34
1
). Analogous expressions can be developed for G
1
= 0,
G
2
= 0.
This general procedure can be extended to handle any type of two dimensional
polygonal coarse cell. For such cases the two dimensional control volume must be
triangulated, which is in general no more dicult than the triangulation of the
quadrilateral coarse cell. In the case of a general polygon, the specication of peri-
odicity may not be appropriate, as a general polygon does not contain four distinct
sides. The simplest approach in this case is probably to impose pressure boundary
conditions linear in y
1
for one solution (i.e., p = y
1
) and boundary conditions linear
in y
2
for the second solution. Then, k

could be computed directly through the use

of Eqs. (19) and (21) or (22).
5 Transmissibility upscaling: global and near-well
methods
In the descriptions above, we considered boundary specications and averaging pro-
cedures for the calculation of upscaled permeability tensors. These approaches can
also be applied to the direct calculation of upscaled transmissibility. This provides
more accurate coarse scale models in many cases because it eliminates the addi-
tional approximations that result when transmissibility is calculated from the grid
block k

. Transmissibility upscaling, which we now consider, is commonly used for

near-well upscaling as well as in global upscaling methods. In this section we rst
consider transmissibility upscaling for regions away from wells and then discuss the
application of this approach to the near-well region.
19
5.1 Calculation of upscaled transmissibility
We consider a two dimensional system in the x y plane (of thickness z) with
permeability everywhere diagonal (k
xy
= k
yx
= 0). Directional permeabilities will
be referred to as k
x
and k
y
. Note that we now designate the coordinate system as
xy rather than y
1
y
2
. We do this because some of the transmissibility upscaling
procedures we will describe are global (or quasi global) methods, so it is no longer
appropriate to represent the system in terms of two distinct scales of variation. In
the numerical discretization of the pressure equation (on either the ne or the coarse
scale), the ow rate q
i+1/2,j
from grid block (i, j) to grid block (i +1, j) is given by:
q
i+1/2,j
= (T
x
)
i+1/2,j
(p
i,j
p
i+1,j
) , (29)
where the transmissibility (T
x
)
i+1/2,j
is dened as
(T
x
)
i+1/2,j
=
2(k
x
)
i+1/2,j
yz
x
i+1,j
+ x
i,j
. (30)
Here x
i,j
denotes the size of grid block (i, j) and (k
x
)
i+1/2,j
, the interface perme-
ability, is given by the weighted harmonic average of (k
x
)
i,j
and (k
x
)
i+1,j
:
(k
x
)
i+1/2,j
=
(x
i,j
+ x
i+1,j
)(k
x
)
i,j
(k
x
)
i+1,j
x
i,j
(k
x
)
i+1,j
+ x
i+1,j
(k
x
)
i,j
. (31)
For the case of full tensor permeabilities, the formation of the discretized equations
is more complex (see, e.g., [2, 30, 52] for a full discussion). We will not consider
transmissibility upscaling for systems of this type here.
Coarse scale transmissibilities can also be computed directly from the local or
extended local problems described in previous sections. To accomplish this, we
focus on a cell edge. Consider the case in which the local problem contains the
ne scale region corresponding to coarse blocks (i, j) and (i + 1, j), which are on
either side of the target edge (i + 1/2, j). We solve for ow over this region using
any of the boundary conditions described above and then compute the upscaled
transmissibility via:
(T

x
)
i+1/2,j
=
q
c
i+1/2,j
< p >
i,j
< p >
i+1,j
, (32)
where T

x
designates the upscaled transmissibility in the x direction, q
c
i+1/2,j
is the
total ow across the interface between coarse blocks (i, j) and (i +1, j); i.e., the sum
of the ne grid ows over this region, and < p > is the volume average of the ne
scale pressure over the designated coarse block region. A similar expression provides
T

y
:
(T

y
)
i,j+1/2
=
q
c
i,j+1/2
< p >
i,j
< p >
i,j+1
, (33)
where all quantities are dened analogously.
20
5.2 Global upscaling procedures
The methods described in 4 local and extended local procedures all require
the specication of (assumed) boundary conditions on the local problem. In global
upscaling methods, by contrast, the intent is to solve a global ow problem and to use
this solution to extract coarse scale quantities. Some examples of global upscaling
methods are those presented by White and Horne [79], Pickup et al. [62], Nielsen and
Tveito [59], Holden and Nielsen [40] and Aarnes [1]. Most of these methods apply
transmissibility upscaling for the calculation of coarse grid quantities. From the ne
grid solution, transmissibilities are computed by averaging over coarse block regions
and then applying Eqs. (32) and (33). This gives a rst estimate for T

x
and T

y
. In
highly heterogeneous models, a signicant fraction of these transmissibilities may
be negative. Iteration is therefore performed until all of the transmissibilities are
positive and a sucient level of agreement between the ne and coarse solutions is
achieved [40]. Various quantities (pressure, velocity, ux) can be considered during
these iterations and dierent approaches focus on dierent variables.
Global upscaling methods can provide very accurate results for a particular set
of wells and boundary conditions. In many cases the model developed in this way
can be used for other ow scenarios. However, it is also possible that the model may
lack robustness with respect to other boundary conditions or well arrangements.
This area is a subject of ongoing research (see, e.g., Holden et al. [41] for some
results addressing this issue).
A technique that shares some similarities with global upscaling methods is multi-
grid upscaling [57]. In multigrid upscaling, coarse scale parameters are determined
at a specied level from the multigrid solution algorithm. It will be of interest
to compare multigrid and global upscaling techniques in detail to obtain a better
understanding of the commonalities and dierences between the two approaches.
5.3 Coupled local-global (quasi global) upscaling
Quasi global upscaling methods attempt to estimate the eects of the global ow
without actually solving a global ne scale problem. Here we will briey outline
the quasi global procedure recently presented by Chen et al. [14], referred to as
a coupled local-global approach. The idea of this method is to use global coarse
scale simulations to estimate the boundary conditions to use in the extended local
calculation of T

(we will describe the method within the context of T

upscaling,
though it is also applicable for k

upscaling). The procedure is iterated until the

upscaled quantity is consistent with the global ow (i.e., self-consistency is enforced).
The technique is illustrated schematically in Fig. 5.
The rst step in this procedure is the extended local calculation of T

through
application of any of the boundary conditions discussed above (e.g., pressure - no
21
Figure 5: Schematic showing coupling of global coarse scale pressure and local ne
scale ow calculations (from [14]).
ow or periodic). This provides the initial estimate of the upscaled quantities. Using
the coarse scale model generated in this way, two generic global simulations are then
performed one with large scale ow driven in the x direction and the other with
large scale ow in y. These simulations provide global coarse scale pressure elds,
which are then used to set boundary conditions for the extended local problems for
the calculation of T

. Iteration proceeds in this way until the upscaled quantities

no longer change with iteration, at which point the global ow eld and T

are
consistent.
In order to actually solve the extended local problems at each iteration, the
coarse global pressure elds must be interpolated to provide local boundary condi-
tions on the ne grid. This is accomplished using permeability weighted harmonic
interpolation in the main ow direction and simple linear interpolation in the sec-
ondary direction. Such boundary conditions are most appropriate for systems with
some degree of local layering on the ne scale, as is often the case for many types
of geological models.
As with any extended local, quasi global or global approach, this procedure can
on occasion provide negative T

. However, because generic global ows driven in the

main coordinate directions are used, the number of negative T

is quite a bit fewer

than can result from global upscaling techniques. In the cases when a negative T

was computed, we replaced this T

with one computed from the diagonal components

of a purely local k

(which was itself computed using periodic boundary conditions

to assure positive deniteness).
22
5.4 Upscaling in the near-well region
The basic assumption in all of the local and extended local procedures considered
previously is that the ow can be described locally as essentially linear; i.e., the
large scale pressure gradient p is approximately constant. This is not assumed in
global upscaling techniques (or in the local-global approach just described), which
determine local velocity and pressure, used to compute T

, from a global solution.

The assumption of a locally linear pressure eld is also not applicable in the near-well
region, as the steady state pressure in the vicinity of a well away from boundaries
varies as log r, where r is radial distance. We now describe an approach for near-well
upscaling which can be classied as an extended local technique, though the local
problem is driven by a well rather than by a large scale linear pressure eld.
Wells are represented in reservoir simulators through use of a well index, here
designated by W
i,j
, which relates the wellbore pressure in block i, j (p
w
i,j
) to the grid
block pressure p
i,j
and well ow rate q
w
i,j
via:
q
w
i,j
= W
i,j
(p
i,j
p
w
i,j
), (34)
in dimensionless terms. The default well index W
d
i,j
in most simulators for a vertical
well in block i, j is given by the Peaceman expression [61]:
W
d
i,j
=
_
2
_
k
x
k
y
z
ln
rw
ro
_
i,j
, (35)
where
r
0
= 0.28
_
_
ky
kx
x
2
+
_
kx
ky
y
2
4
_
ky
kx
+
4
_
kx
ky
. (36)
Here permeability is again assumed to be diagonal. Expressions similar to Eqs. (35)
and (36) are used for horizontal or deviated wells.
If no near-well upscaling is applied, the well index can be computed using
Eqs. (35)-(36), but with the upscaled permeability components k

x
and k

y
replac-
ing k
x
and k
y
. We refer to this well index as W
i,j
(k

). This approach can provide

acceptable results in cases where the system is not very heterogeneous or where per-
meability is reasonably continuous (i.e., high correlation lengths) in the near-well
region.
With more highly heterogeneous permeability elds, the simple W
i,j
(k

) treat-
ment described above can lead to considerable error. In such cases, it is necessary to
compute upscaled well indices and, for additional accuracy, near-well transmissibil-
ities between the well block and adjacent blocks. The rst such near-well upscaling
23
procedure was presented by Ding [21]. Subsequent approaches were presented in
[26, 55, 58]. We now describe the basic ideas behind these near-well upscaling tech-
niques.
We dene a local problem with r = 1 (where r designates the number of bordering
rings of coarse block regions) with the well block in the center, as shown in Fig. 6.
Rather than impose boundary conditions that lead to an approximately linear ow,
we solve the dimensionless ne scale pressure equation with a well source term q
w
in the central block:
(k p) = q
w
. (37)
In the discretized form of this equation, q
w
i,j
is represented using Eq. (34). We specify
a wellbore pressure of p
w
i,j
= 1 and a pressure on the outer boundary of p = 0. The
solution of Eq. (37) can then be computed to determine pressure and velocity in all
of the ne grid blocks in the extended local problem.
Figure 6: Schematic showing ne (ner lines) and coarse (heavier lines) grids in the
near-well region (r = 1).
Following this ne grid solution, we estimate the coarse grid well index, des-
ignated W

i,j
, and coarse scale well block transmissibilities (i.e., transmissibilities
linking the well block to adjacent blocks), designated T

w
, as follows. We rst com-
pute the volume averaged pressures < p >
i,j
in each of the coarse block regions:
< p >
i,j
=
1
V
t
Nt

l=1
p
l
V
l
, (38)
where N
t
is the number of ne blocks corresponding to the single coarse block i, j,
p
l
and V
l
are the ne grid pressures and cell bulk volumes, and V
t
is the total bulk
volume of the coarse block. We use a single subscript for ne scale variables for
24
conciseness. The ow rates over the ne grid cell edges corresponding to coarse
block faces are computed in a similar manner:
< q >
i1/2,j
=
N
f

l=1
u
l
nA
l
, (39)
where < q >
i1/2,j
is the ow rate associated with coarse edge i 1/2, j or i +1/2, j,
N
f
is the total number of ne scale cell edges, u
l
n is the ne scale ux across the
edge and A
l
is the cell area.
We can now estimate the coarse scale parameters W

i,j
and T

w
via:
W

i,j
=
< q >
i,j
< p >
i,j
p
w
i,j
, (40)
and
(T

w
)
i1/2,j
=
< q >
i1/2,j
< p >
i,j
< p >
i1,j
. (41)
A similar expression provides (T

w
)
i,j1/2
.
The use of these parameters in many cases provides signicantly improved coarse
scale representations of well performance. In some cases, it has proved useful [55] to
introduce coarse scale iteration to force the uxes computed from the local coarse
scale problem to agree with the integrated ne scale uxes < q >
i1/2,j
. This is
accomplished by iterating on the coarse scale parameters W

i,j
and T

w
until an ob-
jective function based on the mismatch in ux between the ne and coarse problems
is minimized. This minimization was achieved in [55] using a Gauss-Newton proce-
dure.
6 Flow-based gridding techniques
Until this point we have considered the calculation of the upscaled parameters k

,
T

and W

. The accuracy of the upscaled model can often be improved consider-

ably through the use of specialized gridding techniques. In particular, the use of
ow-based gridding in conjunction with accurate upscaling procedures represents a
powerful approach for the generation of coarse scale models that retain the geolog-
ical realism of the ne grid description. The basic goal of these methodologies is
the introduction of higher levels of grid renement in high ow regions of the model
and coarser descriptions in lower ow regions. This allows the coarsened model to
capture many important eects of the ne scale permeability eld without retaining
a uniformly ne grid.
Flow-based gridding procedures have been developed by a variety of investiga-
tors, both within Cartesian and curvilinear grid frameworks. Within a structured
25
Cartesian or stratigraphic-grid framework, Durlofsky et al. [24, 25] presented a
nonuniform coarsening technique that selectively removes ne scale grid lines in a
manner that retains important high ow regions. This approach, applied in both two
and three dimensions, is appropriate within the context of Cartesian or stratigraphic
grids. However, it is limited because the coarse grid is constrained to be aligned
with (i.e., overlay) the ne grid. More general ow-based gridding procedures are
not constrained in this way because they allow the coarse grid to be non-aligned
with respect to the underlying ne (geocellular) grid.
Within the context of reservoir simulation, a number of procedures for ow-
based grid generation have been presented. Verma and Aziz [73], Edwards et al.
[29], Portella and Hewett [64], Castellini et al. [13, 12] and Wen et al. [74] pre-
sented techniques based on the use of streamlines computed from a single-phase
ow solution of the ne scale problem. An alternate gridding procedure, based on
the grouping of cells of similar permeability, was suggested by Garcia et al. [33].
These researchers introduced the concept of an elastic grid, in which the grid is
adjusted to minimize the variance of permeability within coarse grid cells. Other
investigations along these lines are discussed by Farmer [32]. In later work by Tran
[72] and Wen and G omez-Hernandez [78], the general approach of Garcia et al. [33]
was combined with ow information to generate ow-based grids.
Many of the previous techniques for ow-based grid generation were limited to
two dimensional systems. Of the work cited here, only Castellini et al. [13, 12]
developed procedures for curvilinear grid generation in three dimensional systems.
The generation of ow-based grids has also been addressed within the context
of groundwater hydrology. In this setting, ow-based grids may be applied either
for the solution of the contaminant transport equation or for the ecient solution of
the pressure equation. Relevant papers along these lines include the recent work of
Cirpka et al. [17] and Cao and Kitanidis [11]. Cirpka et al. applied ow-based grids
for the solution of the contaminant transport equation, while Cao and Kitanidis
developed an unstructured ow-based gridding procedure for the solution of the
pressure equation. Their procedure entails the use of an a posteriori error estimate
to guide grid renement. Within an unstructured grid context, there have been a
number of other ow-based grid generation procedures. These include the PEBI
windowing technique of Mlacnik et al. [56] as well as the triangle-based procedure
developed by Edwards [28].
Grid generation is a very broad area, with applications in many areas of scientic
and engineering computing. Extensive discussions of the general area can be found
in the books by Thompson et al. [71] and Knupp and Steinberg [51]. Elliptic and
optimization-based grid generation techniques, such as those described by Knupp
and coworkers [49, 50], can make use of ow information and have the potential
to provide coarse grids with a relatively high degree of control on internal grid
geometry. To date, these methods have not been applied extensively within the
26
context of reservoir simulation, though they appear to be very well suited for this
application.
Methods based directly on the use of streamline information may provide grids
with an extremely high concentration of grid lines in high ow regions, particularly
in highly heterogeneous systems. Here we describe ow-based grid generation tech-
niques that use streamline information, though a grid-smoothing step is introduced
to provide more control over the local grid density, which acts to improve grid qual-
ity. This capability introduces a higher degree of exibility into the overall grid
generation procedure.
6.1 Overview of gridding procedure
We now describe an approach for the generation of ow-based grids. This descrip-
tion follows the presentation in Wen et al. [74]. The rst step in the ow-based
grid generation procedure is the solution of the single-phase incompressible pressure
equation over the ne grid region. The pressure equation, repeated here,
(k p) = 0 , (42)
can be solved over either the entire geocellular model or over only some portion of it.
In the latter case, the global problem must rst be decomposed into a number of sub-
domains or modules. Although the ne grid permeability eld can be described on
any type of grid, it is commonly dened on a Cartesian or stratigraphic grid. The ne
grid permeability may be isotropic, a diagonal tensor or a full tensor quantity. The
coarse grid models will however be characterized by irregular quadrilateral (in two
dimensions) or hexahedral (in three dimensions) cells and full tensor permeabilities
in the general case.
Although the generation of a ow-based grid requires the solution of a ne scale
problem, this calculation often represents a fairly small computational cost relative
to the solution of the two-phase (or multiphase) system on either the ne or coarse
scale. This is because Eq. (42) is only solved once, while the two-phase ow prob-
lem requires the solution of the pressure equation at every time step. If the ne
scale solution of Eq. (42) over the entire ow domain is prohibitive, then the grid
generation can be accomplished in a modular fashion. Following the generation of
the grid in each module, the modules can be recombined into either a single globally
structured grid or a multiblock (or modular) grid (see Jenny et al. [44] for examples
of grids of this type). Here we describe the basic grid generation procedure for a
single domain.
27
6.2 Flow-based gridding in two dimensions
The basic idea of the grid generation procedure is to use streamlines to dene the
high ow paths and to introduce renement in these areas. Although the grid
is determined using ow-based information from a single ne grid solution, it is
appropriate for use in a variety of related ow problems. We describe the grid
generation procedure with reference to a two dimensional ow domain in an x y
coordinate system. The ow-based grid is formed by rst solving Eq. (42) subject
to p = 1 on the left face (x = 0), p = 0 on the right face (x = L
x
) and no ow on
the y = 0 and y = L
y
boundaries, where L
x
and L
y
denote the system lengths in the
x and y directions. Following this solution, streamlines can be generated through
either a particle tracking technique (Thiele et al. [70]) or through the contouring of
the streamfunction . The particle tracking technique is more general, in that it
can be readily extended to three dimensions, while the streamfunction technique is
only appropriate in two dimensions. However, when the ow is driven by boundary
conditions (rather than by wells), as is the case here, the streamfunction can be
easily computed from the velocity eld in a post-processing step using:
=
_
l
u n dl , (43)
where is the increment in streamfunction between two cell corners and l is the cell
edge connecting the two nodes. Because the quantity u n is known accurately (and
conservatively) along cell edges from the ux-continuous nite volume solution, this
is a viable approach for determining streamfunction. Streamlines are then generated
as lines of constant . Some number of streamlines is then selected to provide one
set of coordinate lines.
The other coordinate lines can be obtained from isopotentials (contours of con-
stant pressure) or from streamlines generated from a complementary ow problem
(e.g., ow from y = 0 to y = L
y
with no ow on x = 0 and x = L
x
). Another
approach is to divide each of the streamline coordinate lines into a specied number
of segments of equal arc length and to simply connect these segments.
Grids generated directly from streamlines introduce high levels of resolution in
regions of high ow. This approach can result in grids with highly distorted cells
and an overly high concentration of grid lines in high ow regions (and very few
grid lines in lower ow regions). This can present diculties if the grid is applied
to problems that dier signicantly from that used to dene the streamlines. For
example, if the grid is to be used to study the eects of well placement, then the
grid must be applicable for a variety of dierent ow problems. To provide grids
with more uniformity, it is useful to apply some amount of grid smoothing.
The basic idea of grid smoothing is to control the level of grid line concentration
throughout the domain. The procedure described here is relatively simple, though
more complex variants can be readily dened. Following the solution of Eq. (42)
28
on the ne scale in an x y coordinate system and the determination of the initial
coordinate lines, the x and y locations of all of the grid line intersections can be
determined. These intersections, designated x
i,j
and y
i,j
, provide the locations of
the corner points of each coarse cell.
The grid smoothing now entails some number of Laplacian-type iterations of the
form:
x
k+1
i,j
= (1 )x
k
i,j
+
1
4
(x
k
i1,j
+ x
k
i+1,j
+ x
k
i,j1
+ x
k
i,j+1
) , (44a)
y
k+1
i,j
= (1 )y
k
i,j
+
1
4
(y
k
i1,j
+ y
k
i+1,j
+ y
k
i,j1
+ y
k
i,j+1
) , (44b)
where (x
k
i,j
, y
k
i,j
) with k = 1 designates the initial grid, (x
k+1
i,j
, y
k+1
i,j
) designates sub-
sequent (smoothed) grids and is a relaxation parameter (0 < 1). In the two
dimensional examples below, we used anywhere from 2-5 iterations of Eqs. (44) with
= 0.6.
We now present an example illustrating the smoothing procedure. The perme-
ability eld for a channel system is depicted in Fig. 7 (this permeability eld is
from [53]). The initial coordinate lines, generated as described above, are shown in
Fig. 8. The grid is clearly highly concentrated in some regions and very coarse in
other regions. A more uniform grid, generated using two smoothing iterations, is
displayed in Fig. 9.
Figure 7: Permeability eld for uvial reservoir.
This grid is better suited for ow calculations. The nal step in the grid gen-
eration procedure is to upscale the geocellular permeabilities to the cells of the
ow-based grid. This can be accomplished using either the nite volume procedure
29
Figure 8: Initial ow-based grid for permeability eld of Fig. 7 (from [74]).
Figure 9: Final ow-based grid after smoothing (from [74]).
30
described in 4.4 or the nite element based conforming scale up method described
in 4.5.
6.3 Flow-based gridding in three dimensions
Flow-based grids in three dimensions can be constructed using a similar approach to
that described above. A streamline-equipotential coordinate system is again estab-
lished as the initial grid, though the procedure is more involved in three dimensions.
Smoothing is again applied to improve grid quality if necessary. We now briey
describe the specic steps (see Castellini et al. [13, 12] for more detail).
First, a single-phase ow calculation is performed by solving the pressure equa-
tion subject to constant pressure boundary conditions at the inlet and outlet faces
of the domain. Streamlines are then traced from the ne grid velocity eld using
the procedure of Thiele et al. [70]. A structured (layered) set of streamlines is
then dened over the inlet boundary, where each layer contains the same number of
streamlines. The streamlines are selected within each layer according to streamline
density (for a specied level of coarse grid resolution). Layers of streamlines are
then recomputed according to layer density for a specied level of resolution.
The other coordinate lines are formed by intersecting the streamlines with isopo-
tential surfaces. This is accomplished by projecting the pressure eld onto the
selected streamlines at logical uniform intervals via tri-linear interpolation. The
interpolated pressures are functions of the corresponding eight cell centered pres-
sures. For each layer, isopotential line interpolation is performed across the selected
streamlines, resulting in a set of logically Cartesian streamline-isopotential-line sur-
faces. Isopotential surfaces are then dened by joining the surface layers to form
the three dimensional grid.
Following the generation of the initial grid as described above, smoothing can be
accomplished via the procedure outlined in Eqs. (44), with an additional equation
of the same form introduced for z. Considerable levels of smoothing will likely be
required for three dimensional problems with signicant cross ow, since the grid
generation procedure described above is most appropriate for systems displaying
some degree of layering. For systems with large amounts of cross ow, alternate
procedures such as those of [49, 50] might prove useful for the generation of the
grid. Following the construction of the ow-based grid, the ne scale permeabilities
must be upscaled onto the new coarse grid. Any of the procedures described earlier
can again be applied for this purpose.
31
7 Assessing the quality of the upscaled model
In this section we will assume that the upscaled model is dened by the ow-based
grid and the cell k

or T

, computed using the methods described earlier. The

upscaled model is intended to capture, to the extent possible, the eects of key
geological and geometric features. Because the ne scale multiphase ow model
may be overly time consuming to actually simulate even once, it is often dicult
to assess the quality of the coarse scale description. We now briey discuss some
strategies for assessing the accuracy of the coarse model and describe an approach for
model iteration. In assessing the upscaled model, the idea is to simulate a problem
that is representative of the ne scale problem without simulating the actual ne
scale model. One means of accomplishing this is to simulate unit mobility ratio
displacements, as described in [25].
In a unit mobility ratio displacement, the displacing phase has properties iden-
tical to those of the displaced phase. This model is much faster to run than an
actual two or three-phase ow problem because the pressure and velocity elds do
not change during the course of the simulation. Thus, the pressure equation need be
solved only once on the ne scale. The unit mobility ratio case can be recovered from
the two-phase equations by setting the oil viscosity equal to the water viscosity and
the relative permeabilities equal to the phase saturation; i.e., k
rw
= S, k
ro
= 1 S,
where S is water saturation. In this case, the dimensionless pressure equation is
identical to Eq. (42). The equation describing the movement of the individual phase
fronts within the reservoir, the saturation equation, is then given by:
S
t
+ (uS) = 0 , (45)
where is assumed to be constant. Because the velocity eld does not vary in
time, this equation can be solved very eciently and accurately by integrating along
streamlines (e.g., Prevost et al. [65]). The Darcy velocity must rst be computed
by applying Darcys law to the pressure eld (u = k p).
The solutions of Eqs. (42) and (45) can be used as diagnostics to assess the
accuracy of the upscaled model. Although these equations dier from the actual ow
equations, they do capture many important aspects of two-phase (or multiphase)
ow problems. Specically, the eects of heterogeneity are often similar in the unit
mobility and two-phase ow problems.
Various quantities can be compared between the ne and coarse scale solutions
of Eqs. (42) and (45). The two that are perhaps the most representative of the
general level of accuracy of the upscaled model are the global (overall) ow rate and
the fractional ow of displaced uid (e.g., oil cut) at the outlet (or production well)
as a function of pore volume injected. The global ow rate through the model can
be quantied in terms of the global equivalent permeability, K
g
. For global ow in
32
the x direction, driven by pressure boundary conditions at x = 0 and x = L
x
, this
quantity is given by:
K
g,x
=
Q
x
L
x
L
y
L
z
p
, (46)
where p is the dierence in pressure between x = 0 and x = L
x
and Q
x
is the
total ow rate through the system in the x direction. For global ow in the y or z
directions, similar expressions can be obtained. The grid structure can be iterated
until the agreement in these two (or similar) quantities is acceptable. A number of
dierent coarse grid structures can be compared to the ne grid solution relatively
eciently since the ne grid solution of Eq. (42) need only be computed once.
The overall gridding, upscaling and diagnostic procedure described above is il-
lustrated schematically in Fig. 10 for a two dimensional system. The rst step here
is the solution of a single-phase ne grid ow problem. This solution is used both
as the reference solution (against which coarse scale simulation results will be com-
pared) and in the grid generation procedure. Next, a grid is formed and k

or T

is
computed for the coarse scale cells. Following this, a global coarse scale ow problem
is solved and the result compared to the ne scale reference solution. Iteration on
the grid structure can be used to improve the agreement with the ne scale solution.
These diagnostics are computationally inexpensive within the context of ow-
based gridding procedures since ne scale ows are already computed in the grid
generation step. The ability to perform an ecient and accurate assessment of
the coarse scale model represents a signicant advantage for ow-based gridding
techniques over other grid generation procedures that are not based on ne scale
ow solutions. As a nal check on the accuracy of the coarse model, it is useful
to perform a multiphase ow simulation of a portion of the ne scale model and
compare this solution to that for the corresponding region of the coarse scale model.
This provides an assessment of the accuracy of the coarse scale model in terms of
multiphase ow eects.
8 Numerical upscaling results
We now present upscaling results for a number of systems. In several of these ex-
amples we use the North Sea channelized system presented in Christie and Blunt
[15]. In other cases we use geostatistical models generated using GSLIB algorithms
(Deutsch and Journel [20]). Most of the results discussed in this section were pre-
sented previously in recent papers (e.g., [55, 38, 74, 75, 14]). The results presented
here are not intended to be exhaustive but rather illustrative of the performance
of the various techniques for dierent geological models. Many of these results are
presented in terms of total ow rate through the system for a prescribed pressure
dierence (K
g,x
, as dened in Eq. (46)).
33
Figure 10: Schematic illustrating iterative grid generation procedure and use of ow
diagnostics.
34
8.1 Use of border regions for highly heterogeneous
systems
In the case of two-point geostatistical models (e.g., spherical variogram models),
when the permeability correlation directions are oriented with the simulation grid,
purely local upscaling techniques have been shown to perform well in terms of main-
taining the total ow through the model. For example, for two dimensional systems
of this type, if we upscale by about a factor of 5 in each coordinate direction (giving
a coarse grid with a factor of 25 fewer grid blocks than the ne grid), K
g,x
computed
from the coarse model will typically show errors of less than 10% relative to the
reference ne scale calculation (see results in [25, 76, 38]). In these cases, the use of
extended local procedures (i.e., border regions, as described in 4.3) does not appear
to lead to signicantly improved coarse scale results [76]. For models of this type,
improved transport results (e.g., oil cut), relative to those obtained using uniformly
coarsened grids, can be achieved using ow-based nonuniform grid coarsening within
a Cartesian grid framework [25, 24]. The observations that border regions do not of-
fer much improvement but that Cartesian-based nonuniform coarsening procedures
do provide enhanced accuracy for transport are both related to the fact that the
correlation structure of the permeability eld is aligned with the simulation grid.
Figure 11: Continuous and channelized layers from North Sea reservoir (model from
[15]).
Border regions often do provide enhanced accuracy when the correlation struc-
ture of the permeability eld is not aligned with the grid or when the permeability
35
eld is characterized by multi-point geostatistics. We illustrate the potential im-
provement in coarse model accuracy oered through the use of border regions with
results from [75]. These calculations use the model of [15]. This model is of dimen-
sions 2206085. The upper 35 layers display a somewhat continuous permeability
eld, while the lower 50 layers are highly channelized (representative layers are shown
in Fig. 11). We illustrate the impact of border regions by considering each layer to
be a two dimensional system, of dimensions 220 60, which is upscaled to 22 20.
Results are presented as scatter plots, with K
g,x
for each coarsened layer plotted
against K
g,x
for the same layer simulated using the ne scale permeability eld. A
perfect upscaling would give all of the points on the 45

line.
Figure 12: Cross plot of K
g,x
from coarse model (r = 0) against K
g,x
from ne
model. Open circles correspond to continuous layers, solid circles to channelized
layers (from [75]).
In the rst set of results, the models are upscaled using a purely local procedure
(r = 0) and periodic boundary conditions. Very similar results were obtained using
pressure - no ow boundary conditions, so the eect of the boundary conditions in
this case is small [75]. From Fig. 12, we see that the coarse scale results for the layers
with the continuous permeability elds, indicated by open circles on the gure, are
in close agreement with the ne scale results. Errors for the lower layers, designated
by solid circles, are by contrast quite substantial. For all of the data in Fig. 12, the
average relative error is 46%.
36
Figure 13: Cross plot of K
g,x
from coarse model (r = 1) against K
g,x
from ne model
(from [75]).
Fig. 13 displays results using the extended local upscaling (r = 1) using periodic
boundary conditions (though again very similar results were obtained using pressure
- no ow boundary conditions). We see considerable improvement in the coarse scale
results for the channelized layers. The average relative error for all of the data is
reduced to 27% in this case. This error is still signicant, though it does represent
a clear improvement over that of Fig. 12.
Part of the error in the coarse scale results in Figs. 12 and 13 is due to numerical
discretization eects rather than upscaling eects. We can gauge the relative mag-
nitudes of these two eects by rening the grid (back to 220 60) in the upscaled
models while retaining the 2220 coarse scale permeability eld (i.e., we project the
coarse scale permeability eld onto the ne grid). Results for 22060 models of this
type with the coarse permeability eld generated using r = 1 are shown in Fig. 14.
The error here is reduced considerably relative to that in Fig. 13 (average relative
error is here 16%). This illustrates that a signicant portion of what is considered
to be upscaling error may really be due to numerical discretization eects rather
than to inaccuracy introduced in the calculation of k

.
We note nally that transmissibility upscaling (using an extended local proce-
dure) was found to provide better overall accuracy than permeability upscaling for
37
Figure 14: Cross plot of K
g,x
computed on 22060 grid using k

from coarse model

(r = 1) against K
g,x
from ne model (from [75]).
the dicult channelized cases in [15]. This may be because transmissibility upscal-
ing avoids the additional approximations introduced in computing transmissibilities
from k

via Eqs. (30) and (31). The local-global procedure illustrated below applies
transmissibility upscaling for this reason.
8.2 Coupled local-global upscaling
We next illustrate the improvement in results that can be attained using the coupled
local-global upscaling procedure described in 5.3. We consider channelized layers
from the model in [15], as discussed above. In this case we apply a higher level of
upscaling to generate coarse models of dimension 22 6 (upscaling by a factor of
10 in each coordinate direction). The results presented here apply transmissibility
upscaling. Border regions of size r = 1.5 were applied in the calculation of the T

(recall that pressure boundary conditions are determined from the global solution).
The initial estimate for T

was determined using an extended local calculation with

r = 1.5.
Results for layer 59 (shown in Fig. 11) are presented in Fig. 15. Here we see
that the initial errors in the global ows (Q
x
and Q
y
) of 15% and 25% are reduced
38
to 5.7% and 0.5%, respectively. It is also apparent that most of the correction is
achieved after only a few iterations of the local-global procedure. A more extreme
example is presented in Fig. 16. This example involves ow in the x direction in
layer 48. Here, the initial error in Q
x
of 63% is reduced to 10% after iteration. This
case illustrates the substantial error that can result from the use of local or extended
local upscaling techniques when high degrees of upscaling are applied, as well as the
signicant improvement oered by the local-global upscaling procedure.
The next example (Fig. 17) illustrates the performance of the method for various
degrees of upscaling. Upscaled models of 22 6, 22 15 and 44 20 are considered.
Though there is some variation in the nal results for Q
y
, all of the results are within
10% of the ne solution. It is interesting to note that the nest of the coarse grid
models (44 20) is not the most accurate. This may be because the border region
is the largest for the 22 6 model. This is the case because the border region is
specied in terms of the size of the coarse blocks.
We also observed the method to be fairly insensitive to the initial T

estimate.
This is reassuring, since dierent procedures can give quite dierent initial estimates.
8.3 Upscaling to irregular coarse grid cells
We next consider ow results for a case involving irregular quadrilateral cells. These
results are from [38] and were generated using the conforming scale up method
(described in 4.5), which entails the nite element solution of the purely local
problem over the irregular cell.
We consider a geostatistical permeability eld (Fig. 18), generated using the
GSLIB algorithms [20]. The domain is a unit square and the ne grid is of dimensions
50 50. The permeability eld is characterized by a correlation length in the x
direction of 0.5 and a correlation length in the y direction of 0.1. Permeability
is log-normally distributed, with the variance of log k equal to 4. Permeability is
specied to be isotropic on the ne scale (k
y
= k
x
).
This system is upscaled to 1010 grids of the type shown in Fig. 19. Three such
10 10 grids were considered, each of which was generated by randomly placing
nodes along side boundaries. The upscaled permeability eld shown in Fig. 19
resembles the underlying ne scale permeability eld to some extent (Fig. 18), as
would be expected. Due to the random nature of the coarse grid structure, however,
it is clear that the grid does not accurately resolve the extremes in permeability
evident in the ne scale description.
Global ow results for the ne and coarse grids are computed for ow in both
the x and y directions. These ne grid solutions give Q
x
= 2.59 and Q
y
= 0.820.
We then upscale this ne grid model to a 10 10 uniform coarse grid and solve the
global ow problem on this grid. The results on the 10 10 uniform coarse grid can
39
Figure 15: Global ow results for Q
x
and Q
y
using local-global T

upscaling for
layer 59 of North Sea reservoir (from [14]).
40
Figure 16: Global ow results (Q
x
) using local-global T

upscaling for layer 48 from

North Sea reservoir (from [14]).
be viewed as the reference solution for this case. Results on the irregular 10 10
grids would be expected to be less accurate in general than these results since the
cells vary signicantly in size and some of them are quite irregular. The ow results
for the uniform coarse grid are Q
x
= 2.46 and Q
y
= 0.754 (errors of 5% and 8%).
Flow results on the irregular 10 10 quadrilateral grids are as follows. For ow
in the x direction, the average Q
x
for the three grids was 2.30. For ow in the y
direction, the average Q
y
for the three grids was 0.727. Though these results are
less accurate than the uniform coarse grid results, errors relative to the 10 10
uniform coarse grid results are only 6.5% for Q
x
and 3.6% for Q
y
. The general
level of accuracy of the results compared to the ne grid calculations is also quite
reasonable (errors of about 11% for both Q
x
and Q
y
).
The ne and coarse grid solutions are further compared in Fig. 20, where pressure
contours for the two cases are displayed (the coarse scale results correspond to the
grid shown in Fig. 19). Here, the pressure contours show a general correspondence,
though dierences are apparent, particularly in the high permeability regions of the
domain. In practice, the grid will be rened in important (e.g., high ow) regions and
coarse in less important regions. Thus, errors due to grid irregularity will be more
than compensated for by the enhanced resolution of key aspects of the permeability
eld. This point will be illustrated below when we consider ow-based grids.
41
Figure 17: Eect of upscaling ratio on local-global upscaling results for layer 59 from
North Sea reservoir (from [14]).
8.4 Near-well upscaling
We illustrate the potential impact of near-well heterogeneity, and the importance
of capturing these eects in coarse scale models, through an example from [55].
This example involves a horizontal well in a complex three dimensional sand-shale
reservoir containing an oil zone, an aquifer and a gas cap. This case was considered
previously by Aziz et al. [4] who used this system to illustrate the eects of ne
scale heterogeneity on horizontal well performance.
The ne scale reservoir model contains 100 50 32 cells and is described in
[4]. The coarse model, of dimensions 25 25 12, was generated through a uniform
coarsening of the ne scale model using a local upscaling method, with an additional
layer retained in both the aquifer and gas cap. The single horizontal producer was
specied to produce at a target rate of 5000 bbl/day, with a minimum bottom hole
pressure constraint of 1500 psi.
Results for oil rate for the ne model, the coarse model using a well index com-
puted directly from the upscaled permeability (i.e., W
i,j
(k

) using the terminology

of 5.4) and the coarse model with near-well upscaling (W

i,j
and T

w
) are shown in
Fig. 21. The coarse results without the near-well treatment are clearly in consid-
erable error relative to the reference ne scale results. This is mainly due to the
fact that the coarse model in this case continues to produce at the target rate for
a period of time that is 10 times too long. The coarse model with the near-well
upscaling, by contrast, provides a much more accurate pressure response and, as
42
Figure 18: Geostatistical 50 50 permeability eld with l
x
= 0.5, l
y
= 0.1, = 2
(from [38]).
Figure 19: Upscaled 10 10 grid with permeability eld (from [38]).
43
Figure 20: Pressure contours (p = 0.9, 0.8, 0.7, . . . , 0.1) for ne (solid curves) and
coarse (dotted curves) grid solutions (from [38]).
a result, switches to bottom hole pressure control at about the correct time. This
behavior is evident from the pressure proles shown in Fig. 22. The upscaled model
with the near-well treatment also provides accurate results for water cut. Results
for GOR, though much more accurate than coarse scale results without near-well
upscaling, are not as accurate. Results for both water cut and GOR are presented
in [55].
8.5 Flow-based gridding and upscaling
We illustrate ow-based gridding and upscaling with an example from Wen et al.
[74]. This case involves a permeability eld with oriented layers, generated using
Gaussian sequential simulation [20]. The permeability eld is of dimensions 100100
and displays layering (and principal axes of permeability) oriented at an angle
of 30

relative to the x axis. The correlation length along the direction of the
layering is 0.8L while the correlation length normal to the layering is 0.04L, where
L = L
x
= L
y
is the length of a side of the (square) domain. Permeability is log-
normally distributed, with the variance of log k equal to 4. In each ne scale block we
set k
2
= 0.1k
1
, where k
1
is the principal value of permeability in the direction along
the layering and k
2
is the principal value of permeability in the direction across the
layering. The permeability in the x y coordinate system is therefore a full tensor
44
Figure 21: Oil production rate (three-phase ow) for sand - shale system (from [55]).
quantity, which can be determined from k
1
and k
2
in each block and via:
k(x, y) =
_
k
1
cos
2
+ k
2
sin
2
(k
1
k
2
) sin cos
(k
1
k
2
) sin cos k
1
sin
2
+ k
2
cos
2

_
, (47)
where = 30

. The permeability eld is shown in Fig. 23 (log(k

1
) is the scalar
quantity actually displayed in the gure). The ow-based grid for this case is of
dimensions 20 20. This grid, shown in Fig. 24, was generated using the technique
described in 6.2. Upscaled permeabilities were computed using the nite volume
method described in 4.4.
The simulation in this case involves two-phase ow. Relative permeabilities to
water and oil are specied as k
rw
= S
2
and k
ro
= (1 S)
2
and the viscosity ratio
(
o
/
w
) is 10. The total mobility therefore varies in time, which means that the
pressure equation (10) and the saturation equation (9) must be solved at each time
step (an IMPES procedure is applied). Flow in this case is in the x direction and is
driven by boundary conditions of the same form as those used to generate the initial
streamlines. We x S = 1 at the inlet edge of the model for these simulations.
Results for total ow rate (Q) as a function of time (PVI) are shown in Fig. 25.
Four curves are shown in the gure. These correspond to the 100 100 ne grid
solution (solid curve), the 2020 uniform coarse grid solution generated using r = 1
45
Figure 22: Wellbore pressure for sand - shale system (from [55]).
(dotted curve), the ow-based grid solution with k

computed with r = 0 (dot-dash

curve), and the ow-based grid solution with k

computed with r = 1 (dashed

curve). The ow-based grid results, with k

computed using r = 1, are in close

agreement with the ne grid results (at initial time, the error is about 5%). This is
in contrast to the uniform grid results with r = 1 and to the ow-based grid results
with k

computed using r = 0 (initial time errors for both of these results are more
than 20%).
Results for oil cut (F
o
) versus PVI are shown in Fig. 26. Again we see the
best accuracy using the ow-based grid with k

computed using r = 1, though

breakthrough does occur slightly earlier than in the ne scale solution. This may be
due to numerical dispersion eects. The uniform grid predicts late breakthrough, as
does the ow-based grid with k

computed using r = 0 (though error is less in this

case). The results of Figs. 25 and 26 demonstrate that a grid formed by solving a
single-phase ow problem can still be used within the context of a two-phase ow
simulation. It is important to note that the coarse scale ow-based grid model
required a factor of 60 times less computation time than the ne model. These
savings will be realized each time the coarse model is run. The grid generation
and permeability upscaling calculations required only about 0.4% of the ne grid
simulation time, so the overhead cost is quite small for this problem.
46
Figure 23: Permeability eld with layering oriented at 30

(from [74]).
Figure 24: Flow-based grid after smoothing for permeability eld of Fig. 23 (from
[74]).
47
Figure 25: Total ow rate for two-phase ow in the x direction in oriented system
(from [74]).
Figure 26: Oil cut for two-phase ow in the x direction in oriented system (from
[74]).
48
9 Concluding remarks and future directions
In this paper, techniques for upscaling detailed geocellular models were described
and applied. The methods discussed include local, extended local, global and quasi
global procedures for computing upscaled permeability or transmissibility, tech-
niques for upscaling to irregular quadrilateral cells, upscaling in the vicinity of wells,
and the use of ow-based gridding procedures for capturing the eects of connected
permeability in coarse models.
The various upscaling methods were shown to provide coarse scale results of
varying levels of accuracy. For example, for the North Sea permeability eld consid-
ered in [15], it was shown that purely local permeability upscaling provided accurate
coarse scale results for the upper, more continuous layers. The coarse models gen-
erated in this way were not nearly as accurate, however, for the lower channelized
layers. For these layers, the use of border regions in the upscaling improved the
coarse scale results considerably. The local-global procedure was shown to further
enhance the results for highly heterogeneous reservoirs. Near-well upscaling was
additionally shown to provide improvement in the pressure response of horizontal
wells. The benets of smoothed ow-based grids were illustrated using a permeabil-
ity eld that was oriented relative to the underlying coordinate system. In this case
it was demonstrated that the combined use of the ow-based grid and border region
upscaling provided the most accurate overall results.
The upscaling of two-phase ow functions (e.g., calculation of pseudo relative
permeabilities) was not discussed in this review. As indicated in the Introduction,
such upscaling will be required when high degrees of coarsening are applied. In
addition, this type of upscaling may also be necessary when additional two-phase
ow eects (such as eective coarse scale capillary pressure functions) are required.
A number of techniques were described and illustrated individually in this paper.
It will be of interest to investigate hybrid procedures that combine one or more of the
methods described here to give a more powerful overall upscaling methodology. Po-
tentially useful combinations include the use of local-global upscaling in conjunction
with ow-based gridding and/or near-well upscaling, as well as the coupling of many
of the techniques described here with approaches for two-phase upscaling. The use
of more general grid generation techniques (e.g., [49, 50]) for the grid construction
step may also provide improved coarse scale models. It will also be essential that up-
scaling and gridding techniques be further (or completely) automated, particularly
if many geological models must be considered.
Finally, it is very important that we achieve a better understanding of the error
introduced by the various upscaling procedures. By quantifying this error through
the development of error models, we will be able to determine the appropriate up-
scaling method and level of coarsening to use for a particular problem. Progress
along these lines has been reported within the context of two-phase upscaling (e.g.,
49
[35]), though the techniques have thus far been applied only in idealized settings.
The development of practical techniques to assess upscaling error will allow for the
use of coarse scale simulations with the appropriate level of model complexity.
Acknowledgments
I am grateful to many colleagues, with whom a number of the techniques described
in this paper were developed. I especially acknowledge Dr. Xian-Huan Wen for our
extensive collaboration and for many useful discussions. Thanks also to Yuguang
Chen for her insightful comments on an earlier version of this paper. Much of the
research presented here was supported in part by the U.S. Department of Energy
under contract number DE-AC26-99BC15213 and by the industrial aliates of the
SUPRI-B (Reservoir Simulation) and SUPRI-HW (Nonconventional Wells) research
programs at Stanford University.
Nomenclature
A area
f fractional ow function
F
o
oil cut
G arbitrary pressure gradient
k permeability tensor
k

eective or equivalent permeability tensor

k
rj
relative permeability for phase j
K
g
global equivalent permeability
l dimensionless correlation length
L system length
m mass ow rate
n normal vector
p pressure
q volumetric ow rate
Q global ow rate
r size of border region or radius
S saturation
t time
T transmissibility
u Darcy velocity
V volume
W
i,j
well index for block i, j
x slow (coarse) scale
50
y fast (ne) scale
x, y, z grid block sizes
porosity
mobility
viscosity
layer orientation
density
standard deviation
power averaging exponent, relaxation parameter
streamfunction
Subscripts
0 reference
b bulk
c capillary
i block index
j phase or block index
o oil
t total
w water or well
x, y, z coordinate direction
Superscripts
per unit volume

eective or equivalent
c coarse scale
d default
j ow solution (1 or 2)
k iteration number
w well
51
References
[1] J. E. Aarnes. On numerical methods for multield problems and fast reservoir
performance prediction. Ph.D. thesis, University of Bergen, Norway, 2002.
[2] I. Aavatsmark, T. Barkve, O. Boe, and T. Mannseth. Discretization on non-
orthogonal, quadrilateral grids for inhomogeneous, anisotropic media. Journal
of Computational Physics, 127:214, 1996.
[3] T. Arbogast and S. L. Bryant. A two-scale numerical subgrid technique for
waterood simulations. SPE Journal, 7:446457, 2002.
[4] K. Aziz, S. Arbabi, and C.V. Deutsch. Why is it so dicult to predict the
performance of horizontal wells? Journal of Canadian Petroleum Technology,
38:3745, 1999.
[5] K. Aziz and A. Settari. Petroleum Reservoir Simulation. Applied Science
Publishers, 1979.
[6] C. B. Barber, D. P. Dobkin, and H. Huhdanpaa. The Quickhull algorithm for
convex hulls. ACM Transactions on Mathematical Software, 22:469483, 1996.
[7] J. W. Barker and P. Dupouy. An analysis of dynamic pseudo-relative perme-
ability methods for oil-water ows. Petroleum Geoscience, 5:385394, 1999.
[8] J. W. Barker and S. Thibeau. A critical review of the use of pseudo-relative
permeabilities for upscaling. SPE Reservoir Engineering, 12:138143, 1997.
[9] O. Boe. Analysis of an upscaling method based on conservation of dissipation.
Transport in Porous Media, 17:7786, 1994.
[10] A. Bourgeat. Homogenized behavior of two-phase ows in naturally fractured
reservoirs with uniform fractures distribution. Computer Methods in Applied
Mechanics and Engineering, 47:205216, 1984.
[11] J. Cao and P. K. Kitanidis. Adaptive-grid simulation of groundwater ow in
heterogeneous aquifers. Advances in Water Resources, 22:681696, 1999.
[12] A. Castellini. Flow Based Grids for Reservoir Simulation. M.S. report, Stanford
University, Stanford, CA, 2001.
[13] A. Castellini, M. G. Edwards, and L. J. Durlofsky. Flow based modules for
grid generation in two and three dimensions. 7
th
European Conference on the
Mathematics of Oil Recovery, Baveno, Italy, Sept. 5-8, 2000.
[14] Y. Chen, L. J. Durlofsky, M. Gerritsen, and X. H. Wen. A coupled local-global
upscaling approach for simulating ow in highly heterogeneous formations. Ad-
vances in Water Resources. In review.
52
[15] M. A. Christie and M. J. Blunt. Tenth SPE comparative solution project: a
comparison of upscaling techniques. SPE Reservoir Evaluation & Engineering,
4:308317, 2001.
[16] M.A. Christie. Upscaling for reservoir simulation. Journal of Petroleum Tech-
nology, 48:10041010, 1996.
[17] O. A. Cirpka, E. O. Frind, and R. Helming. Streamline-oriented grid generation
for transport modelling in two-dimensional domains including wells. Advances
in Water Resources, 22:697710, 1999.
[18] N. H. Darman, G. E. Pickup, and K. S. Sorbie. A comparison of two-phase dy-
namic upscaling methods based on uid potentials. Computational Geosciences,
6:527, 2002.
[19] C. V. Deutsch. Calculating eective absolute permeability in sandstone/shale
sequences. SPE Formation Evaluation, 4:343348, 1989.
[20] C. V. Deutsch and A. G. Journel. GSLIB: Geostatistical Software Library and
Users Guide. 2nd edition, Oxford University Press, New York, 1998.
[21] Y. Ding. Scaling up in the vicinity of wells in heterogeneous eld. SPE paper
29137, presented at the SPE Reservoir Simulation Symposium, San Antonio,
TX, Feb. 12-15, 1995.
[22] L. J. Durlofsky. Numerical calculation of equivalent grid block permeability
tensors for heterogeneous porous media. Water Resources Research, 27:699
708, 1991.
[23] L. J. Durlofsky. Coarse scale models of two phase ow in heterogeneous reser-
voirs: volume averaged equations and their relationship to existing upscaling
techniques. Computational Geosciences, 2:7392, 1998.
[24] L. J. Durlofsky, R. A. Behrens, R. C. Jones, and A. Bernath. Scale up of
heterogeneous three dimensional reservoir descriptions. SPE Journal, 1:313
326, 1996.
[25] L. J. Durlofsky, R. C. Jones, and W. J. Milliken. A nonuniform coarsening
approach for the scale up of displacement processes in heterogeneous media.
Advances in Water Resources, 20:335347, 1997.
[26] L. J. Durlofsky, W.J. Milliken, and A. Bernath. Scale up in the near-well region.
SPE Journal, 5:110117, 2000.
[27] M. G. Edwards. Superconvergent renormalization and tensor approximation.
5
th
European Conference on the Mathematics of Oil Recovery, Z. E. Heinmann,
ed., Leoben, Austria, Sept. 3-6, 1996.
53
[28] M. G. Edwards. Unstructured, control-volume distributed, full-tensor nite-
volume schemes with ow based grids. Computational Geosciences, 6:433452,
2002.
[29] M. G. Edwards, R. Agut, and K. Aziz. Quasi K-orthogonal streamline grids:
gridding and discretization. SPE paper 49072, presented at the SPE Annual
Technical Conference and Exhibition, New Orleans, LA, Sept. 27-30, 1998.
[30] M. G. Edwards and C. F. Rogers. Finite volume discretization with imposed ux
continuity for the general tensor pressure equation. Computational Geosciences,
2:259290, 1998.
[31] C. F. Eek-Jensen, I. Aavatsmark, and O. Boe. Upscaling on general quadrilat-
eral grids in 3D with application to eld cases. Proceedings of the EAGE - 10
th
European Symposium on Improved Oil Recovery, Brighton, UK, Aug. 18-20,
1999.
[32] C. L. Farmer. Upscaling: a review. International Journal for Numerical Meth-
ods in Fluids, 40:6378, 2002.
[33] M. H. Garcia, A. G. Journel, and K. Aziz. Automatic grid generation for mod-
eling reservoir heterogeneities. SPE Reservoir Engineering, 7:278284, 1992.
[34] Y. Gautier, M. J. Blunt, and M. A. Christie. Nested gridding and streamline-
based simulation for fast reservoir performance prediction. Computational Geo-
sciences, 3:295320, 1999.
[35] J. Glimm, S. L. Hou, H. J. Kim, Y. Lee, D. H. Sharp, K. Ye, and Q. S. Zou.
Risk management for petroleum reservoir production: a simulation-based study
of prediction. Computational Geosciences, 5:173197, 2001.
[36] J. J. G omez-Hernandez and A. G. Journel. Stochastic characterization of grid
block permeability. SPE Formation Evaluation, 9:9399, 1994.
[37] D. R. Guerillot and S. Verdi`ere. Dierent pressure grids for reservoir simulation
in heterogeneous reservoirs. SPE paper 29148, presented at the SPE Reservoir
Simulation Symposium, San Antonio, TX, Feb. 12-15, 1995.
[38] C. He, M. G. Edwards, and L. J. Durlofsky. Numerical calculation of equivalent
cell permeability tensors for general quadrilateral control volumes. Computa-
tional Geosciences, 6:2947, 2002.
[39] L. Holden and O. Lia. A tensor estimator for the homogenization of absolute
permeability. Transport in Porous Media, 8:3746, 1992.
[40] L. Holden and B. F. Nielsen. Global upscaling of permeability in heterogeneous
reservoirs: the Output Least Squares (OLS) method. Transport in Porous
Media, 40:115143, 2000.
54
[41] L. Holden, B. F. Nielsen, and S. Sannan. Upscaling of permeability using global
norms. 7
th
European Conference on the Mathematics of Oil Recovery, Baveno,
Italy, Sept. 5-8, 2000.
[42] T. Y. Hou and X. H. Wu. A multiscale nite element method for elliptic
problems in composite materials and porous media. Journal of Computational
Physics, 134:169189, 1997.
[43] P. Jenny, S. H. Lee, and H. A. Tchelepi. Multi-scale nite-volume method
for elliptic problems in subsurface ow simulation. Journal of Computational
Physics, 187:4767, 2003.
[44] P. Jenny, C. Wolfsteiner, S. H. Lee, and L. J. Durlofsky. Modeling ow in geo-
metrically complex reservoirs using hexahedral multiblock grids. SPE Journal,
7:149157, 2002.
[45] E. Kasap and L. W. Lake. Calculating the eective permeability tensor of a
gridblock. SPE Formation Evaluation, 5:192200, 1990.
[46] M. J. King, D. G. MacDonald, S. P. Todd, and H. Leung. Application of
novel upscaling approaches to the Magnus and Andrew reservoirs. SPE paper
50643, presented at the SPE European Petroleum Conference, The Hague, The
Netherlands, Oct. 20-22, 1998.
[47] M. J. King and M. Manseld. Flow simulation of geologic models. SPE Reser-
voir Evaluation & Engineering, 2:351367, 1999.
[48] P. R. King. The use of renormalization for calculating eective permeability.
Transport in Porous Media, 4:3758, 1989.
[49] P. Knupp. Jacobian-weighted elliptic grid generation. SIAM Journal of Scien-
tic Computing, 17:14751490, 1996.
[50] P. Knupp, L.G. Margolin, and M. Shashkov. Reference Jacobian optimization-
based rezone strategies for arbitrary Lagrangian Eulerian methods. Journal of
Computational Physics, 176:93128, 2002.
[51] P. Knupp and S. Steinberg. Fundamentals of Grid Generation. CRC Press,
Boca Raton, 1993.
[52] S. H. Lee, L. J. Durlofsky, M. F. Lough, and W. H. Chen. Finite dierence
simulation of geologically complex reservoirs with tensor permeabilities. SPE
Reservoir Evaluation & Engineering, 1:567574, 1998.
[53] S. Mao and A. G. Journel. Generation of a reference petrophysical/seismic data
set: the Stanford V reservoir. Stanford Center for Reservoir Forcasting Report,
1998.
55
[54] L. D. Marini. An inexpensive method for the evaluation of the solution of
the lowest order Raviart-Thomas mixed method. SIAM Journal of Numerical
Analysis, 22:493496, 1985.
[55] O. Mascarenhas and L. J. Durlofsky. Coarse scale simulation of horizontal wells
in heterogeneous reservoirs. Journal of Petroleum Science and Engineering,
25:135147, 2000.
[56] M. J. Mlacnik, A. W. Harrer, and Z.E. Heinemann. Locally streamline-pressure-
potential-based PEBI grids. SPE paper 79684, presented at the SPE Reservoir
Simulation Symposium, Houston, TX, Feb. 3-5, 2003.
[57] J. D. Moulton, J. E. Dendy, and J. M. Hyman. The black box multigrid numer-
ical homogenization algorithm. Journal of Computational Physics, 142:80108,
1998.
[58] A. H. Muggeridge, M. Cuypers, C. Bacquet, and J. W. Barker. Scale-up of well
performance for reservoir ow simulation. Petroleum Geoscience, 8:133139,
2002.
[59] B. F. Nielsen and A. Tveito. An upscaling method for one-phase ow in het-
erogeneous reservoirs. A Weighted Output Least Squares (WOLS) approach.
Computational Geosciences, 2:93123, 1998.
[60] D. W. Peaceman. Fundamentals of Numerical Reservoir Simulation. Elsevier,
New York, 1977.
[61] D. W. Peaceman. Interpretation of well-block pressures in numerical reser-
voir simulation with nonsquare grid blocks and anisotropic permeability. SPE
Journal, 23:531543, 1983.
[62] G. E. Pickup, J. L. Jensen, P. S. Ringrose, and K. S. Sorbie. A method for
calculating permeability tensors using perturbed boundary conditions. 3
rd
Eu-
ropean Conference on the Mathematics of Oil Recovery, Delft, The Netherlands
(1992).
[63] G. E. Pickup, P. S. Ringrose, J. L. Jensen, and K. S. Sorbie. Permeability
tensors for sedimentary structures. Mathematical Geology, 26:227250, 1994.
[64] R. C. M. Portella and T. A. Hewett. Upscaling, gridding, and simulation using
streamtubes. SPE Journal, 5:315323, 2000.
[65] M. Prevost, M. G. Edwards, and M. J. Blunt. Streamline tracing on curvilinear
structured and unstructured grids. SPE Journal, 7:139148, 2002.
[66] M. Ram`e and J. E. Killough. A new approach to the simulation of ows in
highly heterogeneous porous media. SPE paper 21247, presented at the SPE
Reservoir Simulation Symposium, Anaheim, CA, Feb. 17-20, 1991.
56
[67] Ph. Renard and G. de Marsily. Calculating eective permeability: a review.
Advances in Water Resources, 20:253278, 1997.
[68] Y. Rubin and J. J. G omez-Hernandez. A stochastic approach to the problem
of upscaling of conductivity in disordered media: theory and unconditional
numerical simulations. Water Resources Research, 26:691701, 1990.
[69] A. Saez, C. J. Otero, and I. Rusinek. The eective homogeneous behavior of
heterogeneous media. Transport in Porous Media, 4:213238, 1989.
[70] M. R. Thiele, R. P. Batycky, M. J. Blunt, and F. M. Orr. Simulating ow
in heterogeneous systems using streamtubes and streamlines. SPE Reservoir
Engineering, 11:512, 1996.
[71] J. F. Thompson, Z. U. A. Warsi, and C. W. Mastin. Numerical Grid Generation:
Foundations and Applications. North-Holland, New York, 1985.
[72] T. T. Tran. Stochastic Simulation of Permeability Fields and their Scale Up for
Flow Modeling. Ph.D. thesis, Stanford University, Stanford, CA, 1995.
[73] S. Verma and K. Aziz. A control volume scheme for exible grids in reser-
voir simulation. SPE paper 37999, presented at the SPE Reservoir Simulation
Symposium, Dallas, TX, June 8-11, 1997.
[74] X. H. Wen, L. J. Durlofsky, and M. G. Edwards. Upscaling of channel systems in
two dimensions using ow-based grids. Transport in Porous Media, 51:343366,
2003.
[75] X. H. Wen, L. J. Durlofsky, and M. G. Edwards. Use of border regions for
improved permeability upscaling. Mathematical Geology, 35(5), 2003.
[76] X. H. Wen, L. J. Durlofsky, S. H. Lee, and M. G. Edwards. Full tensor upscaling
of geologically complex reservoir descriptions. SPE paper 62928, presented at
the SPE Annual Technical Conference & Exhibition, Dallas, TX, Oct. 1-4, 2000.
[77] X. H. Wen and J. J. Gomez-Hernandez. Upscaling hydraulic conductivities in
heterogeneous media: an overview. Journal of Hydrology, 183:ixxxxii, 1996.
[78] X. H. Wen and J. J. Gomez-Hernandez. Upscaling hydraulic conductivities in
cross-bedded formations. Mathematical Geology, 30:181211, 1998.
[79] C. D. White and R. N. Horne. Computing absolute transmissibility in the
presence of ne-scale heterogeneity. SPE paper 16011, presented at the SPE
Reservoir Simulation Symposium, San Antonio, TX, Feb. 1-4, 1987.
[80] X. H. Wu, Y. R. Efendiev, and T. Y. Hou. Analysis of upscaling absolute
permeability. Discrete and Continuous Dynamical Systems, Series B, 2:185
204, 2002.
57
[81] W. Zijl and A. Trykozko. Numerical homogenization of the absolute perme-
ability using the conformal-nodal and the mixed-hybrid nite element method.
Transport in Porous Media, 44:3362, 2001.
58