Durham E-Theses

Convection in uid and porous media

Carr, Magda

How to cite:

Carr, Magda (2003) Convection in uid and porous media, Durham theses, Durham University. Available
at Durham E-Theses Online: [Link]

Use policy

The full-text may be used and/or reproduced, and given to third parties in any format or medium, without prior permission or
charge, for personal research or study, educational, or not-for-prot purposes provided that:
• a full bibliographic reference is made to the original source
• a link is made to the metadata record in Durham E-Theses
• the full-text is not changed in any way
The full-text must not be sold in any format or medium without the formal permission of the copyright holders.
Please consult the full Durham E-Theses policy for further details.

Academic Support Oce, Durham University, University Oce, Old Elvet, Durham DH1 3HP
e-mail: [Link]@[Link] Tel: +44 0191 334 6107
[Link]
 i

Convection in Fluid and Porous Media

Magda Carr

A copyright of this thesis rests

with the author. No quotation
from it should be published
without his prior written consent
and information derived from it
should be acknowledged.

A thesis presented for the degree of

Doctor of Philosophy

Numerical Analysis
Department of Mathematical Sciences
University of Durham
England
July 2003
N O V 2003
 Dedicated to
My family and the memory of B J Goodwin.
 Convection in Fluid and Porous Media

Magda Carr

Submitted for the degree of Doctor of Philosophy

July 2003

Abstract

The subject of convection in fluid and porous media is investigated. Particular attention

is paid to penetrative convection. The first two chapters are devoted to penetrative con-

vection when fluid overlies and saturates a porous medium. Penetrative convection is

described by a quadratic equation of state in the first instance and via internal heating

in the second. Linear instability analyses are performed in both cases. A surprising and

striking array of streamlines are presented at the onset of convection. The streamlines

exhibit novel behaviour when physical parameters of the problem are varied.

Penetrative convection in a horizontally isotropic porous layer is discussed next. Again

penetrative convection is described by a quadratic equation of state and internal heating.

The internally heated model is dealt with primarily as it yields a global nonlinear stability

bound. The two models are shown to be mathematically adjoint and the nonlinear stabil-

ity results compared with previously published linear ones. Good agreement between the

two is seen.

The effect of convection on the evolution of under-ice meltponds is investigated next.

Linear and nonlinear analyses are employed to yield instability and global stability results

respectively. Discrepancy between the two is found and the region of possible subcritical

instabilities is presented.

Finally convection in a porous medium is investigated via a cubic equation of state.

It is found that unconditional nonlinear stability results can be established if Forchheimer

theory is introduced. The results are compared to previously published linear ones and it

is shown that the linear theory essentially captures the physics involved.
Declaration

The work in this thesis is based on research carried out in the Numerical Analysis Group,

Department of Mathematical Sciences, University of Durham, England. No part of this

thesis has been submitted elsewhere for any other degree or qualification and it is all my

own work with the exception of chapters 2 & 4. Chapter 2 was written in collaboration

with Prof B Straughan of the University of Durham, England, and chapter 4 with Mr S de

Putter of the Technical University of Eindhoven, The Netherlands. In both instances the

collaborating author contributed 50% of the work. The content of chapters 2, 4, 5, & 6

are published in [11-14].

Copyright © 2003 by Magda Carr.

"The copyright of this thesis rests with the author. No quotations from it should be pub-

lished without the author's prior written consent and information derived from it should

be acknowledged".

iv
Acknowledgements

It is with great pleasure and deepest gratitude that I thank my supervisor Prof Brian

Straughan. The guidance, commitment and time which he has freely given me and the

generosity and wisdom with which he has supervised me are immeasurable. I would also

like to thank Prof Straughan for enhancing my knowledge of footballers from the 60's and

70's and being the first Mackem to admit what a good player Alan Shearer is.

I thank everyone in the Department of Mathematical Sciences at the University of

Durham for their help and assistance. In particular I would like to thank Dr J V Armitage

for his encouragement, advice and friendship. I pay a special mention to all members of

the Numerical Analysis Group particularly, Dr JP Coleman for his time and support, and

Imran M for saving myself and others precious time by preparing this IJIgX template.

Special thank also to Bamba (BJ Goodwin R.I.P) and Mr E M Morrison for fostering

my love of mathematics in the early days. Thank to all of my friends, near and far.

Last but not least I would like to thank my family, Mom, Dad, Lucy and Thomas. Your

unconditional love and support was and is my inspiration and rock, thank you.

This work was supported by a research studentship of the Engineering and Physical

Sciences Research Council.

v
Contents

Abstract Hi

Declaration iv

Acknowledgements v

1 Introduction 1

2 Penetrative Convection in a Superposed Porous-Fluid Layer via a Quadratic

Density Law 9
2.1 Governing Equations 13

2.2 Perturbation Equations 16

2.3 Linearised Instability Theory 17

2.4 Boundary Conditions 19

2.5 Numerical Method 21

2
2.5.1 The Chebyshev tau D method 22

2.5.2 The Chebyshev tau D method 26

2.6 Computational Results 27

2.7 Numerical Findings 29

2.7.1 Variation of the neutral curves with d, T v and 6 29

2.7.2 Streamline and isotherm patterns 33

2.7.3 Effect of permeability 42

2.8 Concluding Remark 42

3 Penetrative Convection in a Superposed Porous-Fluid Layer via Internal Heat-

ing 44
vi
Contents vii

3.1 Governing Equations 46

3.2 Perturbation Equations 48

3.3 Linearised Instability Theory and Boundary Conditions 50

2
3.4 The Chebyshev tau D Method 52

3.5 Discussion of the Model and Numerical Results 53

3.5.1 Stability characteristics of Q 56

3.5.2 Variation of the streamlines with Q 58

3.5.3 Stability characteristics of Q m 60

3.5.4 Variation of the streamlines with Q m 64

3.6 Comparison with the Quadratic Density Model 67

3.7 Concluding Remarks 74

4 Penetrative Convection in a Horizontally Isotropic Porous Layer 76

4.1 Internal Heat Sink Model 77

4.2 Linearised Instability Theory and The Principle of Exchange of Stabilities 81

4.3 Global Nonlinear Stability via the Energy Method 84

4.4 Comparison of the Heat Sink Model with a Quadratic Density M o d e l . . . 87

4.5 Numerical Results and Conclusions 89

5 A Model for Convection in the Evolution of Under-ice Melt Ponds 94

5.1 Governing Equations 96

5.2 Linearised Instability Theory 99

5.3 Global Nonlinear Stability via the Weighted Energy Method 101

5.4 Numerical Discussion, Results and Conclusions 106

6 Unconditional Nonlinear Stability for Temperature Dependent Density Flow

in a Porous Medium 110
6.1 Convection in a Porous Medium with a Cubic Density Law 112

6.2 Unconditional Nonlinear Stability 114

6.3 Numerical Solution of the Euler-Lagrange Equations 120

6.4 Numerical Results and Conclusions 122

7 Conclusion 126
Contents viii

Appendix 129

A Numerical Methods 129

A.l The Chebyshev Tau Method 129

A. 2 The Compound Matrix Method 132

B Inequalities 135
B. l Poincare's Inequality 135

B.2 Young's Inequality 135

Bibliography 136
List of Figures

2.1 Schematic diagram of the governing system 13

2.2 Neutral instability curves, Ra m against a . m Tu = 4°C, 6 = 0.002, d varies as

shown on graph 30

2.3 Neutral instability curves, Ra m against a . m 6 = 0.002, d = 0.4, T[/(°C) varies

as shown on graph 31

2.4 Neutral instability curves, Ra m against a . m Tu = 8.2°C, d — 1, S varies as

shown on graph 31

2.5 Streamline plot, T v = 5°C, T = 3.18°C, Ra

0 m = 2.458, a m = 9, d = 0.4,

6 = 0.002 36

2.6 Streamline plot, T v = 5.3°C, T = 3.37°C, Ra0 m = 11.331, a m = 12, d = 0.4,

6 = 0.002 36

2.7 Streamline plot, T[/ = 5.4°C, T = 3.4°C, Ra 0 m = 15.042, a m = 2, d = 0.4,

8 = 0.002 36

2.8 Streamline plot, = 5.8°C, T = 3.69°C, Ra 0 m = 19.620, a m = 2, J = 0.4,

* = 0.002 36

2.9 Streamline plot, Tu = 6°C, T 0 = 3.8°C, i i a m = 21.833, a m = 2, rf = 0.4,

<J = 0.002 37

2.10 Streamline plot, T v = 8°C, T 0 = 3.29°C, Ra m = 5.723, a m = 10.4, d = 1,

J = 0.002 37

2.11 Isotherm, T v = 8°C, T = 3.29°C, i ? a

0 m = 5.723, a m = 10.4, d = 1, 5 = 0.002. 37

2.12 Streamline plot, T v = 8.1°C, T = 3.3°C, Ra 0 m = 8.065, a m = 11.2, d = 1,

J = 0.002 38

2.13 Streamline plot, T v = 8.2375°C, T 0 = 3.39°C, Ra m = 13.296, a m = 12.2,

d = l , 5 = 0.002 38

ix
List of Figures x

2.14 Streamline plot, T v = 8.25°C, T = 3.4°C, Ra 0 m = 13.875, a m = 2, d = 1,

5 = 0.002. 38

2.15 Streamline plot, T v = 8.4°C, T = 3.46°C, Ra

0 m = 15.258, a m = 2, d = 1,

5 = 0.002 38

2.16 Streamline plot, T v = 9.6°C, T = 3.95°C, Ra

0 m = 24.503, a m = 2, d = 1,

5 = 0.002 39

2.17 Streamline plot, = 13°C, T = 3.37°C, Ra

0 m = 10.470, a m = 11.6, d = 2,

5 = 0.002 40

2.18 W eigenfunction, T[/ = 13°C, T 0 = 3.37°C, Ra m = 10.470, a m = 11.6,

d = 2, J = 0.002 40

2.19 Streamline plot, T v = 13.2°C, T = 3.42°C, Ra

0 m = 13.686, a m = 2, d = 2,

S = 0.002 40

2.20 W eigenfunction, T v = 13.2°C, T = 3.42°C, i ? a 0 m = 13.686, a m = 2, d = 2,

5 = 0.002 40

2.21 Streamline plot, T v = 8.2°C, T - 3.38°C, i ? a

0 m = 11.561, a m = 12, d = 1,

(5 = 0.002 43

2.22 Streamline plot, T v = 8.2°C, T = 3.38°C, Ra

0 m = 13.614, a m = 2, d - 1,

J = 0.003 43

2.23 Streamline plot, T v = 8.2°C, T = 3.38°C, Ra

0 m = 13.331, a m = 1.8, d = 1,

S = 0.01 43

3.1 Steady state temperature profile, depth z against temperature T . z tp < 0, zj£ <

-dm 54

3.2 Steady state temperature profile, depth z against temperature T , z t p < 0, —d m <

< 0, T > 0
0 55

3.3 Steady state temperature profile, depth z against temperature T . zt p < 0, —d < m

zg} < 0, T < 0

0 55

3.4 Steady state temperature profile, depth z against temperature T , z tp > 0, zj£ >

0, T < 0
0 56

3.5 Neutral instability curves, Ra m against a . m T\j = 9°C, d = 1, e m = -0.0728,

e varies as shown on graph 58

List of Figures xi

3.6 Streamline plot from the Q model, T v = 9°C, d = 1, 8 = 0.002, e = -0.05,

em = -0.0728, Ra m = 29.155, a m = 2.2 59

3.7 Streamline plot from the Q model, T v = 9°C,rf= 1, 8 = 0.002, e = -0.0275,

e m = -0.0728, Ra m = 14.112, a m = 2 59

3.8 Streamline plot from the Q model, T v = 9°C, d = 1, 8 = 0.002, e = -0.025,

e m = -0.0728, i ? a m = 8.839, a m = 12 59

3.9 Streamline plot from the Q model, T v = 9°C, d = 1, 8 = 0.002, e = -0.015,

e m = -0.0728, Ra m = 0.713, a m = 8.5 59

3.10 Neutral instability curves, Ra m against a . m Tu = 9°C, d = 1, e = —0.0357,

e varies as shown on graph

m 61

3.11 Steady state temperature profiles, depth z against temperature T. Tu = 9°C,

d = 1, e = —0.0357, e m varies as shown on graph 62

3.12 Neutral instability curves, Ra m against a . m Tu = 15°C, d = 1, e = —0.1, e m

varies as shown on graph 62

3.13 Steady state temperature profiles, depth z against temperature T. Tu — 15°C,

d = 1, e = —0.1, e varies as shown on graph.

m 63

3.14 Streamline plot from the Q model, Tu = 9°C, d = 1, 8 = 0.002, e = -0.0357,

e m = -0.25, Ra m = 1.317, a m = 10 65

3.15 Streamline plot from the Q model, T v = 9°C, J = 1, 8 = 0.002, e = -0.0357,

£m = -0.15, Ra m = 5.534, a m = 12 65

3.16 Streamline plot from the Q model, T v = 9°C, d = 1, 8 = 0.002, e = -0.0357,

e m = - 0 . 1 , Ram = 15.262, a m = 2 65

3.17 Streamline plot from the Q model, Tu = 9°C, d = 1, 8 = 0.002, e = -0.0357,

e OT = -0.01, Ra m = 46.107, a m = 2.4 65

3.18 Streamline plot from the Q model, T v = 15°C, d = 1, 8 = 0.002, e = -0.1,

e m = -0.5, i ? a m - 1352.493, a TO = 8.5 67

3.19 Streamline plot from the Q model, T v = 15°C, d = 1, 8 = 0.002, e = -0.1,

e m = -0.01, i ? a m = 57.061, a m = 2.6 67

3.20 W eigenfunctions for the adjoint models, Tu = 9°C, d = 1 71

2
3.21 Streamline plot from the T model, T v = 9°C, d = 1, 8 = 0.002, i ? a m =

19.902, a m = 2 72
List of Figures xii

3.22 Streamline plot from the Q model, T = 9°C, d = 1, 8 = 0.002, e,e take
v m

adjoint values, Ra = 20.233, a = 2

m m 72

2
3.23 Streamline plot from the T model, T v = 9°C, d = 1, 5 = 0.01, # a m =

18.534, am =2 73
2
3.24 Streamline plot from the T model, T[/ = 9°C, d=l,6 = 0.1, Ra m = 18.928,

am = 1.8 73

4.1 Schematic diagram of the governing system 78

4.2 Profile of W(z) normalised over the spatial layer, £ = 1, T varies as shown on
u

graph 92

4.3 Profile of ®(z) normalised over the spatial layer, £ = 1, T varies as shown on
u

graph 92

5.1 Schematic diagram of the governing system 96

5.2 A plot of the linear instability curves for R 2 2

against i ? . r — 100, a 2
=

9.705432, Pr = 13.4, - 7 j varies as shown on graph 107

2
5.3 A plot of the linear instability curves and the nonlinear stability curve for R
2 2
against R . -T
C t = - 1 . 9 ° C , a = 9.705432, Pr = 13.4, r varies as shown on

graph 108

6.1 Schematic diagram of the governing system. 112

List of Tables

2.1 The range and approximate value of T\j (°C) in which a switch in dominance

occurs for d varying, 8 = 0.002 32

2.2 The number of convection cells in the two layer system, d varying, 8 = 0.002,

Tu&Tff" 41

3.1 The critical porous Rayleigh number and wave number when the two models are

adjoint 71

4.1 Critical parameters of linear instability and nonlinear stability with T\ = 4°C. . 90

4.2 Critical parameters of linear instability and nonlinear stability with £ = 1. . . . 91

4.3 The interval in which T\ yields penetrative convection 93

6.1 A comparison of the critical parameters of linear instability and nonlinear stabil-

ity theory, /j, is the best coupling parameter. 125

xiii
Chapter 1

Introduction

The objective of this thesis is to investigate thermal convection in fluid and porous media

via stability and instability analyses. The aim of this chapter is to introduce the underlying

theory and provide a brief outline of the work done. We begin by explaining exactly what

is meant by stability and instability in the context of this work. The governing equations

exhibit stationary solutions which allow some simple flow patterns. If a system is in a

stationary state this simply means that none of the variables describing it are a function

of time. The main purpose of this thesis is to investigate what happens to a system which

is in a stationary state when subjected to some disturbance. If the disturbance gradually

dies and the stationary state is unaffected then the system is said to be stable. On the other

hand if the disturbance grows and the system is affected such that it never reverts back to

its initial state then it is said to be unstable. A system is only considered stable if it can be

shown to be stable with respect to every disturbance to which it may be subject. On the

other hand, a system is considered unstable if it can be shown to be unstable with respect

to any one disturbance.

To clarify this concept and introduce some mathematical notation we begin with a

simple illustrative example. Suppose u is a solution of the linear diffusion equation with a

linear source term and subject to boundary and initial conditions as follows, (cf. Straughan

1
Chapter 1. Introduction 2

[94])

2
du du
2 + au are ( 0 , 1 ) , t>0,
dt ox
u(0,t) = u(l,t) 0, (1.1)

u(x, 0) = u (x).
0

Here t and x are time and spatial point respectively, and a is some real, positive constant.

The stationary solution, u = 0, is clearly a solution of (1.1). To investigate the stability of

this solution we introduce some perturbation (disturbance) to it. If all perturbations decay

to zero with time t, then the solution is said to be stable. If just one disturbance can be

found that grows in amplitude with time, then the solution is unstable.

Perturbations are introduced by a normal mode representation of the form

u(x, t) = e',<rt+ikx (1.2)

where k is a real number and a (constant, possibly complex) is the growth rate. This is

a suitable representation since ( l . l ) i is linear, and (1.2) allows the instability of the zero

solution to be studied against any periodic disturbance in x. A full justification of the

normal mode representation can be found in detail in Chandrasekhar [15], p.3-6. The sign

of Re(a) determines whether the zero solution is stable or not. If Re(a) > 0 then the

perturbation will grow exponentially with time and clearly there will be instability. On

the other hand if Re(a) < 0 the normal modes will decay with time and there is stability.

Imposing boundary conditions (1.1)2, the normal mode representation (1.2), reduces

to the form

u(x, t) = e^sinkx k = mr, n = ±l,±2,....

Substituting this representation into ( l . l ) i yields

2 2 2 2
a = -k + a, where k = n iv , n = ±l,±2,

2 2 2
Therefore, a € R and a < 0 < $ a < k < $ a < k^ n = ir . So for a < ir we say
2
there is linear stability. On the other hand if a > ir then a > 0 for n = 1 and there
2
is linear instability. Hence a = ir is the linear instability-stability boundary. In general

the locus which separates instability and stability is known as the neutral stability curve.

The neutral stability curve represents a marginal state between stability and instability.
Chapter 1. Introduction 3

This marginal state exhibits one of two kinds of motion, stationary (cr G M) or oscillatory

(a G C) . If the motion is stationary, perturbations grow (or are damped) aperiodically

whereas if it is oscillatory they grow (or are damped) periodically. If instability sets in

as stationary motion the principle of exchange of stabilities is said to hold. On the other

hand if instability sets in as oscillatory motion then the system is said to be subject to

overstability. One of the aims of this thesis is to generate neutral stability curves and

investigate their state.

In the work that follows the governing systems under consideration are all nonlinear.

While a linear instability analysis like that described above yields useful information for

such systems, it is important that nonlinear results can be established as well. The energy

method is an excellent means of doing this, cf. Straughan [94]. It can provide precise

and practical information about the stability of a nonlinear system. To illustrate how the

energy method may be implemented the linear system (1.1) is taken as a simple example

again. It is noted at this point that the example is simply illustrative and the method applies

to a nonlinear system in an analogous way though ad hoc techniques may be necessary to

deal with the nonlinearities (see chapters 4-6).

To begin, an energy E(t) is defined by

2
E(t) = IhW ,
2 2
where || • || denotes the norm on L (0,1) (where L (0,1) is the space of square integrable

functions on (0,1)), i.e.,

2 2
||u|| = / u dx.
Jo
Then differentiating E(t) and using ( l . l ) i yields

dE 1 f l
du2
J f 1
du J f 1
«9 u2
f 1
2 j
= o / -^rdx = / u—dx = I M——ote + a I u'dx. (1.3)
~dt 2 Jo dt J0 dt Jo dx 2
J 0

Integrating by parts and noting the boundary conditions, we note that

T 1 /•! / o \ 2
1 |2
f u — d x - 2
\u—1 - C ( — \ dx - -\\ux
J0 dx [ dx\ 0 Jo \dx)
Chapter 1. Introduction 4

where u x = du/dx. Hence (1.3) can be written as

d 2 2
f t = - | K | | + a|H| ,

|2
1 || | U

-a\\u 2
x
a im " ' ' T

2
< - |K|| f--maxJMt j
a ( 1 . 4 )
n u
V° \\ x\\

where H is the space of admissible functions,

2
U = {u e C ( 0 , l ) | u = 0 when x = 0,l},

2
and C ( 0 , 1 ) is the space of twice continuously differentiable functions on (0,1). Suppose

now that R E is defined by

2
i H|
RE = max
« K|| ' 2

then (1.4) becomes

dE „ l l 2 /l 1
— < - o il, - - —
at \a RE

Let c = 1/a - I/RE, then if a < RE, C > 0, and Poincare's inequality (see Appendix

B . l ) can be used to deduce that

2 2 2
^ < -ac7r ||«|| = 11 11
-2n acE.
dt -

This implies
2 7 r 2 a c t
4(e £ ) < 0,
dt

which integrates up to yield

2n2act
E(t) < e- E(0).

2
Recall that E(t) = |||w|| ; therefore

|| ||2
U < E -2^ A C < || U ( ) ||2 ^ Q A S

Hence it has been shown that provided a < RE, \\U\\ —>• 0 at least exponentially as t —> oo,

and so the zero solution to (1.1) is stable.

2 2
To find R E let h = \\u\\ ,1 2 = \\u \\ x and recall that

1 h
— = max —.
RE H h
Chapter 1. Introduction

Suppose u is a maximising solution and consider solutions of the form u + er\ where e is
2
some constant and 77 is an arbitrary C ( 0 , 1 ) function such that 77(0) = 77(1) = 0. Clearly

the maximum occurs at e = 0 so,

•d h
1 dh 11 dh 1 5_h 1
( X \ 0,
Je e=0 T ~ck
2 J f de _ 6=0 h RE h
e=0

where 5 stands for the derivative with respect to e evaluated at e = 0. Hence it can be

deduced that

R Sh
E - 5I 2 = 0, (1.5)

where

5h = — [ (u + e r f f d x = 2 f wqdx,
dt J0 J o ^0 c=

5I 2 = / (u x + er) ) dx
x
2
= 2 / u r) dx.
x x
de JO J =0£ >/o

Therefore (1.5) yields

/ (R ur)
E - u r) )dx
x x = 0,
J0
which can be integrated by parts to give

/ r)(R u E + u )da; = 0. XI

Jo

Apart from the continuity and boundary conditions it must satisfy, 77 is arbitrary hence it

can be concluded that,

2
du
2
+ Ru E = 0, u ( 0 ) = u ( l ) = 0. (1.6)
dx

This is known as the Euler-Lagrange equation for the example at hand. It yields an eigen-

value problem for R.

E

In the work that follows, the Chebyshev tau method (cf. Fox [32] and Orszag [75]), or

the compound matrix method (cf. Drazin & Reid [24] and Ng & Reid [65-67]) are used

in the linear and nonlinear analyses to yield generalised eigenvalue problems of the form,

Ax = cf-Bx,

where A and B are matrices and x is some vector, all of which depend upon the system

under consideration. Straughan & Walker [95] applied these two techniques to linear
Chapter 1. Introduction 6

and nonlinear stability problems for convection in porous media. Their paper provides an

excellent summary of the two aforementioned methods. They compare the techniques and

highlight the advantages and disadvantages of both when investigating stability problems.

Details of the numerical methods as employed here can be found in the relevant chapters

and appendix A.

The general solution of (1.6) is given by

u = A sin ^REX + B cos \JREX,

where A, B are constants. The boundary condition it(0) = 0 implies that B = 0, hence

u = A sin \JR~EX.

For a non arbitrary solution suppose A ^ 0, then = 0 implies that

y/R~E = wr, where n = ±l,±2,

2 2 2
i.e. R E = T T , 4TT , 9TT , . . . .

2
Recall that stability requires a < R . E The minimum value of RE is 7r SO as expected,
2
a < IT , is the stability bound of the zero solution to (1.1).

In general it is found that the energy method yields some critical threshold below

which everything is stable and a linear instability analysis some bound above which ev-

erything is unstable. In the example above the energy method and normal mode analysis

yield the same critical instability-stability boundary. This is due to the fact that the dif-

ferential equation under consideration (1.1) i is linear. If nonlinearities are introduced

then it is found that in general a nonlinear stability analysis will yield a different critical

threshold to one obtained by a linear analysis. A linear analysis assumes that any per-

turbation is small and so neglects terms of quadratic order and higher, hence discrepancy

between linear instability and nonlinear stability results can occur. One of the aims of this

thesis, and energy theory in general, is to try and arrange that the two thresholds are as

close together as possible and so reduce the possibility of subcritical instabilities which

may occur below the linear instability bound and above the nonlinear stability threshold.

Many techniques have been used to do this and the theory applied to a range of important

physical problems. There are many references in the literature and significant advance-

ment has been made in the last 50 years. For reference we mention some recent papers
Chapter 1. Introduction 7

and the citations therein namely, Basurto & Lombardo [5], Lombardo et al. [49,50], and

Mulone [64]. Many more references and a thorough discussion of the energy method and

its subtleties can be found in Straughan [94].

Having established the underlying theory, a brief outline of each chapter is now given.

In chapter 2 an accurate numerical calculation is presented for the onset of thermal con-

vection in a two layer system in which water overlies and saturates a porous layer. The

physical picture is such that water at the density maximum may be encompassed. To ac-

count for the fact that the density may have a maximum, an equation of state is adopted

which expresses the density in the buoyancy force as a quadratic function of temperature.

It is found that the onset of convection may have a bi-modal nature in which convection

can be dominated by the porous medium or by the fluid depending on parameters which

appear in the problem. The important parameters in this instance turn out to be the ratio of

fluid to porous medium depth, the upper surface temperature, and a parameter represent-

ing non-dimensional permeability of the porous matrix. A surprising array of streamline

patterns are found and presented at the onset of convection as one varies the appropriate

parameters.

In chapter 3 a similar system to that described in chapter 2 is investigated. In this

case however an alternative model is employed to describe thermal convection. Instead

of employing a quadratic equation of state, a linear one is used and internal heat sinks

are introduced in both layers to simulate convection. The internal sinks are found to have

a dramatic effect on both the onset of convection and the size of the ensuing convection

cells. It is found that the heat sink Q, in the fluid layer has a destabilising effect on the

system whereas the heat sink Q , m in the porous medium can be either stabilising or desta-

bilising depending on the steady state temperature profile and the strength of Q. A range

of streamlines are presented that exhibit novel behaviour when Q and Q m are varied. The

model is compared with that given in chapter 2 and when the two are mathematically

adjoint they are shown to yield the same critical instability threshold but different eigen-

functions. It is also found that the initiating cell is not necessarily the strongest one. This

curious behaviour is explained and illustrated with a range of streamline plots.

In chapter 4 convection in a horizontally isotropic porous layer is investigated primar-

ily using an internal heat sink model and alternatively, a quadratic density temperature
Chapter 1. Introduction 8

law. Employing the heat sink model, it is shown that the temporal growth rate for the

linearised system is real. A nonlinear energy analysis is also presented, yielding a global

stability threshold i.e. a stability bound which is independent of the intial state of the

system and without constraint. The heat sink model is compared to the quadratic density

model and it is found that the linearised systems are adjoint, implying that the instability

boundaries which can be derived from the two models are the same.

In chapter 5 a model for convection in the evolution of under-ice melt ponds is pre-

sented. The system exhibits two competing effects namely, a temperature gradient which

is destabilising and a salt gradient which is stabilising. Density is assumed to have

a dependence quadratic in temperature and linear in concentration. A linear instabil-

ity analysis and a nonlinear stability analysis are both performed. The standard energy

method does not yield unconditional stability so a weighted energy analysis is employed

to achieve global results. The global stability bound is found to be independent of the salt

field and a presentation of the region of possible subcritical instabilities is given.

In chapter 6 a nonlinear stability analysis of thermal convection in a saturated porous

medium is presented again. In this case however, the medium is assumed isotropic,

the density has a cubic temperature dependence and the equations of flow in the porous

medium incorporate a Forchheimer drag term. Unconditional (global) stability is estab-

3 A 2
lished using L and L norms and it is shown that L theory is insufficient to obtain similar

results.

Finally in chapter 7 some concluding remarks are made, the implications of this thesis

discussed and some suggestions for future work are given.

Chapter 2

Penetrative Convection in a Superposed

Porous-Fluid Layer via a Quadratic
Density Law

The object of this chapter is to provide an accurate numerical calculation for the instability

that arises due to penetrative convection when water overlies and saturates a layer of

porous material. Penetrative convection refers to convective motion which begins in an

unstable layer and penetrates into otherwise stable layer(s), cf. Veronis [102]. Penetrative

convection has many applications in the real world. For example penetrative convection

is believed to play an integral part in the formation of patterned ground, cf. George et

al. [33]. Patterned ground is the natural arrangement of stones into regular shapes such

as circles, stripes or polygons. Typically, it consists of stone borders with soil centres

forming regular hexagonal stone nets. The regularity of the stone nets can be remarkable

and quite striking, for example see the photographs given in Krantz et al. [44]. The size

and shape of patterned ground tends to be the same at any one site but can vary in width

for a single stone net, from about 4cm to 4m depending on location cf. Gleason [37]. In

general, patterned ground tends to form above the timberline in the high mountains of the

temperature zone, in remote locations beyond the tree line in the Arctic and Antarctic,

and even under water in shallow Alpine lakes.

As well as being inherently fascinating, patterned ground is of significant geophysical

interest as it can disclose clues to various climate conditions. For example relict patterns

9
Chapter 2. Penetrative Convection in a Superposed Porous-Fluid Layer via a
Quadratic Density Law 10
can give clues to prehistoric climates, active patterns can be used to monitor mankind's
impact on polar regions, and patterned ground on Mars [55-57] can be compared to that
seen on Earth to further our understanding of climatic conditions in space. The formation
and evolution of patterned ground has been discussed in geological literature for more
than a century and is still a subject of debate, cf. Kessler & Werner [43] and Mann [58].

Gleason [37] and Krantz et al. [44] reported findings of hexagonal stone patterns on

the bed of shallow alpine lakes in Wyoming and Colorado. McKay & Straughan [53]

developed a porous-fluid layer model to study such under water patterned ground forma-

tion. They used a shooting method and assumed stationary convection. In this chapter

we essentially use the model of McKay & Straughan but employ a more accurate nu-

merical technique, namely the Chebyshev tau method. This allows for the possibility of

oscillatory convection and yields many new streamline patterns.

The fundamental model for convection in a porous-fluid layer system is due to Nield

[68], with an important extension in Nield [69]. Chen & Chen [17] also analysed the

problem of thermal convection in a superposed two layer system in which a layer of

fluid overlies and saturates a porous medium. The layer is heated from below with the

bottom of the porous medium and the upper surface of the fluid both assumed fixed. Chen

& Chen made the significant discovery that the linear instability neutral curves for the

onset of convective motion, i.e., the Rayleigh number against wave number curves for

which o = 0, may be bi-modal in the sense that they possess two local maxima. These

maxima correspond to a parameter range where convection is initiated in the fluid layer,

or alternatively in the porous layer. The cell sizes are different depending on whether

convection is dominated by the fluid or the porous layer. Further work on the superposed

porous-fluid layer problem, also due to Chen & Chen [17], is the study of the onset of

instability when both a thermal and a salt field are present. These writers employed an

orthonormal shooting method to numerically calculate the eigenvalues.

2
In recent work Straughan [92,93] has developed a Chebyshev tau D method appli-

cable to the porous-fluid layer thermal convection problem. Straughan [92] focused on

the problem where the upper surface is exposed to the atmosphere and surface tension is

important. Whereas in [93] the effect of the various fluid and porous matrix properties on

the instability criteria were investigated. It was discovered that porous permeability may
Chapter 2. Penetrative Convection in a Superposed Porous-Fluid Layer via a
Quadratic Density Law 11
have a strong effect and the interface condition between the fluid and the porous medium

was crucial.

Since the eigenvalues occurring in porous-fluid convection problems are highly sen-

sitive to changes in parameters, a very accurate numerical method such as a Chebyshev

tau one is necessary. In Straughan [92,93] it was found that all eigenvalues were real,

but in the present work we find that the spectrum also contains complex conjugate pairs.

Spurious eigenvalues may present a problem in hydrodynamical stability studies, see for

example the comprehensive account by Dawkins et al. [22]. We have been careful to

exclude such terms and this point is revisited in section 2.6. Further references to work

on the porous-fluid convection problem can be found in Nield [70-72], and Nield & Be-

jan [73].

In this chapter we consider water saturating a porous medium with the water forming

a layer of finite depth above. The lower surface at the bottom of the porous layer is

maintained at a constant temperature of 0°C while the upper fluid layer is kept at a fixed

temperature greater than this. When the upper temperature, 7V, is greater than 4°C we

have a situation in which the water in the region 0 — 4° C is in a potentially gravitationally

unstable state while that above is stably stratified. This is due to the fact that the density

of water is a maximum at approximately 4°C. Therefore convective instability may arise

in the lower part of the layer and penetrate into the stable upper part. Thus the model is

capable of describing penetrative convection. A careful numerical study of this instability

is the goal of this chapter. This is the first time a highly accurate numerical technique has

been applied to the problem, and we find that an array of complex fluid instability patterns

may occur.

As well as patterned ground formation under water, penetrative convection in a porous-

fluid layer system has many more important applications. For example there is speculation

as to whether convection patterns may have a bearing on accelerated melting of the Arctic

ice sheet, cf. Bogorodskii & Nagumyi [9]. This is investigated in more detail in chap-

ter 5 and the implications discussed at length in chapter 7. In addition there are many

other applications for the porous-fluid flow system including contaminant transport due

to flow of water under the Earth's surface, cf. Allen [2], Allen & Khosravani [3], Cur-

ran & Allen [20], El-Habel et al. [28], Ewing [29], Ewing & Weekes [30], flow of oil
Chapter 2. Penetrative Convection in a Superposed Porous-Fluid Layer via a
Quadratic Density Law 12
in underground reservoirs, cf. Allen [1], Ewing [29], bio-remediation of contaminated
ground, cf. Suchomel et al. [97,98], and hydrothermal synthesis in the growth of crys-
talline materials, cf. Chen et al. [18]. Penetrative convection in a single layer is also a
subject receiving recent attention with very important applications. For example Zhang
& Schubert [107,108] reported a penetrative convection effect which may be important
in planetary and stellar atmospheres and Krishnamurti [45] and Larson [46,47] reported
penetrative convection effects when radiation is added to a standard model for cumulus
convection.

Penetrative convection can be described in several different ways, at least five of which

are discussed in detail in Straughan [91]. In this thesis we concentrate on two of the most

widely employed models, namely a quadratic equation of state and internal heating. In

chapter 6 the effect of a cubic equation of state is also investigated. The strengths and

weaknesses of the various models are discussed in the relevant chapters and a comparison

between models is often made.

In this chapter the penetrative effect is due to the fact that water has a density max-

imum at approximately 4°C. Hence we choose to model penetrative convection by a

quadratic equation of state of the form, cf. Veronis [102],

2
p = A,[l-c*(T-4) ], (2.0.1)

where p, T, and a are density, temperature, and thermal expansion, and p is the density
0

of water at 4°C. Detailed numerical studies of penetrative convection in a fluid layer em-

ploying a law like (2.0.1) are given in Chasnov & Tse [16] and Tse & Chasnov [100].

The buoyancy law (2.0.1) leads to non-Boussinesq effects. In a Boussinesq approxima-

tion p is linear in T and treated as a constant in all terms of the equation of motion

except the term involving the external force. The basis for this is that there are flows in

which the temperature (and hence density) variation are small yet the external force is

not. Hence the variation of density is chosen to be neglected everywhere except in the

external force term. The Boussinesq approximation simplifies the governing equations of

hydrodynamics significantly. It is widely employed and its validity discussed at length

in Chandrasekhar [15], p.16-18. Non-Boussinesq effects however are increasingly com-

mon and have been under much recent investigation, see e.g. Selak & Lebon [89], and

Straughan [91,94]. In this chapter we choose to employ a non-Boussinesq equation of

2.1. Governing Equations 13

state. In the next chapter an alternative is given in which a Boussinesq approximation is

used. A comparison of the two models is discussed in detail in chapter 3.

In the paper of Chen & Chen [17] the important parameter in their analysis is the ratio

d e t h o f fluid l a v e r
d= — = P ( 0 2)
2

dm depth of porous layer

They interpret their findings by showing that for relatively small d ( < 0.13) instability

is initiated in and dominated by the porous medium. Otherwise, the mechanism changes

and instability is dominated by the fluid layer. In the current work we find a very strong

interplay between d and T , the temperature at the surface of the fluid layer. The influence
v

of Tu and the quadratic law (2.0.1) can vastly change the findings as compared to those

of Chen & Chen [17], and of Straughan [92,93]. The non-dimensional permeability

8 — y / K / d , where K is permeability and d

m m is the depth of the porous layer, is also

important in the current situation. In this chapter we find that streamline patterns for

penetrative convection in a fluid overlying a porous layer are very different from what one

finds for penetrative convection in a single layer of fluid or porous medium. We believe

the present chapter contains new findings which will be of interest to anyone working in

convection where a fluid overlies a porous layer.

2.1 Governing Equations

T v > 0°C

Fluid Layer
T0

Porous medium
0°C

Permafrost

Figure 2.1: Schematic diagram of the governing system.

2
Consider a fluid (water) occupying the three dimensional layer {(x, y) 6 E } x {z e
2
(0, d)}, saturating a porous medium superposed below {(x,y) e K } x {z € (—d ,0)}.
m

The interface between the saturated porous medium and the water is at z = 0, see figure
2.1. Governing Equations 14

2.1. The motion in the fluid layer can be described by the Navier-Stokes equations (see

Chandrasekhar [15], p.11-13, for a derivation from first principles),

A v
-j£+Vi-fT- = P,i + » i ~ -9P{T)ki, (2.1.1)
at dxj p0 p 0

where Vi,t,Xi,p,v, and g, are velocity, time, displacement, pressure, kinematic viscos-

ity and gravity respectively and p the density, is assumed to be a quadratic function of

temperature of the form (2.0.1). Standard indicial notation and the Einstein summation
T
convention are employed throughout, and k = ( 0 , 0 , 1 ) . The constant term in (2.0.1) can

be absorbed into the pressure by defining a modified pressure p to be p + gp (l 0 ~ 16a)z.

Then, dropping the hat, (2.1.1) becomes

Equation (2.1.2) together with an incompressibility condition (Chandrasekhar, p.11) and

equation of energy balance (Chandrasekhar, p. 18) yield the following system of governing

equations
dvi dvi 1 dp _ . 2 s

| ^ = 0, (2.1.3)
OXi
k f
— + v — l
= AT
dt dxi (poC )f p

where kf and c are thermal conductivity and specific heat at constant pressure respec-
p

tively. Equations (2.1.3) are assumed to hold for time t > 0, in the spatial domain
2
{(x,y) 6 R , z € (0,G?)}. Subscript (or superscript) / or m denotes fluid or porous

medium respectively, and the symbol A is the Laplace operator.

In the porous layer the equations of motion are described by Darcy's law. A full

derivation, discussion and justification of Darcy's law can be found in Nield & Bejan [73].

In addition, the recent account by Brown [10] provides an interesting insight into Darcy's

life and the making of his law. The buoyancy force is again given by (2.0.1); hence the

governing equations for the porous medium are

m
1 dv v
0 = -=v? + ghat {Tl - ST ),
m

po OXi K
m
dv
! T - = 0, (2.1.4)
OXi
( p 0 c p y ^ + ( p 0 c p ) v r ^ =
f k*ATm- 1
2.1. Governing Equations 15

2
These equations hold on {{x,y) G R } x {z G ( — d , 0 ) } x {t > 0 } , and the variables m

^"\Pm, r m are velocity, pressure and temperature in the porous medium. In (2.1.4)x the
1
acceleration term has been omitted since this is believed to be small and negligible , cf.

Nield & Bejan. Starred quantities are defined in terms of fluid and porous variables as

S* = 4>S f + (1 - <£)S , m

where <f> is the porosity and 5 stands for a physical variable such as thermal conductivity

k.

The steady state solution whose stability is to be investigated is one where there is no

fluid motion in either layer and the temperatures on the upper and lower boundaries are

held fixed at constant temperatures, Ty and T , respectively, with T L v > T ,T L L — 0°C.

The basic steady state solution to equations (2.1.3) and (2.1.4), denoted by (vi,T,p),

(tJf.f^^.isthen

Vi = 0, T = /3z + T , 0 0 < z < d ,

(2.1.5)
v™ = 0, T m = pz m + T, 0 - d m < z < 0 ,

where /? = ( T v - T )/d,/3
0 m = T 0 / d m and T is the temperature at the interface. The
0

temperature at the interface is found for known T v by requiring continuity of temperature

at z = 0 and continuity of the heat flux, viz.

k d T = k t d f r n ^ z = Q

dz dz

If we define = k f / k * then from (2.1.6) the relation, T 0 = e Tu/(e

T T + d), can be

obtained. Furthermore we define S = e r / ( e T + d), £i = 4/7V, and £, £ m by

r _ & ~ g t _hzA
In our problem we need only specify T ^ , and since and d are specified, T does not 0

enter the analysis directly and we know £ and £ . m

'if the inertial term is included we have dependence on the porosity and an extra n eigenvalues are seen
in the spectrum (n is the number of Chebyshev polynomials used in the numerical computation). However,
the extra n eigenvalues do not appear in the upper part of the spectrum and so do not include the critical
eigenvalue. The numerical value of the leading eigenvalue was computed with and without inertial effects
and no significant difference was seen.
2.2. Perturbation Equations 16

We are unaware of a specific existence theory which has been developed for the model

outlined above. However, Rodrigues & Urbano [84] developed a mathematical existence

theory for a general Darcy theory with ice present, and their analysis is a suitable starting

point.

2.2 Perturbation Equations

In order to study the instability of the basic solutions (2.1.5) we introduce perturbations
m
(ui, 9,7r), (it™,9 , 7r ), m to these solutions and set,

Vi = Vi + Ui, T = f + 0, p = p + Tv,

m
v™ = v? + u?, T m = T m + 0, P m = p m + n m-

The perturbation equations which arise from (2.1.3) and (2.1.4) may be shown to be,

dui dui 1 dn . , . -. „ A , „ 9
1 u 2
~ET + UjTT = ^ - + "& i - 2agki(4 - T)9 + agk 9 , t

dt oxj p 0 oxi

P = 0, • (2.2.D
OXi
d d k
M f *n o
dt dx l [PoCp) /

and,

m m
0 = - 5 - - -j7«T - 2 a ^ ( 4 - T ) 9 + agkift,
pO OXi -ft

du™
= 0, (2.2.2)
axi
m m
d8 B9
m m
(Pocp)*-^- + (p c ) u™— 0 p f = k*A9 - (poc ) {J w ,
p f m

m
where w = u and w 3 = u™.

To non-dimensionalise the above equations let,

Ui = Uu*, Xi = dx*i, 7r = Px*, 9 = T*9\ t = Tt*,

m m m m m m n m
u™ = U u™, x™ = d x*, m 7r m
= P ii \ 9 = T *9 *, t = V t *,

where the starred quantities are dimensionless and the fluid and porous scalings are given

by,
2.3. Linearised Instability Theory 17

Pr
Jjm _ pm _ PoVU™ p#
m = u ™ <ym _ <%n
dm dm Y gotd K' m V '
where A = k f / ( p c ) f , Pr 0 p m = u/X m and A m = k*/(p c )f.
0 p In addition the fluid and

porous Rayleigh numbers Ra and Ra m are defined by

1
z^A v

We also define the Darcy number S = V K / d m , the Prandtl number Pr = v/X, and

G m = {poCp)* I'{poCp) / •

Then, omitting the stars, equations (2.2.1) and (2.2.2) can be rewritten in terms of

non-dimensional variables as,

at CTCj S^i

and,

m
du
= 0, (2.2.5)
dxi

2
Equations (2.2.4) hold in the region {(a;, y) G K } x {z e (0,1)} x {t > 0} while (2.2.5)
2 m
hold on {(x, y) £ R } x {z e ( - 1 , 0 ) } x { i > 0 } . The interface is z = 0, the lower

porous surface is now z = — 1 and z = 1 is the (upper) fluid surface.

2.3 Linearised Instability Theory

The linearised perturbation equations are derived from (2.2.4) and (2.2.5) by discard-

ing quadratic terms. Since the resulting system is linear we may seek solutions of the
<7t <Tt <Tt m at m m
form u x = ii (x)e , 0 = 6(x)e ,
l TT = 7r(x)e , uf = u (yi)e , 6 = ^ (x)e^,
m m <Tt
and ix = 7r (x)e . This gives rise to an eigenvalue problem for the growth rate a.

In fact, there are an infinite number of eigenvalues cr;, which may be ordered so that
2.3. Linearised Instability Theory 18

ai > a2 > .. • . If (J\ > 0 then instability occurs. The associated eigenfunction
1] {l
{u\ , TT^\ u™ \ 0 m ( 1 )
, 7r m ( 1 )
} yields the solution at the onset of instability. In par-

ticular, {uf\u™^} yield the streamline curves which are included in section 2.7, to

visualise the flow patterns. Likewise, the functions {9^,9™^} yield the isotherm pat-

terns.

The time scaling t m

= t/cu is introduced, then (2.2.4) and (2.2.5) yield the non-

dimensional linearised perturbation equations

aui = — + Aui — 2R(£ — z)9ki,

Ui,i = 0, (2.3.1)

oPr9 = A9- Rw,

2 m m
6 K™ = -u? - 2R {U - z)9 k u

u™ = 0, (2.3.2)

m m m m
auPr G 9 m m = A9 - Rw,

The method of solution for these equations follows the work of Chandrasekhar [15].
m
The pressures 7r and 7 r are eliminated by taking curlcurl of (2.3.1)i and (2.3.2)i. The

third component of the resulting equations are retained to find

2
a Aw = A w - 2RM{z)A*9, (2.3.3)

and
m m m m
0 = Aw + 2R M (z)A*9 , (2.3.4)

where
m
M ( z ) = £ - z , M (z) = U - z ,

and A* is the horizontal Laplacian,

2 2
A * a d
A = 1 .
2 2
dx dy
m m
A normal mode representation ofw, 9, w and 9 in (2.3.3), (2.3.1) , (2.3.4), and (2.3.2)
3 3

is now employed by setting,

w = W(z)f(x,y), e = G{z)f(x,y),
m m
wm = W m ( z ) f m { x ^ y ^ Qm = e (z)f (x,y),
2.4. Boundary Conditions 19

m
where f ( x , y) and f ( x , y) are horizontal plan forms which satisfy

2 m 2 m
A*/ = - a / , A*/ = -a m f .

The plan forms reflect the horizontal shape of convection cells at the onset of instabil-

ity. Benard [7,8] demonstrated experimentally, that for a horizontal layer of fluid heated

from below the ensuing motion at the onset of convection has a stationary cellular struc-

ture. The cells become equal and align themselves forming a regular horizontal pattern.

Typically, the convection cells are hexagonal for which

1 1
f ( x , y) = cos-a(V3x + y) + cos-a(V3x — y) + cosay,

cf. Straughan [94] and Drazin & Reid [24]. The wavenumber a is a measure of the

"width" of the convection cell to the depth, d. In this way the cells may be periodically

repeated in a pattern which tiles the (x, y) plane. The plan form in the porous medium

has the same shape although a = da . m

The fourth order equation (2.3.3), can be written as two second order equations by
2 2 2 2 2
regarding A = (D — a )W as an independent variable where D = d /dz . Using this
m m
and the normal mode representation of w, 6, w and 9 in (2.3.1) , (2.3.3), (2.3.2) and
3 3

(2.3.4), results in having to solve five coupled second order equations to determine the

critical growth rate a, namely,

2 2
{D - a )W = A,
2 2 2
{D - a ) A + 2RM{z)a e = a A, (2.3.5)

2 2
(D - a )0 -RW = aPr9,

2 m m m 2 m
(D - al)W - 2R M (z)a e = 0,
(2.3.6)
2 m
(D 2
- a je - RW m m
= aujPr G e ,m m
m

where (2.3.5) hold on z € (0,1) while (2.3.6) hold on z G ( - 1 , 0 ) . Coupling is through

the boundary conditions.

2.4 Boundary Conditions

The upper surface of the water (z = d) is open to the atmosphere (tangential stress free)

and held fixed at some constant temperature T v > 0°C. Hence for an incompressible
2.4. Boundary Conditions 20

fluid we have the boundary conditions

w = 0, w =0,
>zz 0 = 0, at z = d. (2.4.1)

Condition (2.4.1) is derived from the fact that an incompressible fluid at a free surface
2

satisfies
u,x + v, + w
y tZ = 0, and u 2 = v >z = 0.

The lower boundary of the porous medium (z = —d ) is held fixed and has tempera- m

ture T = 0°C. Hence the lower boundary conditions are

L

wm = Q ) Qm = Q ) at Z = -dm. (2.4.2)

At the interface z = 0, continuity of the normal components of velocity, temperature and

heat flux are required, so
m
ap f)f)
m m
w = w, 9 = 9, A/— = A - — . m (2.4.3)
(JZ (J Z>YYI

Two additional conditions on z = 0 are needed. The first of these is continuity of normal
stress, cf. Straughan [92],
m
% = 7T- (2.4.4)
az
For the second we employ a Beavers-Joseph boundary condition,

^ = 0 = 1,2, (2.4.5)

where a is a coefficient depending on the fluid and porous medium under consideration.
Further details of the last condition and alternatives are given in Nield & Bejan [73] and
Straughan [93]. Condition (2.4.5) relates tangential stress to the relative velocity at the
interface.
Using the relevant scalings the boundary and interface conditions can be written in
non-dimensional form as:

6 =w = w tZZ = 0, on z = l, (2.4.6)
m m
6 = w = 0, on z = -1, (2.4.7)
2.5. Numerical Method 21

and

3 d z 5
d Vd *r V of ^
(p7r m
= 7T - 2w, , 2 (2.4.8)
p
— = — (u -dui), £ = 1,2, on 2 = 0.

The boundary conditions (2.4.6)-(2.4.8) are now written in terms of their normal mode
expansions to yield,

6 = W = A = 0, on z = l, (2.4.9)
m m
9 = W = 0, on z = -1, (2.4.10)

and

m m
w = dw , m
e = - 4 = e, £>e = s.f^De,
3 5
Vd e T V d
2
r/
2 m
2 a L W - DA - — L W = -aDW, (2.4.11)
2 m
a W + ,4 = — ( £ W - d Z > W ) , on z = 0.
0

Thus our goal is to solve (2.3.5) and (2.3.6) subject to the boundary conditions (2.4.9)-
(2.4.11). Before describing the numerical method of solution note some useful relations,
namely
T *o 1 d 5

LO = -^ = d , Pr m = e Pr, T

'm

2.5 Numerical Method

2
In this section two variants of the Chebyshev tau method, the D and D techniques are
presented, cf. Dongarra et al. [23] and appendix A . l . Apart from numerical accuracy,
the Chebyshev tau method enables us to calculate as many eigenvalues as we need which
is useful in cases when the eigenvalues of interest are changing in parameter space. The
eigenfunctions can also be calculated very easily via this method. The implementation
of a Chebyshev tau method has several other advantages which are discussed in detail in
Straughan [93].
2.5. Numerical Method 22

2
The Chebyshev tau D technique is used to find the majority of the numerical results
presented here so this is considered first. In Dongarra et al. [23] the authors also discussed
A
the D and D methods. The power of D refers to the highest derivative being discretized.
2
Dongarra et al. recommended the D alternative as it has better growth rate properties
4
than the D case, whereas it makes more efficient use of the QZ algorithm than the D
2
method, which requires matrices twice as large. We concentrate on the D method, as
recommended, but also include a description of the D method for completeness as the
2
results obtained in the D method are used to independently check those of the D case.

2
2.5.1 The Chebyshev tau D method
2
Before implementing the Chebyshev tau D method (2.3.5), (2.3.6), and (2.4.9)-(2.4.11)
must be transformed to the Chebyshev domain (—1,1). To do this let z = 2z — 1 and
zm = —2z m — 1 in (2.3.5) and (2.3.6) respectively. Similarly conditions (2.4.9)-(2.4.11)
are transformed by making the appropriate substitution for the fluid and porous terms.
Then the fluid surface z = 1 becomes z = 1 and the porous surface z m = — 1 becomes
zm = 1 while the interface z = 0 = z m is transformed to z = —1 = z . Hence the m

eigenvalue problem (2.3.5)-(2.3.6) on the Chebyshev domain ( - 1 , 1 ) (dropping the hats)

is

2
(AD 2
a )W = A,
2 2
(AD 2
a )A + 2RM(z)a e = aA,
2
(AD 2
a )Q -RW = aPrQ, (2.5.1)
2 m m m 2 m
(AD 2
a JW - 2R M (z)a e m = 0,
2 m
(AD 2
a je - RW m m
= auPr G &
m m

The corresponding boundary conditions are

e =w=A =
m
e = w =0, m
on z =1 (2.5.2)
2.5. Numerical Method 23

and the interfacial terms,

m m m
W = dW , 0 = - 4 = ©, -D@ = 5 [^DO,
x

3 5
Vd e T V d
2

2
d m
2a DW -DA + —DW l
= -aDW, (2.5.3)
o
2 m
a W + A = ^(DW + dDW ), on z =-1,
o
m
where z G ( - 1 , 1 ) , M = £ - 1/2 - z/2 and M =£ m + 1/2 + 2/2.
The five quantities W, A, 0 , W m and 0 m are regarded as being independent variables
and expanded as Chebyshev series, i.e.

N+2 N+2 N+2

W T Z
W = J2 " "( ^ A = Y,A T (z), n n Q=J2e T (z), n n

n=o 71=0 71=0 ^2 ^ ^1

N+2 N+2 V • • )
W T 1 @
W m = Yl n n (Z), 0" = E nTn (z) •
71=0 71 = 0

The series above are truncations of an infinite series. Due to the truncation, the tau method
argues that rather than solving (2.5.1) one instead solves

2 2
{AD - a )W - A = N T N + 1 + r T 2 N + 2 ,

2 2 2
(AD - a )A + 2RM(z)a B -oA = rT 3 N+l + rT ,
4 N+2

2
(AD 2
- o ) 0 - R W - aPr& = T T 5 N + 1 + T T 6 N + 2 ,

2 2 m m m 2 m
(AD - a )W m - 2R M (z)a Q m = rT 7 N+1 + r T s N + 2 ,
2 2 m m m m
(AD - a JQ - RW - aujPr G Q m m = rT 9 N+l + r T w N + 2 ,

where T i , . . . , r , the tau coefficients, maybe used to measure the error associated with
10

the truncation of (2.5.4), cf. Dongarra et ai. [23]. For our purposes, it is not necessary to
calculate the tau coefficients as it is sufficient in practice to simply observe convergence
of the eigenvectors to assess accuracy.
Equations (2.5. l)-(2.5.3) can be expressed as a generalised matrix eigenvalue problem
of the form

Ax = aBx. (2.5.5)

To illustrate this more clearly, and establish the form of the matrices A and B we take
the inner product of each equation in (2.5.1) with some Chebyshev polynomial T (cf. k
2.5. Numerical Method 24

appendix A . l ) . The Chebyshev polynomials have the following orthogonality property,

1
T (x)T (x)
n k n
(T ,T )
n k = - c8 .k nk (2.5.6)
1
y/T X
where

2 i f A; = 0,
Cfc
1 i f A; > 0.

This is exploited as follows. Consider (2.5.1)i, taking inner products with some T we k

find,
2 2
(4D W,T ) k - (a W,T ) k - (A,T ) k = 0,
which can be expanded using (2.5.4), (2.5.6), and the fact that the derivative of a Cheby-
shev polynomial is a linear combination of lower order Chebyshev polynomials to yield,
N+2
2 2
Y,4D W -a W -A n k k = 0. (2.5.7)
71=0

Similarly an entirely analogous argument can be applied to (2.5.1) to deduce that, 2

N+2
2 2 2
4D A n - aA k + (2RM(z)a @, T ) = aA .
k k (2.5.8)
n=0
The third term of (2.5.8) can be written as,
2 2 2
(2RM(z)a Q,T ) k = 2Ra {£ - l / 2 ) ( 0 , T ) - fe Ra (zQ,T ). k (2.5.9)

To expand (zQ, T ) note the following,

k

N+2 N+2 , ~ v
T
Z& = T ^0nT l n = J2[f( l+n+T ^ l n l ))
n=0 "=n
n=0 ^ '

OQTI + - # i ( T + T ) + . . . + - 9 0 2 N + 2 (T N + i + T N + 3 ).

So, taking inner products with T , 7\ and T for k — 2 , . . . , TV + 2, yields,

0 k

e = -(ze,T ),
l Q e + ± = -{zQ T ), Q
d
1 l +
Z 7T
7T I 7T z
which can be expressed in matrix form as,
i (o )
/o 2
0 0 ... o\ 0

1 0 1 0 ... 0
2 01
0 1 0 1
... 0 jf(*e,T )
2 62 2
2
= (2.5.10)

0 1 o \
:
2

0 \0N+2J
V°
2.5. Numerical Method 25

Let the coefficient matrix in (2.5.10) be denoted by Z. Then expressing (2.5.7) and (2.5.8)
in matrix form and applying an enti rely si mi lar argument to (2.5.1) , (2.5.1) 4, and (2.5.1) 5 3

we can write (2.5.1) as the eigenvalue problem,

(AD 2
- aI 2
-I 0 0 0 /w\
2 2
0 AD - aI 0 0 A
-RI 0 AD - 2
aI 2
0 0 0
#
2 2
0 0 0 AD - a m

m
0 0 0 -R I AD 2
0 m
V /
0 0 0 0 0 \ W

0 I 0 0 0
(A)
= a 0 0 PrI 0 0 e
0 0 0 0 0
\0 0 0 0 uPr G I m m

where / is the identity matrix,

2 2
* = 2Ra {^ - -)I - Ra Z, ^ = -2R ai(r m
+ ^)I-R af Z, m
n
£1

and,

T
W = (W ,W ...,0 U W N + 2 f , A = (A , A x , . . . , A
0 N + 2 ) ,
T m T
e = (ft,, 0 i , . . . , o ), N+2 w = (w™, w ™ , . . . , w% ) , +2

0 m -_ (am am am \T
(yo , ^ 1 , • • • , Vn+2) •

2
The operator D is written in matrix form (cf. Dongarra et al. [23] and appendix A . l )
and the m(N + 2) and m(N + 3) rows of the matrix blocks in the system above, m =
1 , . . . , 5, are replaced with the discrete forms of the boundary conditions. The boundary
n
conditions (2.5.2) and (2.5.3) are realised with the aid of the relations T (±l) n = (±l)
- 2
and T ^ ( ± l ) = ( i l ) " ^ . This results in the generalised 5(N + 3) x 5(iV + 3) matrix
eigenvalue problem of the form

Ax = crBx. (2.5.11)

This generalised matrix eigenvalue problem is solved for the eigenvalues , n = 1,2,...
n
and the eigenfunctions x^ ^ with the aid of the QZ algorithm, cf. Moler & Stewart [62],
which we employed via the NAG routine F02BJF.
2.5. Numerical Method 26

2.5.2 The Chebyshev tau D method

In the previous section the transformed eigenvalue problem (2.5.1)-(2.5.3) was written as
a system of 2nd order equations. To implement a Chebyshev tau D method the system is
rewritten in terms of first order equations only,

2DW -A =0, 1 2DA 1 - A = 0,2 2DA 2 - A = 0,

3

m m
2DG -B = 0, 2DS - C = 0, 2DW - E = 0,
2 A 2 2
2DA 3 - 2a A 2 + a W + 2RMa @ = a(A 2 - a W),
(2.5.12)
2
2DB - a e -RW = aPrG,

2DE - a W 2
m
m
- 2R M a ne
m m 2 m
= 0,
m
2DC - a S 2
m
m
- RWm m
= auPr G e . m m

The boundary conditions are now given by

m m
Q = W = A = O 2 = W = 0, on 2 = 1, (2.5.13)

and the interfacial conditions by,

m r
W = dW , G e, C 6J-=-B, 5
d
2
d
-3a Ai 2
+A + — (2.5.14)
35 E = 2
oA u

da
A, (Ai-dE), on z — — 1.

2
Using arguments entirely analogous to those used in the D case, (2.5.12) can be
written as the generalised eigenvalue problem,
/ 2D -I 0 0 \fw\
0 2D -I 0 0 A 1

0 0 2D -I 0 0 A 2

a/4
0 -2a / 2
2D * 0 0 A 3

0 0 0 0 2D -I 0 0 0
-RI 0 0 0 -a I 2
2D 0 0 0 0 B
m
0 0 2D I 0 0 0
0 0 2
a I
m 2D m
RI 0 c
0 0 0 2D I
0 E
\ )
2.6. Computational Results 27

/ 0 0 .. (w\
0 0 0 A,
0 0 0 0 0 A 2

2
-a I 0 10 0 .. A 3

0 0 0 0 e
a
0 PrI 0 0 B
m
0 0 0 0 e
0 -uPr G I
m m 0 0 0 c
0 0 0 0
0 0 0/ \
\ E
)
2 2 2
where * = 2Ra (£ - 1/2)7 - Ra Z and # = 2R a (( m m m + 1/2)1 + RmO^Z. Z is
2
defined as in the D case and the operator D is written in matrix form (cf. Dongarra et
al. [23] and appendix A . l ) .
The m(N + 2) rows of the matrix blocks, m = 1 , . . . , 10, are replaced with the
discrete forms of the 10 boundary conditions (2.5.13) and (2.5.14). This results in the
2
10(iV + 3) x 10(N + 3) (twice as large as the D case) generalised matrix eigenvalue
problem of the form

Ax = crBx. (2.5.15)

n n
Again this is solved for the eigenvalues o^- \ n = 1,2,..., and the eigenfunctions x^ )
with the aid of the QZ algorithm. The results of the D method were then used as an
2
independent check of those in the D case.

2.6 Computational Results

In this section computational results obtained by varying the three parameters d, T , and v

5 are reported. Other parameter variations for the non-penetration case are considered in
Straughan [93]. The novelty of the results presented here is due to the penetration effect
and we highlight this. The Chebyshev tau method allows us to obtain highly accurate
results with very few polynomials. In our computations we employed 30 polynomials.
It is curious to note that for the parameter values selected we always find the critical
2.6. Computational Results 28

eigenvalue to be real. The spectrum is such that e R although other eigen-

values further down the list may occur as complex conjugate pairs. The validity of all
eigenvalues was checked by carefully controlling the convergence of each eigenvector.
This is important since spurious eigenvalues are often found in problems of this nature.
Spurious eigenvalues are numbers which may appear in the eigenvalue list but are not
solutions to the eigenvalue problem. Practical checks as to whether an eigenvalue is spu-
rious involve changing the number of polynomials or examining the convergence of the
associated eigenvector. I f an eigenvalue is spurious then changing the number of poly-
nomials is accompanied by a significant change in the "eigenvalue", and the components
of the eigenvector are not observed to converge. An inspiring account of how spurious
eigenvalues may arise in hydrodynamic stability problems is given by Dawkins et al. [22].
All the eigenvalues we computed, including the complex conjugate pairs, are believed to
be solutions of the problem.

We find that the streamlines and isotherms at the onset of convection may display
striking patterns with a many cellular structure. This effect can only be seen by carefully
computing streamline patterns or by careful analysis of the eigenfunction. The many
cellular structure is not obvious from the eigenfunction with the naked eye (see e.g. figure
2.18). Due to the highly sensitive streamline structure observed we note that it is essential
2
to employ a highly accurate numerical technique like the Chebyshev tau D one to detect
such fine detail.

In our computations we discover striking differences with the findings of Straughan

[92,93] and Chen & Chen [17]. These papers typically found that the neutral curves
showed the instability pattern changed from being dominated by the porous medium to
being dominated by the fluid layer for d in the region of 0.1. In Straughan [93] the
parameter a in the boundary condition (2.4.5) was found to play an important role in that
if a > 0.8 the porous medium dominates the instability. However, for many fluids and
porous media, a lies in the range 0.1 to 0.8 and then the changeover is of order cf = 0.1
(cf = O(0.1)). In this work we discover that the changeover point also depends strongly
on the upper temperature of the fluid layer, T . Thus the penetrative convection effect is
v

very important. For example, with T v = 0(8°C) the changeover occurs for of = 0(1)
whereas when Ty = 0(4°C) the changeover is in the region d = O(0.1).
2.7. Numerical Findings 29

We also observe that penetrative convection in the fluid overlying a porous medium
situation is very different from penetrative convection in a single fluid layer, or penetrative
convection in a layer of saturated porous material. The interface in the two layer situation
plays an important role. For example, in the one layer fluid or porous problem true pen-
etrative convection, i.e. the formation of two convection cells in the vertical, occurs only
when T v > 6.4°C and three cells form only when T v > 0(21°C), cf. Straughan [91].
However in the two layer porous-fluid problem with d = 0.4 we already see two fluid
cells when Tu = 5°C and when Ty = 6°C three cells have formed.

We stress that we have not seen the penetrative convection problem studied in the fluid
overlying a porous medium problem and the numerical findings are quite striking and very
different from anything we have seen in previous works on penetrative convection, or in
convection in a porous-fluid layer system. We also stress that our results can only be
observed with a highly accurate numerical method and we strongly advocate the use of
2
the Chebyshev tau D method for this and related problems.

2.7 Numerical Findings

In this section a selection of our numerical findings are presented. We point out that these
are a small amount of the many graphs computed. The idea is to present a representative
selection highlighting some of the novel features we have found. Throughout all calcula-
tions we select Pr = 6 (the Prandtl number for water) and fix the parameters ey, G m and
a at the values 0.7,10 and 0.1 respectively, cf. Straughan [93].

2.7.1 Variation of the neutral curves with d, Tu and S

In figures 2.2-2.4 we present the linear instability neutral curves, i.e., the porous Rayleigh
number against porous wavenumber curves such that a = 0, for various values of d, T u

at
and S. Since the time representations of u 9, n,
iy U™,6 m
and 7r m
are of the form e , it is
clear that above the neutral curves (a > 0) the system is unstable. For a fixed wavenumber
a , one finds a value for Ra .
m m Since all values of a m are allowed, we let a m vary from
0 —¥ oo and the minimum value of Ra m yields the critical value above which instability
will commence.
2.7. Numerical Findings 30

140-•

120-

100
0.10

80

60

0.1.2,-
40

0.14
20 0.16

0.20

0 15 20 25
%1

Figure 2.2: Neutral instability curves, Ra m against a . T\j = 4°C, 8 = 0.002, d varies as shown
m

on graph.

In figure 2.2 we fix T v and 8 at 4°C and 0.002 respectively and vary d. The bi-modal
nature of the neutral curves is clearly seen. Of course, with an upper temperature of
4°C the whole layer has a tendency to be unstable and penetrative convection will not
be witnessed. This is analogous to the computations of Straughan [93] and we expect to I
find the switch from porous to fluid dominated convection with d in the range we have.
In fact, the absolute minimum on each curve represents where instability commences.
Thus, for d = 0.16,0.18,0.2, the minimum of the curve lies on the fluid dominated part
with a m larger (narrower convection cells confined mostly in the fluid) whereas for d =
0.1,0.12,0.14, the minimum of the curve is in the porous dominated part with a m smaller
(wider convection cells, with convection occurring strongly in the porous medium).

In figure 2.3 we fix d and 8 at 0.4 and 0.002 respectively and show the neutral curves
for T v varying from 5°C through to 5.4°C. Note that for T v = 5°C to 5.3°C the fluid
dominates the convection pattern with large wavenumbers but when T v = 5.4°C the
pattern changes and convection is driven by the porous medium with a corresponding
smaller wavenumber.
In figure 2.4 we fix d and T v at 1 and 8.2°C respectively and show the neutral curves
for 8 varying from 0.001 to 0.01. We note that for 8 = 0.001 and 0.002 the fluid dom-
inates the convection pattern with large wavenumbers but when 8 is increased to 0.003
the pattern changes and convection is driven by the porous medium with a corresponding
2.7. Numerical Findings 31

30 i

20

R
%1

i
10 15 20

Figure 2.3: Neutral instability curves, Ram against a . 6 = 0.002, d = 0.4, T[/(°C) varies as
m

shown on graph.

150 i

0.01

100

50
0.OO3

0.002

0.001
i i
0 15 20 30

Figure 2.4: Neutral instability curves, Ra m against a . T\j = 8.2°C, d = 1, 5 varies as shown
m

on graph.
2.7. Numerical Findings 32

smaller wavenumber.
Having computed and inspected several more instability curves, table 2.1 was con-
structed. Table 2.1 illustrates the range in which TV must be to see the interchange where
instability moves from fluid to porous (or porous to fluid) dominated convection. An ap-
proximate value, T f f " , at which this occurs is listed. Clearly the smaller d the lower the
w
value of Ty (as expected from figures 2.2-2.3). In fact it can be seen from table 2.1 that d
and Tfj° are essentially proportional. A change of 0.2 in d corresponds to an approximate
change of 1°C in T f f ° .

d Range of TV (°C) Approximate value of Tff" (°C)

2.00 1 3 . 0 5 < T ^ <13.10 13.10

1.80 w
12.10;$ T{j <12.15 12.10
w
1.60 11.15;$ T(j <11.20 11.15
1.40 w
10.15< T f j ^10.20 10.20
1.20 9.20< Tff" <9.25 9.20
w
1.00 8.20< T(j <8.25 8.25
0.50 5.80< T f f ° <5.85 5.80
0.40 5.30< T f f ° <5.35 5.35
w
0.30 4.80< T(j <4.85 4.85
0.20 4.35 < Tfj° <4.40 4.40
0.18 4.25< T f f ° <4.30 4.30
0.16 4.15< T™ <4.20 4.15
0.14 3.90< T f f ° ^3.95 3.95

Table 2.1: The range and approximate value of TV (°C) in which a switch in dominance occurs
for d varying, S = 0.002.

Figures 2.2-2.4 show respectively, that as d increases the system becomes less stable
(the instability curves shift down), and as TV and 5 increase the system becomes more
stable (the instability curves shift up). It can be concluded then from table 2.1, and the
neutral curves presented above, that the onset of convection, and indeed which layer dom-
inates it, depends strongly upon d, TV and 5 (for e , G , Pr and a fixed).
T m
2.7. Numerical Findings 33

2.7.2 Streamline and isotherm patterns

The neutral curves already show that a variety of interesting behaviour is found when
penetrative convection is combined with the porous-fluid convection problem. However,
it is by looking at the streamline patterns that the structure of the cell patterns can clearly
be seen.

In the governing model and the plan form of the normal mode expansion, it was as-
sumed that a convection cell must be three dimensional. Reproducing three dimensional
streamlines is very difficult. Instead we choose to simplify the problem by concentrating
on a two dimensional roll cell, cf. McKay & Straughan [53]. To ensure continuity at z = 0
and show graphically the effect on the flow of varying d, we must first dimensionalise the
2
eigenfunctions found via the Chebyshev tau D method. We do this as follows.
Case 1: d < 1 (d < d ) m

m
Fix dm = U=m T* = 1 then

w = \w*, x = dx\ w m = W, m xm = x* , m

d
r (2.7.1)
0 = Q* Qfn gm*

Case 2: d > 1 (d > d ) m

Fix d = U = T* = 1 then

w = w*, x = x*, w m = dW , m xm = \x* ,

m

d (2.7.2)
e = e* 9 m = ^trd? Qm*
' 5 '
where the starred quantities denote non-dimensional terms.
We choose the planform for a two dimensional roll cell to be

7T 7T
g {x) = cosa x,
c c xG

where
a
c = ~, - ,

This takes into account that although the fluid and porous regions have different length
scales, once redimensionalised, both regions may be represented by a single plan-form.
2.7. Numerical Findings 34

We seek a stream function xf)(x, z) satisfying

u
-ip = w,
tX V,* = for z <E [0, d],
m m
= w, V > z = u for z G [ - d , 0),
m

where x G [—7r/a , 7 r / a ] and

c c

m m
u [ ± ^ , z j =u ( ± ^ , z\ = u(0, z) = « ( 0 , z) = 0.

Then the normal mode expansion gives,

w = W(z) cos a x, c z G [0, d];

m
ii) = F ( z ) c o s s i , c ze[-d ,0]. m

Hence the stream function is chosen to be

\ —±-W{z) sin a x, c ze[0,d],

ip{x, Z) — <
m
[~W (z)sma x, c ze[-d ,0).
m

The streamlines are given by the lines in the (x, z) plane where ip(x,z) = constant.
In the remaining figures of this chapter, it is the number of cells in the vertical direction
and the width of cells in the horizontal direction that are of interest. Throughout all
figures, the porous-fluid interface is at z = 0, and the eigenfunctions are all normalised
before plotting. Note that only one isotherm graph is presented since the isotherms were
found to be not dissimilar to the equivalent streamline graphs.
In figures 2.5-2.9 we highlight the streamline patterns for the situation for which fig-
ure 2.3 yields the neutral curves i.e. d and 8 are held fixed at 0.4 and 0.002, respectively.
The lower dashed line marks the interface, z = 0, and the upper dashed line, the con-
duction solution at 4°C. When Tu = 5°C (figure 2.5) two convection cells already exist
in the vertical direction, and are almost exclusively confined to the fluid. The evidence
of a second cell in the fluid indicates that penetrative convection is already occurring. It
must be remembered that water has a density maximum at approximately 4°C. Therefore,
since the temperature at the interface, T = 3.18°C, there is an unstable layer between
0

the porous interface and the conduction solution at 4°C. It is only in this layer and the
2.7. Numerical Findings 35

saturated medium below that instability can be initiated. When convection commences,
the (stable) fluid above is brought into motion and a second cell forms. As T v is in-
creased to 5.3°C (figure 2.6), the convective motion begins to penetrate into the porous
medium, and a third cell is seen in the fluid. Since the upper temperature has increased,
T has increased as well (3.37°C), and the depth of the layer between the 4°C line and
0

the interface is smaller than that seen in figure 2.5. Thus a third cell is formed since
the (stable) fluid above adjusts itself accordingly to accommodate the smaller (unstable)
depth. Once T reaches 5.4°C (figure 2.7) the convection pattern changes to much wider
v

cells (the bi-modal effect cf. figure 2.3). In fact, we see one cell wholly in the fluid and
one cell spanning the fluid and porous medium with circulation in each media. By the
time T(7 = 5.8°C (figure 2.8), again we have one large cell in the fluid and one cell in
which the fluid is circulating between the lower part of the fluid layer and the whole of
the porous medium. Upon increasing Tu further to 6°C (figure 2.9) the structure of figure
2.8 is maintained in the porous-fluid region, but now a third cell has formed at the top
of the fluid layer. Again this is due to the fact that the unstable layer in the fluid is of
smaller depth than that seen in figure 2.8, and the (stable) fluid above has adjusted itself
accordingly. Thus the penetrative effect is inducing the formation of further convection
cells.

In figures 2.10-2.16 we concentrate o n c i = 1, i.e., the fluid and porous medium depths
are the same, and fix 5 = 0.002. Figure 2.10 shows the streamlines when Tu = 8°C while
figure 2.11 shows the equivalent isotherms. Observe that there are six vertical convection
cells with little movement in the porous medium. Figure 2.12 displays the streamlines
for Tu = 8.1°C. Observe that now there are seven vertical convection cells in this case
and again little movement in the porous medium. Figure 2.13 displays the streamlines for
Tu = 8.2375°C, and in this case we see eight vertical convection cells and yet again little
movement in the porous medium. The increase in the number of cells with an increase of
Tu is due to the penetrative effect as previously explained. As Tu increases, the depth of
the unstable layer decreases and the fluid adjusts itself, forming a layered convection cell
structure which is stable for a given upper temperature.
Figures 2.14, 2.15 and 2.16 show what happens to the streamlines when Tu is in-
creased further to 8.25°C, 8.4°C and 9.6°C respectively. There is a switch from con-
2.7. Numerical Findings 36

• 1
r n &• ,"!

? >
(i J J 5
o O
"05

-0.125 -

Z -0.3-

-0.475 -

-'•""I 1 1 1 r ~i
U 1 1 i r
-0.4 -0.2 0.0 0.2 0.4 -0.4 -0.2 0.0 0.2 0.4
x

Figure 2.5: Streamline plot, T v = 5°C, Figure 2.6: Streamline plot, T v = 5.3°C,
T 0 = 3.18°C, Ra m = 2.458, a m = 9, T = 3.37°C, Ra
0 m = 11.331, a m = 12,
d = 0.4, 8 = 0.002. d = 0.4, 6 = 0.002.

i
r

4 • 03
4
•»!•( -0 H 5

-Q47-;
-o

0 4 O 3 0
-0 82?

:
! I u
1 •
• f. s -(! t .2 i :> 1 2 0.4 • 3 1 6
X

Figure 2.7: Streamline plot, Ty = 5.4°C, Figure 2.8: Streamline plot, T v = 5.8°C,
T = 3.4°C, Ra
0 m = 15.042, a m = 2, d = T = 3.69°C, Ra
0 m = 19.620, a m = 2, d =
0.4, 5 = 0.002. 0.4, S = 0.002.
2.7. Numerical Findings .37

p: ::
1.1225

1
-li 125

•0.475 0 0
-0

i 0 I I '' " I " I T I I I

•1-6 -|,2 -0.8 -0-4 0.0 0.4 0.8 1.2 1.6

Figure 2.9: Streamline plot, T v = 6°C,

To = 3.8°C, Ra. m = 21.833, a m = 2,
d = 0.4, 5 = 0.002.

i ii i i
i u

'I s,

-
5?
n n
o o

0 I,
uu
o :•

i n
Uu -0 :•

Figure 2.10: Streamline plot, T v = 8°C, Figure 2.11: Isotherm, T v = 8°C, T =0

T = 3.29°C, Ra
0 m = 5.723, a m = 10.4, 3.29°C, i?a = 5.723, a
m m = 10.4, d = 1,
rf = 1, S = 0.002. J = 0.002.
2.7. Numerical Findings 38

c—)

CD

V
., y

V
0u
U J
1 1
-0.4 -0.2 0.0 0.2 0.4 -0.4 -0.2 0.0 0.2 0.4

Figure 2.12: Streamline plot, T v = 8.1°C, Figure 2.13: Streamline plot, Ty =

T 0 = 3.3°C, Ra m = 8.065, a m = 11.2, 8.2375°C, T = 3.39°C, Ra 0 m = 13.296,
d=l,S = 0.002. am = 12.2, d=l,S = 0.002.

cr
1.0 10 1 1

3Cc 3
C c
3 r

c
0.4

0.0

•02

-0.4

-0.6
0 0
1.0 .0
1 1 1
16 1.2 •0.8 0.8 .2 16 .6 .2 -0.8 •04 00 .2 16
x

Figure 2.14: Streamline plot, T = 8.25°C, v Figure 2.15: Streamline plot, T v = 8.4°C,
T = 3.4°C, Ra
0 m = 13.875, a m = 2, d = T = 3.46°C, Ra
0 m = 15.258, a m = 2, d =
1,6 = 0.002. 1, 5 = 0.002.
2.7. Numerical Findings 39

1.0

0.8
J

0.4

0.2

-02

0
0
r
1.6 1.2 -0.8 12 1.6
x

Figure 2.16: Streamline plot, T v = 9.6°C,

T = 3.95°C, Ra
0 m = 24.503, a m = 2, d =
1, S = 0.002.

vection dominated by the fluid to that dominated by the porous medium between T v =
8.2375°C and 8.25°C. In figures 2.14-2.16 the porous medium has clearly begun to influ-
ence the convection and much wider cells are seen. Since the cells are wider, less form
in the vertical direction. Indeed in figures 2.14 and 2.15 we see one cell spanning the
porous-fluid region with four further cells in the fluid. By the time the upper temperature
is increased to T v = 9.6°C (figure 2.16), one cell still spans the porous-fluid region but
now a further five cells are observed in the fluid. Again this is due to the fact that the
unstable layer in the fluid is of smaller depth to that seen in figure 2.15 and the (stable)
fluid above has adjusted itself accordingly.

In figures 2.17-2.20 the highly sensitive nature of the streamline structure is illustrated
further. We fix d = 2 and present both the streamlines and corresponding eigenfunctions
for Tu = 13°C and 13.2°C. From observing the eigenfunction with the naked eye it
is not immediately evident what the corresponding cellular structure looks like, hence
a streamline plot is necessary and it is essential to employ a highly accurate numerical
technique to detect such fine detail. It would appear that for T v = 13°C we find 21
cells in the vertical direction, in the fluid. Whereas for Tu = 13.2°C the porous medium
influences the convection pattern to yield wider cells, with one cell spanning the porous
medium and fluid and a further 15 cells in the fluid above.
2.7. Numerical Findings 40

1.0

0.90625

0.8125
— •

0.71875 -

0.625

0.53125

0.4375

0.34375

Z 0.25 -

0.15625 -
r ir i
0.0625

u
to) to)
-0.03125

-0.125 J
-0.21875 -

-0.3125

-0.40625
V
-0.5
•0.5 -0.3125 -0.125 0.0625 0.25 0.4375 0.625 0.8125 1.0

Figure 2.17: Streamline plot, T v = 13°C, Figure 2.18: W eigenfunction, T v = 13°C,

T = 3.37°C, Ra
0 m = 10.470, a m = 11.6, T = 3.37°C, Ra
0 m = 10.470, a m = 11.6,
d = 2,S = 0.002. d = 2,6 = 0.002.

.0 1
1.0

0.8125 i 08

06

0625
=3 0.4

04375
2 0.2

0.25 0.0

-0.2
0.0625

-0.4

-0.125
-0.6

-0.3125
•08

-0 5 1.0
-0,9 -0,7 -0.3 ^).l 0.1 0.3 0.5 0.7 0.9 -0.5 -0.3125 0.125 0,0625 0.25 0.4375 0.625 0.8125 1.0
•0.5

Figure 2.19: Streamline plot, T = 13.2°C, v Figure 2.20: W eigenfunction, Tu

T = 3.42°C, Ra 13.686, a 13.2°C, T 3.42°C, Ra 13.686,
0
m m 0

2,6 = 0.002. a 2,d = 2,6 = 0.002.

2.7. Numerical Findings 41

Computing complex streamlines like those illustrated above can be quite time con-
suming. I f we wish to simply count the number of convection cells without computing
the streamlines then a quick alternative is to closely inspect the eigenfunction. Every sign
change in the eigenfunction represents a cell in the system. To ensure that all sign changes
are seen the eigenfunction must be computed at a sufficiently large number of points. The
w
eigenfunction at the critical Rayleigh number and wave number for T v « T v , was com-
puted at 5000 points and inspected to construct table 2.2.

w
d T v <T v (Fluid dominant) T v > T™ (Porous dominant)

2.00 21 16
1.80 21 11
1.60 20 10
1.40 17 8
1.20 13 6
1.00 8 5
0.50 3 3
0.40 3 2
0.30 2 2
0.20 1 1
0.18 1 1
0.16 1 1
0.14 1 1

Table 2.2: The number of convection cells in the two layer system, d varying, 6 = 0.002, Tu ~
rpSW
1
u •

Table 2.2 illustrates how many cells are present just before and just after there is a
switch in dominance. Observe that as d increases the number of convection cells increases
too. For d suitably large ( ^ 1) there are more cells present when the fluid dominates the
convection than there are when the porous medium does. This is to be expected since the
cells are wider when the porous medium dominates convection.
2.8. Concluding Remark 42

2.7.3 Effect of permeability

It has already been shown in figures 2.5-2.19, that the porous medium has a strong effect
on the streamline patterns as T v is varied. We now fix d = 1 and T v — 8.2°C and vary
the permeability. In practice, a value of permeability often gives rise to 5 = \/K/d m =
0(0.002), cf. Straughan [93], and so the findings of figures 2.5-2.19 are expected to be
representative. However, we feel it worthwhile to compute directly the effect of varying
3 3 - 2
K. Values of 8 = 2 x 10~ ,3 x 10~ and 1 0 are used to generate figures 2.21, 2.22 and
3
2.23. Clearly, a striking difference is noticed. When S = 2 x 10~ there are eight cells
in the fluid region with hardly any movement in the porous medium. When 5 is increased
- 3
to 3 x 1 0 the higher permeability induces motion in the porous medium and the total
number of cells decreases to five with a much wider cell structure being observed. On
2
increasing 5 further to 10~ only three cells are observed and the increased permeability
allows the porous medium to play a stronger role in the convection process.

2.8 Concluding Remark

Straughan [93] argued that there is a need for additional experiments to measure the coef-
ficient a in the boundary condition (2.4.5). I f one wishes to model convection in porous-
fluid systems which occur in real life for example, patterned ground formation under
water (cf. Krantz et al. [44], McKay & Straughan [53]) or fast sea ice melting in the
Arctic shelf (cf. Bogorodskii & Nagurnyi [9]), then our calculations indicate that since
penetrative convection coupled with a porous-fluid layer system is so sensitive, there is a
need for much more accurate experimental data in this area also.
2.8. Concluding Remark 43

• i
1.0

0.8
10 :

C_5 CD c

•02
-0 2
rr
-0.6
-06

08

10
1.0
16 12 0 8 12 6
-02

Figure 2.21: Streamline plot, T v = 8.2°C, Figure 2.22: Streamline plot, T v = 8.2°C,

3.38°C, Ra 11.561, a 12, o 3.38°C, Ra m 13.614, a 2, d

0 rn m m
d = 1, J = 0.002 1, 5 = 0.003.

i i
1 0
i
r
3
C3 C3

10 I
I 8 1 4 10 -06 .0 14 18

Figure 2.23: Streamline plot, T v = 8.2°C,

T = 3.38°C, Ra
0 m = 13.331, a m = 1.8,
d = 1, 8 = 0.01.
Chapter 3

Penetrative Convection in a Superposed

Porous-Fluid Layer via Internal
Heating

The purpose of this chapter is to provide an alternative model to that given in chapter 2.
In this instance we concentrate on the superposed two layer system but model penetrative
convection via internal heating as opposed to a quadratic equation of state. An internal
heat source (or sink) can give rise to a situation where one part of a layer is naturally
convecting while the other wishes to remain stable. Hence a heat source (sink) is capa-
ble of inducing penetrative convection. Many references to work in which convection is
described via internal heating can be found in Straughan [91]. One of the most signifi-
cant contributions, from which great advancement has been made, is that of Roberts [83].
Roberts modelled convection in a horizontal layer of fluid cooled from above, thermally
insulated from below and heated uniformly by an internal source. Matthews [60], moti-
vated by convection in an ice covered lake subject to solar radiation, presented a model
for the onset of penetrative convection in a fluid layer via internal heating. Matthews em-
ployed a cubic temperature profile maintained by a linear internal heat source. Similarly
penetrative convection in a porous layer can be described by internal heating, see for ex-
ample Straughan & Walker [96] and chapter 4. Applications of penetrative convection via
internal heating include convection within the Earth's mantle, cf. McKenzie et al. [54],
cumulus convection, cf. Krishnamurti [45] and Larson [46,47], and convection of stellar

44
Chapter 3. Penetrative Convection in a Superposed Porous-Fluid Layer via Internal
Heating 45
and planetary interiors and atmospheres, cf. Ghosal & Spiegel [34], Schubert et al [88],
and Zhang & Schubert [107,108]. In this chapter, penetrative convection in the coupled
porous-fluid system is dealt with via internal heating in both layers. We believe this is the
first time penetrative convection in such a system has been modelled this way.
It is found that the instability of the two layer system and the size of the ensuing
convection cells are very sensitive to change in the strength of the heat sinks. The sink in
the fluid is shown to have a destabilising effect on the system whereas that in the porous
can be stabilising or destabilising depending on the form of the steady state temperature
profile and the strength of the sink in the fluid layer. A strong interplay between the two
sinks is clearly seen and discussed in detail in section 3.5. A selection of streamlines are
presented which exhibit novel behaviour when the strength of the sinks are varied.
It is shown that the linearised forms of the model presented here and that given in
chapter 2 are mathematically adjoint in some instance. When this is the case it is found
that the two models yield the same critical instability thresholds. Hence the results given
here confirm the validity of those in chapter 2. However it is also shown that when the two
models are adjoint the corresponding eigenfunctions are not the same. From a physical
point of view the two models under consideration are very different and we stress this fact.
Further discussion of this can be found in section 3.6 and a similar result is encountered
in chapter 4 where the effect of penetrative convection on an anisotropic porous medium
is investigated by internal heating and a quadratic equation of state.

An outline of this chapter is as follows. In the first two sections the governing system
and corresponding equations are described and the non-dimensional perturbation equa-
tions derived. The boundary conditions are then discussed and a linear instability analysis
is given. A brief description of the numerical method is then provided. An analytical and
numerical discussion of the effect of the heat sinks on the stability of the system is pre-
sented next. A comparison is then made between the model presented here and that given
in chapter 2. The two are shown to be mathematically adjoint in some instance and the
eigenvalues and corresponding eigenfunctions are computed and discussed in this case.
Finally a further array of streamlines are presented which show that contrary to intuition,
the initiating convection cell is not necessarily the strongest one.
3.1. Governing Equations 46

3.1 Governing Equations

2
Consider a fluid occupying the three dimensional layer {(x,y) G R } x {z G (0,d)}
2
saturating a porous medium superposed below {(x,y) G R } x {z G ( — d , 0 ) } . The
m

interface between the saturated porous medium and the fluid is at z = 0, see figure 2.1.
Suppose the density has a linear temperature dependence of the form

p = Po[l-a(T-T )], r (3.1.1)

where p, T, and a are density, temperature, and thermal expansion coefficient respectively,
and p and T are constant reference values. Then the governing equations for the fluid
0 r

can be expressed by the Navier-Stokes equations, with a Boussinesq approximation,

dvi dvi 1 dp .
— +Vj— = — + v&Vi + agTki,
at oxj p oxi
0

dvi
L
o - = 0, (3.1.2)
OXi
at dxi (p Cp)/
0

Equations (3.1.2) and (3.1.2) are the incompressibility condition and balance of energy,
2 3

respectively. The Q term in the energy equation (3.1.2) is some (constant) internal heat 3

source or sink and it's inclusion allows the model to describe penetrative convection in the
fluid layer (see section 3.5). Equations (3.1.2) are assumed to hold for time t > 0, in the
2
spatial domain {{x,y) € R ,z G (0,d)}. In these equations Vi,t,x p, i: and T are veloc-
ity, time, displacement, pressure and temperature, and u, g, k f , c are kinematic viscosity, p

gravity, thermal conductivity and specific heat at constant pressure. Standard indicial no-
tation and the Einstein summation convention are employed throughout. Subscript (or
superscript) / or m denotes fluid or porous medium, respectively. The symbol A is the
T
Laplace operator and k = (0,0,1) .
In the porous medium the motion of the fluid is described by Darcy flow with the
buoyancy force again given by (3.1.1); therefore
rn
1 dp v
0= - f - - - v ? + agT ki,
m

p OXi
0 K
dv™
- ^ - = 0, (3.1.3)
OXi
{p C y
Q p dT m + vmdTrn = k* ^ +
1
(p c )f
0 p dt dxi (p Cp)f
0
3.1. Governing Equations 47

2
These equations hold on {(x, y) e R } x {z € (—rf , 0)} x { i > 0}, and the variables m

v™, p , Tm m are velocity, pressure and temperature in the porous medium. K is the perme-
m
ability and Q is some (constant) internal heat source or sink in the porous layer. Starred
quantities are defined in terms of fluid and porous variables as

S* = <j>S + (1 - </>)S
s mf

where 4> is the porosity and S stands for a physical variable such as thermal conductivity
k.
To establish what values of Q yield a source or sink, consider the integral form of the
energy balance equation (3.1.2) , namely, 3

£ f TdV + I v . f d V = j X r f f ( f ) dV + 2 / QiV, (3.1.4)

dt Jv J dxi v(PoCpjf J dxi \dxij J v v

where V is some periodic cell in the fluid layer z 6 (0, d). Using the divergence theorem,
(3.1.4) can be rewritten as

4- [ TdV+ I niViTdA= — ^ — I m^-dA x

+ 2 [ QdV,
dt J v Jav KPoCpjf Jdv vi Jv

where rii is the unit normal to V and dV the boundary of V. Note that, in the absence of
any heat flux and velocity a positive value of Q will give rise to an increase in temperature
of the layer and similarly, a negative value of Q will give rise to a decrease. Hence a
positive value of Q represents an internal heat source and a negative value of Q an internal
heat sink. An entirely analogous argument applied to (3.1.3) reveals that a positive value 3

of Q m represents an internal heat source in the porous layer and a negative value of Q m

an internal heat sink.

The temperatures on the upper and lower boundaries are held fixed at values T v (>
0°C) and 0°C respectively. The steady state solution to equations (3.1.2) and (3.1.3),
1 m m
denoted by (v,i,T,p), (wf , f , p ) , is obtained from

Ui = 0, f" = 0 < z < d, (3.1.5)

A
2
12^ = 0, T' = - - ^ ,
m -d <z<0,
m (3.1.6)

where A = kf/(p c )f 0 p and A m = k*/(p c )f.

0 p The hydrostatic pressures p andp™ may
1
be found from (3.1.2)i and (3.1.3)i respectively, but since p and p" are eliminated from
3.2. Perturbation Equations 48

our analysis later on, we do not include a derivation of them here. Integrating T" and
twice, and imposing the boundary conditions it can be deduced that

T { z ) = + l 2 + T o
(3.1.7)
- — { - ^ - + T '
m
f (z)= + ( ^ ~ )z + T . 0 (3.1.8)
Xm \ d m X m

The temperature at the interface, T , can be found by requiring continuity of temperature

0

at z = 0 and continuity of the heat flux, viz.

dT„
A ? = A, at z = 0. (3.1.9)
dz dz

If we define er = X / X and S m = CT/(CT + d) then from (3.1.9) one obtains the relation,
2 2
Qd , Qd d m m
S\Tu +
+ (3.1.10)
X X tTm

3.2 Perturbation Equations

To study the instability of the basic solutions (3.1.5)i, (3.1.6)i, (3.1.7) and (3.1.8), per-
turbations (ui, 6,7r), (U™, 8,
m 7r )
m to these solutions are introduced and we set,

Vi = Ui + Ui, T = f + 8, p = p + 7T,

V? = U? + U?, T m = f m + e , m Pm=Pm+^m-

The perturbation equations which arise from (3.1.2) and (3.1.3) are,

dui dui 1 dir _

— + Uj— = h uAm + ag9k u
at oxj p oxi
0

duj
= 0, (3.2.1)
dxi
Tu - T 0 Q
— +Ui— = XAd-w + { d 2 z )
at oxi ~ ^ - -x -

and,
Tl
1 dn
0 = if™ + ag9 ki, m
p dxi
0 K

= 0, (3.2.2)
dxi

{d + 2z)
m
at dxi dm X m
3.2. Perturbation Equations 49

m
where w = u , w
3 = v% and G m = (p c )*/(p c ) .
0 p 0 p f

To non-dimensionalise the above equations the fluid scalings of time, velocity, pres-
sure, and temperature are taken as,

d
r= -, u= ^ P = p^L j T # = ufdM
v a a Ay ga

The respective porous scalings are,

v ' d ' m d m A y m gaK

m
To ensure that T#, T # are real we impose Q, Q m < 0, i.e. we deal explicitly with a
heat sink in both layers. The length scales in the fluid and porous domains are d and d , m

respectively. In addition the fluid and porous Rayleigh numbers Ra and Ra m are defined
by
5 3
?2 _ -Qgad _ 2 -Q gaKd
m m
Ra, — R — ——, 2
Rfim — R,„ — m
uX ' i/A^
The Darcy number is defined to be 8 = \ f K j d and the fluid and porous Prandtl numbers m

m
are defined as Pr — v/X, Pr = v/X m respectively.
In terms of the non-dimensional variables, equations (3.2.1) and (3.2.2) become,

dui dui dir

dm
( 3 2 3 )
a £ = ° - ' '

and,
m
dir
2 m m
5 ~ - = -u™ + R 9 ki,
OXi
du™
7 T " = 0, (3.2.4)
OXi
QQm QQm
P r m G m + u T A 9 m + m m m
[ d ^ ~dx~)= R f (z)™ >
m
where f(z) = l~2z + (T -T )\/Qd v 0
2
and f (z) = T Xm/Qmd0
2
m - 2 z - l . Equations
2
(3.2.3) hold in the region {(x,y) G R } x {z G (0,1)} x {t > 0} while (3.2.4) hold
2 m
on {{x,y) G R } x {z G ( - 1 , 0 ) } x {t > 0}. The interface is z = 0 while the lower
porous surface is now z = — 1 and z = 1 is the (upper) fluid surface.
3.3. Linearised Instability Theory and Boundary Conditions 50

3.3 Linearised Instability Theory and Boundary Condi-

tions
The obtain linearised perturbation equations all quadratic terms in (3.2.3) and (3.2.4) are

discarded. The resulting system is linear so solutions can be sought of the form Uj =
at fft <rt at m m a< m
Ui(-x)e , 0 = 0(x)e , ?T = 7r(x)e , u? = u™(x)e , 6 = 0 ( x ) e , and 7r =
m <Ti
7r (x)e . This yields an eigenvalue problem for the growth rate a. The eigenvalues a , {

are ordered so that o\ > a 2 > . . . . I f o\ > 0 then instability occurs. The associated

eigenfunction {u^, u™^} yields the streamline curves which are included in sections

3.5.3, 3.5.4, and 3.6, to visualise the flow patterns at the onset of instability.
m
The time scale t = t/uj is introduced, then (3.2.3) and (3.2.4) yield the linearised

perturbation equations

aui = — 7Ti + Aui + R9ki,

uiti = 0, (3.3.1)

aPrO = A9 + Rf(z)w,

and,

2 m m
5 n™ = - ™ + u R0k h

u™ = 0, (3.3.2)
m m m m m
auPr G d
m m - A6 +R f (z)w .

The boundary conditions required for the case where the upper surface is open to the

atmosphere (tangential stress free) are, in non-dimensional form,

on z = 1, 8= w = w >zz = 0; (3.3.3)
m m
on z= -l, 9 =w = 0. (3.3.4)

Continuity of the normal components of velocity, temperature, and heat flux, across the

interface z — 0, leads to the conditions

„- » <r = LJl ,
t ° £ = s M » (3.3.5)
d er V d 9z V d dz 3
3.3. Linearised Instability Theory and Boundary Conditions 51

where d = d/d m and A = Q/Q . m In addition at z = 0, continuity of normal stress is

required,

fin™ = 7T - 2w, , z (3.3.6)

and the Beavers-Joseph boundary condition is employed,

^ =f V " ^ ) , 0 = 1,2, (3.3.7)

where a is a coefficient depending on the fluid and porous medium under consideration.
m
The pressures IT and 7 r are eliminated by taking curlcurl of (3.3.l)i and (3.3.2) . The x

third component of the resulting equations is retained to yield

2
a Aw = A w + RA*6, (3.3.8)

and
m m m
0 = Aw - R A*9 . (3.3.9)

where A* is the horizontal Laplacian,

2 2
A* d d
A = 1 .
2 2
dx dy
m m
Next a normal mode representation of w, 6, w and 8 is used in (3.3.8), (3.3.1) , (3.3.9),
3

and (3.3.2)3 by setting,

w = W(z)g(x,y), 9= 0{z)g(x,y),
m
wm = W m { z ) g m { x ^ ^
y Qm = e™( )g (x,
Z y),

m
where g(x, y) and g (x, y) are horizontal plan forms which satisfy,

2 m 2 m
A*g = -a g, A*g = -a g ,
m

where a and a m are the wave numbers associated with convection cells in the fluid and

porous medium, respectively.

The fourth order equation (3.3.8) can be written as two second order equations by
2 2 2 2 2
regarding A = (D — a )W as an independent variable, where D = d jdz . This yields

five coupled second order equations to determine the critical growth rate a, namely,
2 2
(D - a )W = A,
2 2 2
(D -a )A- Ra Q = a A, (3.3.10)
2 2
(D - a )0 + Rf(z)W = oPrO,
 2
3.4. The Chebyshev tau D Method 52

2
(D - al)W m
+ Ra e m 2
m
m
= 0,
(3.3.11)
2 m m
(D - 2
a )e m
m m
+ R f (z)W m
= auPr G e ,
m m

where (3.3.10) hold on z e (0,1) while (3.3.11) hold on z G (-1,0). Coupling is through

the boundary conditions.

The boundary conditions (3.3.3)-(3.3.7) are reduced to yield,

e = W = A = 0, on 2 = 1, (3.3.12)
m m
e = W = 0, on z = -1, (3.3.13)

and,

m m m
w = dw , e = —i/4e, De = 6\^De,

2 m
2 a £ W - D>1 - - M = -a£W, (3.3.14)

2
dot a
m
a W + A = — ( M - dDW ), on z = 0.
o

Hence, our goal is to solve (3.3.10) and (3.3.11) subject to the boundary conditions

(3.3.12)-(3.3.14). Before describing the numerical method we note the useful relations,

2
^- = d , Pr m = e Pr, T Ra =

2
3.4 The Chebyshev tau D Method
2
To solve the eigenvalue problem (3.3.10)-(3.3.14) the Chebyshev tau D method was em-

ployed. It was implemented in an entirely analogous fashion to that described in section

2.5.1. The five quantities W, A, 0, W m and 0 m are regarded as independent variables and

expanded as Chebyshev series, i.e.

N+2 N+2 N+2

W=Y,W T (z),
n n A = Y^A T (z), n n e = ^0 T„(2), n

n=0 n=0 n=0

N+2 N+2

W
Wm = Y , nTn(z), 6™ = £ B™T {z).
n

71 = 0 71=0
2
Equations (3.3.10), (3.3.11), and the operator D are then written in matrix form cf. Don-

garra et al. [23] and chapter 2. The boundary conditions (3.3.12)-(3.3.14) are imposed by
3.5. Discussion of the Model and Numerical Results 53

removing the m(N + 2) and m(N + 3) rows of the resulting matrix, m = 1,..., 5, and re-
placing them by the discrete forms of the boundary conditions. The boundary conditions
n _ 1 2
were realised with the aid of the relations T ( ± l ) = ( ± l ) n and T {±\) n = (±l)" n .
This results in a generalised 5(iV + 3) x 5(iV + 3) matrix eigenvalue problem of the form

Ax = aBx.

This was solved for the eigenvalues a'"), n = 1,2,... , and the eigenfunctions x^ with
the aid of the QZ algorithm which we employed via the NAG routine F02BJF as before.

3.5 Discussion of the Mode! and Numerical Results

Recall from section 3.1 that the steady state temperature profiles in the fluid and porous
layers, and the temperature at the interface are given by

T(z) = -—+ [ — j - + T J ^ o , (3.5.1)

f m { z ) = _Q^_ + (Z° _ 9 j ^ A z + T o > (3.5.2)

T o = e ( T v + 9 f + 9 ^ A . ( 3. .3)
5

^ A Xem T J

m
Differentiating (3.5.1) and (3.5.2) yields the turning points for T and T ,
zt = \ [{Tv - T )i + 1],
P 0 % = y [Totm - 1], (3.5.4)

2
where e = X/Qd and e m = A / Q e ^ . Since Q, Q
m m m < 0 the turning points are both
minimum points (?", > 0). To investigate what type of steady state temperature
3.5. Discussion of the Model and Numerical Results 54

profile can be obtained we begin by using (3.5.3) and (3.5.4) to derive the identities,

1
(3.5.5)
j
1
z <0tp & T> v -- > (3.5.6)
e
1 ( 2d
z™ > -d m & T v < (3.5.7)
e

J _ ( 2d
1
(3.5.8)
i
1 d
T > 0
0 <S> T v > (3.5.9)
e T

1 d
T < 0
0 Tv < (3.5.10)

Equations (3.5.5), (3.5.3) and (3.5.4) can be used to derive the relation
2

ztp > 0 & z%>0 T< 0 0.

Similarly (3.5.6), (3.5.3) and (3.5.4) yield 2

z <0 tp e> z™<0 & T > ~.

0

Hence only the following 4 types of steady state temperature profile are possible.

Case 1: The whole system is stable.

z
l Tu
d

=»-
^ ^ T o T

Figure 3.1: Steady state temperature profile, depth z against temperature T. zt < 0, zj£ < — d . v m

If Zty < —d , then (3.5.8) implies that T > 0 and z

m 0 tp < 0, and a profile like that dis-
played in figure 3.1 is obtained. Recall that density is a linearly, decreasing function of
temperature (see (3.1.1)) i.e., the warmer the fluid the lighter it is. Hence for the profile
3.5. Discussion of the Model and Numerical Results 55

given infigure3.1 the system is always stable.

Case 2: There is an unstable region in the porous layer, T > 0. 0

rf u

e Tio

Potentially unstable

Figure 3.2: Steady state temperature profile, depth z against temperature T, z tp < 0, — dm <
z f < 0, T > 0.
v 0

The profile given in figure 3.2 can be obtained by imposing (3.5.7) and (3.5.9) (note
(3.5.6) follows automatically). Fluid at the turning point z%, is colder and hence heavier
than that at the base z = —d . Therefore there is a potentially unstable region between
m

the turning point and the base of the porous layer.

Case 3: There is an unstable region in the porous layer, T < 0. 0

z \
(
Tu
1
a

T
Potentially unstable

Figure 3.3: Steady state temperature profile, depth z against temperature T. z% < 0, — d
p m <
z% < 0, T < 0.
0

The profile given in figure 3.3 can be obtained by imposing (3.5.6) and (3.5.10) (note
(3.5.7) follows automatically). Again there is a potentially unstable layer between the
turning point and the base since the fluid at the turning point is colder and heavier than
that at z = — d . m
3.5. Discussion of the Model and Numerical Results 56

Case 4: There is an unstable region in the fluid and underlying porous layer, T < 0. 0

Potentially unstable
0-
T,o T

•m

Figure 3.4: Steady state temperature profile, depth z against temperature T, z tp > 0, z^ > 0,
T < 0.
0

If z tp > 0 then (3.5.5) implies that T < 0 and z% > 0, and a profile like that displayed
0

in figure 3.4 is obtained. Fluid at the turning point z , is heavier than that below and
tp

hence there is a potentially unstable layer in the lowerfluidregion and underlying porous
medium (everywhere beneath the dotted line).
From the four cases presented above it is clear that penetrative convection is possible
in cases 2,3, and 4. Hence it is these cases that are of interest in the following analysis. It 11
is evident from the steady state temperature profiles that the position of the turning point \
determines the stability characteristics of the system. To investigate what effect Q and
Q m have on the system we assume that d, d , A, A and e are all fixed. Then e oc 1/Q
m m T

and e m oc 1/Q ,
m and we consider what effect e and i m have on the turning points z , tp z^.

3.5.1 Stability characteristics of Q

Eliminating T , (3.5.4) can be rewritten as

0

d{l-£) 1
+ 1 (3.5.11)
u
m
d£
m
m Tu + 1 (3.5.12)
m

By inspection of (3.5.11) it can be seen that as e gets less negative, z tp increases for fixed
i.
m In other words as the heat sink Q gets stronger the system becomes more unstable.
Similarly by inspection of (3.5.12), as e gets less negative, z^ increases forfixede (e /e m m

gets more positive). So again as Q gets stronger, the system becomes more unstable.
3.5. Discussion of the Model and Numerical Results 57

Hence Q is destabilising in cases 2,3, and 4, i.e. the stronger the heat sink in the fluid
layer the more unstable the system. Physically this makes sense in all three cases.
Consider case 2,fluidat the interface is warmer and lighter than that at the bottom of
the porous medium. The lighter the fluid at the interface, the more stable the system. If the
heat sink in the fluid is made stronger, the temperature at the interface will decrease and
the fluid there will become colder. Hence the heat sink is destabilising in this instance. In
case 3, fluid at the interface is colder (and hence heavier) than that at the base of the porous
medium. Therefore if the heat sink in the fluid layer is made stronger, the temperature at
the interface decreases and the difference between the fluid density at the interface and
that at the base of the porous medium increases. Since the fluid at the interface is heavier
than that at the base this is clearly destabilising. In case 4, fluid at the interface is heavier
than that at the top of the fluid layer. Therefore the whole of the fluid layer is potentially
stable. However if there is fluid within the layer that is colder than that at the interface
then there is a potentially unstable layer between the interface and the colder fluid. The
stronger the heat sink in the fluid, the more likely this is (for fixed Q ). Hence a heat sink
m

in thefluidis destabilising in this instance also.

To be in keeping with chapter 2, we suppose that the saturating fluid is water and
unless otherwise stated take, Pr = 6, e T — 0.7, G m = 10, a = 0.1 and 5 = 0.002. In
figure 3.5 the neutral curves for TV, d and e fixed at 9°C, 1 and —0.0728 respectively,
m

with e varying are presented. As e becomes less negative (Q becomes stronger) the neu-
tral curves shift down and the system becomes more unstable. This is expected and in
agreement with the analytical work presented above. Q clearly has a destabilising ef-
fect. Moreover, the neutral curves exhibit a bimodal nature like thatfirstseen in Chen &
Chen [17]. For e = —0.05, —0.0275 the minimum of the curve occurs at a m « 2 whereas
for e = —0.025, —0.015, a m « 12,8.5, respectively. Therefore as e becomes less nega-
tive the shape of the convection cells switches from wide (smaller a ) to narrow (larger
m

a ). Hence, as Q gets stronger convection switches from porous to fluid dominated. So

m

not only does Q dramatically effect the onset of convection it also effects the nature of
the ensuing convection cells. To illustrate this more clearly consider the corresponding
streamlines presented in the next section.
3.5. Discussion of the Model and Numerical Results 58

80-

75 -

70-

65 -

60-

55 -

50- -0.05
R
w
40-

35-

30-

25 -

20-

15-

10- -0.025

5- -O.015
0-
I
0 0 5 20 25 30
*1

Figure 3.5: Neutral instability curves, Ra m against a . Tu = 9°C, d = 1, i

m m — -0.0728, e
varies as shown on graph.

3.5.2 Variation of the streamlines with Q

In this section the streamlines at criticality, corresponding to figure 3.5 are presented i.e.
Try, d and e are fixed at 9°C, 1 and —0.0728 respectively, and e is varied from —0.05 to
TO

—0.015. Throughout all figures the porous-fluid interface is at z = 0, and the dashed line
marks the turning point of the steady state solution. Everywhere below the dashed line is
potentially unstable and everywhere above it is stable.
When e = —0.05 (figure 3.6) there is one cell bridging the porous-fluid interface and a
further five cells are seen in the fluid above. The cell bridging the interface is the initiating
cell. It is situated mainly below the unstable dashed line and hence drives convection.
The five cells in the (stably stratified)fluidabove are counter cells, a result of penetrative
convection. Note that the centre of the driving cell is situated above the dashed line. This
somewhat curious behaviour is a result of the resistance to movement the porous medium
offers. This is explained in more detail in section 3.6 where an example in which the
driving cell is not as strong as the counter cell above is given.
As Q is made stronger (figure 3.7) the initiating cell moves up slightly and only four
counter cells are seen in the fluid above. This is due to the fact that Q is destabilising.
The depth of the unstable layer has increased and hence the driving cell becomes deeper.
3.5. Discussion of the Model and Numerical Results 59

1 1 1 1 1 1 1
1.0

c
3
0.8

c 3 .

£ ^ '
-0.2

-0.8

10
1.6 1.2 -0.8
J
0.4
J1r
1.2 1.6

Figure 3.6: Streamline plot from the Q Figure 3.7: Streamline plot from the Q
model, T v = 9°C, = 1, 6 = 0.002, model, Tu = 9°C, d = 1, 6 = 0.002, i =
e = -0.05, e = -0.0728, Ra
m m = 29.155, -0.0275, e m = -0.0728, i ? a m = 14.112,
=
"ra 2.

P — i
M
H
to

-04 -0.2 0.0

Figure 3.8: Streamline plot from the Q

0.2 0.4
'-i

Figure 3.9: Streamline plot from the Q

-0.4
1 1
-0.2
B
1
0.0 0.2
1 r
0.4

model, T v = 9°C, d = 1, 8 = 0.002, model, T v = 9°C, d = 1, 6 = 0.002,

e - -0.025, e = -0.0728, Ra
m m = 8.839, e = -0.015, e = -0.0728, Ra m m = 0.713,
a m = 8.5.
3.5. Discussion of the Model and Numerical Results 60

The stably stratified fluid above is shallower and so fewer counter cells form within it. As
the strength of Q is increased again (figure 3.8) the size of the convection cells changes
to much narrower cells (the bimodal effect cf. figure 3.5). In this instance there are seven
cells almost completely confined to the fluid layer with slight movement into the porous
medium. An increase in the strength of Q has clearly resulted in a switch in dominance
from porous to fluid dominated convection. If Q is made even stronger (figure 3.9) there
is even less movement in the porous medium, just four cells are seen in the fluid layer
and the driving cell becomes deeper. Again this is due to the fact that an increase in the
strength of Q results in an increase in the depth of the unstable layer. Hence the initiating
cell gets deeper and the counter cells above adjust themselves accordingly (for a fixed
upper temperature) resulting in fewer counter cells.

In summary, figures 3.6-3.9 have shown that the onset and nature of convection are
very sensitive to change in the strength of Q. As the strength of the heat sink in the
fluid layer is increased, the fluid layer plays a much stronger role in the convection pat-
tern. Convection switches from porous tofluiddominated, the number of counter cells in
the system decreases (for a given mode of dominance) and the depth of the driving cell
increases.

3.5.3 Stability characteristics of Q m

By inspection of (3.5.11) it can be seen that as e gets less negative, z

m tp decreases for fixed
e (—e/e gets more negative). Therefore when the turning point is in thefluidlayer (case
m

4), the stronger Q m the more stable the system. This can be explained as follows. In case
4, there isfluidwithin thefluidlayer which is heavier than that at the interface. There is a
potentially unstable layer therefore between the interface and the colderfluidabove. The
stronger the heat sink in the porous medium the colder thefluidat the interface (for fixed
Q), and the smaller is the density difference between fluid at the interface and the colder
fluid above. Hence the heat sink in the porous medium is stabilising in this instance.

To investigate whether Q m is stabilising or destabilising when the turning point is in

the porous layer is slightly more complicated. Inspection of (3.5.12) reveals that the effect
of i m depends on the sign of (T + 1/e). Suppose (Tu + 1/e) < 0, then as e
v m gets less
negative, z™ decreases (forfixede). So in this case Q
t m is stabilising. On the other hand
3.5. Discussion of the Model and Numerical Results 61

70

160

150

140

30

120

10
0.01
100

90

80

70

60

50 0.1

40

30

20
-0.15
0
-0.25
1 1 1 1 1 1 r
0 5 10 15 20 25 30

Figure 3.10: Neutral instability curves, Ra m against a . Tu = 9°C, d = 1, e = -0.0357, e

m m

varies as shown on graph.

suppose (Tu + 1/e) > 0, then as e gets less negative, z% increases (for fixed e). So in
m

this instance Q m is destabilising. Note that (Tu + 1/e) < 0 43- T < 0, so Q 0 m can only
be stabilising in case 3 such that Tu < —1/e. In all other instances for which the unstable
layer is confined to the porous medium i.e. case 2 and case 3 with Tu > —1/e, Q m is
destabilising.
To illustrate this curious behaviour more clearly consider figures 3.10-3.13. In figure
3.10 the neutral curves for Tu, d and e fixed at 9°C, 1 and —0.0357 respectively, with e m

varying are presented. As e becomes less negative the neutral curves shift up. Therefore
m

as Q m gets stronger the system becomes more stable. This is in agreement with the
analysis presented above since Tu < —1/e, so we expect an increase in the strength of
Qm to have a stabilising effect on the system. In figure 3.11 the corresponding steady
state temperature profiles are displayed. It is clear that as e becomes less negative the
TO

turning point moves down. Note that the steady state temperature profiles are of type 3
and 4 only. Case 2 cannot be obtained since T violates (3.5.9).
v

In figure 3.12 the neutral curves for Tu, d and e fixed at 15°C, 1 and —0.1 respectively,
with e varying are presented. In this instance e has the opposite effect. As e becomes
m m TO

less negative the neutral curves shift down. Therefore as Q m gets stronger the system
becomes more unstable. This is in agreement with the analysis presented above since
3.5. Discussion of the Model and Numerical Results 62

1.0

OS

•0.01

-0.1

-06

-08
-0.25
1.0
-70 -60 50 -40 30 20 10 10

Figure 3.11: Steady state temperature profiles, depth z against temperature T. Ty = 9 C, d = 1,

0.0357, e vanes as shown on graph.
m

3200

2800

2400

2000

600

200

800

400 -0.01

0 0 5 20 25 30

Figure 3.12: Neutral instability curves, Ra m against a .

m T\j = 15°C, d = 1, e = -0.1, e,
varies as shown on graph.
3.5. Discussion of the Model and Numerical Results 63

-0

0-8

0.4

0.01

-0.4

-0 8

1.0 I 1
•65 55 -45 35 25 5 15

Figure 3.13: Steady state temperature profiles, depth z against temperature T. Tu = 15°C,
d = 1, e = —0.1, e varies as shown on graph.
m

Tu > —1/e, so we expect an increase in the strength of Q m to have a destabilising

effect on the system. Infigure3.13 the corresponding steady state temperature profiles
are displayed. It is clear that as e becomes less negative the turning point moves up.
m

Note that the steady state temperature profiles are of type 2 and 3 only. Case 4 cannot be
obtained since Tu violates (3.5.5).
It has been shown analytically and in figures 3.12 and 3.13, that in case 2, Q m is
destabilising. This is explained as follows. In case 2 there is fluid within the porous
layer which is colder than that at the base of the porous medium. There is a potentially
unstable layer therefore between the colder fluid and the base. As the heat sink in the
porous medium is made stronger the temperature of the fluid within the porous decreases
while that at the base remains fixed. Hence as the heat sink in the porous gets stronger
the unstable layer gets deeper (forfixedQ) and the heat sink in the porous medium has a
destabilising effect.

To explain the effect of the porous heat sink in case 3 is not so straight forward. For
fixed Tu, it has been shown that the effect of the porous heat sink depends strongly upon
the strength of the fluid heat sink. If the sink in the fluid is such that e > —1/Tu then
the sink in the porous is stabilising. On the other hand if the sink in the fluid is such
that e < —l/Tu (weaker) then the sink in the porous is destabilising. To clarify this we
3.5. Discussion of the Model and Numerical Results 64

differentiate (3.5.12) with respect to e to yield m

This implies that if e < -l/Tu

dz m
tp
de771

then
Ed,'771
2 ( 1
e )
is an increasing function of e . So as e ( < 0)
m m

increases z^ increases as well and Q m is destabilising. On the other hand if e > —l/Tu
then is a decreasing function of e . So as e m m increases z^ decreases and in this
instance Q m is stabilising. Clearly e = —l/Tu is a local maximum and the critical point
at which Q m switches from one behaviour to another. The strong interplay between the
two heat sinks clearly has a profound effect on the stability characteristics of Q m and
makes a straight forward physical explanation unclear in this case.
In summary, the stability characteristics of Q m depend strongly on the steady state
temperature profile and the strength of Q. In case 2, Q m is destabilising whereas in case 4
it is stabilising. In case 3, if Q is sufficiently weak (e < —l/Tu) then Q m is destabilising
whereas if Q is sufficiently strong (e > -l/Tu) it is stabilising. Figures 3.10 and 3.12
both show that as e becomes less negative (Q
m m stronger) the width of the convection
cell switches from narrow (a m larger) to wide ( a smaller). Therefore like Q, Q
m m has a
dramatic effect on both the onset and nature of instability. This is illustrated more clearly
in the next section where a selection of streamlines corresponding to figures 3.10 and 3.12
at criticality are presented.

In an entirely analogous analysis to that presented above the effect of varying d and
Tu can also be investigated. Inspection of (3.5.11) and (3.5.12) imply that d and Tu
are destabilising and stabilising parameters respectively. This is in agreement with the
numerical results of chapter 2.

3.5.4 Variation of the streamlines with Q rn

In this section the streamlines at criticality, corresponding to figures 3.10 and 3.12 are
presented. In figures 3.14-3.17, Tu, d and e are fixed at 9,1 and -0.0357 respectively, and
e is varied from -0.25 to -0.01 (cf. figure 3.10). When e = -0.25 (figure 3.14) there
m m

are six convection cells almost completely confined to the fluid layer with little movement
into the porous medium. If Q m is made stronger (figure 3.15) the convection cells are
again mostly confined to thefluidlayer but now there are seven cells in total. The increase
3.5. Discussion of the Model and Numerical Results 65

QQ

4 -0.2 0.0 0.2 0.4

Figure 3.14: Streamline plot from the Q Figure 3.15: Streamline plot from the Q
model, T v = 9°C, d = 1, S = 0.002, model, Tu = 9°C, rf = 1, 6 = 0.002,
e = -0.0357, e m = -0.25, Ra m = 1.317, e = -0.0357, e = -0.15, Ra m m = 5.534,
a m = 10. a m = 12.

1
1 1 1 1 1 1 1 =J 1 . 1

•c ^ r ^> "

y
Y

1 J,
•1.6
-i : i
= - -1.2
i -0.8
1 -0.4 r = 0.0
^
1
0.4
1
0.8
1 —
1.2
' -
1.6 -1.6 -1.2 -0.8 -0.4 0,0 0.4 0.8 1.2 1.6

Figure 3.16: Streamline plot from the Q Figure 3.17: Streamline plot from the Q
model, Tu = 9°C, d = 1, S = 0.002, model, Trj = 9°C, d = 1, S = 0.002,
e = -0.0357, i m = -0.1, Ra m = 15.262, e = -0.0357, e m = -0.01, i l a m = 46.107,
a m = 2. a m = 2.4.
3.5. Discussion of the Model and Numerical Results 66

in the number of cells is due to the fact that the depth of the unstable layer (everywhere
below the dashed line) becomes shallower with an increase in Q . m Recall that Q m is
stabilising in this instance. The stably stratified fluid above adjusts itself accordingly
resulting in an extra counter cell.

If the strength of Qm is increased further (figure 3.16) the convection pattern changes
altogether. The width of the cells becomes wider (the bimodal effect cf. figure 3.10)
and the porous medium dominates the convection pattern. In this instance there is one
cell bridging the porous-fluid interface and a further four counter cells are seen in the
fluid above. When the strength of Q m is increased again (figure 3.17), the initiating cell
still bridges the interface but now the centre of the driving cell is in the porous medium
(as opposed to the fluid layer), and there are five counter cells in the fluid above. The
shift in the position of the centre of the cell indicates that an increase in the strength of
Qm corresponds to an increase in the dominance of the porous medium. As previously
explained the number of counter cells has increased since the unstable layer has become
shallower.
Figures 3.14-3.17 clearly show that as the strength of the heat sink in the porous layer
is increased, the porous medium plays a much stronger role in the convection pattern.
Convection switches from fluid to porous dominated, the number of counter cells in the
system increases (for a given mode of dominance) and the depth of the driving cell de-
creases.
Infigures3.18 and 3.19 we concentrate on the behaviour of the streamlines when Q m

is destabilising. In this instance T , d and i are fixed at 15, 1 and -0.1 respectively,
v

and e varies from —0.5 to —0.01 (cf. figure 3.12). Note that in figure 3.18 penetrative
m

convection is witnessed in the porous medium. This is thefirsttime we have seen such be-
haviour in the two layer system and we highlight this novel result. The relatively shallow
depth of the unstable layer explains the occurrence of penetrative convection in the porous
medium. The driving cell in the lower region of the porous medium induces penetrative
convection in the stable porous medium andfluidabove. It is worth noting however that
the counter cells are very weak in comparison to the driving cell. Inspection of the eigen-
function reveals that the strongest counter cell (the one directly above the driving cell) is
approximately 200 times weaker than the driving cell below. The counter cells above it
3.6. Comparison with the Quadratic Density Model 67

0.8-

0.6-

0.4-
r
0.2-

>

w (ffnx\
-0.8

1.0
-I 1——i r - - — f — — i r = — -i 1-
-0.4 -0.2 0.0 0.2 0.4 -1.6 -1.2 -0.8 -0.4 0.0 0.4 0.8 1.2 1.6
X

Figure 3.18: Streamline plot from the Q Figure 3.19: Streamline plot from the Q
model, Tu = 15°C, d = 1, 6 = 0.002, model, T v = 15°C, d = 1, 6 = 0.002,
e = -0.1, e m = -0.5, Ra m = 1352.493, e = -0.1, e m = -0.01, -Ra m = 57.061,
am — 8.5. a m = 2.6.

are even weaker again.

In figure 3.19 the strength of Q m is stronger than that infigure3.18. In this instance
the convection pattern is completely different to that previously illustrated. The cells are
wider, the driving cell spans the porous-fluid interface andfivecounter cells are observed.
This shows that an increase in the strength of the heat sink in the porous layer corresponds
to increased movement in the porous medium. Note that there is no switch in dominance
between the two mediums in this case but the width of the cells does change from narrow
to broad with an increase in Q m (cf. figure 3.12).

3.6 Comparison with the Quadratic Density Model

In chapter 2, penetrative convection in the porous-fluid system was modelled via a quadratic
density profile. The governing equations differed in that those of chapter 2 did not include
an internal heat sink in either layer. Instead penetrative convection was modelled via a
3.6. Comparison with the Quadratic Density Model 68

quadratic density profile of the form

2
p= p [l-a(T-4) ].
0

Recall from chapter 2, the non-dimensional linearised perturbation equations (2.3.1) and
2
(2.3.2) (from now on called the T model),

aui = -itj + Aui - 2R 2M(z)6ki,

T

Ui = 0,
ti (3.6.1)

oPr9 = A0- R 2W, T

and,

2
5 TT™ = -< - 2Rjp M (z)9 k ,
2
m m
i

uZ = 0, (3.6.2)
m m m
auPr G 9 m m = A0 -Bjp w , 2

where

2 gad^mi-Sf 4/Tu - S
ti^2 — —, M(z) = £- z, t, = ,

KT2
- 2 ' M
\> z
^m
~ > z
^m
c •
IS Cs

From section 3.3 recall the non-dimensional linearised perturbation equations (3.3.1) and
3.3.2) (from now on called the Q model),

oU{ = —iXj + Aui + R9ki,

= 0, (3.6.3)

oPr9 = A9 + Rf(z)w,

and,

2 m m
5 n™ = -u? + R9k u

«M = 0, (3.6.4)
m m m m m
auPr G 9 m m = A6 +R f (z)w ,

where
m
f ( z ) = l-2z + (Tu- T )i,
0 f ( z ) = Ti 0 m -2z-l.
3.6. Comparison with the Quadratic Density Model 69

To investigate if and when these two systems are adjoint we follow an analysis similar
to the one given in Tracey [99]. First note that, the adjoint A*, of an n x n matrix A is
defined by the following relation

(x,Ay) = (A*x,y),

for all n vectors x and y. I f two systems are adjoint then the eigenvalues of both systems
are the same.
m m m m m T m m
Let $ = (7r,u,v,w,9,n ,u ,v ,w ,e ) , where u, u and v, v are the i, j -
m
components of u and u respectively. I f we identify 6 with —8 in the heat sink model then
the two systems can be written as A ^ T = crB$ (quadratic density law) and AQ$ = oB$
(heat sink), where the operators A 2, AQ, and B are defined by
T

/ 0 a a a 0 0
dx dy dz
a_ A 0 0 0
dx
_d_ 0 A 0 0
dy
d 0 0 A -2R 2M
dz T

0 0 0 —RT 2
A
A2
T

0 0 2
8 l 6 l 2
0
2
dx dy
: ... ... 5 f 1 0 0 0
ox
• n 0 1 0 0
m

n 0 0 1 2B% M 2

0 0 0 0 pm A

Ihrp2 )

0 d d d 0
dx dy dz
_a_ A 0 0 0
dx
_d_ 0 A 0 0
dy
d 0 0 A
dz -R
0 0 0 -Rf{z) A
AQ
0 0 2
5 f dy U
dz 0
2
ox
: 5 f 1 0 0 0
ox
<5 £ 2
0 1 0 0
dy m

n . 0 0 1 -R
0 0 0 0 m m
R f (z) A
3.6. Comparison with the Quadratic Density Model 70

( 0 0 0 0 0 0 \
0 1 0 0 0 0
0 0 10 0 0
B = 0 0 0 1 0 0

0 0 0 0 Pr 0

m
\ 0 ... coPr G m j
Now, let <&i and $ 2 be independent. It is not difficult to show that (A^t&x, <J>) = 2

{ * 1 , A Q * I ) iff

R =R,
T2 f(z)= 2M(z),
(3.6.5)
m m m
Rjp = R ,
2 f (z) = 2M {z).

Conditions (3.6.5) are satisfied iff

1 rp _ ,1-2^
e -
T

(3.6.6)
1Q — lu H —,

e
m — rp2c2> J
° —
; (3.6.7)
1 fc
- (7° m
Substituting (3.6.6)i into (3.6.6) , (3.6.7)! into (3.6.7) and eliminating T we find that
2 2 0

T v = 9°C.

Then choosing T v — 9°C and e, e such that (3.6.6)i, (3.6.7)i, are satisfied, we have
m

AQ = A 2 (where A*Q is the adjoint of AQ). In this case the two linearised models
T

give rise to the same instability boundary and we expect the two models to yield the same
critical eigenvalues. In table 3.1a comparison of the critical porous Rayleigh numbers and
wave numbers obtained from the Q model with T v = 9°C, and e, e m satisfying (3.6.6)i,
2
(3.6.7)i , with those from the T model with Tu = 9°C is made. Very good agreement
between the two is seen.
2
The eigenvalues from the Q and T models are in good agreement when the two are
adjoint, the corresponding eigenfunctions however are not. For example consider the W
eigenfunctions for the two models when d = 1, Tu = 9°C and i,e m satisfy (3.6.6)i,
(3.6.7) i , as illustrated in figure 3.20. Note that in computing the eigenfunctions we have
normalised over both W and W . The strongest cell therefore is always positive regard-
m

less of it's true circulation. Hence we cannot determine the true direction of circulation
3.6. Comparison with the Quadratic Density Model 71

d RCL™2 Ra% a
Q
0.05 273.311 273.322 5.0 5.0
0.10 225.470 225.631 4.5 4.5
0.50 61.942 62.844 2.5 2.5
1.00 19.902 20.233 2.0 2.0

Table 3.1: The critical porous Rayleigh number and wave number when the two models are
adjoint.
i
.0

0.8

W,W_

-0.2

-0.8
2
T
•"> T 1 1 1 1 1 1 1 1 1 r
-1.0 -0.8 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6 0.8 1.0
z

Figure 3.20: W eigenfunctions for the adjoint models, T = 9°C, d = 1. v

from the eigenfunction. We can conclude however that wherever a sign change occurs a
counter cell is present. Figure 3.20 clearly shows that the eigenfunctions exhibit different
behaviour. The main difference between the two is the position of the strongest cell. In
2
the Q model it is the first cell which is the strongest whereas in the T scenario it is the
second. To illustrate this more clearly consider the corresponding streamlines given in
figures 3.21 and 3.22.
In figure 3.21, T ,d
v and 6 are fixed at 9°C, 1 and 0.002 respectively. A streamline
2
plot obtained via the T model at the onset of convection is presented. In this instance
the dashed line represents the 4°C conduction solution (below which the fluid is unstable
since water has a density maximum at 4°C in this model). The plot shows one cell in the
porous medium just bridging the interface, with a further 5 cells in the fluid above. In
3.6. Comparison with the Quadratic Density Model 72

.0

c J
c
3
ad

2
0.2

a-i

-as
0 0 -0.8

0 i i i i
J
-a4 1.2 .6 .2 •0.8 r 00 0.8 .2 16
x

Figure 3.21: Streamline plot from the T 2

Figure 3.22: Streamline plot from the Q
model, Tu = 9°C, d = 1, 8 = 0.002, model, T v = 9°C, d = 1, S = 0.002, i,e m

i?a = 19.902, a
m m = 2. take adjoint values, Ra rn = 20.233, a m = 2.

figure 3.22, T[/, d and 5 are also fixed at 9°C, 1 and 0.002 respectively and e, i m take the
adjoint values (3.6.6)!, (3.6.7)i. A streamline plot obtained via the Q model at the onset
of instability is presented. In this case the dashed line represents the turning point of the
steady state solution (below which the fluid is unstable, see section 3.5). The two unstable
layers coincide exactly in the steady states. This can be proved analytically by simply
inspecting the steady state solutions of each model and evaluating them at the adjoint
values. The plot in figure 3.22 shows one cell bridging the fluid and porous medium with
2
a further 4 cells in the fluid above. Figures 3.21 and 3.22 clearly show that the Q and T
models generate different streamlines despite being mathematically adjoint and yielding
near exact eigenvalues. The physical processes that determine the depth of the unstable
2
layer (and hence drive convection) are very different in the two models. In the T model
it is the fact that density is a quadratic function of temperature whereas in the Q model it
is the inclusion of the heat sinks. Hence we cannot expect the streamlines to be the same
even when the two models are adjoint. The two models represent very different physical
systems and this must be understood in their application and interpretation at the outset.
Figure 3.21 is curious in that the cell which initiates penetrative convection (the lower
one confined mostly below the dashed line) is not as strong as the counter cell above.
One explanation for this is the resistance to movement the porous medium offers. For
3.6. Comparison with the Quadratic Density Model 73

-0.2 Ol

2 2
Figure 3.23: Streamline plot from the T Figure 3.24: Streamline plot from the T
model, T v = 9°C, d = 1, 6 = 0.01, Ra m = model, T v = 9°C, d = 1, 8 = 0.1, Ram =
18.534, a m = 2. 18.928, a = 1.8.
m

example i f the porous medium were more permeable, the fluid within would be freer to
move and we would expect the initiating cell to be stronger. To illustrate this notion
2
more clearly consider figure 3.23. In figure 3.23 a streamline plot generated via the T
model for Tu,d and 8 fixed at 9°C, 1, and 0.01 respectively, is presented. Recall that
5 — y/K/d , m where K is the permeability of the porous medium. So in figure 3.23 the
porous medium is more permeable than that presented in figure 3.21. In figure 3.23 we
note that the initiating cell is now the strongest and there are only 3 cells in the fluid
above. Clearly the increase in permeability has allowed the lower cell (mainly confined to
the porous medium) to dominate the convection pattern. To reinforce this consider what
happens when 8 is increased further (figure 3.24). In this case 8 = 0.1 and the initiating
cell is definitely the dominant cell. There is only one cell in the fluid above. Clearly an
increase in permeability has allowed the porous medium to have a much stronger role in
the convection pattern.

Conversely i f the porous medium were made less permeable we would expect the
lower (porous dominated) cell to become weaker and the motion in the fluid above to
become stronger. Close inspection of the eigenfunction when 8 = 0.0003 (keeping d = 1
and Tu = 9°C) revealed that in this case, there is one cell in the porous medium just
bridging the interface with a further 10 counter cells in the fluid above. The lowest counter
3.7. Concluding Remarks 74

cell was the strongest with the initiating cell below having a strength approximately 2/5
of that above. Hence as predicted, when the porous medium is made less permeable the
first counter cell can indeed be stronger than the initiating cell and greater movement is
seen within the fluid layer.
If the porous medium is made even less permeable it is found that convective motion
is almost completely confined to the fluid layer and the width of the convection cells
changes. Inspection of the eigenfunction when 8 = 0.0002, revealed that in this instance
there were 21 cells in the fluid layer. The initiating cell was the lowest and strongest in the
layer, and there was only slight movement into the porous medium. The critical porous
wave number was found to be a ro — 30.6 and so the cells much narrower than those
previously seen. It is worth noting at this point that care must be taken when interpreting
data from the neutral instability curves. It is often inferred from the critical mode what the
dominant layer is. For example i f the critical wave number is small, the porous medium
is said to dominate convection whereas if it is large the fluid layer is believed to, cf. Chen
& Chen [17]. To avoid confusion exactly what is meant by dominant must be understood.
The dominant layer refers to the layer in which instability commences. The layer contains
the initiating (driving) cell but as illustrated this may not necessarily be the strongest cell.

Equations (3.6.1)-(3.6.4) clearly show that the Prandtl number Pr, is a parameter of
the governing system. However it is worth noting that when Pr was varied from 0.001-
1000, no significant difference was seen in the critical eigenvalues obtained from either
model.

3.7 Concluding Remarks

In this chapter we have modelled penetrative convection via internal heating for a two
layer system in which fluid overlies and saturates a porous medium. It was found that
internal heating must be via a sink and a sink was included in both layers. The instability
of the system was found to be very sensitive to change in the strength of the sinks. It
was shown analytically and numerically that the sink in the fluid layer had a destabilising
effect on the system. The effect of the sink in the porous medium however depended upon
what type of steady state temperature profile was being considered and on the strength of
3.7. Concluding Remarks 75

the heat sink in the fluid layer. An array of streamlines were presented to illustrate the
effects of varying Q and Q m on the pattern of instability.
The model presented in this chapter was compared to that given in chapter 2. When
mathematically adjoint, the two models were shown to yield the same critical instability
boundaries but different eigenfunctions. Hence the Q model confirmed the instability
boundaries of chapter 2 and a need for caution and understanding, when interpreting
2
physical results was highlighted. Physically, the Q model is very different to the T one
and we stress this point.

The linear instability analysis presented here and that given in chapter 2 provide in-
formation about the nature and onset of instability. However they do not preclude the
possibility of subcritical instabilities. A nonlinear analysis is necessary to assess the va-
lidity of the linear work. Straughan & Walker [96] pointed out that a Q model can make
a nonlinear energy analysis more straightforward than one involving a quadratic equation
of state. In future work it is our intent to develop a nonlinear stability analysis for the
two layer problem. It is hoped that the mathematical properties of the Q model can be
utilised in this task. The favourable properties of a Q model are discussed in more detail
in the following chapter and the possible application to a nonlinear analysis of the two
layer system is discussed in chapter 7.
Chapter 4

Penetrative Convection in a
Horizontally Isotropic Porous Layer

In this chapter penetrative convection in a horizontally isotropic porous layer is investi-

gated. A horizontally isotropic porous medium is one in which the permeability is the
same in all directions in the horizontal plane, but different in the direction of the verti-
cal axis. The effect of anisotropy on convection in porous media is a subject receiving
recent attention. For example, Qin & Kaloni [80] studied the onset of Rayleigh-Benard
convection in a porous layer, using Brinkman's equation and anisotropic permeability.
In general, they allowed different components of permeability in all three principal di-
rections. However, they do briefly investigate the horizontally isotropic porous medium
case as well. Straughan & Walker [96] dealt with a porous medium for which the per-
meability was transversely isotropic with respect to an inclined axis. They found that the
growth rate a was real in the anisotropic case without penetrative convection as well as
in the isotropic case with penetrative convection. In this chapter we show that for a hori-
zontally isotropic medium (anisotropic configuration) the growth rate a is real even with
penetrative convection.

It is our intent to model penetrative convection via internal heating and a quadratic
equation of state. For the majority of the analysis we choose to describe penetrative con-
vection by internal heating. The motivation for this choice is the fact that it allows us to
perform a non weighted energy analysis to obtain a global stability bound. The system
is also modelled via a quadratic density law since the results it produces are easier to in-

76
4.1. Internal Heat Sink Model 77

terpret physically than those obtained via internal heating. The two models are compared
and it is shown that the linearised forms of each are in fact mathematically adjoint and
therefore, the instability boundaries which can be obtained from the models are the same,
cf. Tracey [99] and chapter 3.

Straughan & Walker [96] dealt primarily with the quadratic density model, however,
they also gave a brief development of a model with a uniform heat sink. In this chapter
the same configuration is taken, but it is simplified to a horizontally isotropic medium.
This model is of physical relevance to real life situations since many geothermal rock/soil
structures display anisotropy where the permeability in the vertical direction is different
to that in the horizontal plane. An example of such anisotropy is soil lofting due to frost
heave. This plays an integral part in the formation of patterned ground, cf. George et
al. [33]. Another example is the permeability of sea ice, cf. Eicken et al. [27]. Possible
applications of modelling sea ice as an anisotropic porous medium are discussed at length
in chapter 7. Further anisotropy effects can be found in Nield & Bejan [73].
An outline of this chapter is as follows. To begin with an internal heat sink model
which allows for penetrative convection in a horizontally isotropic porous medium is in-
troduced. It is then shown that for the linearised system the growth rate o is real. Using
this result an instability boundary is found for the linearised system. After this an energy
analysis is performed to obtain a global nonlinear stability bound. The model is then
compared with one in which the density is assumed quadratic and it is shown that the two
linearised systems are mathematically adjoint. Finally, a presentation and discussion of
the numerical results is given in the last section.

4.1 Internal Heat Sink Model

Consider a porous layer saturated with fluid contained between the horizontal planes z =
0 and z = d. The z direction is denoted by the vector k with i, j , k being the standard
Cartesian basis. Gravity acts in the negative z direction and the upper and lower planes are
held fixed at temperatures T and 7} respectively, see figure 4.1. The density is assumed
u

to have a linear temperature dependence of the form

p = [l-a(T-T )},
Po 0 (4.1.1)
4.1. Internal Heat Sink Model 78

where p, T, and a are density, temperature and thermal expansion respectively, and p 0

and T are constant reference values for the density and temperature respectively.
0

z =d

Saturated porous medium

z =0
Ti

Figure 4.1: Schematic diagram of the governing system.

Attention is restricted to the anisotropic case for which permeability in the k direction
is different to that in the i and j directions. Hence the permeability tensor K is chosen to
be of the form
( h 0 0 ^
K
o h o
\ o o k j 2

where clearly ki is the permeability in the i and j directions and k is the permeability in
2

the k direction. The inverse permeability tensor M is given by

/ £ 0 0 ^
l
1
M — K~ = 0 £ 0

V0 0 1

where f = k /k\
2 is the anisotropy parameter.
The fluid motion in the porous layer is assumed to satisfy Darcy's law with a Boussi-
nesq approximation,

1 dp gp(T)ki
vMijVj (4.1.2)
~dt p dxi
0 ' " - p
0

where Vi,t,p,Xi,v, and g, are velocity, time, pressure, displacement, kinematic viscos-
ity, and gravity respectively. Standard indicial notation and the Einstein summation
T
convention are employed throughout, and as previously mentioned k = (0,0,1) . We
choose to neglect inertia again and henceforth discard the v ijt term in (4.1.2). The con-
stant term in (4.1.1) is absorbed into the pressure by defining a modified pressure p to
4.1. Internal Heat Sink Model 79

be p + <7p (l +
0 OLT )Z.
0 Then, dropping the hat and neglecting the inertial term, (4.1.2)
becomes
p = -uMijVj
ti + p agkiT,
0 (4.1.3)

where p, is the dynamic viscosity. Equation (4.1.3) together with the incompressibility
condition and the equation of energy balance yield the following system of governing
equations

p,i = -fiMijVj + p agkiT,

0

viti = 0, (4.1.4)

T + ViTi = KAT + Q.
t

Here, K is the thermal diffusivity, Q is some (constant) internal heat source or sink and A
is the Laplace operator. The upper and the lower boundaries of the system are assumed
fixed, so the boundary conditions v (0) = v (d) = 0, T(0) = T
3 3 h T(d) = T , must be
u

imposed.
Consider the integral form of the energy balance equation (4.1.4) , namely, 3

d dT
I TdV+ [ v dV QdV, (4.1.5)
J\v dx
d V +
dt Jv f dxi idxi) j
v v
V
where V is some periodic cell in the saturated porous layer z £ (0, d). Using the boundary
conditions and divergence theorem, (4.1.5) can be simplified to find

^ [ TdV = K(f n ~ d A + [ QdV,

at Jv J a v dxi Jv
where rij is the unit normal to V and dV the bounding surface of V. Note that, in the
absence of any heat flux, a positive value of Q will give rise to an increase in temperature
of the layer and similarly, a negative value of Q will yield a decrease. Therefore a positive
value of Q represents an internal heat source and a negative value of Q an internal heat
sink.
The basic steady state solutions of (4.1.4) in whose stability we are interested, are
obtained from
,r. n rFitf Q
Vi = l), 1 = .
AC

The hydrostatic pressure p may be found from (4.1.4)j but since p is eliminated from our
analysis later on, a derivation of it is not included here. Integrating T" twice yields
TV ^ Q 2 , (Tu - T, , Qd\
4.1. Internal Heat Sink Model 80

For the system to be in a potentially gravitationally unstable state the condition 7} > T > u

Tmin where T m i n is the minimum temperature of the fluid is imposed. Furthermore, we

take Q < 0, i.e. we deal explicitly with the case of a heat sink. In this situation the lower
part of the layer has a tendency to be gravitationally unstable whereas the upper part is
stable. Hence penetrative convection can arise when the fluid in the unstable layer begins
to convect and induce motion in the upper part.

To study the stability of (4.1.4) perturbations to the steady state solutions are intro-
duced as follows

Vi = Vi + Ui, T = f + 9, p=p+ ix.

The perturbed expressions are then substituted into (4.1.4) and non-dimensionalised with
scalings of

X = d, T = —, U= -7, P = fAK, IT*) = ,

a
K a Po 9
for displacement, time, velocity, pressure and temperature respectively. Then (4.1.4) may
be written in terms of the non-dimensional perturbation quantities as

n i = —MijUj + Rk{9,
7

uiti = 0, (4.1.6)

0,t + Ui0,i = Rf{z)w + A0,

where

T u l ) K 2
f{z) = ± - e - z , e=-( Q J , w = u, 3 x € R x [0,1],

and
R 2 = _Qp agl^ £

defines the so called Rayleigh number. The boundary conditions for the perturbed vari-
ables are
w(0) = w(l) = 0, 9(0) = 9(1) = 0. (4.1.8)
4.2. Linearised Instability Theory and The Principle of Exchange of Stabilities 81

4.2 Linearised Instability Theory and The Principle of

Exchange of Stabilities
at
Neglecting the nonlinear terms in (4.1.6) and assuming a temporal growth rate like e we
obtain the linearised system

7r i = —MijUj + Rktd,

- 0, (4.2.1)

<70 = Rf(z)w + A9.

To eliminate the pressure in (4.2. l ) i , curlcurl of both sides is taken. Then, setting i = 3,
yields
M Aw - M
3 3 j j U j t 3 j - RA*9 = 0, (4.2.2)
2 2 2 2
where A* is the horizontal Laplacian, A* = d /dx + d /dy . Finally, using the incom-
pressibility relation (4.2.1) in (4.2.2), system (4.2.1) can be written as follows
2

Cw = RA*9,
(4.2.3)
{A — a)9 — -Rf(z)w,
2 2
where the operator C is defined as A* + £d /dz .
To show that a e K, let a = ay + oj,, where a and Oi are both real. Applying A* to T

(4.2.3) and substituting the expression for A*9 from (4.2.3)i, the system can be rewritten
2

as a single fourth order equation in w, namely

2
(A - a) Cw = -R f{z)A*w. (4.2.4)

The boundary conditions are w = w tZZ = 0 for z € { 0 , 1 } . Let V now be some periodic
2
cell and introduce the complex L -inner product as follows

(u , v) = / uv*dV,
Jv
where the asterisk here denotes complex conjugation. In general, (.) will mean integra-
tion over V. After multiplying equation (4.2.4) on the right by the conjugate of w and
integrating over V, there results

£ (Aw , tZt w*) + (AA*w, w*) - oi (w , w*) iZZ

2
- a (A*w, w*) = -R (f(z)A*w, w*).
4.2. Linearised Instability Theory and The Principle of Exchange of Stabilities 82

By repeated use of Gauss' theorem and the boundary conditions w = w tZZ =w iX =w >y =
0 on z e {0,1} , it can be shown that

2
e | | V ^ | | + ||V*Vw|| + 2
<7<£|KH 2
+ <r||V*H| 2
= r 2
< / ( * ) V V V*w*). (4.2.5)

Finally, taking the imaginary part of (4.2.5) yields

(ell w \\
2 2

Oi tZ + ||V*w|| ) = 0

which implies that Oi = 0. Hence, a is real and we say that the principle of exchange of
stabilities holds. When this holds convection sets in as stationary convection.
Suppose then o = 0, equations (4.2.3) yield
2
d
2 w RA*9
dz
(4.2.6)
d
2
9 Rf(z)w.
dz
Assuming a normal mode representation for w and 9 of the form w = W(z)g(x, y) and
2
9 = Q(z)g(x, y), where g(x, y) is a horizontal plan form satisfying A*# = -a g, (4.2.6)
becomes
2 2 2
(££> - a ) W = - i ? a 6 ,
(4.2.7)
2 2
(£> - a ) 6 = -Rf(z)W,
2 2 2
where D = d /dz and a is the wavenumber.
This fourth order system for R was solved numerically by two entirely different meth-
ods. Firstly the compound matrix method was employed as described in Straughan [94],
and appendix A.2. Secondly the Chebyshev-tau QZ-algorithm as advocated by Dongarra
et al. [23], and outlined in chapter 2 and appendix A . l was used. Without going into
detail, the basics of both routines are now described for the system at hand.
r
For the compound matrix method, the vector W is defined as W = (W, W, 0 , 0 ' ) .
Let W i and W be independent solutions obtained by replacing the boundary condition
2

T T
at z = 1 by the initial conditions W (0) = (0,1,0,0) and W (0) = (0,0,0,1) . We
x 2

introduce the variables

W W!2 y = we-
2 1 2 we 2

y = e'
3 Wl 2 - w e[, 2

2/6 = 0 x 6 2 - 6 2 6 ; .
4.2. Linearised Instability Theory and The Principle of Exchange of Stabilities 83

Then the compound matrix equations for (4.2.7) are

V\ = - y * t y 2 , Vi = 2/3 + ?/4,
2
a

2
2/3 = 2/5 + a 2/2, 2/4 = 2/5 + -^2/2,

2
2/5 = « Q2/3 + 2/4^ + R (j(z)yi - y2/e
R z
2/e = f ( ) V 2 -

This system was numerically integrated subject to the initial condition y (0) = 1. The 5

value of R was varied such that the condition 2/2(1) = 0 was satisfied to a predefined
2
accuracy. Furthermore, we minimised over a to find the critical Rayleigh number which

is denoted by R\ henceforth.
2
Chebyshev-tau D and D routines were also implemented to solve (4.2.7). First, the

solutions W and O are expanded in a series of Chebyshev polynomials

N N
9
W = J2w T (z), n n 0 = Y, nT (z).
n

n=0 n=0

2 2
The second order differentiation operator obtained from writing out d /dz in the Cheby-
2
shev representation is denoted by D . The boundary conditions for (4.2.7) are imposed
2
by removing rows from the D matrix. Furthermore, let Z denote the Chebyshev matrix

represention of the function z (see chapter 2). Then, the resulting generalised eigenvalue

problem is

2 2 W
4££> -a 1 0 ^ I ^ RI °
0 2
4D -a I 2
J \ & J \eI+\Z
\ eI+\Z 0 j \ 0
(4.2.9)
T T
where W = (w ,WN) 0 and 0 = ( 9 , 6 N ) . The matrix system was solved using
0

the QZ algorithm. The critical Rayleigh number is approximated by the square of the

smallest generalised eigenvalue of (4.2.9). A Chebyshev tau D method, as described in

chapter 2, was also employed as an independent check. A presentation of the numerical

results is delayed until section 4.5. For the moment however, we remark that excellent

agreement was found between the three aforementioned numerical methods.

4.3. Global Nonlinear Stability via the Energy Method 84

4.3 Global Nonlinear Stability via the Energy Method

Consider the nonlinear system (4.1.6)

7T i = —MijUj + Rki6,
t

uiti = 0, (4.3.1)

9, + u e
t i ji = Rf(z)w + A9,

with boundary conditions

« (0) = i* (l) = 0,
3 3 0(0) = 0(1) = 0.

Multiplying equation (4.3.l)i by Ui, (4.3.1) by 9 and integrating both equations over V,
3

yields

R (9w) - (MijUiUj) = 0,

j \j\n J - R ( / ( ^ )
t
2
J + I I V C = O.
Consider the thermal energy of the perturbations

E(t) = ±\\er.

Multiplying (4.3.2\ by some coupling parameter A ( > 0) and adding this expression to
(4.3.2) , the following relation is obtained
2

dE ^
m + D
n = >

where the sign indefinite and definite terms are respectively,

2
I = ( ( f ( z ) + A) w6), D — | | W | | + A (MijUiUj).

Next, RE is defined by
1 I
—— = max —,
R E H D'

where % is the space of admissible solutions over which we seek a maximum, namely,

2
n = {ui,eeC {0,l)\u 3 = 9= 0 when z = 0,l}.

This definition leads to the following relation

4.3. Global Nonlinear Stability via the Energy Method 85

Using Poincare's inequality (appendix B . l ) and the fact that is positive definite, it can
be shown that
2 2 2
D > 7 r | | 0 | | = 2n E. (4.3.4)

Then using (4.3.4) it can be deduced from (4.3.3) that

2 2 7 r 2 a
^ < -2n aE E{t) < £ ( 0 ) e " ',

where a = (R E - R)/R .
E So for all R < R , E(t) -> 0 as t ->• oo, where convergence
E

is at least exponential. Hence R E provides a global nonlinear stability bound.

To find the Euler-Lagrange equations for the maximum, a calculus of variations tech-
nique is employed as follows. Suppose 9 and Ui are the maximising solutions and consider
6+eT] and Ui + iVi, where 77 and i> are arbitrary functions which satisfy the same boundary
f

conditions as 9 and Ui respectively, and e is some constant. It can be easily shown (cf.
(1.5)) that the maximising solution satisfies

R 8IE -6D = 0, (4.3.5)

where

Substituting 9 + er/ and U{ 4- evi into (4.3.5) it can be deduced that

/ [R ( f ( z ) + A) (1077 + v 9) - XMij ( Vj + vtUj) + 2r]A0] dx = 0.

E 3 Ui (4.3.6)
Jv

Apart from the continuity and boundary conditions they must satisfy, 77 and are arbitrary
functions. Therefore adding the zero term j v -K Vidx = 0 in (4.3.6), and letting 77 = 0 in
ti

the first instance, and Vi = 0 in the second yields the Euler-Lagrange equations

= RE { f ( z ) + A) 9ki - 2XM u ,
ij j

(4.3.7)
2A9 = -R E ( f ( z ) + X)w.

The Lagrangian multiplier ir in (4.3.7)i can be eliminated by taking the curlcurl of both
sides. Setting 7 = 3 and introducing normal modes yields the system

2 2 2
2A (£L> - a ) W = -R E ( f ( z ) + A) a 6 ,
(4.3.8)
2 2
2 (D - a ) 0 = -R E ( f ( z ) + A) W,
4.3. Global Nonlinear Stability via the Energy Method 86

with boundary conditions W = 6 = 0, for z G { 0 , 1 } .

The compound matrix equations for (4.3.8) are
2
_ a {f(z) + X)
y = -R
x E ^ y, 2 2/2 = 2/3 + 2/4,
2
a
2
2/3 = 2/5 + « 2/2, 2/4 = 2/5 + y2/2,

2 / 1 1 / 6
2/5 = a ^-2/3 + 2/4J + RE y g 2A^
, f(z)+ A
2/ = -RE-
6 ^ 2/2-

This system was solved in an entirely analogous fashion to (4.2.8), i.e. subject to 2/5(0) =
2
1 and 2/2(1) = 0 and minimised over a . This time however, we also maximised over
the coupling parameter A to make R E as large as possible. When maximising over A, it
is important to have a reasonable initial guess. It is worth noting at this point that it is
expedient to put Uj = \/Xuj and %j = -Kj/y/Xm (4.3.7). Then, with

f(z)+ A
*(*,A) = — y ^ - ,

the Euler-Lagrange equations become (dropping the hats)

7r,i = REgQki - 2MijUj,

(4.3.9)
2A0 = -R gw.
E

By employing parametric differentiation it may be shown that

R E + 9 e w = 0
d x - d x - ) ) -

At the best value of A, 8RE/0X = 0, so we set dg/dX — 0 to obtain a good guess for A
to be -e ( > 0).
2
The Chebyshev-tau D system for (4.3.8) is

2 2
8A£D -2Aa J 0 \ / W
2 2
0 8D -2a I ) \ 0
2
= R ( 0 a {{e-X)I+\Z)
E
\ { e - X ) I + \ Z 0

This system was solved numerically and the results used to check those obtained by the
compound matrix method. Again, excellent agreement was found.
4.4. Comparison of the Heat Sink Model with a Quadratic Density Model 87

4.4 Comparison of the Heat Sink Model with a Quadratic

Density Model
An alternative way to model penetrative convection in the porous medium is via a quadratic

density profile. Consider a porous medium like before, but now fix the temperature at the

lower boundary at 0°C and maintain the temperature at the upper boundary at 7\ ( > 0°C).

Let the density have a quadratic temperature dependence of the form

p(T) = p [1 - a (T - T f ] .
0 0

Then the governing equations are, cf. Straughan & Walker [96],

2
p,i = -fiMijVj + p agki
0 (T - 2T T),0

Vi,i = 0, (4.4.1)

T + ViTi = KAT.
t

The basic steady state solutions of (4.4.1) are

^ = 0, -T = Pz,

where /3 = T i / d . Introducing perturbations 7r) to the velocity, temperature and

pressure and non-dimensionalising with scalings of

X - d , T = £ , U== P = ^ v
(T*f /
= - ^ i - ,
K a 2ap gTi
0

leads to

7r j = —MijUj — R 2 (rj — z) k{6 + ^r9 ki,

T
2

uhl = 0, (4-4.2)

01 + u f o = -R 2W T + Ad,

where r\ = T jT\,
0 and the Rayleigh number is defined as

D2 _ Zpoagd
IUp2 — i 1 .
PK

The boundary conditions for this system are u = 9 — 0, for z £ { 0 , 1 } . Since we want
3

2 2
to compare this system to the heat sink model, we choose T x = —Qd /2K, to give the
Rayleigh number the same interpretation as before (see (4.1.7)).
4.4. Comparison of the Heat Sink Model with a Quadratic Density Model 88

Neglecting the nonlinear terms in (4.4.2) and assuming a temporal growth rate like

e , yields

rr,i = -MijUj - R(r) - z) hrf,

Ui,i = 0, (4.4.3)

o9 = -Rw + AO.

To show that system (4.4.3) is adjoint to that described in (4.2.1) an analysis very similar

to the one given by Tracey [99] for a fluid layer is performed.

T
Let <I> = (ir, U, v, w, 9) , where v is the j-component of u. If we identify 9 with —9 in

the heat sink model, the two systems (4.2.1) and (4.4.3) can be written as AQ$> = oBQ

(heat sink) and A z& T = oB& (quadratic density law) respectively, where the operators

AQ and A 2 are defined by

T

' 0 a_ a_ d
dx dy dz
a_ M n 0 0 0
dx
= d_ 0 0 0
dy M 2 2

a 0 0 R
M 3 3
dz A
K
1 0
0 0
d_
0
d_
!(z)R
d
J
0
dx dy dz
a_ M n 0 0 0
dx
A2 = d_ 0 0 0
T
dy M 2 2

a 0 0 M 3 3 (V - z)R
dz
\ o 0 0 R A

and

Now, let * i and # 2 be independent. It is not difficult to show that (j4Q<I>I, < & )
2 =

(&\,A 2$ )
T 2 iff f ( z ) =V-z, i.e. e = 1 / 2 - 4 / T i . For this choice A 2 = A* (where A*
T Q Q

is the adjoint of AQ), implying that the two linearised models have the same eigenvalues
and hence give rise to the same instability boundary. Putting a = 0 and assuming a
4.5. Numerical Results and Conclusions 89

normal mode representation, (4.4.3) can be rewritten as follows

2 2 2
{ZD - a ) W a R{rj - z)G
(4.4.4)
2 2
(D - a) e RW.

2
The Chebyshev tau D method applied to (4.4.4) gives the following generalised
eigenvalue system

Z z
4£D - aI 0 W

0 AD an 0
2
0 a (( V -\)I-\Z) W
R
0 0

It was found (as expected) that this system yields exactly the same critical Rayleigh num-
bers as those obtained from the linear analysis using the heat sink model.

4.5 Numerical Results and Conclusions

In this section, subscript L will be used for results from the linear analysis (section 4.2),
while E will indicate results obtained via the energy method (section 4.3). In tables 4.1
and 4.2, results are generated using the compound matrix method and checked via the
Chebyshev-tau QZ routine. For the precision used in the tables, no significant difference
could be found between the two sets of results and so only a presentation of the results
obtained via the compound matrix method is given.
Table 4.1 displays the critical values (RL, a|, a | ) for the upper temperature, 7\
fixed and the anisotropy parameter £ varying. T is fixed at 4°C in the quadratic density
X

model, which corresponds to a value of e = —1/2 in the heat sink model. For all values
of £ in the table we found A & 0.54, which is reasonably close to its estimated value of
-e. We conclude from table 4.1 that for this parameter range, the critical values obtained
from the linear analysis are very close to those obtained via the energy analysis i.e. the
linear instability and nonlinear stability thresholds are very close. Hence the linear results
give a good physical representation of what is happening and it is unlikely that subcritical
instabilities exist. However it is worth noting at this point that as £ increases the agreement
between the linear and nonlinear results is not as good as it is for smaller values of £. We
4.5. Numerical Results and Conclusions 90

RL RE «! a\

0.2 6.34800 6.29303 4.58564 4.61381

0.4 7.16423 7.10451 6.46601 6.49883
0.6 7.78941 7.72536 7.91138 7.94873
0.8 8.31594 8.24790 9.13197 9.17390
1.0 8.77950 8.70776 10.2090 10.2555
2.0 10.5962 10.5086 14.4496 14.5197
3.0 11.9886 11.8878 17.7190 17.8129
5.0 14.1946 14.0716 22.9282 23.0691

Table 4.1: Critical parameters of linear instability and nonlinear stability with T\ = 4°C.

also observe that as £ increases the Rayleigh number increases too. This implies that the
system becomes more stable as £ increases. The most interesting behaviour in table 4.1
however, is that as £ increases, the wavenumber increases quite rapidly. This suggests
that the convection cells are becoming narrower as the permeability in the z direction is
increased.

In table 4.2, £ is kept fixed while e (and thus 7\) is varied. Note that as 7\ increases
from 1°C to 8°C, e tends to zero. Physically, as Ti is increased in the quadratic density
model, we could say that in the heat sink model Q becomes stronger, while the boundary
conditions remain fixed. The critical Rayleigh number increases with 7\, and thus the
system becomes more stable as the upper temperature is increased. This is expected and
in good agreement with Straughan & Walker [96]. Note that for e close to zero, the linear
results do not agree with the nonlinear results quite so well. In particular, the differences
in the wavenumber become larger. We also note, that as 7\ increases the estimated value
of A is not as good.
Figure 4.2 shows the Chebyshev expansion of W(z) for the first 30 polynomials in
the isotropic case. The approximation is based on the quadratic density model. We only
show profiles obtained with this model because the physical interpretation of the results
obtained with the heat sink model was not clear. Although for systems (4.2.7) and (4.4.4)
the eigenvalues are the same, the eigenvectors are not (cf. chapter 3). Since the quadratic
density model provides a more tangible representation, we present these results and high-
4.5. Numerical Results and Conclusions 91

Ti(°C) e RL RE 4 a| A

1.0 -3.500000 3.35766 3.35702 9.87671 9.87831 3.50618

2.0 -1.500000 5.12318 5.11797 9.90824 9.91653 1.51434
3.0 -0.833333 6.85262 6.83077 9.99394 10.0177 0.858793
4.0 -0.500000 8.77950 8.70776 10.2090 10.2555 0.541011
5.0 -0.300000 11.1114 10.9037 10.7667 10.7905 0.363036
6.0 -0.166667 14.0823 13.5485 12.3140 11.9279 0.259922
7.0 -0.071429 17.7073 16.6903 16.1375 14.0611 0.201908
8.0 -0.000001 21.7114 20.2574 21.8574 17.3643 0.169258

Table 4.2: Critical parameters of linear instability and nonlinear stability with £ = 1.

light the need for caution when interpreting results found via the heat sink model. For
Ti less than or equal to 6°C, W{z) is positive for all z. In this case there is only one
convection cell, the centre of which is located at the maximum of W(z). For 7\ greater
than or equal to 7°C however, W(z) changes sign indicating the existence of a counter
cell and hence the onset of penetrative convection. As 7\ increases from 7°C to 7.9°C,
the boundary layer effect near z = 0 and the second cell become stronger.

Figure 4.3 shows the Chebyshev expansion of 9(z) for the first 30 polynomials, again
in the isotropic quadratic density case. In this instance there is no sign change and the
isotherms are in agreement with those presented in Moore & Weiss [63].
In table 4.3 the range in which Tf™*, the temperature at which penetrative convection
starts, is for £ varying is presented. It has already been concluded that the system becomes
more stable as both f and 7\ increase. The behaviour seen in table 4.3 agrees with this,
since it clearly shows that for higher values of £ a higher value of T is needed for the
x

onset of penetrative convection .

4.5. Numerical Results and Conclusions 92

.0

0.9

0.8

0.7

0.6

W 0.5

4.0
0.4

0.3

0.2
6.0

0.

7.0
0.0
7.9
0. T
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Figure 4.2: Profile of W{z) normalised over the spatial layer, £ = 1, T varies as shown on u

graph.

i •
.0

0.9

0.8

0.7

0.6
Theta
0.5

0.4

0.3 6.

0.2

0.

0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 .0

Figure 4.3: Profile of @(z) normalised over the spatial layer, £ = 1, T varies as shown on graph.
u
4.5. Numerical Results and Conclusions 93

£ Interval for T f "

0.25 (5.90,5.95)
0.50 (6.20,6.25)
0.75 (6.35,6.40)
1.00 (6.45,6.50)
1.25 (6.55,6.60)
1.50 (6.65,6.70)
1.75 (6.70,6.75)
2.00 (6.80,6.85)
3.00 (7.00,7.05)
5.00 (7.30,7.35)

Table 4.3: The interval in which T\ yields penetrative convection.

Chapter 5

A Model for Convection in the

Evolution of Under-ice Melt Ponds

In this chapter a model for convection in the evolution of under-ice melt ponds is pre-
sented. The evolution of under-ice melt ponds was first explained by Martin & Kauff-
man [59]. They modelled the phenomenon observed in the summer Arctic, where a layer
of fresh water at a temperature of approximately 0°C, lies under a layer of ice and over
a much colder ( ~ -1.6°C), denser layer of salt water. In this scenario the fresh water
layer is convectively unstable since the fresh water diffuses heat to the salt water causing
a density inversion. The supercooled water rises to the overlying ice surface at which
point it nucleates into thin vertical interlocking ice crystals. These grow down to the salt
water interface and then a lateral crystal growth begins and continues until a horizontal
ice sheet forms, trapping fresh water above and forming a so called under-ice melt pond.

The formation of under-ice melt ponds are an important factor in sea ice melting.
In recent literature much attention has been paid to fast sea ice melting and the factors
affecting it. For example Bogorodskii & Nagurnyi [9] showed that under-ice melt ponds
favour melting in the lower boundary of sea ice. This has important applications in global
and regional climate modelling. Schmittner et al. [87] argued that active ice sheets offer
an explanation for the forcing mechanism behind rapid warming events. Rahmstorf [81]
also remarked that the nonlinear dynamics of ice sheets can cause abrupt climatic change.
A comprehensive summary of theoretical glaciology can be found in Hutter [40] and a
detailed study of the thermomechanical behaviour of large ice sheets and their relation to

94
Chapter 5. A Model for Convection in the Evolution of Under-ice Melt Ponds 95

climate variations in Baral et al. [4]. Such work and the references therein highlight the
need for accurate models. Indeed Clark et al. [19] commented that more realistic models
of the hydrological cycle (precipitation, run-off, iceberg discharge, sea ice) are urgently
needed to understand the role of thermohaline circulation in abrupt climate change.
It is the aim of this chapter to accurately model the first stage in the formation of under-
ice melt ponds, i.e. the onset of convection in the unstable fresh water layer. Bogorodskii
& Nagurnyi [9] presented a model for the same situation. In their work however they also
took into account the layer of melt water overlying the ice layer and treat the system as a
unified one. They argued from physical considerations that all layers should lose stability
simultaneously and so restricted their attention to solving the problem for the upper layer
only and then used a symmetry argument to consider the system as a whole. They found
that the major parameters affecting stability were the thickness of the ice and coefficients
of heat conductivity. In this chapter we employ a different model and find that - 7 ] , the
temperature of the salt water, and r , the ratio of the heat diffusivity to solute diffusivity,
are also major parameters affecting stability.

One of the important aspects in the evolution of under-ice melt ponds is the fact that
the layers of fresh water and salt water remain separated. This is due to double diffusion.
Density increases with supercooling for solutions with salinity > 2 3 ° / 00 but decreases
for solutions with salinity < 2 3 ° / , cf. Martin & Kauffman [59]. The salinty of surface
00

sea water ranges from 26 — 2 8 ° / 00 in the Arctic to 3 4 ° / QO in the Antarctic. Hence for the
model in hand, the temperature and salt fields exhibit competing behaviour. It is widely
known that the competing effects in double diffusion mean that the linear instability bound
does not always capture the physics of instability and subcritical instabilities may arise,
cf. Hansen & Yuen [39], Proctor [79], and Shir & Joseph [90]. It is very important
therefore, to obtain global nonlinear stability thresholds which guarantee bounds below
which convection cannot occur. In this chapter we find both the linear instability and
global nonlinear stability thresholds and present the subcritical region bounded by these
two limits.

In this chapter we choose a density profile with quadratic temperature dependence

and linear concentration dependence. As a result the standard energy technique does not
yield a global stability bound. Therefore a weighted technique as introduced in Payne &
5.1. Governing Equations 96

Straughan [76] is used. The two competing effects demand two coupling parameters cf.
Straughan [94]. We find that the unweighted coupling parameter has a critical value such
that the nonlinear stability bound is independent of the salt field.
An outline of this chapter is as follows. We begin by formulating a governing model
and describe a density field which allows for thermal convection and double diffusion. We
then find an instability bound for the linearised system and a global stability bound for the
nonlinear system. Finally we present the numerical results and discuss their implications
in modelling under-ice melt ponds.

5.1 Governing Equations

Ice c o°c
u
z =d

Fresh Water

z = 0 •
Salt Water a —Ti

Figure 5.1: Schematic diagram of the governing system.

Consider a layer of water lying in the infinite three-dimensional region 0 < z < d,
sandwiched between a layer of salt water at z = 0 and a layer of ice at z = d. The upper
and lower planes are held fixed at temperatures 0°C and — TJ(< 0°C), and concentrations
C and C\ respectively, see figure 5.1. Suppose that the layer of water has a salt concen-
u

tration less than that in the layer of salt water lying below. This system is in a potentially
gravitationally unstable state since the warmer water at the upper boundary is denser than
that at the lower with respect to temperature. However the water at z = 0 is saltier than
that at z = d and so denser with respect to the salt field. Hence in this situation we have
two competing effects, a temperature gradient which is destabilising and a salt one which
is stabilising. The z direction is denoted by the vector k with i , j , k being the standard
Cartesian basis. Gravity acts in the negative z direction and we assume that the density p
is constant, everywhere except the body force. Then, the Navier-Stokes equation for the
5.1. Governing Equations 97

fluid motion are

Vi,t + VjVij = —-p,i + uAvi - —gp(T, C)k h (5.1.1)

Po Po
where v t,p, u, g, T, and C are velocity, time, pressure, kinematic viscosity, gravity, tem-
it

perature and concentration respectively and p the density, is assumed to be of the form
2
p(T, C) = P o [ l - a(T - 4 ) + A(C - C)], (5.1.2)

where a and A are the thermal and solute expansion coefficients respectively, and C
and po are some constant reference values. Standard indicial notation and the Einstein
T
summation convention are employed throughout, and k = ( 0 , 0 , 1 ) . The constant term
in (5.1.2) is absorbed into the pressure by defining a modified pressure p to be p+gp (l 0 —
16a — AC)z. Then, dropping the hat, (5.1.1) becomes

2
Vi,t + vjVij = - —p ti + vAvi + gki [a(T - 8T) -AC]. (5.1.3)
Po
Equation (5.1.3) together with the incompressibility condition and the equations of energy
and solute balance yield the following system of governing equations
2
Vi, + v ^ j = -—p,i + vAvi + gki [a(T
t - ST) - AC],
Po
Vi i = 0,
(5.1.4)
T t + ViTi = KAT,

C, + ViCj =
t K AC,
C

where K and K are thermal diffusivity and solute diffusivity, respectively.

C

The boundary conditions are

T = 0°C, C =C U at z = d,
(5.1.5)
T = - T U C =Q at z = 0,

where 7), C , C\ are prescribed constant fields and C; > C and 7} > 0°C. Under these
u u

conditions, the basic steady state solution of (5.1.4) whose stability is to be investigated,
is
Vi = 0, T = Z
- ± - T U C=-jz + C, t

where
Ci - C u

7 = > 0
^ — -
5.1. Governing Equations 98

The hydrostatic pressure p may be found from (5.1.4)i but since p is eliminated from our
analysis later on, a derivation of it is not included here.

To study the stability of (5.1.4) perturbations to the steady state solutions are intro-

duced as follows

Vi = Vi + Ui, T = T + 9, P =p+ TT, C = C + <f>.

The perturbed expressions are substituted into (5.1.4) and non-dimensionalised with scal-
ings of
D 2 V P V 2
Y A T TT P °

v A v 6
' nagd ' K Aga
c

for displacement, time, velocity, pressure, temperature and concentration respectively.

Then (5.1.4) may be rewritten in terms of the non-dimensional perturbation quantities

Uij + UjUij = -ivj + Aui - 2RM(z)9ki + Pr9 k 2

{ - R <j)ki, c

Ui,i = 0,
(5.1.6)
Pr(9, t + 9) Ui ti = A9- Rw,

Sc(^ >t + Ui4>,i) = A(j> + R w,c

where

2
M(z)=£-z, £ = ^ ± ^ , w = u, 3 xel x(0,l),

and
2 7 U
R = ± ^ , Rt = ^ L , Pr= -, Sc=^
f\jpU Ki Kip

define the Rayleigh, solute Rayleigh, Prandtl and Schmidt numbers. The boundary con-
ditions for the perturbed variables are

w(0) = w(l) = 0, 9(0) = 0(1) = 0 , (f)(0) = 0(1) = 0,

and it can be deduced from (5.1.6) that 2

w (0)
>z = w, (l)z = 0.
5.2. Linearised Instability Theory 99

5.2 Linearised Instability Theory

at
Neglecting the nonlinear terms in (5.1.6) and assuming a temporal growth rate like e we

obtain the linearised system

oui = —ix^ + Aui — 2RM(z)9ki — R (j)ki,

c

v>i,i = 0,
(5.2.1)

oPr9 = A9- Rw,

aSc(f) = A(f) + R w.
c

To eliminate the pressure in (5.2.l)x, we take the curlcurl of both sides. Then, setting

i = 3, we obtain
2
a Aw = A w - 2RM(z)A*9 - R A*(/),
C (5.2.2)

2 2 2 2
where A* is the horizontal Laplacian, A* = d /dx + d /dy . Assuming a normal
mode representation for w, 9 and (/>, of the form w — W(z)g(x, y), 9 = Q(z)g(x, y) and
2
$ = $(z)g(x,y) where g(x,y) is some horizontal plan form satisfying A*(? = -a g,
(5.2.1)3,4 and (5.2.2) yield

2 2 2 2 2 2 2
a(D - a )W = (D - a ) W + 2RM(z)a S + R a $,
c

2 2
aPrO = (D - a ) 0 - RW, (5.2.3)
2 2
aSc$ = {D - a ) $ + R W, C

2 2 2
where D = d /dz and a is the wave number. The corresponding boundary conditions
are
W = DW = 9 = $ = 0, for z € { 0 , 1 } . (5.2.4)

2
This eighth order system for a was solved by the Chebyshev tau D and D algorithms
as discussed in chapter 2 and appendix A . l . Without going into detail, the basics of the
2
D method are now briefly described and the interested reader referred to appendix A . l
and Dongarra et al. [23] for further details and an outline of the D method.
2
In the D method, (5.2.3) is rewritten in terms of 2nd order derivatives only. Letting
5.2. Linearised Instability Theory 100

2 2
A = (D — a )W, (5.2.3) can be expressed as the four 2nd order equations

2 2
(D - a )W - 4 = 0,
2 2 2 2
(D - a ) A + 2RM(z)a e + Ra$ c = a A,
(5.2.5)
2 2
(D - a )G -RW = aPrG,
2 2
(D - a ) $ + R W = aSc$. C

This system is then transformed onto the Chebyshev domain (—1,1) and the solutions
W, A, 6 , and $ treated as independent variables and expanded in a series of Chebyshev
polynomials
N N
W = y^w T (z), n n A = y^a T (z),
n n

71=0 n=0
N N

71=0 71=0

The inner product of each equation is taken with some Tjt and the orthogonality of the
Chebyshev polynomials exploited to obtain the following generalised eigenvalue problem

2 2
( AD -a 1 -I 0 0 \
2 2 2
0 4D -a I * RaI
c A
2 2
-RI 0 AD -a 1 0 0
2 2
RJ 0 0 AD - aI
\ J

/0 0 0 0
0 / 0 0 A
= a
0 0 Pr 0 0

\ 0 0 0 Sc
2 2 2 2 2
H e r e D is the Chebyshev matrix representation ofd /dz , * = 2i?a (£—1/2)1 — Ra Z,
T T
Z is the Chebyshev matrix representation of z, W = (w , ...,w^) , a A = (o ,a ) ,
0 N

T T
0 = (d , ...,0 )
0 N and * = (<j> ,cf) ) . 0 N The last two rows of each (N + 1) x (jV + 1)
block are removed and replaced by the discrete form of the eight boundary conditions
n
(5.2.4). The resulting system is solved for the eigenvalues a^ \n = 1,2,..., via the
QZ algorithm. A presentation of the numerical results is delayed until section 5.4. For
the moment however, we remark that excellent agreement was found between the two
aforementioned numerical methods.
5.3. Global Nonlinear Stability via the Weighted Energy Method 101

5.3 Global Nonlinear Stability via the Weighted Energy

Method
Consider the nonlinear system (5.1.6)

2
u i)t + UjUij = —7r j + Aui - 2RM(z)6ki + Pr6 ki - R <f>ki,
c

Ui,i = 0,
(5.3.1)

Pr(9 !t + Ui0 ) = A0 -
ti Rw,

Sc((j) t + Wi^.i)
t = + R w,c

with boundary conditions

w(0) = w(l) = 0, 0(0) = 0(1) = 0,

0(0) = 0(1) = 0, w (0)

tt = w, (l)
z = 0.

2
Let V be some periodic cell and denote the L -inner product as follows

(uv) = / uvdV.
Jv

If the usual energy method is applied to (5.3.1), i.e. multiply equation (5.3.l)i by u it

(5.3.1) by 0, and (5.3.1)4 by <f> and then integrate all equations over V, a cubic nonlinear-
3

2
ity (w9 ) is obtained from (5.3.l)i. To remove the cubic term and hence achieve a global
nonlinear stability result, we use a weighted energy technique as described in Payne &
Straughan [76].
First, multiply (5.3.l)i by Ui and integrate over V to obtain

d_ (\ 2 2 2
Q | | u | | ^ = -\\Vu\\ -2R(M(z)6w) -R (<f>w) + Pr(w6 ).
c (5.3.2)
dt

A weight function is then introduced

jx = n — 2z,

where // > 2 is some constant coupling parameter chosen so that p, > 0. Using (5.3.1) it 3

can be deduced that

d (Pr
d V = P r w & 2 R 9w 5 3 3

dl ( t / f ^ ) ~ ( ) ~ f WAiW ~ (V ) • (--)
5.3. Global Nonlinear Stability via the Weighted Energy Method 102

Finally, multiplying (5.3.1) by <j> and integrating over V yields

4

An energy, E, is defined as

where A > 0 is some constant coupling parameter. Then

where

I = R c [ ^ - l ) (<W - R {(2M(z) + p) 6w),

2 s
D = | | V u | | + J OOfadV + ±\\V0||

Next, define RE by
1 /
—- = max —,
RE n D
where % is the space of admissible solutions, namely

2
n = {ui,d,(j>eC (0, l)|u = 0 = 0 = 0
3 when 2 = 0,1}.

This definition leads to the following relation

Using the fact that z G (0,1) we observe that

a
jx = jji — 2z =>- jl> fi — 2 and — < 1.
A*
Therefore using Poincare's inequality (appendix B . l ) and the information above, D can
be bounded from below as follows
2 z
D > c i | | u | | + c ——— / fi9' dV + c -
2 3
S c
A* Jv
where c\, c , c are positive constants. It is easy to show then
2 3

D > cE, (5.3.6)

5.3. Global Nonlinear Stability via the Weighted Energy Method 103

where c is some positive constant dependent on n, A and Sc. Using (5.3.6), (5.3.5) can be
rewritten as
dE act
V < -acE => E(t) < E(0)e-
at
where a = 1 — I/RE- So for all R E > 1, E(t) —> 0 as t —> oo, where convergence is at
least exponential. Therefore, R E = 1 provides a global nonlinear stability threshold.
Setting R E = 1 and using a calculus of variations technique, the Euler-Lagrange
equations for the maximum are found to be

f{X)R ^ki c - R (2M(z) + p) Oh + 2 A W i = n, >t

BR
- R (2M(z) + p,)w + 2jlA0 - 4 — = 0, (5.3.7)
oz
f(X)R w c + j - ^ = 0,
c

where

/(A) = * = v^^- (5.3.8)

The Lagrange multiplier IT in (5.3.7)i is eliminated by taking the curlcurl of both sides.
Then, setting i = 3 and introducing normal modes w = W(z)g(x, y), 9 = Q(z)g(x, y),
and ^ = $(z)g(x, y) gives

2 2 2 2 2
2{D - a ) W + Ra M(z)e - R a f(\)<f>
c = 0,
2 2
- RM(z)W + 2fi(D - a )Q - 4 D 6 = 0, (5.3.9)
2 2
f(\)R W c + ^-(D -a )<t> = 0,
be
where M(z) = 2M(z) + fx and we have boundary conditions W = DW = 6 = $ =
0, f o r ^ G { 0 , 1 } .
This eighth order system was solved numerically by both the compound matrix method
2
and the Chebyshev tau D method. The basics of the two routines are now briefly de-
scribed for the system in hand.
For the compound matrix method, the vector W is defined as

T
W = (W, W, W", W", 0 , 0 ' , $, $ ' )

Let W i , W , W , and W be independent solutions of (5.3.9) obtained by replacing the

2 3 4

boundary condition at z = 1 by the initial conditions

W!(0) - (0,0,1,0,0,0,0, O f , W (0) = (0,0,0,1,0,0,0, o f ,

2
5.3. Global Nonlinear Stability via the Weighted Energy Method 104

T T
W (0) - (0,0,0,0,0,1,0,0) ,
3 W (0) = (0,0,0,0,0,0,0,1)
4

T
A new seventy vector Y = ( y . . . , yjo) 1 ( is formed from the 4 x 4 minors of the 8 x 4
solution matrix with columns W i , W , W , and W . Using (5.3.9), seventy first order
2 3 4

differential equations in terms of yi,..., y 70 are found and numerically integrated from 0
to 1. The boundary conditions on W are

W(0) = 0(0) = W'(0) = $(0) = 0,

W(l) = 0(1) = W'(l) = $(1) = 0.

We replace the conditions at z = 1 by

< ( o ) = < ( o ) = e' (o) = < ( o ) = i . s

Hence the appropriate initial and final conditions are (given the labelling used)

yeo(0) = l , yn(l)=0. (5.3.10)

R is fixed and the eigenvalue R varied until y n ( l ) = 0 is satisfied to some preassigned

c

2
tolerance. Keeping a , A > 0, and fi > 2, a golden section search routine was employed
to numerically find
2 2
RCIE = max mini? (a , A, //). (5.3.11)
2
A,fi a

When maximising over A and /x, it is important to have initial guesses which are
reasonably good. By using parametric differentiation on the Euler-Lagrange equations
(5.3.7), fixing A and varying ft we may show that

i 2 ( 2 f - / i ) / w9dV = 0,
Jv
and hence deduce an estimate for fi = 2<^. Similarly i f we fix ft and vary A, parametric
differentiation suggests a best value for A = — Sc. This is unsatisfactory since we are
assuming A > 0. It is of interest to note at this point however, that in all numerical
calculations we found \ c r i t = Sc (see section 5.4).
Suppose A = Sc, then system (5.3.9) looses dependence on $ and reduces to the sixth
order system

2 2 2 2
2(£> - a ) W + Ra M{z)e = 0,
(5.3.12)
2 2
- RM{z)W + 2fi{D - a )B - ADS = 0,
5.3. Global Nonlinear Stability via the Weighted Energy Method 105

with boundary conditions W = DW = 0 = 0, for z € { 0 , 1 } . This system was also

solved via the compound matrix method and the results used to check the results of the
eighth order system ( 5 . 3 . 9 ) when A = Sc (or R = 0 ) . c

2 2 2
To implement the Chebyshev tau D method we begin by letting A = (D — a )W.
Then the eighth order system ( 5 . 3 . 9 ) can be rewritten as

2 2
(D - a )W -A = 0,
2 2 2 2
2{D - a )A - i ? a / ( A ) $ = c -Ra M{z)G,
(5.3.13)
2 2
2fi{D - a )0 - 4 D 0 = RM(z)W,
2 2
f(\)R W+j- (D -a )$
c c = 0,
with boundary conditions W = DW = 0 = $ = 0, for z e { 0 , 1 } . To transform
(5.3.13) onto the Chebyshev domain ( — 1 , 1 ) , let z = 2z — 1. Then dropping the hats
(5.3.13) yields

2
(4D 2
- a )W - A = 0,
2 2 2
2(4£> - 2
a )A - i ? a / ( A ) $ = c -Ra M(z)G,
(5.3.14)

2 2
2 / i ( 4 £ > - a ) 0 - SDQ = RM(z)W,

2 2
f(\)R Wc + j- (W c - a )$ = 0,

where M = 2 £ + /x — 2z — 2, /2 = // — z — 1 and now z E ( — 1 , 1 ) . Since p, is a function

2
of z we seek the Chebyshev matrix representation of zD . To construct this matrix we
follow an analysis similar to that for the representation of z as given in section 2 . 5 . 1 .
2
To find ( ^ D 0 , Tk), we begin by noting the following,
N+2 N+2
2
zD G = T ^ O X = Y,*»TiZ
71=0 71 = 0
= 47\02 + 1 2 ( T + T ) 0 0 2 3 + (567\ + 2 4 T ) 0 + (60T + 100T + 4OT )0 + • • • •
3 4 0 2 4 5

So, taking inner products with T , T , T , . . . ,TN , 0 X 2 +2 and using the orthogonality of
Chebyshev polynomials it can be deduced that,

2
( z £ > 0 , T ) = 12TT0 + 6OTT05 + . . . ,
2
0 3 (zD Q, T ) = 2TT0 + 28TT0 + . . . ,
x 2 4

2
{zD G, T ) = 2 67T0 + 507T^ + • • • ,
3 5
5.4. Numerical Discussion, Results and Conclusions 106

which can be expressed in matrix form as,

A) 0 0 12 0 60 0 . . . 0^ e 0 \ ( 2
±(zD 9,T ) 0 ^

0 0 4 0 56 0 0 . . . 0 \(zDH,T ) x

0 0 0 12 0 100 0 ... 0 02
2
l{zD 9,T ) 2
(5.3.15)

2
0 \9 ) \^(zD 9,T )J
\ I N + 2 N+2

2
Let the coefficient matrix in (5.3.15) be denoted by ZD then (5.3.14) can be rewritten as
the eigenvalue problem,

2
4D - a I 2
-I 0 0 \
0 2(4£> -a /) 2 2
0 -fR a I c
2
A
0 0 * 0 0
2 2
fRJ 0 0 £(4D -a /) J
V * /
^ 0 0 0 o ^ f w
0 0 # 0 A
R
$ 0 0 0 0
0 0 0 0
\ V * /
2 2 2
where * = 2(/x - 1)(4D 2
a I) -8D - 8ZD + 2a Z, D is the Chebyshev matrix
2 2
representation of d/dz, # :
- a ( 2 £ + n - 2)1 + 2a Z, and $ = (2f + fjt - 2)1 - 2Z.
This system was solved numerically (via the QZ algorithm) and the results used to check
those obtained by the compound matrix method. Excellent agreement was found between
the three systems outlined above.

5.4 Numerical Discussion, Results and Conclusions

To be in keeping with Joseph [41] the parameter r = Sc/Pr = K/K is introduced.
C

3 2 _1
Typical values for the thermal and solute diffusivities are K = 1.33 x 10~ cm s and
_5 2 _1
K = 0.7 x 1 0 c m s
C respectively cf. Unterberg [101], so we restrict our attention to
2 3
r e [10 ,10 ]. Throughout all calculations we fix Pr = 13.4 which is consistent for
water at low temperatures.
5.4. Numerical Discussion, Results and Conclusions 107

340

330 1.9

320

1.8

310

1.7
300

290-

1 1 1 1 1 1 1 1 1
0 10 20 30 40 50 60 70 80 90 100

2 2 2
Figure 5.2: A plot of the linear instability curves for R against R . r = 100, a = 9.705432,
Pr = 13.4, -T\ varies as shown on graph.

2
Figure 5.2 displays the linear instability curves for r and a fixed, and — 7} varying.
As — 7} decreases the linear instability curves shift up (a change of 0.1°C corresponds
2 2 2
to a change of approximately 15 in R ). Note that R oc T . To see how —7] effects
2 2 2
instability we define a new Rayleigh number R = R /T . Then inspection of figure 5.2
reveals that as the temperature of the salt water becomes cooler the system becomes more
unstable. This is to be expected since the temperature gradient is destabilising.
2
Figure 5.3 displays the linear instability and nonlinear stability curves for — 7} and a
2
fixed and r varying. — 7} is fixed at — 1.9°C and a at the critical value returned from
2 2
energy theory. To produce the linear instability curves, R was fixed and R found such
that the real part of the temporal growth rate oy = 0 (a = o + ioi, o ,Oi e R). We T T

2
then tracked through R to produce the curves. It is worth noting that for r < 10 the
2
tracking routine ran for a very wide range of starting values for R . However for r > 10
the graphing routine failed to locate the root, o r = 0, unless the two initial guess for
2 2
R were very close to the actual value. When R = 0, the critical Rayleigh number is
2
approximately the same for all r . So to find good starting values for R the graphing
routine was initially run for r = 10 and the results used to provide initial values when
5.4. Numerical Discussion, Results and Conclusions 108

331.5

331.0

330.5 -

330.0

329.5 -

329.0

328.5 -

328.0

327.5

Energy bound
327.0
10 20 30 40 50 60 70 80 90 100

2
Figure 5.3: A plot of the linear instability curves and the nonlinear stability curve for R against
2 2
R . -T = -1.9°C, a = 9.705432, Pr = 13.4, r varies as shown on graph.
C t

r > 10.
The behaviour of the linear instability curves is in good agreement with that seen in
Joseph [41]. The kink in the curves represents the point at which convection switches from
steady convection (<r = 0) to oscillatory (a = 0, <7j 7^ 0). Note that as r increases the
r

onset of convection is more likely to be via oscillatory convection as opposed to steady

convection. We also note that as r increases the critical Rayleigh number, for a fixed
solute Rayleigh number, decreases. To establish how r effects the instability of the system
2 2 2
we note that R oc 1/r for fixed R . Hence denning a new Rayleigh number, R =
2
TR and inspecting figure 5.3 shows that, as the thermal diffusivity increases, for fixed
solute diffusivity, the system becomes more stable. Again this is to be expected since the
temperature gradient is destabilising.

The nonlinear stability bound was produced by integrating the compound matrix equa-
tions and then finding the critical Rayleigh number

RaE = maxmini? (a , A,//).

2
A,/* a
5.4. Numerical Discussion, Results and Conclusions 109

We note here that in all calculations we found the critical parameters to be

^crit — Sc, fJ- it ~ 2£.

cr

If A = Sc then, as previously mentioned, the governing system reduces to a sixth order

one, independent of the salt field. Hence the global nonlinear stability bound is inde-

pendent of the salt field. This is in keeping with what Shir & Joseph [90] found for a

Boussinesq fluid layer heated and salted from below.

We observe from figure 5.3 that for R = 0 there is very good agreement between the
c

nonlinear stability and linear instability bounds. We also note that as r increases the linear

and nonlinear results become closer and remark that since the critical Rayleigh numbers

are so close (for r sufficiently large) we expect the linear analysis to have captured the

essential physics at the onset of convection. However as r decreases the agreement be-

tween the two thresholds is not so good and we highlight this area as a region of possible

subcritical instabilities.

Mulone [64], Lombardo et al. [48,49], and Basurto & Lombardo [5] all investigated

such regions of subcritical instabilities. In these cases however the density was assumed

to be linear in temperature and concentration field. Mulone [64] considered a Boussi-

nesq fluid layer heated and salted from below and showed that when r < 1 coincidence

between linear and nonlinear critical parameters can be found. Lombardo et al. [49] and

Basurto & Lombardo [5] investigated the situation when r > 1 and regions of coincidence

between linear and nonlinear results were established. Lombardo et al. [48] concentrated
on the Benard problem for a mixture with superimposed plane parallel shear flows. They

found unconditional stability results for arbitrary r and small Reynolds numbers, and
conditional stability results independent of the Reynolds number up to criticality for any

Schmidt and Prandtl number. A similar analysis of the subcritical region presented in this

chapter proves to be nontrivial. It is stressed that the nonlinear stability bounds derived
here and those presented in the next chapter are all unconditional.
Chapter 6

Unconditional Nonlinear Stability for

Temperature Dependent Density Flow
in a Porous Medium

The object of this chapter is to present a nonlinear stability analysis of thermal convection
in a saturated porous layer. The density of the saturating fluid in this instance however
is assumed to have a cubic temperature dependence as proposed by Merker et al. [61].
Merker et al. also formulated a density law with quintic temperature dependence. They
compared their two relations to the more commonly used quadratic one, as previously
employed, and advocated by Veronis [102]. On the basis of linear theory, Merker et
al. suggested that their cubic law was about 10% more accurate than the quadratic one,
whereas the quintic law yielded an improvement of approximately 3% over the cubic law.
McKay & Straughan [52] used the cubic and quintic laws of Merker et al. to describe
penetrative convection in a fluid layer. They concluded that the cubic density relation was
preferable to the quintic model since the quintic model leads to complicated mathematics
for relatively little gain in accuracy. In fact the cubic law proves to be a popular choice
because it is more accurate than the quadratic law but easier to implement than the quintic
one, cf. McKay & Straughan and the references therein.

A cubic density law is also employed by McKay & Straughan [51]. In that work
McKay & Straughan investigated the influence of a cubic density law on patterned ground
formation. In this chapter we consider a system essentially the same as that given in

110
Chapter 6. Unconditional Nonlinear Stability for Temperature Dependent Density
Flow in a Porous Medium 111
McKay & Straughan but with three notable exceptions. Two of the differences are due to
the fact that McKay & Straughan considered the specific case of patterned ground forma-
tion, an application of the analysis presented here. The third is that McKay & Straughan
did not include a Forchheimer drag term in their governing equations whereas here we
do. McKay & Straughan employed the well known Darcy [21] equations to describe
flow in the porous medium, cf. chapters 2-4. The equations of Darcy are believed to be
valid for flow in a region where the velocity is not too large. For situations where the
fluid velocity is not so small however, Darcy's equations are usually modified by adding a
Forchheimer term, see Forchheimer [31] and Nield & Bejan [73]. Various mathematical
and experimental justifications of Forchheimer theory have recently been given cf. Gilver
& Altobelli [36], Whitaker [105], and Giorgi [35]. In fact Giorgi points out that Forch-
heimer terms were actually used much earlier by Dupuit [25]. A thorough discussion of
Darcy's equations and the Forchheimer modification can be found in Nield & Bejan [73].

To establish unconditional nonlinear results we found it necessary to use Forchheimer

theory and introduce higher order norms. The attractive mathematical features of Forch-
heimer theory are highlighted in Payne & Straughan [77,78]. Payne & Straughan [78]
used Forchheimer theory to establish unconditional nonlinear results for flow in porous
media where viscosity was assumed to vary with temperature and density was assumed
constant. They also obtained unconditional results for which the viscosity was assumed
linear in temperature and the density quadratic. We use a similar analysis to that pre-
sented in Payne & Straughan [78] but address the issue of unconditional stability for flow
in porous media for which the viscosity is assumed constant and the density cubic in
temperature dependence.

McKay & Straughan [51] successfully developed a linear instability and a nonlinear
stability analysis for the onset of convection in the formation of patterned ground. How-
ever, the nonlinear stability results they achieve are conditional i.e. they are restricted to
initial data thresholds. In this chapter we tackle the important question of unconditional
stability and stress that we establish a fully unconditional result i.e. for all initial data. In
section 6.4, a comparison is made between the unconditional results derived here and the
linear results of McKay & Straughan. Very close agreement is found between the critical
parameters of both theories. The nonlinear theory presented here establishes an uncon-
6.1. Convection in a Porous Medium with a Cubic Density Law 112

ditional threshold, below which stability is guaranteed. The linear work of McKay &
Straughan provides an instability threshold above which instability is guaranteed. Since
the two thresholds are so close, we effectively guarantee that there are no subcritical in-
stabilities, and conclude that the linear analysis of McKay & Straughan gives a very good
physical representation of what happens in reality.
An outline of this chapter is as follows. In the next section the equations and boundary
conditions which govern the system are described. It is then shown that the standard en-
2
ergy technique with an L norm is insufficient to establish unconditional results. Higher
order norms are then introduced and an unconditional nonlinear stability bound is estab-
lished. The numerical solution of the nonlinear system is then discussed and finally a
presentation of the numerical results and a comparison with those published in McKay &
Straughan [51] is given.

6.1 Convection in a Porous Medium with a Cubic Density

Law

T > 0°C
u

z =d
Saturated isotropic porous medium

2 = 0
0°C

Figure 6.1: Schematic diagram of the governing system.

Consider an isotropic porous layer saturated with fluid contained between the horizon-
tal planes z = 0 and z = d. The z direction is denoted by the vector k with i, j , k being
the standard Cartesian basis. Gravity acts in the negative z direction and the upper and
lower planes are held fixed at temperatures T > 0°C and 0°C respectively, see figure 6.1.
u

The equation of state is assumed to have a cubic temperature dependence, as advocated

by Merker et al. [61], namely

2 3
p= P o [1 + AT - BT + CT ] , (6.1.1)
6.1. Convection in a Porous Medium with a Cubic Density Law 113

where p is the liquid density, p is the density at 0°C, T(°C) is the temperature and
0

A, B, C, have constant values obtained by curve fitting to data points (cf. Merker et al.).

Neglecting fluid inertia, absorbing the constant term in (6.1.1) into the pressure and

employing Darcy's law with Forchheimer drag terms (cf. Forchheimer [31] and Nield &

Bejan [73]), yields a momentum equation of the form

P,i = - f a - g P o ki (AT - BT 2 3
+ CT ) - b'\v\ Vl - c'\v\\, (6.1.2)

where p, fi,K,Vi,g,b', c' are pressure, dynamic viscosity, permeability, velocity, gravity

and Forchheimer terms respectively. Standard indicial notation and the Einstein summa-
T
tion convention are employed throughout, and as previously mentioned k = (0,0,1) .

Equation (6.1.2) together with the incompressibility condition and the equation of energy

balance yield the following system of governing equations

2
P,i = - % V i - gpoki (AT - BT 2 3
+ CT ) - b'\v\Vi - c'\v\ ,
Vi

viti = 0, (6.1.3)

T t + TiVi = KAT,

where t, K, and A are time, thermal diffusivity and the Laplacian operator respectively.
The upper and lower boundaries of the porous medium are assumed fixed, so the boundary
conditions u (0) = u (cQ = 0, T(0) = 0°C, T(d) = T are imposed.
3 3 U

The basic steady state solutions of (6.1.3) are,

Vi = 0, T = (3z,

where /? = T /d. u The hydrostatic pressure p may be found from (6.1.3)i but since p is

eliminated from our analysis later on, a derivation of it is not included here.
To study the stability of (6.1.3), perturbations (ui, 0, ir) to the steady state solution are

introduced as follows

Vi = Vi + Ui, T = T + 9, p = p + n.

Substituting the perturbed expressions into (6.1.3) and non-dimensionalising with scal-
ings of

(i PQK K W gdA
6.2. Unconditional Nonlinear Stability 114

for displacement, time, velocity, pressure and temperature respectively, yields

7Tj = —Ui — R9F(z)ki 2

4- 9 G(z)ki ——ki — b\\x\ui — c|u| Uj, 2

uhl = 0, (6-1.4)

Pr (6 tt + Ui0,i) = -Rw + A0,

where

2
F(z) = 1 - 2a z + 3a z ,
x 2 G(z) = Pr (oi - 3a z),
2

2
B C b'd dud
A A p0 p^K
2
The Prandtl number Pr, and associated Rayleigh number R are given by

p r = ^ J? = 9dAp KT 0 u

KpoK' KfJ,
2
The spatial domain of the porous layer is now {(x, y) 6 R } x {z 6 (0,1)} and w = u . 3

The perturbed boundary conditions are given by

w(0) = w(l) = 0, 0(0) = 9(1) = 0 . (6.1.5)

It is assumed that (u^, 9, n) have a periodic plan-form shape that tiles the (x, y) plane and

henceforth we let V denote the period cell for the perturbations.

6.2 Unconditional Nonlinear Stability

2
The standard approach to obtain a global nonlinear stability bound in L (V), (where
2
L (V) is the space of square integrable functions on V), proceeds as follows. First multi-

ply (6.1.4)x by Ui, integrate over V and make use of (6.1.5) and some integration by parts

to obtain

2 2 3
0 = - | | u | | - R (F(z)w9) + (G(z)w9 ) - (w9 )
r f (6-2.1)
3 4
-b / |u| dx - c / |u| dx,
Jv Jv
2
where || • || denotes the norm on L (V) and (•) denotes integration over V. Similarly,
multiplying (6.1.4) by 9, integrating over V and making use of (6.1.5) yields (after some
3

integrations by parts)

2
^ ( y W ) =-R(™0)-\\V9\\ - (6-2.2)
6.2. Unconditional Nonlinear Stability 115

An energy equation is then obtained by adding (6.2.1) to (6.2.2), hence

d (Pr R w d W 2 2 2
w V 9 »-» / = - ( ) ~ I I l l " \H -R(F{z)w9) + (G{z)w9 )
V J
2 (6.2.3)
_ ^ L . ( W 9 * ) - H [ | u | d x - c / lu| dx.
3 4

R Jv Jv
However, due to the higher order terms in the right hand side of (6.2.3), a standard energy
2
technique in L (V) fails to provide an unconditional nonlinear stability bound. To see
why this is, we follow an argument similar to that given in Payne & Straughan [78]. We
begin by defining the sign indefinite terms on the right of (6.2.3) by I and the sign definite
dissipation by D. Therefore, assuming b, c > 0,

2 3
/ = -R(w9)-R{F(z)w9) + (G(z)w9 )~^^-(w9 ), (6.2.4)

2 2 4
D = \\V9\\ + ||u|| + b [ l u f d x + c / |u| dx. (6.2.5)
Jv Jv
2
Suppose an energy is defined as E = |Pr||f9|| then from (6.2.3)

dt \ D

Then by the usual energy argument, if

max ^- < 1,
H D

where % is the space of admissible functions, we may conclude nonlinear stability (cf.
section 4.3).
In this case however, the term m a x I/D is not bounded above by 1. To see this we
w

begin by noting that since the space % is linear, Ui may be written as cm; and 9 as (59 for
arbitrary real numbers a, /3. Then
2 3
I_ _ Aap + Ba/3 + Ca/3
2 2 3 4
D~ D/3 + Ea + Fa + Ga

where A,..., G are the appropriate integrals from (6.2.4) and (6.2.5). Suppose Ui and
1+w
9 are fixed, then A,..., G can be treated as constants. Let j3 = e* where u is some
positive constant, then
2+W 3+2w 4 3 2+UJ
/_ _ Aa + Ba + Ca + " _ A/of + Ba + Ca
2 2 2 3 2 2 2 2
D ~ Da + » + Ea + Fa + Go* ~ D + E/a " + Fa/a " + Ga /a "'
6.2. Unconditional Nonlinear Stability 116

Suppose we choose co = 1/2 then

l 2 2 2 2
/ A/a l + Ba + Cot>l Ajo?' + B + Ca^
D D + E/a + F + Ga D/a + E/a + F/a + G —> oo 2 as a —> oo.

Hence the maximum of I/D in W cannot be achieved and we conclude that any straight-
2
forward approach to derive global nonlinear stability using L (V) theory seems to fail.
We note at this point however that a conditional nonlinear stability bound can be found
2
in L (V), cf. McKay & Straughan [51]. In the remainder of this section unconditional
stability results are established by introducing norms of higher order and following an
analysis analogous to that presented in Payne & Straughan [78].
LPiV), where p > 2, is defined as the space of functions on V such that

U>{V) = j v e V\ j p
| v | d x < oo J .

Let the associated norm || • || be given by

p

| v | P d x
H Z } ' '
Using (6.1.4) , the fact that the derivative of sign# is a Dirac delta function and employing
3

properties of the delta function, the following can be derived

2
= -R I wd {sign6)dx-^- [ |0|j|0|fdx. (6.2.6)
Jv 9 Jv
The second term in the right hand side of (6.2.6) can be bounded using Poincare's in-
equality (appendix B . l ) to yield

jt { ^ j M l ) < ~ R
^ 0 ( s i g n 0 ) d x - ^ J \9\ dx.
2
v
3
(6.2.7)

Young's inequality (appendix B.2) is used on the first term in the right hand side of (6.2.7)
to obtain

where ft > 0 is some constant which we are free to choose.

4
In a similar fashion an identity for L is introduced as follows. First observe that

d_
(^PW\) = -R(™0 ) 3
-l^effidx. (6.2.9)
It
6.2. Unconditional Nonlinear Stability 117

Then bounding the first term on the right by Young's inequality and the second via
Poincare's inequality yields

j (T» »0 A (
t
9 r
I - * ) 2
I **< 47 +
X | u | 1 < i x
' ( 6
' - 2 i o )

where 7 > 0 is some constant which we are free to choose. We now define an "energy",
E(t) by
2 T
E(t) = ^ | | 0 | | + ^\\9\\l + -^\\9\\i (6.2.11)
where //, A and r are positive constant coupling parameters at our disposal. Then using
(6.2.1), (6.2.2), (6.2.8) and (6.2.10), it is found that

, 3 d x + | u | 3 d x
f * «•*> - ^ + * ( I - x ) X ' * ^ X

- | | u f - i2 (F(;z)w0) + (G{z)w6 ) 2
- (w0 ) - b f |u| dx - c /" |u| dx.
3 3 4

^ 7v Jv
(6.2.12)
3
To bound the (w9 ) term, Young's inequality is used as follows
2
° 2 J P R 2
/ FL3\ ^
3 Q
2 ^ 2
7 / 4 a Pr 2 f 4

2
To deal with (G(z)w9 ) we maximise (7(2) for z G (0,1) and then use Young's inequality
again to deduce that

2 2 3 3
(G(z)w9 ) < Pr(w9 )
ai < [ \9\ dx+ [ |u| dx. (6.2.14)
op J V 3 Jy
Let
2 2
I=-v(w9)-(F(z)w9), L> = / / | | V 0 | | + | | u | | , (6.2.15)

and define R E by

J - = max L. (6.2.16)

Substituting (6.2.13), (6.2.14) and (6.2.16) into (6.2.12) yields

dE _ ( R E - R
D +j 3
| u | | y {R\ + aiPr)-b^dx.
dt ~ V RE
2
_2_ . „ v 8TT A
3£
(6.2.17)
6.2. Unconditional Nonlinear Stability 118

To deal with the higher order terms, we begin by choosing A, r, 7, /3 such that

A = A' + he u A ( \> + R a i P r ) _ ^ = 0 , (6.2.18)

2
r = r' + k e , 2 2 ^ (i*y - T T ) + = 0, (6.2.19)

where £ 1 , 6 2 , are arbitrary (small) positive numbers and k ly k are free to choose. The
2

following are then minimised

£ (RX' + [Link]), - L (RT' + . (6.2.20)

To do this, ( 6 . 2 . 1 8 ) 2 is solved for A', (6.2.19)2 for r ' and the respective terms substi-
tuted into ( 6 . 2 . 2 0 ) . The resulting terms are then differentiated with respect to f3 and 7
respectively and /3,7 found such that ( 6 . 2 . 2 0 ) are minimised. This yields the selection
2 2
9i2 _3TT _ 2 Prai ,_ 3a Pr
2

2 2
The choice &i = 27/8n , k = 16/37T is then made and, hence (6.2.17) reduces to
2

dE ( R E - R S

^ D - bj |u| rfx - e 3
l j |6>| <ix 3

dt V RE , J V ov (6.2.21)
4 4
- c / \u\ dx - e [ 0 dx, 2

Jv Jv
where

2 2
R V3TT J V3TT ) '

To ensure b, c > 0, b and c are restricted as follows

9 R 9 R
„ f Y f Y a Pr f4R\
2
2 3
(4iT 4

Since ei,e 2 > 0 are at our disposal and we chose them arbitrarily small, the effective
restrictions on b and c are
2 2 3
9R\ a Pr2 / >• » \

Let d = (RE — R)/RE, and suppose b, c satisfy (6.2.22), then provided R < R , (6.2.21) E

reduces to
3 4
— < -aD -e x I |6>j dx - e / 0 d x . 2 (6.2.23)
dt Jv J v
6.2. Unconditional Nonlinear Stability 119

Using Poincare's inequality in (6.2.15) we note that

2 2 2
D > fiir \\d\\ + ||u|| > n<n \\ef. 2
(6.2.24)

Hence using (6.2.24) in (6.2.23), yields

^ < -a^\\0\\ 2
- e J |0| dx-
x
3
e 2 | ^dx,

< -kE,
kt
=>E{t) < E(0)e- , (6.2.25)

where
2
. f 2air 3ci 4e 1 2
k = min < — — , ———, —— > .
\ Pr ' A P r ' TPT J
Therefore, E ( i ) —> 0 as i —)> oo where convergence is at least exponential.
2
Observe that decay in ||u|| ,11xa11§, ||u|j^ can also be established. To see this first note
from (6.2.1) that

2 2
||u|| + b\\u\\l + c\\u\\t = -R (F(z)w9) + (G(z)w9 ) - ( w 9 *). (6.2.26)

To bound the first term on the right hand side of (6.2.26) we minimise —F(z) for z e
(0,1) and use Young's inequality to deduce that

(-RF(z)w6) < J R ^ - - l j 6 w d x ,

Then, using (6.2.27), (6.2.13), and (6.2.14) in (6.2.26) yields

2 2 2
||u|| + &||u||l + ||u||S < R ( J L - l ) ^ | ^ | | + ^
C « + ^ ^ « , (6.2.28)

2 3 2 2
where we have chosen /3 = 3b/2aiPr and j = a Pr /2Rc.
2 Decay in ||u|| , ||uj|f,
||u||4 is then obtained from (6.2.28) since E(t) is decaying at least exponentially. Hence
we have established that, provided R < R E and the non dimensional Forchheimer coeffi-
cients b, c satisfy (6.2.22), then global nonlinear stability is achieved.

It is worth noting at this point that an unconditional stability bound can also be
2 3 4
achieved in L and L i f weights are employed (L is not necessary). Unfortunately
the resulting Euler-Lagrange equations contain the function sign#. It is unclear how to
deal with this function in a numerical method and so the weighted method does not yield
any practical information in this case.
6.3. Numerical Solution of the Euler-Lagrange Equations 120

6.3 Numerical Solution of the Euler-Lagrange Equations

The Euler-Lagrange equations for the maximum in (6.2.16) are found to be

-P,i = RE {F{Z) + / j ) ki9 + 2u h

(6.3.1)
0 = -R E (F(Z) +H)W + 2/iAe,

where here, p is a Lagrange multiplier which we eliminate from (6.3. l ) i by taking curlcurl
of both sides. Then, setting i = 3 and assuming a normal mode representation for w and
6 of the form w = W(z)g(x, y) and 8 = Q(z)g(x, y), where g(x, y) is a horizontal plan
2 2 2 2 2
form satisfying (d /dx + d /dy )g = —a g, yields

2 2 2
2 (D - a ) W - R E (F(z) + //) a 0 = 0,
(6.3.2)
2 2
RE (F(Z) + fi)W-2n (D - a ) 0 = 0,
2 2 2
where D = d /dz and a is the wave number. The corresponding boundary conditions
written in terms of the normal modes are

W(0) = W(l) = 6(0) = 6(1) = 0. (6.3.3)

This fourth order system was solved numerically for R E by the compound matrix
2
method and the Chebyshev tau D algorithm. Without going into detail, the basics of
both routines are briefly described for the system in hand.
T
For the compound matrix method, the vector W is defined as W = (W, W, 6 , 6 ' ) .
Let W i and W be independent solutions of (6.3.2) obtained by replacing the bound-
2

T
ary condition at z = 1 by the initial conditions Wi(0) = (0,1,0,0) and W (0) = 2

r
(0, 0,0, l ) . The variables t / i , . . . , y are formed from the 2 x 2 minors of the 4 x 2
6

solution matrix with columns W l t W , to give

2

yi = w w - w w{,
x 2 2 2/2 = w \ 6 - w e 2 2 1:

y = Wi&
3 2 - W 0[, 2 2/4 = ^ ' 6 2 - ,

y = W[& - W^ei,
5 2 2/6 = 6 6 X 2 - 6,6;.
6.3. Numerical Solution of the Euler-Lagrange Equations 121

Then using (6.3.2) the compound matrix equations are

2
, „ {F(z) + ii)a
y 2
Vi = RE g ' 2/2 = 2/3 + 2/4,

2 2
2/ = « 2 / 2 + 2/5,
3 y'i = a 2/2 + 2/5,
2 6 3 4
0 + 2 , . „ (F(z) + //)a ( - - >
2/5 = 2 / A 1 / 1 + ( ? / 3 +
^ + G
2
Y 6
'

(F(z)+/z)
2/ = -RE
6 ^ 2/2-

This system was integrated numerically subject to the initial condition y (0) = 1 , and 5

the value of R E varied such that 2 / 2 ( 1 ) = 0 was satisfied to some predefined accuracy.
2
Keeping / i , a > 0, a golden section search routine was then used to numerically find

2
Ra = maxmini?!;(/i, a ). (6.3.5)
2
H a

2
A Chebyshev tau D routine was also performed to solve (6.3.2). First, equations

(6.3.2) were transformed to the Chebyshev domain ( — 1 , 1 ) by letting z• — 2z - 1. Then

dropping the hats, (6.3.2), yields

2 2 2
2 (AD - a ) W = R E (F(z) + n) a 0,
(6.3.6)
2 2
2pt (4D - a ) <d = R E (F(z) + n) W,

2
where F(z) = 3a z /4 + (3a /2 - ajz
2 2 + 3a /4 + l - a 2 l t z e ( - 1 , 1 ) and W = 9 = 0
for z = { — 1 , 1 } . Since F is a quadratic function of z we seek the Chebshev matrix
2
representation of z . To construct this matrix we follow an analysis similar to that for the
representation of z as given in section 2.5.1.

The solutions W and 0 are expanded in a series of Chebyshev polynomials

N N
w
W = J2 nTn(z), 9 = ^0 T„(z),
n

n=0 n=0

2
and the inner product of each equation is taken with some T*. To find (z Q, T ), we begin k

by noting the following,

N
2
z2
0 = T J2^T n

n=0
2
= T (9 T 0 0 + 9^ + ... + 9 T ) . N N
6.4. Numerical Results and Conclusions 122

So, taking inner products with T , T i , T , . . . , and using the orthogonality of Chebyshev
0 2

polynomials it can be deduced that,

2
(z e,r ) = ^ 0 0 + ^2,

(^6,70 = ^ + ^ 3 ,
2
( e,T ) = ^
2 2 0 + ^ 2 + ^4,...etc.

This can be expressed in matrix form as,

1
A 0 0 0 e
2 u
4 ( °)
0 \3
4
0 1
4 0 01
0 0 1
4 02 = (6.3.7)

2 2
Let the coefficient matrix in (6.3.7) be denoted by Z . Note that Z is exactly the same as

the matrix product Z x Z, where Z is given in section 2.5.1. Then (6.3.6) can be rewritten

as the eigenvalue problem,

2 2 2
8D - 2a / 0 \ / W \ ._ . / 0 a* \ ( W
2 2 E
0 8fiD - 1\ia l J \ 0 J y * 0 J \ Q

2 2 2
where here D is the Chebyshev matrix representation of d /dz , I is the identity, * =
2 T
3 a Z / 4 + (3a /2 - a )Z
2 2 x + (3a /4 - a + 1 + / / ) / , W = ( w , - ; W ) and 0
2 x 0 N =
T
( 0 , 0 ; v ) - The last two rows of each (N + 1) x (N + 1) block were removed and
O

replaced by the discrete form of the four boundary conditions (6.3.3). The resulting sys-

tem was then solved using the QZ algorithm. The numerical results obtained from the
2
Chebyshev tau D routine are not presented since excellent agreement was found between

them and those obtained via the compound matrix method. A presentation and discussion

of the results found from the compound matrix routine are now given.

6.4 Numerical Results and Conclusions

In this section we compare our numerical results with those of McKay & Straughan [51].
Therefore, we assume that the saturating fluid is water and take the following values for
6.4. Numerical Results and Conclusions 123

A, B, C (cf. Merker et ah [61]),

A = 6.85650 x 10~ (°C)'- l5

6 -2
B = 8.82063 x 10" (°C)"

C = 4.16668 x 10" (°C) - 3 8

To be in keeping with George et ah [33], McKay & Straughan adopted a Rayleigh number
that reflected the layer depth which is destabilising, namely

gapoKdT*
Ra
Kit

6 2
where a = 7.68 x 10" (°C~ ) and x = 4/T . Therefore, in order to compare the
u

nonlinear results calculated here with the linear results of McKay & Straughan [51], we
adopt a Rayleigh number of the form

2
Ra = 2
-rX R 2
E 0.448x i?|-

In McKay & Straughan [51] the governing equations and boundary conditions are es-
sentially the same as those used here but with three notable exceptions. Firstly the Forch-
heimer terms, necessary to establish unconditional stability, are not present in McKay &
Straughan. However, in this work the Forchheimer terms are eliminated from the final
system due to the way the energy analysis is implemented. Therefore provided the Forch-
heimer terms satisfy the constraint (6.2.22) we can compare both works. The second
difference to note is in the permeability term. McKay & Straughan allowed the perme-
ability to have dependence on z of the form

K = k0 (1 + 72), (6.4.1)

where A; was the permeability at the base of the layer and 7 was a positive constant. They
0

choose this relationship since it reflects the physics of what happens in patterned ground
formation. In the results that follow we only compare results with those published in
McKay & Straughan for which they set 7 = 0 and hence K is equivalent to the definition
used here. We remark at this point however, that i f we allow the permeability to take the
same form as (6.4.1) then we can still obtain an unconditional stability bound. To see
this note that i f we employ (6.4.1) then the first term in the non-dimensional perturbation
6.4. Numerical Results and Conclusions 124

equation (6.1.4)! is of the form —«i/(l 4- jzd), where A; has been absorbed into the
0

scales. In deriving an energy identity for u we multiply this term by Ui, integrate over V
and bound above as follows

- f M U l U
\,dx < -JHiLL for z E (0,1).
v ;
(6.4.2)
J (1 + <yzd) -
v (l + d)
7

Since the right hand side of (6.4.2) is sign definite our energy analysis can follow as
before. Finally McKay & Straughan allowed for radiation heating at the upper surface.
Their boundary condition at z = 1 in terms of the perturbed variables was

0 + a6 = 0,
z

where a > 0 was a constant "radiation parameter". Again this boundary condition was
used because it is physically relevant to the specific case of patterned ground formation.
We only compare results with those published in McKay & Straughan for which a — oo
and we recover the boundary condition imposed here.
In table 6.1 we compare the linear results of McKay & Straughan [51] with the non-
linear results of this work. The critical parameters of linear and energy theory are denoted
by subscript L and E respectively. RCLL and denote the critical Rayleigh number and
wave number obtained by McKay & Straughan via a linear instability analysis. Ra E and
a denote the critical Rayleigh number and wave number obtained here via the compound
E

matrix method. From table 6.1 we see that there is excellent agreement between the linear
and nonlinear results. Hence we conclude that the linear analysis of McKay & Straughan
has captured the essential physics of reality. We stress that the unconditional stability
threshold established here is of physical importance since it effectively guarantees there
are no subcritical instabilities.
6.4. Numerical Results and Conclusions 125

T (°C)
U Ra L RaE aE

0.5 1209.546 1209.163 3.142 3.142 0.936

1.0 324.427 324.267 3.143 3.143 0.873
1.5 155.419 155.316 3.145 3.145 0.812
2.0 94.727 94.640 3.148 3.150 0.751
2.5 66.083 65.992 3.153 3.156 0.693
3.0 50.363 50.248 3.162 3.167 0.636
3.5 40.922 40.760 3.176 3.183 0.582
4.0 34.957 34.713 3.197 3.208 0.531
4.5 31.120 30.745 3.230 3.245 0.485
5.0 28.705 28.126 3.283 3.300 0.444

Table 6.1: A comparison of the critical parameters of linear instability and nonlinear stability
theory. \i is the best coupling parameter.
Chapter 7

Conclusion

Throughout this thesis the subject of thermal convection in fluid and porous media has
been investigated. Linear instability and nonlinear stability analyses have been employed
to establish critical thresholds for the onset and type of convection involved. Two layer
and single layer systems were investigated and convection modelled primarily via a quadratic
equation of state and alternatively by internal heating or a cubic density law. Before a dis-
cussion of the work and suggestions for the future are made, a quick summary is provided.

In chapter 2 penetrative convection in a superposed porous-fluid layer was modelled

via a quadratic equation of state. A similar system was then investigated in chapter 3
but represented alternatively using internal heating. It was found that the two models
generated the same eigenvalues when mathematically adjoint but different eigenfunctions.
It was concluded that physically the two models were very different and care must be
taken when interpreting results.
In chapter 4 penetrative convection in a single anisotropic porous layer was mod-
elled by both a quadratic equation of state and internal heating. The porous medium was
assumed to have a horizontally isotropic configuration in which the permeability in the
horizontal plane was different to that in the vertical direction. An unconditional nonlinear
stability bound, as well as a linear instability bound was established and the two shown to
be very close.
In chapter 5 a model for convection in the evolution of under-ice meltponds was pre-
sented. Again linear and nonlinear analyses were given and it was shown that to establish
unconditional nonlinear results a weighted energy must be employed. The linear insta-

126
Chapter 7. Conclusion 127

bility and nonlinear stability bounds did not coincide and a region of possible subcritical
instabilities was presented.
Finally in chapter 6, an unconditional nonlinear stability bound was found for a layer
of isotropic porous medium. Convection in the porous medium was described by a cubic
equation of state and Forchheimer drag terms were included in the equations of motion.
Higher order norms were found necessary to establish an unconditional bound and so
employed to achieve the given result.

In all five chapters a linear instability analysis was performed. The analyses were
effective and efficient in establishing basic results for the convection problems at hand.
For example the numerical results in chapters 2 & 3 were novel and quite striking. To es-
tablish how representative these linear results are however requires a nonlinear approach.
This is highlighted by chapters 4, 5 & 6, where linear and nonlinear results were derived
and compared for a single layer of fluid or porous medium. A nonlinear analysis of the
two layer system however is highly nontrivial. It is hqped that in future work the energy
method may be applied to the two layer problem to form a nonlinear stability bound. The
inclusion of a Q term may simplify the nonlinear analysis, cf. chapter 4. Indeed this
was the motivation behind chapter 3. It is hoped that such work will be of use in future
analyses. The introduction of a weight as outlined in chapter 5, or higher order norms as
given in chapter 6, are further techniques that may also be of benefit here.

In chapter 4 the porous medium was assumed to have a horizontally isotropic configu-
ration in which the permeability in the horizontal plane was different to that in the vertical
direction. Such a configuration is common in the real world. In particular we mentioned
frost heave in the formation of patterned ground. Future work could include coupling the
work done in chapter 4 with that given in chapter 2. In other words penetrative convection
could be modelled in a superposed porous-fluid layer in which the porous medium has an
anisotropic configuration like that described in chapter 4. This would provided a more
refined model for the onset of penetrative convection in the formation of under water pat-
terned ground. In addition a radiative boundary condition at the upper fluid surface could
be included to further reflect the physics involved cf. George et al. [33].
A further example of the anisotropic configuration used in chapter 4 is sea ice. Eicken
et al. [27] suggested that sea ice has a different permeability in the vertical direction to
Chapter 7. Conclusion 128

that given in the horizontal plane. Hence the configuration presented in chapter 4 is an
appropriate representation for the permeability of sea ice in the melt season. This could
be coupled with the work done in chapters 2 & 5 to yield a model with very serious
applications as detailed below.
In chapter 5 a model for convection in the evolution of under-ice meltponds was pre-
sented. While the model employed very accurate numerical techniques and gave inter-
esting and useful results, a far more sophisticated approach could be taken. The analysis
of chapters 2, 4 & 5, could be used in conjunction with one another to build a layered
model which incorporates the overlying melt water, the sea ice, and the layer of fresh
water below.
There is evidence that sea ice in the Arctic region is thinning at a most dramatic
rate, see for example Rothrock et al. [85] and Wadhams & Davis [103,104]. Ice cover
in the Arctic, and anywhere else for that matter, plays a very important role in climate.
The effect of ice on Earth's surface albedo and its role in reducing the amount of heat
exchanged between the atmosphere and the ocean (or land) beneath the ice make it a vital
component of our climate. The amount of ice also affects the density, salinity and flow of
sea water. In recent literature a lot of attention has been paid to fast sea ice melting and
it's implication for polar climates. For example see Goldman [38], Kerr [42], Rignot &
Thomas [82], chapter 5, and the references therein.

One factor which affects sea ice melting but about which relatively little is known, is
the formation of under-ice melt ponds. There is very little field data on under-ice melt
ponds and as far as I am aware, the work of Bogorodskii & Nagurnyi [9], Eicken [26],
Notz et al. [74], and that given in chapter 5, are the only recent attempts at modelling
them. The formation of under-ice melt ponds however are instrumental in the melting of
summer Arctic sea ice. Future research on their formation and evolution is both vital and
worthwhile. A model that combines the analysis of chapters 2, 4, 5, and takes account
of the surface melt water above would be most beneficial in this area. An investigation
of how convection in the fresh water effects the underside of the ice and the pycnocline
(layer of water where the density changes rapidly with depth) below, would undoubtedly
enhance our current understanding of under-ice meltponds.
Appendix A

Numerical Methods

A.l The Chebyshev Tail Method

A full discussion of the Chebyshev tau algorithm applied to hydrodynamic stability prob-
lems can be found in Dongarra et al. [23] and Straughan & Walker [95]. For the purposes
of this appendix consider the boundary value problem as outlined in Dongarra et al. ,

u" + Xu = 0, xe(-l,l)
(AAA)
u(-l) = u ( l ) = 0,

2 2
where u e L (—1,1) and we work in the weighted L (—1,1) space with inner product
l
fg
(f>9)=[_ dx,
l V I - x

and associated norm || • ||.

Let u be written as a truncated series of Chebyshev polynomials,
N+2
u(x) = ^ u T ( x ) .k k (A.
k=0

Instead of solving (A. 1.1), the tau method argues that the following is solved exactly

u" + Au = nT N+1 + TT
2 N+2 (A. 1.3)

where u is given by (A. 1.2) and TI,T2 are the tau coefficients which may be used to
measure the error associated with the truncation of (A. 1.2).

129
A . l . The Chebyshev Tau Method 130

2
The Chebyshev polynomials are orthogonal in the L (—1,1) space, so multiplication

of (A. 1.3) by Ti yields

(Lu,Ti) = 0, 2 = 0,1,...,JV, (A. 1.4)

2
(Lu,T ) N+j = T \\T \\ ,
j N+J J = 1,2, (A.1.5)

where L is the differential operator defined by Lu = u"+Xu. In (A. 1.2), there are (N+3)
unknowns, namely, Ui, i = 0 , . . . , N + 2. Equation (A.1.4) yields (N + 1) conditions for
Ui and (A.1.5) can be used to calculate the r's. To find a further two conditions note that
n
T ( ± l ) = ( ± l ) and use the boundary conditions u(—1) = u ( l ) = 0 to deduce
n

N+2 N+2
B
^ ( - 1 ) « „ = 0, J > „ = 0. (A.1.6)
n=0 n=0

Hence (A. 1.4) and (A.1.6) yield a system of (N + 3) equations for the (N + 3) unknowns
u i = 0 , . . . , N + 2.
h

The derivative of a Chebyshev polynomial is a linear combination of lower order

Chebyshev polynomials and can be expressed as

{ 2nY] -]T ,
^ ~\
2n YTk=2 k +
n

k
k

T n T
o,
neven,

n odd.
(A.1.7)

Then (A. 1.4) can be written as

)
u f + \u i = Q, i = Q,...,N, (A.1.8)

2
where the coefficients u- ^ are given by
-. N+2
2
u^ = j- Yl P ( P - ^ (A. 1.9)
1
p=i+2,p+i even

with c = 2, and Q = 1, i = 1,2,

0 To see why this is, begin by differentiating (A. 1.2)

to yield
N+2 N+2 N+2 N+2

u' = J2 u r (x)
fc=0
k k

fc=0
=^E r=0
uD T k rk r =
r=0
£ x)
4 T, r (A.1.10)

where
JV+2 N+2
A . l . The Chebyshev Tau Method 131

Differentiating (A. 1.10) yields

N+2 N+2 N+2
N+2 N+2

u" = £
r=0
u^T' r
EE
= J 2 Y , ^ D
r = 0 p=0
v r T p = £ uf%,
p=0
(A.1.11)

where
N+2 N+2 N+2
(2)
kUk-

r=0 r=0 k=0

Therefore relabelling (A.l.10) and (A.1.11)
N+2

i=0

Then using (A. 1.4) we deduce (A. 1.8) and it only remains to show that
JV+2 N + 2 _ N+2
(2) 2 i 2 u
U = ^ ^ DDu
ip pk k = - J2 P(P ~ ) v (A.l.12)
1
p=0 fc=0 p = i + 2 , p + i even

2
To do this we use induction and (A. 1.7). The differentiation matrices D and D are

established in the process.

To begin note the following

N+2
DrkTr DokT
T x
k( ) = Yl = ° + D l k T l
+ " • + D +2,kT . N N+2 (A.l.13)
r=0

Using (A. 1.7) we can write

T' (x) = 2k{T

k Y + T + • • • + T*_i)
2 for /c even,
(A. 1.14)
T£(x) = A;T + 2fc(T + T + • • • + T _i)
0 2 3 fc for fc odd,

Therefore comparing the coefficients in (A.l.13) and (A.1.14) it can be deduce that

A>, ; = 0,
2 j> 1, (A.l.15)

A w - i = 2j - 1, j > 1, (A. 1.16)

A,i+2,--i = 2(i + 2; - 1), j > 1, i > 1, (A.1.17)

A,o = 0, t > 1, (A. 1.18)

or, in matrix form

A) 1 0 3 0 5 0 7
0 0 4 0 8 0 12 0
D= 0 0 0 6 0 10 0 14
A.2. The Compound Matrix Method 132

2
To find D we explicitly expand
N+2 N+2

U 2 ) = D D
i 5^ E ip pkUk-
p=0 k=0

Then using (A. 1.15) and comparing the coefficients of u f ' for i = 0 , . . . , N + 2, it can be
seen that
N+2
(2)
\ p\>
p=2,p even
N+2

p=3,p odd

1 JV+2

p=4,p even

Hence by induction (A. 1.12) is proved as required. Further expansion and comparison of
the coefficients of u yields
p

Dl, 2j = \{2j)\ J>1,

Dh+V = (* + + i > 1, j > 1,

which can be expressed in matrix form as

0 0 4 0 32 0 108 .
0 0 0 24 0 120 0
2
D = 0 0 0 0 48 0 192 .

1° /
2
Note that the matrix D = D.D in the sense of matrix multiplication.

A.2 The Compound Matrix Method

To illustrate how the compound matrix method is implemented consider the simple 4th
order eigenvalue problem as detailed in Straughan [94],

2 2 2
DW = a W - R a F(z, E A)6,
(A.2.1)
2 2
De = a Q- R F(z,X)W,
E
A.2. The Compound Matrix Method 133

2
where W, 0 , z are unknown variables, and a , A are constants we wish to minimise and
maximise over respectively. F is an arbitrary function of z and A, and R E is the gener-
alised eigenvalue of the system. Suppose we have boundary conditions

W = 6 = 0, on z = 0,1. (A.2.2)

To solve (A.2.1) subject to (A.2.2) the compound matrix method proceeds as follows.
r
Let W = (W, W, 0 , 0 ' ) , and suppose W 1 and W 2 are solutions to (A.2.1) with
initial values
r
W i ( 0 ) = (0,1,0, O f , W ( 0 ) = (0, 0, 0, l ) .
2 (A.2.3)

A new six vector

T
Y = (yi,y2,2/3,Z/4,2/5,2/6) ,

is defined by the 2 x 2 minors of the 4 x 2 solution matrix ( W i , W ) . Then, 2

2/i = W W^ - X W W[,
2

2/2 = - we
^ 0 2
2 1:

2/3 = W G' -w e[, 1 2 2

(A.2.4)
2/4 = w[e 2

2/5 = w{& 2

= 0 ! 0 ' -- 020'i2

Differentiating (A.2.4) and using the governing equations (A.2.1), the initial value prob-
lem for yi is found to be

2
= -R Fa y ,
E 2

y'2 = 2/3 + J/4,

2
= a2/2 + 2/5,
(A.2.5)
y\ = a 2/2 + 2/5,
2

2 2

y' = R Fy\
5
E + a (y 3 + y) - 4 R Fa y ,
E e

y'e = R Fy .
E 2

System (A.2.5) is integrated numerically from 0 to 1 subject to initial and final conditions.
The initial values (A.2.3) yield 2/5(0) = 1 and the boundary conditions at z — 1 give
A.2. The Compound Matrix Method 134

2/2(1) = 0. The eigenvalue, R ,

E is varied until 2/2(1) = 0 is satisfied to some preassigned
tolerance. In the computational work presented within this thesis the secant method was
used to locate the root and a golden section search routine to find the critical eigenvalue

2
Ra,E = maxmin R^(a ,2
A).
A a

More details and a thorough discussion of the compound matrix method, applied to
energy eigenvalue problems can be found in Straughan [94].
Appendix B

Inequalities

B.l Poincare's Inequality

Let V be a cell in three dimensions. Suppose for simplicity V is the cell 0 < x < 2a\,
0 < y < 2a , and 0 < z < 1. Let u be a function periodic in x, y of period 2a\, 2a ,
2 2

respectively, and suppose u = 0 on z = 0,1. Poincare's inequality may be written as

2 2
f u dV < ^ [ \Vu\ dV. (B.1.1)
71
Jv Jv
2
In general, the constant 1/TT in (B.1.1) depends on the geometry and size of the domain
V. Further discussion of this and Poincare's inequality can be found in Saloff-Coste [86].

B.2 Young's Inequality

Let a, b be arbitrary functions and p, q positive constants such that 1/p + l/q = 1, then
Young's [106] inequality may be written as
q
, aP b
ab<— + - . (B.2.1)
P Q
An elementary derivation of Young's inequality in this form can be found in Beckenbach
& Bellman [6], p.15.

135
Bibliography

[1] Allen M B . 1984. Collocation Techniques for Modeling Compositional Flows in

Oil Reservoirs. Springer-Verlag, New York.

[2] Allen M B . 1986. Mechanics of multiphase fluid flows in variably saturated porous
media. Int. J. Engng. Sci. 24, 339-351.

[3] Allen MB & Khosravani A. 1992. Solute transport via alternating-direction collo-
cation using the modified method of characteristics. Advances in Water Resources.
15,125-132.

[4] Baral DR, Hutter K & Greve R. 2001. Asymptotic theories of large-scale motion,
temperature and moisture distribution in land-based polythermal ice sheets. A crit-
ical review and new developments. Appl Mech Reviews. 54(3), 215-256.

[5] Basurto & Lombardo. 2003. Global nonlinear stability in the Benard problem for a
mixture near the bifurcation point. Continuum Mech. Thermodyn. 15(3), 265-274.

[6] Beckenbach EF & Bellman R. 1965. Inequalities. Springer-Verlag, Berlin.

[7] Benard H. 1900. Les tourbillons cellulaires dans une nappe liquide. Revue generale
des Sciences pures et appliquees. 11, 1261-1271,1309-1328.

[8] Benard H. 1901. Les tourbillons cellulaires dans une nappe liquide transportant de
la chaleur par convection en regime permanent. Annales de Chimie et de Physique.
23, 62-144.

[9] Bogorodskii PV & Nagurnyi AR 2000. Under ice meltwater puddles: A factor of
fast sea ice melting in the Arctic. Doklady Earth Sciences. 373, 885-887.

136
Bibliography 137

[10] Brown GO. 2002. Henry Darcy and the making of a law. Water Resources Re-
search. 38(07), 11.1-11.12.

[11] Carr M . 2003. A model for convection in the evolution of under-ice melt ponds.
Continuum Mech. Thermodyn. 15(1), 45-54.

[12] Carr M . 2003. Unconditional nonlinear stability for temperature dependent density
flow in a porous medium. Math. Models. Methods. Appl Sci. 13(2), 207-220.

[13] Carr M & de Putter S. 2003. Penetrative convection in a horizontally isotropic

porous layer. Continuum Mech. Thermodyn. 15(1), 33-43.

[14] Carr M & Straughan B. 2003. Penetrative convection in a fluid overlying a porous
layer. Advances in Water Resources. 26, 263-276.

[15] Chandrasekhar S. 1981. Hydrodynamic and Hydromagnetic Stability. Dover, New

York.

[16] Chasnov JR & Tse K L . 2001. Turbulent penetrative convection with an internal
heat source. Fluid Dyn. Res. 28, 397-421.

[17] Chen F & Chen CF. 1988. Onset of finger convection in a horizontal porous layer
underlying a fluid layer. J. Heat Transfer. 110,403^109.

[18] Chen Q.-S, Prasad V & Chatterjee A. 1999. Modeling of fluid flow and heat transfer
in a hydrothermal crystal growth system: Use of fluid-superposed porous layer
theory. J. Heat Transfer. 121(4), 1049-1058.

[19] Clark PU, Pisias NG, Stocker TF & Weaver AJ. 2002. The role of thermohaline
circulation in abrupt climate change. Nature. 415, 863-869.

[20] Curran MC & Allen MB. 1990. Parallel computing for solute transport models via
alternating direction collocation. Advances in Water Resources. 13, 70-75.

[21] Darcy H. 1956. Les fontaines publiques de la ville de Dijon. Dalmont, Paris.

[22] Dawkins P, Dunbar S & Douglass R. 1998. The origin and nature of spurious eigen-
values in the spectral tau method. J. Computational Phys. 147,441-462.
Bibliography 138

[23] Dongarra JJ, Straughan B & Walker DW. 1996. Chebyshev tau - QZ algorithm
methods for calculating spectra of hydrodynamic stability problems. Appl. Numer.
Math. 22, 399^135.

[24] Drazin PG & Reid W H . 1981. Hydrodynamic Stability. Cambridge University

Press.

[25] Dupuit J. 1863. Etudes Theoriques et Pratiques sur le Mouvement des Eaux.
Dunod, Paris.

[26] Eicken H. 1994. Structure of under-ice melt ponds in the central Arctic and their
effect on the sea-ice cover. Limnol. Oceanogr. 39(3), 682-694.

[27] Eicken H, Krouse HR, Kadko D & Perovich DK. 2002. Tracer studies of pathways
and rates of meltwater transport through Arctic summer sea ice. J. Geophys. Res.
107(C10), 8046, 22.1-22.20.

[28] El-Habel F, Mendoza C & Bagtzoglou AC. 2002. Solute transport in open channel
flows and porous streambeds. Advances in Water Resources. 25,455-469.

[29] Ewing RE. 1996. Multidisciplinary interactions in energy and environmental engi -
neering. J. Comp. Appl. Math. 74, 193-215.

[30] Ewing RE & Weekes S. 1998. Numerical methods for contaminant transport in
porous media. Computational Mathematics. 202 (Marcel Decker, Inc., New York),
75-95.

[31] Forchheimer P. 1901. Wasserbewegung durch Boden. Zeitschrift fur Vereines

Deutscher Ingnieure. 50,1781-1788.

[32] Fox L. 1962. Chebyshev methods for ordinary differential equations. Computer J.
4,318-331.

[33] George JH, Gunn RD & Straughan B. 1989. Patterned ground formation and pene-
trative convection in porous media. Geophys. Astrophys. Fluid Dyn. 46, 135-158.

[34] Ghosal S & Spiegel EA. 1991. On thermonuclear convection: I . Shellular instabil-
ity. Geophys. Astrophys. Fluid Dyn. 61, 161-178.
Bibliography 139

[35] Giorgi T. 1997. Derivation of the Forchheimer law via matched asymptotic expan-
sions. Transport in porous media. 29, 191-206.

[36] Givler RC & Altobelli SA. 1994. A determination of effective viscosity for the
Brinkman-Forchheimer flow model. J. Fluid Mechs. 258, 355-370.

[37] Gleason KJ. 1984. Nonlinear Boussinesq Convection in Porous Media: Application
to Patterned Ground Formation. M.S. Thesis, Univ. Colorado.

[38] Goldman E. 2002. Even in high Arctic nothing is permanent. Science. 297,1493.

[39] Hansen U & Yuen DA. 1989. Subcritical double-diffusive convection at infinite
Prandtl number. Geophys and Astrophys Fluid Dyns. 47, 199-224.

[40] Hutter K. 1983. Theoretical Glaciology: Material Science of Ice and the Mechanics
of Glaciers and Ice Sheets. Reidel, Dordrecht.

[41] Joseph DD. 1969. Uniqueness criteria for the conduction-diffusion solution of the
Boussinesq equations. Arch. Rational Mech. Anal. 35(3), 169-177.

[42] Kerr R. 2002. A warmer Arctic means change for all. Science. 297, 1490-1493.

[43] Kessler M A & Werner BT. 2003. Self organization of sorted patterned ground.
Science. 299, 380-383.

[44] Krantz WB, Gleason KJ & Caine N. 1988. Patterned ground. Scientific American.
256, 68-76.

[45] Krishnamurti R. 1997. Convection induced by selective absorption of radiation: a

laboratory model of conditional instability. Dynamics of Atmospheres and Oceans.
27, 367-382.

[46] Larson VE. 2000. Stability properties of and scaling laws for a dry radiative-
convective atmosphere. Q. J. Royal Meteorological Soc. 126, 145-171.

[47] Larson VE. 2001. The effects of thermal radiation on dry convective instability.
Dynamics of Atmospheres and Oceans. 34, 45-71.
Bibliography 140

[48] Lombardo S, Mulone G & Rionero S. 2000. Global stability in the Benard problem
for a mixture with superimposed plane parallel shear flows. Math. Methods Appl.
Sci. 23, 1447-1465.

[49] Lombardo S, Mulone G & Rionero S. 2001. Global nonlinear exponential stabil-
ity of the conduction-diffusion solution for Schmidt numbers greater than Prandtl
numbers. J. Math. An. and Appl. 262, 191-207.

[50] Lombardo S, Mulone G & Straughan B. 2001. Non-linear stability in the Benard
problem for a double-diffusive mixture in a porous medium. Math. Methods Appl.
Sci. 24, 1229-1246.

[51] McKay G & Straughan B. 1991. The influence of a cubic density law on patterned
ground formation. Math. Models. Meth. Appl Sci. 1(1), 27-39.

[52] McKay G & Straughan B. 1992. Non-linear energy stability and convection near
the density maximum. Acta Mechanics. 95, 9-28.

[53] McKay G & Straughan B. 1993. Patterned ground formation under water. Contin-
uum Mech. Thermodyn. 5, 145-162.

[54] McKenzie DP, Roberts JM & Weiss NO. 1974. Convection in the earth's mantle:
towards a numerical simulation. J. Fluid Mech. 62,465-538.

[55] Malin Space Science Systems. (11 May 1999). South Melea Planum, By The
Dawn's Early Light. MGS MOC Release No. MOC2-126.

[56] Malin Space Science Systems. (19 July 1999). Indications of Subsurface Ice: Poly-
gons on the Nothern Plains. MGS MOC Release No. MOC2-150.

[57] Malin Space Science Systems. (29 May 2002). Southern Hemisphere Polygonal
Patterned Ground. MGS MOC Release No. MOC2-315.

[58] Mann D. 2003. On patterned ground. Science. 299, 354-355.

[59] Martin S & Kauffman P. 1974. The evolution of under-ice melt ponds, or double
diffusion at the freezing point. J. Fluid Mech. 64(3), 507-527.
Bibliography 141

[60] Matthews PC. 1988. A model for the onset of penetrative convection. J. Fluid
Mech. 188, 571-583.

[61] Merker GP, Waas P & Grigull U. 1979. Onset of convection in a horizontal water
layer with maximum density effects. Int. J. Heat Mass Transfer. 22, 505-515.

[62] Moler CB & Stewart GW. 1973. An algorithm for generalized matrix eigenprob-
lems. SIAM Journal of Numerical Analysis. 10, 241-256.

[63] Moore DR & Weiss NO. 1973. Nonlinear penetrative convection. J. Fluid Mech.
61, 553-581.

[64] Mulone G. 1994. On the nonlinear stability of a fluid layer of a mixture heated and
salted from below. Continuum Mech. Thermodyn. 6, 161-184.

[65] Ng BS & Reid WH. 1979. An intial-value method for eigenvalue problems using
compound matrices. J. Comp. Phys. 30, 125-136.

[66] Ng BS & Reid W H . 1980. On the numerical solution of the Orr-Sommerfeld prob-
lem: asymptotic initial conditions for shooting methods. J. Comp. Phys. 38, 275-
293.

[67] Ng BS & Reid W H . 1985. The compound matrix method for ordinary differential
systems. J. Comp. Phys. 58, 209-228.

[68] Nield DA. 1977. Onset of convection in a fluid layer overlying a layer of a porous
medium. J. FluidMech. 81, 513-522.

[69] Nield DA. 1983. The boundary correction for the Rayleigh-Darcy problem: limi-
tations of the Brinkman equation. J. Fluid Mech. 128, 37-46.

[70] Nield DA. 1991. The limitations of the Brinkman-Forchheimer equation in mod-
elling flow in a saturated porous medium and at an interface. Int. J. Heat & Fluid
Flow. 12, 269-272.

[71] Nield DA. 1998. Instability and turbulence in convective flows in porous media.
In Nonlinear Instability, Chaos and Turbulence, pp. 225-276. Eds. L. Debnath &
D.N. Riahi, WIT Press, Boston, Southampton.
Bibliography 142

[72] Nield DA. 1998. Modelling the effect of surface tension on the onset of natural
convection in a saturated porous medium. Transport in Porous Media. 31, 365-
368.

[73] Nield DA & Bejan A. 1999. Convection in Porous Media. Second Edition. Springer
- Verlag, New York.

[74] Notz D, McPhee MG, Worster MG, Maykut GA, Schlunzen K H & Eicken H. 2003.
Impact of underwater-ice evolution on Arctic summer sea ice. J. Geophys. Res. To
appear.

[75] Orszag SA. 1971. Accurate solution of the Orr-Sommerfeld stability equation. J.
Fluid Mech. 50(4), 689-703.

[76] Payne LE & Straughan B. 1987. Unconditional nonlinear stability in penetrative

convection. Geophys. Astrophys. Fluid Dyn. 39, 57-63 (see also, Corrected and
extended numerical results. Geophys. Astrophys. Fluid Dyn. 43, 307-309).

[77] Payne LE & Straughan B. 1999. Convergence and continuous dependence for the
Brinkman-Forchheimer equations. Stud. Appl. Math. 102,419^439.

[78] Payne LE & Straughan B. 2000. Unconditional Nonlinear Stability in Temperature-

Dependent Viscosity Flow in a Porous Medium. Stud. Appl. Math. 105, 59-81.

[79] Proctor MRE. 1981. Steady subcritical thermohaline convection. J. Fluid Mech.
105, 507-521.

[80] Qin Y & Kaloni PN. 1994. Convective instabilities in anisotropic porous media.
Stud. App. Math. 91, 189-204.

[81] Rahmstorf S. 2001. Abrupt Change. Encyclopedia of Ocean Sciences. Academic

Press, London, p. 1-6.

[82] Rignot E & Thomas RH. 2002. Mass balance of polar ice sheets. Science. 297,
1502-1506.

[83] Roberts PH. 1967. Convection in horizontal layers with internal heat generation,
Theory. J. Fluid Mech. 30, 33^19.
Bibliography 143

[84] Rodrigues JF & Urbano JM. 1999. On a Darcy-Stefan problem arising in freezing
and thawing of saturated porous media. Continuum Mech. Thermodyn. 11, 181-
191.

[85] Rothrock DA, Yu Y & Maykut G. 1999. Thinning of the Arctic sea-ice cover. Geo-
phys Res Letts. 26(23), 3469-3472.

[86] Saloff-Coste L. 2002. Aspects of Sobolev-Type Inequalities. University Press,

Cambridge.

[87] Schmittner A, Yoshimori M & Weaver AJ. 2002. Instability of glacial climate in a
model of the ocean-atmosphere-cryosphere system. Science. 295; 1489-1493.

[88] Schubert G, Turcotte DL & Oxburgh ER. 1969. Stability of planetary interiors.
Geophys. J. R. Astron. Soc. 18,441^460.

[89] Selak R & Lebon G. 1997. Rayleigh-Marangoni thermoconvective instability with

non-Boussinesq corrections. Int. J. Heat Mass Trans. 40, 785-798.

[90] Shir CC & Joseph DD. 1968. Convective instability in a temperature and concen-
tration field. Arch. Rational Mech. Anal. 30(1), 38-80.

[91] Straughan B. 1993. Mathematical Aspects of Penetrative Convection. Longman,

Harlow.

[92] Straughan B. 2001. Surface tension driven convection in a fluid overlying a porous
layer. J. Computational Phys. 170, 320-337.

[93] Straughan B. 2002. Effect of property variation and modelling on convection in a

fluid overlying a porous layer. Int. J. Num. Anal. Meth. Geomech. 26,75-97.

[94] Straughan B. 2003. The Energy Method, Stability, and Nonlinear Convection. Sec-
ond Edition. Springer-Verlag, New York.

[95] Straughan B & Walker DW. 1996. Two very accurate and efficient methods
for computing eigenvalues and eigenfunctions in porous convection problems. J.
Comp. Phys. 127, 128-141.
Bibliography 144

[96] Straughan B & Walker DW. 1996. Anisotropic porous penetrative convection. Proc.
R. Soc. Lond. A 452, 97-115.

[97] Suchomel BJ, Chen B M & Allen MB. 1998. Network model of flow, transport and
biofilm effects in porous media. Transport in Porous Media. 30, 1-23.

[98] Suchomel BJ, Chen B M & Allen MB. 1998. Macroscale properties of porous me-
dia from a network model of biofilm processes. Transport in Porous Media. 31,
39-66.

[99] Tracey J. 1997. Stability Analysis of Multi-Component Convection-Diffusion

Problems. PhD. Thesis, Univ. Glasgow.

[100] Tse K L & Chasnov JR. 1998. A Fourier-Hermite pseudospectral method for pene-
trative convection. J. Computational Phys. 142,489-505.

[101] Unterberg W. 1964. Thermophysical properties of aqueous sodium chloride solu-

tions. DeptEng. UCLA. Rept 64-21.

[102] Veronis G. 1963. Penetrative convection. Astropbys. J. 137, 641-663.

[103] Wadhams P & Davis NR. 2000. Further evidence of ice thinning in the Arctic
Ocean. Geophys Res Letts. 27(24), 3973-3975.

[104] Wadhams P & Davis NR. 2001. Arctic sea-ice morphological characteristics in
summer 1996. Anns. Glaciology. 33, 165-170.

[105] WhitakerS. 1996. The Forchheimer equation: a theoretical development. Transport

in porous media. 25, 27-62.

[106] Young WH. 1912. On classes of summable functions and their Fourier series. Proc.
Roy. Soc. A 87, 225-229.

[107] Zhang K K & Schubert G. 2000. Teleconvection: remotely driven thermal convec-
tion in rotating stratified spherical layers. Science. 290, 1944-1947.

[108] Zhang K K & Schubert G. 2002. From penetrative convection to teleconvection.

Astrophys J. 572,461-476.