Sie sind auf Seite 1von 72

Homogenization Technique Applied to

a Smoldering Combustion Model


Ekeoma Rowland Ijioma
July 2010

Homogenization Technique
Applied to a Smoldering
Combustion Model

by
Ijioma, Ekeoma Rowland

Supervisors:
Dr. Adrian Muntean
Dr. Martijn Anthonissen

July 28, 2010

Contents
1 Introduction
1.1 Smoldering combustion . . . . . . . .
1.2 Choice of microstructure . . . . . . .
1.3 Mechanism of smoldering combustion
1.4 The objective of this thesis . . . . . .
1.5 Outline of this thesis . . . . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

2 Mathematical model of smoldering combustion


2.0.1 Combustion reaction . . . . . . . . . . . .
2.0.2 Conservation of energy in the gas and solid
2.0.3 Conservation of mass in gas . . . . . . . .
2.0.4 Solid product formation . . . . . . . . . .
2.0.5 Boundary and initial conditions . . . . . .
2.1 Reduction to a two-dimensional formulation . . .
2.1.1 Non-dimensionalization . . . . . . . . . . .
3 The
3.1
3.2
3.3
3.4

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

. . . .
phases
. . . .
. . . .
. . . .
. . . .
. . . .

homogenization method
General averaging strategy . . . . . . . . . . . . . .
The microscopic problem . . . . . . . . . . . . . . .
Homogenization of the microscopic problem . . . .
The macroscopic equations and effective coefficients

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.

1
1
2
4
6
6

.
.
.
.
.
.
.

7
7
7
8
8
9
9
14

.
.
.
.

18
18
18
21
29

4 Numerical multiscale homogenization approach


4.1 Computation strategy for the homogenized problem . . . . . . . . .
4.2 Numerical computation of the homogenized coefficients . . . . . . .
4.2.1 Weak formulation of the cell problem . . . . . . . . . . . . .
4.2.2 Finite element approximation to the solutions of the cell problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.3 Existence and uniqueness of weak solutions to the cell problems
4.3 Further properties of the effective coefficients . . . . . . . . . . . . .
4.3.1 Symmetry and positive definiteness of the effective tensors .
4.3.2 A comparison of computed coefficients with various bounds .
4.4 Numerical computation of the macroscopic and microscopic solutions
4.5 Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . .
4.5.1 Smoldering process in a microstructure . . . . . . . . . . . .
4.5.2 One-dimensional comparison of results . . . . . . . . . . . .

ii

31
31
33
33
34
37
40
40
41
42
45
45
49

4.5.3

Error estimates in the FE approximations in the numerical


simulations . . . . . . . . . . . . . . . . . . . . . . . . . . .

50

5 Conclusion and Future work

55

A Matlab code for the discrete L2 error estimate

64

iii

Abstract
We study a semi-linear reaction-diffusion (RD) system modeling reverse smoldering
combustion of a thin cellulose material (e.g. paper) under the influence of oxidizing
wind.
The basic geometry of our material consists of a collection of many small thin
solid cylinders periodically distributed in space that we call microstructures. The
main working assumption is that the microstructure depends on two independent
parameters: the height of the cylinder (oriented in Oz direction) and the
width of a square cell that is copied periodically to cover the x y plane. After a
volume averaging of the RD system along the z coordinate, we obtain the so-called
micro problem.
The main objective is to use periodic homogenization methods to study the asymptotic behavior of the solutions to the micro problem as the parameter goes to zero.
This method gives an upscaled RD system (that we refer to as macro problem) together with explicit formulae for the effective coefficients.
We use COMSOL Multiphysics to solve numerically the micro problem for decreasing values of and compare the obtained results with the solution of the macro
problem ( = 0). We see that our model captures the expected behavior of concentrations and temperature profiles as observed in smoldering combustion experiments.
Keywords: Homogenization, reaction-diffusion system, smoldering combustion, multiscale numerical method MSC 2000: 35B27; 76M50; 35K57; 80A25; 65N99

iv

Acknowledgements
Firstly, I acknowledge the hand of God throughout the period of this thesis, for
keeping me safe and in good health. I will like to express my sincere gratitude to
my supervisors Dr. Adrian Muntean and Dr. Martijn Anthonissen for giving me
this project, and especially to Dr. Adrian Muntean, for his guidance, encouragement and patience during the period of this thesis.
I am also grateful to all my lecturers, both at the University of Technology, Kaiserslautern and the University of Technology, Eindhoven. Your lectures and useful
discussions have inspired me over the years and have contributed to the successful
completion of this work.
Finally, to all my friends and well wishers, especially to those that followed the
Erasmus Mundus Masters program in Industrial and Applied Mathematics, I say a
huge thank you for your encouragements and all the interesting moments we worked
together. More grease to your elbows in your subsequent careers.

Ijioma, Ekeoma Rowland


Eindhoven, July 2010.

Chapter 1
Introduction
1.1

Smoldering combustion

Smoldering is a flameless form of combustion that derives its heat from heterogeneous reactions occurring on the surface of a solid fuel when heated in the presence
of oxygen [1]. It is of interest for the fundamental understanding of combustion
processes. It can also be studied as a practical fire hazard, since the smoldering can
transit to flaming and is usually accompanied by the emission of toxic gases.
In porous media, smoldering is a very complex situation that happens in situations where a reactive part of the porous material is oxidized using heat content
without flaming. Common examples of smoldering combustion in porous media
include the initiation of upholstered furniture fires by weak heat sources and the
flaming combustion of biomass occurring in wild land fires behind the flame front.
Many porous materials that can sustain a smoldering reaction namely coal, cotton,
tobacco, paper, peat, wood and most charring polymers. We look to smoldering
combustion from the perspective of reaction-diffusion-flow phenomena in porous
media.
Smoldering initiation requires the supply of heat flux to the solid. The subsequent temperature increase of the solid triggers its thermal-degradation reactions
(endothermic pyrolysis and exothermic oxidation) until the net heat released is high
enough to balance the heat required for propagation. This net heat released by the
reactions is partially transferred by conduction, convection and radiation ahead of
the reaction and partially lost to the surrounding environment.
Oxygen is transported to the reaction zone by diffusion and convection, and in
turn it feeds the oxidation reactions. Once ignition occurs, the smolder front propagates through the material in a creeping fashion [1].
In this thesis, we present a model of the smoldering combustion of a thin cellulose
material exposed to a flow of oxygen confined in a narrow gap above the material.
The experimental setup (see Fig.1.1) under consideration is one having a particu1

larly simple geometry (see. [2] for more details) in which the sheet is ignited on one
side rising the local temperature by means of a heat pulse. We assume the flow of
oxygen to be parallel to the cellulose material, and in the opposite direction to the
combustion front.

Figure 1.1: Sketch of the experimental setup [2].


We consider the situation where the combustion front proceeds from the ignition
boundary x = L,(see Fig. 1.1) to the opposite boundary at x = 0. The top and
base of the experimental setup is thermally insulated to prevent heat losses to the
surrounding. This means that the transport mechanism is fully controlled within
the closed system, where convective/conductive heat transfer takes place between
the reacting species.

1.2

Choice of microstructure

The material under consideration in this study is a paper sample like that used
in [3]. We consider the paper material to be a porous medium with periodic arrangement of pores distributed around cylindrical structures as shown in Fig. 1.3.
The microstructure we bear in mind is depicted in Fig. 1.2 (left). The paper has
thickness of size . We choose a reference unit cell, Y = (0, 1)n , n {2, 3} as in

Figure 1.2: Left: Volumetric representative cell Y ; Right: Planar representative


unit cell Y .
Fig. 1.2 (the microstructure). We generate a lattice of copies of cells Y that spans
2

Figure 1.3: 3D micro scale geometry of the paper sheet.


the entire paper occupying the region := (0, ), with := (0, L)2 .
Within the unit cell Y , we define a geometrical structure Ys , as the solid part,
i.e. a closed subset of Y and Yg := Y \ Ys , the gas part such that Y = Ys Yg .
Also, we assume that any two neighboring Ys do not touch each other and the two
parts of the unit cell satisfy Ys Yg = . Next, we make a periodic repetition of
Ys all over Rn and set Ysk, := Ys + k, k Zn .
Clearly, the obtained set
Es :=

Ysk,

(1.1)

kZn

is a closed subset of Rn and Eg := Rn \ Es is an open set in Rn .


Moreover, we assume that the paper consists of connected gas parts. We define
by Es Rn , the closed set that is obtained by Y -periodic repetition of Ys in the
entire space Rn and we denote its gas counterpart by Eg . The following hypotheses
have to be satisfied [4]:
i. Yg and Ys have strictly positive measures in Y with Ys Y = .
ii. Eg and the interior of Es are open sets with C 0,1 boundaries. Furthermore Eg
is connected.
iii. Yg is an open set with a local Lipschitz boundary.
The implications of these hypotheses for a paper sheet are as follows. One elementary cell Y of a paper consists of both solid and gas parts as depicted in Fig. 1.2.
Therefore, assumption (i) is fulfilled.
Furthermore, we assume that a paper sheet consist of periodic repeated cells covering . Thus, the solid part Ys , as well as the gas part Yg , of the repeated cell
must be Y -periodic. i.e. Eg , must be connected.

Both Eg and Es have a C 0,1 boundary, and assumption (ii) is fulfilled. Furthermore, we assume that the boundary of Yg is sufficiently regular so that (iii) holds.
Now, we generate a lattice of copies of cells Y , for any > 0, by letting the
domain, Rn , be covered by regular mesh of size as in Fig. 1.3. We denote
each cell by Yi = (0, )n , where n = 2 or 3, with 1 i N (), and N () = ||n
denotes the number of cells along the x- and y- directions, while keeping (0, ) fixed,
i.e we have in general
N (, ) :=

| |
.
n1

(1.2)

Each cell is homeomorphic to Y , by a linear homeomorphism i , with ratio of


magnification 1/, i.e. we re-scale the cell Y by . Hence,
Ys,
:= (i )1 (Ys )
i

(1.3)

(i )1 (Yg )

(1.4)

Yg,
i

:=

denote respectively the solid and the gas part of the unit cell,Y , , of the order .
The gas domain, g , is obtained by removing the periodically distributed solid
parts. That is,
N ()

:= \

= Eg
Ys,
i

(1.5)

i=1

with
N ()

Ys,
.
i

(1.6)

i=1

Also, we define g = s to be the boundary of the gas part of the porous


medium, where s is the solid-gas interface. Subsequently, we will, for simplicity
of notation, denote s as .

1.3

Mechanism of smoldering combustion

Combustion has been studied extensively over the years both experimentally and
numerically. However, few relevant theoretical studies based on smoldering models
have been done. Previous studies in this area include the work of Ohlemiller [5],
who presented a review of the most significant mechanisms involved in the smoldering combustion of polymers. Rein [1] presented a 1-D computational study to
investigate smoldering ignition and propagation in polyurethane foam.
In [6], traveling wave solutions were sought by considering cases of sufficiently large
flow and moderate flow of the air flux. They found that for each Peclet number (Pe)
beyond some lower threshold, there exist two solutions: a fast wave and a slow wave,
4

where both have peculiar qualitative differences.The slow waves are unstable while
the fast waves are stable for large values of Pe. K. Ikeda and M. Mimura (2008) (see
[7]) proposed a model of reaction-diffusion system for theoretical understanding of
pattern formation in smoldering combustion, and then proved the existence and
uniqueness of strong solution to the system using the standard theory of analytic
semigroups. Their numerical simulation exhibits a good qualitative agreement with
the experimental results of Zik and Moses (see Fig. 1.4).
In their fundamental paper [3], Zik and Moses studied the smoldering combustion of a thin cellulose material experimentally and developed a phenomenological
model for the analysis of pattern formation and fingering instability. The primary
use of their model has been to analyze the global behavior of smoldering combustion
using the flow velocity (see also Peclet number) as a control parameter.
Other control parameters considered in [3] include the vertical gap between the
plates and the heat conductivity of the bottom plate of the experimental setup.
The lateral boundaries of the experimental setup are made to create a uniform heat
conduction and to decrease the supply of oxygen from the sides of the sample. This
ensures uniform propagation along the boundaries.

Figure 1.4: Instability of combustion front when the flow velocity (or Peclet number)
decreases from a to e. Oxygen flows downward and the smolder front moves upward
as shown by the arrow in d [6]
The major mechanism of their experiments which is most relevant to our study is
the role of oxygen flow velocity (alternatively the Peclet number) in the observed
pattern formation of the smolder front. The results of their experiment show different behaviors of the smolder front as shown in Fig. 1.4.
In Fig.1.4, the front is initially smooth due to the high flux of oxygen (a). Next,
a sinusoidal pattern develops after the supply of oxygen is limited (b-c). Further
reducing the oxygen flux separates the peaks and a fingering pattern is formed. The
same behavior is also possible with a nondimensionalized system. In this case, the
non-dimensional control parameter is the Peclet number Pe, which measures the
relative importance of advection and molecular diffusion.

1.4

The objective of this thesis

The primary objective of this thesis is to investigate the smoldering combustion of a


thin solid fuel (paper)-a porous solid with periodic microstructure. The theoretical
background of our proposed model is partly based on a previous work by L. Kagan
and G. Sivashinsky [2], where fragmentation1 of the flame front was excluded. We
expect that our proposed upscaled model will capture some of the qualitative features of smoldering combustion as given in [3].
The geometry of the periodic microstructure depends on a small parameter (see
section 1.2). This means that any sought solution of the model depends on . It
is difficult to obtain an exact solution of this problem. Thus, only an approximate
solution depending on is available. In a formal asymptotic analysis we show that
as approaches zero, the approximate microscopic solution converges to the macroscopic solution ( = 0). This procedure becomes computationally expensive with
decreasing values of .
Further, we implement a method that drastically reduces the computational effort involved in solving the microscopic model (see chapter 3 (The homogenization
method)). We derive an averaged (homogenized) model. Although the homogenized model is only an approximation, we show in this study that it is akin to the
physical behavior of the microscopic model without loosing too much accuracy. We
investigate how far the solution of the homogenized model is from the approximate
solution of the microscopic model.

1.5

Outline of this thesis

The outline of this thesis is as follows: Chapter 2 introduces the mathematical


modeling of smoldering combustion, from which the three dimensional microscopic
model was reduced to a two dimensional formulation. We identified a set of governing parameters by using dimensional analysis. The influence of these parameters
is illustrated by direct and detailed numerical simulations on the micro and macro
problems. In Chapter 3, we introduce the homogenization method and approximate
the 2 dimensional microscopic model presented in Chapter 2 by a set of macroscopic
equations with effective coefficients. Chapter 4 begins with discussion on the the
multiscale numerical homogenization approach. We provide solutions to the cell
problems encountered in Chapter 3 as well as a strategy to compute effective coefficients. Finally, we solve the microscopic and macroscopic problems using COMSOL
Multiphysics and give an error estimate in terms of the discrete L2 norm. In chapter
5, we discuss the results of the previous chapters with concluding remarks. Also,
we present some useful insights into some issues that were not covered in the course
of this study as future work.

By fragmentation, we mean the splitting of the smolder front as shown in Fig. 1.4 (center)

Chapter 2
Mathematical model of
smoldering combustion
In order to remain consistent with a possible real life scenario, we adopt the materials and operating parameters, corresponding to the laboratory experiment described
in [3], as a reference case. In this section, we describe the transport and reaction
mechanisms of smoldering combustion of a thin sheet of paper using a 3-dimensional
system of partial differential equations.
Following the ideas in [8], we consider the conservation of various physical quantities, governed by flow equations and combustion equations. The flow equations
express conservation of mass and momentum. The combustion equation express
conservation of the species mass fractions and energy. However, in this present
study, we consider a minimal formulation where we do not consider the dynamics
of the problem due to the gas pressure (see. [6], [2]).

2.0.1

Combustion reaction

The species involved in the chemical model includes C6 H10 O5 , O2 , CO2 , H2 O and
the inert gases such as N2 (cf.[6], [8]). However, we restrict ourselves to the species
entering the reaction. Basically, the combustion reaction takes place between cellulose and oxygen [6]:
C6 H10 O5 + 6O2 6CO2 + 5H2 O,

(2.1)

where the constant coefficients in the reaction species in (2.1) are the stoichiometric
coefficients for the reaction. The reaction takes place at the solid-gas interface. We
can obtain the rate of reaction of the solid at this interface by the application of
the law of mass action, which states that the rate of a reaction is proportional to
the product of the concentrations of the reactants.

2.0.2

Conservation of energy in the gas and solid phases

The balance in energy in the two phases can be described by two heat transport
equations in the gas part, Yg and in the solid part Ys . The transport equation in
7

Figure 2.1: Reaction/transport mechanisms within the microstructure


the gas part is of convection-diffusion type. The oxygen is first preheated and is
passed from the boundary, x = 0.
Oxygen flows through the porous medium and reacts at the reaction zones presented at the solid-gas interface. Heat is carried by the oxygen across the pore
sites of the microstructure, and also the oxygen is being diffused in the process (see
Fig. 2.1). The heat transfer in the solid part of the microstructure involves only
conduction. The governing equations are thus:
g cg

Tg
+ (g cg u Tg g Tg ) = 0 in Yg ,
t

(2.2)

Ts
(s Ts ) = 0 in Ys ,
t

(2.3)

s cs

where T is the temperature (K), , the thermal conductivity (J/(msK)), , the


mass density (kg/m3 ), c the specific heat capacity (J/(kgK)), u = (u, 0, 0) the flow
field (m/s). The subscripts g and s denotes gas and solid parts, respectively.

2.0.3

Conservation of mass in gas

The oxygen concentration, C (dimension kg/m3 ), is governed by a convectiondiffusion equation.


C
+ (uC DC) = 0,
t

in Yg

(2.4)

where D is the diffusion coefficient (dimension m2 /s) following Ficks law that
describes the flow of oxygen from the region above the paper to the region within
the porous medium as in Fig. 2.1.

2.0.4

Solid product formation

The transition from paper to char involves a decay due to the chemical reaction
(2.1). This follows a mass action type law defining W and is given by
R
:= W (T, C)
t

at Ys .

(2.5)

R is the concentration of solid product (kg/m3 ) and the reaction rate W (T, C),
(kg/(m3 s)) is defined by


Ta
.
(2.6)
W (T, C) := AC exp
T
(2.6) describes an Arrhenius type kinetics, with Ta being the activation temperature
(K). The boundary prescribed in (2.5) is at the gas-solid interface in the volumetric
unit cell Y .

2.0.5

Boundary and initial conditions

The transport equations described in (2.2) and (2.3) are coupled by the continuity
of the temperature across the boundary Ys , and also by the source and sink terms
included in the surface reaction.
The reaction induces source and sink terms for the oxygen and the heat, that are
all proportional to the reaction rate. The sources and the sinks, localized on the
surface Ys of the reactive solid part, lead to the following boundary conditions at
the gas-solid interface:

(D C)n = W (T, C)

n (g Tg s Ts ) = QW (T, C)
at Ys
(2.7)

Tg = Ts = T
where T is the interfacial temperature, n is the outward unit normal vector to the
interface, and Q the heat release ( J/kg)
Due to the diffusive-thermal insulation of the system, the boundary conditions at
the walls are:
n C = n Tg = n Ts = 0,

at {z = } {z = 0}.

(2.8)

The system is completed by upstream boundary conditions,i.e., by the gas flow


rate, the entry gas temperature and oxygen concentration. Radiative effects are
not taken into account in this formulation.
The initial conditions are as follows:

Tg = Ts = T0
C = C0

R = R0

2.1

in .

(2.9)

Reduction to a two-dimensional formulation

In order to successfully study the surface smoldering of the material, we reduce1


the dimension of the model equations from 3-dimensions to 2-dimensions. To be
1

This reduction also has the effect of suppressing any effects due to gravity. This brings us in
line with previous studies on smoldering combustion under micro-gravity.

more precise, we do this by averaging along the z-axis of the domain.


Integrating (2.2) over z (0, ), we have
Z 2
Z
Z
Z
Z
Tg

Tg

Tg dz + g
dz = g 4
g cg
Tg dz + g cg u
Tg dz + g cg 0
dz,
2
t 0
0
0 z
0
0 z
(2.10)
where

u 7 (u, 0, 0)


, ,
x y z

with
=
u = (u, 0) and

+0 ,
= u
z


,
x y

(2.11)


.

Similarly, we define
4 7

2
2
2
2
+ ,
+
+
=
4
x2 y 2 z 2
z 2

with
=
4

2 2
,
x2 y 2

(2.12)


.

Now, we define an average operator along the z-axis as follows:


Z
1
Mn (x, y, z, t)dz for all (x, y, z) .
hMn iz :=
0

(2.13)

This is a typical way of working in the averaging technique developed by Whitaker


[9]. By (2.13), (2.10) becomes



g Tg

g iz + g Tg
g iz = g 4hT

(2.14)
g cg
hTg iz + u hT
.
t
z z=
z z=0
Applying the boundary condition (2.8) in (2.14), we have



g Tg

g cg
hTg iz + u hTg iz = g 4hTg iz
.
t
z z=0

(2.15)

Similarly, we obtain the following from (2.3), but changing the path of integration
to z (, 0).

s Ts

s cs hTs iz = s 4hTs iz +

t
z z=0

(2.16)

Note that we have changed the path of integration to maintain the fact that the
solid is defined in a region situated below the gas.
10

Now, adding up (2.15) and (2.16), and using the boundary condition in (2.8), we
obtain



g iz + s 4hT
s iz + Q W (T, C) in Y.

g cg
hTg iz + u hTg iz + s cs hTs iz = g 4hT
t
t

(2.17)
In (2.17), we can also denote W (T, C), by
(T, C) := W (T, C) .
W

(2.18)

Furthermore, integrating (2.4) and using definitions (2.11) and (2.12) as we did for
(2.2), we obtain

2C
dz
z 2

(2.19)

D C
D C

hCiz + u hCi
=
D
4hCi
+


.
z
z
t
z z=
z z=0

(2.20)

Cdz + u

Cdz = D4

Cdz + D
0

Z
0

such that by (2.13), we have

Using the boundary conditions (2.7) and (2.8), we obtain the following:

hCiz + u hCi
z = D 4hCiz W (T, C) in Yg .
t

(2.21)

Finally, the set of equations is completed by


R
= W (T, C) at Ys .
t

(2.22)

However, in order to be consistent with the model reduction in terms of averages,


we must also average the differential equation (2.22), as well as the remaining
boundary and initial conditions . In what comes next, we present closure relations
(strong assumptions) under which we can average the ordinary differential equation.
Claim: In Y , let the temperature of the solid-gas interface be T . Let also the
temperature-dependent Arrhenius exponential factor be given by A(T ). Assuming
that T is below a critical temperature Tc , then A(T ) reduces to
(
0, T < Tc
A(T ) =
(2.23)
, T > Tc ,
where is a positive constant. In particular, if T  Ta in A(T ), then
W (T, C) = AC exp(

11

Ta
) AC.
T

(2.24)

Proof of the claim:


For T 25 C and Tc = 450 C i.e., the auto-ignition temperature of paper, we see
that A(T ) 0 and conversely, for T > Tc , we have that A(T ) = , for some positive
constant . Let k = TTa , such that A(T ) = exp(k). Taking the Taylors expansion
of A around k = 0, we have
A(T ) 1 k +

k2 k3

+ O(k 4 )
2
6

(2.25)

Hence, T  Ta , k 0 implies that A(T ) 1 and thus W (T, C) AC.


Now, assuming that the gas occupies the porous region of the material as shown in
Fig. 1.3, and applying the result of our claim, (2.22) reduces to
R
= AC.
t

(2.26)

Averaging the right hand side of (2.26), leads to


hRiz
= AhCiz .
t

(2.27)

Alternatively, the right hand side of (2.22) can be approximated by the mean value
theorem as follows:
We assume, as before, that the gas occupies the porous region of the paper as
depicted in Fig.1.3. For small variations in the local concentration and temperature
around z = 0, it makes sense to average along 0 z such that,
Z
1
hRiz
W (T, C)dz
(2.28)
=
t
0
where the unknowns C and T depend on (x, y, z, t).
Now, applying the mean value theorem (MVT), the right hand side of (2.28) becomes


Z
Z
1
1
Ta
W (T, C)dz =
AC exp
dz
0
0
T


Ta

= C(x, y, z , t) exp
T (x, y, z , t)
for some z (0, ). By the definition of the average operator, we have
Z
1
hCiz =
C(x, y, z, t)dz C(x, y, z , t) and
0
Z
1
hT iz =
T (x, y, z, t)dz T (x, y, z , t) (by MVT),
0
for some z (0, ). Thus,
W (T, C) W (hCiz , hT iz )
12

at Ys .

(2.29)

Note that the same averaging idea used in (2.29) also applies to the last term
of (2.22). Another possibility of averaging (2.22) will be to consider the FrankKamenetskii approximations (see [10] and [11] for example.).
From the derivations so far, we see that the equations are given in quantities and
parameters associated with a given part of the representative unit cell (Yg or Ys ,
see Fig. 1.2). However, we wish to reduce the number of variables appearing in the
equations by introducing the following definition:
Definition 2.1.1 Let Y be the representative unit cell as shown in Fig. 1.2 (right).
The characteristic function Y` restricted to a part of Y is given by
(
1, (x, y) Y` , ` {g, s}
Y` (x, y) :=
(2.30)
0, otherwise.
Using (2.30) yields
T := Ts Ys + Tg Yg ,

(2.31)

c := s cs Ys + g cg Yg ,
:= s + g .

Yg
Ys

(2.32)
(2.33)

Summarizing, the governing equations and initial/boundary conditions, now formulated in 2-dimensions, are:



4h
Tiz + QW (hTiz , hCiz ) in Y,
Tiz =
c
hT iz + u h
(2.34)
t

hCiz + u hCi
z = D 4hCiz W (hT iz , hCiz )
t
hRiz
= W (hTiz , hCiz )
t

in Yg ,

at Ys .

The corresponding boundary conditions are:

Dn hCiz = 0
Ti
g iz = ns hT
s iz = nh
n g hT

hTs iz = hTg iz = hT iz

(2.35)

(2.36)

at Ys .

(2.37)

The upstream boundary conditions include the following:


hCiz
= 0 at {x = L}
x

hT iz = T0
hCiz = C0

hRiz = 0

at {x = 0}.

13

(2.38)

(2.39)

Finally, we supplement the model problem with initial conditions

hT iz = T0
for {0 < (x, y) < L}.
hCiz = C0

hRiz = R0

(2.40)

Note that the 2-D system of equations in (2.34)-(2.36) needs to be understood in


the distributional sense.
Having obtained the set of model equations in terms of averages, we wish to drop
the notation, h iz , as used for the averages and introduce the notations , , and
R. Also, we drop the tildes and bars on the parameters. The model equations
formulated in the new variables are:



+ u = 4 + QW (, ) in Y,
(2.41)
c
t

+ u = D4 W (, )
t
R
= W (, )
t

in Yg ,

at Ys ,

and the corresponding boundary conditions are:

Dn = 0
n (g g s s ) = 0

s = g =

at Ys ,

(2.42)

(2.43)

(2.44)

where , and R are respectively the temperature, gas concentration, and solid
product(char).

2.1.1

Non-dimensionalization

In this section, we make the model equations dimensionless following the same
strategy given in [2], [6]. Instead of having a large number of physical parameters
and variables, all with dimensional units, we wish to scale all the variables with
characteristic values-values of the size we expect to see, or dictated by the geometry, boundary conditions etc-so that we are left with an equation written in
dimensionless variables.
The idea is to collect all the physical parameters and typical values together into a
smaller number of dimensionless parameters (or dimensionless groups) which, when
suitably interpreted, should tell us the relative importance of the various mechanisms. This means that we are much interested in obtaining relevant dimensionless
numbers, which can be used to interpret the physical mechanisms involved in the
14

balanced laws.
We begin now with (2.17) and rescale the mass concentration of the gas and solid
product with their initial values as follows:
:=

hCiz
,
C0

hRiz
R0

(2.45)

where ` {s, g},

(2.46)

R :=

and the temperature by


` :=

hT` iz T0
,
Tb T0

with Tb = T0 + f racQcg i.e., the temperature of combustion product. Also, we


rescale the independent variables (x, y) and t respectively by L and
tD :=

L2
t
, such that :=
D
tD

(2.47)

where tD is the diffusive time scale. Rewriting (2.17) in terms of the characteristic
scales, we obtain
D

(Tb T0 ) s u(Tb T0 )
g (Tb T0 )
(Tb T0 ) g
+ mD
+
g =
4g
2
2
L

L
g cg
L2
s (Tb T0 )
+
4s
g cg
L2

QAC0 A()
+
g cg
(2.48)

where
s cs
m :=
g cg

(the weighted mass ratio) and A() = exp

Ta
(T0 + (Tb T0 ))


.

Next, we divide through by the coefficient of the term on the left hand side, so
that we have

g
s uL
g
s
QAC0 L2 A()
4g +
4s +
+m
+
g =

D
g cg D
g cg D
D(Tb T0 )g cg
At this moment, we introduce the following the combustion time scale

1
QAC0
tB =
,
(Tb T0 )g cg

(2.49)

(2.50)

where in (2.49), we define


Les :=

s
,
s cs D
15

Leg :=

g
g cg D

(2.51)

are respectively the Lewis number of the solid (i.e. the ratio between the equivalent
heat diffusivity of the paper and oxygen diffusivity) and gas, respectively.
Furthermore,
P e :=

uL
;
D

the Peclet number.

(2.52)

Thus, the equation is given by


g
s
tD
+m
+ P eg = Leg 4g + mLes 4s +
A().

tB

(2.53)

(2.53) can also be written in compact form by introducing the following definition
:= g g + ms s
Le := Leg g + Les s ,

(2.54)
(2.55)

so that, we may write (2.53) as


tD

A().
+ P e = Le4 +

tB

(2.56)

Similarly, the non-dimensional form of (2.21) and (2.36) are:

tD
+ P e = 4
A(),

tG

(2.57)

where the gas combustion time scale tG is defined as


tG := (A)1
R
L2 C0
=
AA()

D R0
tD
= A().
tR
The time scale of solid product tR is given by
1

C0
,
tR :=
A
R0

(2.58)

(2.59)
(2.60)

(2.61)

Considering the boundary condition


n (g g s s ) = 0,

(2.62)

we simplify (2.62) by introducing g cg D and s cs . (2.62) then reduces to the following


g
s s cs
g
s = 0
g cg D
s cs D s cs
n (Leg g Les s ) = 0.
16

(2.63)
(2.64)

Hence, (2.64) can then be written as


[n Le] = 0 on Ys .
The dimensionless boundary conditions are:

Dn = 0
[n Le] = 0

s = g =

on Ys ,

= 0 at {x = L}.
x

(L, t) =


T T0
H( ),
Tb T0
tD

(2.65)

(2.66)

(2.67)

(2.68)

where T > Tc , for some critical temperature Tc .  is some sufficiently small time
and H() is the 2 Heaviside function.

s = g = 0
at {x = 0}
(2.69)
=1

R=0
and the initial conditions are:

g = s = 0
=1

R=1

for {0 < (x, y) < L}.

(2.70)

The introduction of the Heaviside function is to mimic a heat pulse. (see [6] for more details.)

17

Chapter 3
The homogenization method
3.1

General averaging strategy

In this chapter, we wish to obtain a macroscopic model of the 2-dimensional formulation presented in subsection 2.1.1, which is now considered here to be the
micro-problem. We do this by applying the mathematical theory of periodic homogenization, i.e., the averaging of the solution of the microscopic equations for
finding effective equations and coefficients (see [12], [13] for details).
By this averaging procedure, the equations are replaced by similar equations with
constant coefficients, which are easy to solve numerically. Therefore, the idea of
this homogenization technique is to consider the heterogeneous medium as a periodically repeating structure with period as pointed out earlier (see section 1.2).
Then, a sequence of problems for which the length scale , becoming increasingly
small, goes to zero.
The main target in this chapter is to perform an asymptotic analysis as tends
to zero. An effective coefficient (more precisely homogenized (macroscopic) tensor)
in the limit problem ( = 0), describes material properties of the medium such as
its conductivity. We will see later that the effective coefficients typically depend on
the microstructure information.
An advantage of this approach, despite its complexity, is that the homogenized
property is uniquely defined. As a further step, the approximation made by using
only effective properties instead of the true microscopic coefficients can be rigorously justified by quantifying the resulting error (see [12]). We comment more on
this idea in chapter 4 and chapter 5

3.2

The microscopic problem

Let be a region in R2 with boundary containing two phases, gas and solid,
as discussed in section 1.2. The gas and the solid matter possess constant material

18

coefficients1 specific to each of the phases. We denote the coefficient by Lij .


Let the spatial variable be given by x := (x1 , x2 ). We refer to x as the macroscale
(global) variables. Also, let the temporal variable be . We assume that the quantities , , and R defined in chapter 2 are prescribed in the cylindrical domain
Q := (0, T ), where T is a fixed time.
By the assumption of periodicity given in chapter 2, R2 is then divided into periodically repeated unit cells Y. Each unit cell is divided into two parts, the gas and
the solid parts Yg and Ys . The surface porosity of the medium is defined by
:=

meas(Yg )
.
meas(Y )

(3.1)

In (3.1), meas(Y) is the Lebesgue measure (i.e. the area in R2 ) of Y . In what comes
later on, we will write |Y | instead of meas(Y).
The periodicity in variable Y implies that the coefficient Lij must be periodic.
We describe the Y-cell, with the microscale (local) variable, as y := (y1 , y2 ) with
yi := xi / for i {1, 2}. The scale factor suggests that the quantity together
with the coefficients in the different parts of the medium is defined by
:= g g (x/) ms s (x/)
Lij (x)
Pj (x)

(3.2)

:= Leg g (x/) + Les s (x/)

(3.3)

:= P eg (x/) for all x .

(3.4)

We assume that the convective term is divergence free i.e. P~ = 0. Also, we define
Lij (x) := Lij (x/) = L(y)
Pj (x) := Pj (x/) = P (y)

(for y = x )

(3.5)

Y-periodicity means that Lij (a) = Lij (b) whenever a and b are homologous by periodicity, i.e. have the same positions in corresponding cells. In addition, if the points
a and b are close enough with respect to the x variable, we say that the coefficient
is locally periodic. For more details on Y-periodicity and locally periodic, see [14],
[13], [15].
We note that in a cell Y, Y-periodic functions take on the same boundary values twice, but with opposite outer normals n = (n1 , n2 ). Thus, for a Y-periodic
vector field (fi (y)), the following relation holds:
I
Z
fi
dy =
fi ni dS(y) = 0
(3.6)
Y yi
Y
by Gauss theorem. In the subsequent analysis, we will use (3.6) quite frequently.
In (3.6), dS(y) denotes the infinitesimal line element of .

19

Figure 3.1: (a): Deformed cell after time, t; (b) Not structurally deformed cell after
time t.

Also, we define the following Hilbert space for Y -periodic functions:


Z
1
1
vdy = 0}.
W (Y ) := {v H (Y ) : v is Y-periodic;
|Y | Y

(3.7)

Further, we assume that the paper is not structurally deformable, that is, the shape
of the paper from the start of the heating process is not deformed due to changes
resulting from the smoldering of the paper after a time, say t. This assumption and
its implication is illustrated in Fig. 3.1.
Finally, we assume that each Lij belongs to L () and that Lij is positive definite, i.e.
Lij i j i i

(3.8)

for some > 0 and all choices of i and j .


Now, we let vary, thereby obtain a class of initial boundary value problems depending on the parameter :
~
+ (P (y) L(y) ) = f ( , ) in ,

~
+ (P (y) ) = g( , ) in g ,

R
= h( , ) on ,

(3.9)
(3.10)
(3.11)

Coefficients defined differently in solid and gas parts can be interpreted as a single coefficient
discontinuous in the space variable. As a result, the corresponding equations can be interpreted
in the sense of distributions in .

20

where
tD
A( ),
tB
tD
),
g( , ) = A(
tG
tD
.)
h( , ) = A(
tR

f ( , ) =

(3.12)
(3.13)
(3.14)

with


A( ) = exp

Ta
(T0 + (Tb T0 ) )


.

The corresponding boundary and initial conditions are as follows:

n L(y) = 0
on Ys ,
n = 0

g = s =

s = g = 0
g = 1


R =0

at {x = 0},

= 0 at {x = L},
x
(L, t) = H(

(x, 0) = 0
(x, 0) = 1


R (x, 0) = 1


),
tD

for {0 < (x, y) < L}.

(3.15)

(3.16)

(3.17)

(3.18)
(3.19)

(3.20)

We refer to (3.9)-(3.11) as the microscopic problem.

3.3

Homogenization of the microscopic problem

We consider (3.9), (3.10), and (3.11) and assume that the microscopic solutions
, , R together with the right hand side terms f , g and h can be represented
by asymptotic expansions of the form:

(x, ) = 0 (x, y, ) + 1 (x, y, ) + 2 2 (x, y, ) + O(3 )

(x, ) = 0 (x, y, ) + 1 (x, y, ) + 2 2 (x, y, ) + O(3 )

R (x, ) = R0 (x, y, ) + R1 (x, y, ) + 2 R2 (x, y, ) + O(3 )


x
,
for
y
=

0
0
1
0
0
1
0
0
3

f
(
,

)
=
f
(
,

)
+

1 f ( , ) + 2 f ( , ) + O( )

g( , ) = g(0 , 0 ) + 1 1 g(0 , 0 ) + 1 2 g(0 , 0 ) + O(3 )

h( , ) = h(0 , 0 ) + 1 1 h(0 , 0 ) + 1 2 h(0 , 0 ) + O(3 )


(3.21)
21

and the functions i (x, y, ), i (x, y, ) and Ri (x, y, ) are all Y periodic in y for
any choice of i N
The first terms of the series expansions (3.21) will be usually identified with the
typically verified homogenized equations whose unknown effective coefficients can
be exactly computed.
Therefore, numerical computations involving the homogenized equation do not require a fine mesh solution since the heterogenities of size have been averaged out
(see [12] for example).
Now, following the line of argument of [13], we define the following operators:
~ (P (y) L(y)),
A :=
~ (P (y) ).
B :=

(3.22)
(3.23)

By setting (x) = (x, y), y = x , we obtain via the chain rule of differentiation:
1
= x + y .
(3.24)

Next, we insert the homogenization ansatz (3.21) into (3.9), (3.10), and (3.11). In
this way, a geometric series in is obtained by application of the chain rule of differentiation (3.24).
From (3.9), we obtain

~ x + 1
~ y ) (P (y) L(y)((x + 1 y ))(0 + 1 + ))
(0 + 1 + ) + (

= f (0 , 0 ) + 1 1 f (0 , 0 ) + 1 2 f (0 , 0 ) +
(3.25)
(2 A0 + 1 A1 + 0 A2 )(0 + 1 + ) = f (0 , 0 )
+ 1 1 f (0 , 0 )+
1 2 f (0 , 0 ) +

(3.26)

where
~ y (L(y)y ),
A0 :=
~ y (P (y) L(y)x )
~x (L(y)y ),
A1 :=

(3.27)

~x (P (y) L(y)x ).
A2 :=

(3.29)

(3.28)

Now, equating the terms with the same powers of leads us to the following three
lowest order equations:
A0 0 = 0 (2 term),
A0 1 + A1 0 = 0 (1 terms),
0
+ A0 2 + A1 1 + A2 0 = f (0 , 0 ) (0 terms).

22

(3.30)
(3.31)
(3.32)

(3.30), (3.31) and (3.32) are posed in . Likewise, we expand the boundary condition
corresponding to (3.10) and obtain the following expressions:
1
n L(y)(x + y )(0 + 1 + 2 2 + O(2 )) = 0

(3.33)

from which we derive:


n L(y)y 0 = 0 (1 term),
n L(y)(x 0 + y 1 ) = 0 (0 term),
n L(y)(x 1 + y 2 ) = 0 (1 term).

(3.34)
(3.35)
(3.36)

(3.30), (3.31), and (3.32) together with (3.34), (3.35) and (3.36) suggest solving

, corresponding to each j term for j


a succession of differential problems Pj
{0, 1, 2}. To this end, we pose the following problems:

0
~

y (L(y)y ) = 0 in Y

(3.37)
P2
[n L(y)y 0 ] = 0 on Ys ,

0
is Y-periodic.
This implies that 0 does not depend on y. The only periodic solution satisfying
(3.37) is that 0 must be a constant with x and as parameters, i.e.
0 (x, y, ) = 0 (x, ).
Taking into account (3.38), we find the corresponding problem for 1 :

0
1
0
~

y (L(y)(x + y )) + y P (y) = 0 in Y ,

P1
[n L(y)(x 0 + y 1 )] = 0 on Ys ,

1
is Y-periodic.

(3.38)

(3.39)

Since (3.39) is linear, we expect that 1 is a linear function of x 0 . Thus, we


compute 1 in terms of the gradient of 0 :
1 (x, y, ) :=

2
X
0
j=1

xj

wj (y) =

0
0
w1 (y) +
w2 (y),
x1
x2

(3.40)

where wj (y)j {1, 2} are the so-called cell functions.


It is clearly pointed out in [12], [14], that 1 is defined up to an additive func 1 (x, ), depending only on x and . This does not matter since only its
tion
gradient w.r.t y arises in the homogenization equation (3.43).
Also, we recall the definition of the gradient of a scalar field:
0

x =

2
X
0
j=1

xj

~ej =

23

0
0
e1 +
e2 ,
x1
x2

(3.41)

where (~ej )j=1,2 , is the canonical basis in R2 .


We apply (3.40) and (3.41) in (3.39) and due to linearity, we solve the following problem:
Find w = (w1 , w2 ) W (Y ) satisfying

y (L(y)(y wj (y) + ~ej )) = 0 in Y ,


Pj [n L(y)(y wj (y) + ~ej )] = 0 on Ys ,
(3.42)

wj (y) is Y-periodic,
where the term y P (y)0 in (3.39) vanishes due to the fact that P (y) is divergence
free, i.e. P (y) = 0.
Note
that for all j {1, 2}, wj (y) W (Y ) = {v H 1 (Y ) : v is Y-periodic;
R
1
vdy = 0} the function w = (w1 , w2 ) is the unique solution of the so-called
|Y | Y
local or cell problems (for more details, see [14], [15] e.g.).
Finally, we solve the problem for 0 taking into account (3.38), as well as the
solutions of (3.42), and thus (3.39).

~ y (L(y)(x 1 + y 2 ) P (y)1 )
0

~x (L(y)(x 0 + y 1 ) P (y)0 ) = f (0 , 0 ) in Y,
P0

[n L(y)(x 1 + y 2 )] = 0
on Ys ,

2
is Y-periodic.

(3.43)

Now, we obtain the homogenized equation from (3.43) by averaging it with respect
to y, and inserting the definition of 1 given in (3.40) as described below:
Z
Z
1
0
1
1
2
~
~ y P (y)1 dy
y L(y)(x + y )dy +

|Y | Y
|Y | Y
{z
} |
{z
}
|
T1
T2
Z
Z
Z
1
1
1
0
1
0
~
~
x (L(y)x + y )dy +
x P (y) dy =
f (0 , 0 )dy
|Y | Y
|Y | Y
|Y | Y
|
{z
} |
{z
} |
{z
}
T3

T4

T5

(3.44)
We solve the integrals labeled T1 , T2 , , T5 as follows:
For T1 :
Z
Z
1
1
1
2
~
y L(y)(x + y )dy =
n L(y)(x 1 + y 2 ) dS(y)
{z
}
|Y | Y
|Y | Ys |
=0 on Ys
Z
1

n L(y)(x 1 + y 2 )dS(y)
|Y | Y
|
{z
}
=0 by periodicity of Y
24

This implies that the integral T1 = 0.


For T2 :
Z
Z X
2
1
1
0
1
y P (y) dy =
y (P (y)wj (y))dy
|Y | Y
|Y | Y j=1 xj
Z
Z
2
2
1 X 0
1 X 0
y P (y)wj (y)dy =
n (P (y)wj (y))dS(y)

|Y | j=1 xj Y
|Y | j=1 xj Y
= 0 (by the periodicity of P (y)wj (y)).
For T4 :
Z
Z
1
0
0 1
x P (y) dy = x
P (y)dy
|Y | Y
|Y | Y
= P x 0 ,
where

1
P :=
|Y |

(since 0 is y-independent) (3.45)


(3.46)

Z
P (y)dy.
Y

Similarly, the fact that T5 = f (0 , 0 ) follows from the fact that 0 and 0 are
y-independent.
Finally, we have for T3 :
1

|Y |
1

|Y |

Z 
Y

~x (L(y)(x 0 + y 1 ))dy =


 0



0
0
0

~ x L(y)y
~ x L(y)

w1 +
w2
e1 +
e2 dy.
+
x1
x2
x1
x2

Collecting conveniently the termsR and since that the differential operator commutes with the integral operator , we obtain:

 0
Z
Z
0
1

~x
(L(y)y w1 + L(y)e1 )dy +
(L(y)y w2 + e2 )dy

x1 |Y | Y
x2 |Y | Y
= x (A(w)x 0 ).
(3.47)
where, on expanding the integral in (3.47), we obtain the matrix, A(w) as
R
1 R

1
1
2
(L(y) w
+ L(y))dy
(L(y) w
)dy
|Y | Y
y1
|Y | Y
y1

A(w) = R
R
1
1
1
2
(L(y) w
)dy
(L(y) w
+ L(y))dy
|Y | Y
y2
|Y | Y
y2

25

(3.48)

where w = (w1 , w2 ).
Similarly, we follow the same procedure as above to deal with (3.9).

~ x + 1
~ y ) (P (y) ((x + 1 y ))(0 + 1 + ))
(0 + 1 + ) + (

0
0
1
0
0
= g( , ) + 1 g( , ) + 1 2 g(0 , 0 ) +
(3.49)
= (2 B 0 + 1 B 1 + 0 B 2 )(0 + 1 + ) = g(0 , 0 )
+ 1 1 f (0 , 0 ) + 1 2 g(0 , 0 ) +
(3.50)
where
~ y y ,
B 0 :=
~ y (P (y) x ) (
~x y ),
B 1 :=

(3.51)

~x (P (y) x ).
B 2 :=

(3.53)

(3.52)

Now, equating in powers of leads us to the following three lowest order equations:
B 0 0 = 0 (2 term),
B 0 1 + B 1 0 = 0 (1 terms),
0
+ B 0 2 + B 1 1 + B 2 0 = g(0 , 0 ) (0 terms).

(3.54)
(3.55)
(3.56)

Similarly, we expand the boundary conditions corresponding to (3.9) to obtain:


1
n (x + y )(0 + 1 + 2 2 + ) n y 0 = 0 (1 term),

n (x 0 + y 1 ) = 0 (0 term),
n (x 1 + y 2 ) = 0 (1 term).

(3.57)
(3.58)
(3.59)

From the information provided in (3.54), (3.55), (3.56) and (3.57), we construct a
sequence of problems as follows:

0
~

y (y ) = 0 in Yg

P2
(3.60)
n (y 0 ) = 0 on Ys ,

0
is Y-periodic.
(3.60) indicates that 0 does not depend on y. The only periodic solution satisfying
(3.60) is that 0 must be a constant in y and having x and as parameters, i.e.
0 (x, y, ) = 0 (x, ) for all x and 0.

26

(3.61)

Taking into account (3.61), we give the corresponding problem for 1 :

0
1
0
~
~

y (x + y ) + y P (y) = 0 in Yg ,

P1
n (x 0 + y 1 ) = 0 on Ys ,

1
is Y periodic.

(3.62)

Following the same argument as in the case of 1 , we compute 1 in terms of the


gradient of 0 :
1

(x, y, ) =

2
X
0
j=1

xj

j (y) =

0 (x, )
0 (x, )
1 (y) +
2 (y)
x1
x2

(3.63)

for all x , y Y and 0


Also, we use the following definition of a gradient:
0

x =

2
X
0
j=1

xj

~ej =

0
0
e1 +
e2 .
x1
x2

(3.64)

We apply (3.63) and (3.64) in (3.62) and due to linearity of the problem, we obtain
the following cell problem: Find
j (y) = (1 , 2 ) W (Y )
such that the following BVP is solvable, viz.

y (y j (y) + ~ej ) = 0 in Yg ,
Pj n (y j (y) + ~ej ) = 0 on Ys ,

j (y) is Y-periodic.

(3.65)

One easily sees that j (y) is the unique solution of cell problems. A simple calculation shows that the term y (P (y)0 ) in (3.62) vanishes due to y P (y) = 0
Finally, we solve the problem obtained for 0 taking into account (3.61), as well as
the solutions of (3.65), and thus (3.62). We get

~ y ((x 1 + y 2 ))
~x ((x 0 + y 1 ))
0

+ (P (y)0 ) + (P (y)1 ) = g(0 , 0 ) in Y ,


x
y
g

P0
(3.66)
1
2

n (x + y ) = 0
on Ys ,

2
is Y-periodic.
Averaging (3.66) over Y , we have:
Z
Z
1
1
0
1
2
~ y (x + y )dy +
~ y P (y)1 dy

|Y | Yg
|Y | Yg
|
{z
} |
{z
}
S1
S2
Z
Z
Z
1
1
1
0
1
0
~
~
x (x + y )dy +
x P (y) dy =
g(0 , 0 )dy
|Y | Yg
|Y | Yg
|Y | Yg
|
{z
} |
{z
} |
{z
}
S3

S4

S5

(3.67)
27

As before, we solve the integrals labeled S1 , S2 , , S5 as follows:


For S1 :
1
|Y |

~ y (x + y )dy = 1

|Y |
Yg
1

|Y |
|

n (x 1 + y 2 ) dS(y)
{z
}
Ys |
=0 on Ys

n (x 1 + y 2 )dS(y)
Y
{z
}
=0 by periodicity of Y

(3.68)

This implies that S1 = 0.


For S2 :
Z
Z X
2
1
1
0
1
y (P (y)j (y))dy
y P (y) dy =
|Y | Yg
|Y | Yg j=1 xj
Z
Z
2
2
1 X 0
1 X 0
y P (y)j (y)dy =
n (P (y)j (y))dS(y)
This leads to
|Y | j=1 xj Yg
|Y | j=1 xj Y
= 0 (by the periodicity of P (y)j (y)).
For S4 :
Z
Z
1
0
0 1
x P (y) dy = x
P (y)dy
|Y | Yg
|Y | Yg

(since 0 is y-independent)
(3.69)

= P x ,
where

1
P :=
|Y |

Z
P (y)dy.
Yg

|Yg |
g(0 , 0 ) follows from the
|Y |
|Yg |
can be interpreted as the
|Y |

Similarly, S5 =

fact that 0 and 0 are y-independent.

The quantity
medium porosity, i.e. the area of the
gas part of the basic cell to the total area of the cell.
Finally, we have for S3 :
1

|Y |
1

|Y |
"

Z 
Yg

~x (x 0 + y 1 )dy =

Yg

  0

 0

0
0

~ x y
~x
1 +
2
+
e1 +
e2 dy.

x1
x2
x1
x2

0
~ x 1

x1 |Y |

0 1
(y 1 + e1 )dy +
x2 |Y |
Yg

#
~ x (B()x 0 ).
(y 2 + e2 )dy =

Yg

(3.70)
28

where, on expanding the integral in (3.70), we obtain the matrix B() defined by
R 2
1 R 1
1
(
+
1)dy
dy
|Y | Yg y1
|Y | Yg y1

.
B() :=
(3.71)
R 1
R 2
1
1
dy
(
+
1)dy
|Y | Yg y2
|Y | Yg y2

3.4

The macroscopic equations and effective coefficients

Figure 3.2: Left: Heterogeneous medium; Right: Homogeneous representation of


the perforated domain.
Summarizing, we obtain the following effective (upscaled) equations for the required
macroscopic quantities, namely:
(x, ) = 0 (x, ),
(x, ) = 0 (x, ),
R(x, ) = R0 (x, ),
and


+ P A = f (, ) in (0, T ),

+ P B = g(, ) in (0, T ),

R(x, )
= h(, ) in (0,T) and a.e x ,

(3.72)
(3.73)
(3.74)

where

A(w) :=

1
|Y |

1
|Y |

1
+ L(y))dy
(L(y) w
y1

1
(L(y) w
)dy
y2
Y

1
|Y |

29

1
|Y |

R
Y

2
)dy
(L(y) w
y1

2
(L(y) w
y2
Y

+ L(y))dy

(3.75)


B() :=

1
|Y |

1
|Y |

( 1 + 1)dy
Yg y1
Yg

1
dy
y2

1
|Y |

1
|Y |

2
dy
Yg y1

( 2
Yg y2

(3.76)

+ 1)dy

and
1
P :=
|Y |

Z
P (y)dy,

(3.77)

at {x = 0},

(3.78)

Yg

with

= 0
=1

R=0

= 0 at {x = L},
x

(L, t) = H(

(x, 0) = 0
(x, 0) = 1

R=1


),
tD

for {0 < (x, y) < L}.

(3.79)

(3.80)

(3.81)

It turns out that the effective coefficients describe a homogeneous medium that is
independent of the right-hand side terms f, g and h as well as of the prescribed
boundary and initial conditions.
Furthermore, the application of (3.21) in (3.9), (3.10) and (3.11) leads to a rigorous deductive procedure for obtaining the macroscopic equations (3.72), (3.73)
and (3.74). These equations can then be used for predicting the global behavior of
the heterogeneous medium.
Thus, homogenization process gives the passage from a microscopic description
to a macroscopic description of the model problems (3.9), (3.10) and (3.11) [16].
Intuitively, we illustrate in Fig. 3.2 (left) the heterogeneous material (left) that is
homogenized, leading to its homogeneous representation in Fig 3.2, by the application of the homogenization method.

30

Chapter 4
Numerical multiscale
homogenization approach
In this chapter, we wish to explore numerically the solution of the system of equations obtained from the multiscale structure discussed in chapter 2 and chapter 3.

4.1

Computation strategy for the homogenized


problem

In order to compute the solution to the homogenized problem numerically, we consider the following three-steps procedure:

Step I: Computation of the cell problems


For the purpose of the discussion in this section, we write the cell problems described
in (3.42) and (3.65), in the following form:



wj

Lij (y), wj Y periodic,


(4.1)

Lik (y)
=
yi
yk
yi

yi

j
yk


=

ej = 0,
yi

j Y periodic.

(4.2)

We use COMSOL Multiphysics [17] to solve (4.1) and (4.2) posed in the unit cells
depicted in Fig. 4.1 and Fig. 4.2 respectively. We essentially use the Finite Element Method (FEM), since this method is well suited for problems with complex
geometries.
Looking at the geometry depicted in Fig. 4.1, we see that the unit cell has two
disjoint subdomains Yg and Ys , where the subdomains have different material coefficients given by Lij .

31

The coefficient term Lij (y), y Y , is defined as follows:


(
Lgij , y Yg
Lij (y) :=
Lsij , y Ys .

(4.3)

Consequently, we define (4.1) in both Yg and Ys but with their corresponding material coefficients. In particular, we define (4.2) in Yg for the geometry in Fig. 4.2
Further, we compute Lij for the two materials and we assume, in the present study,
that these coefficients are isotropic in the sense that L`ij = L`ij ij for ` {g, s}.

Figure 4.1: Unit cell with two subdomains Yg and Ys .

Figure 4.2: Unit cell with hole: one subdomain.

Step II: Computation of the homogenized (effective) coefficients


We insert the solutions of the cell problems into the following equations of the
homogenized tensors-A and B:
Z
D
wj E
1
wj
aij := Lij (y) + Lik (y)
=
(Lij (y) + Lik (y)
)dy,
(4.4)
yk
|Y | Y
yk
32

j E
1
bij := ij + ik
=
yk
|Y |
D

Z
(ij + ik
Y

j
)dy.
yk

Z
D E
1
P := Pk =
Pk dy.
|Y | Yg

(4.5)

(4.6)

The homogenized coefficient are thus calculated by integration over the unit cell
(precisely with respect to the subdomain(s)).

Step III: Solution of the homogenized problem


With the homogenized coefficients computed in step II, the upscaled model (3.72)(3.74) can then be solved numerically provided the numerical approximation of the
solutions to (4.1)-(4.2) are available.

4.2
4.2.1

Numerical computation of the homogenized


coefficients
Weak formulation of the cell problem

The homogenized coefficients are given by (4.4) and (4.5). In order to compute
these coefficients, we need to solve the cell problems (4.1) and (4.2) respectively for
wj and j . We then compute the integrals in (4.4) and (4.5).
To get the weak formulation of (4.1), we note that



Lik
yj =
Lik kj =
Lij

yi
yk
yi
yi



ej = 0
yi yk
so that (4.1) and (4.2) can be written as:



(wj + yj ) = 0,

Lik
yi
yk



(wj + ej ) = 0,
yi yk

(4.7)
(4.8)

wj is Y periodic.

(4.9)

j is Y periodic.

(4.10)

Now, we recall the space W(Y).


W (Y ) = { H 1 (Y ) : is Y-periodic}.

(4.11)

Assume we have solutions ujk := wj + yj and vj := j + ej to (4.9) and (4.10)


respectively, then (4.9) and (4.10) can be written as



Lik
= 0,
(4.12)
yi
yk



= 0.
(4.13)
yi yk
33

Multiplying (4.12) and (4.13) by W (Y ) and integrating over Y, we have




Z
Z
Z

u
u
u

Lik
Lik
Lik
dy =
dy
ni dS
(4.14)
yi
yk
yk yi
yk
Y
Y
Y


Z
Z
Z

v
v
v

dy =
dy
Lik
ni dS
(4.15)
yi yk
yk
Yg
Y yk yi
Y
by Greens theorem. The boundary integral terms in (4.14) and (4.15) vanish due
to Gauss theorem and periodicity. Hence we have for all W (Y ), the identities
Z
u
Lik
dy = 0
(4.16)
yk yi
Y
Z
v
dy = 0.
(4.17)
Y yk yi
We define the following bilinear form for (4.16) and (4.17) respectively:
Z
u
Lik
aY (u, ) :=
dy,
yk yi
Y
Z
v
bY (v, ) :=
dy for all W (Y ).
Y yk yi

(4.18)
(4.19)

Thus, the weak formulations of (4.1) and (4.2) are then:


Find wj , j W (Y ) such that
aY ((wj + yj ), ) = 0 for all W (Y ),
aY ((j + ej ), ) = 0 for all W (Y ),

(4.20)
(4.21)

or alternatively,
find wj , j W (Y ) such that
aY (wj , ) = FYj (),

(4.22)

aY (j , ) = GjY () for all W (Y ),

(4.23)

where
Z

dy,
Lij (y)
yi
Y
Z

GY () :=
dy.
ej
yi
Yg

FY () :=

4.2.2

(4.24)
(4.25)

Finite element approximation to the solutions of the


cell problems

In order to obtain a FEM formulation of (4.20) or (4.22), we will use the Galerkin
method [18] to describe the idea behind our later implementation of the method
in COMSOL Multiphysis solver. We start by making a subdivision of the unit cell
34

Y in N triangles Kn and M nodes Ni according to [18]. Let Wh (Y ) be a finitedimensional subspace of W (Y ) of dimension . We introduce the finite element
basis functions k (y) Wh (Y ), k = 1, 2, , , where denotes the number of
degrees of freedom.
Generally, is seldom equal M , because in the case of Y-periodicity, nodes at
opposite positions on the boundary of the Y -cell must correspond to the same degrees of freedom. To relate the nodes to the degrees of freedom, we let Q be a
mapping from node Ni to the corresponding degree of freedom l, that is l = Q(Ni ).
We also choose the basis functions to be piecewise linear and define them by
k (Ni ) = kl , where l = Q(Ni ) and k = 1, 2, , .

(4.26)

We assume the existence of an approximate solution to (4.22) of the form


wjh (y) =

kj k (y) for all y Y,

(4.27)

k=1

where kj are constants to be determined. This is the so-called Galerkin ansatz.


These constants correspond to the approximation of wjh at the nodal points. Let
Wh (Y ) be such that has a unique representation
=

i i .

(4.28)

i=1

We now formulate the discrete analogue to the problem (4.22):


Find wjh Wh (Y ) such that
aY (wjh , ) = FYj () for all Wh (Y ).

(4.29)

Because (4.29) must be valid for all Wh (Y ), it makes sense to use =


i , for i = 1, 2, , as test functions for the finite dimensional approximation.
With = i (y), (4.29) becomes

Aik kj = bji , i = {1, 2, , }; j = {1, 2},

(4.30)

k=1

or, in matrix form,


A j = bj ,

(4.31)

where
Z
Aik = aY (k , i ) =

Lij
Y

k i
dy
yj yi

(4.32)

and
bji

FYj (i )

Z
=

Lij (y)
Y

i
dy
yi
35

for all i Wh (Y ).

(4.33)

The finite element approximation of the second problem (4.23) follows in a similar
way.
We can now compute the approximate homogenized coefficients by inserting (4.27)
in (4.4) to get
Z

l
)dy
y
j
Y
l=1
Z
Z

1
1 X j
l
dy.
=
Lij dy +
l
|Y | Y
|Y | l=1
Y yj

1
aij =
|Y |

(Lij + Lik

lj

(4.34)

We note that
Z
Lmj

aY (l , yi ) =
Y

because

yi
ym

l yi
dy
yj ym

(4.35)

= im . Hence, we obtain
1
aij =
|Y |

Z
Y

1 X
Lij dy +
aY (l , yi )lj
|Y | l=1

Suppose Lij is symmetric, (4.24) and (4.33) gives


Z
Z
l
l
dy =
aki
dy = FYi (l ) = bil
aik
yk
yk
Y
Y

(4.36)

(4.37)

so that
1
aij =
|Y |

Z
(Lij dy +
Y

1 X i j
b .
|Y | l=1 l l

(4.38)

In order to implement the FEM discretization of subsection 4.2.1 in COMSOL


Multiphysics, we use the conservative formulation of the cell problems (3.42) and
(3.65). These equations are given in matrix form as:


  
L11 0
1 w1
L
y
+ 11
=0
(4.39)
0 L22 2 w1
0


  
L11 0
1 w2
0
y
+
=0
(4.40)
0 L22 2 w2
L11
(4.39) and (4.40) corresponds to the following PDE Coefficient form application
mode of COMSOL:
(cu u + ) + u + au = f
T

n (cu u + ) = h qu + g

in Y,

(4.41)

on Ys

(4.42)

with the following periodic boundary conditions u(yi ) = u(yi +Y ) on Y for i {1, 2}.
The model problems (4.39) and (4.40) are then obtained by setting a, , f, , g, q
36

Figure 4.3: Sample geometry of the unit cell


and to zero. We then implement the steps outlined in section 4.1 as follows:
Create the geometry and insert equations with corresponding material
coefficients:
The computational geometry is given in Fig. 4.3, where we have inserted, in each
of the subdomains, an equation with its corresponding material coefficient.
Prescribe the interior and periodic conditions and solve the problem:
For the cell problems, we prescribe periodic boundary conditions on the outer
boundaries of Y and use the Neumann condition on the interior boundary of Ys .
We solve the problems and obtain the solutions for the Y periodic cell functions
as depicted in Fig. 4.5 and Fig. 4.7. Fig. 4.5 shows the periodic solution of the
temperature in a unit cell, while in Fig. 4.7, we obtain the periodic solution for
concentration of oxygen in the part of the unit cell (see Fig. 4.2) Similar qualita-

(a)

(b)

Figure 4.4: Solution of cell problems, (a): w1 ( x ) and (b): w2 ( x ) with x .


tive results as depicted in Fig. 4.5Fig. 4.7 can be seen, for example, in [19], [20],
and [21].

4.2.3

Existence and uniqueness of weak solutions to the cell


problems

In this section, we introduce some important definitions and lemmas. These will be
used in the subsequent section to show the existence and uniqueness of weak solutions of the cell problems. Finally, some properties of the homogenized coefficients
e.g. like symmetricity are treated in subsequent sections.
37

(c)

(d)

Figure 4.5: Solution of cell problems, (c): w1 ( x ) and (d): w2 ( x ) with x .


3D visualization of the solution to the cell problems for temperature.

(e)

(f)

Figure 4.6: Solution of cell problem (e): 1 ( x ) and (f): 2 ( x ) with x .

(g)

(h)

Figure 4.7: Solution of cell problem (g): 1 ( x ) and (h): 2 ( x ) with x .


3D visualization of the solution to the cell problems for concentration.

Definition 4.2.1 Let X be a real Hilbert space. A mapping


a(, ) : X X R
is called a bilinear form on X if
i. a(x1 + x2 , y) = a(x1 , y) + a(x2 , y) for all x1 , x2 , y X ,
ii. a(x, y) = a(x, y) for all x, y X and R,
38

iii. a(x, y1 + y2 ) = a(x, y1 ) + a(x, y2 ) for all x, y1 , y2 X ,


iv. a(x, y) = a(x, y) for all x, y X and R.
Definition 4.2.2 Let a(, ) be a bilinear form on the real Hilbert space X. We say
that a(, ) is
i. bounded if there exists a constant > 0 such that |a(x, y)| kxkX kykX for all x, y X,
ii. symmetric if a(x, y) = a(y, x) for all x, y X,
iii. coercive if there exists a constant > 0 such that a(x, x) kxk2X for all x X.
Lemma 4.2.3 (Poincare inequality): Let be an open and bounded set in Rn
which has the Lipschitz property. Then
Z
u u
2
,
(4.43)
kukH 1 ()/R
xi xi
for each u H 1 ()/R and some constant C = C().
Proof. See e.g. [22].
Lemma 4.2.4 (Lax-Milgram): Let X be a real Hilbert space. If the bilinear form
a(, ) : X X R is bounded and coercive and if, in addition, the linear functional
f : X R is bounded, then there exists a unique element x X such that a(x, y) =
(f, y) for all y X.
Proof. See e.g. [23], p.118-119.
Lemma 4.2.5 Let F be square integrable over Y and consider the boundary value
problem
A0 = F in a unit Y-cell, is Y-periodic.

(4.44)

where A0 is similar to the operator (3.30). The following holds true:


i. There exists a weak Y-periodic solution of (4.2.5) if and only if hF i = 0.
ii. If there exists a weak Y-periodic solution of (4.2.5), then it is unique up to an
additive constant.
Proof. We recall the Hilbert space W (Y ) and define on it the bilinear form
Z

aY (, ) =
Lij
dy for all W (Y )
yj yi
Y
and the linear functional
Z
(F, )Y =

F dy.
Y

39

(4.45)

Then the weak formulation of (4.44) is:


Find W (Y ) such that
aY (, ) = (F, )Y for all W (Y ).

(4.46)

Assume that is a solution and choose 1. Then


0 = aY (, 1) = (F, 1)Y = |Y |hF i,

(4.47)

which shows that hF i = 0. Conversely, assume that hF i = 0. It is clear that the


bilinear form aY (, ) is not coercive in W (Y ), since if is a nonzero constant
function, then aY (, ) = 0 but kkW (Y ) > 0. However, according to Lemma 4.2.3,
it is coercive in W (Y ). Moreover, since Lij is bounded, it follows that aY (, ) is
bounded in W (Y ). (F, )Y is a well-bounded continuous linear functional on W (Y ),
since if c is a constant,
Z
Z
F cdy = c F dy = c|Y |hF i = 0,
(4.48)
Y

and therefore
Z
(F, )Y =

Z
F ( c)dy.

F dy =
Y

(4.49)

Hence
|(F, )Y | kF kL2 (Y ) inf k ckL2 (Y ) kF kL2 (Y ) kkW (Y ) .
cR

(4.50)

Now, by Lemma 4.2.4, there exists a unique solution in W (Y ) to (4.46).

4.3

Further properties of the effective coefficients

This section is very much inspired by [13].

4.3.1

Symmetry and positive definiteness of the effective


tensors

Theorem 4.3.1 Let Lij be a symmetric and positively definite tensor. Then the
associated effective tensor aij , defined as in (3.75), is symmetric and positive definite.
Proof. We consider the cell problem (3.42) and recall the weak formulation (4.20):
Find wj W (Y ) such that
aY ((wj + yj ), ) = 0 for all W (Y )

40

(4.51)

where W (Y ) and aY (, ) have their usual meaning. According to (4.4), we have


Z
Z
wj
1
wj
1
(Lij (y) + Lik (y)
)dy =
(Lmk kj im + Lmk
)dy =
aij =
|Y | Y
yk
|Y | Y
yk
Z
1
yj yi
wj
=
(Lmk
+ Lmk
)dy (4.52)
|Y | Y
yk ym
yk
Z
1
wj
yj yi
1
=
Lmk (
+
)
dy =
aY (wj + yj , yi ),
|Y | Y
yk
yk ym
|Y |
which, by using (4.51) with = wi , gives
aij = aij + 0 =

1
1
aY (wj + yj , yi ) +
aY (wj + yj , wi ) =
|Y |
|Y |
1
aY (wj + yj , wi + yi ).
=
|Y |

(4.53)

If Lik = Lki , then aY (u, v) = aY (v, u) and hence by (4.53), aij = aji , i.e., aij is
symmetric.
Now we consider aij i j for an arbitrary Rn . Let := i (wi + yi ). The positive
definiteness of aij and (4.53) yield
D E
1
aY (, )
0.
(4.54)
|Y |
yi yi

= 0, that is, if and only if c


Thus, aij i j 0 with equality if and only if y
i
(constant), or equivalently, i yi = i wi c. Here wi is Y -periodic so that equality
in (4.54) holds if and only if i = 0, thus the theorem is proved.

4.3.2

A comparison of computed coefficients with various


bounds

With the solutions of the cell problems computed, we compute the effective coefficients from section 3.4(3.75)-(3.77) by evaluating the integrals. We compare
afterwards the effective coefficients aij obtained with the different bounds that exist. Let L1 and L2 denote the specific thermal coefficients of the gas and solid parts
of the medium. Moreover, let m1 and m2 be the corresponding surface fractions of
the gas and solid respectively.
Having obtained the effective coefficients a11 , a22 and b11 , b22 , in two specified directions, from (3.75)-(3.77), we show that our estimates of the effective coefficients
satisfy the following inequalities between the harmonic and arithmetic means. We
follow now the strategy presented in [13] Chapter 2, p. 16-18 and [20] p. 29-31.
The inequality is given by
Mh aii Ma , i {1, 2},

(4.55)

where Mh denotes the harmonic mean


Mh :=

m1
L1

41

1
+

m2
L2

(4.56)

and Ma denotes the arithmetic mean


Ma := m1 L1 + m2 L2 .

(4.57)

However, there are stronger bounds, the so-called Hashin-Shtrikman bounds (see
Wendt et al.[24]), which can be obtained for L1 < L2 . The Hashin-Shtrikman
bounds are thus defined through the following relation:
m2
+
m
1
a+ := L2 + 1
+
L1 L2
a := L1 +

1
L2 L1

m1
2L1

m2
2L2

(4.58)
(4.59)

such that
a aii a+ , i {1, 2}.

(4.60)

In Table 4.1, aii for i {1, 2} represents the diagonal entries of the homogenized
L1
L2
Mh
Ma
a
a+
aii

0.01
0.3
0.019451
0.15577
0.027752
0.109482
0.02954

0.721777
1.349831
0.94211
1.037475
0.981312
0.996104
0.9855

Table 4.1: Comparison of the homogenized coefficients with known bounds


matrix for the temperature field. The last two columns of Table 4.1 contain values
obtained for different choices of the material coefficients L1 and L2 . We compare the
harmonic and arithmetic means with the effective coefficients given by aii , i {1, 2}.
We do the same using the lower a and upper a+ , Hashin-Shtrikman bounds. In
all cases, we see that the computed values satisfy the inequalities given in (4.55)
and (4.60). Also, within the limits of discretization errors, the difference between
the computed effective coefficients are relatively small if compared with the means
or the Hashin-Shtrikman bounds.

4.4

Numerical computation of the macroscopic


and microscopic solutions

Having computed the homogenized coefficients as in subsection 4.3.2, we turn now


to the aspect of numerically evaluating the macroscopic model. Since the system
of equations given in (3.72)-(3.74) is convection-dominated for P  1, i.e. the
convection term, it is not efficient to use the standard finite element method for
(3.72)-(3.74) due to stability issues that may arise. We use the streamline diffusion method, (see [25],[18] for details). In [18], the discretization of the transient
42

convection-diffusion equation using streamline diffusion is explained in detail.


Consequently, we adopt the theory explained in [18] and implement this method in
COMSOL Multiphysics (see [17] for more details) with the domain of interest as
= (0, 1)2 . In particular, we use the Streamline-upwind Petrov/Galerkin method
(SUPG). The homogenized matrix are given as follows:




b11 b12
a
11 a
12
(
aij ) =
,
(bij ) =
,
(4.61)
a
21 a
22
b21 b22
and
Pj = (P1 , 0),

(4.62)

where a
11 = a
22 as in Table 4.1 and b11 = b22 = 0.349884. Also, we have a
ij = bij = 0
for i 6= j. For the microscopic problem of section 1.2, we see that it is difficult to
find the exact solution of the original problem, hence we use the standard finite
element solution, implemented in COMSOL, to replace the exact solutions and
so that we can compare our the solution of the macroscopic problem with the
1
ideal exact solution. In COMSOL Multiphysics, we walk through the following
steps:
Step I: Choose a representative physics
Since our model equations are of transient convection-diffusion (i.e. for the transport of oxygen) and conduction-convection (i.e. for heat transport) types, we choose
in COMSOL, representative physics based on (3.72) and (3.73). Similarly, we do
the same for the microscopic equations (3.9) and (3.10)
We customize the chosen multiphysics models by modifying its coefficients suitably.
The reaction rate (3.74) for the macro models is implemented using the subdomain
weak form PDE module of COMSOL. Finally, we select the SUPG from the stabilization tab in the subdomain dialog. We see that for the micro model, 2 each
equation is prescribed appropriately on each of the periodic cells (see Fig. 4.8 (a)).

Step II: Create a geometry


The geometry of the macroscopic problem in consideration is relatively simple. We
create the domain = (0, 1)2 as a unit square in COMSOL (see Fig. 4.8).
We create the geometry of the microscopic problem is as follows: We begin with a
typical unit cell as shown in Fig. 4.3. For any > 0, we rescale the unit cell (Fig.
4.3) and generate a lattice of copies of cells that span the entire domain = (0, 1)2
as described in section 1.2. For example, a typical microscopic domain consisting
of 100 periodic cells is depicted in Fig. 4.8 (a).
Step III: Set the material properties i.e. setting all the constants that
1

A solution that can be used in place of the actual solution.


The microscopic domain has two subdomains in each of its periodic cells. we prescribe equations related to either subdomain of each periodic cell.
2

43

(a) Geometry of the microscopic problem for 10x10 cells

(b) Geometry of the macroscopic problem

Figure 4.8: 2D Geometry of the model problems


appear in the PDE
With the chosen physics in step I, we customize the settings of the parameters of
the model appropriately. First, we declare all constants and expressions that are
part of the model respectively in the constant and expression settings dialog boxes.
We also enter their names correctly in the subdomain settings dialog box under
the physics menu. The reaction rate (3.11) for the micro model is implemented
44

using the boundary weak form PDE module of COMSOL. Here, we prescribe the
equations at each of the interior circular boundaries of the geometry in Fig. 4.8 (a).
Most importantly, we ensure continuity of fluxes across the interior boundaries of
the micro model geometry Fig. 4.8 (a). Table 5.2 contains information of the list
of parameters used.
Step IV: Set the boundary conditions and initial conditions
Here, we set the boundary conditions. For (3.72) and (3.73), the boundary conditions at the top/bottom are of thermal/diffusive insulation type, as stated earlier
in chapter 2, while the side boundaries are either of Dirichlet or Neumann (i.e. for
purely convective outflow) type. The heat pulse maintained at one of the ends,
specifically at x = 1, is to imitate the impulsive onset of combustion [6]. In COMSOL, we implement this using the smoothed heaviside function, f lc2hs (see [17] and
Table 5.2). The boundary and initial conditions are respectively specified under the
boundary settings and the subdomain settings in COMSOL.
Step V: Choose an element type and mesh the geometry
For the element type, we use the quadratic Lagrange elements with a triangular
mesh as shown in Fig. 4.13. The mesh is created with a click on the mesh button.
Step VI: Choose a solver and solve for the unknowns
For the implementation here, we choose the direct linear solver (e.g. UMFPACK)
and click the solve button.
Step VII: Post-process the results to find the information that is required
To postprocess the solution, we export the data to Matlab. The idea is to be able
to compute the error estimate between the microscopic problem with decreasing
value of (i.e the scale parameter) and the macroscopic solution. We can also plot
different types of graph of the computed solutions.

4.5

Results and discussion

In this section, we discuss the results of our numerical simulations in 4.5.

4.5.1

Smoldering process in a microstructure

Smoldering phenomenon has been studied extensively and interpreted from a macroscopic standpoint [3], [6]. It can as well be studied at a microscale level, where
microscopic descriptions are given to smoldering combustion process [26]. Here, we
illustrate the numerical results of our simulation, giving insight into the microscopic
behavior of smoldering process of a paper sample.
For the plots depicted in Fig. 4.16 and Fig. 4.17, we have used the Peclet number in
the numerical simulation as a control parameter. At P e = 10 (see Fig. 4.16 (left)),
the smolder front spreads uniformly on each cell from the ignition line at x = 1,
45

(a) Mesh consists of 14261 elements.

(b) Mesh consists of 24042 elements.

Figure 4.9: 2D mesh of the model poblems.


and proceeding away from this line. We then cut the flux of oxygen to P e = 8 and
P e = 5 respectively as illustrated in Fig. 4.16 (right) and Fig. 4.17 (left). It is
observed that the front spreads towards the ignition line, and gradually generates
concentration gradients and lateral diffusion currents across adjacent cells. (see [3]
p.2817, Fig. 4.18 (right) and Fig.4.18 (left)) This observation is expected since it
describes a gradual drop in the advective dominance over diffusion.
46

Figure 4.10: Qualitative comparison of the temperature distribution of a smoldered


paper. Left: Macroscopic solution h0 (P e = 2); Right: Microscopic solution (P e =
1
2, = 20
).

Figure 4.11: 3D plot of the temperature distribution of a smoldered paper. Left: Macroscopic solution h0 (P e = 2); Right: Microscopic solution (P e = 2, =

1
20 ).

Figure 4.12: One dimensional plot of the temperature distribution of a smoldered paper.
Left: Averaged solution h0 (P e = 2) without oscillations; Right: Microscopic solution
1
(P e = 2, = 20
) with oscillations.

47

Figure 4.13: Qualitative comparison of the concentration distribution on a smoldered


paper. Left: Macroscopic solution h0 (P e = 2); Right: Microscopic solution (P e =
1
2, = 20
).

Figure 4.14: 3D plot of the concentration distribution on a smoldered paper. Left:


Macroscopic solution h0 (P e = 2); Right: Microscopic solution (P e = 2, =

1
20 ).

Figure 4.15: One dimensional plot of the concentration distribution on a smoldered


paper. Left: Averaged solution h0 (P e = 2) without oscillations; Right: Microscopic
1
solution (P e = 2, = 20
) with oscillations.

As the Peclet number is reduced further to P e = 2, we notice an increase in lateral

48

Figure 4.16: Smoldering process of a paper. Left: numerical solution for P e = 10; Right:
numerical solution for P e = 8.

Figure 4.17: Smoldering process of a paper. Left: numerical solution for P e = 5; Right:
numerical solution for P e = 2.

diffusion currents that eventually spreads across the entire


sinusoidal pattern ahead of the front.

4.5.2

reaction region with a

One-dimensional comparison of results

We illustrate here a one-dimensional analysis of the results obtained from the microscopic and macroscopic problems. These results were achieved by considering a
set of points on the two-dimensional domain and computing the solutions at these
points.
In Fig. 4.19, we see that the reaction term, which depends on the oxygen flux, is
a decreasing function. it decreases fastest at R1 close to the reaction region which
suggest that the oxygen is continually consumed at this region. At R5 , we observe
an almost steady flow with no much effect on the oxygen flux.
3

The reaction region is a region where the transport and reaction mechanisms have the most
critical influence. [26]

49

Diffusive flux (arrow) of oxygen on a temperature field at P e = 2 and 49 cells.

Concentration gradient (arrow) of oxygen on a temperature field at P e = 2 and 49 cells.

Figure 4.18: Numerical solution for P e = 2 and 49 cells.

4.5.3

Error estimates in the FE approximations in the numerical simulations

In this subsection, we present the relative error estimates describing the quantitative
behavior of solutions between the FE approximations in the multiscale numerical
homogenization approach introduced in 4.5 and the direct computation of the micro
problem. The direct computation provides an ideal exact solution. This is done by
consecutively increasing the number of periodic cells or decreasing the value of .
50

Macroscopic solution of reaction term arising in (3.74).

Microscopic solution of the reaction term arising in (3.11).

Figure 4.19: Numerical computation of the reaction term.


We will see later that as we make the value of smaller and smaller in the direct
computational approach, the computational time of the 4 CPU increases.
The error estimate we wish to adopt here is called the discrete L2 norm and is
4

An acronym for Central Processing Unit that represents the part of a computer (a microprocessor chip) that does most of the data processing

51

presented, for example, in [27]. It is defined as follows:


! 21
kuh kh =

hd

(uhi )2

(4.63)

where d is the dimension of the domain. We adapt this to our problem to approximate the continuous L2 error estimate as follows:
! 12
X
( hi )2
(4.64)
k h kh = hd
i

where is the FE approximation of the micro solution of the temperature field


and h is the FE approximation of the homogenization solution of the temperature. We obtain similar estimates for , the concentration. Obviously, we see that
the estimate depends on any chosen . As mentioned in 4.4, step VII, we export
from COMSOL the solutions of the macro problem and of the micro problem for
decreasing .
However, looking back at the solutions of the macro and the micro problems, we see
that these solutions are obtained in an 5 unstructured mesh. Therefore, to enable us
compare the solutions in the two problems, we map the solutions by linear interpolation, to a structured mesh of 500 500 data points on the (x,y)-plane. We then
implement (4.64) in Matlab for the relative error estimates. The error estimates
are summarized in table 4.2 and depicted in Fig. 4.20.

1
1
2
1
4
1
7
1
10
1
14
1
20

L2 error estimates in
3.6727
0.7886
0.2720
0.1492
0.1138
0.0940
0.0790

1
1
2
1
4
1
7
1
10
1
14
1
20

L2 error estimates in
0.3153
0.2691
0.2126
0.2053
0.2045
0.2031
0.2061

Table 4.2: Left: Error estimates in the FE approximation of k h kh for various


; Right: Error estimates in the FE approximation of k h kh for various . In
all cases, P e = 10.
In Fig. 4.20 (left), we see that the error estimates for the temperature drops exponentially with decreasing . This shows that the solution gets better with decreasing
. However, the estimate for the concentration behaves poorly. This is due to the
wrong implementation of the boundary reactive terms in the microscopic geometry. The results is also shown in Table 4.2. From Fig. 4.21 and Table 4.3, it
is now evident that the computational time of the direct implementation of the
5

By unstructured mesh, we mean the values of the solutions are obtained in a randomly generated space variables

52

Error estimates in the FE approximation of k h kh for various

Error estimates in the FE approximation of k h kh for various

Figure 4.20: Error estimates in the temperature and concentration. In all cases,
P e = 10. The plots correspond to values in Table 4.2.

1
1
2
1
4
1
7
1
10
1
14
1
20

No. of Mesh
7577
7838
8560
11879
24042
43120
52000

No. of solved d.o.f


31196
32578
35778
50412
101610
182674
221602

Computational time (s)


39.312
152.144
198.047
306.801
597.398
1399.9
2020.342

Table 4.3: Left: Comparing the computational time with decreasing . In all cases,
P e = 10

53

Figure 4.21: Comparing the computational time with decreasing . In all cases,
P e = 10. The plots correspond to values in table 4.3.
micro problem increases quadratically with decreasing . This is one important
reason why we consider the numerical multiscale homogenization method, since the
computational time is drastically reduced. For example, the computation time of
our homogenization numerical simulation is 64.466 s for P e = 10 and 59.105s for
P e = 2.

54

Chapter 5
Conclusion and Future work
In this thesis, we modeled the smoldering combustion of a paper assuming a periodic distribution of microstructures. The realized model is a system of semi-linear
reaction-diffusion equations. Due to the usual difficulties involved in solving systems
of this nature with periodic microstructures, we applied the homogenization technique that yields a macroscopic description of the problem. This is one of the main
achievements of this thesis. The numerical implementation of the homogenization
technique results in a model that drastically reduces the CPU time compared to
the direct computation of the micro model with the periodic structure. The results
of our analysis have shown that the macroscopic solutions are in agreement with
the predictions of the microscopic ones. Furthermore, we point out that the qualitative description of our numerical simulation is not a direct consequence of the
experimental observations in [3]. A similar observation, however, can be accounted
for in a microscale smoldering simulations like those conducted in [26]. These were
the subject of subsection 4.5.2.
For the purpose of further research in this subject, we present other areas of interest
which were not covered in the limited time frame: In the construction of the microstructure, we assumed, for simplicity, a structure consisting of cylindrical solids
distributed periodically over the volume of paper. However, it will be of interest,
in subsequent analysis of the techniques implemented in the course of this thesis,
to consider the effect of other geometries on the solution of the homogenized model
(see [20] more details and related questions). Another possibility is to consider a
design where the pore space is disconnected and the solid matrix connected as
shown in Fig. 5.1.

Figure 5.1: An alternative geometry with connected solid matrix


55

The proposed model can be modified to address other issues such as modeling
the velocity of the smolder front and the effect of heat losses due to radiation [6].
Stokes model of flows in porous media can also be adopted in computing the micro
velocity field of the gas. We have used a one-temperature model in this thesis. A
further implementation could be to consider a two-temperature model accounting
for the temperature fields in the solid and in the oxidizer. This model (i.e the twotemperature model) can be used to address the questions:
(Q1 ) : Under which conditions will the two-temperature model be in diffusivethermal equilibrium?
(Q2 ) : Under which conditions will the instantaneous oxidation of the gas be
achieved?
On the other hand, the reaction rate (3.11) can be reformulated to include a charred
or uncharred part of the material (see [6], for instance). Also, we see that the right
hand side of (3.11) has some regularity issue that was not immediately addressed in
the present study. We have approximated this term by considering (2.23). Finally,
in the homogenization technique implemented in this study, a formal asymptotic
expansion (3.21) was considered. The approximate solutions obtained in this case
were of order zero i.e. 0 , 0 and R0 . These gives rise to an amount of error in our
numerical solutions, since the method only approximate the solutions with the first
terms of their corresponding asymptotic expansions.
The question here is two fold:
(Q3 ) : Can we show the convergence of , , R to 0 , 0 , R0 ?
(Q4 ) : How large is the error when replacing , , R by 0 , 0 , R0 ?
A possible answer to (Q4 ) is to introduce an asymptotic expansion of order one
(see [25], [19] for example). This expansion has the following form:
1

= +

2
X
l=1

wl (y)

0
,
xl

(5.1)

where wl are the solutions of the cell problems e.g. (3.42) and (3.65). It was shown
in [25] for example, that 1 is an approximation of . The enhancement of the
error estimate of the two-scale finite element solution will then be to compute the
terms of correction h1 (x, x , ) and h1 (x, x , ). These terms will improve the
approximation of and . Thus, the relative error estimate can that can be
achieved in the L2 norm is :k (h0 + h1 )kL2 ( ) C, with C independent of
.
In addition, a detailed procedure on how to get this error estimate is given for
instance in [25]. Obviously, the enhanced approximation (5.1) recovers the small
oscillation which is not present in h0 .

56

Nomenclature
Quantity

Y
Y
s
g
Ys , Ys
Yg , Yg
Ys
Ys
g
s
T
C
R
cg,s
g,s
g,s
D
W
u = (u, 0, 0)
Q
Ta
A

T0
W (Y )

Description (units).
paper thickness [m]
3D outer domain
boundary of outer domain
planar outer domain
boundary of planar outer domain
volumetric representative unit cell
planar representative unit cell
solid part
gas part
solid parts of representative unit cell
gas parts of representative unit cell
inner boundary of 3D representative unit cell
inner boundary of planar representative unit cell
pore space
solid space
temperature[K]
volumetric mass fraction in gas part [kg/m3 ]
volumetric mass fraction of the solid product [kg/m3 ]
specific heat capacities [Jkg 1 K 1 ]
densities[kg/m3 ]
thermal conductivities[J/m s]
molecular diffusivity[m2 /s]
reaction rate with Arrhenius temperature dependence[kg/m3 s]
prescribed flow-field [m/s]
heat release [J/kg]
activation temperature[K]
pre-exponential factor[s1 ]
non-dimensional gas concentration
non-dimensional temperature
initial temperature[K]
Hilbert space of Y-periodic functions
Table 5.1: Notation.

57

L
D
v
P
Pe
Ta
A
T0
Q
C0
g
s
cg
cs
R0
Tc

hs
m
tD
tR
tG
Tb
tB
A
f
g
h
T
s
g
Leg
Les

1 102 [m]
characteristic length scale
5
2
2.5 10 [m /s]
diffusivity of oxygen [3]
[m/s]
velocity of gas
(1 0.42 )vL/D
homogenized parameter
vL/D
Peclet number
723.15[K]
activation temperature (in this study)
1.545[1/s]
pre-exponential factor (in this study)
297.15[K]
ambient temperature (in this study)
13.37[J/kg]
heat release (in this study)
3
0.23[kg/m ]
initial conc. of oxygen (in this study)
1376.3[kg/m3 ]
mass density of oxygen[3]
3
540[kg/m ]
mass density of paper[3]
923[J/kg/K]
specific heat of gas
1270[J/kg/K]
specific heat of paper[28]
1.4[kg/m3 ]
initial conc. of unburnt solid (in this study)
656.1[K]
critical temperature (in this study)
0.5[s]
small time
0.01
scale for resolution of step function
(cs s )/(cg g )
density ratio
2
(L )/D
diffusive time scale
R0 /(C0 A)
solid product time scale
1/A
gas time scale
T0 + (Q/cg )
temperature combustion product
1/((QAC0 )/((Tb T0 )g cg ))
combustion time scale
0.86
Arrhenius exponential factor

tD A/t
temperature source term
B

tD A/tG
gas source term

tD A/tR
reaction term
(460 T0 )H(/tD t, hs )/(Tb T0 )
ignition source
0.07[W/m/K]
thermal conductivity of solid [28]
0.0238[W/m/K]
thermal conductivity of gas
g /(g cg D)
Lewis number for gas
s /(s cs D)
Lewis number for solid
Table 5.2: Physical parameters

58

List of Figures
1.1
1.2
1.3
1.4

Sketch of the experimental setup [2]. . . . . . . . . . . . . . . . . .


Left: Volumetric representative cell Y ; Right: Planar representative
unit cell Y . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3D micro scale geometry of the paper sheet. . . . . . . . . . . . . .
Instability of combustion front when the flow velocity (or Peclet number) decreases from a to e. Oxygen flows downward and the smolder
front moves upward as shown by the arrow in d [6] . . . . . . . . .

2.1

Reaction/transport mechanisms within the microstructure . . . . .

3.1

(a): Deformed cell after time, t; (b) Not structurally deformed cell
after time t. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Left: Heterogeneous medium; Right: Homogeneous representation of
the perforated domain. . . . . . . . . . . . . . . . . . . . . . . . . .

3.2
4.1
4.2
4.3
4.4
4.5
4.6
4.7
4.8
4.9
4.10

Unit cell with two subdomains Yg and Ys . . . . . . . . . . . . . .


Unit cell with hole: one subdomain. . . . . . . . . . . . . . . . .
Sample geometry of the unit cell . . . . . . . . . . . . . . . . . .
Solution of cell problems, (a): w1 ( x ) and (b): w2 ( x ) with x .
Solution of cell problems, (c): w1 ( x ) and (d): w2 ( x ) with x .
Solution of cell problem (e): 1 ( x ) and (f): 2 ( x ) with x . .
Solution of cell problem (g): 1 ( x ) and (h): 2 ( x ) with x . .
2D Geometry of the model problems . . . . . . . . . . . . . . .
2D mesh of the model poblems. . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

Qualitative comparison of the temperature distribution of a smoldered paper. Left: Macroscopic solution h0 (P e = 2); Right: Microscopic solution
1
(P e = 2, = 20
). . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.11 3D plot of the temperature distribution of a smoldered paper. Left:
Macroscopic solution h0 (P e = 2); Right: Microscopic solution (P e =
1
2, = 20
). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.12 One dimensional plot of the temperature distribution of a smoldered paper. Left: Averaged solution h0 (P e = 2) without oscillations; Right:
1
Microscopic solution (P e = 2, = 20
) with oscillations. . . . . . . . .
4.13 Qualitative comparison of the concentration distribution on a smoldered
paper. Left: Macroscopic solution h0 (P e = 2); Right: Microscopic solu1
tion (P e = 2, = 20
). . . . . . . . . . . . . . . . . . . . . . . . . .

59

2
2
3

5
8
20
29
32
32
37
37
38
38
38
44
46

47

47

47

48

4.14 3D plot of the concentration distribution on a smoldered paper. Left:


Macroscopic solution h0 (P e = 2); Right: Microscopic solution (P e =
1
2, = 20
). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.15 One dimensional plot of the concentration distribution on a smoldered
paper. Left: Averaged solution h0 (P e = 2) without oscillations; Right:
1
) with oscillations. . . . . . . . .
Microscopic solution (P e = 2, = 20
4.16 Smoldering process of a paper. Left: numerical solution for P e = 10;
Right: numerical solution for P e = 8. . . . . . . . . . . . . . . . . . .
4.17 Smoldering process of a paper. Left: numerical solution for P e = 5; Right:
numerical solution for P e = 2. . . . . . . . . . . . . . . . . . . . . . .

48

48
49
49
50
51

4.18 Numerical solution for P e = 2 and 49 cells. . . . . . . . . . . . . . .


4.19 Numerical computation of the reaction term. . . . . . . . . . . . . .
4.20 Error estimates in the temperature and concentration. In all cases,
P e = 10. The plots correspond to values in Table 4.2. . . . . . . . .
4.21 Comparing the computational time with decreasing . In all cases,
P e = 10. The plots correspond to values in table 4.3. . . . . . . . .

54

5.1

55

An alternative geometry with connected solid matrix . . . . . . . . . .

60

53

List of Tables
4.1
4.2

4.3
5.1
5.2

Comparison of the homogenized coefficients with known bounds . .


Left: Error estimates in the FE approximation of k h kh for
various ; Right: Error estimates in the FE approximation of k
h kh for various . In all cases, P e = 10. . . . . . . . . . . . . . . .
Left: Comparing the computational time with decreasing . In all
cases, P e = 10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

42

Notation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Physical parameters . . . . . . . . . . . . . . . . . . . . . . . . . . .

57
58

61

52
53

Bibliography
[1] G. Rein, Computational Model of Forward and Opposed Smoldering Combustion with Improved Chemical Kinetics, Ph.D. thesis, Dept of Mechanical
Engineering, Univ. of California of Berkeley (2005).
[2] L. Kagan, G. Sivashinsky, Pattern formation in flame spread over thin solid
fuels, Combustion Theory and Modelling. 12 (2) (2008) 269 281.
[3] O. Zik, E. Moses, Fingering instability in combustion, Physical Review E.
60 (1) (1999) 114.
[4] M. Espedal, A. Fasano, A. Mikelic, Filtration in Porous Media and Industrial
Application, Springer, Berlin, 2000.
[5] T. Ohlemiller, Modeling of smoldering combustion propagation, Progress in
Energy and Combustion Science. 11 (1985) 277 310.
[6] A. Fasano, M. Mimura, M. Primicerio, Modelling a slow smoldering combustion
process, Mathematical Methods in the Applied Sciences. (2009) 111.
[7] K. Ikeda, M. Mimura, Mathematical treatment of a model for smoldering combustion, Hiroshima Math. J. 38 (2008) 349361.
[8] M. J. H. Anthonissen, Local Defect Correction Techniques:Analysis and Application to Combustion, Ph.D. thesis, Eindhoven University of Technology
(2001).
[9] S. Whitaker, The Method of Volume Averaging, Springer-Verlag New York,
1999.
[10] W. Gill, A. B. Donaldson, A. R. Shouman, The frank-kamenetskii problem
revisited. part i. boundary conditions of first kind, Combustion and Flame. 36
(1979) 217232.
[11] W. Gill, A. B. Donaldson, A. R. Shouman, The frank-kamenetskii problem
revisited, part ii: Gradient boundary conditions, Combustion and Flame. 41
(1981) 99105.
[12] G. Allaire, Shape Optimization by the Homogenization Method, Springer,
2002.

62

[13] L. E. Persson, L. Persson, N. Svanstedt, J. Wyller, The Homogenization


Method: An Introduction, Chartwell-Bratt, 1993.
[14] E. Sanchez-Palencia, Non-Homogeneous Media and Vibration Theory, Vol. 127
of Lecture Notes in Physics, 1980.
[15] A. Bensoussan, J. Lions, G. Papanicolaou, Asymptotic Analysis for Periodic Structures, Vol. 5 of Studies in mathematics and its application, NorthHolland, 1978.
[16] E. Sanchez-Palencia, A. Zaoui, Homogenization Techniques for Composite Media, Vol. 272 of Lecture Notes in Physics, Springer, 1985.
[17] http://www.comsol.com.
[18] C. Johnson, Numerical Solution of Partial Differential Equations by the Finite
Element Method, Studentlitteratur, Lund, 1987.
[19] G. Allaire, K. E. Ganaoui, Homogenization of a conductive and radiative heat
transfer problem, simulation with cast3m, Proceedings of 2005 ASME Summer
Heat Transfer Conference. (2005) 16.
[20] J. Bystrom, The Homogenization Mathod Applied to the Computation of effective Thermal Conductivities of Composite Materials, Lulea University of
Technology,S-971 87 Lulea, Sweden.
[21] J. Bystrom, Some mathematical and engineering aspects of the homogenization theory, Ph.D. thesis, Department of Mathematics, Lulea University of
Technology, SE-97187, Lulea, Sweden (2002).
[22] J. Necas, Les methodes directes dans la theorie des equations elliptiques, Ed.
de lAcad. Tech. des Sciences, Prague, 1967.
[23] H. Alt, Lineare Funktionalanalysis, Springer, Berlin, 1992.
[24] F. Wendt, H. Liebowitz, N. Perrone, Mechanics of Composite Materials, Pergamon Press, Oxford, 1970.
[25] W. Zhihua, Y. Ningning, Numerical simulation for convection-diffusion problem with periodic micro-structure, Acta Mathematica Scientia 28B (2) (2008)
236252.
[26] G. Debenest, V. Mourzenko, J. Thovert, Smouldering in fixed beds of oil shale
grains. a three-dimensional microscale numerical model, Combustion Theory
and Modelling 9 (1) (2005) 113135.
[27] W. L. Briggs, V. E. Henson, S. F. McCormick, A multigrid tutorial, Society
for Industrial and Applied Mathematics, 2000.
[28] R. Dinwiddie, M. A. White, D. L. McElroy, Thermal Conductivity 28, DEStech
Publications. Inc., 2006.

63

Appendix A
Matlab code for the discrete L2
error estimate
% Computation of the Discrete L^2 error of the solution of the homogenized
% problem
% Load data
clear all
% load into Matlab the set of solutions for temperature and concentration
% (macro)
load U.txt;
load C.txt;
%***********************************************************************
% load set of solutions computed from COMSOL for the temperatures (micro)
load
load
load
load
load
load
load

Ue2D20.txt;
Ue2D10.txt;
Ue2D14.txt;
Ue2D7.txt;
Ue2D4.txt;
Ue2D2.txt;
Ue2D1.txt;

% load set of solutions computed from COMSOL for the concentration (micro)
load Ce2D20.txt;
load Ce2D14.txt;
load Ce2D10.txt;
load Ce2D7.txt;
load Ce2D4.txt;
load Ce2D2.txt;
load Ce2D1.txt;
64

%***********************************************************************
N=length(U_h);
h=(0.002)^2; % mesh size
%***********************************************************************
% variable assignment
U_h=U(:,3);
C_h=C(:,3);
C_eps20=Ce2D20(:,3);
C_eps14=Ce2D14(:,3);
C_eps10=Ce2D10(:,3);
C_eps7=Ce2D7(:,3);
C_eps4=Ce2D4(:,3);
C_eps2=Ce2D2(:,3);
C_eps1=Ce2D1(:,3);

U_eps20=Ue2D20(:,3);
U_eps14=Ue2D14(:,3);
U_eps10=Ue2D10(:,3);
U_eps7=Ue2D7(:,3);
U_eps4=Ue2D4(:,3);
U_eps2=Ue2D2(:,3);
U_eps1=Ue2D1(:,3);
%***********************************************************************
% compute the difference Ue - Uh
Cerr20=C_eps20-C_h;
Cerr14=C_eps14-C_h;
Cerr10=C_eps10-C_h;
Cerr7=C_eps7-C_h;
Cerr4=C_eps4-C_h;
Cerr2=C_eps2-C_h;
Cerr1=C_eps1-C_h;
err20=U_eps20-U_h;
err14=U_eps14-U_h;
err10=U_eps10-U_h;
err7=U_eps7-U_h;
err4=U_eps4-U_h;
err2=U_eps2-U_h;
err1=U_eps1-U_h;
65

%***********************************************************************
% computation of the discrete L^2 Error
Cerror20=sqrt(h*(sum(Cerr20.^2)));
Cerror14=sqrt(h*(sum(Cerr14.^2)));
Cerror10=sqrt(h*(sum(Cerr10.^2)));
Cerror7=sqrt(h*(sum(Cerr7.^2)));
Cerror4=sqrt(h*(sum(Cerr4.^2)));
Cerror2=sqrt(h*(sum(Cerr2.^2)));
Cerror1=sqrt(h*(sum(Cerr1.^2)));
error20=sqrt(h*(sum(err20.^2)));
error14=sqrt(h*(sum(err14.^2)));
error10=sqrt(h*(sum(err10.^2)));
error7=sqrt(h*(sum(err7.^2)));
error4=sqrt(h*(sum(err4.^2)));
error2=sqrt(h*(sum(err2.^2)));
error1=sqrt(h*(sum(err1.^2)));
%***********************************************************************
% Output results.
figure(1),hold on
H=bar([error1,error2,error4,error7,error10, error14 ,error20]);
plot([error1,error2,error4,error7,error10, error14 ,error20],r,...
LineWidth,2);
set(get(H(1),BaseLine),LineWidth,4)
colormap summer % Change the color scheme
xlabel(increasing no of cells); ylabel(value of error);
title(Error estimates in temperature \Theta for increasing cell sizes);
hold off;
%**********************************************************************
figure(2), hold on
G=bar([Cerror1, Cerror2,Cerror4,Cerror7,Cerror10,Cerror14,Cerror20]);
plot([Cerror1, Cerror2,Cerror4,Cerror7,Cerror10,Cerror14,Cerror20],r,...
LineWidth,2);
set(get(G(1),BaseLine),LineWidth,4)
colormap summer % Change the color scheme
xlabel(increasing no of cells); ylabel(value of error);
title(Error estimates in concentration \Psi for increasing cell sizes);
hold off;
%***********************************************************************

66

Das könnte Ihnen auch gefallen