You are on page 1of 87

Reservoir simulation with imposed flux

continuity conditions on heterogeneous and


anisotropic media for general geometries, and
the inclusion of hysteresis in forward modeling
Dr.Scient Thesis, Reservoir Mechanics

Geir Terje Eigestad


Department of Mathematics
University of Bergen

24th April 2003

For my sister Monica

Preface
This thesis constitutes the work that I have done during the time as a Dr. Scient. student at the Department of Mathematics, University of Bergen (UiB). It is
written in accordance with the criteria stated by the University Board. The work
has been done in collaboration with the research group at Norsk Hydro and the
Department of Mathematics at University of Bergen.
Magne Espedal (UiB) and Ivar Aavatsmark (UiB/Norsk Hydro Research
Centre) have been advisors for me, and Norsk Hydro has provided the funding
for the work.
The thesis main research part is the five papers A-E, and four of these have
been published in conference proceedings and journals. The main focus of these
papers is on issues related to reservoir simulation with improved reliability, and in
particular discretization methods which should be valid for complex geometry.
The governing equations for fluid flow in oil and gas reservoirs can be
classified according to conventional theory of applied mathematics. This does
not mean that all issues regarding reservoir simulation have been solved. Reliable
reservoir simulation is a complicated field, and many issues are still not known to a
satisfactory level. The oil and gas reservoirs are very complex physical media, and
the level of heterogeneity seen in real sandstones can sometimes overwhelm the
novice mathematician. Sandstones may have rapidly varying structure, porosity
and permeability on a small scale. It can sometimes be hard to convince oneself
that it is possible to model enormous amounts of fluids on a large scale when being
faced with such complexity.
Regardless of these complexities, applied reservoir simulation does study
movement of fluids in complex media. The goal for scientists is to include as much
information as possible in the models for both the fluid flow as well as the physical
media themselves. At multi discipline research centres, geologists, geophysicists,
physicists and mathematicians work together with this drive. Each group may
work with their own small problem, but all these pieces should eventually be
woven together. For the most complex reservoirs one might never be able to
perform accurate and detailed predictions of the future performance. But being
able to see the main trends is very useful, and this is really the goal even the

scientist should aim for.


The contributions in this thesis are mainly in the mathematical and numerical modelling of fluid flow in oil and gas reservoirs. For a person not directly
involved, even the field of numerical modelling may seem like a very narrow
issue to focus on, and merely a detail in the big picture of oil exploration and
production. We, however argue that this is not the case. For future prediction
of the performance of an oil reservoir, one has to build a mathematical model.
This model should be reliable for a wide range of input, such as rock types, well
data, fluid composition etc. Although there will always be uncertainty involved
with such parameters, the bottom line for numerical modelling should be that the
models give us reliable output for cases that can be verified.
For most of the work in this thesis, the physical parameters that describe the
reservoir are assumed to be given. The issue of reliable parameter determination
is a field of its own, and when these are known, the techniques of numerical
modelling/discretisation might be applied to more complex input.
Part I of the thesis gives an introduction to reservoir simulation and the
issues that I have been working with. Part II is a collection of research papers that
have been written during the PhD period.

Outline of Thesis
The thesis is divided into two main parts. Part I gives an overview and summary
of the theory that lies behind the flow equations and the discretisation principles
used in the work. Part II is a collection of research papers that have been written
by the candidate (in collaboration with others).
The main objective of this thesis is the discretisation of an elliptic PDE which
describes the pressure in a porous medium. The porous medium will in general
be described by permeability tensors which are heterogeneous and anisotropic. In
addition, the geometry is often complex for practical applications. This requires
discretization approaches that are suited for the problems in mind. The discretisation approaches used here are based on imposed flux and potential continuity,
and will be discussed in detail in Chapter 3 of Part I. These methods are called
Multi Point Flux Approximation Methods, and the acronym MPFA will be used
for them. Issues related to these methods will be the main issue of this thesis.
The rest of this thesis is organised as follows:
Part I: Chapter 1 gives a brief overview of the physics and mathematics
behind reservoir simulation. The standard mass balance equations are presented,
and we try to explain what reservoir simulation is. Some standard discretizations
methods are briefly discussed in Chapter 2.
The main focus in Part I is on the MPFA discretisation approach for various
geometries, and is given in Chapter 3. Some details may have been left out in the
papers of Part II, and the section serves both as a summary of the discretisation
method(s), as well as a more detailed description than what is found in the papers.
In Chapter 4, extensions to handle time dependent and nonlinear problems are
discussed. Some of the numerical examples presented in Part II deal with two
phase flow, and are based on the extension given in this chapter.
Chapter 5 discusses numerical results that have been obtained for the MPFA
methods for elliptic problems, and Chapter 6 deals with issues related to properties
of the discrete set of (one phase) pressure equations.
Chapter 7 contains summaries of the research papers found in Part II.

Part II: This part contains 5 research papers. The papers mainly deal with
MPFA methods, and issues related to the discrete set of equations that are
obtained by these discretisation methods. Also, one of the papers (Paper E) deals
with the inclusion of hysteresis for forward simulation of two phase flow. The
theme in Paper E is therefore also related to discretisation issues.
The papers included in Part II are:
Paper A: Symmetry and M-matrix Issues for the MPFA O-method on
an Unstructured Grid. Published in Computational Geosciences, Volume 6,
Editors M. G. Edwards, R. D. Lazarov, I. Yotov.
Paper B: MPFA applied to Irregular grids and Faults. Published in Developments in Water Sciences Volume 47, Editors: Hassanizadeh, Schotting,
Gray, Pinder.
Paper C: MPFA for Faults with Crossing Layers and Zig-zag Patterns.
In Proceedings of ECMOR VIII, 2002.
Paper D: A note on convergence of the MPFA O-method; numerical experiments on some 2D and 3D grids. Draft manuscript.
Paper E: Numerical Modeling of Capillary Transition Zones. SPE 64374, in
Proceedings of the SPE APOGCE, 2000.

Acknowledgements
This work would never have been completed without the help of my advisors
Magne Espedal and Ivar Aavatsmark. I owe them a big thank for their guidance
and encouragement throughout this time. I would especially thank Magne for his
positiveness and ability to make me perform more even in the hardest situations.
Ivar needs a big thank for sharing his deep insight in many issues in the field
of applied mathematics. They also both deserve my gratitude for their ability to
handle my moods during times of intense stress.
During the time as a PhD student many people have been helpful and good
to talk to. At Norsk Hydro Research Centre, I would like to thank the following
people: Edel Reiso, Hilde Reme, Rune Teigland and Tor Barkve. Formerly employed Johne Alex Larsen has also been good support. The management of Norsk
Hydro needs a big thank for funding the work.
At CSIRO Division of Petroleum Resources, Dr. Chris Dyt and Dr. Cedric
Griffiths deserve my thanks for letting me stay there for 6 months in 2000/2001.
I would also thank Professor Thomas Russell for letting me stay at University of
Colorado at Denver for three weeks during spring 2002.
At the Mathematical Institute I have had very good colleagues. I want to thank
Erlend ian for many interesting discussions regarding reservoir simulation and
implementation. Jan Martin Nordbotten is also a good source of inspiration. Hans
Fredrik Nordhaug has been of tremendous help in all Latex issues, and all general
computer questions. His attitude to questions is admirable even when the answer
is trivial. I would also like to thank Torbjrn Aadland and Jarle Haukas for letting
me use parts of their C++ code.
Helge Dahle is a very nice person to have at the department; both academically
and personally. I want to thank him for good discussions and taking me and other
PhD students hunting and hiking near his cabin.
Runhild Klausen has been a good friend and colleague. I would like to thank
her for a good cooperation in the academic field and her friendliness at all times.
My family has been fantastic to me all my life, including my period as PhD
student. They have always supported me, and it has been good to retire to them
when the times have been hard.

My loving wife Melanie is the one who made this possible. Without her I
would never even have started doing a PhD. I owe her all my love.

Geir Terje Eigestad


Bergen, April 2003.

Contents
I Introduction
1

Basics
1.1 Darcys law . . . . . . . . . . . . . . . .
1.2 Multiphase flow . . . . . . . . . . . . . .
1.3 Capillary pressure; inclusion of hysteresis
1.4 Geology . . . . . . . . . . . . . . . . . .

1
.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

3
4
5
7
9

Discretisation methods; various formulations


11
2.1 Finite Differences . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Finite element methods . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Mixed finite element methods . . . . . . . . . . . . . . . . . . . 14

Multi Point Flux Approximation methods


3.1 Transmissibility calculations . . . . . . . .
3.2 Unstructured grids; 2D polygonal CVs . . .
3.3 Faults with crossing grid lines . . . . . . .
3.4 Treatment of Dirichlet boundary conditions

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

17
17
25
28
34

Time dependent problems


39
4.1 Nonlinearities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

Numerical results
41
5.1 Analytical solutions . . . . . . . . . . . . . . . . . . . . . . . . . 43

Properties of discrete system


45
6.1 M-matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
6.2 Notes on violation of maximum principle . . . . . . . . . . . . . 46

Summary of the papers


49
7.1 Summary of Papers A and B . . . . . . . . . . . . . . . . . . . . 49
7.2 Summary of paper C . . . . . . . . . . . . . . . . . . . . . . . . 52

0
7.3
7.4

Summary of paper D . . . . . . . . . . . . . . . . . . . . . . . .
Summary of paper E . . . . . . . . . . . . . . . . . . . . . . . .

54
55

8 Further work

57

Bibliography

59

II Published and submitted work

65

A Symmetry and M-matrix issues for the O-method on an Unstructured


Grid
67
B MPFA applied to Irregular Grids and Faults

69

C MPFA for Faults with Crossing Layers and Zig-zag Patterns

71

D A note on convergence of the MPFA O-method; numerical experiments on some 2D and 3D grids
73
E Numerical Modeling of Capillary Transition Zones

75

Part I
Introduction

Chapter 1
Basics
Reservoir simulation is the study of how fluids flow or behave in a reservoir. A
reservoir is a porous medium where hydrocarbons exist in the pore space, and in
the oil industry the goal is to determine how hydrocarbons and water flow/behave
in a reservoir under different conditions.
To do so, one has to derive mathematical and physical models for the processes that occur in the reservoir. Often the physical processes themselves are
of interest, and many physical phenomena (such as interface-interaction when
different phases are present, and complexities associated with hysteresis) are not
fully understood. This thesis approaches reservoir simulation from a mathematical point of view, and some physical assumptions must be made to complete the
mathematical approach. The theory that leads to the partial differential equations
(PDEs) that describe the fluid flow in a porous medium, is standard in many textbooks in the area of reservoir simulation, and further theory details may be found
in for example Aziz [39].
For simulation of fluid flow in a hydrocarbon reservoir (eg. to quantify how
much of a specific fluid flows through a specific area of a reservoir due to pressure
differences), one has to solve a set of coupled PDEs. This set of equations is often
the starting point for applied mathematics, and discretisation of the equations is
a field that has received a lot of attention in the past decades. The PDEs arise
from the principle of mass conservation and a physical description of the velocity/flux dependency of the pressure (Darcys law). Commonly studied are the
mass conservation equations for each of the fluid phases present in a reservoir. In
its original form, the mass conservation equation for a fluid with mass density m
(mass per volume) and fluid density on a domain is given by
Z
Z
Z

mdV +
v ndS = qdV .
(1.1)
t

Here v is the volumetric flow density of the fluid, n is the outward normal

Basics

vector of the surface , and q is a source term. In other words; Eq. (1.1) says
that the accumulation of mass of the fluid within a volume is balanced by the
fluxes (across the edges of the volume) and possible sources q. Eq. (1.1) is an
integral equation. The integral form of the mass balance equation(s) will be used
throughout this thesis. A discretisation approach based on the integral form is
called a Control Volume (CV) formulation. We will use this form because of the
advantage that local mass conservation is satisfied.
When applied to a porous medium, we need to define the porosity of the medium/rock. This is the fraction of the volume of the rock where mobile fluid may
exist, and will be denoted . Hence, we may formally define it by
=

Vp
,
Vc

(1.2)

where Vp is the effective pore volume available for the fluid, and Vc is the total
volume of the rock. The total pore volume of a rock may be larger than Vp , but
this extra volume is not connected to the pore space where fluids may flow. It
should be mentioned that a typical sand stone found in reservoirs from the North
Sea has a porosity of 10-20 percent.
The mass density for a representative volume is hence . Applying this
together with the divergence theorem, we obtain the integral form of the mass
conservation equation for a single phase fluid within a volume V of a porous
medium
Z
()
+ (v) q) dV = 0.
(1.3)
(
t
V
This equation is referred to as the single phase flow equation, or the single
phase pressure equation. The reason for its name is that with some further assumptions discussed below (Darcys law + density function of pressure), this is a
integral PDE where only the pressure needs to be treated as an unknown.

1.1 Darcys law


The French engineer Henry Darcy experimented with flow through different types
of sand during the 19th century. In a range of experiments with flow vertically
through samples of sand, he concluded that the flow through the sands was proportional to the pressure difference between the top and bottom pressure. This law
is now known as Darcys law. In its most primitive form (1D, single phase, gravity neglected), the relation says that the volumetric flow density (Darcy velocity)
v is proportional to the gradient of the pressure p:

1.2 Multiphase flow

p
.
(1.4)
x
The proportionality constant k is called the permeability or conductivity of
the medium. For flow in higher dimensions, the permeability will be a (spatially
varying) tensor K, and when gravity is included, Darcys law for single phase flow
reads
v = k

v = K (p + gd).

(1.5)

This may now be inserted in Eq. (1.3), and pressures will hence enter the
equation. If it is assumed that the fluid at all times occupy the pore space, we
obtain a parabolic PDE which states how the pressure varies in space and time:
Z
()
(K (p + gd)) q)dV = 0.
(1.6)
(
t
V
If the fluid further is assumed to be incompressible, the first term will vanish.
When no sources or sinks are present, one hence obtains a purely elliptic pressure
equation (time independent)
Z

(Ku) = 0.
(1.7)

where we have defined the potential u of the fluid through the relation
u = p + gd.

(1.8)

When suitable boundary conditions are posed, Equation (1.7) has a unique solution.
Whereas reservoir simulation rarely is as simple as to solve time independent
problems, many discretisation methods that are applied for time dependent problems, are based on the purely elliptic equation. Time discretisation is briefly discussed in Chapter 4, and is often included by a standard backward Euler scheme
for time dependent terms.
Most of the work in this thesis (papers A-D) uses Eq. (1.7) as the basis for the
discretisations to be dealt with.

1.2

Multiphase flow

When several phases or components are present in a porous medium, mass conservation must be posed for each of the phases or components. In reservoir simulation, there are usually 3 phases that may exist; oil, gas and water. In the Black

Basics

Oil formulation, hydrocarbons will be grouped, and either be classified as gas or


oil. Heavy components will be grouped as oil, and lighter components will be
grouped as gas components. Depending on temperature, density and pressure,
lighter hydrocarbons may exist in both oil and gas phase. This means that the gas
components may exist in either oil or gas phase. The water component will only
exist in the water phase.
For the oil and water component respectively, the mass conservation equation
is given by.
Z
Z
(i Si )
kri
qi , i = o, w.
(1.9)
(
( i Kui ))dV =
t
i
V
V
For the gas component, the mass conservation equation is slightly different
because the component may exist both in oil and gas phase, and we refer to for instance [40] for details. Eq. (1.9) is a parabolic type of equation, which contains an
elliptic term (second term on left hand side). Although each mass balance equation is parabolic, the total system of mass balance equations has a more hyperbolic
character. This is for instance discussed in [52]. The new quantities Si , kri and i
of Eq. (1.9) will be explained in the following: If the three components/phases
oil, gas and water exist in our system, three mass conservation equations must be
solved in the region we investigate. The efficient pore space will be assumed to
be filled with fluid at all times. Depending on the processes, the different phases
will compete to occupy this space. It is then natural to define phase saturations.
By this we mean the ratio of the available pore space the different phases occupy.
Formally, the saturation of phase i in a representative volume V is defined as
Vi
,
(1.10)
V
where Vi is the volume occupied by the specific phase. Hence, the saturations vary
between 0 and 1. Since the phases compete to occupy pore space, the effective
permeabilities will be reduced, and this is reflected in the occurrence of the entities
kri , of Eq. (1.9). These entities are called relative permeabilities, kri , and will be
reduction factors of the permeability K. The relative permeabilities are functions
which depend on the saturation of phase i, ie. kri = kri (Si ). They take values
between 0 and 1.
For hysteresis models, the relative permeabilities also depend on the direction
of the flow processes, as discussed in the next subsection.
The entity i is called the viscosity of the phase i. This is a weakly nonlinear
term, and for many applications the viscosities of pure fluids will be treated as
constants. For gases, simple models use the linear relationship
Si =

g = bp.

(1.11)

1.3 Capillary pressure; inclusion of hysteresis

The quantity b will be a gas dependent constant. For a three phase system the three
saturations will be unknowns of the three equations of the form (1.9). In addition
to this, each of the three phase pressures will be unknowns. As densities, i , are
also assumed to depend on the pressures, the total number of unknowns in the
three coupled mass conservation equations is 6. To close the system of equations,
three closure relations are needed.
One closure relation is intuitive if the pore space is filled with fluids at all
times:
X

Si = 1, i = w, o, g,

(1.12)

where the different subscripts denote water, oil and gas respectively. To get the
two remaining closure conditions, one needs to know more about the physics of
multiphase interactions. Moreover, to make the system of mass balance equations
solvable, the physics must be quantified.

1.3

Capillary pressure; inclusion of hysteresis

In black oil models, it is commonly assumed that the phase pressures are related through a quantity called the capillary pressure. When only two phases
are present, the capillary pressure is given by the pressure difference between the
non-wetting fluid and the wetting fluid:
Pc = pnw pw .

(1.13)

For two phase problems in the oil industry this may be oil and water, or gas
and water. If all three phases (water, oil and gas) exist, two capillary pressures will
be used; the capillary pressure between oil and water and the capillary pressure
between gas and oil:
Pcow = po pw ,

Pcgo = pg po .

(1.14)

Experimental results show that Pcow is (mainly) a function of the water saturation (Sw ), and that Pcgo is (mainly) a function of the gas saturation (Sg ). The two
capillary pressure relationships then close the system of equations obtained when
modelling three phase flow. Since the capillary pressure and the relative permeabilities of Eq. (1.9) depend on the solutions Si , the set of mass balance equations
are nonlinear (with respect to solution variables).
Much emphasis is made in experimental reservoir physics to describe the capillary pressure behaviour. In addition to being dependent on the water saturation,
the capillary pressure description also depends on whether a phase is entering or

Basics

Figure 1.1: Primary drainage capillary pressure curve with imbibition capillary
pressure curves starting out from different points on the primary drainage curve.
exiting the pore space. In physics this is referred to as irreversible processes or
hysteresis.
When water (wetting phase) enters the pore space, this is called imbibition,
and when oil (non-wetting phase) enters the pore space this is called drainage.
In Paper E in Part II of this thesis, a consistent capillary pressure and relative
permeability hysteresis model is included for forward modelling of two phase flow
in a reservoir. The idea is that the fluid flow simulator should dynamically update
capillary pressure and relative permeabilities according to what direction the flow
has. For reasons explained in the paper, the reservoir will be initialised with a
(single) primary drainage capillary pressure curve prior to production from the
reservoir. Figure 1.1 illustrates possible imbibition curves that may start out from
the primary drainage curve.
If pure imbibition takes place in a reservoir, the capillary pressure time evolution will be described by imbibition curves. When gravity is included in a model,
and the system is assumed to be in vertical equilibrium, the capillary pressure
will vary linearly with height, and different grid cells at different height start out
with different saturations. This means that even for pure imbibition throughout
the reservoir, different capillary pressure descriptions apply for different parts of
the reservoir. Such cases have been investigated in Paper E. The model allows for
different capillary pressure curves for all grid cells. In addition to (primary and
secondary) imbibition curves, arbitrary secondary drainage curves may apply for
the different cells throughout the simulation. Examples of such curves are visualised in Fig. 1.2a. Here we depict examples of capillary pressure scanning curves
for cells that have their origin at some point of the primary drainage curve. In
Fig. 1.2b relative permeability scanning curves are visualised, and are explained
as follows: A cell may have an initial water saturation corresponding to the saturation at point (1). If imbibition starts for this cell, the relative permeability curve

1.4 Geology

2
a. Capillary pressure.

b. Relative permeability

Figure 1.2: a: Capillary pressure scanning curves. Scanning curves are closed
loops, and arbitrary reversals may occur within the loops. b: Relative permeability
scanning curves.
will follow the lowermost path to point (2). If the direction of the flow is reversed
at point (2), the relative permeability curve will follow a different path back to
(1). Notice that the loop between (1) and (2) is closed; this is an experimental
observation, and is included in the scanning curve generation.
Similar explanations may be given for the curves of Fig. 1.2a, and details may
be found in Paper E.

1.4

Geology

Reservoir simulation has become very advanced over the past decades. This is due
to at least two main reasons. The first reason is of course the computer revolution, which allows large models to be implemented for computing. The second
reason is that the geology of a North Sea reservoir may often be known quite accurately (compared to other porous media). Simulation grids may then be very
large, and the level of details can be very high. Figure 1.3 shows the simulation
grid for a realistic field example. One of the reasons for the high level of details is the considerable effort which is put into seismic measuring prior to drilling
exploration wells. Also, the history may be well known for a reservoir that has
been producing hydrocarbons for many years (5-15 years) through logging and
measured production etc. This again may be used to verify or recalculate geophysical data. Parameter estimation is an important area in its own, and production data is essential to recalculate/calibrate a model with respect to porosity and
permeability. Because the geology may be known at such a detailed level, one
may be required to model flow on grids that include faults and general complex

10

Basics

Figure 1.3: Simulation grid for realistic North Sea field example

Figure 1.4: Extracted part of sandstone found in the Oslo Fault.


geometry. Complex geometry has been a key issue in the work of this thesis. Papers B and C discuss MPFA discretisation for simulation grids where faults have
been included. Faults may be observed and studied in analogues to North Sea
reservoirs. One examples of such an analogue may be found in the Oslo Fault.
Such analogues may be useful for understanding the flow mechanisms in faults
and fractures. Also, they support the need to model fluid flow on simulation grids
which contain faults, have large aspect ratios and grid cells that may degenerate
to triangles/tetrahedra.
A part of some sandstone found in the Oslo Fault is visualised in Fig. 1.4,
where the scale is a few meters. Even on this small scale the permeability may
vary extensively.

Chapter 2
Discretisation methods; various
formulations
Discretisation of the governing equations for fluid flow has been the issue of many
researchers involved in porous media flow. Finite difference methods (FD) have
been used extensively in commercial reservoir simulation, whereas finite element
methods (FEM) have a strong support in applied mathematics.
We will here briefly describe the following methods: (Control volume) finite
differences, finite element- and mixed finite element methods.
The starting point for the discretisation methods to be discussed is the purely
elliptic equation, which is obtained when a possible source q has been added in
Eq. (1.7):
(Ku) = q.
(2.1)
Suitable boundary conditions will also apply, and for reservoir simulation these
will usually be no-flow at boundaries, ie. Ku n at the boundary of some
domain .

2.1

Finite Differences

Finite difference based discretisation (FD) is used extensively in conventional


reservoir simulation. The basic idea is to approximate derivatives by discrete differences. In 1D the use of finite difference approximations for the elliptic equation
(2.1) is justified since the permeability K = k then is a scalar entity.
For 1D, the second order problem u00 = f may be discretised by applying the
forward discretisation for the first derivative u0 (x) = (u(x + h) u(x))/h, which
leads to a similar treatment of u00 (x). For multi dimensional problems with nonorthogonal grids and general tensor permeability description, such differences are

12

Discretisation methods; various formulations

not directly applicable to Eq. (2.1), as the differentiation involves the permeability
as well as the pressure. The 1D version of Eq. (2.1) is

u(x)
(k(x)
) = 0,
(2.2)
x
x
so that differentiation of the permeability must also be included. Denoting F =
u
k x
, the discrete version of (2.2) takes the form (F (x + h) F (x))/h. Without
going into detail, there are possible grids to use for finite difference discretisations:
point- or cell distributed grids. Cell centred grids (non-uniform) are preferable
to use because they yield local mass conservation, and have consistent fluxes.
This is enforced through harmonic averages of permeabilities obtained at media
discontinuity points. It should be noted that the FD discretisation on cell centred
grid in 1D is second order convergent [34], even though the discretisation scheme
is inconsistent.
For 2D orthogonal grids where the principal axes of the permeability tensors
are aligned with the coordinate axes, the 1D principles may be utilised to obtain discretisation schemes on cell centred grids. Such grids will be termed Korthogonal grids. Convergence on cell centred grids is proven in [34] and [26].
For general non-orthogonal grids in 2D, and generally varying permeability
tensors, direct finite difference discretisation does not lead to convergent schemes.
The numerical solutions convergence to the wrong solution, and this may in some
cases be interpreted as convergence to solutions for continuous problems where
the permeability tensors are not symmetric. The MPFA discretisation methods
to be discussed in Chapter 3 will be able to handle full tensor representation of
the permeabilities and non-orthogonal grids. Our numerical results (eq. Paper
D) indicate/show convergence for examples with skew grids and full permeability
tensor description.

2.2 Finite element methods


We will here briefly discuss the variational formulation of the problem (2.1), and
the use of finite element methods to discretise the equation.
The equation (Ku) = q can be rewritten in order to obtain a weak formulation of the equation: take the original equation, multiply by a suitable test
function of some function space V [47], and integrate over a domain :
Z
Z
( (Ku)vd = qvd.
(2.3)

Eq. (2.3) may be rewritten by using the rules for partial integration:
Z
Z
Z
Ku vd
Ku nvdS = qvd.

(2.4)

2.2 Finite element methods

13

If the boundary conditions are Ku n = 0 at , the second term of the left hand
side vanishes, and the variational equation reads
Z

Z
Ku vd =

qvd,

v V .

(2.5)

This type of boundary condition is called a natural boundary condition, whereas


it is possible to define boundary conditions for the functions v V by v = 0 on
. The latter boundary condition is called an essential boundary condition. It is
common to use the notation (, ) for inner products in some inner product space,
see for example [47]. Letting (, ) here denote the inner product in the inner
product space L2 (), we may rewrite Eq. (2.5) as
(Ku, v) = (q, v),

v H 1 (),

(2.6)

The solution of Eq. (2.3) will be denoted a weak solution of the original equation
(2.1). One hence wants to solve this equation for u. The equivalence of the weak
solution and the original solution for smooth K is proven in various textbooks
on finite element analysis, see for example [47] and [50]. The
R equivalence Rof
the weak solution and the solution of the integral equation (Ku) = q
is also valid for discontinuous K. Although not discussing discontinuous K, the
equivalence of the solutions is discussed in [48] for solutions with regularity less
than H 2 .
The term (Ku, v) of Eq. (2.6) is a bilinear form, see [47]. Since K is symmetric, this bilinear form will be symmetric. The bilinear form will be bounded (or
equivalently continuous) provided that K is bounded (K L ()). Further, the
bilinear form will be coercive because of the positive definiteness of K. Existence
and uniqueness of a solution u of Eq. (2.6) is then proven by the Lax-Milgram
theorem [47].
To solve a discrete problem of the form (2.6), one may choose a finitedimensional subspace Vh H 1 . The discrete problem then reads: find uh Vh
such that
(Kuh , v) = (q, v),
v Vh .
(2.7)
This method is called the (Continuous) Galerkin Finite Element Method. Under
the same conditions as above (Vh closed subspace of H 1 , (Kuh , v) bounded,
symmetric bilinear form that is coercive on Vh ), there exists a unique uh that solves
(2.7). The Galerkin Finite Element Method is an example of a method where the
discrete solution is sought in a subspace Vh V , whereas nonconforming methods
do not seek a solution in a subspace of V . Discontinuous Galerkin Methods [37]
are also methods where jumps in the pressure are allowed.

14

Discretisation methods; various formulations

2.3 Mixed finite element methods


Fluxes are not calculated directly by the Galerkin Finite Element Method. A post
processing needs to be done in order to find the fluxes from the discrete pressures,
and it is known [52] that there is a loss of accuracy when fluxes are found from
approximate pressures. For discontinuous coefficient problems, the fluxes found
from the Galerkin FEM are not continuous. This was one of the reasons that mixed
finite element methods (and other locally conservative methods) were developed
for simulation of fluid flow in porous media.
The starting idea behind the mixed finite element method is to do a splitting
of the second order elliptic pressure equation into two equations, where the two
equations express the physics of the pressure, u, and the velocity, v, respectively.
One then expresses Eq. (2.1) as a system of two first order equations
v = Ku,

(2.8)

v = q.

(2.9)

and

The two equations (2.8) and (2.9) will both be used for weak formulations.
To obtain a weak formulation for the two equations, one multiplies by suitable
test functions, and integrates over some domain . For the first equation we obtain
Z
Z
Z
Z
1
K v wd = u wd = wud +
w nuds

The boundary condition u = 0 on is chosen (natural boundary condition here),


and the last term on the right hand side then cancels.
Using the same inner product notation as in Sec. 2.2, the weak formulation of
the mixed system reads: Find (v, u) H (div) L2 such that
(K 1 v, w) + (u, w) = 0,
( v, z) = (q, z),

w H (div),
z L2 .

(2.10)

The discrete problem reads: Find (v h , uh ) V h Ph H (div) L2 such that


(K 1 v h , w) + (uh , w) = 0,

w V h ,

( v h , z) = (q, z),

z Ph .

(2.11)

2.3 Mixed finite element methods

15

The discrete space Ph is the space of piecewise constants (which is a subspace


of L), and the discrete space V h is in many applications the lowest order Raviart
Thomas space [27], [28]
The power of the mixed methods lies in the way they are able to handle media
discontinuities. The methods may be shown to be locally mass conservative when
the media have discontinuities. One of the drawbacks of the method is that the
discrete system of equations is a saddle point problem, which is more costly to
solve than symmetric positive definite systems. In addition, more unknowns need
to be solved for.
The mixed finite element framework has also been extended to account for a
control volume formulation by Russell [29], [30] and coworkers. The Expanded
Mixed Finite Element Method was developed by Wheeler, Yotov and coworkers
[35], [36]. The relationship between this method and the MPFA O-method was
recently found by Klausen [32].
The next chapter deals with the MPFA discretisation methods, and are alternative discretisation methods to the methods discussed in this chapter.

Chapter 3
Multi Point Flux Approximation
methods
The main work in this thesis has been on various Multi Point Flux Approximation
methods, and discretisation issues related to these. It is important for a discretisation method to handle complex geometry in order for it to be applied for realistic
cases. In addition, the discretisation methods should be applicable in physical
space if this is required. The MPFA methods have this advantage.
MPFA methods have also been the issue in work by Edwards and coworkers
[17], [18] and [19].
We will here give a thorough description of the methods for the geometries
that have been considered in the papers: 2D quadrilaterals, 2D polygonal control volumes, 3D quadrilaterals with crossing grid lines and handling of general
Dirichlet boundary conditions.

3.1

Transmissibility calculations

The basic idea behind the MPFA methods is to define discrete fluxes for edges
of control volumes by a reasonable restriction of the medium. The ideas can be
carried out for several types of geometries, and can be illustrated in 2D by the
general quadrilateral grid visualised in Fig. 3.1.
In practical reservoir simulation, the geometry may very well have similar
trends as in this figure. Fluxes should therefore be calculated for each cell edge,
where the orientation of the different edges may vary highly throughout the grid.
Furthermore, the permeability tensor K will be heterogeneous and anisotropic,
and the principal directions are likely to vary from grid cell to grid cell. However, K is assumed to be symmetric and positive definite. A good discretisation
method should take this physical picture into account, and produce reliable fluxes

18

Multi Point Flux Approximation methods

Figure 3.1: Possible local gridding situation; grid cells need not be orthogonal.

a. Polygonals

b. Triangles

Figure 3.2: Unstructured, conforming 2D grids.


for all cell edges that are encountered.
The MPFA approach differs from methods such as FEM (continuous Galerkin
method) [47], and Finite Differences [39], [40] by the fact that the MPFA methods discretise the flux itself, whereas the methods mentioned above discretise the
pressure. But the MPFA methods also yield pressure equations to solve instead of
flux equations. Mixed finite element methods consider both pressures and fluxes
as unknowns, as discussed in Sec. 2.3.
We will first try to give a unified treatment of the ideas that lie behind the
MPFA discretisation method for various conforming grids. By conforming grids
we mean grids where corners of grid cells meet other corners. The grid shown in
Fig. 3.1 is an example of such grids. Some other examples are shown in Fig. 3.2.
As mentioned above, the basic ideas behind the MPFA control-volume formulation is to determine fluxes across all edges of all grid cells in our grid. Hence,
all grid cells are control volumes for which we pose mass conservation. By this
approach local mass conservation will be satisfied. Further, we want to express
the edge fluxes by some local formula/molecule. In doing so, we define a corner
of degree n at corners where n grid cells meet. This definition goes for both triangular, quadrilateral and polygonal control volumes, see Fig. 3.3. The MPFA
approach has also been discussed for non-conforming grids in [6], [7] and [12].

3.1 Transmissibility calculations

19

Figure 3.3: Left: Interaction region and associated corner for 2D quadrilateral
control volumes, degree of corner is 4. Right: Interaction region for polygonal
control volumes, degree of corner is 3.

The degree of the corner is defined to be the number of grid cells that meet in
the corner. Around each corner an interaction region is defined, and the interaction
regions make up a dual grid to the original control volume grid. The control
volume grid and associated dual grid for a general conforming 2D quadrilateral
grid is visualised in Fig. 3.4

Figure 3.4: Control volume grid (fully drawn lines) and associated interaction
regions (dotted lines).

Here, the interaction regions (indicated by dotted lines) may in general have
8 edges, and a particular interaction region is visualised on the left hand side of
Fig. 3.3. For some special polygonal control volumes the interaction region may
be a triangle as in the rhs. of Fig. 3.3. The reason why the interaction region has
8 edges for the quadrilateral case is explained by the algorithm for constructing
the interaction regions in Fig. 3.3: From each node we draw (dotted) lines to
midpoints of cell edges, and an octagon arises around each corner. The midpoints
of the edges are denoted dividing points or continuity points. The dividing points
need not be midpoints of the edges, but for the 2D discussion in this thesis, the

20

Multi Point Flux Approximation methods

midpoints of edges are chosen as dividing points/continuity points.


For the interaction depicted on the left hand side of Fig. 3.3, four grid cells
meet at the corner associated with the interaction region. There will be four cell
edges that meet in the corner, and parts of these cell edges will be defined to belong
to the interaction region. These parts will be defined sub-interfaces or half-edges,
and it is these we want to calculate fluxes for within the interaction region. A
whole cell edge for both quadrilaterals and general polygonals will be covered by
two neighbouring interaction regions. The quadrilateral case is depicted in Fig.
3.5.
II

I
E

Figure 3.5: Whole cell edge is covered by two neighbouring interaction regions.
Total flux obtained by summing fluxes for two half edges. Interaction regions for
quadrilateral grid cells depicted.

If the degree of the corner associated with an interaction region is n, we say


that n grid cells interact in the interaction region. The number of interacting cells
determine how the flux-molecule for a sub-interface will be. For each sub-interface
the flux will be a weighted sum of the potentials at the nodes of the cells that
interact:
fi =

n
X

tij uj .

(3.1)

j=1

These weights will be denoted transmissibilities, and the discretisation problem for the MPFA control volume method is to determine these. When the grid
is quadrilateral, n = 4 in Eq. (3.1), and for the 2D polygonals studied in Paper A,
n = 3. We may express the n sub-interface fluxes of an interaction region by the
general matrix notation
f = T nn u,

(3.2)

where f and u are n 1 vectors, and T nn is an n n matrix. The flux across the
whole cell edge E of Fig. 3.5 is found by summing two half edge fluxes of two

3.1 Transmissibility calculations

21

neighbouring interaction regions. In total this yields a flux that is a weighted sum
of the 6 (node) potential values of the cells that touch upon the whole cell edge.
Similarly, the flux for a whole cell edge of the polygonal control volumes (with
triangular interaction regions) discussed in Paper A, is given by a weighted sum
of 4 potential values.
Looking at the continuous case, fluxes across cell edges are given by
Z
f = Ku n dS.
(3.3)
S

Assume that the flux across sub-interface i of the left interaction region of Fig
3.3 is to be considered. If the flux is to be approximated by using the information
to the left of the sub-interface, the gradient of the potential for this cell is needed.
By the assumption that the potential is linear for the part of the grid cell contained
in the interaction region, u is hence constant, and the (discrete) flux across the
sub-interface will also be constant (see Eq. (3.4) below).
Likewise, the flux may be approximated by the information to the right of the
sub-interface, and the two discrete flux expressions should then be equal. The
discrete flux across the cell edge can hence be expressed as
fi = nTi K i ui ,

(3.4)

where the index i denotes if the flux is approximated by the information of either
the cell in the positive or anti-positive direction of the edge (positive direction
defined by positive direction of normal vector). If, for instance, the potential was
assumed to be bilinear, fluxes would not be constant across cell edges.
Formally, the flux continuity condition for each sub-interface i of an interaction region reads
fi,i = fi,i+ ,

(3.5)

The gradient of the potentials in each of the sub-cells contained in an interaction region will now be derived in order to use explicit expressions for Eq. (3.4).
Since the potential U is linear on each sub-cell j of the control volume, the
equation that describes a potential plane applies:
Uj (x) = Uj (x x0 ) + U0 ,

(3.6)

where x is the position vector, x0 is the node of grid cell j, and U0 is the potential
value at the node.
The potentials at the continuity points may be used as the basis for determining the gradient of the potential in each sub-cell of an interaction region. This
information may be inserted in Eq. (3.6) to obtain

22

Multi Point Flux Approximation methods

x0

x1

Figure 3.6: Variational triangle associated with part of quadrilateral control


volume contained in interaction region; x0 is node of grid cell, x 1 and x 2 are
corresponding dividing points/continuity points

U (xk x0 ) = u k u0 ,

k = 1, 2.

(3.7)

This system may be written as




u 1 u0
u 2 u0

(x 1 x 0 )T
(x 2 x 0 )T

XU =

(3.8)

(3.9)

where X is the matrix



X=

The inverse of X must hence be found. Inward normal vectors of the variational triangle are introduced:
1 = R(x2 x0 ), 2 = R(x1 x0 ),

(3.10)

where

R=

0 1
1 0


.

(3.11)

The vectors and points used here are illustrated in Fig. 3.6, where the gradient
of the potential is to be determined for a sub-cell of a quadrilateral grid cell. The
matrix R has the property that for vectors a and b, aRb is equal to the third
component of the cross product between a and b. Moreover, the determinant of
X is
T = det(X) = (x 1 x0 )T R(x 2 x0 ).

(3.12)

3.1 Transmissibility calculations

23

which is equal to twice the area spanned by the points x0 , x 1 and x 2 (in a right
handed system seen from x0 ). To find the inverse, observe that by Eq. (3.10)
matrix multiplication yields


(x 1 x 0 )T
(x 2 x 0 )T




( 1 , 2 )


=

(x 1 x 0 )T 1 (x 1 x 0 )T 2
(x 2 x 0 )T 1 (x 2 x 0 )T 2

2F 0
0 2F


.


.

(3.13)

Hence,

1 
1, 2 .
2F
And finally, the gradient of the potential reads
X 1 =


U = X

u 1 u0
u 2 u0

(3.14)


.

(3.15)

Explicit fluxes across sub-interfaces are now obtained by inserting the expression for the gradients in (3.4).
Having found an expression for the gradient of the potential of each sub-cell of
an interaction region, it is seen that the discretisation principles may be applied
to a whole range of grids. In the case of conforming quadrilaterals in 2D, four
flux conditions of the form (3.5) are posed. In [8] scalar coefficients ijk are
introduced, and they are defined by
nTi K j jk
ijk =
,
Tj

(3.16)

where ni is the positive normal vector of sub-interface i, K j is the permeability of


cell j, and jk is local outward normal vector k of cell j. The entity Tj is equal
to twice the area of the triangle for which the two normal vectors jk , k = 1, 2 are
legs of.
The flux continuity conditions (3.5) lead to a local system of equations that
will define the transmissibilities of Eq. (3.1). Each sub-cell uses two continuity
points (in addition to the node) to determine u. Inserting the explicit expressions
for the gradients yields discrete fluxes that depend on the potentials at continuity
points, and are hence not on the form (3.1). However, the potentials at the continuity points can be eliminated so that discrete fluxes can be given that are only
dependent on the potentials at the nodes of the control volumes.

24

Multi Point Flux Approximation methods

This is illustrated by rewriting the system of flux continuity equations (3.5) in


matrix form, where the information about the potentials at the continuity points
and the potentials at the nodes has been split
Av = Bu.

(3.17)

The vector v = [u1 , u 2 , u 3 , u 4 ] then contains the potentials at all continuity


points, and u = [u1 , u2 , u3 , u4 ] contains the information of the potentials at the
nodes. The 2 matrices A and B are 4 4, and their elements consist of the scalar
coefficients of (3.16). Details may be found in [8]. To complete the derivation,
the left hand side of (3.5) is rewritten in terms of potentials at continuity points
and nodes:
f = Cv Du.

(3.18)

The matrices C and D also have elements that consist of coefficients (3.16).
If A is non-singular, v can now be eliminated from (3.17); v = A1 Bu. Inserting this in (3.18), the flux expressions of the desired form are obtained
f = T u = CA1 Bu Du.

(3.19)

We now briefly comment on the solvability of the system, and hence the determination of the transmissibility matrix T of Eq. (3.19). This discussion is also
valid for the 2D unstructured grids discussed in the next section.
In an interaction region where n grid cells interact, 3n degrees of freedom are
originally available to determine the n piecewise linear potential planes (as functions of the position vector x). Since each potential plane must honour the nodes
(of the sub-cell in question), n degrees of freedom will be locked. This means
that 2n degrees of freedom are left for flux and potential continuity. There are n
sub-interfaces contained in the interaction region, and from Eq. (3.5) there are n
flux continuity equations. Now, there are n degrees of freedom left for potential
continuity. To derive explicit expressions for the gradients of the potentials, the
continuity points were used as a basis, see Eq. (3.8). The vector v of Eq. (3.17)
contains the n unique potential values at the continuity points. This is how the last
n degrees of freedom have been used; namely that the potentials are continuous at
the dividing points.
Full potential continuity can in general not be achieved by the assumption of
piecewise linear potentials. This is the price that has to be paid to get (piecewise)
constant fluxes across sub-interfaces.

3.2 Unstructured grids; 2D polygonal CVs

3.2

25

Unstructured grids; 2D polygonal CVs

A discussion of the transmissibility calculation for the sub-interfaces of triangular


interaction regions will now be given. In Paper A, the MPFA O-method is discussed for polygonal control volumes, and it is the transmissibility calculations
that apply for this geometry that will be discussed.

Figure 3.7: Left: Polygonal control volume grid with associated triangular interaction regions. Right: Whole cell edge, E, covered by two neighbouring triangular
interaction regions.
The geometry may be illustrated by Fig. 3.7, and the notation used for this
geometry will be the same as in [3], [4] and in Paper A.
The interaction regions of the polygonal control volumes discussed in Paper
A, will be triangles. Such an interaction region was visualised in Fig. 3.3. For
this case, fluxes across sub-interfaces will also be given by the general formula
(3.1). Now 3 grid cells will interact, so that n = 3 in Eq. (3.1). Further, 3 flux
continuity equations of the form (3.5) must hold. Similar coefficients as (3.16)
were introduced in [3], but the notation is slightly different. More specifically, the
scalar coefficients for the case of polygonal control volumes are given by

ik

nTi K k
k
2Fk

(3.20)

The quantities used in the above coefficients are explained by Fig. 3.8.
Note that parts of the triangle edges are used in the definition of the scalar
coefficients (3.20). In Paper A, the continuity points/dividing points of the triangle edges are midpoints of the triangular edges, such that the outward normal

vectors +
i and p(i) are equal. However, as the method was derived for arbitrary
continuity points along the triangle edges, we have kept the original notation.
The transmissibilities for the sub-interfaces of this triangular interaction region are derived using exactly the same principles as for quadrilateral control
volumes. Flux continuity is posed across the sub-interfaces, and potential continuity is posed at dividing points. As for the quadrilateral case, explicit expressions

26

Multi Point Flux Approximation methods


+
3

n2
n3 n1
+
1

+
2

Figure 3.8: Triangular interaction region with quantities used in scalar coefficients.

are obtained for the gradients of the potentials for each sub-cell in the interaction
region. For each sub-cell, the potentials at the node and the dividing points are
used as a basis for expressing the gradient. This is illustrated by Fig. 3.9, which
shows the corresponding picture to Fig. 3.6, where inward normal vectors of the
interaction region were used for expressing the gradients of the potentials.
m(k)
xm(k)

xk
k

xk

Figure 3.9: Variational triangle

Using these points and vectors, one obtains the general expression for the
gradient of the potential of sub-cell k, k = 1, .., 3 in a triangular interaction region:
1 +
( (u m(k) uk ) +
k (u k uk )),
2F k
where m(k) is a backward integer function defined by
(
k 1, for k > 1,
m(k) =
3,
for k = 1.
Uk =

The flux continuity conditions are with this notation stated by

(3.21)

(3.22)

3.2 Unstructured grids; 2D polygonal CVs

fi,i = fi,p(i) ,

i = 1, .., 3.

The indexing p(i) denotes the forward integer function


(
i + 1, for i < 3,
p(i) =
3,
for i = 3.

27

(3.23)

(3.24)

When inserting the scalar coefficients in Eq. (3.23) and the flux expression
fi = nTi K j Uj for sub-interface i, one obtains a similar system of equations as
Eq. (3.17). In the original paper [3], the system of flux continuity equations was
written as
Av + Bu = Cv + Du.

(3.25)

As in the discussion of the transmissibility calculations for the case of quadrilateral grids, v denotes pressure values at the continuity points, and u denotes the
pressures at the nodes of the cells that interact.
The above system of equations is written differently from the system (3.17)
by the fact that the information there was split, so that matrices acting on v and
u respectively were defined. This is purely an issue of definition. The matrix
difference (A C) of Eq. (3.25) serves as the matrix A of system (3.17), and the
matrix difference (D B) of (3.25) serves as the matrix B of system (3.17).
The matrices A, B, C and D of the system of equations (3.25) are defined,
and can be found on p.6 of Paper A. Their elements consist of scalar coefficients
(3.20). As in the discussion of the transmissibility calculations for the quadrilateral case, the potentials at the continuity points may be eliminated. From Eq.
(3.17) one obtains
v = (A C)1 (D B)u.

(3.26)

Together with, for example, the left hand side expression for the flux of (3.17),
the transmissibilities are found:
f = A(A C)1 (D B)u + Bu.

(3.27)

Hence, the transmissibility matrix is


T = A(A C)1 (D B) + B.

(3.28)

Having found the transmissibilities, the coefficient matrix for the total system
of discrete pressure equations may be found by an assembly procedure. This is
also discussed in Paper A.

28

Multi Point Flux Approximation methods

3.3 Faults with crossing grid lines


3D structured grids are the most commonly used grids for field reservoir simulation. This is mostly because of the frontier commercial simulator Eclipse [57],
where structured, orthogonal, conforming quadrilaterals were used in the first released versions. Although the grids are structured, one is faced with several challenges if for instance faults need to be taken into account. Grid cells need no
longer be conforming, and the discretisation of fluxes across cell edges must be
re-investigated.
Paper C of Part II extends the MPFA methods to handle cases where lateral
grid lines are allowed to intersect. Examples of such grids are encountered when
using (Geophysical) models for real fields. In particular, the North Sea reservoir
Visund uses a reservoir description where such faults are found. The simulator
Eclipse allows reservoir simulation for this grid type, but the flux discretisation
uses a standard two-point flux expression, see [6] and [57] for details.
MPFA Transmissibility calculations to handle faults with crossing lateral grid
lines are presented in Paper C, and the discretisation uses the same principles as
for previously investigated grid types. Note, however, that some additional issues
need to be taken care of because of the under-determined character the system of
equations that determines the transmissibilities has.
Before discussing these special 3D faults, we briefly discuss standard faults in
2D and 3D. The extension of the MPFA method to handle faults was first done for
2D quadrilateral grids in [6]. Here the concept of non-conforming grids was used,
and interaction regions were defined around such corners. As seen in Fig. 3.10,
three grid cells meet in a corner.
Fault line

Figure 3.10: Corners when faults are encountered in 2D


A natural interaction region to use here is the triangular interaction region
which was introduced in Sec. 3.2, or a similar interaction region where three grid
cells interact. It has been reported that this interaction region has an apparent disadvantage when one of the grid cells of the interaction region is inactive [6], where
it was shown that there will be no flow across one of the three sub-interfaces for

3.3 Faults with crossing grid lines

29

certain situations. Instead, in [6] it was suggested to use an extended interaction

4.

.3

1.
.

.2
5.
Figure 3.11: Interaction region associated with 2D faulted corners. Five grid cells
interact
among grid cells. The suggested interaction region will contain information from
5 grid cells instead of 3. Hence, the fluxes across each of the sub-interfaces i are
expressed as
fi =

5
X

tij uj .

(3.29)

j=1

This is visualised in Fig. 3.11. The MPFA method was extended to faulted
corners in 3D in [7]. Similar ideas as for the 2D faults were used to derive transmissibilities for the extended interaction regions, and details may be found in [7].
Five different situations may occur, and the flux molecules will contain information from 8 to 11 grid cells. One possible situation is illustrated in Fig. 3.12, where
7 grid cells meet in a corner. For this case the flux molecules contain information
from 9 grid cells. These types of corners will be termed regular corners, and 5-8

Figure 3.12: Ordinary faulted corners in 3D


grid cells will touch upon such corners. However, in [7] the case where grid lines
intersect was not discussed. This situation was discussed in Paper C. Lateral grid
lines may for some simulation grids cross, and the flux across general polygonal
flux interfaces must be discretised. The point for which lateral grid lines intersect,
will be termed an irregular corner, and an interaction region will be associated

30

Multi Point Flux Approximation methods

with it. How these new corners arise is depicted in Fig. 3.13. For this corner, 4
grid cells will touch upon it, and these 4 grid cells will be the interacting grid cells
in the associated interaction region. The sub-interfaces for which the transmissibilities need to be calculated are visualised in Fig. 3.14. Here, 6 sub-interfaces
will be associated with the interaction region. To understand this, consider Fig.
3.13. The local numbering of the four grid cells is as follows: Local grid cell
1 is the front cell, grid cell 2 is the back cell. Above cell 1 there will be a grid
cell (not drawn), and this cell has local number 3. Above cell 2, grid cell 4 will
be located (also not drawn). A cell interface is a a sharing of a common surface
between two neighbouring grid cells. From Fig. 3.13 it is seen that there will be
6 cell interfaces touching upon the irregular corner. Denoting these sharings by
a tuple (a1, a2) (where a1 and a2 correspond to the local numbering of the grid
cells), the 6 sharings are given by (1, 3), (2, 4), (1, 2), (1, 4), (3, 2) and (3, 4). The
6 sub-interfaces will be parts of these 6 cell interfaces, and this is discussed below.
For each of the 6 sub-interfaces we want to find fluxes that are linear combinations of these 4 interacting grid cells:

fi =

4
X

tij uj ,

i = 1, ..., 6.

(3.30)

j=1

4
2
3
1

Figure 3.13: Crossing grid lines arising from faults

As for the 2D case, U is needed in order to derive discrete fluxes and transmissibilities for the sub-interfaces belonging to this new type of interaction region.
This is done by using the same methodology as in [8]. For three-dimensional
cases (ie. both this situation and the situation with ordinary faulted corners), U
for each sub-cell of an interaction region, can be defined implicitly by the values
of the potentials at three corresponding potential continuity points:

3.3 Faults with crossing grid lines

31

5
1

Figure 3.14: Sub-interfaces associated with irregular corner

u 1 u0
XU = u 2 u0 .
u 3 u0

(3.31)

The derivation of the inverse of X follows exactly the same ideas as for the
2 dimensional case. We introduce variational tetrahedra spanned by the cell
centre x0 (node) and three local continuity points x 1 , x 2 , x 3 that are located on
the three sub-interfaces seen from the sub-cell in question. These points are locally numbered cyclically. A variational tetrahedron used for the faulted corners in
3D discussed in [7] is visualised in the left hand side of Fig. 3.15. The continuity
points are three of the corners of the tetrahedron, and are located on 3 separate
surfaces of the cell in question. Due to the special geometry when lateral grid
lines are allowed to intersect, the variational tetrahedra will be somewhat different from the regular case. This is depicted in the right hand side of Fig. 3.15. Two
of the tetrahedron corners will be located on the same cell interface (clarified by
Fig. 3.17).
4 of the 6 sub-interfaces are parts of side surfaces of the grid cells that interact,
whereas the two remaining sub-interfaces are parts of the top/bottom surfaces of
the cells of the interaction region. This is visualised in Fig. 3.14. The 4 subinterfaces that lie in the surface generated by the side surfaces of the grid cells
that interact are numbered 3-6 in the figure. In Paper C it was assumed that the
side surfaces of the cells are planar, and the surfaces 3-6 will hence also be planar.
To better see how these four sub-interfaces arise, it may be useful to visualise
the situation orthogonally on the intersection, as in Fig. 3.16.
Different situations may occur, and the whole cell interfaces that touch upon
the corner may have from 3 to 6 corners. Each of these corners will correspond
to an interaction region. For a complete discretisation of the fluxes across the
whole cell interfaces, the cell interface will be shared among the 3-6 interaction
regions which the corners of the whole flux interfaces correspond to. The sharing
of flux-interfaces among the interaction regions may easily be defined by dividing
a whole cell interface with n corners into n segments/sub-interfaces. This is done

32

Multi Point Flux Approximation methods


x3

x3

x2

x1

x1

x2

Figure 3.15: Variational tetrahedra for a) regular case b) crossing grid lines.

Front
Back
Figure 3.16: Intersection of lateral grid lines seen orthogonally on the intersection
P
by finding the geometric centre of the cell interface (x0 = 1/n xi ), and drawing
straight lines to the midpoint of all edges. All sub-interfaces will then have 4
edges, and the approach is easy to implement.

4
6

3
5

Figure 3.17: Sub-interfaces belonging to side surfaces of grid cells


Fig. 3.17 explains how sub-interfaces 3-6 of Fig. 3.14 arise. At each subinterface there will be one point for which the potentials of neighbouring grid cells
are continuous, ie. a continuity point. These points are marked with dots, and
their numbering corresponds to the numbering of sub-interfaces. This explains

3.3 Faults with crossing grid lines

33

why two of the corners of the variational tetrahedra used here are located on the
same side surface (of the cell in question).
For all 6 sub-interfaces, flux continuity will be posed, and (together with potential continuity at 6 points) this leads to a local system of equations to determine
the transmissibilities of the interfaces. This system is written out in Paper C, and
the redundancy of the system is discussed. The redundancy implies that more
conditions are needed in order to determine the transmissibilities. The determination of transmissibilities for the sub-interfaces was partially answered by using a
criterium for no-flow and a criterium for uniform flow in Paper C. Details may be
found in the paper.
Linear flow assumes a homogeneous medium, and it is not clear whether the
transmissibility calculations can be generalised to arbitrary heterogeneous media
without additional information.

34

Multi Point Flux Approximation methods

Figure 3.18: Gridding of region of interest; inner grid cells are sketched by continuous lines. Grid cells surrounding the area of interest are indicated by dotted
lines.

3.4 Treatment of Dirichlet boundary conditions


The first published papers on MPFA methods [1]-[8] handled boundaries with
homogeneous Neumann conditions. The implementation of such conditions were
straight forward since the transmissibility coefficients could be calculated by the
same ideas as for the fully active interaction region. Details may be found in [3].
Paper D investigates convergence of the O-method on some quadrilateral grids
in 2D. Such a study requires more general conditions than homogeneous Neumann conditions. All our analytical examples required using fixed potentials at
some boundary.
We now discuss various ways of how to extend the MPFA O-method for 2D
quadrilateral grids to account for general Dirichlet conditions.
The simplest way to include fixed potentials at boundaries is to add an encircling layer of artificial grid cells around the medium which we study. If the
analytical pressure solution is known, we may fix the potentials at these outer
cells, and solve the discrete pressure equations for all the inner grid cells (at the
nodes). At each refinement level one solves the pressure equation on the same
physical medium, as illustrated by Fig. 3.18.
This approach does not incorporate the boundary conditions (Dirichlet and
Neumann) for the active area exactly, but rather approximately. This can be
seen in Figure 3.19, where an extended interaction region is shown. The points
x3a and x4a are nodes of the artificial grid cells, and at these points we use exact
pressures. The transmissibility calculations then yield approximate fluxes f1d and
f2d and corresponding discrete pressures at the continuity points x 2d and x 4d .
As the approach above requires that the analytical solution is known outside
the area of interest, a better approach of handling general Dirichlet boundary conditions is to actually incorporate fixed potentials/pressures at the actual boundaries
of the medium we study. Utilising the ideas from the original transmissibility de-

3.4 Treatment of Dirichlet boundary conditions

x4a

35

x3a

x4d f1d

x
f2d

x2d
x2

x1

Figure 3.19: Incorporation of approximated boundary conditions

x1D

x1

x2D

xF

x2

Figure 3.20: Treatment of Dirichlet boundary conditions


rivations, this can be done by studying an interaction region which covers a part
of the boundary.
Consider the extracted part of a grid visualised in Fig. 3.20. The nodes of the
two grid cells are x1 and x2 respectively. The points xD1 and xD2 are points on
the boundary, and the potentials at these point will be specified by the Dirichlet
boundary conditions. To derive the transmissibilities for the sub-interface going
from xF to the boundary, we pose continuity of the flux at the sub-interface and
continuity of potential at the dividing point x F . When the potential is assumed
to be piecewise linear in each of the sub-cells for the cells that interact in the
interaction region, the local system is closed with respect to degrees of freedom.
Hence, we may derive the transmissibilities through the equation
K 1 U1 n = K 2 U2 n.

(3.32)

The gradients can be calculated by using variational triangles defined on each


sub-cell and the notation used in Sec. 3.1. Hence,
1X
ik (u ik ui ),
(3.33)
Ti
for the each of the cells i = 1, 2. The vectors ik are as before the inward normal
vectors of the edges (x ij xi ) of the variational triangles (j, k cyclic, k local numbering of dividing points of triangle). Ti is equal to twice the area of variational
Ui =

36

Multi Point Flux Approximation methods

triangle i. The potentials u ik are the potentials at the points x ik (ie. either potential
at boundary or at continuity point x F ).
Having expressed the gradients, the flux continuity equation (3.33) now reads

1
K 1 ( 11 (u1D u1 ) + 12 (u F u1 ))n =
T1

1
K 2 ( 21 (u F u2 ) + 22 (u2D u2 ))n.
T2

(3.34)

This may be written as


av = bu + k 1 u0 ,

(3.35)

where a (scalar), b and k 1 (1x2 vectors) depend on geometry and permeability.


Further, v = u F is the unknown potential at the dividing/continuity point x F , u =
[u1 , u2 ] represents the vector of node potentials, and u0 = [u1D , u2D ] is the vector
involving the pressures u1D and u2D specified by the boundary conditions.
The entities a, b and k 1 of Eq. (3.35) can be given explicitly:
a=
b = [

1
1
K 1 12 n + K 2 21 n
T1
T2

1
1
K 1 ( 11 + 12 )n, K 2 ( 21 + 22 )n]
T1
T2

1
1
K 1 11 n, K 2 22 n]
(3.36)
T1
T2
To express the transmissibilities for the sub-interface, we may write the left hand
side of (3.34) as
k1 = [

f = cv du + k2 u1D ,

(3.37)

where c (scalar), d (1x2 vector) and k2 (scalar) depend on geometry and permeability. These can also be given explicitly:
c=
d = [

1
K 1 11 n
T1

1
K 1 ( 11 + 12 )n, 0]
T1

1
K 1 11 n
T1
The unknown u F may now be eliminated from Eq. (3.35):
k2 =

(3.38)

3.4 Treatment of Dirichlet boundary conditions

x1D e1D

e2D
e1

x1

37

x2D
x2

xF

Figure 3.21: Boundary interaction region, used to include Dirichlet boundary conditions.

v = a1 (bu + k 1 u0 ),

(3.39)

and inserting this in Eq. (3.37) yields


f = (ca1 b d)u + ca1 k 1 u0 + k2 u1D .

(3.40)

Hence, the flux is expressed as


f = T u + T 0 u0 .

(3.41)

The transmissibilities from T will enter the system matrix A for the discrete
set of (pressure) equations MU = R, and the transmissibilities from T 0 will enter
the right hand side R of the system. To construct the full system matrix, a looping
over all interaction regions will be done, and for inner interaction regions we
calculate transmissibilities in the standard way, discussed in Sec. 3.1.
The above method to take Dirichlet boundary conditions into account seem
to be a reasonable way to handle the problem. However, for K-orthogonal grids,
some information seems to be lost from the boundary because of the orthogonality
of the grid. Moreover, the elements of T 0 of Eq. 3.41 will be zero for the case
depicted in Fig. 3.22. It is possible to move the continuity point and the boundary
points x1D and x2D in such a way that the same points as in Edwards [17] MPFA
method are obtained. When this is done, the elements of T 0 in (3.41) will in
general be nonzero.
The third way to deal with Dirichlet boundary conditions for the MPFA
method is to somehow include fluxes on the boundary as well.
The information about the pressure on the boundary will as before be included
for the transmissibility calculations, as illustrated in figure 3.21. The potential
values at the boundary are used in the expressions for the gradient of potentials

38

Multi Point Flux Approximation methods

Ky
Kx

Figure 3.22: K-orthogonal grid; Dirichlet transmissibilities will be zero


for each cell. From this, the discrete fluxes for the sub-interfaces are
f11 = K 1 U 1 ne1
f12 = K 2 U 2 ne1
f1D = K 1 U 1 ne1D
f2D = K 2 U 2 ne2D

(3.42)
(3.43)
(3.44)
(3.45)

The gradient, U i , i = 1, 2 will use the potential values at the points xi , xDi and
xF . The flux from cell 1 into the flux interface e1 , calculated by the information
from cell 1, is denoted f11 , and equals the flux calculated for interface e1 into cell
2, (information from cell 2), f12 . The potential at the dividing point xF may hence
be eliminated, from Eq. (3.42) and Eq. (3.43). If we next use the potential at the
dividing point xF in (3.43)-(3.45), we are left with the flux f = (f12 , f1D , f2D )T as
a function of the potentials of the active cells 1 and 2, u = (u1 , u2 )T , and potentials
at the boundary, u0 = (u(xD1 ), u(xD2 ))T .
The fluxes can now be expressed as:
f = T u + T u0
The transmissibilities of T will be incorporated in the the discrete version of
(2.1), while T u0 incorporates the boundary conditions on the right hand side of
the corresponding discrete system.

Chapter 4
Time dependent problems
Reservoir simulation models are used to predict future performance of reservoirs.
For multiphase problems the flow is hardly ever stationary, so in order to make
forward simulations (in time), the time dependent terms of the governing PDEs
must also be discretised.
Several options are possible, but it is common to use backward Euler time
discretisation to obtain fully implicit discrete equations. This means that all unknowns are coupled at a given time level, and the equations for the unknowns need
to be solved simultaneously. A semi discrete version of the mass balance equation
for phase then reads
mn mn1 + tn

n
fj
= qn .

(4.1)

The entities above are: m accumulated mass of phase , n and n 1 time levels,
fj phase fluxes through surface number j of a control volume, and q a possible
source term. Nonlinearities and how these affect the numerical solution strategy
are discussed next.

4.1

Nonlinearities

The discretisation of multiphase flow equations/problems yields sets of non-linear


equations to solve (with respect to solution variables). The non-linearities arise
because
of the appearance of relative mobilities (see Chapter 1) of the phase fluxes
R kri
i Kui n, and the inclusion of capillary pressure. Relative mobilities will be
handled by standard upstream evaluation across edges. A possible modification
of this has not been an issue for the MPFA O-method yet. This means that the
the transmissibilities derived for single phase flow will be scaled by relative mobilities according to the flow direction. By an implicit solution strategy, relative

40

Time dependent problems

permeabilities will be updated at each iteration level.


At each time level, indicated by Eq. (4.1), nonlinear algebraic equations must
be solved. Our in-house research simulator solves these algebraic equations by
using a Newton discretisation approach, see for instance [39], [40].
Based on the time stepping, an outer iteration is defined, and the goal is to
solve the equation based on Eq. (4.1), which determines the change in solution
variables from one time step to the next time step
X
r(x) = m(x) mn t
fk (x) tq(x) = 0.
(4.2)
k

Here x denotes the vector of the (primary) solution variables in all grid cells.
For three-phase flow these may be the three quantities po , Sw and Sg , and the
remaining 3 solution variables (So , pw , pg ) are found by the previously discussed
closure relations.
The solution by the Newton discretisation procedure is defined by
r 1 ()
) [x x(1) ] = r (1) ,
(4.3)
x
where the superscripts and 1 denote iteration levels. Eq. (4.3) is solved
until x() x(1) has reached some desired limit (ie. specified convergence criteria). The data structure of the research simulator allows for general geometry,
r
). Moreover, there will be more
and this will be reflected in the Jacobi matrix ( x
couplings in the Jacobi matrix when a control volume interacts with more control
volumes than when two point flux schemes are used for flux discretisation. This
generalisation is accounted for in, for example, the handling of simulation grids
with crossing lateral grid lines in Paper D. Details of the Newton strategy may be
found in [39].
(

Chapter 5
Numerical results
Paper D investigates convergence numerically for the MPFA O-method. In the literature, there are numerous reference examples to be found for the purely elliptic
problem (1.7) with suitable boundary conditions. Linear pressure solutions (in
space) are perhaps the easiest reference solutions to investigate. Although such
a reference solution is very simple, it is a good first test to examine the discrete
solution obtained by a discretisation method when the grid is challenging.
The discrete flux operator is examined for various non-orthogonal grids in
both 2D and 3D in Paper D. It is shown that the discrete fluxes (from the MPFA
O-method) across sub-interfaces of interaction regions are exact when applied to
linear potentials. Intuitively we should expect that the discrete pressure solution
obtained from the corresponding system Au = b also should be exact since the
discrete set of pressure equations is constructed by assembly of transmissibilities.
The pressure solution on a randomly perturbated grid is investigated for the
Support Operator Method (SOM) for a homogeneous medium in [44], where the
analytical flow is uniform in a given direction. Their discrete pressure solution is
exact for that case.
Fig 5.1 shows a random 8 8 grid, and the underlying medium is homogeneous. The discrete pressure solution is calculated by the MPFA O-method, where
the analytical pressure solution is linear (driven by specified (Dirichlet) pressures
on the left and right hand side of the medium). The right hand plot of Fig. 5.1
reveals that the pointwise pressures are linear and exact for this case.
A similar test can be carried out for a discontinuous coefficient problem, where
the analytical pressure solution is piecewise linear. The medium is divided into
two parts, and the permeability ratio is 1/10 across the medium discontinuity. The
discontinuity line is here given by rx + sy = 0, where r = tan(/3)/(1 + tan(/3))
and s = 1/(1 + tan(/3)). This is visualised in Fig. 5.2. The permeabilities are
kx = ky = 10 for rx + sy < 0 and kx = ky = 1 for rx + sy > 0. The true solution is
given by

42

Numerical results

0.8
1.2

0.6
1

0.4
0.8

0.2

Pressure

0.6

0.4

0.2
0.2

0.4
0

0.6
0.2
0.2

0.2

0.4

0.6

0.8

1.2

xcoordinate

0.8
0.8

0.6

0.4

0.2

0.2

0.4

0.6

0.8

a. Random grid;.

b. Pressure vs. x-coord

Figure 5.1: Numerical pressure solution; uniform flow in x-direction.


5

4
0.5

0.4

0.3

2
0.2

1
0.1

0.1

1
0.5

0.2

0.3

0
0.4

0.5
0.5

0.5
0.4

0.3

0.2

0.1

0.1

0.2

0.3

0.4

0.5

a. Random grid.

0.5

0.4

0.3

0.2

0.1

0.1

0.2

0.3

0.4

0.5

b. Discrete pressure solution

Figure 5.2: Numerical pressure solution; discontinuous coefficient problem.


Piecewise linear reference solution.

(
u(x, y) =

rx + sy,
10(rx + sy),

for rx + sy < 0,
for rx + sy > 0.

(5.1)

The numerical results are shown in Fig. 5.2, where the pressures at the grid nodes
are employed. The discrete pressure solution on our 16 16 grid is piecewise
linear and exact also for this case.
More sophisticated cases are tested in Paper D. Especially interesting test
cases are cases where the permeability of the medium has sharp jumps, and the
analytical solution exhibits singularities in the velocity field. Such reference solutions are discussed in [46].

5.1 Analytical solutions

43

10

10
0.5
0.5
0
0
0.5

0.5

Figure 5.3: Plot of analytical solution on [0.5, 0.5] [0.5, 0.5]. Solution exhibits singularity at the origin.

5.1

Analytical solutions

The problems dealt with in Paper D are special cases where analytical solutions
of the elliptic equation (1.7) are known. The cases are discussed because they
regard important physics which discretisation methods should handle. In reservoir
simulation, non-orthogonal grids are used extensively, and the permeabilities vary
from grid cell to grid cell. Four grid cells will meet at each corner for a conforming
grid. If the grid cells have different permeability, the pressure may be less regular
near these corners than in areas where the medium is homogeneous.
In particular, cases where the analytical solutions exhibit singularities, should
be discussed. Such a solution is depicted in Fig. 5.3 on a square domain. Using
polar coordinates, this is a solution of the form
ui (r, ) = r (acos() + bsin()),

(5.2)

where the index i denotes regions with different permeability. The parameter
defines the regularity of the solution, and will depend on heterogeneity and the
angles of the grid cells that meet at a corner. These functions are known to not
belong to the function space H 2 (), but rather the fractional space H 1+ (),
where < , see [46]. The solutions are derived based on separation of variables,
and details may be found in Strang and Fix [46].

Chapter 6
Properties of discrete system
There has recently been work done to investigate the relationships between MFEtypes of methods and MPFA methods, [30] and [32]. Furthermore, these relations
are used to analyse convergence of the MPFA O-method on a general quadrilateral
grid (with continuously varying heterogeneity) theoretically [32].
As the MPFA method does not directly build on a variational formulation of
the governing (elliptic) equation, the approach in Paper A was not to investigate convergence but rather to investigate the coefficient matrix that arises when
discretising Eq. 2.1. This matrix is also frequently referred to as the Discrete
operator (corresponding to the continuous operator).
Various works in the literature, [41], [42] etc., indicate that convergence of a
discretisation method is related to three important issues; conservation of mass,
consistency of fluxes (see discussion in [4]) and stability. MPFA-methods are
locally mass conservative, so this is not an issue to be discussed. The concept of
consistent flux means that the discrete fluxes are reasonably approximated. We
claim that this is the case for fluxes obtained by the MPFA approach, since these
allow for general geometry, and general tensor representation of permeabilities.
Exceptions are cases where the solutions are far from H 2 -regular, as indicated by
the last numerical example of Paper D.
Stability issues are among the reasons for studying the discrete operator. Certain important properties of the continuous elliptic operator should be inherited for
the discrete operator. The operator (K ) is a self adjoint operator, and the
discrete operator should then ideally be symmetric [43]. This was one of the issues
discussed in Paper A (see discussion of Paper A below). Stability of a method is
assured if the discrete system satisfies the coercivity condition [8], [51]:
uT Au uT Du,

(6.1)

where A denotes the coefficient matrix of the system, and u denotes the discrete
solution vector. For the coercivity condition to be satisfied, A should be positive

46

Properties of discrete system

definite [49].
A special class of matrices that are positive definite, are symmetric M-matrices:

6.1 M-matrices
The following definition is taken from [51].
Definition: A matrix A is called an M-matrix if the following is satisfied:
a) The diagonal elements aii are positive whereas the off-diagonal terms
aij , i 6= j are non-positive.
b) A is non-singular and A1 0.
Such matrices are appealing when looking at the continuous problem
(Ku) = q. For positive sources q, the solution u should also be positive, and this is assured if the coefficient matrix that arises from the discretisation
of (K ) is an M-matrix. Moreover, the positiveness of the solution is
assured if the coefficient matrix has a non-negative inverse. Such matrices are
called monotone (operators), see also [8] for a discussion.
The M-matrix property is examined in Paper A. However, this property may
be too restrictive for a method in order to say something about the monotonicity.
It is well known that a matrix may have a non-negative inverse without being an
M-matrix. This means that the matrices one really is looking for, are matrices
that are symmetric, positive definite, and have non-negative inverses. As far as we
know, it is not known how such matrices can be classified based on their elements,
but work has been done recently to analyse the monotonicity of the O-method.

6.2 Notes on violation of maximum principle


A specific case has been investigated in [15] where it was of interest to find out
whether a non-negative inverse still could be obtained even though the conditions
for the matrix to be an M-matrix were violized. The O-method was examined on a
parallelogram grid (2D) for homogeneous media. The coefficient matrix will then
be described by the legs of the grid cells (and the corresponding inner angles of
the grid cells). The elements of the coefficient matrix are given by the cell stencil
for the central cell (cell 1) of Fig. 6.1. These coefficients are denoted m1 .....m9 ,
and correspond to the local numbering of grid cells in Fig. 6.1. For the O-method
on a parallelogram grid on a homogeneous medium, these coefficients are given

6.2 Notes on violation of maximum principle

47

by [8]

2
9

Figure 6.1: Local numbering of grid cells

a = a1 T Ka1 ,

b = a2 T Ka2 ,

c = a1 T Ka2 ,

(6.2)

and
m2 = a +

c2
c c2
c2
c c2
, m3 = , m4 = b + , m5 = ,
d
2 2d
d
2 2d

(6.3)

and
P mi+4 = mi , i = 2, ..., 5. The variable d is given by 2ab/(a + b), and m1 =
i=2 9mi . The permeability K is positive definite such that a > 0, b > 0, and
c2 < ab. It is shown that two criteria are sufficient for the coefficient matrix A to
have a non-negative inverse. These criteria are
(1) m2 m4 m3 m1 0, (2) m2 m4 m5 m1 0,
2

(6.4)

c
) d. Since these criteria only are sufficient, the
which are the same as |c|(2 ab
inverse will be non-negative even if (6.4) is not satisfied. In [15] it is stated that for
grids with approximately unity aspect ratios, grid skewness of about 45 degrees
may be allowed. For stronger aspect ratios and/or anisotropy ratios, the limitations
on grid skewness are more severe. The reader may notice that grid skewness of
45 degrees has been tested in Paper D for various reference solutions.
When the inverse is investigated numerically on an n n grid, it is found that
angles down to /6 yield positive inverses for large n (n approximately 20) for
aspect ratios in the interval (1/2, 2).
This result may be interpreted as an unfortunate way of choosing neighbouring
information for the fluxes across cell edges when the grid cells are very skewed.
When the aspect ratio of the grid is close to 1, it is possible to construct a scheme
where the bandwidth is reduced. This should intuitively lead to a more diagonally
dominant matrix than the 9-point scheme does. Such a scheme can be illustrated
by Fig. 6.2.

48

Properties of discrete system

3
2
4
1

Figure 6.2: Possible 7-point cell molecule. Combination of different types of


interaction regions
This molecule can be obtained by combining two types of interaction regions.
Interaction region 2 and 4 are triangular interaction regions which are special cases
of the triangles investigated in Paper A. Interaction region 1 and 3 are standard interaction regions associated with (conforming) corners for 2D quadrilaterals. This
scheme has not been implemented by us, but for the homogeneous case the scheme
may easily be generated by the use of explicit expressions for transmissibilities of
sub-interfaces of the two different types of interaction regions.

Chapter 7
Summary of the papers
Summaries of the papers which Part II of this thesis consists of, will now be given.
Some issues may also have been left out of the papers themselves, and some of
these are discussed here. Details are also to be found in the previous sections.

7.1

Summary of Papers A and B

This Subsection will give an overview of Paper A and Paper B:


Paper A: Symmetry and M-matrix Issues for the O-method on an Unstructured Grid
Authors: G.T. Eigestad, I. Aavatsmark, M. Espedal
Paper B: MPFA applied to Irregular grids and Faults.
Authors: G.T. Eigestad, I. Aavatsmark, E. Reiso, H. Reme, R. Teigland.
The theme is similar for both papers. The MPFA O-method is discussed
for irregular 2D grids in Paper A, and an extension to account for certain gridding
situations which arise when faults occur in 2D, is discussed in Paper B.
Paper A discusses the O-method for polygonal control volumes. These will
be control volumes for which the interaction regions are triangles. The MPFA Omethod for such cases is also referred to as the Triangle method in the literature.
It is of interest to study the coefficient matrix that arises when the elliptic
term (Ku) is discretised. Because the underlying continuous operator is
self-adjoint, this should ideally be inherited for a corresponding discrete operator
(coefficient matrix). This means that the matrix should be symmetric and positive
definite (since K is positive definite). There exist alternative discretisation methods (Support Operator Methods, [43]), which actually use the preservation of such

50

Summary of the papers

properties for the construction of discrete schemes. Secondly, the physics behind
the purely elliptic problem (2.1) implies that positive sources q should lead to
positive pressures (for homogeneous Dirichlet boundary conditions). Translated
to the discrete problem, this means that the coefficient matrix should be monotone,
see Chapter 6 and [8].
The interaction regions are as mentioned above triangular, and three grid cells
interact in such an interaction region. This was visualised in Fig. 3.8. The derivation of the transmissibilities are discussed in Sec. 3.2.
The first case studied in the paper is the case with homogeneous permeability
(for the whole medium). For this case, the coefficient matrix is shown to be symmetric as long as the dividing points (see Sec. 3.2) are midpoints of triangle edges.
Further, the dual triangular grid should satisfy the Delauney condition in order for
the coefficient matrix to be an M-matrix. This means that opposite angles of triangles which share an edge, should satisfy the condition 1 + 2 < .
For general permeabilities, symmetry of the discrete operator is not satisfied directly for the MPFA O-method applied to polygonal control volumes. This was
further examined in the paper, and the question raised was whether it is possible
to obtain a symmetric, discrete operator. To answer this question, symmetry equations were defined locally for the interaction regions. The dividing points were
fixed (midpoints of triangle edges), and the equations were solved with respect
to the location of the triangle points inside the interaction regions. The control
volumes change with this procedure, but the goal of the experiment is to find geometrical situations for which symmetry of the discrete operator is satisfied. Solutions of the symmetry equations may be sought for arbitrary heterogeneity and
anisotropy ratios inside the interaction regions. The algorithm for finding triangle
points that yield symmetric coefficient matrices, may easily be implemented for
numerical purposes. We were not able to obtain solutions for all cases we tested
(anisotropic cases), but we were able to extract important information for layered
media (two cells of an interaction region have the same permeability). For these
cases it is possible to construct grids a priori that honour the media discontinuities,
and at the same time yield coefficient matrices that are symmetric. For isotropic
cases it was found that the triangle point should be located on a straight line which
connects two dividing points, and we refer to Paper A for details. If the anisotropy
ratios are different, but the principal axes of the permeability tensors are aligned
with the discontinuity line, the triangle point should also be located on this straight
line.
This result was found using the symmetry algorithm numerically, and is later
proven analytically in the paper.
The next question raised in the paper is how the dual triangular should be
in order for the local contributions from the interaction regions to the coefficient
matrix to have the correct signs. This is referred to as local M-matrix property

7.1 Summary of Papers A and B

51

in the paper. It is found that this local M-matrix property yields a constraint on
angles of the dual triangular grid, heterogeneity and anisotropy combined.
The local M-matrix property does not fully answer the question whether the
total coefficient matrix is an M-matrix, as a neighbouring interaction region may
compensate for unwanted signs/contributions from one interaction region. The
M-matrix property derived for homogeneous media in Paper A is an example of
this. Such an analysis may be carried out in the same manner as in Paper A to
answer this for specific cases.
Paper B is an overview paper, and the symmetry results found in Paper A are
used in a discussion of practical gridding. Paper B also discusses the fully
heterogeneous case on a right angled triangle. Symmetry may be obtained for
this case when the media are isotropic. Further, faults in 2D are illustrated, and
are seen to be generalisations of the interactions of a triangular interaction region.
Main results of the papers:
The main results of the papers may be summarised:
Operator fully analysed on homogeneous media
A general symmetry algorithm was investigated
Layered, isotropic media can always be gridded to obtain symmetry and
local M-matrix property for the triangle method.
Layered media with varying anisotropy ratios can be handled for special
cases
A 90 degree triangle has been found to have good properties regarding
symmetry for cases where all three grid cells have different permeability
Practical gridding based on symmetry results has been sketched
An overview of gridding issues regarding 2D faults is given

52

Summary of the papers

7.2 Summary of paper C


This subsection is a summary of Paper C:
MPFA for faults with crossing grid lines and zig-zag patterns.
Authors: G.T. Eigestad, I. Aavatsmark, E. Reiso, H. Reme.
This paper discusses a possible extension of the MPFA method for faults
to cases where the lateral grid lines of control volumes/grid cells are allowed to
intersect. At the point where the grid lines intersect we define an irregular corner,
and we associate an interaction region to it. Such cases are occasionally seen for
grids that are used for reservoir simulation of North Sea fields. The geometry for
such faults was illustrated in Sec. 3.3. The idea behind the work was to be able
to simulate fluid flow on grids with general faults in the corner point geometry
setting.
A simulation grid may then contain both faulted corners as discussed in [7],
in addition to this new type of corners. In the paper we have shown how to calculate transmissibilities for the sub-interfaces in the interaction region associated
with the irregular corner. To do so, we used the same framework as for ordinary
corners [7] in addition to using the principle of exact discrete fluxes for linear
flow. Linear flow assumes a homogeneous medium, and it is not clear whether the
transmissibility calculations can be generalised to arbitrary heterogeneous media
without additional information.
A structured 3D (quadrilateral) grid where all corners are matching has a reasonable simple data-structure. Fluxes across cell surfaces will in general be 18point flux molecules, and this is only modified when inactive grid cells exist or
when the grid cell is on the boundary.
When faults are allowed, the flux molecules may be much more general, as a
grid cell may have many more and different neighbours that for the matching grid.
This has an impact on the Jacobian matrix which is needed when a fully implicit
problem is solved.
The data structure gets more complicated when crossing grid lines are allowed.
Surrounding all corners, suitable interaction regions are defined. This means that
different parts of a cell surface is shared between several interaction regions, and
transmissibilities will be calculated for sub-interfaces of the different interaction
regions.
The paper contains only one example where we have simulated two phase
flow on a simulation grid which contains irregular corners. The grid is, however,
a rather extreme example as nearly all cells have crossing lateral grid lines. This
is not likely to occur for a realistic grid.
The paper further investigates planar faults vs. zig-zag faults in 3D by a

7.2 Summary of paper C

53

selection of numerical examples.


Notes on data-structure, crossing grid lines
In the work of implementing this new type of geometry, a substantial amount
of time was spent on generalising the data structure in the in-house reservoir
simulator. To sum up, the following data structure needs to be defined in order to
compute transmissibilities, calculate phase fluxes and define the Jacobian matrix
properly:
All irregular corners must be found
Cells that touch upon corners must be known
Whole flux-interfaces must be divided and shared among different interaction regions
Sub-interfaces in different types of interaction regions must be defined
Corners that touch upon whole flux-interfaces must be known
All corners (both regular and irregular) must know their neighbour corners
All flux-interfaces must know their neighbouring flux-interfaces
The main results of Paper C can be summarised as:
Main results of the paper:
New geometry handled by the MPFA methods
Data structure made more general; simulation grid may contain general
faults
Transmissibilities derived for interaction region associated with corner
arising when lateral grid lines cross
Additional physical constraints discussed in the derivation of transmissibilities
Zig-zag vs. planar faults in 3D investigated numerically

54

Summary of the papers

7.3 Summary of paper D


This subsection gives a summary of Paper D:
A note on the convergence of the MPFA O-method; numerical experiments
on some 2D and 3D grids
Authors: G.T. Eigestad, R. Klausen.
Paper D investigates convergence of the MPFA O-method on general quadrilaterals in 2D and some special 3D grids.
The question of whether the O-method converges was raised by referees in the
reviewing process of Paper A. We tried to answer this question by some numerical
experiments in this paper. In parallel, Klausen [33] has analysed this through
relationships found between Mixed Finite Element methods and MPFA methods.
The results show 1st order convergence for problems with smooth coefficients
K. Numerical results for MPFA methods have been published by Edwards [18]
and Jeannin et al. [38]. Edwards method, however, uses different dividing points
than what is used for the MPFA O-method (see Sec. 3.1). In addition to this, the
numerical examples in the above references do not have the same complexities
as for instance the examples found in Riviere et al [37]. We therefore perform a
broader spectrum of tests for the MPFA O-method in Paper D. These problems
ranged from the simple case where the flow is (piecewise) uniform, to cases
where the derivatives of the analytical pressure solutions may become infinitely
large. These cases are likely to occur in real reservoir simulation because sharp
heterogeneity contrasts (with corners) must be handled. The convergence is
observed for the examples that have been tested, and the results are summarised
below.
Main results of the paper:
Convergence of the O-method on quadr. 2D grids investigated numerically
Flux operator investigated analytically on quadrilateral 2D and 3D grids,
and found to be exact for linear problems.
2nd order convergence observed for the H 2 -regular problems we tested.
h2 -order convergence of pressure indicated for problems tested with analytical solutions belonging to H 1+ () for > 0.39. Lower convergence
for case with = 0.127.
Flux is roughly h -order convergent for all examples except for case with
= 0.127.

7.4 Summary of paper E

7.4

55

Summary of paper E

The concluding paper has the title:


Numerical Modeling of Capillary Transition Zones.
Authors: G.T. Eigestad, J.A. Larsen.
In this paper, a consistent hysteresis model was included in a reservoir
simulator for forward simulation of two phase flow. Hysteresis does apply for
both capillary pressure and for relative permeabilities. The hysteresis models
are taken from work done by researchers at Stavanger College [53], [54], and
they suggested a hysteresis model to describe capillary pressure (P c) and relative
permeability (kr ) for reservoirs with varying degree of wettability. The entities
P c and kr will be described by general curves/functions dependent on water
saturations, Sw , as well the history of reservoir locally (the history of each
individual grid cell in a discrete model).
Regardless of the generality of the curves, they will usually be described by
only a few parameters. For instance, the capillary pressure for a primary drainage
curve will be described by the expression
Pcd (S) =

cwd
(SSwr ) awd
( (1Swr ) )

(7.1)

The parameters used here are: cwd is the threshold pressure, S is the water
saturation, Swr is the irreducible water saturation. The parameter awd is the inverse of the pore-size distribution index, and is a parameter related to the specific
sandstone. Further details are given in the paper and its references. The expression (7.1) is the foundation for the generation of general capillary pressure curves
for media with varying wettability. The wettability of a sandstone says something
about the preferred glueing/clinging of water to the sandstone compared to the
same effect for oil. Sandstones that are classified as mixed-wet, will contain both
pores that are water-wet and oil-wet. From the discussions in [53] and [56], this
implies capillary pressure curves with both positive and negative capillary pressure. For details we refer to [53] and [56] and references herein.
The capillary pressure curves used in Paper D will be combinations of two
terms of the form (7.1), where one of the terms will be the capillary pressure
description in a completely water-wet system, and the second term will be the
capillary pressure description in a completely oil-wet system. The bounding
secondary imbibition and drainage curves will be on the form
Pc (S) =

cw
wr ) aw
( (SS
(1Swr ) )

co
or ) ao
( (1SS
(1Sor ) )

(7.2)

56

Summary of the papers

The as and cs are constants, and there is one set each for imbibition curves
and drainage curves respectively. Irreducible water saturation is denoted Sor . The
as are all positive numbers, cw is a positive number, whereas co is a negative
number. The curves will cross Pc = 0 for some saturation value between the irreducible water saturation and the irreducible oil saturation. Furthermore, the
capillary pressure curves will have asymptotes at Swr and Sor .
A hysteresis loop logic is introduced in [53], and this allows for arbitrary
imbibition and drainage capillary pressure curves to be defined. Furthermore,
similar logics will apply for relative permeability, so that our numerical model
will include a complete description of both capillary pressure and relative
permeability. In paper D, all examples initialise the reservoirs by primary
drainage curves given by Eq. (7.1). The initial saturation distribution in the
reservoir is hence a function of height, and the gradual change in saturation
defines the vertical transition zone between oil and water. Water flooding is
modelled by using a water injector at the bottom of the reservoir, and an oil
producer at the bottom of the reservoir. After a specific amount of fluids has been
injected/produced, the wells will be shut off, and the redistribution of the fluids is
modelled. Different rate regimes are investigated, and the redistribution is highly
dependent on the state at the time of well shut off. A comparison was also made
with a hysteresis option available in the commercial simulator Eclipse. The main
results of the paper are summarised below.
Main results of the paper:
Inclusion of consistent capillary pressure and relative permeability models
in forward reservoir simulator
Numerical handling of solution dependent curves with non-smooth derivatives in implicit solver
Investigation of examples where processes change directions throughout
simulations
Comparison of new hysteresis model and hysteresis model available in commercial simulator.

Chapter 8
Further work
The work of Paper A dealt with symmetry and M-matrix issues for the MPFA
O-method on a polygonal grid. This only partially answers the question about the
monotonicity of the method. As mentioned in Chapter 6, analysis has recently
been done for the O-method on 2D structured grids in regards to monotonicity
issues. These ideas can possibly be extended to expand the analysis of the Omethod on polygonal grids. As a starting point one should consider polygonal
grids with set patterns, so that there is a certain structure of the elements of the
coefficient matrix. A natural case to analyse is layered media.
The work with faults in 3D is by no means finished. The paper which deals
with faults for quadrilateral grids with crossing grid lines (Paper C) was initially
started because of specific field needs. However, we had certain problems with
the first version of the code that accounts for these types of faults. At the time of
writing this thesis, we have not been able to successfully simulate flow for real
field cases. This is one of the issues that we will work on at a later stage. We
then hope to be able to do field studies of a field with corresponding reservoir/grid
model depicted in Figure 8.1.
In addition to this, there are unresolved issues with the transmissibility calculations. Because of the special character the local system of continuity equations
to determine transmissibilities has, additional principles had to be used to determine the transmissibilities. As the discussion in the paper reveals, one degree of
freedom was left undetermined even after the principle of exact fluxes for linear
flow was used. The underlying local system is still subject to more analysis, and
more information may be used to bind the last degree of freedom.
The work of Paper D dealt with numerical convergence results. So far, only
matching quadrilateral grids have been tested, and most of the examples deal with
2D examples. Since the MPFA methods have been developed for general faults,
this may certainly be used to perform tests with local grid refinement. Results
should then be compared with those of Discontinuous Galerkin Methods and En-

58

Further work

FloViz 2001A_1

Cells

473

945

1418

1890

2362

2834

3306

3779

4251

4723

5195

5667

6140

6612

7084

Figure 8.1: Extracted part of real field case with exaggerated faults.

hanced Mixed Finite Element Methods.


Streamline issues are currently being considered for the MPFA O-method on
irregular grids [16]. Uniform flow results for the Control Volume Mixed Finite
Element Method have been presented in [31]. This method gives a continuous
representation of velocities within grid cells (interpolation). Because of complex
geometry, uniform flow will not be simulated correctly for all shape function [31],
but the problem is overcome with different choices of shape functions. For the
MPFA O-method, interpolation issues associated with bilinear cell surfaces have
not been considered yet. This is a field where we hope to do more analysis and
numerical tests in the future.

Bibliography
[1] I. Aavatsmark, T. Barkve, . Be, T. Mannseth: Discretization on nonorthogonal, curvilinear grids for multi-phase flow, Proc. 4th European Conference on the Mathematics of Oil Recovery, vol. D, Rros, Norway (1994),
17pp.
[2] I.Aavatsmark, T.Barkve, T.Mannseth, . Be: Discretization on nonorthogonal, quadrilateral grids for inhomogeneous, anisotropic media, J.
Comput. Phys. 127 (1996), 2-14.
[3] I.Aavatsmark, T.Barkve, .Be and T.Mannseth: Discretization on Unstructured Grids for Inhomogeneous, Anisotropic Media. Part I: Derivation of the
methods, SIAM J. Sci. Comput.19 (1998), pp. 1700-1716.
[4] I.Aavatsmark, T.Barkve, .Be and T.Mannseth: Discretization on Unstructured Grids for Inhomogeneous, Anisotropic Media. Part II:Discussion and
numerical results, SIAM J. Sci. Comput.19 (1998), pp. 1717-1736.
[5] I.Aavatsmark, T.Barkve, .Be and T.Mannseth: A Class of Discretization
Methods for Structured and Unstructured Grids in Anisotropic, Inhomogeneous Media, Proc. 5th European Conference on the Mathematics of Oil Recovery, Leoben, Austria, 1996.
[6] I. Aavatsmark, E. Reiso and R. Teigland: Control-volume discretization
method for quadrilateral grids with faults and local refinement, Computational Geosciences 5, 2001, pp. 1-23.
[7] I. Aavatsmark, E. Reiso, H. Reme, R. Teigland: MPFA for Faults and Local
Refinement in 3D Quadrilateral Grids With Application to Field Simulations,
SPE 66356, Proceedings of SPE Reservoir Simulation Symposium, Houston, Texas, 2001.
[8] I. Aavatsmark: An introduction to Multipoint Flux Approximations for
Quadrilateral Grids, Computational Geosciences 6, 2002, pp. 405-432.

60

Bibliography

[9] G. T. Eigestad, I. Aavatsmark, M. Espedal: Symmetry and M-matrix issues


for the O-method on an Unstructured Grid, Computational Geosciences 6,
2002, pp. 381-404.
[10] G.T. Eigestad, I. Aavatsmark, E. Reiso, H. Reme, R. Teigland: MPFA methods applied to irregular grids and faults, Developments in water science 47;
Editors Hassanizadeh, Schotting, Gray, Pinder.
[11] G.T. Eigestad, I. Aavatsmark, E. Reiso, H. Reme: MPFA for faults with
crossing grid lines and zig-zag patterns ECMOR VIII, September 3-6.,
2002.
[12] G.T. Eigestad, R. Klausen: A note on the convergence of the MPFA Omethod; numerical experiments on some 2D and 3D grids Draft manuscript.
[13] G.T. Eigestad, J. A. Larsen: Numerical modeling of capillary transition
zones, SPE 64376 APOGCE 2000, 16.-18. October 2002.
[14] G. T. Eigestad: On the Symmetry of a Discrete Operator; Discretizing an
Elliptic term by the O-method, Cand. Scient Thesis, Department of Applied
Mathematics, University of Bergen, Norway.
[15] J. M. Norbotten, I. Aavatsmark: Monotonicity Conditions for Control
Volume Methods on Uniform Parallellogram Grids in Homogeneous Media.
Submitted.
[16] H. Hgland: Streamline Tracing on Irregular Grids, Ongoing Master Thesis
at Department of Mathematics, University of Bergen.
[17] M.G. Edwards and C.F. Rogers: A flux continuous scheme for the full
tensor pressure equation, Proceedings of the 4th European Conference on
the Mathematics of Oil Recovery, Norway, June 1994.
[18] M.G. Edwards and C.F. Rogers: Finite volume discretization with imposed flux continuity for the general tensor pressure equation Computational
Geosciences 2, 1999.
[19] A.F. Ware, K. Parrott, C. Rogers: A Finite Volume Discretization for Porous
Media Flows Governed by Non-Diagonal Permeability Tensors Proceedings
of CFD95, Banff, Canada, 1995.
[20] S. Verma, K. Aziz: A Control Volume Scheme for Flexible Grids in Reservoir Simulation, SPE 37999, 14th SPE Reservoir Symposium, Dallas, Texas,
June 1997

Bibliography

61

[21] Z.E. Heinemann, C.W. Brand, M. Munka, Y.M. Chen: Modelling Reservoir
Geometry with Irregular Grids, SPE Reservoir Engineering, May 1991, 225232.
[22] Z.E. Heineman, C.W. Brand Gridding techniques in Reservoir Simulation
First Intl. Forum on Reservoir Simulation, Alpbach, 1988, pp 339-425.
[23] A.F. Ware, A.K. Parrott, C. Rogers: A Finite Volume Discretisation for Porous Media Flows Governed by Non-Diagonal Permeability Tensors, Proceedings of CFD, Banff, Canada, 1995
[24] L.S.K. Fung, A.D. Hiebert, L.X. Nghiem: Reservoir Simulation With a
Control-Volume Finite-Element Method, SPE Reservoir Engineering 7 1992,
349-357
[25] P.A. Forsyth: A Control-Volume, Finite-Element Method for Local Mesh Refinement in Thermal Reservoir Simulation, SPE Reservoir Engineering, SPE
18415, Nov 1990, pp. 561-566
[26] P.A. Forsyth, P.A. Sammon: Quadratic Convergence for Cell Centered
Grids, Applied Numerical Mathematics 4 (1988), pp- 377-394.
[27] P.A. Raviart and J.M. Thomas: A mixed finite element method for 2nd order
elliptic problems, Mathematical Aspects of Finite Element Methods, Eds. I.
Galligani and E. Magenes, Springer Verlag, 1977, pp.292-315.
[28] J. Shen: Mixed Finite Element Methods on Distorted Rectangular Grids,
Technical report, Institute for Scientific Computation, Texas A/M University,
1994.
http://citeseer.nj.nec.com/shen94mixed.html
[29] Z. Cai, J.E. Jones, S.F. McCormick and T.F. Russell: Control-volume Mixed
Finite Element Methods, Computational Geosciences 1 1997, 289-315
[30] Thomas F. Russell: Relationships Among Some Conservative Discretization
Methods, Lecture Notes in Physics, Chen, Ewing and Shi (eds.), 1999, pp
1-16.
[31] R. L. Naff, T. F. Russell, J. D. Wilson: Shape functions for three-dimensional
control-volume mixed finite-element methods on irregular grids, Developments in Water Science 47, pp. 359-366.
[32] R. A. Klausen and T. Russell. Relationships among some conservative discretization method that handle discontinuous coefficients Draft Manuscript.

62

Bibliography

[33] R. A. Klausen. Convergence of Multi Point Flux Approximation on Quadrilateral Grids Draft Manuscript.
[34] A. Weiser, M.F. Wheeler On convergence of block-centered finite differences
for elliptic problems, SIAM J. Numer. Anal. 25, pp 351-375, 1988.
[35] J. A. Wheeler, M. F. Wheeler, I. Yotov: Enhanced velocity mixed finite element methods for flow in multiblock domains, TICAM report 01-27, Sept.
2001.
[36] I. Yotov: Mixed Finite Element Methods for Flow in Porous Media PhD
thesis, Rice University (1996).
[37] B. Riviere, M. F. Wheeler, K. Banas: Discontinuous Galerkin method applied
to a single phase flow in porous media, Computational Geosciences 4, pp
337-349, 2000.
[38] L. Jeannin, I. Faille and T.Gallouet How to model Compressible Two-Phase
Flows on Hybrid Grids?
Oil and Gas Science and Technology - Rev. IFP, Vol. 55 (2000), No. 3.
[39] Aziz, Settari: Petroleum Reservoir Simulation, Elsevier Applied Science
Publishers, London 1979
[40] H. Kleppe: Reservoir Simulation, Lecture compendium, Stavanger College,
Norway, Dec. 1995.
[41] A.A Samarskij: Theorie der Differenzenverfahren, Akademische Verlagsgesellschaft Geest und Portig, Leipzig 1984
[42] R. Sanders: On convergence of monotone finite difference schemes with
variable spatial differencing, Math. Comp. 40, pp 91-106, 1983.
[43] M. Shashkov: Conservative Finite-Difference Methods on General Grids,
CRC Press, 1996
[44] M. Shashkov and S. Steinberg: Solving Diffusion Equations with Rough
Coefficients in Rough Grids,
J. Computational Physics 129, pp. 383-405 (1996).
[45] J. Hyman, M. Shashkov and S. Steinberg: The Numerical Solution of Diffusion Problems in Strongly Heterogeneous Non-isotropic Materials, J. Computational Physics 132, pp. 130-148 (1997).

Bibliography

63

[46] G. Strang and G. Fix: An Analysis of the Finite Element Method Wiley, New
York, 1973.
[47] S. C. Brenner, L. Ridgway Scott: The Mathematical Theory of Finite Element Methods, Springer
[48] A. J. Chorin, J. E. Marsden: A mathematical Introduction to Fluid Mechanics Springer Verlag, New York, 1979.
[49] L. N. Trefethen D. Bau: Numerical Linear Algebra, SIAM press
[50] D. Braess: Finite Elements, Cambridge University Press
[51] W. Hackbusch: Theorie und Numerik elliptischer Differentialgleichungen,
1986 Teubner Studienbucher
[52] T. F. Russell and M. F. Wheeler: Finite Element and Finite Difference Methods for Continuous Flows in Porous Media In The mathematics of reservoir
simulation, Editor: Ewing, R. SIAM
[53] S. Skjaeveland, L. M. Siqveland, A. Kjosavik, W. L. Hammersvold, G. A.
Virnovsky: Capillary Pressure for Mixed-Wet Reservoirs SPE 39497, Proceedings of 1998 SPE India Oil and Gas Conference.
[54] A. Kjosavik: Integrated modelling of Relative Permeability and Capillary
Pressure, M.Sc Thesis, Stavanger College, Norway (1999).
[55] R. H. Brooks, A. T. Corey: Hydraulic Properties of Porous Media, Hydraulic
Paper no. 3, 1964, Colorado State U.
[56] J. A. Larsen: Vannmetningsmodeller og fasemobiliteter for reservoarer med
residuell oljesone; systemer med blandet fukting, Report, Norsk Hydro Research Center, (1999).
[57] Eclipse 100, Reference Manual 2001, Sclumberger Geoquest (2001)

Part II
Published and submitted work

Paper A
Symmetry and M-matrix issues for
the O-method on an Unstructured
Grid
A person who is looking for something is not travelling very fast
In Stuart Little by E. B. White

Published in Computational Geosciences, vol. 6, 2002, pp. 381404.

Paper B
MPFA applied to Irregular Grids and
Faults
One day I shall reach the summit, and see how small the other mountains are
Du Fu, Chinese Philosopher

Published in Developments in Water Science 47, Elsevier, 2002, pp. 413-420.

Paper C
MPFA for Faults with Crossing
Layers and Zig-zag Patterns
I was programmed for etiquette, not destruction
3CPO, in Star Wars Episode II

Paper D
A note on convergence of the MPFA
O-method; numerical experiments
on some 2D and 3D grids
An idea that remains an idea is not a good idea
Unknown

Paper E
Numerical Modeling of Capillary
Transition Zones
Wer reitet so spat durch Nacht und Wind? Es ist der Vater mit seinem
Kind
Goethe

You might also like