thesis
TRANSCRIPT
Partitioned Analysis of
Nonlinear Soil-Structure Interaction
Hamid ZOLGHADR ZADEH JAHROMI
BSc, MSc
August 2009
A thesis submitted in fulfilment
of the requirements for the degree of
Doctor of Philosophy of the University of London
and the Diploma of Imperial College London
Department of Civil and Environmental Engineering
Imperial College London
London, SW7 2AZ
2
Abstract
This work has been primarily motivated by the lack of sophisticated monolithic tools
for modelling nonlinear soil–structure interaction problems, while recognising the
existence of advanced tools for nonlinear analysis of structure and soil in isolation.
Although coupled modelling of soil-structure interaction problems may be achieved
using a monolithic treatment, the partitioned treatment has been advocated as
offering major benefits in the context of coupled modelling of nonlinear soil-
structure interaction. Accordingly, the aim of this work has been to develop
advanced numerical methods for nonlinear coupling of soil-structure interaction
problems, where the partitioned approach is adopted as a framework for coupling
field-specific tools with minimal intrusion into codes.
In this respect, the partitioned approach for soil-structure interaction analysis has
been fully investigated in this work. Various coupling techniques are developed in
the context of soil-structure interaction analysis, and their computational
characteristics are discussed. Novel formulations for coupling soil-structure systems,
based on relaxation coupling methods and also utilizing the tangent stiffness matrix
of the partitioned sub-domains at the interface, are proposed, and their relative
performance is evaluated. The proposed approaches are believed to possess superior
convergence characteristics in comparison with existing coupling methods, rendering
these methods more general procedures for modelling soil-structure interaction
problems using coupled field-specific tools.
Based on the presented coupling algorithms, a novel simulation environment is
developed, utilising discipline-oriented solvers for nonlinear structural and
geotechnical analysis. Using the developed simulation environment in several
examples in which nonlinearity arises in both structure and soil, its applicability and
Abstract
3
high potential are demonstrated. Indeed, it is shown that the partitioned treatment is a
feasible and realistic approach for coupled modelling of nonlinear soil-structure
interaction problems, providing an integrated interdisciplinary computational
approach which combines the advanced features of both structural and geotechnical
modelling for a variety of challenging soil-structure interaction problems.
4
Acknowledgements
First and foremost I would like to express my deep and sincere gratitude to my
supervisors, Prof. Bassam A. Izzuddin and Dr. Lidija Zdravkovic. I owe a great deal
of appreciation for this thesis to Prof. Bassam A. Izzuddin for his insightful
supervision, invaluable guidance and unfailing support from very early stages of this
research, without which this thesis would not have been possible. I thank him for his
suggestions, refinements and constructive comments on full drafts. Likewise, I am
deeply grateful to Dr. Lidija Zdravkovic for her continuous encouragement and
support, substantive guidance, valuable comments on full drafts, and her overall
contribution to the final outcome.
Being part of the research groups in the Systems & Mechanics section and the Soil
Mechanics section at Imperial College has been a great experience. Thanks are due
to all academic staff and students for creating such an enjoyable and unique
environment. Special thanks are due to Mrs. Fionnuala Donavan the Postgraduate
Administrator, whose kindness will always be remembered. Many thanks are owed
to my friends and colleagues for the remarkable moments I have shared with them.
I am very grateful and deeply indebted to my parents for their inseparable love and
support throughout my life. Their support has been instrumental in all my
achievements thus far. I would like to thank my brother for being supportive and
caring.
Words fail me to express my deep appreciation to my wife Ahoura for her dedication
and persistence confidence in me. Without her love and understanding I would not
have been able to complete this thesis.
5
Table of Contents
Abstract 2
Acknowledgements 4
List of Tables 9
List of Figures 11
Notation 19
Chapter 1
1 Introduction 21
1.1 Soil-Structure Interaction 21
1.2 Coupled Systems 23
1.3 Treatment of Coupled Soil-Structure Interaction 24
1.4 Partitioned vs. Monolithic Approaches 26
1.5 Aims and Scope of Research 27
1.6 Layout of Thesis 29
Chapter 2
2 Literature Review 32
2.1 Background 32
2.2 Modelling of Soil-Structure Interaction 34
2.2.1 Field Elimination 35
2.2.2 Finite Element Method (FEM) 40
2.2.3 Integrated Modelling 42
2.2.4 Partitioned Analysis 46
2.2.5 Staggered Approach 48
2.2.6 Iterative Coupling 52
2.3 Concluding Remarks 60
Chapter 3
3 Staggered Approach 63
Table of Content
6
3.1 Introduction 63
3.2 Dynamic Analysis Formulation 65
3.3 Partitioning 67
3.4 Staggered Coupling Procedure 70
3.5 Predictors 71
3.6 Stability Analysis by Amplification Method 72
3.7 Stability Analysis of Staggered Coupling Scheme for a Test System 76
3.8 Accuracy 84
3.9 Example 84
3.10 Conclusion 91
Chapter 4
4 Iterative Coupling 92
4.1 Introduction 92
4.2 Iterative Coupling 96
4.2.1 Sequential Dirichlet-Neumann Iterative Coupling 97
4.2.2 Sequential Neumann-Dirichlet Iterative Coupling 98
4.2.3 Parallel Dirichlet-Neumann Iterative Coupling 100
4.2.4 Parallel Neumann-Dirichlet Iterative Coupling 102
4.2.5 Parallel Dirichlet-Dirichlet Iterative Coupling 103
4.2.6 Parallel Neumann-Neumann Iterative Coupling 105
4.3 Treatment of Interactive Boundary Conditions 106
4.4 Convergence of Iterative Coupling 107
4.4.1 Convergence of Sequential D-N with Trivial Update 107
4.4.2 Example 112
4.5 Simulation Environment 117
4.6 ADAPTIC 117
4.7 ICFEP 119
4.8 INTERFACE 121
4.9 Simulation Environment Architecture 122
4.10 Data Communication 125
4.11 Concluding Remarks 127
Table of Content
7
Chapter 5
5 Interface Relaxation 129
5.1 Introduction 129
5.2 Constant Relaxation 131
5.2.1 General Convergence Analysis 134
5.2.2 Convergence Studies 142
5.2.2.1 Example 1: Dynamic FEM-FEM Coupling 142
5.2.2.2 Example 2: Static FEM-FEM Coupling 154
5.3 Adaptive Relaxation 159
5.3.1 Convergence Analysis 159
5.3.2 Convergence Studies 163
5.3.2.1 SDOF Test System 163
5.3.2.2 Example 1: Dynamic FEM-FEM Coupling 164
5.3.2.3 Example 2: Static FEM-FEM Coupling 170
5.4 Soil-Structure Interaction Analysis 172
5.5 Concluding Remarks 184
Chapter 6
6 Reduced Order Method 187
6.1 Introduction 187
6.2 Condensed Interface Tangent Stiffness 188
6.2.1 Numerical Example 193
6.2.2 Analogy between Interface Relaxation and Condensed Interface Stiffness Approaches 198
6.3 Approximation of the Condensed Tangent stiffness 201
6.3.1 Condensed Interface Secant Stiffness Matrix 202
6.3.1.1 Example 1: Static FEM-FEM Coupling 204
6.3.1.2 Example 2: Dynamic FEM-FEM Coupling 208
6.3.1.3 Discussion on Nonlinear Analysis 213
6.3.2 Reduced Order Method 214
6.3.2.1 Singularity of .T
T TU U and .T
B BF F 217
6.3.2.2 Selective Addition or Replacement of the
Displacement/Force Vectors 219
Table of Content
8
6.3.2.3 Singularity of Approximated Stiffness/Flexibility
Matrices 227
6.3.2.4 Example 1: Static FEM-FEM Coupling 228
6.3.2.5 Example 2: Dynamic FEM-FEM Coupling 233
6.3.2.6 Example 3: Linear Soil-Structure Interaction 239
6.3.3 Mixed Reduced Order Method 242
6.4 Case Study: Nonlinear Soil-Structure Interaction Problem 249
6.5 Conclusion 253
Chapter 7
7 Case Studies 255
7.1 Introduction 255
7.2 Nonlinear Behaviour of Pitched-Roof Frame on Flexible Soil 256
7.3 Settlement Analysis of Multi-storey Five-bay Steel Frame 264
7.4 Building Response to an Adjacent Excavation 274
7.5 Coupled Modelling of Retaining Steel Sheet Piles 289
7.6 Conclusion 296
Chapter 8
8 Conclusion 297
8.1 Introduction 297
8.2 Conclusions 298
8.2.1 Staggered Approach 299
8.2.2 Iterative Coupling 300
8.2.3 Interface Relaxation 301
8.2.4 Reduced Order Method 302
8.2.5 Case Studies 304
8.3 Recommendations for Future Works 304
References 307
Appendix A: Structure of INTERFACE Data File 316
Appendix B: Structure of Communication Data File 317
Appendix C: Iterative Coupling Algorithms 318
Appendix D: Numerical Example 322
9
List of Tables
Table 2.1: Basic solution requirements satisfied by various methods of
geotechnical analysis 41
Table 2.2: Design requirements satisfied by various methods of
geotechnical analysis 41
Table 3.1: Single-step predictors for second-order dynamic systems 72
Table 5.1: Range of suitable and optimal relaxation parameter for different
1
2
m
m 145
Table 5.2: Range of suitable and optimal relaxation parameter for different
T
B
E
E 147
Table 5.3: Range of suitable and optimal relaxation parameter for different
/T BE E 155
Table 5.4: Number of coupling iterations with adaptive and optimum relaxation
for 1000 time-steps ( 0.01t s ) with a tolerance of 10-4 m 165
Table 5.5: Number of required coupling iterations with adaptive and optimum
relaxation 171
Table 5.6: Geometric and material properties of the partitioned soil-structure
system 175
Table 5.7: Convergence characteristics of constant and adaptive relaxation
schemes 177
Table 6.1: Coupling Procedure 193
List of Tables
10
Table 6.2: Required coupling iterations for different coupling schemes 206
Table 6.3: Number of required coupling iterations for 500 time-steps
( 0.01t s ) with a tolerance of 10-4 m 210
Table 6.4: Required coupling iterations for different coupling schemes 229
Table 6.5: Number of required coupling iterations for 500 time-steps
( 0.01t s ) with a tolerance of 1e-4 m 235
Table 6.6: Mixed reduced order method coupling procedures 248
Table 6.7: Geometric and material properties of the partitioned soil-structure
system 250
Table 6.8: Number of required coupling iterations for different coupling
schemes 250
Table 7.1: Geometric and material properties of the partitioned soil-structure
system 266
Table 7.2: Comparison of different coupling methods 267
Table 7.3: Geometric and material properties of the partitioned soil-structure
system 277
Table 7.4: Loading scenarios 278
Table 7.5: Comparison of different coupling methods 285
Table 7.6: Comparison of maximum bending moment in fully coupled
analysis 289
Table 7.7: Loading scenario in different incremental stages of the analysis 290
Table 7.8: Loading scenario in different incremental stages of the analysis 293
11
List of Figures
Figure 1.1: Field elimination treatment of soil-structure interaction 25
Figure 1.2: Partitioned treatment of soil-structure interaction 26
Figure 2.1: Illustrative SSI problem 33
Figure 2.2: Schematic diagram of a Winkler foundation 35
Figure 2.3: Schematic view of a) elasto-plastic b) visco-elastic foundation 38
Figure 2.4: Field elimination model for SSI analysis of lateral vibration of an elevated water tank 39
Figure 2.5: Idealization of jack-up models for SSI 43
Figure 2.6: a) Schematic illustration of the pile and soil model b) Schematic diagram of the structure model 44
Figure 2.7: a) Idealisation of the structure sub-domain b) Idealisation of the foundation system 44
Figure 2.8: a) Soil-structure coupled system b) Idealization of the coupled system 45
Figure 2.9: Idealization of the coupled soil-structure interaction system 45
Figure 2.10: Monolithic treatment of soil-structure interaction problem 46
Figure 2.11: Partitioned treatment of soil-structure interaction 47
Figure 2.12: Staggered coupling approach 48
Figure 2.13: Schematic representation of the staggered coupling approach 50
Figure 2.14: Staggered coupling of BEM-FEM for soil-structure interaction simulation of high speed train induced vibrations 50
Figure 2.15: Schematic of sequential iterative coupling approach 52
Figure 2.16: Coupled dam-reservoir-soil system 55
List of Figures
12
Figure 2.17: Flowchart of iterative coupling method 56
Figure 3.1: Staggered approach in dynamic soil-structure interaction 64
Figure 3.2: Partitioning and discretisation 68
Figure 3.3: Z-transformations 75
Figure 3.4: Example: geometric configuration and material response 85
Figure 3.5: Acceleration at the base 85
Figure 3.6: Problem partitioning 86
Figure 3.7: Rotation at the interface 87
Figure 3.8: Horizontal displacement at the interface 87
Figure 3.9: Horizontal displacement at the tip of the cantilever 88
Figure 3.10: Rotation at the interface 89
Figure 3.11: Horizontal displacement at the interface 89
Figure 3.12: Horizontal displacement at the tip of the cantilever 90
Figure 3.13: Rotation at the tip of the cantilever 90
Figure 4.1: Partitioned treatment of soil-structure interaction 93
Figure 4.2: Schematics of iterative coupling algorithms 95
Figure 4.3: Schematics of sequential D-N iterative coupling 98
Figure 4.4: Schematics of sequential N-D iterative coupling algorithms 100
Figure 4.5: Schematics of Parallel D-N iterative coupling 101
Figure 4.6: Schematics of Parallel N-D iterative coupling 103
Figure 4.7: Schematics of Parallel D-D iterative coupling 104
Figure 4.8: Schematics of Parallel N-N iterative coupling 106
Figure 4.9: a) coupled mass-spring system, b) partitioned sub-domain B , and c)
partitioned sub-domain T 113
Figure 4.10: Communication and synchronization between ADAPTIC and ICFEP, via INTERFACE 122
List of Figures
13
Figure 4.11: Schematics of the interaction sequence between the INFERCAE, ADAPTIC and ICFEP 124
Figure 4.12: Data exchange structure 126
Figure 5.1: Variation of error reduction factor against relaxation parameter 133
Figure 5.2: Convergence range and optimum relaxation parameter 141
Figure 5.3: Dynamic FEM-FEM coupling 143
Figure 5.4: Acceleration at the base 143
Figure 5.5: Partitioned sub-domains 144
Figure 5.6: Influence the effective mass ratio on convergence 146
Figure 5.7: Influence the effective stiffness ratio on convergence 147
Figure 5.8: Error reduction for different relaxation schemes (Time = 3.6s) 148
Figure 5.9: Error reduction for different relaxation schemes (Time = 3.5s) 149
Figure 5.10: Error reduction for different relaxation schemes (Time = 4.18s) 149
Figure 5.11: Rotation at the interface of sub-domain T 150
Figure 5.12: Horizontal displacement at the interface of T 151
Figure 5.13: Rotation at the interface of sub-domain B 151
Figure 5.14: Horizontal displacement at the interface of B 152
Figure 5.15: Rotation at the tip of the cantilever 152
Figure 5.16: Horizontal displacement at the tip of the cantilever 153
Figure 5.17: Rotation at the tip of the cantilever 5( 10 )Tolerance 153
Figure 5.18: a) Plane strain problem, b) Problem partitioning 154
Figure 5.19: Influence of relaxation parameter on convergence properties 156
Figure 5.20: Error reduction for different relaxation schemes (model M4) 157
Figure 5.21: Different discretatzions of model M4 158
Figure 5.22: Effect of mesh density on convergence 158
List of Figures
14
Figure 5.23: Error reduction at t=3.05s for model S8 166
Figure 5.24: Error reduction at t=3s for model S8 167
Figure 5.25: Error reduction at t=2.1s for model S8 167
Figure 5.26: Error reduction at t=4.9s for model S8 168
Figure 5.27: Error reduction at t=3.05s for model K1 168
Figure 5.28: Error reduction at t=1.85 s for model K1 169
Figure 5.29: Error reduction at t=4.94 s for model K1 169
Figure 5.30: Error reduction at t=4.98 s for model K1 170
Figure 5.31: Error reduction for adaptive and constant relaxation schemes 171
Figure 5.32: Linear soil-structure interaction 172
Figure 5.33: Monolithic vs. Partitioned Approach 173
Figure 5.34: Plane frame resting on soil 174
Figure 5.35: Plan view of the analysed building frame 174
Figure 5.36: Geometric configuration of considered frame 176
Figure 5.37: Influence of relaxation parameters on convergence properties 177
Figure 5.38: Error reduction for different relaxation schemes for the first load increment 178
Figure 5.39: Error reduction for different relaxation schemes for the fifth load increment 178
Figure 5.40: Error reduction for different relaxation coupling schemes for the last load increment 179
Figure 5.41: Convergence performance over full range of response 180
Figure 5.42: Vertical displacement of the soil surface 181
Figure 5.43: deformed mesh of the soil partitioned sub-domain 182
Figure 5.44: Vector plot of displacements in the soil partitioned sub-domain 182
Figure 5.45: Contours of stress level in soil partitioned sub-domain 183
List of Figures
15
Figure 5.46: Deformed shape and bending moment (kN.m) of the frame (scale=5) 183
Figure 6.1: Coupled spring system 194
Figure 6.2: a) Linear FEM-FEM coupled problem b) Partitioned sub-domain B c)
Partitioned sub-domain T 205
Figure 6.3: Error reduction of different coupling schemes for model A1 206
Figure 6.4: Error reduction of different coupling schemes for model A4 207
Figure 6.5: Error reduction of different coupling schemes for model A6 207
Figure 6.6: a) Coupled dynamic FEM-FEM problem b) Partitioned sub-domain Bc) Partitioned sub-domain T 209
Figure 6.7: Acceleration at the base 209
Figure 6.8: Horizontal displacements at the interface of T 211
Figure 6.9: Rotation at the interface of T 211
Figure 6.10: Comparison between different coupling schemes for Model C1 212
Figure 6.11: Comparison between different coupling schemes for Model C5 213
Figure 6.12: Error reduction for Model A1 230
Figure 6.13: Error reduction for Model A2 230
Figure 6.14: Error reduction for Model A3 231
Figure 6.15: Error reduction for Model A4 231
Figure 6.16: Error reduction for Model A5 232
Figure 6.17: Error reduction for Model A6 232
Figure 6.18: Error reduction of different schemes (Time=2.74s) for Model C1 236
Figure 6.19: Error reduction of different schemes (Time=3.63s) for Model C1 236
Figure 6.20: Error reduction of different schemes (Time=4.92s) for Model C1 237
Figure 6.21: Error reduction of different schemes (Time=4.42s) for Model C1 237
Figure 6.22: Error reduction of different schemes (Time=1.19s) for Model C1 238
List of Figures
16
Figure 6.23: Error reduction of different scheme (Time=2.74s) for Model C1 238
Figure 6.24: Coupled soil-structure interaction problem 239
Figure 6.25: Error reduction for different coupling schemes 240
Figure 6.26: Error reduction for different coupling schemes 241
Figure 6.27: Error reduction for different coupling schemes 247
Figure 6.28: Plane frame resting on soil 249
Figure 6.29: convergence behaviour over full range of response 251
Figure 6.30: Error reduction in the first load step 251
Figure 6.31: Error reduction in the 5th load step 252
Figure 6.32: Error reduction in the 6th load step 252
Figure 7.1: Pitched-roof steel frame resting on soil 256
Figure 7.2: Vertical displacement of the footing A 258
Figure 7.3: Vertical displacement of the footing E 258
Figure 7.4: Horizontal displacement of node B 259
Figure 7.5: Horizontal displacement of node D 259
Figure 7.6: Horizontal displacement of node C 260
Figure 7.7: Vertical displacement of node C 260
Figure 7.8: Variation of moment at node E 261
Figure 7.9: Variation of moment at node A 261
Figure 7.10: Variation of moment at node C 262
Figure 7.11: Deformed shape (scale=5.0) and bending moment (kN-m) in final load step for a) non-interactive case b) interactive case 263
Figure 7.12: Contours of stress level in soil sub-domain (at final increment) 263
Figure 7.13: Multi-Storey five-bay steel frame 265
Figure 7.14 : vertical settlement of Column C2 268
List of Figures
17
Figure 7.15: Vertical displacement profile of the soil surface 269
Figure 7.16: Displacement vectors in soil partitioned sub-domain 270
Figure 7.17: Contours of stress level in soil partitioned sub-domain 270
Figure 7.18: Bending moment (kN-m) at the base of C1 for different load-steps 271
Figure 7.19: Bending moment at the middle of beam B1 for different load-steps 271
Figure 7.20: Deformed shape (scale=5.0) and bending moment (kN-m) in final load step for a) linear Winkler foundation b) nonlinear partitioned analysis 272
Figure 7.21: Variation of bending moment along beam B1 273
Figure 7.22: Variation of bending moment along Column C1 273
Figure 7.23: Variation of vertical displacement along the beam B1 274
Figure 7.24: Plane frame resting on soil subject to ground excavation 275
Figure 7.25: Plan view of considered building 276
Figure 7.26: Geometric configuration of considered frame 276
Figure 7.27: Vertical settlement of the left footing for different load cases 279
Figure 7.28: Vertical settlement of the middle footing for different load cases 279
Figure 7.29: Vertical settlement of the right footing for different load cases 280
Figure 7.30: Cumulative vertical displacement of the ground surface 280
Figure 7.31: Cumulative horizontal displacement of excavation wall 281
Figure 7.32: Vertical displacement of ground surface for different excavation depths (Case 6) 282
Figure 7.33: Horizontal displacement of the excavation wall for different excavation depths (Case 6) 282
Figure 7.34: Vertical settlement of the left footing for different Le (Case 6) 283
Figure 7.35: Vertical settlement of the right footing for different Le (Case 6) 284
Figure 7.36: Vertical settlement of the right footing for different Le (Case 6) 284
Figure 7.37: Vectors of displacement in soil sub-domain in increment 6 (Case 6) 286
List of Figures
18
Figure 7.38: Contour plots of stress levels and plasticity induced in soil sub-domain in increment 6 (Case 6) 286
Figure 7.39: Vectors of displacement in soil sub-domain in increment 12 (Case 6) 287
Figure 7.40: Contour plots of stress levels and plasticity induced in soil sub-domain in increment 12 (Case 6) 287
Figure 7.41: Deformed shape (scale=5) and bending moment (kN-m) of structure for (a) 1st, (b) 6rd , (c) 7th and (d) 12th increment 288
Figure 7.42: Schematic diagram of the cantilever retaining wall 290
Figure 7.43: Vectors of accumulated displacements in soil sub-domain (final increment) 292
Figure 7.44: Deflection (scale: 5) and bending moment (kN-m/m) distribution of the retaining wall for (a) 1st, (b) 4th, (c) 7th, (d) 8th and (e) 9th increment 292
Figure 7.45: Schematic diagram of the propped retaining wall 293
Figure 7.46: Vectors of accumulated displacements in soil sub-domain (final increment) 294
Figure 7.47: Deflection (scale: 5) and bending moment (kN-m/m) distribution of the retaining wall for (a) 1st, (b) 2nd , (c) 3rd , (d) 4th and (e) 5th increment 295
Figure 7.48: Horizontal reaction at the lateral support 295
19
Notation
All symbols used in this thesis are defined where they first appear. The reader is
cautioned that some symbols denote more than one quantity; in such cases the
meaning should be clear when read in the context.
- Symbols of matrices and vectors are represented by [ ] and operators
respectively, with right hand side superscript and subscript (e.g. CBK , i
XU ).
- Equations are identified by numbers located on the right-most margin and
composed of two entries with the first entry indicating the Chapter in which the
equations appear.
A summary of operators and symbols is given in the following:
ΩT : partitioned structure sub-domain
ΩB : partitioned soil sub-domain
B : right-side subscript, denotes the partitioned sub-domain ΩB
T : right-side subscript, denotes the partitioned sub-domain ΩT
T : right-side superscript, denotes the transpose sign
I : right-side superscript, denotes the iteration number
n : right-side subscript, denotes the time/load increment number
[ ] : encloses terms of a matrix
: encloses terms of a vector
a : encloses absolute value of a
a : encloses second norm of a
I : identity matrix
Notation
20
XXU : displacement vector of the non-interface degrees of freedom in sub-
domain ΩX
XXF : force vector of the non-interface degrees of freedom in sub-domain ΩX
iXU : displacement vector of the interface degrees of freedom in ΩX
iXF : force vector of the interface degrees of freedom in sub-domain ΩX
CTK : condensed tangent stiffness matrices at the interface of the structure
sub-domain
CBK : condensed tangent stiffness matrices at the interface of the soil sub-
domain
Chapter 1
Introduction
1.1 Soil-Structure Interaction
There are numerous problems in Civil Engineering construction that require the use
of realistic models for the structure, the supporting soil and the soil-structure
interface. Examples include the assessment of various structures under earthquake
loading, the analysis of offshore jackup structures under extreme wave loading and
the evaluation of the structural damage due to excavations to name but a few. This
interaction can sometimes modify the stresses and deflections of the whole structural
system significantly. In fact, the structure with its loading conditions imposes
stresses and forces on the ground, which in turn deforms and as a consequence
transmits back additional forces and deformation to the structure. This process
continues until full equilibrium of the whole soil-structure system is satisfied, or
until both the soil and the structure fail in the case of excessive loading and
deformations of the system.
Numerical analysis, typically using the finite element or finite difference method, is
currently the most advanced tool available to facilitate soil and structure analysis.
Chapter 1 Introduction
22
With the development of these numerical tools, there are advanced techniques
employed for addressing the behaviour of structure and soil. Although there are
common techniques for structural and geotechnical modelling, the distinct demands
of the two fields meant that modelling has evolved differently in each domain, thus
leading to two modelling disciplines with distinctive high-level features.
In structural analysis, nonlinear modelling has evolved to address relevant issues
including i) the variety in structural form (e.g. frame, shell, membrane, whole
building), ii) the influence of geometric nonlinearity (e.g. large displacements/
rotations, buckling), iii) the nonlinear constitutive response of structural materials
(e.g. steel, concrete, composites) under serviceability and extreme loading conditions
(e.g. cracking, creep, plasticity, elevated temperature, high strain-rate). In
geotechnical analysis, on the other hand, developments have focused on i) the
constitutive modelling of different soils including pertinent nonlinear phenomena
(e.g. highly nonlinear elasto-plastic soil behaviour), ii) coupling of
mechanical/hydraulic/thermal/chemical processes in soils, iii) modelling of special
boundary conditions (e.g. excavation, construction, pore fluid pressure, non-
reflective boundary), iv) time dependent process such as consolidation and creep,
and v) problem reduction (e.g. Fourier series aided elements).
The above distinct challenges in the two fields have been reflected in the
development of discipline-oriented computational tools, which offer sophisticated
nonlinear modelling for their respective domain but, at best, a crude approximation
of the other, leading to poor representation of soil-structure interaction. This usually
means that structural analysis simplifies soil behaviour, while geotechnical analysis
simplifies structural behaviour using field elimination techniques. It is therefore a
real challenge to achieve the same amount of sophistication in modelling both the
soil and the structure in a single soil-structure interaction analysis.
In this respect, existing advanced discipline-oriented computational tools are
inadequate, on their own, for modelling a soil-structure interaction problem that
involves considerable nonlinearity in both the structure and the soil; rather such a
Chapter 1 Introduction
23
problem demands an integrated interdisciplinary computational model combining the
features of both structural and geotechnical modelling.
1.2 Coupled Systems
Here, the term system is identified as a set or assemblage of interacting or
independent entities, real or abstract, forming an integrated whole. The term
“system” is mainly used in this thesis for “physical systems” and in particular those
of importance in civil engineering, specifically soil-structure interaction problems.
Frequently there are number of functionally distinct physical systems interacting
with each other, where finding a solution of any one system is completely impossible
without considering the interaction with the other systems. Such systems are
categorized as “Coupled Systems” (Zienkiewicz & Taylor, 1991).
In general, coupled systems are analyzed by breakdown of the considered system
(domain) into several subsystems (sub-domains), where each sub-domain is
modelled according to its specific characteristics. This breakdown can be dictated
by: a) physical characteristics, b) functional characteristics and c) computational
characteristics (Fellippa et al., 2001).
One of the generic classes of coupled systems in the field of Civil Engineering
relates to the modelling of soil-structure interaction problems, where due to the
interaction between the physically distinct sub-domains, soil and structure, at the
interface, neither domain can be solved separately from the other without undue
simplifications, especially in the nonlinear range of response.
The multi-physics nature of soil-structure interaction modelling requires choosing
among a wide range of models, algorithms, and implementations to deal with
modelling of the soil and structure sub-domains, as well as implementing a coupling
procedure which can vary significantly in terms of the employed techniques in
Chapter 1 Introduction
24
treatment of the interaction effects at the interface of the coupled soil-structure
system.
In the following, general procedures available for coupled modelling of soil-structure
interaction are presented, considering that models and algorithms already exist for
nonlinear analysis of soil and structure on their own.
1.3 Treatment of Coupled Soil-Structure Interaction
In principle there are three possible approaches for dealing with soil-structure
interaction problems, namely: i) field elimination, ii) direct/monolithic/
simultaneous, and iii) domain decomposition/partitioned analysis (Rugonyi & Bathe,
2001).
In the field elimination treatment, one or more sub-domains of the coupled problem
are eliminated using a simple reduction technique, and the remaining sub-domain(s)
are considered under appropriate boundary conditions representing the eliminated
sub-domain(s). An example of implementing the field elimination technique in
modelling soil-structure interaction problems is illustrated in Figure 1.1, where the
soil sub-domain is replaced by spring type boundary conditions in order to model the
interaction effects.
Field elimination techniques are mostly restricted to simple problems that permit
efficient decoupling. In the context of nonlinear soil-structure interaction problems,
the simplification made in the modelling brings with it a loss of generality and
accuracy, due to the complexity of numerical modelling of this phenomenon.
Chapter 1 Introduction
25
Figure 1.1: Field elimination treatment of soil-structure interaction
In contrast with the field elimination treatment, the monolithic and partitioning
treatments as discussed in the following are general in nature. In the monolithic
treatment of soil-structure interaction, the whole problem is modelled as a single
computational entity and the solution of the complete system of equations of the
coupled problem is attenuated in one analysis scheme. Alternatively, in partitioned
analysis as shown in Figure 1.2, the soil-structure coupled system is physically
partitioned into soil and structure sub-domains. These partitioned sub-domains are
then computationally treated as isolated entities, and the response of each sub-
domain is calculated using already developed soil and structural solvers. In this
procedure, the interaction effects are viewed as force/displacement effects at the
interface of soil-structure system which are communicated between the individually
modelled soil and structure sub-domains using prediction, substitution and
synchronization techniques.
Chapter 1 Introduction
26
Figure 1.2: Partitioned treatment of soil-structure interaction
1.4 Partitioned vs. Monolithic Approaches
Although coupled modelling of soil-structure interaction problems may be achieved
using a monolithic treatment (as is likely with developing new software and solution
methods for different soil-structure interaction applications), a partitioned treatment
with different partitioned sub-domains modelled as separate computational entities,
amongst which interaction effects are exchanged, offers major benefits in the context
of nonlinear soil-structure interaction. Such benefits include (Felippa et al., 2001;
Rugonyi & Bathe, 2001; Lai, 1994; Hagen & Estorff, 2005):
Allowing field-specific discretisation and solution procedures that have
proven performance for each partitioned domain. In this respect, the
partitioned soil and structure sub-domains can be modelled and solved to the
desired degree of sophistication and accuracy with techniques and algorithms
that are known to perform well for each individual sub-domain.
Facilitating the reuse of existing nonlinear discipline oriented soil and
structural analysis solvers (private, public or commercial), with all the
resource savings that this brings.
Chapter 1 Introduction
27
Project breakdown advantages in the analysis of complex systems.
Suitable for research environments, where software development is more
likely to be cyclical. In other words, the existing discipline-oriented
programs developed for individual soil and structural sub-domains, can be
reused and developed further, while all these new developments could be
easily incorporated in soil-structure interaction analysis.
Facilitating parallel computing thorough problem partitioning.
However, despite the significant potential benefits of the partitioned approach, this
method should be formulated and implemented with great care, since the stability
and convergence characteristics of coupling algorithms are typically conditional and
problem dependent. Moreover, gain in computational efficiency using the partitioned
approach over monolithic treatment is not guaranteed.
1.5 Aims and Scope of Research
This work aims at developing an advanced generic numerical method for nonlinear
coupling of static and dynamic soil-structure interaction problems. In this respect,
this work is primarily motivated by the lack of sophisticated monolithic tools for
modelling nonlinear soil–structure interaction. Recognising the existence of
advanced tools for nonlinear analysis of structure and soil, in isolation, the
partitioned approach is adopted as a framework for coupling field-specific tools with
minimal intrusion. Accordingly, in adopting the partitioned treatment, the focus is on
providing an advanced capability for modelling nonlinear soil–structure interaction
with existing field-specific codes rather than on achieving superior computational
performance in comparison with the monolithic treatment.
Due to the aforementioned advantages of the partitioned approach in soil-structure
interaction analysis (especially large-scale nonlinear problems), this method is fully
investigated and developed. It is therefore the aim of this thesis to develop and
investigate the mathematical and computational characteristics of various domain
Chapter 1 Introduction
28
decomposition methods in the context of nonlinear soil-structure interaction analysis,
where particular emphasis is given to utilizing the finite element method in
discretising the partitioned sub-domains.
In this regard, novel formulations for coupling soil-structure systems, based on
relaxation coupling methods and also utilizing the tangent stiffness matrix of the
partitioned sub-domains at the interface, are proposed. The proposed approaches are
believed to possess superior convergence characteristics in comparison with existing
coupling methods, rendering these methods more general procedures for modelling
real soil-structure interaction problems using coupled field-specific tools.
Based on the presented coupling algorithms a novel simulation environment,
utilising discipline-oriented solvers for nonlinear structural and geotechnical
analysis, is developed. This is carried out through coupling of two powerful in-house
nonlinear analysis programs developed at Imperial College, ADAPTIC (Izzuddin,
1991) and ICFEP (Potts & Zdravkovic, 1999). Each of the programs has been
developed for sophisticated analysis of nonlinear structural and soil behaviour,
respectively, and has been used both in research and consulting. Although the
developed coupling method will be applied to the coupling of ADAPTIC and ICFEP,
a higher level objective is for this method to be generally applicable to the coupling
of other existing nonlinear soil and structural software. As a result, a generic
numerical tool for coupling nonlinear finite element systems, focusing on nonlinear
soil-structure interaction problems, is developed.
The developed simulation environment is subsequently used to demonstrate the
performance characteristics and merits of the various presented and proposed
algorithms, with particular reference to nonlinear-soil structure interaction problems.
The main focus of the initial numerical studies is on the criteria for algorithmic
stability, accuracy and convergence as well as computational costs, where some of
the presented techniques are investigated first for linear analysis prior to their
exclusion/adoption/generalisation for nonlinear analysis.
Chapter 1 Introduction
29
Finally, the developed tool is used in a number of case studies involving nonlinear
soil-structure interaction in which nonlinearity arises in both the structure and the
soil, thus leading to important conclusions regarding the adequacy and applicability
of the proposed coupling approaches for such problems, as well as the prospects for
further enhancements.
1.6 Layout of Thesis
Chapter 1 (Introduction): This chapter describes the general background and
provides the basic motivation behind investigation and development of the
partitioned treatment of soil-structure interaction problems in this thesis.
Chapter 2 (Literature Review): In this chapter an extensive critical review of
various available modelling approaches, applicable to soil-structure interaction
analysis, is presented. The advantages and disadvantages of different methods are
discussed. Moreover, a comprehensive overview of current coupling algorithms is
provided. Staggered and iterative coupling algorithms and their respective origins are
introduced. The shortcomings and benefits of current iterative coupling techniques
such as interface relaxation and reduced order method are identified. The need for
further enhancement and development of iterative coupling algorithms is established
in the context of soil-structure interaction.
Chapter 3 (Staggered Approach): The applicability of the staggered approach for
coupling dynamic FEM-FEM problems is investigated, with particular focus on soil-
structure interaction. The mathematical procedure for evaluating the stability of the
staggered approach is presented. Stability is conditional for this approach, where the
convergence characteristics may be enhanced by employing modifications in the
formulation of the partitioned sub-domains or change in the governing equations by
augmentations. It is concluded that stability and accuracy requirements of the
staggered approach typically demand excessively small time steps for the types of
problem under consideration, rendering this scheme computationally prohibitive.
Chapter 1 Introduction
30
Chapter 4 (Iterative Coupling): In this chapter, different iterative coupling
methods in partitioned analysis of soil-structure interaction problems are presented.
It is assumed that the overall domain is divided into physical partitions consisting of
soil and structure sub-domains, and the coupling of separately modelled sub-domains
is undertaken via the presented iterative algorithms. It is established that an
important feature of the proposed coupling approaches that needs to be addressed
comprehensively is the convergence behaviour of the scheme, which is directly
dictated by the chosen update technique during successive iterations. The developed
coupling simulation environment, utilising discipline-oriented solvers for nonlinear
structural and geotechnical analysis, is explained in detail. The software architecture
of the developed simulation environment, which is based on Sequential Dirichlet-
Neumann iterative coupling algorithms, is described and the structure of the data
exchange between the field-specific codes is elaborated.
Chapter 5 (Interface Relaxation): The use of interface relaxation coupling
techniques in domain decomposition analysis of soil-structure interaction is
investigated in this chapter. The general mathematical convergence characteristics of
constant and adaptive interface relaxation schemes are established. The applicability
and merits of the presented coupling techniques are highlighted trough various case
studies. It is shown that the adaptive scheme improves the convergence
characteristics in both linear and nonlinear analysis significantly, though the need for
further enhancements is also established.
Chapter 6 (Reduced Order Method): In this chapter, various domain
decomposition methods for nonlinear analysis of soil-structure interaction problems
based on approximating the condensed interface stiffness matrix are proposed. It is
shown that by using the condensed tangent interface stiffness matrices of the
partitioned sub-domains in the update of boundary conditions, superior convergence
characteristics could be achieved. This brings the performance of the proposed
coupling approach close to the monolithic treatment. It is proposed that suitable
approximations for the condensed stiffness matrices can be constructed during
successive coupling iterations to avoid explicit assemblage of the stiffness matrices.
Chapter 1 Introduction
31
In this respect, various novel reduced order method coupling algorithms are
developed and their performance is examined. The applicability of the presented
coupling techniques is demonstrated for nonlinear soil-structure interaction analysis
via an example investigating the interactive behaviour of a plane frame and a
supporting soil system.
Chapter 7 (Case Studies): The applicability of the partitioned treatment and its
great potential towards performing realistic nonlinear soil-structure interaction
analysis are illustrated through various case studies in this chapter.
Chapter 8 (Conclusion): In this chapter, detailed conclusions on the applicability
and efficiency of the developed coupling approaches as well as the benefits of the
developed soil-structure interaction approach are drawn. Several issues are also
identified for future research.
Chapter 2
Literature Review
2.1 Background
Soil-structure interaction problems relate to two-way interactive coupled systems,
where the state of structural deformations and stresses depend on the earth pressures
and movements, while simultaneously the earth pressures depend on the loading and
deformations of the structure. Accordingly, analysing such problems requires
simultaneous modelling of the structure and retaining/retained soil, using reasonably
accurate and computationally efficient techniques. Examples of such problems
include the assessment of various structures under static and dynamic loading,
offshore jackup structures under extreme wave loading, structures subject to ground
movement or excavations, reinforced slopes and embankments, retaining walls, and
tunnels to name but a few.
For example, consider the frame of Figure 2.1 supported on soil. As the loading on
the frame is symmetric, it is clear that there is a higher concentration of load over the
middle column and its support. This results in the soil below the middle footing
sustaining greater vertical settlements than the adjacent footings. In turn, framing
Chapter 2 Literature Review
33
action causes redistribution of the loads to the end columns due to the generated
differential settlements. These differential settlements among various parts of the
structure alter both the axial forces and the moments in the structural members
considerably, where the amount of redistribution of loads depends upon the rigidity
of the structure and the load-settlement characteristics of soil (Dutta & Roy, 2002).
This considerable influence of the structural rigidity on the interactive response has
been qualitatively explained in the literature long back (Taylor, 1964).
Figure 2.1: Illustrative soil-structure interaction (SSI) problem
Indeed, such interaction between structure and soil at their common interface leads
to an actual response for both physical sub-domains, which is considerably different
from what is obtained for each sub-domain in isolation. In view of the
interdependence between the responses of the two sub-domains, the frame structure,
its foundations and the soil on which it rests constitute together a coupled system.
The effect of soil–structure interaction under both static and dynamic loading has
attracted significant research interest over many years due to its important role in the
analysis of real civil engineering problems. In this respect, research by
Noorzaei et al. (1993), Allam et al. (1991), Roy & Dutta (2001) and Potts &
Zdravkovic (2001) on static soil-structure interaction analysis, and also research by
Chapter 2 Literature Review
34
Stewart et al. (1999) and Dutta et al. (2004) on dynamic soil-structure interaction
analysis, demonstrates the significant effect of soil-structure interaction and its
importance in predicting the overall coupled response. There are also numerous
reported case histories in the literature which demonstrate the significant influence of
considering soil-structure interaction effects in civil engineering design and practice,
such as the destruction of buildings in Carcas earthquake in 1967 (Dowrick, 1977),
the enormous damage to structures in the Hanshin and Awaji areas during Kobe
earthquake in 1999 (Inaba et al., 2000), the case of leaning tower of Pisa (Burland &
Potts, 1994), and the reported damage to several building structures due to
differential settlements (Charles & Skinner, 2004). All of these researches deliver an
important message that considering soil-structure interaction effects can be essential
for a multitude of real civil engineering construction problems.
The primary expected results from a coupled soil-structure interaction (SSI) analysis
are the stresses and displacements of the structure and the soil sub-domains. SSI
analysis tools can be used in design to calculate the stresses and deformations of the
structure and soil sub-domain to compare with the allowable values and, if
necessary, to modify the system configuration so as to meet specific design criteria
of serviceability, safety and economy. In addition, SSI tools can be used for the
assessment of existing coupled systems, enabling the prediction of potential damage
and the proposal of strengthening and repair measures.
2.2 Modelling of Soil-Structure Interaction
Towards this end, modelling the material behaviour of both structure and soil,
modelling the geometric nonlinearity of structure under extreme events, modelling of
special boundary conditions such as excavation in soil, etc., are amongst the most
central issues that should be considered in a reliable SSI analysis.
Works by Jardine et al. (1986), Wrana (1993), Noorzaei et al. (1993), Noorzaei et al.
(1995a), Inaba et al. (2000), Estorff & Firuzian (2000), Krabbenhoft et al. (2005)
and Bourne-Webb et al. (2007), concentrating on the nonlinear modelling of soil-
Chapter 2 Literature Review
35
structure interaction problems, have emphasised the importance of employing
accurate material modelling in nonlinear SSI analysis. Indeed, completely
misleading predictions may be obtained, unless the interactive study of the soil–
structure is conducted by considering accurate models of both soil and structure for
the problem under consideration.
There exist various models, algorithms and implementations for modelling SSI
problems. These all depend on the availability and limitations of computational tools
to carry out the SSI analysis, the physical effects that are to be captured, the desired
degree of accuracy, etc. In the following some of these modelling techniques are
presented and their advantages and disadvantages are briefly discussed.
2.2.1 Field Elimination
For a long time, field elimination techniques have been favoured over fully coupled
analysis of soil-structure interaction problems, largely due to their computational
simplicity. In this respect, numerous idealisation and elimination techniques
representing either soil or structure, depending on their relative significance for the
problem under consideration, have been employed and evaluated. In most cases,
however, the lack of accuracy of field elimination approaches is apparent.
Figure 2.2: Schematic diagram of a Winkler foundation
One of the most common field elimination techniques for idealization of the soil sub-
domain is Winkler model (Winkler, 1867), where the soil medium is removed and
the soil-structure interface is modelled by a system of independent and closely
spaced linear springs as shown in Figure 2.2. Clearly, in this type of modelling the
Chapter 2 Literature Review
36
deformation at the interface is only confined to the loaded region, which is simply
incorrect.
Numerous studies for SSI analysis have been carried out using the Winkler
hypothesis, including studies by Terzaghi (1955), Vesic (1961), Brown et al. (1977)
and Bowles (1996). Despite the simplicity and low computational cost of the
Winkler idealization, the fundamental problem is the determination of the stiffness of
the associated elastic springs replacing the soil sub-domain. As a coupled problem,
the value of the sub-grade reaction is not only dependent on the sub-grade but also
on the parameters of the loaded area as well. However, the sub-grade reaction is the
only parameter in Winkler idealization, thus great care is required in determination
of the sub-grade parameter (using plate load test, consolidation test, triaxial test or
CBR tests (Dutta & Roy, 2002)).
As expected with field elimination techniques, there are serious limitations with
Winkler type idealizations. The major obvious pitfall of such an approach is that it
provides no or very little information regarding the stress and deformation state
within the soil mass. Dutta and Roy (2002) summarises some other primary
limitations of such an idealization:
‘…the basic limitation of Winkler hypothesis lies in the fact that this model
cannot account for the dispersion of the load over a gradually increasing influence
area with increase in depth. Moreover, it considers linear stress–strain behaviour of
soil. The most serious demerit of Winkler model is the one pertaining to the
independence of the springs. So the effect of the externally applied load gets
localized to the sub-grade only to the point of its application. This implies no
cohesive bond exists among the particles comprising soil medium.’ (Dutta & Roy,
2002: p.1582)
Beside the Winklerian approach, there is also a conceptual approach for physical
idealization of the infinite soil sub-domain using the theory of continuum mechanics
(Harr, 1966). The continuum idealization of the soil sub-domain, originally
introduced by Odhe and the research work of Boussinesq (Bowles, 1996), is used to
Chapter 2 Literature Review
37
analyse the problem of a semi-infinite, homogeneous, isotropic, linear elastic solid
subjected to a concentrated force acting normal to the plane boundary, using the
theory of elasticity. Although in the continuum idealization, the soil sub-domain is
typically considered as a semi-infinite and isotropic medium, the effect of soil
layering and anisotropy may be conveniently accounted for in the analysis (Carrier &
Christian, 1973; Stavridis, 2002).
Such approaches in principle overcome the inadequacy of the Winkler spring model
to transmit lateral shear stresses, where a load acting on the soil surface cannot
produce any settlement except over its specific loading area. In addition, compared
to the Winklerian approach, the continuum approaches provide much more
information on the stress and deformation state within the soil mass. Furthermore,
these approaches are more realistic as they are based on explicit data of geotechnical
investigations, in contrast to the Winkler method, where the sub-grade reaction
modulus does not represent a soil property.
Nevertheless, there are major drawbacks associated with the elastic continuum
approach such as its inaccuracy in reactions calculated at the edges of the foundation.
Furthermore, although solutions for practical problems idealizing the soil sub-
domain as elastic continuum are available for some limited cases, this more
sophisticated method has never gained popularity among the designers, mainly
because of it leading to various mathematical intricacies and its inherent inability for
a direct analytical implementation for design purposes. Adding to these, the inability
of evaluating complex material behaviours, such as the nonlinear and elasto plastic
behaviour of soil, severely limits the application of this model in practice (Dutta &
Roy, 2002).
The mathematically and computationally attractive but physically inadequate
Winkler hypothesis has attracted several attempts over time to develop modified
models to overcome its shortcomings. Amongst many are Hetenyi’s foundation
(Hetenyi, 1946), Pasternak foundation (Wang et al., 2001), where the continuity in
the soil medium is modelled by introducing some sort of structural elements, and
Chapter 2 Literature Review
38
that of Kurian et al. (2001), where the springs are intermeshed so that the
interconnection is automatically achieved. Similarly, different types of continuum
models for foundation modelling have been proposed (Harr et al., 1969; Nogami &
Lain, 1987; Vallabhan & Das, 1991). There are also models that take into account
the elasto-plastic behaviour (Zeevrat, 1972) of foundation soil as illustrated in Figure
2.3a or the visco-elastic behaviour (Noda et al., 2000) of soil as presented by Figure
2.3b. A brief summary of such foundation models could be found in the review paper
of Dutta & Roy (2002).
Figure 2.3: Schematic view of a) elasto-plastic b) visco-elastic foundation
Although conceptually the abovementioned foundation models are useful, little
evidence has been produced to verify the computational accuracy of the various
models in studies representing the soil medium in soil–structure interaction analysis.
Moreover, problems typically occur depending on the choice of the various
parameters as well as the proper adjustment of foundation elements. On the other
hand, field elimination techniques idealising the system more rigorously with fewer
parameters deviate more in predicting the overall response. Most importantly, such
approaches still provide no or very little information regarding the stress and
deformation state within the soil mass.
The search for physically close and mathematically simple models in soil-structure
interaction problems has not been limited to field elimination techniques
representing the soil sub-domain. There are several field elimination techniques
which tend to simplify the structural sub-domain both computationally and
Chapter 2 Literature Review
39
mathematically. The work of Standing et al. (1998) falls into this category, where the
movements of the Treasury building in London arising from the construction of twin
tunnels are studied using a finite element model for the soil, while the building
structure is simply modelled as an equivalent raft. Also in the work by Dutta et al.
(2004) on dynamic soil-structure interaction modelling of elevated tanks, as shown
in Figure 2.4, it can be seen that both the structure and soil sub-domains are
simplified.
Figure 2.4: Field elimination model for SSI analysis of lateral vibration of an
elevated water tank (Dutta et al., 2004: p.828)
Field elimination techniques cannot deal accurately with geometric and material
nonlinearity in the replaced sub-domain, hence modelling the nonlinear response of
both soil and structure becomes complex for which more sophisticated modelling
approaches would be required. With the increasing availability of powerful
computers and the wider applicability of numerical methods compared to analytical
approaches, the use of the finite element method has become a common means for
modelling such complex interactive behaviour.
Chapter 2 Literature Review
40
2.2.2 Finite Element Method (FEM)
The finite element method is a special form of matrix analysis, where the whole
continuum is discretized into a finite number of elements connected at different
nodal points. The general principles and use of the finite element method are well
documented (e.g. Desai & Abel, 1987; Zienkiewicz et al., 2005).
In structural analysis, nonlinear modelling using finite element analysis has evolved
to address complex issues of various structural forms (Izzuddin, 1991; Zienkiewicz
et al., 2005; Zienkiewicz & Taylor, 2005). In fact, due to the universal nature of the
finite element method in modelling real-life complex conditions including geometric
and material nonlinearity in the response (Izzuddin, 1991), this numerical technique
has been widely used in design and assessment of complex structures (Bull, 1988;
Smith & Coull, 1991; Rombach, 2004).
On the other hand, utilising numerical methods in geotechnical engineering has also
provided geotechnical engineers with an extremely powerful analysis and design tool
(Potts & Zdravkovic, 1999; Potts & Zdravkovic, 2001; Potts, 2003). According to
Potts & Zdravkovic (1999), the numerical methods in modelling the soil medium are
far superior to conventional analytical methods, which tend to relax one or more of
the basic solution requirements (see Table 2.1).
Chapter 2 Literature Review
41
Method of Analysis
Solution Requirements
Equilibrium Compatibility Constitutive Behaviour
Boundary Conditions Force Displacement
Closed Form Limit equilibrium Stress Field
S. S. S.
S. N.S. N.S.
Linear elastic Rigid with Failure Criterion
S. S. S.
S. N.S. N.S.
Limit Analysis
Lower Bound Upper Bound
S. N.S.
N.S. S.
Ideal Plasticity with Associated Flow Rule
S. N.S.
N.S. S.
Full Numerical Analysis
S. S. Any S. S.
S. = Satisfied ; N.S. = Not Satisfied
Table 2.1: Basic solution requirements satisfied by various methods of geotechnical
analysis (Potts & Zdravkovic, 1999)
Although the application of numerical methods in geotechnical design is not as
widespread as in structural design, recent work by Gaba et al. (2002) and Ravaska
(2002) has demonstrated that the use of numerical analysis, as opposed to other
conventional methods, can lead to more accurate and economical design. As a result,
the use of this type of analysis in design applications is bound to increase in the
future, where a single simulation can provide all required information for design
purposes (see Table 2.2).
Method of Analysis Design requirements Stability Movements Adjacent structures
Closed form (linear-elastic) Limit equilibrium
Stress field
No Yes Yes
Yes No No
Yes No No
Limit analysis Lower bound Upper bound
Yes Yes
No Crude estimate
No No
Full numerical analysis Yes Yes Yes
Table 2.2: Design requirements satisfied by various methods of geotechnical analysis
(Potts and Zdravkovic, 1999)
Chapter 2 Literature Review
42
2.2.3 Integrated Modelling
In view of the above, the finite element method currently stands as the most
powerful and versatile tool for solving complex soil-structure interaction problems.
A finite element procedure for the general problem of soil-structure interaction
involving nonlinearities due to material behaviour, geometric changes and interface
behaviour is presented by Desai et al. (1982). A three-dimensional visco-elastic
finite element formulation has also been proposed by Viladkar et al. (1993) for
studying the interactive behaviour of a space frame with the supporting soil sub-
domain. The interactive behaviour of a plane frame-footing-soil system with elastic-
perfectly plastic soil has been investigated by Viladkar et al. (1991) and Noorzaei et
al. (1995a). The influence of soil strain-hardening on the elasto-plastic soil-structure
interaction of framed structures has also been undertaken by Noorzaei et al. (1995b).
Several studies have been carried out to model the discontinuous behaviour that may
occur at the interface of soil and structure, including the work of Beer (1985) and
Viladkar et al. (1994) who developed interface elements to model this discontinuity.
All of the above mentioned research has two major conceptual characteristics in
common. Firstly, the finite element method is used as a tool to obtain the complex
response of soil-structure interaction, and secondly the soil and structure sub-domain
are modelled simultaneously in a single computational model.
Although numerical analysis, and particularly finite element analysis, is currently the
most advanced approach for modelling soil-structure interaction, the application of
finite element modelling to soil-structure interaction problems is often limited by the
availability of a simulation environment offering advanced modelling capabilities for
both soil and structure sub-domains. Despite the existence of common techniques for
nonlinear structural and geotechnical analysis, the distinct demands of the two fields
in terms of material modelling and solution procedures have led to differently
evolved modelling techniques in each field. In fact, this has led to development of
two modelling disciplines with distinctive high-level features. In turn, this has been
reflected in the development of discipline-oriented computational tools, which offer
Chapter 2 Literature Review
43
sophisticated nonlinear modelling for their respective sub-domain (structure/soil)
and, if available, an approximation of the other (soil/structure).
This usually means that structural analysis simplifies soil behaviour, while
geotechnical analysis simplifies structural behaviour using field elimination
techniques. For example, in the work of Vlahos et al. (2006) on soil-structure
interaction of spudcan footings on clay soil subjected to cyclic loading, the structural
model is connected to a hybrid footing model (developed based on available
experimental data) representing the soil medium, as illustrated in Figure 2.5.
Figure 2.5: Idealization of jack-up models for SSI (Vlahos et al., 2006: p.217)
Similarly in the work by Jin et al. (2005) on evaluation of damage to offshore
platform structures, the structure sub-domain is modelled in 3D using FEM, while
the effect of the soil reactions on each pile element is simplified into three kinds of
non-linear springs, as illustrated in Figure 2.6. Further similarity of approach can be
seen in the work of Dutta et al. (2004) on soil-structure interaction analysis of low
rise buildings under seismic ground excitation (Figure 2.7) and that of Bhattacharya
Chapter 2 Literature Review
44
et al. (2004) on the assessment of the influence of soil flexibility on the dynamic
behaviour of building frames (Figure 2.8).
Figure 2.6: a) Schematic illustration of the pile and soil model b) Schematic diagram
of the structure model (Jin et al., 2005: p.1324)
Figure 2.7: a) Idealisation of the structure sub-domain b) Idealisation of the
foundation system (Dutta et al., 2004: p.897)
Chapter 2 Literature Review
45
Figure 2.8: a) Soil-structure coupled system b) Idealization of the coupled system
(Bhattacharya et al., 2004: p.117)
On the other hand, Tian & Li (2008) undertook work on the dynamic response of
building structures subject to ground shock from a tunnel explosion, where the multi-
story building structure is simplified and represented by a shear lumped mass model
as depicted in Figure 2.9.
Figure 2.9: Idealization of the coupled soil-structure interaction system (Tian & Li,
2008: p.1171)
In summary, with the current discipline-oriented structural and soil solvers, it is a
real challenge to achieve the same amount of sophistication in numerical modelling
of both the soil and the structure in a single soil-structure interaction analysis. There
are different ways to tackle this issue. One possibility is to develop new software or
augment existing software such that both soil and structure may be modelled to an
Chapter 2 Literature Review
46
equivalent level of sophistication. Such an approach is referred to as the monolithic
or direct treatment. Example applications of the monolithic approach in modelling
soil-structure interaction are the work of Viladkar et al. (2006) on static soil-
structure interaction response of hyperbolic cooling towers to symmetrical wind
loads, and the work of Noorzaei et al. (2006) on nonlinear interactive analysis of
cooling tower–foundation–soil interaction under unsymmetrical wind load using the
finite element method (Figure 2.10).
As discussed in Chapter 1, the monolithic treatment has the disadvantage that it
requires massive software development resources and it does not allow for
employing models, new techniques and already developed softwares in a modular
fashion. These disadvantages can be addressed by adopting a partitioned treatment,
as reviewed in the following section.
Figure 2.10: Monolithic treatment of soil-structure interaction problem (Noorzaei et
al., 2006: p.1001)
2.2.4 Partitioned Analysis
In the partitioned treatment, the partitioned sub-domains of the coupled mechanical
system are computationally treated as isolated entities, and the response of the
Chapter 2 Literature Review
47
coupled system is calculated using already developed solvers. A partitioned
treatment with different partitioned sub-domains modelled as separate computational
entities, amongst which interaction effects are exchanged (see Figure 2.11), can offer
major benefits in the context of nonlinear soil-structure interaction analysis.
Figure 2.11: Partitioned treatment of soil-structure interaction
Such benefits include i) allowing field-specific discretisation and solution procedures
that have proven performance for each partitioned sub-domain, ii) facilitating the
reuse of existing nonlinear analysis solvers with all the resource savings that this
brings, and iii) enabling parallel computations through problem partitioning (Lai,
1994; Felippa et al., 2001)
According to Felippa et al. (2001), the development and application of partitioned
analysis of coupled systems involving structures goes back to 1970. In their paper
(Felippa et al., 2001) the authors state:
‘The partitioned treatment of coupled systems involving structures emerged
independently in the mid 1970s at three locations: Northwestern University by T.
Belytschko and R. Mullen, Cal Tech University by T.J.R. Hughes and W.K. Liu, and
Lockheed Palo Alto research laboratories (LPARL) by J.A. Deruntz, C.A. Felippa,
T.L. Geers and K.C. Park’ (Fellipa et al, 2001: P. 3255).
The above research groups were focused on different applications and pursued
different types of problem-decomposition methodology. For Instance, Belytschko &
Mullen (1978) studied node by node partitioning and sub-cycling. Hughes & Liu
Chapter 2 Literature Review
48
(1978) developed element by element implicit-explicit partitions. While the work of
these groups focused on structure-structure and fluid-structure interaction treated by
FEM discretisation, in LAPRL the above mentioned researchers concentrated on the
coupling of a finite element computational model of submerged structures to a
boundary element model of the exterior fluid (Geers & Felippa, 1980). A staggered
solution procedure was developed for the resulting fluid-structure interaction
problem by Fellippa & Park (1980) and was included in the general class of
partitioned methods (Park, 1980; Park & Felippa, 1980).
2.2.5 Staggered Approach
In general, the partitioned analysis is mainly carried out by using staggered or
iterative sub-structuring (Quarteroni & Valli, 1999) methods. The staggered
approach is suited to transient dynamic analysis only, where the governing equations
of the partitioned domains are solved independently at each time step using predicted
boundary conditions at the interface (force, displacement, velocity or acceleration)
obtained from previous time step(s) by a predictor.
Figure 2.12: Staggered coupling approach
The staggered coupling approach is illustrated in Figure 2.12 for a soil-structure
interaction problem involving two physically partitioned and independent sub-
domains (i.e. soil and structure). In this respect, sub-domain B stands for the soil,
while T represents the structure. The prediction stage typically stands for prediction
of displacements from the soil model at the soil-structure interface, while the
Chapter 2 Literature Review
49
substitution stage typically stands for the substitution of the reaction forces from the
structural model into the soil model at the same interface.
This approach is approximate in nature due to the fact that the predicted interface
displacements, obtained using displacements, velocities and accelerations from
previous steps, are invariably different from the displacements evaluated following
force substitution into the soil model, thus violating the compatibility conditions at
the interface.
The application of the staggered approach to coupled mechanical systems has
attracted significant research interest over the past years (Piperno, 1997; Huang &
Zienkiewicz, 1998; Farhat & Lesoinne, 2000). Although the focus of these
researchers has been mainly on fluid-structure interaction, there are a few examples
of employing this technique in soil-structure interaction coupling. One such example
is the work of Rizos & Wang (2002), who developed a partitioned method for soil-
structure interaction analysis in the time domain based on the staggered approach
(Figure 2.13), where a standard Finite Element Method (FEM) model, representing
the structure domain, was coupled to a Boundary Element Method (BEM) model,
representing the soil domain as an elastic half-space. The advantages of this
staggered approach (Rizos & Wang, 2002) are that the derivation of a global system
matrix for the combined FE–BE domain is not required, special solution strategies
do not need to be developed, and smaller systems of simultaneous equations are
encountered.
The aforementioned advantages make the proposed approach appealing for the
solution of large coupled systems. Further to this, O’Brien & Rizos (2005)
considered the application of the staggered approach for the simulation of high speed
train induced vibrations. The Boundary Element Method is used to model the soil-tie
system, while Finite Element Method, along with Newmark’s integration, is used for
the modelling of the rail system. The two methods are coupled at the tie-rail interface
and the solution is obtained following a staggered, time marching scheme
(Figure 2.14).
Chapter 2 Literature Review
50
Figure 2.13: Schematic representation of the staggered coupling approach (Rizos &
Wang, 2002: p.878)
Figure 2.14: Staggered coupling of BEM-FEM for soil-structure interaction
simulation of high speed train induced vibrations (O’brien & Rizos, 2005: p.290)
Besides the inapplicability of staggered approach to static analysis, there are major
issues in relation to the staggered approach that need to be addressed in coupled
dynamic analysis, namely stability and accuracy. A staggered solution procedure
should not degrade the numerical stability of any individual system nor the overall
Chapter 2 Literature Review
51
coupled problem. The accuracy degradation of the solution procedure is also another
concern in a staggered coupling procedure. In general, for nonlinear mechanical
problems the stability and accuracy of a staggered solution procedure are intertwined
and should be studied concurrently.
Considerable research has been undertaken on the stability and accuracy of the
staggered approach for different coupled problems (Farhat & Park, 1991; Huang &
Zienkiewicz, 1998). For Instance, in the work of Rizos & Wang (2002) on staggered
coupling of soil-structure interaction, it is concluded that in the absence of elastic
(restoring) forces in the structure (e.g. rigid massive foundations on elastic soils), the
FEM solver in their staggered scheme becomes unstable; therefore, a stabilisation
should be augmented into the FEM equations with an equivalent stiffness of the soil
region under the foundation.
In this regard, as rightly discussed by Felippa et al. (2001), achieving stability in a
staggered solution procedure is extremely difficult and in many cases impossible
without reformulation (mainly by augmentation) of the field equations of the
original partitioned sub-domains. Furthermore, a general theory of stability of
discrete and semi-discrete nonlinear coupled systems is yet to be developed.
Although the stability of staggered treatment may be ensured by reformulation of the
partitioned sub-domains’ field equations, this would be a backward step in terms of
modular use of the existing nonlinear structural/soil discipline oriented software. In
fact, such reformulation requires access to the soil/structural solvers solution
procedure codes.
This shortcoming is discussed further in Chapter 3, where the staggered coupling of
linear dynamic FEM-FEM coupled problems is presented, and where it is shown that
without reformulation of the field equations only conditional stability may be
achieved. Indeed in such a scenario, both accuracy and stability will be dependent on
the time-step size, while the choice of the predictor and the partitioning strategy will
also have a significant effect on achieving conditional stability. However, achieving
conditional stability in staggered coupling of linear FEM-FEM problems requires
Chapter 2 Literature Review
52
choosing small time-steps which makes such coupling procedures computationally
inefficient. As a result, with only conditional stability, the staggered approach should
be used with great care.
2.2.6 Iterative Coupling
To address the above shortcomings, coupling algorithms have been developed that
are stable and accurate for a wider range of time-step size, which is mainly achieved
by introducing corrective iterations into the staggered approach, hence the name
iterative sub-structuring/coupling methods (Quarteroni & Valli, 1999). In addition to
the enhancement of stability and accuracy, iterative coupling approaches facilitate
parallel computing through problem partitioning which can lead to much greater
computational efficiency. In addition, iterative coupling approaches can be applied to
both static and dynamic problems.
Figure 2.15: Schematic of sequential iterative coupling approach
The general procedure of iterative coupling is illustrated in Figure 2.15, which refers
to a soil-structure interaction coupled system decomposed into the soil and structure
sub-domains (T, B). The governing equations of the partitioned sub-domains are
solved independently at each load increment (or time step in the case of dynamic
analysis), using predicted boundary conditions (either force or displacement) at the
interface. These predicted boundary conditions are then successively updated using
Chapter 2 Literature Review
53
corrective iterations until convergence to equilibrium and compatibility is achieved
at the interface and within the partitioned sub-domains.
Unlike the staggered approach, the compatibility and equilibrium conditions at the
interface of partitioned sub-domains are enforced in the iterative coupling approach,
provided convergence is achieved, thus the stability and accuracy concerns
associated with the staggered approach are no longer relevant.
Despite the significant potential benefits of iterative coupling, a major issue relates
to whether convergence to equilibrium and compatibility at the interface can always
be achieved through successive iterative substitutions. In this respect, the utilised
technique in successive update of boundary conditions at the interface of the
partitioned coupled system is the most critical algorithmic stage. In fact, it is the
update technique that dictates the convergence behaviour of the algorithm rather than
the time/load step size.
In this regard, if a trivial update of boundary conditions in FEM-FEM iterative
coupling is utilised, only a conditional convergence to compatibility and equilibrium
could be achieved. Due to this conditional convergence, a relaxation of the updated
boundary conditions is often augmented to the iterative coupling procedure in order
to improve the convergence characteristics, hence the term interface relaxation
(Marini & Quarteroni, 1989).
Interface relaxation is a traditional method in domain decomposition methods which
follows Southwell’s relaxation (Hoffman, 2001). A general framework for solving
composite PDEs based on interface relaxation can be found in the work of Mu
(1999).
Interface relaxation iterative coupling algorithms vary significantly in terms of the
adopted computational procedure. Concerning the computational method, the
algorithms can be categorized into forms of sequential and parallel coupling. Here
the term parallel coupling refers to a form of partitioned computation in which
obtaining the response of each independently modelled soil and structure sub-
Chapter 2 Literature Review
54
domains is carried out simultaneously during every coupling iteration. Unlike the
parallel coupling, in sequential coupling the partitioned sub-domains are solved one
after the other at each iterative stage. In addition to the parallel or sequential nature
of iterative coupling procedures, these algorithms also differ in relation to the
treatment of prescribed Dirichlet (Displacement) and Neumann (Force) effects at the
interface of the partitioned sub-domains. Elleithy & Tanaka (2003) presented and
categorized a range of different iterative coupling algorithms according to the
aforementioned characteristics for elasto-static BEM-BEM and FEM-BEM coupling
(for example, Sequential Dirichlet–Neumann coupling algorithm, Parallel Neumann–
Neumann coupling algorithm, Parallel Dirichlet-Neumann coupling algorithms, etc.).
Such algorithms will be described in detail in a more general form in the context of
static and dynamic soil-structure interaction FEM-FEM coupling in Chapter 4.
An example application of parallel Neumann-Neumann and Dirichlet-Dirichlet
iterative coupling algorithms is that of Kamiya et al. (1996). In this research the
authors studied the parallel implementation of the boundary element method using a
domain decomposition method on a cluster computing system. As concluded by the
authors, the main advantage of the proposed domain decomposition technique is that
it avoids the distinct formulation for constructing global matrices for the problems
with different domain decompositions. In the presented method, the global matrices
of the whole domain are not constructed and the boundary element analysis is
performed on individual sub-domains. More recent application of sequential
Dirichlet-Neumann iterative coupling approach for coupling BEM and FEM in
elasto-static analysis is given by Elleithy et al. (2001).
The application of iterative coupling algorithms in partitioned analysis has not been
limited to BEM-FEM coupling or fluid-structure interaction. In fact, these
algorithms are general in nature and with some modification could be applied to any
desired type of discretization technique in partitioned analysis of multi-physics
interaction problems. For instance, Collenz et al. (2004) utilised an FEM-FEM
coupling method based on a sequential approach for their work on modelling the
interactive nonlinear behaviour of the micro-beams under electrostatic loading.
Chapter 2 Literature Review
55
The significant potential of the partitioned approach, as highlighted before, was
recognised more recently for coupled modelling of soil-structure interaction
problems. Hagen & Estorff (2005b) presented a domain decomposition approach for
the transient analysis of arbitrary three-dimensional soil–structure interaction
problems where the coupling of the sub-domains is performed in an iterative manner.
Based on such techniques, Hagen & Estorff (2005a) presented a hybrid approach for
transient dynamic investigation of Dam-Reservoir-soil problem, where different
discretisation techniques were utilized for different partitioned sub-domains as
shown in Figure 2.16. Elleithy et al. (2004) presented an interface relaxation FEM–
BEM coupling method for elasto-plastic analysis, which was applied to investigate
the stresses developed in tunnel structures. A schematic representation of their
presented coupling scheme is shown in Figure 2.17.
Figure 2.16: Coupled dam-reservoir-soil system (Hagen & Estorff, 2005a: p.10)
Chapter 2 Literature Review
56
Figure 2.17: Flowchart of iterative coupling method by Elliethy et al. (2004: p.852)
Although the superiority of the iterative coupling partitioned approach has been
recognised more recently for coupled modelling of soil-structure interaction, there
remain significant technical challenges related to algorithmic and computational
issues, particularly with reference to convergence.
In all of the abovementioned iterative coupling algorithms, the update technique used
in the successive iterations is the constant interface relaxation scheme. As shown
later in this work, convergence to compatibility and equilibrium at the interface of
the partitioned sub-domains could be guaranteed through the choice of a suitable
value for the relaxation parameter, though the number of required iterations may be
unrealistically large.
Despite the relative simplicity of iterative relaxation schemes, there are several
issues that need to be carefully considered: i) determination of the range of suitable
relaxation parameters for the specific problem under consideration in order to
achieve convergence, and ii) selection of the optimum relaxation parameter in order
to achieve maximum computational efficiency.
Chapter 2 Literature Review
57
In this respect, work by El-Gebeily et al. (2002) on convergence analysis of static
coupling of BEM-FEM has shown that the convergence behaviour is very sensitive
to the value of relaxation parameter. Moreover, it has been shown that the value of
relaxation parameter is problem dependent and varies considerably with respect to
the employed combination of the partitioned sub-domains, mesh density, material
properties and the adopted coupling scheme (parallel-sequential).
These results have also been confirmed by the research of Estorff & Hagen (2005)
on dynamic analysis of coupled BEM-FEM problems, where the significant
influence of different problem parameters on the choice of suitable relaxation
parameter is demonstrated. It has also been shown that, depending on the
characteristics of the problem under consideration, there is a range of applicable
relaxation parameters outside of which convergence to compatibility and equilibrium
could not be guaranteed. Moreover, there exists an optimum relaxation parameter in
the convergent range that holds the highest convergence rate and computational
efficiency. In this regard, there is a lack in the literature with regard to establishing
the general mathematical convergence behaviour with relation to the relaxation
parameter. Moreover, the determination of the optimum constant relaxation
parameter would typically require a process of trial and error for every coupling case
under consideration, which is prohibitive for real large scale soil-structure interaction
problems.
To address the above shortcomings with constant relaxation, adaptive relaxation
approaches were proposed by Funaro et al. (1998) for iterative coupling of
partitioned second-order elliptic problems, Wall et al. (2007) and Kuttler & Wall
(2008) focusing on iterative coupling of fluid-structure interaction problems, Elliethy
et al. (2004) for static iterative coupling of BEM-FEM, and Soares (2008) for
iterative coupling of FEM-BEM in dynamic analysis, which avoid the trial and error
process in constant relaxation.
Building on the above, as shown later in this work, the adaptive coupling technique
in the context of nonlinear soil-structure interaction is proposed. In this respect, the
Chapter 2 Literature Review
58
need for establishing the general convergence behaviour of the scheme for both
dynamic and static FEM-FEM coupled problems is demonstrated mathematically
and illustrated through various case studies in Chapter 5. In this regard, it is shown
that the performance of interface relaxation in iterative coupling of soil-structure
interaction problems is enhanced significantly through the use of an adaptive
relaxation, using error minimization techniques.
In general, due to the lack of general convergence studies on the application of such
methods and the effects of nonlinearity on the convergence behaviour, specifically in
the context of soil-structure interaction, there remain significant issues to be covered
in relation to the convergence characteristics and performance of constant and
adaptive relaxation methods.
Furthermore, all the presented adaptive relaxation schemes have been employed
within the sequential Dirichlet-Neumann form of coupling algorithm, thus an
adaptive method for evaluating the relaxation parameter in the parallel forms of
coupling algorithms is still largely absent from current research on iterative coupled
modelling via interface relaxation approach.
Albeit, it is possible to extend the existing iterative coupling methods to enhance the
computational efficiency of both adaptive and constant relaxation coupling
algorithms, while overcoming the problematic issues regarding the trial and error
process embedded in evaluation of the constant relaxation parameter.
Kamiya & Iwase (1997), in their research on parallel elasto-static BEM-FEM
coupling, proposed a conjugate gradient scheme to overcome the drawback of the
relaxation scheme in selection of parameters for renewal iterations by users ( which,
as stated by them, require ‘trial and error and deep experience’). The presented
conjugate gradient method was built on the conjugate gradient method as a tool for
solution of simultaneous equations and application in domain decomposition
problems such as potential analysis by Glowinski et al. (1983). In the work by
Kamiya & Iwase (1997) it is shown that the proposed conjugate gradient method
Chapter 2 Literature Review
59
coupling scheme is computationally efficient and has a higher convergence rate
compared to constant relaxation coupling algorithms.
In the view of above and considering the need for developing iterative coupling
methods with unconditional convergence characteristics and high convergence rate, a
new approach in coupled nonlinear soil-structure interaction analysis is proposed
later in this thesis. This approach is based on utilizing the condensed interface
tangent stiffness matrices of the structure and soil models, depending on the variant
coupling algorithms under consideration (Chapter 6).
Although the application of adaptive relaxation has been restricted to sequential
Dirichlet-Neumann coupling algorithm, by utilizing the condensed tangent interface
stiffness of the partitioned soil and structure sub-domains, various parallel and
sequential iterative coupling algorithms could be employed effectively without any
need for a trial and error procedure. In this respect, regardless of the employed type
of coupling algorithm (parallel or sequential) and partitioned problem parameters,
superior convergence behaviour compared to relaxation schemes is guaranteed to be
achieved.
Although the condensed tangent stiffness matrix can be determined with current
nonlinear field modelling tools, its implementation for nonlinear soil-structure
interaction coupling using discipline oriented softwares would normally necessitate
direct access to the source codes. In this respect, it is proposed in this work that the
condensed tangent interface stiffness matrices of the partitioned soil and structural
sub-domains may be reasonably approximated by constructing reduced order models
of the structure and soil sub-domains. The benefit of such an approach is that it does
not require the explicit assembly of the stiffness matrices. Indeed, with minimal
intrusion to the source codes of the solvers, a potentially efficient coupling technique
for coupled modelling of soil-structure interaction will be achieved. This builds on a
previous approach presented by Vierendeels (2006) and Vierendeels et al. (2007),
who utilized a procedure for constructing the reduced order model of partitioned sub-
domains throughout the coupling iterations for implicit coupling of fluid–structure
Chapter 2 Literature Review
60
interaction problems, although some significant modification has been proposed in
this research for adopting and improving the scheme in the context of soil-structure
interaction.
In this regard, there is a pitfall associated with the traditional reduced order method.
For instance, in FEM-FEM iterative coupling problems, where the number of
coupling iterations exceeds the number of interface degrees of freedoms, due to
existence of linearly dependent displacement/force modes, the reduced order method
approximation for the condensed interface tangent stiffness matrix will be poor. This
shortcoming is discussed and addressed comprehensively in Chapter 6, where a
significant modification in reduced order method formulations is proposed to
overcome this shortcoming. Moreover, an advanced technique for approximating the
reduced order method (mixed reduced order method) which possesses a very high
convergence rate, especially in nonlinear problems, is further proposed.
2.3 Concluding Remarks
A review of various soil-structure interaction modelling techniques, along with their
advantages and disadvantages, has been presented. These include field elimination
techniques, such as the Winkler and continuum approaches, and numerical modelling
techniques, such as the finite element and boundary element methods. The distinct
superiority of numerical methods in SSI analysis, specifically the finite element
method, is discussed.
It is shown that, with the current discipline oriented structural and soil solvers,
achieving the same amount of sophistication in numerical modelling of both soil and
structure in a single soil-structure interaction analysis is still a challenge. Since
sophisticated discipline-oriented structural and geotechnical solvers are readily
available, and continue to be used for either application, the partitioned treatment of
coupled soil-structure interaction problems is introduced as an alternative to the
monolithic approach for tackling interaction problems.
Chapter 2 Literature Review
61
The background and origin of the partitioned approach are presented. The staggered
coupling procedure, including its previous application to several transient dynamic
coupled problems, is discussed, and its main disadvantages are highlighted. These
disadvantages will be further demonstrated in Chapter 3, where the staggered
solution procedure for coupled modelling of FEM-FEM problems is
comprehensively examined. To address these shortcomings, iterative coupling
procedures have been introduced as an alternative to the staggered approach. Not
only do these iterative methods address the stability and accuracy issues of the
staggered approach, but they are applicable to both static and dynamic analysis.
Notwithstanding, a major issue associated with iterative coupling algorithms remains
their convergence to compatibility and equilibrium within a single load/time-step. In
this respect, several iterative update techniques such as using interface relaxation,
condensed interface tangent stiffness matrix and reduced order method are discussed.
In the context of relaxation approaches, the shortcomings of constant relaxation and
its enhancement in the form of adaptive relaxation are discussed. Due to the lack of
general convergence studies on the application of such methods and the effects of
nonlinearity on the convergence behaviour, specifically in the context of soil-
structure interaction, the convergence characteristics and performance of constant
and adaptive relaxation methods will be considered in detail in Chapter 5.
It is also pointed out that further enhancement of iterative coupling beyond the use of
relaxation schemes may be achieved through the use of the condensed tangent
stiffness matrices at the soil–structure interface. As a special case of such an
approach, the condensed tangent stiffness matrix may be approximated via reduced
order models of the partitioned sub-domains, where various highly efficient reduced
order methods will be presented in Chapter 6. This builds on previous work by
Vierendeels (2006) and Vierendeels et al. (2007) concerned with implicit coupling of
fluid–structure interaction problems, though major modifications are proposed in this
research for coupled modelling of nonlinear soil-structure interaction.
Chapter 2 Literature Review
62
In summary, despite significant research on the partitioned treatment of coupled
systems, there remain significant technical challenges related to algorithmic and
computational issues, especially in the case of nonlinear soil-structure interaction
analysis. The focus of previous research has been mainly on fluid-structure coupling
and FEM-BEM coupled modelling, and where soil-structure interaction problems
were considered the focus has been mainly on the coupling procedure formulation
(again mainly on FEM-BEM coupling) rather than development of an advanced soil-
structure analysis platform. In this work, a novel simulation environment, which
embodies the proposed coupling algorithms and utilises discipline-oriented solvers
for nonlinear structural and geotechnical analysis, is developed and discussed in
Chapter 4. This environment is based on the coupling of two powerful in-house
programs at Imperial College, ADAPTIC (Izzuddin, 1991) and ICFEP (Potts &
Zdravkovic, 1999), for nonlinear structural and geotechnical analysis, respectively.
The developed environment is employed in this work to demonstrate the
performance characteristics and merits of the various presented algorithms, and to
illustrate the practical application of coupled modelling for realistic modelling of
nonlinear soil-structure interaction problems.
Chapter 3
Staggered Approach
3.1 Introduction
One of the earliest approaches proposed for coupling partitioned mechanical systems
is the staggered approach (Felippa & Park, 1980), which is particularly suited to
transient dynamic analysis. In the staggered solution approach, the equations for
each sub-domain are solved once at each time-step and predicted values of the
coupling boundary conditions are used to obtain the response of the individually
modelled sub-domains.
This approach is illustrated in Figure 3.1 for a soil-structure interaction problem
involving two physical partitions. In the context of soil-structure interaction, the
prediction stage typically stands for prediction of displacements from the soil model
at the soil-structure interface, while the substitution stage typically stands for the
substitution of the reaction forces from the structural model into the soil model at the
same interface.
Chapter 3 Staggered Approach
64
A typical approach in the staggered partitioning of soil-structure interaction, as
illustrated in Figure 3.1, would be to fix the contact degrees of freedom of the
structure sub-domain using predicted values obtained by a predictor. The structure
solver then computes the response of the structure as well as the forces at the
interface nodes. Based on these forces, the soil solver evaluates the displacements at
the interface nodes, which are used by the predictor to establish the new initial
conditions for the structure as the solution moves to the next time step.
Figure 3.1: Staggered approach in dynamic soil-structure interaction
The two most important issues of the staggered approach that should be addressed in
any scheme are the stability and the accuracy of the algorithms. The stability of
staggered algorithms depends on the suitable choice of the predictor operator, the
selected time integration scheme in each domain, the employed time-step for the
time integration schemes and the utilized discretization techniques at the interface
and through each field.
Once satisfactory stability is achieved, the next concern is accuracy. The predicted
interface displacements, obtained using velocities/accelerations at the beginning of
the step, are invariably different from the displacements evaluated following force
Chapter 3 Staggered Approach
65
substitution into the soil model, thus leading to compatibility defaults at the
interface. The accuracy of the staggered approach may be improved by reducing the
time-step and using an optimal predictor, though stability requirements may dictate
step sizes that are too small rendering this approach prohibitively expensive for
large-scale soil-structure interaction problems.
In the following, a staggered approach for dynamic soil-structure interaction
problems is discussed, and the stability and accuracy issues are addressed in the
context of linear analysis. The proposed approach employs the Newmark integration
scheme for analysing the dynamic behaviour of the structure and soil sub-domains,
as described in the next section. In this context, the general stability conditions
required for staggered coupling of FEM-FEM are established. It is worth
mentioning, that unconditional stability may be achieved in staggered coupling by
employing significant modifications in the formulation of the partitioned sub-
domains for different interaction applications or by changes to the governing
equations by augmentations, such as those proposed by Farhat & Park (1991) and
Huang & Zienkiewicz (1998) for fluid-structure interaction analysis.
Notwithstanding the benefits of such methods, these are not considered here since
they do not permit the modular use of structural and soil solvers as black box
solvers, as they require major change in formation of the stiffness matrices and
reformulation (mainly by augmentation) of the field equations of the partitioned sub-
domains.
3.2 Dynamic Analysis Formulation
One of the most used techniques in assessing the dynamic response of mechanical
systems is the direct method, where the dynamic equilibrium equations are integrated
directly at the overall structural level, typically using a step-by-step time integration
scheme. This type of analysis is the most general and is readily applicable to
nonlinear analysis.
Chapter 3 Staggered Approach
66
The dynamic equilibrium equations for a discrete linear structural system can be
expressed as:
M U C U K U F (3.1)
where M is the mass matrix, C is the damping matrix, K is the stiffness
matrix, U is the acceleration vector, U is the velocity vector, U is the
displacement vector and F is the applied load vector.
Equation (3.1) represents a semi-discrete problem since only spatial discretization is
undertaken, but not discretization in the time domain. Step-by-step integration
methods deal with the latter issue by considering Equation (3.1) at specific points of
time, and obtaining the solution at time tn+1 utilising the displacement, velocity and
acceleration history over previous discrete times. Consideration is restricted here to
the Newmark method (Zienkiewicz & Taylor, 1991), which is a single-step method
(i.e. only the displacement, velocity and acceleration at time nt are employed in
determining the corresponding entities at time tn+1).
Considering the discrete dynamic problem expressed by:
1 1 1 1n n n nM U C U K U F (3.2)
the Newmark method utilises the following parametric single-step difference
equations for obtaining a solution:
21 1
1
2n n n n nU U t U t U U
(3.3)
1 11n n n nU U t U U (3.4)
in which β and γ are algorithm parameters.
Chapter 3 Staggered Approach
67
The above equations, allow the three unknowns 1nU , 1nU and 1nU
to be
determined as:
1
1 12
1 ˆ ˆn n n nU M C K F M U C U
t t
(3.5)
1 12
1 ˆn n nU U U
t
(3.6)
1 1ˆ
n n nU U Ut
(3.7)
where:
22
1 1ˆ2n n n nU U U t t U
t
(3.8)
ˆ ˆ1n n n nU U t U U (3.9)
3.3 Partitioning
Before developing the general staggered approach framework for coupling, it is
beneficial to illustrate the domain decomposition strategy and discretised
representation of the partitioned sub-domains in coupling algorithms using FEM.
Consider a coupled system governed by the semi-discrete equation of dynamic
equilibrium:
M U C U K U F (3.10)
Assume that the above coupled system is composed of ΩT and ΩB sub-domains with
the interface (ΩΓ= ΩT∩ ΩB), as shown in Figure 3.2 and formulated by:
Chapter 3 Staggered Approach
68
11 12 11 12
21 22 11 12 21 22 11 12
21 22 21 22
11 12
21 22 11 12
21 22
0 0
0 0
0
0
T T T T T TT T
T T B B i T T B B i
B B B B B BB B
T T TT
T T B B i
B B BB
M M U C C U
M M M M U C C C C U
M M U C C U
K K U
K K K K U
K K U
TT
i
BB
F
F
F
(3.11)
In the above, vectors XXU , X
XU , XXU and X
XF correspond to the displacement
vectors and external loads, respectively, for the non-interface degrees of freedom in
sub-domain ΩX, while vectors iU , iU
, iU and iF correspond to
displacement vectors and external loads at the interface degrees of freedom. In
addition, XijM , X
ijC and XijK correspond to mass, damping and stiffness matrices in
sub-domain ΩX, respectively.
Figure 3.2: Partitioning and discretisation
Assuming that the above coupled system is partitioned into two independently
modelled sub-domains (ΩT, ΩB) and discretised by FEM using a step-by-step time
Chapter 3 Staggered Approach
69
integration method, such as the Newmark method, the governing discrete equations
of each independently modelled partitioned sub-domain can be formulated as
follows:
For sub-domain ΩT:
11 12 11 12 11 12
21 22 21 22 21 22
T T T T T T T TT TT TT T
T T T T T T i ii iT TT T
M M C C K K U FU U
M M C C K K U FU U
(3.12)
For sub-domain ΩB:
11 12 11 12 11 12
21 22 21 22 21 22
B B B B B B B BB BB BB T
B B B B B B i ii iB BB B
M M C C K K U FU U
M M C C K K U FU U
(3.13)
In the above, vectors iXU and i
XF are the displacements and external loads,
respectively, for the interface degrees of freedom in sub-domain ΩX, whereas:
11 12
21 22
X X
X X
K K
K K
= stiffness matrix of sub-domain ΩX,
11 12
21 22
X X
X X
M M
M M
= mass matrix of sub-domain ΩX, and
11 12
21 22
X X
X X
C C
C C
= damping matrix of sub-domain ΩX.
Clearly, the time marching Equations (3.12) and (3.13) cannot be solved
independently. This is due to existence of unknown displacement (i.e. ,i iT BU U )
and force vectors (i.e. ,i iT BF F ) at the interface. However, by employing a
coupling procedure in which the response of the partitioned sub-domains is obtained
for certain interface boundary conditions and united at the interface level, the
partitioned domains can be solved independently. Here this task is achieved using a
staggered coupling procedure as described hereafter.
Chapter 3 Staggered Approach
70
3.4 Staggered Coupling Procedure
Consider the partitioning strategy presented in the previous section, where ΩT refers
to the partitioned structure sub-domain, while ΩB corresponds to the partitioned soil
sub-domain. The general algorithmic framework of staggered coupling in the context
of soil-structure interaction analysis using FEM can be described by the following
procedure, in which n denotes the current time-step number.
For n = 1, 2,...
STEP 1: At the start of each time-step, the ΩT sub-domain (structure) is loaded by
the external forces TT n
F , while the displacements at the interface nodes, iT n
U , are
prescribed by a predictor in accordance with the initial conditions.
STEP 2: The response of ΩT sub-domain (structure) is obtained using Equation
(3.12) for TT n
U and iT nF .
STEP 3: The corresponding interface forces iB nF at ΩB sub-domain (soil) are
determined from equilibrium:
0i iT Bn n
F F (3.14)
STEP 4: Based on the interface forces, iB nF , and the external loading applied to the
soil sub-domain, BB n
F , the soil response iB n
U and BB n
U is obtained using
Equation (3.13).
STEP 5: Before moving to the next time-step, a predictor will evaluate 1
iT n
U
:
1, , ,i i i i
T B B Bn n n nU U U U t
(3.15)
Chapter 3 Staggered Approach
71
while the compatibility of the initial condition for the structure domain and soil
domain is enforced:
i iT Bn n
U U (3.16)
i iT Bn n
U U (3.17)
The above staggered coupling scheme couples the response of the partitioned sub-
domains by enforcing equilibrium condition at the interface level, while the
compatibility default is minimized by choosing a suitable predictor function and
small time step in order to level the response of the coupled sub-domains.
3.5 Predictors
As mentioned before, a suitable choice of the predictor is very important since it
influences the stability and accuracy of the staggered scheme. Emphasis is placed on
Linear Multi Step (LMS) integrators, since they include a large number of
integration formulas used in practice, and both the functions and their derivatives are
evaluated at the same times. The general expression of a LMS predictor for constant
step size t is (Felippa & Park, 1980):
2
11 1 11 1 1
m m mP P P P
i i in in n i n ii i i
U U t U t U
(3.18)
where Pi , P
i and Pi are numeric coefficients of the predictor. Since for second-
order governing equations (as considered here), historical information involves at
most computation of second derivatives, LMS predictors continuing beyond the
1n i
U
terms are not considered. Note also that Equation (3.18) involves m past
terms, where m is the number of steps in the LMS integrator.
Chapter 3 Staggered Approach
72
Trivial Predictors
General Form:
1
Ppredicted nn
U U
Trivial:
1predicted nn
U U
First Order Predictors
General Form:
1
P Ppredicted nn n
U U t U
Central Difference Method:
1predicted nn n
U U t U
Second Order Predictors
General Form:
2
1
P P Ppredicted nn n n
U U t U t U
Central Difference Method:
2
1 2predicted nn n n
tU U t U U
Table 3.1: Single-step predictors for second-order dynamic systems
Some examples of the single-step predictors that can be used here for second-order
dynamic systems are presented in Table 3.1.
3.6 Stability Analysis by Amplification Method
The term “stable” informally means resistant to change. For technical use, the term
has to be defined more precisely in terms of the mathematical model, but the same
connotation applies. In mathematics, stability theory is typically concerned with
whether a given function is sensitive to a small perturbation (Felippa & Park, 2004).
Considering a dynamic problem and supposing that the system is undergoing
periodic motion with a period (T), then:
( ) ( )t T t U U (3.19)
Stability of the system requires that, in studying the behaviour of the system after
application of an arbitrary perturbation in the initial displacement or velocity, the
motion remains within small prescribed limits of the unperturbed motion.
Chapter 3 Staggered Approach
73
One of the widely used methods in stability analysis is the amplification method,
which is also called von Neumann stability analysis (Zienkiewicz & Taylor, 1991;
Felippa & Park, 2004). The approach is based on decomposition of motion into
normal modes and checking the growth or decay of perturbations from one step to
the next, and it can be implemented using standard linear algebra procedures.
Considering the application of a time approximation recurrence algorithm such as
the Newmark method to an initial-value problem, it can be shown that the
homogeneous form (i.e. free vibration) takes the form of (Zienkiewicz & Taylor,
1991):
1n nU U
A (3.20)
in which [A] is known as the amplification matrix. Since methods such as Newmark
represent approximations that are used to derive equations of type (3.20), error is
introduced into the solution 1nU
at each time step. Since the solution 1n
U
at time
1nt depends on the solution nU at time nt the error can grow with time. It can be
clearly observed that any error presented in the solution will of course be subjected
to amplification by precisely the same factor:
k n
k nU U
A (3.21)
The algorithm is considered to be stable if the error introduced in nU does not
grow unbounded as Equation (3.20) is solved repeatedly.
Knowing that the general modal solution of any recurrence algorithm can be written
as (Zienkiewicz & Taylor, 1991):
1n nU U
(3.22)
and by substituting Equation (3.22) into Equation (3.20), it can be observed that is
given by the eigenvalues of the amplification matrix as:
Chapter 3 Staggered Approach
74
0n
U A I (3.23)
Clearly if any eigenvalue, , of the amplification matrix is greater than one in
absolute value, 1 , all initially small errors will increase without limit and the
solution will be unstable. Therefore the stability of the recurrence algorithm requires
that 1 (Zienkiewicz & Taylor, 1991). In the case of complex eigenvalues, the
requirement is modified to the modulus of being less than or equal to one.
The calculation of the eigenvalues in such problems is not trivial. Therefore, two
general procedures are employed here to make the task of stability checks more
practical.
The first procedure is named z transformation (Zienkiewicz & Taylor, 1991). The
determinant equation 0 A I provides the characteristic polynomial yielding
the eigenvalues of the amplification matrix and, as mentioned for stability, it is
sufficient and necessary that the module of all the eigenvalues 1 . By using the
following variable transformation in the obtained characteristic polynomial:
1
1
z
z
(3.24)
where z and are in general complex numbers, it is easy to show that the
requirement of 1 is identical to that demanding the real part of z to be negative,
as illustrated in Figure 3.3 (Zienkiewicz & Taylor, 1991).
Chapter 3 Staggered Approach
75
Figure 3.3: Z-transformations (Zienkiewicz & Taylor, 1991)
The second transformation is the well-known Routh-Hurwitz condition (Routh,
1877; Zienkiewicz & Taylor, 1991) which states that for a polynomial
10 1 ... 0n n
nC z C z C , the real part of all roots will be negative if, for 0 0C ,
1 0C ,
1 3
0 2
0C C
C C
1 3 5
0 2 4
1 3
0
0
C C C
C C C
C C
and so on up to,
1 3 5
0 2 4
1 3
2
...
...
0 ...0
0 0 ...
... ... ... ...
0 n
C C C
C C C
C C
C
C
I
R
R Ii R Iz z iz
Rz
Iz
Chapter 3 Staggered Approach
76
By utilizing the above tools, the stability of the proposed staggered algorithms can
be assessed by implementing the general procedure described in the next section.
3.7 Stability of Staggered Coupling Scheme for a Test System
The stability of the presented staggered approach can be established by considering a
test system where both partitioned sub-domains are assumed to be condensed at the
interface (i.e. all the freedoms are at interface level). The stability analysis consists
of the following steps:
1. Time-discretization of the partitioned sub-domains of a test system, while the
applied force term is dropped.
2. Construction of the amplification matrix through following the algorithmic
steps of the staggered approach under consideration and obtaining the
characteristic equation of the amplification matrix.
3. Converting the characteristic equation of the amplification matrix into a
Routh-Hurwitz polynomial and applying stability check of the system.
Consider a coupled system which is partitioned into the sub-domains (ΩB) and (ΩT),
with each sub-domain discretised in the time domain using Newmark method, and
assume that both domains are condensed to a single degree of freedom at the
interface, neglecting the effects of damping for simplicity. The governing
equilibrium equations for the partitioned sub-domains of the test system are given
by:
Governing equilibrium condition for sub-domain (ΩT):
i i iT T T T TM U K U F (3.25)
Governing equilibrium condition for sub-domain (ΩB):
i i iB B B B BM U K U F (3.26)
Chapter 3 Staggered Approach
77
where iTF and i
BF are the interface forces applied to the structure and soil sub-
domains, respectively.
Following the staggered approach algorithmic steps presented in Section 3.4, the
response of the partitioned sub-domains of the above test system can be coupled as
discussed in the following, yielding to the construction of the amplification matrix:
STEP 1- The initial condition in the ΩB and ΩT sub-domains is prescribed, while at
the interface the initial conditions are:
i iT Bn n
U U (3.27)
i iT Bn n
U U (3.28)
i iT Bn n
U U (3.29)
Choosing the second-order central difference predictor from Table 3.1, the
prescribed displacement in partitioned sub-domain ΩT at time 1nt t is:
2
1 2T T T Tn n n n
tU U t U U
(3.30)
STEP 2 - Solving the ΩT sub-domain for 1
iT n
F
using Newmark method (see
Section 3.2):
2 1 1
ˆ 0i i iTT T T T Tn n
MK U F M U
t
2 2
1
21
ˆi iT T T T Tni
T n
M t K U t M UF
t
(3.31)
where:
Chapter 3 Staggered Approach
78
22
1 1ˆ2
i i i iT T T Tn n n
U U t U t Ut
(3.32)
STEP 3 - Applying equilibrium at the interface:
2 2
1
21 1
ˆi iT T T T Tni i
B Tn n
M t K U t M UF F
t
(3.33)
STEP 4 - Solving the ΩB for 1
iB n
U
using Newmark method:
2 1 1
ˆ 0i i iBB B B B Bn n
MK U F M U
t
2
1
21
ˆi iB B Bni
B nB B
t F M UU
M K t
(3.34)
where:
22
1 1ˆ2
i i i iB B B Bn n n
U U t U t Ut
(3.35)
The velocity and acceleration in sub-domain ΩB at time tn+1 would be:
11
ˆi i iB B Bnn
U U Ut
(3.36)
2 11
1 ˆi i iB B Bnn
U U Ut
(3.37)
where:
ˆ ˆ1i i i iB B B Bn n
U U t U U (3.38)
Chapter 3 Staggered Approach
79
STEP 5 - Prescribing the initial condition of the sub-domain T compatible with
that of sub-domain B for next time step:
1 1
i iT Bn n
U U (3.39)
1 1
i iT Bn n
U U (3.40)
Using a central difference predictor from Table 3.1, the prescribed displacements for
ΩT at time 2nt t is:
2
2 1 1 12T B B Bn n n n
tU U t U U
(3.41)
Substituting Equations (3.35) and (3.33) in Equation (3.34), the displacements of
sub-domain ΩB at time tn+1 takes the form of:
2
21
2
2
2 2
2
2 21
2
i iB TB Bn n
B B
B T iB n
B B
B T T T iB n
B B
M K tU U
M K t
M K t tU
M K t
M M M K t tU
M K t
(3.42)
The velocity at the interface of sub-domain ΩB at time tn+1 can also be calculated by
expanding Equation (3.36), as:
21
2 2
2
2 2 2
2
+
2 2 21 -
2
T Bi iB B nn
B B
B B B T iB n
B B
T B T B B iB n
B B
K K tU U
M K t
M K t K K tU
M K t
K t K t M M K t tU
M K t
(3.43)
Chapter 3 Staggered Approach
80
Similarly the acceleration at the interface of sub-domain ΩB at time tn+1 can also be
expressed in terms of displacement, velocity and acceleration of the previous time
step (tn ) as:
21
2
2 2
2
2 21
2
i iT BB B nn
B B
T B iB n
B B
T B T B iB n
B B
K KU U
M K t
K K tU
M K t
K K t M K tU
M K t
(3.44)
Now, the amplification matrix of the presented test system can be constructed in the
form of 1n n U A U , by knowing Equations (3.42), (3.43) and (3.44), with
A being the amplification matrix of the proposed staggered coupling scheme. The
amplification matrix can be derived as:
11 12 13
21 22 23
31 32 331
B B
B B
B Bn n
U A A A U
U A A A U
U A A A U
(3.45)
The equation 0 A I provides the characteristic polynomial yielding the
eigenvalues of the above amplification matrix as:
2 23 2
2
2 2
2
2
2
2 4 1 (2 2 1) 2 41
2
2 2 1 2 4 1 4 21
2
0
B T T B
B B
B T T B
B B
T T
B B
K t K t M M
M K t
K t K t M M
M K t
M K t
M K t
(3.46)
As mentioned before it is sufficient and necessary for stability that the modulus of all
the eigenvalues is less than or equal to one ( i.e. 1i ).
Chapter 3 Staggered Approach
81
Implementing the variable transformation as discussed earlier, the polynomial of
Equation (3.46) can be converted into a Ruth-Hurwitz polynomial given by Equation
(3.47) through 1
1
z
z
mapping, leading to:
2 2 32
2 2 22
22
22
14 2 4 2 4 4
14 1 4 1 4 4
12
10
B T B TB B
B T T BB B
T BB B
T BB B
K t K t M M zM K t
K t K t M M zM K t
K K t zM K t
K K tM K t
(3.47)
Therefore, the stability condition of 1i will simplify to the following Ruth-
Hurwitz conditions:
0
2
2
0
4 2 4 2 4( )0B T B T
B B
C
K K t M M
M K t
(3.48)
1
2
2
0
4 1 4 1 40T B B T
B B
C
K K t M M
M K t
(3.49)
1 3
0 2
2
22
0
2 1 2 1 2 1 2 10T B B T
B B
C C
C C
K K t M M
M K t
(3.50)
Chapter 3 Staggered Approach
82
1 3
0 2
1 3
2
32
0
0 0
0
2 1 2 1 2 1 2 10T B B T
B B
C C
C C
C C
K K t M M
M K t
(3.51)
Considering Equations (3.48)-(3.51), it can be concluded that the presented
staggered approach is only conditionally stable, as elaborated next.
Consider the above stability conditions under the assumption of very small time
steps (i.e. 0t ):
00
lim( ) 0 0B Tt
C M M
(3.52)
10
lim( ) 0 0B Tt
C M M
(3.53)
1 3
1 30 2
0 00 4
1 3
0
lim 0 lim 0 0 2 1 2 1 0
0B T
t t
C CC C
C C M MC C
C C
(3.54)
The Newmark method is unconditionally stable in linear analysis for all modes
regardless of the choice of t, if:
12 Newmark method unconditionally stable
2 (3.55)
The above conditions for the stability clearly leads to 0B TM M as 0t .
Therefore, it can be concluded that the stability of the presented algorithm is
guaranteed, provided that the mass of the coupled system at the interface level is
distributed in accordance to the previous condition. However, the above sufficient
criterion for stability is valid based on the assumption of a very small time step
( 0t ). In the more general case where this assumption is not accurate, the full
Chapter 3 Staggered Approach
83
conditions given by Equations (3.48)-(3.51) must be satisfied in order to achieve
stability.
The fact that the stability in the presented staggered procedure is conditional and is
guaranteed provided that 0B TM M as 0t , can be further demonstrated by
considering the following lemma:
If [A] is a square n×n matrix with real or complex entries and if ( 1 2, , , n ) the
eigenvalues of [S] are listed according to their multiplicities, then:
1 1det . n A (3.56)
Accordingly, for stability of the system of equations presented by Equation (3.45), it
is sufficient and necessary that the modulus of all the eigenvalues, i , of the
amplification matrix [A] satisfy:
1 ( 1,..., )i i n (3.57)
Considering Equations (3.56) and (3.57), it is clear that a necessary (but not
sufficient) condition for the stability of Equation (3.45) would be:
1 1det . 1n A (3.58)
Using Equation (3.45), it can be shown that the above necessary stability condition
simplifies to:
2
2det 1T T
B B
M K t
M K t
A (3.59)
which confirms the conditional stability of the scheme and shows that conditional
stability may be achieved provided that 0B TM M as 0t . Moreover, it is
important to note in relation to the above discussions that if 0B TM M , then
the use of a small time-step may be counterproductive in relation to stability.
Chapter 3 Staggered Approach
84
3.8 Accuracy
In addition to stability requirements of the staggered scheme, another major issue
related to staggered coupling is achieving the required level of accuracy. Unlike
stability, the lack of which has an obvious effect on progressive escalation of error
with time, assessing accuracy is not so straightforward. While the stability of a
solution is a measure of the boundedness of the approximate solution with time, the
accuracy of the staggered scheme is a measure of the closeness between the
approximate solution obtained by the staggered scheme and the exact solution. This
means that the degradation of the staggered approach with respect to the results
obtained by monolithic approach determines the level of accuracy. The accuracy of a
staggered scheme might be increased by reducing the time step, which as discussed
before also influences the stability. However, in nonlinear problems, stability and
accuracy are interlinked and considered together. In fact, the previously discussed
stability is based on the assumption of a linear dynamic response, and therefore the
derived stability conditions should only be employed as a rough guide. In general,
the accuracy of a staggered scheme can be established by performing a new analysis
that is undertaken with a smaller time step for comparison purposes.
3.9 Example
Here we consider a representative FEM-FEM coupled problem treated both by the
monolithic and staggered approaches. Consider the cantilever column of Figure 3.4
subject to an excitation acceleration signal applied to its bottom support, as given in
Figure 3.5. The length of the cantilever is 20m, and it utilises two rectangular cross
sections. The mass of the system is modelled with two concentrated masses of 2000
kg and 800 kg, in the middle and the free end of the cantilever, respectively. A
bilinear kinematic material model is employed, as illustrated in Figure 3.4, where
9 2210 10E Nm (elastic modulus), 0.01 (strain hardening factor) and
6 2300 10y Nm (yield stress).
Chapter 3 Staggered Approach
85
Figure 3.4: Example: geometric configuration and material response
Figure 3.5: Acceleration at the base
The above system is partitioned into two sub-domains namely T and B , as
shown in Figure 3.6, with three degrees of freedom at the interface (one rotational
and two translational). The partitioned problem of Figure 3.6 is analysed employing
a staggered approach using a central difference predictor, where sub-domain T is
Chapter 3 Staggered Approach
86
treated by Dirichlet boundary conditions (displacement) and sub-domain B is
treated by Neumann boundary conditions (force). The obtained results by the
staggered approach are compared to the corresponding results of the monolithic
problem. The analyses show that the scheme is not stable for 0.01 st and
0.001 st , which makes the staggered algorithm extremely computationally
expensive as the stability requirements dictate very small time steps B TM M .
This would be clearer when compared to the monolithic treatment which is stable for
0.01 st .
Figure 3.6: Problem partitioning
A comparison between the results of the staggered and monolithic approaches is
undertaken, where a time-step of 0.0005 st is considered in the staggered
coupling analysis. Figure 3.7 shows the variation with time of the rotational degree
of freedom at the interface for both the monolithic and staggered approaches.
Despite the very small chosen time step, it can be seen that at around 5 sec, the
staggered approach exhibits high frequency oscillations. The lack of
accuracy/stability can be further demonstrated by comparing the variation with time
of the horizontal (X) displacement at the interface and at the top of the cantilever as
depicted in Figures 3.8 and 3.9, respectively. Again, at around 5 sec the staggered
approach starts to show poor performance.
Chapter 3 Staggered Approach
87
Figure 3.7: Rotation at the interface
Figure 3.8: Horizontal displacement at the interface
Chapter 3 Staggered Approach
88
Figure 3.9: Horizontal displacement at the tip of the cantilever
The stability/accuracy of the staggered approach with the specified time step size
( .0005 st ) is clearly not satisfactory, hence another analysis was carried out
reducing the time step to ( .0001 st ). Although this is an extremely small time
step that makes the staggered approach extremely computational expensive, the
results of staggered analysis with .0001 st confirms that by reducing the time
step size a full coupling of the partitioned sub-domains can be achieved. The fact
that the stability and accuracy are achievable by the staggered coupling approach,
provided that small time steps are employed, is confirmed by comparing the results
against those of the monolithic treatment.
Figures 3.10 and 3.11 show the variation with time of the rotation and horizontal
displacement, respectively, at the interface for both the staggered and monolithic
approaches. The variation with time of the horizontal displacement and rotation at
the free end of the cantilever is also depicted in Figures 3.12 and 3.13, respectively.
Chapter 3 Staggered Approach
89
Figure 3.10: Rotation at the interface
Figure 3.11: Horizontal displacement at the interface
Chapter 3 Staggered Approach
90
Figure 3.12: Horizontal displacement at the tip of the cantilever
Figure 3.13: Rotation at the tip of the cantilever
Chapter 3 Staggered Approach
91
3.10 Conclusion
Coupling of partitioned sub-domains can be achieved using the staggered approach,
though this approach should be used with great care in relation to both stability and
accuracy. In fact, stability and accuracy considerations typically demand excessively
small time steps rendering this scheme computationally prohibitive for many
coupled problems. In general the stability of the staggered scheme is conditional on
the time step and also on the equivalent mass and stiffness on both sides of the
interface, though this depends on the formulation of the partitioned sub-domains and
the employed predictors. Although some progress may be achieved by modifying the
specifics of the staggered scheme, this is not considered here since the ultimate
performance will vary with the coupling application under consideration, and often
its implementation would conflict with the modular use of the structural and soil
solvers as black box solvers.
Clearly therefore, there is a need for enhancement of the performance of the
staggered coupling schemes to be stable and accurate for a wider range of time steps,
whilst maintaining the practical and computational benefits of the partitioned
treatment. This has led to the development of iterative coupling algorithms which
are mainly carried out by introducing corrective iterations in the staggered approach,
hence the name iterative methods. These methods, in addition to the aforementioned
enhancement of the stability and accuracy issues of the staggered approach, are
applicable to static as well as dynamic problems and facilitate parallel computing
through problem partitioning. Such methods are introduced and discussed in detail in
the following chapters.
Chapter 4
Iterative Coupling
4.1 Introduction
As shown in the previous chapter, in staggered coupling the compatibility condition
at the interface of the partitioned sub-domains may be violated, which significantly
affects the stability and accuracy of the coupling scheme. This is due to the inherent
approximation of the prediction stage, which consists of prescribing the predicted
interface displacements at the structure interface level (obtained using displacements,
velocities and accelerations from previous steps), and these are invariably different
from the displacements evaluated following force substitution into the soil model.
Consequently, the staggered approach should be used with great care, since its
stability is conditional and depends greatly on the size of time step. Stability and
accuracy issues related to the staggered approach typically demand excessively small
time steps, rendering this scheme computationally prohibitive for many coupled
problems. This has led to the development of coupling algorithms that are stable and
accurate for a wider range of time step size, which is mainly achieved by introducing
corrective iterations, hence the name iterative sub-structuring methods (Quarteroni &
Valli, 1999). It is worth mentioning that iterative coupling approaches, in addition to
Chapter 4 Iterative Coupling
93
the aforementioned treatment of the stability and accuracy issues of the staggered
approach, facilitate parallel computing through problem partitioning which could
lead to much greater computational efficiency. The other advantage of iterative
coupling algorithms is that, unlike the staggered approach, they could be readily
employed in both static and dynamic analysis of soil-structure interaction problems.
Consider an arbitrary soil-structure system with arbitrary boundary conditions as
shown in Figure 4.1. A common practice, that allows field-specific discretisation and
solution procedures in a partitioned treatment of the system, would be to decompose
the coupled system into two sub-domains according to their physical and material
properties, namely soil and structure sub-domains.
Figure 4.1: Partitioned treatment of soil-structure interaction
Assuming that each sub-domain is independently discretised by FEM using a step by
step time integration method such as Newmark, the governing equilibrium
conditions for the soil and structure sub-domains can be formulated independently as
given by Equations (4.1) and (4.2), where, without loss of generality, the response of
individual sub-domains is assumed to be linear elastic:
Structure sub-domain:
Chapter 4 Iterative Coupling
94
2
1
1 T T TT TT T TT T
Ti i ii iT T TT T TT Tn n n
U F UU UM C K
t t U F UU U
(4.1)
Soil sub-domain:
2
1
1 B B BB BB B BB B
Bi i ii iB B BB B BB Bn n n
U F UU UM C K
t t U F UU U
(4.2)
In the above, XM , X
C and XK are the mass, damping and the stiffness
matrices of the partitioned sub-domain X . XXU , X
XU , XXU and X
XF
correspond to displacement, velocity, acceleration and external load vectors for the
non-interface degrees of freedom in sub-domain X , while iXU , i
XU , iXU and
iXF correspond to displacement, velocity, acceleration and external load vectors
for the interface degrees of freedom in sub-domain X , respectively. and are
Newmark algorithm parameters, and is a function of displacement, velocity and
acceleration at time nt employed for determining the corresponding entities at time
1nt . It should be noted that the suggested discretisation of the partitioned sub-
domains is just for illustration, where the proposed method is in fact also applicable
to different desired types of discretisation techniques. Moreover, the static analysis
of the presented coupled system can be considered as a special case of the presented
dynamic analysis.
Although Equations (4.1) and (4.2) of the partitioned soil and structure sub-domains
cannot be solved independent of each other, by applying an iterative coupling
scheme and coupling the response of the partitioned soil and structure sub-domains
at the interface level, the partitioned sub-domains can be analysed separately. The
proposed solution scheme couples the response of the soil and structure sub-domains
by enforcing explicitly compatibility and equilibrium conditions at the interface.
Chapter 4 Iterative Coupling
95
Figure 4.2 shows a soil-structure interaction coupled system, decomposed into the
soil and structure sub-domains treated by iterative coupling algorithms. The
governing equations of the partitioned sub-domains are solved independently at each
time step (or load increment in the case of static analysis), using predicted boundary
conditions (either force or displacement) at the interface. These predicted boundary
conditions are then successively updated using corrective iterations, until
convergence to equilibrium and compatibility is achieved at the interface and within
the partitioned sub-domains. This enables the coupling procedure to have an
effectively similar overall accuracy and stability to the monolithic treatment.
Figure 4.2: Schematics of iterative coupling algorithms
The compatibility and equilibrium requirements that should be satisfied at the
interface of a soil-structure system can be generally defined in view of Equations
(4.1) and (4.2) as:
Compatibility condition:
0i iB Tn n
U U (4.3)
Chapter 4 Iterative Coupling
96
Equilibrium condition:
0i iB Tn n
F F (4.4)
Although in the above the assumption is that the soil and the structure always remain
in contact at the interface, the treatment of separation and slip can be a simple
extension through the use of interface elements that may be considered to be either
part of one of the sub-domains or even part of the interface model.
4.2 Iterative Coupling Algorithms
Iterative coupling algorithms, as discussed in Chapter 2, can vary significantly in
terms of the adopted computational procedure. Concerning the computational
method, the algorithms can be categorized into forms of sequential and parallel
coupling. Here, the term parallel coupling refers to a form of partitioned computation
in which obtaining the response of each independently modelled soil and structure
sub-domain is carried out simultaneously during coupling iterations. Unlike parallel
coupling, in sequential coupling the partitioned sub-domains are solved one after the
other at each iteration stage.
In addition to the parallel or sequential nature of the iterative coupling procedures,
these algorithms also differ in relation to the treatment of prescribed Dirichlet
(Displacement) and Neumann (Force) conditions at the interface of the partitioned
sub-domains in order to achieve convergence.
In the following, the various types of iterative coupling algorithms (Elleithy &
Tanaka, 2003) are adapted and categorized in the context of soil-structure interaction
analysis. Throughout this discussion, subscript n and superscript I denote the
time/load increment and the iteration number, respectively.
Chapter 4 Iterative Coupling
97
4.2.1 Sequential Dirichlet-Neumann Iterative Coupling
The sequential Dirichlet–Neumann coupling algorithmic steps are described as
follows and illustrated in Figure 4.3.
For n=1,2,…(number of load/time increments)
I=0 (iteration number)
STEP 1: At the start of each increment, the structure domain is loaded by the
external forces, TT n
F , while the displacements (Dirichlet data) at the interface
nodes, IiT n
U , are prescribed in accordance with the initial conditions:
IiT n
U U
STEP 2: The structural solver computes the response of the structure using Equation
(4.1), for non-interface displacements, ITT n
U , and interaction forces at the interface,
IiT n
F .
STEP 3: The corresponding interface forces at the soil domain can be calculated by
applying equilibrium:
0I Ii i
B Tn nF F
STEP 4: Based on these forces, IiB n
F , and the external loading applied to the soil
sub-domain, BB n
F , the soil solver computes the response of the soil domain for
IiB n
U and IBB n
U .
By comparing the initial prescribed interface displacement and the obtained interface
displacements in STEP 4, the following two scenarios can occur:
STEP 5: If convergence to compatibility has not been achieved, the new estimation
of the interface displacements according to the compatibility condition of Equation
Chapter 4 Iterative Coupling
98
(4.3) is applied to the structure domain, and iteration continues (I=I+1) from STEP 2
until convergence to compatibility is achieved.
STEP 6: If convergence to compatibility at the interface of partitioned domains has
been achieved, the solution proceeds to the next time/load increment (n=n+1).
Figure 4.3: Schematics of sequential D-N iterative coupling
4.2.2 Sequential Neumann-Dirichlet Iterative Coupling
The sequential Neumann-Dirichlet coupling algorithmic steps are described as
follows and illustrated in Figure 4.4.
For n=1,2,…(number of load/time increments)
I=0 (iteration number)
STEP 1: At the start of each increment, the structure domain is loaded by the
external forces, TT n
F , while the forces (Neumann data) at the interface nodes,
IiT n
F , are prescribed in accordance with the initial conditions:
IiT n
F F
0IiT n
U
IiB n
F 1IiT n
U
0
1
IiT n
U
Chapter 4 Iterative Coupling
99
STEP 2: The structural solver computes the response of the structure using Equation
(4.1), for non-interface displacements, ITT n
U , and displacements at the interface,
IiT n
U .
STEP 3: The corresponding interface displacements at the soil domain can be
calculated by applying compatibility:
0I Ii i
B Tn nU U
STEP 4: Based on these displacements, IiB n
U , and the external loading applied to
the soil sub-domain, BB n
F , the soil solver computes the response of the soil domain
for IiB n
F and IBB n
U .
By comparing the initial prescribed interface forces and the obtained ones in STEP 4
the following two scenarios can occur:
STEP 5: If convergence to equilibrium has not been achieved, the new estimation of
the interface forces according to the equilibrium condition of Equation (4.4) is
applied to the structure domain, and iteration continues (I=I+1) from STEP 2 until
convergence to equilibrium is achieved.
STEP 6: If convergence to equilibrium at the interface of partitioned domains has
been achieved the solution proceeds to the next time/load increment (n=n+1).
Chapter 4 Iterative Coupling
100
Figure 4.4: Schematics of sequential N-D iterative coupling algorithms
4.2.3 Parallel Dirichlet-Neumann Iterative Coupling
The Parallel Dirichlet–Neumann coupling algorithmic steps are described as follows
and illustrated in Figure 4.5.
For n=1,2,…(number of load/time increments)
I=0 (iteration number)
STEP 1: At the start of each increment, the structure sub-domain is loaded by the
external forces, TT n
F , while the displacements (Dirichlet data) at the interface
nodes, IiT n
U , are prescribed in accordance with the initial conditions:
IiT n
U U
Concurrently, the soil sub-domain is loaded by the external forces, BB n
F , while the
forces (Neumann data) at the interface nodes, IiB n
F , are prescribed in accordance
with the initial conditions:
IiB n
F F
0IiT n
F
IiB n
U 1IiT n
F
0
1
IiT n
F
Chapter 4 Iterative Coupling
101
STEP 2: The structural solver computes the response of the structure using Equation
(4.1), for non-interface displacements, ITT n
U , and interaction forces at the interface,
IiT n
F . Simultaneously, the soil solver computes the response of the structure using
Equation (4.2), for non-interface displacements, IBB n
U , and interaction
displacements at the interface, IiB n
U .
By comparing the initial prescribed interface displacements and forces and the
obtained interface displacements and forces in STEP 2, the following two scenarios
can occur:
STEP 3: If convergence to compatibility and equilibrium has not been achieved, the
new estimation of the interface displacements and forces according to the
compatibility and equilibrium conditions of Equations (4.3) and (4.4) is applied to
the structure and soil sub-domains respectively, and iteration continues (I=I+1) from
STEP 2 until convergence to compatibility and equilibrium is achieved.
STEP 4: If convergence to compatibility and equilibrium at the interface of
partitioned sub-domains has been achieved, the solution proceeds to the next
time/load increment (n=n+1).
Figure 4.5: Schematics of Parallel D-N iterative coupling
0IiT n
U
0IiB n
F
1IiB n
F 1Ii
T nU
0
1
IiB n
F
0
1
IiT n
U
Chapter 4 Iterative Coupling
102
4.2.4 Parallel Neumann-Dirichlet Iterative Coupling
The Parallel Neumann-Dirichlet coupling algorithmic steps are described as follows
and illustrated in Figure 4.6.
For n=1,2,…(number of load/time increments)
I=0 (iteration number)
STEP 1: At the start of each increment, the structure sub-domain is loaded by the
external forces, TT n
F , while the forces (Neumann data) at the interface nodes,
IiT n
F , are prescribed in accordance with the initial conditions:
IiT n
F F
Concurrently, the soil sub-domain is loaded by the external forces, BB n
F , while the
displacements (Dirichlet data) at the interface nodes, IiB n
U , are prescribed in
accordance with the initial conditions:
IiB n
U U
STEP 2: The structural solver computes the response of the structure using Equation
(4.1), for non-interface displacements, ITT n
U , and interaction displacements at the
interface, IiT n
U . Simultaneously, the soil solver computes the response of the
structure using Equation (4.2), for non-interface displacements, IBB n
U , and
interaction forces at the interface, IiB n
F .
By comparing the initial prescribed interface displacements and forces and the
obtained interface displacements and forces in STEP 2, the following two scenarios
can occur:
STEP 3: If convergence to compatibility and equilibrium has not been achieved, the
new estimation of the interface displacements and forces according to the
compatibility and equilibrium conditions of Equations (4.3) and (4.4) is applied to
Chapter 4 Iterative Coupling
103
the structure and soil sub-domains respectively, and iteration continues (I=I+1) from
STEP 2 until convergence to compatibility and equilibrium is achieved.
STEP 4: If convergence to compatibility and equilibrium at the interface of
partitioned domains has been achieved the solution proceeds to the next time/load
increment (n=n+1).
Figure 4.6: Schematics of Parallel N-D iterative coupling
4.2.5 Parallel Dirichlet-Dirichlet Iterative Coupling
The Parallel Dirichlet–Dirichlet coupling algorithmic steps are described as follows
and illustrated in Figure 4.7.
For n=1,2,…(number of load/time increments)
I=0 (iteration number)
STEP 1: At the start of each increment, the structure sub-domain and the soil sub-
domain are loaded by the external forces, TT n
F and BB n
F , while the displacements
(Dirichlet data) at the interface nodes of the partitioned sub-domains IiT n
U and
0IiT n
F
0IiB n
U
1IiB n
U 1Ii
T nF
0
1
IiB n
U
0
1
IiT n
F
Chapter 4 Iterative Coupling
104
IiB n
U are prescribed in accordance with the initial conditions (preferably enforcing
compatibility):
0 0I Ii iT Bn n
U U U
STEP 2: The structural and soil solver compute the response of the structure and soil
concurrently using Equations (4.1) and (4.2) for non-interface displacements, ITT n
U
and IBB n
U , and interaction forces at the interface, IiT n
F and IiB n
F , simultaneously.
By comparing the obtained forces at the interface of soil and structure sub-domains
in STEP 2, the following two scenarios can occur:
STEP 3: If convergence to equilibrium has not been achieved, a new estimation of
the interface displacements is applied to the structure and soil sub-domains, and
iteration continues (I=I+1) from STEP 2 until convergence to equilibrium is
achieved.
STEP 4: If convergence to equilibrium at the interface of partitioned domains has
been achieved, the solution proceeds to the next time/load increment (n=n+1).
Figure 4.7: Schematics of Parallel D-D iterative coupling
0 0I Ii iT Bn n
U U
1IiB n
U 1Ii
T nU
0 0I Ii iB Tn n
U U
0 0
1 1
I Ii iT Bn n
U U
0 0
1 1
I Ii iT Bn n
U U
Chapter 4 Iterative Coupling
105
4.2.6 Parallel Neumann -Neumann Iterative Coupling
The Parallel Neumann–Neumann coupling algorithmic steps are described as follows
and illustrated in Figure 4.8.
For n=1,2,…(number of load/time increments)
I=0 (iteration number)
STEP 1: At the start of each increment, the structure sub-domain and the soil sub-
domain are loaded by the external forces, TT n
F and BB n
F , while the forces
(Neumann data) at the interface nodes of the partitioned sub-domains IiT n
F and
IiB n
F are prescribed in accordance with the initial conditions (preferably enforcing
equilibrium):
0 0I Ii iT Bn n
F F F
STEP 2: The structural and soil solvers compute the response of the structure and
soil concurrently using Equations (4.1) and (4.2) for non-interface displacements,
ITT n
U and IBB n
U , and interaction displacements at the interface, IiT n
U and IiB n
U ,
simultaneously.
By comparing the obtained displacements at the interface of soil and structure sub-
domains in STEP 2, the following two scenarios can occur:
STEP 3: If convergence to compatibility has not been achieved, a new estimation of
the interface forces is applied to the structure and soil sub-domains respectively, and
iteration continues (I=I+1) from STEP 2 until convergence to compatibility is
achieved.
STEP 4: If convergence to compatibility at the interface of partitioned sub-domains
has been achieved the solution proceeds to the next time/load increment (n=n+1).
Chapter 4 Iterative Coupling
106
Figure 4.8: Schematics of Parallel N-N iterative coupling
4.3 Treatment of Interactive Boundary Conditions
The presented sequential and parallel coupling algorithms can be used in partitioned
treatment of the soil-structure interaction problems. However, the algorithms in
which the partitioned structure sub-domain is considered as subjected to Dirichlet
(displacement) boundary conditions at the interface, while the partitioned soil sub-
domain is subjected to a Neumann (force) or Dirichlet (displacement) boundary
condition at the same interface, are more suitable in the context of static soil-
structure interaction analysis. This is purely due to the fact that in static problems
only essential Dirichlet boundary conditions can be imposed on the interface of the
structure sub-domain, where applying the natural Neumann boundary conditions at
this interface results in singularity of the equilibrium equations for the structural sub-
domain. In fact, treating the structure interface with Neumann boundary conditions
in such a problem requires some additional conditions to make the structure sub-
domain solvable.
The above requirement narrows the choice of coupling algorithms in a general soil-
structure interaction analysis to Dirichlet-Neumann or Dirichlet-Dirichlet family of
Chapter 4 Iterative Coupling
107
iterative coupling schemes. Therefore the subject of this research is to present
domain decomposition methods for coupled analysis of soil-structure interaction
problems, with particular emphasis on Dirichlet-Neumann types of iterative
schemes, specifically the sequential D-N algorithms.
4.4 Convergence of Iterative Coupling
Despite the significant potential benefits of iterative coupling methods, a major issue
relates to whether convergence to equilibrium and compatibility at the interface can
always be enforced through successive iterations. Thus, besides choosing a suitable
iterative scheme it is important to address the issues related to the convergence
behaviour of the selected iterative scheme.
Achieving convergence and reasonable computational efficiency in the iterative
coupling procedures is directly dictated by the chosen update technique of interface
displacements and/or forces. This is demonstrated by considering the convergence
behaviour of an FEM-FEM coupled test system treated by a sequential D-N iterative
scheme with a trivial update of boundary conditions.
4.4.1 Convergence of Sequential D-N with Trivial Update
The objective of this section is to elaborate the basic procedure used for convergence
analysis of sequential D-N iterative coupling algorithm in the case of trivial update
of boundary conditions. Consider the following iterative coupling solution form:
New Oldcte (4.5)
in which λ and cte are constant parameters.
The subsequent iterations for the above equation, in which the solution process is
started by prescribing some arbitrary initial value 0I , would take the
following form (with variable I denoting the iteration number):
Chapter 4 Iterative Coupling
108
1C (4.6)
2 1C (4.7)
3 2C (4.8)
and in general:
1I IC (4.9)
Considering Equations (4.6), (4.7) and (4.8) the solution error, , in different
iteration stages can be written as:
2 2 1 1 (4.10)
3 3 2 2 1 12 (4.11)
Therefore, generally the solution error after I iterations would take the following
form:
11I I (4.12)
Considering Equation (4.12) the convergence of the presented coupling algorithm
can now be related to which can be viewed as the error reduction indicator.
According to Equation (4.12) for the successive approximation 1I Icte
to converge, 0n for any initial value of , it is necessary and sufficient
that is less than one. This can be extended to the matrix form as, for a successive
approximation 1I Icte to converge for any initial vector , it is
Chapter 4 Iterative Coupling
109
necessary and sufficient that the moduli of all eigenvalues of the error reduction
matrix, , are less than one.
Using this theorem we can investigate the convergence behaviour of the iterative
coupling algorithms by constructing the matrix during successive iterations. This
can be illustrated by considering a soil-structure interaction SDOF test system,
decomposed into the soil sub-domain ( B ) and the structure sub-domain ( T ),
treated by the sequential D-N iterative coupling procedure. Assume that each sub-
domain is discretised in time using the Newmark step-by-step integration method
and that both sub-domains are condensed at the interface. The governing equilibrium
equations for the partitioned soil and structure domains of the above mentioned
system are given as:
Structure sub-domain ( T ):
i i i iT T T T T T TM U C U K U F (4.13)
Soil sub-domain ( B ):
i i i iB B B B B B BM U C U K U F (4.14)
Following the steps of the sequential D-N coupling algorithm presented in Section
4.2.1 and considering a trivial update of boundary conditions at the interface in
successive iterations ( 1I Ii iT Bn n
U U ), the error reduction factor of this scheme can
be defined by considering a typical iteration:
STEP 1: Set the initial interface displacements, IiT n
U , for the structure sub-domain.
STEP 2-3: Compute the response of the structure forces at the interface and apply
equilibrium:
Chapter 4 Iterative Coupling
110
221 1
1 ˆ ˆI Ii i i iT T T T T T T T Tn n n n
F M K t C t U M U C Ut
(4.15)
1 1
I Ii iB Tn n
F F (4.16)
STEP 4: Solving soil domain for displacements at the interface:
2
1
21
ˆ ˆ ˆ ˆIi i i i iT T T T T T B B T T B BnI n n n ni
B nB B B
M K t C t U M U M U C U C U
UM K t C t
(4.17)
STEP 5: If convergence has not been achieved (i.e. 1 1
0I Ii i
B Tn nU U
) the new
estimation of the displacements will be applied to the structure sub-domain:
1
1 1
I Ii iT Bn n
U U
(4.18)
Now, Equation (4.17) can be rewritten in a new form as:
2
21 1
I Ii iT T TB Tn n
B B B
M K t C tU U cte
M K t C t
(4.19)
It can be shown that the compatibility default at the interface after K iterations can
be obtained from the following difference equation:
0 0.
I K I K I II K i i K i iB T B Tn n n n
U U U U U (4.20)
with:
2
2
1/ ( / )
1/ ( / )
T T T
B B B
K t M t C
K t M t C
(4.21)
Chapter 4 Iterative Coupling
111
The term 2 21 / ( / )K t M t C in the above equation is usually referred to
as the effective stiffness ( effectiveK ).
Comparing Equations (4.20) and (4.12), the necessary and sufficient condition for
convergence of the presented coupled test system can be written as:
2
2
1/ ( / )1
1/ ( / )
T T T
B B B
K t M t C
K t M t C
(4.22)
In the case of static analysis, this condition would simplify to:
1T
B
K
K (4.23)
Equation (4.22) shows that for the presented sequential D-N iterative coupling
algorithm with trivial update of boundary conditions to converge for any initial
prescribed displacement value, the condensed effective stiffness of the sub-domain
treated by Dirichlet data at the interface should be less than the condensed effective
stiffness of the sub-domain treated by Neumann boundary conditions. Here,
condensation is a process by which some of the degrees of freedom are eliminated
from the overall equilibrium providing a reduced set of equilibrium equations for the
remaining degrees of freedom.
To elaborate on the condensation procedure, consider the following system of
equations in which the condensed stiffness matrix corresponding to the interface
degrees of freedom iU is required:
11 12 x x
21 22 i i
K K U F=
K K U F (4.24)
Expansion of Equation (4.24) gives:
11 x 12 i xK U K U F (4.25)
Chapter 4 Iterative Coupling
112
21 x 22 i iK U K U F (4.26)
Substituting Equation (4.25) in (4.26) gives:
* *iK U = F (4.27)
where
* -122 21 11 12K = K - K K K (4.28)
1 *i 21 11 xF F K K F (4.29)
Matrix *K is called the condensed stiffness matrix.
The fact that the convergence behaviour is influenced by the ratio of the condensed
effective stiffness of the partitioned sub-domains can be further established by
constructing the general convergence behaviour of the example presented in the next
section.
4.4.2 Example
Here we examine the convergence behaviour and the effect of mass and stiffness of
coupled domains in a partitioned interaction problem treated by sequential D-N
iterative coupling algorithm using a trivial update of boundary conditions at the
interface. Consider the dynamic problem of the mass-spring system illustrated in
Figure 4.9a, where the problem is partitioned into two sub-domains B and T as
shown in Figures 4.9b and 4.9c respectively.
Chapter 4 Iterative Coupling
113
Figure 4.9: a) coupled mass-spring system, b) partitioned sub-domain B , and c)
partitioned sub-domain T
The governing equilibrium condition for the partitioned domain T can be written
in the form of:
1 1
ˆ0
ˆ
TT TTeffective T T
T Ti ii
T Tn n T n
UU FK M
U F U
(4.30)
with
1
2
0
0T
MM
M
(4.31)
1 2 2
2 2T
K K KK
K K
(4.32)
iBU
iBF iTU
iTF
Chapter 4 Iterative Coupling
114
2
1effectiveT T TK K M
t
(4.33)
Similarly the governing equilibrium condition for partitioned domain B can be
given by:
1
ˆ
ˆ
ii iBeffective B B
B BB BB
B Bn B n
UU FK M
U F U
0
(4.34)
with
3
4
0
0B
MM
M
(4.35)
3 3
3 3 4B
K KK
K K K
(4.36)
2
1effectiveB B BK K M
t
(4.37)
Following the steps of the sequential Dirichlet-Neumann iterative algorithm for
coupling the above partitioned sub-domains, where the Dirichlet boundary condition
is applied to the sub-domain T and the Neumann boundary condition on sub-
domain B , the following recurrence relationship is obtained:
1 1
I Ii iB Tn n
U U cte (4.38)
where:
2 2 4 23 4 4 2 1 2 2 1 2 2 1 2 1
2 2 4 21 2 1 4 3 3 3 3 4 4 3 4 3
K K t M K K t M K M K M K t M M
K K t M K K t M K M K M K t M M
(4.39)
Chapter 4 Iterative Coupling
115
The essential and sufficient requirement for the above system to converge using the
sequential Dirichlet-Neumann algorithm as demonstrated before would be:
2 2 4 23 4 4 2 1 2 2 1 2 2 1 2 1
2 2 4 21 2 1 4 3 3 3 3 4 4 3 4 3
1K K t M K K t M K M K M K t M M
K K t M K K t M K M K M K t M M
(4.40)
The convergence condition of Equation (4.40) can also be derived by constructing
the following ratio between the condensed effective stiffness of the partitioned sub-
domains as obtained before:
Condensed effective stiffness of sub-domain at the interface1
Condensed effective stiffness of sub-domain at the interfaceT
B
(4.41)
Considering Equations (4.33) and (4.37), and applying the condensation process of
Equation (4.28) the exact convergence condition of Equation (4.40) can be achieved:
22 2
2 21
1 2 2
23 3
3 24
3 4 2
1
M KK
Mt K Kt
M KK
Mt K Kt
(4.42)
Considering the above convergence condition for very small time steps (i.e. 0t ):
22 3
03
lim 1t
MM M
M
(4.43)
Equation (4.43) shows that if the mass at the interface is assigned such that the mass
of the domain treated by Dirichlet boundary condition is less than that of the sub-
domain treated by Neumann boundary condition, the convergence criteria of the
algorithm is satisfied in the limit as the time step becomes very small. It is important
to point out that the above condition is valid only if 22
0lim 0
tK t
and
Chapter 4 Iterative Coupling
116
23
0lim 0
tK t
, which would not be the case if one of the corresponding springs is
rigid. For example, if the spring associated with 2K is rigid 2K , it can be
shown mathematically that:
2
1 21 2 3
03
lim lim 1 ( )t K
M MM M M
M
(4.44)
In fact due to the infinite stiffness 2K , the mass participating at the interface level at
partitioned sub-domain T would be 1 2iTM M M , which according to
Equation (4.43) should be less than 3iBM M (i.e. i i
B TM M ). The same argument
holds for the case of 3K , where in this case 1iTM M should be less than
3 4iBM M M :
3
22 3 4
03 4
lim lim 1 ( )t K
MM M M
M M
Considering the above conditions, it can be concluded that in the sequential D-N
iterative coupling algorithm with trivial update of boundary conditions a conditional
coupling convergence can be achieved. In fact, the convergence can only be
achieved provided that the problem under consideration allows for suitable
partitioning strategies at the interface that satisfy the condition of Equation (4.41).
The conditional convergence behaviour of the scheme employing trivial update of
boundary conditions is unsatisfactory, since the effective stiffness for the structure
domain at the interface is larger than that of the soil domain for numerous real
problems regardless of the time step, while for other problems unrealistically small
time steps would be required to satisfy Equation (4.41). This illustrates the need for
further improvements to the sequential D-N iterative coupling scheme, where
different update techniques are proposed in Chapters 5 and 6 (relaxation and reduced
order methods), and their convergence characteristics are addressed
comprehensively. The assessment of these proposed techniques is carried out
Chapter 4 Iterative Coupling
117
through the development and application of an iterative coupling simulation
environment which is described in the next section.
4.5 Simulation Environment
In order to deal with different coupling methods utilizing robust update techniques
for nonlinear-soil structure interaction problems, as proposed in Chapters 5, 6 and 7,
a novel iterative coupling simulation environment is developed, utilizing discipline-
oriented solvers for nonlinear structural and geotechnical analysis. The developed
simulation environment is used to demonstrate the relative performance
characteristics and merits of various presented algorithms. This tool is also applied in
a number of case studies involving nonlinear soil-structure interaction with
nonlinearity in both structure and soil, thus leading to important conclusions
regarding the adequacy and applicability of various coupling techniques for such
problems as well as the prospects for further enhancements.
The simulation of soil-structure interaction via the partitioned approach is carried out
in this work through the coupling of two powerful FEM codes, ADAPTIC (Izzuddin,
1991) and ICFEP (Potts & Zdravkovic, 1999) that have been developed at Imperial
College London for advanced nonlinear structural and geotechnical analysis,
respectively.
It should be emphasised that although the approach developed in this work is applied
to the coupling of ADAPTIC and ICFEP, the related algorithms are general in nature
and equally applicable to the coupling of other existing nonlinear soil and structural
analysis tools.
4.6 ADAPTIC
ADAPTIC (Izzuddin, 1991) is an adaptive static and dynamic structural analysis
program which has been developed to provide an efficient tool for the nonlinear analysis
Chapter 4 Iterative Coupling
118
of steel and composite frames, slabs, shells and integrated structures. The program
features are described briefly hereafter.
The development of ADAPTIC was initially driven by the needs of the offshore industry
for accurate and efficient, nonlinear analysis of offshore structures subject to extreme
loading conditions. This motivated the development of pioneering adaptive nonlinear
dynamic analysis techniques for framed structures, accounting for geometric and
material nonlinearity. The program has been extensively developed to deal with different
extreme loading, such as earthquake, fire and blast, as well as numerous additional
structural forms, such as reinforced concrete and steel-decked composite slabs, cable and
membrane structures, and curved shells.
In ADAPTIC, inelastic analysis of steel frames may be performed by either of two
methodologies. The first is an approximate solution using ideal plastic hinge elements,
while the second is a more accurate solution employing elements which account for the
spread of plasticity across the section depth and along the member length. For reinforced
concrete and composite frames, inelastic analysis is performed using the second
approach only. The ADAPTIC library includes a number of uniaxial material models
which can be used to model steel, concrete and other materials with similar behavioural
characteristics.
In ADAPTIC the loading can be either in the terms of applied forces or prescribed
displacements/accelerations at nodal points. The loads can vary proportionally under
static conditions, or can vary independently in the time or pseudo-time domains. The
latter variation can be utilised for static or dynamic analysis. Different types of analysis
in ADAPTIC are namely:
static analysis with proportional loading,
static analysis with time-history loading,
dynamic analysis,
eigenvalue analysis.
Chapter 4 Iterative Coupling
119
The following show some of the various assumptions that can be made using
ADAPTIC in structural modelling:
elastic modelling,
geometric and material nonlinearity modelling,
novel frame, plate, membrane and shell elements,
extreme loading modelling (e.g. static, dynamic and fire),
plastic hinge modelling,
elasto-plastic modelling,
adaptive elasto-plastic modelling,
joints and boundary conditions.
4.7 ICFEP
ICFEP (Potts & Zdravkovic, 1999) is a powerful finite element program specifically
written for the analysis of geotechnical engineering problems. It has been used for
many numerical analysis research projects at Imperial College and is continually
being developed. ICFEP has successfully been applied to numerous practical
engineering projects. The comprehensive formulation of ICFEP makes it possible to
consider a very wide range of problems. The most common applications are the
analysis of deep basement excavations, embankments, slopes and tunnelling.
The behaviour of soil is in general highly nonlinear. In many practical situations
some part of the soil is in a state of yield. The global stiffness equation is nonlinear
and therefore its accurate solution is not simple. The ability of a finite element
program to obtain accurate results (in particular accurate failure or ultimate loads) is
very dependent on its solution technique. Commonly used solution methods are the
Tangent stiffness, Visco-plastic, and Newton-Raphson methods. ICFEP uses an
Chapter 4 Iterative Coupling
120
accelerated form of the Modified Newton-Raphson method incorporating a sub-
stepping stress point algorithm with automatic error control. This ensures the stress
at any point does not drift from the yield surface and violate the constitutive law. The
solution obtained is independent of the size of the load increment applied.
Convergence of displacements, rotations, pore fluid flux, forces, moments and pore
pressures is required before the solution is accepted.
The following types of analyses can be performed with ICFEP:
two-dimensional plane strain,
two-dimensional plane stress,
axi symmetric,
full 3D (Conventional and Fourier series aided) analysis,
drained and un-drained,
fully coupled consolidation,
large displacement,
cyclic,
partial saturation,
dynamics.
A number of different element types are available in order to model a particular
problem such as: four or eight nodded quadrilateral elements, eight or twenty nodded
hexahedral elements, four or sixteen nodded interface or joint elements and etc. The
followings show some of various constitutive modelling assumptions that can be
made using the ICFEP elements:
linear material behaviour (isotropic, transversely isotropic, anisotropic),
nonlinear and elasto-plastic material behaviour,
Mohr Coulomb - with/without strain hardening /softening,
Tresca - with/without strain hardening /softening,
Von Mises - with/without strain hardening /softening,
Chapter 4 Iterative Coupling
121
Drucker Prager,
Cam Clay - with/without Hvorslev surface, general or Mohr Coulomb shape
yield surface,
Modified Cam Clay - with/without Hvorslev surface, general or Mohr
Coulomb shape yield surface,
Dramen Clay - undrained and drained,
2 surface and 3 surface ‘bubble’ models,
Partly saturated soil models,
Boundary surface models for both clays and sand (e.g. MIT-E3, Papadimitriu
and Bouckovalas, respectively).
4.8 INTERFACE
Coupling of the ADAPTIC and ICFEP is carried out using a coupling program called
INTERFACE which utilises a sequential Dirichlet-Neumann type of iterative
coupling algorithms. The INTERFACE program is written in FORTRAN 95.
ADAPTIC and ICFEP run on separate processors as independent black box solvers,
where the task of communication and synchronization between the two individual
codes is achieved via INTERFACE that implements the iterative coupling methods
utilizing various update techniques. In this respect, the interface program manages
the retrieval, manipulation and passing the necessary data between the two field
programs during coupled analysis (See Figure 4.10).
Chapter 4 Iterative Coupling
122
Figure 4.10: Communication and synchronization between ADAPTIC and ICFEP,
via INTERFACE
4.9 Simulation Environment Architecture
The developed simulation environment is based on a client-server software
architecture, communicating over a computer network. Client-server architecture
refers to the relationship between a number of computer programs, in which the
client program makes a service request from other programs, the servers, which
execute the request.
The client software, here the INTERFACE program, can send data requests to the
connected servers, ADAPTIC and ICFEP, which in turn accept these requests,
process them, and return the requested information to the client. The important
characteristics of the INTERFACE as a client program are: i) initiation and
synchronisation of the requests to ADAPTIC and ICFEP, and ii) waiting for,
receiving and processing the replies from ADAPTIC and ICFEP. On the other hand,
Chapter 4 Iterative Coupling
123
ADAPTIC and ICFEP as server programs never initiate any activity without
receiving a request from the INTERFACE program.
In the developed simulation tool, ADAPTIC and ICFEP run on separate processors
as independent black box solvers and are started separately. After start-up, the two
programs wait until receiving a run request and prescribed interaction data from the
INTERFACE program in a specific sequence which is managed by the INTERFACE
program. Then the programs start to solve their own partitioned sub-domain as black
box solvers, during which the INTERFACE is waiting for the either program to
solve for equilibrium in its respective sub-domain under the prescribed interface
conditions and to return the complementary displacement/force conditions at the
interface boundary.
Using the obtained results, the state of coupling convergence at the interface of the
partitioned sub-domain is checked. If convergence is not achieved at the interface,
the INTERFACE program calculates the new estimates of the prescribed boundary
conditions and initiates the iterative coupling procedure within the same time/load
step in ADAPTIC and ICFEP. If convergence is achieved, the interface sends a
request of updating the initial condition for the next time-load increment to
ADAPTIC and ICFEP and assigns new prescribed initial boundary condition for the
structure partitioned sub-domains, moving the solution to the next time/load
increment. The sequence of interaction between the INTERFACE, ADAPTIC and
ICFEP programs is detailed in Figure 4.11.
Chapter 4 Iterative Coupling
124
Figure 4.11: Schematics of the interaction sequence between the INFERFACE,
ADAPTIC and ICFEP
Chapter 4 Iterative Coupling
125
4.10 Data Communication
As elaborated before, the communication system is based on client-server
architecture; therefore, only a few small subroutines are needed to link the
INTERFACE program to ADAPTIC and ICFEP. The task of these subroutines is to
receive requests from and send the required data to the interface. These subroutines,
provide a code-independent way for specification of relevant coupling parameters,
configurations and initiation of different procedures. This exchange of data between
the coupled codes is carried out using direct access formatted data files generated by
INTERFACE and accessible by both ADAPTIC and ICFEP.
For a given problem, the user specifies the input data files for ADAPTIC and ICFEP
corresponding to structure and soil partitioned sub-domains, respectively. These
contain the typical relevant information such as the finite element model, material
models, analysis type, etc. In addition to these, an input file for the INTERFACE
program is also specified which defines the coupling region, where the coupling
interaction occurs (soil-structure interface), for both the structure and soil partitioned
models. Moreover, the required number of time/load increments, the coupling
convergence criteria at the interface and the type of update technique required for the
coupling analysis are specified in the same data file. A sample structure of the
INTERFACE data file is given in Appendix A. After launching all three programs,
while ADAPTIC and ICFEP are waiting for the initiation request and the prescribed
interactive values to be posted by the INTERFACE, the INTERFACE starts
generating a formatted direct access FORTRAN data communication file based on
its own input data file. This file is just for communication purposes between the
programs, and is read-write accessible to all programs, where a sample structure is
given in Appendix B. After the communication data file has been generated by the
INTERFACE, the first calls of the augmented subroutines in ADAPTIC and ICFEP
codes identify the coupling regions taking part in the coupling process as described
in Section 4.9 (see Figure 4.12).
Chapter 4 Iterative Coupling
126
Figure 4.12: Data exchange structure
The coupling procedure continues until the required time/load steps are completed.
At this stage, both ADAPTIC and ICFEP produce the required results of their own
solved partitioned sub-domain in separate output files, which can be accessed
directly via their own post processing tools for analysing the results. In addition, the
effect of different coupling parameters on the convergence characteristics, including
the achieved compatibility and equilibrium defaults at the interface and the
convergence rate, can be analysed by considering the output file generated by the
INTERFACE program.
Chapter 4 Iterative Coupling
127
4.11 Concluding Remarks
In this chapter, different iterative coupling methods are presented for partitioned
analysis of soil-structure interaction problems, where it is assumed that the overall
domain is divided into physical partitions consisting of soil and structure sub-
domains. The advantage of the iterative coupling schemes in which the partitioned
structure sub-domain is treated by Dirichlet boundary was discussed, and particular
emphasis is therefore given in this work to D-N iterative coupling methods.
It is shown that an important feature of the proposed approach that needs to be
addressed comprehensively is the convergence behaviour of the scheme, which is
directly dictated by the chosen update technique during successive iterations. This is
further illustrated by means of an example in which a trivial update of boundary
conditions is assumed during the coupling iterations. The obtained convergence
behaviour of the sequential D-N iterative coupling scheme employing trivial update
of boundary conditions demonstrates the need for further improvement in the update
of boundary conditions. Accordingly, the convergence behaviour of D-N iterative
coupling methods is addressed in the following chapters, where more powerful
methods for the update of boundary conditions are discussed and developed.
Finally, a coupling simulation environment that has been developed utilizing
discipline-oriented solvers for nonlinear structural and geotechnical analysis is
described. The software architecture of the developed simulation environment,
which is based on sequential D-N iterative coupling algorithms is outlined, and the
structure of the data exchange between the various codes is elaborated.
The developed tool is used in this research to assess the adequacy and applicability
of various coupling methods in nonlinear soil-structure interaction analysis,
including the new methods proposed in this work. In this context, the developed
approach is believed to offer great potential towards providing an integrated
interdisciplinary computational framework which combines the advanced features of
Chapter 4 Iterative Coupling
128
both structural and geotechnical modelling for a variety of challenging problems in
the field of nonlinear soil-structure interaction.
Chapter 5
Interface Relaxation
5.1 Introduction
In Chapter 4 it was discussed that the critical algorithmic stage of an iterative
coupling algorithm such as sequential Dirichlet-Neumann (D-N) is the evaluation of
new estimates of the interface Dirichlet data during the coupling iterations. In fact, it
was shown that this determines the convergence characteristics to compatibility at
the interface (noting that the equilibrium condition at the interface is always satisfied
by prescribing the corresponding interaction forces). It was also demonstrated that in
sequential D-N using a trivial update of boundary conditions for FEM-FEM
coupling, depending on the status of the relative stiffness of the partitioned sub-
domains, only a conditional convergence could be achieved.
In order to improve the unsatisfactory convergence behaviour of using trivial update
of boundary conditions, frequently a relaxation of the iteratively updated Dirichlet
boundary conditions is augmented to the iterative coupling algorithms, hence the
term interface relaxation (Marini & Quarteroni, 1989; Elleithy & Tanaka, 2003).
Employing a relaxation scheme is one of the most common techniques used for
Chapter 5 Interface Relaxation
130
successive update of the boundary conditions in iterative coupling algorithms. In this
method, the convergence to compatibility of the presented D-N coupling algorithm is
accelerated and ensured by employing a suitable relaxation parameter. In this
respect, considering the sequential D-N iterative coupling algorithm presented in
Chapter 4, the update of boundary conditions at the interface level (in STEP 5) using
the relaxation technique can be expressed for a specific iteration I as:
11
I I Ii i iT I T I Bn n n
U U U (5.1)
In the above, I is a real positive parameter, which is called the relaxation parameter
and can improve convergence of the iterative coupling scheme.
It is clear from Equation (5.1) that the interface relaxation method possesses the
benefits of being simple to implement and having an undemanding process of
estimating new updates of boundary conditions in corrective iterations.
Although the superiority of the interface relaxation coupling, as highlighted in
Chapter 2, has been recognised more recently for coupled modelling of soil-structure
interaction (Hagen & Estorff, 2005a; Hagen & Estorff, 2005b; Elleithy et al., 2004)
there remain significant technical challenges related to algorithmic and
computational issues, particularly with reference to convergence issues.
In this chapter, the convergence characteristics of algorithms based on the sequential
Dirichlet-Neumann iterative coupling method are investigated for soil-structure
interaction problems. The efficiency of employing a constant relaxation scheme to
enhance the convergence characteristics and the effect of the partitioned sub-domain
properties on these characteristics are demonstrated through a number of case
studies. Furthermore, it is proposed that the performance of iterative coupling
methods for FEM-FEM coupling may be effectively enhanced through the use of an
adaptive relaxation scheme. Similar approaches have been introduced by Funaro et
al. (1998) for iterative coupling of partitioned second-order elliptic problems and by
Wall et al. (2007) focusing on fluid-structure interaction. In this respect, an
Chapter 5 Interface Relaxation
131
analogous procedure for evaluating the adaptive relaxation parameter in the context
of soil-structure interaction using FEM-FEM coupling is proposed here, and its
convergence characteristics is investigated. It is shown that, in contrast to the
traditional relaxation scheme, in which the relaxation parameter is typically
evaluated by trial and error, an adaptive relaxation scheme offers improved prospects
for achieving convergence and computational efficiency in complicated large scale
nonlinear problems. The evaluation of such prospects and the comparison of
different relaxation schemes are therefore primary objectives of this chapter,
particularly considering soil-structure interaction problems with nonlinearity in both
structure and soil.
5.2 Constant Relaxation
As mentioned before, the use of a relaxation scheme as in Equation (5.1) can
enhance convergence. Here, it is assumed that the relaxation parameter is a constant
positive real parameter during the coupling iterations (i.e. 1,..., I ), that should
be specified prior to the start of coupling analysis. Since the constant relaxation
parameter should be determined in advance, there are two major issues regarding the
applicability of such a technique, namely: i) determination of the range of suitable
relaxation parameters for the specific problem under consideration in order to
achieve convergence, and ii) selection of the optimum relaxation parameter in order
to achieve maximum computational efficiency.
Previous convergence analysis of sequential D-N coupling with trivial update of
boundary conditions for a SDOF test system at the interface, as presented in Chapter
4 (Section 4.4.1), demonstrated that convergence to compatibility is guaranteed for
any initial value, if:
2 2
Structure2 2
Soil
1/ ( / ) (Effective Stiffness)1
(Effective Stiffness)1/ ( / )
effectiveT T T T
effectiveBB B B
K t M t C K
KK t M t C
(5.2)
Chapter 5 Interface Relaxation
132
It can be shown that by introducing a suitable constant relaxation parameter, not only
an unconditionally convergent algorithm may be achieved, but also significant
improvement in the convergence rate may be attained by prescribing an optimum
relaxation parameter.
This can be illustrated by considering the same SDOF soil-structure interaction test
system and following the sequential D-N coupling algorithmic steps, while
employing the constant interface relaxation:
11
I I Ii i iT T Bn n n
U U U (5.3)
Assume that the following relationship holds at every iteration stage:
I Ii iB Tn n
U U cte (5.4)
with,
2
2
1/ ( / )
1/ ( / )
T T T
B B B
K t M t C
K t M t C
(5.5)
Substituting subsequently Equation (5.4) for K iterations while employing Equation
(5.3) for renewal of boundary conditions, the compatibility error at the interface
would take the following form of:
0 01 ( 1)
K K I IKi i i iB T B Tn n n n
U U U U
(5.6)
Now, it can be easily shown that for this relaxation scheme to converge,
1 1 should be less than one:
1 1 1 1 1 1 1 (5.7)
The above convergence condition restricts the relaxation parameter to be a positive
value less than two, as shown below:
Chapter 5 Interface Relaxation
133
1 1 1 0
0 221 1 1
1
(5.8)
Figure 5.1 shows the 1 1 value against the relaxation parameter for
different assumed values. Considering Equations (5.8) and (5.7), it can be shown
that for all possible partitioned soil and structure sub-domains stiffness ratios,
0, , there exists a range of relaxation parameters 0, 2 that
guarantees convergence to compatibility at the interface.
Figure 5.1: Variation of error reduction factor against relaxation parameter
The figure also illustrates that there exists an optimum relaxation value that ensures
convergence at the highest rate:
0 ( ) 1Optimum (5.9)
In fact, the optimum relaxation parameter corresponds to 0 , leading to:
Chapter 5 Interface Relaxation
134
1 1 ( )
11
EffectiveB
Effective Effective EffectiveT B TEffectiveB
KOptimum
K K KK
(5.10)
5.2.1 General Convergence Analysis
Discretization of a linear coupled soil-structure system can be described in general
by:
K U = F (5.11)
where K , U and F represent the global stiffness matrix (effective stiffness
matrix for dynamic analysis), the displacement vector and external force vector
(effective force vector for dynamic analysis) of the coupled soil-structure interaction
system.
Assuming that the coupled soil-structure system is composed of soil B and
structure T sub-domains, Equation (5.11) can then be re-written in its general
form as:
11 12
21 22 22 21
12 11
0
0
T T T TT T
T T B B i iext
B B B BB B
K K U F
K K K K U F
K K U F
(5.12)
where vectors XXU and X
XF correspond to displacements and external loads
respectively, for the non-interface degrees of freedom in sub-domain X , while
vectors iU and iextF correspond to displacements and external loads
respectively, for the interface degrees of freedom.
Chapter 5 Interface Relaxation
135
Henceforth for the sake of simplicity, but without loss of generality, it is assumed
that there is no external load applied at the interface of the coupled system (i.e.
0iextF ).
Decomposing the coupled system of Equation (5.12) into two sub-domains
according to their physical and material properties, namely soil and structure sub-
domains, the soil and structure sub-domains can be formulated independently,
where, without loss of generality, the response of individual sub-domains are
assumed to be linear elastic:
Governing equilibrium conditions for partitioned structure sub-domain:
11 12
21 22
T T T TT T
T T i iT T
K K U F
K K U F
(5.13)
Governing equilibrium conditions for partitioned soil sub-domain:
11 12
21 22
B B B BB B
B B i iB B
K K U F
K K U F
(5.14)
In the above, vectors iXU and i
XF correspond to displacements and external loads
for the interface degrees of freedom in sub-domain X , respectively.
The response of the above separately modelled sub-domains can now be coupled by
enforcing compatibility and equilibrium conditions at the interface using a sequential
D-N iterative coupling scheme, as follows:
STEP 1: At the start of each increment, the structure domain is loaded by the
external forces TT n
F , while the displacements at the interface nodes, iT n
U , are
prescribed in accordance with the initial conditions:
IiT n
U U (5.15)
Chapter 5 Interface Relaxation
136
where superscript I and subscript n denote iteration and increment numbers,
respectively.
STEP 2: The structural solver computes the response of the structure, using
Equation (5.13), for ITT n
U and IiT n
F :
-1
11 12
I I IT T T T iT T Tn n n
U K F K U (5.16)
21 22
I I Ii T T T iT T Tn n n
F K U K U (5.17)
It should be noted here that although Equations (5.16) and (5.17) are for a linear
response, the same entities can also be readily obtained for a nonlinear response from
the model of the structural sub-domain.
STEP 3: The corresponding interface forces at the soil sub-domain can be calculated
by applying equilibrium:
0I Ii i
B Tn nF F (5.18)
STEP 4: Based on these forces IiB n
F and the external loading applied to the soil
domain IBB n
F , the soil solver computes the response of the soil domain, using
Equation (5.14), for IiB n
U and IBB n
U :
1
11 12
21 22
I IB BB B
B Bn n
B BI Ii iB Bn n
U FK K
K KU F
(5.19)
Again, it is noted that although Equation (5.19) is for a linear response, the same
entities can also be readily obtained for a nonlinear response from the model of the
soil sub-domain.
Chapter 5 Interface Relaxation
137
STEP 5: If convergence to compatibility has not been achieved, the new estimation
of the displacements will be applied to the structure domain:
If 2
2
I Ii iB Tn n
U U , then
11
I I Ii i iT T Bn n n
U U U
1I I , go to STEP2
where is a real positive constant relaxation parameter
STEP 6: If convergence to compatibility at the interface of partitioned sub-domains
has been achieved, the solution will proceed to the next increment:
If I Ii iB Tn n
U U , then 1n n , go to STEP 1
The convergence characteristics of the method can be now established by
considering the above algorithmic steps and constructing the error reduction matrix.
Using Equations (5.17), (5.16) and (5.19), the interface displacements of the soil
sub-domain at iteration I , IiB n
U , can be written as:
I Ii iB Tn n
U U cte (5.20)
where,
1B TC CK K
(5.21)
1 1 1 -1
21 11 21 11
I IB B B B B T T TC B C Tcte K K K F K K K F
(5.22)
It is worth mentioning, that [λ] in Equation (5.21) is positive, where λ in Equation
(5.5) is assumed to be negative.
Chapter 5 Interface Relaxation
138
It is noted that, as evident from Equations (5.21) and (5.22), cte would be constant
during the coupling iterations at a specific time/load increment, while TCK and
BCK represent the condensed stiffness matrix of the soil and structure sub-
domains, respectively, corresponding to the interface degrees of freedom.
-1
22 21 11 12T T T T TCK K K K K (5.23)
-1
22 21 11 12B B B B BCK K K K K (5.24)
Since the equilibrium condition is automatically satisfied in the above algorithm,
only convergence to compatibility should be checked during the coupling iterations.
Considering Equation (5.20), the interface displacement values at the structure and
soil sub-domains at different iteration stages are obtained as:
0
0 0Iteration no. 0
iT n
i iB Tn n
U U
U U cte
(5.25)
1 0 0
1 1
1Iteration no. 1
i i iT B Tn n n
i iB Tn n
U U U
U U cte
(5.26)
2 1 1
2 2
1Iteration no. 2
i i iT B Tn n n
i iB Tn n
U U U
U U cte
(5.27)
Accordingly, the compatibility defaults at each of the above iteration stages are
given by:
0 0 0i i iB T Tn n n
U U U cte I (5.28)
1 1 1i i iB T Tn n n
U U U cte I (5.29)
Chapter 5 Interface Relaxation
139
2 2 2i i iB T Tn n n
U U U cte I (5.30)
By substituting the value of 1iT n
U from Equation (5.26) in Equation (5.29) the
compatibility default at iteration number 1 takes the following form:
1 1 0 01i i i i
B T B Tn n n nU U U U cte I (5.31)
Expanding Equation (5.31) leads to:
1 1
0 0 0
i iB Tn n
i i iB T Tn n n
U U
U U U cte
I I (5.32)
Substituting Equation (5.28) in the above gives:
1 1
0 0 0 0
i iB Tn n
i i i iB T B Tn n n n
U U
U U U U
I (5.33)
The above equation can be further simplified to:
1 1 0 0i i i iB T B Tn n n n
U U U U I I (5.34)
Similarly it can be shown that the compatibility default at iteration number 2 takes
the following form:
2 2 1 1i i i iB T B Tn n n n
U U U U I I (5.35)
and in general:
1 1I I I Ii i i iB T B Tn n n n
U U U U
I I (5.36)
Chapter 5 Interface Relaxation
140
Consequently, at iteration number K using a constant relaxation parameter, the
compatibility default at the interface of the decomposed soil-structure system, can be
obtained by the following difference equation:
0 0KK Ki i i iB T B TU U U U I I (5.37)
In view of Equation (5.37), the convergence of the presented coupling algorithm can
now be related to the eigenvalues of matrix K I I , where the matrix
can be regarded as the kth error reduction factor.
For the above successive iteration process to converge for any initially prescribed
value, it is necessary and sufficient that all eigenvalues of the matrix
I I be less than one in modulus:
1 1 1, 1,...,i i p (5.38)
where, i is the i th eigenvalue of matrix .
Assuming that eigenvalues of are positive, which is true for a positive definite
as would be the case in linear analysis, the above condition simplifies to:
max1 1 1 (5.39)
where, max is the largest eigenvalue of matrix .
Considering Equations (5.39) and (5.21), it can be shown that for all possible
partitioned soil and structure sub-domains, there exists a range of relaxation
parameters that guarantee convergence to compatibility of the coupling algorithm:
max
20 2
1
(5.40)
Chapter 5 Interface Relaxation
141
Moreover, the best choice of relaxation parameter, opt , as depicted in Figure 5.2, is
that for which:
max min1 1 1 1opt opt (5.41)
where, min is the smallest eigenvalue of matrix .
Therefore, the optimum relaxation value that ensures the convergence and holds the
best convergence rate can be defined as:
max min
1 10 1
1 12
optav
(5.42)
Figure 5.2: Convergence range and optimum relaxation parameter
The above convergence analysis, illustrated in Figure 5.2, clearly indicates that
convergence to compatibility in sequential D-N iterative coupling algorithms can be
ensured by using the interface relaxation update technique, at least if the coupled
system under consideration is relatively linear. Indeed, by predefining the relaxation
parameter in the range given by Equation (5.40), the reduction of compatibility error
at the interface is guaranteed. Moreover, Equation (5.42) demonstrates the existence
of an optimum relaxation parameter in the convergent range for which not only the
Chapter 5 Interface Relaxation
142
convergence to equilibrium is guaranteed, but also the convergence rate is optimum.
Furthermore, the theoretical aspects regarding the convergence behaviour as
discussed above show that the convergence characteristics of the constant relaxation
scheme are highly sensitive to the partitioned domain parameters, specifically the
condensed stiffness of the partitioned sub-domains at the interface.
Notwithstanding, in practice finding the suitable range of relaxation parameters and
the optimum constant relaxation parameter in soil-structure interaction problems
with multi degrees of freedom at the interface is difficult, since it requires first
constructing the matrix and then solving the associated eigenvalue problem, not
to mention additional complications for nonlinear coupled soil-structure systems.
Therefore, in order to avoid this process, the determination of an optimum constant
relaxation parameter must rely on trial and error, which may be applied to the initial
part of the coupled simulation after which full coupling analysis may be carried out.
5.2.2 Convergence Studies
The previous theoretical study has shown that the convergence of the constant
relaxation scheme for D-N iterative coupling is highly sensitive to the partitioned
sub-domain parameters, specifically the condensed stiffness of the partitioned sub-
domains at the interface. Here, these findings are demonstrated through some
illustrative FEM-FEM coupling examples.
5.2.2.1 Example 1: Dynamic FEM-FEM Coupling
The first example is a representative dynamic FEM-FEM coupled problem treated by
D-N iterative coupling scheme with a constant relaxation, where consideration is
given to a cantilever beam (Figure 5.3) under dynamic loading applied to its fixed
end (Figure 5.4). The length of the cantilever beam is 20m (L=10m), with a
rectangular cross section of 20.2 0.2 m . Mass of the system is modelled with two
concentrated masses of M1=2000 kg and M2=1000 kg, in the middle and the free
end of the cantilever respectively. Material modelling of the beam is described in
Chapter 5 Interface Relaxation
143
Figure 5.3, where 9 2210 10E Nm , 0.01 and 6 2300 10y Nm . The above
system is partitioned into two sub-domains, namely T and B , as shown in Figure
5.5, with three degrees of freedom at the interface (one rotational and two
translational).
Figure 5.3: Dynamic FEM-FEM coupling
Figure 5.4: Acceleration at the base
y
y
E
E
E
Chapter 5 Interface Relaxation
144
The partitioned problem of Figure 5.5 is analysed employing a constant relaxation
coupling algorithm, where sub-domain T is treated by Dirichlet boundary
conditions (displacements) while sub-domain B is treated by Neumann boundary
conditions (forces).
Figure 5.5: Partitioned sub-domains
Firstly, this problem is analysed to illustrate the trial and error process embedded in
the coupling procedure via constant relaxation scheme, for finding the convergence
range and the optimum relaxation parameter. In this regard, in order to demonstrate
the effect of the condensed effective stiffness matrix on the convergence behaviour,
the proposed problem is analysed for different problem partitioning types. This is
achieved by analysing the same problem with different mass ratios at the interface of
the partitioned sub-domains T and B ( 1 2/m m ). The different analysed models,
including the considered partitioned mass ratios at the interface, range of suitable
relaxation parameters and the optimum relaxation parameter associated with the least
computational cost, are presented in Table 5.1. These values have been obtained by
an initial process of trial and error for each model under consideration for 100 steps
with 0.01st , where the relaxation parameter is chosen as a real value in the range
of ]0, 2[ according to Equation (5.8).
Chapter 5 Interface Relaxation
145
Model 1
2
m
m Range of Suitable
Relaxation parameters Optimum Relaxation
S1 200 kg
0. 11800 kg
]0,1.14] [0.8,0.9]
S2 400 kg
0.251600 kg
]0,1.13] [0.7,0.8]
S3 600 kg
0.431400 kg
]0,1.1] [0.6,0.7]
S4 800 kg
0.61200 kg
]0,1] 0.6
S5 1000 kg
1.01000 kg
]0,0.9] 0.5
S6 1200 kg
1.5800 kg
]0,0.7] 0.4
S7 1400 kg
2.3600 kg
]0,0.5] 0.3
S8 1600 kg
4.0400 kg
]0,0.35] 0.2
S9 1800 kg
9.0200 kg
]0,0.15] 0.1
Table 5.1: Range of suitable and optimal relaxation parameter for different 1
2
m
m
Figure 5.6 shows the total number of coupling iterations required for various
constant relaxation parameters and the applicable range of relaxation parameters
ensuring convergence to compatibility. A tolerance of ε = 10-4 m was set for the error
based on the norm of compatibility defaults:
I Ii iB Tn n
U U M (5.43)
where M corresponds to the number of coupled degrees of freedom at the interface,
which is equal to 3 in this case.
Chapter 5 Interface Relaxation
146
Figure 5.6: Influence the effective mass ratio on convergence
The results confirm that the convergence characteristics of the D-N iterative scheme
using constant relaxation is very sensitive to the chosen relaxation parameter,
rendering its selection a very difficult task, as evidenced by the significant increase
in number of iterations for model S8, between α0.2 with 126 coupling iterations,
α0.1 with 307 coupling iterations and α0.3 with 331 coupling iterations.
Considering Figure 5.6 more closely, the optimum relaxation parameter for each
model is defined as the value corresponding to the minimum number of iterations
(see Table 5.1). The fact that the optimum relaxation parameter varies and is
sensitive to the partitioned model stiffness ratios is confirmed by the obtained
results.
The dependence of optimum relaxation and convergence rate on the partitioned
model characteristics (effective mass/stiffness) can be further demonstrated by
considering the same problem with different ratios of elastic modulus ( /T BE E )
for model S5 ( i.e. 1 2 1000m m kg ). The different analysed models, including the
Chapter 5 Interface Relaxation
147
considered elastic modulus ratios at the interface, range of suitable relaxation
parameters and the optimum relaxation parameter associated with the least
computational cost, are presented in Table 5.2.
Model Range of Suitable Relaxation parameters
Optimum Relaxation
K1 5.0
]0,0.27] 0.2
K2 4.0
]0,0.45] 0.35 K3 2.0
]0,0.75] [0.4,0.5]
K4 1.0 ]0,0.9] 0.5
Table 5.2: Range of suitable and optimal relaxation parameter for different T
B
E
E
Figure 5.7: Influence the effective stiffness ratio on convergence
As shown in Figure 5.7, the sensitivity of the convergence rate and value of the
optimum relaxation parameter to the problem characteristics is even more critical for
cases where the condensed effective stiffness at the interface of the sub-domain
treated by Dirichlet boundary condition, is much greater than that of the other
domain treated by Neumann boundary condition. In general, as this ratio increases
Chapter 5 Interface Relaxation
148
the optimum relaxation parameter tends to smaller values, while the range of an
applicable relaxation parameters reduces significantly; moreover, the convergence
rate is also considerably reduced, leading to a significant increase in computational
cost.
The results from the two previous parametric studies clearly indicate that by
employing an optimum relaxation parameter the convergence rate of the coupling
method can be considerably enhanced. This fact is demonstrated in Figures 5.8 , 5.9
and 5.10 for model S7 from Table 5.1, where it is clear that the by using an optimum
relaxation, the prescribed tolerance on compatibility defaults is achieved with far less
iterations than those with non-optimum constant relaxation parameters.
Figure 5.8: Error reduction for different relaxation schemes (Time = 3.6s)
0
0.002
0.004
0.006
0.008
0.01
0.012
0.014
0.016
0.018
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
Compatibility error at Interface (m
)
Iteration Number
Relaxation=0.1
Relaxation=0.2
Optimum Relaxation=0.3
Relaxation = 0.45
Convergence Criterion = 0.00017
Chapter 5 Interface Relaxation
149
Figure 5.9: Error reduction for different relaxation schemes (Time = 3.5s)
Figure 5.10: Error reduction for different relaxation schemes (Time = 4.18s)
Despite the shortcomings of the constant relaxation scheme and the difficulties of
determining the optimum relaxation parameter, the superiority of this method over
0
0.002
0.004
0.006
0.008
0.01
0.012
0.014
0.016
0.018
0.02
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Compatibility Error at Interface (m
)
Iteration Number
Relaxation = 0.1
Relaxation = 0.2
Optimum Relaxation = 0.3
Relaxation =0.4
Relaxation = 0.5
Convergence Criterion = 0.00017m
0
0.001
0.002
0.003
0.004
0.005
0.006
0 1 2 3 4 5 6 7 8 9 10
Compatibility Error at Interface (m)
Iteration Number
Relaxation = 0.1
Relaxation = 0.2
Relaxation = 0.3
Relaxation = 0.4
Relaxation = 0.5
Convergence Criterion = 0.00017 m
Chapter 5 Interface Relaxation
150
the staggered approach and its verification against the monolithic treatment are
demonstrated next for Model S7. For iterative coupling, the relaxation parameter is
taken at the optimal value of 0.3, and in all cases the dynamic analysis is undertaken
over 10s with a time step of 0.01t s .
Figures 5.11 and 5.12 show the variation with time of the rotation and horizontal
displacement, respectively, at the interface of the sub-domain T for both the
optimum relaxation and monolithic approaches. Similarly Figures 5.13 and 5.14
show the variation with time of the rotation and horizontal displacement,
respectively, at the interface of the sub-domain B for both the optimum relaxation
and monolithic approaches. The variation with time of the rotation and horizontal
displacement at the free end of the cantilever is also depicted in Figures 5.15 and
5.16, respectively. The graphs confirm that the results obtained from coupled
partitioned analysis match very well with those obtained from the monolithic
treatment within the prescribed compatibility tolerance of 410 m .
Figure 5.11: Rotation at the interface of sub-domain T
Chapter 5 Interface Relaxation
151
Figure 5.12: Horizontal displacement at the interface of T
Figure 5.13: Rotation at the interface of sub-domain B
Chapter 5 Interface Relaxation
152
Figure 5.14: Horizontal displacement at the interface of B
Figure 5.15: Rotation at the tip of the cantilever
Chapter 5 Interface Relaxation
153
Figure 5.16: Horizontal displacement at the tip of the cantilever
Figure 5.17: Rotation at the tip of the cantilever 5reduced tolerance to 10
Chapter 5 Interface Relaxation
154
Considering Figure 5.15 it can be seen, that there are slight differences between the
values obtained by the partitioned treatment and those obtained via monolithic
approach. This slight inaccuracy is due to the prescribed tolerance for convergence
to compatibility. In fact, by prescribing a smaller convergence tolerance a closer
match to the monolithic approach can be achieved, as demonstrated in Figure 5.17
where a smaller tolerance of 510 is used.
5.2.2.2 Example 2: Static FEM-FEM Coupling
Here, the static plane strain problem of Figure 5.18a is considered and discretised
using 8-noded quadrilateral elements. The presented system is partitioned into two
sub-domains, namely T and B , where each partitioned domain has 5 interface
nodes (10 DOFs) as shown in Figure 5.18b. The resulting partitioned problem is
treated by the D-N iterative coupling technique. As pointed out before, prescribing
Neumann boundary condition on T will result in singularity of the matrices for
static analysis; therefore, the T and B partitioned sub-domains are treated by
Dirichlet and Neumann boundary conditions, respectively.
Figure 5.18: a) Plane strain problem, b) Problem partitioning
Chapter 5 Interface Relaxation
155
Iterative coupling is undertaken with constant relaxation for different stiffness ratios
of partitioned sub-domains T and B ( /T BE E ). The results obtained from
coupled partitioned analysis match very well with those obtained from the
monolithic treatment within the prescribed compatibility tolerance.
Details of the analysed models, including the elastic modulus ratios, range of suitable
relaxation parameters and the optimum relaxation parameter associated with the least
computational cost, are presented in Table 5.3. These values have been obtained by a
process of trial and error for each model, where the relaxation parameter is chosen as
a real value in the range of ]0, 2[.
Model /T BE E Range of Relaxation Optimum
Relaxation
M1 8.0 (0-0.14] 0.13
M2 4.0 (0-0.26] 0.2
M3 2.0 (0-0.47] 0.4
M4 1.0 (0-0.75] 0.6
M5 0.5 (0-1.1] [0.6-0.8]
M6 0.2 (0-1.29] [0.8-0.9]
Table 5.3: Range of suitable and optimal relaxation parameter for different /T BE E
Figure 5.19 shows the number of coupling iterations required for various constant
relaxation parameters for each of the considered coupled systems, where
convergence is assumed at a tolerance of 41 10 L for the compatibility defaults
(see Equation (5.43), with L=1m being the characteristic element size. The results
confirm that the convergence behaviour is significantly influenced by the stiffness
ratios of the partitioned sub-domains.
Chapter 5 Interface Relaxation
156
Figure 5.19: Influence of relaxation parameter on convergence properties
Considering Figure 5.19, the optimum relaxation parameter for each model can be
easily defined as the one corresponding to the minimum number of iterations. The
fact that the optimum relaxation parameter varies and depends on the partitioned
model stiffness ratio has been illustrated earlier, and is further confirmed by the
currently obtained results. The sensitivity of the convergence rate and value of the
optimum relaxation parameter to the problem parameters is even more critical for
cases where the domain treated by the Dirichlet boundary condition is relatively
stiffer than the other domain treated by the Neumann boundary condition. In general,
as this stiffness ratio increases, the range of convergent relaxation parameters
significantly reduces, and the optimum relaxation parameter tends to smaller values,
leading to deterioration in the convergence rate and a significant increase in
computational cost. The convergence rates of the different relaxation schemes are
illustrated in Figure 5.20 for model M4, which demonstrate that the optimum
relaxation provided a much faster convergence rate than other non-optimum
relaxation schemes.
0
10
20
30
40
50
60
70
80
90
100
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3
Number of Iterations
Relaxation Parameter
M1 M2 M3 M4 M5 M6
Chapter 5 Interface Relaxation
157
Figure 5.20: Error reduction for different relaxation schemes (model M4)
The effect of mesh size on convergence characteristics of the relaxation scheme is
investigated next, where different discretisations of the problem are considered with
/ 1.0T BE E . The effect of the different mesh densities shown in Figure 5.21 on
convergence is shown in Figure 5.22.
The results show that as the discretisation of the partitioned sub-domains becomes
finer and as a result more interface degrees of freedom are employed, the range of
applicable relaxation parameter ensuring convergence becomes smaller. Moreover,
as evident from Figure 5.22, the optimum relaxation parameter changes noticeably.
For instance, considering the case of Mesh 0.5 the range of applicable relaxation
parameters is ]0, 1.1] and this would significantly be reduced to ]0, 0.6] for Mesh
2.0. In addition, the value of the optimum relaxation parameter for Mesh 0.5 is about
0.8, which is out of the applicable convergence range of Mesh 2.0 for which the
optimum relaxation parameter is about 0.5, highlighting the significant sensitivity of
the constant relaxation scheme to the partitioned sub-domains parameters.
Chapter 5 Interface Relaxation
158
Figure 5.21: Different discretatzions of model M4
Figure 5.22: Effect of mesh density on convergence
0
10
20
30
40
50
60
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1
Number of Total Iterations
Relaxation Number
Mesh 0.5
Mesh 1.0
Mesh 2.0
Chapter 5 Interface Relaxation
159
5.3 Adaptive Relaxation
As discussed earlier and demonstrated in the examples, finding the optimum
relaxation parameter based on Equation (5.42) will require solving the eigenvalue
problem of the matrix , which is not readily available using black box field
solvers for large multi-degree of freedom problems. Therefore, the optimum constant
relaxation parameter is usually found by a process of trial and error for every case
under consideration. Convergence conditions clearly indicate that convergence of the
Dirichlet-Neumann iterative coupling method depends on the stiffness ratio of the
partitioned soil and structure sub-domains. In linear analysis, the stiffness matrix
remains constant at all computational steps. However, for nonlinear analysis the
effective stiffness depends on the deformation state. This change of the stiffness of
the partitioned domains will have a significant effect on the convergence
characteristics of the coupling method at the interface level, where the optimum
relaxation parameter will change over the load/time increments. Accordingly,
ensuring convergence and computational efficiency in large scale nonlinear
problems, where the optimum relaxation parameter is to be determined over the full
range of response by trial and error, would be very difficult if not impossible.
Instead, such a scenario requires a dynamic change of relaxation parameter to ensure
optimal convergence and associated computational efficiency. This leads to the
concept of adaptive relaxation, where the relaxation parameter is determined during
coupling iterations, using error minimization techniques. With adaptive relaxation,
the use of trial and error for determining the relaxation parameter is avoided, whilst
leading to significant improvement in the convergence rate of iterative coupling, as
discussed hereafter.
5.3.1 Convergence analysis
Consideration is given here to an iterative coupling algorithm utilising an adaptive
relaxation parameter that changes during successive iterations. In this respect the
relaxation parameter is obtained, for a specific coupling iteration, from minimising
Chapter 5 Interface Relaxation
160
the compatibility and equilibrium defaults at the next iteration by using the
compatibility and equilibrium default history of the previous iterations.
Considering again Equation (5.20) for a linear coupled system with constant [],
while assuming an adaptive change of relaxation parameter, the interface
displacement values at the structure and soil sub-domains at different iteration stages
can be expressed as:
0
0 0Iteration no. 0
iT n
i iB Tn n
U U
U U cte
(5.44)
1 0 0
1 1
1 1
1Iteration no. 1
i i iT B Tn n n
i iB Tn n
U U U
U U cte
(5.45)
2 1 1
2 2
2 2
1Iteration no. 2
i i iT B Tn n n
i iB Tn n
U U U
U U cte
(5.46)
Accordingly, the compatibility default at each of the above iteration stages is given
by (see Equation (5.36)):
1 1 0 0
1i i i iB T B Tn n n n
U U U U I I (5.47)
Similarly it can be shown that:
2 2 1 1
2i i i iB T B Tn n n n
U U U U I I (5.48)
Therefore, the compatibility default at iteration number 2 can be rewritten as:
2 2 0 0
2 1i i i iB T B Tn n n n
U U U U I I I I (5.49)
Chapter 5 Interface Relaxation
161
Consequently at iteration number K, using an adaptive relaxation parameter, the
compatibility default at the interface of the decomposed soil-structure system, can be
obtained by the following difference equation:
0 0
1
K Ki i i iB T K B Tn n n n
U U U U I I I I (5.50)
Accordingly, at any iteration 0I :
1 1 1I I I IE E I I I I (5.51)
with:
I Ii iI B Tn n
E U U (5.52)
Considering Equation (5.51), the proposed method defines the adaptive relaxation
parameter at each coupling iteration stage ( 1I ), based on minimizing the
following:
2
1 1 12
I I IE I I (5.53)
Equation (5.53) is related to the interface compatibility error at previous coupling
iterations (clearly here the assumption is that convergence to compatibility has not
been achieved yet, that is 0I Ii i
B TU U ).
Expanding Equation (5.53) leads to:
2
1 1 1 1 1 1 2I I I I I IE E E (5.54)
Given that:
1I I IE E I I (5.55)
Chapter 5 Interface Relaxation
162
it can be shown that:
1 1 1
1 1I I I I
I I
E E E E (5.56)
Substituting Equation (5.56) back into Equation (5.54) leads to the following:
22 2 21
1 1 1 12 2 2
211 12
2
2
II I I I I I
I
II I I
I
E E E E E
E E E
(5.57)
Minimizing 1I with respect to 1I leads to:
11
2
1 12
1 2
1 2
0II
I I I I
I
I I
d
d
E E E
E E
(5.58)
Considering the positive sign of the second derivative:
22
12 21
20I I
II
dE E
d
(5.59)
it is evident that the value of 1I obtained by Equation (5.58) corresponds to the
minimisation of 1I .
Now, defining:
1I I Ii i iT T Tn n n
U U U
(5.60)
1I I Ii i iB B Bn n n
U U U
(5.61)
it can be shown that:
Chapter 5 Interface Relaxation
163
1
1
I I Ii i iT T T I In n n
U U U E
(5.62)
1I Ii i
B T I In nU U E E (5.63)
Therefore, Equation (5.58) for the adaptive relaxation parameter can be further
simplified to:
2
21 2
2
I I Ii i iT T Bn n n
II Ii i
T Bn n
U U U
U U
(5.64)
The automatic choice of relaxation parameter from either Equation (5.64) or (5.58)
can be used for successive evaluation of the interface displacements using the
iterative coupling algorithm for 0I . For the first iterative stage (i.e. 0I ), the
relaxation parameter can be chosen as an arbitrary real value to start the coupling
process. Although the choice of the first relaxation parameter does not have
influence on the convergence, a choice which is close to the optimal value of
Equation (5.42) will result in better error reduction for the first coupling iteration,
and hence to a fewer coupling iterations with adaptive relaxation.
5.3.2 Convergence studies
The previous theoretical study has shown that the relaxation parameter in D-N
iterative coupling may be adaptively adjusted for every iteration (I≥1), leading to
enhanced convergence and improved efficiency. Here, these findings are
demonstrated through some illustrative FEM-FEM coupling examples.
5.3.2.1 SDOF Test System
The coupled test system considered in Section 5.2 for constant relaxation is
investigated here with adaptive relaxation. In this case, since the interface of the
Chapter 5 Interface Relaxation
164
coupled test system consists of a single degree of freedom, Equation (5.64) can be
rewritten in the form of:
2
21 2
2
I I I Ii i i iT T Bn n n T n
I I II I i ii iT Bn nT Bn n
U U U U
U UU U
(5.65)
Expanding Equation (5.65) using Equations (5.61) and (5.4), it can be shown that:
1 1
1
1 1
11
I I Ii iB Bn n
I Ii iT Tn n
U U
U U
(5.66)
where in the above is defined by Equation (5.5).
Considering Equation (5.66), it is immediately observed that this equation
corresponds to the optimum relaxation parameter introduced by Equation (5.10).
However, in contrast with the constant relaxation approach, the selection of the
relaxation parameter does not require trial and error. In addition, the method holds an
optimum convergence rate as demonstrated by examining the scalar test system
presented in Section 5.2 for convergence analysis.
5.3.2.2 Example 1: Dynamic FEM-FEM Coupling
In this section, the efficiency of the adaptive relaxation scheme is evaluated through
considering the dynamic FEM-FEM coupling problem of Example 1 in Section
5.2.2.1. The problem is analysed for the different models presented in Tables 5.1 and
5.2, where comparison is made between the results of constant and adaptive
relaxation.
As before, the results obtained from partitioned analysis are in excellent agreement
with the monolithic approach, as expected. A comparison between the number of
Chapter 5 Interface Relaxation
165
coupling iterations using an optimum constant relaxation and adaptive relaxation is
provided in Table 5.4.
Model Optimum Relaxation
Average Iterations per time-step
Adaptive Relaxation
Average Iterations per time- step
S1 1850 1.8 1824 1.8
S2 1788 1.8 1783 1.7
S3 1657 1.7 1688 1.7
S4 1700 1.7 1698 1.7
S5 1598 1.6 1620 1.6
S6 1806 1.8 1772 1.8
S7 2092 2.1 1883 1.9
S8 3638 3.6 1887 1.9
K1 5230 5.2 2607 2.6
K2 2975 3.0 2546 2.5
K3 2180 2.2 1999 2.0
Table 5.4: Number of coupling iterations with adaptive and optimum relaxation for
1000 time-steps ( 0.01t s ) with a tolerance of 10-4 m
These results clearly indicate that the adaptive relaxation method is far superior to
the constant relaxation scheme, since not only does it avoid the process of trial and
error for finding the relaxation parameter, but it also enhances the convergence rate
of iterative coupling significantly. This fact is further demonstrated in Figures 5.23-
5.30, for the most critical cases with regard to number of coupling iterations required
for convergence with optimum relaxation, namely Model K1 and Model S8. These
figures present the convergence rate of the adaptive and corresponding optimum
relaxation in different randomly selected time steps of the analysis. The automatic
change of the relaxation parameter value is also illustrated in these figures, where it
Chapter 5 Interface Relaxation
166
is clear that adaptive relaxation method achieves the prescribed tolerance on
compatibility defaults with less iteration than with optimum or non-optimum
constant relaxation. In some exceptional cases, the optimum relaxation approach
achieves the prescribed compatibility tolerance in less iteration. For example, at
t=4.98s as illustrated in Figure 5.30 the adaptive relaxation converges in 3 iterations
while the optimum relaxation converges within 2 iterations. However it is important
to note that adaptive relaxation can overcome the initial poor convergence rate
arising from a poor initial relaxation parameter, subsequently achieving a superior
convergence rate to that obtained with optimum relaxation.
Figure 5.23: Error reduction at t=3.05s for model S8
Chapter 5 Interface Relaxation
167
Figure 5.24: Error reduction at t=3s for model S8
Figure 5.25: Error reduction at t=2.1s for model S8
Chapter 5 Interface Relaxation
168
Figure 5.26: Error reduction at t=4.9s for model S8
Figure 5.27: Error reduction at t=3.05s for model K1
Chapter 5 Interface Relaxation
169
Figure 5.28: Error reduction at t=1.85 s for model K1
Figure 5.29: Error reduction at t=4.94 s for model K1
Chapter 5 Interface Relaxation
170
Figure 5.30: Error reduction at t=4.98 s for model K1
5.3.2.3 Example 2: Static FEM-FEM Coupling
Here the static FEM-FEM coupled problem of example 2 in Section 5.2.2.2 is also
analysed using the adaptive relaxation technique with the same prescribed tolerance,
where a comparison between the number of coupling iterations using optimum
relaxation and adaptive relaxation is provided in Table 5.5.
As before, these results confirm that the adaptive relaxation method is far superior to
the constant relaxation scheme. This fact is further demonstrated in Figure 5.31 for
the most critical case with regard to convergence, namely model M1, where it is
clear that adaptive relaxation method achieves the prescribed tolerance for
compatibility defaults with 10 coupling iterations compared to the optimum constant
relaxation with 31 coupling iterations. Although the compatibility error after the first
iteration is significantly increased, this is only due to the initial choice for the
relaxation parameter at I=0 ( 0 1.0 ). In fact, if the constant relaxation approach is
applied with 1.0 , convergence would not be achieved. However, this is not a
Chapter 5 Interface Relaxation
171
serious problem for the adaptive relaxation approach, since the compatibility default
decreases significantly in subsequent iterations.
Model Optimum relaxation Adaptive relaxation
M1 31 10
M2 20 7
M3 11 6
M4 7 4
M5 6 5
M6 4 4
Table 5.5: Number of required coupling iterations with adaptive and optimum
relaxation
Figure 5.31: Error reduction for adaptive and constant relaxation schemes (Model
M1)
Chapter 5 Interface Relaxation
172
5.4 Soil-Structure Interaction Analysis
In this section two examples of soil-structure coupling are presented, utilising both
the constant and adaptive relaxation approaches. The first example is a simple linear
soil-structure interaction problem, as shown in Figure 5.32, which is aimed at
verifying the relaxation approaches for iterative coupling against the monolithic
treatment.
Figure 5.32: Linear soil-structure interaction
The problem consists of linear static plane strain analysis of a concrete cantilever
wall, resting on a flexible soil, loaded at the top with a horizontal force. In
partitioned analysis of the problem, the adaptive relaxation shows reasonable
convergence characteristics and converges within 3 coupling iterations to a
prescribed tolerance of 10-4 m. With the same problem also modelled monolithically,
the obtained results from both monolithic and partitioned analysis are in excellent
agreement, as illustrated in Figure 5.33. This figure shows the horizontal
displacement of the beam obtained for the non-interactive case (rigid base) and for
the interactive case by both the partitioned and monolithic approaches. This further
demonstrates that by enforcing convergence to equilibrium and compatibility in the
Chapter 5 Interface Relaxation
173
sequential D-N iterative coupling algorithm, a strong coupling of the partitioned sub-
domains can be achieved at the interface.
Figure 5.33: Monolithic vs. Partitioned Approach
The second example represents a typical urban situation, where due to foundation
settlements the force quantities in the structural members often are revised. The
example considers a steel frame resting on soil and subjected to static loading
(Figure 5.34), where nonlinear elasto-plastic constitutive behaviour of the soil as
well as geometric and material nonlinearity of the structure are taken into account.
The considered frame is taken from a building, which is designed for office purposes
and assumed to be loaded equally on each floor, where plan and elevation views are
shown in Figure 5.35 and 5.36, respectively. The soil-structure interaction analysis is
carried out assuming plane strain conditions in the soil using an effective out-of-
plane width of 1m, where the developed domain decomposition approach is
employed utilising ADAPTIC and ICFEP. The main objective of this study is to
establish the applicability and efficiency of the presented coupling algorithm using
the adaptive relaxation technique, highlighting its merits compared to constant
relaxation scheme.
Chapter 5 Interface Relaxation
174
Figure 5.34: Plane frame resting on soil
Figure 5.35: Plan view of the analysed building frame
Chapter 5 Interface Relaxation
175
According to the partitioned treatment, the considered soil-structure system is
partitioned physically into two sub-domains, soil and structure, where each sub-
domain is discretised separately according to its characteristics as shown in Table
5.6. The frame structure is modelled with ADAPTIC using cubic elasto-plastic
beam-column elements (Izzuddin & Elnashai, 1993), which enable the modelling of
geometric and material nonlinearity. The frame is discretised using 10 elements per
member for both columns and beams, and the material behaviour is assumed to be
bilinear elasto-plastic with kinematic strain hardening. The footings are discretised
using 4 elements per member.
The soil sub-domain is modelled with ICFEP using an elasto-plastic Mohr-Coulomb
constitutive model, with parameters chosen to represent the behaviour of London
clay (Table 5.6). The nonlinear solution procedure employed for analysing the soil
sub-domain is based on the Modified Newton-Raphson technique, with an error
controlled sub-stepping stress point algorithm. The soil continuum is discretised
using 8-noded isoparametric quadrilateral elements. The loading is modelled in the
structure partitioned sub-domain and applied in 10 load increments (5kN/m2). The
interface degrees of freedom are assumed to be at nodes that belong to both the
footings and soil underneath, totalling 30 interface freedoms for this case.
Structure Sub-domain Material Properties All beams and columns (steel)
Steel Grade = S355 Elastic Modules = 210 GPa Strength = 355 MPa Bilinear elasto-plastic with strain hardening factor = 1%
Foundation Beam (concrete)
Elastic Modulus = 30 GPa Linear material
Soil Sub-domain Material Properties Soil Angle of Shear resistance ( 𝛷′) = 22° Dilation angle (𝜈 ) = 11° Effective out of plane depth = 1m Cohesion = 20 kPa
Young’s modulus varies linearly with depth from 10000 kPa at the ground surface (dE/dZ=5000 kPa/m)
Elasto-plastic Mohr-Coulomb constitutive model
Table 5.6: Geometric and material properties of the partitioned soil-structure system
Chapter 5 Interface Relaxation
176
Figure 5.36: Geometric configuration of considered frame
To assess the merits of the presented relaxation coupling approaches, both constant
and adaptive relaxation are employed. For the constant relaxation approach, the
considered problem is analysed using different relaxation parameters ( 𝛼 ∈]0,1] ). Figure 5.37 shows the total number of coupling iterations required for various
constant relaxation parameters and the applicable range of relaxation parameters
ensuring convergence to compatibility over the full range of response consisting of
ten increments. A tolerance ε = 10-4 m was set for the compatibility error of each
coupled degree of freedom at the interface (see Equation (5.43)).
The results confirm that the convergence behaviour of the iterative scheme using
constant relaxation is very sensitive to the chosen relaxation parameter, rendering its
selection a very difficult task, as evidenced by the significant increase in number of
iterations between α=0.55 with 64 coupling iterations and α0.65 with 180
coupling iterations.
Chapter 5 Interface Relaxation
177
Figure 5.37: Influence of relaxation parameters on convergence properties
Table 5.7 presents the range of constant relaxation parameters which guarantee
convergence, the optimum relaxation parameter and the number of coupling
iterations required for convergence to the prescribed tolerance of 410 m for both
adaptive and constant relaxation schemes.
Relaxation Range
Optimum relaxation
Number of iterations with optimum relaxation
Number of iterations with adaptive relaxation
(0,0.65] 0.55 64 46
Table 5.7: Convergence characteristics of constant and adaptive relaxation schemes
The convergence rates of the three approaches in the first, fifth and last load
increments are illustrated in Figures 5.38, 5.39 and 5.40 respectively. This
demonstrates that the adaptive relaxation achieves a faster convergence rate than the
relaxation scheme, whether optimal or not.
Chapter 5 Interface Relaxation
178
Figure 5.38: Error reduction for different relaxation schemes for the first load
increment
Figure 5.39: Error reduction for different relaxation schemes for the fifth load
increment
Chapter 5 Interface Relaxation
179
Figure 5.40: Error reduction for different relaxation coupling schemes for the last
load increment
Figure 5.41 shows the number of coupling iterations required in each increment for
the different coupling schemes, which demonstrates that the optimum relaxation
parameter should be determined over the full range of response in nonlinear
problems.
In fact, in nonlinear problems, finding the optimum relaxation parameter by trial and
error is not computationally efficient. This is evidenced by considering the behaviour
of coupling with α0.6, where a relatively good convergence rate is observed in the
first three load increments (even better than the optimum relaxation with α0.55),
but as the coupled system presents more nonlinearity the convergence rate decreases
significantly.
Chapter 5 Interface Relaxation
180
Figure 5.41: Convergence performance over full range of response
These results demonstrate the superiority of the adaptive relaxation scheme, which
achieves much faster convergence than the constant relaxation scheme, whether
optimal or not. It is, however, worth observing that the adaptive relaxation scheme
on average still requires about 5 iterations per load step (for 10 increments), which is
relatively large in comparison with what would be necessary in a typical monolithic
treatment, thus highlighting the need for further enhancement in iterative coupling
algorithms.
Notwithstanding, the benefits of the developed simulation environment in the
practical assessment of nonlinear soil-structure interaction problems can now be
demonstrated by considering the results of this case study. The frame structure with
its applied loads transmits the loading to the soil, which in turn deforms due to
cumulative action of these loads. As a consequence, the soil deformation beneath the
structure transmits back additional deformations and corresponding forces to the
structure. This interactive process is continued until the whole coupled system
Chapter 5 Interface Relaxation
181
reaches a compatible equilibrium state. In the following, the deformation and stress
states of the coupled problem at the end of load application are briefly described.
The vertical deformation profile of the soil surface, with respect to the distance from
the middle column of the frame, obtained from coupled analysis is given in Figure
5.42, where the three troughs correspond to the locations of the three footings
showing their vertical settlement. Clearly, the generated level of vertical settlement,
of the order of 20cm and vertical differential settlement of the order of 10 cm,
requires the structural analysis model to account for geometric nonlinearity.
The deformed shape of the mesh of the soil partitioned sub-domain is also depicted
in Figure 5.43.
Figure 5.42: Vertical displacement of the soil surface
Chapter 5 Interface Relaxation
182
Figure 5.43: deformed mesh of the soil partitioned sub-domain
A vector plot of displacements in the soil sub-domain in the vicinity of the structure
foundations is also shown in Figure 5.44, where the relative magnitude of these
vectors reflects the mechanism of ground deformation.
Figure 5.44: Vector plot of displacements in the soil partitioned sub-domain
Figure 5.45 shows contours of stress level at the end of analysis under the applied
load on the structure. The stress level is the ratio, at the same mean effective stress,
of the current deviatoric stress to the deviatoric stress at failure. It therefore varies
from 0 to 1, where 1 indicates full plasticity and failure. It is evident from Figure
5.45 that the applied loading conditions have mobilised an extensive plastic zone
Chapter 5 Interface Relaxation
183
underneath the building. This zone is, however, bigger and deeper under the middle
footing, which is in agreement with the previous figures that show most of the
deformation and load concentration under the middle column.
Figure 5.45: Contours of stress level in soil partitioned sub-domain
Figure 5.46: Deformed shape and bending moment (kN.m) of the frame (scale=5)
Finally, the deformed shape and the bending moment contours of the partitioned
structure sub-domain are shown in Figure 5.46. It is evident from the deformed
Chapter 5 Interface Relaxation
184
shape of the structure and also the displacement vectors underneath each of the three
footings in Figure 5.44 that they experience rigid tilting (indicated previously in
Figure 5.42) and significant vertical settlements.
5.5 Concluding Remarks
This chapter investigates the use of interface relaxation coupling techniques in FEM-
FEM domain decomposition analysis of soil-structure interaction. In this respect, the
overall domain is divided into physical partitions consisting of soil and structure sub-
domains. Coupling of the separately modelled sub-domains is undertaken with
various interface relaxation algorithms based on the sequential iterative D-N sub-
structuring method, which ensures compatibility and equilibrium at the interface
boundaries of the sub-domains. This enhancement of the convergence characteristics
is achieved by employing a relaxation of the interface Dirichlet entities in successive
iterations. Various mathematical and computational characteristics of the coupling
method, including the governing convergence rate and choice of relaxation
parameter, are studied, where it is demonstrated that the convergence behaviour of
the constant relaxation scheme is very sensitive to the stiffness ratio of the
partitioned sub-domains. This renders the selection of an optimum relaxation
parameter very difficult, leading to considerable computational inefficiency,
especially for realistic large-scale nonlinear problems.
An adaptive relaxation scheme is also considered for enhancing the performance of
iterative coupling algorithms, where the choice of the relaxation parameter is easily
guided by the iterative corrections of Dirichlet entities at the interface. It is shown
that the adaptive scheme improves the convergence characterises in both linear and
nonlinear analysis significantly, where the number of coupling iterations required for
convergence could be reduced by over 50%. Moreover, the adaptive scheme has the
advantage of avoiding trial and error for the selection of an optimum, even adequate,
constant relaxation parameter. In general, adaptive relaxation schemes enjoy great
popularity in multiphysics coupling due to their simplicity and low computational
Chapter 5 Interface Relaxation
185
cost of determining the adaptive relaxation parameter, as evidenced by the simple
vector algebra involved.
In addition to demonstrating the key convergence characteristics of the considered
coupling algorithms with interface relaxation, the developed coupling simulation
environment has been used in a number of case studies, and where applicable
comparison with available monolithic treatment has been made. The results verify
the applicability of the developed simulation environment and also demonstrate that
by enforcing convergence to equilibrium and compatibility in the sequential D-N
iterative coupling algorithm, a strong coupling of the partitioned sub-domains can be
achieved at the interface.
It should be emphasised that the considered methods are generally applicable to the
coupling of various computational procedures that are used for nonlinear structural
and geotechnical analysis. In this context, these coupling methods have the potential
to provide an integrated interdisciplinary approach which combines the advanced
features of both structural and geotechnical modelling for a variety of problems in
soil-structure interaction analysis.
Finally, it should be noted that although using the adaptive scheme removes
significant difficulties in the conventional relaxation iterative coupling scheme, there
is a pitfall associated with both constant and adaptive relaxation schemes when the
partitioned sub-domain parameters dictate very small values of the relaxation
parameter for convergence. In such cases, relaxation methods breakdown with poor
convergence rates and significant computational inefficiency.
An example of such problems in the context of soil-structure interaction analysis is
presented in next chapter where both adaptive and constant relaxation techniques
become extremely computationally inefficient. In this respect, it is proposed that the
performance of iterative coupling methods may be effectively enhanced for
nonlinear analysis through the use of the condensed interface stiffness matrices of
the structure and soil partitioned sub-domains, providing an effective first-order
guide to iterative displacements at the soil-structure interface. As discussed in
Chapter 5 Interface Relaxation
186
Chapter 6, these techniques require the determination of the condensed stiffness
matrix, which typically involves more interface-related computations than the
relaxation approach. Notwithstanding, their superior convergence rate can be such
that their use is easily justified, if not essential, for iterative coupling of nonlinear
soil-structure interaction problems.
Chapter 6
Reduced Order Method
6.1 Introduction
The performance of interface relaxation iterative coupling schemes can be
effectively enhanced for nonlinear analysis through the use of the condensed
interface tangent stiffness of both the structure and soil models, depending on the
variant coupling algorithm under consideration. The condensed tangent stiffness is
either readily available or easily determined with current nonlinear field modelling
tools, and provides an effective first-order guide to iterative forces/displacements at
the soil-structure interface. Although such an approach involves more interface-
related computations than the relaxation approach, these computations are relatively
minor in comparison with those undertaken in the soil and structure models, and
therefore superior overall computational efficiency should be obtained due to the
enhanced convergence of iterative coupling. In this respect, convergence to
compatibility occurs for linear problems immediately at the first coupling iteration.
Although this immediate convergence does not normally occur for nonlinear
problems, employing this method for the update of interface Dirichlet boundary
conditions ensures a high convergence rate. Indeed, the enhanced approach should
Chapter 6 Reduced Order Method
188
bring the numerical performance of iterative coupling closer to the monolithic
treatment, whilst maintaining the practical and computational benefits of the
partitioned treatment.
Although the condensed tangent stiffness may be readily available with some
nonlinear field modelling tools, a more general approximation for the condensed
stiffness matrices is desirable and can indeed be achieved during the course of
coupling iterations, thus avoiding the need for explicit determination and extraction
of the stiffness matrices. In this respect, the condensed tangent stiffness matrix is
approximated via reduced order models, building on a previous approach by
Vierendeels et al. (2007) concerned with fluid–structure interaction problems.
Various significant modifications to this approach are proposed in this chapter,
leading to a versatile and efficient approach for coupled modelling of nonlinear soil-
structure interaction problems. In this regard, a major pitfall associated with the
original method causing divergence is overcome by a new selective
addition/replacement procedure of force and displacement mode vectors. Moreover,
in order to achieve better approximation of the condensed tangent stiffness matrix in
the initial stages of coupling iterations, a mixed reduced order method is proposed,
which achieves a more robust coupling technique than the conventional reduced
order method. In the following such approaches are discussed in detail.
6.2 Condensed Interface Tangent Stiffness
Consider domain decomposition of a soil-structure interaction problem as presented
below:
Governing equilibrium conditions for partitioned structure sub-domain:
11 12
21 22
T T T TT T
T T i iT T
K K U F
K K U F
(6.1)
Governing equilibrium conditions for partitioned soil sub-domain:
Chapter 6 Reduced Order Method
189
11 12
21 22
B B B BB B
B B i iB B
K K U F
K K U F
(6.2)
In the above, vectors XXU and X
XF correspond to the displacements and external
forces for the non-interface degrees of freedom, while iXU and i
XF correspond to
displacements and forces for the interface degrees of freedom, respectively.
Assume that an iterative coupling method is employed for coupling of the above
partitioned soil and structure sub-domains, where the compatibility and equilibrium
defaults at the interface of the structure and soil sub-domains for iteration number I
of load/time step n take the form:
I II i iU B Tn n n
U U (6.3)
I II i iF B Tn n n
F F (6.4)
Consider the new iterative estimation of the interface displacements and forces in the
successive iteration I+1 expressed incrementally in the general form:
1I I Ii i iT T Tn n n
U U U (6.5)
1I I Ii i iB B Bn n n
U U U (6.6)
1I I Ii i iT T Tn n n
F F F (6.7)
1I I Ii i iB B Bn n n
F F F (6.8)
where to a first order:
I Ii C iT T Tn n
F K U (6.9)
Chapter 6 Reduced Order Method
190
I Ii C iB B Bn n
F K U (6.10)
in which CTK and C
BK are the condensed tangent stiffness matrices at the
interface of the structure and soil sub-domains.
Clearly in order to achieve convergence to compatibility and equilibrium at the
interface of the partitioned soil-structure system at iteration I+1, the following
compatibility and equilibrium conditions should be satisfied:
1 10
I Ii iB Tn n
U U (6.11)
1 10
I Ii iT Bn n
F F (6.12)
Expanding Equation (6.12) using Equations (6.4), (6.7) and (6.8) gives:
0I II i i
F T Bn n nF F (6.13)
Similarly expanding Equation (6.11) using Equations (6.3), (6.5) and (6.6) gives:
0I II i i
U B Tn n nU U (6.14)
Substituting Equations (6.9) and (6.10) in Equation (6.13) gives:
0I II C i C i
F T T B Bn n nK U K U (6.15)
Now by substituting IiB n
U from Equation (6.14) in Equation (6.15) the following
is obtained:
0I II IC i C i
F T T B T Un nn nK U K U (6.16)
Therefore the incremental value of IiT n
U takes the following form of:
Chapter 6 Reduced Order Method
191
1I I Ii C C CT T B B U Fn nn
U K K K
(6.17)
Similarly, by substituting IiT n
U from Equation (6.14) in Equation (6.15) the
incremental value of IiB n
U can be obtained as:
1I I Ii C C CB T B T U Fn nn
U K K K
(6.18)
Substituting Equations (6.13), (6.9) and (6.10) in Equation (6.14) the incremental
value of IiT n
F can be obtained as:
11 1 1I I Ii C C CT B T U B Fn nn
F K K K
(6.19)
Similarly, the incremental value of IiB n
F can be obtained as:
11 1 1I I Ii C C CB B T U T Fn nn
F K K K
(6.20)
The above equations can be used in successive update of boundary conditions of the
different sequential and parallel coupling algorithms presented in Chapter 4 (see
Appendix C). Depending on the utilized coupling algorithm type, the following
equations could provide the new estimation of the iterative Dirichlet/Neumann
boundary conditions ensuring convergence.
11I I I Ii i C C CT T T B B U Fn nn n
U U K K K
(6.21)
11I I I Ii i C C CB B T B T U Fn nn n
U U K K K
(6.22)
11 1 1 1I I I Ii i C C CT T B T U B Fn nn n
F F K K K
(6.23)
Chapter 6 Reduced Order Method
192
11 1 1 1I I I Ii i C C CB B B T U T Fn nn n
F F K K K
(6.24)
Although, as evidenced by the above equations, this method requires the access to
the stiffness/flexibility matrices of the partitioned sub-domains and poses more
interface related computations compared to interface relaxation schemes, there are
major advantages associated with the above update technique. More specifically,
these advantages are, i) avoiding the need for using predefined relaxation parameters
and trial and error for ensuring optimal convergence in the coupling procedure ii)
offering a high convergence rate compared to relaxation schemes, and iii)
facilitating the development of different coupling algorithms (see Appendix C). In
this respect, as further shown in Appendix C, this approach provides richer and more
general update technique in iterative coupling algorithms compared to both adaptive
and constant relaxation, leading to a versatile and efficient approach for partitioned
modelling of coupled systems.
The general algorithmic steps of the sequential Dirchlet-Neumann iterative coupling
scheme, using the above update technique are presented in Table 6.1. In the
following sections this coupling algorithm will be examined for illustrative coupled
FEM-FEM examples in order to demonstrate its merits.
Chapter 6 Reduced Order Method
193
For n=1,2,…(number of load/time increments)
For I=0,1,…(number of iterations)
STEP 1: At the start of each increment, sub-domain T is loaded by the external
forces, while the displacements at the interface nodes are prescribed in accordance
with the initial conditions: IiT n
U U
STEP 2: The structural solver computes the response of the structure for : IiT n
F
STEP 3: The corresponding interface forces at the soil domain can be calculated by
applying equilibrium: 0I Ii i
T Bn nF F
STEP 4: Based on these forces and the external loading applied to the soil domain,
the soil solver computes the response of the soil domain for: IiB n
U
STEP 5: If convergence to compatibility has been achieved the solution proceeds to the next time/load increment (n=n+1) with:
STEP 6: If convergence to compatibility has not been achieved, the following new estimation of the displacements will be applied to the structure sub-domain and the iteration will continue (I=I+1) until convergence:
11I I Ii i C C CT T T B B U nn n
U U K K K
Table 6.1: Coupling Procedure
6.2.1 Numerical Examples
Here, the coupling procedure and convergence behaviour of the sequential Dirichlet-
Neumann algorithm are examined using the condensed interface stiffness matrix for
updating the Dirichlet boundary conditions. Consider the linear coupled spring
system consisting of eight spring elements associated with different stiffness as
illustrated in Figure 6.1. In order to perform the partitioned analysis procedure, the
coupled system is decomposed into two separately modelled partitioned sub-domains
T and B (Note that nodes 5 and 4 are the interface nodes corresponding to 2
interface degrees of freedom).
Chapter 6 Reduced Order Method
194
Figure 6.1: Coupled spring system
To facilitate the verification of the various coupling schemes presented hereafter, the
coupled problem is initially modelled and solved monolithically. Performing global
structural analysis using a monolithic approach results in the formation and solution
of the following global system of equations:
2
3
4
5
6
40 10 0 0 0 0
10 40 10 20 0 20
0 10 25 0 15 0
0 20 0 35 15 0
0 0 15 15 40 60
(6.25)
where in the above j denotes the displacement of node j.
Solving Equation (6.25) results in the following nodal displacements:
i i
j j
PK K
PK K
Chapter 6 Reduced Order Method
195
2
3
4
5
6
67
231268
231541
231463
231241
77
(6.26)
Clearly any result obtained by the partitioned analysis of the coupled spring system
should mach the results obtained by the monolithic approach given by Equation
(6.26). In order to perform the partitioned analysis using the condensed interface
stiffness approach, the partitioned sub-domains must be modelled in isolation, and
their corresponding condensed interface stiffness matrices must be obtained.
Considering sub-domain T in isolation, its governing equilibrium conditions can be
written in the form of Equation (6.27). In the following iX T
P and iX T
correspond
to the interface forces and displacements of node X in the partitioned sub-domain
ΩT, respectively.
4 4
5 5
6
15 0 15
0 15 15
15 15 40 60
i i
T T
i i
T T
P
P
(6.27)
Applying condensation on interface nodes 4 and 5, the condensed stiffness matrix at
the interface of the partitioned sub-domain T can be determined as:
75 45
15 0 15 1 8 815 150 15 15 45 7540
8 8
CTK
(6.28)
Chapter 6 Reduced Order Method
196
Similarly, the governing equilibrium condition for partitioned sub-domain B can be
written as:
2
3
4 4
5 5
040 10 0 02010 40 10 20
0 10 10 0
0 20 0 20
i i
B B
i i
B B
P
P
(6.29)
In the above iX B
P and iX B
correspond to the forces and displacement of the
interface node X in the partitioned sub-domain B , respectively.
Again by applying condensation on interface nodes 4 and 5, the condensed stiffness
matrix at the interface of sub-domain B is obtained as:
122 16
10 0 0 10 40 10 0 0 3 30 20 0 20 10 40 10 20 16 28
3 3
CBK
(6.30)
Utilizing the obtained condensed interface stiffness matrices of Equations (6.28) and
(6.30) in the update of boundary conditions, the partitioned sub-domains can be
effectively coupled at the interface using various coupling algorithms. In the
following, the presented example is coupled using the sequential Dirichlet-Neumann
coupling algorithm which is the main emphasis of this work. Demonstration of the
applicability of the presented update technique for other types of parallel/sequential
coupling algorithms can be found in Appendix D for the same example.
6.2.1.1 Sequential DirichletNeumann Iterative Coupling
Prescribing an initial interface Dirichlet data, 0
4 0Ii
T
and 0
5 0Ii
T
at the
interface of T at iteration (I=0), and solving the partitioned sub-domain T gives:
Chapter 6 Reduced Order Method
197
0 0
4 4 6
0 0 0
5 5 4
06
5
30 215 0 15
450 15 15 0
215 15 40 60 45
2
I Ii i
T T
I I Ii i i
T T T
Ii
T
P
P P
P
(6.31)
Prescribing the interaction Neumann data at the interface of B , by applying
equilibrium (i.e. 0 0
4 5
45 45 and
2 2
I Ii i
B BP P
), and solving the partitioned sub-
domain B gives:
2 2
3 3
0 0 0
4 4 4
0 0 0
5 5 5
5
6040 10 0 0 102010 40 10 20 3
45 / 20 10 10 0 67
120 20 0 20 45 / 2107
24
I I Ii i i
B B B
I I Ii i i
B B B
P
P
(6.32)
Comparing 0
4
Ii
B
and 0
5
Ii
B
with 0
4
Ii
T
and 0
5
Ii
T
, it is clear that convergence to
compatibility at the interface is not achieved at this iterative stage. Therefore, new
estimates of Dirichlet data should be calculated according to Equation (6.21) to
enforce convergence in the next iteration:
1
1
4
1
5
75 45 22 16 22 16 670 8 8 3 3 3 3 12.0 45 75 16 28 16 28 107
8 8 3 3 3 3 24
541
231463
231
Ii
T
Ii
T
(6.33)
Chapter 6 Reduced Order Method
198
Prescribing the above estimate for interface Dirichlet data at the interface of T , at
iteration I=1, and solving the partitioned sub-domain T :
11
44 6
1 1 1
5 5 4
1
6 5
541 241
231 7715 0 15463 130
0 15 15231 11
15 15 40 60 1300
77
Ii IiT
T
I I Ii i i
T T T
Ii
T
P
P P
P
(6.34)
Prescribing the interaction Neumann data at the interface of B by applying
equilibrium and solving the partitioned sub-domain B gives:
2 2
3 311 1
44 4
1 115 55
670 231
40 10 0 0 2682010 40 10 20 2311300 10 10 0 54111
23113000 20 0 2046377231
II Iii iBB B
I IIi iiB BB
P
P
(6.35)
Clearly the interface displacement values of 1
4
Ii
B
and 1
5
Ii
B
exactly match 1
4
Ii
T
and 1
5
Ii
T
, thus at the first coupling iteration (I=1) convergence to compatibility is
achieved. Moreover, the obtained results by the above coupling procedure given by
Equations (6.34) and (6.35) are identical to those obtained with the monolithic
treatment given by Equation (6.26).
6.2.2 Analogy between Interface Relaxation and Condensed Interface Stiffness
Approaches
Consider the sequential Dirichlet-Neumann coupling of partitioned sub-domains T
and B represented by Equations (6.1) and (6.2). Considering Equation (6.21) and
Chapter 6 Reduced Order Method
199
knowing that the equilibrium condition at the interface is already satisfied for each
coupling iteration in STEP 6 (i.e. 0I
F n ), the renewal of the boundary
conditions can be expressed as:
11I I Ii i C C CT T T B B U nn n
U U K K K
(6.36)
Rewriting the above equation in the form of:
11 1I I Ii i C C
T T B T U nn nU U K K
I (6.37)
and substituting Equation (6.3) in the above gives:
1I I Ii i iT B Tn n n
U U U β I β (6.38)
where in the above [I] is the identity matrix, with:
1β I + λ (6.39)
and:
1C CB TK K
λ (6.40)
It can be immediately noted that Equation (6.38) resembles the interface relaxation
scheme presented by Equation (5.1). However, 1β I + λ is a fully populated
relaxation matrix, as opposed to the single relaxation parameter of Equation (5.1),
which ensures convergence to compatibility at the interface of the partitioned sub-
domains while holding an optimum convergence rate. Indeed, the proposed approach
brings the numerical performance of iterative coupling approach close to the
monolithic treatment, whilst maintaining the practical and computational benefits of
the partitioned treatment. Moreover, in contrast to the constant relaxation coupling
algorithms, this method does not require the definition of certain parameters by a
Chapter 6 Reduced Order Method
200
process of trial and error for each case under consideration. Although the proposed
approach involves more interface-related computations than the relaxation approach,
overall computational efficiency is envisaged due to the enhanced convergence of
iterative coupling.
The convergence of the presented sequential D-N algorithms can now be established
for a linear case by following the algorithmic steps of sequential D-N, while
assuming that the iteratively updated boundary conditions have the following form:
1I I Ii i iT B Tn n n
U U U X I X (6.41)
where X is an unknown matrix which is to be determined to ensure convergence of
the iterative coupling algorithm.
Using Equation (6.41) for renewal of the interface Dirichlet data, in coupling of
Equations (6.1) and (6.2), the compatibility default after K successive iterations at
the interface of the decomposed soil-structure system, can be obtained by the
following difference equation:
0 0KK Ki i i iB B B BU U U U I - X I + λ (6.42)
with:
1C CB TK K
λ (6.43)
1
22 21 11 12C B B B BBK K K K K
= condensed stiffness matrix of the structure
sub-domain corresponding to the interface degrees of freedom, and
1
22 21 11 12C T T T TTK K K K K
= condensed stiffness matrix of the soil sub-
domain corresponding to the interface degrees of freedom.
Chapter 6 Reduced Order Method
201
For the successive iteration process of Equation (6.42) to converge for any initial
value, it can be clearly shown that X should take the form of:
1 X β I + λ (6.44)
in which case convergence to compatibility occurs for linear problems immediately
at the first coupling iteration as shown in the previous example. It is worth noting
that, in the previous example of Section 6.2.1, the adaptive relaxation update
technique (with an initial relaxation of 1.0 ) requires 7 coupling iterations for
convergence to compatibility with a tolerance of 410 . This highlights the high
convergence rate of the above method compared to both adaptive and constant
relaxation schemes. Of course, this immediate convergence does not normally occur
for nonlinear problems, though use of Equation (6.36) to update the interface
Dirichlet boundary conditions ensures a superior convergence rate to the relaxation
approach.
6.3 Approximation of the Condensed Tangent Stiffness
As evident from Equation (6.37), update of boundary conditions in successive
iterations using the condensed interface stiffness matrices, requires the determination
of these matrices during coupling iterations, which involves more interface-related
computations than the relaxation approach and requires direct access to the field
specific solver codes. Alternatively, suitable approximations for the condensed
stiffness matrices can be obtained, thus avoiding the need for obtaining these
matrices via the field codes. This builds on a previous approach by Vierendeels et al.
(2007) who utilized a procedure for constructing the reduced order model of
partitioned sub-domains for iterative coupling of fluid–structure interaction
problems, though various modifications are proposed here to provide a superior
approach for coupled modelling of nonlinear soil-structure interaction problems.
Chapter 6 Reduced Order Method
202
6.3.1 Condensed Interface Secant Stiffness Matrix
Consideration is given here to approximating the condensed interface tangent
stiffness matrix for nonlinear analysis using an initial secant stiffness matrix that
would be exact for linear analysis. In this respect, at the beginning of the coupled
analysis procedure the linear secant condensed stiffness matrices are obtained, and
these are used thereafter throughout the coupling iterations for the update of
boundary conditions. Assume that the partitioned sub-domain T is treated by
Dirichlet boundary conditions at the interface, while the partitioned sub-domain B
is treated by Neumann boundary conditions at the interface. The general
formulations for establishing the initial condensed secant stiffness matrices of the
partitioned sub-domains T and B are presented in the following.
Assume that there are N degrees of freedom at the interface of the coupled problem.
The following N+1 displacement vectors are prescribed at the interface of the sub-
domain treated by Dirichlet boundary conditions ( T ):
1 , 0,...,,..., ,...,Ti
T i ij iNiU i Nu u u (6.45)
where:
0 or 0
and 1iju
i j iu
i j i
(6.46)
and εu is a predefined incremental displacement, which should be relatively small in
comparison with the anticipated interface displacements in nonlinear analysis.
By obtaining the interface force vectors of the partitioned sub-domain T
corresponding to the above prescribed displacement vectors:
1 , 0,...,,..., ,...,Ti
T i ij iNiF i Nf f f (6.47)
Chapter 6 Reduced Order Method
203
the initial secant condensed interface stiffness matrix of the above partitioned sub-
domain corresponding to the chosen incremental displacement εu can be obtained as:
1 01 1 01
0 01
, 1,...,
i N
u u
CT
iN N NN
u u
f f f f
K i N
f f f f
(6.48)
Similarly in order to obtain the initial condensed flexibility matrix of the sub-domain
treated by Neumann boundary conditions, the following N+1 force vectors are
prescribed at the interface of the sub-domain ( B ):
1 , 0,...,,..., ,...,Ti
B i ij iNiF i Nf f f (6.49)
where:
0 or 0
and 1fij
i j if
i j i
(6.50)
and f is a predefined incremental force, which should be relatively small in
comparison with the anticipated interface forces in nonlinear analysis.
By obtaining the interface displacement vectors of the partitioned sub-domain B
corresponding to the above prescribed force vectors as:
1 , 0,...,,..., ,...,T
iB i ij iNiU i Nu u u (6.51)
the initial secant condensed flexibility matrix of the above partitioned sub-domain
corresponding to the chosen incremental displacement f can be obtained as:
Chapter 6 Reduced Order Method
204
1 01 1 01
1
0 01
, 1,...,
i N
f f
CB
iN N NN
f f
u u u u
K i N
u u u u
(6.52)
Considering Equations (6.48) and (6.52), a robust sequential Dirichlet-Neumann
coupling algorithm with high convergence rate for FEM-FEM coupling problems
can be achieved. Indeed, the use of the secant condensed stiffness matrix offers
convergence within one iteration for linear analysis, and can significantly enhance
convergence in nonlinear analysis in comparison with relaxation methods, depending
on the extent of overall system nonlinearity.
6.3.1.1 Example 1: Static FEMFEM Coupling
This example demonstrates the high performance and applicability of using the
initial secant condensed stiffness matrix in coupling linear static FEM-FEM
problems. Consider the plane strain problem of Figure 6.2a, which is discretized
using four noded elements. The presented system is partitioned into two sub-
domains, namely T and B , where each partitioned sub-domain has 3 interface
nodes at its interface (i.e. a total of 6 degrees of freedom), as shown in Figures 6.2b
and 6.2c.
The resulting partitioned problem is treated by the D-N iterative coupling algorithm.
Since, prescribing Neumann boundary conditions on T results in singularity of the
corresponding partitioned system of equations in static analysis, the T and B
partitioned sub-domains are treated by Dirichlet and Neumann boundary conditions,
respectively. The proposed problem is analysed using the Equations (6.48) and
(6.52) in the update of iterative boundary conditions and the results are compared
with the adaptive relaxation scheme. The problem is analysed for different stiffness
Chapter 6 Reduced Order Method
205
ratios of partitioned sub-domains T and B , which is achieved by considering
different ratios of elastic modulus ( /T BE E ).
The results obtained from coupled partitioned analysis match very well with those
obtained from the monolithic treatment within the prescribed compatibility tolerance.
The results corresponding to the analysed models, including the considered elastic
modulus ratios, compatibility tolerance and the number of required coupling
iterations for different considered schemes, are presented in Table 6.2.
Figure 6.2: a) Linear FEM-FEM coupled problem b) Partitioned sub-domain B c)
Partitioned sub-domain T
These results clearly show that using the condensed interface stiffness matrices of
the partitioned sub-domains is a far superior update technique compared to adaptive
relaxation, since it enhances the convergence rate of the coupling method
significantly. In fact, as expected, employing the condensed interface stiffness
matrices for this linear problem leads to convergence after only one iteration. This
fact is further demonstrated in Figures 6.3, 6.4 and 6.5 for models A1, A4 and A6,
respectively.
Chapter 6 Reduced Order Method
206
Model T
B
E
E
Adaptive
310 L
Adaptive
410 L
Adaptive
510 L
Adaptive
610 L
Condensed
610 L
A1 8.0 6 10 13 16 1
A2 4.0 5 6 10 13 1
A3 2.0 4 6 6 10 1
A4 1.0 3 5 6 6 1
A5 0.5 3 3 5 6 1
A6 0.2 1 3 3 4 1
Table 6.2: Required coupling iterations for different coupling schemes
Figure 6.3: Error reduction of different coupling schemes for model A1
Chapter 6 Reduced Order Method
207
Figure 6.4: Error reduction of different coupling schemes for model A4
Figure 6.5: Error reduction of different coupling schemes for model A6
Chapter 6 Reduced Order Method
208
6.3.1.2 Example 2: Dynamic FEMFEM Coupling
In this example, a representative dynamic FEM-FEM coupled problem is treated by
a sequential D-N iterative coupling scheme in which the effective condensed
interface stiffness matrices of the partitioned sub-domains are approximated initially
using the procedure outlined in Section 6.3.1 (It is worth noting that this approach
requires some adaptations for dynamic problems where the Dirichlet boundary
conditions are prescribed in terms of acceleration rather than displacements.
Moreover, here it is assumed that the time-step does not change during the analysis,
so it is ensured that the approximated effective stiffness does not change during the
analysis).
Consider a cantilever beam of Figure 6.6a subject to an excitation acceleration signal
applied to its bottom support, as given in Figure 6.7. The length of the cantilever
beam is 20m (L=10m) with a rectangular cross section of 20.1 0.1 m . The mass of
the system is modelled with two concentrated masses of M1=2000 kg and M2=2000
kg, at the middle and the free end of the cantilever respectively. A bilinear material
model is assumed, as depicted in Figure 6.6a, where 9 2210 10E Nm , 0.01
and 6 2300 10y Nm . The above system is partitioned into two sub-domains
namely T and B , as shown in Figures 6.6b and 6.6c, with three degrees of
freedom at the interface (one rotational and two transitional).
The above partitioned problem is coupled with sub-domain T treated by Dirichlet
boundary conditions (displacement) and sub-domain B treated by Neumann
boundary conditions (forces). The proposed problem is analysed for different
problem partitioning, where different mass ratios are assigned at the interface to the
partitioned sub-domains T and B ( 1 2/m m ), as listed in Table 6.3. The problem is
analysed for duration of 5s with 0.01t s using the initial linear elastic condensed
interface stiffness matrices, as outlined in Section 6.3.1.The obtained results for a
prescribed compatibility tolerance of 10-4m, are compared to adaptive relaxation and
summarised in Table 6.3.
Chapter 6 Reduced Order Method
209
Figure 6.6: a) Coupled dynamic FEM-FEM problem b) Partitioned sub-domain B
c) Partitioned sub-domain T
Figure 6.7: Acceleration at the base
Chapter 6 Reduced Order Method
210
Model 1
2
m
m
Number of Iterations using
condensed stiffness matrix
Number of iterations using
adaptive relaxation
C1 200 kg
0. 11800 kg
529 1624
C2 400 kg
0.251600 kg
476 1027
C3 600 kg
0.431400 kg
471 879
C4 800 kg
0.61200 kg
456 824
C5 1000 kg
1.01000 kg
451 758
C6 1200 kg
1.5800 kg
448 741
C7 1400 kg
2.3600 kg
448 752
C8 1600 kg
4.0400 kg
446 801
C9 1800 kg
9.0200 kg
444 821
Table 6.3: Number of required coupling iterations for 500 time-steps ( 0.01t s )
with a tolerance of 10-4 m
The results obtained from coupled partitioned analysis match very well with those
obtained from the monolithic treatment within the prescribed compatibility tolerance.
Considering Model C1, Figures 6.8 and 6.9 show the variation with time of the
rotation and horizontal displacement, respectively, at the interface of the sub-domain
T for both the partitioned approach with initial stiffness approximation and the
monolithic approaches.
Chapter 6 Reduced Order Method
211
Figure 6.8: Horizontal displacements at the interface of T
Figure 6.9: Rotation at the interface of T
Chapter 6 Reduced Order Method
212
The results show that the update of boundary conditions via the condensed interface
stiffness matrix approach provides a much better convergence rate than via the
adaptive relaxation scheme.
Considering the critical analysed model, C1, it is apparent that the update via
condensed interface stiffness matrix achieves superior convergence rate to adaptive
relaxation by more than 50%. In general, it is evident that the proposed procedure
outperforms the adaptive relaxation scheme in linear FEM-FEM coupling, as
expected. Figures 6.10 and 6.11 show the number of coupling iterations required
over the full 5s duration for Models C1 and C5, where it is clear that the condensed
stiffness approach converges typically after one coupling iteration when the response
is linear elastic.
Figure 6.10: Comparison between different coupling schemes for Model C1
Chapter 6 Reduced Order Method
213
Figure 6.11: Comparison between different coupling schemes for Model C5
6.3.1.3 Discussion on Nonlinear Analysis
It has been shown in the previous examples that the condensed interface stiffness
approach in iterative coupling enables convergence to compatibility within one
iteration for linear problems. However, this immediate convergence does not
normally occur for nonlinear problems through the use of the secant matrices given
by Equations (6.48) and (6.52). In nonlinear analysis, the effective stiffness depends
on the deformation state. In such cases the rate of convergence depends on the
change in the effective stiffness. This change of the stiffness of the partitioned sub-
domains can have a significant effect on the convergence characteristics of the
coupling method at the interface level, thus requiring the reformation of the tangent
stiffness matrices during iteration in order to ensure convergence to compatibility
using Equation (6.37), as discussed hereafter.
Chapter 6 Reduced Order Method
214
6.3.2 Reduced Order Method
Although the condensed tangent stiffness matrix at the interface could be determined
with existing nonlinear field modelling tools, this might require significant
modification of such computational tools, the extent and nature of which would vary
between one tool and another. It is therefore proposed that the condensed tangent
stiffness matrix may be reasonably and generally approximated by constructing
reduced order models of the structure and soil sub-domains. The benefit of such an
approach is that it does not require the explicit assembly of the stiffness matrices,
thus providing a general yet potentially efficient coupling technique.
Considering a sequential Dirichlet-Neumann coupling algorithm, the interface
condensed tangent stiffness matrices of the partitioned soil and structure sub-
domains can be approximated, at each coupling iteration stage 1I of a particular
time/load increment, by constructing the following reduced order models of the
partitioned sub-domains.
For the partitioned structure sub-domain, the following displacement mode matrix
can be constructed at coupling iteration 1I :
0 1,..., ,...,
I J I I Ii i i i i iT T T T T T
M IU U U U U U
TU (6.53)
where M corresponds to the number of coupled degrees of freedom at the interface
and 0,..., 1J I .
Similarly, the variation of the interface forces corresponding to TU can be
constructed as:
0 1,..., ,...,
I J I I Ii i i i i iT T T T T T
M IF F F F F F
TF (6.54)
Considering Equation (6.5), any IiTU can be projected as a linear combination of
the displacement modes given by Equation (6.53):
Chapter 6 Reduced Order Method
215
IiT n
U TU .δ (6.55)
where:
0 1,..., ,...,T
J I δ (6.56)
Thus the variation of the interface forces can also be approximated in a similar
manner as:
IiT n
F TF .δ (6.57)
The coefficients 𝛅 which provide the minimum error in approximating any IiTU
according to Equation (6.55) can be obtained as:
1
.IT T i
T nU
T T Tδ U U U (6.58)
Now, the reduced order model of the partitioned structure sub-domain can be
determined from combining Equations (6.57) and (6.58):
1
.I IT Ti i
T Tn nF U
T T T TF U U U (6.59)
Comparing the reduced order model of the structure given by Equation (6.59) with
Equation (6.9), it is evident that the condensed tangent stiffness matrix of the
partitioned structure sub-domain can be approximated as:
1
.T TC
TK
T T T TF U U U (6.60)
In a similar manner, the condensed tangent flexibility matrix of the partitioned soil
sub-domain can be obtained by constructing the following force and displacement
mode matrices at coupling iteration 1I :
Chapter 6 Reduced Order Method
216
0 1,..., ,...,
I J I I Ii i i i i iB B B B B B
M IF F F F F F
BF (6.61)
0 1,..., ,...,
I J I I Ii i i i i iB B B B B B
M IU U U U U U
BU (6.62)
where M corresponds to the number of coupled degrees of freedom at soil-structure
interface and 1,..., 1J I .
Now projecting IiB n
F as a linear combination of force modes given by Equation
(6.61):
IiB n
F BF .η (6.63)
where:
0 -1,..., ,...,T
J I η (6.64)
the variation of the interface displacement can also be approximated as:
IiB n
U BU .η (6.65)
Accordingly, the reduced order model of the partitioned soil sub-domain can be
determined as:
1
.I IT Ti i
B Bn nU F
B B B BU F F F (6.66)
Comparing the reduced order model of the soil partitioned sub-domain given by
Equation (6.66) with Equation (6.10), it is evident that the condensed tangent
flexibility matrix of the partitioned soil sub-domain can be approximated as:
11
.T TC
BK
B B B BU F F F (6.67)
Chapter 6 Reduced Order Method
217
The abovementioned approximations of the condensed tangent stiffness matrices at
the interface, as given by Equations (6.60) and (6.67), can be constructed through
coupling iterations and used in successive updates of the boundary conditions
utilising Equation (6.37).
It is worth noting that, in the original reduced order scheme proposed by Vierendeels
et al. (2007) for coupling fluid-structure interaction problems, the prescribed
Neumann data are calculated as a result of solving the reduced order models of the
partitioned sub-domains in consecutive iterations, (except for the first two iterations,
where equilibrium is applied). Here, the reduced order models are solved only once
to obtain the Dirichlet data prescribed to the structure sub-domain, while the
Neumann data prescribed to the soil sub-domain is obtained by enforcing
equilibrium at all iterations.
6.3.2.1 Singularity of .T
T TU U and .T
B BF F
Consider both the condensed interface stiffness and flexibility matrices of the
structure and soil partitioned sub-domains as approximated by Equations (6.60) and
(6.67), respectively. These approximations are only valid if and only if the square
matrices .T
T TU U and .T
B BF F are invertible. In fact, if .T
T TU U or
.T
B BF F become singular, Equations (6.60) and (6.67) cannot be used to
approximate the condensed interface stiffness/flexibility matrices during coupling
iterations.
In the following, the general conditions for which .T
T TU U and .T
B BF F are
non-singular are established, by considering that any square matrix n nΤ is
invertible if and only if all its rows and columns are independent.
A column/row vector is said to be independent of other column/row vectors if it
cannot be expressed as a linear combination of these vectors. The number of linearly
Chapter 6 Reduced Order Method
218
independent columns/rows is called the column/row rank of a matrix (Horn &
Johnson, 1999). The column rank and row rank are always the same, and therefore
they are simply called the rank of a matrix. In general, the rank of any p q matrix is
at most min( , )p q , and a matrix that has a rank as large as possible is said to have
full rank. If a square matrix has a full rank, none of its column/row vectors can be
expressed as a linear combination of the remaining vectors, and the matrix is
therefore non-singular.
In view of the above, a sufficient condition for non-singularity of n nΤ is for the
square matrix n nΤ to have a full rank (Horn & Johnson, 1999). Since both TU
and BF are M K matrices, with M denoting the number of interface degrees of
freedom and K the current coupling iteration number, it is clear that matrices
.T T TA U U and .
T
B BB = F F would be square K K matrices.
Given that:
( ) ( ) ( )Rank Rank Rank T TX X X X (6.68)
it is evident that:
.T
K K M KRank Rank
T T TΑ U U U (6.69)
.T
K K M KRank Rank
B B BΒ F F F (6.70)
Assuming that K<M, the requirement of .T T TΑ U U and .
T B BΒ F F to be
invertible dictates TU and BF to have rank of K (full column rank). Considering
these matrices as given by Equations (6.53) and (6.61), for TU and BF to have
full column rank, it is essential that in all coupling iterations the newly obtained
displacement and force increments at the interface nodes for T and B are linearly
independent of the previous corresponding incremental vectors.
Chapter 6 Reduced Order Method
219
In coupling iterations for which the coupling iteration number is less than the
number of interface freedoms, there is only the possibility of having linearly
dependent incremental displacement/force vectors. However, in problems where the
required number of coupling iterations for achieving convergence exceeds the
number of interface degrees of freedom, any newly obtained incremental
displacement/force vectors is undoubtedly a linear combination of the previous
corresponding independent vectors. As a result, the construction of the condensed
tangent stiffness matrix at the interface would not be possible in this case.
These facts highlight a major shortcoming of the original reduced order method
proposed by Vierendeels et al. (2007). In the following section, a new approach is
proposed for selective addition and replacement of the incremental
displacement/force vectors in TU and BF , thus overcoming the above
shortcoming and offering a robust reduced order model for approximating the
condensed interface tangent stiffness/flexibility matrices.
6.3.2.2 Selective Addition or Replacement of the Displacement/Force
Vectors
As mentioned above, in constructing the reduced order method for approximating the
condensed interface tangent stiffness matrix of the partitioned sub-domains, there is
a need for a procedure to make sure that any newly obtained interface
displacement/force modes added to TU and BF matrices are linearly independent.
In this respect, the procedure proposed in this section ensures that any newly
constructed interface incremental displacement vector, hereafter referred to as
displacement mode, is linearly independent of the previous modes:
1, 1
0 2 with , , , ,I I I Ti
T J IU
TU ζ ζ (6.71)
Consider the reduced order model for partitioned sub-domain T which is treated by
Dirichlet boundary conditions at the interface. Following the algorithmic procedure
Chapter 6 Reduced Order Method
220
of sequential D-N coupling method, at the starting coupling iteration (I=0), the
relaxation of the interface boundary conditions is applied. At the first coupling
iteration (I=1), TU and TF matrices would be single column matrices as shown
by Equations (6.72) and (6.73):
0,11I iT TU
U (6.72)
0,11I iT TF
F (6.73)
where:
,I J I Ji i iT T TU U U (6.74)
,I J I Ji i iT T TF F F (6.75)
In the next coupling iteration (I=2), there will be a new set of interface displacement
modes:
0,2 1,22,
I i iT TU U
TU (6.76)
Considering Equation (6.76) it is clear that, in order to avoid the singularity of
.T
T TU U , TU should be a full column rank matrix. In general at any iteration
2I , the column rank of I
TU given by Equation (6.77) should be checked:
0, , 1,,..., ,...,
I J I I II i i iT T TU U U
TU (6.77)
Given that TU of previous iteration (I-1) is of the form:
0, 1 , 1 2, 11,..., ,...,
I J I I II i i iT T TU U U
TU (6.78)
Chapter 6 Reduced Order Method
221
adding the currently obtained interface displacement mode, 1,I IiTU
, to all the
previously obtained interface displacement modes given by Equation (6.78) leads to:
1 0, 1 1, 2, 1 1,ˆ ,...,I I I I I I I Ii i i i
T T T TU U U U TU (6.79)
which simplifies to:
1 0, , 2,ˆ ,..., ,...,I I J I I Ii i i
T T TU U U TU (6.80)
Assuming the following normalization of the displacement modes:
,
I Ji iI J T Ti
T I Ji iT T
U UU
U U
(6.81)
1ˆ I TU could be written in the normalised form of:
1 0, , 2,ˆ ,..., ,...,
I I J I I Ii i iT T TU U U
TU (6.82)
At this point, it can be established whether the currently obtained displacement
mode, 1,I IiTU
, is a linear combination of the previous vectors in
1ˆ I TU . If
1,I IiTU
is a linear combination of
1ˆ I TU :
11, ˆ .
II IiTU
TU ζ (6.83)
with
0 2, , , ,T
J I ζ (6.84)
solving Equation (6.83) for ζ gives:
Chapter 6 Reduced Order Method
222
1,.
I IiTU
ζ ω (6.85)
where:
1
1 1 1ˆ ˆ ˆT TI I I
T T Tω U U U (6.86)
Substituting (6.85) into (6.83) gives:
11, 1,ˆ . .
II I I Ii iT TU U
TU ω (6.87)
Constructing the following conditions:
1, 1,, 1
I I I II i iT TU U
(6.88)
11, 1,1 ˆ, . .
II I I II i iT TU U
TU ω (6.89)
it can be clearly shown that if 1,I IiTU
is a linear combination of
1ˆ I TU as
assumed in Equation (6.83), the value of 1I should be equal to I .
Therefore, if in coupling iteration 2I the value of 1I I is greater than a
prescribed tolerance, , 1,I IiTU
cannot be expressed as a linear combination of
1ˆ I TU . In this case, 1,I Ii
TU
can be added to 1ˆ I
TU to construct a new linearly
independent I
TU :
0, , 1,,..., ,...,
I J I I II i i iT T TU U U
TU (6.90)
Chapter 6 Reduced Order Method
223
where the corresponding force modes at the interface of sub-domain T take also
the following form:
0, , 1,,..., ,...,
I J I I II i i iT T TF F F
TF (6.91)
On the other hand, when 1I I is less than the prescribed tolerance, 1,I IiTU
is a linear combination of 1ˆ I
TU . In such a case, the previous 1I TU and 1I
TF
can still be used for approximating the condensed stiffness matrix at the interface.
However, a better approach for nonlinear analysis would be to replace one of the
displacement modes in 1ˆ I
TU with the newly obtained 1,I IiTU
to form a new
full rank I
TU .
The replacement procedure considers the outcome of Equation (6.85):
1,
0 2, , , , .I IT i
J I TU
ζ ω (6.92)
It can be shown that if 1,I IiTU
replaces a displacement mode of
1ˆ I TU which
corresponds to the maximum absolute value of ( 0,..., 2)J j I , a full rank matrix
I
TU can be constructed which incorporates the latest incremental mode vectors.
Assuming that 0,..., 2( )J J I mMax , a new I
TU can be constructed where an
intermediate vector is removed from the assembled matrix of Equation (6.90):
0, ,, 1,
,..., ,..., ,...,I J I I II i i i
T T T
Taken
IiT
t
m
ou
U U UU
TU
(6.93)
and the corresponding force mode matrix at the interface of sub-domain T takes
the form:
Chapter 6 Reduced Order Method
224
0, ,, 1,
,..., ,..., ,...,I J I I II i i i
T T T
Taken
IiT
t
m
ou
F F FF
TF
(6.94)
In a very similar manner, an enhanced reduced order model is developed for
partitioned sub-domain B which is treated by Neumann boundary conditions at the
interface. In general, at any iteration 2I , the column rank of I
BF given by
Equation (6.95) should be checked:
0, , 1,,..., ,...,
I J I I II i i iB B B BF F F
F (6.95)
Given that BF of previous iteration (I-1) is of the form:
0, 1 , 1 2, 11,..., ,...,
I J I I II i i iB B B BF F F
F (6.96)
adding the currently obtained interface force mode, 1,I IiBF
, to all the previously
obtained interface force modes, given by Equation (6.96) leads to:
1 0, , 2,ˆ ,..., ,...,I I J I I Ii i i
B B B BF F F
F (6.97)
Assuming the following normalization of the force modes:
,
I Ji iI J B Bi
B I Ji iB B
F FF
F F
(6.98)
1ˆ I
B
F could be written in the normalised form of:
1 0, , 2,ˆ ,..., ,...,
I I J I I Ii i iB B B BF F F
F (6.99)
Chapter 6 Reduced Order Method
225
At this point, it can be established whether the currently obtained displacement
mode, 1,I IiBF
, is a linear combination of the previous vectors in
1ˆ I
B
F . If
1,I IiBF
is a linear combination of
1ˆ I
B
F :
11, ˆ .
II IiB BF
F ζ (6.100)
With
0 2, , , ,T
J I ζ (6.101)
solving Equation (6.100) for ζ gives:
1,.
I IiBF
ζ ω (6.102)
where:
1
1 1 1ˆ ˆ ˆT TI I I
B B B
ω F F F (6.103)
Substituting (6.102) into (6.100) gives:
11, 1,ˆ . .
II I I Ii iB B BF F
F ω (6.104)
Constructing the following conditions:
1, 1,, 1
I I I II i iB BF F
(6.105)
11, 1,1 ˆ, . .
II I I II i iB B BF F
F ω (6.106)
Chapter 6 Reduced Order Method
226
If in coupling iteration 2I the value of 1I I is greater than a prescribed
tolerance, , 1,I IiBF
cannot be expressed as a linear combination of
1ˆ I
B
F . In
this case, 1,I IiBF
can be added to
1ˆ I
B
F to construct a new independent I
BF :
0, , 1,,..., ,...,
I J I I II i i iB B B BF F F
F (6.107)
where the corresponding displacement modes at the interface of sub-domain B will
take the following form:
0, , 1,,..., ,...,
I J I I II i i iB B B BU U U
U (6.108)
On the other hand, when 1I I is less than the prescribed tolerance, 1,I IiBF
is a linear combination of 1ˆ I
B
F . In such a case, the previous 1I
B
F and 1I
B
U
can still be used for approximating the condensed stiffness matrix at the interface.
However, a better approach for nonlinear analysis would be to replace one of the
displacement modes in 1ˆ I
B
F with the newly obtained 1,I Ii
BF
to form a new full
rank I
BF .
The replacement procedure considers the outcome of Equation (6.102):
1,
0 2, , , , .I IT i
J I BF
ζ ω (6.109)
It can be shown that if 1,I IiBF
replaces a force mode of
1ˆ I
B
F which corresponds
to the maximum absolute value of ( 0,..., 2)J j I , a full rank matrix I
BF can be
constructed which incorporates the latest incremental mode vectors.
Chapter 6 Reduced Order Method
227
Assuming that 0,..., 2( )J J I mMax a new I
BF can be constructed where an
intermediate vector is removed from the assembled matrix of Equation (6.107):
0, , , 1,
,..., ,..., ,...,I J I I II i i i
B B B B
Taken out
m IiBF F FF
F
(6.110)
and the corresponding the corresponding displacement mode matrix at the interface
of sub-domain B takes the following form:
0, , , 1,
,..., ,..., ,...,I J I I II i i i
B B B B
Taken out
m IiBU U UU
U
(6.111)
6.3.2.3 Singularity of Approximated Stiffness/Flexibility Matrices
Considering the presented reduced order method, Equations (6.60) and (6.67) are re-
written as:
1
.T TC
T M K K M M K K MM MK
T T T TF U U U (6.112)
11
.T TC
B M K K M M K K MM MK
B B B BU F F F (6.113)
In view of the discussion in section 6.3.2.2, it is assumed that for coupling iterations
K<M both M KTU and M KBF have full column rank (with M and K denoting the
number of interface degrees of freedom and the current coupling iteration number,
respectively).
Given that for any X and Y matrices:
( ) min ( ), ( )Rank Rank RankX.Y X Y (6.114)
and considering Equations (6.112) and (6.113), it can be easily shown that:
Chapter 6 Reduced Order Method
228
CT M M
Rank K K
(6.115)
1CB M M
Rank K K
(6.116)
As previously mentioned, the condition for non-singularity of CTK and
1CBK
requires these square matrices to have a full rank (M). Since K<M, it is clear from
Equations (6.115) and (6.116) that both CTK and
1CBK
approximated by reduced
order method are singular matrices in nature. This is not a problematic issue in the
view of updating the boundary conditions in successive iterations using
Equation (6.37), since there is no need for inverting CTK and
1CBK
. In fact, the
reason for constructing the flexibility condensed interface matrix of the partitioned
sub-domain B directly by the reduced order method in Section 6.3.2 is due to the
singularity of the reduced order method approximation of CBK and not for
computational efficiency. In this context, it is very important to note that in coupling
procedures via reduced order method using any of Equations (6.21) to (6.24), where
an inverse of condensed stiffness matrices is required, the corresponding flexibility
matrix should be constructed instead.
6.3.2.4 Example 1: Static FEMFEM Coupling
Here the static FEM-FEM coupled problem of 6.3.1.1 is also analysed using the
presented reduced order scheme with and without the selective addition and
replacement procedure. A comparison between the number of coupling iterations
using adaptive relaxation and reduced order method is also provided in Table 6.4.
These results clearly show that the reduced order method is far superior to the
adaptive relaxation scheme, since it enhances the convergence rate of the coupling
method significantly. This high convergence rate is further demonstrated in Figures
6.12 to 6.17 for models A1 to A6 respectively.
Chapter 6 Reduced Order Method
229
Furthermore, it is shown in Figures 6.12 and 6.13 that without employing the
proposed selective addition/replacement procedure of section 6.3.2.2, if the number
of required coupling iterations for convergence to a prescribed tolerance exceeds the
number of independent modes, the iterative scheme will diverge. This fact highlights
the significance of employing the selective addition/replacement procedure proposed
in Section 6.3.2.2.
Model T
B
E
E
Reduced Order Method
with Add/Rep
Reduced Order Method
without Add/Rep
Adaptive
Relaxation
A1 8.0 5 Not Converged 16
A2 4.0 5 Not Converged 13
A3 2.0 5 7 10
A4 1.0 5 6 6
A5 0.5 5 5 6
A6 0.2 4 4 4
Table 6.4: Required coupling iterations for different coupling schemes
Chapter 6 Reduced Order Method
230
Figure 6.12: Error reduction for Model A1
Figure 6.13: Error reduction for Model A2
Chapter 6 Reduced Order Method
231
Figure 6.14: Error reduction for Model A3
Figure 6.15: Error reduction for Model A4
Chapter 6 Reduced Order Method
232
Figure 6.16: Error reduction for Model A5
Figure 6.17: Error reduction for Model A6
Chapter 6 Reduced Order Method
233
6.3.2.5 Example 2: Dynamic FEMFEM Coupling
The dynamic FEM-FEM coupling example presented in Section 6.3.1.2 is
considered here using the reduced order method, and the obtained results are again
compared with the adaptive relaxation. Two slightly different types of adaptive
relaxation and reduced order method are considered here, depending on the first
coupling iteration update procedure.
In the first set of methods, named as ‘Adaptive’ and ‘Reduced’ in Table 6.5, the
relaxation parameter is initialised at the start of every load/time step to a constant
value, 1.0 , and thereafter the adaptive relaxation or reduced order method is
applied in the normal way.
On the other hand, the second set of coupling methods, named as ‘Adaptive*’ and
‘Reduced*’ in Table 6.5, continue with the parameter or condensed matrices from
the previous step. For the ‘Adaptive*’ approach, the starting relaxation parameter is
considered to be equal to the last computed adaptive relaxation parameter of the
previous time/load step. Similarly in the case of ‘Reduced*’, the last approximation
of the condensed stiffness/flexibility matrices obtained in previous time/load steps is
used instead of a constant relaxation parameter for the update of boundary conditions
in the first iteration.
The dynamic analysis problem is considered for the different problem partitioning
types. This is achieved by analysing the same system with different mass ratios at the
interface of the partitioned sub-domains T and B ( 1 2/m m ). The different
analysed models are presented in Table 6.5 for a 5s duration of response with
0.01t s using different coupling techniques.
The presented results in Table 6.5 again show that the reduced order method
possesses a higher convergence rate than adaptive relaxation. Moreover, it is shown
that the performance of ‘Adaptive*’/‘Reduced*’ techniques are far better than their
corresponding ‘Adaptive’/‘Reduced’ counterparts (see Figure 6.23). However, it
should be noted that, unlike the problem under consideration, in problems where
Chapter 6 Reduced Order Method
234
there is a considerable or abrupt change in the state of problem (significant stiffness
change or high nonlinearity), it is not guaranteed that ‘Adaptive*’ and ‘Reduced*’
techniques outperform ‘Adaptive’ and ‘Reduced’ coupling methods.
The higher convergence rate of the reduced order method compared to adaptive
relaxation is further demonstrated in Figures 6.18, to 6.22, where the error reduction
of the two schemes, Adaptive*’/‘Reduced*, are compared in different arbitrarily
chosen time steps.
In all the analyses, the procedure introduced in Section 6.3.2.2 for selective
addition/replacement of the displacement and force vectors is utilised. As discussed
earlier, if such a procedure is not employed, any new obtained displacement/force
modes undoubtedly will be a linear combination of the previous independent
displacement/force modes, since the required number of coupling iterations exceeds
the number of independent modes. As a result, the approximation for the condensed
tangent stiffness matrix at the interface will be poor, if numerically possible due to
round-off errors, thus leading to a very low convergence rate and even divergence.
This shortcoming of the conventional reduced order method is illustrated in Figure
6.23, where it is shown that without employing the selective addition/replacement
procedure the coupling scheme starts to diverge after the third coupling iteration.
Chapter 6 Reduced Order Method
235
Model 1
2
m
m
Number of Coupling Iterations
Adaptive Reduced
Method Adaptive*
Reduced
Method*
C1 200 kg
0. 11800 kg
2133 1611 1624 744
C2 400 kg
0.251600 kg
1706 1279 1027 730
C3 600 kg
0.431400 kg
1432 1224 879 696
C4 800 kg
0.61200 kg
1291 1183 824 676
C5 1000 kg
1.01000 kg
1177 1170 758 702
C6 1200 kg
1.5800 kg
1149 1147 741 671
C7 1400 kg
2.3600 kg
1118 1121 752 668
C8 1600 kg
4.0400 kg
1056 1062 801 666
C9 1800 kg
9.0200 kg
972 975 821 644
*using the relaxation parameter/condensed interface stiffness matrix of the last time-
step in which convergence was achieved for the first coupling iteration of the current
time-step
Table 6.5: Number of required coupling iterations for 500 time-steps ( 0.01t s )
with a tolerance of 1e-4 m
Chapter 6 Reduced Order Method
236
Figure 6.18: Error reduction of different schemes (Time=2.74s) for Model C1
Figure 6.19: Error reduction of different schemes (Time=3.63s) for Model C1
Chapter 6 Reduced Order Method
237
Figure 6.20: Error reduction of different schemes (Time=4.92s) for Model C1
Figure 6.21: Error reduction of different schemes (Time=4.42s) for Model C1
Chapter 6 Reduced Order Method
238
Figure 6.22: Error reduction of different schemes (Time=1.19s) for Model C1
Figure 6.23: Error reduction of different scheme (Time=2.74s) for Model C1
Chapter 6 Reduced Order Method
239
6.3.2.6 Example 3: Linear SoilStructure Interaction
Consideration is given here to a linear static plane strain analysis problem of a
concrete cantilever wall resting on a flexible soil, which is loaded at the top with a
horizontal force (Figure 6.24). First, the problem is analysed using both the adaptive
relaxation and reduced order methods, while the stiffness of the foundation beam is
partitioned into two and is partly modelled in the structure sub-domain and partly in
soil sub-domain. This would increase the convergence rate of the iterative coupling
schemes, since by adding rigidity to the interface of soil the number of Dirichlet
degrees of freedom at the interface is effectively reduced during coupling iterations.
Using partitioned analysis, both the adaptive relaxation scheme and the reduced
order method show similar convergence characteristics as illustrated in Figure 6.25
Figure 6.24: Coupled soil-structure interaction problem
Chapter 6 Reduced Order Method
240
Figure 6.25: Error reduction for different coupling schemes
On the other hand, with the same problem modelled such that the foundation beam is
only modelled in the partitioned structure sub-domain, the convergence
characteristics become poor, as expected and as shown in Figure 6.26. However, it is
clear that the reduced order method with the selective addition/replacement
outperforms adaptive relaxation by far. Moreover, without employing the selective
addition/replacement procedure, the reduced order coupling scheme starts to diverge
after the 10th coupling iteration (noting that there are 10 degrees of freedom at the
soil-structure interface).
According to Equations (6.60) and (6.67), the reduced order method converges when
a good approximation of the condensed tangent stiffness is achieved. Considering
Figure 6.26, it can be seen that this is achieved through eleven cycles of
displacement/force history data. By using the reduced order method, it is guaranteed
that reasonable approximations of the condensed interface stiffness matrices can be
constructed during coupling iterations, and indeed the more displacement/force
Chapter 6 Reduced Order Method
241
history data (i.e. the more coupling iterations) the better the approximation of the
tangent stiffness matrix.
Figure 6.26: Error reduction for different coupling schemes
It is however worth observing that the required number of coupling iterations for
convergence using the reduced order method is still relatively large in comparison
with what would be necessary in a typical monolithic treatment. Furthermore,
considering the convergence behaviour of the reduced order method, it is clear that
as a result of having a poor approximation of condensed interface stiffness matrices
at the first coupling iterations, the compatibility errors are very big. This could be a
problematic issue in highly nonlinear problems where the high compatibility error of
the starting coupling iterations could cause divergence in the nonlinear solution
procedure. Although in problems such as the one presented this could be avoided by
reallocating the interface stiffness between the two sub-domains, in problems where
such partitioning strategy cannot be employed (such as modelling of a retaining wall)
there is a need for a more robust coupling technique with improved convergence
Chapter 6 Reduced Order Method
242
characteristics. Indeed, if a better approximation of the condensed tangent stiffness
matrix could be achieved, the coupling algorithm becomes even more robust, thus
highlighting the potential for further enhancement.
Towards this end, a mixed reduced order method is proposed in the next section. The
approach is based on approximating initial linear condensed stiffness and flexibility
matrices, which are then constantly updated using the iterative displacement and
force modes obtained during successive coupling iterations.
6.3.3 Mixed Reduced Order Method
A further enhancement of the reduced order method is proposed here, which utilises
an initial linear condensed stiffness matrix of the partitioned soil and structure sub-
domains at the beginning of soil-structure interaction coupling procedure. During
subsequent coupling iterations, these initial approximated matrices are continuously
updated based on the iterative history of the interface displacement/force increments,
thus achieving improved approximation in nonlinear analysis. The general
formulation for establishing the initial stiffness matrix is similar to the procedure
presented in Section 6.3.1. In the following, these initial stiffness matrices are
obtained via a reduced order method formulation, which gives identical initial
quantities.
Considering the structure partitioned sub-domain and assuming that there are N
degrees of freedom at the interface of the soil-structure interaction problem, the
following N+1 displacement vectors are prescribed at the interface of the structure
sub-domain:
1,..., ,..., , 0,...,Ti
T i ij iNiU u u u i N (6.117)
where:
0 or 0
and 1iju
i j iu
i j i
(6.118)
Chapter 6 Reduced Order Method
243
and εu is a predefined displacement value in the required range of response.
By obtaining the corresponding response of the structure sub-domain to the above
prescribed displacement vectors as:
1,..., ,..., , 0,...,Ti
T i ij iNiF f f f i N (6.119)
the initial condensed stiffness matrix of the structure sub-domain can be formulated
as:
1
0.
T TCTK
T T T TF U U U (6.120)
where:
1
,..., , 0,...,i i i iT T T T Ti N N N
U U U U i N
U (6.121)
1
,..., , 0,...,i i i iT T T T Ti N N N
F F F F i N
F (6.122)
The condensed tangent stiffness matrix of the structure sub-domain at iteration I is
obtained using a correction to the initial stiffness matrix:
0
C C CT T TI
K K K
(6.123)
such that:
I J I JC i i i iT T T T TI n n n n
K U U F F (6.124)
where I is iteration number and J=0,…,I-1.
Taking:
,I J I Ji i iT T TU U U (6.125)
Chapter 6 Reduced Order Method
244
,I J I Ji i iT T TF F F (6.126)
Equation (6.124) leads to:
, , ,
0
I J I J I JC i i C iT T T T TK U F K U
(6.127)
By constructing the following incremental displacement and force vectors:
0, , 1,, , , ,
I J I I Ii i iT T T TU U U
U (6.128)
0, , 1,, , , ,
I J I I I
T T T TF F F
F (6.129)
with:
, , ,
0
J I J I J Ii C iT T T TF F K U (6.130)
the corrective stiffness matrix of Equation (6.123) can now be obtained in the normal
way from the reduced order method according to the following:
1 1I IC iT T TF K U
(6.131)
Assuming that the displacement/load increments at iteration I+1 can be
approximated as a linear projection of the calculated displacement/load vectors of
Equations (6.128) and (6.129):
1IiT TU
U δ (6.132)
1I
T TF F δ (6.133)
With:
0 1, , , ,T
J I δ (6.134)
Chapter 6 Reduced Order Method
245
and solving Equation (6.132) for δ through minimisation of the error norm, the
following can be obtained from Equation (6.133):
11 1I IT T i
T T T T T TF U F U U U (6.135)
Comparing Equation (6.135) with Equation (6.131), the updated condensed stiffness
matrix of the partitioned structure sub-domain given by Equation (6.123) takes the
following form:
1
0
T TC CT T T T T TI
K K
F U U U (6.136)
In a similar manner, by prescribing the following N+1 force vectors given by
Equation (6.137) at the interface of the soil sub-domain, and obtaining the
corresponding response of the soil sub-domain to the above prescribed force vectors:
1,..., ,..., , 0,...,Ti
B i ij iNiF f f f i N (6.137)
where:
0 or 0
and 1ijf
i j if
i j i
(6.138)
and εf is a predefined incremental force value in the required range of response.
1,..., ,..., , 0,...,Ti
B i ij iNiF f f f i N (6.139)
the initial condensed flexibility matrix of the soil partitioned sub-domain can be
formulated as:
11
0.
T TCBK
B B B BU F F F (6.140)
Now the condensed tangent flexibility matrix of the soil sub-domain at iteration I can
be approximated by using a correction to the initial flexibility matrix:
Chapter 6 Reduced Order Method
246
1 1 1
0
C C CB B BI
K K K
(6.141)
such that:
I J I JC i i i iB B B B BI n n n n
K F F U U (6.142)
where I is iteration number and J=0,…,I-1.
Taking:
,I J I Ji i iB B BU U U (6.143)
,I J I Ji i iB B BF F F (6.144)
Equation (6.142) leads to:
, , ,1 1
0
I J I J I JC i i C iB B B B BK F U K F
(6.145)
By constructing the following incremental force and displacement vectors during
coupling iterations:
0, , 1,, , , ,
I J I I Ii i iB B B BF F F
F (6.146)
0, , 1,, , , ,
I J I I I
B B B BU U U
U (6.147)
with:
, , ,
0
J I J I J Ii C iB B B BU U K F (6.148)
the updated condensed flexibility matrix of the partitioned soil sub-domain can be
obtained as:
Chapter 6 Reduced Order Method
247
11 1
0
T TC CB B B B B BI
K K U F F F (6.149)
The algorithmic framework of the presented modified reduced order method is
illustrated in Table 6.6.
Using the above presented mixed reduced order method coupling technique, the
same problem of Section 6.3.2.6 is considered with the foundation beam completely
modelled in the structure sub-domain, which is more onerous than the case where the
foundation beam is apportioned between the two sub-domains. Figure 6.27,
compares the convergence rate of the mixed reduced order method with the
conventional reduced order method, highlighting its potential superiority in iterative
coupling of FEM-FEM coupled problems.
Figure 6.27: Error reduction for different coupling schemes
Chapter 6 Reduced Order Method
248
STEP 0: Calculate the initial linear stiffness/flexibility matrices at the beginning of
soil-structure coupling: 1
0 0,C C
T BK K
For n=1,2,…(number of load/time increments)
For I=0,1,…(number of iterations)
STEP 1: At the start of each increment, the structure domain is loaded by the external forces, while the displacements at the interface nodes are prescribed in
accordance with the initial conditions: IiT n
U U
STEP 2: The structural solver computes the response of the structure for : IiT n
F
IF I ≥1 then :
i. Selective addition/replacement of displacement/force vectors ii. Construct TU and T
F
iii. 1
0
T TC CT T T T T TI
K K
F U U U
STEP 3: The corresponding interface forces at the soil domain can be calculated by
applying equilibrium: 0I Ii i
T Bn nF F
STEP 4: Based on these forces and the external loading applied to the soil domain,
the soil solver computes the response of the soil domain for: IiB n
U
IF I ≥1 then :
i. Selective addition/replacement of displacement/force vectors ii. Construct BF and B
U
iii. 11 1
0
T TC CB B B B B BI
K K U F F F
STEP 5: If convergence to compatibility has been achieved the solution proceeds to the next time/load increment (n=n+1) with:
1 1
0 0,C C C C
T T B BI IK K K K
STEP 6: If convergence to compatibility has not been achieved, the following new estimation of the displacements will be applied to the structure sub-domain and the iteration will continue (I=I+1) until convergence:
11 1I I I Ii i C C i iT T B T B TI In n n n
U U K K U U
I
Table 6.6: Mixed reduced order method coupling procedures
Chapter 6 Reduced Order Method
249
6.4 Case Study: Nonlinear Soil-Structure Interaction Problem
In this section the presented nonlinear soil-structure interaction problem in Chapter 5
(Section 5.4), which was analysed using the relaxation scheme, is modelled and
analysed using the various reduced order formulations proposed in this chapter. The
example considers a steel frame resting on a soil under static loads, where nonlinear
elasto-plastic constitutive behaviour of the soil as well as geometric and material
nonlinearity of the structure are taken into account. The problem under consideration
is detailed in Figure 6.28 and Table 6.7.
Figure 6.28: Plane frame resting on soil
The building is designed for office purposes and is assumed to be loaded equally on
each floor. The footing is partly modelled in soil sub-domain and partially in the
structure sub-domain and the interface degrees of freedom are assumed to be at
nodes that belong to both of the partially modelled footings with a total number of
degrees of freedom equal to 30. The soil-structure interaction analysis is carried out
assuming plane strain conditions, where the developed domain decomposition
approach is employed utilising ADAPTIC and ICFEP.
Chapter 6 Reduced Order Method
250
Structure Sub-domain Material Properties
All beams and columns (steel)
Steel Grade = S355
Elastic Modules = 210 GPa
Strength = 355 MPa
Bilinear elasto-plastic with strain Hardening Factor = 1%
Foundation Beam (concrete)
Elastic Modulus = 30 GPa
Linear material
Soil Sub-domain Material Properties
Soil Angle of Shear resistance ( 𝛷′) = 22°
Dilation angle (𝜈 ) = 11°
Cohesion = 20 kPa
Young’s modulus varies linearly with depth from 10000 kPa at the ground surface (dE/dZ=5000 kPa/m)
Elasto-plastic Mohr-Coulomb constitutive model
Table 6.7: Geometric and material properties of the partitioned soil-structure system
Table 6.8 presents the required number of coupling iterations for convergence to the
prescribed tolerance for the reduced order method, mixed reduced order method and
adaptive relaxation scheme. Figure 6.29 shows the number of coupling iterations
required in each increment for the different coupling schemes. The convergence rates
of the three approaches in the first, fifth and sixth load increments are also illustrated
in Figures 6.30 to 6.32, respectively. This demonstrates that the mixed reduced order
method achieves a faster convergence rate than other coupling schemes.
Number of iterations
with adaptive relaxation
Number of iterations
with reduced order
Number of iterations
with mixed reduced order
46 31 22
Table 6.8: Number of required coupling iterations for different coupling schemes
Chapter 6 Reduced Order Method
251
Figure 6.29: convergence behaviour over full range of response for different schemes
Figure 6.30: Error reduction in the first load step
Chapter 6 Reduced Order Method
252
Figure 6.31: Error reduction in the 5th load step
Figure 6.32: Error reduction in the 6th load step
Chapter 6 Reduced Order Method
253
6.5 Conclusion
In this Chapter various domain decomposition methods for nonlinear analysis of
soil-structure interaction problems, based on approximating the condensed interface
stiffness matrix, are proposed. The overall domain is divided into physical partitions
consisting of soil and structure sub-domains, and coupling of the separately
modelled sub-domains is undertaken via sequential iterative Dirichlet-Neumann sub-
structuring method.
It is shown that by using the condensed tangent interface stiffness matrices of the
partitioned sub-domains in the update of boundary conditions, superior convergence
characteristics could be achieved. In this respect, convergence to compatibility
occurs for linear problems immediately at the first iteration. Although this immediate
convergence does not normally occur for nonlinear problems, employing this method
for the update of interface Dirichlet boundary conditions ensures a high convergence
rate. This brings the performance of the proposed coupling approach close to the
monolithic treatment.
Although the condensed tangent stiffness may be readily available with some
nonlinear field modelling tools, a more general approximation for the condensed
stiffness matrices is desirable and can indeed be achieved during the course of
coupling iterations, thus avoiding the need for explicit determination and extraction
of the stiffness matrices. In this respect, the condensed tangent stiffness matrix is
approximated via reduced order models, building on a previous approach by
Vierendeels et al. (2007) concerned with fluid–structure interaction problems.
Various significant modifications to this approach are proposed here, leading to a
versatile and efficient approach for coupled modelling of nonlinear soil-structure
interaction problems. In this regard, a major pitfall associated with the original
method causing divergence is overcome by a new selective addition/replacement
procedure of force and displacement mode vectors, where the applicability and
advantages of this modification are demonstrated by means of several examples.
Moreover, different to the method proposed by Vierendeels et al. (2007), the reduced
Chapter 6 Reduced Order Method
254
order method proposed here is solved only once to obtain the prescribed Dirichlet
data to structure sub-domain, while the prescribed Neumann data to the soil sub-
domain is obtained by enforcing equilibrium at all iterations.
In order to achieve better approximation of the condensed tangent stiffness matrix in
the initial stages of coupling iterations, a mixed reduced order method is proposed,
which achieves a more robust coupling technique than the conventional reduced
order method,. The approach is based on approximating an initial linear condensed
stiffness and flexibility matrix, which is then continuously updated using the
iterative displacement and force modes during successive coupling iterations.
The applicability of the presented coupling techniques is demonstrated for nonlinear
soil-structure interaction analysis via a case study consisting of a plane frame
supported on soil foundations. In this context, it is shown that the mixed reduced
order method achieves a faster convergence rate than the other coupling schemes,
demonstrating a great potential towards providing an integrated interdisciplinary
computational approach for nonlinear soil-structure interaction problems.
Chapter 7
Case Studies
7.1 Introduction
In order to illustrate the application of the proposed coupling techniques in soil-
structure interaction analysis and provide a better grasp of the involved concepts,
several illustrative static nonlinear soil-structure interaction problems are presented
and discussed in this chapter.
The problems under consideration are treated by partitioned analysis, where the
coupling is carried out through coupling of discipline-oriented solvers, ADAPTIC
and ICFEP, for nonlinear structural and geotechnical analysis as outlined in
Chapter 4.
Using the developed simulation environment for a number of problems in which
nonlinearity arises in both the structure and the soil, the advantages of the proposed
partitioned analysis are demonstrated. Moreover, a comparison between various
proposed coupling algorithms is performed in some cases to further highlight their
relative performance characteristics.
Chapter 7 Case Studies
256
7.2 Nonlinear Behaviour of Pitched-Roof Frame on Flexible Soil
In this example, the in-plane nonlinear behaviour of the unbraced single-bay pitched-
roof steel frame of Figure 7.1, resting on a flexible soil is modelled. The effect of
soil-structure interaction at the foundation level is taken into account using the
partitioned treatment. The effect of the soil-structure interaction on the overall
response is demonstrated by comparing the results obtained by the partitioned
treatment to those of a non-interactive case (i.e. rigid soil base).
Figure 7.1: Pitched-roof steel frame resting on soil
The considered steel frame (Figure 7.1) utilises the following material characteristics
and member cross-sections: E= 210 GPa (Young’s modulus), fy=350 MPa (yield
strength), HEB280 columns, and HEB240 rafters. The frame is modelled with
ADAPTIC using elasto-plastic cubic beam-column elements (Izzuddin & Elnashai,
1993), which enable the modelling of geometric and material nonlinearity, while the
material behaviour is assumed to be bilinear elasto-plastic with kinematic strain
hardening of 1%. The soil sub-domain is modelled with ICFEP using an associated
elasto-plastic Mohr-Coulomb constitutive model and discretised using 8-noded
isoparametric quadrilateral elements. Young’s Modulus E varies with depth z
Chapter 7 Case Studies
257
according to (E = 10,000 + 5,000z kN/m2), Poisson’s ratio μ=0.2, the bulk unit
weight γ = 20 kN/m3, the angle of shearing resistance, φ′=20° and cohesion c′=20
kPa.
The interface degrees of freedom are assumed to be at nodes that belong to the
footings with a height of 0.5 m and an effective out of plane depth of 1m. There are
20 degrees of interface freedoms in total. The loading system consists of a constant
horizontal force of H=22 kN and a varying vertical load of magnitude 0.25w ,
which are modelled in the structure partitioned sub-domain and applied in 34 load
increments.
The results from the partitioned treatment are presented in Figures 7.2 and 7.3, which
depict the nonlinear variation of the vertical settlements under footings A and E of
the steel frame with respect to the vertical load factor. Although the generated level
of the vertical settlements is relatively small, different settlements of the left end and
right end of each foundation generate significant rigid foundation tilting (Figures 7.2,
7.3). This can provide different equilibrium paths and change the force distribution in
the structure compared to non-interactive case. As a result the response of the
interactive case is significantly different compared to the non-interactive case, for
which the rotation of the base is not taken into account. This is demonstrated in
Figures 7.4 and 7.5, showing the horizontal displacement of nodes B and D of the
steel frame with respect to the load factor, for both interactive and non-interactive
cases, respectively. The horizontal and vertical displacements of the roof top of the
frame (node C) is also depicted in Figures 7.6 and 7.7, showing the significant
difference in the response of the interactive case compared to the non-interactive
case. It is worth noting that due to the small indeterminacy degree of the steel frame
and high distance between the foundations, the horizontal and vertical reactions at
the foundations in both interactive and non-interactive case are almost similar.
However, due to the rotation of the foundations the generated moment at the base as
shown in Figures 7.8 and 7.9 varies significantly. The variation of the moment at
node C with respect to the load factor for both interactive and non-interactive case is
also depicted in Figure 7.10.
Chapter 7 Case Studies
258
Figure 7.2: Vertical displacement of the footing A
Figure 7.3: Vertical displacement of the footing E
Chapter 7 Case Studies
259
Figure 7.4: Horizontal displacement of node B
Figure 7.5: Horizontal displacement of node D
Chapter 7 Case Studies
260
Figure 7.6: Horizontal displacement of node C
Figure 7.7: Vertical displacement of node C
Chapter 7 Case Studies
261
Figure 7.8: Variation of moment at node A
Figure 7.9: Variation of moment at node E
Chapter 7 Case Studies
262
Figure 7.10: Variation of moment at node C
In Figure 7.11, the deformed shape and the bending moment contours of the
partitioned structure sub-domain in the final load step are shown and compared for
both interactive and non interactive case.
Finally, Figure 7.12 shows contours of stress level at the end of analysis under the
applied load on the structure (load step = 34). The stress level is the ratio, at the same
mean effective stress, of the current deviatoric stress to the deviatoric stress at
failure. It therefore varies from 0 to 1, where 1 indicates full plasticity and failure. It
is evident from Figure 7.12 that the applied loading conditions have mobilised a
plastic zone underneath the footings.
Chapter 7 Case Studies
263
Figure 7.11: Deformed shape (scale=5.0) and bending moment (kN-m) in final load
step for a) non-interactive case b) interactive case
Figure 7.12: Contours of stress level in soil sub-domain (at final increment)
Chapter 7 Case Studies
264
The spread of plasticity under both footings (Figure 7.12) shows the significant rigid
tilting of both footings A and E. This is further evident from the deformed shape of
the structure which shows the footings experiencing rigid tilting. As a consequence,
there is no plastic hinge formation at the base of right column, as depicted in
Figure 7.11, compared to the non-interactive case.
An alternative treatment to partitioned approach in problems such as the above,
where there is no significant interaction between the influence zones of the soil
beneath the foundations due to the significant distance between them, is to capture
this behaviour using a field elimination treatment. This involves modelling the soil
beneath the foundation sub-grade as a beam on Winkler foundation with a system of
discrete, mechanistic, uncoupled springs. This type of model is useful in engineering
practice due to its simplicity and ease of implementation in a general purpose finite
element platform. However, the nonlinear soil behaviour underneath the shallow
footings cannot be captured by linear spring elements. Moreover, even if nonlinear
springs were to be used, it would involve careful calibration of both the rotational
and horizontal nonlinear springs against field tests. On the other hand, by using a
partitioned treatment, as demonstrated above, a fully coupled nonlinear soil-structure
interaction analysis can be readily carried out, where both the behaviour of the
structure sub-domain and soil sub-domain are evaluated using proven field specific
analysis tools.
7.3 Settlement Analysis of Multi-storey Five-bay Steel Frame
Assessment of the potential construction and settlement-induced damage to building
structures is of significant importance in civil engineering practice (Boone, 1996;
Boone, 2001; Charles & Skinner, 2004), and this can require fully coupled models
for capturing the behaviour of both soil and structure. Not only do settlements
impinge on the performance of structure and soil, but also they are typically
considered in the design and damage assessment of non-structural elements, such as
infill panels.
Chapter 7 Case Studies
265
In this section, the response of a multi-storey five-bay steel frame (Figure 7.13)
resting on a flexible soil is modelled using the partitioned analysis. The benefits of
utilising a fully coupled soil-structure analysis via the partitioned approach are
demonstrated, and the results are compared to those from a field elimination
technique where the soil sub-domain is modelled with linear transitional and
rotational springs (Winkler foundation).
Figure 7.13: Multi-Storey five-bay steel frame
The considered soil-structure system is partitioned physically into two sub-domains,
soil and structure, where each sub-domain is discretised separately according to its
characteristics as shown in Table 7.1. The framed structure is modelled with
ADAPTIC using cubic elasto-plastic beam-column elements, which enable the
modelling of geometric and material nonlinearity. The frame is discretised using 10
elements per member for both columns and beams, and the material behaviour is
assumed to be bilinear elasto-plastic with kinematic strain hardening. The footings
are discretised using 4 elements per member. The soil sub-domain is modelled with
ICFEP using an elasto-plastic Mohr-Coulomb constitutive model, with parameters
chosen to represent the behaviour of London clay (Table 7.1). The loading system is
modelled in the structure partitioned sub-domain and applied in 10 load increments
(uniformly distributed load of 5kN/m applied on all beams). The interface degrees of
Chapter 7 Case Studies
266
freedom are assumed to be at nodes that belong to both the footings and soil
underneath. The total number of interface degrees of freedom is 60 for this case.
Structure Sub-domain Material Properties Columns Beams All beams and columns (steel)
UC 203×203×46 UB 305×102×25 Steel Grade = S355 Elastic Modules = 210 GPa Strength = 355 MPa Bilinear elasto-plastic with strain Hardening Factor = 1%
Foundation Beam (concrete)
Elastic Modulus = 30 GPa Linear material Size: 2m×0.5m
Soil Sub-domain Material Properties Soil Angle of Shear resistance ( 𝛷′) = 22° Dilation angle (𝜈 ) = 11° Effective out of plane depth = 1m Cohesion = 20 kPa
Young’s modulus varies linearly with depth from 10000 kPa at the ground surface (dE/dZ=5000 kPa/m)
Elasto-plastic Mohr-Coulomb constitutive model
Table 7.1: Geometric and material properties of the partitioned soil-structure system
To further assess the merits of the various coupling algorithms presented in this
work, the above problem is analysed using different update techniques for a
tolerance of 410 m, where the total number of required coupling iterations (for 10
increments) is listed in Table 7.2. As expected, the constant relaxation scheme has
the worst convergence rate, while the proposed mixed reduced order coupling
scheme has the highest convergence rate.
However, the mixed reduced order scheme requires a greater number of calls to soil
and structural solvers in this case, since it requires the determination of the initial
condensed stiffness matrix prior to the coupling iterations, which involves more
interface-related computations than the reduced order method. Hence, one might
consider the trade-off for different types of problems under consideration, where the
Chapter 7 Case Studies
267
proposed mixed reduced order scheme attains even greater superiority for problems
involving more load steps and/or greater nonlinearity.
Coupling method Number of required coupling iterations
Increase in convergence rate compared to optimum
relaxation Optimum constant relaxation 85 - Adaptive relaxation 64 25% Reduced order method 51 40% Mixed reduced order method 43 50%
Table 7.2: Comparison of different coupling methods
Notwithstanding, the benefits of the developed simulation environment in the
practical assessment of nonlinear soil-structure interaction problems can be
demonstrated by considering the results of this example.
The frame structure with its applied loads transmits the loading to the soil, which in
turn deforms due to cumulative action of these loads. As a result, the soil below the
footings goes under vertical and differential settlements. Consequently, the
deformation of soil surface beneath the foundation could cause significant
redistribution of the loads in the frame structure. Moreover, in cases where the
differential settlements are considerable this could cause significant damage to the
infill walls in the frame structure.
As mentioned before, a common approach in capturing this behaviour, benefitting
from simplicity and ease of implementation, is to model the foundation sub-grade as
a beam-on Winkler foundation with a system of discrete and uncoupled springs.
Clearly, however, this type of modelling requires careful calibration of the spring
elements using different experimental test results. Moreover, due to the uncoupled
nature of the spring elements beneath the foundations, the soil sub-domain is often
modelled with gross inaccuracy.
On the other hand, it is shown here that by using a fully coupled partitioned
approach, both the structure and soil behaviour could be effectively captured to the
Chapter 7 Case Studies
268
desired degree of accuracy. This is demonstrated in the following, where the
deformation and stress states of the coupled problem are briefly described.
Comparison is made between the fully coupled partitioned approach and the field
elimination approach, where linear springs with an approximated stiffness of 10,000
kPa are used in the latter approach to represent the soil domain.
The results of both approaches for the variation of the vertical settlement under
column C2 (where the maximum vertical settlement is observed) with respect to the
load factor are presented in Figure 7.14. Clearly, except for the first increment where
the response of both models is linear, the real settlement of the footing is expected to
be much higher (up to 500%) than predicted by the simple field elimination approach
when the nonlinearity of the soil is taken into account.
Figure 7.14 : Vertical displacement (m) of Column C2
The vertical deformation profile of the soil surface, for different load-steps, obtained
from coupled analysis is given in Figure 7.15, where it is clear that the six troughs
correspond to the locations of the footings showing their vertical settlement.
Considering the generated level of vertical settlement and the rigid tilting of the
Chapter 7 Case Studies
269
footings, it is evident that geometric nonlinearity should be considered in the
structural analysis.
Figure 7.15: Vertical displacement profile of the soil surface
A vector plot of displacements in the soil sub-domain in the vicinity of the structure
is also shown in Figure 7.16. The absolute magnitudes of these vectors are not
important, though their relative magnitude shows the mechanism of ground
deformation. Contours of stress level in the soil partitioned sub-domain for the final
load-step are also depicted in Figure 7.17, where it is evident that the applied loading
conditions have mobilised an extensive plastic zone underneath the building. This
zone is, however, smaller and shallower under the right hand side footing, which is
in agreement with the previous figures that show most of the deformation and load
concentration nearer the left side footings.
Chapter 7 Case Studies
270
Figure 7.16: Displacement vectors in soil partitioned sub-domain
Figure 7.17: Contours of stress level in soil partitioned sub-domain
In Figure 7.18, the variation of bending moment at the base of column C1 for
different load levels is shown and compared with that of the field elimination
technique. This shows the significant effect of taking into account the nonlinear
behaviour of soil sub-domain in soil-structure interaction analysis. It can be clearly
observed that the bending moment of the structural elements in the fully coupled
analysis is significantly higher than that of field elimination analysis. This fact is
further demonstrated in Figure 7.19, where the bending moment in the middle of
beam B1 obtained by partitioned approach is compared with that of field elimination
for different load-steps.
Chapter 7 Case Studies
271
Figure 7.18: Bending moment (kN-m) at the base of C1 for different load-steps
Figure 7.19: Bending moment at the middle of beam B1 for different load-steps
Chapter 7 Case Studies
272
The deformed shape and the bending moment contours of the partitioned structure
sub-domain at the final load step are shown in Figure 7.20b. The same quantities are
also obtained and presented in Figure 7.20a using the field elimination approach. It is
evident from the deformed shape of the structure in fully coupled interaction analysis
and from the vectors underneath each of the three footings in Figure 7.16 that these
experience rigid tilting and significant vertical settlements.
Figure 7.20: Deformed shape (scale=5.0) and bending moment (kN-m) in final load
step for a) linear Winkler foundation b) nonlinear partitioned analysis
Chapter 7 Case Studies
273
The significant support settlements predicted by the more realistic partitioned
approach lead to much greater bending moments in the beams and columns than
predicted by the field elimination approach. This is illustrated in Figures 7.21 and
7.22 in which the bending moment diagram of beam B1 and column C1 at the final
load-step are respectively shown.
Figure 7.21: Variation of bending moment along beam B1
Figure 7.22: Variation of bending moment along Column C1
Chapter 7 Case Studies
274
Finally, the deformed shapes of beam B1 at the final load-step are presented in
Figure 7.23 for both partitioned and field elimination approaches, where
significantly larger deflections are confirmed with the partitioned approach.
Figure 7.23: Variation of vertical displacement along the beam B1
7.4 Building Response to an Adjacent Excavation
Open cuts and excavations in a limited urban space are gradually increasing in
frequency because of the development and upgrade of infrastructures and the
construction of new buildings. At the same time, public concerns have risen over the
effects of excavation-induced ground movements on adjacent structures and utilities.
Excavation inevitably results in deformation of the adjacent ground and settlement of
adjacent buildings behind an excavation wall, causing problems such as loss of
invaluable historic property, third party impact, construction delay, and substantial
increase of project cost (Son et al., 2005; Aye et al., 2006; Boone et al., 1999; Seok
et al., 2001).
Chapter 7 Case Studies
275
The case study under consideration in this section represents a typical urban
situation, where ground excavation can often induce significant movements and
damage to the nearby structures. It is shown that by utilizing a fully coupled soil-
structure interaction model using the partitioned treatment, such nonlinear behaviour
of both structure and soil could be accurately captured. This shows the high potential
of using a fully coupled soil-structure model towards providing reliable assessment
and minimizing the associated damage in such problems. The example considers a
steel frame resting on a soil subjected to ground excavation, where nonlinear elasto-
plastic constitutive behaviour of the soil, as well as geometric and material
nonlinearity of the structure, are taken into account. Figure 7.24 depicts the problem,
where the left hand side boundary is assumed to be consistent with an axis of
symmetry. The plan view of the analysed building frame is also shown in Figure
7.25.
Figure 7.24: Plane frame resting on soil subject to ground excavation
Chapter 7 Case Studies
276
Figure 7.25: Plan view of considered building
Figure 7.26: Geometric configuration of considered frame
Chapter 7 Case Studies
277
The soil-structure interaction analysis is carried out assuming plane strain conditions
in the soil using an effective out-of-plane width of 1m, where the developed domain
decomposition approach is employed utilising ADAPTIC and ICFEP. The
considered soil-structure system is partitioned physically into two sub-domains, soil
and structure, where each sub-domain is discretised separately according to its
characteristics as shown in Table 7.3.
Structure Sub-domain Material Properties All beams and columns (steel)
Steel Grade = S355 Elastic Modules = 210 GPa Strength = 355 MPa Bilinear elasto-plastic with strain Hardening Factor =1%
Foundation Beam (concrete)
Elastic Modulus = 30 GPa Linear material Size: 2m×0.5m
Soil Sub-domain Material Properties Soil and excavation Angle of Shear resistance ( 𝛷′) = 22° Dilation angle (𝜈 ) = 11° Effective out of plane depth = 1m Cohesion = 20 kPa
Young’s modulus varies linearly with depth from 10000 kPa at the ground surface (dE/dZ=5000 kPa/m)
Excavation width=20m He (excavation depth) (=1, 2, 3, 4, 5, 6, 7, 8, 9, 10 (m) Le (distance from the structure =2, 4, 6, 8, 10 (m)
Elasto-plastic Mohr-Coulomb constitutive model
Table 7.3: Geometric and material properties of the partitioned soil-structure system
The frame structure is modelled with ADAPTIC using cubic elasto-plastic beam-
column elements (Izzuddin & Elnashai, 1993) using 10 elements per member for
both columns and beams, and the material behaviour is assumed to be bilinear elasto-
plastic with kinematic strain hardening. The footings are discretised using 4 elements
per member.
The soil sub-domain and the un-braced excavation are modelled with ICFEP using
an elasto-plastic Mohr-Coulomb constitutive model, with parameters chosen to
represent the behaviour of London clay (Table 7.3). The nonlinear solution
procedure employed for analysing the soil sub-domain is based on a Modified
Chapter 7 Case Studies
278
Newton-Raphson technique, with an error controlled sub-stepping stress point
algorithm. The soil continuum is discretised using 8-noded isoparametric
quadrilateral elements. The interface degrees of freedom are assumed to be at nodes
that belong to both the footings and soil underneath. The total number of interface
degrees of freedom is 30 for this case.
The above problem is analysed for various scenarios with respect to the loading
applied to the structure (which is assumed to be loaded equally on each floor with a
total gravity load equal to λ×5 kN/m2), the excavation depth (He) and the distance of
the structure from the excavation wall (Le).
Table 7.4 lists various loading scenarios considered for analysing the above problem
with respect to the load factor (λ) applied in structure sub-domain, and the
excavation depth (He) in the soil sub-domain. Considering Table 7.4, the loads on
the structure are applied in the first six increments, and from increment 7 to 16 the
soil is excavated while the loading in the structure is assumed to be constant.
Model Case
Increment number of the coupled analysis (16 increments in total) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
values for λ (no excavation)
Values for He (m) (λ=constant)
Case 1 1 1 1 1 1 1 1 2 3 4 5 6 7 8 9 10Case 2 1 2 2 2 2 2 1 2 3 4 5 6 7 8 9 10Case 3 1 2 3 3 3 3 1 2 3 4 5 6 7 8 9 10Case 4 1 2 3 4 4 4 1 2 3 4 5 6 7 8 9 10Case 5 1 2 3 4 5 5 1 2 3 4 5 6 7 8 9 10Case 6 1 2 3 4 5 6 1 2 3 4 5 6 7 8 9 10
Table 7.4: Loading scenarios
Figures 7.27, 7.28 and 7.29, show the vertical settlement at the centre of the left,
middle and right footings of the analysed frame (see Figure 7.24), respectively, for
the various loading scenarios, where it is assumed Le=2m. The vertical displacement
of the ground surface is also depicted in Figure 7.30.
Chapter 7 Case Studies
279
Figure 7.27: Vertical settlement of the left footing for different load cases
Figure 7.28: Vertical settlement of the middle footing for different load cases
Chapter 7 Case Studies
280
Figure 7.29: Vertical settlement of the right footing for different load cases
Figure 7.30: Cumulative vertical displacement of the ground surface (last increment)
Chapter 7 Case Studies
281
A common practice in estimating the ground movement adjacent to an excavation is
to regard both the settlement of the ground and the building as identical to avoid
complications in geotechnical analysis. Considering the footing settlements depicted
for various cases in Figures 7.27, 7.28 and 7.29, it is clear from Figure 7.30 that
these are considerably underestimated by the free field response.
The effect of the structural loads on the horizontal displacement of the excavation
wall is also depicted in Figure 7.31. It is a consequence of these lateral ground
movements that the structure undergoes additional settlements due to its weight.
Figure 7.31: Cumulative horizontal displacement of excavation wall (last increment)
The effect of the excavation depth on the vertical deformation profile of the soil
surface and the horizontal displacement of the excavation wall is also depicted in
Figures 7.32 and 7.33, respectively, for Case 6 (Table 7.3).
Chapter 7 Case Studies
282
Figure 7.32: Vertical displacement of ground surface for different excavation depths
(Case 6)
Figure 7.33: Horizontal displacement of the excavation wall for different excavation
depths (Case 6)
Chapter 7 Case Studies
283
The effect of distance of the structure from the excavation (Le) on the vertical
settlements of the footings is also depicted in Figures 7.34, 7.35 and 7.36 for model
Case 6.
As expected by increasing Le the additional settlements of the structure due to the
excavation decreases. The results further emphasise the importance of using a fully
coupled soil-structure interaction analysis for such cases.
Figure 7.34: Vertical settlement of the left footing for different Le (Case 6)
Chapter 7 Case Studies
284
Figure 7.35: Vertical settlement of the right footing for different Le (Case 6)
Figure 7.36: Vertical settlement of the right footing for different Le (Case 6)
Chapter 7 Case Studies
285
All the above coupled analyses were carried out using reduced order method with a
tolerance of 10-4 m. However, to further assess the merits of different coupling
algorithms presented in this thesis, Case 6 is analysed using different update
techniques for the same tolerance, where the corresponding total number of required
coupling iterations is listed in Table 7.5. As expected, the constant relaxation
scheme has the worst convergence rate, while the proposed mixed reduced order
coupling scheme outperforms the other coupling schemes.
Coupling method Number of required coupling
iterations Constant relaxation 158 Adaptive relaxation 96
Reduced order method 88 Mixed reduced order method 67
Table 7.5: Comparison of different coupling methods
The benefits of the developed simulation environment in the practical assessment of
nonlinear soil-structure interaction problems is further demonstrated by considering
the results obtained for Case 6. In this regard, vector plots of displacements in the
soil sub-domain in the vicinity of the structure and excavation for increment number
6 (before excavation) and increment 12 (He=6m) are shown in Figures 7.37 and
7.39, respectively.
Figure 7.38 and 7.40 show contours of stress level before and after the full
excavation to 6m depth under the applied load on the structure. It is evident from
Figure 7.40 that the applied loading conditions have mobilised an extensive plastic
zone underneath the building. This zone is, however, smaller and shallower under
the right hand side footing, which is in agreement with the previous figures that
show most of the deformation and load concentration nearer the excavation.
Chapter 7 Case Studies
286
Figure 7.37: Vectors of displacement in soil sub-domain in increment 6 (Case 6)
Figure 7.38: Contour plots of stress levels and plasticity induced in soil sub-domain
in increment 6 (Case 6)
Chapter 7 Case Studies
287
Figure 7.39: Vectors of displacement in soil sub-domain in increment 12 (Case 6)
Figure 7.40: Contour plots of stress levels and plasticity induced in soil sub-domain
in increment 12 (Case 6)
Chapter 7 Case Studies
288
The deformed shape and bending moment variation in the structure sub-domain for
different increments are shown in Figure 7.41. It can be clearly observed that the
maximum bending moments of the structural elements after the excavation are
significantly higher than before the excavation. The comparison of the bending
moment values for four selected regions A, B, C and D, as shown in Figure 7.41d, is
presented for both the 6th and 12th load increments in Table 7.6. It is evident from the
deformed shape of the structure after excavation and also from the vectors
underneath each of the three footings (Figure 7.39), that after 6m of excavation the
footings experience rigid tilting and significant vertical settlements. However, the
footing nearest to the excavation has the smallest tilting, as its deformation is also
dominated by the horizontal movement towards the unsupported excavation.
Figure 7.41: Deformed shape (scale=5) and bending moment (kN-m) of structure for
(a) 1st, (b) 6rd , (c) 7th and (d) 12th increment
Chapter 7 Case Studies
289
RegionMz (kN-m)Increment 6
Mz (kN-m)Increment 12
% Increase
A 120 215 80% B -120 -191 60% C -21 -72 340% D 21 153 720%
Table 7.6: Comparison of maximum bending moment in fully coupled analysis
7.5 Coupled Modelling of Retaining Steel Sheet Piles
With the new Eurocodes coming into force, and their introduction of unified
procedures for limit state design of structural elements, plastic bending of steel sheet
piles will be allowed. According to Bourne-Webb et al. (2007), the limit state design
approach is well established for conventional building structures but has not
generally been applied to the design of earth-retaining structures — at least not in a
unified manner. One area of interest is the use of limit state principles in the design
of steel sheet pile retaining walls and, in particular, whether it is safe to allow the
formation of plastic hinges at the ultimate limit state and, if so, how to verify that the
behaviour of the wall zone undergoing plastic deformation is within acceptable
limits. In this section, the soil-structure interaction problem under consideration is a
2D simulation of a cantilever steel sheet pile retaining wall, as depicted in Figure
7.42, where the left hand side boundary of the problem is assumed to be consistent
with an axis of symmetry.
Following the partitioned treatment, the considered soil-structure system is
partitioned physically into two sub-domains, soil and structure, where each sub-
domain is discretised separately according to its characteristics. The wall is modelled
with ADAPTIC using cubic elasto-plastic 2D beam-column elements, which enable
the modelling of geometric and material nonlinearity, while the material behaviour is
assumed to be bilinear elasto-plastic with kinematic strain hardening. The steel sheet
pile retaining wall is modelled (I section) with a second moment of area I = 2346
cm4/m, plastic section modulus of Wpl = 410 cm3/m, and a yield stress of fy=270
MPa. The soil sub-domain is modelled with ICFEP using an associated elasto-plastic
Chapter 7 Case Studies
290
Mohr-Coulomb constitutive model and discretised using 8-noded isoparametric
quadrilateral elements with a Young’s Modulus E varying with depth z according to
(E = 4000 + 5000z kN/m2), Poisson’s ratio μ=0.2, the bulk unit weight γ = 20 kN/m3,
the angle of shearing resistance, φ′=20°. The retaining wall is assumed to be
embedded 2 m into the ground, and the retained surface is surcharged by a load W.
The problem is analysed in nine increments as shown in Table 7.7.
Figure 7.42: Schematic diagram of the cantilever retaining wall
Increment Number
1 2 3 4 5 6 7 8 9
He (m) W(kPa)
He=8 W=20
He=8 W=40
He=8 W=60
He=8 W=80
He=8 W=100
He=8 W=120
He=8 W=140
He=8 W=160
He=8 W=180
Table 7.7: Loading scenario in different incremental stages of the analysis
Chapter 7 Case Studies
291
In this example, employing the relaxation scheme is very expensive since the
relaxation parameter tends to very small values, in order of 10-3, to ensure
convergence, and as a result both constant and adaptive schemes become extremely
computationally inefficient. Moreover, by employing the standard reduced order
method in this case, appropriate approximation of the condensed stiffness matrix,
and thus convergence, can only be achieved when the coupling iterations exceed the
number of interface degrees of freedom which in this case is 42.
In contrast, by utilizing the proposed mixed reduced order method, superior
convergence characteristics are achieved. In fact, an average of 10 coupling
iterations per increment is sufficient to achieve convergence to a tolerance of 10-3 m
set for the compatibility error of each coupled degree of freedom at the interface.
Indeed, the mixed reduced order method presented in Chapter 4, outperforms other
coupling techniques in numerous problems where the interaction effects at the
interface become more pronounced and the number of interface degrees of freedoms
is large.
The applicability and benefits of the developed coupling technique in nonlinear soil-
structure interaction analysis can be established by considering the results of this
case study. These results show that the excavation has mobilized a significant plastic
deformation near the retaining wall.
A vector plot of displacements in the soil sub-domain in the vicinity of the
excavation is shown in Figure 7.43 for the final increment. The relative magnitude of
these vectors shows the mechanism of ground deformation.
Importantly in the final increment, the soil deformations due to excavation have
caused the retaining wall to reach its plastic moment as illustrated in Figure 7.44,
where the bending moment distribution and the deflected shape of the retaining wall
are depicted for the different incremental stages of the analysis given in Figure 7.44.
Chapter 7 Case Studies
292
Figure 7.43: Vectors of accumulated displacements in soil sub-domain (final
increment)
Figure 7.44: Deflection (scale: 5) and bending moment (kN-m/m) distribution of the
retaining wall for (a) 1st, (b) 4th, (c) 7th, (d) 8th and (e) 9th increment
Chapter 7 Case Studies
293
The same soil-structure interaction problem of Figure 7.42 is now simulated as a
propped steel sheet pile retaining wall, as depicted in Figure 7.45. Again, the
retaining wall is assumed to be embedded 2 m into ground and the retained surface is
surcharged by a load of W. The above problem is analysed in five increments as
given in Table 7.8.
Increment number
1 2 3 4 5
He (m) W(kPa)
He=8 W=20
He=8 W=40
He=8 W=60
He=8 W=80
He=8 W=100
Table 7.8: Loading scenario in different incremental stages of the analysis
Figure 7.45: Schematic diagram of the propped retaining wall
Similarly, by utilizing the mixed reduced order method, superior convergence
performance is achieved, and an average of 10 coupling iteration per increment is
sufficient to achieve convergence to the same tolerance of 10-3 m.
Chapter 7 Case Studies
294
The obtained results for the proposed wall show that the excavation has mobilized a
significant plastic deformation near the retaining wall. The vector plot of
displacements in the soil sub-domain in the vicinity of the excavation is presented in
Figure 7.46, which shows a completely different mechanism in the partitioned soil
sub-domain compared to the previous case.
Importantly in the final increment, the soil deformations due to excavation cause the
retaining wall to reach its plastic moment as illustrated in Figure 7.47, where the
bending moment distribution and the deflected shape of the retaining wall are
depicted for different incremental stages of the analysis outlined in Table 7.8.
Finally, the change in the horizontal force in the lateral support is also depicted in
Figure 7.48.
Figure 7.46: Vectors of accumulated displacements in soil sub-domain (final
increment)
Chapter 7 Case Studies
295
Figure 7.47: Deflection (scale: 5) and bending moment (kN-m/m) distribution of the
retaining wall for (a) 1st, (b) 2nd , (c) 3rd , (d) 4th and (e) 5th increment
Figure 7.48: Horizontal reaction at the lateral support
Chapter 7 Case Studies
296
7.6 Conclusion
The applicability of the partitioned treatment of nonlinear soil-structure interaction
analysis and its great potential towards capturing the nonlinear behaviour of both soil
and structure are illustrated in this chapter through various case studies. Towards this
end, iterative coupling is carried out using various existing and proposed procedures
using the discipline-oriented solvers, ADAPTIC and ICFEP, for nonlinear structural
and geotechnical analysis, respectively. It is shown that the proposed mixed reduced
order method provides superior convergence characteristics in comparison with
existing coupling methods.
Using the developed simulation environment in a number of case studies in which
nonlinearity arises in both structure and soil, the applicability and high potential of
the developed simulation environment is demonstrated towards providing an
integrated interdisciplinary computational approach which combines the advanced
features of both structural and geotechnical modelling for a variety of challenging
soil-structure interaction problems.
Chapter 8
Conclusion
8.1 Introduction
This work has been primarily motivated by the lack of sophisticated monolithic tools
for modelling nonlinear soil–structure interaction problems, while recognising the
existence of advanced tools for nonlinear analysis of structure and soil in isolation.
Although coupled modelling of soil-structure interaction problems may be achieved
using a monolithic treatment, the partitioned treatment has been advocated as
offering major benefits in the context of coupled modelling of nonlinear soil-
structure interaction.
Accordingly, the aim of this work has been to develop advanced numerical methods
for nonlinear coupling of static and dynamic soil-structure interaction problems,
where the partitioned approach is adopted as a framework for coupling field-specific
tools with minimal intrusion into codes.
The partitioned approach in soil-structure interaction analysis is fully investigated in
this work. In this respect, various coupling techniques in the context of soil-structure
interaction analysis are developed, and their computational characteristics are
Chapter 8 Conclusion
298
discussed. In this regard, novel formulations for coupling soil-structure systems,
based on relaxation coupling methods and also utilizing the tangent stiffness matrix
of the partitioned sub-domains at the interface are proposed, and their relative
performance is evaluated.
Based on the presented coupling algorithms, a novel simulation environment,
utilising discipline-oriented solvers for nonlinear structural and geotechnical
analysis, has been developed, which couples two advanced nonlinear finite element
solvers, ADAPTIC (Izzuddin, 1991) and ICFEP (Potts & Zdravkovic, 1999), for
structural and geotechnical analysis respectively. Although the developed methods
are applied to the coupling of ADAPTIC and ICFEP, they are also generally
applicable to the coupling of other existing nonlinear soil and structural software.
In this respect, the developed approach is believed to offer great potential towards
providing an integrated interdisciplinary computational framework for coupled
modelling of soil-structure interaction problems.
The developed simulation environment is used in this work to demonstrate the
performance characteristics and merits of the various presented algorithms.
Accordingly, the developed tool is employed for a number of numerical examples
involving nonlinear soil-structure interaction analysis in which nonlinearity arises in
both the structure and the soil, leading to important conclusions regarding the
adequacy and applicability of the alternative coupling methods as well as the
prospects for further enhancements.
8.2 Conclusions
This work has shown that the partitioned treatment is a feasible and realistic
approach for coupled modelling of nonlinear soil-structure interaction problems.
Unlike the monolithic approach, the partitioned approach offers major benefits,
including i) allowing field-specific discretisation and solution procedures that have
proven performance for each partitioned domain, and ii) facilitating the reuse of
Chapter 8 Conclusion
299
existing nonlinear analysis software with all the resource savings that this brings.
The success of the partitioned approach, however, has been shown to hinge on the
adopted coupling algorithm. Indeed, a principal contribution of this work has been
the investigation of the performance of existing coupling algorithms, and the
development of more powerful algorithms that address convergence issues,
particularly in the context of nonlinear soil-structure interaction problems. Towards
this end, this work has successfully developed novel iterative coupling approaches,
which are based on the reduced order method, yet offering major improvements of
the convergence characteristics and computational performance. Furthermore, the
applicability of the developed partitioned approaches to soil-structure interaction
problems, exhibiting significant nonlinearity in both structure and soil, has been
demonstrated by means of several case studies.
Hereafter, the main conclusions from various parts of this work are summarised.
8.2.1 Staggered Approach
It has been shown that coupling of partitioned sub-domains may be achieved using
the staggered approach, though this approach should be used with great care in
relation to both stability and accuracy. The main conclusions in this regard are:
The staggered approach is applicable only to transient dynamic problems.
The stability of the staggered scheme is conditional on the time step and also
on the equivalent mass and stiffness on both sides of the interface, though
this depends on the formulation of the partitioned sub-domains and the
employed predictors.
Stability and accuracy considerations typically demand excessively small
time steps rendering this scheme computationally prohibitive for typical
coupled problems.
In general, achieving stability in a staggered solution procedure is extremely difficult
and in many cases impossible without reformulation of the field equations of the
Chapter 8 Conclusion
300
original partitioned sub-domains. Although some further progress may be achieved
by modifying the specifics of the staggered scheme, these are not considered here,
since the ultimate performance will vary with the application problem under
consideration, and often their implementation would conflict with the modular use of
the structural and soil solvers as black box solvers.
8.2.2 Iterative Coupling
Iterative coupling methods have been shown to offer major enhancement over the
staggered approach in the context of partitioned analysis of soil-structure interaction
problems. Considering the applicability requirements for both dynamic and static
analysis, the algorithms that are considered to be more suitable for soil-structure
interaction coupling are:
Sequential/Parallel Dirichlet-Neumann.
Parallel Dirichlet-Dirichlet.
Particular emphasis has been placed in this work on Dirichlet-Neumann (D-N) type
of iterative schemes, specifically the sequential D-N algorithms.
Within the family of sequential D-N algorithms, the convergence performance over a
load/time step has been identified as the most important feature. In this respect, it has
been demonstrated that:
Convergence is directly dictated by the chosen update technique during
successive iterations.
By employing trivial update of boundary conditions, only conditional
convergence could be achieved, which is unsatisfactory.
Based on a generalised sequential Dirichlet-Neumann iterative coupling algorithm, a
novel simulation has been developed which utilises the discipline-oriented solvers,
ADAPTIC and ICFEP, for nonlinear structural and geotechnical analysis. The
software architecture of the developed simulation environment has been outlined,
Chapter 8 Conclusion
301
where the structure of the data exchange between the various codes is also
elaborated.
8.2.3 Interface Relaxation
The use of interface relaxation update techniques in FEM-FEM domain
decomposition analysis of soil-structure interaction has been studied, where coupling
of the separately modelled sub-domains is undertaken on the sequential iterative D-N
sub-structuring method. In this respect, the convergence characteristics of iterative
coupling algorithms are enhanced by employing a relaxation of the interface
Dirichlet entities in successive iterations.
Due to the lack of general convergence analysis for the relaxation update technique
in FEM-FEM coupling, various mathematical and computational characteristics of
the coupling method, including the governing convergence rate and choice of
constant relaxation parameter, have been established. The work undertaken has
shown that:
Convergence to compatibility in sequential D-N iterative coupling algorithms
can be ensured by employing an appropriate constant relaxation parameter.
There exists a range of relaxation parameters that guarantee convergence to
compatibility.
There exists an optimum relaxation parameter in the convergent range for
which not only is convergence guaranteed, but also the convergence rate is
optimum.
The choice of a suitable constant relaxation scheme is problem dependent
and highly sensitive to the parameters of the partitioned sub-domains,
specifically the condensed stiffness at the interface of the partitioned sub-
domains.
In practice, the determination of an optimum constant relaxation parameter
must rely on trial and error, which may be applied to the initial part of the
coupled simulation after which full coupling analysis may be carried out.
Chapter 8 Conclusion
302
In nonlinear problems, finding the optimum relaxation parameter by trial and
error is computationally prohibitive, since it should be determined over the
full range of response.
An adaptive relaxation scheme has been developed for enhancing the performance of
iterative coupling algorithms, where the choice of the relaxation parameter is guided
by the iterative corrections of Dirichlet entities at the interface. It has been shown
that the adaptive relaxation scheme:
Avoids the trial and error procedure for the selection of an optimum, even
adequate, constant relaxation parameter.
Improves the convergence rate of constant relaxation in both linear and
nonlinear analysis significantly.
Finally, although using the adaptive scheme removes significant difficulties in the
conventional relaxation iterative coupling scheme, there is a pitfall associated with
both constant and adaptive relaxation schemes when the partitioned sub-domain
parameters dictate very small values of the relaxation parameter for convergence. In
such cases, relaxation methods break down with poor convergence rates and
significant computational inefficiency.
8.2.4 Reduced Order Method
It has been proposed in this work that the performance of iterative coupling methods
may be effectively enhanced for nonlinear analysis through the use of the condensed
interface stiffness matrices of the structure and soil partitioned sub-domains,
providing an effective first-order guide to iterative displacements at the soil-structure
interface. This would bring the performance of the proposed coupling approach very
close to the monolithic treatment.
Various domain decomposition methods for nonlinear analysis of soil-structure
interaction problems based on approximating the condensed interface stiffness
matrix have been considered and proposed. It has been shown that by using the
Chapter 8 Conclusion
303
condensed tangent interface stiffness matrices of the partitioned sub-domains in the
update of boundary conditions:
Unconditional and problem independent convergence characteristics could
be achieved.
Much greater convergence rate could be achieved compared to adaptive
relaxation scheme.
Convergence to compatibility occurs for linear problems immediately at the
first iteration. Although this immediate convergence does not normally occur
for nonlinear problems, employing this method ensures a high convergence
rate.
Moreover, in the absence of an adaptive method for evaluating the relaxation
parameter in parallel forms of coupling algorithms, utilising the above method could
be employed within various iterative coupling algorithms.
Although the condensed tangent stiffness may be readily available with some
nonlinear field modelling tools, more general approximation for the condensed
stiffness matrices is desirable. Towards this end, the condensed tangent stiffness
matrix has been approximated in this work via reduced order models, building on a
previous approach by Vierendeels et al. (2007). Nevertheless, major modifications to
this approach have been proposed in this work, leading to a more versatile and
efficient approach for coupled modelling of nonlinear soil-structure interaction
problems. In this respect:
A major pitfall associated with the originally introduced reduced order
method causing divergence is overcome by a new selective
addition/replacement procedure of force and displacement mode vectors,
where the applicability and advantages of this modification are demonstrated
by means of several examples.
In order to achieve better approximation of the condensed tangent stiffness
matrix in the initial stages of coupling iterations, a mixed reduced order
Chapter 8 Conclusion
304
method is proposed, which has a much higher convergence rate than the
conventional reduced order method.
The applicability of the presented coupling techniques has been demonstrated for
nonlinear soil-structure interaction, where the superior convergence rate of the
presented reduced order schemes, particularly the proposed mixed reduced order
method, compared to relaxation scheme is highlighted.
8.2.5 Case Studies
The applicability of the partitioned treatment of nonlinear soil-structure interaction
analysis and its great potential towards capturing the nonlinear behaviour of both soil
and structure have been illustrated through various examples. Iterative coupling has
been carried out using various existing and proposed procedures employing the
discipline-oriented solvers, ADAPTIC and ICFEP. It has been shown that the
proposed mixed reduced order method provides superior convergence characteristics
in comparison with existing coupling methods. The successful application of the
partitioned approach with iterative coupling based on the proposed coupling method
has demonstrated the applicability and great potential of the developed simulation
environment and the underlying methods towards providing an integrated
interdisciplinary computational tool for nonlinear soil-structure interaction analysis.
8.3 Recommendations for Future Works
Although iterative coupling method proposed in this work, utilising the condensed
tangent stiffness matrices of the partitioned sub-domains, possesses a high
convergence rate, its performance may be further enhanced through future research
in the following areas:
Combined coupling/field iterations: Since the field models for structure and
soil are nonlinear, iterations are typically performed in such models to
determine the state of the corresponding sub-domains. Therefore, there are
Chapter 8 Conclusion
305
potentially considerable computational benefits to be gained from combining
the field and coupling iterations. Further research is suggested on the
development of a generalised approach, which could allow combined
coupling/field iterations at arbitrary ratios of iteration (e.g. 1 (coupling):3
(structure):2 (soil)), where the optimal ratio would be established adaptively
during the specific soil-structure simulation.
Interface boundary conditions: The emphasis of this work has been on
iterative coupling methods based on a Dirichlet-Neumann algorithm. It is
suggested that other coupling algorithms could be considered in detail as part
of future research. While, the application of such algorithms with the
relaxation approach is fraught with problems, the proposed approach of
utilising the condensed interface tangent stiffness could readily transform
displacement incompatibilities at the interface to boundary forces, and is in
fact applicable to both Dirichlet-Neumann and Dirichlet-Dirichlet algorithms.
Indeed, the use of the latter algorithm with the proposed approach using a
single coupling/field iterative loop should be identical in terms of numerical
performance to the standard monolithic treatment.
Parallel processing: The partitioned treatment offers a natural framework for
parallel computations, provided the coupling algorithm is parallel over the
various partitions. It remains to be established for soil-structure interaction
problems whether parallelisation over multi-processor machines, including
optimal load balancing, is most effectively undertaken through low-level
procedural parallelisation of the field models with only two physical
partitions or through coupling-level parallelisation with additional
computational partitions. In both cases, it would be important to employ an
effective communication protocol between interface and sub-domain models,
which is equally applicable to single processor, multi-processor and
distributed computing simulations.
Chapter 8 Conclusion
306
In addition to the above, the following extensions to the originally developed
simulation environment for coupled modelling of soil-structure interaction,
specifically the INTERFACE code, would enhance its applicability:
Interface modelling: The coupling operations undertaken at the interface in
the current work assume that the adjacent sub-domains have matching
meshes, in which case the INTERFACE operations consist mainly of
transferring displacements/forces from one side of the soil/structure interface
to the other. In practice, it would be very useful to allow for non-matching
soil and structural meshes, either due to different element sizes or different
element types, thus requiring more involved INTERFACE models that
address nonlinear compatibility and equilibrium between the non-matching
sides. Another related issue is the treatment of phenomena such as contact,
sliding and friction between adjacent sub-domains, which can potentially add
further demands on the INTERFACE model.
Multiple sub-domains: Attention has been focussed in this work on coupled
modelling of soil-structure interaction problems that involve two sub-
domains only, often dictated by physical partitioning. However, the iterative
coupling algorithms could be generalized to multiple sub-domains
partitioning, to realise the additional benefits of computational, physical and
functional partitioning, including parallelisation at the coupling level.
References
Allam, M. M., Subba-Rao, K. S. & Subramanya, B. V. 1991. Frame soil interaction and winkler model Proceedings of Institution of Civil Engineers, 91, 477-494.
Aye, Z. Z., Karki, D. & Schulz, C. 2006. Ground movement prediction and building damage risk-assesment for the deep excavations and tunneling works in bangkok subsoil. International Symposium on Underground Excavation and Tunneleing. Bangkok, Thailand.
Beer, G. 1985. An isoparametric joint/interface element for finite element analysis International Journal for Numerical Methods in Engineering, 21, 585-600.
Belytschko, T. & Mullen, R. 1978. Stability of explicit-implicit mesh partitions in time integration. International Journal for Numerical Methods in Engineering, 12,1575-1586.
Bhattacharya, K., Dutta, S. C. & Dasgupta, S. 2004. Effect of soil-flexibility on dynamic behaviour of building frames on raft foundation. Journal of Sound and Vibration, 274, 111-135.
Boone, S. J. 1996. Ground-movement-related building damage. Journal of Geotechnical Engineering, 122, 886-896.
Boone, S. J. 2001. Assesing construction and settlement-induced building damage: A return to fundamental principles. In: Proceedings, Underground Construction, Institution of Mining and Metalurgy, London 559-570.
References
308
Boone, S. J., Westland, J. & Nusink, R. 1999. Comparative evaluation of building responses to an adjacent braced excavation. Canadian Goetechnical Journal, 36, 210-223.
Bourne-Webb, P. J., Potts, D. M. & Rowbottom, D. 2007. Plastic bending of steel sheet piles. Geotechnical Engineering, 160, 129-140.
Bowles, J. E. 1996. Foundation analysis and design, McGraw-Hill International Editions.
Brown, C. B., Tilton, J. R. & Laurent, J. M. 1977. Beam-plate system on winkler foundation. Journal of the Engineering Mechanics, ASCE, 103, 589-600.
Bull, J. W. 1988. Finite element analysis of thin-walled structures, Taylor & Francis.
Burland, J. B. & Potts, D. M. 1994. Development and application of a numerical model for the leaning tower of pisa. In: SHIBUYA, S., MITACHI, T. & MIURE, S. (eds.) In Prefailure deformation of geomaterials. Rotterdam: Balkema.
Carrier, W. D. & Christian, J. T. 1973. Rigid circular plate resting on a non-homogeneous elastic half space. Geotechnique, 23, 67-84.
Charles, J. A. & Skinner, H. D. 2004. Settlement and tilt of low-rise buildings. Proceedings of the Institution of Civil Engineers: Geotechnical Engineering, 157, 65-75.
Collenz, A., Bona, F. D., Gugliotta, A. & Soma, A. 2004. Large deflections of microbeams under electrostatic loads. Journal of Micromechanics and Microengineering 14, 365-373.
Desai, C. S. & Abel, J. F. 1987. Introduction to the finite element method: A numerical method for engineering analysis., CBS Publisher and Distributors.
Desai, C. S., Phan, H. V. & Perumpral, J. V. 1982. Mechanics of threedimensional soil–structure interaction. Journal of Engineering Mechanics, ASCE, 108, 731-747.
Dowrick, D. J. 1977. Earthquake resistance design: A manual for engineers and architects, New York, John Wiley and Sons Ltd.
Dutta, S., Mandal, A. & Dutta, S. C. 2004. Soil–structure interaction in dynamic behaviour of elevated tanks with alternate frame staging configurations. Journal of Sound and Vibration, 277, 825-853.
Dutta, S. C. & Roy, R. 2002. A critical review on idealization and modeling for interaction among soil–foundation–structure system. Computers & structures, 80, 1579-1594.
References
309
Dutta, S. C., Bhattacharya, K. & Roy, R. 2004. Response of low-rise buildings under seismic ground excitation incorporating soil–structure interaction. Soil Dynamics and Earthquake Engineering, 24, 893-914.
El-Gebeily, M., Elleithy, W. M. & Al-Gahtani, H. J. 2002. Convergence of the domain decomposition finite element–boundary element coupling methods. Computer Methods in Applied Mechanics and Engineering, 191, 4851-4867.
Elleithy, W. M., Al-Gahtani, H. J. & El-Gebeily, M. 2001. Iterative coupling of BE and FE methods in elastostatics. Engineering Analysis with Boundary Elements, 25, 685-695.
Elleithy, W. M. & Tanaka, M. 2003. Interface relaxation algorithms for BEM–BEM coupling and FEM–BEM coupling. Computer Methods in Applied Mechanics and Engineering, 192, 2977-2992.
Elleithy, W. M., Tanaka, M. & Guzik, A. 2004. Interface relaxation FEM–BEM coupling method for elasto-plastic analysis. Engineering Analysis with Boundary Elements, 28, 849-857.
Estorff, O. V. & Firuziaan, M. 2000. Coupled BEM/FEM approach for nonlinear soil/structure interaction. Engineering Analysis with Boundary Elements, 24, 715-725.
Estorff, O. V. & Hagen, C. 2005. Iterative coupling of FEM and BEM in 3D transient elastodynamics. Engineering Analysis with Boundary Elements, 29, 775-787.
Farhat, C. & Lesoinne, M. 2000. Two efficient staggered algorithms for the serial and parallel solution of three-dimensional nonlinear transient aeroelastic problems. Computer Methods in Applied Mechanics and Engineering, 182, 499-515.
Farhat, C. & Park, K. C. 1991. An unconditionally stable staggered algorithm for transient finite element analysis of coupled thermoelastic problems. Computer Methods in Applied Mechanics and Engineering, 85, 349-365.
Felippa, C. A. & Park, K. C. 1980. Staggered transient analysis procedures for coupled-field mechanical systems: Formulation. Computer Methods in Applied Mechanics and Engineering, 24, 61-111.
Felippa, C. A. & Park, K. C. 2004. Synthesis tools for structural dynamics and partitioned analysis of coupled systems. Multi-Physics and Multi-Scale Computer Models in Nonlinear Analysis and Optimal Design of Engineering Structures under Extreme Conditions. Ljubliana, Slovenia: Proceedings NATO-ARW PST ARW980268.
References
310
Felippa, C. A., Park, K. C. & Farhat, C. 2001. Partitioned analysis of coupled mechanical systems. Computer Methods in Applied Mechanics and Engineering, 190, 3247-3270.
Funaro, D., Quarteroni, A. & Zanolli, P. 1988. An iterative procedure with interface relaxation for domain decomposition methods. SIAM Journal on Numerical Analysis, 25, 1213-1236.
Gaba, A. R., Simpson, B., Powrie, W. & Beadman, D. R. 2002. Embedded retaining walls: Guidance for economic design, research project: 629. London: Construction Industry Information and Research Association.
Geers, T. L. & Felippa, C. A. 1980. Doubly asymptotic approximations for vibration analysis of submerged structures. Journal of the Acoustical Society of America, 73, 1152-1159.
Glowinski, R., Dinh, Q. V. & Periaux, J. 1983. Domain decomposition methods for nonlinear problems in fluid dynamics. Computer Methods in Applied Mechanics and Engineering, 40, 27-109.
Hagen, C. & Estorff, O. V. 2005a. Transient dynamic investigation of 3D dam-reservoir-soil problems using an iterative coupling approach. International Conference on Computational Methods for Coupled Problems in Science and Engineering, COUPLED PROBLEMS 2005. Barcelona: M. Papadrakakis, E. Onate and B. schrefler ( Eds)
Hagen, C. & Estorff, O. V. 2005b. Transient dynamic investigation of nonlinear 3D soil–structure interaction problems using an iterative coupling approach. PAMM · Proc. Appl. Math. Mech, 5, 399-400.
Harr, M. E. 1966. Foundation of theoretical soil mechanics, New York, Mcgraw Hill book co.
Harr, M. E., Davidson, J. L., Ho, H. O., Pombo, L. E., Ramaswamy, S. V. & Rosner, J. C. 1969. Euler beams on a twoparameter elastic foundation model. Journal of Geotechnical and Geoenvironmental Engineering, 95,933-948.
Hetenyi, M. 1946. Beams on elastic foundation University of Michigan Press.
Hoffman, J. D. 2001. Numerical methods for engineers and scientists, CRC Press.
Horn, R. A. & Johnson, C. R. 1999. Matrix analysis, Cambridge University Press.
Huang, M. & Zienkiewicz, O. C. 1998. New unconditionally stable staggered solution procedures for coupled soil-pore fluid dynamics problems. International Journal for Numerical Methods in Engineering, 43, 1029-1052.
References
311
Hughes, T. J. R. & Liu, W. K. 1978. Implicit-explicit finite elements in transisent analysis: I. Stability theory; II. Implementations and numerical examples. Journal of Applied Mechanics, 45, 371-378.
Inaba, T., Dohi, H., Okuta, K., Sato, T. & Akagi, H. 2000. Nonlinear response of surface soil and NTT building due to soil-structure interaction during the 1995 Hyogo-ken Nanbu (Kobe) earthquake. Soil Dynamics and Earthquake Engineering, 20, 289-300.
Izzuddin, B. A. 1991. Nonlinear dynamic analysis of framed structures. Ph.D., Imperial College, University of London.
Izzuddin, B. A. & Elnashai, A. S. 1993. Adaptive space frame analysis. 2. A. Distributed plasticity approach. ICE Proceedings, Structures and Buildings, 99, 317-326.
Jardine, R. J., Potts, D. M., Fourie, A. B. & Burland, J. B. 1986. Studies of the influence of non-linear stress-strain characteristics in soil-structure interaction. Geotechnique, 36, 377-396.
Jin, W.-L., Song, J., Gong, S.-F. & Lu, Y. 2005. Evaluation of damage to offshore platform structures due to collision of large barge. Engineering Structures, 27, 1317-1326.
Kamiya, N. & Iwase, H. 1997. BEM and FEM combination parallel analysis using conjugate gradient and condensation. Engineering Analysis with Boundary Elements, 20, 319-326.
Kamiya, N., Iwase, H. & Kita, E. 1996. Parallel implementation of boundary element method with domain decomposition. Engineering Analysis with Boundary Elements, 18, 209-216.
Krabbenhoft, K., Damkilde, L. & Krabbenhoft, S. 2005. Ultimate limit state design of sheet pile walls by finite elements and nonlinear programming. Computers & structures, 83, 383-393.
Kurian, N. P. & Manjokumar, N. G. 2001. A new continuous winkler model for soil-structure interaction. Journal of Structural Engineering, 27, 269-276.
Kuttler, U. & Wall, W. A. 2008. Fixed-point fluid–structure interaction solvers with dynamic relaxation. Computational Mechanics, 43, 61-72.
Lai, C. H. 1994. Diakoptics, domain decomposition and parallel computing. The Computer Journal, 37, 841-846.
Marini, L. D. & Quarteroni, A. 1989. A relaxation procedure for domain decomposition methods using finite elements. Numerische Mathematik, 55, 575-598.
References
312
Mu, M. 1999. Solving composite problems with interface relaxation. SIAM Journal on Scientific Computing, 20, 1394-1416.
Noda, T., Fernado, G. S. K., Asaoka, A. 2000. Delayed failure in soft clay foundations. Journal of the Japanese Geotechnical Society of Soils and Foundations, 40, 85-97.
Nogami, T. & Lain, Y. C. 1987. Two parameter layer model for analysis of slab on elastic foundation. Journal of Engineering Mechanics, ASCE, 1131, 1279-1291.
Noorzaei, J., Godbole, P. N. & Viladkar, M. N. 1993. Nonlinear soil-structure interaction of plane frames-a parametric study. Computers & Structures, 49, 561-566.
Noorzaei, J., Naghshineh, A., Kadir, M. R. A., Thanoon, W. A. & Jaafar, M. S. 2006. Nonlinear interactive analysis of cooling tower–foundation–soil interaction under unsymmetrical wind load. Thin-Walled Structures, 44, 997-1005.
Noorzaei, J., Viladkar, M. N. & Godbole, P. N. 1995a. Elasto-plastic analysis for soil–structure interaction in framed structures. Computers & Structures, 55, 797-807.
Noorzaei, J., Viladkar, M. N. & Godbole, P. N. 1995b. Influence of strain hardening on soil-structure interaction of framed structures. Computers & Structures, 55, 789-795.
O’brien, J. & Rizos, D. C. 2005. A 3D BEM-FEM methodology for simulation of high speed train induced vibrations. Soil Dynamics and Earthquake Engineering, 25, 289-301.
Park, K. C. 1980. Partitioned transient analysis procedures for coupled-field problems: Stability analysis. Journal of Applied Mechanics, 47, 370-376.
Park, K. C. & Felippa, C. A. 1980. Partitioned transient analysis procedures for coupled-field problems: Accuracy analysis. Journal of Applied Mechanics, 47, 919-926.
Piperno, S. 1997. Explicit/implicit fluid/structure staggered procedures with a structural predictor and fluid subcycling for 2D inviscid aeroelastic simulations. International Journal for Numerical Methods in Fluids, 25, 1207-1226.
Potts, D. M. 2003. Numerical analysis: A virtual dream or practical reality? Géotechnique, 53, 535-573.
References
313
Potts, D. M. & Zdravkovic, L. 1999. Finite element analysis in geotechnical engineering: Theory, London, Thomas Telford.
Potts, D. M. & Zdravkovic, L. 2001. Finite element analysis in geotechnical engineering: Application, London, Thomas Telford.
Quarteroni, A. & Valli, A. 1999. Domain decomposition methods for partial differential equations, Oxford, Clarendon press.
Ravaska, O. 2002. A sheet pile wall design according to Eurocode 7 and Plaxis. In: MESTAT, P. (ed.) Numerical methods in geotechnical engineering. Paris: Presses de l’ENPC/LCPC.
Rizos, D. C. & Wang, Z. 2002. Coupled BEM–FEM solutions for direct time domain soil–structure interaction analysis. Engineering Analysis with Boundary Elements, 26, 877-888.
Rombach, G. A. 2004. Finite element design of concrete structures: Practical problems and their solution, Thomas Telford.
Routh, E. J. 1877. A treatise on the stability of a given state or motion, London, Macmillan.
Roy, R. & Dutta, S. C. 2001. Differential settlement among isolated footings of building frames: the problem, its estimation and possible measures. International Journal of Applied Mechanics and Engineering 6, 165-186.
Rugonyi, S. & Bathe, K. J. 2001. On finite element analysis of fluid flows fully coupled with structural interactions. Computer Modelling in Engineering & Sciences, 2, 195-212.
Seok, J. W., Kim, O. Y., Chung, C. K. & Kim, M. M. 2001. Evaluation of ground and building settlement near braced excavation sites by model testing. Canadian Goetechnical Journal, 38, 1127-1133.
Smith, B. S. & Coull, A. 1991. Tall building structures: Analysis and design, Wiley-Interscience.
Soares-Jr., D. 2008. An optimised fem-bem time-domain iterative coupling algorithm for dynamic analyses. Computers & Structures, 86, 1839-1844.
Son, M. & Cording, E. J. 2005. Estimation of building damage due to excavation-induced ground movements. Journal of Geotechnical and Geoenvironmental Engineering 131, 162-177.
Standing, J. R., Farina, M. & Potts, D. M. 1998. The prediction of tunnelling buidling settlements-a case study. In: NEGRO, A. & FERREIRA, A. (EDS),
References
314
Proceedings of the world tunnel congress 98 on tunnels and metropolises. Sao Paulo, Brazil.
Stavridis, L. T. 2002. Simplified analysis of layered soil-structure interaction. Journal of Structural engineering, 128, 224-230.
Stewart, J. P., Fenres, G. L. & Seed, R. B. 1999. Seismic soil-structure interaction in buildings. I: Analytical methods. Journal of Geotechnical and Geoenvironmental Engineering 125, 26-37.
Taylor, D. W. 1964. Fundamentals of soil mechanics, New York, John Wiley and Sons.
Terzaghi, K. V. 1955. Evaluation of coefficient of subgrade reaction. Geotechnique, 5, 297-326.
Tian, L. & Li, Z.-X. 2008. Dynamic response analysis of a building structure subjected to ground shock from a tunnel explosion. International Journal of Impact Engineering, 35, 1164-1178.
Vallabhan, C. V. G. & Das, Y. C. 1991. Modified vlasov model for beams on elastic founadtions. Journal of Geotechnical Engineering, 117, 956-966.
Vesic, A. B. 1961. Bending on beams resting on isotropic elastic solid. Journal of Engineering Mechanic, ASCE, 2, 35-53.
Vierendeels, J. 2006. Implicit coupling of partitioned fluid–structure interaction solvers using reduced-order models. In: BUNGARTZ, S. (ed.) Fluid–structure interaction, modelling, simulation, optimisation, lecture notes in computational science and engineering. Springer.
Vierendeels, J., Lanoye, L., Degroote, J. & Verdonck, P. 2007. Implicit coupling of partitioned fluid–structure interaction problems with reduced order models. Computers & Structures, 85, 970-976.
Viladkar, M. N., Godbole, P. N. & Noorzaei, J. 1991. Soil–structure interaction in plane frames using coupled finite infinite elements. Computers & Structures, 39, 535-546.
Viladkar, M. N., Godbole, P. N. & Noorzaei, J. 1994. Modelling of interface for soil-structure interaction studies. Computers & Structures, 52, 765-779.
Viladkar, M. N., Karisiddappa, Bhargava, P. & Godbole, P. N. 2006. Static soil–structure interaction response of hyperbolic cooling towers to symmetrical wind loads. Engineering Structures, 28, 1236-1251.
Viladkar, M. N., Ranjan, G. & Sharma, R. P. 1993. Soil–structure interaction in time domain. Computers & Structures, 46, 429-442.
References
315
Vlahos, G., Cassidy, M. J. & Byrne, B. W. 2006. The behaviour of spudcan footings on clay subjected to combined cyclic loading Applied Ocean Research, 28, 209-221.
Wall, W. A., Genkinger, S. & Ramm, E. 2007. A strong coupling partitioned approach for fluid structure interaction with free surfaces. Computers & Fluids, 36, 169-183.
Wang, C. M., Xiang, Y. & Wang, Q. 2001. Axisymmetric buckling of reddy circular plates on pasternak foundation. Journal of Engineering Mechanics, ASCE, 127, 254-259.
Winkler, E. 1867. Die lehre von der elasticitaet und festigkeit [the systematic teaching of elasticity and strength], Prag,Czechoslovakia, H. Domenicus, ed.
Wrana, B. 1993. Nonlinear elastic-plastic model of soil-structure interaction in time domain Transactions on the Built Environment.
Zeevrat, L. 1972. Foundation engineering for difficult sub-soil conditions, New York Van Nostrand Reinold Company.
Zienkiewicz, O. C., Taylor R. L., Jhu J. Z. 2005. The finite element method: Its basis and fundamentals, Butterworth-Heinemann.
Zienkiewicz, O. C. & Taylor, R. L. 1991. Finite element method solid and fluid mechanics: Dynamics and nonlinearity, New York, MCGraw-Hill.
Zienkiewicz, O. C. & Taylor, R. L. 2005. The finite element method for solid and structural mechanics, Butterworth-Heinemann.
316
Appendix A
Structure of INTERFACE Data File
The INTERFACE data file includes the following information:
1- NINOD : Total number of nodes at the interface of soil-structure system
2- STGSTP: Total number of load-time increments
3- DTIM: Time step size (for dynamic analysis)
4- CV : Convergence tolerance
5- RELAX: Predefined relaxation parameter
6- TYPE: 1-Adaptive*/Condensed* 2- Adaptive/Condensed
7- METHOD : 1- Constant relaxation 2- Adaptive relaxation 3- Reduced Order
Method
8- REDTYPE: 1- Secant approximation 2- Reduced Order Method 3- Mixed
Reduced Order Method
9- LCV: Tolerance for the ADD/REP procedure
10- EPSU: Predefined incremental displacement
11- EPSF: Predefined incremental force
12- INODA: Interface node numbers of the structure partitioned sub-domain
13- INODI: Interface node numbers of the soil partitioned sub-domain
317
Appendix B
Structure of Communication Data File
A sample structure of the communication data file is presented in the following:
Record Number Variable Description 1 TGSTP Total number of load/time increments 2 TIME Current time (for dynamic analysis) 3 DTIM Time step (for dynamic analysis) 4 NINOD Total number of nodes at the interface 5 RUNA Run switch for ADAPTIC (structure solver) 6 RUNI Run switch for ICFEP (soil solver) 7 EQA Equilibrium switch (ADAPTIC) 8 EQI Equilibrium switch (ICFEP) 9 CONA Interface convergence switch (ADAPTIC) 10 CONI Interface convergence switch (ICFEP) 11 to 20
- Reserved space
21 to 20+NINOD
INODA Node numbers at the interface of structure sub-
domain 21+NONOD to 20+2×NINOD
INODI Node numbers at the interface of soil sub-domain
21+2×NINOD to 20+3×NINOD
UA Prescribed Dirichlet data at the interface of
structure 21+3×NINOD to 20+4×NINOD
FA Obtained Neumann data at the interface of
structure sub-domain 21+4×NINOD to 20+5×NINOD
FI Prescribed Neumann data at the interface of soil
sub-domain 21+5×NINOD to 20+6×NINOD
UI Obtained Dirichlet data at the interface of soil
sub-domain
Appendix C
Iterative Coupling algorithms
In the following the critical algorithmic step (the update of iterative boundary
conditions at the interface) of various iterative coupling schemes (presented in
Chapter 4) are presented, where the condensed interface tangent stiffness matrices
of partitioned sub-domains are used for the update of boundary conditions in
successive iterations enforcing convergence to compatibility/equilibrium at the
interface.
Consider domain decomposition of a soil-structure interaction problem, into structure T and soil B partitioned sub-domains, as presented below:
Governing equilibrium conditions for partitioned structure sub-domain:
11 12
21 22
T T T TT T
T T i iT T
K K U F
K K U F
(C.1)
Governing equilibrium conditions for partitioned soil sub-domain:
11 12
21 22
B B B BB B
B B i iB B
K K U F
K K U F
(C.2)
Appendix C
319
In the above, vectors XXU and X
XF correspond to the displacements and external
forces for the non-interface degrees of freedom, while iXU and i
XF correspond to
displacements and forces for the interface degrees of freedom, respectively.
Assume that the compatibility and equilibrium defaults at the interface of the
structure and soil sub-domains for iteration number I of load/time step n take the
form:
I II i iU B Tn n n
U U (C.3)
I II i iF B Tn n n
F F (C.4)
Assume that CTK and C
BK are the condensed tangent stiffness matrices at the
interface of the structure and soil sub-domains.
C.1 Sequential Neumann-Dirichlet Iterative Coupling
This algorithm is presented in Chapter 4 (Section 4.2.2).
Using the aforementioned update technique, the update of boundary conditions in
successive iterations takes the following form of:
STEP 5: If convergence to equilibrium has not been achieved, the new estimation of
the interface forces according to the Equation (C.5) is applied to the structure
domain, and iteration continues (I=I+1) from STEP 2 until convergence to
equilibrium is achieved.
11 1 1 1I I I Ii i C C CT T B T U B Fn nn n
F F K K K
(C.5)
C.2 Parallel Dirichlet-Neumann Iterative Coupling
This algorithm is presented in Chapter 4 (Section 4.2.3).
Appendix C
320
Using the aforementioned update technique, the update of boundary conditions in
successive iterations takes the following form of:
STEP 3: If convergence to compatibility and equilibrium has not been achieved, the
new estimation of the interface displacements and forces according to Equations
(C.6) and (C.7) is applied to the structure and soil sub-domains respectively, and
iteration continues (I=I+1) from STEP 2 until convergence to compatibility and
equilibrium is achieved.
11I I I Ii i C C CT T T B B U Fn nn n
U U K K K
(C.6)
11 1 1 1I I I Ii i C C CB B B T U T Fn nn n
F F K K K
(C.7)
C.3 Parallel Neumann-Dirichlet Iterative Coupling
This algorithm is presented in Chapter 4 (Section 4.2.4).
Using the aforementioned update technique, the update of boundary conditions in
successive iterations takes the following form of:
STEP 3: If convergence to compatibility and equilibrium has not been achieved, the
new estimation of the interface displacements and forces according to Equations
(C.8) and (C.9) is applied to the soil and structure sub-domains respectively, and
iteration continues (I=I+1) from STEP 2 until convergence to compatibility and
equilibrium is achieved.
11I I I Ii i C C CB B T B T U Fn nn n
U U K K K
(C.8)
11 1 1 1I I I Ii i C C CT T B T U B Fn nn n
F F K K K
(C.9)
Appendix C
321
C.4 Parallel Dirichlet-Dirichlet Iterative Coupling
This algorithm is presented in Chapter 4 (Section 4.2.5).
Using the aforementioned update technique, the update of boundary conditions in
successive iterations takes the following form of:
STEP 3: If convergence to equilibrium has not been achieved, a new estimation of
the interface displacements, according to Equations (C.10) and (C.11) is applied to
the structure and soil sub-domains, and iteration continues (I=I+1) from STEP 2
until convergence to equilibrium is achieved.
11I I I Ii i C C CT T T B B U Fn nn n
U U K K K
(C.10)
11I I I Ii i C C CB B T B T U Fn nn n
U U K K K
(C.11)
C.5 Parallel Neumann -Neumann Iterative Coupling
This algorithm is presented in Chapter 4 (Section 4.2.6).
Using the aforementioned update technique, the update of boundary conditions in
successive iterations takes the following form of:
STEP 3: If convergence to compatibility has not been achieved, a new estimation of
the interface forces according to Equations (C.12) and (C.13) is applied to the
structure and soil sub-domains respectively, and iteration continues (I=I+1) from
STEP 2 until convergence to compatibility is achieved.
11 1 1 1I I I Ii i C C CT T B T U B Fn nn n
F F K K K
(C.12)
11 1 1 1I I I Ii i C C CB B B T U T Fn nn n
F F K K K
(C.13)
Appendix D
Numerical Example
In the following, the presented example in Chapter 4 (Figure D.1) is coupled using
various coupling algorithms presented in Appendix C.
Figure E.1: Coupled spring system
i i
j j
PK K
PK K
Appendix D Numerical Example
323
To facilitate the verification of the various coupling schemes presented hereafter, the
coupled problem is initially modelled and solved monolithically. Performing global
structural analysis using a monolithic approach results in the formation and solution
of the following global system of equations:
2
3
4
5
6
40 10 0 0 0 0
10 40 10 20 0 20
0 10 25 0 15 0
0 20 0 35 15 0
0 0 15 15 40 60
(D.1)
Where in the above j denotes the displacement of node j.
Solving Equation (D.1) results in the following nodal displacements:
2
3
4
5
6
67
231268
231541
231463
231241
77
(D.2)
In order to perform the partitioned analysis using the condensed interface stiffness
approach, the partitioned sub-domains must be modelled in isolation, and their
corresponding condensed interface stiffness matrices must be obtained.
Considering sub-domain T in isolation, its governing equilibrium conditions can be
written in the form of Equation (D.3). In the following iX T
P and iX T
correspond
to the interface forces and displacements of node X in the partitioned sub-domain T
, respectively.
Appendix D Numerical Example
324
4 4
5 5
6
15 0 15
0 15 15
15 15 40 60
i i
T T
i i
T T
P
P
(D.3)
Applying condensation on interface nodes 4 and 5, the condensed stiffness matrix at
the interface of the partitioned sub-domain T can be determined as:
75 45
15 0 15 1 8 815 150 15 15 45 7540
8 8
CTK
(D.4)
Similarly, the governing equilibrium condition for partitioned sub-domain B can be
written as:
2
3
4 4
5 5
040 10 0 02010 40 10 20
0 10 10 0
0 20 0 20
i i
B B
i i
B B
P
P
(D.5)
In the above iX B
P and iX B
correspond to the forces and displacement of the
interface node X in the partitioned sub-domain B , respectively.
Again by applying condensation on interface nodes 4 and 5, the condensed stiffness
matrix at the interface of sub-domain B is obtained as:
122 16
10 0 0 10 40 10 0 0 3 30 20 0 20 10 40 10 20 16 28
3 3
CBK
(D.6)
Appendix D Numerical Example
325
D.1 Parallel Dirichlet-Neumann
Prescribing an initial guess for the interface Dirichlet data at the interface of T ,
0
4 0Ii
T
and 0
5 0Ii
T
, for the first iteration (I=0):
0 0
4 4
0 0
5 5
6
015 0 15
0 15 15 0
15 15 40 60
I Ii i
T T
I Ii i
T T
P
P
6
0
4
0
5
3
245
245
2
Ii
T
Ii
T
P
P
(D.7)
Prescribing an initial guess for the interface Neumann data at the interface of B (
0
4 0Ii
BP
and 0
5 0Ii
BP
), for the first iteration (I=0):
2 2
3 3
0 0 0
4 4 4
0 0 0
5 5 5
040 10 0 0 2 / 3
2010 40 10 20 8 / 3
00 10 10 0 8 / 3
0 20 0 20 8 / 30
I I Ii i i
B B B
I I Ii i i
B B B
P
P
(D.8)
Comparing 0
4
Ii
BP
and 0
5
Ii
BP
with 0
4
Ii
TP
and 0
5
Ii
TP
, and comparing 0
4
Ii
B
and
0
5
Ii
B
with 0
4
Ii
T
and 0
5
Ii
T
, it is clear that convergence to either equilibrium or
compatibility at the interface is not achieved. Therefore new estimates of Neumann
and Dirichlet data should be calculated enforce convergence in the next iteration:
Appendix D Numerical Example
326
1
4
1
5
1
0
0
75 45 22 16 22 16 8 54145
8 8 3 3 3 3 3 2312.45 75 16 28 16 28 8 45 463
8 8 3 3 3 3 3 2 231
Ii
T
Ii
T
(D.9)
1
4
1
5
11 1 1
0
0
75 45 22 16 8 75 45 13045
8 8 3 3 3 8 8 112.130045 75 16 28 8 45 75 45
8 8 3 3 3 8 8 2
Ii
B
Ii
B
P
P
77
(D.10)
Prescribing the new estimate for the interface Dirichlet data at the interface of T at
iteration I=1, and solving the partitioned sub-domain T gives:
11
44 6
1 1 1
5 5 4
1
6 5
541 241
231 7715 0 15463 130
0 15 15231 11
15 15 40 60 1300
77
Ii IiT
T
I I Ii i i
T T T
Ii
T
P
P P
P
(D.11)
Prescribing the new estimate for the interface Neumann data at the interface of B
at iteration I=1, and solving the partitioned sub-domain B gives:
Appendix D Numerical Example
327
2 2
3 311 1
44 4
1 115 55
670 231
40 10 0 0 2682010 40 10 20 2311300 10 10 0 54111
23113000 20 0 2046377231
II Iii iBB B
I IIi iiB BB
P
P
(D.12)
It is clear that at the first coupling iteration (I=1) convergence to equilibrium and
compatibility is achieved. Moreover, the obtained results by the above coupling
procedure are identical to those obtained by the monolithic treatment.
D.2 Parallel Dirichlet-Dirichlet
Prescribing an initial guess for the interface Dirichlet data at the interface of T (
0
4 0Ii
T
and 0
5 0Ii
T
), for the first iteration (I=0) gives:
0 0
4 4 6
0 0 0
5 5 4
06
5
30 215 0 15
450 15 15 0
215 15 40 60 45
2
I Ii i
T T
I I Ii i i
T T T
Ii
T
P
P P
P
(4.13)
Prescribing an initial guess for the interface Dirichlet data at the interface of B (
0
4 0Ii
B
and 0
5 0Ii
B
), for the first iteration (I=0):
Appendix D Numerical Example
328
2 2
3 3
0 0 0
4 4 4
0 0 0
5 5 5
2
15040 10 0 0 82010 40 10 20 15
00 10 10 0 16
30 20 0 20 032
3
I I Ii i i
B B B
I I Ii i i
B B B
P P
P P
(D.14)
Comparing 0
4
Ii
BP
and 0
5
Ii
BP
with 0
4
Ii
TP
and 0
5
Ii
TP
, it is clear that convergence
to equilibrium at the interface is not achieved. Therefore new estimates of Neumann
data at the interface of T and B should be calculated to enforce convergence in
the next coupling iteration:
1
1
4
1
5
22 16 75 45 103 5410 3 3 8 8 6 2310 16 28 45 75 71 463
3 3 8 8 6 231
Ii
T
Ii
T
(D.15)
Prescribing the new estimate for the interface Dirichlet data at the interface of T at
iteration I=1, and solving the partitioned sub-domain T gives:
00
44 6
0 0 0
5 5 4
0
6 5
541 241
231 7715 0 15463 130
0 15 15231 11
15 15 40 60 1300
77
Ii IiT
T
I I Ii i i
T T T
Ii
T
P
P P
P
(4.16)
Prescribing the new estimate for the interface Dirichlet data, by applying
compatibility, at the interface of B at iteration I=1 and solving the partitioned sub-
domain B gives:
Appendix D Numerical Example
329
22
33
0 0 04 4 4
0 005 55
67
231040 10 0 0 2682010 40 10 20 2315410 10 10 0 130231
110 20 0 20 4631300231
77
Ii I Ii iB B B
I II i iiB BB
P P
P P
(4.17)
It is clear that at the first coupling iteration (I=1) convergence to equilibrium is
achieved. Moreover, the obtained results by the above coupling procedure given by
are identical to those obtained by the monolithic treatment.
D.3 Parallel Neumann-Neumann
Prescribing an initial guess for the interface Neumann data at the interface of T (
0
4 0Ii
TP
and 0
5 0Ii
TP
), for the first iteration (I=0) gives:
0 0 0
4 4 4
0 0 0
5 5 5
6 6
015 0 15 6
0 15 15 0 6
15 15 40 660
I I Ii i i
T T T
I I Ii i i
T T T
P
P
(D.18)
Prescribing an initial guess for the interface Neumann data at the interface of B (
0
4 0Ii
BP
and 0
5 0Ii
BP
), for the first iteration (I=0) gives:
Appendix D Numerical Example
330
2 2
3 3
0 0 0
4 4 4
0 0 0
5 5 5
2
3040 10 0 0 82010 40 10 20 3
00 10 10 0 8
30 20 0 20 08
3
I I Ii i i
B B B
I I Ii i i
B B B
P
P
(D.19)
Comparing 0
4
Ii
B
and 0
5
Ii
B
with 0
4
Ii
T
and 0
5
Ii
T
, it is clear that convergence to
compatibility at the interface is not achieved. Therefore new estimates of Neumann
data should be calculated to enforce convergence in the next iteration:
11 11
4
1
5
22 16 75 45 26 1300 3 3 8 8 13 11
13000 16 28 45 75 26
773 3 8 8 13
Ii
T
Ii
T
P
P
Prescribing the new estimate for the interface Neumann data at the interface of T
at iteration I=1, and solving the partitioned sub-domain T gives:
10 1
44 4
0 1 1
5 5 5
6 6
541130
2311115 0 151300 463
0 15 1577 231
15 15 4060 241
77
IiI Ii iT
T T
I I Ii i i
T T T
P
P
(D.20)
Prescribing the new estimate for the interface Neumann data at the interface of B
at iteration I=1, and solving the partitioned sub-domain B gives:
Appendix D Numerical Example
331
2 2
3 311 1
44 4
1 115 55
670 231
40 10 0 0 2682010 40 10 20 2311300 10 10 0 54111
23113000 20 0 2046377231
II Iii iBB B
I IIi iiB BB
P
P
(D.21)
It is clear that at the first coupling iteration (I=1) convergence to equilibrium is
achieved. Moreover, the obtained results by the above coupling procedure are
identical to those obtained by the monolithic treatment.