exploring conceptual geodynamic models€¦ · introduction 1.1 preamble the mainstream role of...
TRANSCRIPT
Exploring Conceptual Geodynamic Models
Numerical Method and Application to Tectonics and Fluid Flow
Christopher P. Wijns, B.Sc., M.Sc.
Submitted for the degree of Doctor of Philosophy of Geophysicsat the University of Western Australia
School of Earth and Geographical Sciences
September 2004
This page is blank.
DECLARATION
This dissertation is my own composition. All co-authors involved in previous publications
have read and given their consent for inclusion of the relevant material. Any contributions from
co-authors are clearly detailed in section 1.5 of the Introduction.
Chris Wijns Date
Prof. David Groves (principal supervisor) Date
5
ABSTRACT
Geodynamic modelling, via computer simulations, offers an easily controllable method for
investigating the behaviour of an Earth system and providing feedback to conceptual models of
geological evolution. However, most available computer codes have been developed for engi-
neering or hydrological applications, where strains are small and post-failure deformation is not
studied. Such codes cannot simultaneously model large deformation and porous fluid flow. To
remedy this situation in the face of tectonic modelling, a numerical approach was developed to
incorporate porous fluid flow into an existing high-deformation code calledEllipsis. The resulting
software, with these twin capabilities, simulates the evolution of highly deformed tectonic regimes
where fluid flow is important, such as in mineral provinces.
A realistic description of deformation depends on the accurate characterisation of material
properties and the laws governing material behaviour. Aside from the development of appropriate
physics, it can be a difficult task to find a set of model parameters, including material properties and
initial geometries, that can reproduce some conceptual target. In this context, an interactive system
for the rapid exploration of model parameter space, and for the evaluation of all model results,
replaces the traditional but time-consuming approach of finding a result via trial and error. The
visualisation of all solutions in such a search of parameter space, through simple graphical tools,
adds a new degree of understanding to the effects of variations in the parameters, the importance
of each parameter in controlling a solution, and the degree of coverage of the parameter space.
Two final applications of the software code and interactive parameter search illustrate the power
of numerical modelling within the feedback loop to field observations. In the first example, vertical
rheological contrasts between the upper and lower crust, most easily related to thermal profiles and
mineralogy, exert a greater control over the mode of crustal extension than any other parameters.
A weak lower crust promotes large fault spacing with high displacements, often overriding initial
close fault spacing, to lead eventually to metamorphic core complex formation. In the second
case, specifically tied to the history of compressional orogenies in northern Nevada, exploration
of model parameters shows that the natural reactivation of early normal faults in the Proterozoic
basement, regardless of basement topography or rheological contrasts, would explain the subse-
quent elevation and gravitationally-induced thrusting of sedimentary layers over the Carlin gold
trend, providing pathways and ponding sites for mineral-bearing fluids.
CONTENTS
List of Figures v
List of Tables vii
Acknowledgements ix
1 INTRODUCTION 1
1.1 Preamble . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Aims . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Terminology and Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.5 Published Content of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.6 Digital Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 PHYSICAL AND NUMERICAL MODEL 7
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.1 History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.2 Code Overview. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Governing Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.1 Momentum Balance. . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.2 Constitutive Law. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.3 Mass Conservation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.4 Yield Law. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.5 Pore Flow. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Permeability Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3.1 Cubic Lattice. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3.2 Kozeny-Carmen Model. . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3.3 Dilation Angle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3.4 Permeability Tensor Rotation. . . . . . . . . . . . . . . . . . . . . . . 15
2.4 Physical Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Scaling Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.6 Numerical Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.6.1 Pore Pressure Solution. . . . . . . . . . . . . . . . . . . . . . . . . . . 17
i
CONTENTS Wijns Ph.D. Thesis
2.6.2 Pore Pressure Gradient in Momentum Equation. . . . . . . . . . . . . . 19
2.6.3 Darcy Velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.6.4 Interface Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.7 Analytic Solution Benchmarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.7.1 One-Dimensional Compaction of a Porous Halfspace. . . . . . . . . . . 21
2.7.2 One-Dimensional Consolidation of a Porous Halfspace. . . . . . . . . . 23
2.8 Miscellaneous Test Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.8.1 Deflection of Flow Field Due to Permeability Contrasts. . . . . . . . . . 25
2.8.2 Rotation of Anisotropic Permeability. . . . . . . . . . . . . . . . . . . 26
2.9 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3 INTERACTIVE INVERSE MODELLING 27
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.3 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3.1 Model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3.2 Inversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4 MODEL SPACE VISUALISATION 39
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2 Target, Model, and Inversion Results . . . . . . . . . . . . . . . . . . . . . . . . 40
4.3 Parallel Axis Display . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3.1 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3.2 Visualisation of the Inversion Data. . . . . . . . . . . . . . . . . . . . . 43
4.4 Multi-Dimensional Projections . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.4.1 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.4.2 Projections of the Inversion Data. . . . . . . . . . . . . . . . . . . . . . 45
4.5 Self-Organising Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.5.1 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.5.2 Visualisation of the Inversion Data. . . . . . . . . . . . . . . . . . . . . 49
4.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.7 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5 MODES OF CRUSTAL EXTENSION 55
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.2 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.3.1 Constant Viscosity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.3.2 Temperature-Dependent Viscosity. . . . . . . . . . . . . . . . . . . . . 59
ii
Wijns Ph.D. Thesis CONTENTS
5.3.3 Field Validation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.4.1 Continuum Between Modes. . . . . . . . . . . . . . . . . . . . . . . . 62
5.4.2 Comparison with Other Modelling and with Nature. . . . . . . . . . . . 65
5.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6 CRUSTAL-SCALE FLUID FLOW: CARLIN SYSTEM 69
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.2 Geological Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.3 Numerical Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.3.1 Whole crust tectonics. . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.3.2 Upper crust tectonics. . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6.3.3 Fluid flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.3.4 Local scale. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
7 CONCLUSION 79
REFERENCES 81
A Pore Pressure Approximation 91
B Mesh Dependence of Model Results 93
C Crustal Viscosity Profile 95
C.1 Viscosity at Brittle to Ductile Transition . . . . . . . . . . . . . . . . . . . . . . 95
C.2 Integrated Crustal Strength . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
D Digital Appendix on CDROM
iii
LIST OF FIGURES
2.1 Strain weakening function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Permeability – strain relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Ellipsisflowchart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.4 Interface boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.5 1D dry compaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.6 1D saturated compaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.7 1D saturated consolidation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.8 Flow deflection due to permeability contrasts . . . . . . . . . . . . . . . . . . . 25
2.9 Anisotropic permeability rotation . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.1 Inversion target and initial geometry . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2 Inversion generation 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.3 Inversion generation 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4 Inversion generation 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.1 Target, initial geometry, and inversion results . . . . . . . . . . . . . . . . . . . 41
4.2 Sample parallel axis display . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3 Parallel axis display of inversion data . . . . . . . . . . . . . . . . . . . . . . . 43
4.4 Simulation rank versus generation . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.5 Variable parameters versus rank . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.6 2D crossplots with rank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.7 4D projection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.8 Sample SOM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.9 SOM of inversion data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.10 Simulations mapped on to SOM . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.11 Individual parameter SOM plots . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.12 SOM vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.13 Six variations on the SOM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.1 Initial extension model and schematic strength profile . . . . . . . . . . . . . . . 56
5.2 Constant viscosity extension . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.3 Metamorphic core complex mode with temperature dependent viscosity . . . . . 60
5.4 Metamorphic core complex mode with strong faults . . . . . . . . . . . . . . . . 60
v
LIST OF FIGURES Wijns Ph.D. Thesis
5.5 Cooling curves for exhumed lower crust . . . . . . . . . . . . . . . . . . . . . . 61
5.6 Endmember faulting behaviour . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.7 Number of faults versusrτ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.8 Fault spacing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6.1 Tectonic timeline for western U.S. . . . . . . . . . . . . . . . . . . . . . . . . . 70
6.2 Gravity and magnetic images over northern Nevada . . . . . . . . . . . . . . . . 71
6.3 Carlin conceptual model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6.4 Carlin initial models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6.5 Effects of rheology and fault strength . . . . . . . . . . . . . . . . . . . . . . . . 74
6.6 Upper crust . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.7 Upper crust with low basement step . . . . . . . . . . . . . . . . . . . . . . . . 75
6.8 Upper crust with low basement step . . . . . . . . . . . . . . . . . . . . . . . . 76
6.9 Local-scale folding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
A.1 Hydrostatic pore pressure approximation test . . . . . . . . . . . . . . . . . . . 91
B.1 Mesh and accuracy dependence of model results . . . . . . . . . . . . . . . . . . 93
vi
LIST OF TABLES
1.1 Mathematical notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.1 Variable model parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.1 Variable model parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.1 Natural values for extension model parameters. . . . . . . . . . . . . . . . . . . 57
6.1 Natural values for Nevada model parameters. . . . . . . . . . . . . . . . . . . . 72
vii
ACKNOWLEDGEMENTS
Some of the work in this dissertation began during my tenure as an invited researcher at the
University of Western Australia (UWA) and the Commonwealth Scientific and Industrial Research
Organisation (CSIRO) of Australia. During this time, I was financially supported by an initial grant
from Prof. David Groves, and a living stipend from Dr. Alison Ord.
My doctoral work was supported by a scholarship funded jointly between UWA and CSIRO
(WACUP Scholarship), and, soon after my enrolment, I began to receive a top-up scholarship
from the Predictive Mineral Discovery Cooperative Research Centre (pmdCRC). Placer Dome
Asia Pacific Ltd. provided partial tuition funding during my Ph.D., and Mr. Greg Hall and Dr.
John Muntean were specifically involved in directing the modelling of the Carlin gold trend.
I received additional travel funding for conferences from the J.H. Lord Foundation, the Aus-
tralian Institute of Geoscientists, the European Union Young Scientists Fund, Placer Dome, and
the Society of Exploration Geologists.
My thanks go to my primary supervisors, Prof. David Groves and Dr. Alison Ord, for scientific
instruction, financial support, and, not least, for administrative support in securing me funding,
travel opportunities, and the attendant exposure to a world-wide community of geological science.
Valuable day-to-day scientific mentoring came from two sources: Dr. Louis Moresi, who left
CSIRO shortly after I began my Ph.D., and Dr. Roberto Weinberg at UWA, who also left halfway
through my studies. The time I did spend with Dr. Weinberg convinced me to quickly nominate
him as my third supervisor, and both he and Dr. Moresi continued their involvement in my research
efforts.
Drs. Boschetti and Klaus Gessner, at CSIRO, were my closest scientific collaborators during
the latter part of my Ph.D., for, respectively, interactive inversion and the extensional tectonic
modelling. Mr. Thomas Poulet, applied mathematician at CSIRO, helped me to learn much about
self-organising maps and several other data visualisation techniques.
My final and undoubtedly most pressing thanks go to my wife Jennifer, who patiently (most
of the time) put up with my irregular work hours and numerous missed dinners, while helping to
support my Ph.D.
ix
Chapter 1
INTRODUCTION
1.1 Preamble
The mainstream role of numerical modelling in mineral exploration today is forward and in-
verse modelling of potential field and electromagnetic data, and, to a much lesser degree, seismic
inversion. This type of modelling is relatively mature, but belongs to the realm of direct detection
of ore-bearing or ore-related geology. Exploration interest also encompasses the understanding of
physical and chemical processes in ore deposit formation, reflected in geological and laboratory
geochemical research. The recent entry of geophysics into this aspect of exploration is through
numerical modelling of physical and chemical processes involved in deformation and fluid flow.
The understanding of processes underpins direct geological, geochemical, and geophysical ob-
servation. It supplements empirical knowledge for the evaluation of mineral potential. Numerical
modelling of rock deformation, thermal effects, chemical reactions, and fluid flow contributes by
adding to this store of knowledge beyond what field and laboratory studies can deliver. Conceptual
models can be tested for feasibility, many variations can be modelled by altering input parameters,
specific questions can be answered in relation to the geological evolution of a model, and, most
importantly, results can be suggested that lie outside the envelope of the geoscientist’s initial ideas.
The focus in this thesis on the tectonic evolution of the crust is the culmination of numerical
code development (Chapter 2) linked to an interactive system for exploring model space (Chap-
ters 3 and 4). This “exploration system” for geodynamic modelling provides the environment for
research into extensional tectonics (Chapter 5), and the compressional evolution of the Carlin gold
trend in northern Nevada, U.S.A. (Chapter 6).
1.2 Aims
The purpose of developing and applying numerical process modelling to hypothetical and ac-
tual field situations is to predict the behaviour of an Earth system; in relation to mineral explo-
ration, it is to aid in targeting prospective areas. The scope of modelling in this dissertation is
tectonic deformation and associated fluid flow at crustal scales. In all cases, the aim is to deter-
mine, through intelligent exploration of model parameter space, the initial geometries and physical
properties that produce some hypothesised behaviour of the system.
The predictive capabilities of numerical process modelling are limited by two main factors:
“distance” from the initial to the final model state, and accuracy of the model description. The first
point refers to the simple fact that errors accumulate as calculations accumulate. Compounding
1
CHAPTER 1. INTRODUCTION Wijns Ph.D. Thesis
this, the evolution of a simulation is often deterministically chaotic, so that, with increasing time,
or other factors such as deformation, small differences in the initial state may lead to large differ-
ences in the final state. Many simulations should thus be run, with variations on model parameters,
before drawing conclusions about characteristic outcomes. Accuracy, the second factor, refers to
the level and consistency of input detail with respect to the desired outcome, in terms of geometry,
material properties, rheological descriptions, etc. An investigation of fluid flow through first- and
second-order faults, for example, depends on having an accurate representation of all faults with
influence on the chosen scale. In other words, the quality of output depends on the quality of
input. The input – for example, mapped faults – is often sparse, which may lead to output that,
although correct with respect to the model, is incorrect with respect to the actual field situation,
and therefore misleading for site-specific analysis such as in exploration.
It is crucial to understand the accuracy limit in numerical modelling. It decides the degree
to which a model is useful as a predictive tool in a field situation, versus an empirical tool for
understanding field observations. Returning to the above example of fluid flow through faults, a
model that includes all relevant second-order faults can be used to predict fluid flux in a mineral
district. A model that is incomplete can be used to predict fluid flux in relation to the geometries
and permeabilities of the faults in the model, that is, generic information can be compared to
similar field observations, and conclusions drawn in context. A model that relates fault spacing to
rock rheology (Chapter 5) is akin to chemical titration experiments that provide reaction products
based on chemical components. Such information does not locate drill holes in the case of mineral
exploration, but is used empirically by an explorationist in deciphering field observations.
High-deformation numerical modelling, of the kind undertaken in this dissertation, might never
be able to predict the full complexity of the real Earth, where the true heterogeneities are unknown.
Modelling should provide an envelope of behaviour, and this should be checked within the scope
of initial condition perturbations. In light of the above discussion, the numerical results in this
thesis should be considered as empirical aids, albeit based on sound physics, which can provide
new insight into the behaviour of Earth systems and their controlling parameters.
1.3 Methodology
Field observations provide a basis for constructing conceptual models of Earth systems. Nu-
merical models act as laboratories for testing the conceptual models, and, by comparing the results
to the original observations, the conceptual models can be validated or discredited. An iterative
process ensues whereby numerical results act as feedback to field interpretations, which can be
used to construct new hypotheses for numerical testing.
The modelling carried out for this dissertation can be considered the numerical equivalent of
sandbox (or sand and putty) modelling, calledanaloguemodelling. Numerical modelling is poised
to surpass analogue modelling as an investigative technique within the next decade. Analogue
modelling still has the advantage of resolution at fine scales, but, with increasing computer power,
this will disappear. Supercomputers today do allow equal detail, but they are not in common use
2
Wijns Ph.D. Thesis CHAPTER 1. INTRODUCTION
for geological modelling. Analogue models do not have the control over material behaviour that
numerical models have. Although some researchers take great pains in using well-documented ma-
terials and experimental conditions, the practice of mixing analogue materials may lead to proper-
ties that are poorly known outside strict limits of temperature, strain, and strain rate. Furthermore,
ambient conditions such as temperature are sometimes not controlled – for example, lighting over
the experimental setup – which leads to unconstrained material behaviour. Numerical models have
the distinct advantage of absolute control and the possibility of continuous observation at every
point. The weakness of present-day numerical modelling lies in the simplistic constitutive models
used to represent Earth materials. These do not allow the same wealth of deformation styles seen
in analogue models. This is the most serious impediment, as computer power for model resolution
will progress independently of the geosciences.
As indicated above, the strength and duty of numerical modelling is to investigate many vari-
ations on a hypothesis and to delimit the physical parameter space, and the resulting processes,
that control the behaviour of an Earth system. Variation of parameters by trial and error is the
traditional, but inefficient, method for exploring parameter space. Chapters 3 and 4 explain a new
method of interactive modelling that is an efficient system for exploring some target behaviour of
a system. Through various methods of visualising the resulting parameter space, a user can de-
termine limits on parameter values and draw conclusions about controlling factors. A unique and
powerful system for exploring conceptual geodynamic models results from coupling a computer
simulation code to the interactive technology.
1.4 Terminology and Notation
The wordmodelis used throughout this thesis to refer to a particular representation of nature
by mathematical means. This would include, for example, the underlying equations, initial ge-
ometries, material properties, and imposed deformations. The concept of a single model includes
variations of the above components, which are parameterised. Thus, one model gives rise to a
family of results,simulations, or solutions.
Mathematical notation follows conventional norms (e.g., Arfken, 1985). Table 1.1 contains a
list and explanation of mathematical symbols used throughout the text, including the subscript
notation known asindex notationor Einstein notation, which is a compact notation used for vector
and matrix operations.
1.5 Published Content of Thesis
Disregarding the editing required for continuity of the thesis, and revisions made in the course
of new learning, some of the following chapters have been published or submitted, in whole or
in part, as journal articles. Coauthors did not perform the modelling, nor write the articles: their
contributions are summarised explicitly for each article.
3
CHAPTER 1. INTRODUCTION Wijns Ph.D. Thesis
Symbol Explanation
b scalar
b vector
bi componenti of vector
b tensor (matrix)
bij componentij of tensor
xi unit vectors defining a coordinate system (e.g.,i=1,2,3 in 3D)ddx complete derivative with respect to variablex∂∂x partial derivative with respect to variablex
a · b scalar producta1b1 + a2b2 + ... + anbn (in nD)
∆b change in quantityb
∇ “del” operator (vector)[
∂∂x1
∂∂x2
... ∂∂xn
](in nD)
∇b gradient ofb
∇ · b divergence ofb
b,j∂b∂xj
for all values ofj (index notation)
bi,j∂bi∂xj
for all values ofi, j (index notation) (e.g.,bi,i = ∇ · b)
bij,j∂bij
∂xjfor all values ofi, j (index notation)
δij 1 if i=j, 0 otherwise (“Dirac delta”)
X ⇒ Y X impliesY
a → b a tends towardsb
a ≡ b a is equivalent tob
a ≈ b a is approximately equal tob
a ¿ b a is much less thanb
f(b) function ofb
|b| absolute value ofb∑ni=1 bi sum of allbi over valuesi=1 ton
Table 1.1: Mathematical symbols and notation used in the text.
Chapter 3 is published as:
Wijns, C., Boschetti, F., Moresi, L., 2003. Inverse modelling in geology by interactive evolution-
ary computation. J. Struct. Geol. 25 (10), 1615–1621.
F. Boschetti pioneered the use of this interactive method in the geosciences, and contributed to the
text of the Introduction and Method sections. L. Moresi was involved in discussions.
Chapter 4, apart from an expanded introduction and the sections on parallel axis display and multi-
dimensional projections, is published as:
Boschetti, F., Wijns, C., Moresi, L., 2003. Effective exploration and visualisation of geological
parameter space. Geochem. Geophys. Geosys. 4 (10), 1086.
4
Wijns Ph.D. Thesis CHAPTER 1. INTRODUCTION
Although he did not perform most of the work, F. Boschetti is first author because of his idea to
use a self-organising map (SOM) in this context. He produced only the raw graphical SOM output
(no figures) and portions of the text on SOM theory and visualisation.
Chapter 5 is submitted as:
Wijns, C., Weinberg, R., Gessner, K., Moresi, L. Mode of crustal extension determined by rheo-
logical layering. Earth Planet. Sci. Lett., submitted April 2004.
No coauthor contributed any text or figures, but all three coauthors participated in geological dis-
cussions of extensional tectonics, and provided reviews of the manuscript, with R. Weinberg most
heavily involved.
The results of Chapter 6 will be submitted as a paper to an economic geology journal. To date,
none of the eventual coauthors have provided anything more than geological discussion and ideas
for modelling.
1.6 Digital Appendix
A CDROM is included as a digital appendix to this disseration. It contains some code tests
from Chapter 2, the complete sequence of generations for the interactive inverse modelling of
Chapter 3, and selected animations of some simulations for Chapters 5 and 6. All publications and
abstracts related to the thesis research are also included.
5
Chapter 2
PHYSICAL AND NUMERICAL MODEL
2.1 Introduction
2.1.1 History
The predecessor to the present modelling code was a finite element computer code named
Citcom (California Institute ofTechnologyCOnvection in theMantle), developed by Dr. Louis
Moresi at the California Institute of Technology (Moresi, 1992). This code solved the equations
of viscous fluid dynamics, and, as the name implies, was geared towards modelling convection in
the Earth’s mantle. Moresi and Solomatov (1995) provide a detailed description of the multigrid
finite element algorithm. After this and during his time at CSIRO, Dr. Moresi modified the code to
include mobile integration points. The code became a particle-in-cell finite-element code named
CIItcom, which was able to track and evolve material properties on particles. Recent work at
CSIRO has resulted in a new generation of the software, namedEllipsis. The version developed
for this doctoral research incorporates porous fluid flow within the viscous matrix.
The decision to develop porous flow capabilities inEllipsis arose out of the need to fill a gap
in fluid flow modelling capabilities within the former Structural Controls on Mineralisation group
at CSIRO in Nedlands, Western Australia. Previous modelling was restricted to mechanical-fluid
flow models involving limited deformation, using finite-difference software (FLAC, Cundall and
Board, 1988), or large-deformation mechanical models using a particle code (PFC, Guest and
Cundall, 1994; Itasca Consulting Group, 2003) without a porous flow component. In many cases,
FLAC is sufficient for the problems at hand, but modelling of intense folding and thrusting, large
extension, and subduction, with the resultant fluid flow, is beyond the capabilities of this software
at the present time. The incorporation of pore fluid flow into a code allowing unlimited deforma-
tion (Ellipsis) provides a new modelling capability for research, with direct applications to mineral
and petroleum exploration.
2.1.2 Code Overview
Ellipsis is used to model crustal geodynamics by treating rocks as viscous fluids over long time
scales. This thesis does not describe the particulars of the various algorithms in the code, except
those that relate directly to the modification to porous flow. Details of the otherEllipsisalgorithms
are in Moresi et al. (2001, 2002).
The governing equations, solved by a finite-element scheme, are those for a coupled system of
mechanical deformation, porous flow, and thermal transport. The role of particles in the code is
7
CHAPTER 2. PHYSICAL AND NUMERICAL MODEL Wijns Ph.D. Thesis
that of integration points over an element, during the construction of the finite-element matrices.
Thus, the accuracy of the solution depends on the accuracy of this integration scheme and the
ability of the chosen shape functions to represent the variation of quantities across an element.
Bilinear shape functions are used over quadrilateral elements. Physical quantities are attached
to these integration points, or particles, such as material density, strain rate, and porosity. In
particular, the particles may carry time-integrated quantities such as accumulated strain, which
allows history-dependent material behaviour to be modelled. The fact that these time-integrated
quantities are not stored at nodal mesh points means that the underlying computational grid does
not have to deform with the material. There is no limit to the amount of deformation that can
be modelled, as the grid will not become tangled or otherwise distorted beyond the realm of
computational feasibility. An additional advantage is not having to remesh upon large deformation.
Particle positions are updated at the end of every time step based on the calculated nodal velocities.
2.2 Governing Equations
The system under study is that of a fluid-saturated porous matrix. Although the matrix repre-
sents a solid, such as the crust of the Earth, on the time scales involved, it behaves as a viscous
fluid. This assumption is justified by a comparison of typical viscous and elastic parameters for
the crust. The ratio of viscous to elastic shear modulus, which is the ratio of viscosityη to (ap-
proximately) Young’s modulusE, defines the Maxwell time
τMaxwell =η
E.
In problems with a timescale greater than the Maxwell time, the crust can be considered viscous.
For a (lower) crust viscosity of 1021 Pa·s, andE varying between 10 and 100 GPa (Goodman,
1989; Schultz, 1996), the Maxwell time varies from 300 to 3000 years. The scaled time steps in
the simulations of Chapters 5 and 6 are 7500 and 30,000 years, respectively. The Maxwell time
is thus better respected in Chapter 6. A viscoelastic model would be more appropriate for many
problems in geodynamics.
The viscous fluid that is the crust may also deform plastically, that is, it fails at sufficiently high
shear stresses. The pore fluid is single-phase, and considered inviscid except to the extent that the
viscosity controls the pore fluid velocity through Darcy’s law. The dynamics of the system are
governed by the conservation of momentum, mass, and energy. The effect of the pore pressure is
coupled into the stress state of the matrix, and the temperature feeds back into the rheology. In
the following development, the matrix is referred to as the solid, in order to distinguish it from the
fluid that fills the pores. As such, any variable with the superscripts is associated with the matrix,
and the superscriptp denotes the pore fluid.
The physics that describe the flow of a viscous fluid are generally known as the Navier-Stokes
equations, after the independent work of the French engineer Claude-Louis Navier, in 1822, and
the Irish mathematician George Stokes, in 1845 (Lamb, 1895, p. 515). The full Navier-Stokes
equations are difficult to solve, but can be considerably simplified for Earth systems that describe
the slow flow of geological materials. The addition of the pore component results in a two-phase
8
Wijns Ph.D. Thesis CHAPTER 2. PHYSICAL AND NUMERICAL MODEL
fluid dynamic system, described for geophysical applications by, among others, McKenzie (1984),
Scott and Stevenson (1984), and Bercovici et al. (2001).
2.2.1 Momentum Balance
The complete equation for the conservation of momentum of a representative elementary vol-
umeV of material with densityρ, as written by McKenzie (1984), is
d
dt
∫
VρvidV =
∫
VρgidV +
∫
VIidV +
∫
SσijnjdS −
∫
SρvivjnjdS .
In the order of the above terms, the time rate of change of momentum depends on body forces,
interaction with other materials, surface forces, and the advection of momentum. The velocity of
the volume isv, g is the acceleration due to gravity,I is an interaction force per unit volume,
andσ is the stress tensor over the surfaceS of the volume.S has a unit normaln. The material
velocities involved in this problem, as well as their accelerations, are sufficiently small for both
the time rate of change and the advection of momentum to be neglected. The above equation, for
the combined system of matrix and pore fluid, reduces to the simplified form
∂
∂xjσT
ij + ρbgi = 0 ,
where the divergence of the total stressσT is balanced by the body force due to the bulk density
ρb of the material. This bulk density depends on the porosityφ according to
ρb = (1− φ)ρs + φρp .
The total stress is a combination of the matrix and pore fluid stresses:
σTij = (1− φ)σs
ij + φσpij
= (1− φ)σsij − φppδij ,
wherepp is the pore pressure andδij is the Dirac delta function described in section 1.4. The
(inviscid) pore fluid does not support any shear stress.
In a porous medium, the pressure of the pore fluid modifies the stress regime. Biot (1941)
proposes the following relation to describe the effective stressσe due to the presence of a pore
pressure:
σeij = σs
ij + αppδij .
The parameterα is meant to account for grain compressibility. In the most general of cases,
this may be anisotropic andα would be a tensor. For the following development,α = 1, which
corresponds to incompressibility on a grain scale. Through this simplification, it is understood that
compressibility due to pore space dominates the system.
Since it is the effective stress that is related to the matrix strain rate (see below), the total stress
must be recast in terms of the known quantitiesσe andpp.
σTij = (1− φ)
(σe
ij − ppδij
)− φppδij
= (1− φ)σeij − ppδij .
9
CHAPTER 2. PHYSICAL AND NUMERICAL MODEL Wijns Ph.D. Thesis
The equation for the conservation of momentum is then
∂
∂xj
[(1− φ)σe
ij
]− ∂pp
∂xi+ [(1− φ)ρs + φρp] gi = 0 . (2.1)
2.2.2 Constitutive Law
The derivation of mass conservation, in the next section, depends on the materialconstitutive
law, the equation that describes the strain in response to applied stress. A porous material is
considered compressible even though the solid material itself may be incompressible, as in the
present case. Compaction or dilation of the pore space leads to a bulk volume change of the solid.
For this reason, a compressible formulation is required for the constitutive law of the Newtonian
viscous fluid that represents the solid. At high stresses, the solid deforms in a non-viscous manner,
described later, but at low stresses, there exists a linear relationship between stress and strain rate.
The bulk viscosityζ and the shear viscosityη of the solid matrix relate the effective stress to the
velocity according to (McKenzie, 1984)
σeij = ζ∇ · vsδij + η
(∂vs
i
∂xj+
∂vsj
∂xi− 2
3∇ · vsδij
)(2.2)
in three dimensions. In two dimensions the factor2/3 becomes unity.
The bulk viscosityζ depends on porosity as described in the next section. The shear viscosityη
incorporates the effect of the porosity throughη = (1− φ)ηm, whereηm is the true (non-porous)
material parameter. The temperature dependenceηm = ηoe−cT , whereT is temperature, is a
feature of the original code, and is used and explained more fully in section 5.2.
The stress decomposes into the deviatoric partτ and the isotropic partq, which incorporates
the solid pressureps.
σeij = τ e
ij − qδij = 2ηDij − qδij , (2.3)
whereD is the deformation tensor. In the incompressible case,
Dij =12
(∂vs
i
∂xj+
∂vsj
∂xi
)
has zero trace (the diagonal tensor components add to zero), andq becomes the true solid thermo-
dynamic pressure.
Compressible formulation 1. In order to adopt the same formulation for the compressible case,
D is replaced by
D∗ij = Dij − 1
3Dkkδij = Dij − 1
3∇ · vsδij
so thatD∗ has zero trace. In two dimensions, the factor1/3 becomes1/2. The stress decomposi-
tion
σeij = 2ηD∗
ij − psδij (2.4)
is now equally applicable to both compressible and incompressible cases, since in the limit of
incompressibility,
∇ · vs → 0 ⇒ D∗ → D .
10
Wijns Ph.D. Thesis CHAPTER 2. PHYSICAL AND NUMERICAL MODEL
Comparing the stress decomposition (2.4) with the constitutive equation (2.2), the solid pressure
is given by−ζ∇ · vs. In the incompressible case,ζ → ∞ as∇ · vs → 0, and this term becomes
the true thermodynamic pressure. This formulation meets with limited numerical success in the
code, and has therefore been superceded by the next case.
Compressible formulation 2. The following approach continues to use the incompressible form
of D. The compressible part, reflected in the divergence of the solid velocity, is contributed through
the parameterq that contains the remaining terms of the constitutive equation (2.2):
q = −(
ζ − 23η
)∇ · vs . (2.5)
This allows the original form (2.3) of the stress decomposition to be maintained. The finite element
stiffness matrix is constructed as for the incompressible case, and the effect of compressibility is
treated separately when satisfying the continuity, or mass conservation, equation.
The final form of the momentum equation (2.1), using the above stress decomposition (2.3), is
∂
∂xj[2(1− φ)ηDij ]− ∂
∂xi[(1− φ)q]− ∂pp
∂xi+ [(1− φ)ρs + φρp] gi = 0 .
2.2.3 Mass Conservation
The conservation of mass for a medium of densityρ, over the same representative elementary
volume used above, is∂
∂t
∫
VρdV = −
∫
SρvinidS +
∫
VΨdV ,
whereΨ is a volumetric source term. Disregarding any source term, the matrix of porosityφ obeys
∂
∂t[(1− φ)ρs] +
∂
∂xi[(1− φ)ρsvs
i ] = 0 .
In the present formulation, the density of the solid is constant except during the calculation of
buoyancy forces (Boussinesq approximation), so that
∂φ
∂t+ vs
i
∂φ
∂xi=
dφ
dt= (1− φ)
∂vsi
∂xi. (2.6)
The time derivatived/dt represents the material derivative, and the evolution of porosity is calcu-
lated on the moving particles.
The form of the continuity equation that is coupled to the momentum equation comes from the
constitutive law (2.2), and is apparent in the compressible formulation (2.5). Usingλ = ζ− 2/3η,
mass conservation is∂vs
i
∂xi+
q
λ= 0 . (2.7)
The only difference with the original non-porous version ofEllipsis is that the bulk viscosity now
depends on the porosity:
ζ = ζoφo
φ.
11
CHAPTER 2. PHYSICAL AND NUMERICAL MODEL Wijns Ph.D. Thesis
The system of solid plus pore fluid is a compressible system overall, although the solid itself is
considered incompressible. Thus, when all pore space has disappeared through compression of
the solid, the material becomes incompressible. The bulk viscosity, and henceλ, tends to infinity
under this condition, and the divergence of the solid velocity is zero through the above continuity
equation (2.7).
2.2.4 Yield Law
Plastic strain is the proxy for brittle failure inEllipsis. It is a feature of the original code,
but a description here is warranted because brittle failure, in the form of faulting, underpins the
results of all studies in this dissertation. Plastic strain occurs when the viscous stress (equation
2.2) reaches a specified yield stress. At such a point, the viscosity of the material is reduced in
order to respect the limiting stress. Zones of highly localised plastic strain represent faults. The
yield stressτyield may depend on pressure (depth) and previously accumulated plastic strainεp.
τyield = (co + cp p ) f(εp) , (2.8)
whereco is the cohesion, or yield stress at zero pressure, andcp is the pressure dependence of the
yield stress, equivalent to the friction coefficient in Byerlee’s law (Byerlee, 1968). The power law
functionf(εp) mimics the phenomenon ofstrain weakening, which is the progressive weakening
of a zone of plastic failure due to continued deformation.
f(εp) =
1− a (εp/εo)
n εp < εo
1− a εp ≥ εo
,
in whichεp is measured as the second invariant of the deviatoric plastic strain tensor. The roles of
the parametersa, εo, andn are illustrated graphically in Figure 2.1. The “saturation” strainεo is
the accumulated plastic strain beyond which no further weakening takes place, and, at this point,
the yield stress has been reduced by a proportiona. For example, there is no strain weakening
whena = 0. The exponentn describes the shape of the decay curve.
2.2.5 Pore Flow
The fluid filling the pores flows in response to pore pressure gradients according to Darcy’s
law (Darcy, 1856). Modified to include gravitational forces explicitly, the volumetric flow rate per
unit area of pore space, relative to the matrix, is
vpi − vs
i ≡ vpsi = −kij
φµ
(∂pp
∂xj− ρpgj
), (2.9)
in whichk is the permeability tensor andµ is the pore fluid viscosity. This is valid for laminar flow
and is thus restricted to relatively small flow velocities. Within the pore fluid mass conservation
equation∂
∂t(φρp) +
∂
∂xi(φρpvp
i ) = 0 ,
12
Wijns Ph.D. Thesis CHAPTER 2. PHYSICAL AND NUMERICAL MODEL
Figure 2.1: The strain weakening functionf(εp) is determined by the ratio1 − a of minimum to original
yield stress, the “saturation strain”εo after which there is no further weakening, and the exponentn, which
dictates the curvature of the weakening relation.
substitution of Darcy’s law in terms ofvpsi yields the following relation describing the pore pres-
sure:
− ∂
∂xi
[ρpkij
µ
(∂pp
∂xj− ρpgj
)]+ ρp∇ · vs = 0 . (2.10)
The second term represents compaction or dilation of the bulk matrix (i.e., changes in pore space),
and acts as a pressure source term. The density is not eliminated because the pore fluid cannot be
considered incompressible over the crustal scales involved in geological problems.
2.3 Permeability Models
The relationship between porosity and permeability is a field of study in itself. Permeability
models included inEllipsis have been either derived from basic geometry or else adopted from
published literature. Both porosity and strain may affect permeability.
2.3.1 Cubic Lattice
The following derived porosity-permeability law is an extension of an example by Turcotte and
Schubert (1982). A cubic lattice is composed of pores connected by cylindrical pipes of diameter
δ and lengthb. In physical terms,b would be the grain diameter andδ is a measure of the space
between grains. The scalar permeabilityk depends on flow through the connecting pipes only. The
pores at the vertices of the lattice act only to increase fluid storage capacity. In two dimensions,
the cylindrical porosityφk = 2δ/b for δ ¿ b. The 2D average channel (pipe) velocity
u = − δ2
12µ∇pp
gives the volumetrically averaged fluid flow through the lattice, or Darcy velocity
φkvps =
δu
b= − δ3
12bµ∇pp .
13
CHAPTER 2. PHYSICAL AND NUMERICAL MODEL Wijns Ph.D. Thesis
Comparing with the familiar form of Darcy’s law (2.9), the two-dimensional permeability
k2D =δ3
12b=
b2
96φ3
k .
The total porosity including the vertex pores is greater thanφk. Settingφk = αφ,
k2D =α3
96b2φ3 . (2.11)
A similar procedure in three dimensions yields
k3D =πδ4
128b2=
α2
72πb2φ2 .
2.3.2 Kozeny-Carmen Model
The Kozeny-Carmen relation (Carmen, 1956), modified by Mavko and Nur (1997), accounts
for a percolation threshold porosityφc, below which the porosity is effectively no longer connected
and the permeability becomes zero.
k =B
S2(φ− φc)
3 , (2.12)
in whichS is the specific surface area (pore surface area per volume of rock) andB is an empirical
geometric factor. For packed spheres of diameterb (Mavko and Nur, 1997),
S =3(1− φ)
2d.
This leads to
k = B(φ− φc)
3
(1 + φc − φ)2d2 ≈ B (φ− φc)
3 d2 ,
where the factor3/2 has been included inB. Two-dimensional packed spheres obeyS = 4(1 −φ)/d, which leads to the same relation as above with a different value forB. For packed cubes
(3D) with a small separation between faces,S = 6(1−φ)/d, and for squares (2D),S = 4(1−φ)/d.
2.3.3 Dilation Angle
A simple permeability increase with accumulated plastic strain mimics the effect of a dilation
angle in the material.
kij =
ko
ij (1 + ϕε) ε < εo
koij (1 + ϕεo) ε ≥ εo
, (2.13)
whereko is the initial permeability andϕ is the dilation factor (not the actual dilation angle). This
law may be modified by having the permeability begin to decrease after a specified strain has been
reached, as illustrated schematically in Figure 2.2. In this way, older faults may eventually become
less permeable than younger faults.
14
Wijns Ph.D. Thesis CHAPTER 2. PHYSICAL AND NUMERICAL MODEL
Figure 2.2: Schematic illustration of permeability change with increasing accumulated strain. The insert
shows the corresponding (schematic) dilation angleθd (not ϕ from equation 2.13). A dilation angle that
becomes negative (dotted line) results in a permeability decrease after some threshold.
2.3.4 Permeability Tensor Rotation
Anisotropic permeability due to rock fabric, for example, must follow any material rotation.
For this purpose, the orientation of a particle is tracked by means of a unit vectorn, called a
director, with some prescribed initial orientation. In 2D, after some rotation through an angleθ, n
becomesn′, with components
n′1 = − sin θ , n′2 = cos θ .
The rotation matrix
r =
(cos θ sin θ
− sin θ cos θ
)=
(n′2 −n′1n′1 n′2
)
transforms the initial permeability tensor through
k′ = rTkr .
Such a rotation is illustrated in Figure 2.9 in the test cases at the end of this chapter.
2.4 Physical Assumptions
Various assumptions, some of which are noted during the development of the relevant equa-
tions, render the mathematics more tractable. A connected pore network is assumed, such that
isolated pores that hold trapped fluid are not considered, in the sense that they do not affect ma-
terial properties. The system is fully saturated. The porosity is isotropic but the permeability is
not: it may depend on specified anisotropy in the material and/or become anisotropic as a result
of accumulated strain. Capillary pressures are ignored, as are thermal strains in the solid. There is
no individual grain compression in the solid.
15
CHAPTER 2. PHYSICAL AND NUMERICAL MODEL Wijns Ph.D. Thesis
Pore fluid density does vary with pressure and temperature, since water and other possible pore
fluids, over the scale of the crust, do not meet the incompressibility condition
|∆ρ|ρ
¿ 1 . (2.14)
Over a range of pressuresp, this implies
| ∫p ρβdp|ρ
= β|∆p| ¿ 1
in terms of the compressibilityβ, which is about10−9/Pa. For a depth of 20 km, the hydrostatic
pressure alone is 200 MPa, leading toβ∆p = 0.2. In terms of temperatureT , condition (2.14)
implies that| ∫T ραdT |
ρ= α|∆T | ¿ 1 .
For a thermal expansion coefficientα = 10−3/ C and a temperature of 750C at the bottom of
the crust,α∆T = 0.75. These incompressibility conditions are strictly valid for steady flow only.
There is no viscous energy dissipation accounted for in the solid since the Reynolds number is
negligible.
Re =UL
ν≈ 10−24
for velocitiesU varying by 1 cm/year over a length L = 1 km, and a (lower) crustal kinematic
viscosity ofν = 1015 m2/s. Shear heating in zones of plastic strain is treated empirically through
the strain softening equation (2.9).
2.5 Scaling Method
Modelling studies must accurately represent nature if they are to be useful tools. The issue
of scale is crucial to this accuracy. In analogue models, materials used to represent crustal com-
ponents will of necessity have properties, such as viscosity or density, that are not equal to their
corresponding natural properties. In numerical approaches, stability of the code may demand that
mathematical properties are also different than their natural counterparts.Dynamic similarityre-
quires that combinations of properties and dimensions produce similar measures between model
and nature.
A scale for a parameterx is defined by
x∗ =xmodel
xnature.
Dynamic similarity in terms of the equivalence of lithostatic versus viscous stresses demands that
σ∗ = ρ∗g∗L∗ = η∗ε∗ . (2.15)
Throughout this thesis, model densities are always equal to natural densities, and the gravitational
acceleration is also the true value, so that both the density scaleρ∗ and gravitational scaleg∗ are
16
Wijns Ph.D. Thesis CHAPTER 2. PHYSICAL AND NUMERICAL MODEL
equal to unity. The viscosity scaleη∗ is thus uniquely dependent on the length and strain rate
scales, and holds for the solid bulk viscosity as well.
In the yield law (2.8), the cohesionco scales as stress, that is,c∗o = σ∗, but the friction co-
efficient cp and the parameters in the weakening relationf(εp) are dimensionless and therefore
unscaled.
The Peclet number Pe= UL/κ, a measure of advective versus diffusive heat transport, must
be equal for both systems, leading to a thermal diffusivity scale
κ∗ = U∗L∗ = ε∗L∗2 (2.16)
in terms of the chosen independent length and strain rate scales.
TemperatureT can be scaled independently because it drops out of the heat equation:
κ∇2T = v · ∇T
κ∗T ∗
L∗2= U∗T
∗
L∗.
Parameters in a temperature-dependent viscosity law will scale in accordance with model temper-
atures. The thermal expansion coefficient has a scaleα∗ = 1/T ∗.The Darcy velocity must scale as the solid velocityU∗:
k∗
µ∗∇p∗ =
k∗
µ∗σ∗
L∗= U∗ ,
considering that solid and pore pressure scales (σ∗) must also be equal. The permeability scale
then follows:
k∗ =U∗µ∗
ρ∗g∗. (2.17)
Since the liquid viscosity is independent of the solid viscosity,µ∗ can be taken as unity, so that
k∗ = ε∗L∗.
2.6 Numerical Method
This section explains only the numerical solution of the equations relating to pore flow, devel-
oped in two dimensions. Figure 2.3 contains a flowchart for the code. The original multigrid loops
in the Ellipsis code, which solve for solid pressure and velocity, have been disrupted as little as
possible by consigning the pore pressure gradient (c.f. equation 2.1) to the right-hand-side force
vector, which is then updated iteratively within one time step.
There is greater detail in Hughes (1987) on the finite element formulations presented below,
and multigrid loops are explained in a tutorial by Briggs (1987).
2.6.1 Pore Pressure Solution
Following Hughes (1987), the computational domainΩ is bounded by surfacesΓh and Γd
having prescribed pore fluxesh and pore pressuresd, respectively. Ifqi ≡ φρpvpsi represents the
17
CHAPTER 2. PHYSICAL AND NUMERICAL MODEL Wijns Ph.D. Thesis
Figure 2.3: Solution flowchart for the porous flow version ofEllipsis.
Darcy mass flux, the boundary value problem can be stated as follows, using index notation:
qi,i = f ∈ Ω
pp = d onΓd
−qini = h onΓh .
Any source, or forcing terms, are represented byf , andni is the normal to the surface. Although
an exact solution to the problem is defined only at the nodes of the computational grid, the finite
element formulation begins by requiring that areally integrated quantities (overΩ) satisfy the fluid
mass conservation equation (2.10). This produces theweak form
−∫
Ωw,i
ρkij
µ
(pp
,j − ρpgj
)dΩ =
∫
Ωwρp∇ · vsdΩ−
∫
Γh
whdΓh ,
in which w is a weighting function and the source termf is given byρp∇ · vs, that is, the com-
paction or dilation rate of pore fluid mass.
Numerical integration over the domainΩ is made possible byshape functionsthat describe the
variation of quantities between nodes. Each nodeA in the grid has an associated shape function
NA, with the property thatNA is nonzero only within the elements that include nodeA. Ellipsis
uses bilinear shape functions on a rectangular grid of four-noded elements.
NA =4∑
e=1
14
(1− (xe − xA)) (1− (ze − zA)) ,
summed over the four elementse that touch nodeA at coordinates(xA, zA).
18
Wijns Ph.D. Thesis CHAPTER 2. PHYSICAL AND NUMERICAL MODEL
The solutionpp is composed of the unknown and prescribed partsp andd, respectively. When
the weighting function is taken equal to the element shape functionsNA (Galerkin method),
w =∑
A
NAcA , p =∑
A
NApA , d =∑
A
NAdA .
The functionsw, p, andd are linear combinations of the nodal shape functions with constant
coefficientscA, pA, anddA. Letting a permeability factork′ ≡ ρpk/µ , the weak form expands
as∑
A,B
∫
ΩNA,ik
′ijNB,jpBdΩ = −
∑
A,B
∫
ΩNAρp∇ · vsdΩ +
∑
A,B
∫
ΩNA,ik
′ijρ
pgjdΩ
−∑
A,B
∫
Γd
NA,ik′ijNB,jdBdΓg +
∑
A,B
∫
Γh
NAhdΓh . (2.18)
The material propertiesρp andk′ are calculated at the integration points, that is, the particle lo-
cations, and the solid divergence is calculated on an element basis, this being the quantity satisfied
by the solid continuity equation (2.7). In matrix form, equation (2.18) is represented by
Mp = f ,
whereM is the stiffness matrix,p is the vector of unknown nodal pore pressures, andf is the
force vector that holds all the known terms on the right-hand side of equation (2.18). In practice,
the stiffness matrix is built on an element by element basis. In a 2D elemente with coordinate
directions 1 and 2,
meab =
(Na,1 Na,2
)(k′11 k′12
k′21 k′22
)(Nb,1
Nb,2
)
is the element stiffness matrix entry for local node numbersa andb. The quantitiesk′ij are built
by averaging over all particle locations within the element. The entrymeab is added into a global
matrix according to the global numbersA andB of the four local nodes of the element. The force
vector entry
fea = −Naρ
p (∇ · vs)e +(
Na,1 Na,2
)(k′11 k′12
k′21 k′22
)(0
ρpg
)+ Nah
e
usesρp averaged over particles in the element. Solutions onΓd are left out because they are known
and solved separately. The fluxhe is included when prescribed.
2.6.2 Pore Pressure Gradient in Momentum Equation
The pore pressure gradient feeds back into the right-hand side of the governing momentum
equation (2.1) through the term ∑
A,B
∫
ΩNANB,ip
pBdΩ ,
where the nodal pore pressuresppB are known from the previous solution of pore pressure. The
element force contribution for directioni is
fea = NaNb,ip
pb .
19
CHAPTER 2. PHYSICAL AND NUMERICAL MODEL Wijns Ph.D. Thesis
Figure 2.4: The numerical interface approximates the particle interface according to the nodes on which
boundary conditions (BC) are applied.
2.6.3 Darcy Velocity
The pore fluid velocity is the end product of the final pore pressure field. Following a similar
procedure as above, each component of the Darcy flow is calculated independently using equation
(2.9), Darcy’s law.
∑
A,B
∫
ΩNANB (φvps
i )B dΩ = −∑
A,B
∫
ΩNA
kij
µ
(NB,jp
pB − ρpgj
)dΩ .
The element stiffness matrix for componentφvpsi is simply
meab = NaNb
and the force vector is
fea = −Na
1µ
(ki1 ki2
)(Nb,1p
pb
Nb,2ppb − ρpg
)
using known pore pressuresppb .
2.6.4 Interface Tracking
In the event that boundary conditions must be imposed on a material interface, the following
method locates and tracks any interface between two materials in the system. A nearest node algo-
rithm renders the interface only approximate and reliant on the density of the computational mesh
(Figure 2.4). This method is the only way of circumventing the need to mix both quadrilateral
and triangular elements if nodes were to be moved to interfaces. The alternative, which is to have
interfaces crossing through elements, is undesirable when integrating contrasting material proper-
ties within an element. Both pore pressure and temperature boundary conditions may be imposed.
20
Wijns Ph.D. Thesis CHAPTER 2. PHYSICAL AND NUMERICAL MODEL
Fluxes and stresses are far more complicated as boundary conditions because the problem arises
of actually defining the interface and its orientation, so as to integrate along it.
Nodes are deemed to belong to the material defined by the closest particle. The algorithm
works by running through nodes in each column of the mesh and noting boundaries. After the first
pass, a list is kept of nodes that are close to the interface, so that the search does not run through
the entire mesh during subsequent time steps. The list of searcheable nodes is determined by the
maximum displacement of the interface according to the computed time step.
2.7 Analytic Solution Benchmarks
Any numerical code must reproduce, to sufficient accuracy, analytic solutions for simple test
cases. The analytic solutions below have been either derived or taken from published literature.
All cases involve a viscous, porous matrix that is fully saturated with a pore fluid. The governing
equations of momentum and mass conservation assume appropriate simplified forms under the
assumptions of the following problems.
2.7.1 One-Dimensional Compaction of a Porous Halfspace
Compactionis defined as a bulk volume change of the matrix under the influence of applied
stresses, but in the absence of gravitational body forces. When gravitational forces are included,
the phenomenon isconsolidation. Compaction involves the simplified force balance equation
(1− φ)σeij − pp = σo , (2.19)
in which σo is the applied stress on a column of material resting against a rigid, impermeable
boundary.
2.7.1.1 Infinitely compressible pore fluid
In the case of an infinitely compressible pore fluid, for example, vacuum-filled pores, the pore
pressure remains zero. A constant compacting stress of magnitudeσo is applied to the matrix in
the vertical direction. The bulk viscosityζ changes as a result of changing porosityφ according to
ζ = ζoφo
φ,
whereφo is the initial porosity of the column,ζo = ζ(φo), andζ(φ → 0) → ∞. The shear
viscosityη remains constant. The momentum or force balance equation (2.19) reduces to
σo = (1− φ)σezz
= (1− φ)[(ζ − η)∇ · vs + 2η
∂vsz
∂z
]
= (1− φ)(
ζoφo
φ+ η
)∂vs
z
∂z.
21
CHAPTER 2. PHYSICAL AND NUMERICAL MODEL Wijns Ph.D. Thesis
0.0 0.2 0.4 0.6 0.8 1.0t
0.0
0.2
0.4
0.6
0.8
1.0
φ / φ
ο
φο = 0.05
0.020.01
0.20 analyticnumerical
Figure 2.5: 1D compaction of a dry porous column showing the evolution of porosity with time. Initial
porosities are 0.01, 0.02, 0.05, and 0.2.
Recalling that the porosity evolves in time according to
dφ
dt= (1− φ)
∂vsz
∂z,
the final force balance equation is
σo =(
ζoφo
φ+ η
)dφ
dt.
No spatial boundary conditions are required, since no spatial terms are involved in the equation.
Lettingφ′ = φ/φo and integrating,
∆t =[ζo lnφ′ + η(φ′ − 1)
] φo
σo. (2.20)
Computational results compare well with solutions to equation (2.20) for different initial porosities
(Figure 2.5). The relevant parameters areη = 102 Pa·s,ζo = 104 Pa·s, andσo = −103 Pa.
2.7.1.2 Incompressible pore fluid
Since there is no direct relationship for casting pore pressure in terms of porosity, it is difficult to
obtain a solution to the compaction equation (2.19) for porosity through time. Instead, Figure 2.6
contains solutions for the pore pressure with depth at fixed porosities. Parameters are identical to
those for dry compaction.
For vertical compaction, the pore pressure equation (2.10) simplifies to
∂2pp
∂z2=
µ
k
∂vsz
∂z
=µ
k
σo
ζ + η,
22
Wijns Ph.D. Thesis CHAPTER 2. PHYSICAL AND NUMERICAL MODEL
0.00 0.01 0.02 0.03 0.04 0.05 0.06
p p
0.0
0.2
0.4
0.6
0.8
1.0
z
φ = 0.01
0.10
0.05
analyticnumerical
Figure 2.6: 1D compaction of a saturated porous column showing pressure profiles at fixed porosities of
0.01, 0.05, and 0.1.
in which the constitutive equation (2.2) is used to replace the solid velocity gradient. Changes
in porosity change the bulk and shear viscosity values. At fixed porosity (or time), the above
pore pressure equation becomes a second-order ordinary differential equation with the quadratic
solution
p(z) =12
µ
k
σo
ζ + η
(z2 − h2
)
for a column of heighth and basez = 0. The boundary conditions arep(h) = 0, where the fluid
escapes freely out the top, and, since the fluid velocity is zero at the base,dp/dz = 0 atz = 0.
2.7.2 One-Dimensional Consolidation of a Porous Halfspace
2.7.2.1 Incompressible pore fluid
As above, there is no analytic solution for the variation of porosity with time in the case of
consolidation under gravitational forces. Ricard et al. (2001) use a finite difference solver to
calculate porosity versus depth at different times, for an infinite layer from0 ≤ z ≤ lo, and zero
matrix and fluid velocities at these boundaries. Their solution is for the case where the matrix
viscous term dominates in balancing the gravity term, that is, the matrix viscosity is much larger
than the fluid propertyµ/k, leading to a compaction length much greater thanL.
Figure 2.7 compares porosity profiles fromEllipsis with those from Ricard et al. (2001). Be-
cause of different formulations (Ricard et al. (2001) do not use a bulk viscosity), the shear and
bulk viscosity parameters inEllipsis were tuned so thatφ(z=lo) is equivalent att = 10 in both
approaches. Solutions appear to match very well at all times.
In the porous modelling in this dissertation, the effect of consolidation is negligible, because
23
CHAPTER 2. PHYSICAL AND NUMERICAL MODEL Wijns Ph.D. Thesis
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8φ
0.0
0.2
0.4
0.6
0.8
1.0
z / l
o
t=0
t=2.5
t=5
t=10
t=20
a b
Figure 2.7: 1D consolidation of a saturated porous layer showing porosity profiles at different timest. The
initial porosity is 0.05, and all velocities are zero at the basez = 0 and the topz = lo. (a)Ellipsissolutions
are similar to (b) solutions from Fig. 13 of Ricard et al. (2001).
solid compression or dilation, which drives fluid flow, is dominated by boundary velocities that
are always much greater than the natural consolidation velocity.
2.7.2.2 Equal matrix and pore fluid densities
When the matrix and pore fluid are of equal density, the equilibrium momentum equation is
(1− φ)σeij,j − pp,i + ρgi = 0
at time zero when the porosity is constant. The divergence of the effective stress is balanced by
the gradient of the pore pressure and the body force due to gravity.
In the case of one-dimensional consolidation of a porous column on to an impermeable base,
the stress tensor is isotropic and diagonal, the components being equal to the solid pressure pa-
rameterq, which is only the thermodynamic pressure if the system is at rest:
(1− φ)q,i − pp,i + ρgi = 0 .
The pore pressure may be split into a hydrostatic part and a superhydrostatic partp′, in which case
the hydrostatic gradient is equal to the body force term, leaving
(1− φ)q,i = p′,i . (2.21)
McKenzie (1984) derives an analytic solution for the initial consolidation of a porous halfspace,
that is, at timet = 0 (constant porosity), on to an impermeable surface atz = 0. The solid vertical
velocity
vsz = −vp
oφ(1− e−z/δc
),
wherez is the height above the impermeable base,δc is the compaction length, andvpo , the initial
pore fluid velocity at largez, is a constant given by
vpo =
k
µ
(1− φ)φ
(ρs − ρp) g . (2.22)
24
Wijns Ph.D. Thesis CHAPTER 2. PHYSICAL AND NUMERICAL MODEL
Figure 2.8: Flow deflection due to permeability contrasts. A higher permeability center (a) causes the
flow field (black arrows with white tips) to deviate towards the center, whereas a lower permeability center
(b) causes the flow field to be deflected around the center. The profiles on top trace the pore pressure
horizontally through the center of the box.
From McKenzie’s inital pore fluid velocity (2.22),vpo = 0 when the matrix and fluid densities
are equal. Thus, the driving force for flowp′,i = 0 in Equation 2.21, and from this the solid
pressure gradient is zero also. Since the solid pressure is zero at the surface, it is zero everywhere
and the matrix is fully supported by the pore pressure. By the continuity equation,
vsi,i = − q
λ= 0 ,
and by the fact thatvs = 0 at the impermeable base, the matrix velocity is zero everywhere. This
zero velocity result is borne out by code tests.
2.8 Miscellaneous Test Cases
This section contains results for which analytic solutions have not been calculated, but which
may be compared with reasonable expectations. In all cases, the porous medium is fully saturated.
2.8.1 Deflection of Flow Field Due to Permeability Contrasts
Prescribed pore pressures applied to each side of a box maintain a flow field through a porous
material. The permeability is isotropic but there is a permeability contrast in the center of the
box. When the center is more permeable, the flow field, shown by arrows, is deviated towards the
center (Figure 2.8a). In the opposite case (Figure 2.8b), the flow field is deflected around the less
permeable center. The results are similar to those illustrated in Phillips (1991, p. 69).
25
CHAPTER 2. PHYSICAL AND NUMERICAL MODEL Wijns Ph.D. Thesis
Figure 2.9: Anisotropic permeability follows material rotation. The initial state (a) is rotated by approx-
imately 90 clockwise (b), as recorded by the stripes of marker particles. Fluid flow arrows (black with
white tips) from the high-pressure center demonstrate the permeability rotation.
2.8.2 Rotation of Anisotropic Permeability
A box is filled with a porous material having anisotropic permeability. Figure 2.9a illustrates
the initial configuration. The crossed stripes are for visualising the rotation. A constant pore
pressure in the center forces fluid to flow towards the edges of the box, which are held at zero pore
pressure. The moving walls of the box cause the material to rotate, and the anisotropic pore fluid
flow, shown by the arrows, follows the clockwise rotation (Figure 2.9b).
2.9 Chapter Summary
TheEllipsisfinite element code uses a fixed, rectangular computation grid, inside which parti-
cles move, representing different materials and carrying physical information such as porosity and
time-integrated quantities such as accumulated strain. The governing equations of momentum and
mass conservation describe a viscous fluid that represents a solid matrix, such as the crust, with
a porosity filled by an inviscid pore fluid that flows according to Darcy’s law. The solid matrix
can yield plastically, which is an approximation to brittle failure, and zones of concentrated failure
represent faults. Faults may continue to weaken with accumulated strain. Deformation causes the
porosity to evolve, and this, together with topographic differences or fluid sources, provides pore
pressure gradients that drive fluid flow. Simple tests against analytic solutions demonstrate the
accuracy of the code.
Numerical models represent nature when they are accurately scaled, which means that certain
parameter combinations must respect the values they have for real Earth systems. Parameters such
as length, strain rate, and permeability can then be translated into meaningful natural values in
order to assess the model results, and enter into the iterative process of refinement between field
observations and numerical tests.
26
Chapter 3
INTERACTIVE INVERSE MODELLING
3.1 Introduction
The sophisticated modelling of geological processes, made possible by a code such asEllip-
sis, is due to the recent advent of powerful computers. Tracking evolving material properties on
particles, as described in Chapter 2, is computationally intensive, but allows a degree of deforma-
tion not otherwise possible. Various codes in existence today can treat subduction, faulting and
folding, mantle convection, and fluid flow in a rigorous numerical fashion, much in the same way
as traditional geophysical applications such as seismic or potential field problems. As modelling
capabilities have increased, so has the number of parameters that enter into a model, and this has
led to the increasingly difficult task of discriminating the influence of these parameters. In the
non-linear geological systems under investigation today, predicting model behaviour through a
knowledge of individual parameter influences may often not be feasible.
In general,forward modellinganswers questions such as “What response should be expected
from this distribution of material properties under these initial conditions?” (e.g., “What faults will
be generated by this stress field in this particular material?”). The answer is obtained by providing
a computer code with certain input parameters and running the code for a number of time steps.
In many cases, the final result is a geological section or 3D representation.
Most real geological problems require an answer to a question that goes in the opposite di-
rection, that is, “What combination of material properties and initial conditions may result in this
geological response?” (e.g., “What was the stress field that generated these faults in this rock?”).
This is a more complicated problem, for which the answer must be found by iterative numerical
trial and error methods. This is computationally intensive, and the manner in which the search for
an answer is optimised is calledinverse theory, or simplyinversion(Tarantola, 1987).
Inversion is the natural step forward in geological modelling. Reconstructing initial configura-
tions from their geological responses is very much what geology is about. It is an implicit inverse
problem tackled on a daily basis by every geologist. The basic difficulties encountered in an in-
verse problem are: the lack of a guaranteed solution, or the probable existence of many solutions
giving the same answer (non-uniqueness); the efficiency with which a solution can be found that
matches a target within a given tolerance; the sensitivity of a given solution to changes in the initial
conditions. Parker (1977) and Tarantola (1987) explain the mathematical and technical aspects of
inversion.
To overcome the difficulties of inversion, geologists use their intuition and experience to focus
only on the “geologically reasonable” model results that explain the particular features they record.
27
CHAPTER 3. INTERACTIVE INVERSE MODELLING Wijns Ph.D. Thesis
This qualitative knowledge is often difficult to quantify, and the resulting conceptual geological
targets cannot be adequately described by numerical data. This is in contrast to, for example, the
inversion of a gravity profile, where the gravity data are the objective measure to be reproduced. To
compensate for the lack of a quantitative target, the method explained below combines the formal
methodology of mathematical inversion with the knowledge held in observational experience. This
inverse modelling technique can help every time a problem needs visual appraisal of the results or
a priori expert knowledge. The only requirement is a code that allows the user to forward-model
a process and view its result.
3.2 Method
At present, geodynamic modelling is often confined to the forward-modelling stage. Success
has been achieved in quantitative inversion of sedimentary basin models (e.g., Cross and Lessen-
ger, 1999; Bellingham and White, 2000), demonstrating cases where available quantitative data
(i.e., borehole logs and stratigraphic horizons from seismic interpretations) can be used for a direct
measure of misfit. Kaus and Podladchikov (2001) were able to invert a Rayleigh-Taylor instability
to restore initial conditions, but only for very restricted cases of initial geometry. In many applica-
tions of geological modelling, a forward solution is judged visually according to its resemblance
to patterns in the field, to the fact that it does not contradict basic geological principles, or simply
according to the modeller’s expectations.
If one accepts the fact that much of a geologist’s expertise is difficult to quantify, then it is
necessary to incorporate human interaction in directing the inversion process. Recently, research
in artificial intelligence has resulted in systems to support such human interaction in optimisation
problems (Takagi, 2001). They have been used in such diverse fields as graphic design, music
composition, and the engineering of hearing aids. These systems are known collectively under the
term interactive evolutionary computation(IEC).
This chapter extends the use of IEC to geological applications in which visual judgment is
necessary to evaluate model results in the absence of sufficient constraints. The system repre-
sents an advance on traditional, time consuming, trial and error approaches by providing a formal
role in the inversion process for geological experience that cannot be transformed into data. The
traditional numerical measure of data mismatch is replaced by the user’s subjective evaluation.
Humans find it hard to express subjective judgment with absolute values, while they generally find
it much easier to compare different instances of the same process and rank them according to cer-
tain criteria. Consequently, interactive inversion works by producing different possible solutions
and presenting them to the user for judgment and ranking.
Genetic algorithms (GAs) are one search method suitable for the inversion of non-linear func-
tions. Starting with a set of random solutions, these algorithms progressively modify the solution
set by imitating the evolutionary behavior of biological systems (selection, cross-over, and mu-
tation) until an acceptable result is achieved. GAs belong to a class of algorithms that work by
optimising multiple solutions, unlike other classes that optimise one single solution. They are
28
Wijns Ph.D. Thesis CHAPTER 3. INTERACTIVE INVERSE MODELLING
Figure 3.1: (a) Conceptual inversion target showing faults with a spacing on the order of the thickness of the
upper crust. (b) Initial geometry of the numerical model. The upper crust contains light-coloured marker
units with no material differences.
also suitable for handling ranked suites of solutions, which makes them an obvious choice as the
internal engine for interactive inversion applications.
GAs are an established technique today, with a wide range of applications to both theoretical
and industrial problems. Goldberg (1989) gives a basic description of GAs, and Boschetti et al.
(1996) provide a more detailed description of the specific GA implementation used in this work.
The IEC system works by linking a geological forward model (Ellipsis) to the GA. A geologist
runs the forward-modelling code with the aim of producing a geological simulation that matches
a conceptual target. A number of selected parameters is allowed to vary within given ranges. The
GA initially generates a suite of different simulations using randomly picked parameter values.
These simulations could be static geological images or animations showing time evolution. In
the example below, there is no automated method for discriminating between geologically ap-
propriate results, so the geologist ranks each of them according to criteria founded in his or her
experience and knowledge. A relative target misfit is now contained within these rankings. When
this stage is complete, the GA applies parameter swapping between highly ranked results to gen-
erate a new set of simulations that progressively converges towards the target geological section.
As in biological evolution, an element of randomness exists in the generation of new simulations,
so that unexpected results may suggest new possibilities outside the experience or expectation of
the geologist.
3.3 Application
3.3.1 Model
Fault spacing during crustal extension serves as an example to illustrate the inversion process.
The goal is to find a set of material parameters that gives rise to fault spacing on the order of the
thickness of the upper crust, while nowhere enabling the upper crust to be completely pulled apart.
The results of forward numerical simulations are ranked by comparing them to the simplified
line sketch of Figure 3.1(a). It is important to note that the method can proceed without any
29
CHAPTER 3. INTERACTIVE INVERSE MODELLING Wijns Ph.D. Thesis
Table 3.1: Variable model parameters.
Parameter Initial range Increment “Best” value Range
η l (×1.56× 1019Pa·s) 10 - 500 25 400 400
a 0.0 - 0.9 0.1 0.4 0.4
εo 0.0 - 1.0 0.1 0.7 0.7 - 0.9
n 0.4 - 2.0 0.2 1.0 0.6 - 1.0
Four model parameters, described in the text, are free to vary during the inversion, within the ranges and
by the increments indicated. The “best” values give rise to the top-ranked simulation of the final (sixth)
generation. The last column gives the range of parameter values for the five best simulations of the final
generation.
actual target image. The target is included as a guide, and the exact number, location, and dip of
the faults need not be reproduced. In fact, the location of faults is extremely sensitive to initial
perturbations, both in nature and in the numerical simulations, so that it is all the more appropriate
to look for general behaviour and relative spacing. Rheological controls on fault spacing are the
subject of many analogue and numerical modelling studies (e.g., Benes and Davy, 1996; Spadini
and Podladchikov, 1996; Thibaud et al., 1999; Bai and Pollard, 2000). The results generated below
are fairly simple and are intended to illustrate the method.
The model is composed of two initially homogeneous crustal layers. The upper layer behaves
as a visco-plastic material, yielding according to equation (2.8), and the bottom layer has a Newto-
nian viscosity. On top of these is a low-density, low-viscosity background material (“air”), which
does not interfere with the mechanics of the problem. The fact that the mantle is not included is
akin to specifying a strong mantle that does not appreciably deflect during extension. This initial
configuration is illustrated in Figure 3.1(b). Light-coloured horizons in the upper crust are simply
marker units with no physical differences. The upper crust has strain weakening properties (equa-
tion 2.9) that cause initial strain perturbations to localise. The fault geometry and successive fault
spacing arise naturally from the initial conditions of the problem.
The manner in which faults actually weaken is poorly understood, so the three parametersa, εo,
andn in the strain weakening law (2.9) are allowed to vary, in an attempt to gauge the influence of
weakening. The fourth parameter under study is the viscosity of the lower crust. These variables
are listed in Table 3.1, together with their ranges of investigation.
Numerous studies of the Basin and Range area in the western U.S.A. provide data on an envi-
ronment that most agree is extensional in nature, and so many of the parameter values are from the
literature for this region. The overall amount of extension is arguable, from a few tens of percent,
to greater than 100% (e.g., Jones et al., 1992) in certain regions. Total extension in the simulations
below is 50%. The initial3 × 1 model box of Figure 3.1(b) represents natural dimensions of 150
by 50 km. The upper and lower crust are each 20 km thick initially, and after extension the total
thickness of the crust is about 25 km, which is representative of large areas in the Basin and Range,
based on the regional crustal thickness model of Chulick and Mooney (2002).
30
Wijns Ph.D. Thesis CHAPTER 3. INTERACTIVE INVERSE MODELLING
The rate of extension is taken from data typical of the western U.S.A. Based on GPS data
through time, Thatcher et al. (1999) have determined velocities of between 2.8 and 6.5 mm/yr
for various domains of the Basin and Range, and Murray and Segall (2001) find similar rates of
between 2.3 and 3.6 mm/yr. Over the model length scale of 150 km, 5 mm/yr is equivalent to a
strain rateε =1.06×10−15/s.
Model densities for the upper and lower crust are 2700 and 3000 kg/m3. The choice of viscosity
ηu for the upper crust fulfills the maximum yield stress criterion at the interfacezu between upper
and lower crust (see Appendix C.1).
τyield(zu) = ηu ε
The rock cohesionco is 10 MPa (e.g., Suppe, 1985, p. 155), which is close to the value of zero,
relative to the stresses in the model, employed in Byerlee’s law for the upper 10 km of the crust,
but which avoids a cohesionless surface material that would allow plastic strain along the entire
air-rock interface. Numerous laboratory experiments by Byerlee (1968) resulted in a universal
friction coefficientcp of 0.6 – 0.85 for most rock types. However, this is for dry samples. As-
suming an average hydrostatic pore pressure, the solid pressure is reduced by more than one third,
so that an equivalent dry friction coefficient of 0.7 is reduced to 0.44 in the model. A test in
Appendix A shows this hydrostatic pore pressure approximation to be valid in terms of initial
brittle failure depth, although the relationship is certainly more complicated when deformation is
involved. Using these values ofco andcp , the maximum shear strength is about 100 MPa at the
brittle to ductile transition.
3.3.2 Inversion
Each step of the inversion involves running six forward simulations, and the four parameters
in Table 3.1 are allowed to vary. Initially, these parameter values are chosen randomly by the GA,
within the bounds specified. Extension proceeds by applying a uniform velocity to the right-hand
boundary. Figures 3.2, 3.3, and 3.4 illustrate the evolution of results using the IEC algorithm.
Bands of high localised plastic strain represent faults. Accumulated strain in the upper crust is
indicated by darkened material, and the degree of shading is a measure of the amount of strain.
The first generation (Figure 3.2) produces no satisfactory simulation. The image ranked first has
the most desirable fault spacing out of the choices presented, and also exhibits the most clearly
defined faults. The third and fourth simulations are ranked at the bottom because they have resulted
in complete dissection of the upper crust, which, as mentioned above, is an undesirable solution.
The GA now uses these rankings as a measure of relative misfit.
Figure 3.3 contains the second generation of the inversion. From this generation onward, results
are always compared with the best-ranked simulation of the previous generation, in order to assess
convergence towards the target. In this example, there are no second generation results that have
improved on the best first generation simulation. This is still a legitimate and useful result that
tells the GA that it has explored parameter space in the wrong direction. However, overall results
are better in the sense that there are no longer any simulations with a dissected upper crust. The
31
CHAPTER 3. INTERACTIVE INVERSE MODELLING Wijns Ph.D. Thesis
Figure 3.2: First generation of the inversion process. Dark areas in the upper crust represent accumulated
plastic strain. Model results are ranked according to their similarity to the concepts embodied in the target
image of Figure 3.1. (Faint white lines in the lower crust represent uneven particle distribution after de-
formation, and are not a visualisation of any physical process. They are interference patterns between the
deformed distribution of particles and the plotting routine that allocates colours to a regular array of pixels
based on nearby particles. Although particles reproduce in order to accomodate minimum requirements for
computation, they do not necessarily reproduce sufficiently for aesthetic purposes.)
simulation ranked third is so chosen in order to encourage denser fault spacing.
After iterating the ranking process a total of six times, five of the resulting images are qualita-
tively similar (Figure 3.4), and show no evidence of substantial improvement over the best-ranked
result of generation five. It is in fact difficult to assign different rankings to the first five images.
The simulations have a regular fault spacing that corresponds to the target concept. It is appar-
ent that, although the features of the final simulations are satisfactory in concept, they would not
32
Wijns Ph.D. Thesis CHAPTER 3. INTERACTIVE INVERSE MODELLING
Figure 3.3: Second generation of the inversion process. Results are compared with the best-ranked simula-
tion of the first generation.
match up well with the target image through any quantitative direct image comparison.
The outcome of this experiment is a set of the four chosen parameters that leads to the qual-
itative behaviour that was targeted. Table 3.1 lists the final (“best”) values that give rise to the
highest-ranked simulation of the sixth and final generation (Figure 3.4, second simulation). The
table also shows the range of each parameter for the five top-ranked simulations of the final gener-
ation, as a measure of variability within visually equivalent results. The final values permit some
physical conclusions about the system. The lower crust has an optimal viscosity that is about
33
CHAPTER 3. INTERACTIVE INVERSE MODELLING Wijns Ph.D. Thesis
Figure 3.4: Sixth and final generation of the inversion process. Almost all results are qualitatively similar,
with no substantial improvement over the best-ranked simulation of the previous generation.
6.2 × 1021 Pa·s. This is an order of magnitude greater than the upper limit determined by Pollitz
et al. (2001), based on geodetic measurements of post-seismic velocity fields after the 1999 Hec-
tor Mine earthquake in California. Flesch et al. (2000) determine a 100 km vertically averaged
crustal viscosity of between5× 1021 and5× 1022 Pa·s for most of the western U.S.A. Depending
on mantle viscosity and the method of vertical averaging, the inversion result may lie within this
range. The calculation of an effective viscosity for the upper crust influences the result for the
lower crust, so that rather than discuss absolute viscosity values, it would be fairer to state that the
34
Wijns Ph.D. Thesis CHAPTER 3. INTERACTIVE INVERSE MODELLING
optimal fault spacing occurs when the lower crust is about one fifteenth the strength of the upper
crust. Upon examination of all results, it is apparent that smaller viscosities for the lower crust
lead to the dissection of the upper crust (simulations 3 and 4 of generation 1) and exhumation of
the lower crust akin to metamorphic core complex formation. Higher viscosities lead to more even
stretching of the upper crust and probably finely spaced faulting that is below the model resolution
(e.g., simulation 3 of generation 2). The results for the strain weakening law settle on a value
of a = 0.4 that reflects a 40% drop in the yield strength after a total strain accumulationεo of
0.7 to 0.9. Major faults may undergo significantly more weakening than this, according to both
numerical experiments (e.g., Bird and Kong, 1994) and field-based heat flow measurements (e.g.,
Lachenbruch and Sass, 1980, 1992). The less pronounced weakening of the model faults in this
study may contribute to the phenomenon of distributed faulting, whereas greater weakening may
promote a higher degree of localisation on single faults. The last parametern varies between 0.6
and 1.0, indicating that the shape of the weakening curve is less important than the other parame-
ters in controlling fault spacing.
3.4 Discussion
The aim of the inversion in the example above is to arrive at some particular behaviour of the
crust during deformation: behaviour that cannot be sufficiently described by numerical measures.
Finding a suitable combination of parameters that gives rise to this behaviour would previously
have involved one of two more laborious approaches: the manual selection of parameters by trial
and error, or an exhaustive coverage of all parametric space. Trial and error may succeed with a
limited number of parameters, but depends on the user’s knowledge of the coupling and feedback
between parameters, which, in highly non-linear problems involving complex crustal rheologies,
may be impossible. A parametric sweep quickly becomes unfeasible due to the sheer number of
simulations that must be run as the number of parameters is increased. In this example, in excess
of 20 000 simulations would have to be run in order to cover all possible parameter combinations,
and each forward simulation takes a few hours to run on a 935 MHz desktop computer with 500
MB of RAM.
The IEC technique finds multiple solutions with only 36 simulations being run. This vast re-
duction in the number of individual simulations can be attributed to the fact that visual ranking
provides more information in this type of search than numerical misfit in a non-interactive inver-
sion. A result containing one or more features of paramount importance, but with a potentially
large numerical misfit because of, for example, spatial discrepancy in feature locations, is ranked
highly and provides a significant step forward in the search through parameter space. In fact, be-
cause of the combinatorial nature of the GA progression, two images that each contain a different
feature of importance can both be ranked highly in order to increase the likelihood of producing a
new result containing both desired features. Neither trial and error, nor a parametric sweep, takes
full advantage of the expert knowledge of a user, which in this case is the observational experi-
ence of a field geologist. Both the number of variable parameters and the number of individual
35
CHAPTER 3. INTERACTIVE INVERSE MODELLING Wijns Ph.D. Thesis
simulations in each generation is limited by available computer speed. The number of simulations
per generation would ultimately be limited by the capacity of a human to distinguish between, and
rank, the results.
Although the single GA search produces satisfactory results, it should be pointed out that, since
a GA depends on an element of randomness, the most correct procedure involves many GA trials.
In a statistical sense, the greater the number of trials, the more likely the optimal solution will
be found. In practice, the computational intensity of geodynamic modelling would not generally
allow this.
The power of inversion lies in demonstrating the range of non-uniqueness of a solution. A very
simple and quicka posterioriinvestigation highlights the sensitivity of the results to changes in the
variables. For example, referring to the five best simulations (Table 3.1), the viscosity of the lower
crust has settled on a unique value, which suggests that it exerts a strong influence on the outcome
of the problem. The same is true of the final amount of strain weakeninga. A parameter such
as the saturation strainεo has a weaker influence on the outcome, since it varies slightly without
affecting the final qualitative picture. The shape of the strain weakening function, determined
by n, has the least influence. With the inversion approach in general, most late-stage results are
close to the target in parameter space, and so a back-analysis through all generations is instructive
for looking at the sensitivity of solutions. However, such conclusions can only be drawn with
confidence on the strength of sufficient sampling of the model parameter space. To this end, and
to facilitate an understanding of the physical controls on the behaviour of the system, Chapter 4
concentrates on the effective visualisation of the whole multi-dimensional parameter space.
An important component of this interactive inversion technique is its ability to embrace un-
forseen results. An intuitive approach relies entirely on the experience of the modeller, and may
miss realistic targets that lie outside the realm of modelling space that is envisaged. The GA,
although converging to a specific area in parameter space, also provides for random solutions. If
ranked highly, such a random solution may open up an entirely different class of simulations that
also yield realistic results.
3.5 Chapter Summary
Conceptual models in geology are often difficult to describe numerically. They mostly reflect
the translation of qualitative field knowledge into verbal descriptions or geological cross-sections.
This usually precludes implementing the traditional notion of comparing a numerical simulation to
a target via one or more numbers. The interactive inversion system circumvents this difficulty by
using a visual comparison between forward model results and a target, in order to guide a genetic
algorithm to the parameters that generate a preferred solution. At each step, ranking between
interim results acts as a proxy for a numerical misfit.
The technique of interactive evolutionary computation considerably diminishes the effort re-
quired to explore parameter space during the inversion of conceptual models in geology. The
method is particularly geared towards cases of highly non-linear interactions between material
36
Wijns Ph.D. Thesis CHAPTER 3. INTERACTIVE INVERSE MODELLING
parameters, where the resultant behaviour is difficult to predict. This visual approach exploits
the unquantifiable aspects of an expert user’s knowledge in a situation where this is currently an
under-utilised resource.
37
Chapter 4
MODEL SPACE VISUALISATION
4.1 Introduction
As alluded to in Chapter 3, the end product of most inversion procedures is a single best solu-
tion. However, many solutions are generated during the course of the process. The focus of this
chapter is on maximising the information to be gained from an analysis of all solutions. Since the
user ranks each solution in the genetic algorithm search described in Chapter 3, this information
should be used to obtain some understanding of the parameter space, rather than be discarded at
the end of the inversion run. This, in turn, leads to a better understanding of the dynamical prob-
lem. From the statistical tools that are available to perform such an analysis, visual approaches
are preferable. The reason is twofold: first, visual displays are in keeping with the qualitative de-
scription and evaluation process for each simulation result, and second, this can provide a further
level of interactivity between the user and the inversion. If intermediate displays of the cumulative
solution space are prepared, the user not only judges the quality of each individual solution, but
follows the inversion progress, and, eventually, may control the direction of parameter exploration.
The three different techniques used below serve as complementary methods for visualising the
accumulated results of an inversion. These are aparallel axis display, multi-dimensional projec-
tions, and aself-organising map. In each case, a few simple plots reveal which combinations of
parameter values lead to “good” solutions, and what role individual parameters play in contributing
to “bad” solutions. The plots also offer straightforward assessments of the non-uniqueness of the
problem, and the extent to which the inversion has covered the specified ranges of investigation.
The example in this chapter is a modification of the previous crustal extension problem.
Before presenting the different visualisation techniques, it is imperative to caution on two
points. First, any reduction from a high-dimensional problem to a lower-dimensional visuali-
sation results in information loss. Of the three techniques presented, only the parallel axis display
does not involve reduction of dimensionality, that is, all data points are faithfully reproduced. Sec-
ond, any statistical method improves with the number of data points. The intensive computation
involved in geodynamic modelling results in relatively few simulations, and hence a paucity of
data for statistical manipulation. This cannot be avoided, and the challenge is to draw conclusions
from sparse data, keeping in mind that the addition of new data will usually change the working
view of parameter space. To reinforce an appreciation of the poor coverage of model space in a
typical study, the following problem has six variables per simulation and a total of 48 simulations.
The number of points needed to simply define a 6D hypercube is 26 = 64. The available data are
therefore equivalent to fewer than two points in 1D – sparse data by any stretch of the imagination.
39
CHAPTER 4. MODEL SPACE VISUALISATION Wijns Ph.D. Thesis
Table 4.1: Variable model parameters.
Parameter Initial range Increment
viscosityηu (× 6.25×1018Pa·s) 5000 - 10000 1000
cohesionco (× 0.025 MPa) 0 – 2000 200
friction coefficientcp 0.1 – 1.0 0.1
tensile stress limitσc (× 0.025 MPa) 100 – 1000 100
strain weakeninga 0.1 - 0.9 0.1
saturation strainεo 0.1 - 1.0 0.1
Six model parameters, described in the text, are free to vary during the inversion, within the ranges and by
the increments indicated.
In spite of this, interactive inversion does produce good results, which may indicate that the non-
linearity of the problems is not great. “Non-linearity” in this case refers to a simple shape of the
solution space as defined by the rankings of the user, not the actual non-linearity of the physics.
This chapter illustrates methods of data visualisation that represent the best attempts in an area
of ongoing research. Wherever possible, different techniques, with different and complementary
strengths, should be used in parallel.
4.2 Target, Model, and Inversion Results
In this variation on the inversion in Chapter 3, the goal is to find one or more sets of material
parameters that give rise to fault spacingwider than the thickness of the upper crust, as sketched
in Figure 4.1a. The model and initial geometry (Figure 4.1b) are similar to the previous example.
The visco-plastic upper crust, with strain-weakening behaviour, is 15 km thick. This lies over
6 km of weaker, ductile lower crust. Above these is a highly compressible layer of low density,
low viscosity background material (“air”), which does not interfere with the mechanics of the
problem. The right-hand boundary extends with a uniform velocity equivalent to an average strain
rate of10−15/s. The thin lower crust would not in general be as representative of natural conditions
as the thicker model in Chapter 3. Although it may describe specific geological settings with a
very thinned lower crust, such as the extended area around Naxos in the Aegean (Jolivet, 2001),
or the crustal structure under parts of Antarctica (Mishra et al., 1999), there is no intent to link this
model to a particular field example.
The six model parameters that are allowed to vary are the viscosityηu of the upper crust, and
five yield parameters that describe the plastic failure of the upper crust (equations 2.8 and 2.9).
These are the cohesionco, the pressure dependence or friction coefficientcp, the strain weakening
parametera, the “saturation strain”εo, and a new parameterσc, the tensile limit of the crust. If the
solid pressure is negative (i.e., tensile stress) and has a magnitude greater thanσc, the viscosity is
reduced threefold to simulate tensile failure. Table 4.1 lists the variables together with their ranges
of investigation.
40
Wijns Ph.D. Thesis CHAPTER 4. MODEL SPACE VISUALISATION
Figure 4.1: (a) Conceptual target and (b) initial geometry of the crust. Panels (i) to (iii) hold the first two and
the last (sixth) generations of the inversion algorithm. Images are ranked within each generation according
to the spacing of faults (dark bands) in the upper crust. Unranked simulations have not extended because of
numerical non-convergence. The third image of panel (iii) shows the greatest fault spacing.
Six generations of simulations are run, with eight simulations per generation. After each gener-
ation, the user ranks the output images against the conceptual target (Figure 4.1a). As in Chapter
3, ranking is based on general behaviour and relative fault spacing, rather than trying to repro-
41
CHAPTER 4. MODEL SPACE VISUALISATION Wijns Ph.D. Thesis
Figure 4.2: Schematic diagram of a parallel axis display for a three-parameter problem. Individual param-
eter values are joined to indicate the two solution vectorss1 ands2.
duce the image exactly. “Multi-criteria” human evaluation, where solutions are ranked according
to separate, but simultaneously evaluated criteria, is more effective than trying to devise a single
numerical target. The bottom panels of Figure 4.1 show the progression of the outputs and the
rankings accorded to each image for the first, second, and sixth (last) generations. By the sixth
generation, several images appear very similar, and represent “satisfactory” solutions. Unranked
solutions do not converge numerically, meaning that the computer code fails to calculate any result.
As discussed below, this is due to unrealistically high yield strengths.
At the end of the inversion, each solution is ranked in an absolute fashion, against all other
solutions, from 1 to 10, with 1 being the best. This ranking is distinct from the intra-generational
ranking that guides the genetic algorithm (GA), as displayed below each image in Figure 4.1, and
provides a classification fora posteriorianalysis of the inversion results.
4.3 Parallel Axis Display
4.3.1 Theory
The parallel axis display is a basic means of plotting all parameter values for all solutions. Each
parameter that makes up a component of the solution vector is assigned to a successive point on
the abscissa, and its normalised values are plotted above it on a common ordinate axis. Parameter
values belonging to a common solution vector are joined together (Figure 4.2). Unlike the next
two methods, the parallel axis display is quickly limited by the number of solutions that can be
plotted together before overcrowding makes the graph illegible.
One method for mitigating the inevitable obscuring of information that arises through the su-
perposition of similar solutions is to offset plot points, by either spreading them around each
parameter location (along the abscissa) or adding random noise to each parameter value. A com-
bination of both techniques proves useful for visualising the inversion data below.
42
Wijns Ph.D. Thesis CHAPTER 4. MODEL SPACE VISUALISATION
Figure 4.3: Parallel axis display for all inversion solutions, ranked by colour from red (best) to purple
(worst). For easier visualisation of overlapping solution segments, points are spread slightly around param-
eter locations and some parameter values have small random noise added.
4.3.2 Visualisation of the Inversion Data
The inversion results consist of 48 points in a 6D space, the six dimensions corresponding to
the six model variablesηu, co, cp, σc, a, andεo. These data represent the extent to which the
inversion procedure has sampled the parameter space. The parallel axis display in Figure 4.3 was
produced with the free softwareggobi, available from http://www.ggobi.org. The best solutions
(red) are the result of a high viscosity, a low cohesion, friction coefficient, and tensile limit, and
pronounced fault weakening. The saturation strain varies.
It is apparent that the sampling has fallen short of the alotted range for the four parameters
co, cp, a, andεo, since points are not distributed over their full spectra. More simulations could
be run in order to fill in some of the missing data and increase the confidence that good results
have not been overlooked in unexplored space. In the most rudimentary manner, new parameter
combinations can be chosen manually by referring to blank areas in Figure 4.3. Based on the user
rankings of these new outputs, the GA would incorporate or cull this additional parameter space.
Highly ranked solutions, in orange and red (Figure 4.3), are in general quite clustered. They
occupy single values forco andcp, a small range of values forσc, and show greater variability for
the remaining three parameters. A possible conclusion is that model behaviour is strictly controlled
by co andcp, whereas good solutions exhibit some degree of non-uniqueness with respect to the
other parameters. The caveat is that a GA can possibly become locked in to particular values for
non-optimal reasons, including “the luck of the draw” in the first random distribution of inputs
to the model. One safeguard against this is to increase the frequency of mutations, leading to
more simulations that are not created according to the ranking. However, too high a mutation
rate may detract from the convergence to target behaviour. Increasing the number of individuals
per generation can also guard against locked values, but computation time and the challenge of
effectively ranking all solutions may preclude this option. In this example, the difference between
good (orange) solutions and the best (red) solutions lies in the parametersηu anda.
The worst-ranked solutions, in purple, including those simulations that failed to converge nu-
43
CHAPTER 4. MODEL SPACE VISUALISATION Wijns Ph.D. Thesis
Figure 4.4: Simulation rank versus generation showing overall progress of the inversion towards better
results (1= best rank, 10= worst). Small random noise is added to each rank for easier visualisation of
close results. Eight simulations sit within each generational interval from 1 to 6.
merically, are widely spread in parameter space. This suggests that no single parameter is re-
sponsible for bad solutions. Highco andcp values characterise many of the numerically invalid
simulations. Fortunately, these high values translate into unrealistically high yield stresses, mean-
ing that this part of parameter space can be safely neglected.
4.4 Multi-Dimensional Projections
4.4.1 Theory
The simplest method for projecting data from high dimensions into 2D is to neglect all but two
parameters (dimensions) at a time. This is called anorthogonal projection, that is, all neglected
dimensions have their axes orthogonal to the viewing direction and therefore invisible. Such linear
projections are also known ascrossplots. Labels or colours can add a third dimension. As for the
parallel axis plot, small random noise is added to parameter values to avoid losing sight of multiple
points that are superimposed in 2D.
Perspective views of 3D data can be well represented in 2D, to which labels or colours can
again be added to achieve a fourth dimension. Beyond this it is necessary to use non-linear trans-
formations of the data to reflect, in 2D, relative distances between points innD. One suchmulti-
dimensional scaling, pioneered by Sammon (1969), is applied by Wijns et al. (in press) to evaluate
the effectiveness of an interactively guided 9D inversion.
44
Wijns Ph.D. Thesis CHAPTER 4. MODEL SPACE VISUALISATION
Figure 4.5: Orthogonal projection of each variable parameter versus simulation rank (1= best rank). Small
random noise added to each parameter value and rank prevents strict superposition of points.
4.4.2 Projections of the Inversion Data
All projection plots in this section were produced with theggobisoftware used for the parallel
axis display. Figure 4.4 offers a simple way to judge the overall progress of the inversion towards
better-ranked simulations. With each successive generation, there are more solutions with good
ranks (low numbers), but the spread of rankings shows that the GA is still exploring far from the
well-ranked parameter space.
The series of six plots in Figure 4.5 illustrates the entire space explored by the GA as well as
any convergence to specific parameter values with respect to rank. The same conclusions hold as
were derived from the parallel axis display: highηu, low co, cp, andσc, high a, and variableεo
produce good results. The rapid convergence ofco (Figure 4.5b) is easy to visualise. Based on
a relatively complete coverage ofco space, a high cohesion invariably leads to bad results. Gaps
in the exploration of the parameter ranges are even more obvious than in the parallel axis plot.
The GA has missed a swath of intermediate values forcp, σc, and especiallyεo. Extremities have
not been sampled forcp, a, andεo. This type of analysis is essential for determining the strength
of any conclusions about parameter controls on the system. A highηu andco near zero can be
confidently concluded to lead to good simulations. It is unlikely that a highcp will produce good
solutions, but the true bounds between 0.0 and 0.4 have not been established (Figure 4.5c). The
range of good values forσc is similarly inconclusive (Figure 4.5d). The GA has not explored the
possibility of strong faults (lowa), but the evidence points to significant weakening as essential
45
CHAPTER 4. MODEL SPACE VISUALISATION Wijns Ph.D. Thesis
Figure 4.6: 2D crossplots with a third dimension – rank – indicated by colour, from red (best) to purple
(worst). Small random noise added to each parameter value prevents strict superposition of points.
for good solutions (Figure 4.5e). Finally, no conclusions should be drawn about the control ofεo
(Figure 4.5f).
Colour represents the third dimension of rank in the selected crossplots of Figure 4.6. The
best-ranked results (red and orange) show combinations of high viscosity and low cohesion (Fig-
ure 4.6a), low cohesion and low friction coefficient (Figure 4.6b), and low tensile limit with gen-
erally large strain weakening (Figure 4.6c). The larger spread of good points in Figure 4.6a and c
shows that a range of high viscosity, and a range of large strain weakening, can produce acceptable
results. Referring only to the good points in Figure 4.6a, if the value ofco is fixed near zero, all the
variability is expressed throughηu, which is then known as theprincipal component(e.g., Jolliffe,
1986). A principal component rotation (PCR) seeks to express a maximum of the variation in data
through a minimum of axes. Like orthogonal projections, a PCR is a linear transformation that
may not always be appropriate for analysing non-linear data sets.
Figure 4.7 shows the result of manipulating four variables projected into 2D plus rank (i.e., 3D).
The four parameter axes are rotated into the position that provides the most distinct clustering of
similar solutions ranked in the top half. The nearly verticalσc axis indicates that this parameter is
extraneous in distinguishing between the better clusters. The displayed variability is orthogonally
resolved between thea axis and a new axis, for example,c′, that is a combination ofco andcp. All
mediocre (yellow) to good points are due to a lowc′ value, in keeping with previous conclusions
(e.g., Figure 4.6b). The three better clusters are aligned along thea axis, revealing that fault
weakening is responsible for differences between more highly ranked simulations.
4.5 Self-Organising Map
4.5.1 Theory
A self-organising map (SOM) is a transformation of high-dimensional (nD) data into a lower-
dimensional (usually 2D) plot (Kohonen, 2001). Broadly speaking, its aim is to spreadnD points
over a plane in such a way that topology is respected, that is, two points lying close to one an-
other in the higher-dimensional space should lie close in the 2D plot. In doing so, it acts as a
46
Wijns Ph.D. Thesis CHAPTER 4. MODEL SPACE VISUALISATION
Figure 4.7: Projection of four variables plus rank into 3D. Rank is indicated by colour, from red (best) to
purple (worst). Small random noise added to each parameter value prevents strict superposition of points.
Axes have been rotated so as to isolate three groups of better simulations. The box contains normalised
projection distances for the four axes, along with total axis length in brackets. Theσc axis is almost vertical.
classification algorithm that assigns the input data to cells according to similarity.
The description and analysis of the mathematics underlying a SOM can be found both in the
neural network literature and in the framework of statistical clustering tools. The following de-
scription of the algorithm highlights the simplicity of the technique. Given the task of visualising
a set ofm data points innD:1. Generate a (2D) grid. This grid can have different shapes and sizes, and its nodes can also
have different shapes. The example in this section uses a grid with hexagonal nodes, which
allow contact with more nearest neighbours than rectangular nodes. The optimal number of
nodes to representm data points is determined via a heuristic calculation, in the absence of
any established theory.2. Initialise each node by assigning it a randomnD vector of the same dimensionality as the
data points. These are the SOM vectors.3. Pick one data point and find the SOM vector that is closest to the data point according to a
certain metric (usually a simple Euclidean distance after normalisation of the dataset). This
node is thebest matching unit(BMU). Modify the SOM vector belonging to the BMU by a
certain amount so that it is closer (more similar) to the data point.4. Choose a neighbourhood of the BMU over the grid. The SOM vectors belonging to this
neighbourhood are also modified to be closer to the data point, by an amount inversely pro-
portional to their grid distance to the BMU, that is, the farther the location of the vector on
the SOM grid, the less it is incremented.5. Repeat items 3 and 4 for each point in the dataset.6. Iterate items 3 to 5 several times (usually a few thousand) to ensure convergence.The result of this algorithm is that SOM vectors belonging to nearby grid locations will tend to
be similar, since they will tend to converge towards similar data points. Conversely, data points that
47
CHAPTER 4. MODEL SPACE VISUALISATION Wijns Ph.D. Thesis
Figure 4.8: Illustration of nodes in a SOM. Two 3D SOM vectors(T1, P1, t1) and(T2, P2, t2), separated
by a distanced, belong to two SOM data nodes. The distance node that connects the data nodes is coloured
according to the magnitude ofd. Data nodes are also coloured, using an average of surrounding distance
nodes, for a more continuous display.
are close to one another will tend to fall in grid nodes close to one another, or in the same grid node.
This should achieve the dual purpose of clustering similar points and giving a best approximation
of the original topology in thenD space. In practice, all parameter values are normalised according
to their data extents before any calculations are made. Thus, each component is equally weighted
in the determination of distance. This can be both desirable or not. In the event that comparisons
are to be made between multiple GA trials, the data extents may be different. Computed distances
can therefore not be compared, leading to very different views of the parameter space. A better
approach, in this case, is to normalise according to a subset of the “best” results, which will provide
a more stable set of limits. This presupposes that at least two parameter values are available per
dimension, to provide a minimum and maximum. Very large and very small normalised values
will indicate outliers from the subset of best data points.
Clearly, what a SOM tries to achieve (mappingnD points into 2D) is impossible except for
trivial cases. Consequently, the final result is often non-optimal. The SOM depends on a number
of parameters that control the algorithm described above. These include the number of grid points,
the initialisation of the SOM vectors, the amount by which SOM vectors are updated towards the
data point, the size of the BMU neighbourhood, and the metric used to calculate distances. There
are no standard values for these parameters in the literature, and their choice is problem specific
and often left to the user’s discretion. Also, because the SOM grid is not tied to any specific refer-
ence location in thenD space, for example, the center of the SOM is not the center of thenD space,
the appearance of maps with different initialisations may be very different. This fact is demon-
strated in the next section, and the website http://davis.wpi.edu/∼matt/courses/soms/applet.html
also illustrates the effect of different parameter choices. Provided satisfactory convergence has
been achieved, the relationship between adjacent clusters should still be represented in a relatively
homogeneous way, and a trained user will recognise this.
The SOM display in Figure 4.8 employs two different types of nodes:data nodesanddistance
nodes. The data nodes represent the SOM vectors at the grid locations described above. The
distance nodes connect the data nodes, and indicate the relativenD distance between adjacent
SOM vectors. The distance nodes are coloured to show the magnitude of the distance between
adjacent data nodes. Data nodes are also coloured, according to an average of surrounding distance
48
Wijns Ph.D. Thesis CHAPTER 4. MODEL SPACE VISUALISATION
Figure 4.9: SOM of the extensional faulting data. (a) Data nodes, indicated by superimposed dots, are
separated by distance nodes that indicate the difference between neighbouring SOM vectors. The colour bar
shows the distance scale. (b) Only data nodes are present, labelled with the absolute rank of the associated
simulations. Unlabelled nodes contain no input vectors. The dashed line encircles the domain of best model
results. Arrows illustrate three equivalent data nodes between the two displays.
nodes, to produce a more continuous map. Colours representing short distances indicate clusters
of close points, while colours representing large distances show cluster borders, which can also
be interpreted as steep ridges dividing valleys of similar points. This relationship between SOM
nodes is valid in a statistically averaged sense: over many maps of the same data, the colours will
reflect the distances between data points in a fair manner, but it is possible that, in a non-optimal
map, points that are close in data space are not close in the SOM. The only guaranteed relationship
in a SOM is that immediate neighbours represent points that are close in data space.
SOMs have been extensively employed in recent years in both scientific and engineering ap-
plications in order to visualise high-dimensional data and highlight data structure. The SOM plots
below have been produced with the use of the MatlabTM SOM Toolbox, written by Juha Vesanto.
More details about SOMs, and the specific implementation used in this work, can be obtained at
the SOM Toolbox website http://www.cis.hut.fi/projects/somtoolbox.
4.5.2 Visualisation of the Inversion Data
The input data for the SOM visualisation consist of 48 points in a 6D space, or 48 vectors of
six components each (η, co, cp, σc, a, εo). Figure 4.9a contains the main SOM plot, in which
thirty-five data nodes, indicated by superimposed dots, are considered optimal to classify the 48
input vectors. Depending on the convergence, each SOM data node may correspond to one or
more of the 48 output images from the simulations, or it may be “empty”, meaning that no data
point is close to it. A cluster in the SOM may be recognised as any collection of darker blue nodes
that is bounded by light blue to red nodes. Such an area should represent a domain of similarly
49
CHAPTER 4. MODEL SPACE VISUALISATION Wijns Ph.D. Thesis
Figure 4.10: Mapping of model outputs on to the SOM grid of Figure 4.9a. Data nodes (every second node)
represent outputs from the inversion process, as illustrated by the mapped images. Images are labelled with
their absolute rankings (c.f. Figure 4.9b).
parameterised simulations.
The original data vectors can be represented on the SOM by plotting each point at the grid
location of its corresponding BMU (the grid node that is most similar to it). Each SOM node
is then labelled with the rank of its associated simulations (Figure 4.9b). Well ranked solutions
(closer to 1) are similar to the target. Unlabelled nodes are empty, as described above. It is clear
that the top left domain, encircled with a dashed line, contains the best cluster of model images.
Figure 4.10 illustrates the mapping of the output images on to the SOM. Both data nodes
and distance nodes are included, as in Figure 4.9a, and the output images are labelled with their
absolute rank. It is again clear that the area in parameter space at the top left corner contains the
solutions most closely resembling the target image. The simulation ranked 5.5 in Figure 4.10 is
far from the “best” corner, being separated by a ridge, but it nevertheless shares characteristics
of the best results. This may be indicative of either the non-uniqueness of the inversion problem
(simulations with very different parameter inputs may produce similar outputs) or of a non-optimal
SOM clustering. Further analysis will prove that this point is indeed far from the best simulations
in parameter space, and is thus well represented in the SOM.
In order to analyse the influence of each model parameter on the solution, the individual di-
mensions are mapped on to the SOM network of data nodes. The images in Figure 4.11 show
the (non-dimensional) magnitude of each model parameter at each node in the SOM. Viscosity
50
Wijns Ph.D. Thesis CHAPTER 4. MODEL SPACE VISUALISATION
Figure 4.11: Individual parameter plots (data nodes only) showing the magnitude of each separate variable
at each data node in the SOM.
shows increasing values from the bottom right corner towards the top left (best corner). Similar
regular patterns occur for other parameters. When compared to the labelled SOM of Figure 4.9b,
these parameter plots show the component values associated with each solution obtained during
the inversion. There is a clear correlation, for example, between the best model images and a
high viscosity. However, because of the averaging effect of associating multiple input vectors
with one SOM node, some details are lost. The projection in Figure 4.5a shows that badly ranked
simulations exist with high viscosities, but this information is missing in the SOM plot forηu
(Figure 4.11). Such inconsistencies between visualisation methods illustrate the benefit of using a
diversity of approaches.
Figure 4.12 contains a combined representation of the individual parameter plots of Figure 4.11.
The height of each bar indicates the normalised magnitude of each component of the SOM vector.
The order of the bars at each node is the order of the dimensions in Figure 4.11. It is easy to
understand the distance between SOM data nodes (c.f. Figure 4.9a) by comparing bar charts. The
simulation ranked 5.5 in Figure 4.10, which shares characteristics of the best solutions, is the result
of a set of very different parameter values than those of the best solutions, confirming the SOM
representation of this point, and the non-uniqueness of the problem. The worst solutions (bottom)
are not dictated by any one component, for example, a high cohesionco, by itself, does not imply
a bad solution. An unrealistically high yield stress, through a combination of high cohesion and
pressure dependence, is responsible for the simulations that fail to converge numerically (labelled
“10” in Figure 4.10). Either of Figures 4.11 or 4.12 allows a quick scan of the parameter controls
on model outputs.
As specified earlier, different initialisations for the SOM will result in different final maps.
Figure 4.13 compares the SOM used above, with five variations, each labelled according to the
absolute ranking of solutions. In general, a greater number of isolated large-distance nodes (green
to red) reveals poor clustering, or a non-optimal SOM. The original map (Figure 4.13a) represents
all the better solutions, in the top half, with acceptable clustering, but the bottom half of the map
is not optimal. Figure 4.13e is the worst representation of the data, and Figure 4.13f is the best.
Labels are not consistent across all maps because, where more than one point occupies the same
SOM node, the label belongs to the last simulation added to that node. The solution ranked 5.5,
51
CHAPTER 4. MODEL SPACE VISUALISATION Wijns Ph.D. Thesis
Figure 4.12: Visualisation of the normalised SOM vectors at each data node. The order of the bars refers to
ηu, co, cp, σc, a, andεo. The arrow indicates the best model output.
when labelled, is consistently far from the best corner, again reinforcing the non-uniqueness that
was signalled in earlier plots.
4.6 Discussion
Geoscientific problems, and geodynamic ones in particular, are among the most difficult inverse
problems. Their full solution is beyond not only current computational power, but also existing
mathematical tools. The combination of interactive inversion and visualisation is not a panacea,
but offers a number of advantages, some quite obvious, some less so.
The SOM is a successful method for displaying high-dimensional data, and provides a way to
deal with the many solutions that are generated by an inversion process. It produces relatively
simple plots that are useful for determining which parameters most affect the process under study.
Related to this is the determination of possible mechanical behaviour, given a certain modelling
scenario. The drawback of a SOM is that it cannot represent data that are nonexistent, meaning
that unsampled space is unrepresented. A parallel axis display or multi-dimensional projection
incorporates the full extent of the space under investigation, so data gaps can be easily detected.
The SOM is a clustering technique that is used to best advantage when a user is confident of
adequate sampling over the parameter space.
Keeping in mind the weaknesses in the coverage of parameter space that are highlighted by
the parallel axis display and multi-dimensional projections, all plots in combination could be sum-
marised as follows, converting to natural parameter values: the best matches to the target image
are achieved with a high viscosityηu > 5×1022 Pa·s, and a low cohesionco < 5 MPa. The low
52
Wijns Ph.D. Thesis CHAPTER 4. MODEL SPACE VISUALISATION
Figure 4.13: (a) Original SOM from Figure 4.9a. (b) – (f) Five variations due to different random initial-
isations. Labels indicate simulation ranks. Labels are not consistent across all maps because, where more
than one point occupies the same SOM node, the label belongs to the last simulation added to that node.
friction coefficientcp (0.2 – 0.3), significant strain weakeninga (0.6 – 0.7), and fairly rapid rate of
weakening toεo (0.3 – 0.4) imply overall weak faults, as has been suggested for major structures
(Lachenbruch and Sass, 1980, 1992; Bird and Kong, 1994; Bird, 1995). Physically, this would
promote large fault spacing by ensuring that, once a fault is initiated, extensional stresses to either
side quickly drop to below yield values, so that new faults do not form nearby. The tensile limit
σc (2 – 6 MPa) may vary somewhat in the lower part of its range, again indicating weak rocks,
although this should only have an effect near the surface where tensile stresses are more likely.
Such a summary emerges quite naturally from the three sets of visualisation plots.
The approximate understanding of the search space can be used to further interact with the
inversion itself. One obvious way is to remove parameters once they have been “optimised”, or
parameters that are shown to have little impact on the results. This reduces the dimensionality
of the problem. Another approach is to focus the search into smaller domains, where the “good”
simulations lie. This can result in faster convergence as well as the opportunity to increase the
search resolution of certain parameters. A user with a good knowledge of the inversion algorithm
(a GA in this case) can also use the information of the current state of convergence in order to tune
certain parameters controlling the inversion itself. The opportunity to do this during the course
of an inversion can considerably improve performance. In the case of a GA, this may involve
changing the cross-over or mutation operator, or changing the number of individuals.
This inversion and visualisation system can be a vehicle to facilitate the transfer of technology
to applied modellers. The user-friendliness inherent in the approach (which removes the user from
53
CHAPTER 4. MODEL SPACE VISUALISATION Wijns Ph.D. Thesis
the underlying mathematics and allows him orher to concentrate on the geological aspects of the
problem) has been very well received by industry representatives. It can be a crucial factor in the
acceptance of geological modelling as, for example, a mineral or oil exploration tool.
4.7 Chapter Summary
During any inverse modelling, it is in the interest of the modeller to capitalise on the availability
of accumulated solutions in order to understand the dynamics and non-uniqueness of the problem
as fully as possible. This is easy to do within the interactive inversion system, in contrast to most
inverse procedures that return a single best solution. Combined with the systematic ranking of
trial solutions, visualisation methods such as parallel axis displays, multi-dimensional projections,
and self-organised maps represent the high-dimensional parameter space in a clear and simple
2D visualisation environment. In the first instance, methods that represent the entire parameter
space, such as the parallel axis display and multi-dimensional projections, provide an overview of
the sampling by the inversion, and expose any gaps in coverage. The graphical summaries also
offer a platform from which to draw conclusions regarding the controlling physical factors and the
connections between them.
54
Chapter 5
MODES OF CRUSTAL EXTENSION
5.1 Introduction
Chapters 3 and 4 give some preliminary insights into rheological controls on fault spacing
during extension of the crust. This chapter builds on those results, using more sophisticated and
realistic representations of crustal strength profiles.
Continental lithosphere may be highly extended without entirely rifting to a new ocean basin.
For example, total Cenozoic strain estimates of up to 100% (β = 2) have been proposed for parts
of the Basin and Range in the western U.S.A., which have not been rifted (e.g., Lachenbruch
and Sass, 1978; Jones et al., 1992). A total of up to 80% extension is suggested on a very large
scale of 300–500 km (Wernicke et al., 1982), and Niemi et al. (1999) have even suggested 500%
extension in the Death Valley region, California, since the middle Miocene. Stretching of the
crust may then be accommodated by two contrasting phenomena: distributed, closely spaced, and
limited-slip normal faulting over a large area, or localised, large-strain normal faulting that can
result in the complete dissection of the upper crust and exhumation of the lower crust. Examples
of distributed faulting exist in various parts of the North Sea basin (e.g., Fossen and Rœrnes, 1996;
Viejo et al., 2002), where the faults are characterised by a relatively steep angle and small offset.
Normal faults that exhibit a low angle and very large displacement, juxtaposing exhumed high-
grade metamorphic rocks against near-surface rocks, are the hallmarks of a metamorphic core
complex (MCC), as in the Basin and Range (e.g., Axen et al., 1990; Duebendorfer et al., 1990;
Hodges et al., 1990; Spencer and Reynolds, 1991) or the Aegean (e.g., Gauthier and Brun, 1994;
Forster and Lister, 1999; Burchfiel et al., 2000; Gessner et al., 2001).
Why should extending lithosphere in places form a MCC rather than fail in a distributed man-
ner? One possibility is that major lateral discontinuities in the strength of either the upper crust
(e.g., pre-existing faults) or the lower crust (e.g., partial melt zones) focus stresses and localise
extension. Christiansen and Pollard (1997) document field observations where shear zones have
nucleated from pre-exisiting dikes that act as weaknesses, although these shear zones are not as-
sociated with a MCC. In their conceptual models of extensional tectonics, both Wernicke (1985)
and Wernicke and Axen (1988) assume an immediately active horizontal or low-angle detachment
surface through the upper crust, which then controls faulting and exhumation of the lower crust. In
some analogue modelling experiments, initial faults end up controlling fault spacing and the mode
of extension (e.g., Koyi and Skelton, 2001). Brun et al. (1994) conclude that a weak rheological
zone at the base of the upper crust, such as a magma body, is required to trigger a MCC. Other ana-
logue modellers (e.g., Mulugeta and Ghebreab, 2001; Corti et al., 2001) employ a pre-weakened
55
CHAPTER 5. MODES OF CRUSTAL EXTENSION Wijns Ph.D. Thesis
Figure 5.1: (a) Initial geometry and boundary conditions for the numerical model. Stripes in the upper
and lower crust are marker materials for visualising deformation. (b) Representative maximum shear stress
profile (solid line) through the crust for a given strain rateε. Neglecting localised strain weakening, strength
increases with pressurep from co at the surface to a maximum value at the basezu of the brittle upper
crust, at which point the yield curve intersects the viscous, temperature-dependent flow law. To avoid any
extremely low viscosity, the value at some depthzc is adopted as a minimum value and is constant to the
basezl of the crust.
lower lithosphere to localise stresses and influence the resulting extension, and Corti et al. (2001)
add an imposed basal velocity discontinuity to this. Some numerical modelling experiments also
assume initial weaknesses, and proceed to investigate related factors that vary the results of ex-
tension (e.g., Dunbar and Sawyer, 1989). Although such weak features do serve to focus stresses,
in the absence of lateral heterogeneities, vertical contrasts in rheology dictate whether one or the
other mode of extension will result. In fact, small lateral heterogeneities are often insufficient to
trigger a MCC mode when the rheological layering promotes distributed faulting.
The mode of extension is thus determined by the combined ability of both lateral and vertical
rheological structure to focus stresses in the brittle upper crust. The numerical modelling in this
chapter concentrates on the role of vertical rheological contrast in dictating the spacing between
fault zones and the accompanying mode of distributed faulting or MCC formation, building on
the results of the inversion in Chapter 3. The 2D vertical sections are readily compared to much
published analogue modelling, including that of Brun et al. (1994).
5.2 Model
As in previous models, the upper crust behaves viscoplastically, which gives rise to the local-
isation of deformation and necking of layers. However, the transition from brittle behaviour to
ductile flow in the lower crust is now governed by temperature, which creates the possibility of
ductile material cooling and entering the brittle (plastic) field. The initial geometry of the model
(Figure 5.1) corresponds to 20 km of upper crust above 40 km of lower crust, along a length of
160 km. Horizontal and vertical stripes are simply marker materials to enhance visualisation (as in
Brun et al., 1994). Above these two crustal layers is the “air” layer that does not influence the me-
chanics of the problem. All walls of the closed bounding box are free-slip, and vertical walls have
zero heat flux. Extension proceeds by applying a uniform velocity to the right-hand boundary,
equivalent to 100% strain in 5 Ma (Gessner et al., 2001), or 6.3×10−15/s. This boundary velocity
56
Wijns Ph.D. Thesis CHAPTER 5. MODES OF CRUSTAL EXTENSION
Table 5.1: Natural values for extension model parameters.
Parameter Value
depthzu, upper crust 20 km
depthzl, lower crust 60 km
velocityU , boundary 3.1 cm/yr
strain rateε, initial 6.3×10−15/s
densityρu, upper crust 2700 kg/m3
densityρl, lower crust 3000 kg/m3
gravityg 10 m/s2
cohesionco 16 MPa
friction coefficientcp 0.44
strain weakeninga 0.8 (0.2, simulation D)
“saturation” strainεo 0.5 (1.0, simulation D)
thermal diffusivityκ 10−6 m2/s
temperatureTs, surface 0oC
is low enough to not create tensional stresses, so that gravity, in effect, drives the deformation.
The yield law (2.8) for brittle crust uses a rock cohesionco = 16 MPa and a pressure depen-
dencecp = 0.44, similar to the model in Chapter 3 The strain weakening parameters,a andεo
(equation 2.9), vary in order to investigate the influence of fault weakening on the mode of ex-
tension, but the exponent is fixed atn = 1, based on the previous conclusion that it is the least
important parameter for controlling the outcome (Chapter 3).
The boundary between upper and lower crust is defined at all times by a fixed density contrast.
Initially, this corresponds to the depth at which the temperature dictates a change from brittle to
ductile behaviour. With increasing extension, the initially Newtonian lower crust may be exhumed
into the brittle field and undergo faulting. It behaves mechanically as the upper crust does, but
retains its original density (i.e., metamorphic grade and mineralogy). There is no thermal expan-
sion, which provides a minor buoyancy effect compared to compositional density contrasts; nor is
melting modelled.
Viscosity η varies with temperatureT according to the Frank-Kamenetskii relation (Frank-
Kamenetskii, 1969, Chapter 6, p. 340–41)
η(T ) = ηo e−cT . (5.1)
This is a simplification of an Arrhenius rheology, where the constantsηo andc are chosen such
that the viscosity at the interfacezu between upper and lower crust, which is the initial brittle to
ductile transition, satisfies the maximum yield stress
τyield(zu) = η (T (zu)) ε ,
that is, the maximum shear stress profile through the crust is continuous. Appendix C.1 contains
a full development of this relation. To avoid an extremely low viscosity at the base of the crust,
57
CHAPTER 5. MODES OF CRUSTAL EXTENSION Wijns Ph.D. Thesis
η(T ) > η(Tc) for some maximum temperatureTc. The constants used in equation (5.1) then result
in a maximum viscosity variation of two orders of magnitude from the top of the lower crust to the
point of minimum viscosity. In a constant velocity extension model, the crustal strength profile
(Figure 5.1b) evolves with a changing strain rate, but provides a convenient starting point from
which to characterise the crust.
The surface of the upper crust is maintained at 0C. Since the temperature scale can be chosen
independently of other variables, it is fixed only for comparison with field data in a later section.
For reference, an initial temperature gradient of 17.5C/km results in a brittle to ductile transition
at 350 C, in the range discussed by Brace and Kohlstedt (1980) and references therein, and
McKenzie and Fairhead (1997). This transition temperature also corresponds to theoretical flow
laws derived by Handy et al. (1999) for various crustal components.
Table 5.1 contains parameter values for the natural system, which apply to all simulations
unless explicitly stated otherwise. Values for the viscosity law (5.1) are not listed because they
change with every simulation, but the discussion section offers some natural viscosity equivalents.
5.3 Results
As in previous simulations, bands of localised plastic strain are proxies for fault zones in the
continuum code. Unless otherwise noted, a maximum of 80% strain weakening (a = 0.8) occurs
after an accumulated plastic strainεo = 0.5. These values reflect evidence, from both numerical
experiments (e.g., Bird and Kong, 1994) and field-based heat-flow measurements (e.g., Lachen-
bruch and Sass, 1980, 1992), that major faults may undergo significant weakening.
The different behaviours of the crust under extension are partly parameterised by the ratiorτ
of integrated upper to lower crustal strength. Appendix C.2 contains the calculation of integrated
strengths, based on the maximum sustainable shear stress at a given strain rate.
5.3.1 Constant Viscosity
The two contrasting modes of crustal extension are first illustrated for a constant viscosity
lower crust (Figure 5.2), for comparison with the results of Brun et al. (1994). When the system
is characterised by a smallrτ = 0.53 (i.e., relatively high-viscosity lower crust), the result is
distributed faulting (simulation A, Figure 5.2i). The upper crust develops many closely-spaced
steep faults, each of which accomodates limited strain, and the interface between upper and lower
crust remains relatively flat. This result is very similar to the tilted block mode of Brun et al.
(1994), although, according to the information supplied,rτ ≈ 1.8 for their analogue model. This
discrepancy is probably due, in large part, to differences in material properties for the experiments,
and is addressed below in the discussion. Even with more than 80% extension (β > 1.8), the upper
crust in simulation A is never completely dissected. New steep faults form soon after older faults
have accommodated a small amount of shear strain, breaking the upper crust into a series of small
blocks that undergo limited rotation.
Although Brun et al. (1994) propose that a weakness is needed in the lower crust in order to
58
Wijns Ph.D. Thesis CHAPTER 5. MODES OF CRUSTAL EXTENSION
i ii
Figure 5.2: Simulations with constant viscosity lower crust. (i) Simulation A: evolution of distributed
faulting mode withrτ = 0.53. (ii) Simulation B: evolution of MCC mode caused by a uniformly weak
lower crust andrτ = 2.52. Total extension of (a) 25% (1.2 Ma), (b) 53% (2.7 Ma), and (c) 83% (4.2 Ma).
localise stresses and trigger a MCC mode of extension, simulation B (Figure 5.2ii) shows that
a uniformly weak lower crust (highrτ = 2.52) is sufficient to achieve this behaviour. Defor-
mation is accommodated by only a few normal fault zones. Displacement is large within each
fault zone, which, through fault weakening, remains active even after rotating to a shallow dip.
The lower crust flows easily to isostatically compensate for localised thinning of the upper crust,
which enhances block rotation and continued strain on low-angle faults. Despite the Newtonian
rheology of the lower crust, strain is localised as a result of kinematic interaction with the plastic
(non-Newtonian) upper crust. The lower crust is first exhumed atβ = 1.6.
5.3.2 Temperature-Dependent Viscosity
The main limitation in using a constant viscosity for the lower crust, and therefore a limitation
of the analogue modelling, is that lower crustal material that rises towards the surface remains
very weak, although it should cool and strengthen, even becoming brittle. This phenomenon does
not affect the case of distributed faulting, where the lower crust remains buried below a laterally
stable brittle to ductile transistion (Figure 5.2i). In MCC mode, however, exhumation and cooling
will influence the tectonics by changing the rheological behaviour of lower crust material.
A weak lower crust with a temperature-dependent rheology, in a laterally uniform system,
gives rise to the MCC mode in Figure 5.3. The rising lower crust enters the brittle domain, but
cooling is slow enough, relative to exhumation, that brittle deformation of the lower crust does not
penetrate to great depth. The propagation of faults from the upper crust into the exhuming lower
crust allows continued strain along the original structures, until they become very gently dipping
or even flat-lying detachment surfaces accomodating large displacements. There is evidence in the
vertical lower crust markers that these faults are spatially connected to diffuse, high-shear zones
59
CHAPTER 5. MODES OF CRUSTAL EXTENSION Wijns Ph.D. Thesis
Figure 5.3: Simulation C: evolution of MCC mode (rτ = 2.15) with both a viscosity and a brittle to ductile
transition that are temperature dependent. Total extension of (a) 16% (0.8 Ma), (b) 53% (2.6 Ma), and (c)
82% (4.1 Ma). Faults continue to propagate through the newly brittle lower crust as it nears the surface, and
are connected to diffuse, high-shear zones in the ductile region, indicated with pairs of solid lines in (c).
Figure 5.4: Simulation D: MCC mode withrτ = 2.15, but a maximum fault weakening of 20%, rather than
80% as in simulation C. Total extension of 80% (4.0 Ma).
that extend to the base of the entire crust. These deep shear zones are traced in Figure 5.3c, and
show much greater localisation than in the case of constant viscosity (Figure 5.2ii). Such brittle
fault to ductile shear continuity has been suggested in an intraplate setting by Zoback et al. (1985).
Despite a decrease inrτ relative to the constant viscosity MCC of simulation B, fewer fault
zones are generated in simulation C. The greater tendency to MCC mode is due to the variable
viscosity profile that, although higher in an integrated sense, allows much lower viscosity material
near the base of the model to dominate crustal flow and tectonic expression. This is apparent when
comparing the strain visualisation markers in the lower crust (Figures 5.2(i)c and 5.3c).
In order to determine the importance of fault weakness with respect to MCC formation, simu-
lation D (Figure 5.4) has a maximum strain weakening of only 20% (a = 0.2) instead of 80% as in
simulation C, and this weakening accumulates more slowly throughεo = 1.0 instead of 0.5. The
resulting extension still produces a MCC mode, although the characteristics are different. Fault
zones are more numerous and more diffuse, and the upper crust tends to neck rather than produce
planar shear zones. Although these stronger fault zones initially evolve as for the case of simula-
tion C, they are less effective in accomodating continued strain, so that higher inter-fault stresses
60
Wijns Ph.D. Thesis CHAPTER 5. MODES OF CRUSTAL EXTENSION
Figure 5.5: (a) Temperature-time paths for an exhumed footwall of a MCC from fission-track ther-
mochronology and Ar40/ Ar39 ages (from Gessner et al., 2001) and from numerical simulation C. The
numerical data are scaled to match the 300C initial field temperature. (b) Schematic cross section of
Central Menderes MCC from Gessner et al. (2001). The asterisk denotes the location, in high-grade meta-
morphic material, of the laboratory analyses. (c) Part of simulation C with an asterisk showing the final
location of the numerical temperature sampling point.
develop, causing new faults to form.
5.3.3 Field Validation
A test of the physical validity of MCC simulation C is the comparison of numerical results
with cooling data from the Kuzey detachment in the Central Menderes. After 100% extension,
the crust is approximately 30 km thick, in agreement with a seismic interpretation of western
Turkey by Saunders et al. (1998). Both field and numerical temperature-time curves for exhumed
footwall material are similar (Figure 5.5a). The field data, from Hetzel et al. (1995) and Gessner
et al. (2001), reflect measurements from apatite and zircon fission-track thermochronology, and40Ar/ 39Ar ages for the higher temperatures. Temperature in the numerical model is scaled to
match the initial 300C of the exhumed field sample. The numerical curve does not reach the
surface temperature of 0C because, due to the continuum nature of the code, the upper crust is
never completely removed from above the exhumed lower crust, so that the lower crust remains
marginally buried and begins to stabilise at approximately 75C.
The match between cooling rates for the numerical and field data suggests that the model for-
mulation provides an adequate physical description for the investigation of extensional tectonics.
Parsons et al. (2001) provide observations that could further argue the case that weak lower crust
61
CHAPTER 5. MODES OF CRUSTAL EXTENSION Wijns Ph.D. Thesis
Figure 5.6: (a) Endmember of MCC mode (rτ → ∞) represented by the brittle upper crust over a free-
slip lower boundary. Total extension of 3% (0.15 Ma). The profile above is the normalised basal velocity.
(b) Same simulation as (a) at a total extension of 38% (1.9 Ma). Because of the continuum nature of the
code, there is some stress transmitted across the graben. At this stage, gravitational slumping also begins to
affect the simulation because there is no lower crust isostatic compensation. (c) Endmember of distributed
faulting mode (rτ → 0) represented by the brittle upper crust over a no-slip boundary. Total extension of
40% (2.0 Ma).
promotes MCC formation. A low seismic-velocity middle crust (hotter and/or weaker) is present
under the Buckskin-Rawhide core complex in Arizona, U.S.A. Further west, the middle crust
shows high velocities (colder and/or stronger) under the steep, normal faults of the Salton Trough.
5.4 Discussion
5.4.1 Continuum Between Modes
Pre-existing faults or thermal and rheological heterogeneities are not required to produce a
MCC mode of extension. A large ratio of upper to lower crustal strength is necessary, often
overcoming distributed initial faults. Increasing the ratiorτ eventually leads to a state where a
single fault is nucleated and remains the only zone of failure. This is illustrated by the endmember
caserτ → ∞ of a single brittle layer extended over a free-slip lower boundary (Figure 5.6a and
b), which is the limit of very weak lower crust that effectively decouples the upper crust from the
tectonics below. Once a fault or symmetric graben is nucleated, boundary stresses can no longer
be communicated laterally through the layer, and the velocity gradient approaches zero away from
the fault, as apparent in the basal velocity profile plotted above the section. If the fault was truly
a discontinuity (impossible with theEllipsis continuum code), the velocity gradient would be
62
Wijns Ph.D. Thesis CHAPTER 5. MODES OF CRUSTAL EXTENSION
2
3
76
3
6
6
88 8 9
3
4 4
5 6
77
0 1 2 3 4 5strength ratio rτ
0
2
4
6
8
10
12
num
ber
of fa
ults
alo
ng 1
60 k
m
η = cstr
h = 0.5
rh
= 0.4
rh
= 1.0
rρ = 1.0
wf = 0.1
Figure 5.7: Number of major fault zones as a function ofrτ . Data labels indicate a subjective evaluation
of the model result between the distributed endmember (0) and the MCC endmember (10). Unless other-
wise indicated, the standard simulations (filled circles) have a temperature-dependent lower crust viscosity,
thickness ratiorh = 0.5, density ratiorρ = 0.9, and fault weakness factorwf = 0.53. Whilerτ controls
the overall trend, changes in the other variables, especiallyrh, move results vertically for equalrτ , and
horizontally for equal mode or number of faults.
zero everywhere except at the discontinuity, unlike the profile in Figure 5.6b, where some stress
continues to be transmitted. The lower crust, where present, fulfills the role of stress distributor
and ensures non-zero traction at the base of the brittle crust. For the hypothetical case of a lower
crust that is so strong as to maintain a constant stress at this interface,rτ → 0 and a zero-slip basal
condition exists (Figure 5.6c). The resulting uniform yield of the upper crust is the endmember of
distributed faulting.
Therτ → ∞ case in Figure 5.6b will not necessarily result in MCC formation. The presence
of a volume of mobile lower crust may be essential for block rotation, shallowing of fault dips, and
lower crust exhumation. Fault zone localisation and significant block rotation are complementary
expressions of the weak lower-crust endmember. These two phenomena might not be separable in
nature.
Simulations A to C, together with the endmember cases in Figure 5.6, suggest that the number
of major fault zones that develop to accomodate extension correlates inversely with the tendency
towards MCC mode. This is borne out in Figure 5.7, in which the number of faults and mode of
extension are plotted against the strength ratio of the crustrτ . Faults are counted if they transect
the upper crust, and a graben is counted as one fault zone. To illustrate, in Figure 5.6b, ther
is only one fault zone. The trend is that fewer fault zones develop with increasingrτ . More
simulations have been run than are shown in the previous section, in order to explore the phase
space of extension modes. Results are labelled with a subjective rank from 0 to 10, corresponding
to position between the endmembers of, respectively, distributed faulting and MCC mode. In
63
CHAPTER 5. MODES OF CRUSTAL EXTENSION Wijns Ph.D. Thesis
34
3
6
6
8
4
8
5 6
8 9
0 1 2 3 4 5 6 7 8 9rτ / rh
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1.8
2.0
(∆f
β /z
l )
Figure 5.8: Spacing∆f of major fault zones (normalised by final crustal thicknesszl/β) as a function of
rτ/rh for temperature-dependent simulations withrρ = 0.9 andwf = 0.53. Data labels indicate mode as
in Figure 5.7. The trend should pass through the origin (c.f. Figure 5.6c forrτ/rh → 0) and towards infinite
spacing forrτ/rh → ∞, although these cases would not be reached for natural parameter ranges for the
crust.
addition torτ , the following parameters simplify the classification of results: the ratiorh of upper
to lower crust thickness, the ratiorρ of upper to lower crust density, and the fault weakness factor
wf = a/(1 + εo). A largerwf implies greater and/or more rapid fault weakening. According to
Figure 5.7, variations inrτ exert the greatest control on the mode of extension. Of secondary but
significant importance isrh, while rρ andwf have a minor influence. The extension rate in these
simulations may be too high for density contrasts to appreciably affect exhumation of the lower
crust. Buoyancy forces will be more evident for lower extension rates, at some point dictating a
transition to diapirism forrρ > 1.
Numerical work by Bai and Pollard (2000) may explain why strength contrasts control the
mode of extension. They relate fracture spacing in layered rocks to the stress distribution between
fractures, which depends on the physical properties of neighbouring layers. The observations on
the numerical simulations in this chapter are that:
1. at the onset of extension, many faults are nucleated in both modes;
2. whenrτ is small, many of the nascent faults zones are activated;
3. whenrτ is large, few fault zones are activated.
For a fractured layer sandwiched between two unfractured layers, Bai and Pollard (2000) find
that there is a critical fracture spacing to layer thickness below which the interfracture stress is
everywhere compressive and new fractures will not form. At greater fracture spacings, the stress
becomes tensile, and eventually overcomes the tensile strength of the layer, leading to new in-
64
Wijns Ph.D. Thesis CHAPTER 5. MODES OF CRUSTAL EXTENSION
fill fractures. This critical spacing to thickness ratio increases with increasing ratio of Young’s
modulus between the fractured and neighbouring layers. Although the Bai and Pollard (2000)
analysis is for linear elastic media, there is an exact analogue in the results for normal faulting of
viscoplastic layers. As the strength ratiorτ increases, the spacing between active faults increases.
Thus, the magnitude of lateral stress transfer from the unfaulted lower crust to the upper crust is
key to the mode of extension, where the mode is specifically controlled by the number of active
fault zones. A relatively strong lower crust provides greater traction at the base of the upper crust,
which results in the yield stress being reached at shorter spacings between faults. The existence
of few fault zones in the presence of a weak lower crust is simultaneously linked to the ability of
crustal blocks to rotate because of lower crustal mobility, leading to MCC formation.
The effect of crustal thickness ratios is more straightforward. For equal values of the integrated
upper crustal strength, a thicker upper crust (largerrh) simply requires more strain until it is
dissected. A thin lower crust (largerrh) relies largely on lateral flow to compensate for unloading,
and thus, at the same integrated viscosity as a thick lower crust, does not accomodate a MCC mode
of extension as easily.
When the different thickness ratiosrh are accounted for, the standard deviation of model be-
haviour is largely reduced. Figure 5.8 shows the more or less direct variation between normalised
fault spacing andrτ/rh for the temperature-dependent simulations withrρ = 0.9 andwf = 0.53.
The greatest fault spacing achieved with the model setup is about 1.6 times the total crustal thick-
ness. The deviation from the trend for the two simulations with highestrτ/rh is most likely due
to boundary conditions, that is, the box is finite and there will always be at least one fault dur-
ing the experiment. The simplistic empirical prediction of fault spacing and extension mode in
Figure 5.8 would be difficult to relate to field observations in a quantitative sense, for example,
through measurements of crustal thickness orβ factor. Other physical formulations for the model
would likely change the absolute values of the defining parameters, but there is a continuum of
behaviour in extension that depends mostly upon the contrast between upper and lower crustal
strength, modulated by crustal thickness ratios. Furthermore, after significant extension (β > 1.8),
there are only two endmember modes. A smallrτ results in a contiguous upper crust even after
extreme stretching, whereas a largerτ initiates more localised strain that evolves into a MCC.
5.4.2 Comparison with Other Modelling and with Nature
As suggested above, the value ofrτ = 1.8 for the distributed faulting mode of Brun et al. (1994)
should lead to an intermediate mode of extension, according to the numerical results above. The
thicker lower crust (rh = 0.33) in Brun et al. (1994) would push the model even further into
the MCC field (c.f. Figure 5.8). A sand upper crust (not viscoplastic), less extension, periodic
extension due to the experimental set-up, and insufficient basal lubrication may all be factors that
hinder meaningful comparison between these numerical and analogue results.
The crustal model in this chapter does not incorporate upper mantle, which precludes direct
comparison with the modes of extension of Buck (1991), for example. However, an equivalent
“narrow rift” mode in the simulations in this chapter is simply a juvenile MCC mode in terms of
65
CHAPTER 5. MODES OF CRUSTAL EXTENSION Wijns Ph.D. Thesis
total extension. Strain is localised in one or a few zones, but the lower crust is not yet exhumed.
Analogue modelling by Benes and Davy (1996), in which they distinguish model behaviour using
a similar measure of layer strength ratios as here, purports to uphold the modes of Buck (1991).
They conclude that very weak lower crust gives rise to a MCC mode of extension, and results
above are broadly consistent with theirs. While Benes and Davy (1996) also classify a narrow
rift mode, they change the upper to lower crust thickness ratio with respect to their core complex
models to achieve this. A narrow rift, which they classify after only 9.3% extension, is the result of
a thicker upper crust and a thinner lower crust, which both contribute to hindering exhumation of
the lower crust. Given greater extension, the narrow rifts of Benes and Davy (1996) are probably
core complexes, whereas the “wide rifts” will remain as such, equivalent to the distributed faulting
models in this chapter. In this sense, narrow rifts are an intermediate stage and not an endmember
behaviour.
The upper mantle will have some influence on the behaviour of the crust during extension.
The thinner or stronger the lower crust, the more likely this influence will manifest itself. The
lower crust is sufficiently thick and ductile in the model above that it effectively decouples the
crust from the upper mantle. If the upper mantle is strong, it will then remain relatively flat, as
the lower crust will flow more quickly than the mantle in response to any unloading. This case is
implicit in the use of a rigid, zero-traction bottom boundary, and corresponds to the observations
of a flat Moho boundary under areas of high extension in Arizona (Hauser et al., 1987). On the
other hand, if the upper mantle is of comparable viscosity to the lower crust, it will participate in
isostatic compensation. In the present model, this would be equivalent to increasing the thickness
of the lower crust.
The geometric origin of low-angle detachment faults remains controversial. According to Scott
and Lister (1992) and Livaccari et al. (1995), for example, structural field relations favour initiation
and activation of some faults at a shallow dip. Others, such as Buck (1988) and Wernicke and
Axen (1988), propose initial high-angle faults that subsequently rotate to shallow dip. In all the
simulations above, initially high-angle normal faults form as planar features that evolve to lower
angles because of lower crustal flow, or isostatic compensation, during unloading of the footwall.
The faults are not listric but initially dip at approximately 45 due to the nature of the implemented
rheology. Listric faults should enhance block rotation, so that the trend from distributed faulting to
MCC mode will be moved towards lower values ofrτ in systems of listric faults. The rotation of
high-angle faults first manifests itself at the base of the upper crust, where the changes in the local
stress field are greatest. Continued unroofing and lower crustal flow can eventually produce very
gently-dipping faults such as those of Figure 5.3c. These results do not contradict the hypothesis
that detachment faults may form at low angles. There is simply no condition present for an initial
stress field rotated from the vertical, and new faults form at a specific angle toσ1. However, the
results do illustrate and support the hypothesis that initial high-angle faults can rotate to low angle
and develop into flat-lying detachment faults that continue to slip due to significant weakening.
It is worth noting a Coulomb failure analysis by Wills and Buck (1997) that determines the
likelihood of fault slip at shallow dips due to specific boundary or loading conditions. They con-
66
Wijns Ph.D. Thesis CHAPTER 5. MODES OF CRUSTAL EXTENSION
clude that slip will not occur on low-angle surfaces unless there is an unlikely combination of
localised, near-lithostatic pore pressure, with unsustainably high tensile stresses in the upper 5 km
of the crust. On the other hand, if normal faults form at shallow dips because of pre-existing weak-
nesses, for example, earlier, gently-dipping thrust faults (e.g., Horvath, 1993), such structures can
alter the stress field and also override the control of vertical rheological strength contrasts on fault
spacing.
Through numerical experiments on a single elasto-plastic layer, Lavier et al. (2000) find that
both the layer thickness and the amount of fault weakening can affect the degree of offset per
fault. However, because the Lavier et al. (2000) experiments are over an inviscid substrate, they
nucleate a single fault zone only (this includes a single graben or horst with secondary faults in
the same zone), that is, the endmember of MCC mode (Figure 5.6a). They acheive fault rotation
by allowing influx of filler material from the base. By including the important role of the lower
crust in distributing stresses, results in this chapter determine fault weakening to be less important
than the rheological contrast between upper and lower layers of the crust.
A more sophisticated model by Lavier and Buck (2002) includes a visco-elastic lower crust.
While their purpose is to investigate the influence of cooling rates on faulting style, their results fit
within the description of modes given above. Estimated values ofrτ for the simulations presented
by Lavier and Buck (2002) are all very high, such that the experiments are well within a MCC
mode and consequently nucleate single fault zones, albeit sometimes composed of more than one
individual fault.
Translation ofrτ values into natural viscosity for the lower crust allows a comparison with
independent estimates. It is difficult to express the integrated viscosity as a single value, since the
profile varies considerably with depth. Observational estimates based on, for example, post-glacial
rebound or post-seismic relaxation of crustal velocities are probably reflective of the most mobile
part of the lower crust. The minimum viscosity of the lower crust (belowzc) is therefore used
for crude comparison. The distributed faulting mode withrτ = 0.53 has a lower crust viscosity
of 1022 Pa·s. The MCC mode in simulation C returns a viscosity of 4.5×1020 Pa·s. An estimate
of lower crustal viscosity is lacking for western Turkey, but, in the western U.S.A., Pollitz et al.
(2001) calculate an upper limit of 5×1020 Pa·s based on geodetic measurements of post-seismic
velocity fields after the 1999 Hector Mine earthquake in California. This value is very close to
the model value for the MCC mode, but the distributed mode viscosity appears very high. For
a better comparison with the western U.S.A., the initial lower crust thickness is scaled to 20 km
(rh = 1), leaving a total crustal thickness of between 20 and 25 km after extension, consistent
with the regional crustal thickness model of Chulick and Mooney (2002) for the Basin and Range.
Now, the distributed faulting mode occurs for a viscosity of 6×1021 Pa·s, and the MCC mode for
a viscosity of 2.5×1020 Pa·s or lower.
Byerlee’s law (Byerlee, 1968), the result of numerous laboratory experiments, postulates a
profile of brittle yield strength versus depth that is relatively independant of rock type. Pore
pressure also exerts a strong influence on rock strength, but departures from hydrostatic level are
usually not sustained over the length and time scales of the modelling above. The thickness of the
67
CHAPTER 5. MODES OF CRUSTAL EXTENSION Wijns Ph.D. Thesis
upper crust, or, equivalently, the depth to the brittle to ductile transition, will therefore determine
the integrated strength of the brittle layer. The temperature gradient will be the first-order control
on this thickness, and it will also govern the viscosity of the lower crust, so that the mode of
extension can reflect the thermal regime at the time of deformation. The mineral composition
(including grain size) and water content of the lower crust are similarly responsible for its viscosity
(e.g., Wilks and Carter, 1990; Ross and Wilks, 1995; Rybacki and Dresen, 2004).
5.5 Chapter Summary
Numerical modelling shows that the mechanical stratification of the crust provides the funda-
mental control on fault spacing, and, ultimately, the mode of extension. Pre-existing structures
and weaknesses are often thought to govern the behaviour of continental crust under extension,
and many prior studies have focussed on the effect of heterogeneities in triggering faulting and ex-
humation of lower crustal material. The role played by such features is, in fact, subordinate to that
exerted by the rheological contrast from upper to lower crust, quantified by the ratio of integrated
upper to lower crustal strength. This ratio dictates fault spacing in the upper crust through the
stress transfer from ductile lower crust to brittle upper crust. The fault spacing, naturally linked to
the ability of the lower crust to flow, controls the subsequent evolution of the normal fault systems.
A small ratio of upper to lower crustal strength leads to a distributed mode of faulting, where many
faults take up limited strain, and the upper crust is never completely pulled apart. A large strength
ratio results in few active fault zones, each accomodating a large amount of strain. This leads
to block rotation and complete dissection of the upper crust, with the consequent exhumation of
lower crustal rocks. Examples of this mode of extension may be the metamorphic core complexes
of the Western U.S.A. and the Aegean. The actual critical strength ratio for the transition between
modes will depend upon such factors as the relative thickness of the lower crust with respect to
the upper crust, and the degree of fault weakening. These are secondary factors that do not alter
the primary importance of vertical rheological contrasts in determining the mode of extension.
Changes in the relative strengths of crustal layers can most easily be related to heat flow, and the
mineral composition and water content of the lower crust.
68
Chapter 6
CRUSTAL-SCALE FLUID FLOW: CARLIN SYSTEM
6.1 Introduction
From the generic modelling of Chapter 5, the present chapter pushes the use of numerical sim-
ulations towards practical integration into a geologically-based industry. The Carlin gold trend, in
northern Nevada, U.S.A., serves as the example of crustal-scale tectonic and fluid flow modelling
applied to a mineral exploration problem.
The results in Chapter 5 demonstrate that the vertical stratification of strength controls fault
spacing and the large-scale tectonic behaviour of the crust. However, pre-existing structures are
often important in actually localising the strain patterns that describe this large-scale behaviour.
During a compressional orogeny, a crustal weakness may serve as the locus of a thrust fault. Rhe-
ological contrasts between basement blocks can also influence the degree and location of faulting
and relative uplift. In northern Nevada, basement architecture in the form of early rifted con-
tinental margins, formed during pre-Cambrian extension, may dictate the subsequent structural
geometry of overlying sedimentary sequences during large-scale compression. Within the region
of the Carlin gold trend, specific anticlinal fold and thrust geometries in the sedimentary rocks,
involved in various orogenies up until the Laramide, may focus fluid movement and provide effec-
tive traps to the system, resulting in the unique gold endowment of the area. Most mineralisation is
situated less than 100 m below the Roberts Mountain thrust, which defines the lower boundary of
the sequence of deep-water sedimentary rocks that has ridden over both the basement and younger
sedimentary layers.
Muntean et al. (2003) argue that the Carlin and Battle Mountain–Eureka (BME) gold trends
correspond to reactivated normal faults that likely had their origins in Proterozoic rifting. Numer-
ical modelling offers a way to test the basic hypothesis by which old normal faults, or “steps”
that are relics of continental rifting, control the subsequent location of upper crustal faults and
anticlinal structures during compression.
6.2 Geological Setting
Following continental rifting in the pre-Cambrian through to Devonian, present-day north-
ern Nevada has been subject to a number of compressional episodes of varying duration (Fig-
ure 6.1). These range from the Antler orogeny, approximately 340 Ma ago, through to the
Laramide orogeny, which ended with the onset of Basin and Range extension about 50 Ma ago
(Miller et al., 1992). The Roberts Mountain thrust, which defines a regional cap to the mineralisa-
69
CHAPTER 6. CRUSTAL-SCALE FLUID FLOW: CARLIN SYSTEM Wijns Ph.D. Thesis
Figure 6.1: Geological history of northern Nevada. Arrows indicate eastward (right) or westward (left)
orientation of driving forces. Timing and direction of events from Link et al. (1993), Bird (2002), Wyld
et al. (2003), and J. Muntean (pers. comm).
tion and probably acted as a permeability seal, occurred during the Antler orogeny. The deep-water
sedimentary units that were thrust overtop of inland units are commonly referred to as the “upper
plate”. The “lower plate” refers to the inland units that actually have their origin at higher strati-
graphic levels. Orogenic events subsequent to the Laramide emplaced more sedimetary sequences
over the Roberts Mountain allochthon. Although the direction of maximum stress changed for dif-
ferent orogenic events, the overall evolution of northern Nevada was dominated by east-directed
compression.
The linear arrangements of gold deposits along the BME and Carlin trends have prompted
many researchers to look for evidence of large-scale structural controls, especially in geophysical
data (e.g., Rodriguez, 1998; Grauch et al., 2003). The demarcation between ancient continental
crust and younger oceanic crust is well established through Pb and Sr isotope ratios (Wooden
et al., 1998; Grauch et al., 2003), but this boundary, although close, is not coincident with the
major mineral trends. Processing of gravity and magnetic data by Grauch et al. (2003) has revealed
features that align with mineral occurences; these are more persuasive for the BME than the Carlin
trend (Figure 6.2). A 2D inversion of magnetotelluric data also shows narrow, vertically extensive,
electrically conductive zones under the two trends, which Rodriguez (1998) interprets as crustal
faults.
If the geophysical data are highlighting major crustal faults that control the locations of the
mineral trends, these may be expressions of the reactivation of early normal rift faults at even
deeper levels (Figure 6.3). The relative offsets between reactivated normal faults and their prop-
agated thrusts in overlying sedimentary rocks are likely to be complicated by multiple orogenies,
gravitational slumping, and widespread extension in the Eocene and later.
70
Wijns Ph.D. Thesis CHAPTER 6. CRUSTAL-SCALE FLUID FLOW: CARLIN SYSTEM
a b
Figure 6.2: Gold deposits (circles and squares) on top of (a) 5 km upward-continued Bouguer gravity
anomaly with basin effects removed, and (b) magnetic potential with Pb and Sr isotope ratio boundaries.
Images from Grauch et al. (2003).
Figure 6.3: Schematic conceptual model, slightly modified from Muntean (2000), for the structural setting
of the Carlin trend above a basement normal fault that is reactivated during compression. The steps in the
basement represent the tops of pre-Cambrian extensional faults that are the product of a rifted continen-
tal margin. The actual cross-sectional distance between the Shoshone Mountains and the Carlin trend is
approximately 50 km.
71
CHAPTER 6. CRUSTAL-SCALE FLUID FLOW: CARLIN SYSTEM Wijns Ph.D. Thesis
a
bWest East
Figure 6.4: Initial numerical model, consisting of three basement blocks (Archaean in the east, Proterozoic
in the center, oceanic in the west) separated by faults and overlain by flat-lying sedimentary sequences.
Stripes in the sedimentary layers are marker units for visualising deformation. All bounding surfaces are
free-slip, and compression is from the west. (a) Entire crust, initially 20km thick. (b) Upper crust, 10 km
thick, where the top of the oceanic basement block is represented by the bottom boundary of the model.
Table 6.1: Natural values for Nevada model parameters.
Parameter Value
Whole crust model
depth of crust 20 km
compression velocity 10 mm/yr
Upper crust model
depth of crust 10 km
compression velocity 6 mm/yr
All models
densityρ of sedimentary rocks 2500 kg/m3
densityρ of basement rocks 3000 kg/m3
permeabilityk of sedimentary rocks, initial 10−14 m2
permeabilityk of basement rocks, initial 10−16 m2
cohesionco 16 MPa
friction coefficientcp of sedimentary rocks 0.6
friction coefficientcp of basement rocks 0.8
strain weakeninga 0.2
“saturation” strainεo 0.5
strain rateε, initial 1×10−15/s
thermal diffusivityκ 10−6 m2/s
gravityg 10 m/s2
72
Wijns Ph.D. Thesis CHAPTER 6. CRUSTAL-SCALE FLUID FLOW: CARLIN SYSTEM
6.3 Numerical Modelling
Section models in 2D follow an approximate ENE – WSW transect running from Archaean
cratonic crust in present-day western Nevada, U.S.A, to the oceanic crust along the protomargin
of western North America. The numerical model represents a number of sedimentary layers de-
posited over three basement crustal blocks – Archaean, transitional or Proterozoic, and oceanic –
that are the product of continental rifting during the Proterozoic through Devonian (Figure 6.4).
These blocks are separated by normal faults that provide zones of weakness that will reactivate.
Conceptually, the western rift fault represents the setting under the BME trend, whereas the east-
ern fault lies under the present-day Carlin trend. The model distance of 50 km between rift faults
is approximately the true distance today.
Basement blocks are much stronger than sedimentary layers. Relevant parameter values appear
in Table 6.1, along with the boundary conditions. All units are fluid saturated, but the permeability
of the sedimentary layers is initially two orders of magnitude greater than that of the basement
blocks. With continuing compression, this permeability evolves, but never reaches the low values
of the basement. Under accumulated shear, a dilation factor (c.f. equation 2.13 and Figure 2.2)
causes the permeability to increase up to a maximum of 50 times its original value, enhancing flow
along fault zones. Temperature is not modelled, therefore fluid flow is a result of mechanical and
gravitational forces only.
6.3.1 Whole crust tectonics
In the first instance, the entire crust, of initial thickness 20 km, undergoes compression over a
length of 200 km (Figure 6.4a). In the case of successively weaker basement blocks from east to
west (Figure 6.5a), most slip occurs along the first fault, and the eastern fault is barely reactivated.
The oceanic crust is thrust up against the transitional crust, providing the elevation for contin-
ued transport of western sedimentary rocks over eastern ones, mostly by gravitational slumping.
Eastern sedimentary layers are relatively isolated from deformation.
With the objective of determining the importance of basement rheological contrasts, the oceanic
and transitional blocks in Figure 6.5b are made the same strength. The lack of rheological contrast
does not affect the outcome in terms of strain partioning between basement faults. However,
the more competent oceanic crust undergoes faulting rather than homogeneous thickening, which
provides locally greater elevation for the sedimentary rocks in the vicinity of the basement thrust.
The eastern fault is only significantly reactivated when the western fault is considerably stronger
(Figure 6.5c). More compressive stress is transferred through the western fault, deforming the sed-
iments over the transitional crust. There is noa priori reason for a difference in fault strength, and
the greater elevation of the inland sedimentary rocks hinders the gravitational transport of western
units.
73
CHAPTER 6. CRUSTAL-SCALE FLUID FLOW: CARLIN SYSTEM Wijns Ph.D. Thesis
Figure 6.5: Simulations after 15% compression of initial model in Figure 6.4a. (a) Successively less com-
petent basement blocks from east to west, (b) equally competent oceanic and transitional blocks, and (c)
stronger western fault. Dark bands indicate accumulated plastic shear strain (faults).
6.3.2 Upper crust tectonics
Simulations on the upper crust alone assume that the top of the oceanic basement block sits
under the western sedimentary layers at the lower boundary of the model in Figure 6.4b. This
mimics the situation where oceanic crust subducts below the transitional and Archaean blocks. An
erosion surface near the top of the model box keeps the maximum elevation of material, relative
to the undeformed eastern level, at 5 km. Thus, gravitational slumping is not as pronounced as in
the whole-crust simulations, where erosion is neglected.
Figure 6.6 contains three snapshots of the model at different stages of shortening. Once again,
the eastern fault is markedly inactive, all strain being accomodated by the western fault. An early
backthrust forms off the asperity between oceanic and transitional crust (Figure 6.6a), and is car-
ried up and over the eastern sedimentary rocks. As this backthrust steepens due to continued
compression, a new one forms in the same original location (Figure 6.6b). This behaviour contin-
ues, with successive backthrusts becoming upright and inactive towards the east (Figure 6.6c), in
similar fashion to the results of Erickson et al. (2001). Only minor gravitational spreading occurs
in the thrust foreland.
When the basement step from oceanic to transitional crust is smaller, greater direct compressive
stress can be transmitted across the western fault. This leads to slightly more deformation of
eastern sedimentary units (Figure 6.7). However, even with 50% shortening (Figure 6.7c), the
deep-water sedimentary rocks (upper plate) are not thrust as far as the eastern fault, as is the case
postulated for the Carlin trend. This is probably dependent on the steep thrust angle, which is
discussed below.
74
Wijns Ph.D. Thesis CHAPTER 6. CRUSTAL-SCALE FLUID FLOW: CARLIN SYSTEM
Figure 6.6: Simulation based on initial model in Figure 6.4b, after (a) 15% shortening, (b) 25% shortening,
and (c) 34% shortening. An erosion surface limits the maximum elevation.
Figure 6.7: Simulation based on initial model in Figure 6.4b, but with a lower basement step from oceanic
to transitional crust. (a) 17% shortening, (b) 30% shortening, and (c) 50% shortening. An erosion surface
limits the maximum elevation.
6.3.3 Fluid flow
Fluid flow patterns are first dictated by the evolution of more permeable shear zones, and
initial compression moves fluid upwards along the reactivated western fault (Figure 6.8a). As the
thrust elevation builds, compressively driven upward flow competes with topographically driven
downward flow (Figure 6.8b). The isolation of the eastern sedimentary rocks to both deformation
and high fluid flux remains constant throughout the simulation, with eastward fluid flow, from both
compression and topography, directed along the main thrust shear.
75
CHAPTER 6. CRUSTAL-SCALE FLUID FLOW: CARLIN SYSTEM Wijns Ph.D. Thesis
Figure 6.8: Average representative fluid flow vectors, in white, for upper crust simulation from Figure 6.6.
(a) 5% shortening, before topographic build-up, (b) 15% shortening, with topographically driven flow, and
(c) 34% shortening.
6.3.4 Local scale
Figure 6.9 shows the effect on folding, at a more local scale, of compression against the ramp
between two basement blocks. Folding depends to a great degree on anisotropic strength (vis-
cosity) contrasts within layers, rather than strength contrasts between isotropic layers (Muhlhaus
et al., 2002a,b). The version ofEllipsis used for these simulations cannot simultaneously model
material anisotropy and failure, so yielding is omitted in this test. Results concentrate on local
folding that is beyond the resolution of the crustal-scale simulations above, and show that a ramp,
such as exists between basement blocks, is effective in localising an anticline in the sedimentary
rocks above it. This behaviour would be expected in the vicinity of both reactivated basement
faults.
6.4 Discussion
The structural evolution of the sedimentary sequences that host Carlin-type deposits depends
less on rheological differences between basement blocks (Archaean, transitional, oceanic) than
the ability to reactivate deep faults, which then propagate into the overlying sedimentary rocks.
Where erosion is neglected in the whole-crust model, high topography promotes upper plate mo-
tion across the lower plate by gravitational slumping rather than thrusting due to the driving com-
pressive stress. In the upper crust model that limits topography through an erosion surface, the
movement of the upper plate is more a result of the compressive stress. In neither case does the
upper plate reach the eastern fault that represents the structural setting of the Carlin trend. This
shortfall in the thrust distance is probably due mostly to the steep angle of failure (initial 45) that
is an outcome of the constitutive model in the numerical code. Successive thrust faults should
form at low angles, and this would facilitate eastward transport of the upper plate.
76
Wijns Ph.D. Thesis CHAPTER 6. CRUSTAL-SCALE FLUID FLOW: CARLIN SYSTEM
20 km
a b
Figure 6.9: (a) Initial model for testing folding at a local scale, seeded with small random anisotropy per-
turbations in the dark layer. (b) Final state after 33% compression. Compression velocity as in Figure 6.4b.
Gravitational collapse causes the formation of extensional normal faults while overall compres-
sion is still active (most apparent in Figure 6.6a and b). Uplift is important for promoting slumping
over great distances, so the erosional setting plays a large part in the effectiveness of this mecha-
nism as a contributor to transport of upper plate units. It is likely that such a gravitational driver
is responsible for the widespread extension apparent in the Basin and Range when compression
ceased after the Laramide orogeny (Axen et al., 1993).
Apart from the propagation of the reactivated rift fault, a ubiquitous feature of both crustal
models is the early backthrust formed off the asperity between basement blocks. The implication
is that an early backthrust should exist in such a field situation, but, depending on the amount
of shortening, it, and subsequent backthrusts, will be transported eastward along the main thrust
while rotating towards the vertical. A backthrust, always present at the location of the oceanic
to transitional contact of the model, can explain anomalous west-vergent deformation associated
with a geanticline that is locally, but not everywhere, subparallel to the Sr 0.706 line – the inferred
edge of the original continent (Madden-McGuire and Marsh, 1991; Saucier, 1997).
Field evidence shows that anticlines are important for hosting Carlin-type gold deposits, and
these would act as natural fluid ponding sites if seals are present. A ramp anticline often devel-
ops when there is compression across existing faults (e.g., Cooke and Pollard, 1997), or, as in
the model, against a ramp between two basement blocks. The fluid ponding potential of such
structures is enhanced if they remain unbroken by thrusting.
The fluid flow paths in Figure 6.8 illustrate the potential for mixing between fluids from deeper
sources that travel up the main thrust, and meteoric fluids that are driven by topography. The
maximum depth of meteoric fluid penetration is approximately 5 km, equal to the maximum rela-
tive topographic elevation. Thus, erosion will also be critical to determining the depth of mixing
between these fluids from different sources. The simplicity of Darcy flow and the homogeneous
nature of the sedimentary layers establish this determination of mixing level as a first-order con-
clusion. Probable anisotropic permeability from sedimentary layering during deposition should
move the level towards the surface.
Fluid flow through fault zones is naturally more complicated than what is modelled. There is no
provision in these simulations for the development of anisotropic permeability during shear. Such
a phenomenon might act as a barrier to flow perpendicular to fault zones, and this would serve
to further reduce fluid flux between upper and lower plate rocks. Darcy’s Law, which assumes
a laminar flow field, may not be appropriate for high fluxes along faults. Shear zones in the
brittle regime will evolve with extremely variable slip rates, periods of complete quiescence, and,
77
CHAPTER 6. CRUSTAL-SCALE FLUID FLOW: CARLIN SYSTEM Wijns Ph.D. Thesis
possibly, fault healing. All of these factors contribute to the great difficulty in modelling flow
along shear zones at long time scales.
Elevated pore pressure is one of the most effective ways to change the stress regime and pro-
mote failure and fracturing where it would not otherwise occur. A low-permeability cap in the
sedimentary layers, perhaps in the form of strong, shear-parallel, anisotropic permeability devel-
opment in the Roberts Mountain thrust, or layers of low-permeability shale units, would lead to
elevated pore pressures in the rock column below as more material is thrust above (e.g., Hubbert
and Rubey, 1959). This has the immediate effect of extending the depth of faulting towards the
basement, possibly reaching more deeply sourced fluids and enhancing the control that basement
features exert on fault locations. If a permeability seal can be kept contiguous, deeper faulting will
be allied to an effective fluid trap until a significant change in the stress regime allows venting and
mineral precipitation.
6.5 Chapter Summary
The reactivation of basement normal faults during subsequent compression leads to propagated
thrust faults in overlying, initially intact rock units. In northern Nevada, pre-Cambrian rift faults
dating from the stages of continental break-up may serve as the loci for thrust faults formed during
a series of compressional orogenies from 340 to 50 Ma. These thrusts carried deep-water sedimen-
tary rocks (upper plate) from the west over younger units to the east (lower plate). The failure of
the model simulations to carry the upper plate as far as the proposed location of the Carlin trend is
most likely due to a high angle of thrust initiation (45). This angle is determined by the material
constitutive law in the numerical code, and does not adequately reflect nature. In all simulations,
the western fault accomodates the vast majority of strain, and the eastern (Carlin) fault is protected
from reactivation unless it is initially much weaker than the western fault. The importance of (ex-
tensional) gravitational slumping in transporting the upper plate eastwards will be a function of
maximum topography, hence of erosion. Fluid flow is equally driven by the compressional stress
and by topography. Mixing of fluids from deep versus surface sources will occur at a level that
is a function of topographic relief, and which is probably not deeper than the maximum relative
surface elevation.
78
Chapter 7
CONCLUSION
Numerical modelling is a basic tool for tectonic studies. In the obvious absence of experi-
ments at a field scale, analogue models and computer simulations are two means by which to
extrapolate laboratory rock mechanics to a scale of interest for geology. With both techniques,
the greatest challenge is the accurate representation of the behaviour of geologic materials: in the
analogue case, this requires confidence in the validity of using proxy materials, and in the numer-
ical case, confidence in the validity of the physical equations. As limitations on computational
loads disappear with technical advances in the computer hardware industry, the richness in detail
of three-dimensional numerical modelling will equal that of analogue modelling. This will leave
the numerical approach as the preferred technique for geological simulations, due to the control
over measurements and material properties, the ease of visualisation of results, and the rapidity
with which parameters can be varied in order to test new scenarios. Analogue models will remain
valuable for testing numerical codes on the physics of known materials.
In line with the vision of expanding the scope of numerical modelling, the inclusion of porous
fluid flow into a high-deformation code such asEllipsis opens the door to new areas of modelling
not possible with the analogue technique. In Appendix A, a test with coupled pore fluid pressure
verifies the Byerlee approximation used in the non-porous modelling of Chapters 3 through 5.
Pathways for possible mineral-bearing fluids are calculated in Chapter 6 during the evolution of a
compressional orogen. The application of ever-increasing computational sophistication, however,
must be tempered by constant referral to field observations. Chapters 5 and 6 compare results to
cooling curves, seismic sections, viscosity estimates, and field mapping, and make suggestions
for guiding new observations. This feedback loop is crucial for establishing numerical modelling
as a real tool in applied geology, from generic studies to specific cases. It should influence both
the collection of observations and subsequent refinements to the initial conceptual and numerical
models.
The application of computer modelling to any problem, but especially to specific field situa-
tions, comes with a warning against over-interpretation. Although this warning is based on the
self-evident truth that a result is only as good as its input, the pitfalls may not always be obvi-
ous, especially to those less aware of the numerics behind whatever software is being used. As
an example, Appendix B illustrates different results for the exact same geological problem, due
simply to changes in computational grid density and solution accuracy in theEllipsis algorithms.
More obvious candidates for influencing the results are the material parameter values of the model.
Changes in some parameters will influence a result more than changes in other parameters – this
is referred to as model sensitivity. It is essential to explore the model sensitivity with respect to all
79
CHAPTER 7. CONCLUSION Wijns Ph.D. Thesis
parameters being used. This allows conclusions about model behaviour to be formulated with due
consideration for uncertainties.
A method for handling the often daunting task of exploring model sensitivity is available
through the interactive evolutionary computation (IEC) and visualisation system of Chapters 3
and 4. Not only is the navigation through parameter space towards one or more final solutions
made simple, but associated techniques for visualising all solutions become automatic sensitivity
maps. The effectiveness of the IEC system declines with an increase in the number of model pa-
rameters, eventually reaching a stage where the number of required outputs for ranking cannot be
matched by either the available computer time or the ability of a human to evaluate all solutions,
and the reduction of dimensionality for visualisation is equally strained. This breakdown of the
method, however, is a reminder of the point at which the usefulness of a model is probably suffer-
ing from the inclusion of too many parameters, to the detriment of a firm grasp of the fundamentals
of the problem.
The development of porous flow capabilities within theEllipsis code, the use and refinement
of the IEC system, and the adoption of various visualisation techniques for parameter space all
culminate in the ability to perform meaningful numerical modelling for new problems in tecton-
ics and fluid flow, with an understanding of the expanded solution space around the conceptual
model. In the larger geological modelling community, the transition to three dimensions, minus
porous flow, has already been made in a successor toEllipsisand in independent high-deformation
codes; the coupling of chemical reactions to stress and fluid flow is a reality with other software;
and research is being undertaken on the physics of translating laboratory-scale rock mechanics
phenomena to ever larger scales. The growing computational complexity made possible by this
research will demand continued attention to the basic concepts of responsible modelling that have
been outlined throughout this thesis, and the adoption of systems, such as IEC plus visualisation,
that will render tractable the task of analysing model results and formulating viable conclusions.
80
REFERENCES
Arfken, G., 1985. Mathematical Methods for Physicists, 3rd Edition. Academic Press, Orlando,
985 pp.
Axen, G., Taylor, W., Bartley, J., 1993. Space-time patterns and tectonic controls of Tertiary
extension and magmatism in the Great Basin of the western United States. Geological Society
of America Bulletin 105 (1), 56–76.
Axen, G., Wernicke, B., Skelly, M., Taylor, W., 1990. Mesozoic and Cenozoic tectonics of the
Sevier thrust belt in the Virgin River Valley area, southern Nevada. In: Wernicke, B. (Ed.),
Basin and Range Extensional Tectonics Near the Latitude of Las Vegas, Nevada. Memoir 176.
Geological Society of America, Boulder, Colorado, pp. 123–153.
Bai, T., Pollard, D., 2000. Fracture spacing in layered rocks: a new explanation based on the stress
transition. Journal of Structural Geology 22 (1), 43–57.
Bellingham, P., White, N., 2000. A general inverse method for modelling extensional sedimentary
basins. Basin Research 12, 219–226.
Benes, V., Davy, P., 1996. Modes of continental lithospheric extension: experimental verification
of strain localization processes. Tectonophysics 254 (1–2), 169–187.
Bercovici, D., Ricard, Y., Schubert, G., 2001. A two-phase model for compaction and damage 1.
general theory. Journal of Geophysical Research 106 (B5).
Biot, M., 1941. General theory of three-dimensional consolidation. Journal of Applied Physics 12,
155–164.
Bird, P., 1995. Lithosphere dynamics and continental deformation. Reviews of Geophysics 33,
supplement (July), 379–383.
Bird, P., 2002. Stress direction history of the western United States and Mexico since 85 Ma.
Tectonics 21 (3).
Bird, P., Kong, X., 1994. Computer simulations of California tectonics confirm very low strength
of major faults. Geological Society of America Bulletin 106 (2), 159–174.
Boschetti, F., Dentith, M., List, R., 1996. Inversion of seismic refraction data using genetic algo-
rithms. Geophysics 61 (6), 1715–1727.
81
REFERENCES Wijns Ph.D. Thesis
Brace, W., Kohlstedt, D., 1980. Limits on lithospheric stress imposed by laboratory experiments.
Journal of Geophysical Research 85 (B11), 6248–6252.
Briggs, W., 1987. A Multigrid Tutorial. Society for Industrial and Applied Mathematics, Lan-
caster, Pennsylvania, 88 pp.
Brun, J.-P., Sokoutis, D., van den Driessche, J., 1994. Analogue modelling of detachment fault
systems and core complexes. Geology 22, 319–322.
Buck, W., 1988. Flexural rotation of normal faults. Tectonics 7 (5), 959–973.
Buck, W., 1991. Modes of continental lithospheric extension. Journal of Geophysical Research
96 (B12), 20161–20178.
Burchfiel, B., Nakov, R., Tzankov, T., Royden, L., 2000. Cenozoic extension in Bulgaria and
northern Greece: the northern part of the Aegean extensional regime. In: Bozkurt, E., Winch-
ester, J., Piper, J. (Eds.), Tectonics and Magmatism in Turkey and the Surrounding Area. Special
Publication 173. Geological Society, London, pp. 325–352.
Byerlee, J., 1968. Brittle-ductile transition in rocks. Journal of Geophysical Research 73, 4741–
4750.
Carmen, P., 1956. Flow of Gases Through Porous Media. Academic Press, New York, 182 pp.
Christiansen, P., Pollard, D., 1997. Nucleation, growth and structural development of mylonitic
shear zones in granitic rock. Journal of Structural Geology 19 (9), 1159–1172.
Chulick, G., Mooney, W., 2002. Seismic structure of the crust and uppermost mantle of North
America and adjacent oceanic basins: a synthesis. Bulletin of the Seismological Society of
America 92 (6), 2478–2492.
Cooke, M., Pollard, D., 1997. Bedding-plane slip in initial stages of fault-related folding. Journal
of Structural Geology 19 (3–4), 567–581.
Corti, G., Bonini, M., Innocenti, F., Manetti, P., Mulugeta, G., 2001. Centrifuge models simulating
magma emplacement during oblique rifting. Journal of Geodynamics 31 (5), 557–576.
Cross, T., Lessenger, M., 1999. Construction and application of a stratigraphic inverse model. In:
Harbaugh, J., Watney, W., Rankey, E., Slingerland, R., Goldstein, R., Franseen, E. (Eds.), 62
Numerical Experiments in Stratigraphy: Recent Advances in Stratigraphic and Sedimentologic
Computer Simulations. SEPM (Society for Sedimentary Geology), pp. 69–83.
Cundall, P., Board, M., 1988. A microcomputer program for modelling large-strain plasticity prob-
lems. In: Swobada, C. (Ed.), Numerical Methods in Geomechanics. Proceedings of the 6th
International Conference on Numerical Methods in Geomechanics. Balkema, Rotterdam, pp.
2101–2108.
82
Wijns Ph.D. Thesis REFERENCES
Darcy, H., 1856. Les Fontaines Publiques de la Ville de Dijon (The Public Fountains of the City
of Dijon). Dalmont, Paris, 647 pp. and atlas.
Duebendorfer, E., Sewall, A., Smith, E., 1990. The Saddle Island detachment: an evolving shear
zone in the Lake Mead area, Nevada. In: Wernicke, B. (Ed.), Basin and Range Extensional
Tectonics Near the Latitude of Las Vegas, Nevada. Memoir 176. Geological Society of America,
Boulder, Colorado, pp. 77–97.
Dunbar, J., Sawyer, D., 1989. How preexisting weaknesses control the style of continental
breakup. Journal of Geophysical Research 94 (B6), 7278–7292.
Erickson, S., Luther, M., Suppe, J., 2001. Initiation and reactivation of faults during movement
over a thrust-fault ramp: numerical mechanical models. Journal of Structural Geology 23, 11–
23.
Flesch, L., Holt, W., Haines, A., Shen-Tu, B., 2000. Dynamics of the Pacific-North American
plate boundary zone in the western United States. Science 287, 834–836.
Forster, M., Lister, G., 1999. Detachment faults in the Aegean core complex of Ios, Cyclades,
Greece. In: Ring, U., Brandon, M., Lister, G., Willett, S. (Eds.), Exhumation Processes: Normal
Faulting, Ductile Flow and Erosion. Geological Society, London, pp. 305–323.
Fossen, H., Rœrnes, A., 1996. Properties of fault populations in the Gullfaks Field, northern North
Sea. Journal of Structural Geology 18 (2–3), 179–190.
Frank-Kamenetskii, D., 1969. Diffusion and Heat Transfer in Chemical Kinetics, 2nd Edition.
Plenum Press, New York, 574 pp.
Gauthier, P., Brun, J.-P., 1994. Crustal-scale geometry and kinematics of late-orogenic extension
in the central Aegean (Cylcades and Evvia Island). Tectonophysics 238, 399–424.
Gessner, K., Ring, U., Johnson, C., Hetzel, R., Passchier, C., Gungor, T., 2001. An active biver-
gent rolling-hinge detachment system: Central Menderes metamorphic core complex in western
Turkey. Geology 29 (7), 611–614.
Goldberg, D., 1989. Genetic Algorithms in Search, Optimization, and Machine Learning.
Addison-Wesley, Reading, Massachusetts, 412 pp.
Goodman, R., 1989. Introduction to Rock Mechanics, 2nd Edition. John Wiley and Sons, New
York, 576 pp.
Grauch, V., Rodriguez, B., Wooden, J., 2003. Geophysical and isotopic constraints on crustal
structure related to mineral trends in north-central Nevada and implications for tectonic history.
Economic Geology 98, 269–286.
Guest, A., Cundall, P., 1994. A three-dimensional particle flow code for block cave modelling. In:
The Application of Numerical Modeling in Geo-technical Engineering. ISRM-SANG, Pretoria,
pp. 15–18.
83
REFERENCES Wijns Ph.D. Thesis
Handy, M., Wissing, S., Streit, L., 1999. Frictional-viscous flow in mylonite with varied biminer-
alic composition and its effect on lithospheric strength. Tectonophysics 303 (1–4), 175–191.
Hauser, E., Gephart, T., Latham, T., Bown, L., Kaufman, S., Oliver, J., Lucchitta, I., 1987. CO-
CORP Arizona transect: strong crustal reflection data. Geological Society of America Bulletin
99, 833–844.
Hetzel, R., Ring, U., Akal, C., Troesch, M., 1995. Miocene NNE-directed extensional unroofing
in the Menderes massif, southwestern Turkey. Journal of the Geological Society 152, 639–654.
Hodges, K., McKenna, L., Harding, M., 1990. Structural unroofing of the central Panamint Moun-
tains, Death Valley region, southeastern California. In: Wernicke, B. (Ed.), Basin and Range
Extensional Tectonics Near the Latitude of Las Vegas, Nevada. Memoir 176. Geological Soci-
ety of America, Boulder, Colorado, pp. 377–388.
Horvath, F., 1993. Towards a mechanical model for the formation of the Pannonian Basin. In:
Cloetingh, S., Sassi, W., Horvath, F. (Eds.), The Origin of Sedimentary Basins: Inferences from
Quantitative Modelling and Basin Analysis. Elsevier, Amsterdam, pp. 333–357.
Hubbert, M., Rubey, W., 1959. Role of fluid pressure in mechanics of overthrust faulting. Geolog-
ical Society of America Bulletin 70, 115–166.
Hughes, T., 1987. The Finite Element Method. Prentice-Hall, Englewood Cliffs, 803 pp.
Itasca Consulting Group, I., 2003. PFC3D (Particle Flow Code in Three Dimensions). Minneapo-
lis, 2nd Edition.
Jolivet, L., 2001. A comparison of geodetic and finite strain pattern in the Aegean, geodynamic
implications. Earth and Planetary Science Letters 187 (1–2), 95–104.
Jolliffe, I., 1986. Principal Component Analysis. Springer-Verlag, New York, 271 pp.
Jones, C., Wernicke, B., Farmer, G., Walker, J., Coleman, D., McKenna, L., Perry, F., 1992.
Variations across and along a major continental rift: an interdisciplinary study of the Basin and
Range Province, western U.S.A. Tectonophysics 213, 57–96.
Kaus, B. J., Podladchikov, Y. Y., 2001. Forward and reverse modeling of the three-dimensional
viscous Rayleigh-Taylor instability. Geophysical Research Letters 28 (6), 11095–11098.
Kohonen, T., 2001. Self-Organizing Maps, 3rd Edition. No. 30 in Series in Information Sciences.
Springer, New York, 501 pp.
Koyi, H., Skelton, A., 2001. Centrifuge modelling of the evolution of low-angle detachment faults
from high-angle normal faults. Journal of Structural Geology 23, 1179–1185.
Lachenbruch, A., Sass, J., 1978. Models of extending lithosphere and heat flow in the Basin and
Range province. In: Smith, R., Eaton, G. (Eds.), Cenozoic Tectonics and Regional Geophysics
84
Wijns Ph.D. Thesis REFERENCES
of the Western Cordillera. Memoir 152. Geological Society of America, Boulder, Colorado, pp.
209–250.
Lachenbruch, A., Sass, J., 1980. Heat flow and energetics of the San Andreas fault zone. Journal
of Geophysical Research 85 (B11), 6185–6222.
Lachenbruch, A., Sass, J., 1992. Heat flow from Cajon Pass, fault strength, and tectonic implica-
tions. Journal of Geophysical Research 97, 4995–5015.
Lamb, H., 1895. Hydrodynamics. Cambridge University Press, Cambridge, 604 pp.
Lavier, L., Buck, W., 2002. Half graben versus large-offset low-angle normal fault: importance of
keeping cool during normal faulting. Journal of Geophysical Research 107 (B6).
Lavier, L., Buck, W., Poliakov, A., 2000. Factors controlling normal fault offset in an ideal brittle
layer. Journal of Geophysical Research 105 (B10), 23431–23442.
Link, P., Christie-Blick, N., Devlin, W., Elston, D., Horodyski, R., Levy, M., Miller, J., Pearson,
R., Prave, A., Stewart, J., Winston, D., Wright, L., Wrucke, C., 1993. Middle and Late Pro-
terozoic stratified rocks of the western U.S. Cordillera, Colorado Plateau, and Basin and Range
province. In: Reed, J., Bickford, M., Houston, R., Link, P., Rankin, D., Sims, P., Schmus, W. V.
(Eds.), Precambrian: Conterminous U.S.: The Geology of North America. Vol. C-2. Geological
Society of America, Boulder, Colorado, pp. 463–595.
Livaccari, R. F., Geissman, J. W., Reynolds, S. J., 1995. Large-magnitude extensional deformation
in the South Mountains metamorphic core complex, Arizona: evaluation with paleomagnetism.
Geological Society of America Bulletin 107 (8), 877–894.
Madden-McGuire, D. J., Marsh, S. P., 1991. Lower Paleozoic host rocks in the Getchell gold
belt: Several distinct allochthons or a sequence of continuous sedimentation? Geology 19 (5),
489–492.
Mavko, G., Nur, A., 1997. The effect of a percolation threshold in the Kozeny-Carman relation.
Geophysics 62 (5), 1480–1482.
McKenzie, D., 1984. The generation and compaction of partially molten rock. Journal of Petrology
25 (3), 713–765.
McKenzie, D., Fairhead, D., 1997. Estimates of the effective elastic thickness of the continen-
tal lithosphere from Bouger and free air gravity anomalies. Journal of Geophysical Research
102 (B12), 27523–27552.
Miller, E., Miller, M., Stevens, C., Wright, J., Madrid, R., 1992. Late Paleozoic paleogeographic
and tectonic evolution of the western U.S. Cordillera. In: Burchfiel, B., Lipman, P., Zoback, M.
(Eds.), The Cordilleran Orogen: Conterminous U.S.: The Geology of North America. Vol. G-3.
Geological Society of America, Boulder, Colorado, pp. 57–106.
85
REFERENCES Wijns Ph.D. Thesis
Mishra, D., Chandra Sekhar, D., Venkata Raju, D., Vijaya Kumar, V., 1999. Crustal structure based
on gravity-magnetic modelling constrained from seismic studies under Lambert Rift, Antarctica
and Godavari and Mahanadi rifts, India and their interrelationship. Earth and Planetary Science
Letters 172 (3–4), 287–300.
Moresi, L., 1992. The relation of surface observables to convection in a fluid with variable viscos-
ity. Ph.D. thesis, Oxford University.
Moresi, L., Dufour, F., Muhlhaus, H.-B., 2002. Mantle convection modeling with viscoelas-
tic/brittle lithosphere: Numerical methodology and plate tectonic modeling. Pure and Applied
Geophysics 159 (10), 2335–2356.
Moresi, L., Muhlhaus, H.-B., Dufour, F., 2001. Viscoelastic formulation for modelling of plate
tectonics. In: Muhlhaus, H.-B., Dyskin, A., Pasternak, E. (Eds.), Bifurcation and Localization
in Soils and Rocks. Balkema, Rotterdam, pp. 337–344.
Moresi, L., Solomatov, V. S., 1995. Numerical investigation of 2d convection with extremely large
viscosity variations. Physics of Fluids 7, 2154–2162.
Muhlhaus, H.-B., Dufour, F., Moresi, L., Hobbs, B., 2002a. A director theory for viscoelastic fold-
ing instabilities in multilayered rock. International Journal of Solids and Structures 39 (13/14),
3675–3691.
Muhlhaus, H.-B., Moresi, L., Hobbs, B., Dufour, F., 2002b. Large amplitude folding in finely
layered viscoelastic rock structures. Pure and Applied Geophysics 159 (10), 2311–2333.
Mulugeta, G., Ghebreab, W., 2001. Modeling heterogeneous stretching during episodic or steady
rifting of the continental lithosphere. Geology 29 (10), 895–898.
Muntean, J., Sep. 2000. Models for Carlin-type gold deposits, unpublished internal Placer Dome
Ltd. report.
Muntean, J., Coward, M., Tarnocai, C., 2003. Paleozoic normal faults in north-central Nevada:
deep crustal conduits for Carlin-type gold deposits, unpublished paper presented at Marco T.
Einaudi Symposium, April 3, 2003, Colorado School of Mines, 20 pp.
Murray, M. H., Segall, P., 2001. Modeling broadscale deformation in northern California and
Nevada from plate motions and elastic strain accumulation. Geophysical Research Letters 28,
4315–4318.
Niemi, N., Wernicke, B., Brady, R., Saleeby, J., Dunne, G., 1999. Magnitude and timing of ex-
treme continental extension, central Death Valley region, California. In: Slate, J. (Ed.), Pro-
ceedings of Conference on Status of Geologic Research and Mapping, Death Valley National
Park. Open-File Report 99-153. U.S. Geological Survey, Denver, Colorado, pp. 33–35.
Parker, R., 1977. Understanding inverse theory. Annual Review of Earth and Planetary Sciences
5, 35–64.
86
Wijns Ph.D. Thesis REFERENCES
Parsons, T., McCarthy, J., Thompson, G., 2001. Very different response to extreme extension in
the southern Basin and Range and Colorado Plateau transistion. In: Erskine, M., Faulds, J.,
Bartley, J., Rowley, P. (Eds.), American Association of Petroleum Geologists Pacific Section
Guidebook 78. American Association of Petroleum Geologists, pp. 291–303.
Phillips, O., 1991. Flow and reactions in permeable rocks. Cambridge University Press, New York,
295 pp.
Pollitz, F., Wicks, C., Thatcher, W., 2001. Mantle flow beneath a continental strike-slip fault:
postseismic deformation after the 1999 Hector Mine earthquake. Science 293, 1814–1818.
Ricard, Y., Bercovici, D., Schubert, G., 2001. A two-phase model for compaction and damage 2.
applications to compaction, deformation, and the role of interfacial surface tension. Journal of
Geophysical Research 106 (B5), 8907–8924.
Rodriguez, B., 1998. Regional crustal structure beneath the Carlin trend, Nevada based on deep
electrical geophysical measurements. In: Tosdal, R. (Ed.), Contributions to the Gold Metal-
logeny of Northern Nevada. Open-File Report 98-338. U.S. Geological Survey, Denver, Col-
orado, pp. 15–19.
Ross, J., Wilks, K., 1995. Effects of a third phase on the mechanical and microstructural evolution
of a granulite. Tectonophysics 241 (3–4), 303–315.
Rybacki, E., Dresen, G., 2004. Deformation mechanism maps for feldspar rocks. Tectonophysics
382 (3–4), 173–187.
Sammon, Jr., J., 1969. A nonlinear mapping for data structure analysis. IEEE Transactions on
Computers C-18, 401–409.
Saucier, A., 1997. The Antler thrust system in northern Nevada. In: Perry, A., Abbot, E. (Eds.),
The Roberts Mountains Thrust, Elko and Eureka Counties, Nevada. Nevada Petroleum Society,
pp. 1–16.
Saunders, P., Priestley, K., Taymaz, T., 1998. Variations in the crustal structure beneath western
Turkey. Geophysical Journal International 134 (2), 373–389.
Schultz, R., 1996. Relative scale and the strength and deformability of rock masses. Journal of
Structural Geology 18 (9), 1139–1149.
Scott, D., Stevenson, D., 1984. Magma solitons. Geophysical Research Letters 11, 1161–1164.
Scott, R., Lister, G., 1992. Detachment faults: evidence for a low-angle origin. Geology 20 (9),
833–836.
Spadini, G., Podladchikov, Y., 1996. Spacing of consecutive normal faulting in the lithosphere:
a dynamic model for rift axis jumping (Tyrrhenian Sea). Earth and Planetary Science Letters
144 (1–2), 21–34.
87
REFERENCES Wijns Ph.D. Thesis
Spencer, J., Reynolds, S., 1991. Tectonics of mid-Tertiary extension along a transect through west
central Arizona. Tectonics 10, 1204–1221.
Suppe, J., 1985. Principles of Structural Geology. Prentice-Hall, New Jersey, 537 pp.
Takagi, H., 2001. Interactive evolutionary computation: fusion of the capacities of EC optimiza-
tion and human evaluation. Proceedings of the IEEE 89 (9), 1275–1296.
Tarantola, A., 1987. Inverse Problem Theory. Elsevier, Amsterdam, 613 pp.
Thatcher, W., Foulger, G., Julian, B., Svarc, J., Quilty, E., Bawden, G., 1999. Present day defor-
mation across the Basin and Range Province, western United States. Science 282, 1714–1718.
Thibaud, R., Dauteuil, O., Gente, P., 1999. Faulting pattern along slow-spreading ridge segments:
a consequence of along-axis variation in lithospheric rheology. Tectonophysics 312 (2–4), 157–
174.
Turcotte, D., Schubert, G., 1982. Geodynamics. John Wiley and Sons, New York, 450 pp.
Viejo, G., Laigleb, M., Raneroc, C., 2002. Pre-Permian sedimentary basins in the North Sea: im-
ages from reprocessed and pre-stack depth migrated MONA LISA data. Marine and Petroleum
Geology 19 (5), 519–526.
Wernicke, B., 1985. Theory of large-scale, uniform-sense normal simple shear of the continental
lithosphere. Canadian Journal of Earth Science 22, 108–125.
Wernicke, B., Axen, G., 1988. On the role of isostasy in the evolution of normal fault systems.
Geology 16, 848–851.
Wernicke, B., Spencer, J., Burchfiel, B., Guth, P., 1982. Magnitude of crustal extension in the
southern Great Basin. Geology 10 (10), 499–503.
Wijns, C., Poulet, T., Boschetti, F., Dyt, C., Griffiths, C., 2004. Interactive inverse methodology
applied to stratigraphic forward modelling. In: Curtis, A., Wood, R. (Eds.), Geological Prior
Information. Special Publication 239. Geological Society, London, pp. 147–156, in press.
Wilks, K., Carter, N., 1990. Rheology of some continental lower crustal rocks. Tectonophysics
182 (1–2), 57–77.
Wills, S., Buck, W., 1997. Stress-field rotation and rooted detachment faults: a Coulomb failure
analysis. Journal of Geophysical Research 102 (B9), 20503–20514.
Wooden, J., Kistler, R., Tosdal, R., 1998. Pb isotopic mapping of crustal structure in the north-
ern Great Basin and relationships to Au deposit trends. In: Tosdal, R. (Ed.), Contributions to
the Gold Metallogeny of Northern Nevada. Open-File Report 98-338. U.S. Geological Survey,
Denver, Colorado, pp. 20–33.
88
Wijns Ph.D. Thesis REFERENCES
Wyld, S. J., Rogers, J. W., Copeland, P., 2003. Metamorphic evolution of the Luning-Fencemaker
fold-thrust belt, Nevada: illite crystallinity, metamorphic petrology, and40Ar/39Ar geochronol-
ogy. Journal of Geology 111, 17–38.
Zoback, M., Prescott, W., Krueger, S., 1985. Evidence for lower crustal ductile strain localization
in southern New York. Nature 317, 705–707.
89
APPENDIX A
PORE PRESSURE APPROXIMATION
As mentioned in Chapter 3, from laboratory experiments by Byerlee (1968), there appears to be
a universal friction coefficientcp of 0.6 – 0.85 for dry rock. In a saturated rock column, assuming
an average hydrostatic pore pressure, the solid pressure is reduced, so that the zone of failure is
deeper. For dry rock of densityρs, the pressurepsdry = ρsgz (without deformation). In rock
saturated with water of densityρp, the solid pressurepssat is related to the effective pressurepe by
pssat = pe − pp ≈ ps
dry − ρpgz
usingpe ≈ psdry for small porosity. As a fraction of the dry pressure,
pssat
psdry
=(ρs − ρp) gz
ρsgz=
ρs − ρp
ρs.
For the trial in Figure A.1, the pore pressure is kept hydrostatic through negligible pore space
compression and a permeability that does not evolve. Extension velocity, dimensions, and all
other material parameters are equal to those for the IEC experiment of Chapter 3. Usingρs = 2700
kg/m3 andρp = 1000 kg/m3, a Byerlee coefficient of 0.6 for dry rock becomes 0.38 for saturated
rock. Although this approach seems valid in terms of the initial depth of yielding, it is certainly
too simplistic to deal with the effect of pore pressure during deformation of the rock.
Figure A.1: Extension experiment (33% extension) for a single crustal layer. (a) with a Byerlee coefficient
of 0.6 for dry rock, (b) with fully coupled pore pressure, and (c) with a Byerlee coefficient of 0.38, to
account for an average hydrostatic pore pressure. The brittle zone is the same depth as in (b).
91
APPENDIX B
MESH DEPENDENCE OF MODEL RESULTS
A numerical modelling code that has no explicit internal length scale for deformation, such as
the version ofEllipsis used for this dissertation, is callednon-regularised. The results of such a
code exhibit a dependence on the resolution of the computational grid. In effect, the grid spac-
ing defines an implicit length scale for deformation. It also controls the width of shear bands.
This phenomenon can sometimes be used to advantage in choosing a grid resolution, but, at the
same time, requires that conclusions with respect to the results, particularly in a spatial sense, be
drawn with caution. Figure B.1a contains an extension experiment, against which are compared
the results of two other simulations with identical input files, save the finite element mesh density
(Figure B.1b), and computational accuracy of the finite element loops (Figure B.1c). The differ-
ences are due to the nucleation of shear zones at sites of small perturbations in the stress field, and
minute changes in calculated velocities are sufficient to displace these perturbations.
Figure B.1: (a) Extension experiment similar to simulation C from Chapter 5 (c.f. Figure 5.3). (b) Same
experiment at a finite element grid resolution of 83% of original. (c) Same experiment with a computational
accuracy twice as large as original.
The above cases underline the importance of running the same simulation at different grid spac-
ings or accuracy. This will highlight the results that can be adopted with confidence. For example,
the location of fault zones changes dramatically, and should therefore not form part of any mod-
elling conclusions. The relatively large spacing between fault zones is quite consistent, showing
that a weak crust does indeed promote widely spaced faulting. The most robust conclusion is that,
in all cases, the upper crust is completely pulled apart, exhuming the lower crust.
All the modelling in this dissertation heeds the caution just given. Any conclusions are based
on multiple runs at different resolutions, and sometimes different accuracies.
93
APPENDIX C
CRUSTAL VISCOSITY PROFILE
C.1 Viscosity at Brittle to Ductile Transition
In order to create a continuous maximum shear stress profile with depth, the viscosityη at the
brittle to ductile transitionzu must satisfy both the brittle yield equation (2.8) and the temperature-
dependent viscosity equation (5.1).
η(zu) = ηu =τyield(zu)
ε= ηoe
−cT (zu) .
Having fixed the strain rate and bothco andcp, ηu is derived from equation (2.8) for the yield
stress before any strain softening, noting that the total overburden pressurep is the lithostatic
stress reduced by the extensional stress.
co + cp (ρugzu − ηu ε) = ηuε
ηu =co + ρugzucp
(1 + cp) ε.
The constantsηo and c control the viscosity profile through the ductile region, and are related
through the equations above.
e−cT (zu) =ηu
ηo
−c∂T
∂zzu = ln
(co + ρugzucp
(1 + cp) ηoε
)
c = −(
∂T
∂zzu
)−1
ln(
co + ρugzucp
(1 + cp) ηoε
).
C.2 Integrated Crustal Strength
Referring to Figure 5.1b, the integrated strengthτint of the crust is simply the area between 0
and the maximum shear stress. This maximum stress is defined by the yield envelope in the brittle
zone (equation 2.8), and by the viscous stressη ε in the ductile zone. The crustal response to
extension is defined by the ratiorτ of the integrated strength of the upper crust (initially all brittle)
to that of the lower crust (initially all ductile):
rτ =τuint
τ lint
.
95
APPENDIX C. CRUSTAL VISCOSITY PROFILE Wijns Ph.D. Thesis
For the upper crust at the onset of extension,
τuint =
∫ zu
0τyield dz
= cozu +12zu (τyield(zu)− co)
= cozu +12zu
(co + ρgzucp
1 + cp− co
)
=zu
2
(co +
co + ρgzucp
1 + cp
).
Integrating the viscous stress profile for the lower crust at an initial constant strain rate and linear
temperature profile,
τl(int) =∫ zl
zu
η(T )ε dz
= ε
∫ zl
zu
ηoe−cT dz
= ε
∫ zc
zu
ηoe−c ∂T
∂zz dz + ε
∫ zl
zc
ηoe−cT (zc) dz
= − ηoε
c∂T∂z
[e−c ∂T
∂zz]zc
zu
+ ηoεe−cT (zc) (zl − zc)
= − ηoε
c∂T∂z
[e−cT (zc) − e−cT (zu)
]+ ηoεe
−cT (zc) (zl − zc) .
96
Appendix D (on CD-ROM) is available with the print copy of the thesis, which is held in the University Library.