See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/301351333

PhD Thesis(Xue Zhang): Particle Finite Element Method in Geomechanics

Thesis · October 2014

CITATIONS READS

13 3,825

1 author:

Xue Zhang
University of Liverpool
63 PUBLICATIONS 1,525 CITATIONS

SEE PROFILE

All content following this page was uploaded by Xue Zhang on 18 April 2016.

The user has requested enhancement of the downloaded file.

Particle Finite Element Method
in Geomechanics

by

Xue Zhang
BEng, MEng
Faculty of Engineering and Built Environment
The University of Newcastle, Australia

A thesis submitted in partial fulfilment of requirement for the

degree of
Doctor of Philosophy

September 2014
 Declaration

The thesis contains no material which has been accepted for the award
of any other degree or diploma in any university or other tertiary in-
stitution and, to the best of my knowledge and belief, contains no
material previously published or written by another person, except
where due reference has been made in the text. I give consent to
the final version of my thesis being made available worldwide when
deposited in the University’s Digital Repository∗∗ , subject to the pro-
visions of the Copyright Act 1968.
∗∗
Unless an Embargo has been approved for a determined period.

(signed)

i
ii
 Dedication

To Defu, Shouzhen and Ya for their endless love, support and

encouragement.

iii
iv
 Acknowledgements

I am extremely grateful to the many wonderful people who have

shaped my experience at The University of Newcastle (UoN) and at
the ARC Centre of Excellence for Geotechnical Science and Engineer-
ing (CGSE), without whom this work would not have been possible.
First and foremost, I would like to express my heartfelt gratitude
to my principle supervisor A/Prof. Kristian Krabbenhoft who
has continually been a source of energy, ideas, as well as enthusi-
asm throughout my time at the UoN and at the CGSE. I am also
graciously indebted to my co-supervisor Prof. Daichao Sheng for
his constant support, availability and constructive suggestions, which
were determinant for the accomplishment of the work presented in
this thesis.
I am obliged to Prof. Xikui Li (Dalian University of Technology,
Dalian, China), my BEng and MEng supervisor as well as the lecturer
for my postgraduate course “Continuum Mechanics” at Dalian, who
introduced me into computational mechanics.
Appreciation also goes to Dr. Dong Wang (The University of West-
ern Australia, Perth, Australia), Dr. Dorival Pedroso (The Univer-
sity of Queensland, Brisbane, Australia), and Dr. Mario Vicente
da Silva (Universidade Nova de Lisboa, Lisbon, Portugal) for their
numerous beneficial suggestions for the development of the PFEM,
as well as Prof. Jose E. Andrade (California Institute of Technol-
ogy, Pasadena, USA), Prof. Yves Leroy (Ecole Normale Superieure,
Paris, France) and Dr. Pierre-Yves Lagree (Universit Pierre et
Marie Curie - Paris 6, Paris, France) for the helpful discussion on the
simulation of granular media.

v
Thanks are also extended to all the staff members at the Geotechni-
cal group of The University of Newcastle for providing a warm and
friendly environment.
Last but not least, I am eternally grateful to my parents Defu Zhang
and Shouzhen Cheng, as well as my fiancee Ya Gao, for their
endless love, support and encouragement.

vi
 Abstract

Despite the wide application of the finite element method (FEM) in

geotechnical engineering, the numerical analysis usually stops at the
point when soil flow occurs and results in overall ‘failure’. In many
cases, the so-called failure only represents a specific time point of the
deformation process and the soil flow itself is of interest as well. A
typical example is a landslide in which a transition of the soil be-
haviour is experienced from solid-like to liquid-like, and then back to
solid-like. For such problems, a correct understanding of the trigger-
ing mechanism is important. However, the prediction of the sliding
process as well as the estimation of the final deposit are also of great
concern. Unfortunately, the traditional Lagrangian FEM cannot han-
dle problems involving both solid-like and liquid-like behaviour. This
is to a large extent, due to the following two issues:
• Severe mesh distortion and boundary evolution as a result of large
changes in geometry.
• Difficulties in solving the highly nonlinear and non-smooth discrete
governing equations in an efficient and robust manner.
In this thesis, a new continuum approach that addresses the above two
issues explicitly is proposed for handling problems involving the solid-
liquid transitional behaviour in geomechanics. More specifically, the
first issue is solved via the so-called Particle Finite Element Method
(PFEM) originally proposed for the solution of fluid dynamics prob-
lems involving free surfaces. The key feature of the PFEM is that
mesh nodes are treated as a cloud of particles which can move freely
and even separate from the domain to which they originally belong.
At each time step, the computational domain is detected based on

vii
those particles; then, the conventional FEM is used to solve the prob-
lem on the identified domain. Regarding the second issue, mathemat-
ical programming formulations for the dynamic analysis of elastoplas-
tic behaviour are developed with a wide utilisation of the Hellinger-
Reissner variational theorem. The resulting formulations can be cast
as a second-order cone program and solved via appropriate optimiza-
tion methods. Unlike the conventional Newton-Raphson based FE
scheme, the convergence of the solution of the scheme developed is
guaranteed regardless of the quality of the initial solution. Formula-
tions for both plane strain and axisymmetric problems are developed.
Moreover, the contact between deformable bodies and rigid bound-
aries is also taken into account.
A number of challenging problems in plane strain cases are solved
successfully, which demonstrates the capabilities of the proposed ap-
proach. Furthermore, the approach is used to reproduce laboratory
tests involving the collapse of axisymmetric granular columns. A
quantitative comparison between the simulated results and the ex-
isting experimental data is conducted. Finally, an actual natural dis-
aster event, the Yangbaodi landslide, is considered and analysed in
some detail.

viii
 Preface

The research work presented in this thesis was conducted in the De-
partment of Civil, Surveying and Environmental Engineering, School
of Engineering, at The University of Newcastle, Australia, under the
supervision of A/Prof. Kristian Krabbenhoft and Prof. Daichao
Sheng from August 2010 to September 2014. During the term of
the candidature, the following papers were published:

1. X. Zhang, K. Krabbenhoft, D. M. Pedroso, A. V. Lyamin, D.

Sheng, M. V. da Silva and D. Wang (2013). Particle finite ele-
ment analysis of large deformation and granular flow problems.
Computers and Geotechnics, 54:133-142.
2. X. Zhang, K. Krabbenhoft and D. Sheng (2014). Particle fi-
nite element analysis of the granular column collapse problem.
Granular Matter, 16:609-619.
3. X. Zhang, K. Krabbenhoft and D. Sheng (2014). Particle finite
element simulation of granular media. Applied Mechanics and
Materials, 553:410-415. (1st Australasian conference on compu-
tational mechanics, Sydney, 2013)
4. X. Zhang, K. Krabbenhoft, D. Sheng and W. Li (2014). Numer-
ical simulation of a flow-like landslide using the particle finite
element method. Computational Mechanics. (Accepted)

ix
x
Contents

Contents xi

List of Figures xv

List of Tables xix

1 Introduction 1
1.1 Motivation and objectives . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Review of research status . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Numerical approaches in geomechanics . . . . . . . . . . . 3
1.2.2 Solution algorithms . . . . . . . . . . . . . . . . . . . . . . 20
1.3 Contributions and outlines . . . . . . . . . . . . . . . . . . . . . . 23

2 Fundamentals of Particle Finite Element Method 26

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2 Particle finite element method . . . . . . . . . . . . . . . . . . . . 27
2.3 Domain detection . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3.1 Alpha-shape method . . . . . . . . . . . . . . . . . . . . . 28
2.3.2 Choice of α . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4 Mesh generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.5 Variable mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

3 Mathematical Programming Formulations: Statics 37

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2 Governing equations . . . . . . . . . . . . . . . . . . . . . . . . . 38

xi
 CONTENTS

3.2.1 Kinematic equations and equilibrium equations . . . . . . 38

3.2.2 Boundary conditions . . . . . . . . . . . . . . . . . . . . . 39
3.2.3 Constitutive equations . . . . . . . . . . . . . . . . . . . . 40
3.3 Variational formulation . . . . . . . . . . . . . . . . . . . . . . . . 42
3.4 Finite element discretisation . . . . . . . . . . . . . . . . . . . . . 43
3.5 Solution algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.6 Preliminary results: strip footings . . . . . . . . . . . . . . . . . . 50
3.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

4 Mathematical Programming Formulations: Dynamics 56

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.2 Time discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.3 Variational formulation . . . . . . . . . . . . . . . . . . . . . . . . 58
4.4 Spatial discretisation . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.5 Karush-Kuhn-Tucker conditions . . . . . . . . . . . . . . . . . . . 62
4.6 Verifications: cantilever beam . . . . . . . . . . . . . . . . . . . . 63
4.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

5 Mathematical Programming Formulations: Contact 67

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.2 Contact scheme in CD method . . . . . . . . . . . . . . . . . . . . 68
5.3 Extension to FEM . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.4 Verifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.4.1 Sliding rigid box . . . . . . . . . . . . . . . . . . . . . . . 72
5.4.2 Deeply buried pipeline . . . . . . . . . . . . . . . . . . . . 72
5.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

6 PFEM for Large Deformation Plane-Strain Problems 76

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.2 Error estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.3 Preliminary results for large deformation analysis . . . . . . . . . 78
6.3.1 An inclined axial rod . . . . . . . . . . . . . . . . . . . . . 78
6.3.2 Trussed frame . . . . . . . . . . . . . . . . . . . . . . . . . 80
6.4 Analyses of plane strain problems . . . . . . . . . . . . . . . . . . 82

xii
 CONTENTS

6.4.1 Cylinder-soil interaction . . . . . . . . . . . . . . . . . . . 83

6.4.2 Accretionary wedge . . . . . . . . . . . . . . . . . . . . . . 87
6.4.3 Collapse of granular columns . . . . . . . . . . . . . . . . . 89
6.4.4 Silo discharge . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

7 Axisymmetric Collapse of Granular Columns 99

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.2 Governing equations . . . . . . . . . . . . . . . . . . . . . . . . . 100
7.2.1 Momentum conservation equations . . . . . . . . . . . . . 101
7.2.2 Strain-displacement relations . . . . . . . . . . . . . . . . . 102
7.2.3 Constitutive relations . . . . . . . . . . . . . . . . . . . . . 103
7.2.4 Frictional contact conditions . . . . . . . . . . . . . . . . . 104
7.3 Discretisation and solution . . . . . . . . . . . . . . . . . . . . . . 106
7.3.1 Time discretisation . . . . . . . . . . . . . . . . . . . . . . 106
7.3.2 Euler-Lagrangian equations . . . . . . . . . . . . . . . . . 107
7.3.3 Spatial discretisation . . . . . . . . . . . . . . . . . . . . . 109
7.4 Numerical simulations . . . . . . . . . . . . . . . . . . . . . . . . 111
7.4.1 Collapse evolution . . . . . . . . . . . . . . . . . . . . . . 112
7.4.2 Details of collapse with large aspect ratio . . . . . . . . . . 114
7.4.3 Normalised final height and radius . . . . . . . . . . . . . 116
7.4.4 Influence of basal roughness . . . . . . . . . . . . . . . . . 119
7.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

8 A Case Study: Yangbaodi Landslide 121

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
8.2 Yangbaodi landslide . . . . . . . . . . . . . . . . . . . . . . . . . . 122
8.3 Model setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
8.4 Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
8.4.1 Calibration of friction coefficient . . . . . . . . . . . . . . . 125
8.4.2 Influence of material density . . . . . . . . . . . . . . . . . 127
8.4.3 Kinematic of typical material points . . . . . . . . . . . . 129
8.4.4 Flow evolution . . . . . . . . . . . . . . . . . . . . . . . . 132

xiii
 CONTENTS

8.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

9 Conclusions 136

A Second-Order Cone Programming 138

B Formulations of Bar Element 142

C Cartesian and Cylindrical Representations 146

C.1 Converting points between Cartesian and cylindrical bases . . . . 146
C.2 Converting vectors between Cartesian and cylindrical bases . . . . 147
C.3 Converting tensors between Cartesian and cylindrical bases . . . . 150

References 151

xiv
List of Figures

1.1 Numerical approaches in geomechanics. . . . . . . . . . . . . . . . 4

1.2 A typical assembly of grains considered in DEM simulations. . . . 5
1.3 Interparticle force model for the classical DEM. . . . . . . . . . . 6
1.4 Computational sequence of the DEM by Bicanic [20]. . . . . . . . 7
1.5 Non-penetration conditions for two particles in the CD method. . 8
1.6 Lagrangian and Eulerian meshes. . . . . . . . . . . . . . . . . . . 9
1.7 Approximation schemes in (a) the FEM, and (b) the Meshfree
Particle Method. . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.8 Weight functions in the SPH. . . . . . . . . . . . . . . . . . . . . 13
1.9 The Moving Least Square (MLS) scheme. . . . . . . . . . . . . . . 15
1.10 Spatial discretisation in the Material Point Method (MPM). . . . 19
1.11 The Newton-Raphson iteration scheme. . . . . . . . . . . . . . . . 22
1.12 The modified Newton-Raphson iteration scheme. . . . . . . . . . . 23

2.1 Steps in the Particle Finite Element Method. . . . . . . . . . . . . 29

2.2 Alpha-shape method. . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.3 Boundary recognition via the scheme of Cremonesi et al. [36]:
cloud of points (a), Delaunay triangulation (b), after deletion of
triangles with the diameter of its circumcircle greater than αh (c). 31
2.4 Domain recognition via the alpha-shape method. . . . . . . . . . . 32
2.5 Variable mapping by Inverse Distance Algorithm (IDA). . . . . . 34
2.6 Variable mapping by Unique Element Method (UEM). . . . . . . 35

3.1 Quadratic displacement/linear stress isotropic triangular element

utilised in the simulation. . . . . . . . . . . . . . . . . . . . . . . 45

xv
 LIST OF FIGURES

3.2 Surface foundations. . . . . . . . . . . . . . . . . . . . . . . . . . 51

3.3 Setup of the strip footing. . . . . . . . . . . . . . . . . . . . . . . 52
3.4 Load-displacement curves for strip footing. . . . . . . . . . . . . . 53
3.5 Spatial discretisation using (a) medium meshes, and (b) fine meshes
at the edge of the foot. . . . . . . . . . . . . . . . . . . . . . . . . 53
3.6 Load-displacement curves for strip footing with different spatial
discretisations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.7 Bearing capacity versus friction angles for strip footing. . . . . . . 54

4.1 A cantilever beam subjected to an end load F. . . . . . . . . . . . 63

4.2 Force history. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.3 Vertical displacements of point A versus time. . . . . . . . . . . . 64
4.4 Simulated vertical displacements with and without numerical damp-
ing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

5.1 Granular contact scheme. . . . . . . . . . . . . . . . . . . . . . . . 68

5.2 Contact specification. . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.3 Moving rigid box on a frictional surface with a total of (a) 10 (b)
20 (c) 40 (d) 60 time steps. . . . . . . . . . . . . . . . . . . . . . 73
5.4 Deeply buried pipeline. . . . . . . . . . . . . . . . . . . . . . . . . 74
5.5 Failure mode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

6.1 A rotated bar. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

6.2 Geometry of a single-axial rod. . . . . . . . . . . . . . . . . . . . 79
6.3 Load deflection behaviour. . . . . . . . . . . . . . . . . . . . . . . 80
6.4 Geometry of a trussed frame (dimensions in mm). . . . . . . . . . 81
6.5 Deformed shapes with the vertical displacement at 20 mm, 40 mm
and 60 mm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.6 Load-displacement curves. . . . . . . . . . . . . . . . . . . . . . . 82
6.7 Interaction of a rigid cylinder with a Tresca soil. Colours are pro-
portional to the equivalent plastic strain increment. . . . . . . . . 84
6.8 Interaction of a rigid cylinder with a Tresca soil. Colours are pro-
portional to the equivalent plastic strain increment (continued). . 85
6.9 Horizontal and vertical resultant forces acting on the cylinder. . . 86

xvi
 LIST OF FIGURES

6.10 Accretionary wedge problem. Colours are proportional to the

equivalent plastic strain increment. . . . . . . . . . . . . . . . . . 88
6.11 Granular column problem: initial column and final deposit. . . . . 89
6.12 Collapse of granular columns (dots represent the finite element
nodes). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.13 Normalised final height and width of granular columns as function
of aspect ratio. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6.14 Convergence of final deposit as function of time step for a = 1
(dots represent the finite element nodes). . . . . . . . . . . . . . . 93
6.15 Volume change as function of time for a = 1. . . . . . . . . . . . . 93
6.16 Silo discharge problem. . . . . . . . . . . . . . . . . . . . . . . . . 94
6.17 Discharge of smoothed walled silo (dots represent the finite element
nodes). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
6.18 Discharge of rough walled silo (dots represent the finite element
nodes). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.19 Silo volume versus time for different time steps. . . . . . . . . . . 97

7.1 Global cylindrical coordinate system. . . . . . . . . . . . . . . . . 100

7.2 Contact specification. . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.3 Quadratic displacement/linear stress element utilised in the simu-
lation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
7.4 Granular column problem: initial column and final deposit. . . . . 112
7.5 Velocity distribution of granular columns with initial aspect ratio
a at time t̄. (a) a = 0.5, t̄ = 1.0 (b) a = 1.0, t̄ = 1.0 (c) a = 5.0,
t̄ = 0.8 (d) a = 5.0, t̄ = 1.2. The grey line corresponds to the initial
configuration while the black line is the final deposit. Colours are
proportional to the magnitude of velocities (cm/s). . . . . . . . . 113
7.6 Comparison of experimental and simulated dynamic data for the
column with a = 3.6 and r0 = 3.9 cm. The dash line corresponds
to the simulated result while the solid line is the experimental result.114

xvii
 LIST OF FIGURES

7.7 Configurations of granular columns with initial aspect a = 8.0 at

time (a) t̄ = 0, (b) t̄ = 1.0, (c) t̄ = 1.2, (d) t̄ = 1.28, (e) t̄ = 2.0, (f)
t̄ = 4.0. Colours are proportional to the magnitude of velocities
(cm/s). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.8 Normalised final radius of granular columns as a function of aspect
ratio. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.9 Normalised final height of granular columns as a function of aspect
ratio. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.10 Normalised final radius and height of deposit profiles as functions
of aspect ratio. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
7.11 Deposit profiles normalised to the initial radius. . . . . . . . . . . 119

8.1 Post event topography (after Li et al. [92]). . . . . . . . . . . . . 122

8.2 Yangbaodi flowslide: (a) topographic map showing the extent of
the flowslide; and (b) slope profiles before and after the flowslide
(after Li et al. [92]). . . . . . . . . . . . . . . . . . . . . . . . . . 123
8.3 Setup for the numerical simulation of the Yangbaodi landslide. . . 125
8.4 Final deposit obtained with friction coefficient (a) µ = tan 8◦ , (b)
µ = tan 10◦ , and (c) µ = tan 12◦ . . . . . . . . . . . . . . . . . . . 126
8.5 The horizontal position of the leading edge versus time. . . . . . . 127
8.6 Final deposit obtained with density (a) ρ = 0.5 t/m3 , (b) ρ =
1.25 t/m3 , and (c) ρ = 2.5 t/m3 . . . . . . . . . . . . . . . . . . . . 128
8.7 Horizontal velocity of material points. Green lines denote the ve-
locity obtained from the DEM simulation in [92], while blue lines
from the developed PFEM simulation. . . . . . . . . . . . . . . . 130
8.8 Vertical velocity of material points. Green lines denote the velocity
obtained from the DEM simulation in [92], while blue lines from
the developed PFEM simulation. . . . . . . . . . . . . . . . . . . 131
8.9 Snapshots of flow evolutions with velocity contour. The unit for
velocities used is m/s. . . . . . . . . . . . . . . . . . . . . . . . . 133
8.10 Snapshots of flow evolutions with velocity contour (continued).
The unit for velocities used is m/s. . . . . . . . . . . . . . . . . . 134

B.1 Truss element. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

xviii
List of Tables

3.1 Parameters for strip footing. . . . . . . . . . . . . . . . . . . . . . 51

6.1 Parameters for the inclined axial rod. . . . . . . . . . . . . . . . . 79

8.1 Physical properties of the soil from [92]. . . . . . . . . . . . . . . 124

xix
Chapter 1

Introduction

1.1 Motivation and objectives

Usually, the problems in geomechanics can be described by means of algebraic,
differential, or integral equations. It is ideal to obtain the closed-form solutions
of these equations; unfortunately, regarding practical engineering problems, those
equations are often considerably complex. In most cases, obtaining their analyti-
cal solutions is impossible. As a consequence, it is inevitable to seek approximate
solutions using various numerical methods.
The finite element method (FEM) is considered a dominant numerical tech-
nique for analysing nonlinear problems in geotechnical engineering. Nonetheless,
the conventional Lagrangian finite element analysis usually stops at the point
when the large strain is emerged, leading to the so-called limit loads that trigger
a failure (undesirable deformation). Typical examples are the collapse load of a
footing on an elastoplastic soil and the limit force (gravity, rainfall, excavation,
etc.) that triggers a landslide. In many cases, the so-called failure only represents
a specific time point of the deformation process and the failure process itself or
the post-failure deformation is of interest as well. These include problems like the
penetration of various devices such as cones and torpedo anchors into the ground;
the interaction, during installation or under operating conditions, of off-shore oil
and gas infrastructures with the seabed; the prediction of sliding processes or the
estimation of hazard zones of landslides, and so on. A common characteristic of

1
 INTRODUCTION

these problems is that the geomaterial undergoes transitions between solid-like

and liquid-like behaviour. In other words, when the critical state is achieved,
the deformation pattern of the geomaterial resembles a fluid flow more than a
solid that leads to extremely large deformations. Despite its dominant role in
the numerical analyses of geotechnical problems, the classical Lagrangian FEM
cannot handle these problems due to two fundamental challenges.
The first one relates to geometry in the sense that the magnitude of the defor-
mations is bound to lead not only to severe mesh distortion, but also to situations
where the boundaries of the problem change from one time step to the next. Of
these two separate but related issues, the former has received by far the most at-
tention. Indeed, for many problems the original boundaries are maintained even
after relatively large distortion. Perhaps the most common approach to avoid-
ing or alleviating mesh distortion is the Arbitrary Eulerian Lagrangian (ALE)
method. This method utilises the respective advantages of purely Eulerian and
purely Lagrangian formulations and has been used quite successfully in geotechni-
cal applications [125, 126, 127] as well as other solid and fluid dynamics problems
[17, 44]. Another popular method for geotechnical applications is the so-called
Remeshing and Interpolation Technique with Small Strain (RITSS) technique
proposed by Hu and Randolph [58, 184, 196]. In the RITSS, the computational
domain is remeshed at each incremental step to avoid the mesh distortion and
the state variables (such as displacements, stresses, strains, etc.) are mapped
from the old mesh to the new mesh. The incremental elastoplastic analysis is
conducted based on the small deformation theory. While both the ALE and the
RITSS have been used to solve problems involving relatively large deformations,
they both have shortcomings in the case where the original boundaries change in
the course of the deformation process. For instance, they fail in the case where
an initially contiguous solid separates into two or more parts as a result of ex-
ternal actions. These two issues have motivated the appearance of a class of
meshfree particle methods [16, 91, 128, 192] that in one way or another circum-
vent the problems stemming from the mesh. However, these methods introduce
complications of their own, both practically and fundamentally, and are from a
mathematical point of view much less established than the finite element method.
The second challenge, which in many ways is the more serious one (though it

2
 CHAPTER 1

remains much less explored), is that of solving the governing equations – compris-
ing momentum balance, strain-displacement relations and constitutive relations
that usually are highly nonlinear and may give rise to ill-posedness, localization
of deformations, etc. The Newton-Raphson scheme, which is commonly used to
minimise the out-of-balance in the framework of the conventional finite element
method, relies crucially on the quality of the initial solution, i.e. the availability
of an initial state that is close to the new state. In cases where large changes
occur over small time steps, for example as a result of localization of deforma-
tions, the performance of the Newton-Raphson scheme decreases markedly and
divergence is often experienced. This is a problem that has a tendency to become
more pronounced as the finite element mesh is refined.
The above described challenges motivate the present work. More specifically,
the objective of this work is to:
• develop a robust numerical method capable of dealing with both solid-like
and liquid-like behaviour of geomaterials,
• develop an efficient solution algorithm capable of handling the highly non-
linear governing equations.

1.2 Review of research status

1.2.1 Numerical approaches in geomechanics
Benefiting from the sustained development of computing power, numerical sim-
ulations have become standard tools in geomechanics and its related fields. To
date, many kinds of numerical approaches have been developed; they have been
reported, more or less, to tackle some stimulating problems. In general, those
numerical approaches can be roughly categorised as shown in Figure 1.1. In this
section, some typical numerical approaches are concerned with their basic idea
being detailed. Notably, it is not the intention of this section to give an exhaustive
survey of the numerical methods for nonlinear analyses in geomechanics. Rather,
it is the intention to introduce some numerical approaches, with special attention
paid to their abilities of tackling problems with extremely large deformations.
In those discontinuous numerical approaches, the Discrete Element Method

3
 INTRODUCTION

Numerical Approaches

Disconnuous Connuous Coupled Connuous-

Approaches Approaches Disconnuous Approaches
(MD, CD, etc.) (FEM-DEM, etc.)

Mesh-based Meshfree Parcle Mesh-based Parcle

Approaches Approaches Approaches
(FEM, etc.) (MPM, PFEM, etc.)

Strong Forms Weak Forms

(SPH, etc.) (EFG, etc.)
Abbreviaons:
MD Molecular Dynamics CD Contact Dynamics
DEM Discrete Element Method FEM Finite Element Method
PFEM Par!cle Finite Element Method MPM Material Point Method
SPH Smoothed Par!cle Hydrodynamics EFG Element Free Galerkin

Figure 1.1: Numerical approaches in geomechanics.

(DEM) can be viewed as a representative that is usually utilised to consider an

assembly of independent elements which are also called particles or grains (see
Figure 1.2). As a matter of fact, DEM is a generic name and there are differ-
ent types of the DEM. In this paragraph and the one follows, two of them will
be introduced. The first one is the so-called classical DEM pioneered by Cun-
dall and Strack in 1979 [40]. This method, often referred to as a Molecular
Dynamics (MD) method, makes use of Newton’s laws of motion to govern the
grains in consideration. Interaction between grains is accounted for by relating
the overlaps that may occur between grains to contact forces via an appropriate
force-displacement relation, a model of which is illustrated in Figure 1.3. The

4
 CHAPTER 1

simulation is typically executed in an explicit manner where the positions of the

particles are first advanced by a suitably small time increment after which the
contact forces are calculated explicitly on the basis of the particle overlaps. This
procedure is repeated until the simulation is achieved. A typical computational
sequence for the classical DEM is shown in Figure 1.4. Since this method is
constructed based on the interaction between individual grains, it is of crucial
importance to correctly evaluate the relations between the interacting forces and
the relative motions of two particles in contact. There are a number of models pro-
posed to account for the normal and tangential contact forces, as well as the rolling
resistance. Those include the models documented in [4, 67, 82, 95, 176, 187, 199]
among others. Heretofore, the application of the DEM has been proven success-
fully for micromechanical analyses [46, 65, 136, 194], as well as the simulations of
large deformation problems, i.e. deep penetrations [69, 70, 178], rock avalanches
[119, 173, 183], and landslides [92, 97, 179].

Figure 1.2: A typical assembly of grains considered in DEM simulations.

The second type of the DEM considered here is the so-called Non-Smooth
Contact Dynamics (CD) Method developed by Moreau, Jean and their co-
workers [68, 121, 122, 123]. In the original version of the CD method, the grains

5
 INTRODUCTION

Spring

Slider
Dashpot Spring

Dashpot

Normal direction Tangential direction

Figure 1.3: Interparticle force model for the classical DEM.

are considered perfectly rigid, which means no inter-particle penetration is per-

mitted (see Figure 1.5). Unlike those in the classical DEM where explicit time
discretisation is utilised, the CD method usually employs the implicit time dis-
cretisation which makes the use of larger time steps possible. Recently, a new
formulation of the CD method, in which the grains can be rigid or even elastic,
was proposed by Krabbenhoft et al. [59, 79, 81]. The new formulation makes
extensive use of variational concepts; this not only ensures the existence and
uniqueness of solutions but also paves the way for the application of efficient
mathematical programming. Even though, the CD method is currently not as
prevalent as the classical DEM, there are several examples of the application of
CD method for the simulation of granular media undergoing large deformation
[48, 85, 145, 147, 174, 175, 177]. Despite that, a crucial shortcoming of both the
classical DEM and the CD method that limits their application lies in the high
computing cost. Indeed, the number of particles considered in typical simulations
is usually far less than the actual number of particles, even though, in principle,
it is possible to account for all the particles within the domain of interest.
In contrast to the discontinuous methods that predict material behaviour
based on the contact interactions between discrete grains, the continuous meth-
ods predict material behaviour by the conservation equations of mass, momentum

6
 CHAPTER 1

Figure 1.4: Computational sequence of the DEM by Bicanic [20].

and energy associated with displacement-strain relations and appropriate consti-

tutive models. According to whether or not a mesh is used, the continuous ap-
proaches may be categorised into two classes: mesh-based methods and meshfree
particle methods. Perhaps the most successful mesh-based method developed in
the last century is the so-called Finite Element Method (FEM) [14, 17, 210],
which by now has become the most popular and widely used numerical method

7
 INTRODUCTION

rj

xj

ri
xi x i − x j ≥ ri + rj

Figure 1.5: Non-penetration conditions for two particles in the CD method.

in geotechnical analysis [143]. For the finite element analysis of large deforma-
tion problems, a fundamentally important consideration is the selection of an
appropriate kinematic description for continuum, i.e. whether a Lagrangian or
an Eulerian description is used, which determines the relationship between the
deforming continuum and the finite mesh of computing zones.
In the Eulerian description, (Eulerian) meshes do not deform with the mate-
rial. That means, regardless of the magnitude of the deformation in the process,
the finite meshes retain their original shapes; consequently, no mesh distortion oc-
curs in the Eulerian description. However, such a characteristic makes it difficult
to track material points and free surfaces. As a result, the Eulerian description
has not been used much for solving problems in solid mechanics, especially those
involving history-dependent materials or free boundaries. Another shortcoming of
the Eulerian description lies in the occurrence of numerical oscillations of the stan-
dard Galerkin formulation in the case that the convective term is dominant which
arise from the difference in the velocity of the mesh (coordinate system) and the
medium. To overcome this limitation, many methods have been proposed, such
as Streamline Upwind/Petrov-Galerkin methods [5, 55], Taylor-Galerkin methods
[45, 186], Characteristic Galerkin methods [94, 142] etc.
In the Lagrangian description, the nodes of (Lagrangian) meshes remain co-
incident with material points. The positions of the mesh nodes are described
in terms of material configuration (original material configuration in the total
Lagrangian description and current material configuration in the updated La-

8
 CHAPTER 1

Lagrangian Mesh

Eulerian Mesh

Figure 1.6: Lagrangian and Eulerian meshes.

grangian description). Figure 1.6 illustrates the difference between Lagrangian

and Eulerian meshes. The Lagrangian description is the most popular one in
solid mechanics stemming from its ease to track complicated boundaries and
follow material points (so that history-dependent materials can be treated accu-
rately). However, the Lagrangian description cannot handle the problem with
very large strain within the body because of the occurrence of excessive mesh
distortion. To overcome the mesh distortion problem, a technique named Ar-
bitrary Lagrangian-Eulerian (ALE) description is proposed to combine the
advantages of both Lagrangian and Eulerian descriptions while avoiding their dis-
advantages [17]. The basic idea of ALE is that the nodes of the finite mesh may
move with the material in normal Lagrangian fashion, or be held in Eulerian man-
ner, or move in some arbitrarily specified way [44]. With this feature, the ALE
method not only performs well for fluid dynamics problems, it is also robust in the
analysis of problems with large strain and distortion in solid mechanics. In the
past few years, the ALE method was used in a wide range of applications such as
simulation of forming and machining processes [61, 130], fluid-structure interac-
tion [168, 189], as well as large deformation problems in geotechnical engineering
[102, 126, 127, 156]. One shortcoming of the ALE method is the appearance
of convective terms in governing equations like the Eulerian method. Some ap-

9
 INTRODUCTION

proaches overcoming the convection-dominant problems have been listed in the

previous paragraph. Besides, the ALE method does not allow new surfaces to be
created and is limited to the case that material deformations are relatively pre-
dictable. Another popular method for large deformation analysis in geotechnical
engineering is the Remeshing and Interpolation Technique with Small
Strain (RITSS) technique proposed by Hu and Randolph [58, 184, 196]. In the
RITSS, Lagrangian meshes are used. The computational domain is remeshed at
each incremental step to avoid the mesh distortion, after which state variables, as
well as history-dependent material parameters, are mapped from the old mesh to
the new mesh. Then, an incremental elastoplastic analysis is conducted based on
the small deformation theory. Such a procedure is looped in the RITSS until the
simulation is achieved. While both the ALE and the RITSS have been used to
solve problems involving relatively large deformations, they both fail in the case
where the original boundaries change in the course of the deformation process.
One example is that of an initially contiguous solid separating into two or more
parts as a result of external actions.
In order to circumvent those issues stemming from the meshes with the pre-
defined connectivity, a number of Meshfree Particle Methods were created
over the past few decades [16, 91, 128]. As shown in Figure 1.7, unlike the clas-
sical Lagrangian-based FEM, where the computational domain is represented by
a number of finite meshes and field variables (i.e. displacements and/or stresses)
at an arbitrary point x are interpolated via the shape functions constructed on
the predefined mesh where the point is located, meshfree particle methods de-
note the computational domain via a cloud of particles and approximate the field
function by shape functions established on a local domain, termed the support
domain of the point. In the meshfree particle methods, the shapes of the sup-
port domains can be different from point to point; the shapes most often used in
two-dimensional cases are circles and rectangles. According to the computational
formulation developed, meshfree particle methods can be roughly classified into
two sub-categories: (1) those that serve as approximations of the strong forms of
the partial differential equations (PDEs), i.e. the Smoothed Particle Hydrody-
namics, the Vortex Method, and the Generalized Finite Difference Method among
others; and (2) those that serve as approximations of the weak forms of the PDEs,

10
 CHAPTER 1

such as the Diffuse Element Method, the Element Free Galerkin Method, the Re-
producing Kernel Particle Method, the Meshless Local Petrov-Galerkin Method
and so on [91].

Support domain

(a) (b)

Figure 1.7: Approximation schemes in (a) the FEM, and (b) the Meshfree Particle
Method.

In the first sub-category of meshfree particle approaches, the Smoothed Par-

ticle Hydrodynamics (SPH) method is viewed as a typical representative. Ac-
tually, the SPH method was originally proposed by Lucy [106] and Gingold and
Monaghan [51] in 1977 for astrophysical problems without boundaries, such as
the formation and evolution of proto-stars or galaxies. The SPH technique con-
structs shape functions via a finite integral representation method, and then the
shape functions are used to interpolate field values in local support domains. The
construction of shape functions in the SPH scheme can be described as follows.
Consider a field function of u(x) at any point x. Its integral representative
can be expressed as Z +∞
u(x) = u(ξ)δ(x − ξ) dξ (1.1)
−∞

where δ(x) is the Dirac delta function. In the SPH scheme, the above function

11
 INTRODUCTION

is approximated by Z
h
u (x) = u(ξ)w(x − ξ, h) dξ (1.2)
Ω

where uh (x) is the approximation of the function u(x), w(x−ξ, h) is the so-called
kernel or weight or smoothing or window function which is usually bell-shaped,
and h is the smoothing length that controls the size of the support domain Ω.
The above integral can then be evaluated via numerical quadrature techniques
which gives
X X mI
uh (x) ≈ w(x − xI , h)uI ∆VI = w(x − xI , h)uI (1.3)
I I
ρI

where xI is the coordinate of the particle I located within the support domain
of point x, and uI is the function value at particle I. VI , mI and ρI denote the
volume, the mass and the density of particle I, respectively. Then Eq. (1.3) can
be re-written as
X
uh (x) ≈ NI (x)uI (1.4)
I

where NI (x) are the shape functions in the SPH method given by

mI
NI (x) = w(x − xI , h) (1.5)
ρI

It should be emphasised that, unlike those in the FEM, the generated shape
functions above do not possess the Kronecker property, which means uI does not
equal uh (xI ). Thus, uI is regarded as a parameter instead of a true value of the
field variable at particle I. The true value of the field variable should be evaluated
based on Eq. (1.4) after uI is determined.
Apparently, the choice of the kernel function plays a significant role on the
above approximation procedure. Some commonly used weight functions are plot-
ted in Figure 1.8 with the equations listed below.

12
 CHAPTER 1

1.2

1 cubic
quartic
exponential
0.8

0.6
w

0.4

0.2

0
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
r

Figure 1.8: Weight functions in the SPH.

• the cubic spline weight function:


2 1
− 4r 2 + 4r 3 , r≤


3 2



4 4 1

w(r) = − 4r + 4r 2 − r 3 , <r≤1 (1.6)


 3 3 2

 0

, r>1

• the quartic spline weight function:


 1 − 6r 2 + 8r 3 − 3r 4

, r≤1
w(r) = (1.7)
 0

, r>1

13
 INTRODUCTION

• the exponential weight function:


 e−(r/α)2

, r≤1
w(r) = (1.8)
 0

, r>1

with
k xI − x k
r= (1.9)
h
In the above exponential weight function, α is a constant equal to 0.3 in Figure
1.8. More details on those kernel functions can be found in [90, 100, 101]. Due to
its distinct advantages of being a meshfree particle method, the SPH method has
not only been extensively used in astrophysics [154, 171], but also widely adopted
in computational fluid dynamics [91, 120]. In geomechanics, Bui et al. [25, 26, 27]
developed a two-dimensional version of the SPH scheme for modelling elastoplas-
tic behaviour of geomaterials. Chen et al. [33] proposed a three-dimensional SPH
model for the study of granular column collapse. Huang et al. applied the SPH
approach for the failure simulation of geo-disasters [60]. Other applications of the
SPH method for simulating large deformation problems in solid mechanics can
be found in [21, 56, 71, 98, 99, 146]. For all this, the SPH method still has not
been viewed as an accurate numerical tool, stemming from the fact that it lacks
a rigorous convergence theory as well as a successive refinement procedure. Some
shortcomings or pathologies of the SPH approach (including tensile instability,
lack of interpolation consistency or completeness, zero-energy mode) and the dif-
ficulty in enforcing essential boundary conditions still impede its application in
practice even though various corrective SPH methods have been developed [91].
The strong-form system equations are very simple, but usually they are less
stable than the weak-form system equations in numerical computations [100].
At the beginning of the 1990s, people turned their attention to develop meshfree
particle methods, which served as approximations of the weak forms of the PDEs.
As a result, the second class of the meshfree particle methods (meshfree Galerkin
methods) emerged. In all those meshfree Galerkin algorithms, either the classical
Moving Least Square (MLS) method or its variants are used to construct the
shape functions. Thus, first of all, the construction of shape functions by means

14
 CHAPTER 1

of the classical MLS is detailed in the following.

Support domain of

Figure 1.9: The Moving Least Square (MLS) scheme.

Suppose we have a point x̄ with its support domain represented by Ω as shown

in Figure 1.9. Let u(x) be the function of the field variable. This function is then
approximated by the MLS method in the support domain of the point x̄ as:

u(x) ≈ uh (x, x̄) = P T (x)c(x̄) (1.10)

where P T (x) is the complete polynomial basis given by

P T (x) = {P0 (x), P1 (x), P2 (x), . . . , Pm (x)} (1.11)

in which m is the number of terms of polynomial basis in use. In a one-dimensional

case, we have
P T (x) = {1, x, x2 , . . . , xm } (1.12)

In two-dimensional and three-dimensional cases, the Pascal triangle and Pascal

pyramid of monomials can be used to build P T (x), i.e. we have

P T (x) = {1, x, y, xy, x2, y 2 . . . , xm , y m } (1.13)

15
 INTRODUCTION

in a two-dimensional case. In a three-dimensional case, we have

P T (x) = {1, x, y, z, xy, yz, zx, x2 , y 2, z 2 . . . , xm , y m , z m } (1.14)

In Eq. (1.10), c(x̄) is a vector of coefficients given by

c(x̄) = {c0 (x̄), c1 (x̄), c2 (x̄), . . . , cm (x̄)} (1.15)

Note that, in the support domains of different points x̄, the vector of coefficients
c(x̄) is usually different. Now, we are in the position to define the functional of
the weighted residual
X
J(c) = w(xI , x̄)[uh (xI , x̄) − u(xI )]2 (1.16)
I∈S h

where S h is the set of indices of particles located within the support domain Ω
of the point x̄ (i.e. the blue particles in Figure 1.9), and w(xI , x̄) is the weight
function. The coefficients c are obtained by minimising the functional J with the
minimisation condition given by

∂J
=0 (1.17)
∂c

Substituting Eq. (1.16) into Eq. (1.17) leads to

!
X X
w(xI , x̄)P (xI )P T (xI ) c(x̄) − w(xI , x̄)P (xI )u(xI ) = 0 (1.18)
I∈S h I∈S h

which can then be re-written as

X
c(x̄) = M −1 w(xI , x̄)P (xI )u(xI ) (1.19)
I∈S h

where
X
M= w(xI , x̄)P (xI )P T (xI ) (1.20)
I∈S h

16
 CHAPTER 1

Then the substitution of Eq. (1.19) back into Eq. (1.10) results in
X
uh (x, x̄) = P T (x)M −1 w(xI , x̄)P (xI )u(xI ) (1.21)
I∈S h

By re-arranging the above equation, we have

X
uh (x, x̄) = NI (x, x̄)u(xI ) (1.22)
I∈S h

with
NI (x, x̄) = w(xI , x̄)P T (x)M −1 P (xI ) (1.23)

being the shape functions for the support domain of point x̄. After deriving
the shape function, the implementation of the meshfree Galerkin method then is
quite similar to the finite element procedure except for the evaluation of the in-
tegrals. In the FEM, the integration is usually calculated based on the Gaussian
quadrature whose solutions are usually of high accuracy, whereas the calculation
of integration in meshfree Galerkin methods is somewhat annoying and usually
evaluated by means of the nodal integration, cell quadrature, or background finite
element mesh quadrature etc. [16]. Nayroles et al. [124] were the first to used
the MLS approximation in a Galerkin weak form and call it the Diffuse Ele-
ment Method (DEM∗ ). Since then, a class of meshfree Galerkin methods have
been invented. Among these methods Element Free Galerkin (EFG) method,
originally developed by Belytschko et al. [15] for predicting crack growth in a lin-
ear elastic solid, is viewed as a representative. Other types of meshfree Galerkin
methods and their applications can be found in [16, 90, 91, 100, 128]. Although,
in contrast to the SPH method, those meshfree Galerkin methods inherit both
the flexibility to interpolate scattered data and the solid mathematical founda-
tion in approximation theory, they have some drawbacks as well. One of them is
the difficulty of enforcing essential boundary conditions due to the fact that the
shape function in those meshfree particle methods does not possess the Kronecker
properties and also does not vanish at the essential boundaries in the Galerkin
formulation. Another one is the inaccuracy in the integration of the weak form,
which may lead to significant error in the solution. In the EFG method, for ex-

17
 INTRODUCTION

ample, to ensure the accuracy of the integration, delicate background cells and
a large number of quadrature points are required, which make the EFG method
much more expensive than the FEM.
Besides these purely mesh-based and meshfree particle methods, there are a
number of approaches that are considered combinations of these two. For such
methods, particles are used to represent the continuum whereas shape functions
are constructed based on finite meshes. Here we call them Mesh-based Par-
ticle Methods. A typical example of those methods is the so-called Mate-
rial Point Method (MPM) which was originally named the Particle-in-Cell
(PIC) method for the simulation of problems in fluid dynamics [54]. As shown in
Figure 1.10, the MPM uses both particles and meshes. More specifically, the con-
tinuum is represented by a cloud of Lagrangian particles, called material points,
which carry all the information of physical properties and state variables (i.e.
mass, material parameters, strains, stresses) as well as external loads, whereas
the fixed Eulerian background meshes carry no permanent information and are
only used to determine incremental displacements and strain increments at the
material points. Thus, at the beginning of each time step, the information should
be transferred from the material points to the background meshes, while at the
end of the time step, the solutions should be mapped from the background meshes
back to the material points. In solid mechanics, the MPM has been used suc-
cessfully for solving a number of extremely large deformation problems, such as
silo discharging, high-velocity impacts, cutting process, crack propagation, etc.
[6, 42, 107, 190, 191, 203]. The application of the MPM in geomechanics includes
[3, 8, 19, 41, 197, 202, 207] where large deformation problems, e.g. landslides,
retaining wall movement, anchor pullout etc., are simulated. Because the MPM
makes use of both background mesh and material point data, one shortcoming is
that it is much more expensive in terms of storage than other continuum methods.
Besides, spurious oscillation may occur while material points cross the boundaries
of the mesh, the effect of which can be alleviated by using some generalised in-
terpolation methods [157].
In the last decade, another mesh-based particle method called the Particle
Finite Element Method (PFEM) was proposed by Onate et al. [131] and is the
one adopted in the present work. The basic idea of this approach is that the mesh

18
 CHAPTER 1

Material point
Background mesh

Figure 1.10: Spatial discretisation in the Material Point Method (MPM).

nodes in the FEM are considered as separated particles which can freely move
and even separate from the main computational domain to which they originally
belong. At each time step, a so-called ‘alpha-shape’ method is employed to recog-
nise the computational domain based on the positions of those particles; then,
Delaunay triangulation is applied to connect the particles to generate meshes.
By doing this, not only can the mesh distortion be avoided, but the free-surface
evolution can also be captured. After the mesh is generated, the PFEM solves the
problems in a standard finite element procedure to enjoy the solid mathematical
foundation of the traditional FEM. Up to date, the PFEM has been successfully
applied for modelling a large number of challenging problems in numerous dis-
ciplines including dam engineering, harbour engineering, industrial engineering,
and so on [64, 131, 132]. In contrast to the wide application of the PFEM in fluid
dynamics, there are few contributions on applying the PFEM in solid mechanics,
especially geomechanics. The exceptions include [30, 129] where the possibilities
of using the PFEM for simulating the riveting process, the machining process,

19
 INTRODUCTION

the powder filling process and the ground excavation are shown. In this thesis,
we focus on developing and applying the PFEM for solving problems involving
extremely large deformations in geomechanics. More details of the PFEM will be
presented in Chapter 2.
Besides the above-mentioned approaches, some researchers devote to develop
coupled discontinuous-continuous approaches [28, 93, 96, 152]. One example
[93, 152] is the coupled FEM-DEM method, in which the DEM is employed
to consider the computational domain undergoing large deformations whereas
the FEM takes care of the domain experiencing relatively small deformations to
reduce the computational cost. Although the robustness and potential of such
coupled methods have been demonstrated by academic exercises, they have not
found their way successfully into general practical applications.

1.2.2 Solution algorithms

In the standard finite element procedure, the solution of elastoplastic boundary
value problems typically proceeds by first obtaining a displacement increment
via solving the discrete weak forms of PDEs at the global structural level. Then,
at the local material point level, the stress increments are calculated via solving
the local constitutive relations with either the fully implicit or explicit integration
schemes. Those typical implicit and explicit integration algorithms for elastoplas-
tic constitutive relations can be found in [23, 24, 135, 162, 163, 165, 167, 206].
After that, an updated stiffness, which possibly deviates from the initially as-
sumed one, may be evaluated based on the updated stress states. The deviation
between the initially assumed and updated stiffness results in a residual in the
form of a global out-of-balance force, which is subsequently sought minimised
by a Newton-Raphson procedure that alternates between the global equilibrium
iterations and local constitutive updates. The Newton-Raphson scheme can be
described as follows.
Let vectors F ext and F int be the externally applied nodal loads and the inter-
nal nodal point forces equivalent to the element stress, respectively. Then, in a
typical time increment [t, t + ∆t], the nonlinear finite element analysis turns into

20
 CHAPTER 1

seeking variables U (i.e. displacements) that fulfil the following equations

R(U ) = F ext int

t+∆t − F t+∆t (U ) = 0 (1.24)

where R represent the residual forces. Suppose we have evaluated U i−1 ; then
R(U ) can be expanded by Taylor series as

∂R(U )
R(U ) = R(U i−1 )+ (U −U i−1 )+higher-order terms = 0 (1.25)
∂U U =U i−1

Substituting Eq. (1.24) into the above equation gives

∂F int
t+∆t (U )
F ext int
t+∆t −F t+∆t (U i−1 )− (U −U i−1 )+ higher-order terms = 0
∂U U =U i−1
(1.26)
Neglecting the higher-order terms, the calculation of displacement increments
∆U i can be conducted as

ext int
∆U i = U i − U i−1 = K −1
i−1 (F t+∆t − F t+∆t (U i−1 )) (1.27)

where
∂F int
t+∆t (U )
K i−1 = (1.28)
∂U U =U i−1

is the tangential stiffness matrix. Then the improved displacement can be ob-
tained by
U i = U i−1 + ∆U i (1.29)

The above algorithm can be illustrated as in Figure 1.11. When large-order sys-
tems are in consideration, the major computational cost in the Newton-Raphson
iteration lies in the estimation of the stiffness matrix K. To decrease the time
consumption in each iteration, a modified Newton-Raphson iteration scheme can
be applied, in which the stiffness matrix K is calculated only at the first iteration
in each time increment [tn , tn+1 ] (see Figure 1.12). This modification indeed saves
time in the calculation of the stiffness matrix, but usually more iterations will
be needed to get convergence. Despite the wide use of both classical and modi-
fied Newton-Raphson schemes in the FEM, their convergence behaviour heavily

21
 INTRODUCTION

depends on whether the initial solution is close to the unknown state or not.

Figure 1.11: The Newton-Raphson iteration scheme.

Alternatively, the elastoplastic finite element problems also can be considered

as extreme problems and solved via those methods established based on mathe-
matical programming. Some examples of solving the elastoplastic problems in op-
timisation manners (e.g. quadratic programming methods, semi-smoothed New-
ton Methods etc.) can be found in [29, 35, 52, 108, 109, 110, 164, 172, 195, 200,
201] and the reference therein. Recently, Krabbenhoft et al. [76, 77, 78] cast the
problem of elastoplastic static analysis as a Second-order Cone Program-
ming (SOCP) problem, with an extensive use of variational principle, which
can then be solved using sophisticated optimisation tools. Unlike the Newton-
Raphson based finite element method that iterates between a local integration
point where the constitutive relations are resolved and a global structure level
where the out-of-balance is minimised, the developed method operates simulta-
neously on all governing equations, local as well as global. One advantage of
this approach is that no special consideration is required for the treatment of
yield surfaces with singularities. Besides that, the convergence behaviour of this

22
 CHAPTER 1

Figure 1.12: The modified Newton-Raphson iteration scheme.

approach is always guaranteed regardless of whether or not the initial state is

close to the new, unknown state. Last but not least, the formulation of Contact
Dynamics has been cast as a SOCP problem (see [59, 79, 81]) as well, which
means both the discontinuous analysis by the CD method and the continuous
elastoplastic analysis by the FEM can be solved using an unified solution tool.
This is in agreement with the modern tendency to unify the analysis tool for
different problems. More details regarding this approach will be presented in the
chapters followed.

1.3 Contributions and outlines

In this thesis, a new scheme is developed to tackle the following issues: (1) se-
vere mesh distortion and free-boundary evolution resulting from extremely large
change in geometry; (2) difficulties in solving the highly nonlinear and non-smooth
discrete governing equations in an efficient and robust manner. More specifi-
cally, issues related to geometry are handled by means of the Particle Finite Ele-

23
 INTRODUCTION

ment Method (PFEM) while the solution of the governing equations is addressed
by means of variational and mathematical programming methods. Other issues
taken into account in the present thesis include the dynamics and the frictional
contact between deformable bodies and rigid surfaces.
The scheme is applicable to general large deformation problems. In other
words, both problems where the deformation patterns resemble fluid flows and
those that merely involve the deformation of solids slightly beyond the small
deformation limit can be handled. A number of challenging problems in geome-
chanics are addressed successfully with the proposed scheme. Furthermore, the
scheme is used to reproduce a laboratory test involving the collapse of axisymmet-
ric granular columns. A quantitative comparison between the simulated results
and the existing experimental data is conducted as well. Finally, an actual nat-
ural disaster event, the Yangbaodi landslide, is considered and analysed in some
detail. The outline of this thesis that reflects the above contributions is as follows.
In Chapter 2, a detailed description of the PFEM is provided. The possibility
of using the PFEM for solving extremely large deformation problems in solid
mechanics is explored. Some critical issues, including domain detection, mesh
generation, and variable remapping, are addressed as well.
In Chapter 3, the mathematical programming formulations for elastoplastic
static problems published in literatures are revisited. This is conducted by first
summarising the governing equations of elastoplastic static problems. Then, the
variational formulations are given followed by the finite element discretisation
in the spatial domain. The equivalence between the mathematical programming
formulations and those in the standard FEM is proven by showing that the Euler-
Lagrange equations associated with the optimisation problem do indeed repro-
duce the discrete governing equations in standard FEM. Numerical tests are also
performed to verify the formulations.
Based on those formulations for static problems described in Chapter 3, math-
ematical programming formulations further accounting for dynamic effects are
developed in Chapter 4. This is achieved by first discretising the momentum
conservation equations in the time domain using the θ-method. Afterwards, vari-
ational formulations are exhibited and discretised in the spatial domain following
the finite element procedure. The correctness of the developed mathematical pro-

24
 CHAPTER 1

gramming formulations is verified through the dynamic analysis of an end-loaded

cantilever beam.
The frictional contact between deformable bodies and rigid boundaries is taken
into account in Chapter 5. Inspired by the approaches for handling the contact in
the DEM-type simulation, a new contact scheme is developed in the finite element
context. The scheme is developed within the general variational framework as
those in the previous chapters; thus, the mathematical programming formulations
proposed for elastoplastic static or dynamic analyses need only minor revision to
account for the frictional contact. Two numerical examples are given to verify
the formulations.
In Chapter 6, the possibility of using a sequence of incremental analysis based
on small-deformation theory for solving large deformation problems is discussed
by considering truss problems. Both analytical and numerical analyses are carried
out. Then, a number of plane strain dynamic problems involving extremely large
deformations are solved successfully. This is achieved by combining the PFEM
scheme described in Chapter 2 and the mathematical programming formulations
proposed in Chapter 4 and Chapter 5.
Chapter 7 concerns the numerical investigation of a laboratory test, the ax-
isymmetric collapse of granular columns. To this end, the mathematical pro-
gramming formulations for dynamic analyses are first of all developed within the
cylindrical coordinate system. Then an axisymmetric version of the PFEM are
proposed to analyse the collapse of granular columns. A quantitative comparison
between the simulated and available experimental results is conducted. The effect
of basal roughness and material properties on the final deposit are also discussed.
A case study is carried out in Chapter 8, where the developed PFEM is
used to analyse an actual natural disaster, Yangbaodi landslide, that occurred in
Southern China.
Conclusions are drawn in Chapter 9.

25
Chapter 2

Fundamentals of Particle Finite

Element Method

2.1 Introduction
Since its invention [131], the Particle Finite Element Method (PFEM) has been
employed to solve numerous challenging problems in fluid dynamics and its related
disciplines. Despite its wide application, the computational details of the PFEM
are rarely given, especially on the boundary detection technique, ‘alpha-shape’
method, which plays a key role in this scheme. Meanwhile, unlike the case in
computational fluid dynamics, there have been few attempts to use the PFEM
for solving problems in solid mechanics. The exceptions include the modelling of
ground excavation by Carbonell et al. [30] and the simulation of forming process
by Oliver et al. [129].
In this chapter, the key features and the fundamental steps of the PFEM for
solving extremely large deformation problems in solid mechanics are described in
detail in Section 2.2. After that, three critical issues are handled. The first one
is the domain detection from a cloud of points, which is addressed in Section 2.3.
The second one is the mesh generation for a detected computational domain in
Section 2.4. The last one is the variable mapping from the old mesh to the new
mesh, which will be tackled in Section 2.5. Conclusions are drawn in Section 2.6.

26
 CHAPTER 2

2.2 Particle finite element method

The particle finite element method (PFEM), as discussed in Chapter 1, is a mesh-
based particle approach. In fact, it can be considered as a combination of the
meshfree particle method and the finite element method. First developed for
fluid dynamics applications [63, 64, 131], the PFEM makes use of a Lagrangian
description to account for the motion of nodes of the finite element mesh, and also
takes the advantage of the meshless method. The key feature of the method is
that nodes are viewed as ’particles’ that can freely move and even separate from
the continuum to which they originally belong. On the basis of the resulting
cloud of points, the continuum and void domains are identified by the so-called
‘alpha-shape’ method [47]; then, a standard finite element discretisation is used to
advance the simulation by a given time step. In this way, the PFEM not only takes
the advantages of the meshfree particle method by using particles to circumvent
mesh distortion but also inherits the solid mathematical foundation of the finite
element method. Over the past decades, the PFEM has entered something of
a ‘Golden Age’ in fluid dynamics community. Its capabilities of tackling the
mesh distortion and following the free surface evolvement have been displayed
through various applications, such as the modelling of free surface flows, breaking
waves, fluid-solid interactions, multi-fluid flows, etc. [63, 64, 131, 132, 133, 134].
Moreover, a number of experiments have been carried out to validate the PFEM
in fluid dynamics [88] as well.
In this section, a general framework of the PFEM is presented for solving
problems in solid mechanics. More specifically, considering a time increment
tn → tn+1 , the fundamental steps of the PFEM in solid mechanics are as follows
(see also Figure 2.1):

0. A cloud of particles, C n , is given at time tn .

1a. On the basis of C n , identify the computational domain, V n .

1b. Mesh the domain and discretise the governing equations on M n .

1c. Map the state variables (velocities, stresses etc.) from the old mesh, M n−1 ,
to the new mesh, M n .

27
 FUNDAMENTALS OF PARTICLE FINITE ELEMENT METHOD

2a. Solve the discrete governing equations to obtain the displacement of the
nodes.

2b. Update the positions of the nodes to arrive at C n+1 and repeat.

It is clear that the procedures of the PFEM in solid mechanics are quite
similar to those in fluid dynamics. The only difference is that, for solving solid
mechanics problems, the mapping of state variables and physical properties from
the old mesh to the new mesh is required once the remeshing operation is finished.
This is due to the fact that material models considered in solid mechanics are
usually history-dependent. To date, numerous schemes have been proposed for
the remapping of states [58, 139, 140, 159, 185, 188, 209], which will be detailed
shortly in this chapter.

2.3 Domain detection

In the above-mentioned steps of the PFEM, a critical issue is the identification of
the computational domain on the basis of a cloud of points. For the general case,
there is no unique solution to this problem. The solution originally proposed by
Idelsohn et al. [64] and subsequently adopted as a standard feature of the PFEM
was to use the alpha-shape method previously developed by Edelsbrunner and
Mucke [47] for computer graphics applications. In this section, the alpha-shape
method is described and the choice of the parameter α is estimated.

2.3.1 Alpha-shape method

In computational geometry, the shape identification can be conducted via the
alpha-shape method. The recognised shape is a linear approximation of the orig-
inal shape [18]. The basic principle of the alpha-shape method is as follows.
Consider a cloud of points with a characteristic spacing h. Then for some prede-
fined value of a parameter α, all nodes on an empty sphere with a radius greater
than αh are considered boundary nodes. In other words: for each point in the
domain, examine whether it is possible to place a sphere with radius αh such
that it contains only that point. If possible, the point is a boundary point and if

28
 CHAPTER 2

Cloud of parcles C n

Computaonal domain V n

Mesh M n

Cloud of parcles C n +1

Computaonal domain V n +1

Mesh M n +1

n+2
Cloud of parcles C

Figure 2.1: Steps in the Particle Finite Element Method.

29
 FUNDAMENTALS OF PARTICLE FINITE ELEMENT METHOD

not (i.e. if the sphere inevitably will contain more than the one point), it is an
internal point. Figure 2.2 shows an example of the boundary detection using the
alpha-shape method.

Figure 2.2: Alpha-shape method.

While a number of algorithms for recognising boundaries by means of the

alpha-shape method are available, another (and in many ways more straight-
forward) possibility has been proposed by Cremonesi et al. [36]. The steps in
this scheme for two-dimensional problems are as follows. Consider the cloud of
particles as shown in Figure 2.3(a). A Delaunay triagulation is first performed
to generate the convex domain shown in Figure 2.3(b). Next, the radius of the
circumcircle of each triangle is examined, and triangles with a circumcircle ra-
dius greater than αh are deleted. For the example at hand, this leads to the
final configuration shown in Figure 2.3(c). As shown by Cremonesi et al. [36],
this procedure is equivalent to the original alpha-shape approach owing to the
property that the circumcircles of all triangles generated by the Delaunay trian-
gulation are empty. It is clear that the implementation of this scheme is rather
simple; more importantly, the extension of this algorithm to three-dimensional
cases is straightforward.
Regarding the above-mentioned algorithm, there are three points that should

30
 CHAPTER 2

Figure 2.3: Boundary recognition via the scheme of Cremonesi et al. [36]: cloud
of points (a), Delaunay triangulation (b), after deletion of triangles with the
diameter of its circumcircle greater than αh (c).

be emphasised. First of all, even though Figure 2.3(c) shows that the final com-
putational domain can be obtained as well as the finite element mesh, the mesh
resulting from the algorithm might be of low quality due to the large distortion
of the continuum. Consequently, a regular remeshing operation for the identi-
fied domain is inevitable. Secondly, the utilisation of sophisticated constitutive
models in geomechanics in addition to ensuring the accuracy of variable mapping
demand the employ of higher-order finite elements. In our work, the six-node
rather than three-node triangular element is utilised for two-dimensional prob-
lems. In such a case, while identifying computational domains, the cloud of points
includes only the vertex mesh nodes with the rest being omitted. Thirdly, using
non-uniform meshes in the simulation is possible, although the above procedure
is for the problem with uniform meshes. When non-uniform meshes are in use,
the characteristic spacing h of the cloud of particles differs from one domain to
another. For instance, a large h is set for the domain discretised by large size
meshes, and vice versa. An example of using non-uniform meshes is shown in
Section 6.4.1.

2.3.2 Choice of α
It is clear that the alpha-shape method involves an element of subjectivity. In-
deed, the resulting domain is a direct function of the value of the parameter α.

31
 FUNDAMENTALS OF PARTICLE FINITE ELEMENT METHOD

Cloud of points α = 1.3

α = 0.5 α = 0.4

α = 1.5 α = 100

Figure 2.4: Domain recognition via the alpha-shape method.

This is illustrated in Figure 2.4. A value of α = 1.3 here produces a set of bound-
aries that in many cases would be deemed reasonable. In fact, any value of α in
the interval 0.9 ≤ α ≤ 1.3 produces this set of boundaries. Decreasing α below
0.9 leads first to internal voids (α = 0.5) and subsequently to a disintegration of
the external boundaries as well (α = 0.4). On the other hand, increasing α above
1.3 leads first to a coalescence of the two distinct sets of points (α = 1.5) and
then, for large values of α, to a solid defined by the convex hull inscribing the

32
 CHAPTER 2

cloud of points.
The conclusion of this example, that a value of α slightly greater than 1 is
appropriate, is consistent with experience from application to actual physical
problems. For example, for a wide range of coupled fluid-solid interaction prob-
lems, Onate et al. [134] conclude that a value of α in the range of 1.3 to 1.5 is
appropriate. For the examples of the present paper that cover static and dynamic
deformation processes of purely cohesive and purely frictional solids, we find the
appropriate range of α to be approximately 1.4 to 1.6. Thus, while the ‘optimal’
value of α is problem dependent, the range of possible values is in practice rather
limited and requires a relatively minor effort to establish by simple trial and error.
Nevertheless, regardless of the choice of α, the total volume (and thereby for
incompressible materials, the total mass) is bound to fluctuate somewhat in the
course of a typical simulation. However, as will be demonstrated via examples,
this error is rather small. In principle, it could be controlled in various ways, for
example by adjusting α in the course of the simulations. At present no general
scheme is available for this purpose although a particular scheme, applicable to
fluid dynamics, has recently been proposed [155].

2.4 Mesh generation

As mentioned, the remeshing of the detected domain should be conducted due
to the extremely large deformation. In the present work, the task of mesh gen-
eration is achieved by an open source code called Triangle [160, 161], which is
a two-dimensional quality mesh generator and Delaunay Triangulator with high
efficiency. More information regarding Triangle can be found on its homepage
(www.cs.cmu.edu/∼quake/triangle.html).

2.5 Variable mapping

It is notable that the remeshing of the detected computational domain is in-
evitable while using the PFEM for solving extremely large deformation problems.
Once a new mesh is generated, the remapping of state variables from the old finite
element mesh to the new one is usually required for solid mechanics problems.

33
 FUNDAMENTALS OF PARTICLE FINITE ELEMENT METHOD

For example, the stress tensor should be transferred from the old Gauss points to
the new Gauss points while material models utilised are history-dependent. Also,
velocities and accelerations should be mapped from the old mesh nodes to the
new mesh nodes while dynamic effect is taken into account. Up to date, a number
of variable remapping schemes have been proposed [139, 140, 159, 188, 209] and
comparisons among them have been conducted [58, 185]. In this section, two
popular remapping schemes are described in some detail.
The first remapping scheme considered here is the so-called Inverse Dis-
tance Algorithm (IDA) [159], the implementation of which is rather straight-
forward. Suppose we have the stress state σ at a collection of old Gauss points
(black points in Figure 2.5(a)). Then the stress state at the new Gauss point x
(the cross in Figure 2.5) can be approximated simply by using the stress states
at its neighbouring old Gauss points (blue points in Figure 2.5(b)) as:

Old Gauss Points

Neighboring Gauss Points
New Gauss Point

(a) (b)
Figure 2.5: Variable mapping by Inverse Distance Algorithm (IDA).

P
wi σ i
σ(x) = P (2.1)
wj

where σ i are the stresses at the ith neighboring Gauss point, and wi is its weight-
ing function defined as
wi = d−c
i (2.2)

34
 CHAPTER 2

in which di is the distance between the ith neighbouring Gauss point and the
new Gauss point, and c is a constant parameter usually taken between 2 and
4, with 3.5 recommended [34, 50]. Alternatively, the approximation can also be
constructed using a polynomial basis function or a radial basis function on the
base of the Moving Least Square method which is now widely used in meshfree
particle approaches to construct shape functions [16, 90, 100].

Old Gauss points used

Old Meshes Old Gauss Points for interpolaon or
extrapolaon

New Gauss Point

(a) (b)

Figure 2.6: Variable mapping by Unique Element Method (UEM).

The second remapping scheme concerned here is the Unique Element Method
(UEM) which is reported to be more stable than the IDA scheme [58]. To make
it simple, suppose the six-node triangular element is used. The transferring of
stresses from the old Gauss point to the new Gauss point by the classical UEM
comprises the following two steps: First we have to identify the old finite element
mesh in which the new Gauss point (red cross in Figure 2.6(a)) is located. Then
the stress at the new Gauss point can be interpolated or extrapolated by the
stresses at the old Gauss points of the identified old finite element mesh (blue
crosses in Figure 2.6(b)), since the stress varies linearly within the six-node trian-
gular element. Rather than transferring stresses directly from the old Gauss point
to the new Gauss point, some improved versions of the UEM prefer recovering
stresses at the old mesh nodes first by the superconvergent patch recovery scheme
[209] or averaging schemes [140]. Then stress mapping is carried out from the

35
 FUNDAMENTALS OF PARTICLE FINITE ELEMENT METHOD

old mesh node to the new Gauss point. However, the computational cost of this
additional step is usually considerable; meanwhile, the improvement regarding
the results is not warranted [58]. Thus, in our simulation, we employ the classical
UEM for the variable mapping. It is worth noting that, for dynamic analysis,
the mapping of velocities and accelerations is also required and can be conducted
similarly, except they are transferred from old mesh nodes to new mesh nodes.

2.6 Conclusions
This chapter explores the possibility of using the PFEM for analysing large de-
formation problems in solid mechanics. The fundamental steps of the PFEM for
solving solid mechanics problems are similar to those for handling fluid dynamics
problems. The only difference is that the variable remapping is required for the
analysis of solid mechanics problems.
Regarding the domain detection, the technique proposed by Cremonesi et al.
[36] is more straightforward to implement. The quality of the identified domain
depends on whether or not an appropriate value of parameter α is chosen. Even
though the optimal value of α is problem-dependent, the range of possible values
is rather limited (i.e. from 1.3 to 1.6).
In the proposed PFEM, an open source code called Triangle is used to generate
new meshes, and the Unique Element Method (UEM) is applied to map the
variables.

36
Chapter 3

Mathematical Programming
Formulations: Statics

3.1 Introduction
The possibility of conducting nonlinear finite element analysis of boundary con-
dition problems via the mathematical programming approaches has been shown
in [29, 35, 76, 77, 78, 108, 109, 110, 164, 164, 195, 200, 201]. Compared with
the traditional Newton-Raphson based finite element analysis, the convergence
behaviour is warranted in the mathematical programming approaches regardless
of whether or not the available initial solution is close to unknown, new states.
Besides, some burdensome problems that plague the traditional finite element
method are solved naturally. One example is the treatment of the singularity of
the Mohr-Coulomb criterion. In the traditional FEM, smooth approximations to
the Mohr-Coulomb are required to remove the singularity at the tip of the surface
where gradients are discontinuous [1, 2, 138, 166, 208], whereas in the mathemat-
ical programming approach, no special consideration is required in connection
with the yield surface singularities when an appropriate method of solution is
chosen [77]. Moreover, recently both the elastoplastic finite element and the con-
tact dynamics formulations are cast as standard second-order cone programming
forms by Krabbenhoft and his co-workers [77, 79, 81], which means the contin-
uum finite element analysis and the discontinuum contact dynamics analysis can

37
 MATHEMATICAL PROGRAMMING FORMULATIONS: STATICS

be conducted using an unified optimisation tool.

In this chapter, the method of casting the elastoplastic static finite element
analysis as a second-order cone program (SOCP) documented in [77] is revisited.
The governing equations of the elastoplastic static analysis are first summarised
in Section 3.2 and their variational formulations are given in Section 3.3. The
spatial discretisation is conducted using the finite element interpolation in Sec-
tion 3.4. Then, Section 3.5 shows the casting of the resulting problem in terms
of a standard form of the second-order cone program, which can then be solved
by an appropriate optimisation engine. At last, numerical examples are given in
Section 3.6 to verify the mathematical programming formulations.

3.2 Governing equations

In this section, the governing equations of rate-independent elastoplastic static
problems are briefly summarised.

3.2.1 Kinematic equations and equilibrium equations

Assuming infinitesimal deformations, the kinematic equations are given by

ε = ∇u (3.1)

where u = (ux , uy , uz )T are the displacements; the strains ε are given by

 ∂ux 
   ∂x 
εx 
 ∂uy 

  
 εy   ∂y 
∂uz

   
 ε   
 =  ∂ux ∂z ∂uy 
 z  
ε= (3.2)

2εxy  
  ∂y + ∂x 


2ε   
 yz   ∂uy ∂uz 
 ∂z + ∂y 
 
2εzx
∂uz ∂ux
 
+
∂x ∂z

38
 CHAPTER 3

and ∇ is the usual linear strain-displacement operator taking the following form:

∂
 
 ∂x 0 0
∂
 
 0 0
 
 ∂y 
∂ 
 

 0 0 
∇=
 ∂z  (3.3)
∂ ∂

0

 ∂x ∂y


∂ ∂ 
 
 0


 ∂y ∂z 
∂ ∂
 
0
∂x ∂z

The general (three-dimensional) differential equations of equilibrium are given by

∇T σ + b = 0, in V (3.4)

where σ = (σx , σy , σz , σxy , σyz , σzx )T are the stresses, b = (bx , by , bz )T are the
body forces stemming for example from self weight, and V is the domain under
consideration.

3.2.2 Boundary conditions

Boundary conditions are of the either essential (geometric) or natural (traction)
type. Essential boundary conditions are prescribed displacements while natural
boundary conditions are prescribed tractions expressed as

N σ = t, on S (3.5)

where S is the boundary, t are the given tractions, and

 
nx 0 0 ny 0 nz
 
N =  0 ny 0 nx nz 0  (3.6)
 
 
0 0 nz 0 ny nx

39
 MATHEMATICAL PROGRAMMING FORMULATIONS: STATICS

with nx , ny , nz being the cosine directions of the outward unit normal vector on
the traction surface.

3.2.3 Constitutive equations

Material constitutive equations give the relations between the stresses σ and
the strains ε. Here, the constitutive relations of rate-independent elastic-perfect
plasticity are briefly summarised for later reference.
The fundamental characteristic of elastoplastic materials is the existence of a
yield criterion that effectively limits the magnitude of the stresses:

F (σ) ≤ 0 (3.7)

where F is the yield function. F (σ) < 0 corresponds to stress states within the
elastic domain, while F (σ) = 0 corresponds to yielding, i.e. the current stress
point is on the yield surface.
Considering only infinitesimal deformations, a standard assumption is that
the total strains can be decomposed additively according to

ε = εe + εp (3.8)

where ε are the total strains, εe are the elastic strains, and εp are the plastic
strains. For an isotropic material whose elastic property is linear, its elastic
strains are related to the stresses via

εe = Cσ (3.9)

where C is the elastic compliance modulus given by

 
1 −ν −ν 0 0 0
 
−ν 1 −ν 0 0 0 
 
1 −ν −ν 1 0 0 0 
C=  (3.10)
 
E 0

 0 0 2(1 + ν) 0 0  
 0 0 0 0 2(1 + ν) 0 
 
0 0 0 0 0 2(1 + ν)

40
 CHAPTER 3

where the constant E and ν are known as Young’s modulus and Poisson’s ratio,
respectively. The corresponding rate form of Eq. (3.9) is given by

ε̇e = Cσ̇ (3.11)

where a superposed dot indicates differentiation with respect to pseudo-time (only

rate-independent processes are considered in this chapter).
The plastic strain rates are often assumed to be derivable from a plastic po-
tential G such that
ε̇p = λ̇∇σ G(σ) (3.12)

which is a compact form of

 
∂G
 ∂σx 
 
   ∂G 
ε̇px  
 ∂σy 
 
 p
 ε̇y 
 ∂G 
 
 
 ε̇p 
 ∂σ 
 
 z
 p  = λ̇  z  (3.13)
ε̇xy   ∂G 
 ∂σxy 
   
ε̇p 
 yz 
 ∂G 
 
p
ε̇zx  
 ∂σyz 
 
 ∂G 
∂σzx

where λ̇ is the plastic multiplier that satisfies the complementarity conditions:

λ̇F (σ) = 0, λ̇ ≥ 0 (3.14)

For associated flow rule, the yield function F and the plastic potential G take the
same form. The above equations may be summarised in the following compact
format:
F (σ) ≤ 0
ε̇ = Cσ̇ + λ̇∇σ G(σ) (3.15)
λ̇F (σ) = 0, λ̇ ≥ 0

41
 MATHEMATICAL PROGRAMMING FORMULATIONS: STATICS

3.3 Variational formulation

The incremental elastoplastic analysis of the above governing equations can be
stated in terms of the following optimisation problem, or variational principle:

min max hσn+1 , ∇(∆u)iV − hb, ∆uiV − ht, ∆uiS

∆u σ n+1
− 12 h∆σ n+1 , C∆σn+1 iV (3.16)
subject to F (σn+1 ) ≤ 0

where we have used the notation

Z
hx, yiA = xT y dA (3.17)
A

The equivalence between the above optimisation problem and the governing equa-
tions follows by showing that the Euler-Lagrange equations associated with (3.16)
do indeed reproduce the governing equations. To this end, first we convert the
yield inequality into an equality by addition of a variable s which is positively
restricted:
F (σ n+1 ) + s = 0 (3.18)

Next, the following functional is defined:

J = hσn+1 , ∇(∆u)iV − hb, ∆uiV − ht, ∆uiS

(3.19)
− 21 h∆σ n+1 , C∆σ n+1 iV + β ln s − ∆λ(F (σ n+1 ) + s)

where β is an arbitrarily small constant and ∆λ is a Lagrange multiplier. The

logarithmic barrier function eliminates the need to make explicit reference to the
fact that s > 0. Next, taking functional derivatives, the following equations
appear:
δJ n T
σ n+1 + b = 0, in V
= ∇
δ∆u N σ n+1 = t, on S
δJ
= ∇(∆u) − C∆σ n+1 − ∆λ∇σ F (σn+1 ) = 0
δσ n+1 (3.20)
δJ
= F (σn+1 ) + s = 0
δ∆λ
δJ β
= − ∆λ = 0 ⇒ s∆λ = β
δs s

42
 CHAPTER 3

For β → 0, these equations are easily verified as being the governing equations
for the static problems at hand. Note that the flow rule in the above equations
is associated, i.e. the flow potential is the same as the yield potential F . This
is a consequence of the variational formulation and is not possible to circumvent
directly. To consider the non-associated flow rule, the approach suggested in [80]
can be used. This approach replaces the original yield function by an ‘effective’
one which, when used as a flow potential, yields the correct plastic strains. So
far this approach has been shown to produce reliable results for a range of static
small deformation problems.

3.4 Finite element discretisation

In this section, the finite element discretisation of the governing equations is con-
sidered. The principle (3.16) is discretised in space by postulating finite element
approximations to the state variables σ and u. Using standard finite element
notation, we have
σ ≈ N σ σ̂ (3.21)

u ≈ N u û (3.22)

where σ̂ and û are the nodal variables, N u and N σ matrices contain the shape
functions for displacements and stresses respectively. Here we restrict our dis-
cussion to plane strain problems; however, it is straightforward to extend the
equations to general three-dimensional cases. In a plane strain case, we have

σx
 
 
 σy  u
 σz  ,
σ=  u= x (3.23)
uy
σxy

and
εx

 εy 
ε = ∇u = 
 εz 
 (3.24)
2εxy

43
 MATHEMATICAL PROGRAMMING FORMULATIONS: STATICS

where
∂
 
 ∂x 0 

 0 ∂ 

∂y 
 
∇=
  (3.25)
 0 0 
 
∂ ∂ 
∂y ∂x
In the simulations, quadratic displacement/linear stress isoparametric trian-
gular elements are utilised. As illustrated in Figure 3.1, the stress interpolation
1 4 1
points are located at (λj−1, λj , λj+1 ) = ( , , ), j = 1, 2, 3, where λj are the
6 6 6
area coordinates. This choice leads stress shape functions in N σ to be continuous
within the elements and discontinuous between elements. The displacements u(x)
are interpolated from the corner and midside nodes, which makes shape functions
in N u be continuous between elements and one polynomial degree higher than
those in N σ . More specifically, the stresses are given by
 
σ̂x1
 
 σ̂ 1 
 y
 1
 σ̂ 
 z
 1 
σ̂ 
   xy 
N1 0 0 0 Nσ2 0 0 0 Nσ3 0 0 0  σ̂x2 
 
 σ  
 0 N1 0 0 0 Nσ2 0 0 0 Nσ3 0 0  2
σ   σ̂y 
σ ≈

 2 
0
 0 Nσ1 0 0 0 Nσ2 0 0 0 Nσ3 0  σ̂z 
 
0 Nσ1 0 0 Nσ2 0 Nσ3 
 2 
0 0 0 0 0 σ̂xy 

 3
 σ̂x 
 
 σ̂ 3 
 y
 3
 σ̂z 
 
3
σ̂xy
= N σ σ̂
(3.26)
where

44
 CHAPTER 3

3 Stress nodes
Displacement nodes

6 5

ξ
1 4 2
Figure 3.1: Quadratic displacement/linear stress isotropic triangular element
utilised in the simulation.
 
Nσ1 0 0 0 Nσ2 0 0 0 Nσ3 0 0 0
 
1
 0 Nσ 0 2 3
0 0 Nσ 0 0 0 Nσ 0 0 
 
Nσ = 
  (3.27)
0 1 2 3
0 Nσ 0 0 0 Nσ 0 0 0 Nσ 0 

 
1 2 3
0 0 0 Nσ 0 0 0 Nσ 0 0 0 Nσ

is the matrix of shape functions for the stress interpolation, and

h iT
σ̂ = σ̂x1 σ̂y1 σ̂z1 σ̂xy
1 2 2 2 2 3 3 3 3
σ̂x σ̂y σ̂z σ̂xy σ̂x σ̂y σ̂z σ̂xy (3.28)

is the vector containing stress states at the stress interpolation points, i.e. Gauss
integration points. The shape functions in Eq. (3.27) are given as below:

2
Nσ1 = 2ξ − ,
6
2 (3.29)
Nσ2 = 2η − ,
6
2
Nσ3 = 2ζ −
6

45
 MATHEMATICAL PROGRAMMING FORMULATIONS: STATICS

where ξ and η are natural coordinates, and ζ = 1 − ξ − η. In contrast, the

displacements u are approximated by higher-order shape functions as below:
 
1
ûx 
 
û1 
 y
 
û2x 

 
Nu1 0 Nu2 0 · · · Nu6 0  
u ≈ 2
ûy 
 
0 Nu1 0 Nu2 · · · 0 Nu6  . 
 (3.30)
 .. 

 
 
û6 
 x
 
û6y

= N u û

where  
1 2 6
Nu 0 Nu 0 · · · Nu 0
Nu =   (3.31)
1 2 6
0 Nu 0 Nu · · · 0 Nu

is a matrix of shape functions for the displacement interpolation, and

h iT
1 1 2 2 6 6
û = ûx ûy ûx ûy · · · ûx ûy (3.32)

is a vector consisting of nodal displacements. The shape functions in (3.31) are

given by
Nu1 = (2ξ − 1)ξ,

Nu2 = (2η − 1)η,

Nu3 = (2ζ − 1)ζ,

(3.33)
Nu4 = 4ξη,

Nu5 = 4ηζ,

Nu6 = 4ζξ

46
 CHAPTER 3

Substituting Eqs. (3.22) and (3.21) into the kinematic equation (3.1) results in

ε = ∇u = B u û (3.34)

where  
1
 ∂Nu ∂Nu2 ∂Nu6
0 0 ··· 0 

 ∂x ∂x ∂x


 
∂Nu1 ∂Nu2 6
 

 0 ∂Nu 
0 ··· 0
 ∂y ∂y ∂y 
Bu = 


 (3.35)
 
 0 0 0 0 ··· 0 0 
 
 
 
 ∂Nu1 ∂Nu1 ∂Nu2 ∂Nu2 ∂Nu6 6
 
∂Nu
···
∂y ∂x ∂y ∂x ∂y ∂x
Since the shape functions Nui (i = 1, . . . , 6) are functions of the natural coordi-
nates ξ and η, implicit differentiation should be used to evaluate the derivatives
in Eq. (3.35). Writing the chains rule for shape functions in a matrix form gives:
   
i i
 ∂Nu   ∂Nu 
 ∂ξ 
 = J  ∂x 
 

    (3.36)
 ∂Nui   ∂Nui 
∂η ∂y

where J is the Jacobian matrix given by

 
x1 y 1 
 
∂Nu1 ∂Nu2 ∂Nu6 
   
∂x ∂y 
 ∂ξ ... x2 y 2 
J = ∂ξ 
 =  ∂ξ
 ∂ξ ∂ξ 

 
 (3.37)
 ∂x ∂y   ∂N 1 ∂Nu2 6
∂Nu  .. .. 
u
... . .

∂η ∂η ∂η ∂η ∂η  
 
6 6
x y

In above, (xi , y i ) are the coordinates of the ith node of the isoparametric element

47
 MATHEMATICAL PROGRAMMING FORMULATIONS: STATICS

in the global coordinate system and

∂Nu1 ∂Nu1
= 4ξ − 1, =0
∂ξ ∂η
∂Nu2 ∂Nu2
= 0, = 4η − 1
∂ξ ∂η
∂Nu3 ∂Nu3
= −3 + 4(ξ + η), = −3 + 4(ξ + η)
∂ξ ∂η (3.38)
∂Nu4 ∂Nu4
= 4η, = 4ξ
∂ξ ∂η
∂Nu5 ∂Nu5
= −4η, = 4 − 4ξ − 8η
∂ξ ∂η
∂Nu6 ∂Nu6
= 4 − 8ξ − 4η, = −4ξ
∂ξ ∂η

By multiplying both side of Eq. (3.36) by the inverse of the Jacobian matrix J,
we obtain    
∂Nui ∂Nui
 = J −1  ∂ξ 
   
 ∂x 
(3.39)
 

 ∂Nui   ∂N i 
u
∂y ∂η
Then the substitution of Eqs. (3.26), (3.30) and (3.34) into the variational prin-
ciple (3.16) leads to the following discrete principle:

min max σ̂ Tn+1 B∆û − f T ∆û − 21 σ̂ Tn+1 C σ̂ n+1

∆û (σ̂,r̂)n+1
(3.40)
subject to F (σ̂ jn+1 ) ≤ 0, j = 1, . . . , nσ

where the yield function is imposed at a finite number of stress integration points
nσ , i.e. Gauss integration points, σ̂ jn+1 represent the stress states at the jth stress
integration point at time t = tn+1 , and
Z
B= N Tσ B u dV (3.41)
V

48
 CHAPTER 3

Z Z
T
f= N u b dV + N Tu t dS (3.42)
V S

Z
C= N Tσ CN σ dV (3.43)
V

The minimisation part of (3.40) may be solved first to yield

B T σ̂ n+1 − f = 0 (3.44)

which is included as a constraint in the remaining maximisation part of the prob-

lem:
maximise − 21 σ̂ Tn+1 C σ̂ n+1
(σ̂,r̂)n+1

subject to B T σ̂ n+1 − f = 0 (3.45)

F (σ̂ jn+1 ) ≤ 0, j = 1, . . . , nσ

This is usually the final problem solved numerically in each time step. Although
it only involves stresses, the displacement increments and plastic multipliers are
recovered as the dual variables, or Lagrange multipliers, associated with the dis-
crete equilibrium constraints which will be shown shortly.
Solution of the optimisation problem (3.45) implies satisfaction of the first-
order Karush-Kuhn-Tucker conditions. These may be seen as the discrete counter-
part to the Euler-Lagrange equations associated with the continuous functional.
Similarly, the Karush-Kuhn-Tucker conditions of problem (3.45) can be derived
by first defining the functional:

1
J = − σ̂ Tn+1 C σ̂ n+1 + ∆ûT (B T σ̂ n+1 − f )
2 (3.46)
+β Σj∈C ln sj − Σj∈C ∆λj (F (σ̂jn+1 ) + sj )

In above, β is an arbitrarily small constant, and C is the set of stress integration

points. ∆λj and sj are the Lagrange multiplier and non-negative variable at
the jth stress integration point, respectively. The logarithmic barrier function
eliminates the need to make explicit reference to the fact that sj > 0. Taking

49
 MATHEMATICAL PROGRAMMING FORMULATIONS: STATICS

functional derivatives of (3.46) gives:

δJ
= B T σ̂ n+1 − f = 0
δ∆û
δJ
= B(∆û) − C∆σ̂ n+1 − ∆λ∇σ̂ F (σ̂n+1 ) = 0
δ σ̂ n+1
(3.47)
δJ
= F (σ̂ jn+1 ) + sj = 0
δ∆λj
δJ β
j
= j − ∆λj = 0 ⇒ sj ∆λj = β
δs s

which are easily identified as the approximate, spatially discrete, counterparts to

the governing equations (3.20).

3.5 Solution algorithm

The optimisation problem (3.45) generated by the scheme in each time step are
solved by means of the second-order cone programming (SOCP) solver MOSEK
[7]. Such solvers have become increasingly more powerful and robust in recent
years and require a computational effort similar to conventional implicit Newton-
Raphson based schemes. The SOCP standard form operates with a linear ob-
jective function [as opposed to the quadratic one in (3.45)] and with inequality
constraints that are slightly different from those of the Mohr-Coulomb/Drucker-
Prager criterion. However, a transformation of (3.45) into SOCP standard form
is straightforward and is documented in Appendix A.

3.6 Preliminary results: strip footings

In practical engineering, the shapes and sizes of surface foundations are often dif-
ferent, but usually they can be categorised as either strip, circular or rectangular
footings as shown in Figure 3.2. While the foundations are subjected to vertical
loading only, a strip footing and a circular footing can be idealised as plane strain

50
 CHAPTER 3

and axisymmetric problems respectively, whereas a rectangular footing remains

a three-dimensional problem. To verify the mathematical programming formula-
tions, here we consider a rigid strip footing placed on the surface of a weightless
undrained clay as shown in Figure 3.3, the left and right sides of the soil column
are restrained in the horizontal direction, while the bottom of the column is fixed.
A total of 107 quadratic displacement/linear stress finite elements are used, and
only half of the domain is accounted for due to the symmetry of the problem.
The soil is assumed to be an elastic-perfectly plastic material with material pa-
rameters given in Table 3.1. The analysis is performed under the displacement
control that the soil surface below the position of the footing is forced to move
vertically.

Strip Circle Rectangle

Figure 3.2: Surface foundations.

Young’s modulus E 100 MPa

Poisson’s ratio γ 0.49
Tresca yield surface
Shear strength Su 100 kPa

Table 3.1: Parameters for strip footing.

The results are shown in Figure 3.4, in which Nc is the mobilised bearing
capacity defined as
F
Nc = (3.48)
Su B
where F is the sum of the vertical reaction forces below the footing. It is clear
that a satisfactory agreement is obtained even for the coarse incremental steps.
Nevertheless, the value of the simulated maximum mobilised bearing capacity is

51
 MATHEMATICAL PROGRAMMING FORMULATIONS: STATICS

B=1m

107 elements

Figure 3.3: Setup of the strip footing.

higher than Prandtl’s solution Nc = π + 2. This is due to the fact that a coarse
mesh is used in the simulation. To ensure accuracy, finer meshes should be used
to discretise the computational domain, especially at the area around the edge
of the footing. The spacial discretisation by medium and fine meshes is shown in
Figure 3.5 with the simulated results given in Figure 3.6 from which we can see
that improved results are obtained.
Next, the soil is considered as a cohesive-frictional material. The setup of the
problem is the same as that in the simulation of the undrained clay, except the
Mohr-Coulomb yield criterion is applied. According to Prandtl’s solution, the
bearing capacity for this problem is calculated by

π φ π tan φ
Nc = (tan2 ( + )e − 1) cot φ (3.49)
4 2

In our simulation, the cohesion is Su = 100 kPa and a series of friction angles
ranging from φ = 5◦ to 45◦ are tested. The comparison between the simulated

52
 CHAPTER 3

Prandtl’s solution
5
Bearing capacity N C

3 5 increments

2 10 increments

1 100 increments

0
0 0.01 0.02 0.03 0.04 0.05 0.06
Footing displacement (m)

Figure 3.4: Load-displacement curves for strip footing.

390 elements 1144 elements

(a) (b)

Figure 3.5: Spatial discretisation using (a) medium meshes, and (b) fine meshes
at the edge of the foot.

53
 MATHEMATICAL PROGRAMMING FORMULATIONS: STATICS

Prandtl’s solution
5
Bearing capacity N C

Coarse mesh
3

2 Medium mesh

1 Fine mesh

0
0 0.01 0.02 0.03 0.04 0.05 0.06
Footing displacement (m)

Figure 3.6: Load-displacement curves for strip footing with different spatial dis-
cretisations.

160

140

120
Prandtl's solu!on
100
Simulated results
Nc

80

60

40

20

0
0 10 20 30 40 50
Friction angle (degree)

Figure 3.7: Bearing capacity versus friction angles for strip footing.

54
 CHAPTER 3

bearing capacity and those from Prandtl’s solution is illustrated in Figure 3.7.
As shown, a satisfactory agreement is achieved.

3.7 Conclusions
The mathematical programming formulations for elastoplastic static problems
are summarised. The governing equations are first given, then their variational
formulations are derived. The finite element discretisation is conducted using the
quadratic displacement/linear stress elements, and the equivalence between the
variational formulation and the governing equations at hand is verified by showing
the Euler-Lagrangian equations associated with the variational principle. It is
clear that the singularity of the yield surface can be tackled naturally, without
any special treatment, in the mathematical programming.
The strip footing problem is simulated using those mathematical program-
ming formulations. The soil in the simulations is considered as purely cohesive
and cohesive-frictional materials, respectively. In both cases, close agreements
between the simulated and Prandtl’s solutions are obtained, even with coarse
increments, which verifies the mathematical programming formulations.

55
Chapter 4

Mathematical Programming
Formulations: Dynamics

4.1 Introduction
In the numerical modelling of many geotechnical problems, the effect of dynam-
ics cannot be neglected, especially when the post-failure process is of significant
concern. One example is the analysis of landslides, in which both the under-
standing of their triggering mechanism and the estimation of the hazard zone
are very important. Perhaps the most general approach for the dynamic analysis
of geo-structural systems is the direct numerical integration of the momentum
conservation equations:
∇T σ + b = ρü, in V
(4.1)
N σ = t, on S
where ∇ is the nabla operator, σ are the stresses, b are the body forces stemming
for example from self weight, ρ is the material density, u are the displacements,
and a superposed dot denotes differentiation with respect to time. V is the
domain under consideration and S is its boundary, t are the tractions, and
 
nx 0 0 ny 0 nz
N =  0 ny 0 nx nz 0  (4.2)
 

0 0 nz 0 ny nx

56
 CHAPTER 4

with nx , ny , nz being the cosine directions of the outward unit normal vector on
the traction surface. The basic idea behind the direct numerical integration is to
satisfy the momentum conservation equations at discrete time points. Based on
those formulations for the static analysis described in Chapter 3, mathematical
programming formulations further accounting for dynamic effects are developed
in this chapter.
The outline of the chapter is as follows. In Section 4.2, the governing equations
are discretised in time using the θ-method. Then the variational formulation of
the dynamic problem is given in Section 4.3. After that, the spatial discretisation
using finite element interpolation is conducted in Section 4.4, while Karush-Kuhn-
Tucker conditions of the discrete variational principle are derived in Section 4.5.
The verification of the proposed mathematical programming formulations is con-
ducted in Section 4.6, and conclusions are drawn in Section 4.7.

4.2 Time discretisation

Introducing velocities, v, the first part of the momentum balance equations may
be written in terms of two coupled first-order equations:

∇T σ + b = ρv̇
(4.3)
v = u̇

These equations are then approximated by means of the θ-method to yield:

v n+1 − v n
∇T [θ1 σ n+1 + (1 − θ1 )σ n ] + b = ρ
∆t (4.4)
un+1 − un
θ2 v n+1 + (1 − θ2 )v n =
∆t

where subscripts n and n + 1 refer to the known and new, unknown states re-
spectively and ∆t = tn+1 − tn is the time step. Rearranging the above equations
gives the following equation in the stresses and the displacements

∆u
∇T σ n+1 + b̃ = ρ̃ (4.5)
∆t2

57
 MATHEMATICAL PROGRAMMING FORMULATIONS: DYNAMICS

where ∆u = un+1 − un and

ρ
ρ̃ =
θ1 θ2 (4.6)
1 v n 1 − θ1 T
b̃ = b + ρ̃ + ∇ σn
θ1 ∆t θ1

After the displacements at tn+1 have been determined, the velocities are computed
from  
1 ∆u
v n+1 = − (1 − θ2 )v n (4.7)
θ2 ∆t
The natural boundary conditions are approximated in an analogous manner lead-
ing to
N [θ1 σ n+1 + (1 − θ1 )σ n ] = t, on S (4.8)

or
N σ n+1 = t̃, on S (4.9)

where
1 1 − θ1
t̃ = t− N σn (4.10)
θ1 θ1
1
The above time integration scheme is unconditionally stable provided that θ1 ≥ 2
and θ2 ≥ 21 . For θ1 = θ2 = 21 , it coincides with the Newmark average acceleration
scheme; while, for θ1 = 0, a fully explicit and conditionally stable scheme is
obtained. In the following, we will only consider implicit and unconditionally
stable schemes. Regarding the constitutive model in Section 3.2.3, we require
that all relations are satisfied at time tn+1 :

F (σn+1 ) ≤ 0
∇(∆u) = C∆σ n+1 + ∆λ∇σ G(σ n+1 ) (4.11)
∆λF (σn+1 ) = 0, ∆λ ≥ 0

4.3 Variational formulation

We are now in a position to postulate that the above time-discrete governing
equations in Section 4.2 can be stated in terms of the following time-discrete

58
 CHAPTER 4

optimisation problem, or variational principle:

min max hσ n+1 , ∇(∆u)iV + 21 h∆σ n+1 C∆σ n+1 iV − hb̃, ∆uiV
∆u (σ,r)n+1
−ht̃, ∆uiS − 12 ∆t2 hrn+1 , ρ̃−1 r n+1 iV + hrn+1 , ∆uiV (4.12)
subject to F (σ n+1 ) ≤ 0

where r n+1 are new variables whose significance will become apparent shortly.
Likewise, to show the equivalence between the above optimisation problem
and the governing equations, first we convert the yield inequality into an equality
by addition of a variable s which is positively restricted:

F (σ n+1 ) + s = 0 (4.13)

Next, we define the following functional:

J = hσ n+1 , ∇(∆u)iV + 21 h∆σ n+1 , C∆σ n+1 iV − hb̃, ∆uiV − ht̃, ∆uiS
− 21 ∆t2 hrn+1 , ρ̃−1 rn+1 iV + hr n+1 , ∆uiV
+β ln s − ∆λ(F (σ n+1 ) + s)
(4.14)
where ∆λ is the Lagrange multiplier. Taking functional derivatives gives the
following equations:


δJ  ∇T σ n+1 + b̃ = r n+1 , in V

=
δ∆u 
 N σ n+1 = t̃, on S


δJ
= −∆t2 ρ̃−1 r n+1 + ∆u = 0
δrn+1 (4.15)
δJ
= ∇(∆u) − C∆σ n+1 − ∆λ∇σ F (σ n+1 ) = 0
δσ n+1
δJ
= F (σ n+1 ) + s = 0
δ∆λ
δJ β
= − ∆λ = 0 ⇒ s∆λ = β
δs s

For β → 0, these equations are easily verified as being the governing equations for

59
 MATHEMATICAL PROGRAMMING FORMULATIONS: DYNAMICS

the elastoplastic dynamic analysis at hand. From the second set of conditions,
we see that the variables r n+1 are the dynamic forces r n+1 = ρ̃∆u/∆t2 . The
momentum balance equations are thus verified by the substitution of the second
set of conditions into the first set.

4.4 Spatial discretisation

The spatial discretisation can be carried out following the procedure in Section
3.4. In addition to the stresses σ(x) ≈ N σ (x)σ̂ and the displacements u(x) ≈
N u (x)û, the problem (4.12) also contains a new set of variables r, which are
approximated by
r(x) ≈ N r (x)r̂ (4.16)

where r̂ are the nodal variables given by

h iT
r̂ = r̂x1 r̂y1 r̂x2 r̂y2 · · · r̂x6 r̂y6 (4.17)

and N r are the new set of finite element shape functions for the dynamic forces
r, which are taken to be identical to those of the displacements but are discon-
tinuouse between elements. Thus, we have
" #
Nr1 0 Nr2 0 ... Nr6 0
Nr = (4.18)
0 Nr1 0 Nr2 ... 0 Nr6

where
Nr1 = (2ξ − 1)ξ,

Nr2 = (2η − 1)η,

Nr3 = (2ζ − 1)ζ,

(4.19)
Nr4 = 4ξη,

Nr5 = 4ηζ,

Nr6 = 4ζξ

60
 CHAPTER 4

Substituting Eq. (4.18) and the previously defined shape functions (3.27) and
(3.31) for stresses and displacements into the variational principle (4.12), we
obtain
min max σ̂ Tn+1 B∆û − f T ∆û − 12 σ̂ Tn+1 C σ̂ n+1
∆û (σ̂,r̂)n+1

− 21 ∆t2 r̂ Tn+1 Dr̂ n+1 + r̂ Tn+1 A∆û (4.20)

subject to F (σ̂jn+1 ) ≤ 0, j = 1, . . . , nσ

where the new matrices D and A are given by

R
D= V
N Tr ρ̃−1 N σ dV
(4.21)
R T
A= V
N u N r dV

The minimisation part of problem (4.20) may be solved first to yield

B T σ̂ n+1 + AT r̂ n+1 − f = 0 (4.22)

which is included as a constraint in the remaining maximisation part of the prob-

lem:
maximise − 21 σ̂ Tn+1 C σ̂ n+1 − 21 ∆t2 r̂ Tn+1 Dr̂ n+1
(σ̂,r̂)n+1

subject to B T σ̂ n+1 + AT r̂ n+1 = f (4.23)

F (σ̂jn+1 ) ≤ 0, j = 1, . . . , nσ

The above problem is the one solved at each time increment for the dynamic
analysis. In the course of solving the problem, the displacement increments ap-
pear as the dual variables, or Lagrange multipliers, associated with the discrete
equilibrium constraints. The above problem is similar in structure to those aris-
ing from limit analysis [75], static elastoplasticity [76, 77, 78, 80], and granular
contact dynamics [59, 79, 81].

61
 MATHEMATICAL PROGRAMMING FORMULATIONS: DYNAMICS

4.5 Karush-Kuhn-Tucker conditions

Similarly, the Karush-Kuhn-Tucker conditions of optimisation problem (4.23) can
be derived by first defining the functional:

1 1
J = − σ̂ Tn+1 C σ̂ n+1 − ∆t2 r̂ Tn+1 Dr̂ n+1 + ∆ûT (B T σ̂ n+1 + AT r̂ n+1 − f )
2 2
+β Σj∈C ln sj − Σj∈C ∆λj (F (σ̂jn+1 ) + sj )
(4.24)
Same as in Section 4.3, β is an arbitrarily small constant, C is the set of stress
integration points, and ∆λj and sj are respectively the Lagrange multiplier and
the positively restrained variable at the jth stress integration point, i.e. the Gauss
integration point. The logarithmic barrier function eliminates the need to make
explicit reference to the fact that sj > 0. Taking functional derivatives of (4.24)
gives:
δJ
= B T σ̂ n+1 − AT r̂ n+1 − f = 0
δ∆û
δJ
= −∆t2 Dr̂n+1 + A∆û = 0
δr̂ n+1
δJ (4.25)
= B(∆û) − C∆σ̂ n+1 − ∆λ∇σ̂ F (σ̂n+1 ) = 0
δ σ̂ n+1
δJ
j
= F (σ̂ jn+1 ) + sj = 0
δ∆λ
δJ β
j
= j − ∆λj = 0 ⇒ sj ∆λj = β, j ∈ C
δs s
It is worth noting that the discrete momentum conservation equation in terms of
stresses and displacements are recovered:

1
B T σ̂ n+1 + 2
[AT D −1 A]∆û = f (4.26)
∆t

by combining the first two optimality conditions in (4.25). While the shape
functions for the dynamic forces r are taken to be identical as those for the
displacements, the matrix AT D −1 A takes the following form:
Z
T −1
A D A= N Tu ρ̃N u dV (4.27)
V

62
 CHAPTER 4

which is equivalent to the mass matrix in standard finite element formulations.

This result is independent of whether r are considered to be continuous or discon-
tinuous between elements. However, before solving problems of the type (4.23),
the quadratic terms in the objective function are often accounted for by imposing
equivalent quadratic inequalities as documented in Appendix A. For this purpose,
the matrix square root of D is required; and a discontinuous approximation of r,
which makes D block-diagonal, is then much more convenient although it leads
to a higher number of variables in the final problem.

4.6 Verifications: cantilever beam

To verify the proposed mathematical programming formulations, the dynamic
response of an end-loaded elastic cantilever beam is analysed. The beam is 10 m
in length and 1 m in width. Its domain is discretised using 500 quadratic dis-
placement/linear stress elements as shown in Figure 4.1, while the load history
at point A is illustrated in Figure 4.2. In the first stage, the force increases to
10 kN from zero in 1 second, and then the value of the force is maintained for 6
second. After that, the force reduces linearly to zero in 1 second. The material
parameters of the beam are as follows: density ρ = 0.01 t/m3 , Young’s modulus
E = 3 × 105 kN/m2 , and Poisson’s ratio γ = 0.3. The gravitational acceleration
is set to be zero, which means the weight of the beam is neglected.

Point A
F

Figure 4.1: A cantilever beam subjected to an end load F.

1
Firstly, the simulation is carried out with θ1 = θ2 = to eliminate the
2
numerical damping effect. The time step used is ∆t = 0.1 s. The changes of

63
 MATHEMATICAL PROGRAMMING FORMULATIONS: DYNAMICS

12

10

8
Force F (kN)

0
0 2 4 6 8 10 12 14 16
Time (s)

Figure 4.2: Force history.

0.16

Quasi-static solution
Vertical displacment of point A (m)

0.12

Abaqus solu!on
0.08
Mathema!cal
programming solu!on
0.04

0
0 2 4 6 8 10 12 14
Time (s)
-0.04

Figure 4.3: Vertical displacements of point A versus time.

64
 CHAPTER 4

vertical displacements at point A against time are shown in Figure 4.3 where
the comparison between the results obtained from the proposed mathematical
programming formulations and the commercial finite element analysis software,
Abaqus, is illustrated. The dash line represents the quasi-static solution with the
end load F = 10 kN. As shown in Figure 4.3, a satisfactory agreement is obtained.

0.16

Quasi-static solution
Vertical displacment of point A (m)

0.12

With damping
0.08

Without damping

0.04

0
0 2 4 6 8 10 12 14
Time (s)
-0.04

Figure 4.4: Simulated vertical displacements with and without numerical damp-
ing.

1
Secondly, we simulate the problem with θ1 = 1 and θ2 = to consider the
2
numerical damping effect. As expected (see Figure 4.4), point A first reaches the
highest position in the vertical direction because of the imposition of the force
at the end of the beam. Then the energy of the system is dissipated due to the
numerical damping, and point A stops exactly at the position obtained from the
quasi-static analysis. While the force is removed, the beam begins to vibrate and
finally stops at its initial position, again, due to the numerical damping.

65
 MATHEMATICAL PROGRAMMING FORMULATIONS: DYNAMICS

4.7 Conclusions
Based on the variational principle, the mathematical programming formulations
are extended further for considering dynamic problems in this chapter. The θ-
method is used to discretise the momentum conservation equations in the time
domain. The proposed time integration scheme is unconditionally stable provided
1 1
that θ1 ≥ and θ2 ≥ , while for θ1 = 0 it turns to a fully explicit and
2 2
conditionally stable scheme.
The finite element method is then applied to discretise the governing equa-
tions in the spatial domain. In addition to the stress and the displacement, a
new set of variables r, the physical meaning of which is the dynamic force, ap-
pear in the variational principle and need to be approximated by shape functions
as well. The dynamic forces r are assumed to be discontinuous between the fi-
nite element aiming to make the converting of quadratic terms in the objective
function to equivalent quadratic inequalities more convenient. Equivalence be-
tween the proposed formulations and those in the traditional FEM are proved.
The dynamic analysis of a cantilever beam is carried out to verify the proposed
formulations.

66
Chapter 5

Mathematical Programming
Formulations: Contact

5.1 Introduction
A proper treatment of the contact is an important component in many geotechni-
cal problems. One example is that the grade of the roughness of the basal surface
is of crucial significance to the granular collapse problems [86, 87]. Another one
is that different friction conditions lead to different failure modes in the problem
of accretionary wedges [170].
Inspired by the recently proposed scheme utilised in the granular contact
dynamic analysis [59, 79, 81], a new strategy is developed in the finite element
context for analysing the frictional contact between deformable bodies and rigid
boundaries in this chapter. Using such a strategy, the mathematical programming
formulations developed for static or dynamic analysis need only minor revision
to account for the frictional boundaries; moreover, it is straightforward to extend
this strategy to consider the contact between rigid and deformable bodies.
The outline of this chapter is as follows. Section 5.2 presents the contact
scheme utilised in the granular contact analysis, while its extension to finite el-
ement analysis is shown in Section 5.3. Two numerical examples are given to
verify the proposed contact scheme in Section 5.4, and conclusions are drawn in
Section 5.5.

67
 MATHEMATICAL PROGRAMMING FORMULATIONS: CONTACT

5.2 Contact scheme in CD method

For simplicity, we consider the system consisting of one rigid disc that has poten-
tial to contact with a rigid boundary as shown in Figure 5.1. Neglecting rotation
effects, the equations of motion for the rigid disc can be written as

n
q

p
g0
x2

x1

Figure 5.1: Granular contact scheme.

mv̇ = f ext
(5.1)
v = u̇

in which m is the mass, f ext = (fx , fy )T are the external forces, v = (vx , vy )T are
the velocities, and u = (ux , uy )T are the displacements. Using the θ-method, the
above equations can be discretised as below:

v n+1 − v n
m = f ext
∆t (5.2)
un+1 − un
θv n+1 + (1 − θ)v n =
∆t

where 0 ≤ θ ≤ 1, ∆t = tn+1 − tn is the time step, and subscripts n and n + 1

refer to the known and new, unknown states respectively. Rearranging the above
equations gives
m̄
∆u = f̄
∆t2 (5.3)
1 ∆u
v n+1 = ( − (1 − θ)v n )
θ ∆t

68
 CHAPTER 5

where
m
m̄ =
θ (5.4)
m̄
f̄ = f ext + vn
∆t
and ∆u = un+1 − un are the displacement increments.
Suppose that the disc cannot penetrate into the frictional rigid boundary;
then, the contact conditions for such a problem are stated as follows:

nT ∆u − g0 ≤ 0
p(nT ∆u − g0 ) = 0 (5.5)
|q| − µp ≤ 0

where n is the outward normal vector of the boundary, g0 is the gap between
the disc and the boundary at time t = tn as shown in Figure 5.1, p is normal
friction force, q is tangential friction force, and µ is the friction coefficient. Now,
the variational formulations for such a system can be written as

1
min max { ∆uT m̄∆u − ∆uT f̄ } + {∆uT (np + n̂q) − g0 p}
∆u (p,q)n+1 2∆t2 (5.6)
subject to |q| − µp ≤ 0

Same as in Section 4.3, the equivalence between the above optimisation problem
and the governing equations (i.e. Eqs. (5.3) and (5.5) ) can be proven by showing
the first-order Karush-Kuhn-Tucker optimality conditions associated with the
optimisation problem (5.6), the detail of which has been documented in [81].
Alternatively, the variational principle (5.6) can be cast as below:

∆t2 T
min max {− r m̄−1 r n+1 + ∆uT r n+1 + ∆uT f̄ } + {∆uT (np + n̂q) − g0 p}
∆u (p,q,r)n+1 2 n+1
subject to |q| − µp ≤ 0
(5.7)
by the introduction of new variables

∆u
r n+1 = m̄ (5.8)
∆t2

69
 MATHEMATICAL PROGRAMMING FORMULATIONS: CONTACT

which can be interpreted as dynamic forces. Solving the minimisation part of

problem (5.6) gives

∆t2 T
maximise − r m̄−1 r n+1 − g0 p
(p,q,r)n+1 2 n+1
(5.9)
subject to r n+1 + ∆uT (np + n̂q) = −∆uT f̄
|q| − µp ≤ 0

As shown, the above optimisation problem is quite similar in structure to (4.23)

which is for the dynamic analysis.

5.3 Extension to FEM

Although originally conceived for the contact dynamics of discrete element simu-
lations, the basic principles described in the above section are equally applicable
in a finite element context. The procedure is as follows.
Consider a body discretised by finite elements as shown in Figure 5.2. Initially,
each boundary node of the system is viewed as a weightless particle with zero
radius. Next, the particles that are likely to come into contact with the rigid
surface (in Figure 5.2 taken to be a single segment) are identified. Consider
one of those particles (i.e. identified mesh nodes) which has a distance of g0
from the boundary and with potential contact forces p and q as shown in Figure
5.2. Following the procedure in Section 5.2, a variational formulation of the
corresponding frictional contact problem is given by:

maximise −g0 p
(5.10)
subject to |q| − µp ≤ 0

Compared with principle (5.9) for the CD method, the above optimisation princi-
ple includes neither the term that represents the dynamic energy of the particle in
the objective function nor the equality constrain that corresponds to momentum
conservation of the particle. This is due to the assumption that the particle is
weightless and its radius is zero. The friction forces p and q will be added to the
discretised equation of momentum conservation for the deformable body latter.

70
 CHAPTER 5

ρy

ρx
n

q
x2
p
g0 Potential contact
x1 particles

Figure 5.2: Contact specification.

The principle (5.10) implies the following kinematics:

∆uN = g0 − µλ
(5.11)
∆uT = ±λ

where λ ≥ 0 is a Lagrange multiplier such that λ(|q| −µp) = 0 and ∆uN and ∆uT
are the normal and tangential displacement increments of the particle respectively.
This means that the sliding rule is associated in the sense that the sliding is
accompanied by a dilation of relative magnitude λ. As in the continuum case,
this dilation is not desirable and can rarely be justified physically. However, as
discussed in [81], the dilation arising at frictional contacts may be viewed as an
artifact of the time discretisation that gradually decreases as the time step is
reduced, which will be shown by the simulation of a simple sliding rigid box on
a rough surface in Section 5.4.
Considering all potential contacts and incorporating the above contact prob-
lem into the dynamic problem (4.23) leads to a final problem of the type
Pnc
maximise − 12 σ̂ Tn+1 C σ̂ n+1 − 21 ∆t2 r̂Tn+1 Dr̂ n+1 − j=1 g0,j pj
(σ̂,r̂)n+1

subject to B T σ̂ n+1 + AT r̂ n+1 + E T ρ = f

F (σ̂ in+1 ) ≤ 0, i = 1, . . . , nσ (5.12)
pj = −nTj ρj , j = 1, . . . nC
qj = −n̂Tj ρj
|qj | − µpj ≤ 0

71
 MATHEMATICAL PROGRAMMING FORMULATIONS: CONTACT

with ρ = (ρ1 , ρ2 )T being the nodal forces, n = (n1 , n2 )T the normal of the rigid
boundary and n̂ = (−n2 , n1 )T . Furthermore, E is an index matrix of zeros and
ones and nC is the number of potential contacts. Similarly, the problem 5.12 can
be transferred into a standard SOCP form following Appendix A and then solved
via an available optimisation tool.

5.4 Verifications
5.4.1 Sliding rigid box
Generally, in the context of frictional problems, the principle of maximum plastic
dissipation is problematic in that it predicts a dilation proportional to the friction
coefficient while, for real materials, the dilation is significantly smaller and often
zero. However, for the present application of time-discrete contact dynamics,
the issue is much less pronounced and the dilation that does take place will be
eliminated in the limit of the time step tending to zero, which can be illustrated
in the following test.
Suppose we have a rigid box which is located on a rigid frictional surface and
travels with an initial horizontal velocity v0 = 0.98 m/s. The friction coefficient
of the surface is µ = 0.1. The results for a total of 10, 20, 40, and 60 time steps
(corresponding to ∆t = 0.1 s, 0.05 s, 0.025 s, and 0.0167 s) are shown in Figure
5.3. Here, we observe that the dilation tends to be zero as the time step is
decreased and the final travelling distance converges to 0.49 m, which coincides
with the result obtained from the Newton’s law of motion. In other words, despite
the physical separation between the rigid box and the surface, contact forces
still exist and convergences can be obtained by decreasing the magnitude of the
incremental time steps. This rather surprising property has previously been noted
and utilised by Anitescu, Tasora and their co-workers [10, 11, 12, 141, 180, 181].

5.4.2 Deeply buried pipeline

As the last example in this section, a pipeline penetration problem is focused on.
Due to the symmetry, half of the domain is considered, as shown in Figure 5.4.

72
 CHAPTER 5

(a)

(b)

(c)

(d)

Figure 5.3: Moving rigid box on a frictional surface with a total of (a) 10 (b) 20
(c) 40 (d) 60 time steps.

The radius of the pipeline is r = 0.5 m, and the geometry of the soil column is
10 m×15 m. The pipeline is considered as a rigid body with a smooth surface,
while the soil is represented by a rigid-perfectly plastic model with the Tresca
yield criterion. The material parameters are as follows: density ρ = 2 t/m3 , and
shear strength Su = 50 kPa. The gravitational acceleration is g = −9.8 m/s2 .
According to [114, 149], the upper and lower bounds for the normalised average
F
pressure Nc = (F is the resistance force on the pile) are 9.20 and 9.14,
rSu
respectively.
The above-mentioned problem is simulated using the developed mathematical
programming formulation. In the simulation, the parameters for the dynamic
analysis are θ1 = θ2 = 1. A total of 2812 quadratic displacement/linear stress
elements are utilised, the distribution of which is shown in Figure 5.4. The
normalised average pressure from the simulation is 9.29 which is 1.64% higher

73
MATHEMATICAL PROGRAMMING FORMULATIONS: CONTACT

Figure 5.4: Deeply buried pipeline.

Figure 5.5: Failure mode.

74
 CHAPTER 5

than the lower bound and 0.98% higher than the upper bound. The contour of the
equivalent plastic strain is illustrated in Figure 5.5. As shown, a localised failure
mechanism is experienced for the deeply buried pipeline. A more challenging
problem regarding the pipeline-soil interaction will be considered in the next
chapter.

5.5 Conclusions
In this chapter, the frictional contact between deformable bodies and rigid bound-
aries is taken into account within the general variational framework. The resulting
formulations are similar to those developed for static and dynamic analyses; thus,
they can be solved conveniently by using an unified solution engine.
Even though an artificial dilation is associated with the proposed contact
scheme while the sliding occurs, its magnitude can be minimised by decreasing
the time increment. Two numerical examples are given to show the correctness
of the proposed scheme.

75
Chapter 6

PFEM for Large Deformation

Plane-Strain Problems

6.1 Introduction
In this chapter, a number of challenging problems involving extremely large de-
formations are handled by the PFEM described in Chapter 2, associated with the
mathematical programming formulations for developed in Chapters 4 and 5.
Apparently, the mathematical programming formulations in Chapters 3 and
4 are developed based on the small deformation theory. Such formulations may
lead to several errors in large deformation analysis of which the generation of
strains as a result of rigid body motions is the most serious one. However, we
have found that this and related errors, for all practical purposes, are negligible
for the kind of time steps used in typical simulations. As such, the price to pay for
the convenience of being able to operate with the usual small deformation theory
appears to be very small. In fact, such large deformation analysis by a sequence
of incremental analysis based on the small deformation theory is quite normal.
One representative is the previously mentioned RITSS method [58, 184, 196],
which is rather popular in geomechanics. Another way that is even cheaper is
the so-called sequential limit analysis [31, 74, 89, 148, 193], which is commonly
used for structural analysis as well as the modelling of metal plasticity.
In this chapter, the error estimation of using the incremental small defor-

76
 CHAPTER 6

mation theory for large deformation analysis is first conducted analytically with
the focus on the rotation of a bar in Section 6.2. Section 6.3 then presents the
large deformation analysis of two truss systems with the small deformation the-
ory, in which neither remeshing nor variable remapping is required; thus, their
influences on the results are eliminated. The first example is an inclined axial
rod in which the simulated result is compared to the analytical one, while the
second one is a truss frame in which the comparison between the results from
the small deformation and large deformation theories is conducted. After that,
four numerical examples of plane strain problems are explored in Section 6.4 to
verify the proposed PFEM scheme and demonstrate its capabilities of tackling
extremely large deformation problems. Those examples include cylinder-soil in-
teraction, accretionary wedge, collapse of granular columns and discharge of silo.
Finally, conclusions are given in Section 6.5.

6.2 Error estimation

In this section, the error estimation of using incremental small deformation theory
for large deformation analysis is carried out. To this end, we consider a purely
rotated bar as shown in Figure 6.1. The bar with an initial length L0 is originally
oriented horizontally. Suppose it rotates anticlockwise with an angle θ. Definitely,
this action should not create any strain. However, if we simulate this problem
based on the small deformation theory with one incremental step, the strain is
produced definitely and can be calculated by

L0 cos θ − L0
ε= = cos θ − 1 (6.1)
L0

The above equation can then be expanded by Taylor series as

θ2 θ2
ε≈1− −1=− (6.2)
2 2

with the higher-order terms being neglected. If we use N steps to simulate the
θ
problem, the incremental rotation angle for each step is δθ = . Then the strain
N

77
 PFEM FOR LARGE DEFORMATION PLANE-STRAIN PROBLEMS

L 0

Figure 6.1: A rotated bar.

generated in a total of N incremental small deformation analyses is calculated as

θ2 θ2
ε = Nδε = −N = − (6.3)
2N 2 2N

As shown in the above equation, the norm of the generated strain, that results
from the small deformation theory, for a bar with a rotation angle θ is proportional
1
to . For N → ∞, the strain tends to be zero.
2N

6.3 Preliminary results for large deformation anal-

ysis
6.3.1 An inclined axial rod
In this section, the possibility of analysing large deformation problems with a
sequence of incremental small deformation analyses is presented by considering
an inclined elastic axial rod subjected to a described displacement, u, as shown in
Figure 6.2. Such a geometrically nonlinear problem does have an analytical solu-
tion. Assuming the material is incompressible, the relation between the vertical

78
 CHAPTER 6

displacement u and the corresponding reaction F resulting from such a motion

can be expressed as
u

D=100 mm H=100 mm

Figure 6.2: Geometry of a single-axial rod.

E Young’s modulus 210 kN/mm2

√
L Initial length D√2 + H2

V Initial volume 100

p 2
l Loaded length D 2 + (H + u)2
u Vertical displacement

Table 6.1: Parameters for the inclined axial rod.

EV (H + u) l
F = ln (6.4)
l2 L
l
by using the logarithmic strain εl = ln and its conjugated Cauchy stress σ
L
according to [22]. The meaning of the parameters and variables in the above
equation is listed in Table 6.1.
The problem is simulated by a sequence of incremental analyses. The mathe-
matical programming formulations for linear bar elements are shown in Appendix
B. The simulation is carried out under the displacement control. A vertical dis-
placement u = 5L is enforced on the end of the bar within a total of 100, 200,

79
 PFEM FOR LARGE DEFORMATION PLANE-STRAIN PROBLEMS

0.5

0.4

0.3

0.2 100 increments

F/EA

200 increments
0.1 400 increments
800 increments
0 Analytical solution

−0.1

−0.2
0 1 2 3 4 5 6
u/L

Figure 6.3: Load deflection behaviour.

400, and 800 increments respectively. Figure 6.3 exhibits the resulting normalised
reaction force F/EA obtained from the simulations as well as the analytical solu-
tion. It shows that, the small deformation theory does lead to errors with coarse
increments (i.e. 100 increments). But an increase of the total incremental num-
ber leads to a decrease of the error. A close agreement can be obtained with
increments no less than 400 for the present problem.

6.3.2 Trussed frame

As the second example in this section, a trussed frame consisting of 238 elasto-
plastic bar elements is considered. The geometry of the problem is shown in
Figure 6.4. The material is assumed to be incompressible with Young’s modulus
E = 210 kPa, and the yield stress σs = 2.5 kN/mm2 . The frame is subjected to a
nominal downward vertical force F .
The simulation is conducted under the displacement control, which means
a downward vertical displacement uV = 65 mm is enforced at the loaded point

80
 CHAPTER 6

24

120

120

Figure 6.4: Geometry of a trussed frame (dimensions in mm).

120

20 mm
100
40 mm

80
60 mm

60

40

20

0
−20 0 20 40 60 80 100 120

Figure 6.5: Deformed shapes with the vertical displacement at 20 mm, 40 mm

and 60 mm.

81
 PFEM FOR LARGE DEFORMATION PLANE-STRAIN PROBLEMS

within 650 increments. The deformed configurations of the frame with the dis-
placement uV = 20 mm, 40 mm, and 60 mm are plotted in Figure 6.5, while the
force-displacement curves are shown in Figure 6.6. The blue line in Figure 6.6
is the simulated result obtained by the mathematical programming formulations
with configuration updating at each time increment, while the red dot gives the
simulated result from the large deformation theory presented in [22]. It is clear
that a satisfactory agreement is obtained.

400

Bonet's solu!on
300
Mathema!cal programming solu!on
Force F (N)

200

100

0
0 20 40 60 80
Vertical displacement (mm)

Figure 6.6: Load-displacement curves.

6.4 Analyses of plane strain problems

In this section, four numerical examples of plane strain problems are simulated
by the PFEM procedure associated with the mathematical programming formu-
lations for dynamics as well as frictional contact. For all examples, a reasonable
value of α was first determined by trial and error for a coarse discretisation (in

82
 CHAPTER 6

both time and space) and for a very limited range of the parameter set involved
in each problem. As discussed previously, this procedure is less than ideal but
at present the only feasible one. Also, in practice, it is not particularly arduous
and is usually completed in a fraction of the time that it takes to solve a given
problem (using more refined temporal and spatial discretisations) with a fixed
value of α. By this initial calibration, a value of α = 1.4 was adopted for the first
two examples, while α = 1.6 for the last two.
It is worth noting that, in all the plane strain problems concerned in this
section, both the yield conditions and plastic potential are given in terms of the
Mohr-Coulomb relation. While this constitutive model is very simple, it is also
quite appropriate for many problems involving very large deformations. Indeed,
the model verifies the conditions at the critical state; with the initial transient
phase leading up to this state thus being ignored, it may be thought of as a
reasonably accurate first-order model. In spite of that, it is worth noting that
the consideration of the elasticity, the softening/hardening material behaviour is
possible in the developed method. Besides failing to capture the conditions prior
to fully developed plastic flow, the only obvious flaw in the model is the absence
of rate effects. While such effects usually can be ignored at the microscopic level,
the processes at this level nevertheless manifest themselves in a what, at the
macroscopic level, appears to be a rate effect, essentially implying an increase in
friction angle with shear strain rate [9, 72]. While such an enhancement is entirely
possible within the general framework to be presented, it will not be considered
in the present work.

6.4.1 Cylinder-soil interaction

The first plane strain example concerns the interaction between a rigid weightless
cylinder and a Tresca soil. The cylinder first penetrates vertically to a depth
of 0.8 m after which it is displaced horizontally for a total distance of 2 m (see
Figure 6.7 and 6.8). This type of problem involving the combined horizontal and
vertical motion of a rigid cylindrical object resembles that of a pipeline subjected
to either external loads due to waves and current, or to internal temperature and
pressure loadings.

83
 PFEM FOR LARGE DEFORMATION PLANE-STRAIN PROBLEMS

D = 1m

Initial

uH=0, uV=0.2m

uH=0, uV=0.4m

uH=0, uV=0.8m

0 0.05

0.025

Figure 6.7: Interaction of a rigid cylinder with a Tresca soil. Colours are propor-
tional to the equivalent plastic strain increment.

84
 CHAPTER 6

uH=0.5m, uV=0.8m

uH=1.0m, uV=0.8m

uH=1.5m, uV=0.8m

uH=2.0m, uV=0.8m

0 0.05

0.025

Figure 6.8: Interaction of a rigid cylinder with a Tresca soil. Colours are propor-
tional to the equivalent plastic strain increment (continued).

The analysis conducted uses 280 displacement increments of magnitude equal

85
 PFEM FOR LARGE DEFORMATION PLANE-STRAIN PROBLEMS

to 0.01 m (80 for the penetration and 200 for the subsequent horizontal mo-
tion). The cylinder-soil interface is assumed purely frictional with a friction
coefficient µ = tan 30◦ ≃ 0.577. Other parameters are as follows: soil density
ρ = 1.6 g/cm3 , undrained shear strength c = cu = 1 kPa, gravitational accel-
eration g = −9.8 m/s2 , and cylinder diameter D = 1 m. The analysis is fully
dynamic with a time step of ∆t =1 s. Under these conditions, inertial effects are
negligible and the simulation may be regarded as quasi-static.

20
Normalized resultant force, F (kN/m)

15
FV
10

FH
5

0
0 0.5 1.0 1.5 2.0 uH

−5
0 0.4 0.8 uV
Displacment, u (m)

Figure 6.9: Horizontal and vertical resultant forces acting on the cylinder.

The results of the simulation in terms of load-displacement curves are shown

in Figure 6.9. During the initial penetration phase, the resultant horizontal force
remains at zero while the vertical force increases approximately linearly up to
around Fy = 18 kN/m. The cylinder is then displaced horizontally while being
fixed in its vertical position of uy = 0.8 m. This change in motion brings about
an immediate increase in the horizontal force and a decrease in the vertical force.
This is followed by a more gradual change of the forces until the motion becomes
steady at around ux = 1.5 m. At this point, the resultant forces are Fx ≃ 9 kN/m
and Fy = 7.5 kN/m respectively. Thus, an appreciable vertical force is required

86
 CHAPTER 6

to prevent the cylinder from moving towards the surface in response to the purely
horizontal motion. These findings echo those of [43] where a vertically constrained
cylinder was plowed through a granular material resulting in a similar ratio of
vertical to horizontal force.
Finally, we note that the fundamental change in geometry associated with the
self-contact at the back of the cylinder between uH = 1.0 m and uH = 1.5 m is
handled seamlessly, without resorting to any additional rules or procedures.

6.4.2 Accretionary wedge

The second problem concerns the deformation of an accretionary wedge as shown
in Figure 6.10. This is a common model problem in structural geology used to
study, among other things, mountain building via folding and thrusting events
[39, 111, 169, 170]. The wedge has a height of 8 cm and a length of 140 cm, and is
composed of a purely frictional material with φ = 30◦ and ψ = 0 and a density of
2 g/cm3 . The gravitational acceleration is g = −9.8 m/s2 . The left vertical and
bottom horizontal boundaries are rigid with friction coefficients of µ = 30◦ and
µ = 15◦ respectively.
The simulation is carried out by displacing the left wall horizontally with a
velocity of 0.4 cm/s. A total of 1,000 steps, with a constant time step of 0.16 s,
were used to yield a total wall displacement of 16 cm. Snapshots of the defor-
mations at different times are shown in Figure 6.10. Contrary to what might be
expected, the deformation do not proceed by evolution of one of more dominant
shear bands. Rather, the location of the shear bands changes in a somewhat
unpredictable manner: the classic V-band at the initial phases first propagates in
the direction of loading (∆u = 8.0 cm). Further loading then gives way to a new
shear band much like the initial one (∆u = 10 cm). This again gives way to an-
other family of shear bands further down the direction of loading. The end result
(when viewed as a continuous process) is a seemingly random appearance of shear
bands. This phenomenon has recently been studied in some detail by Mary et al.
[115] using a semi-analytical slip line type model. The conclusion is here that the
system displays deterministic chaos of a nature similar to more common chaotic
dynamical systems. Without repeating this analysis in the present context, we

87
 PFEM FOR LARGE DEFORMATION PLANE-STRAIN PROBLEMS

do note that the system displays behaviour consistent with that observed and
analysed by Mary et al. [115].

3°
g
∆u = 0.4cm

∆u = 8 cm

∆u = 10 cm

∆u = 12 cm

0.003
0.0027
0.0024 ∆u = 14 cm
0.0021
0.0018
0.0015
0.0012
0.0009
0.0006
0.0003
∆u = 16 cm
0

Figure 6.10: Accretionary wedge problem. Colours are proportional to the equiv-
alent plastic strain increment.

88
 CHAPTER 6

6.4.3 Collapse of granular columns

The next example concerns the collapse of granular columns. The setup is shown
in Figure 6.11. A column of granular material of width d0 and height h0 is
allowed to collapse under the action of gravity by quickly removing the right wall.
This type of experiment, which may be seen as a model landslide, has received
significant attention in recent years, both from an experimental point of view
[86, 103, 104] as well as in terms reproducing the experiments using both particle
methods [59, 81, 83, 84, 85, 174, 175, 198] and continuum models [73, 85, 113]. The

Figure 6.11: Granular column problem: initial column and final deposit.

two-dimensional plane strain version of the experiment considered in the following

is due to Lube et al. [104]. On the basis of a rather extensive experimental
program, the following fits to the height, h∞ , and diameter, d∞ , of the final
deposit were determined:

h∞  a , a ≤ 1.15
= (6.5)
d0  1.1a 25 , a > 1.15

89
 PFEM FOR LARGE DEFORMATION PLANE-STRAIN PROBLEMS
a = 0.5
a=1
a=7

−
t = 4.0
−

t = 1.0

t = 0.0

Figure 6.12: Collapse of granular columns (dots represent the finite element
nodes).

90
 CHAPTER 6


 1.6a , a < 1.8


d∞ − d0 
= transition region , 1.8 ≤ a ≤ 2.8 (6.6)
d0 

 2.2a 23

, a > 2.8

where a = h0 /d0 is the initial aspect ratio. We note that there is a transition
region at 1.8 ≤ a ≤ 2.8 for which no expressions for the final width were given.
To validate the proposed scheme, a series of experiments involving columns
with different aspect ratios were conducted. In all simulations, the following
parameters were used: density ρ = 2 g/cm3 , cohesion c = 0, friction angle φ = 31◦
(as quoted by Lube et al.) or φ = 35◦ , dilation angle ψ = 0◦ , and gravitational
acceleration g = −9.8 m/s2 . The simulations proceed from time t = 0 and are
p
terminated at t̄ = t/ h0 /g = 4.0. Figure 6.12 shows three columns, with aspect
ratios a = 0.5, 1.0, and 7.0, at different time instants of the collapse process
using a time step of ∆t̄ = 0.01. The internal deformation patterns observed, in
particular the apparent ‘buckling’ of internal columns, correspond very well with
what can be observed experimentally and in DEM-type simulations [81].
As expected, and predicted from (6.5), the two short columns collapse only
partially while the spread of the tallest one is much wider. A more complete
comparison with the experimental fits of Lube et al. is shown in Figure 6.13. We
here see that the columns essentially lack resistance, even for a friction angle of
35◦ . That is, the final height is underpredicted while the final run-out distance
is overpredicted, and with the deviation between experiment and simulation in-
creasing as the initial aspect ratio increases. As discussed previously, the key
feature missing in the constitutive model adopted is a rate-dependence of the
friction angle on the shear strain rate. With such an enhancement, for example
following the model of Andrade et al. [9], better fits to the experiments can be
expected.
To gauge the influence of the time step, the a = 1 simulation was run with
different time steps. The resulting final deposits are shown in Figure 6.14. As
seen, the shape and dimensions of the final deposit are relatively insensitive to the
time step. Indeed, the final height changes very little between the three different
time steps while the run-out distance does increase somewhat between the first

91
 PFEM FOR LARGE DEFORMATION PLANE-STRAIN PROBLEMS

3.0
Normalized final height, h∞/d0
2.5

2.0

1.5

1.0 Best-fit curves

to experimental data
0.5 Simulated results (φ = 31° )
Simulated results (φ = 35° )
0
0 2 4 6 8 10 12
Aspect ratio, a = h0/d0

12
Normalized final width, (d∞-d0)/d0

10

4 Best-fit curves
to experimental data
2 Simulated results (φ = 31° )
Simulated results (φ = 35° )
0
0 2 4 6 8 10 12
Aspect ratio, a = h0/d0

Figure 6.13: Normalised final height and width of granular columns as function
of aspect ratio.

two simulations involving 100 and 200 time steps respectively. A similar trend is
observed for other aspect ratios.
Finally, we investigate the mass balance properties of the scheme as function

92
 CHAPTER 6

100 time steps

200 time steps

400 time steps

Figure 6.14: Convergence of final deposit as function of time step for a = 1 (dots
represent the finite element nodes).

2
Volume change (%)

−2

100 time steps

−4
200 time steps
400 time steps
−6
0 1 2 3 4
−
Normalized time, t

Figure 6.15: Volume change as function of time for a = 1.

of the time step. The results, shown in Figure 6.15, reveal the loss of mass to be
around 2 to 6% and decreasing as the time step is decreased (the number of time
steps increased).

93
 PFEM FOR LARGE DEFORMATION PLANE-STRAIN PROBLEMS

6.4.4 Silo discharge

The final example concerns the discharge of a two-dimensional plane strain silo
with the dimensions shown in Figure 6.16. The material parameters are: density
ρ = 1.5 g/cm3 , cohesion c = 0, friction angle φ = 31◦ , dilation angle ψ = 0◦ , and
gravitational acceleration g = −9.8 m/s2 . The solid mass confined by the silo
wall is initially at rest after which, at time t = 0, the bottom wall is removed.

100 cm

35 cm

8 cm

50 cm

Figure 6.16: Silo discharge problem.

Two different sets of simulations were run: assuming perfectly smooth walls
and assuming perfectly rough walls (corresponding to a wall-solid friction coef-
ficient of µ = tan φ). The time step utilised in the simulations is ∆t = 0.01 s.
Snapshots of the discharge in the two cases are shown in Figures 6.17 and 6.18.
As can be seen, the wall-solid interface properties have a significant influence on
the overall flow pattern. Again, these patterns are consistent with those observed

94
 CHAPTER 6

in both experiments and particle simulations [53]. Moreover, the wall-solid in-
terface conditions influences the total discharge time, from approximately 5 s for
the smooth silo to almost twice that for the rough silo.

0.0 s 1.0 s 2.0 s

3.5 s 4.0 s 4.5 s

Figure 6.17: Discharge of smoothed walled silo (dots represent the finite element
nodes).

95
 PFEM FOR LARGE DEFORMATION PLANE-STRAIN PROBLEMS

0.0 s 3.0 s 5.0 s

6.0 s 7.0 s 8.0 s

Figure 6.18: Discharge of rough walled silo (dots represent the finite element
nodes).

The convergence of the results in terms of discharge versus time for different
time steps is shown in Figure 6.19 for the smooth walled silo. As seen, some 500
time steps (corresponding to ∆t = 0.01 s) appears to be sufficient to ensure a

96
 CHAPTER 6

reasonable degree of accuracy. This trend is the same for the rough-walled silo.

7
∆ t = 0.04s
6
Volume in silo (1000 cm )

∆ t = 0.02s
3

∆ t = 0.01s
5
∆ t = 0.005s
4

0
0 1 2 3 4 5 6 7
Time (s)
Figure 6.19: Silo volume versus time for different time steps.

6.5 Conclusions
In this chapter, a number of large deformation problems are analysed by using
the PFEM procedure associated with the developed mathematical programming
formulations for dynamic and contact analyses.
Before handling the plane strain problems, error estimation of using small
deformation theory for large deformation analysis is first conducted with the
focus on the truss system. It shows that the strain of a bar element arising from
θ2
the use of the small deformation theory for a pure rotation of θ equals . In
2N
other words, the increasing of the total number of analysis increments N leads
to the decrease of the error. The result of the followed numerical analysis of the
truss-related problems also agrees with this conclusion.
Then, four plane strain problems are concerned to exhibit the capability of
the proposed PFEM scheme associated with the developed mathematical pro-

97
 PFEM FOR LARGE DEFORMATION PLANE-STRAIN PROBLEMS

gramming formulations. The numerical examples show that the proposed scheme
is capable of handling extremely large deformation problems, especially those
involving both solid-like and fluid-like behaviour.

98
Chapter 7

Axisymmetric Collapse of
Granular Columns

7.1 Introduction
The problem of granular flow is of relevance in a large number of natural and
industrial processes and has received significant attention in recent years. A
particular problem that has been studied extensively is the collapse of granular
columns under the action of gravity. Since the basic experiment of an initially
cylindrical column on a flat horizontal surface was introduced [86, 103], a variety
of similar problems have been formulated and studied. These include ‘plane
strain columns’ [13, 87, 104], columns fluidised with air [151], columns in a fluid
[153, 182], columns on inclined planes [57, 105, 112], and quasi-static collapse of
columns [116, 137].
From a numerical point of view, the problem has been studied using both par-
ticle [59, 81, 84, 137, 174, 175, 198] and continuum approaches [33, 37, 73, 83, 85,
113, 204]. While both approaches have been shown to be capable of reproducing
typical experiments qualitatively, there have been few attempts (especially with
continuum approaches) at reproducing available experimental results quantita-
tively. Furthermore, the effect of material properties and basal roughness on the
final profiles is still not well understood. In this chapter, an axisymmetric version
of the PFEM is adopted to simulate the collapse of granular columns and address

99
 AXISYMMETRIC COLLAPSE OF GRANULAR COLUMNS

these issues.
The Chapter is organised as follows. In Section 7.2, the governing equations
for the axisymmetric granular flow problem are summarised. Section 7.3 details
the discretisation of the governing equations in both time and space. These equa-
tions are then cast in terms of a standard mathematical programming problem.
Finally, the resulting axisymmetric version of the PFEM is utilised to reproduce
the collapse of granular columns in Section 7.4 before conclusions are drawn in
Section 7.5.

7.2 Governing equations

While analysing the problem of axisymmetric structures, it is natural to use a
global cylindrical coordinate system (r, z, θ) to simplify the governing equations.
As shown in Figure 7.1, r is the radial coordinate, z is the axial coordinate, and
θ is the circumferential coordinate. The conversion of points, vectors and tensors
between the Cartesian coordinate system and the cylindrical coordinate system
is documented in Appendix C for reference.

Point ( r, z, θ )

θ
r

Figure 7.1: Global cylindrical coordinate system.

100
 CHAPTER 7

7.2.1 Momentum conservation equations

In the cylindrical coordinate system, the general momentum conservation equa-
tions for a body subjected a time-varying motion are given by

1 ∂(rσrr ) 1 ∂σrθ ∂σrz σθθ

+ + − + br = ρü2r
r ∂r r ∂θ ∂z r
1 ∂(rσzr ) 1 ∂σzz (7.1)
+ σzθ + + bz = ρü2z
r ∂r r ∂z
1 ∂(r 2 σθr ) 1 ∂σθθ ∂σθz
+ + + bθ = ρü2θ
r 2 ∂r r ∂θ ∂z

If the axisymmetric structures are subjected to axisymmetric loadings, the re-

sponse of the structure is axisymmetric and the problem can also be considered
axisymmetric. For such an axisymmetric problem, the above equations reduce to

1 ∂(rσrr ) ∂σrz σθθ

+ − + br = ρü2r
r ∂r ∂z r (7.2)
1 ∂(rσzr ) 1
+ σzθ + bz = ρü2z
r ∂r r

since the stress components σrθ and σzθ vanish and σθθ is independent of θ.
Rearranging (7.2) gives

S T σ + b = ρü, in Ω (7.3)

where the operator matrix S takes the form

 
∂ 1
 ∂r + r 0 
 
 −1
 
0 
S=
 r 
(7.4)
∂ 

 0



 ∂ ∂z 
∂ 1
+
∂z ∂r r

101
 AXISYMMETRIC COLLAPSE OF GRANULAR COLUMNS

and σ are the stresses given by

 


 σrr 


σ 
 
θθ
σ= (7.5)


σzz 

σ 
 
rz

In Eq. (7.3), Ω represents the domain under consideration, b = (br , bz ) are

the body forces, ρ denotes the material density, and a superposed dot represents
differentiation with respect to time.

7.2.2 Strain-displacement relations

Assuming infinitesimal deformations, the relationships between the strain and
displacement for the axisymmetric problem are

∂ur
εrr =
∂r
ur
εθθ =
r (7.6)
∂uz
εzz =
∂z
∂ur ∂uz
εrz = +
∂z ∂r

which can be written in the matrix form as

ε = Lu (7.7)

where u = (ur , uz )T , and ε = (εrr , εθθ , εzz , 2εrz )T are the displacements and the
strains in the cylindrical coordinate system, respectively, and operator L takes
the form
∂
 
 ∂r 0
1 
0
 
r

L=  (7.8)

0 ∂ 
 ∂z 
∂ ∂
 
∂z ∂r

102
 CHAPTER 7

7.2.3 Constitutive relations

The constitutive relation for granular materials is a subject of continued research
and there is, as yet, no unified framework or model that accounts satisfactorily
for the entire range of granular behaviour, from solid-like to fluid-like states.
Traditionally, most effort has been focused on the solid-like states, which are
the conditions under which most civil infrastructure operates, while less attention
has been devoted to the fluid-like states encountered in granular flows. However,
in the last decade or so, a number of models for the latter scenario have been de-
veloped [49, 72, 117, 118]. The basic idea is to first of all distinguish between solid-
and fluid-like states by a pressure-dependent Drucker-Prager or Mohr-Coulomb
type criterion involving a friction coefficient as the sole material parameter. Once
flow commences, the friction coefficient is taken to vary with some measure of the
rate of deformation. There are a number of candidates for both the measure of
the rate of deformation and the variation of the friction coefficient with this mea-
sure. However, over the last decade, the µ(I) rheology [72, 117] has emerged as
the most widely used approach for specifying the variation of friction coefficient
with flow rate. This model assumes that the friction coefficient, µ, increases with
the inertial number, I, from an initial value µs under static conditions to some
limiting value µmax for I tending to infinity. While many aspects of the µ(I) rhe-
ology have been verified experimentally, it is not without complications. Indeed,
as shown recently by Lagree et al. [85], the rheology appears to be somewhat
problem-dependent in the sense that a particular expression for µ(I) calibrated
to reproduce one experiment may need recalibration to reproduce another. Also,
from a more practical point of view, the inertial number involves division by the
pressure, which at free surfaces is zero leading potentially to some uncertainty
if not treated carefully. However, the question is whether rate-dependence is at
all necessary to reproduce the salient features of relevant natural and industrial
flows. That is, while the omission of rate-dependence such that only a constant
friction coefficient is considered certainly provides less fitting capability, in many
cases it may be enough. The simulations of [85] for the two-dimensional granular
column problem would seem to suggest that a rate-independent model is quite
reasonable. Consequently, in the modelling of granular columns collapse, a simple

103
 AXISYMMETRIC COLLAPSE OF GRANULAR COLUMNS

rate-independent model will be adopted.

More precisely, in our simulations, the rigid-plastic model is used to represent
the granular media, the constitutive equations of which are as follows:

F (σ) ≤ 0
ε̇ = λ̇∇σ G(σ) (7.9)

λ̇F (σ) = 0, λ̇ ≥ 0

where F is the yield function, G is the plastic potential, and λ̇ is a plastic mul-
tiplier. In this chapter, both the yield conditions and plastic potential are given
in terms of the Drucker-Prager model as

2 sin φ √ 6c cos φ
F =√ I1 + J 2 − √
3(3 − sin φ) 3(3 − sin φ) (7.10)
2 sin ψ √
G= √ I1 + J 2
3(3 − sin ψ)

in which φ is the friction angle (matched to the Mohr-Coulomb criterion for

plane strain), c is the cohesion, ψ is the dilation angle (with ψ = 0 implying
incompressibility), I1 and J2 are the first principal invariant of the stress tensor
and the second principal invariant of the deviatoric stress tensor in the cylindrical
coordinate system given by

I1 = σrr + σθθ + σzz

(7.11)
1
J2 = [(σrr − σθθ )2 + (σθθ − σzz )2 + (σzz − σrr )2 ] + σrz
2
6

It is worth noting that, although the rate-independent model is utilised here, the
rate-dependent model can also be implemented in the variational framework.

7.2.4 Frictional contact conditions

For the simulation of granular flows proper treatment of contact between the
material and the boundaries is essential. In this chapter, the contact between

104
 CHAPTER 7

the flowing granular medium and the rigid basal surface is accounted for via
non-penetration types of conditions (essentially implying that the basal surface
is perfectly rigid) as well.

Figure 7.2: Contact specification.

Consider a continuum body that is likely to come into contact with the rigid
surface as shown in Figure 7.2. For all material points, in the cylindrical coordi-
nate, the following frictional contact conditions should be fulfilled

∆uN − g0 ≤ 0, p ≥ 0

p(∆uN − g0 ) = 0 (7.12)

|q| − µp ≤ 0

where ∆uN is the normal displacement increment, g0 is the gap between the
material point and the surface, and p and q are the normal and shear contact
forces as shown in Figure 7.2. After spatial discretisation by finite element meshes,
the above constraints are enforced only on the mesh nodes of potential contact
with the rigid surface, i.e. those located on the computational boundary. We
note that the treatment of frictional contact outlined above follows that recently

105
 AXISYMMETRIC COLLAPSE OF GRANULAR COLUMNS

proposed by Krabbenhoft et al. [59, 79, 81] in the context of particle simulations.

7.3 Discretisation and solution

7.3.1 Time discretisation
Likewise, the θ-method is used to discretise the momentum conservation equa-
tions shown in (7.3). Subscripts n and n+1 refer to the known and new, unknown
states respectively, and ∆t = tn+1 − tn is the time step. The discretised linear
momentum equations are given by

v n+1 − v n
S T [θ1 σ n+1 + (1 − θ1 )σ n ] + b = ρ
∆t
(7.13)
un+1 − un
θ2 v n+1 + (1 − θ2 )v n =
∆t

where velocities, v, have been introduced. Rearranging the above equations leads
to
∆u
S T σ n+1 + b̃ = ρ̃
∆t2 (7.14)
1 ∆u
v n+1 = [ − (1 − θ2 )v n ]
θ2 ∆t
where the displacement increment is ∆u = un+1 − un and

ρ
ρ̃ =
θ1 θ2 (7.15)
1 v n 1 − θ1 T
b̃ = b + ρ̃ + S σn
θ1 ∆t θ1

The traction boundary conditions are approximated in an analogous manner lead-

ing to
N T σ n+1 = t̃, on Γ (7.16)

106
 CHAPTER 7

with
1 1 − θ1 T
t̃ = t− N σn (7.17)
θ1 θ1
where t are tractions, N = L(nxT ) with n = (nr , nz )T being the outward normal
to the boundary. Γ represents the boundary of the domain in consideration. The
above time integration scheme is unconditionally stable provided that θ1 ≥ 21
and θ2 ≥ 21 . For θ1 = θ2 = 21 the unconditionally stable and energy-preserving
Newmark average acceleration scheme is recovered, while, for θ1 = 0, a fully
explicit and conditionally stable scheme is obtained. In the following, we will
only consider implicit and unconditionally stable schemes.
Following Chapter 4, the above problem can be stated in terms of the fol-
lowing time-discrete optimisation problem (or variational principle) where the
constitutive relations are satisfied at time tn+1 ,

min max hσ n+1 , L(∆u)iΩ − hb̃, ∆uiΩ − ht̃, ∆uiΓ

∆u (σ,r)n+1

− 21 ∆t2 hrn+1 , ρ̃−1 rn+1 iΩ + hrn+1 , ∆uiΩ (7.18)

subject to F (σ n+1 ) ≤ 0

where we have used the notation

Z
hx, yiA = xT y dA (7.19)
A

and r n+1 are a set of variables whose physic meanings are the dynamic forces
r n+1 = ρ̃∆u/∆t2 .

7.3.2 Euler-Lagrangian equations

In this section, the equivalence between the optimisation problem (7.18) and
the governing equations described in Section 7.2 are verified by showing that
the Euler-Lagrange equations associated with (7.18) do indeed reproduce the
governing equations at hand.
To this end, the inequality constraints in (7.18) are first converted into equality

107
 AXISYMMETRIC COLLAPSE OF GRANULAR COLUMNS

constraints by subtraction of positively restricted slack variables s:

min max hσ n+1 , L(∆u)iΩ − hb̃, ∆uiΩ − ht̃, ∆uiΓ

∆u (σ,r)n+1

− 21 ∆t2 hrn+1 , ρ̃−1 rn+1 iΩ + hrn+1 , ∆uiΩ

(7.20)
+β ΣI∈C ln sI
subject to F (σ n+1 ) + s = 0

where β is an arbitrarily small constant and C is the set of stress integration

points. The logarithmic barrier function eliminates the need to make explicit
reference to the fact that s > 0. The standard Lagrange multiplier technique
then applies, i.e. the solution to (7.20) is found by requiring stationarity of the
following functional:

J = hσ n+1 , L(∆u)iΩ − hb̃, ∆uiΩ − ht̃, ∆uiΓ

− 21 ∆t2 hrn+1 , ρ̃−1 r n+1 iΩ + hr n+1 , ∆uiΩ (7.21)

+β ln s − ∆λ(F (σ n+1 ) + s)

where ∆λ is a Lagrange multiplier. The stationary conditions are given by


δJ  S T σ n+1 + b̃ = r n+1 , in Ω

=
δ∆u   N Tσ
n+1 = t̃, on Γ

δJ
= −∆t2 ρ̃−1 r n+1 + ∆u = 0
δr n+1
δJ (7.22)
= ∇u − ∆λ∇σ F (σ n+1 ) = 0
δσ n+1
δJ
= F (σn+1 ) + s = 0
δ∆λ
δJ β
= − ∆λ = 0 ⇒ s∆λ = β
δs s

For β → 0, these equations are easily verified as being the governing equations

108
 CHAPTER 7

for the problem at hand. Since the rigid-plastic model is used, the third set of
the above conditions consists only of plastic strains. Note, again, that the flow
rule in the above equations is associated, i.e. the flow potential is equal to the
yield potential F . This is a consequence of the variational formulation and is
not possible to circumvent directly. The approach suggested in [80] of replacing
the original yield function with an ‘effective’ one can be used to consider non-
associated plastic flow. This approach was shown to produce reliable results for a
range of static small deformation problems, and its equal reliability for dynamic
problems involving very large deformations has been shown in Chapter 6 as well.

7.3.3 Spatial discretisation

Now, we are in the position to introduce the following finite element approxima-
tions for the state variables σ, r and u:

Figure 7.3: Quadratic displacement/linear stress element utilised in the simula-

tion.

σ(x) ≈ N σ (x)σ̂
r(x) ≈ N r (x)r̂ (7.23)
u(x) ≈ N u (x)û, Lu ≈ B u (x)û

where σ̂, r̂ and û are stresses, dynamic forces and displacements at the corre-
sponding interpolation nodes, the N matrices contain the shape functions, and

109
 AXISYMMETRIC COLLAPSE OF GRANULAR COLUMNS

B u = LN u . In our simulations, a linear stress/quadratic displacement trian-

gular element is utilised. As illustrated in Figure 7.3, the stresses interpolation
1 4 1
points are located at (λj−1, λj , λj+1) = ( , , ), j = 1, 2, 3, where λj are the area
6 6 6
coordinates. The stresses are assumed to be continuous within each element and
discontinuous between elements while the displacements vary quadratically and
are continuous between elements. The shape functions for the dynamics forces in
N r are taken to be identical to those of the displacements but are discontinuous
between elements. More discussion on the properties of different elements can
be found in [77, 78]. Substituting the above approximations into the variational
principle (7.18) leads to the following discrete principle

min max σ̂ Tn+1 B∆û − f T ∆û − 12 ∆t2 r̂ Tn+1 Dr̂ n+1

∆û (σ̂,r̂)n+1

+r̂ Tn+1 A∆û (7.24)

subject to F (σ̂ jn+1 ) ≤ 0, j = 1, . . . , nσ

where the yield function is imposed at a finite number of points nσ and

Z
B= N Tσ B u dΩ (7.25)
Ω

Z Z
T
f= N u b̃ dΩ + N Tu t̃ dΓ (7.26)
Ω Γ

Z
D= N Tr ρ̃−1 N σ dΩ (7.27)
Ω

Z
A= N Tu N r dΩ (7.28)
Ω

The minimisation part of (7.24) may be solved first to arrive at the maximisation
problem
maximise − 12 ∆t2 r̂ Tn+1 Dr̂ n+1
(σ̂,r̂)n+1

subject to B T σ̂ n+1 + AT r̂ n+1 = f (7.29)

F (σ̂ jn+1 ) ≤ 0, j = 1, . . . , nσ

110
 CHAPTER 7

At this stage, the contact conditions mentioned in Section 7.2 are imposed on
all potential contact nodes, i.e. mesh nodes located on the computing boundary,
which leads to a final problem of the type

Pnc
maximise − 21 ∆t2 r̂ Tn+1 Dr̂ n+1 − j=1 g0j pj
(σ̂,r̂)n+1

subject to B T σ̂ n+1 + AT r̂ n+1 + E T ρ = f

F (σ̂ in+1 ) ≤ 0, i = 1, . . . , nσ
(7.30)
pj = −nj ρj , j = 1, . . . nc
T

qj = −n̂Tj ρj

|qj | − µpj ≤ 0

with ρ = (ρ1 , ρ2 )T being the nodal forces, n the normal of the rigid boundary and
n̂ = (−n2 , n1 )T . Furthermore, E is an index matrix of zeros and ones and nc is
the number of potential contacts. The above problem is similar in structure to
those arising from limit analysis [75], static elastoplasticity [76, 77, 78, 80], and
granular contact dynamics [59, 79, 81].
Similarly, the problem (7.30) can be transformed into a standard second-
order cone program (SOCP) following Appendix A and then solved using the high
performance optimisation solver MOSEK [7] as shown in Chapter 4. In the course
of solving the problem, the kinematic variables (displacement increments and
plastic multipliers) are recovered as the dual variables, or Lagrange multipliers,
associated with the discrete equilibrium constraints.

7.4 Numerical simulations

The spreading of granular columns, both axisymmetric and plane-strain, on a
horizontal plane has received significant attention in recent years both from an
experimental point of view as well as in terms of reproducing the experiments
using both discrete models and continuum models. In this section, the axisym-
metric case, the experiments of which have been conducted by Lube et al. [103]

111
 AXISYMMETRIC COLLAPSE OF GRANULAR COLUMNS

and Lajeunesse et al. [86, 87], is studied numerically using the proposed PFEM
scheme with the developed mathematical programming formulations for axisym-
metric problems.
The setup of the experiment is shown in Figure 7.4. A cylindrical column of
granular material of initial radius r0 and height h0 is allowed to collapse under
the action of gravity by quickly removing the container. This leads to a deposit
with final radius r∞ and height h∞ . In the following, if not mentioned explicitly,
the following material parameters are used: density ρ = 2.0 g/cm3 , friction angle
φg = 30◦ (as quoted by Lube et al. [103]), friction coefficient of the basal surface
µ = tan φs with φs = φg , and gravitational acceleration g = −9.8 m/s2 . The
p
simulations proceed from time t = 0 and are terminated at t̄ = t/ h0 /g = 4.0.

Figure 7.4: Granular column problem: initial column and final deposit.

7.4.1 Collapse evolution

In all simulations, the columns begin to collapse the moment the container is
removed. However, the patterns of collapse are quite different for columns with

112
 CHAPTER 7

different initial aspect ratio a = h0 /r0 . Figure 7.5 shows three typical collapse
patterns observed in the simulations. The grey line corresponds to the initial
configuration while the black line is the final deposit. The intermediate profiles
are also plotted with distributions of velocities. As seen, increasing the aspect
ratio leads to a more complete collapse [(a), (b) vs (c),(d)]. Also note that while
the upper surface of the column with the lower aspect ratio deforms from the
very beginning, the upper surface of the more slender column remains essentially
undisturbed until some way into the collapse. These collapse patterns are very
similar to those observed by Lube et al. [103] and Lajeunesse et al. [86, 87].

Figure 7.5: Velocity distribution of granular columns with initial aspect ratio a
at time t̄. (a) a = 0.5, t̄ = 1.0 (b) a = 1.0, t̄ = 1.0 (c) a = 5.0, t̄ = 0.8 (d) a = 5.0,
t̄ = 1.2. The grey line corresponds to the initial configuration while the black
line is the final deposit. Colours are proportional to the magnitude of velocities
(cm/s).

The experimental and simulated dynamic data are compared for a column
with a = 3.6 and r0 = 3.9 cm in Figure 7.6. The instantaneous height and radius

113
 AXISYMMETRIC COLLAPSE OF GRANULAR COLUMNS

h(t ) r (t ) − r0
3 r (t ) r0

0
0 1 2 3
t

Figure 7.6: Comparison of experimental and simulated dynamic data for the
column with a = 3.6 and r0 = 3.9 cm. The dash line corresponds to the simulated
result while the solid line is the experimental result.

are here given by h(t) and r(t) respectively. In this simulation, the friction angle
of the granular medium is φg = 28◦ , which approximately coincides with the
avalanche angle (27.4 ± 0.5◦ ) of the beads used by Lajeunesse et al. [87]. As
seen, for the normalised radius, the experimental result displays a slight delay
compared to the simulations. This may be due to the effect of a finite lifting
velocity of the container. At the final state, the simulated final radius is slightly
higher than that of the experiment. Regarding the ratio between instantaneous
height over radius, good agreement is achieved.

7.4.2 Details of collapse with large aspect ratio

In [103], it is reported that, for columns with large initial aspect ratios, a thicken-
ing flow front is observed during the collapse and bulges are preserved in the final
deposits. In this section, the collapse of a column with a relatively large aspect

114
 CHAPTER 7

ratio, a = 8.0, is studied in detail with the aim of reproducing the experimentally
observed characteristics.

Figure 7.7: Configurations of granular columns with initial aspect a = 8.0 at time
(a) t̄ = 0, (b) t̄ = 1.0, (c) t̄ = 1.2, (d) t̄ = 1.28, (e) t̄ = 2.0, (f) t̄ = 4.0. Colours
are proportional to the magnitude of velocities (cm/s).

The initial and final configurations as well as some intermediate profiles are
shown in Figure 7.7. The arrows here represent the flow direction of the material
with the colour proportional to the norm of resultant velocities. As illustrated,
after lifting the container, the flow is divided into three regimes as shown in Figure
7.7(b). The material under curve A remains static whereas the material located
between curves A and B flows along the surface of the undisturbed region. The
material above curve B undergoes purely vertical motion at a relatively higher
speed. As the free upper surface approaches the static region, it deforms to
become a dome as shown in Figure 7.7(c). As a result, a wave is formed at the
margin of the collapsing column and will propagate outwards. Due to its greater

115
 AXISYMMETRIC COLLAPSE OF GRANULAR COLUMNS

velocity, the wave pushes the materials in front; consequently, a thickening flow
front is formed (Figure 7.7(d),(e)) and a bugle is preserved at the final deposit
(Figure 7.7(f)).

7.4.3 Normalised final height and radius

On the basis of a rather extensive experimental program with different granular
materials characterised by a friction angle ranging between 30◦ and 35◦ . The
following fits to the height, h∞ , and radius, r∞ , of the final deposit were concluded
by Lube et al.[103]:

10
(r∞-r0)/r0

best-fit curve
Lube et al.
Lajeunesse et al.
simulated results
0.1
0.1 1 10 100
Aspect ra"o a

Figure 7.8: Normalised final radius of granular columns as a function of aspect

ratio.


r∞ − r0  1.24a , a < 1.7
= (7.31)
r0  1.6a 21 , a > 1.7

116
 CHAPTER 7


h∞  a , a<1
= (7.32)
r0  0.88a 61 , 1.7 < a < 10

In the experiment of Lajeunesse et al. [87], glass beads with a repose angle of
22 ± 0.5◦ and an avalanche angle of 27.4 ± 0.5◦ were used. The scaled final height
and radius are reported to obey the following laws:

r∞ − r0  a , a<3
∝ (7.33)
r0  a 21 , a>3

h∞  a , a < 0.74
= (7.34)
r0  0.74 , a > 0.74

10
h∞/r0

best-fit curve
Lube et al.
Lajeunesse et al.
simulated results
0.1
0.1 1 10 100
Aspect ra"o a

Figure 7.9: Normalised final height of granular columns as a function of aspect

ratio.

117
 AXISYMMETRIC COLLAPSE OF GRANULAR COLUMNS

10 2
(a) (b)
(r∞-r0)/r0

h∞/r0
5 1

best-fit curve
φg = 20°
φg = 30°
φg = 40°

0 0
0 5 10 15 20 0 5 10 15 20
Aspect ra"o a Aspect ra"o a

Figure 7.10: Normalised final radius and height of deposit profiles as functions of
aspect ratio.

In our simulations, the friction angle of the granular material is φg = 30◦

(as quoted for sands by Lube et al. [103]). Figure 7.8 and Figure 7.9 show
the comparison between our simulated results by the developed PFEM and the
experimental results [87, 103] of the normalised radius and height of the final
deposits. The solid lines correspond to (7.31) and (7.32). As seen, the simulated
results, both in terms of the final radius and height, agree well with those from
Lube et al. [103] and the best-fit curves. It is noted that for columns with
initial aspect ratio a > 6, the normalised height obtained by both simulations
and the experiments is significantly less than that estimated from (7.32). Quite
surprisingly, even though the friction angle of the granular medium used in the
experiments of Lajeunesse et al. [87] is much smaller, there are few differences
between the normalised final radius as compared to the results of Lube et al.
However, the normalised final height obtained by Lajeunesse et al. is rather

118
 CHAPTER 7

lower than those from Lube et al. and our simulations.

Next, the influence of the friction angle φg on the final radius and height is
studied. To this end, three different friction angles φg = 20◦ , 30◦ , and 40◦ are
used. In all cases, the basal friction coefficient is µ = tan φg . In contrast to what
was suggested by [103], the friction angle of the granular materials does influence
both the final radius and height. Indeed, increasing the friction angle leads to an
increase of the final height and decrease of the final radius (see Figure 7.10).

7.4.4 Influence of basal roughness

While influence of basal friction to some extent has been studied experimentally,
there is no unified opinion regarding its exact effect on the final profiles. Lube
et al. suggest that the roughness of the basal surface has no effect on the final
profile while as Lajeunesse et al. report that it does have a non-negligible effect.
0.6
φs = 20°
0.4 φs = 25°
h/r0

φs = 30°
0.2
a=0.5
0.0
0.0 1.0 2.0 3.0 4.0
1.5
φs = 20°
1.0 φs = 25°
h/r0

φs = 30°
0.5
a=4.0
0.0
0.0 2.0 4.0 6.0 8.0
1.5
φs = 20°
1.0 φs = 25°
h/r0

φs = 30°
0.5
a=8.0
0.0
0.0 2.0 4.0 6.0 8.0
r/r0

Figure 7.11: Deposit profiles normalised to the initial radius.

119
 AXISYMMETRIC COLLAPSE OF GRANULAR COLUMNS

In this section, the collapse of three columns with aspect ratios a = 0.5, 4.0, 8.0
on horizontal surfaces with different basal friction coefficients µ = tan 20◦ , tan 25◦ ,
tan 30◦ are simulated. The normalised final profiles are illustrated in Figure 7.11.
From these we see that for the column with the smallest aspect (a = 0.5), there
is little effect of the roughness on the final profile. However, as the aspect ratio
increases, i.e. a = 4.0 and a = 8.0, the increase of roughness leads to an increase
of the final height and a decrease of the final radius. In other words, the higher
the aspect ratio, the more pronounced the influence of basal friction is.

7.5 Conclusions
In this chapter, an axisymmetric version of the PFEM is developed for the sim-
ulation of the collapse of granular columns on a horizontal plane. The granular
material is modelled as a rate-independent rigid-plastic material model. The
frictional contact between the granular material and the rigid basal surface is
accounted for. A variational approach has been taken whereby the discrete gov-
erning equations of the axisymmetric problems give rise to a second-order cone
program that can be solved in an efficient and robust manner using readily avail-
able optimisation tools.
The numerical simulations reproduce the three typical collapse patterns for
granular columns with small, intermediate and large initial aspect ratios. The
forming of thickening flow front is reproduced granular columns with relatively
large aspect ratios. A rather good agreement is achieved for the simulated and ex-
perimental results regarding the normalised radius and height of the final profile.
Moreover, from the simulation we can see that the friction angle of the granular
material, as well as the roughness of the basal surface, do impact the final deposit
profiles.

120
Chapter 8

A Case Study: Yangbaodi

Landslide

8.1 Introduction
A landslide is the movement of a mass of rock, debris, or earth (soil) down a
slope (under the influence of gravity) [38]. A large landslide has great destruc-
tiveness, and may result in extensive damages and casualties. For example, the
1963 landslide in Vajont, Italy led to nearly 2000 deaths; likewise, the 2010 mud-
slide in Zhouqu County, China resulted in 1287 deaths and 457 missing. Thus,
the post-failure analysis of landslides (e.g. the assessment of run-out distance
of catastrophic mass movements) is of crucial significance to protect lives and
properties.
The approaches for estimating the run-out distance can be roughly categorised
into empirical-statical and dynamic methods [150]. To date, a number of studies
on large mass movement have been conducted using either of them. Representa-
tive works include [32, 62, 144, 158] and the references cited in. The empirical-
statical approaches are easy to use; nevertheless, they should only be used in
cases where the conditions are similar to those under which the empirical-statical
formulations are developed. The dynamic methods, on the other hand, are devel-
oped based on the conservation of momentum and/or energy of flows; moreover,
they are able to provide the complete description of flows. The main drawback

121
 A CASE STUDY: YANGBAODI LANDSLIDE

of the dynamic methods is the estimation of friction parameters and material pa-
rameters [62, 66]. In this chapter, an actual landslide that occurred in Southern
China is studied using the developed PFEM scheme.

8.2 Yangbaodi landslide

The Yangbaodi landslide (see Figure 8.1) occurred in a hilly region at an altitude
of 104-385 m in Southern China, on September 18th, 2002. It resulted in 5 deaths
and 31 injuries [92]. The plane view of the landslide is shown in Figure 8.2(a)
from which we can see that the slope failed at the middle-low level (130-199 m)
of the hill. The cross-sectional profile is shown in Figure 8.2(b). As illustrated,
the slope consisted of the fills that were composed of completely and intensely
decomposed granite from quarries. The weathering grade of the granite deceased
with the depth.

Figure 8.1: Post event topography (after Li et al. [92]).

122
 CHAPTER 8

Figure 8.2: Yangbaodi flowslide: (a) topographic map showing the extent of the
flowslide; and (b) slope profiles before and after the flowslide (after Li et al. [92]).

123
 A CASE STUDY: YANGBAODI LANDSLIDE

The on-site and laboratory tests were conducted after the failure of the filled
slope. It was concluded [205] that the prolonged rainfall led to the failure of
the static liquefaction of the soil, which resulted in the sliding. The physical
parameters of the sliding soil obtained from the tests are listed in Table 8.1.

Physical properties
Dry density (103 kg/m3 ) 1.25-1.40
Void ratio 0.79-1.36
Specific gravity 2.67-2.68
Peak friction angle ( ◦ ) 28
Cohesion (kPa) 0

Table 8.1: Physical properties of the soil from [92].

8.3 Model setup

In this chapter, the landslide is modelled as a two-dimensional plane strain prob-
lem (which means the lateral changes in slide mass geometry and channel shape
are not taken into account), with constant physical mechanical properties that
reflect the average flow properties during its entire development. The setup of
the problem is illustrated in Figure 8.3, in which the slip surface is predefined
from in situ evidence and treated as a rigid frictional boundary. The sliding mass
is represented by a non-associated rigid-plastic model with Mohr-Coulomb yield
criterion, while the effect of the liquid is neglected. In other words, there is no
deformation taking place below yield; once the yield condition is satisfied, plas-
tic deformation occurs. The material parameters reported in [92] are used. The
friction angle is φ = 28◦ and cohesion is c = 0. The dilatancy angle is ψ = 0,
which means the material is incompressible, and the density of the sliding mass is
ρ = 1.25 t/m3 . The friction coefficient of the slip surface will be calibrated by the
back-analysis later, and the gravitational acceleration is g = −9.8 m/s2 . In the
simulation, all the displacement degrees of mesh nodes on the surface are fixed
to generate the initial stress of the sliding mass. After 10 seconds, those imposed
zero displacement conditions are removed to trigger the sliding process.
The quadratic displacement/linear stress triangular finite element described

124
 CHAPTER 8

in Chapter 4 is used to discretise the computational domain as shown in Figure

8.3. Since, in the simulation, the remeshing operation of the computational do-
main is required frequently, the total number of finite elements is time-variant,
the value of which fluctuates around 2000. A total of 600 incremental finite ele-
ment analyses are carried out to simulate the entire sliding procedure with time
increment ∆t = 0.05 s. The parameters for dynamic analysis are θ1 = θ2 = 1.0,
which means an implicit time integration scheme with numerical damping is ap-
plied. The parameter α for the boundary detection at a typical time step is
estimated automatically through αn = m0 α0 /mn to ensure the mass conservation
in this study, where mn and m0 are the assessed total mass at time tn and t0 ,
respectively. In such a way, the change of the total mass is always within 2% for
the simulation of the present problem.

250

225
Sliding mass
200
Elevation (m)

175
Toe of
150 deposition

125
Slip surface

100
0 50 100 150 200 250 300 350
Horizontal distance (m)

Figure 8.3: Setup for the numerical simulation of the Yangbaodi landslide.

8.4 Simulation results

8.4.1 Calibration of friction coefficient
As illustrated in the simulation of the collapse of granular columns shown in
Chapter 7, the friction coefficient µ of the slip surface has a significant influence

125
 A CASE STUDY: YANGBAODI LANDSLIDE

on the shape of the final deposit. In this section, the roughness of the basal
surface is first calibrated by a series of back-analyses.

250
(a)
225

200

175

150

125

100
0 50 100 150 200 250 300 350
250
(b)
225

200

175

150

125

100
0 50 100 150 200 250 300 350
250
(c)
225

200

175

150

125

100
0 50 100 150 200 250 300 350

Figure 8.4: Final deposit obtained with friction coefficient (a) µ = tan 8◦ , (b)
µ = tan 10◦ , and (c) µ = tan 12◦ .

126
 CHAPTER 8

The simulation is conducted with different µ ranging from µ = tan 0◦ to

tan 28◦ . Figure 8.4 presents three selected final deposits obtained with surface
friction coefficient µ = tan 8◦ , tan 10◦ , and tan 12◦ respectively. As expected,
a decrease of the coefficient µ leads to an increase of the deposit length and a
decrease of the deposit thickness. For µ = tan 10◦ , the simulated deposit shape
and the position of the deposit toe are closest to the reported ones. Thus we set
the friction coefficient µ = tan 10◦ hereafter.

8.4.2 Influence of material density

Note that the mass density obtained from the experimental tests [92] is 1.25 ∼
1.40 t/m3 . We here investigate the effect of the density on the sliding process.
This is conducted by simulating the problem with ρ being 0.5 t/m3 , 1.25 t/m3 ,
and 2.5 t/m3 , respectively. The second value is the one located in the reported
arrange (i.e. 1.25 ∼ 1.40 t/m3 ), while the first and the last values represent two
artificial extremes to gauge the sensitivity to the exact value of the density.

350

300
Horizontal posi!on (m)

ρ=0.5 t/m3
250
ρ=1.25 t/m3

ρ=2.5 t/m3
200

150
0 5 10 15 20 25 30 35
Time (s)

Figure 8.5: The horizontal position of the leading edge versus time.

127
 A CASE STUDY: YANGBAODI LANDSLIDE

250
(a)
225

200

175

150

125

100
0 50 100 150 200 250 300 350
250
(b)
225

200

175

150

125

100
0 50 100 150 200 250 300 350
250
(c)
225

200

175

150

125

100
0 50 100 150 200 250 300 350

Figure 8.6: Final deposit obtained with density (a) ρ = 0.5 t/m3 , (b) ρ =
1.25 t/m3 , and (c) ρ = 2.5 t/m3 .

The position of the flow front versus time obtained with these different mass

128
 CHAPTER 8

densities is plotted in Figure 8.5. As shown, there is little difference between

various results for such a wide range of the density. When the flow starts, the front
moves with its horizontal velocity being nearly constant. At around 20 s after the
initiation of the sliding, the front reaches the maximum horizontal distance at
325 m. Furthermore, we plot the final deposits of the landslide in Figure 8.6
where no apparent differences appear regarding their shapes. These observations
imply that the mass density plays only a minor role in the current landslide. Such
a conclusion echoes the observations in physical model tests of granular collapse
in [86, 87, 103, 104] that the final deposits of granular collapse are independent
of the density of granular medium.

8.4.3 Kinematic of typical material points

Next, we focus on the kinematics of three typical material points that are initially
located at the rear, the middle, and the front parts of the sliding mass, as shown
in Figures 8.7 and 8.8. The grey districts represent the initial configuration of the
sliding mass and its final deposit. Red curves denote the trace of the movements
of the points, while blue curves are the corresponding velocities in X and Y
directions.
As shown, the traces of all three points are very similar to the geometry of the
predefined slip surface. When the flow ceases, points A, B and C are still located
at the rear, the middle and the front of the final deposit, respectively. Rather
than maintaining within the domain as points B and C, point A is located on
the surface of the final deposit. The simulation of such a phenomenon, which is
impossible to achieve by the traditional finite element method, can be tackled by
the proposed PFEM scheme.
The variations of the point velocities during the sliding process are also plot-
ted. The horizontal (X) velocity is illustrated in Figure 8.7(a), while the ver-
tical (Y) in Figure 8.8. For point A, the horizontal velocity obtained from the
PFEM simulation increases considerably when the flow starts; then it fluctuates
around 10 m/s when its horizontal distance is between 50 m and 120 m. After-
wards, the velocity increases to its maximum value (15 m/s) and then remains
constant before decreasing sharply to zero in the course of about 10-meter hor-

129
 A CASE STUDY: YANGBAODI LANDSLIDE

25

Velocity in X direction (m/s)

A 20
15
10
5
0
−5

250 25

Velocity in X direction (m/s)

225 20

200 15
Y (m)

175 B 10

150 5

125 0

100 −5
0 50 100 150 200 250 300 350
X (m)
250 25
Velocity in X direction (m/s)

225 20

200 15
Y (m)

175 10

150 C 5

125 0

100 −5
0 50 100 150 200 250 300 350
X (m)

Figure 8.7: Horizontal velocity of material points. Green lines denote the velocity
obtained from the DEM simulation in [92], while blue lines from the developed
PFEM simulation.

130
 CHAPTER 8

250 5

Velocity in Y direction (m/s)

225 A
0
200
Y (m)

175 −5

150
−10
125
100 −15
0 50 100 150 200 250 300 350
X (m)
250 5

Velocity in Y direction (m/s)

225
0
200
Y (m)

175 B −5

150
−10
125
100 −15
0 50 100 150 200 250 300 350
X (m)
250 5 Velocity in Y direction (m/s)

225
0
200
Y (m)

175 −5

150 C
−10
125
100 −15
0 50 100 150 200 250 300 350
X (m)

Figure 8.8: Vertical velocity of material points. Green lines denote the velocity
obtained from the DEM simulation in [92], while blue lines from the developed
PFEM simulation.

131
 A CASE STUDY: YANGBAODI LANDSLIDE

izontal movement. The situations for points B and C are fairly similar. Both
of their horizontal velocities increase sharply from zero to their maximum values
(16 m/s for point B and 12.5 m/s for point C), and then start to decrease when
the points reach the toe of the slip surface. Regarding the evolution of the vertical
velocity illustrated in Figure 8.8, a negative value represents a downward motion
of the material point, while a positive denotes an upward motion. For points A
and B, the vertical velocities remain negative, whereas for point C, the velocity
in the Y direction turns to be positive when it collides with the horizontal slip
surface for the first time, representing a bounce of the material. Those curves
in Figure 8.8 indicate that the change of the surface geometry can result in an
intense fluctuation on the vertical velocities, especially at the point where a sharp
change in gradient of the surface exists. Also, we note that the velocity evolution
obtained from the PFEM simulation is rather similar to that from the DEM mod-
elling [92]. Nevertheless, the result from the PFEM simulation does not oscillated
severely as that from the DEM simulation; and, moreover, the computation cost
of the PFEM simulation is significantly lower than that of the DEM.

8.4.4 Flow evolution

Snapshots of the flow evolution are displayed in Figures 8.9 and 8.10 with the
contour of the velocity being plotted. The results are presented at a relatively
short time interval of 5 s. As illustrated, after an elapsed time of 5 s, the sliding
mass (especially the rear part) has deformed severely, and the front of the mass
almost reaches the toe of the slip surface. At t =10 s, about 20% of the sliding
mass has reached the flat area, and at that time the middle part of the sliding
mass travels at relatively high speed (≥ 16 m/s). The next 5 s leads to a great
restoration of the mass (about 70%) on the flat area. In the meantime the velocity
of the leading front slows down during this period (also see Figure 8.6). At
t = 20 s the rear part of the mass stops, while the front and middle parts travel
horizontally at very low speed. The entire sliding process is finished at t = 25 s
resulting in a hazard zone in the length of 140 m.

132
 CHAPTER 8

|Velocity|
250 16
225 t=0 s
12
200
Y(m)

175 8
150
4
125
100 0
0 50 100 150 200 250 300 350
X(m)
|Velocity|
250 16
225 t=5 s 12
200
Y(m)

175 8
150
4
125
100 0
0 50 100 150 200 250 300 350
X(m)
|Velocity|
250 16
225
t=10 s 12
200
Y(m)

175 8
150
4
125
100 0
0 50 100 150 200 250 300 350
X(m)

Figure 8.9: Snapshots of flow evolutions with velocity contour. The unit for
velocities used is m/s.

133
 A CASE STUDY: YANGBAODI LANDSLIDE

|Velocity|
250 16
225 t=15 s 12
200
Y(m)

175 8
150
4
125
100 0
0 50 100 150 200 250 300 350
X(m)
|Velocity|
250 16
225 t=20 s 12
200
Y(m)

175 8
150
4
125
100 0
0 50 100 150 200 250 300 350
X(m)
|Velocity|
250 16
225
t=25 s 12
200
Y(m)

175 8
150
4
125
100 0
0 50 100 150 200 250 300 350
X(m)

Figure 8.10: Snapshots of flow evolutions with velocity contour (continued). The
unit for velocities used is m/s.

134
 CHAPTER 8

8.5 Conclusions
In this chapter, the application of the proposed PFEM scheme for the simulation
of an actual natural disaster event is presented. The rate-independent rigid-
plastic model is used to represent the sliding mass. The simulated results show
that the friction coefficient between the sliding mass and its basal surface has a
crucial influence on the forming of final deposit, whereas the density of the sliding
mass is less so. Such an observation in our simulations coincides with those in the
physical model testing of the collapse of granular columns. Besides, the velocity
evolutions of three typical material points are focused on as well. The changes of
the simulated velocities are quite similar to those from the DEM modelling.
It should be noted that the rate-independent model is utilised in the simu-
lation which can reproduce only the average process of the sliding; and the real
flowslide involves a complicated mechanism that requires a more sophisticated
model. However, the proposed method is still valid for the estimation of the
flow evolution and its final run-out distance, which are essential for the hazard
assessment.

135
Chapter 9

Conclusions

In this thesis, a new continuum approach is proposed for the simulation of geotech-
nical problems involving both solid-like and liquid-like behaviour. More specif-
ically, the Particle Finite Element Method (PFEM) is used to tackle geometry
related problems that plague the traditional Lagrangian FEM, i.e. severe mesh
distortion and boundary evolution, while mathematical programming formula-
tions are developed to obtain an efficient and robust numerical solution strategy.
Some conclusions of the present work are drawn as follows.
• Capability of the PFEM
The capability of the PFEM is not just limited to overcoming mesh distortion
in large deformation problems. Rather, the PFEM can capture the severe bound-
ary evolution of the computational domain, i.e. in the case where an initially
contiguous solid separates into two or more parts as a result of external actions.
Indeed, the numerical examples presented show that the PFEM is capable of han-
dling large deformation problems with no real limitations on the magnitude of the
deformation, and particularly suitable for the simulation of problems involving
both solid-like and fluid-like behaviour.
• Boundary identification
The key feature of the PFEM is that the computational domain is identified
at each time increment by the so-called alpha-shape method. The quality of the
detected domain depends on the choice of the parameter α, the ‘optimal’ value
of which is problem-dependent. Fortunately, the range of possible values of α is
in practice rather limited (1.0-2.0), and the ‘optimal’ value can be obtained by

136
 CHAPTER 9

simple trial and error.

• Mathematical programming formulations
Mathematical programming formulations for the dynamics analysis, as well
as for the contact analysis between deformable bodies and rigid boundaries, are
developed. Compared to the Newton-Raphson based FEM, the approaches es-
tablished in the mathematical programming ensure the convergence of the solu-
tion regardless of whether the available initial solution is close to unknown, new
states or not. Moreover, the singularity of the Mohr-Coulomb yield surface can be
treated naturally without any special consideration when an appropriate solution
scheme is applied.
• Sequential small deformation analysis for large deformation
problems
The mathematical programming formulations developed in this thesis are
based on the small deformation theory. Such formulations do lead to several
errors when they are used for analysing large deformation problems. However,
through both an error analysis and numerical tests, it is shown that this and
related errors can be minimised by decreasing the time step. For the kind of time
steps used in the simulations in this thesis, the errors appear to be negligible. As
such, the price to pay for the convenience of being able to operate with the usual
small deformation theory appears to be very small.
• Constitutive models for granular media
The proposed scheme is used to reproduce laboratory tests involving the col-
lapse of axisymmetric granular columns as well as an actual natural disaster,
the Yangbaodi landslide. In both simulations, the rate-independent rigid-plastic
model is used. Despite its simplicity, the simulated results agree well with the
reported ones, especially regarding the final deposit of the granular collapse prob-
lem. Compared to rheological models that appear to be somewhat problem de-
pendent, the simple rate-independent model appears quite suitable in practice.

137
Appendix A

Second-Order Cone
Programming

A standard second-order conic optimisation problem can be represented as fol-

lowing
minimise cT x + cf
subject to Ax = b, (A.1)
x i ∈ Ki , i = 1, 2, · · · , n

where the solution vector, x, is partitioned into n subvectors xi such that x =

(x1 , ..., xn )T , and each cone Ki has one of the following forms:

Quadratic cone

 v 
 um+2 
u X
Kq = x ∈ Rm+2 |x1 ≥ t x2j (A.2)
 
j=2

138
 APPENDIX A

Rotated quadratic cone

( m+2
)
X
Kr = x ∈ Rm+2 |2x1 x2 ≥ x2j , x1 , x2 ≥ 0 (A.3)
j=3

In contrast, the problems considered are cast in the following standard form:

1
maxmise cT x − xT Qx
2
subject to Ax = b, (A.4)
F (xi ) ≤ 0, i = 1, 2, · · · , n

which may be rewritten as

1
minimise − cT x + xT Qx
2
subject to Ax = b, (A.5)
F (xi ) ≤ 0, i = 1, 2, · · · , n

The first step in bringing the above problem into a standard SOCP form is to
recast the quadratic term in the objective function as a linear function subjected
to a quadratic constraint. Suppose Q is block-diagonal; the above equation is
equivalent to:

n
tj − cT x
P
minimise
j=1
subject to Ax = b, (A.6)
F (xi ) ≤ 0, i = 1, 2, · · · , n
1 T
x Q xj − tj ≤ 0, j = 1, 2, · · · , n
2 j j

By introducing a new set of variables that

1
uj = 1, y j = Qj2 xj (A.7)

139
 SECOND-ORDER CONE PROGRAMMING

we arrive n
tj − cT x
P
minimise
j=1
subject to Ax = b,
F (xi ) ≤ 0, i = 1, 2, · · · , n (A.8)
y Tj y j ≤ 2tj uj ,
1
y j = Qj2 xj ,
uj = 1, j = 1, 2, · · · , n

It is obvious that the second set of inequality constrains in problem (A.8) are all
rotated quadratic cones. Next, we need to cast the yield inequalities in terms of
second-order cones. In the following, the recasting of the Mohr-Coulomb criterion
and Drucker-Prager criterion in a plane strain case as conic quadratic constraints
is listed.
• Plane strain Mohr-Coulomb criterion

q
f (σx , σy , τxy ) = (σx − σy )2 + 4τxy
2 + (σ + σ ) sin φ − 2c cos φ ≤ 0
x y (A.9)

where x and y refer to the in-plane directions, φ is the friction angle, and c
represents the cohesion. This constraint can be cast in terms of a conic quadratic
constraint:  q 
3 2 2
ρ ∈ Kq = ρ ∈ R |ρ1 ≥ ρ2 + ρ3 (A.10)

with the transformation between ρ = (ρ1 , ρ2 , ρ3 ) and σ = (σx , σy , σz , σxy ) being

defined by
ρ = Dσ + d (A.11)

where    
− sin φ − sin φ 0 0 2c cos φ
D= 1 −1 0 0 and d =  0  (A.12)
   

0 0 0 2 0

• Plane strain Drucker-Prager criterion

p
f (I1 , J2 ) = αI1 + J2 − k ≤ 0 (A.13)

140
 APPENDIX A

where I1 is the first invariant of the stress:

I1 = σx + σy + σz (A.14)

and J2 is the second invariant of the deviatory stress:

1
J2 = [(σx − σy )2 + (σy − σz )2 + (σz − σx )2 ] + σxy
2 2
+ σyz 2
+ σxz (A.15)
6

In the plane strain case, we have

1
J2 = [(σx − σy )2 + (σy − σz )2 + (σz − σx )2 ] + σxy
2
(A.16)
6

The two material parameters α and k in (A.13) represent the friction coefficient
and the cohesion for the Drucker-Prager criterion, and are conveniently defined
as
tan φ 3c
α=p , k=p (A.17)
2
9 + 12 tan φ 9 + 12 tan2 φ
Similarly, the constraint (A.13) can be cast in terms of a conic quadratic con-
straint:  v 
 u 5
uX 
ρ ∈ Kq = ρ ∈ R3 |ρ1 ≥ t ρ2j (A.18)
 
j=2

with a new set of variables ρ = (ρ1 , ρ2 , ρ3 , ρ4 , ρ5 ) being defined by

ρ = Dσ + d (A.19)

where D and d take the form of

−α −α√ −α 0
   
√ k
1/ 6 −1/ 6 0√ 0 0
 √   
D= 0 1/ 6 −1/
√6 0, 0
d= (A.20)

 √
− 6 0 1/ 6 0 0
0 0 0 1 0

141
Appendix B

Formulations of Bar Element

Consider a linear bar element as shown in Figure B.1. Suppose the initial length
and the cross-sectional area of the bar are L and A respectively. The transfor-

η 2
ξ φ

Figure B.1: Truss element.

mation of nodal displacements between the local ξ η-coordinate system and the

142
 APPENDIX B

global x y-coordinate system is given by

    
1
u cos φ − sin φ 0 0 u1
 ξ    x
 1 
uη   sin φ cos φ 0 0  u1y 
 
 =   (B.1)
 2 
uξ   0 0 cos φ − sin φ u2x 
 
    
2
uη 0 0 sin φ cos φ u2y

where
x2 − x1
cos φ = p
(x2 − x1 )2 + y2 − y1 )2
(B.2)
y2 − y1
sin φ = p
(x2 − x1 )2 + y2 − y1 )2
Hooke’s law states
σ = Eε (B.3)

In above, σ is the axial stress, E is the Young’s modulus, and ε is the infinitesimal
axial strain given by
∂uξ
ε= (B.4)
∂ξ
where uξ is the displacement along the axis of the bar which can be approximated
linearly as
ξ ξ
uξ = (1 − )u1ξ + u2ξ = Nu1 u1ξ + Nu2 u2ξ (B.5)
L L
in which the shape functions are

ξ ξ
Nu1 = 1 − , Nu2 = (B.6)
L L

Note that the lateral displacement uη does not contribute to the stretch of the
bar within the linear theory. Substituting Eqs. (B.1) and (B.5) into Eq. (B.3)
gives

143
 FORMULATIONS OF BAR ELEMENT

 
u1
   x
 1
1 cos φ − sin φ 0 0 u 
 
1   y
ε = 1−
L L 0 0 cos φ − sin φ u2x 
 
 
u2y
 
u1 (B.7)
 x
  1
 uy 

1 1 1 1
= (1 − ) cos φ −(1 − ) sin φ cos φ − sin φ  
L L L L  2
ux 
 
u2y
= B û

where  
u1
 x
 1
 uy 
û = 
 2
 (B.8)
ux 
 
u2y

represent the displacements at the node 1 and node 2 of the bar element and
 
1 1
B = (1 − ) cos φ −(1 − ) sin φ 1 1 (B.9)
cos φ − sin φ
L L L L

Neglecting the self-weight of the bar, the incremental elastoplastic static analysis
of the bar system shown in Figure B.1 can be stated as

1 1
min max σn+1 B∆û − f T ∆û − σn+1 σn+1
∆û σn+1 2 E (B.10)
subject to F (σn+1 ) ≤ 0

The minimisation part of problem (B.10) may be solved first to yield

B T σn+1 − f = 0 (B.11)

144
 APPENDIX B

which is included as a constraint in the remaining maximisation part of the prob-

lem:
1 1
maximise − σn+1 σn+1
σn+1 2 E
subject to B T σn+1 − f = 0 (B.12)
F (σn+1 ) ≤ 0

The above optimisation problem is the one solved in each time step for the bar
system.

145
Appendix C

Cartesian and Cylindrical

Representations

C.1 Converting points between Cartesian and

cylindrical bases
Let P be a point. In a Cartesian coordinate system, the point can be identified
by specifying its coordinates (x, y, z), while in a cylindrical coordinate system its
coordinates are denoted by (r, θ, z). The relations between these two representa-
tions are given by
x = r cos θ
y = r sin θ (C.1)
z=z

146
 APPENDIX C

C.2 Converting vectors between Cartesian and

cylindrical bases
Let a be an arbitrary vector. In a Cartesian coordinate system, the vector is
denoted by
a = ax ex + ay ey + az ez (C.2)

where ex , ey and ez are unit basis vectors, and (ax , ay , az ) are components. When
we use the cylindrical bases (er , eθ , ez ), the vector can be expressed as

a = ar er + aθ eθ + az ez (C.3)

where (ar , aθ , az ) are the components in the cylindrical coordinate system. The
two sets of components are related by
   
a a
 x  r
ay  = R aθ  (C.4)
   
   
az az

where  
cos θ − sin θ 0
 
R =  sin θ cos θ 0 (C.5)
 
 
0 0 1

Derivation: To prove (C.4), we first relate the cylindrical bases (er , eθ , ez )

to the Cartesian bases (ex , ey , ez ). Recall the point P (x, y, z) in the Cartesian
coordinate system, the position vector of which is expressed as:

p = xex + yey + zez (C.6)

147
 CARTESIAN AND CYLINDRICAL REPRESENTATIONS

Substituting Eq. (C.1) into the above equation gives:

p = r cos θex + r sin θey + zez (C.7)

Then the bases er , eθ and ez can be calculated by

∂p ∂p ∂p
er = ∂r eθ = ∂θ ez = ∂z (C.8)
∂p ∂p ∂p
| | | | | |
∂r ∂θ ∂z

Using Eq. (C.7), we have

∂p
= cos θex + sin θey
∂r
∂p (C.9)
= −r sin θex + r cos θey
∂θ
∂p
= ez
∂z

and then
∂p
| |= 1
∂r
∂p (C.10)
| |= r
∂θ
∂p
| |= 1
∂z
Substituting Eqs. (C.9) and (C.10) into Eq. (C.8) gives

er = cos θex + sin θey

eθ = − sin θex + cos θey (C.11)

ez = ez

148
 APPENDIX C

Now, Eq. (C.3) can be re-written as

a = ar (cos θex + sin θey )

+ aθ (− sin θex + cos θey )

+ az ez
(C.12)
= (ar cos θ − aθ sin θ)ex

+ (ar sin θ + aθ cos θ)ey

+ az ez

Comparing Eq. (C.2) and Eq. (C.12) gives

ax = ar cos θ − aθ sin θ
ay = ar sin θ + aθ cos θ (C.13)
az = az

which can be re-written in the matrix form as

    
ax cos θ − sin θ 0 ar
ay  =  sin θ cos θ 0 aθ  (C.14)
    

az 0 0 1 az

149
 CARTESIAN AND CYLINDRICAL REPRESENTATIONS

C.3 Converting tensors between Cartesian and

cylindrical bases
Consider a second-order tensor σ. In a Cartesian coordinate system, the tensor
can be represented in a matrix form as

 
σxx σxy σxz
σ = σyx σyy σyz  (C.15)
 

σzx σzy σzz

which can also be expressed as

 
σrr σrθ σrz
σ = σθx σθθ σθz  (C.16)
 

σzx σzθ σzz

in a cylindrical coordinate system. The relation between Eq. (C.15) and Eq.
(C.16) is given by

   
σxx σxy σxz σrr σrθ σrz
 T
σyx σyy σyz  = R σθx σθθ σθz  R (C.17)
  

σzx σzy σzz σzx σzθ σzz

The above equation can be derived following the procedure in the previous section
by substituting Eq. (C.11) into the following dyadic form of the tensor

σ = σrr er ⊗ er + σrθ er ⊗ eθ + σrz er ⊗ ez

+σθr eθ ⊗ er + σθθ eθ ⊗ eθ + σθz eθ ⊗ ez (C.18)
+σzr ez ⊗ er + σzθ ez ⊗ eθ + σzz ez ⊗ ez

150
References

[1] Abbo, A.J. & Sloan, S.W. (1995). A smooth hyperbolic approximation
to the Mohr-Coulomb yield criterion. Computers and Structures, 54, 427–
441. 37

[2] Abbo, A.J., Lyamin, A.V., Sloan, S.W. & Hambleton, J.P. (2011).
A C2 continuous approximation to the Mohr-Coulomb yield surface. Inter-
national Journal of Solids and Structures, 48, 3001–3010. 37

[3] Abe, K., Soga, K., ASCE, M. & Bandara, S. (2014). Material point
method for coupled hydromechanical problems. Journal of Geotechnical and
Geoenvironmental Engineering, 140, 04013033. 18

[4] Ai, J., Chen, J., Rotter, J.M. & Ooi, J.Y. (2011). Assessment of
rolling resistance models in discrete element simulations. Powder Technol-
ogy, 206, 269–282. 5

[5] Akin, J.E. & Tezduyar, T.E. (2004). Calculation of the advective limit
of the supg stabilization parameter for linear and higher-order elements.
Computer Methods in Applied Mechanics and Engineering, 193, 1909–1922.
8

[6] Ambati, R., Pan, X., Yuan, H. & Zhang, X. (2012). Application
of material point methods for cutting process simulations. Computational
Materials Science, 57, 102–110. 18

151
 REFERENCES

[7] Andersen, E.D., Roos, C. & Terlaky, T. (2003). On implement-

ing a primal-dual interior–point method for conic quadratic optimization.
Mathematical Programming, 95, 249–277. 50, 111

[8] Andersen, S. & Andersen, L. (2010). Modelling of landslides with the

material-point method. Computational Geosciences, 4, 137–147. 18

[9] Andrade, J.E., Chen, Q., Lee, P.H., Avila, C.F. & Evans, T.M.
(2012). A constitutive law for dense granular flows. Journal of the Mechanics
and Physics of Solids, 60, 1122–1136. 83, 91

[10] Anitescu, M. (2006). Optimization-based simulation of nonsmooth rigid

multibody dynamics. Mathematical Programming A, 105, 113–143. 72

[11] Anitescu, M. & Hart, G.D. (2004). A constraint-stabilized time-

stepping approach for rigid multibody dynamics with joints, contact and
friction. International Journal for Numerical Methods in Engineering, 60,
2335–2371. 72

[12] Anitescu, M. & Tasora, A. (2010). An iterative approach for cone

complementarity problems for nonsmooth dynamics. Computational Opti-
mization and Applications, 47, 207–235. 72

[13] Balmforth, N.J. & Kerswell, R.R. (2005). Granular collapse in two
dimensions. Journal of Fluid Mechanics, 538, 399–428. 99

[14] Bathe, K.J. (1996). Finite Element Procedures. Prentice Hall. 7

[15] Belytschko, T., Lu, Y.Y. & Gu, L. (1994). Element free Galerkin
methods. International Journal for Numerical Methods in Engineering, 37,
229–256. 17

[16] Belytschko, T., Krongauz, Y., Organ, D., Fleming, M. & Krysl,
P. (1996). Meshless methods: An overview and recent developments. Com-
puter Methods in Applied Mechanics and Engineering, 139, 3–47. 2, 10, 17,
35

152
 REFERENCES

[17] Belytschko, T., Liu, W.K. & Moran, B. (2000). Nonlinear Finite
Elements for Continua and Structures. Wiley. 2, 7, 9

[18] Bernardini, F. & Bajaj, C. (1997). Sampling and reconstructing man-

ifolds using alpha-shapes. Tech. Rep. CSD-TR-97-013, Purdue University,
Department of Computer Science. 28

[19] Beuth, L., Wieckowski, Z. & Vermeer, P.A. (2011). Solution of

quasi-static large-strain problems by the material point method. Interna-
tional Journal for Numerical and Analytical Methods in Geomechanics, 35,
1451–465. 18

[20] Bicanic, N. (2004). Discrete element methods. In E. Stein, R. de Borst

& T.J.R. Hughes, eds., Encyclopedia of Computational Mechanics, vol. 1,
Wiley. xv, 7

[21] Bonet, J. & Kulasegaram, S. (2000). Correction and stabilization of

smooth particle hydrodynamics methods with applications in metal forming
simulations. International Journal for Numerical Methods in Engineering,
47, 1189–214. 14

[22] Bonet, J. & Wood, R.D. (2008). Nonlinear Continuum Mechanics for
Finite Element Analysis. Cambridge, 2nd edn. 79, 82

[23] Borja, R.I. (1991). Cam-clay plasticity. part ii: implicit integration of
constitutive equations based on a non-linear elastic stress predictor. Com-
puter Methods in Applied Mechanics and Engineering, 88, 225–240. 20

[24] Borja, R.I. & Lee, S.R. (1990). Cam-clay plasticity. part i: implicit
integration of elasto-plastic constitutive relations. Computer Methods in
Applied Mechanics and Engineering, 78, 49–72. 20

[25] Bui, H.H., Fukagawa, R., Sako, K. & Ohno, S. (2008). Lagrangian
mesh-free particles method (SPH) for large deformation and failure flows
of geomaterial using elastic-plastic soil constitutive model. International
Journal for Numerical and Analytical Methods in Geomechanics, 32, 1537–
1570. 14

153
 REFERENCES

[26] Bui, H.H., Fukagawa, R., Sako, K. & Wells, J.C. (2011). Slope
stability analysis and discontinuous slope failure simulation by elasto-plastic
smoothed particle hydrodynamics (SPH). Geotechnique, 61, 565–574. 14

[27] Bui, H.H., Sako, K. & Fukagawa, R. (2013). An improved SPH

method for saturated soils and its application to investigate the mechanisms
of embankment failure: Case of hydrostatic pore-water pressure. Interna-
tional Journal for Numerical and Analytical Methods in Geomechanics, 37,
31–50. 14

[28] Cai, M., Kaiser, P.K., Morioka, H., Minami, M., Maejima, T.,
Tasaka, Y. & Kurose, H. (2007). Flac/PFC coupled numerical simu-
lation of AE in large-scale underground excavations. International Journal
of Rock Mechanics and Mining Sciences, 44, 550–564. 20

[29] Capurso, M. & Maier, G. (1970). Incremental elastoplastic analysis

and quadratic optimization. Meccanica, 5, 107–116. 22, 37

[30] Carbonell, J., Onate, E. & Suarez, B. (2010). Modeling of ground

excavation with the Particle Finite-Element Method. Journal of Engineer-
ing Mechanics, 136, 455–463. 19, 26

[31] Chaaba, A. (2010). Plastic collapse in presence of non-linear kinematic

hardening by the bipotential and the sequential limit analysis approaches.
Mechanics Research Communications, 37, 484–488. 76

[32] Chen, R., Kuo, K.J., Chen, Y. & Ku, C. (2011). Model tests for
studying the failure mechanism of dry granular soil slopes. Engineering
Geology, 119, 51–63. 121

[33] Chen, W. & Qiu, T. (2012). Numerical simulations for large deforma-
tion of granular materials using smoothed particle hydrodynamics method.
International Journal of Geomechanics, 11, 127–135. 14, 99

[34] Cheng, J.H. & Kikuchi, N. (1986). A mesh rezoning technique for nite
element simulations of metal forming processes. International Journal for
Numerical Methods in Engineering, 23, 219–228. 35

154
 REFERENCES

[35] Contrafatto, L. & Ventura, G. (2004). Numerical analysis of Aug-

mented Lagrangian algorithms in complementary elastoplasticity. Interna-
tional Journal for Numerical Methods in Engineering, 60, 2263–2287. 22,
37

[36] Cremonesi, M., Frangi, A. & Perego, U. (2010). A Lagrangian finite

element approach for the analysis of fluid-structure interaction problmes.
International Journal for Numerical Methods in Engineering, 84, 610–630.
xv, 30, 31, 36

[37] Crosta, G.B., Imposimato, S. & Roddeman, D.G. (2009). Numerical

modeling of 2D granular step collapse on erodible and nonerodible surface.
Journal of Geophysical Research: Earth Surface, 114, F03020. 99

[38] Cruden, D.M. (1991). A simple definition of a landslide. In Bulletin of

the International Association of Engieering Geology, vol. 43, 27–29. 121

[39] Cubas, N., Leroy, Y.M. & Maillot, B. (2008). Prediction of thrusting
sequences in accretionary wedges. Journal of Geophysical Research: Solid
Earth, 113, B12412. 87

[40] Cundall, P.A. & Strack, O.D.L. (1979). A discrete numerical model
for granular assemblies. Geotechnique, 29, 47–65. 4

[41] Daphalapurkar, C.J., Vermeer, P.A. & H.Basson, A. (2005). The

modelling of anchors using the material point method. International Jour-
nal for Numerical and Analytical Methods in Geomechanics, 29, 879–895.
18

[42] Daphalapurkar, N.P., Lu, H., Coker, D. & Komanduri, R. (2007).

Simulation of dynamic crack growth using the generalized interpolation
material point (GIMP) method. International Journal of Fracture, 143,
79–102. 18

[43] Ding, Y., Gravish, N. & Goldman, D.I. (2011). Drag induced lift in
granular media. Physical Review Letters, 106, 028001. 87

155
 REFERENCES

[44] Donea, J., Huerta, A., Ponthot, J.P. & Rodriguez-Ferran, A.

(2004). Arbitrary Lagrangian Eulerian method. In E. Stein, R. de Borst
& T.J.R. Hughes, eds., Encyclopedia of Computational Mechanics, vol. 1,
Wiley. 2, 9

[45] Donea, J.C. (1984). A Taylor-Galerkin method for convective transport

problems. International Journal for Numerical Methods in Engineering, 20,
101–119. 8

[46] Duran, O., Kruyt, N.P. & Luding, S. (2010). Micro-mechanical anal-
ysis of deformation characteristics of three-dimensional granular materials.
International Journal of Solids and Structures, 47, 2234–2245. 5

[47] Edelsbrunner, H. & Mucke, E.P. (1994). Three dimensional alpha

shapes. ACM Transactions on Graphics, 13, 43–72. 27, 28

[48] Estrada, N., Azema, E., Radjai, F. & Taboada, A. (2011). Identifi-
cation of rolling resistance as a shape parameter in sheared granular media.
Physical Review E , 84, 011306. 6

[49] Forterre, Y. & Pouliquen, O. (2008). Flows of dense granular media.

Annual Review of Fluid Mechanics, 40, 1–24. 103

[50] Gadala, M.S., Oravas, G.A.E. & Dokainish, M.A. (1983). A con-
sistent eulerian formulation of large deformation problems in statics and
dynamics. International Journal of Non-Linear Mechanics, 18, 21–35. 35

[51] Gingold, R.A. & Monaghan, J.J. (1977). Smoothed particle

hydrodynamics-theory and application to non-spherical stars. Monthly No-
tices of the Royal Astronomical Society, 181, 375–389. 11

[52] Gitterle, M., Popp, A., Gee, M.W. & Wall, W.A. (2010). Finite
deformation frictional mortar contact using a semi-smooth newton method
with consistent linearization. International Journal for Numerical Methods
in Engineering, 84, 543–571. 22

156
 REFERENCES

[53] Gonzalez-Montellano, C., Ayuga, F. & Ooi, J.Y. (2011). Discrete

element modelling of grain flow in a planar silo: influence of simulation
parameters. Granular Matter , 13, 149–158. 95

[54] Harlow, F.H. (1964). The particle-in-cell computing method in fluid dy-
namics. Methods in Computational Physics, 3, 319–343. 18

[55] Heinrich, J.C., Huyakorn, P.S. & Zienkiewicz, O.C. (1997). An up-
wind finite element scheme for two-dimensional convective transport equa-
tion. International Journal for Numerical Methods in Engineering, 11, 131–
143. 8

[56] Herreros, M.I. & Mabssout, M. (2011). A two-steps time discretiza-

tion scheme using the SPH method for shock wave propagation. Computer
Methods in Applied Mechanics and Engineering, 200, 1833–1845. 14

[57] Hogg, A.J. (2007). Two-dimensional granular slumps down slopes.

Physics of Fluids, 19, 093301. 99

[58] Hu, Y. & Randolph, M.F. (1998). A practical numerical approach for
large deformation problems in soil. International Journal for Numerical and
Analytical Methods in Geomechanics, 22, 327–350. 2, 10, 28, 34, 35, 36, 76

[59] Huang, J., da Silva, M.V. & Krabbenhoft, K. (2013). Three-

dimensional granular contact dynamics with rolling resistance. Computers
and Geotechnics, 49, 289–298. 6, 23, 61, 67, 89, 99, 106, 111

[60] Huang, Y. & Dai, Z. (2014). Large deformation and failure simulations
for geo-disasters using smoothed particle hydrodynamics method. Engineer-
ing Geology, 168, 86–97. 14

[61] Huetink, J., Vreede, P.T. & Lugt, J.V.D. (2005). Progress in mixed
Eulerian-Lagrangian finite element simulation of forming processes. Inter-
national Journal for Numerical Methods in Engineering, 30, 1441–1457.
9

157
 REFERENCES

[62] Hungr, O. (1995). A model for the runout analysis of rapid flow slides,
debris flows, and avalanches. Canadian Geotechnical Journal , 32, 610–623.
121, 122

[63] Idelsohn, S., Onate, E. & Pin, F.D. (2003). A Lagrangian meshless
finite element method applied to fluid structure interation problems. Com-
puter and Structures, 81, 655–671. 27

[64] Idelsohn, S.R., Onate, E. & Pin, F.D. (2004). The particle finite
element method: a powerful tool to solve incompressible flows with free-
surfaces and breaking waves. International Journal for Numerical Methods
in Engineering, 61, 964–989. 19, 27, 28

[65] Issa, J.A. & Nelson, R.B. (1992). Numerical analysis of micromechani-
cal behaviour of granular materials. Engineering Computations, 9, 211–223.
5

[66] Iversion, R.M. (1997). The physics of debris flows. Review of Geophysics,
35, 245–296. 122

[67] Iwashita, K. & Oda, M. (1998). Rolling resistance at contacts in simula-

tion of shear and development by DEM. Journal of Engineering Mechanics,
124, 285–292. 5

[68] Jean, M. (1999). The non–smooth contact dynamics method. Computer

Methods in Applied Mechanics and Engineering, 177, 235–257. 5

[69] Jiang, M., Yu, H.S. & Harris, D. (2006). Discrete element modelling
of deep penetration in granular soils. International Journal for Numerical
and Analytical Methods in Geomechanics, 30, 335–361. 5

[70] Jiang, M., Zhu, H.H. & Harris, D. (2008). Classical and non-classical
kinematic fields of two-dimensional penetration tests on granular ground
by discrete element method analyses. Granular Matter , 10, 439–455. 5

[71] Johnson, G.R. & Beissel, S.R. (1996). Normalized smoothing func-
tions for SPH impact computations. International Journal for Numerical
Methods in Engineering, 9, 2725–2741. 14

158
 REFERENCES

[72] Jop, P., Forterre, Y. & Pouliquen, O. (2006). A constitutive law

for dense granular flows. Nature, 441, 727–730. 83, 103

[73] Kerswell, R.R. (2005). Dam break with Coulomb friction: a model for
granular slumping? Physics of Fluids, 17, 057101. 89, 99

[74] Kim, K.P. & Hu, H. (2006). Dynamic limit analysis formulation for im-
pact of structural members. International Journal of Solids and Structures,
43, 6488–6501. 76

[75] Krabbenhoft, K. & Damkilde, L. (2003). A general nonlinear opti-

mization algorithm for lower bound limit analysis. International Journal
for Numerical Methods in Engineering, 56, 165–184. 61, 111

[76] Krabbenhoft, K. & Lyamin, A.V. (2012). Computational cam clay

plasticity using second-order cone programming. Computer Methods in Ap-
plied Mechanics and Engineering, 209–212, 239–249. 22, 37, 61, 111

[77] Krabbenhoft, K., Lyamin, A.V. & Sloan, S.W. (2007). Formulation
and solution of some plasticity problems as conic programs. International
Jounal of Solids and Structures, 44, 1533–1549. 22, 37, 38, 61, 110, 111

[78] Krabbenhoft, K., Lyamin, A.V., Sloan, S.W. & Wriggers, P.

(2007). An interior-point method for elastoplasticity. International Journal
for Numerical Methods in Engineering, 69, 592–626. 22, 37, 61, 110, 111

[79] Krabbenhoft, K., Huang, J., da Silva, M.V. & Lyamin, A.V.
(2012). Granular contact dynamics with particle elasticity. Granular Mat-
ter , 14, 607–619. 6, 23, 37, 61, 67, 106, 111

[80] Krabbenhoft, K., Karim, M.R., Lyamin, A.V. & Sloan, S.W.
(2012). Associated computational plasticity schemes for nonassociated fric-
tional materials. International Journal for Numerical Methods in Engineer-
ing, 89, 1089–1117. 43, 61, 109, 111

[81] Krabbenhoft, K., Lyamin, A.V., Huang, J. & da Silva, M.V.

(2012). Granular contact dynamics using mathematical programming meth-

159
 REFERENCES

ods. Computers and Geotechnics, 43, 165–176. 6, 23, 37, 61, 67, 69, 71, 89,
91, 99, 106, 111

[82] Kruggel-Emden, H., Simsek, E., Rickelt, S., Rickelt, S., Wirtz,
S. & Scherer, V. (2007). Review and extension of normal force models
for the discrete element method. Powder Technology, 171, 57–173. 5

[83] Lacaze, L. & Kerswell, R.R. (2009). Axisymmetric granular collapse:

A transient 3D flow test of viscoplasticity. Physical Review Letters, 102,
108305. 89, 99

[84] Lacaze, L., Phillips, J.C. & Kerswell, R.R. (2008). Planar col-
lapse of a granular column: Experiments and discrete element simulations.
Physics of Fluids, 20, 063302. 89, 99

[85] Lagree, P.Y., Staron, L. & Popinet, S. (2011). The granular column
collapse as a continuum: validity of a two-dimensional Navier-Stokes model
with a µ(i)-rheology. Journal of Fluid Mechanics, 686, 378–408. 6, 89, 99,
103

[86] Lajeunesse, E., Mangeney-Castelneau, A. & Vilotte, J.P.

(2004). Spreading of a granular mass on an horizontal plane. Physics of
Fluids, 16, 2731–2381. 67, 89, 99, 112, 113, 129

[87] Lajeunesse, E., Monnier, J.B. & Homsy, G.M. (2005). Granular
slumping on a horizontal surface. Physics of Fluids, 17, 103302. 67, 99,
112, 113, 114, 117, 118, 129

[88] Larese, A., Rossi, R., Onate, E. & Idelsohn, S. (2008). Validation
of the particle finite element method (PFEM) for simulation of free surface
flows. Engineering Computations, 25, 385–425. 27

[89] Leu, S.Y. (2005). Convergence analysis and validation of sequential limit
analysis of plane-strain deformation. International Journal fro Numerical
Methods in Engineering, 64, 322–334. 76

[90] Li, S. & Liu, G.R. (2004). Meshfree Particle Methods. Springer. 14, 17,
35

160
 REFERENCES

[91] Li, S. & Liu, W.K. (2002). Meshfree and particle methods and their
applications. Applied Mechanics Review , 55, 1–34. 2, 10, 11, 14, 17

[92] Li, W.C., Li, H.J., Dai, F.C. & Lee, L.M. (2012). Discrete element
modeling of a rainfall-induced flowslide. Engineering Geology, 149–150,
22–34. xviii, xix, 5, 122, 123, 124, 127, 130, 131, 132

[93] Li, X. & Wan, K. (2011). A bridging scale method for granular materials
with discrete particle assembly-Cosserat continuum modeling. Computers
and Geotechnics, 38, 1052–1068. 20

[94] Li, X., Wu, W. & Zienkiewicz, O.C. (2000). A third-order taylor-
galerkin models for the simulation of two-dimensional unsteady free surface
flows. International Journal for Numerical Methods in Engineering, 47,
1689–1708. 8

[95] Li, X., Chu, X. & Feng, Y.T. (2005). A discrete particle model and
numerical modeling of the failure modes of granular materials. Engineering
Computations: International Journal for Computer-Aided Engineering and
Software, 22, 894–920. 5

[96] Li, X., Zhang, X. & Zhang, J. (2010). A generalized Hill’s lemma and
micromechanically based macroscopic constitutive model for heterogeneous
granular materials. Computer Methods in Applied Mechanics and Engineer-
ing, 199, 3137–3152. 20

[97] Li, X., He, S., Luo, Y. & Wu, Y. (2012). Simulation of the sliding pro-
cess of Donghekou landslide triggered by the Wenchuan earthquake using a
distinct element method. Environmental Earth Sciences, 65, 1049–1054. 5

[98] Libersky, L.D. & Petschek, A.G. (1991). Smoothed particle hydrody-
namics with strength of materials. Lecture Notes in Physics, 395, 248–257.
14

[99] Libersky, L.D., Randles, P.W., Carney, T.C. & Dickinson, D.L.
(1997). Recent improvements in SPH modelling of hypervelocity impact.
International Journal of Impact Engineering, 20, 525–532. 14

161
 REFERENCES

[100] Liu, G.R. (2009). Mesh Free Methods: Moving beyond the Finite Element
Method . CRC Press, 2nd edn. 14, 17, 35

[101] Liu, M.B. & Liu, G.R. (2010). Smoothed Particle Hydrodynamics (SPH):
an overview and recent developments. Archives of Computational Methods
in Engineering, 17, 22–76. 14

[102] Liyanapathirana, D.S. (2009). Arbitrary Lagrangian Eulerian based

finite element analysis of cone penetration in soft clay. Computers and
Geotechnics, 36, 851–860. 9

[103] Lube, G., Huppert, H.E., Sparks, R.S.J. & Hallworth, M.A.
(2004). Axisymmetric collapses of granular columns. Journal of Fluid Me-
chanics, 508, 175–199. 89, 99, 111, 112, 113, 114, 116, 118, 119, 129

[104] Lube, G., Huppert, H.E., Sparks, R.S.J. & Freundt, A. (2005).
Collapses of two-dimensional granular columns. Physical Review E , 72,
041301. 89, 99, 129

[105] Lube, G., Huppert, H.E., Sparks, R.S.J. & Freundt, A. (2011).
Granular column collapses down rough, inclined channels. Journal of Fluid
Mechanics, 675, 347–368. 99

[106] Lucy, L.B. (1977). A numerical approach to the testing of the fission
hypothesis. Astronomical Journal , 82, 1013–1024. 11

[107] Ma, X., Zhang, D.Z., Giguere, P.T. & Liu, C. (2013). Axisymmetric
computation of Taylor cylinder impacts of ductile and brittle materials us-
ing original and dual domain material point methods. International Journal
of Impact Engineering, 54, 96–104. 18

[108] Maier, G. (1968). A quadratic programming approach for certain classes

on nonlinear structural problems. Meccanica, 3, 121–130. 22, 37

[109] Maier, G. (1968). Quadratic programming theory for elastic perfectly

plastic structures. Meccanica, 3, 31–39. 22, 37

162
 REFERENCES

[110] Maier, G. (1984). Mathematical programming applications to structural

mechanics: some introductory thoughts. Engineering Structures, 6, 2–6.
22, 37

[111] Maillot, B. & Leroy, Y. (2006). Kink-fold onset and development

based on the maximum strength theorem. Journal of the Mechanics and
Physics of Solids, 54, 2030–2059. 87

[112] Mangeney, A., Roche, O., Hungr, O., Mangold, N., Faccanoni,
G. & Lucas, A. (2010). Erosion and mobility in granular collapse over
sloping beds. Journal of Geophysical Research: Earth Surface, 115, F03040.
99

[113] Mangeney-Castelnau, A., Bouchut, F., Vilotte, J.P., Laje-

unesse, E., Aubertin, A. & Pirulli, M. (2005). On the use of Saint
Venant equations to simulate the spreading of a granular mass. Journal of
Geophysical Research: Solid Earth, 110, B09103. 89, 99

[114] Martin, C.M. & Randolph, M.F. (2006). Upper-bound analysis of

lateral pile capacity in cohesive soil. Geotechnique, 56, 141–145. 73

[115] Mary, B.C.L., Maillot, B. & Leroy, Y.M. (2013). Deterministic

chaos in frictional wedges revealed by convergence analysis. International
Journal for Numerical and Analytical Methods in Geomechanics, 37, 3036–
3051. 87, 88

[116] Meriaux, C. (2006). Two dimensional fall of granular columns controlled

by slow horizontal withdrawal of a retaining wall. Physics of Fluids, 18,
093301. 99

[117] MiDi, G.D.R. (2004). On dense granular flows. The European Physical
Journal E , 14, 341–365. 103

[118] Mills, P., Loggia, D. & Tixier, M. (1999). Model for a stationary
dense granular flow along an inclined wall. Planetary and Space Science,
45, 733–738. 103

163
 REFERENCES

[119] Mollon, G., Richefeu, V., Villard, P. & Daudon, D. (2012). Nu-
merical simulation of rock avalanches: Influence of a local dissipative con-
tact model on the collective behavior of granular flows. Journal of Geophys-
ical Research: Earth Surface, 117, 1–19. 5

[120] Monaghan, J.J. (2012). Smoothed particle hydrodynamics and its diverse
applications. Annual Review of Fluid Mechanics, 44, 323–346. 14

[121] Moreau, J.J. (1987). Bounded variation in time. In P.D. Panagiotopoulos

& G. Strang, eds., Topics in Nonsmooth Mechanics, vol. 1, 1–74. 5

[122] Moreau, J.J. (1988). Unilateral contact and dry friction in finite free-
dom dynamics. In J.J. Moreau & P. Panagiotopoulos, eds., Non-Smooth
Mechanics and Applications, vol. 302 of CISM Courses and Lectures, 1–82,
Springer. 5

[123] Moreau, J.J. (1994). Some numerical methods in multibody dynamics:

application to granular materials. European Journal of Mechanics A/Solids,
13, 93–114. 5

[124] Nayroles, B., Touzot, G. & Villon, P. (1992). Generating the finite
element method: Diffuse approximation and diffuse elements. Computa-
tional Mechanics, 10, 307–318. 17

[125] Nazem, M., Sheng, D.C. & Carter, J.P. (2006). Stress integration
and mesh refinement for large deformation in geomechanics. International
Journal for Numerical Methods in Engineering, 65, 1002–1027. 2

[126] Nazem, M., Sheng, D.C., Carter, J.P. & Sloan, S.W. (2008). Arbi-
trary Lagrangian-Eulerian method for large-strain consolidation problems.
International Journal for Numerical and Analytical Methods in Geomechan-
ics, 32, 1023–1050. 2, 9

[127] Nazem, M., Carter, J.P. & Airey, D. (2009). Arbitrary Lagrangian-
Eulerian method for dynamic analysis of geotechnical problems. Computers
and Geotechnics, 36, 549–557. 2, 9

164
 REFERENCES

[128] Nguyen, V.P., Rabczuk, T., Bordas, S. & Duflot, M. (2008).

Meshless methods: A review and computer implementation aspects. Math-
ematics and Computers in Simulation, 79, 763–813. 2, 10, 17

[129] Oliver, J., Cante, J.C., Weyler, R., Gonzalez, C. & Hernandez,
J. (2007). Particle finite element methods in solid mechanics problems. In
E. Onate & D.R.J. Owen, eds., Computational Plasticity, 87–103, Springer.
19, 26

[130] Olovsson, L., Nilsson, L. & Simonsson, K. (1999). An ALE formula-

tion for the solution of two-dimensional metal cutting problems. Computer
and Structures, 72, 497–507. 9

[131] Onate, E., Idelsohn, S.R., Pin, F.D. & Aubry, R. (2004). The par-
ticle finite element method. an overview. International Journal Computa-
tional Methods, 1, 267–307. 18, 19, 26, 27

[132] Onate, E., Idelsohn, S.R., Celigueta, M.A. & Rossi, R. (2008).
Advances in the particle finite element method for the analysis of fluid-
multibody interaction and bed erosion in free surface flows. Computer Meth-
ods in Applied Mechanics and Engineering, 197, 1777–1800. 19, 27

[133] Onate, E., Celigueta, M.A., Idelsohn, S., Salazar, F. & Suarez,
B. (2011). Possibilities of the particle finite element method for fluid-soil-
structure interaction problems. Journal of Computational Mechanics, 48,
307–318. 27

[134] Onate, E., Idelsohn, S.R., Celigueta, M.A., Rossi, R., Marti,
J., Carbonell, J.M., Ryzhakov, P. & Suarez, B. (2011). Advances
in the particle finite element method (PFEM) for solving coupled problems
in engineering. In Particle-Based Methods. Fundamentals and Applications,
vol. 25 of Computational Methods in Applied Sciences, 47–54, Springer. 27,
33

[135] Ortiz, M. & Simo, J.C. (1986). An analysis of a new class of integration
algorithms for elastoplastic constitutive relations. International Journal for
Numerical Methods in Engineering, 23, 353–366. 20

165
 REFERENCES

[136] O’Sullivan, C., Bray, J.D. & Riemer, M.F. (2002). The influence
of particle shape and surface friction variability on macroscopic frictional
strength of rod-shaped particulate media. Journal of Engineering Mechan-
ics, 128, 1182–1192. 5

[137] Owen, P.J., Cleary, P.W. & Meriaux, C. (2009). Quasi-static fall of
planar granular columns:comparison of 2D and 3D discrete element mod-
elling with laboratory experiments. Geomechanics and Geoengineering: An
International Journal , 4, 55–77. 99

[138] Panteghini, A. & Lagioia, R. (2014). A single numerically efficient

equation for approximating the Mohr-coulomb and the Matsuoka-Nakai
failure criteria with rounded edges and apex. International Journal for Nu-
merical and Analytical Methods in Geomechanics, 38, 349–369. 37

[139] Peric, D., Hochard, C., Dutko, M. & Owen, D.R.J. (1996). Trans-
fer operators for evolving meshes in small strain elasto-plasticity. Computer
Methods in Applied Mechanics and Engineering, 137, 331–344. 28, 34

[140] Peric, D., Vaz, M. & Owen, D.R.J. (1999). On adaptive strategies for
large deformations of elasto-plastic solids at finite strains: computational
issues and industrial applications. Computer Methods in Applied Mechanics
and Engineering, 176, 279–312. 28, 34, 35

[141] Petraa, C., Gavreab, B., Anitescu, M. & Potraa, F. (2009). A

computational study of the use of an optimization-based method for sim-
ulating large multibody systems. Optimization Methods and Software, 24,
871–894. 72

[142] Pironneau, O. (1982). On the transport-diffusion algorithm and its appli-

cations to the navier-stokes equations. Numerische Mathematik , 38, 309–
332. 8

[143] Potts, D.M. & Zdravkovic, L. (2001). Finite Element Analysis in

Geotechnical Engineering: Application. Thomas Telford. 8

166
 REFERENCES

[144] Pudasaini, S.P. & Hutter, K. (1973). On the prediction of the reach
and velocity of catastrophic landslides. Rock Mechanics, 59, 153–177. 121

[145] Quezada, J.C., Breul, P., Saussine, G. & Radjai, F. (2014). Pen-
etration test in coarse granular material using contact dynamics method.
Computers and Geotechnics, 55, 248–253. 6

[146] Rabczuk, T. & Eibl, J. (2003). Simulation of high velocity concrete frag-
mentation using sph/mlsph. International Journal for Numerical Methods
in Engineering, 56, 1421–444. 14

[147] Radjai, F. & Richefeu, V. (2009). Contact dynamics as a nonsmooth

discrete element method. Mechanics of Materials, 41, 715–728. 6

[148] Raithatha, A. & Duncan, S.R. (2009). Rigid plastic model of incremen-
tal sheet deformation using second-order cone programming. International
Journal fro Numerical Methods in Engineering, 78, 955–979. 76

[149] Randolph, M.F. & Houlsby, G.T. (1984). The limiting pressure on a
circular pile loaded laterally in cohesive soil. Geotechnique, 34, 613–623. 73

[150] Richenmann, D. (2005). Runout prediction methods. In M. Jakob &

O. Hungr, eds., Debris-Flow Hazards and Related Phenomena, 305–324,
Springer. 121

[151] Roche, O., Attali, M., Mangeney, A. & Lucas, A. (2011). On the
run-out distance of geophysical gravitational flows: Insight from fluidized
granular collapse experiments. Earth and Planetary Science Letters, 311,
375–385. 99

[152] Rojek, J. (2007). Multiscale analysis using a coupled discrete/finite ele-

ment model. Interaction and Multiscale, 1, 1–31. 20

[153] Rondon, L., Pouliquen, O. & Aussillous, P. (2011). Granular col-

lapse in a fluid: Role of the initial volume fraction. Physics of Fluids, 23,
073301. 99

167
 REFERENCES

[154] Rosswog, S. (2009). Astrophysical smoothed particle hydrodynamics.

New Astronomy Reviews, 53, 78–104. 14

[155] Ryzhakov, P., Onate, E., Rossi, R. & Idelsohn, S. (2012). Improv-
ing mass conservation in simulation of incompressible flows. International
Journal for Numerical Methods in Engineering, 90, 1435–1451. 33

[156] Sabetamal, H., Nazem, M., Carter, J.P. & Sloan, S.W. (2014).
Large deformation dynamic analysis of saturated porous media with appli-
cations to penetration problems. Computers and Geotechnics, 55, 117–131.
9

[157] Sadeghirad, A., Brannon, R.M. & Burghardt, J. (2011). A con-

vected partivle domain interpolation technique to extend applicability of
the material point method for problems involving massive deformations.
International Journal for Numerical Methods in Engineering, 86, 1435–
1456. 18

[158] Scheidegger, A.E. (1973). On the prediction of the reach and velocity
of catastrophic landslides. Rock Mechanics, 59, 153–177. 121

[159] Shepard, D. (1968). A two-dimensional interpolation function for

irregularly-spaced data. In Proceedings of the 23rd ACM national confer-
ence, 517–524. 28, 34

[160] Shewchuk, J.R. (1996). Triangle: Engineering a 2D quality mesh gener-

ator and delaunay triangulator. In M.C. Lin & D. Manocha, eds., Applied
Computational Geometry: Towards Geometric Engineering, vol. 1148, 203–
222, Springer-Verlag. 33

[161] Shewchuk, J.R. (2002). Delaunay refinement algorithms for triangular

mesh generation. Computational Geometry: Theory and Applications, 21–
74. 33

[162] Simo, J.C. & Hughes, T.J.R. (1998). Computational Inelasticity.

Springer–Verlag. 20

168
 REFERENCES

[163] Simo, J.C. & Taylor, R.L. (1985). Consistent tangent operators for
rate-independent elastoplasticity. Computer Methods in Applied Mechanics
and Engineering, 48, 101–118. 20

[164] Sivaselvan, M.V., Lavan, O., Dargush, G.F., Kurino, H., Hyodo,
Y., Fukuda, R., Sato, K., Apostolakis, G. & Reinhorn, A.M.
(2009). Numerical collapse simulation of large-scale structural systems using
an optimization-based algorithm. Earthquake Engineering and Structural
Dynamics, 38, 655–677. 22, 37

[165] Sloan, S.W. (1987). Substepping schemes for the numerical integration
of elastoplastic stress-strain relations. International Journal for Numerical
Methods in Engineering, 24, 893–911. 20

[166] Sloan, S.W. & Booker, J.R. (1986). Removal of singularities in Tresca
and MohrCoulomb yield functions. Communications in Applied Numerical
Methods, 2, 173–179. 37

[167] Sloan, S.W., Abbo, A.J. & Sheng, D. (2001). Refined explicit inte-
gration of elastoplastic models with automatic error control. Engineering
Computations, 18, 121–154. 20

[168] Souli, M., Ouahsine, A. & Lewin, L. (2000). ALE formulation for flu-
idstructure interaction problems. Computer Methods in Applied Mechanics
and Engineering, 190, 659–675. 9

[169] Souloumiac, P., Leroy, Y.M., Maillot, B. & Krabbenhoft, K.

(2009, B09404). Predicting stress distributions in fold-and-thrust belts and
accretionary wedges by optimization. Journal of Geophysical Research, 114,
B09404. 87

[170] Souloumiac, P., Krabbenhoft, K., Leroy, Y.M. & Maillot, B.

(2010). Failure in accretionary wedges with the maximum strength theorem:
numerical algorithm and 2D validation. Computational Geosciences, 14,
793–811. 67, 87

169
 REFERENCES

[171] Springel, V. (2010). Smoothed Particle Hydrodynamics in astrophysics.

Annual Review of Astronomy and Astrophysics, 48, 391–430. 14

[172] Stadler, G. (2004). Semismooth newton and augmented lagrangian meth-

ods for a simplified friction problem. SIAM Journal Optimisation, 15, 39–
62. 22

[173] Staron, L. (2008). Mobility of long runout rock flows: a discrete numerical
investigation. Geophysical Journal International , 172, 455–463. 5

[174] Staron, L. & Hinch, E.J. (2005). Study of the collapse of granular
columns using two-dimensional discrete-grain simulation. Journal of Fluid
Mechanics, 545, 1–27. 6, 89, 99

[175] Staron, L. & Hinch, E.J. (2007). The spreading of a granular mass: role
of grain properties and initial conditions. Granular Matter , 9, 205–217. 6,
89, 99

[176] Stevens, A.B. & Hrenya, C.M. (2005). Comparison of soft-sphere mod-
els to measurements of collision properties during normal impacts. Powder
Technology, 154, 99–109. 5

[177] Taboada, A., Chang, K.J., Radjai, F. & Bouchette, F. (2005).

Rheology force transmission and shear instabilities in frictional granular me-
dia from biaxial numerical test using the contact dynamics method. Journal
of Geophysical Research: Solid Earth, 110, 1–24. 6

[178] Tanaka, H., Momozu, M., Oida, A. & Yamazaki, M. (2000). Simu-
lation of soil deformation and resistance at bar penetration by the distinct
element method. Journal of Terramechanics, 37, 41–56. 5

[179] Tang, C., Hu, J., Lin, M., Angelier, J., Lu, C., Chan, Y. & Chu,
H. (2009). The Tsaoling landslide triggered by the Chi-Chi earthquake,
Taiwan: Insights from a discrete element simulation. Engineering Geology,
106, 1–19. 5

170
 REFERENCES

[180] Tasora, A. & Anitescu, M. (2010). A convex complementarity ap-

proach for simulating large granular flows. Journal of Computational and
Nonlinear Dynamics, 5, 031004. 72

[181] Tasora, A. & Anitescu, M. (2011). A matrix-free cone complemen-

tarity approach for solving large-scale, nonsmooth, rigid body dynamics.
Computer Methods in Applied Mechanics and Engineering, 200, 439–453.
72

[182] Thompson, E.L. & Huppert, H.E. (2007). Granular column collapses:
further experimental results. Journal of Fluid Mechanics, 575, 177–186. 99

[183] Thompson, N., Bennett, M.R. & Petford, N. (2009). Analyses on

granular mas movement mechanics and deformation with distinct element
numerical modeling: implications for large-scale rock and debris avalanches.
Acta Geotechnica, 4, 233–247. 5

[184] Tian, Y., Cassidy, M.J., Randolph, M.F., Wang, D. & Gaudin,
C. (2014). A simple implementation of RITSS and its application in large
deformation analysis. Computers and Geotechnics, 56, 160–167. 2, 10, 76

[185] Vavourakis, V., Loukidis, D., Charmpis, D.C. & Papanastasiou,

P. (2013). Assessment of remeshing and remapping strategies for large de-
formation elastoplastic finite element analysis. Computers and Structures,
114-115, 133–146. 28, 34

[186] Venutelli, M. (1998). A third-order Taylor-Galerkin models for the sim-

ulation of two-dimensional unsteady free surface flows. Applied Mathmatical
Modelling, 22, 641–656. 8

[187] Vu-Quoc, L., Zhang, X. & Lesburg, L. (2001). Normal and tangential
force-displacement relations for frictional elasto-plastic contact of spheres.
International Journal of Solids and Structures, 38, 6455–6489. 5

[188] Wang, J.G. & Liu, G.R. (2002). A point interpolation meshless method
based on radial basis functions. International Journal for Numerical Meth-
ods in Engineering, 54, 1623–1648. 28, 34

171
 REFERENCES

[189] Wang, S., Khoo, B.C., Liu, G.R. & Xu, G.X. (2013). An arbitrary
LagrangianEulerian gradient smoothing method (GSM/ALE) for interac-
tion of fluid and a moving rigid body. Computers and Fluids, 71, 327–347.
9

[190] Wieckowski, Z. (2003). Modelling of silo discharge and filling problems

by the material point method. Task Quarterly, 7, 701–721. 18

[191] Wieckowski, Z. (2004). The material point method in large strain engi-
neering problems. Computer Methods in Applied Mechanics and Engineer-
ing, 193, 4417–4438. 18

[192] Yagawa, G. & Furukawa, T. (2000). Recent developments of free mesh

method. International Journal for Numerical Methods in Engineering, 47,
1419–1443. 2

[193] Yang, W.H. (1993). Large deformation of structures by sequential limit

analysis. International Journal of Solids and Structures, 30, 1001–1013. 76

[194] Yang, Z.X., Yang, J. & Wang, L.Z. (2012). On the influence of inter-
particle friction and dilatancy in granular materials: a numerical analysis.
Granular Matter , 14, 433–447. 5

[195] Yonekura, K. & Kanno, Y. (2012). Second-order cone programming

with warm start for elastoplastic analysis with von Mises yield criterion.
Optimization Engineering, 13, 181–218. 22, 37

[196] Yu, L., Hu, Y.X., Liu, J., Randolph, M.F. & Kong, X.J. (2012).
Numerical study of spudcan penetration in loose sand overlying clay. Com-
puters and Geotechnics, 46, 1–12. 2, 10, 76

[197] Zabala, F. & Alonso, E.E. (2011). Progressive failure of Aznalcollar

dam using the material point method. Geotechnique, 61, 795–808. 18

[198] Zenit, R. (2005). Computer simulations of the collapse of a granular col-

umn. Physics of Fluids, 17, 031703. 89, 99

172
 REFERENCES

[199] Zhang, D. & Whiten, W.J. (1999). A new calculation method for par-
ticle motion in tangential direction in discrete element simulations. Powder
Technology, 102, 235–243. 5

[200] Zhang, H.W., Xu, W.L., Di, S.L. & Thomson, P.F. (2002).
Quadratic programming method in numerical simulation of metal form-
ing process. Computuer Methods in Applied Mechanics and Engingeering,
191, 5555–5578. 22, 37

[201] Zhang, H.W., Zhong, W.X., Wu, C.H. & Liao, A.H. (2006). Some
advances and applications in quadratic programming method for numer-
ical modeling of elastoplastic contact problems. International Journal of
Mechanical Sciences, 48, 176–189. 22, 37

[202] Zhang, H.W., Wang, K.P. & Chen, Z. (2009). Material point
method for dynamic analysis of saturated porous media under external
contact/impact of solid bodies. Computer Methods in Applied Mechanics
and Engineering, 198, 1456–1472. 18

[203] Zhang, X., Sze, K.Y. & Ma, S. (2006). An explicit material point
finite element method for hyper-velocity impact. International Journal for
Numerical Methods in Engineering, 66, 689–706. 18

[204] Zhang, X., Krabbenhoft, K., Pedroso, D.M., Lyamin, A.V.,

Sheng, D., da Silva, M.V. & Wang, D. (2013). Particle finite ele-
ment analysis of large deformation and granular flow problems. Computers
and Geotechnics, 54, 133–142. 99

[205] Zhao, C.H. & Dai, F.C. (2007). Study on failure mechanism of a fill
slope in Shenzhen. The Chinese Journal of Geological Hazard and Control ,
18, 1–8. 124

[206] Zhao, J., Sheng, D., Rouainia, M. & Sloan, S.W. (2005). Explicit
stress integration of complex soil models. International Journal for Numer-
ical and Analytical Methods in Geomechanics, 12, 1209–1229. 20

173
 REFERENCES

[207] Zhou, S., Stormont, J. & Chen, Z. (1999). Simulation of geomembrane

response to settlement in landfills by using the material point method. Inter-
national Journal for Numerical and Analytical Methods in Geomechanics,
23, 1977–994. 18

[208] Zienkiewicz, O.C. & Pande, G.N. (1977). Some useful forms of
isotropic yield surfaces for soil and rock mechanics. In G. Gudehus, ed.,
Finite Elements in Geomechanics, 179–190, Wiley. 37

[209] Zienkiewicz, O.C. & Zhu, J.Z. (1992). The superconvergent patch re-
covery and a posteriori error estimates. International Journal for Numerical
Methods in Engineering, 33, 1331–1364. 28, 34, 35

[210] Zienkiewicz, O.C., Taylor, R.L. & Zhu, J.Z. (2005). The Finite
Element Method: Its Basis and Fundamentals. Butterworth-Heinemann,
6th edn. 7

174

View publication stats