Attribution Non-Commercial (BY-NC)

Просмотров: 451

Attribution Non-Commercial (BY-NC)

- ISLAMIC WORK ETHICS AND ORGANZATION CULTURE
- OpenSees Tutorial
- 118812445-AS-1170-1-2002
- Frac l09 Xfem
- Finite Element Analysis in Geotechnical Engineering-Vol2
- Introduction to Data Mining
- Dynamic Soil-Structure Interaction, Wolf
- Vesic (1973) Analysis of Ultimate Loads of Shallow Foundations
- skel gaji 2014
- Worked Examples to Eurocode 2
- Soil Mechanics by Lambe and Whitman
- bs8110-part-2
- Soil Liquefaction During Earthquakes - Idriss and Boulanger - 2008
- Drucker DC_1952_Soil Mechanics and Plastic Analysis or Limit Design
- Shallow Foundation
- Designers' Guide to en 1994-2_Eurocode 4_Design of Composite Steel and Concrete Structures
- FoundationCode2004
- Computational Geomechanics
- SPTN60-skempton
- 71.Theoretical Soil Mechanics (Karl Terzaghi)

Вы находитесь на странице: 1из 629

University of California, Davis, California, U.S.A.

University of Alaska, Anchorage, AK, U.S.A.

Itasca International Inc. Minneapolis, MN, U.S.A.

Wells Fargo Securities, New York, NY, U.S.A.

University of Akron, Akron, OH, U.S.A.

University of British Columbia, Vancouver, BC, Canada

Ecole Polytechnique Fdrale de Lausanne, Lausanne, Suisse

University of California, Davis, CA, U.S.A.

National Technical University of Athens, Greece

Lecture Notes

under

You are free: to Share to copy, distribute and transmit the work to Remix to adapt the work Under the following conditions: Attribution You must attribute parts or all (whatever used) of this work to Boris Jeremi (with link to his web site). Noncommercial You may not use this work for commercial purposes. Share Alike If you alter, transform, or build upon this work, you have to distribute the resulting work, and you have to distribute it under the same or similar license to this one, which must have the above Attribution, Noncommercial and this Share Alike clauses. With the understanding that: Waiver Any of the above conditions can be waived if you get written permission from the copyright holder (Boris Jeremi). Public Domain Where the work or any of its elements is in the public domain under applicable law, that status is in no way aected by the license. Other Rights In no way are any of the following rights aected by the license: Your fair dealing or fair use rights, or other applicable copyright exceptions and limitations; The authors moral rights; Rights other persons may have either in the work itself or in how the work is used, such as publicity or privacy rights. Notice For any reuse or distribution, you must make clear to others the license terms of this work. The best way to do this is with a copy of this license and a link to original work at the web site of Boris Jeremi.

Jeremi et al.

Lecture Notes

The use of the modeling and simulation system F E I (these lecture notes and accompanying modeling, computational and visualization tools) for teaching, research and professional practice is strictly encouraged. Copyright and Copyleft are covered by GPL1 and Woodys (Guthrie) license (adapted by B.J.): This work is Copylefted and Copyrighted worldwide, by the Authors, for an indenite period of time, and anybody caught using it without our permission, will be mighty good friends of ourn, cause we dont give a darn. Read it. Learn it. Use it. Hack it. Debug it. Run it. Yodel it. Enjoy it.

However, to prevent misuse (if any) of this work, we resorted to custom developed license on page 2. This new opensource license gives us a bit more (legal) control.

1 http://www.gnu.org/copyleft/gpl.html

Jeremi et al.

Lecture Notes

Purpose

The main purpose of the F E I system (these lecture notes and accompanying modeling tools, computational libraries and visualization tools) is to help us at the Computational Geomechanics Group at the University of California, Davis, research and teach numerical solution techniques for civil engineering mechanics problems. Focus is on the development and use of methods that reduce Kolmogorov complexity and modeling uncertainty. These lecture notes, in particular, are being developed to document some of the research, teaching and practical problem solving work within the Computational Geomechanics Group at the University of California at Davis, as well as to serve as the main reading material for a number of courses. Work on these lecture notes was motivated by a number of books and lecture notes ( Bathe and Wilson (1976), Felippa (1993), Lubliner (1990), Criseld (1991), Chen and Han (1988), Zienkiewicz and Taylor (1991a,b), Malvern (1969), Dunica and Kolundija (1986), Koji (1997), Hjelmstad (1997), Oberkampf et al. (2002)), that I have enjoyed over many years.

Why OpenSource?

To allow interested readers from UCDs Computational Geomechanics research group and around the world to access, use and contribute to a knowledge base (these notes and accompanying software system) that is managed, organized and quality controlled.

Jeremi et al.

Lecture Notes

Comments

Comments are much appreciated! Special thanks to (in chronological order): Miroslav Robert Roche, The best way to send a comment on lecture notes is by email, however please read the following NOTE about sending an email to me. It would be great if you can place the following in the subject line of your email: Lecture Notes. This will be much appreciated as it will help me lter your email and place it in LectureNotes email-box that I regularly read. ivkovi (Miroslav ivkovi), Dmitry J. Nicolsky, Andrzej Niemunis, Robbie Jaeger, (Yiorgos Perikleous),

Jeremi et al.

Lecture Notes

Jeremi et al.

I Theoretical and Computational Formulations 23

1 Introduction

(1996-2003-)

25

(1989-1994-)

29

(1994-2002-2010-)

55

(1991-1994-2002-2006-2010-)

57

(2004-2006-2009-)

123

(1996-2004-)

125

(1994-)

175

(1989-2006-)

189

(1999-2005-)

193

(1998-2000-2005-)

211

Lecture Notes

II

299

III

301

IV

(1999-2002-)

303

305 367 391 479 515 539

11 Static Soil-Pile and Soil-Pile Group Interaction in Single Phase Soils 12 Earthquake-Soil-Structure Interaction, General Aspects 13 Earthquake-Soil-Structure Interaction, Bridge Structures 14 Earthquake-Soil-Structure Interaction, Nuclear Power Plants 15 Cyclic Mobility and Liquefaction

(1989-2002-2009-2010-2011-)

(2003-2007-2011-)

(2010-2011-2012)

(2002-2006-2009-)

(1999-2010-)

References

541

VI

Appendix

(1985-1989-1993-...)

565

567 577 597 613 621

A Useful Formulae

(2005-2001-2010-2011-)

(1993-1995-1996-1999-)

(1993-1994-)

(1995-1996-)

Jeremi et al.

Contents

I Theoretical and Computational Formulations 23

25 26 26 26 27 27 27 27 27 27

(1996-2003-)

Chapter Summary and Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Specialization to Computational Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tour of Computational Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.1 Special Equilibrium Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.1.1 1.3.1.2 1.3.1.3 1.3.2 1.3.3 Critical Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Turning Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Failure Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.3

2 Finite Element Formulation for Single Phase Material 2.1 2.2 2.3 2.4

(1989-1994-)

29 30 30 34 43 43 44 46 47 50

Chapter Summary and Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Formulation of the Continuum Mechanics Incremental Equations of Motion . . . . . . . . . . . . Finite Element Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Isoparametric Solid Finite Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.1 2.4.2 2.4.3 2.4.4 8 Node Brick . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 Node Brick . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 Node Brick . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Isoparametric 8 20 Node Finite Element . . . . . . . . . . . . . . . . . . . . . . . . . .

2.5

Lecture Notes

10

2.6 2.7

Triangular Shell Finite Element with 6DOFs per Node . . . . . . . . . . . . . . . . . . . . . . . . Two Node, 3D, Frictional Contact Finite Element . . . . . . . . . . . . . . . . . . . . . . . . . . 2.7.1 Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

50 50 51 53 53 53 55 56 56 56 56 57 58 58 61 61 63 64 65 71 80 83 83 84 85 87 88 89 92

2.8

Seismic Isolator Finite Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.8.1 2.8.2 Two Node, 3D, Rubber Isolator Finite Element . . . . . . . . . . . . . . . . . . . . . . . Two Node, 3D, Frictional Pendulum Finite Element . . . . . . . . . . . . . . . . . . . . .

(1994-2002-2010-)

(1991-1994-2002-2006-2010-)

Chapter Summary and Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Elasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Elastoplasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.1 4.3.2 4.3.3 Constitutive Relations for Innitesimal Plasticity . . . . . . . . . . . . . . . . . . . . . . . On Integration Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Midpoint Rule Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.3.1 4.3.3.2 4.3.4 4.3.5 Accuracy Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Numerical Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Crossing the Yield Surface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Singularities in the Yield Surface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.5.1 4.3.5.2 4.3.5.3 Corner Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Apex Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Inuence Regions in Meridian Plane . . . . . . . . . . . . . . . . . . . . . . . .

4.4 4.5

A Forward Euler (Explicit) Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A Backward Euler (Implicit) Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.1 4.5.2 Single Vector Return Algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Backward Euler Algorithms: Starting Points . . . . . . . . . . . . . . . . . . . . . . . . .

University of California, Davis

Jeremi et al.

Lecture Notes

11

4.5.2.1 4.5.3

92 94 95 96 97 99

4.5.4

Gradients to the Potential Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.4.1 4.5.4.2 Analytical Gradients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Finite Dierence Gradients . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

4.6

ElasticPlastic Material Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 4.6.1 4.6.2 4.6.3 4.6.4 4.6.5 Yield Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 Plastic Flow Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 HardeningSoftening Evolution Laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 Tresca Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 von Mises Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 4.6.5.1 4.6.5.2 4.6.6 Yield and Plastic Potential Functions: von Mises Model (form I) . . . . . . . . . 106 Yield and Plastic Potential Functions: von Mises Model (form II) . . . . . . . . 107

Drucker-Prager Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 4.6.6.1 4.6.6.2 4.6.6.3 Yield and Plastic Potential Functions: Drucker-Prager Model (form I) . . . . . . 109 Yield and Plastic Potential Functions: Drucker-Prager Model (form II) . . . . . 110 Hardening and Softening Functions . . . . . . . . . . . . . . . . . . . . . . . . 111

4.6.7

Modied Cam-Clay Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 4.6.7.1 Yield and Plastic Potential Functions: Cam-Clay Model . . . . . . . . . . . . . 115

4.6.8

(2004-2006-2009-)

123 125

(1996-2004-)

Chapter Summary and Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 Continuum Mechanics Preliminaries: Kinematics . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 6.2.1 6.2.2 6.2.3 Deformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 Deformation Gradient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 Strain Tensors, Deformation Tensors and Stretch . . . . . . . . . . . . . . . . . . . . . . 129

University of California, Davis Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

12

6.2.4 6.3

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

Constitutive Relations: Hyperelasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 6.3.1 6.3.2 6.3.3 6.3.4 6.3.5 6.3.6 6.3.7 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 Isotropic Hyperelasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 VolumetricIsochoric Decomposition of Deformation . . . . . . . . . . . . . . . . . . . . 139 SimoSerrins Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 Stress Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 Tangent Stiness Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 Isotropic Hyperelastic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 6.3.7.1 6.3.7.2 6.3.7.3 Ogden Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 NeoHookean Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 MooneyRivlin Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146 Logarithmic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 SimoPister Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

Finite Deformation HyperelastoPlasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 6.4.1 6.4.2 6.4.3 6.4.4 6.4.5 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 Kinematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 Constitutive Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 Implicit Integration Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

6.5

Material and Geometric NonLinear Finite Element Formulation . . . . . . . . . . . . . . . . . . 164 6.5.1 6.5.2 6.5.3 6.5.4 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164 Equilibrium Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164 Formulation of NonLinear Finite Element Equations . . . . . . . . . . . . . . . . . . . . 165 Computational Domain in Incremental Analysis . . . . . . . . . . . . . . . . . . . . . . . 167 6.5.4.1 6.5.5 Total Lagrangian Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

University of California, Davis Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

13

6.5.5.3 6.5.5.4

7 Solution of Static Equilibrium Equations 7.1 7.2 7.3 7.4 7.5 7.6

(1994-)

Chapter Summary and Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 The Residual Force Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 Constraining the Residual Force Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 Load Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178 Displacement Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178 Generalized, HyperSpherical Arc-Length Control . . . . . . . . . . . . . . . . . . . . . . . . . . 178 7.6.1 Traversing Equilibrium Path in Positive Sense . . . . . . . . . . . . . . . . . . . . . . . . 182 7.6.1.1 Positive External Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 Angle Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182

Predictor step . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184 Automatic Increments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185 Convergence Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185 The Algorithm Progress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186 189

(1989-2006-)

Chapter Summary and Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190 The Principle of Virtual Displacements in Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . 190 Direct Integration Methods for the Equations of Dynamic Equilibrium . . . . . . . . . . . . . . . 190 8.3.1 8.3.2 Newmark Integrator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190 HHT Integrator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 193

9 Finite Element Formulation for Porous Solid-Pore Fluid Systems 9.1 9.2

(1999-2005-)

Chapter Summary and Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194 General form of upU Governing Equations 9.2.1 9.2.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194

Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194 Governing Equations of Porous Media . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194 9.2.2.1 The equilibrium equation of the mixture . . . . . . . . . . . . . . . . . . . . . . 195

University of California, Davis Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

14

The equilibrium equation of the uid . . . . . . . . . . . . . . . . . . . . . . . 195 Flow conservation equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196

Modied Governing Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196 9.2.3.1 9.2.3.2 9.2.3.3 Solid part equilibrium equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 196 Fluid part equilibrium equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 Mixture balance of mass . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

9.3

Numerical Solution of the upU Governing Equations . . . . . . . . . . . . . . . . . . . . . . . 199 9.3.1 9.3.2 9.3.3 9.3.4 9.3.5 Numerical Solution of solid part equilibrium equation . . . . . . . . . . . . . . . . . . . . 200 Numerical Solution of uid part equilibrium equation . . . . . . . . . . . . . . . . . . . . 201 Numerical Solution of ow conservation equation . . . . . . . . . . . . . . . . . . . . . . 203 Matrix form of the governing equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 204 Choice of shape functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

9.4

u-p Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205 9.4.1 9.4.2 Governing Equations of Porous Media . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205 Numerical Solutions of the Governing Equations . . . . . . . . . . . . . . . . . . . . . . . 207 9.4.2.1 9.4.2.2 9.4.2.3 Numerical solution of the total momentum balance . . . . . . . . . . . . . . . . 207 Numerical solution of the uid mass balance . . . . . . . . . . . . . . . . . . . 208 Matrix form of the governing equations . . . . . . . . . . . . . . . . . . . . . . 209 211

(1998-2000-2005-)

10.1 Chapter Summary and Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212 10.2 Plastic Domain Decomposition Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212 10.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212 10.2.2 Inelastic Parallel Finite Element . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214 10.2.2.1 Adaptive Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215 10.2.2.2 Multiphase Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216 10.2.2.3 Multiconstraint Graph Partitioning . . . . . . . . . . . . . . . . . . . . . . . . . 216 10.2.2.4 Adaptive PDD Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218 10.2.3 Adaptive Multilevel Graph Partitioning Algorithm . . . . . . . . . . . . . . . . . . . . . . 218 10.2.3.1 Unied Repartitioning Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 222

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

15

10.2.3.2 Study of ITR in ParMETIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223 10.3 Performance Studies on PDD Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223 10.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223 10.3.2 Parallel Computers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224 10.3.3 Soil-Foundation Interaction Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225 10.3.4 Numerical Study for ITR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226 10.3.5 Parallel Performance Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235 10.3.5.1 Soil-Foundation Model with 4,035 DOFs . . . . . . . . . . . . . . . . . . . . . 236 10.3.5.2 Soil-Foundation Model with 4,938 Elements, 17,604 DOFs . . . . . . . . . . . . 240 10.3.5.3 Soil-Foundation Model with 9,297 Elements, 32,091 DOFs . . . . . . . . . . . . 245 10.3.6 Algorithm Fine-Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250 10.3.7 Fine Tuning on Load Imbalance Tolerance . . . . . . . . . . . . . . . . . . . . . . . . . . 250

10.3.8 Globally Adaptive PDD Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254 10.3.8.1 Implementations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255 10.3.8.2 Performance Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257 10.3.9 Scalability Study on Prototype Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261 10.3.9.1 3 Bent SFSI Finite Element Models . . . . . . . . . . . . . . . . . . . . . . . . 261 10.3.9.2 Scalability Runs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263 10.3.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264 10.4 Application of Project-Based Iterative Methods in SFSI Problems . . . . . . . . . . . . . . . . . 271 10.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271 10.4.2 Projection-Based Iterative Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271 10.4.2.1 Conjugate Gradient Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 272 10.4.2.2 GMRES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273 10.4.2.3 BiCGStab and QMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274 10.4.3 Preconditioning Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274 10.4.4 Preconditioners . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276 10.4.4.1 Jacobi Preconditioner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276 10.4.4.2 Incomplete Cholesky Preconditioner . . . . . . . . . . . . . . . . . . . . . . . . 276

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

16

10.4.4.3 Robust Incomplete Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . 277 10.4.5 Numerical Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280 10.4.6 Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286 10.5 Performance Study on Parallel Direct/Iterative Solving in SFSI . . . . . . . . . . . . . . . . . . . 290 10.5.1 Parallel Sparse Direct Equation Solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291 10.5.1.1 General Techniques SPOOLES . . . . . . . . . . . . . . . . . . . . . . . . . . 291 10.5.1.2 Frontal and Multifrontal Methods MUMPS . . . . . . . . . . . . . . . . . . . 292 10.5.1.3 Supernodal Algorithm SuperLU . . . . . . . . . . . . . . . . . . . . . . . . . 294 10.5.2 Performance Study on SFSI Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296 10.5.2.1 Equation System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296 10.5.2.2 Performance Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297 10.5.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297

II

299

III

301

IV

(1999-2002-)

303

305

11.1 Chapter Summary and Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306 11.2 Numerical Analysis of Pile Behavior under Lateral Loads in Layered ElasticPlastic Soils . . . . . 306 11.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306 11.2.2 Constitutive Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307 11.2.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308 11.2.3.1 Pile Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308 11.2.3.2 Plastic Zones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309 11.2.3.3 p y Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310 11.2.3.4 Comparisons of Pile Behavior in Uniform and Layered Soils . . . . . . . . . . . . 315 11.2.3.5 Comparison to Centrifuge Tests and LPile Results

Jeremi et al. University of California, Davis

. . . . . . . . . . . . . . . . 322

Version: March 21, 2012, 17:15

Lecture Notes

17

11.2.4 Summary

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326

11.3 Study of soil layering eects on lateral loading behavior of piles . . . . . . . . . . . . . . . . . . . 326 11.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326 11.3.2 Finite Element Pile Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328 11.3.3 Constitutive Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329 11.3.4 Comparison of py Behavior in Uniform and Layered Soil Deposits . . . . . . . . . . . . . 330 11.3.4.1 Uniform Clay Deposit and Clay Deposit with an Interlayer of Sand. . . . . . . . 330 11.3.4.2 Uniform Sand Deposit and Sand Deposit with an Interlayer of Soft Clay. . . . . 332 11.3.5 Parametric Study for the Lateral Resistance Ratios in Terms of Stiness and Strength Parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335 11.3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341

11.4 Numerical Study of Group Eects for Pile Groups in Sands . . . . . . . . . . . . . . . . . . . . . 343

11.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343 11.4.2 Pile Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345 11.4.3 Summary of Centrifuge Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345 11.4.4 Finite Element Pile Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345 11.4.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347 11.4.6 Plastic Zone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347 11.4.7 Bending Moment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349 11.4.8 Load Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354 11.4.9 p y Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359 11.4.10 Comparison with the Centrifuge Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359 11.4.11 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365

(1989-2002-2009-2010-2011-)

367

12.1 Chapter Summary and Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368 12.2 Seismic Energy Propagation and Dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368 12.2.1 Seismic energy input into SFS system . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368 12.2.2 Seismic energy dissipation in SFS system . . . . . . . . . . . . . . . . . . . . . . . . . . . 368 12.3 Free Field Ground Motions in 3D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

18

12.3.1 Seismic Motion Lack of Correlation (Incoherence) . . . . . . . . . . . . . . . . . . . . . . 372 12.3.2 Lack of Correlation Modeling and Simulation . . . . . . . . . . . . . . . . . . . . . . . . 373 12.4 Multi-Directional and Seismic Input Coming in at Inclined Angle . . . . . . . . . . . . . . . . . . 374 12.5 Dynamic Soil-Foundation-Structure Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . 375 12.6 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377 12.7 The Domain Reduction Method 12.7.1 Method Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378

12.7.2 Method Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381 12.8 Numerical Accuracy and Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382 12.8.1 Grid Spacing h . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383 12.8.2 Time Step Length t . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383 12.8.3 Nonlinear Material Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383

12.9 Domain Boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384 12.9.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386 12.9.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387 13 Earthquake-Soil-Structure Interaction, Bridge Structures 391

(2003-2007-2011-)

13.1 Chapter Summary and Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392 13.2 Case History: Earthquake-Soil-Structure Interaction for a Bridge System . . . . . . . . . . . . . . 393 13.2.1 Prototype Bridge Model Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393 13.2.1.1 Soil Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393

13.2.1.2 Element Size Determination . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394 13.2.1.3 Time Step Length Requirement . . . . . . . . . . . . . . . . . . . . . . . . . . 398 13.2.1.4 Domain Reduction Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 399 13.2.1.5 Structural Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400 13.2.1.6 Simulation Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403 13.2.2 Earthquake Simulations - 1994 Northridge . . . . . . . . . . . . . . . . . . . . . . . . . . 405 13.2.2.1 Input Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405 13.2.2.2 Displacement Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407 13.2.2.3 Acceleration Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

19

13.2.2.4 Displacement Response Spectra . . . . . . . . . . . . . . . . . . . . . . . . . . 414 13.2.2.5 Acceleration Response Spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . 417 13.2.2.6 Structural Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 420

13.2.3 Earthquake Simulations - 1999 Turkey Kocaeli . . . . . . . . . . . . . . . . . . . . . . . . 424 13.2.3.1 Displacement Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426 13.2.3.2 Acceleration Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 430 13.2.3.3 Displacement Response Spectra . . . . . . . . . . . . . . . . . . . . . . . . . . 433 13.2.3.4 Acceleration Response Spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . 437 13.2.3.5 Structural Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440

13.2.4 Earthquake Soil Structure Interaction Eects . . . . . . . . . . . . . . . . . . . . . . . . 443 13.2.4.1 How Strength of Soil Foundations Aects ESSI . . . . . . . . . . . . . . . . . . 443 13.2.4.2 How Site Non-Uniformity Aects ESSI . . . . . . . . . . . . . . . . . . . . . . . 449

13.2.4.3 How Input Motion Aects ESSI . . . . . . . . . . . . . . . . . . . . . . . . . . 470 14 Earthquake-Soil-Structure Interaction, Nuclear Power Plants 479

(2010-2011-2012)

14.1 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 480 14.2 Finite Element Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482 14.3 Slipping behavior of SFSI models by considering 1D wave propagation . . . . . . . . . . . . . . . 484 14.3.1 Morgan Hill earthquake . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485 14.3.2 Ricker wave . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491 14.4 Slipping behavior of SFSI models by considering 3D wave propagation . . . . . . . . . . . . . . . 497 14.4.1 Ricker wave, with fault located at 45 towards the top middle point of the model . . . . . 499 14.4.2 Ricker wave, with fault located at 34 towards the top middle point of the model . . . . . 506 14.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513 15 Cyclic Mobility and Liquefaction 515

(2002-2006-2009-)

15.1 Chapter Summary and Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 516 15.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 516 15.3 Liquefaction of Level and Sloping Grounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 518 15.3.1 Model Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519 15.3.2 Behavior of Saturated Level Ground . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 520

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

20

15.3.3 Behavior of Saturated Sloping Ground . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523 15.4 Pile in Liqueed Ground, Staged Simulation Model Development . . . . . . . . . . . . . . . . . . 524 15.4.1 First Loading Stage: Self Weight . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525 15.4.2 Second Loading Stage: PileColumn Installation . . . . . . . . . . . . . . . . . . . . . . . 527 15.4.3 Third Loading Stage: Seismic Shaking . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528 15.4.4 Free Field, Lateral and Longitudinal Models . . . . . . . . . . . . . . . . . . . . . . . . . 529 15.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529 15.5.1 Pore Fluid Migration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529 15.5.2 Soil Skeleton Deformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533

15.5.3 Pile Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536 15.5.4 Pile Pinning Eects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 538 16 Slope Stability in 2D and 3D 539

(1999-2010-)

References

541

VI

Appendix

(1985-1989-1993-...)

565

567

A Useful Formulae

A.1 Chapter Summary and Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 568 A.2 Stress and Strain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 568 A.2.1 Stress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 568 A.2.2 Strain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 570 A.3 Derivatives of Stress Invariants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573 B Body and Surface Wave Solutions 577

(2005-2001-2010-2011-)

B.1 Matlab code body wave solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 578 B.2 Matlab code surface wave solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 586 B.3 Matlab code Ricker wavelet as an input motion . . . . . . . . . . . . . . . . . . . . . . . . . . 588 C The nDarray Programming Tool 597

(1993-1995-1996-1999-)

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

21

C.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 598 C.3 nDarray Programming Tool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599 C.3.1 C.3.2 Introduction to the nDarray Programming Tool . . . . . . . . . . . . . . . . . . . . . . . 599 Abstraction Levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 600 C.3.2.1 C.3.2.2 C.3.2.3 C.3.2.4 nDarray_rep class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 600 nDarray class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601 Matrix and Vector Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 602 Tensor Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 603

C.4 Finite Element Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 604 C.4.1 C.4.2 C.4.3 Stress, Strain and Elastoplastic State Classes . . . . . . . . . . . . . . . . . . . . . . . . 604 Material Model Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 604 Stiness Matrix Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 605

C.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 606 C.5.1 C.5.2 C.5.3 C.5.4 C.5.5 C.5.6 Tensor Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 606

Fourth Order Isotropic Tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607 Elastic Isotropic Stiness and Compliance Tensors . . . . . . . . . . . . . . . . . . . . . . 608 Second Derivative of Stress Invariant . . . . . . . . . . . . . . . . . . . . . . . . . . . . 608 Application to Computations in Elastoplasticity . . . . . . . . . . . . . . . . . . . . . . . 610 Stiness Matrix Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 611

C.6 Performance Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 611 C.7 Summary and Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612 D Closed Form Gradients to the Potential Function 613

(1993-1994-)

D.1 Chapter Summary and Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 614 E Hyperelasticity, Detailed Derivations 621

(1995-1996-)

E.1 Chapter Summary and Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 622 E.2 SimoSerrins Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 622 E.3 Derivation of 2 volW /(CIJ CKL ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 624 E.4 Derivation of 2 isoW /(CIJ CKL ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 625 E.5 Derivation of wA

Jeremi et al.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 626

University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

22

Jeremi et al.

Part I

23

Chapter 1

Introduction

(1996-2003-)

25

Lecture Notes

26

1.1

1.2

In this section we start from general mechanics and specialize our interest toward the eld of computational mechanics...

1.2.0.0.1

Mechanics

1.2.0.0.2

Computational Mechanics

1.2.0.0.3

1.2.0.0.4

1.2.0.0.5

1.2.0.0.6

Discretization Methods

1.2.0.0.7

1.2.0.0.8

1.3

In this section we describe various examples of equilibrium path and set up basic terminology.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

27

1.3.0.0.9

Equilibrium Path

1.3.1

1.3.1.1 1.3.1.1.1 1.3.1.1.2 1.3.1.2 1.3.1.3

Critical Points Limit Points Bifurcation Points Turning Points Failure Points

1.3.2 1.3.3

Tonti Diagrams

Jeremi et al.

Lecture Notes

28

Jeremi et al.

Chapter 2

(1989-1994-)

29

Lecture Notes

30

2.1 2.2

Chapter Summary and Highlights Formulation of the Continuum Mechanics Incremental Equations of Motion

Consider1 the motion of a general body in a stationary Cartesian coordinate system, as shown in Figure (2.1), and assume that the body can experience large displacements, large strains, and nonlinear constitutive response. The aim is to evaluate the equilibrium positions of the complete body at discrete time points 0, t, 2t, . . . , where t is an increment in time. To develop the solution strategy, assume that the solutions for the static and kinematic variables for all time steps from 0 to time t inclusive, have been obtained. Then the solution process for the next required equilibrium position corresponding to time t + t is typical and would be applied repetitively until a complete solution path has been found. Hence, in the analysis one follows all particles of the body in their motion, from the original to the nal conguration of the body. In so doing, we have adopted a Lagrangian ( or material ) formulation of the problem.

x2 , tx2 , x2

t+ t 0 0 0

t P( t+tx1t+ , x2 , x3 ) t+ t

t t

t+t t+t

A V

A V

A 0 V

Configuration at time 0

Configuration at time t

n+1

0

t+t t

x3 ,tx3 , x3 t+ t

xi = x i + u i x i = 0xi + t ui t+t xi = t xi + ui

t+t

x1 ,t x1,t+ tx1

i=1,2,3

In the Lagrangian incremental analysis approach we express the equilibrium of the body at time t + t using the

1 detailed

Jeremi et al.

Lecture Notes

31

principle of virtual displacements. Using tensorial notation2 this principle requires that:

t+t

t+t V

ij

t+t eij

t+t

dV = t+t R

(2.1)

where the t+t ij are Cartesian components of the Cauchy stress tensor, the t+t eij are the Cartesian components of an innitesimal strain tensor, and the means "variation in" i.e.:

t+t eij

1 2

ui t+t x

+

j

uj t+t x

=

i

1 2

ui t+t x

+

j

uj

t+t x i

(2.2)

It should be noted that Cauchy stresses are "body forces per unit area" in the conguration at time t + t, and the innitesimal strain components are also referred to this as yet unknown conguration. The right hand side of equation (2.1), i.e. time t + t:

t+t

R is the virtual work performed when the body is subjected to a virtual displacement at

t+t

R=

t+t V

t+t B fi

+t t+t u t ui i

t+t

dV +

t+t S

t+t S fi

+t ut i

t+t

dS

(2.3)

where

t+t B fi

and

spectively, and

t+t S fi t+t u i is

are the components of the externally applied body and surface force vectors, rethe inertial body force that is present if accelerations are present3 , ui is the ith

component of the virtual displacement vector. A fundamental diculty in the general application of equation (2.1) is that the conguration of the body at time t + t is unknown. The continuous change in the conguration of the body entails some important consequences for the development of an incremental analysis procedure. For example, an important consideration must be that the Cauchy stress at time t + t cannot be obtained by simply adding to the Cauchy stresses at time t a stress increment which is due only to the straining of the material. Namely, the calculation of the Cauchy stress at time t + t must also take into account the rigid body rotation of the material, because the components of the Cauchy stress tensor are not invariant with respect to the rigid body rotations4 . The fact that the conguration of the body changes continuously in a large deformation analysis is dealt with in an elegant manner by using appropriate stress and strain measures and constitutive relations. When solving the general problem5 one possible approach6 is given in Simo (1988). Previous discussion was oriented toward small deformation, small displacement analysis leading to the use of Cauchy stress tensor ij and small strain tensor eij .

2 Einsteins

summation rule is implied unless stated dierently, all lower case indices (i, j, p, q, m, n, o, r, s, t, . . . ) can have values

of 1, 2, 3, and values for capital letter indices will be specied where need be. 3 This is based on DAlemberts principle. 4 However, that problem will not be addressed in this work since this work deals with MaterialNonlinearOnly analysis of solids, thus excluding large displacement and large strain eects. 5 That is, large displacements, large deformations and nonlinear constitutive relations. 6 This is still a "hot" research topic!

Jeremi et al.

Lecture Notes

32

In the following, we will briey cover some other stress and strain measures particularly useful in large strain and large displacement analysis. The basic equation that we want to solve is relation (2.1), which expresses the equilibrium and compatibility requirements of the general body considered in the conguration corresponding to time t + t. The constitutive equations also enter (2.1) through the calculation of stresses. Since in general the body can undergo large displacements and large strains, and the constitutive relations are nonlinear, the relation in (2.1) cannot be solved directly. However, an approximate solution can be obtained by referring all variables to a previously calculated known equilibrium conguration, and linearizing the resulting equations. This solution can then be improved by iterations. To develop the governing equations for the approximate solution obtained by linearization we recall that the solutions for time 0, t, 2t, . . . , t have already been calculated and that we can employ the fact that the 2nd PiolaKirchho stress tensor is energy conjugate to the GreenLagrange strain tensor:

0V

t 0Sij

0 t 0 ij dV =

0V

0 0 t x mn 0 t xj,n t t i,m

t 0 xk,i

t tekl

t 0

xl,j

dV =

0V

0 t mn t

0 t 0emn dV (2.4)

because:

t 0 0 xk,l t xl,m

= km

and since:

dV = t t dV

we have:

0V

t 0Sij

0 t 0 ij dV =

t

0V

t mn t temn dV

(2.5)

t 0Sij

0 0 t x mn 0 t xj,n t t i,m

(2.6)

University of California, Davis Version: March 21, 2012, 17:15

Jeremi et al.

0 xi t xm

0

Lecture Notes

33

where 0 t xj,n =

and

represents the ratio of the mass density at time 0 and time t, and the GreenLagrange

t 0 ij

1 2

t 0 ui,j

(2.7)

Then, by employing (2.5) we refer the stresses and strains to one of these known equilibrium congurations. The choice lies between two formulations which have been termed total Lagrangian and updated Lagrangian formulations. In the total Lagrangian formulations, also termed Lagrangian formulation, all static and kinematic variables are referred to the initial conguration at time 0, while in the updated Lagrangian formulation all static and kinematic variables are referred to the conguration at time t. Both the total Lagrangian and updated Lagrangian formulations include all kinematic nonlinear eects due to large displacement, large rotations and large strains, but whether the large strain behavior is modeled appropriately depends on the constitutive relations specied. The only advantage of using one formulation rather than the other lies in its greater numerical eciency. Using (2.5) in the total Lagrangian formulations we consider this basic equation:

0V

t+t Sij 0

+t 0 t+t t R ij dV = 0

(2.8)

tV

t+t Sij t

+t t t+t t R ij dV = t

(2.9)

in which t+t R is the external virtual work as dened in (2.3). Approximate solution to the (2.8) and (2.9) can be obtained by linearizing these relations. By comparing the total Lagrangian and updated Lagrangian formulations we can observe that they quite analogous and that, in fact, the only theoretical dierence between the two formulations lies in the choice of dierent reference congurations for the kinematic and static variables. If in the numerical solution the appropriate constitutive tensors are employed, identical results are obtained. Coupling of large deformation, large displacement and material nonlinear analysis is still the topic of research in the research community. Possible direction may be the use of both Lagrangian and Eulerian formulation intermixed in one scheme.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

34

2.3

Consider the equilibrium of a general threedimensional body such as in Figure (2.2) (Bathe, 1996). The external forces acting on a body are surface tractions fiS and body forces fiB . Displacements are ui and strain tensor7 is eij and the stress tensor corresponding to strain tensor is ij .

x3

f3B r3 f1B f2B f1

f3

S

f2

r1

r2

x2

x1

Figure 2.2: General three dimensional body Assume that the externally applied forces are given and that we want to solve for the resulting displacements, strains and stresses. One possible approach to express the equilibrium of the body is to use the principle of virtual displacements. This principle states that the equilibrium of the body requires that for any compatible, small virtual displacements8 imposed onto the body, the total internal virtual work is equal to the total external virtual work. This statement can be mathematically expressed using equation (2.10) for the body at time t + t, and since we are using the incremental approach let us drop the time dimension, so that all the equations are imposed for the given increment9 , at time t + t. The equation is now, using tensorial notation10 :

ij eij dV =

V V

fiB u i ui dV +

S

fiS ui dS

(2.10)

The internal work given on the left side of (2.10) is equal to the actual stresses ij going through the virtual

small strain tensor as dened in equation: eij = 1 (ui,j + uj,i ). 2 satisfy the essential boundary conditions. 9 t + t will be dropped from now one in this chapter. 10 Einsteins summation rule is implied unless stated dierently, all lower case indices (i, j, p, q, m, n, o, r, s, t, . . . ) can have values

7 8 which

of 1, 2, 3, and values for capital letter indices will be specied where need be.

Jeremi et al.

Lecture Notes

35

strains eij that corresponds to the imposed virtual displacements. The external work is on the right side of (2.10) i going through the virtual displacements and is equal to the actual (surface) forces fiS and (body) forces fiB u ui . It should be emphasized that the virtual strains used in (2.10) are those corresponding to the imposed body and surface virtual displacements, and that these displacements can be any compatible set of displacements that satisfy the geometric boundary conditions. The equation in (2.10) is an expression of equilibrium, and for dierent virtual displacements, correspondingly dierent equations of equilibrium are obtained. However, equation (2.10) also contains the compatibility and constitutive requirements if the principle is used in the appropriate manner. Namely, the displacements considered should be continuous and compatible and should satisfy the displacement boundary conditions, and the stresses should be evaluated from the strains using appropriate constitutive relations. Thus, the principle of virtual displacements embodies all requirements that need be fullled in the analysis of a problem in solid and structural mechanics. The principle of virtual displacements can be directly related to the principle that total potential of the system must be stationary. In the nite element analysis we approximate the body in Figure (2.2) as an assemblage of discrete nite elements with the elements being interconnected at nodal points on the element boundaries. The displacements measured in a local coordinate system r1 , r2 and r3 within each element are assumed to be a function of the displacements at the N nite element nodal points:

uu a = HI u Ia

(2.11)

where I = 1, 2, 3, . . . , n and n is number of nodes in a specic element, a = 1, 2, 3 represents a number of dimensions (can be 1 or 2 or 3), HI represents displacement interpolation vector, u Ia is the tensor of global generalized displacement components at all element nodes. The use of the term generalized displacements means that both translations, rotations, or any other nodal unknown are modeled independently. Here specically only translational degrees of freedom are considered. The strain tensor is dened as:

eab =

1 (ua,b + ub,a ) 2

(2.12)

and the by using (2.11) we can dene the approximate strain tensor:

eab e ab

= = =

University of California, Davis

(2.13)

Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

36

0 ab = Eabcd e cd e0 cd + ab

(2.14)

where ab is the approximate Cauchy stress tensor, Eabcd is the constitutive tensor12 , e cd is the innitesimal

0 approximate strain tensor, e0 cd is the innitesimal initial strain tensor and ab is the initial Cauchy stress tensor.

Using the assumption of the displacements within each nite element, as expressed in (2.11), we can now derive equilibrium equations that corresponds to the nodal point displacements of the assemblage of nite elements. We can rewrite (2.10) as a sum13 of integrations over the volume and areas of all nite elements:

ab e ab dV m =

m V

m

B fa u a u a dV m + m

Sm

S m fa u S a dS

(2.15)

where m = 1, 2, 3, . . . , k and k is the number of elements. It is important to note that the integrations in (2.15) are performed over the element volumes and surfaces, and that for convenience we may use dierent element

coordinate systems in the calculations. If we substitute equations (2.11), (2.12), (2.13) and (2.14) in (2.15) it follows:

Vm

0 Eabcd e cd e0 cd + ab

1 (HI,b u Ia + HI,a u Ib ) 2

Sm

dV m = (2.16)

Vm

B fa (HI u Ia ) dV m m

HJ u Ja (HI u Ia ) dV m +

Vm m

S fa (HI u Ia ) dS m

or:

Eabcd

m V

m

1 (HJ,d u Jc + HJ,c u Jd ) e0 cd 2

Vm

0 + ab

1 (HI,b u Ia + HI,a u Ib ) 2

Sm

dV m =

=

m Vm

B fa (HI u Ia ) dV m m

HJ u Ja (HI u Ia ) dV m +

m

S fa (HI u Ia ) dS m

(2.17)

We can observe that in the previous equations represents a virtual quantity but the rules for are quite similar to regular dierentiation so that can enter the brackets and "virtualize" the nodal displacement14 . It thus follows:

11 in

terms of exact stress and strain elds but it holds for approximate elds as well. tensor can be elastic or elastoplastic constitutive tensor. 13 Or, more correctly as a union m since we are integrating over the union of elements. 14 since they are driving variables that dene overall displacement eld through interpolation functions

12 This

Jeremi et al.

Lecture Notes

37

Eabcd

m Vm

1 (HJ,d u Jc + HJ,c u Jd ) e0 cd 2

m

0 + ab

1 (HI,b u Ia + HI,a u Ib ) 2

Sm

dV m =

=

m V

m

B fa (HI u Ia ) dV m m V

HJ u Ja (HI u Ia ) dV m +

m

S fa (HI u Ia ) dS m

(2.18)

Let us now work out some algebra in the left hand side of equation (2.18): Eabcd

m V

m

(HJ,d u Jc + HJ,c u Jd ) 2

B fa (HI u Ia ) dV m m

0 Eabcd e0 cd + ab

(HI,b u Ia + HI,a u Ib ) 2

Sm

dV m =

=

m Vm

HJ u Ja HI u Ia dV m +

Vm m

S fa (HI u Ia ) dS m

(2.19)

and further:

Vm

1 (HJ,d u Jc + HJ,c u Jd ) 2 +

m Vm

Eabcd

Vm

dV m + dV m + dV m =

Eabcd e0 cd

0 ab

+

m V

m

B fa (HI u Ia ) dV m

m Vm

HJ u Ja HI u Ia dV m +

m Sm S fa (HI u Ia ) dS m

(2.20)

Several things should be observed in the equation (2.20). Namely, the rst three lines in the equation can be simplied if one takes into account symmetries of Eijkl and ij . In the case of the elastic stiness tensor Eijkl major and both minor symmetries exist. In the case of the elastoplastic stiness tensor, such symmetries exists if a ow a rule is associated. If ow rule is nonassociated, only minor symmetries exist while major symmetry is destroyed15 . As a matter of fact, both minor symmetries in Eijkl are the only symmetries we need, and the rst line of (2.20) can be rewritten as:

15 for

more on stiness tensor symmetries see sections (4.2, 4.4 and 4.5)

Jeremi et al.

Lecture Notes

38

Vm

1 (HJ,d u Jc + HJ,c u jd ) 2

Eabcd

1 (HI,b u Ia + HI,a u Ib ) 2

dV m =

=

m V

m

m Vm

(2.21)

Similar simplications are possible in second and third line of equation (2.20). Namely, in the second line we can use both minor symmetries of Eijkl so that:

Eabcd e0 cd

1 (HI,b u Ia + HI,a u Ib ) 2

dV m = (2.22)

Vm

Eabcd e0 Ia ) dV m cd (HI,b u

and the third line can be simplied due to the symmetry in Cauchy stress tensor ij as:

0 ab

1 (HI,b u Ia + HI,a u Ib ) 2

dV m = (2.23)

=

m Vm

0 ab (HI,b u Ia ) dV m

m V

m

+

m Vm

Eabcd e0 Ia ) dV m + cd (HI,b u

m B fa

Vm

=

m V

m

(HI u Ia ) dV

m V

m

HJ u Ja HI u Ia dV m +

m Sm

or if we leave the unknown nodal accelerations16 u Jc and displacements u Jc on the left hand side and move all the known quantities on to the right hand side:

16 It

is noted that u Jc = ac u Ja relationship was used here, where ac is the Kronecker delta.

Jeremi et al.

Lecture Notes

39

HJ ac u Jc HI u Ia dV m +

m V

m

m V

m

=

m Vm

B fa (HI u Ia ) dV m + m

Sm

S fa (HI u Ia ) dS m + 0 ab (HI,b u Ia ) dV m

+

m V

m

Eabcd e0 cd

(HI,b u Ia ) dV m

m V

m

(2.25)

To obtain the equation for the unknown nodal generalized displacements from (2.25), we invoke the virtual displacement theorem which states that virtual displacements are any, non zero, kinematically admissible displacements. In that case we can factor out nodal virtual displacements u Ia so that equation (2.25) becomes:

HJ ac u Jc HI dV m +

m Vm B fa HI dV m u Ia + m m Vm

S fa HI dS m u Ia +

u Ia =

=

m V

m

Sm

+

m Vm

m Eabcd e0 u Ia cd HI,b dV m

Vm

0 ab HI,b dV m u Ia

(2.26)

Jc dV m + HJ ac HI u

m Vm

m Vm B fa HI dV m + m S fa HI dS m + 0 ab HI,b dV m

=

m Vm

Sm

+

m Vm

m Eabcd e0 cd HI,b dV m

(2.27)

Vm

One should also observe that in the rst line of equation (2.27) generalized nodal accelerations u Jc and generalized nodal displacements u Jc are unknowns that are not subjected to integration so they can be factored out of the integral:

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

40

HJ ac HI dV m u Jc

m Vm

+

m V

m

B fa HI dV m + m S fa HI dS m + 0 ab HI,b dV m

=

m Vm

Sm

+

m V

m

Eabcd e0 cd

HI,b dV m

m V

m

(2.28)

(m)

MIacJ =

Vm

HJ ac HI dV m

(2.29)

(m)

KIacJ =

Vm

(2.30)

(m) B FIa

=

Vm

B fa HI dV m

(2.31)

(m) S FIa

=

Sm

S fa HI dS m

(2.32)

=

Vm

m Eabcd e0 cd HI,b dV

(2.33)

0 (m) mn FIa

=

V

m

0 ab HI,b dV m

(2.34)

Jeremi et al.

(m)

Lecture Notes

(m) B FIa (m) S FIa

41

where

(m)

(m)

is the tensor

FIa

e0 mn

FIa

0 mn

(m) (m)

MIacJ u Jc +

(m)

(m)

KIacJ u Jc =

m

(m) B FIa

+

m

(m) S FIa

+

m

0 (m) mn FIa

(2.35)

MAacB u Bc + KAacB u Bc A, B

= =

(2.36)

a, c =

where:

MAacB =

m

(m)

MIacJ

KAacB =

m

(m)

KIacJ

(2.37)

are the system mass and stiness tensors, respectively, u Bc is the tensor of unknown nodal accelerations, and u Bc is the tensor of unknown generalized nodal displacements, while the load tensor is given as:

FAa =

m

(m) B FIa

+

m

(m) S FIa

+

m

0 (m) mn FIa

(2.38)

After assembling the system of equations in (2.37) it is relatively easy to solve for the unknown displacements u Lc either for static or fully dynamic case. It is also very important to note that in all previous equations, omissions of inertial force term (all terms with ) will yield static equilibrium equations. Description of solutions procedures for static linear and nonlinear problems are described in some detail in chapter 7. In addition to that, solution procedures for dynamic, linear and nonlinear problems are described in some detail in chapter 8. A note on the nal form of the tensors used is in order. In order to use readily available system of equation solvers equation (2.37) will be rewritten in the following from:

17 Summation

of the element volume integrals expresses the direct addition of the element tensors to obtain global, system tensors.

This method of direct addition is usually referred to as the direct stiness method.

Jeremi et al.

Lecture Notes

42

MP Q u P + KP Q u P = FQ

(2.39)

where MP Q is system mass matrix, KP Q is system stiness matrix and FQ is the loading vector. Matrix form of equation 2.37, presented as equation 2.39 is obtained attening the system mass tensor MAacB , system stiness tensor KAacB , unknown acceleration tensor u Bc , unknown displacement tensor u Bc and the system loading tensor FAa . Flattening from the fourth order mass/stiness tensors to two dimensional mass/stiness matrix is done by simply performing appropriate (re) numbering of nodal DOFs in each dimension. Similar approach is used for unknown accelerations/displacements and for loadings.

2.3.0.0.1

(fIa )int =

(m)

(m)

KIacJ u Jc =

m Vm

ab HI,b dV m

(2.40)

(fIa )ext =

m

(m) B FIa

+

m

(m) S FIa

+

m

0 (m) mn FIa

(2.41)

where (fIa )int is the internal force tensor and (fIa )ext is the external force tensor. Equilibrium is obtained when residual: rIa ( uJc , ) = (fIa ( uJc ))int (fIa )ext (2.42)

(2.43)

Jeremi et al.

Lecture Notes

43

2.4

2.4.1

8 Node Brick

Ni

(e)

1 (1 + i )(1 + i )(1 + i ) 8

(2.44)

Jeremi et al.

Lecture Notes

44

Table 2.2: Values of , , and at each of the 9th to 20th nodes Node 9 10 11 12 13 14 15 16 17 18 19 20 0 -1 0 +1 0 -1 0 +1 +1 -1 -1 +1 +1 0 -1 0 +1 0 -1 0 +1 +1 -1 -1 +1 +1 +1 +1 -1 -1 -1 -1 0 0 0 0

2.4.2

20 Node Brick

Shape function of the 8 corner nodes (1 to 8) which i indicates the node number:

Ni

( e)

University of California, Davis

(2.45)

Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

45

Shape function of the node numbers 9, 11, 13, and 15 which i indicates the node number:

Ni

(e)

1 (1 2 )(1 + i )(1 + i ) 4

(2.46)

Shape function of the node numbers 10, 12, 14, and 16 which i indicates the node number:

Ni

(e)

1 (1 2 )(1 + i )(1 + i ) 4

(2.47)

Shape function of the node numbers 17, 18, 19, and 20 which i indicates the node number:

Ni

(e)

1 (1 2 )(1 + i )(1 + i ) 4

(2.48)

Jeremi et al.

Lecture Notes

46

2.4.3

27 Node Brick

Table 2.3: Values of , , and at each of the 21th to 27th nodes Node 21 22 23 24 25 26 27 0 0 -1 0 +1 0 0 0 +1 0 -1 0 0 0 0 0 0 0 0 +1 -1

Shape function of the 8 corner nodes (1 to 8) which i indicates the node number:

Ni

( e)

(2.49)

Shape function of the node numbers 9, 11, 13, and 15 which i indicates the node number:

Ni

( e)

University of California, Davis

(2.50)

Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

47

Shape function of the node numbers 10, 12, 14, and 16 which i indicates the node number:

Ni

(e)

(2.51)

Shape function of the node numbers 17, 18, 19, and 20 which i indicates the node number:

Ni

(e)

(2.52)

N21 = (1 2 )(1 2 )(1 2 ) Shape function of the node numbers 22 and 24 which i indicates the node number:

(e)

(2.53)

Ni

(e)

(2.54)

Shape function of the node numbers 23 and 25 which i indicates the node number:

Ni

(e)

(2.55)

Shape function of the node numbers 26 and 27 which i indicates the node number:

Ni

(e)

(2.56)

2.4.4

The basic procedure in the isoparametric18 nite element formulation is to express the element coordinates and element displacements in the form of interpolations using the local three dimensional19 coordinate system of the element. Considering the general 3D element, the coordinate interpolations, using indicial notation20 are:

xi = HA (rk ) x Ai

18 name

(2.57)

isoparametric comes from the fact that both displacements and coordinates are dened in terms of nodal values. Super-

parametric and subparametric nite elements exists also. 19 in the case of element presented here, that is isoparametric 8 20 node nite element. 20 Einsteins summation rule is implied unless stated dierently, all lower case indices (i, j, p, q, m, n, o, r, s, t, . . . ) can have values of 1, 2, 3, and values for capital letter indices will be specied where need be.

Jeremi et al.

Lecture Notes

48

where A = 1, 2, . . . , n and n is the total number of nodes associated with that specic element, x Ai is the i-th coordinate of node A, i = 1, 2, 3, k = 1, 2, 3 and HA are the interpolation functions dened in local coordinate system of the element, with variables r1 , r2 and r3 varying from 1 to +1. The interpolation functions HA for the isoparametric 820 node are the so called serendipity interpolation functions mainly because they were derived by inspection. For the nite element with nodes numbered as in Figure (2.6) they are given21 in the following set of formulae:

r3

3 4 11 12 15 16 1 17 13 5 19 7 9 14 18 6 10 2

r 2

x2 r1 20

8

x1 x3

Figure 2.6: Isoparametric 820 node brick element in global and local coordinate systems

H20 =

2 isp (20 ) (1 + r1 ) (1 r2 ) 1 r3 4 2 isp (18 ) (1 r1 ) (1 + r2 ) 1 r3 H18 = 4 2 isp (16 ) (1 + r1 ) 1 r2 (1 r3 ) H16 = 4 2 (1 r3 ) isp (14 ) (1 r1 ) 1 r2 H14 = 4 2 isp (12 ) (1 + r1 ) 1 r2 (1 + r3 ) H12 = 4 2 isp (10 ) (1 r1 ) 1 r2 (1 + r3 ) H10 = 4

isp (19 ) (1 r1 ) (1 r2 ) 4 isp (17 ) (1 + r1 ) (1 + r2 ) H17 = 4 2 isp (15 ) 1 r1 (1 r2 ) H15 = 4 2 (1 + r2 ) isp (13 ) 1 r1 H13 = 4 2 isp (11 ) 1 r1 (1 r2 ) H11 = 4 2 isp (9 ) 1 r1 (1 + r2 ) H9 = 4 H19 =

2 1 r3 2 1 r3

(1 r3 ) (1 r3 ) (1 + r3 ) (1 + r3 )

21 for

Jeremi et al.

Lecture Notes

49

H8 =

(1 + r1 ) (1 r2 ) (1 r3 ) h15 h16 h20 + 8 2 (1 r1 ) (1 r2 ) (1 r3 ) h14 h15 h19 H7 = + 8 2 (1 r1 ) (1 + r2 ) (1 r3 ) h13 h14 h18 H6 = + 8 2 (1 + r1 ) (1 + r2 ) (1 r3 ) h13 h16 h17 H5 = + 8 2 (1 + r1 ) (1 r2 ) (1 + r3 ) h11 h12 h20 + H4 = 8 2 (1 r1 ) (1 r2 ) (1 + r3 ) h10 h11 h19 H3 = + 8 2 (1 r1 ) (1 + r2 ) (1 + r3 ) h10 h18 h9 H2 = + 8 2 (1 + r1 ) (1 + r2 ) (1 + r3 ) h12 h17 h9 + H1 = 8 2

where r1 , r2 and r3 are the axes of natural, local, curvilinear coordinate system and isp (nod_num) is boolean function that returns +1 if node number (nod_num) is present and 0 if node number (nod_num) is not present. To be able to evaluate various important element tensors22 , we need to calculate the straindisplacement transformation tensor23 . The element strains are obtained in terms of derivatives of element displacements with respect to the local coordinate system. Because the element displacements are dened in the local coordinate system, we need to relate global x1 , x2 and x3 derivatives to the r1 , r2 and r3 derivatives. In order to obtain derivatives with respect to global coordinate system, i.e. form:

xa

(2.58)

xa = = Jak rk rk xa xa

(2.59)

where Jak is the Jacobian operator relating local coordinate derivatives to the global coordinate derivatives:

0 0 22 i.e. (m)K (m)F B , (m)F S , (m)F mn , (m)F mn , IacJ , Ia Ia Ia Ia 1 23 from the equation HI,b u Ia + HI,a u Ib ab = 2

Jeremi et al.

Lecture Notes

50

Jak

x 1 r1 x xa 1 = = r2 rk x1 r3

x2 r1 x2 r2 x2 r3

x3 r1 x3 r2 x3 r3

(2.60)

The existence of equation (2.58) requires that the inverse of Jak exists and that inverse exists provided that there is a onetoone24 correspondence between the local and the global coordinates of element. It should be pointed out that except for the very simple cases, volume and surface element tensor25 integrals are evaluated by means of numerical integration26 Numerical integration rules is quite a broad subject and will not be covered here27 .

2.5

2.6

Bergan and Felippa (1985); Alvin et al. (1992); Felippa and Militello (1992); Felippa and Alexander (1992); Militello and Felippa (1991)

2.7

Haraldsson and Wriggers (2000); Sheng et al. (2007a) Almost in all the soil-structure systems, there are interfaces between dierent materials which are in contact. These contact areas can get detached during the loading so there will be gaps in the model. The focus of this section is on the contact of concrete and soil. In fact transferring the loads to structure is dened based on the contact areas which will aect eh response of the soil-structure systems (Kikuchi and Oden (1988)). The main objective at the contact areas is to dene normal and shear forces, and displacements at the common surfaces. The idea of contact element has become from elements so called joint elements which was initially developed for modeling of rock joints. Typically normal and tangential stiness were used to model the pressure and friction at the interface (Sheng et al. (2007b); Haraldsson and Wriggers (2000); Desai and Siriwardane (1984)).

24 unique. 25 as

dened in chapter (2.3) by equations (2.30), (2.31), (2.32), (2.33) and (2.34). NewtonCoates, Lobatto are among the most used integration rules. 27 nice explanation with examples is given in Bathe (1982).

26 GaussLegendre,

Jeremi et al.

Lecture Notes

51

The study of two dimensional and axisymmetric benchmark examples have been done by Olukoko et al. (1993) for linear elastic and isotropic contact problems. Study was done considering Coulombs law for frictional behavior at the interface. In many cases the interaction of soil and structure is involved with frictional sliding of the contact surfaces, separation, and reclosure of the surfaces. These cases depend on the loading procedure and frictional parameters. Sheng et al. (2007b) discussed how frictional contact is important for structural foundations under loading, pile foundations, soil anchors, and retaining walls. Khoei and Nikbakhti (2006) used the extended nite element method (X-FEM) for modeling of frictional contact problems. In order to be able to simulate discontinuities in model during the analysis without considering the boundary conditions of domain, special functions should be included within standard nite element framework. A nite element method for analysis of frictional contact problems was developed by Chandrasekaran et al. (1987). The nonlinear problem is solved by imposing geometric constraints on equilibrium conguration which compatibility conditions are violated. Dierent conditions happening at contact area, such as sticking and slipping, are determined from the relative magnitudes of the normal and tangential nodal forces.

Developing contact nite element for analysis of connections between laminated composite plates has been done by Barbero et al. (1995). Such element is compatible with a three-dimensional plate element based on constant shear theory. The contact element formulation is based on a simple regularization of the unilateral contact with a Coulomb friction problem. Two-dimensional frictional polynomial to segment contact elements are developed by Haraldsson and Wriggers (2000) based on non-associated frictional law and elastic-plastic tangential slip decomposition. Several benchmarks are presented by Konter (2005) in order to verify the the results of the nite element analyses performed on 2D and 3D modelings. In all proposed benchmarks the results were approximated pretty well with a 2D or an axisymmetric solutions. In addition, 3D analyses were performed and the results were compared with the 2D solutions. Hjiaj et al. (2004) presented an algorithm for solving frictional contact problems where sliding rule is nonassociated. The algorithm is a combination of the classical unilateral contact law and an anisotropic friction model with a non-associated slip rule. It has been observed that slip rule has a strong inuence on the frictional behavior.

2.7.1

Formulation

One of the methods of deriving frictional contact law is using the procedure of computational homogenization. In this method, numerical simulation is done using homogeneous boundary conditions for a representative volume element. Figure (2.7 -a) shows the two-dimensional state of this representative element. Real contact geometry with known proles from the micro-mechanical surfaces has to be included. It has to be large enough to resolve the micro-structure of the surface geometries of contact area. When the boundary conditions are such that only deformations are applied in the normal direction, then the computational homogenization procedure yields a

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

52

homogenized contact law for the normal pressure (pN ) (Figure (2.7 -b)). If relative tangential motions are described, then a friction law for contact stresses can be obtained. A general contact law for the contact stresses, along normal and tangential directions of numerical model, including the micro-geometry has to be provided. Using a micro-structure in the contact surface, the normal contact stresses on the rough surface are computed. These stresses only exist in parts of the model which are in contact (s ) (Figure 2.8). Taking the summation of stresses (pNs ns ) over the contact areas will result in a resultant force on entire surface. Equation (2.61) shows the contact stress with respect to the basis (a m , nm ) of the contact surface (shown in Figure (2.8)).

Ns ns d = fN + fT = fN nm + fT a m

s s

(2.61)

Then the average of stresses can be taken over the contact plane (which in case shown in Figure (2.8) is straight)

pN m nm =

1 fN n m A

pN m =

1 fN A

tT m nm =

1 m fT a A

tT m =

1 fT A

A general procedure for the homogenized friction law is derived and reported by Haraldsson and Wriggers (2000) and Bandeira et al. (2002). This procedure is used by Haraldsson and Wriggers (2000) to derive a contact law for a concrete-soil interface. The simplest formulation for contact is a discretization which establishes contact interface equations on a nodal basis. A node-to-node contact can only be used for problems with linear geometry. This is due to the fact that large relative tangential movement of the nodes is not allowed in the contact area with assumption of small strains. In order to model the contact between two nodes at the same location, contact element should be dened with specic normal and tangential stiness in order to come up with contact forces. Considering the contact of the surfaces as shown in Figure (2.9), the normal contact force can only be zero or compressive. The normal contact constraints can be represented as:

dn = 0 dn > 0 dn Fn = 0

when when

Fn > 0 Fn = 0 (2.62)

Jeremi et al.

Lecture Notes

53

where dn is the relative displacement in the normal direction and Fn is the normal force. For frictional contact with Coulomb-type law, the tangential constraints can be written as:

dt = 0

when when

where dt is the relative displacement in the tangential direction and Ft is the tangential force at contact zone, and is the friction coecient at the interface. The most commonly used method to represent contact problems is the penalty method. Considering that kn and kt are normal and tangential stiness of the contact element, normal and tangential forces at the contact area can be expressed as:

F =k d n n n Ft = kt dt And in case of slipping in the contact zone, the tangential force is:

(2.64)

Ft = Fn

(2.65)

In theory, normal and tangential stiness should be large enough to prevent surfaces to penetrate and elastic sliding under compressive normal force. Naturally this cannot be satised in numerical implementation. Therefore, a nite surface penetration might occur (Sheng et al. (2007b)). It also should be mentioned that solving the frictional contact problems can be done using explicit or implicit scheme. In implicit method the obtained state of the contact (e.g. detaching or sliding) is used to update the solution of the current step of analysis while in explicit method it will be considered in next step.

2.8

2.8.1

Two Node, 3D, Rubber Isolator Finite Element

Kelly (1991a,b)

2.8.2

Jeremi et al.

Lecture Notes

54

a)

T S

b)

t p

Jeremi et al.

Chapter 3

(1994-2002-2010-)

55

Lecture Notes

56

Mindlin and Deresiewicz (1953) Hashin (1983) Rowe (1962) COSSERAT (2008)

3.4

Dilatancy

Jeremi et al.

Chapter 4

(1991-1994-2002-2006-2010-)

(In collaboration with Prof. Zhaohui Yang, Prof. Zhao Cheng, Prof. Mahdi Taiebat and Mr. Nima Tafazzoli)

57

Lecture Notes

58

4.1 4.2

kl

In linear elasticity the relationship between the stress tensor ij and the strain tensor the following form:

can be represented in

ij = (

ij )

(4.1)

ij )

ij =

(4.2)

The introduction of the strain energy density function into elasticity is due to Green, and elastic solids for which such a function is assumed to exist are called Green elastic or hyperelastic solids. Linearization of an elastic continuum is carried out with respect to a reference conguration which is stress free at temperature T0 , so that 0ij = 0. If we denote as Eijkl an isothermal modulus tensor, then under isothermal conditions, we obtain the generalized Hookes law: (4.3)

ij = Eijkl

kl

where Eijkl is the fourth order elastic stiness tensor with 81 independent components in total. The elastic stiness tensor features both minor symmetry Eijkl = Ejikl = Eijlk and major symmetry Eijkl = Eklij (Jeremi and Sture, 1997). The number of independent components for such elastic stiness tensor is 21 (Spencer, 1980). 2W ij kl 2W kl ij

Eijkl =

=

=0

(4.4)

=0

We will restrain our considerations to the isotropic case. The most general form of the isotropic tensor of rank 4 has the following representation: (4.5)

I4 = ij kl + ik jl + il jk

Jeremi et al.

Lecture Notes

59

If Eijkl has this form then in order to satisfy the symmetry condition2 Eijkl = Ejikl we must have = . The symmetry condition3 Eijkl = Eklji is then automatically satised. The elastic constant tensor has the following form: Eijkl = ij kl + (ik jl + il jk ) where and are the Lam coecients: = E (1 + ) (1 2 ) ; = E 2 (1 + ) (4.7) (4.6)

and E and are Youngs Modulus and Poissons ratio respectively. The symmetric part of the fourth order unit tensor is :

sym Iijkl =

1 (ik jl + il jl ) 2

(4.8)

and can be found as multiplier of in equation (4.6). Equation (4.6) can be written in terms of E and as: Eijkl = E 2 (1 + ) 2 ij kl + ik jl + il jk 1 2 (4.9)

The same relation in terms of bulk modulus K and shear modulus G is:

(4.10)

(4.11)

kl

kl

2 3

Jeremi et al.

Lecture Notes

60

where Dklpq is the elastic compliance fourth order tensor, dened as: Dklpq = 1 kl pq + (kp lq + kq lp ) 2 (3 + 2) 4 (4.13)

It is worthwhile noting that the part adjacent to the inverse of the bulk modulus K : (kl pq ) controls the volumetric response and that the part adjacent to the inverse of the shear modulus G:

1 1 kl pq + (kp lq + kq lp ) 3 2 controls the shear response! This note will prove useful later on. Linear transformation of the stress tensor pq into itself, i.e. ij is dened as: ij = Eijkl

kl

= Eijkl Dklpq pq

(4.16)

4.2.0.0.2

Linear elastic law is the simplest one and assumes constant Youngs modulus E and constant Poissons Ration .

4.2.0.0.3

This nonlinear model (Janbu, 1963), (Duncan and Chang, 1970) assumes dependence of the Youngs modulus on the minor principal stress 3 = min in the form E = Kpa 3 pa

n

(4.18)

Here, pa is the atmospheric pressure in the same units as E and stress. The two material constants K and n are constant for a given void ratio.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

61

4.2.0.0.4

If Youngs modulus and Poissons ratio are replaced by the shear modulus G and bulk modulus K the nonlinear elastic relationship can be expressed in terms of the normal eective mean stress p as G and/or K = AF (e, OCR)pn (4.19)

where e is the void ratio, OCR is the overconsolidation ratio and p = ii /3 is the mean eective stress (Hardin, 1978).

4.2.0.0.5

Lade and Nelson (1987) and Lade (1988a) proposed a nonlinear elastic model based on Hookes law in which Poisson ratio is kept constant. According to this model, Youngs modulus can be expressed in terms of a power law as: I1 pa

2

E = M pa

1+ + 6 1 2

J2D p2 a

(4.20)

where I1 = ii is the rst invariant of the stress tensor and J2D = (sij sij )/2 is the second invariant of the deviatoric stress tensor sij = ij kk ij /3. The parameter pa is atmospheric pressure expressed in the same unit as E , I1 and J2D and the modulus number M and the exponent are constant, dimensionless numbers.

4.3

4.3.1

Elastoplasticity

Constitutive Relations for Innitesimal Plasticity

A wide range of elastoplastic materials can be characterized by means of a set of incremental constitutive relations of the general form: d

ij

=d

e ij

+d

p ij

(4.21)

dij = Eijkl d

e kl

(4.22)

p ij

= d

Q = d mij (ij , q ) ij

(4.23)

dq = d h (ij , q )

Jeremi et al. University of California, Davis

(4.24)

Version: March 21, 2012, 17:15

e ij ij ),

Lecture Notes

p ij

62

ij 4

ij ,

and

of internal variables . The asterisk in the place of indices in q replaces n indices5 . Equation (4.21) expresses the commonly assumed additive decomposition of the innitesimal strain tensor into elastic and plastic parts. Equation (4.22) represents the generalized Hookes law6 which linearly relates stresses and elastic strains through a stiness modulus tensor Eijkl . Equation (4.23) expresses a generally associated or non-associated ow rule for the plastic strain and (4.24) describes a suitable set of hardening laws, which govern the evolution of the plastic variables. In these equations, mij is the plastic ow direction, h the plastic moduli and d is a plastic parameter to be determined with the aid of the loadingunloading criterion, which can be expressed in terms of the KarushKuhnTucker condition (Karush, 1939; Kuhn and Tucker, 1951) as: F (ij , q ) 0 (4.25)

d 0

(4.26)

F d = 0

(4.27)

In the previous equations F (ij , q ) denotes the yield function of the material and (4.25) characterizes the corresponding elastic domain, which is presumably convex. Along any process of loading, conditions (4.25), (4.26) and (4.27) must hold simultaneously. For F < 0, equation (4.27) yields d = 0, i.e. elastic behavior, while plastic ow is characterized by d > 0, which with (4.27) is possible only if the yield criterion is satised, i.e. F = 0. From the latter constraint, in the process of plastic loading the plastic consistency conditions7 is obtained in the form: dF = where : nij = F ij (4.29) F F dij + dq = nij dij + dq = 0 ij q (4.28)

F q

(4.30)

Equation (4.28) has the eect of conning the stress trajectory to the yield surface8 . It is worthwhile noting that nij and are normals to the yield surface in stress space and the plastic variable space respectively.

4 In

the simplest models of plasticity the internal variables are taken as either plastic strain components (W p ).

p ij

or the hardening

variables dened, for example as a function of inelastic (plastic) work, i.e. = f See Lubliner (1990) page 115. p 5 for example ij if the variable is ij , or nothing if the variable is a scalar value, i.e. . 6 also Eq. 4.3 7 rst order accuracy condition. 8 Since it is only linear expansion stress trajectory is conned to the tangential plane only.

Jeremi et al.

Lecture Notes

63

An interesting alternative way of representing nonassociated ow rules can be found in Runesson (1987). A ctitious plastic strain derived from associated ow rule, ep ij is introduced. This ctitious plastic strain is assumed to be related to the real plastic strain transformation: ep ij = Aijkl

p kl p ij ,

(4.31)

Linear transformation tensor Aijkl may be state dependent in general case, and it reduces to the symmetric part of the fourth order identity tensor10 for the case of associated plasticity. It is often of interest to model deviatoric strains by an associated ow rule while the volumetric part is non associated. For this case, Aijkl can be formulated as:

Aijkl =

1 1 (ij kl ) + (ik jl + il jk ) 3 2

(4.32)

1 A ijkl =

1 1 (ij kl ) + (ik jl + il jk ) 1+ 3 2

(4.33)

and it is obvious that the nonassociated ow rule is obtained with = 0 and the associated ow rule with = 0. It is useful to choose 0 and retain nice, positive denite properties of adjusted constitutive tensors later. Let the

2

ij

1 where Dijkl is the elastic compliance tensor ( Dijkl = Eijkl ), and let us introduce the adjusted complementary

A

ij

= ij (Aijkl Dklmn ) mn = ij

Dijmn mn

(4.35)

where ADijmn is the elastic compliance tensor transformed with respect to the nonassociativity involved. It is

sym clear that when Aijkl Iijkl = A ij 2

ij

4.3.2

On Integration Algorithms

In the section Constitutive Relations for Innitesimal Plasticity we have summarized constitutive equations that are capable of representing a wide variety of elastoplastic materials. The problem in Computational Elasto plasticity is to devise accurate and ecient algorithms for the integration of such constitutive relations. In the context of nite element analysis using isoparametric elements, the integration of constitutive equations is carried out at Gauss points. In each step the deformation increments are given or known, and the unknowns to be found are updated stresses and plastic variables. According to Ortiz and Popov (1985) an acceptable algorithm should satisfy:

10 A 9 as

in equation 4.23. sym ik jl + il jk . Iijkl 1 2 11 This norm will be reintroduced later on!

ijkl

Jeremi et al.

Lecture Notes

64

consistency with the constitutive relations to be integrated or rst order accuracy, Numerical stability, incremental plastic consistency A nonrequired but desirable feature to be added to the above list is: higher12 order accuracy First two conditions are needed for attaining convergence for the numerical solution as the step or increment becomes vanishingly small. The third condition is the algorithmic counterpart of the plastic consistency condition and requires that the state of stress computed from the algorithm be contained within the elastic domain.

4.3.3

A class of algorithms for integrating constitutive equations with potential to satisfy the above mentioned conditions are the Generalized Midpoint rule algorithms. They are given in the following form:

(4.36)

n p n+ d(n+1 p mij ij ) = d( ij ) + d

(4.37)

(4.38)

Fn+1 = 0 where:

n+

(4.39)

n+1

ij , (1 ) nq +

n+1

(4.40)

n+

h = h (1 ) nij +

n+1

ij , (1 ) nq +

n+1

(4.41)

It is quite clear that the case = 0 corresponds to the Forward Euler approach13 , the case = 1 corresponds to the Backward Euler approach14 , and the case = 1/2 to the Crank Nicholson scheme. Equations (4.36), (4.37), (4.38), (4.39), (4.40) and (4.41) are the nonlinear algebraic equations to be solved for the unknowns

n+1 d(n+1ij ), d(n+1 p q ) and d. From the Figure (4.1)15 it can be seen that the Generalized Midpoint ij ), d(

12 at

13 explicit

Jeremi et al.

Lecture Notes

n+ cross

65

m ij

m ij

predictor ij

n+1

m ij

n+1

Q=0

n+1 n+

cross ij

n+1

ij

ij

ij

ij

ij

2 1 3

F=0

n+

F=0 F=0

n+1

Figure 4.1: integration algorithms in elastoplasticity rule may be regarded as a returning mapping algorithm in which the elastic predictor updated yield surface along the ow direction evaluated at the midpoint (

n+ pred

ij is projected on the

ij ,

n+

q ).

4.3.3.1

Accuracy Analysis

Bearing in mind the context of the displacement based nite element analysis the integration of constitutive equations is performed for the given strain increment. The updated strains n+1 as the known function of the step size t. The remaining updated variables

ij

ij

n+1 p n+1 q , ij ,

n+1

ij ,

as well as n

ij ,

the incremental plastic parameter become functions of t implicitly dened through equations (4.36), (4.37), (4.38) and (4.39). It should be clear from (4.36), (4.37), (4.38) and (4.39) that as t 0 than and thus the limiting values of

n+1 n+1 ij

ij ,

n+1 p n+1 q ij ,

t0 t0 t0

lim

n+1

ij

= = = =

ij

lim

n+1 p ij n+1

n p ij n

lim

q (4.42)

t0

lim

15 it

should be pointed out that the vectors, as drawn on this gure, are pointing in the right direction only if we assume that

Eijkl Iijkl . For any general elasticity tensor Eijkl all vectors are dened in the Eijkl metric, so the term normal, as we are used to it, does not apply here.

Jeremi et al.

Lecture Notes

66

It can also be argued that, by virtue of the implicit function theorem ((Abraham et al., 1988) Chapter 2.5), n+1ij ,

n+1 p n+1 q ij ,

and are dierentiable functions of t, if the functions n+mij , n+h and F are suciently smooth.

4.3.3.1.1

First order accuracy16 of the algorithm, dened by the equations (4.36), (4.37), (4.38) and (4.39) with the constitutive equations given by (4.21), (4.22), (4.23) and (4.24) necessitates that the numerically integrated variables

n n+1

ij ,

n+1 p ij

and

n+1

n

p ij

(t + t) and q (t + t) to

n p ij

within second order terms in the Taylors expansion around the initial state ij = ij (t), q = q (t) in t. First order accuracy can be written in the following form:

p ij

(t) and

t0

lim

n+1

p d (n ij ) d n ij d (t) d (t)

(4.43)

t0

lim

n+1 p ij

d (t)

(4.44)

t0

lim

n+1

(4.45)

(4.46)

and the plastic parameter d (n) /d (t) is determined with the aid of the plastic consistency condition at t:

d (nF ) (nF ) dij (nF ) dq dij dq = + = nnij + n =0 d (t) ij d (t) q d (t) d (t) d (t)

(4.47)

It is now rather straightforward to check whether the Generalized Midpoint rule satises the consistency conditions as given by (4.43), (4.44), (4.45) and (4.46). We can proceed further on by dierentiating (4.36), (4.37), (4.38) and (4.39) with respect to t17 :

16 rst 17

order consistency. bearing in mind that values at t are constants and that only variables at t + t are changing with respect to t.

Jeremi et al.

Lecture Notes

67

ij = Eijkl d (t)

n+1

n+1

kl

d (t)

n+1 p kl

d (t)

(4.48)

d d d (t)

n+1 p ij

d (t)

n+

d d (t) d

n+1

n+

mij +

n+1

mij +

mij ij

n+1

ij mij + d (t) q

(4.49)

n+1

n+1 p q

d (t) d d (t)

n+

= d

d d (t)

n+1

n+

h +

n+1

h +

h ij

n+1

ij h + d (t) q

(4.50)

n+1

F n+1F d n+1ij n+1F d n+1n+1q = + n+1 =0 n +1 d (t) ( ij ) d (t) ( q ) d (t) mij and

n+

n+1

(4.51)

where

n+

By taking t to the limit value, t 0, in the (4.48), (4.49), (4.50) and (4.51) and using the relations from (4.42) one nds:

t0

lim

ij = Eijkl d (t)

n+1

d (n kl ) d d (t)

n+1 p kl t=0

d (t)

(4.52)

t0

lim

n+1 p ij

d (t)

Jeremi et al.

Lecture Notes

68

t0

lim

n+1

t0

lim

F (nF ) = d (t) ij

n+1

t0

lim

ij d (t)

n+1

(nF ) q

t0

lim

q d (t)

n+1

=0

(4.55)

In the previous equations it is quite clear that since t = 0, then equations (4.42) hold and since the variables

n

ij ,

n p ij

and n q are constant with respect to the change in t, the result follows readily, i.e. the Midpoint rule

4.3.3.1.2

To investigate second order accuracy of the algorithm given by (4.36), (4.37), (4.38) and (4.39) together with the constitutive equations given by (4.21), (4.22), (4.23) and (4.24) we shall proceed in the following manner. Second order accuracy actually means that the numerically integrated variables with their exact values ij (t + t),

n p ij n+1

ij ,

n+1 p ij

and

n+1

q agree

n p ij

expansion around the initial state ij = ij (t), be written in the following mathematical form:

p ij

t0

lim

d2

n+1

ij

2

d (t)

2

= (4.56)

Eijkl

t0

lim

d2

n+1

d (t)

kl 2

lim

d2 (n p kl ) d (t)

2

t0

= Eijkl

d2 (n kl ) d (t)

d2 (n p kl ) d (t)

2

t0

lim

d2

d (t)

n+1 p ij 2

d2 d (t)

2 t0

lim

n+1

mij + lim d2

2

2 n

(nmij ) +

Jeremi et al.

Lecture Notes

69

t0 2 lim d (t) t0

lim

d2

d (t)

n+1 p q 2

d2

n+1

h + lim d

2

t0

d n+1h = d t d (t)

n+1

d (t) d2 d (t)

2

(nh ) +

n+1

(nh ) +

d t

n+1

t0

lim

d2 () d (t)

2

d2 (n ) d (t)

2

(4.59)

2

and the plastic parameter d2 (n ) /d (t) is determined with the aid of the second order oscillatory satisfaction

d2 (nF ) d (t)

2

dnij dt

+

n

d ( ) d t

(4.60)

Now we can proceed by taking the second derivative of the equations (4.36), (4.37), (4.38) and (4.39) or use the already derived rst derivatives from equations (4.48), (4.49), (4.50) and (4.51), and then dierentiate them again so that we get:

d2

n+1

ij

2

d (t)

= Eijkl

d2

n+1

d (t)

kl 2

d2

d (t)

n+1 p kl 2

(4.61)

n+1 2

d (t)

n+

n+1 p ij 2

mij + + (4.62)

ij mij + d (t) q d

n+1

n+1

d

n+1

q d (t) d q d (t)

n+1

n+1

mij ij

n+1

ij mij + d (t) q

n+1

Jeremi et al.

Lecture Notes

70

n+1 2

d (t)

n+

n+1 p q 2

h + + (4.63)

ij h + d (t) q d

n+1

n+1

d

n+1

q d (t) d q d (t)

n+1

n+1

h ij

n+1

ij h + d (t) q

n+1

d2 d

n+1

n+1

F

2

d (t)

nij d

dij + d

n+1

n+1

n+1

dij

(4.64)

If we drive t to the limit, namely by taking limt0 and keeping in mind equations (4.42) and the assumed consistency of the algorithm18 as given by the equations (4.43), (4.44), (4.45) and (4.46) one nds:

t0

lim

d2

n+1

ij

2

d (t)

= Eijkl

d2 (n kl ) d (t)

2

lim

d2

t0

d (t)

n+1 p kl 2

(4.65)

t0

lim

d2

d (t)

n+1 p ij 2

t0

lim

d2

n+1

n+

d (t)

mij + 2

d (n ) d (t)

mij ij

18 actually

Jeremi et al.

Lecture Notes

71

t0

lim

d2

d (t)

n+1 p q 2

t0

lim

d2

n+1

n+

d (t)

h + 2

d (n ) d (t)

h ij

d (nij ) h + d (t) q

t0

lim

d2

n+1

F

2

d (t)

d2 n+1ij d (n nij ) d (nij ) n + + nij lim 2 t0 dij d (t) d (t) + d2 n+1q d (n ) d (nq ) n + lim =0 t0 d (t)2 dij d (t) (4.68)

By comparing equations (4.65), (4.66), (4.67) and (4.68) with the second order accuracy condition stated in equations (4.56), (4.57), (4.58) and (4.59) it is quite clear that the second order accuracy is obtained i19 = 1/2 ! The conclusion is that the Midpointrule algorithm is consistent20 for all [0, 1] and it is second order accurate for = 1/2. However, one should not forget that these results are obtained for the limiting case t 0, i.e. the strain increments are small and tend to zero.

4.3.3.2

Numerical stability of an algorithm plays a central role in approximation theory for initial value problems. In fact, it can be stated that consistency and stability are necessary and sucient conditions for convergence of an algorithm as the time step tends to zero. In the approach presented by Ortiz and Popov (1985) a new methodology is proposed by which the stability properties of an integration algorithm for elastoplastic constitutive relations can be established. Our attention is conned to perfect plasticity and a smooth yield surface. The purpose of the following stability analysis is to determine under what conditions a nite perturbation in the initial stresses is diluted by the algorithm. In other words:

d

19

n (2) n (1) ij , ij

(4.69)

20 rst

Jeremi et al.

Lecture Notes

72

(2) n+1 (1) ij and n+1ij are two sets of updated n (2) ij , respectively, and all of the previous stress

where d (, ) is some suitable distance on the yield surface and stresses corresponding to arbitrary initial stress values nij and

(1)

values are assumed to lie on the yield surface. Stability in the sense of equation (4.69) is referred to as large scale stability. It is shown in Helgason (1978)21 that for nonlinear initial value problems dened on Banach manifolds, consistency and large scale stability with respect to a complete metric are sucient for convergence. The task of directly establishing estimates of the type expressed in (4.69) is rather dicult, and so despite the conceptual appeal of large scale stability, simplied solutions are sought. It should be recognized that attention can be restricted to innitesimal perturbation in the initial conditions of the type nij nij + d (nij ). This simplication is founded on the fact that the dilution or attenuation, by the algorithm of innitesimal perturbations:

n+1

ij d nij

(4.70)

with respect to some suitable norm , of small scale stability, implies large scale stability in the sense of equation (4.69).

Let the

ij

= ij Dijkl kl

(4.71)

1 where Dijkl is the elastic compliance tensor (Dijkl = Eijkl ), and let the distance d (, ) on the yield surface be

dened as

d ij , ij

(1)

(2)

= inf

ij (s) ds

(4.72)

where the innum is taken over all possible stress paths on the yield surface that are joining two stress states, namely ij and ij . It can be found in Helgason (1978) that for a smooth yield surface, equation (4.72) denes the geodesic distance which endows the yield surface with a complete metric structure. Suppose that we have any two initial states of stress nij and nij and let

(1) (2) n+1 (1) ij (1) (2)

and

(1)

n+1 (2) ij

be the corre-

sponding updated values, respectively, and all the previous stress states are assumed to lie on the yield surface. Then, according to Helgason (1978), there exists a unique geodesic curve that joins nij and nij for which the innum in equation (4.72) is attained. If n is such a curve, then by denition:

(2)

d

21 the

n (1) n (2) ij , ij

=

n

ij (s) ds

(4.73)

Jeremi et al.

Lecture Notes

73

Let the new curve n+1 be the transform of curve n by the algorithm. By denition n+1 lies on the yield surface and joins two stress states

n+1 (1) ij

and

n+1 (2) ij .

=

n+1

ij (s) ds

(4.74)

Under the assumption of small scale stability of the algorithm one can write:

ij (sn+1 ) ds = dij (sn+1 ) dij (sn ) = ij (sn ) ds for every pair of corresponding points sn and sn+1 on n and n+1 respectively, so it follows:

(4.75)

ij (sn+1 ) ds

n+1 n

ij (sn ) ds

(4.76)

n (1) n (2) ij , ij

(4.77)

which proves large scale stability. The main conclusion of the above argument may be stated as follows: small scale stability in the energy norm is equivalent to large scale stability in the associated geodesic distance. The previous result is of practical importance, since it shows that the stability analysis for the integration algorithm in elastoplasticity can be carried out by the assessment of small scale stability. The small scale stability analysis of the Generalized Midpoint rule is necessary to determine how the algorithm propagates innitesimal perturbations in the initial conditions. By dierentiating equations (4.36), (4.37), (4.38) and (4.39)and considering that we are dealing with perfectly plastic case here so that

n+1

n+1

ij = Eijkl d

n+1 p kl

(4.78)

d (n ij ) = Eijkl d (n p kl )

(4.79)

n+1 p ij

n p ij

=d

n+

mij + d

n+

mij

(4.80)

Jeremi et al.

Lecture Notes

74

n+1

F =

F ij

d

n+1

n+1

ij = n+1 nij d

n+1

ij = 0

(4.81)

Let us now examine the shape of d (n+mij ) having in mind the original denition22 given in equation (4.40):

n+

n+1

ij , (1 ) nq +

n+1

n+

mij = (1 )

mij kl

d (nkl ) +

n+

mij kl

d

n+

n+1

kl

Mijkl =

mij kl

d d

n+

n+1 p ij

n p ij n+1

mij + (1 )

n+

Mijkl d (nkl ) +

n+

Mijkl d

kl (4.82)

1 By using equations (4.78) and (4.79) and knowing that Eijkl = Dijkl one can write:

n+1 p ij

= Dijkl d

n+1

kl

22

the remark about restraining analysis to perfectly plastic case still holds, so that

n+1q

Jeremi et al.

Lecture Notes

75

n p ij

= Dijkl d(nkl )

Dijkl d d

n+

n+1

kl + Dijkl d (nkl ) =

n+

mij + (1 )

n+

Mijkl d (nkl ) +

Mijkl d

n+1

kl

n+1

kl :

Dijkl + Dijkl (1 )

n+

n+

Mijkl

n+1

kl = mij

Mijkl

d (nkl ) d

n+

and by denoting :

ijkl = Dijkl (1 )

n+

Mijkl

ijkl = Dijkl +

n+

Mijkl

it follows:

n+1

1 n kl = ijkl ijkl d ( kl ) d

n+

mij

(4.83)

Jeremi et al.

n+1

University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

76

n+1

F = n+1 nkl d

n+1

kl = 0

one gets:

d

n+1 1 n nkl ijkl ijkl d ( kl ) d n+

n+1

F = =0 (4.84)

mij

(4.85)

or23 :

n+1

d =

(4.86)

n+1

kl from (4.83) and the solution for d from (4.86) one can nd:

n+1

n+1

n+m

d (nrs ) (4.87)

n+1

n+1

1 kl = ijkl ijkl

ks rl

n+m

d (nrs )

(4.88)

23

Jeremi et al.

Lecture Notes

77

n+1

klrs = ks rl

n+m

(4.89)

n+1

(4.90)

In order to derive the estimate of the type (4.70) from (4.90) we shall proceed in the following way. The norm of a tensor is dened as:

Aijkl = sup

Aijkl kl kl

(4.91)

Aijkl kl Aijkl

kl

Bijkl

(4.92)

it follows:

n+1

kl

(4.93)

n+1

kl

1 ijkl ijkl

klrs

d (nrs )

(4.94)

Jeremi et al.

Lecture Notes

78

1 n+ mij ijkl

klrs

n+1

1 (klrs ) ijkl

n+

mij =

(4.95)

(klrs ) (rs ) = rs

(4.96)

n+1

nrs . From these properties and the denition in equation (4.91) it follows

klrs 1

(4.97)

is symmetric and positive denite everywhere on the yield surface. The assumption is valid, if the ow direction mij is derived from the convex potential function, which is a rather common feature among yield criteria. It is now clear that :

1 ijkl ijkl =

max

ij max ij

(4.98)

where

max

(ijkl ijkl ) kl = 0

(4.99)

Jeremi et al.

Lecture Notes

79

max

ij

max

ij = maxij Dijkl

max

kl = 1

(4.100)

If we denote:

n+

= maxij

n+

Mijkl

max

kl

(4.101)

n+

from equations (4.98), (4.100), (4.101) and from the denition25 of ijkl and ijkl , it follows:

1 ijkl ijkl =

1 (1 ) (n+ ) 1 + (n+ )

(4.102)

n+1

kl

1 (1 ) (n+ ) 1 + (n+ )

d (nrs )

(4.103)

n+

1 (1 ) (n+ ) 1 1 + (n+ )

n+

n+

(4.104)

n+1

kl

d (nrs )

(4.105)

24 because n+M 25

ijkl is derived from a convex potential function. ijkl = Dijkl (1 ) n+Mijkl and ijkl = Dijkl +

n+M

ijkl

Jeremi et al.

Lecture Notes

80

1 1

(4.106)

min =

1 2

(4.107)

The conclusion is that the Generalized Midpoint rule is unconditionally stable for 1/2. In the case when < 1/2 the Generalized Midpoint rule is only conditionally stable. To obtain a stability condition for 1/2 one has to go back to equation (4.103), and we conclude that:

1 (1 ) (n+ ) 1 1 + (n+ )

max

1 2 for (1 2) 2

(4.108)

critical

4.3.4

Midpoint rule algorithms in computational elastoplasticity require27 the evaluation of the intersection28 stress. Despite the appeal of the closed form solution, as found in Biani (1989), and numerical iterative procedures as found in Marques (1984) and Nayak and Zienkiewicz (1972), for some yield criteria29 the solution is not that simple to nd. Special problems arises, even with the numerical iterative methods in the area of a apex. The apex area problems are connected to the derivatives of yield a function. Having in mind the before mentioned problems, a dierent numerical scheme, that does not need derivatives, was sought for solving this problem. One possible solution was found in Press et al. (1988b) in the form of an excellent algorithm that combines root bracketing, bisection, and inverse quadratic interpolation to converge from a neighborhood of a zero crossing. The algorithm was developed in the 1960s by van Wijngaarden, Dekker and others at the Mathematical Center in Amsterdam. The algorithm was later improved by Brent, and so it is better known as Brents method. The method is guaranteed to converge, so long as the function30 can be

26

that is d n+1ij d nij except for the fully implicit Backward Euler algorithm. 28 contact, penetration point, i.e the point along the stress path where F = 0 or the point where stress state crosses from the

27

elastic to the plastic region. 29 namely for the MRS-Lade elastoplastic model. 30 in our case yield function F ( ). ij

Jeremi et al.

Lecture Notes

81

F(

contact

ij)=0

F( ij )

predictor

ij ij F=0

predictor

F( ij)

1 elastic region

3

Figure 4.2:

which must be resolved for the Forward and Midpoint schemes evaluated within the initial interval known to contain a root. While the other iterative methods that do not require derivatives31 assume approximately linear behavior between two prior estimates, inverse quadratic interpolation uses three prior points to t an inverse quadratic function32 , whose value at y = 0 is taken as the next estimate of the root x. Lagranges classical formula for interpolating the polynomial of degree N 1 through N points y1 = f (x1 ), y2 = f (x2 ), . . . y3 = f (x3 ) is given by:

P ( x) = + +

(4.109)

31

32 x

Jeremi et al.

Lecture Notes

82

If the three point pairs are [a, f (a)], [b, f (b)], [c, f (c)], then the interpolating formula (4.109) yields:

x= + +

(y f (a)) (y f (b)) c+ (f (c) f (a)) (f (c) f (b)) (y f (b)) (y f (c)) a+ (f (a) f (b)) (f (a) f (c)) (y f (c)) (y f (a)) b (f (b) f (a)) (f (b) f (a))

(4.110)

By setting y = 0, we obtain a result for the next root estimate, which can be written as:

x=b+

f (b) f (a)

f (a) f (c)

f (a) f (c)

(c b) 1

f (b) f (c)

f (b) f (c)

(b a) (4.111)

f (b) f (a)

In practice b is the current best estimate of the root and the term:

f (b) f (a)

f (a) f (c)

f (a) f (c)

( c b) 1

f (b) f (c)

f (b) f (c)

(b a)

f (b) f (a)

is a correction. Quadratic methods33 work well only when the function behaves smoothly. However, they run serious risk of giving bad estimates of the next root or causing oating point overows, if divided by a small number

f (a) 1 f (c)

f (b) 1 f (c)

f (b) 1 f (a)

Brents method prevents against this problem by maintaining brackets on the root and checking where the interpolation would land before carrying out the division. When the correction of type (4.112) would not land within bounds, or when the bounds are not collapsing rapidly enough, the algorithm takes a bisection step. Thus, Brents method combines the sureness of bisection with the speed of a higher order method when appropriate.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

83

cone

ij

m ij

cap

predictor

cone

nij

nij =capm

ij

corner

ij

Qcone =0 =0 Fcone

n+1

ij

start

Figure 4.3: Pictorial representation of the corner point problem in computational elastoplasticity: Yield surfaces with singular points

4.3.5

4.3.5.1

Corner Problem

Some yield criteria are dened with more that one yield surface34 . We will restrict our attention to a twosurface yield criterion35 . Koiter has shown in Koiter (1960) and Koiter (1953) that in the case when two yield surfaces are active, the plastic strain rate from equation (4.23) can be derived as follows:

d where

p ij

= dcone

cone

cap

cap

mij (ij , q )

(4.112)

cone

mij (ij , q ) are normals to the potential functions at a corner, which belongs to

the yield functions that are active, i.e. Fcone and Fcap . We now observe that we have two nonnegative plastic multipliers dcone and dcap instead of one. We must require that at the end of the loading step36 , neither of

33

Newtons method for example. example MRS-Lade yield criterion has two surfaces. 35 having in mind MRS-Lade cone-cap yield criterion. 36 after stress correction, i.e. return to the yield surface(s).

34 for

Jeremi et al.

Lecture Notes

84

the two yield functions is violated. These multipliers dcone and dcap can be determined from the conditions:

Fcone

n+1

ij , n+1q = 0

(4.113)

Fcap

n+1

ij , n+1q = 0

(4.114)

noting that by virtue of equation (4.112) we have at the corner singular point:

n+1

cone

cap

mkl

(4.115)

4.3.5.2

Apex Problem

Fcone=0

ij

predictor n

ij

p

n+1

ij

Figure 4.4:

The pictorial representation of the apex point problem in computational elastoplasticity: Yield

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

85

The apex problem, as depicted in Figure (4.4) is solved in an empirical fashion. Rather than facing the complexity of solving a complex dierential geometry problem37 the stress point that is situated in the gray apex region is immediately returned to the apex point. In the case when the hardening rule for the cone portion has developed to the stage that it aects the size of that cone portion of the yield criterion and not the position of intersection with the hydrostatic axis, then all stress returns from any part of apex gray region will be to the apex point itself. This strategy was used by Criseld (1987). Nevertheless, the problem of integrating the rate equations in the apex gray region is readily solvable for the piecewise at yield criteria38 by using Koiters conditions as found in Koiter (1960) and Koiter (1953). The apex problem for yield criteria that are smooth and dierentiable everywhere except at the apex point, is solvable by means of dierential geometry. Further work is needed for solving the problem, when the yield surface is not piecewise at in the apex vicinity.

4.3.5.3

q q

([p,q] v [p,q])

+90o p

tran

q

tran

p 2 p p q

1

p

3

p 1 p0 pcap pcap p4 p q4

Figure 4.5: Inuence regions in the meridian plane for the cone/cap surface of the MRS-Lade material model. In order to dene which surface is active and which is not for the current state of stress, a simple two dimensional analysis will be conducted. The fortunate fact for the MRS-Lade material model is that such an analysis can be

37 using

Koiters work described in Koiter (1960) and Koiter (1953) and the fact that the sum d

p ij

p ij

close to, but in the vicinity of the apex point. 38 Mohr - Coulomb for example.

Jeremi et al.

Lecture Notes

86

conducted in the p q meridian plane, only, i.e. the value for can be frozen. The concept is to calculate the stress invariants p, q and for the current state of stress39 , calculate the position of the apex and corner points in p q space for given the , calculate the two dimensional gradients at these points, perform linear transformation of the current stress state40 to the new coordinate systems, and then check for the values of p i , i = 1, 2, 3, 4, where p i is the transformed pi axis. The angle is dened as the angle between the p axis and the tangent to the potential function. The gradients to the cone portion of the potential surface are dened as:

Qcone = n cone p

q Qcone = g () 1 + q qa

g () m q 1 + + qa

q qa

1+m

The gradients of the cap portion of the yield/potential surface is dened as:

Qcap 2 ( p pm ) = p p2 r

Qcap = q

2 g ()2 q 1 + fr 2

q qa

2m

2 g ()2 m q 2 1 + + fr 2 qa

q qa

1+2 m

Q p Q q

(4.116)

39 by 40 now

using equations (4.158) and (4.159) as dened in section (4.5.4). in p, q and space.

Jeremi et al.

Lecture Notes

87

= arctan

Q p Q q

90 (4.117)

Care must be exercised with regard to which potential function is to be used in angle calculations. It should be mentioned that for the cap portion, the angle at the corner is = 0 , while at the tip of the cap, the angle is = 90 . If a new denition, as found in Ferrer (1992), is used for the cone potential function, where n is variable and n 0 as p pcap , then the corner gray region is empty. The linear transformation41 between coordinate systems p q and p q is dened as:

p q

cos sin

sin cos

p q

tran tran

p q

(4.118)

and by using that linear transformation, one can check the region where our current stress state, in p, q and space, belongs. Figure (4.5) depicts the transformation scheme and the new coordinate systems at three important points42 .

4.4

n The explicit algorithm (Forward Euler) is based on using the starting point (the state stress ij and internal n on the yield surface) for nding all the relevant derivatives and variables. variable space q n n The Explicit algorithm can be derived by starting from a rst order Taylor expansion about starting point (ij , q ):

F new

= F old + =

n

nmn

F F d (nmn ) + mn n q dmn + h d = 0

dq =

n

(4.119)

pq ) fE

mn = Emnpq d (

pq )

pq )

p pq

Emnpq d

p pq

= Emnpq d (

Emnpq d (crossmpq )

41 translation 42 at

and rotation. the apex point, corner point and the cap tip point.

Jeremi et al.

Lecture Notes

88

n

nmn Emnpq d

pq

n

crosn

With this solution for d one can obtain the increments in stress tensor and internal variables as

n

dmn = Emnpq d

pq

Emnpq

nn

ab

(4.120)

dqA =

crosn

hA

(4.121)

where n() denotes the starting elasticplastic point for that increment. It should be noted that the explicit algorithm performs only one step of the computation and does not check on the equilibrium of the obtained solutions. This usually results in the slow drift of the stressinternal variable point from the yield surface for monotonic loading. It also results in spurious plastic deformations during elastic unloading for cyclic loadingunloading.

4.4.0.3.1

ep ) is obtained from the explicit (forward Euler) integration The continuum tangent stiness tensor (contEpqmn

cont ep Epqmn

= Epqmn

nn

(4.122)

ep ep It is important to note that continuum tangent stiness (contEpqmn ) posses minor symmetries (contEpqmn = cont ep Eqpmn ep ep ep = contEpqnm ), while major symmetry (contEpqmn = contEmnpq ), is only retained for associated elastic

4.5

In previous sections, the general theory of elastoplasticity was presented. The accuracy and stability for the general Midpoint rule algorithm has been shown. In this chapter, the focus is on the Backward Euler algorithm, which is derived from the general Midpoint algorithm by setting = 1. The advantage of the Backward Euler

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

89

scheme over other midpoint schemes is that the solution is sought by using the normal43 at the nal stress state. By implicitly assuming that such a stress state exists, the Backward Euler scheme is guaranteed to provide a solution, despite the size of the strain step44 . However, it was shown in section (4.3.3.1) that the Backward Euler algorithm is only accurate to the rst order. The full implicit Backward Euler algorithm is based on the equation:

n+1

ij = predij Eijkl

kl

n+1

mkl

(4.123)

Q kl

is the elastic trial stress state, Q is the plastic potential function and n+1mkl =

n+1

is the gradient to the plastic potential function in the stress space at the nal stress position, and

pred

ij = nij + Eijkl

pred

kl

(4.124)

is the elastic predicted (trial) stress state. An initial estimate for the stress surface.

n+1

does not satisfy the yield condition, so some kind of iterative scheme is necessary to return the stress to the yield

4.5.1

pred

45

ij is not in a corner or apex gray regions, a single vector return to the yield surface is

possible. In order to derive such a scheme for a single vector return algorithm, a tensor of residuals rij will be :

pred

rij = ij

ij Eijkl mkl

(4.125)

This tensor represents the dierence between the current stress state ij and the Backward Euler stress state

pred

ij Eijkl mkl .

pred

ij is kept xed during the iteration process. The rst order Taylor series expansion can

new

new

old

mkl mn dmn

where dij is the change in ij , d() is the change in , and The goal is let

new

mkl qA dqA

43 m = Q/ ij ij 44 large strain step increments 45 By

(4.128)

were tested, the scheme converged to the solution even for deviatoric strain steps of 20% in magnitude. default at increment n + 1, and n+1 () is omitted for simplicity.

Jeremi et al.

Lecture Notes

90

rA will be dened as: rA = qA (nqA + hA ) and nqA is kept xed during iteration, that = oldrA + dqA d() hA From equation 4.127 and 4.130, one obtains

mkl s Iijmn + Eijkl mn

(4.129)

hA hA dij + dqB ij qB

(4.130)

kl Eijkl m qA

(4.131)

From equations 4.131 and 4.132, d() = 1 mkl mkl oldr s E Iijmn + Eijkl ijkl ij qA mn B hA hA oldrA AB qB ij 1 mkl mkl E m s E Iijmn + Eijkl ijkl ijkl kl qA mn hA hA AB hA

ij qB

old

nmn

(4.133)

nmn

The iterative procedure is continued until the yield criterion f = 0, rij = , and rA = are satised within some tolerances at the nal stress state

46

mkl s Iijmn + Eijkl mn hA ij

kl Eijkl m qA A AB h qB

1 (4.135)

plays an important role in the implicit algorithm. It should be mentioned here that the above denition is a simplied expression for very general model with various isotropic and kinematic hardening. Specically, if there is no hardening, C=

46

(4.136)

Jeremi et al.

Lecture Notes

91

mkl s Iijmn + Eijkl mn h ij

kl Eijkl m q

1 (4.137)

1 h q

mkl s + Eijkl Iijmn mn

mn h ij

1 (4.138)

mkl s Iijmn + Eijkl mn h ij

mn h ij kl Eijkl m q

1 (4.139)

1 h q

mn h q

mn h ij

C=

z

z

z

1 (4.140)

If we dene n mn n= B

(4.141)

E m ijkl kl m= hA

(4.142)

old

r=

old

ij

old

(4.143)

rA

old

d() = and

f nT C old r nT C M

(4.144)

d mn = C dqB

Jeremi et al.

old

r + d()m

(4.145)

Lecture Notes

92

4.5.2

Some remarks are necessary in order to clarify the Backward Euler Algorithm. It is a well known fact that the rate of convergence of the Newton - Raphson Method , or even obtaining convergence at all, is closely tied to the starting point for the iterative procedure. Bad initial or starting points might lead our algorithm to an oscillating solution, i.e. the algorithm does not converge. In the following, starting points for the Newton - Raphson iterative procedure will be established for one and twovector return algorithms.

4.5.2.1

pred

One of the proposed starting points (Criseld, 1991) uses the normal at the elastic trial point47 order Taylor expansion about point

pred new pred

ij . A rst

ij yields: d

pred pred

= =

pred old

+ +

F mn

pred

mn +

F qA

dqA =

pred

pred old

nmn dmn + A hA d = 0

(4.146)

kl

kl

pred

ij , i.e.

pred

ij =

so that any further stress "relaxation" toward the yield surface takes place under zero total strain

kl

condition d

pred

pred

mn = Emnpq d

pred p pq

pq

pred p pq pred

= mpq

= Emnpq d

= Emnpq d

pred old

pred

and it follows:

pred old

d =

predn

mn

Emnpq

predm

pq

A hA

With this solution for d we can obtain the starting point for the Newton-Raphson iterative procedure

pred old start

mn = Emnpq

pred

pq

Emnpq

pred pq

predn

mn

Emnpq

predm

A hA

mpq

(4.147)

47 I

Jeremi et al.

Lecture Notes

93

This starting point in six dimensional stress space will in general not satisfy the yield condition F = 0, but it will provide a good initial guess for the upcoming Newton-Raphson iterative procedure. It should be mentioned, however, that this scheme for returning to the yield surface is the well known Radial Return Algorithm , if the yield criterion under consideration is of the von Mises type. In the special case the normal at the elastic trial point

pred

n+1

ij , the return is

exact, i.e. the yield condition is satised in one step. Another possible and readily available starting point can be obtained by applying one Forward Euler step48 . To be able to use the Forward Euler integration scheme, an intersection point has to be found. The procedure for calculating intersection points is given in section (4.3.4). A rst order Taylor expansion about intersection point F new = F old + =

cross

ij yields: dqA =

cross

(4.148)

pq ) fE

mn = Emnpq d (

pq )

pq )

p pq

Emnpq d

p pq

= Emnpq d (

Emnpq d (crossmpq )

cross

pq

and it follows

cross

d =

crosn

mn

With this solution for d we can obtain the starting point for the Newton-Raphson iterative procedure

cross start

mn = Emnpq d

pq

Emnpq

crossn

ab

cross

mpq

(4.149)

48 or

more steps for really large strain increments, for example over 10% in deviatoric direction. What has actually been done is to

divide the region into several parts and depending on the curvature of the yield surface in deviatoric plane, use dierent schemes and dierent number of subincrements ( the more curved, the more subincrements) to get the rst, good initial guess. In the region around = 0, one step of the semi Backward Euler scheme is appropriate, but close to = /3 the Forward Euler subincrementation works better.

Jeremi et al.

Lecture Notes

94

This starting point in sixdimensional stress space will again not satisfy the yield condition49 F = 0, but will provide a good initial estimate for the upcoming Newton-Raphson iterative procedure.

4.5.3

The nal goal in deriving the Backward Euler scheme for integration of elastoplastic constitutive equations is to use that scheme in nite element computations. If the Newton Raphson iterative scheme is used at the global

ep equilibrium level then the use of the so called traditional tangent stiness tensor50 Eijkl destroys the quadratic

rate of asymptotic convergence of the iterative scheme. In order to preserve such a quadratic rate, a consistent, also called algorithmic, tangent stiness tensor is derived. The consistent tangent stiness tensor make use of derivatives of direction51 normal to the potential function, and they are derived at the nal, nal at each iteration, that converges to the nal stress point on the yield surface, stress point. The traditional forward scheme has a constant derivative, mij that is evaluated at the intersection point. It appears that Simo and Taylor (1985) and Runesson and Samuelsson (1985) have rst derived the consistent

tangent stiness tensor. Other interesting articles on the subject can be found in Simo and Taylor (1986), Simo and Govindjee (1988), Jetteur (1986), Braudel et al. (1986), Criseld (1987), Ramm and Matzenmiller (1988) and Mitchell and Owen (1988). As a consequence of consistency, the use of the consistent tangent stiness tensor signicantly improves the convergence characteristics of the overall equilibrium iterations, if a Newton - Raphson scheme is used for the latter. Use of the consistent tangent stiness tensor yields a quadratic convergence rate of Newton - Raphson equilibrium iterations. In what follows, two derivations are given, namely the consistent tangent stiness tensor for single and twovector return algorithms. The concept of consistent linearization was introduced by Hughes and Pister (1978), while detailed explanation is given by Simo and Hughes (1998). The consistent tangent stiness leads to quadratic convergence rates at global level. It should be mentioned that there are various equivalent forms of consistent tangent stiness depending on the specic implicit algorithm equations. For instance, Simo and Hughes (1998), and Belytschko et al. (2001) derived the consistent tangent stiness by taking current plastic strain as unknown and seeking its derivatives in the stress space; Prez-Foguet and Huerta and Prez-Foguet et al. (2000) used the numerical dierentiation to calculate the consistent tangent stiness in a compact matrix-vector form; Choi (2004) adopted the compact matrix-vector form by Prez-Foguet and Huerta and Prez-Foguet et al. (2000) but taking current plastic strain as unknown and seeking its derivatives in the elastic strain space. Slightly dierent from the above strategies, in this work the implicit algorithm is adopting the traditional form but taking current stress as unknown and seeking its derivatives in the stress space. Provided these dierences, the consistent tangent stiness in this work is slightly dierent from those in the above work.

49 except

for the yield criteria that have at yield surfaces ( in the stress invariant space) so that the rst order Taylor linear

expansion, is exact. 50 the one obtained with the Forward Euler method, i.e. where parameter = 0. 51 m = Q/ , i.e. m / 2 ij ij ij kl = Q/ij kl .

Jeremi et al.

Lecture Notes

95

4.5.3.1

In implicit algorithm, a very important advantage is that it may lead to consistent (algorithmic) tangent stiness (Equation 4.157). The concept of consistent linearization was introduced in Hughes and Pister (1978), more details on consistent tangent stiness were explained in Simo and Hughes (1998). The consistent tangent stiness leads to quadratic convergence rates at global level. It should be mentioned that there are various equivalent forms of consistent tangent stiness depending on the specic implicit algorithm equations. For instance, Simo and Hughes (1998), and Belytschko et al. (2001) derived the consistent tangent stiness by taking current plastic strain as unknown and seeking its derivatives in the stress space; Prez-Foguet and Huerta and Prez-Foguet et al. (2000) used the numerical dierentiation to calculate the consistent tangent stiness in a compact matrix-vector form; Choi (2004) adopted the compact matrix-vector form by Prez-Foguet and Huerta and Prez-Foguet et al. (2000) but taking current plastic strain as unknown and seeking its derivatives in the elastic strain space. Slightly dierent from the above strategies, in this work (section 4.5) the implicit algorithm is adopting the traditional form but taking current stress as unknown and seeking its derivatives in the stress space. Provided these dierences, the consistent tangent stiness in this work is slightly dierent from those in the above work. The detail derivation will be followed.

When seeking the algorithmic tangent stiness, we look into the explicit expression of dij /d the time step when seeking the algorithmic tangent stiness. Linearize Equation 4.123, one obtains dij = Eijkl d

pred kl

pred mn .

At the same

time, the internal variables are initialized the values at the previous time step, in other words, they are xed within

(4.150)

Similarly, linearize Equation 4.128, one obtains dqA = d() hA + hA hA dij + dqB ij qB (4.151)

mkl s Iijmn + Eijkl mn

kl Eijkl m qA

(4.152)

If one use the denitions of 4.135, 4.142 and 4.141, Equation 4.152 can be simplied to pred d E d mn ijkl kl C 1 + d()m = dqB 0 Linearize the yield function f (ij , qA ) = 0, one obtains nmn dmn + B dqB = 0

Jeremi et al. University of California, Davis

(4.153)

(4.154)

Version: March 21, 2012, 17:15

Lecture Notes

96

or in a simplied form d mn nT =0 dqB From Equations 4.153 and 4.155, one obtain nT C Eijmn d d() = T n Cm 0

pred mn

(4.155)

(4.156)

Substitute expression 4.156 into 4.153, one obtains d ij = dqA CmnT C C T n Cm E ijmn d 0

pred mn

pred mn

(4.157)

This equation gives the explicit expression of the consistent tangent stiness dij /d

From section 4.5, if there are interactions between internal variables, the implicit algorithm will become very

complicated. Simple models (e.g. Von-Mises model, or sometimes termed as J2 model) have been proved ecient and good performance by the implicit algorithm (Simo and Hughes, 1998). Evidently, the implicit algorithm is mathematically based on the Newton-Raphson nonlinear equation solving method as well as the Eulerian backward integration method. Theoretically, the Newton-Raphson method may have quadratic convergence rate. However, Newton-Raphson method is not unconditional stable, and sometimes the iteration will diverge (Press et al., 1988a). Any bad starting point, non-continuous derivatives around solution, high nonlinearity, and interactions between internal variables, will deteriorate the implicit algorithm performance. A complicated model cannot guarantee good performance or quadratic convergence by the implicit algorithm Criseld (1997). The task to obtain the analytical expressions (Equations 4.136 to 4.140) may prove exceeding laborious for complicated plasticity models Simo and Hughes (1998).

4.5.4

In the derivation of the Backward Euler algorithm and the Consistent Tangent Matrix it is necessary to derive the rst and the second derivatives of the potential function. The function Q is the function of the stress tensor ij and the plastic variable tensor qA . Derivatives with respect to the stress tensor ij and plastic variable tensor qA are given here. It is assumed that any stress state can be represented with the three stress invariants p, q and given in the following form:

1 p = I1 3

q=

3J2D

cos 3 =

3 3 2

J3D (J2D )3

(4.158)

Jeremi et al.

Lecture Notes

97

I1 = kk

J2D =

1 sij sij 2

J3D =

1 sij = ij kk ij 3

(4.159)

Stresses are here chosen as positive in tension. The denition of Lodes angle in equation (4.158) implies that = 0 denes the meridian of conventional triaxial extension (CTE), while = /3 denotes the meridian of conventional triaxial compression (CTC). The Potential Function is given in the following form:

Q = Q(p, q, ) The complete derivation of the closed form gradients is given in Appendix D.

(4.160)

4.5.4.1

Analytical Gradients

Q Q p Q q Q = + + ij p ij q ij ij

(4.161)

and subsequently the rst derivatives of the chosen stress invariants are

p 1 = ij ij 3

(4.162)

q 3 1 = sij ij 2 q

(4.163)

(4.164)

Jeremi et al.

Lecture Notes

98

where:

tij =

J3D ij

2Q pq mn 2 Q p 2 Q q 2 Q + + 2 p mn pq mn p mn + + 2 Q p 2 Q q 2 Q + + 2 qp mn q mn q mn 2 Q q 2 Q 2 Q p + + p mn q mn 2 mn p Q 2p + pq p pq mn 2q q Q + pq q pq mn 2 Q + pq pq mn

= + + (4.165)

2p = pq mn

(4.166)

2q 31 = pq mn 2q

1 pm nq pq nm 3

9 1 smn spq 4 q3

(4.167)

2 = pq mn + 9 cos 3 27 cos 3 81 1 + spq smn + spq tmn + 2 q 4 sin (3) 4 q 4 sin3 3 4 q 5 sin3 3 81 1 81 cos2 3 243 cos 3 + tpq smn tpq tmn + 4 q 5 sin 3 4 q 5 sin3 3 4 q 6 sin3 3 3 cos (3) 9 1 + 2 ppqmn wpqmn 2 q sin (3) 2 q 3 sin (3)

(4.168)

Jeremi et al.

Lecture Notes

99

where:

wpqmn =

and:

ppqmn =

spq = mn

1 mp nq pq mn 3

4.5.4.2

After having developed the closed form, analytical derivatives52 the author of this thesis asked himself: is there a simpler way of nding these derivatives? One of the proposed ways to check the analytical solution is found in Dennis and Schnabel (1983). Dennis and Schnabel proposes the nite dierence method for approximating derivatives if these derivatives are not analytically available and as a tool to check your analytical derivatives if they are derived. Another good reason for developing alternative gradients is that for = 0, /3 gradients are not dened, i.e. indenite terms as 0/0 are appearing. One possible solution is the use of lHospitals rule. This has been done in Peri (1991). The solution to the problem in this work went in a dierent direction, i.e. instead of aiming for the analytical form, numerical derivatives are derived. We should recall that for a function f of a single variable, the nite dierence approximation to f (x), by using forward nite dierence approach, is given by:

52 see

Appendix (D).

Jeremi et al.

Lecture Notes

100

a=

f (x + h) f (x) h

(4.170)

where h is a vanishingly small quantity. The same denition was used in deriving the nite dierence approximation for the rst derivative of the yield function F and potential function Q. The rst derivative of F ( or Q ) with respect to the stress tensor ij for diagonal elements is53 :

approx.

F,ii =

(4.171)

approx.

F,ij =

(4.172)

where hij is the step size which, because of nite precision arithmetic, is a variable55 . The accuracy of the nite dierence approximation to the analytical derivatives is closely bound to the step size hij . It is suggested in Dennis and Schnabel (1983)[section 5.4.] that for functions given by the simple formula, the number h should be h = macheps, while for more complicated functions that number should be larger. Here macheps is the so called machine epsilon. It is dened as the smallest distinguishable positive number56 , such that 1.0 + macheps > 1.0 on the given platform. For example57 , on the Intel x86 platform58 macheps = 1.08E 19 while on the SUNSparc and DEC platforms macheps = 2.22E 16. It has been found that in the case of yield or potential functions the best approximation of analytical gradients is obtained by using h = macheps 103 . The three order of magnitude increase in the nite dierence step is due to a rather complicated59 formula for yield and potential functions. The error in the approximation,

approx.

where N is the order of macheps, i.e. macheps = O(N ). Second derivative approximations for one variable function are given in the form:

a=

53 no

(f (x + hi ei + hj ej ) f (x + hi ei )) (f (x + hj ej ) f (x)) hi hj

(4.173)

54 since

sum convention implied, just the position of the element. the stress tensor ij is symmetric, change in one non-diagonal element triggers the other to be changed as well. 55 it is actually one small number, h, that is multiplied with the current stress value so that the relative order of magnitude is

retained. 56 in a given precision, i.e. oat ( real*4 ), double ( real*8 ) or long double ( real*10 ). 57 the precision sought was double ( real*8 ). 58 PC computers. 59 One should not forget that we work with six dimensional tensor formulae directly.

Jeremi et al.

Lecture Notes

101

If the rst derivatives are available in closed form, one could use equations (4.171) and (4.172) just by replacing the function values with tensor values for analytical derivatives60 . However, if the analytic derivatives are not available, one has to devise a formula that will create a fourth order tensor from the changes in two dimensional stress tensors, ij and kl . Using the scheme employed in equation (4.173) the following scheme has been devised:

Q,ijkl = (Q(mn + hij + hkl ) Q(mn + hij )) (Q(mn + hkl ) Q(mn )) hij hkl (4.174)

approx.

Special considerations are necessary in order to retain symmetry of the fourth order tensor. At the moment it has not been possible to gure out how to build the nite dierence approximation to the second derivatives of yield/potential functions for a general stress state. The only nite dierence approximation of the second derivatives that appears to have worked was the one devised in principal stress space. Namely, diagonal elements of the analytical and the approximate gradients matched exactly, but development of non-diagonal elements, and the whole scheme of symmetrizing the fourth order approximation, still remain a mystery. However, some pattern was observed in nondiagonal elements, and the work on symmetrizing it is in progress. For many dierent potential functions (or yield functions) the only task left would be the derivation of the rst derivatives of F and Q and the second derivatives of Q with respect to p, q and , namely the rst derivatives

Q Q q and 2 Q 2 . If the Q p ,

and

Q Q p , q

and

Q Q Q Q Q Q Q Q p2 , pq , p , qp , q 2 , q , p , q

and

potential function is twice dierentiable with respect to the stress tensor ij , and if it is continuous

4.6

In this section we present elements of general elasticplastic material models for geomaterials. We describe various forms of the yield functions, plastic ow directions and hardening and softening laws.

4.6.1

Yield Functions

The typical plastic behavior of frictional materials is inuenced by both normal and shear stresses. It is usually assumed that there exists a yield surface F in the stress space that encompasses the elastic region. States of stress inside the yield surface are assumed to be elastic (linear or nonlinear). Stress states on the surface are assumed to produce plastic deformations. Yield surfaces for geomaterials are usually shaped as asymmetric tar drops with smoothly rounded triangular cross sections. In addition to that, simpler yield surfaces, based on

60 see

Jeremi et al.

Lecture Notes

102

the DruckerPrager cone or MohrCoulomb hexagon can also be successfully used if matched with appropriate hardening laws. Yield surface shown in Figure 4.6 Lade (1988b) represent typical meridian plane trace for an isotropic granular material. Line BC represents stress path for conventional triaxial compression test. Figure 4.7

Figure 4.6: Yield surface patterns in the meridian plane for isotropic granular materials (from Lade (1988b)) represents the view of the yield surface traces in the deviatoric plane.

2

low confinment trace high confinment trace

Figure 4.7: Deviatoric trace of typical yield surface for pressure sensitive materials.

4.6.2

Plastic ow directions are traditionally derived from a potential surface which to some extent reassembles the yield surface. Potential surfaces for metals are the same as their yield surfaces but experimental evidence suggests that it is not the case for geomaterials. The nonassociated ow rules, used in geomechanics, rely on the potential

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

103

surface, which is dierent from the yield surface, to provide the plastic ow directions. It should be noted that the potential surface is used for convenience and there is no physical reason to assume that the plastic strain rates are related to a potential surface Q (Vardoulakis and Sulem, 1995). Instead of dening a plastic potential, one may assume that the plastic ow direction is derived from an tensor function which does not have to possess a potential function.

4.6.3

The change in size and/or shape of the yield and potential surfaces is controlled by the hardeningsoftening evolution laws. Physically, these laws control the hardening and/or softening process during loading. Depending on the evolution type they control, these laws can be in general separated into isotropic and kinematic (also called anisotropic). The isotropic evolution laws control the size of the yield surface through a single scalar variable. This is usually related to the Coulomb friction or to the mean stress values at isotropic yielding. The nonisotropic evolution laws can be further specialized to rotational, translational kinematic and distortional. It should be noted that all of the kinematic evolution laws can be treated as special case of the general, distortional laws (Baltov

and Sawczuk, 1965). Figure 4.8 depicts various types of evolution laws (for the control of hardeningsoftening) in the meridian plane61 .

q

a)

q

p c

b)

q

c)

d)

Figure 4.8: Various types of evolution laws that control hardening and/or softening of elasticplastic material models: (a) Isotropic (scalar) controlling equivalent friction angle and isotropic yield stress. (b) Rotational kinematic hardening (second order tensor) controlling pivoting around xed point (usually stress origin) of the yield surface. (c) Translational kinematic hardening (second order tensor) controlling translation of the yield surface. (d) Distortional (fourth order tensor) controlling the shape of the yield surface.

61 The

meridian plane is chosen just for illustration purposes, similar sketch can be produced in deviatoric plane as well.

Jeremi et al.

Lecture Notes

104

4.6.4

Tresca Model

The rst yield criteria in the metal plasticity is Tresca yield criteria. Tresca yield criteria states that when the maximum shear stress or, the half dierence of the maximum and minimum principal stresses, reaches the shear strength, s , the material will begin yielding. It is can be expressed by the yield function f = |max | s = 1 |1 3 | s = 0 2 (4.175)

Tresca yield surface in the principal stress space is a regular hexagonal cylinder. It is implied that the intermediate principal stress plays no role in the yielding for Tresca yield criteria.

4.6.5

Experimental data showed that for most metals, von Mises yield criteria is more accurate than Tresca criteria. von Mises yield function can be expressed by f = 3J2 k 2 = 0 (4.176)

where k is the scalar internal variable; its initial value is the uniaxial tension strength. ij is the tensor internal variable termed back stress. Similar to sij , ij is also a deviatoric symmetric tensor. Although von Mises model is mainly for the metal plasticity analysis, for undrained analysis in geomechanics, von Mises model can be approximately used to simulate the undrained behaviors, (Yang and Jeremi, 2002), (Yang and Jeremi, 2003). The stress derivative of the yield function is f = 3(sij ij ) ij From Equation 4.178, it is easily to derive that f = 3(sij ij ) ij and f = 2 k k If the associated plastic ow rule g = f is assumed, then mij = g = 3(sij ij ) ij (4.181) (4.180) (4.179) (4.178)

mij s = 3Iijmn ij mn mn

Jeremi et al. University of California, Davis

(4.182)

Version: March 21, 2012, 17:15

Lecture Notes

105

mij s = 3Iijmn mn

s is the symmetric unit rank-4 tensor. where Iijmn

(4.183)

It is interesting that from the Equation 4.181, von Mises model gives

p p v = ii = mii = 3(sii ii ) = 0

(4.184)

which accords with the phenomena that no plastic volumetric strain occurs for metals. It is implied that the isotropic stress (hydrostatic pressure) can never make the metal yield for this yield criteria. von Mises model is therefore pressure-independent. If k is assumed a linear relation to the equivalent plastic strain = Hs p = H s k eq 2 dev dev m mij 3 ij

0.5 p q,

where Hs is the linear hardening/softening modulus to the equivalent plastic strain, the corresponding hA is then

h = Hs

0.5

(4.186)

where mdev ij is the deviatoric plastic ow, and if it is associated plasticity, h = 2Hs k If ij is assumed a linear relation to the plastic strain tensor ij = Ht p ij = Ht mij if it is associated plasticity, t (sij ij ) ij = 3H (4.189)

p ij ,

where Ht is the linear hardening/softening modulus to plastic strain tensor, the corresponding hA is then hij = Ht mij if it is associated plasticity, hij = 3Ht (sij ij ) (4.191) (4.190)

A saturation-type kinematic hardening rule is the Armstrong-Frederick hardening (Armstrong and Frederick, 1966),

ij =

Jeremi et al.

2 p ha p ij cr eq ij 3

University of California, Davis

(4.192)

Version: March 21, 2012, 17:15

Lecture Notes

106

if it is associated plasticity, [ha (sij ij ) 2cr kij ] ij = where ha and cr are material constants. The corresponding hA is then hij = 2 ha mij cr meq ij 3 (4.194) (4.193)

where meq is the equivalent plastic ow, and if it is associated plasticity, hij = 2ha sij 2(ha + cr k )ij (4.195)

4.6.5.1 4.6.5.1.1 f=

Yield and Plastic Potential Functions: von Mises Model (form I) Yield function and related derivatives 3 [(sij ij ) (sij ij )] k 2 = 0 2 (4.196)

(4.199)

mij =

(4.200)

mij = 3im jn ij mn mn

(4.201)

mij =0 k

Jeremi et al. University of California, Davis

(4.202)

Version: March 21, 2012, 17:15

Lecture Notes

107

mij = 3im jn mn 4.6.5.2 4.6.5.2.1 Yield and Plastic Potential Functions: von Mises Model (form II) Yield function and related derivatives

0.5

(4.203)

f = [(sij ij ) (sij ij )]

3 k=0 2

(4.204)

f skl 0.5 = (skl kl ) [(smn mn ) (smn mn )] ij ij 1 0.5 = ki lj kl ij (skl kl ) [(smn mn ) (smn mn )] 3 = (sij ij ) [(smn mn ) (smn mn )]

0.5

(4.205)

(4.206)

f = k 4.6.5.2.2

3 2

(4.207)

Plastic ow (associated plasticity) and related derivatives f 0.5 = (sij ij ) [(smn mn ) (smn mn )] ij (4.208)

mij =

mij = mn

mij =0 k

(4.210)

mij 0.5 1.5 = im jn [(srs rs ) (srs rs )] +(sij ij ) (smn mn ) [(srs rs ) (srs rs )] (4.211) mn

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

108

4.6.6

Drucker-Prager Model

Drucker and Prager (1950) proposed a right circle cone to match with the Mohr-Coulomb irregular hexagonal pyramid, which can be expressed by f = I1 + J2 = 0 (4.212)

1 1 f = I1 + [ (sij pij )(sij pij )] 2 = 0 2

(4.213)

where and are material constants. By coinciding Drucker-Prager cone with the outer apexes of the Mohr-Coulomb hexagon locus, we get the compressive cone of Drucker-Prager model, with the constants as = 2 sin 6 cos , = c 3(3 sin ) 3(3 sin ) (4.214)

By coinciding Drucker-Prager cone with the inner apexes of the Mohr-Coulomb hexagon locus, we get the tensile

cone of Drucker-Prager model, with the constants as = 6 cos 2 sin , = c 3(3 + sin ) 3(3 + sin ) (4.215)

We can also get the mean cone of the compressive and tensile cone, with the constants as 3 sin 2 3 cos = , = c 9 sin2 9 sin2 Another inner-tangent cone to the the Mohr-Coulomb pyramid, with the constants as = tan 9 + 12 tan2 , = 3c 9 + 12 tan2

(4.216)

(4.217)

Obviously, in practice and are not directly obtained from experiments. They are functions of Mohr-Coulomb parameters, the cohesion c and the friction angle , which can be determined by experiments. The shape of Drucker-Prager yield surface has dierent types. They only partially satisfy the above requirements for locus in the plane: they do not coincide with both compressive and tensile experimental points. A useful formulation on Equation 4.212 is f sij = ij + ij 2 J2 For cohesionless sands, k = 0, Drucker-Prager yield function can thus be simplied as f = I1 + J2 = 0 (4.219) (4.218)

or in terms of p and q , f = q Mp = 0

Jeremi et al. University of California, Davis

(4.220)

Version: March 21, 2012, 17:15

Lecture Notes

109

If Equation 4.214 is adopted, then M can be easily derived as M= 6 sin 3 sin (4.221)

If the kinematic hardening is taken account, Equation 4.220 can be extended into f= 3 [(sij pij )(sij pij )] M 2 p2 = 0 2 (4.222)

(4.224)

mij =

2 g = 3 sij + s mn mn + M 2 p ij ij 3

(4.225)

4.6.6.1 4.6.6.1.1 f=

Yield and Plastic Potential Functions: Drucker-Prager Model (form I) Yield function and related derivatives 3 [(sij pij ) (sij pij )] k 2 p2 = 0 2 (4.227)

3 smn 3 p p f = 2 (smn pmn ) + 2mn (smn pmn ) 2k 2 p ij 2 ij 2 ij ij 1 1 2 = 3 mi nj mn ij (smn pmn ) + 3 mn ij (smn pmn ) + k 2 pij 3 3 3 2 = 3 (sij pij ) + mn (smn pmn ) ij + k 2 pij 3

(4.228)

f = 3p (sij pij ) ij

(4.229)

f = 2kp2 k

Jeremi et al. University of California, Davis

(4.230)

Version: March 21, 2012, 17:15

Lecture Notes

110

4.6.6.1.2

Plastic ow (associated plasticity) and related derivatives f 2 = 3 (sij pij ) + rs (srs prs ) ij + k 2 pij ij 3 (4.231)

mij =

mij =3 mn

4 mij = kpij k 3

(4.233)

mij = 3pim jn + rm sn (srs prs ) ij rs prm sn ij mn = 3pim jn + (smn pmn ) ij mn pij = 3pim jn + smn ij 2pmn ij (4.234)

4.6.6.2 4.6.6.2.1

Yield and Plastic Potential Functions: Drucker-Prager Model (form II) Yield function and related derivatives

0.5

2 kp = 0 3

(4.235)

f = ij =

smn p mn ij ij

0.5

2 p k 3 ij 2 kij 27 (4.236)

f 0.5 = pmi nj (smn pmn ) [(srs prs ) (srs prs )] ij = p (sij pij ) [(srs prs ) (srs prs )]

0.5

(4.237)

f = k

Jeremi et al.

2 p 3

University of California, Davis

(4.238)

Version: March 21, 2012, 17:15

Lecture Notes

111

4.6.6.2.2

Plastic ow (associated plasticity) and related derivatives f 1 0.5 = (sij pij ) + pq ij (spq ppq ) [(srs prs ) (srs prs )] + ij 3 2 kij (4.239) 27

mij =

mij = mn

1 1 mi nj mn ij + mn ij 3 3

0.5

1 1 1 + pq ij mp nq mn pq + mn pq 3 3 3 1 1 mr ns mn rs + mn rs 3 3

1.5

[(srs prs ) (srs prs )] 1 (sij pij ) + pq ij (spq ppq ) 3 (srs prs ) [(stu ptu ) (stu ptu )]

(4.240)

mij = k

2 ij 27

(4.241)

mij 1 1 0.5 = pmi nj + mp nq ij (spq ppq ) ppq ij mp nq [(srs prs ) (srs prs )] mn 3 3 1 1.5 (sij pij ) + pq ij (spq ppq ) [prm sn (srs prs )] [(stu ptu ) (stu ptu )] 3 (4.242) 4.6.6.3 4.6.6.3.1 4.6.6.3.2 Hardening and Softening Functions Isotropic Hardening and related derivatives Linear Isotropic Hardening (Linear_Eeq) 2 mij mij 3

0.5

= Hmequivalent = H k

(4.243)

2 k mpq = Hmpq ij 3 ij

2 mmn mmn 3

0.5

(4.244)

k 2 mpq = Hmpq k 3 k

2 mmn mmn 3

0.5

(4.245)

k 2 mpq = Hmpq ij 3 ij

Jeremi et al.

2 mmn mmn 3

0.5

(4.246)

Lecture Notes

112

4.6.6.3.3 4.6.6.3.4

Kinematic Hardening and related derivatives Linear Kinematic Hardening (Linear_Eij) (4.247)

ij =H mn

mij 1 mkl kl ij mn 3 mn

(4.248)

ij =H k

mij 1 mkl kl ij k 3 k

(4.249)

ij =H mn

mij 1 mkl kl ij mn 3 mn

(4.250)

4.6.6.3.5 ij =

0.5

ij

(4.251)

2 mkl mkl 3

0.5

ij

(4.252)

ij mij = Ht mn mn

(4.253)

4.6.7

The pioneering research work on the critical state soil mechanics by the researchers in Cambridge University (Roscoe et al., 1958), (Roscoe and Burland, 1968), (Wood, 1990)) has made great contribution on the modern soil elastoplastic models. The original Cam Clay model (Roscoe et al., 1963), and later the modied Cam Clay model (Schoeld and Wroth, 1968) were within the critical state soil mechanics framework. We focus on only the modied Cam Clay model and herein the word modied is omitted to shorten writing.

4.6.7.0.6

Critical State

Jeremi et al. University of California, Davis

(4.254)

Version: March 21, 2012, 17:15

Lecture Notes

62

113

where ec is the critical void ratio at the critical mean eective stress pc c is the normal consolidation slope.

The critical state soil mechanics assumes that the normal consolidation line (NRL) is parallel to the CSL, which is expressed by e = e ln p (4.255)

where e is the intercept on the NRL at p = 1. is the normal consolidation slope or the elastoplastic slope of e ln p relation, and c = . The unloading-reloading line (URL) take the similar form but with dierent slope by e = e ln p (4.256)

where e is the intercept on the URL at p = 1. c is the normal consolidation slope or the elastoplastic slope of e ln p relation.

4.6.7.0.7

Elasticity

The elastic bulk modulus K can be directly derived from the Equation 4.256 and takes the form K= (1 + e)p (4.257)

If a constant Poissons ratio is assumed, since the isotropic elasticity needs only two material constants, the shear elastic modulus can be obtained in terms of K and by G= 3(1 2 )(1 + e) 3(1 2 ) K= p 2(1 + ) 2(1 + ) (4.258)

Alternatively, a constant shear elastic modulus G can be assumed and then the Poissons ratio is expressed in terms of K and G as = 3K 2G 2(G + 3K ) Yield Function (4.259)

4.6.7.0.8

2 f = q 2 Mc [p(p0 p)] = 0

(4.260)

where Mc is the critical state stress ration in the q p plane, and the p0 is the initial internal scalar variable, which is controlled by the change of the plastic volumetric strain. The gradient of the yield surface to the stress can be obtained as f q p 1 2 2 = 2q Mc (2p p0 ) = 3sij + Mc (p0 2p)ij ij ij ij 3

62 In

(4.261)

this chapter, only single-phase (dry phase) is studied, the total and eective stresses are thus identical, e.g. pc = pc .

Jeremi et al.

Lecture Notes

114

where q/ij and p/ij are independent of the yield function. The gradient of the yield surface to p0 will be used in the integration algorithm, and can be expressed by f 2 = Mc p p0 4.6.7.0.9 Plastic Flow (4.262)

The plastic ow of the Cam Clay model is associated with its yield function, in other words, the plastic ow is dened by the potential function, g , which is assumed the same as the yield function, f .

2 g = f = q 2 Mc [p(p0 p)] = 0

(4.263)

The stress gradient to the yield surface can be obtained as mij = g q p 1 2 2 (p0 2p)ij = 2q + Mc (2p p0 ) = 3sij + Mc ij ij ij 3 (4.264)

It can dene the plastic dilation angle , which is related to the ratio of plastic volumetric and deviatoric strain

p v p q

2 (p0 2p) Mc 2q

(4.265)

It is interesting to nd that from Equation 4.265, when p < p0 /2, the plastic dilation angle is positive; when p > p0 /2, the plastic dilation angle is negative. If p = p0 /2, the plastic dilation angle is zero, which is corresponding to the critical state. This is evidently more realistic than Drucker-Prager model, whose associated plastic ow always gives positive plastic dilation angle.

4.6.7.0.10

Evolution Law

The evolution law of the Cam Clay model is a scalar one, which can be expressed by p 0 = (1 + e)p0 p v (4.266)

With this scalar evolution law, the change of p0 is decided by the change of plastic volumetric strain. When it reaches the critical state, or when there is no plastic volumetric strain, the evolution of p0 will cease. From Equation 4.266, one gets (1 + e)p0 mii p 0 = so if using Equation 4.264 further, one obtains h= (1 + e)p0 2 Mc (2p p0 ) (4.268) (4.267)

University of California, Davis

(4.269)

Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

115

4.6.7.1 4.6.7.1.1

Yield and Plastic Potential Functions: Cam-Clay Model Yield function and related derivatives (4.270)

2 f = q 2 Mc [p(p p0 )] = 0

(4.271)

f 2 = Mc p p0 4.6.7.1.2 Plastic ow (associated plasticity) and related derivatives 1 2 f (p0 2p) ij = 3sij + Mc ij 3

(4.272)

mij =

(4.273)

(4.274)

(4.275)

1 ii ), a minus sign appears in from of the evolution of p0 Note the due to the current denition of p (i.e. p = 3

p 0 (1 + e) p0 2 = Mc ij

2 ij 3

(4.278)

p 0 2 (1 + e) 2 = Mc (p p0 ) p0

Jeremi et al. University of California, Davis

(4.279)

Lecture Notes

116

4.6.8

Dafalias-Manzari Model

Within the critical state soil mechanics framework, Manzari and Dafalias (1997) proposed a two-surface sand model. This model considered the eects of the state parameter on the behaviors of the dense or loose sands. The features of this model include successfully predicting the softening at the dense state in drained loading, and also softening at the loose state but in the undrained loading. Dafalias and Manzari (2004) later presented an improved version. This version introduced the fabric dilatancy tensor which has a signicant eect on the contraction unloading response. It is also considered the Lodes angle eect on the bounding surface, which produces more realistic responses in non-triaxial conditions. Here only the new version is summarized. The compression stress is assumed negative here, which is dierent from the original reference by Dafalias and Manzari (2004).

4.6.8.0.4

Critical State

Instead of using the most common linear line of critical void ration vs. logarithmic critical mean eective stress, the power relation recently suggested by Li and Wang (1998) was used:

ec = ec,r c

pc Pat

(4.280)

where ec is the critical void ratio at the critical man eective stress pc , ec,r is the reference critical void ratio, c and (for most sands, = 0.7) are material constants, and Pat is the atmospheric pressure for normalization.

4.6.8.0.5

Elasticity

The elastic incremental moduli of shear and bulk, are following Richart et al. (1970): G = G0 (2.97 e)2 (1 + e) p Pat

0.5

Pat , K =

2(1 + ) G 3(1 2 )

(4.281)

where G0 is a material constant, e is the void ratio, and is the Poissons ratio. The isotropic hypoelasticity is then dened by e e ij = s ij p , e v = 2G K Yield Function (4.282)

4.6.8.0.6

where sij is the deviatoric stress tensor, ij is the deviatoric back stress-ratio tensor, m is a material constant, and || = sij pij = [(sij pij )(sij pij )]0.5

Jeremi et al. University of California, Davis

(4.284)

Version: March 21, 2012, 17:15

Lecture Notes

117

The gradient of the yield surface to the stress can be obtained as f 1 = nij + (pq npq + ij 3 2 m)ij 3 (4.285)

where rij = sij /p is the normalized deviatoric stress tensor, and nij is the unit gradient tensor to the yield surface dened by nij = sij pij || (4.286)

It is evident that nii 0 and nij nij 1. The gradient of the yield surface to ij can be easily obtained as f = pnij ij The tensor of nij is to dened n , the Lodes angle of the yield gradient, by the equation cos 3n = 6nij njk nki (4.288) (4.287)

where 0 n /6 and n = 0 at triaxial compression and n = /6 at triaxial extension. The critical stress ratio M at any stress state can be interpolated between Mc , the triaxial compression critical stress ratio, and Me , the triaxial extension critical stress ratio. M = Mc g (n , c), g (n , c) = 2c Me , c= (1 + c) (1 c) cos 3n Mc (4.289)

The line from the origin of the plane parallel to nij will intersect the bounding, critical and dilation surfaces at

b c d three image back-stress ratio tensor ij , ij , and ij respectively (Figure 4.9), which are expressed as

b ij =

2 b 3

nij

(4.290)

c ij =

2 [M m]nij = 3

2 c 3

nij

(4.291)

d ij =

2 d 3

nij

(4.292)

4.6.8.0.7

Plastic Flow

Jeremi et al. University of California, Davis

(4.293)

Version: March 21, 2012, 17:15

Lecture Notes

118

Figure 4.9:

Schematic illustration of the yield, critical, dilatancy, and bounding surfaces in the -plane of

deviatoric stress ratio space (after Dafalias and Manzari 2004). The deviatoric plastic ow tensor is 1 Rij = Bnij + C (nik nkj ij ) 3 where B =1+ 31c g cos 3n , C = 3 2 c 31c g 2 c (4.295) (4.294)

d D = Ad (ij ij )nij = Ad

2 d ij nij 3

(4.296)

where Ad = A0 (1 + zij nij ) A0 is a material constant, and zij is the fabric dilation tensor. The Macauley brackets if x > 0 and x = 0, if x 0. (4.297) is dened that x = x,

4.6.8.0.8

Evolution Laws

This model has two tensorial evolution internal variable, namely, the back stress-ratio tensor ij and the fabric dilation tensor zij . The evolution law for the back stress-ratio tensor ij is [ 2 h(b ij )] ij = ij 3

Jeremi et al. University of California, Davis

(4.298)

Version: March 21, 2012, 17:15

Lecture Notes

119

where ij is the initial value of ij at initiation of a new loading process and is updated to the new value when the denominator of Equation 4.299 becomes negative. b0 is expressed by b0 = G0 h0 (1 ch e) p Pat

0.5

(4.300)

where h0 and ch are material constants. The evolution law for the fabric dilation tensor zij is (zmax nij + zij ) z ij = cz D where cz and zmax are material constants. (4.301)

4.6.8.0.9

When implemented into an implicit algorithm for the Dafalias-Manzari model, some complicated additional analytical derivatives are needed. This section gives the analytical derivatives expressions based on the tensor calculus. Analytical expression of mij : kl

mij mn

= B

(4.302)

D mn and B mn

Ad mn

2 d ab nab 3

Ad

d 2 nab ab 3 mn mn

(4.304)

3 2

1c c

g cos 3 cos 3 + g mn mn

(4.305)

C mn

Jeremi et al.

3 2

1c c

d mn

(4.306)

University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

120

(4.307)

mn

c 3Pat

p Pat

( 1)

mn

(4.308)

g mn

= g2

1c 2c

cos 3 mn

(4.309)

cos 3 mn Ad mn

(4.310)

= A0 zab

(4.311)

Analytical expression of

mij : kl

mij mn

(4.312)

D mn

Ad mn

2 d ab nab 3

Ad

d 2 nab nmn ab 3 mn mn

C mn

3 2

1c c

d mn

(4.316)

d mn

Mc exp (nd )

g mn

University of California, Davis

(4.317)

Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

121

g mn

g2

1c 2c

cos 3 mn

(4.318)

cos 3 mn

(4.319)

Ad mn

A0 zab

(4.320)

Analytical expression of

mij : zmn

1 D ij 3 zmn

(4.321)

Ad zmn

2 d ab nab 3

(4.322)

(4.323)

Analytical expression of

Aij : mn

Aij mn

h 2 3 mn

2 b nij ij 3

b 2 b nij h nij + 3 mn mn

(4.324) where

b = Mc exp (nb ) mn

g nb g mn mn

(4.325)

Jeremi et al. University of California, Davis

(4.326)

(4.327)

Version: March 21, 2012, 17:15

Lecture Notes

122

Analytical expression of

Aij : mn

Aij mn

2 3 +

2 b nij ij 3

h mn (4.328)

where

b g = Mc exp (nb ) mn mn

(4.329)

(4.330)

Analytical expression of

Aij zmn

(4.331)

Analytical expression of

Zij : mn

Zij mn

nij D + zmax D mn mn

(4.332)

Analytical expression of

Zij : mn

Zij mn

D nij + zmax D mn mn

(4.333)

Analytical expression of

Zij : zmn

Zij zmn

D zmn

(4.334)

Jeremi et al.

Chapter 5

(2004-2006-2009-)

For more details on work in this area, please consult the following papers: Jeremi et al. (2007), Sett et al. (2007a), Sett et al. (2007b), Jeremi and Sett (2007), Jeremi and Sett (2006), Sett and Jeremi (2007), Jeremi and Sett (2009), Sett and Jeremi (2010), Sett et al. (2011b), Sett et al. (2011a), Jeremi and Sett (2010).

123

Lecture Notes

124

Jeremi et al.

Chapter 6

(1996-2004-)

125

Lecture Notes

126

6.1 6.2

6.2.1

Deformation

In modeling the material nonlinear behavior of solids, plasticity theory is applicable primarily to those bodies that can experience inelastic deformations considerably greater than the elastic deformation. If the resulting total deformation, including both translations and rotations, are small enough, we can apply small deformation theory in solving these problems. If, however strains and rotations are nite, one must resort to the theory of large deformations. In doing so, we will be using two sets of representations1 , namely: Material coordinates in the undeformed conguration, also called Lagrangian coordinates, Spatial coordinates in the deformed conguration, also called Eulerian coordinates.

Figure 6.1 shows the displacement of a particle from its initial position XI to the current position xi , dened by the deformation equation: xi = xi (X1 , X2 , X3 , t) (6.1)

d xi d XI

XI

ui

xi

Figure 6.1: Displacement, stretch and rotation of material vector dXI to new position dxi . The initial position XI of the particle now occupying the position xi is given by the Eulerian equation:

XI = XI (x1 , x2 , x3 , t)

1 See

(6.2)

Malvern (1969).

Jeremi et al.

Lecture Notes

127

xi = XI + ui

XI = xi ui

(6.3)

6.2.2

Deformation Gradient

The deformation gradients are the gradients of the functions on the righthand side of equations (6.1) and (6.2). To emphasize the dierence between the material, Lagrangian setting and the spatial, Eulerian setting, we will use capital letters for the material coordinate indices and lower case letters for the spatial coordinate indices. We limit our work to the rectangular Cartesian coordinates, thus simplifying the tensor notation to the covariant indices only. The deformation gradient is dened as the twopoint tensor whose rectangular Cartesian components are the partial derivatives:

FkK =

xk = xk,K XK

(6.4)

The deformation gradient FkK transforms (convects) on an arbitrary innitesimal material vector dXI at XI to associate it with a vector dxi at xi :

(6.5)

The the spatial deformation gradients are tensors referred to the deformed, Eulerian conguration: (FKk )

1

XK = XK,k xk

1

(6.6)

Similarly to the deformation gradient FkK , spatial deformation gradient (FKk ) nitesimal material vector dxi at xi to associate it with a vector dXI at XI : dXK = (FKk )

1

dxk =

1

(6.7)

1

= ik

and

(FIj )

FjK = IK

The Jacobian of the mapping (6.4) can be represented as: J = det (FkK ) =

Jeremi et al.

University of California, Davis

(6.9)

Version: March 21, 2012, 17:15

Lecture Notes

128

The relative deformation gradient fkm is the gradient for the relative motion function:

= t (xi , )

(6.10)

If the xed reference position XI , the current position xi and the variable position i are all referred to the rectangular Cartesian coordinate system, the chain rule of dierentiation yields:

k k xm = XI xm XI

or

(6.12)

(6.13)

where Ukj , vik are positive denite symmetric tensors, called right stretch tensors and left stretch tensors, respectively, and Rkj is an orthogonal tensor such that:

Rik Rjk = ij

and also

Rki Rkj = ij

(6.14)

Equation (6.13), as well as Figure 6.2.2 demonstrate that the motion and deformation of an innitesimal volume element at Xi consist of consecutive applications of: a stretch by Ukj , a rigid body rotation by Rik , a rigid body translation to xi or alternatively: a rigid body translation to xi a rigid body rotation by Rkj , a stretch by vik ,

2 referring

xi and Xi to the same reference axes and using lower case indices for both. This reference to the same coordinate

system will be applied only for the polar decomposition example presented here.

Jeremi et al.

Lecture Notes

129

d xi d XI

XI

ui

xi

Figure 6.2: Illustration of the equation Fij = Rik Ukj = vik Rkj .

6.2.3

The strain tensors EIJ and eij are dened so that they give the change in the square length of the material vector dXI . For the Lagrangian formulation we write:

(6.15)

(6.16)

The deformation tensors CIJ and cij are connecting the squared lengths in Lagrangian and Eulerian congurations. The Green deformation tensor3 CIJ , referred to the undeformed conguration, gives the new squared length (ds)2 of the element into which the given element dXI is deformed: (ds)2 = dXI CIJ dXJ

1

(6.17)

The Cauchy deformation tensor cij , sometimes also denoted as4 (bij ) an element dxi identied in the deformed conguration: (dS )2 = dxi cij dxj

3 Also 4 Another

(6.18)

called right CauchyGreen tensor. name for bij is Finger deformation tensor or left CauchyGreen tensor.

Jeremi et al.

Lecture Notes

130

2EIJ = CIJ IJ

(6.19)

2eij = ij cij

(6.20)

(ds)2 = dxk dxk = (FkI dXI )(FkJ dXJ ) = (xk,I dXI )(xk,J dXJ ) = dXI (FkI FkJ )dXK = dXI CIJ dXK (6.21)

so we have obtained the connection between the deformation tensor CIJ and the deformation gradient FkI in the form:

(6.22)

Similarly, by using equation (6.7) and the expression for (dS )2 we can establish the connection between the deformation tensor cij and the deformation gradient FKi as: (dS )2 = dSK dXK = (FKi dxi )(FKj dxj ) = (XK,i dxi )(XK,j dxJ ) = dxi (FKi FKj )dxk = dxi cij dxk cij = (FKi )

1 1 FKj

(6.23)

The expressions for the strain tensors in Lagrangian and Eulerian description5 is obtained from equations (6.19) and (6.20): 1 ((FkI FkJ ) IJ ) 2 1 1 1 ij (FKi ) (FKj ) 2

L: EIJ =

E: eij =

(6.24)

If one starts from the displacement equation (6.3), referenced to the same axes for both XI and xi

xI = XI + uI

5 Lagrangian

XI = xI uI

Jeremi et al.

Lecture Notes

131

the general expression for the Lagrangian strain tensor EIJ in terms of displacements is: EIJ = 1 ((FKI FKJ ) IJ ) = 2

1 ((KI + uK,I ) (KJ + uK,J ) IJ ) = 2 1 (KI KJ + KI uK,J + uK,I KJ + uK,I uK,J IJ ) = 2 1 (IJ + uI,J + uJ,I + uK,I uK,J IJ ) = 2 1 (uI,J + uJ,I + uK,I uK,J ) 2 Similarly, the general expression for the Eulerian strain tensor eij in terms of displacements is: eij = 1 1 1 ij (Fki ) (Fkj ) = 2

(6.25)

1 (ij (ki uk,i ) (kj uk,j )) = 2 1 (ij ki kj + ki uk,j + uk,i kj uk,i uk,j ) = 2 1 (ij ij + ui,j + uj,i uk,i uk,j ) = 2 1 (ui,j + uj,i uk,i uk,j ) 2

(6.26)

It is worthwhile noting that equations (6.25) and (6.26) represent the complete nite strain tensor. They involve only linear and quadratic terms in the components of displacement gradients. The stretch is a measure of extension of an innitesimal element and is a function of direction of an element, in either deformed or undeformed conguration. By denoting NI a unit vector in the undeformed conguration and ni a unit vector in the deformed conguration, we denote material stretch as (N ) of those elements whit initial direction NI and spatial stretch (n) of those elements with initial direction ni . By dividing equations (6.15) and (6.16) by (ds)2 and (dS )2 respectively and by using: NI = dXI dS and ni = dxi ds (6.27)

we obtain the Cartesian form of stretch in the Lagrangian and Eulerian descriptions: L: 2 (N ) = dXI dXJ CIJ dS dS and E: 2 (n) = dxi dxj cij ds ds (6.28)

General strain tensors can be dened by considering a scale function (Hill, 1978) for the stretch. Scale function is any smooth, monotonic function of stretch f () such that: f () ; [0, ) subject to f (1) = 0, f (1) = 1 (6.29)

Scale function is often taken in the form (2m 1)/2m, where m may have any value. If we choose m to be an

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

132

2m UIJ IJ 2m

where

(6.30)

Table 6.1 shows dierent Lagrangian strain measures obtained for a particular choice of parameter m. Table 6.1: Dierent Lagrangian strain measures. Strain measure name GreenLagrange Almansi Biot Hencky parameter m 1 -1 1/2 0

m expression for EIJ

GL 2 EIJ = UIJ IJ /2

2 A EIJ = IJ UIJ /2

B = (UIJ IJ ) EIJ

H EIJ = ln (UIJ )

2m ij vij 2m

(6.31)

6.2.4

The rate of deformation tensor6 describes the tangent motion in terms of velocity components vi = dxi /dt. The spatial coordinates are: vi = vi (x1 , x2 , x3 , t) (6.32)

In Figure 6.3 the dashed lines represents the trajectories of particles P and Q. The velocity vectors vi at p and vi + dvi at q are tangent to the two trajectories. The relative velocity components dvi of particle at q relative to the particle at p are given by: dvk = vk dxm = vk,m dxm = Lkm dxm xm (6.33)

The spatial gradient of the velocity Lkm can be decomposed as the sum of the symmetric, rate of deformation tensor Dkm , and a skew symmetric spin tensor Wkm as follows: Lkm =

6 Also

(6.34)

Jeremi et al.

Lecture Notes

133

q Q d xi d Xi P ui

vi+ d vi

p vi d vi

Figure 6.3: Relative velocity dvi of particle Q at point q relative to particle P at point p. where: Dkm = 1 (Lkm + Lmk ) = Dmk 2 and Wkm = 1 (Lkm Lmk ) = Wmk 2 (6.35)

An alternate way of deriving the rate of deformation tensor goes as follows. The rate of change of squared length

(ds) is given as: d (ds) d (ds) =2 ds dt dt since (ds) = dxk dxk it follows: d (dxk ) d (ds) =2 dxk dt dt and with dxk = (xk /Xm )dXm it follows: d d (dxk ) = dt

xk Xm dXm 2 2 2

(6.36)

(6.37)

d =

xk Xm

dt

dt

dXm +

xk d X d (dXm ) xk m dXm = dt Xm dt

(6.38)

since d (dXm ) /dt 0, because the initial relative position vector dXm does not change with time. By interchanging the order of dierentiation we get: d d (dxk ) = dvk = dt

xk XM

dt

dXm =

vk dXm Xm

where

vk =

xk dt

(6.39)

From equation (6.33) dvk = Lkm dxm and equation (6.39) it follows that: vk dXm = Lkm dxm Xm d (dxk ) = dvk = Lkm dxm = vk,m dxm dt (6.40)

and then the equation (6.37) becomes: d (dxk ) d (ds) =2 dxk = 2dxk vk,m dxm dxk = 2dxk Lkm dxm dxk = dt dt = 2dxk Dkm dxm dxk + 2dxk Wkm dxm dxk = 2dxk Dkm dxm dxk

Jeremi et al. University of California, Davis

2

(6.41)

Version: March 21, 2012, 17:15

Lecture Notes

134

since dxk dxm dxm dxk and Wkm is skew symmetric such that Wkm = Wmk . Finally we obtain: d (ds) = 2dxk Dkm dxm dt

2 2

(6.42)

and thus it follows that the rate of change of the squared length (ds) of the material instantaneously occupying any innitesimal relative position dxk at point p is determined by the tensor Dkm at point p. In order to compare the strain rate to the rate of deformation, we dierentiate equation (6.15) with respect to time: d (ds)2 (dS )2 d (dXI EIJ dXJ ) =2 = dt dt d (ds)2 d (EIJ ) = = 2dXI dXJ dt dt since (dS )2 and dXI are constant through time. From the equations (6.42) and (6.43) it follows that: d (ds) = 2dxk Dkm dxm d = 2 (dXI FIk ) Dkm (FmJ dXJ ) = 2dXI (FIk Dkm FmJ ) dXJ dt (6.44) and from equations (6.43) and (6.44) it follows that: dEIJ = FIk Dkm FmJ dt or inversely: Dkm = (FIk )

1 2

(6.43)

(6.45)

dEIJ 1 (FmJ ) dt

(6.46)

To obtain the rate of change of the deformation gradient we start from equations (6.4) and dierentiate it with respect to time:

xk d X dFkK K = dt dt

k dx dt XK

Jeremi et al. University of California, Davis

(6.48)

Version: March 21, 2012, 17:15

Lecture Notes

135

6.3

6.3.1

Introduction

A material is called hyperelastic or Green elastic, if there exists an elastic potential function W , also called the strain energy function per unit volume of the undeformed conguration, which represents a scalar function of strain of deformation tensors, whose derivatives with respect to a strain component determines the corresponding stress component. The most general form of the elastic potential function, is described in equation 6.49, with restriction to pure mechanical theory, by using the axiom of locality and the axiom of entropy production7 : W = W (XK , FkK ) (6.49)

By using the axiom of material frame indierence8 , we conclude that W depends only on XK and CIJ , that is: W = W (XK , CIJ ) or: W = W (XK , cij ) (6.50)

By assuming hyperelastic response, the following are the constitutive equations for the material stress tensors: 2. PiolaKirchho stress tensor: SIJ = 2 W CIJ (6.51)

and the spatial, Kirchho stress tensor is dened as: Kirchho stress tensor ij = 2

7 See 8 See

(6.54)

Marsden and Hughes (1983) pp. 190. Marsden and Hughes (1983) pp. 194.

Jeremi et al.

Lecture Notes

136

Material tangent stiness relation is dened from: dSIJ = 2 where LIJKL = 4 2W CIJ CKL (6.56) 2W 1 dCKL = LIJKL dCKL CIJ CKL 2 (6.55)

The spatial tangent stiness tensor Eijkl is obtained by the following pushforward operation with the deformation gradient: Eijkl = FiI FjJ (FkK )t (FlL )t LIJKL (6.57)

6.3.2

Isotropic Hyperelasticity

In the case of material isotropy, the strain energy function W (XK , CIJ ) belongs to the class of isotropic, invariant scalar functions. It satises the relation: W (XK , CKL ) = W XK , QKI CIJ (QJL )

t

(6.58)

where QKI is the proper orthogonal transformation. If we choose QKI = RKI , where RKI is the orthogonal rotation transformation, dened by the polar decomposition theorem in equation (6.13), then: W (XK , CKL ) = W (XK , UKL ) = W (XK , vkl ) (6.59)

Right and left stretch tensors, UKL , vkl have the same principal values9 i ; deformation tensor: W = W (XK , 1 , 2 , 3 , ) = W (XK , I1 , I2 , I3 ) where: I1 I2 I3

def

function W can be represented in terms of principal stretches, or similarly in terms of principal invariants of

(6.60)

= =

def

def

(6.61)

9 Principal

stretches.

Jeremi et al.

Lecture Notes

137

Left and right CauchyGreen tensors were dened by equations (6.22) and (6.23), respectively as: CIJ = (FkI )t FkJ ; (c1 )ij = bij = FiK (FjK )t (6.62)

The spectral10 decomposition theorem for symmetric positive denite tensors11 states that: CIJ = 2 A NI

(A)

NJ

(A) A

where

A = 1, 3

(6.63)

and NI are the eigenvectors12 of CIJ . Values 2 A are the roots of the characteristic polynomial

6 6 4 P (2 A ) = A + I1 A I2 A + I3 = 0 def

(6.64)

It should be noted that no summation is implied over indices in parenthesis13 . The mapping of the eigenvectors can be deduced from equation (6.5) and is given by (A) ni

(A)

= FiJ NJ

(A)

(6.65)

where ni

(A)

(A) (A) A

FiJ = A ni NJ

(6.66)

RiJ =

A=1

n i NJ

(A)

(A)

(6.67)

bij = 2 A ni nj

(A) (A) A

(6.68)

Spectral decomposition from equation (6.63) is valid for the case of nonequal principal stretches, i.e. 1 = 2 = 3 . If two or all three principal stretches are equal, we shall introduce a small perturbation to the numerical values for principal stretches in order to make them distinct. The case of two or all three values of principal stretches being equal is theoretically possible and results for example from standard triaxial tests or isotropic compression tests. However, we are never certain about equivalence of two numerical numbers, because of the nite precision arithmetics involved in calculation of these numbers. From the numerical point of view, two number are equal if the dierence between them is smaller than the machine precision (macheps) specic to the computer platform on which computations are performed. Our perturbation will be a function of the macheps.

Simo and Taylor (1991). tensor CIJ for example. 12 So that N I = 1.

11 CauchyGreen 13 For 10 See

(A) (A) NI NJ A

(A)

A=3 A=1 (A) (A) 2 NJ . (A) NI

(A)

, N2

(A)

and N3

(A)

CIJ = 2 A

in this work, we shall make an eort to represent all the tensorial equations in indicial form.

Jeremi et al.

Lecture Notes

138

1 A = 3 where = arccos

I1 + 2

2 3I cos I1 2

+ 2A 3

(6.69)

(6.70)

2 3I ) (I1 2

Recently, Ting (1985) and Morman (1986) have used Serrins representation theorem in order to devise a useful

m representation for generalized strain tensors15 EIJ and eij through CIJ and bm ij . Morman (1986) has shown that

bm ij can be stated as

2m bm ij = A 2 (b2 )ij (I1 2 (A) bij + I3 (A) ij 2 2 24 (A) I1 (A) + I3 (A)

(6.71)

By comparing equations (6.71) and (6.68) it follows that the Eulerian eigendiad ni nj

2 (b2 )ij I1 2 (A) bij + I3 (A) ij 2 2 24 (A) I1 (A) + I3 (A)

(A) (A)

can be written as

ni nj

(A) (A)

(6.72)

(A)

(A)

2 1 CIJ I1 2 )IJ (A) IJ + I3 (A) (C 2 2 24 (A) I1 (A) + I3 (A)

(A) (A) NI NJ

2 (A)

(6.73)

where it was used that: CIJ = (FiI )1 (b2 )ij (FjJ )t (6.74)

(6.75)

(6.76)

It should be noted that the denominator in equations (6.72) and (6.73) can be written as:

2 2 2 2 24 (A) I1 (A) + I3 (A) = (A) (B )

14 See

2 2 (A) (C )

def

= D(A)

(6.77)

15 Dened

also Morman (1986) and Schellekens and Schellekens and Parisch (1994). by equations (6.30) and (6.31).

Jeremi et al.

Lecture Notes

139

where indices A, B, C are cyclic permutations of 1, 2, 3. It follows directly from the denition of D(A) in equation (6.77) that 1 = 2 = 3 D(A) = 0 for equations (6.72) and (6.73) to be valid. Similarly to equations (6.63) and (6.68) we can obtain:

2 (C 1 )IJ = NI A (A)

NJ

(A) A

(6.78)

2 (b1 )ij = ni nj A

(A) (A) A

(6.79)

6.3.3

It proves useful to separate deformation in volumetric and isochoric parts by a multiplicative split of a deformation gradient as i FiI = F

vol

FI

where

1 i = FiI J 3 F

vol

FI = J 3 I

(6.80)

where x represents an intermediate conguration such that deformation XI x is purely volumetric and I and FiI have the same eigenvectors. x xi is purely isochoric. It also follows from equation (6.80) that F

x

vol

F I F iI

iso

F i xi

XI

Figure 6.4: Volumetric isochoric decomposition of deformation. The isochoric part of the Green deformation tensor CIJ , dened in equation (6.63) can be dened as

(A) (A) 2 IJ = J 2 3C C IJ = A NI NJ

(6.81)

A

while the isochoric part of the Finger deformation tensor bij can be dened similarly as

2 2 n(A) n(A) bij = J 3 bij = A i j

(6.82)

A

1 1 A = J 3 A = (1 2 3 ) 3 A

(6.83)

University of California, Davis Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

140

The free energy W is then decomposed additively as: (A) + volW (XK , J ) W XK , (A) = isoW XK , (6.84)

6.3.4

SimoSerrins Formulation

In Section (6.3.2) we have presented the most general form of the isotropic strain energy function W in terms of of principal stretches: W = W (XK , 1 , 2 , 3 , ) (6.85)

It was also shown in Section (6.3.1) that it is necessary to calculate the gradient W /CIJ in order to obtain

2. PiolaKirchho stress tensor SIJ and accordingly other stress measures. Likewise, it was shown that the material tangent stiness tensor LIJKL (as well as the spatial tangent stiness tensor Eijkl ) requires second order derivatives of strain energy function 2 W /(CIJ CKL ). In order to obtain these quantities we introduce16 a second order tensor MIJ

(A)

MIJ

(A)

def

2 (A) NI

(A)

NJ

(A)

(6.86) (FjJ )t

from (6.73)

= =

(FiI )1 1 D(A)

ni nj

(A) (A)

where D(A) was dened by equation (6.77). With MIJ dened by equation (6.86), we get from equation (6.63) that: CIJ = 4 A MIJ

(A) A

(A)

(6.87)

and also from equation (6.78) it follows that: (C 1 )IJ = MIJ + MIJ + MIJ It can also be concluded that:

2 2 2 IJ = 2 (1) MIJ + (2) MIJ + (3) MIJ = A MIJ

16 See

(1)

(2)

(3)

(6.88)

(1)

(2)

(3)

(A) A

(6.89)

Runesson (1996).

Jeremi et al.

Lecture Notes

141

3 (A) (A) (A) A (A) A

IJ =

A=1

NI

NJ

= NI

NJ

(6.90)

We are now in a position to dene the SimoSerrin fourth order tensor MIJKL as:

(A) def

MIJKL

2 1 I3 )IJ (C 1 )KL + (A) (C 2 (A) I3 (A)

(A)

(A) (A) (A)

(6.91)

6.3.5

Stress Measures

In Section (6.3.1) we have dened various stress measures in terms of derivatives of the free energy function W . With the free energy function decomposition, as dened in equation (6.84) we can appropriately decompose all the previously dened stress measures: 2. PiolaKirchho stress tensor: SIJ = = 2 isoW volW W =2 +2 CIJ CIJ CIJ iso SIJ + volSIJ

(6.92)

iso

TIJ + volTIJ

Jeremi et al.

SIJ (FiI )t = 2

iso

University of California, Davis Version: March 21, 2012, 17:15

PiJ + volPiJ

Lecture Notes

142

Kirchho stress tensor W isoW volW = FaI (FbJ )t SIJ = 2FaI (FbJ )t + 2FaI (FbJ )t eij CIJ CIJ ab + volab

ab

iso

(6.95)

The derivative of the volumetric part of the free energy function is volW (J ) J 1 volW (J ) volW (J ) = = J (C 1 )IJ CIJ J CIJ 2 J

(6.96)

where equation (E.9) was used, while the derivative of the isochoric part of the free energy function yields (A) ) (A) ) (A) isoW ( isoW ( 1 1 isoW ((A) ) (A) (A) (A) (MIJ )A = wA (MIJ )A = = CIJ (A) CIJ 2 (A) 2 (6.97) where equation (E.7) was used and wA is derived in Appendix E.5 as:

iso iso B isoW ((A) ) (A) = 1 W ((A) ) (A) B + W ((A) ) B B (A) (A) 3

wA =

(6.98)

vol

SIJ

= =

(6.99)

The derivative of the free energy is then: W ((A) ) CIJ volW ((A) ) isoW ((A) ) + CIJ CIJ vol 1 W (J ) 1 (A) J (C 1 )IJ + wA (MIJ )A 2 J 2

= =

(6.100)

It is obvious that the only material dependent parts are derivatives in the form volW /J and wA , while the rest is independent of which hyperelastic material model we choose.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

143

6.3.6

The free energy function decomposition (6.84) is used together with the appropriate denitions made in section (6.3.1) toward the tangent stiness operator decomposition LIJKL = vol LIJKL + iso LIJKL = 4 2 volW 2 isoW +4 CIJ CKL CIJ CKL (6.101)

vol

1 4

J2

2 volW volW +J JJ J

and the complete derivation is again given in appendix E.3. The isochoric part 2

iso

2 isoW ((A) ) 1 1 (B ) (A) (A) = YAB (MKL )B (MIJ )A + wA (MIJKL )A CIJ CKL 4 2 and the complete derivation is given in the appendix (E.4). Finally, one can write the volumetric and isochoric parts of the tangent stiness tensors as:

vol

LIJKL =

J2

(B )

(A)

(A)

(6.105)

In a similar manner to the stress denitions it is clear that the only material model dependent parts are YAB and wA . The remaining second and fourth order tensors MIJ and MIJKL are independent of the choice of the material model. This observation has a practical consequence in that it is possible to create a template derivations for various hyperelastic isotropic material models. Only rst and second derivatives of strain energy function with A ) and Jacobian (J ) are needed in addition to the independent tensors, respect to isochoric principal stretches ( for the determination of various stress and tangent stiness tensors.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

(A) (A)

Lecture Notes

144

6.3.7

The strain energy function for isotropic solid in terms of principal stretches is represented as: W = W (1 , 2 , 3 ) (6.106)

The only restriction is that W is a symmetric function of 1 , 2 , 3 , although an appropriate natural conguration condition requires that: W (1, 1, 1) = 0 and W (1, 1, 1) =0 i (6.107)

The strain energy function W can either be regarded as a function of principal stretches or the principal invariants of stretches17 : I1 I2

2 2 = 2 1 + 2 + 3 2 2 2 2 2 = 2 2 3 + 3 1 + 1 2 2 2 = 2 1 2 3

I3

(6.108)

A slightly more general formulation is obtained by using principal stretches in the strain energy function denition. A widely exploited family of compressible hyperelastic models18 are dened (Ogden, 1984) as an innite series in powers of (I1 3), (I2 3) and (I3 1) as:

N p q r

W =

p,q,r =0

(6.109)

The regularity condition that W is continuously dierentiable an innitely number of times is satised. The requirement that energy vanishes in the reference conguration is met provided c000 = 0. Reference conguration is stress free i c100 + 2c010 + c001 = 0. Isochoric deviatoric decoupling is possible by setting cpqr = 0 (r = 1, 2, 3, ...) and cpqr = 0 (p, q = 1, 2, 3, ..) to obtain: W = isoW + volW where:

N iso

(6.110)

W =

p,q =0 N

r

vol

W =

r =0

c00r (I3 1)

(6.111)

In what follows, we will present a number of widely used strain energy functions for isotropic elastic solids.

17 See 18 Used

Jeremi et al.

Lecture Notes

145

6.3.7.1

Ogden Model

A very general set of hyperelastic models was dened by Ogden (1984). The strain energy is expressed as a function of principal stretches as:

N

W =

r =1

cr r r r ( 1 + 2 + 3 3) r

(6.112)

N iso

W =

r =1

cr r

r 1 r + r + 3 2 1

(6.113)

Derivatives needed for building tensors wA and YAB are given by the following formulae: isoW = A

N

cr

r =1

r 1

(6.114)

iso

2 A

=

r =1

cr (r 1)

r 2

(6.115)

(6.116)

The general isotropic hyperelastic model dened in terms of invariants of principal stretches contains the Neo Hookean model as special cases. The isochoric part of NeoHookean isotropic elastic model can be obtained by selecting N = 1, q = 0, cp00 = G/2, to get:

iso

W =

G 2 2 2 1 + 2 + 3 3 2

(6.117)

while the volumetric part can be dened by choosing N = 2, c001 = 0, c002 = Kb /2, as:

vol

W =

Kb 2 2 2 1 2 3 1 2

Kb 2 J 1 2

(6.118)

where G and Kb are the shear and bulk moduli respectively. Derivatives needed for building tensors wA and YAB are given by the following formulae: isoW A =G A

Jeremi et al. University of California, Davis

(6.119)

Version: March 21, 2012, 17:15

Lecture Notes

146

iso

2 A

=G

(6.120)

(6.121)

Mooney proposed a strain energy function for isochoric behavior of the form:

N iso

=

n=0 N

2n + 2n + 2n 3 + an 2n + 2n + 2n 3 an 1 2 3 1 2 3 2n + 2n + 2n 3 an 1 2 3

n=0

= + an

2 3

2n

3 1 +

2n

1 2 +

2n

(6.122)

with an and bn being the material parameters and a volume preserving constrain a = 1/(b c , and a, b, c are cyclic permutations of (1, 2, 3). A more general form was proposed by Rivlin:

N iso

W =

p,q =0

(6.123)

which is actually quite similar to the isochoric part of the general isotropic representation from equation (6.111). Both Mooney and Rivlin strain energy functions become similar, if one chooses to set N = 1 and c10 = C1 and c01 = C2 to obtain:

iso

= =

2 + 2 + 2 3 + C2 2 + 2 + 2 3 C1 1 2 3 1 2 3 1 3 + C2 I 2 3 C1 I (6.124)

Derivatives needed for building tensors wA and YAB are given by the following formulae: isoW A 2 C2 3 = 2 C1 A A (6.125)

W 4 = 2 C1 + 6 C2 A 2 A

iso

(6.126)

2 isoW =0 A B

Jeremi et al. University of California, Davis

(6.127)

Lecture Notes

147

6.3.7.4

Logarithmic Model

ln 1 I

= =

1 2 ln ln 1

2

2 + 2 ln

2

3 + 2 ln

2

ln + 2

2

ln + 3

2

ln 2 I

= =

2 4 ln ln 2

2

3 ln

2

3 + 4 ln

2

1 ln

1 + 4 ln ln 2

2

2 ln

ln 3

ln + 3

ln 1

ln + 1

(6.128)

ln and I ln was proposed The general representation of the isochoric part of the strain energy function in terms of I 1 2

by Simo and Miehe (1992). A somewhat simpler isochoric strain energy function can be presented in the form:

iso

W =G

1 ln

2 + ln

3 + ln

(6.130)

vol

W =

Kb 2 (ln J ) 2

(6.131)

Derivatives needed for building tensors wA and YAB are given by the following formulae: isoW =2G A A

1

(6.132)

W = 2 G 2 A

iso

(6.133)

2 isoW =0 A B

(6.134)

vol

W = Kb J 1 ln J dJ

(6.135)

d2

vol

dJ 2

Jeremi et al.

= Kb J 2 Kb J 2 ln J

University of California, Davis

(6.136)

Version: March 21, 2012, 17:15

Lecture Notes

148

6.3.7.5

SimoPister Model

Another form or a volumetric part of strain energy function was proposed by Simo and Pister (1984) in the form: Wvol (J ) = 1 Kb J 2 1 2 ln J 4 (6.137)

The rst and second derivatives with respect to J are then given as:

2 + 2 J Kb dvolW (J ) = J dJ 4 2 + J22 Kb d2 volW (J ) = dJ 2 4

(6.138) (6.139)

6.4

6.4.1

Introduction

The mathematical structure and numerical analysis of classical small deformation elastoplasticity is generally well established. However, development of large deformation elasticplastic algorithms for isotropic and anisotropic material models is still a research area. Here, we present a new integration algorithm, based on the multiplicative decomposition of the deformation gradient into elastic and plastic parts. The algorithm is novel in that it is designed to be used with isotropic as well as anisotropic material models. Consistent derivation is based on the idea from the innitesimal strain algorithm developed earlier by Jeremi and Sture (1997). The algorithm is not an extension of earlier developments, but rather a novel development which consistently utilizes Newtons method for numerical solution scheme for integrating pertinent constitutive equations. It is also shown that in the limit, the proposed algorithm reduces to the small strain counterpart. In what follows, we briey introduce the multiplicative decomposition of the deformation gradient and pertinent constitutive relations. We then proceed to present the numerical algorithm and the algorithmic tangent stiness tensor consistent with the presented algorithm.

6.4.2

Kinematics

An appropriate generalization of the additive strain decomposition is the multiplicative decomposition of displacement gradient. The motivation for the multiplicative decomposition can be traced back to the early works of Bilby et al. (1957), and Krner (1960) on micromechanics of crystal dislocations and application to continuum modeling. In the context of large deformation elastoplastic computations, the work by Lee and Liu (1967), Fox (1968) and Lee (1969) stirred an early interest in multiplicative decomposition. The appropriateness of multiplicative decomposition technique for soils may be justied from the particulate nature of the material. From the micromechanical point of view, plastic deformation in soils arises from slipping,

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

149

crushing, yielding and plastic bending19 of granules comprising the assembly20 . It can certainly be argued that deformations in soils are predominantly plastic, however, reversible deformations could develop from the elasticity of soil grains, and could be relatively large when particles are locked in high density specimens.

dx

Current Configuration

Fe

F e

-1

F p

X x

0

dX

X

dx

Figure 6.5: Multiplicative decomposition of deformation gradient: schematics. The reasoning behind multiplicative decomposition is a rather simple one. If an innitesimal neighborhood of a body xi , xi + dxi in an inelastically deformed body is cutout and unloaded to an unstressed conguration, it would be mapped into x i , x i + dx i . The transformation would be comprised of a rigid body displacement21 and purely elastic unloading. The elastic unloading is a ctitious one, since materials with a strong Baushingers eect, unloading will lead to loading in an other stress direction, and, if there are residual stresses, the body must be cutout in small pieces and then every piece relieved of stresses. The unstressed conguration is thus incompatible and discontinuous. The position x i is arbitrary, and we may assume a linear relationship between

19 For

plate like clay particles. also Lambe and Whitman (1979) and Sture (1993). 21 Translation and rotation.

20 See

Jeremi et al.

Lecture Notes

150

e dx k = (Fik )

dxi

(6.140)

e where (Fik )

is not to be understood as a deformation gradient, since it may represent the incompatible, dis-

continuous deformation of a body. By considering the reference conguration of a body dXi , then the connection to the current conguration is23 :

Fij dXj

(6.141)

p e Fkj = (Fik ) def 1 e p Fij Fij = Fki Fkj def

(6.142)

p The plastic part of the deformation gradient, Fkj represents micromechanically, the irreversible process of slipping, e crushing dislocation and macroscopically the irreversible plastic deformation of a body. The elastic part, Fki

represents micromechanically a pure elastic reversal of deformation for the particulate assembly, macroscopically a linear elastic unloading toward a stress free state of the body, not necessarily a compatible, continuous deformation but rather a ctitious elastic unloading of small cut outs of a deformed particulate assembly or continuum body.

6.4.3

Constitutive Relations

e ) represents a suitable hyperelastic model in terms of the elastic right deformation tensor C e , where W e (C ij ij whereas W p ( ) represents the hardening. It has been shown elsewhere (Runesson, 1996), that the pertinent dissipation inequality becomes ij L p + D=T ij

K 0

(6.144)

ij is the Mandel stress24 and L p is the plastic velocity gradient dened on . where T ij We now dene elastic domain B as ij , K | (T ij , K ) 0} B = {T

22 referred 23 See

(6.145)

Jeremi et al.

Lecture Notes

151

ij (which is the case here) in conjunction with elastic isotropy, we can conclude that T ij When is isotropic in T ij by ij in . is symmetrical and we may replace T As to the choice of elastic law, it is emphasized that this is largely a matter of convenience since we shall be dealing with small elastic deformations. Here, the NeoHookean elastic law is adopted. The generic situation ij = T ij (U e , J e ), where we have used the isochoric/volumetric split of the elastic right stretch tensor as is T kl e = U e (J e )1/3 . U

kl kl

1

(6.146) ij = = M Tij

p := F p L ij ik

p Fjk

(6.147)

= K ( K ) K

(6.148)

(0) = 0

(6.149)

6.4.4

The incremental deformation and plastic ow are governed by the system of evolution equations (6.147) and (6.149): p F ik

p Fjk 1

ij = M

(6.150)

(0) = 0

(6.151)

n+1 p Fij

ik nF p = exp n+1M kj

(6.152)

p e e ik ik Fij = F Fkj F = Fij p Fkj 1

(6.153)

n+1 e Fij

= =

n+1

p Fim (nFmk )

Version: March 21, 2012, 17:15

kj exp n+1M

University of California, Davis

Jeremi et al.

Lecture Notes

152

p = n+1Fim (nFkm )

(6.155)

n+1 e Cij

def

n+1 e T n+1 e Fmj Fim T ir exp n+1M T ir exp n+1M n+1 e,tr T n+1 e,tr Frk Fkl n+1 e,tr Crl

= =

lj exp n+1M

By recognizing that the exponent of a tensor can be expanded in Taylor series25 1 ls n+1M 2

sj + n+1M

(6.157)

and by using the rst order expansion in the equation (6.156), we obtain

n+1 e Cij

= = =

ir ir n+1M

n+1 e,tr Cil n+1 e,tr Cij

n+1

ir n+1M

ir n+1M ir M

lj M (6.158)

+2

n+1

Remark 6.4.1 The Taylors series expansion from equation (6.157) is a proper approximation for the general lj . That is, the approximate solution given by equation (6.158) is valid for a general nonsymmetric tensor M anisotropic solid. This contrasts with the spectral decomposition family of solutions26 which are restricted to isotropic solids.

lj . This is Remark 6.4.2 Taylors series expansion27 is proper for small values of plastic ow tensor n+1M indeed the case for small increments, when 0 which are required for following the equilibrium path for pathdependent solids.

25 See 26 See

for example Pearson (1974). Simo (1992). 27 It should be called MacLaurins series expansion, since expansion is about zero plastic ow state (no incremental plastic deformation).

Jeremi et al.

Lecture Notes

153

Remark 6.4.3 In the limit, when the displacements are suciently small, the solution (6.158) collapses to lim ij + 2n+1 ij = + ij + 2n+1 e,tr ij ir rj + 2n+1 e,tr n+1M rj il + 2n+1 e,tr il + = 2

n+1 n+1

Fij ij

lj M

n+1

ir rl + 2n+1 e,tr M rl

lj M

n+1 n+1

lj M ir M

n+1 tr n+1 Mlj rl

il M

n+1

lj + 22 M

n+1

n+1 ij

n+1 tr ij

ij n+1M

(6.159)

which is a small deformation elastic predictorplastic corrector equation in strain space. In working out the small deformation counterpart (6.159) it was used that

Fij ij

lim

n+1 e Cij

= ij + 2n+1 ij

n+1

2 n+1 tr il

lj M

n+1

ij M (6.160)

By neglecting the higher order term with 2 in equation (6.158), the solution for the right elastic deformation e can be written as tensor n+1C ij

n+1 e Cij

e,tr = n+1C ij

n+1

ir M

e,tr + n+1C il

n+1

lj M

(6.161)

n+1

= n +

(6.162)

n+1

Remark 6.4.4 It is interesting to note that equation (6.161) resembles the elastic predictorplastic corrector equation for small deformation elasticplastic incremental analysis. That resemblance will be used to build an iterative solution algorithm in the next section. The incremental problem is dened by equations (6.161), (6.162), and the constitutive relations

n+1 SIJ

=2

W CIJ

(6.163)

n+1

n+1

K =

(6.164)

n+1

Jeremi et al.

Lecture Notes

154

n+1

n+1 = 0

(6.165)

ij can be obtained from the second PiolaKirchho stress tensor S kj Remark 6.4.5 The Mandel stress tensor T e as and the right elastic deformation tensor C

ik e ij = C ik T Skj

(6.167)

This set of nonlinear equations will be solved with a Newton type procedure, described in the next section. For e,tr , the upgraded quantities n+1S IJ and n+1K can be found, then the appropriate a given n+1Fij , or n+1C ij pullback to B0 or pushforward to B will give

n+1

SIJ and

n+1

ij (6.168)

n+1

SIJ =

n+1 p FjJ

n+1

e ij = n+1F iI

n+1 SIJ

n+1 e 1 FjJ

(6.169)

n+1 e Cij

= =

n+1

ir M

e,tr + n+1C il

n+1

lj M (6.170)

n+1Zij

is used as a starting point for a Newton iterative algorithm. In previous equation, we have introduced tensor Zij to shorten writing. The trial right elastic deformation tensor is dened as

n+1 e,tr Cij

p FrM (nFiM )

FrS

n p FjS

(6.171)

e ij Rij = C current n+1 e,tr Cij

n+1Zij

(6.172)

BackwardEuler

Tensor Rij represents the dierence between the current right elastic deformation tensor and the Backward Euler e,tr is maintained xed during the right elastic deformation tensor. The trial right elastic deformation tensor n+1C

ij

Jeremi et al.

Lecture Notes

155

iteration process. The rst order Taylor series expansion can be applied to the equation (6.172) in order to obtain

old new from the old Rij the iterative change, the new residual Rij

(6.173)

By using that

mn = C e S T mk kn

e C sk

sn = S kn T

(6.174)

we can write

mn dT

(6.175)

new Rij =

n+1Zij e e mk sk dC C mn T n+1Zij + dK K

e e L e pq sn + 1 C dC T 2 mk knpq

= +

e sk C

sn + T

e e L knpq dCpq +

= +

(6.176)

Version: March 21, 2012, 17:15

Jeremi et al.

new The goal is to have Rij = 0 so one can write old e ij = Rij + dC + d() n+1Zij +

Lecture Notes

156

+ + +

= +

1

sk dC e + T ij

e mk e e C L knij dCij

= +

(im nj +

n+1Zmn ik T

e sj C 1 n+1 e e pk sk + 1 Zmn C Lkqij T pq 2 T

(6.178)

1 old Rmn d() n+1Zmn

n+1Zmn dK K

(6.179)

e pq dC = (Tmnpq ) 1 old Rmn d() n+1Zmn

n+1Zmn dK K

(6.180)

e pq dC = (Tmnpq ) 1 old Rmn d() n+1Zmn +

n+1Zmn Q d() H K K

(6.182)

Jeremi et al.

Lecture Notes

157

new

ij , K ) = (T + + = + + = + +

old

ij , K ) 1 (T e mk e e C L knpq dCpq mn 2 T ij , K ) (T dK + K old ij , K ) + = (T ij , K ) ij , K ) (T 1 (T e 1 e sq mk e C C L + Tsn + knpq mn 2 Tpn T ij , K ) (T + dK K By using (6.181), equation (6.183) becomes

new

e pq dC

(6.183)

ij , K ) = (T +

old

ij , K ) + (T ij , K ) (T e C sq pn T

e L e sn + 1 (Tij , K ) C T mk knpq mn 2 T

e dC pq (6.184)

d()

ij , K ) (T H K K

e sq C 1

Fpq =

e sn + 1 (Tij , K ) C mk e T L knpq mn 2 T

(6.185)

new

ij , K ) = old (T ij , K ) + (T (Tmnpq )

1 old Rmn d() n+1Zmn + d()

+ Fpq d()

Jeremi et al.

n+1Zmn H K K (6.186)

Version: March 21, 2012, 17:15

ij , K ) (T H K K

University of California, Davis

new

Lecture Notes

158

After setting

d() =

1 old Fpq (Tmnpq ) Rmn n+1 Zmn 1 n+1Z H + H mn Fpq (Tmnpq ) K K K K old 1

Fpq (Tmnpq )

(6.187)

Remark 6.4.6 In the perfectly plastic case, the increment inconsistency parameter d() is

old

d() =

Fpq (Tmnpq )

old Rmn mn

Fpq (Tmnpq )

1 n+1 Z

(6.188) Remark 6.4.7 In the limit, for small deformations, isotropic response, the increment inconsistency parameter d() becomes

old

(nmn Emnpq )

pm nq +

mmn Eijpq ij

n+1

1 old Rmn

mpq Eijmn mp qn + ij

H mmn + K K

(6.189)

since in the limit, as deformations are getting small Tmnpq Fpq Zpq Rpq pm nq + mmn Eijpq ij

1 Emnpq 2 mn 2 mpq 2

pq

(6.190)

Upon noting that residual Rpq is dened in strain space, the increment inconsistency parameter d() compares exactly with its small strain counterpart (Jeremi and Sture, 1997). The procedure described below summarizes the implementation of the return algorithm.

6.4.4.0.1

Trial State

e pq Given the right elastic deformations tensor nC and a set of hardening variables nK at a specic quadrature

n+1

n+1

n+1

fij = ij + ui,j

University of California, Davis

(6.191)

Version: March 21, 2012, 17:15

Jeremi et al.

n+1 e,tr Cij T

Lecture Notes

T T

159

n+1

e fir nFrk

n+1

n+1

fir

n+1

e fkl nFlj

(6.192)

Compute the trial elastic second PiolaKirchho stress and the trial elastic Mandel stress tensor

n+1 e,tr Sij n+1 e,tr Tij

il lj n+1 tr

(6.193) (6.194)

n+1 tr

n+1 e Cij n+1

e,tr , K ). If (T ij

= = =

K Tij

n+1

n e,tr Tij

6.4.4.0.2

Return Algorithm

If yield criteria has been violated (n+1tr > 0) proceed to step 1. step 1. k th iteration. Known variables

n+1 e(k) Cij n+1 (k) n+1 (k ) K n+1 (k) Tij n+1

(k)

(k)

= =

(n+1T ij

e(k) n+1

(k) K )

n+1(k) n+1Zij

(k )

n+1 e Cij n+1 n+1

(k)

= = = =

K Tij

n+1 n+1

(k)

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

(k)

160

step 3.28 If convergence is not achieved, i.e. (k) > N T OL or Rij stiness tensor Lijkl (k) = 4 L ijkl 2W C ij

e( k )

C kl

e(k)

(6.195)

(k ) mn F (k ) mn ( k ) F Rmn (k) (k) (k) mn (k) Zmn (k ) + (k) Zmn (k) (k) F H H K K

(6.196)

where

,(k) (k) = H (k) H K

Tmnpq (k)

Fpq =

, K ) (T ij pn T

(k)

(k)

e,(k) sq C

(k)

(k)

Tmnij = im nj + (k)

Zmn (k) T

ik

(k)

e,(k) C sj

(k )

step 5. Updated the inconsistency parameter (k+1) (k+1) = (k) + d((k+1) ) (6.197)

step 6. Updated the right deformation tensor, the hardening variable and the Mandel stress

e,(k+1) = dC pq

(k) Tmnpq 1 (k) (k ) Rmn d((k+1) ) n+1Zmn + (k)

(k)

k+1) d( = d((k+1) )

,(k) K

(6.199)

(k+1) dK = d((k+1) ) H

28 From

(k)

,(k) K

(6.200)

step 3. to step 9. all of the variables are in intermediate n + 1 conguration. For the sake of brevity we are omitting

superscript n + 1.

Jeremi et al.

1

Lecture Notes

161

(k+1) mn e,(k+1) dT = dC mk

e,(k) C sk

(6.201)

e,(k+1) (k+1) (k+1) pq mn step 7. Update right deformation tensor C , hardening variable K and Mandel stress T e,(k+1) pq C k+1) ( (k+1) K (k+1) mn T e,(k) e,(k+1) pq pq = C + d(C ) k) (k+1) = ( ) + d( (k) (k+1) = K + d(K ) (k) (k+1) mn mn = T + d(T )

(6.202)

step 8. evaluate the yield function and the residual e(k+1) , K (k+1) ) ; R(k+1) = C e,(k+1) C e,tr (k+1) Z (k+1) (k+1) = (T ij ij ij ij ij step 9. Set k = k + 1 (k) = = = = = (k+1)

e,(k+1) pq C k+1) ( (k+1) K (k+1) mn T

(6.203)

e,(k) pq C k) ( (k ) K (k ) mn T

(6.204)

6.4.5

n+1 e Cij

(6.205)

e ij dC

= =

e,tr d() Zij dC ij e,tr d() Zij dC ij Zij e e mn dCmk Csk T Zij dK K

1

sn + 1 C e L e e T dC pq 2 mk knpq

from (6.175)

(6.206)

1

University of California, Davis

(6.207)

Jeremi et al.

Lecture Notes

162

or as

e ij e,tr d() Zij + d() Zij H dC (Tmnij ) = dC ij K K

(6.208)

ik (k)

e,(k) C sj

(k )

The solution for the increment in right elastic deformation tensor is then

e ij dC = (Tmnij ) 1

Zij H K K

(6.209)

ij , K ) = 0 We next use the rst order Taylor series expansion of yield function d(T mn dTmn + K dK T 1 e e e e 1 e mk sk dK dC C Tsn + C mk Lknpq dCpq + 2 K Tmn 1 e e e 1 e e sq pq pq C Tsn dC + Cmk Lknpq dC + dK 2 Tmn K Tpn 1 e e e 1 e sq d ( ) H C C L Tsn + d C pq mn mk knpq 2 T K K e pq Fpq dC d() H K K = = = = = 0 (6.210) where Fpq = pn T e C sq

1

pn T

sn + 1 C e e T mn mk Lknpq 2 T

(6.211)

1 e,tr mn dC d() Zmn + d()

Zmn H K K d() H K K

0 (6.212)

We are now in the position to solve for the incremental inconsistency parameter d() d() = Fpq (Tmnpq )

1

e,tr dC mn

(6.213)

1 n+1

n+1Zmn H + H K K K K (6.214)

Jeremi et al.

Lecture Notes

163

e pq dC =

(Tmnpq ) Then

mv nt

1

Fop (Trsop )

rv st

Zmn +

e,tr vt dC

Fop (Trsop )

rv st Zij H K K

(6.216)

e,tr e = P pqvt dC vt dC pq

(6.217)

where

pqvt = P + =

(Tmnpq )

mv nt

1

Fop (Trsop )

(Tmnpq )1

mv nt

ijkl (in intermediate conguration ) is then dened as Algorithmic tangent stiness tensor L

S AT e L knvt = Lknpq Ppqvt

(6.219)

Pullback to the reference conguration 0 yields the algorithmic tangent stiness tensor Lijkl in reference conguration 0

n+1 AT S Lijkl p = n+1Fim n+1 p n+1 p n+1 p n+1 AT S Fjn Fkr Fls Lmnrs

(6.220)

Remark 6.4.8 In the limit, for small deformations, isotropic response, the Algorithmic Tangent Stiness tensor

S LAT ijkl becomes 1 ncd Ecdab vtab Hmn

1 Eknpq mnpq mv nt

= = =

Jeremi et al.

Eknpq

1 vtpq

1 1 Eknpq mrpq ncd Ecdab vtab Hmr ncd Rcdvt Rknmr Hmr Rknvt 1 Eknpq vtpq

(6.221)

Version: March 21, 2012, 17:15

Lecture Notes

164

since mnpq = mnpq = lim T = pm nq + Zmn e 1 1 Zmn e e Csq Tsk + C L 2 Tpk Trs rk kspq mmn e pm nq + E rs kspq

(6.222)

lim Fab = =

lim

e ad Csb T

sd + 1 C e e T cd ck Lkdab 2 T (6.223)

Hmn = mmn

(6.224)

lim = = = =

n+1Zmn H + H K K K K m mn 1 1 nab Eabpq H + H mnpq mmn nab Eabpq mnpq K K K K m mn 1 nab Eabpq H + H mnpq mmn K K K K H (6.225) nab Rabmn Hmn + K K

1 n+1

It is noted that the Algorithmic Tangent Stiness tensor given by 6.221 compares exactly with its small strain counterpart (Jeremi and Sture, 1997).

6.5

6.5.1

Introduction

We here present a detailed formulation of a material and geometric nonlinear static nite element system of equations. The conguration of choice is material or Lagrangian. Eulerian and mixed EulerianLagrangian conguration will be mentioned as need be.

6.5.2

Equilibrium Equations

The local form of equilibrium equations in material format (Lagrangian) for static case can be written as: PiJ,J 0 bi = 0 (6.226)

where PiJ = SIJ (FiI )t and SIJ are rst and second PiolaKirchho stress tensors, respectively and bI are body forces.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

165

Weak form of equilibrium equations is obtained by premultiplying 6.226 with virtual displacements ui and integrating by parts on the initial conguration B0 (initial volume V0 ): ui,j Pij dV =

V0 V0

0 ui bi dV

S0

i dS ui t

(6.227)

It proves benecial to rewrite Lagrangian format of weak form of equilibrium equilibrium by using symmetric second PiolaKirchho stress tensor Sij : ui,j Fjl Sil dV

V0

= = = = = (6.228)

V0

V0

1 (ui,l + ui, juj,l ) + (ul,i + ul,j uj,i ) Sil dV V0 2 1 ((ui,l + ul,i ) + (ui,j uj,l + ul,j uj,i )) Sil dV V0 2

where we have used the symmetry of Sil , denition for deformation gradient Fki = ki + uk,i . We have also il (ui , ui ) conveniently dened dierential operator E

(6.229)

6.5.3

Consider the motion of a general body in a stationary Cartesian coordinate system, as shown in Figure (6.6), and assume that the body can experience large displacements, large strains, and nonlinear constitutive response. The aim is to evaluate the equilibrium positions of the complete body at discrete time points 0, t, 2t, . . . , where t is an increment in time. To develop the solution strategy, assume that the solutions for the static and kinematic variables for all time steps from 0 to time t inclusive, have been obtained. Then the solution process for the next required equilibrium position corresponding to time t + t is typical and would be applied repetitively until a complete solution path has been found. Hence, in the analysis one follows all particles of the body in their motion, from the original to the nal conguration of the body. In so doing, we have adopted a Lagrangian ( or material ) formulation of the problem. Weak format of the equilibrium equations can be obtained by premultiplying 6.226 with virtual displacements ui and integrating by parts. We obtain the virtual work equations in the Lagrangian format: ui,j Pij dV =

V0 V0

0 ui bi dV

S0

i dV ui t

University of California, Davis

(6.230)

Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

166

x2 , tx2 , x2

t+ t 0 0 0

t P( t+tx1t+ , x2 , x3 ) t+ t

t t

t+t t+t

A V

A V

0 0

A V

Configuration at time t

n+1

Configuration at time 0

0

t+t t

x3 ,tx3 , x3 t+ t

xi = x i + u i x i = 0xi + t ui t+t xi = t xi + ui

t+t

x1 ,t x1,t+ tx1

i=1,2,3

Figure 6.6: Motion of body in stationary Cartesian coordinate system Virtual work equations can also be written in terms of second PiolaKirchho stress tensor SIJ as: ui,j Fjl Sil dV =

V0 V0

0 ui bi dV +

S0

i dV ui t

(6.231)

which after some algebraic manipulations, and after observing that SIJ = SJI yields (SEE ABOVE!) By intro (u1 , u2 ) as: ducing a dierential operator E il (1ui , 2ui ) = 1 E 2

1

ui,l + 1ul,i +

1 2

(6.232)

V0

=

V0

0 ui bi dV +

S0

i dV ui t

(6.233)

(k) (k )

(6.234)

=

c

ij (ui , n+1u(k) ) n+1S (k) dV E 0 i 0 ij ((uj,i + ui,j ) + (uj,r ur,i + ui,r ur,j )) Sij dV

c (k)

(6.235)

=

Jeremi et al.

Lecture Notes

167

W ext (ui )

=

c

0 ui

n+1 0 bi

dV

c

ui

n+1 0 ti

dS

(6.236)

6.5.4

In this Chapter we elaborate on the choice of Total Lagrangian (TL) formulations as computational domain. We also choose Newton type procedure for satisfying equilibrium, i.e. virtual work for a given computational domain. Given the displacement eld ui (Xj ), in iteration k , the iterative change ui ui

(k+1) (k )

= ui

(k)

+ ui

(6.237)

(k+1)

W (ui , ui ) + W (ui , ui ; ui )

(k)

(k)

(6.238)

Here, W (ui , ui ) is the virtual work expression W (ui , ui ) = W (ui , ui )int + W ext (ui )

(k ) (k)

(k )

(6.239)

(k )

=

c

(6.240)

W ext (ui )

=

c (k )

0 ui

n+1 0 bi

dV

c

ui

n+1 0 ti

dS

(6.241)

(k )

= =

lim

c

=

c

c

kl (ui , ui ). Here we have used dSij = 1/2 Lijkl dCkl = Lijkl E In order to obtain expressions for stiness matrix we shall work on 6.242 in some more details. To this end, as (6.242) can be rewritten by expanding denitions for E W (ui , ui ; ui ) = 1 4 ((uj,i + ui,j ) + (uj,r ur,i + ui,r ur,j )) Lijkl ((uk,l + ul,k ) + (uk,s us,l + ul,s us,k )) dV +

c (k )

+

c

Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

168

Or, by conveniently splitting the above equation we can write 1W (ui , ui ; ui ) = 1 4 ((uj,i + ui,j ) + (uj,r ur,i + ui,r ur,j )) Lijkl ((uk,l + ul,k ) + (uk,s us,l + ul,s us,k )) dV

c (k)

(6.244)

2W (ui , ui ; ui ) =

c

(k)

(6.245)

c (k)

+

c

+

c

+

c

1 (uk,l + ul,k ) dV 2 1 Lijkl (uk,s us,l + ul,s us,k ) dV 2 1 1 (uj,r ur,i + ui,r ur,j ) Lijkl (uk,s us,l + ul,s us,k ) dV 2 2 1 1 (uj,r ur,i + ui,r ur,j ) Lijkl (uk,l + ul,k ) dV 2 2 1 (uj,i + ui,j ) 2 1 (uj,i + ui,j ) 2 Lijkl

(6.246)

It should be noted that the Algorithmic Tangent Stiness (ATS) tensor Lijkl poses both minor symmetries (Lijkl = Ljikl = Lijlk ). However, Major symmetry cannot be guaranteed. Nonassociated ow rules in elastoplasticity lead to the loss of major symmetry (Lijkl =Lklij ). Moreover, it can be shown (i.e. Jeremi and Sture (1997)) that there is algorithmic induced symmetry loss even for associated ow rules. With the minor symmetry of Lijkl one can write (6.246) as: 1W (ui , ui ; ui ) =

c (k)

c

c

+ +

c

(6.247)

Similarly, by observing symmetry of second PiolaKirchho stress tensor Sij we can write 2W (ui , ui ; ui ) =

c (k)

(6.248)

Weak form of equilibrium expressions (i.e. (6.236) and (6.236) ) for internal (W int ) and external (W ext ) virtual work, with the above mentioned symmetry of Sij can be written as W int (ui , n+1 0 ui ) =

c (k)

ui,j Sij dV +

c

(6.249)

Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

169

W ext (ui )

=

c

0 ui bi dV

c

ui ti dS

(6.250)

where u i is the approximation to exact, analytic (if it exists) displacement eld ui , HI are standard FEM shape functions and u Ii are nodal displacements. With this approximation, we have: 1W (ui , ui ; ui ) =

c (k )

c

(HI,j u Ii ) Lijkl (HJ,k u Js ) (HQ,s uQl ) dV (HI,r u Ii ) (HJ,j u Jr ) Lijkl (HJ,k u Js ) (HQ,s uQl ) dV

c

+ +

c

(6.252)

2W (ui , ui ; ui ) =

c

(k)

(6.253)

c

(k)

(HI,j u Ii ) Sij dV +

c

(6.254)

W ext (ui )

0 (HI u Ii ) bi dV

c

(HI u Ii ) ti dS

(6.255)

Upon noting that virtual nodal displacements uIi are any nonzero, continuous displacements, and since they occur in all expressions for linearized virtual work (from Equations (6.238), (6.239), (6.240), (6.241) and (6.242)) they can be factored out so that we can write (while remembering that W 1 + W 2 + W ext + W int = 0: (HI,j ) Lijkl (HQ,k uQl ) dV

c

+

c

(HI,j ) Lijkl (HJ,k u Js ) (HQ,s uQl ) dV (HI,r ) (HJ,j u Jr ) Lijkl (HJ,k u Js ) (HQ,s uQl ) dV

c

+ +

c

c

+ +

c

(HI,j ) Sij dV +

c

c

=

c

0 (HI ) bi dV +

(6.256)

Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

170

c c

c

+

c

c

uQl

+

c

(HI,j ) Sij dV +

c

c

=

c

0 (HI ) bi dV +

(6.257)

The vectors of external and internal forces are (W int (ui , n+1 0 ui )) (ui ) (W ext (ui )) (ui )

(k )

fint fext

= =

(6.258) (6.259)

S The Algorithmic Tangent Stiness (ATS) tensor LAT ijkl is dened as a linearization of second PiolaKirchho

stress tensor Sij with respect to the right deformation tensor Ckl 1 kl (dui , ui ) Lijkl dCkl with dCkl = 2 E 2

dSij =

(6.260)

Then, the global algorithmic tangent stiness matrix (tensor) is given as (W (ui , ui ; ui )) (ui )

(k)

Kt =

(6.261)

The iterative change in displacement vector ui is obtained by setting a linearized virtual work to zero

(k+1) (k ) (k)

W (ui , ui

Jeremi et al.

) = 0 W (ui , ui ) = W (ui , ui ; ui )

University of California, Davis

(6.262)

Version: March 21, 2012, 17:15

Lecture Notes

171

6.5.4.1

The undeformed conguration 0 is chosen as the computational domain (c = 0 ). The iterative displacement ui is obtained from the equation W (ui , n+1ui ) = W (ui , ui ; n+1ui ) where W (ui , n+1ui )

(k) (k) (k )

(6.263)

=

c

c

0 ui

n+1

bi dV

c

ui

n+1

ti dS

(6.264)

(k)

=

c

c (k)

(6.265)

constitutive law, described in Chapter 6.4. It should be noted that by integrating in the intermediate conguration, ij and subsequently29 the second PiolaKirchho stress S kj . The ATS tensor L ijkl we obtain Mandel stress n+1T kj . In order to obtain second PiolaKirchho stress Skj and ATS tensor in initial is then obtained based on S conguration we need to perform a pull-back from the intermediate conguration to the initial one

n+1 p Sij = n+1Fip n+1 p n+1 Fjq Spq

(6.266)

n+1

p Lijkl = n+1Fim

(6.267)

6.5.5

Presented here is a slightly dierent approach to developing large deformation FEM in total Lagrangian form. Lower case indices are used for variables in current conguration, while the capital case indices are used for the reference conguration.

6.5.5.1

Strong Form

29 S kj

(6.268)

ik = C

ij T

Jeremi et al.

Lecture Notes

172

where ij is the Cauchy stress, rho is the material density, bi is the material body force. Here we use so-called Total Lagrangian formulation that is based on the reference conguration. The strong form of momentum balance can be expressed in the reference conguration PiJ + 0 bi = 0 XJ where PiJ is the rst Pialo-Kirchho stress, 0 is the material Lagrangian density and 0 = J. (6.269)

6.5.5.2

Weak Form

0

PiJ + 0 bi dV = 0 XJ

(6.270)

where ui is some arbitrary virtual displacement, 0 is the concerned domain of the reference conguration. Using the partial integration rule, the above equation can be alternatively expressed as (PiJ ui ) dV XJ PiJ

0

ui dV + XJ

0 bi ui dV = 0

0

(6.271)

The rst term of Equation 6.271 can be rewritten in terms of the surface traction (PiJ ui ) dV = XJ ui PiJ NJ dA =

0 0

ti ui dA

(6.272)

with the traction ti = PiJ NJ , where NJ the unit surface normal vector in the reference conguration, and 0 is the boundary of the reference domain 0 . The second term of Equation 6.271 can be rewritten as PiJ

0

ui dV = XJ

PiJ FiJ dV =

0 0

SIJ EIJ dV

(6.273)

where SIJ is the second Pialo-Kirchho stress and EIJ is the Lagrangian-Green strain. FiJ = ui /XJ is used. The overall weak form in the reference conguration is now SIJ EIJ dV =

0 0

ti ui dA +

0

0 bi ui dV

(6.274)

6.5.5.3

Linearized Form

To utilize the iterative algorithm for incremental strategy, we still need the linearized form the governing equation 6.274. The rst term linearization of Equation 6.274 is

0

SIJ EIJ

=

0

0

=

Jeremi et al.

(6.275)

Version: March 21, 2012, 17:15

Lecture Notes

173

where LIJKL is the Lagrangian stiness linked the second Pialo-Kirchho stress SIJ and the Lagrangian-Green strain EKL by the relation SIJ = LIJKL EKL The linearization of EKL is EKL = Sym FK ub XL (6.277) (6.276)

where Sym is the operator of tensor symmetry, dened as Sym(Aij ) := (1/2)(Aij + Aji ). Similarly, EIJ = Sym FaI and (EIJ ) = Sym uc uc XI XJ (6.279) ua XJ (6.278)

Note that SIJ is a symmetric tensor, and LIJKL is a tensor with major and minor symmetries, Equation 6.275 can be expressed as

0

SIJ EIJ

=

0

LIJKL FbK

ub XL

FaI

ua XJ

+ SIJ

uc uc XI XJ

dV (6.280)

=

0

0

ti ui dA +

0 0

0 bi ui dV

(6.281)

6.5.5.4

In nite element form, the displacements ui are interpolated from the element nodal displacements u Ai : ui = HA u Ai where HA is the element shape function of the node A, u Ai is the node A displacements, and ua HA = u Aa , XJ XJ ub HB = uBb XL XL (6.283) (6.282)

ex (KAaBb uBb fAa ) u Aa = 0

(6.284)

where KAaBb

ex fAa

=

0

0

(6.285) (6.286)

=

0

Due to the arbitrariness of the virtual nodal displacements, the expression in the parentheses should be zero in Equation 6.284, which gives the incremental nite element form:

ex KAaBb uBb = fAa

(6.287)

University of California, Davis Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

174

Jeremi et al.

Chapter 7

(1994-)

175

Lecture Notes

176

7.1 7.2

In previous Chapters we have derived the basic equations for (material and/or geometric) nonlinear analysis of solids. Discretization of such problems by nite element methods results in a set of nonlinear algebraic equations called residual force equations: r(u, ) = fint (u) fext = 0 (7.1)

where fint (u) are the internal forces which are functions of the displacements, u, the vector fext is a xed external loading vector and the scalar is a loadlevel parameter that multiplies fext . Equation (7.1) describes the case of proportional loading in which the loading pattern is kept xed. All solution procedures of practical importance are strongly rooted in the idea of "advancing the solution" by continuation. Except in very simple problems, the continuation process is multilevel and involves hierarchical breakdown into stages, incremental steps and iterative steps. Processing a complex nonlinear problem generally involves performing a series of analysis stages. Multiple control parameters are not varied independently in each stage and may therefore be characterized by a single stage control parameter . Stages are only weakly coupled in the sense that end solution of one may provide the starting point for another.

7.3

Various forms of path following methods1 have stemmed from the original work of Riks (1972), Riks (1979) and Wempner (1971). They aimed at nding the intersection of equation (7.1) with s = constant where s is the arc-length , dened as2 :

s= where:

ds

(7.2)

ds =

(7.3)

a = (s)2 (l)2 =

1 also 2A

2 u 2 uT Su + 2 f 2 uref

(l)

(7.4)

called arc-length methods with various methods of approximating the exact length of an arc. bit dierent form in that it is scaled with scaling matrix S, introduced by Felippa (1984).

Jeremi et al.

Lecture Notes

177

where l is the radius of the desired intersection3 and represents an approximation to the incremental arc length. Scaling matrix S is usually diagonal non-negative matrix that scales the state vector u and uref is a reference value with the dimension of uT Su. It is important to note that the vector u and scalar are incremental and not iterative values, and are starting from the last converged equilibrium state.

Load f

1 f ext 1 f ext 2 fext ( u2 , 2 fext) ( u3 , 3 fext )

2 f ext 3 fext

0 f ext

Equilibrium Path

Displacement

u u 1

0

u 2

u 3

Figure 7.1: Spherical arc-length method and notation for one DOF system. The main essence of the arc-length methods is that the load parameter becomes a variable. With load parameter variable we are dealing with n + 1 unknowns4 . In order to solve this problem we have n equilibrium equations (7.1) and the one constraint equation (7.4). We can solve the augmented system of n + 1 equations by applying the Newton-Raphson5 method to equations (7.1) and (7.4)

rnew (u, )

= = =

Figure (7.1). unknown displacement variables and on extra unknown in the form of load parameter. 5 By using a truncated Taylor series expansion.

4n

3 See

Jeremi et al.

Lecture Notes

178

anew = aold + 2

2 u 2 uT S u + 2 f =0 u2 ref

(7.6)

where Kt =

r(u,) u

is the tangent stiness matrix. The aim is to have rnew (u, ) = 0 and anew = 0 so the

Kt

ref

fext

2 2 u2u uT S 2 f

rold aold

(7.7)

Kt 2 u2u uT S

ref

fext

2 2 f

rold aold

(7.8)

K=

2

Kt 2 u2u uT S

ref

fext

2 2 f

(7.9)

= K1

rold aold

(7.10)

It should be mentioned that the augmented stiness matrix remains non-singular even if Kt is singular.

In section (7.3) we have introduced an constraining equation that is intended to reduce the so called drift error in the incremental nonlinear nite element procedure. The constraining equation was given in a rather general

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

179

Load f f = u

f < u

f << u

Load f f = u

f >> u f > u

Equilibrium Path Constraint Hypersurfaces

Equilibrium Path

l

Constraint Hypersurfaces

Displacement u

Displacement u

form. Some further comments and observations are in order. By assigning various numbers to parameters u , f , S and uref one can obtain dierent constraining schemes from (7.4). Coecients u and f may not be simultaneously zero. Useful choices for S are I, Kt and diag (Kt ). If S = I and uref = 1 the method is called the arclength method7 . If we choose S = diag (Kt ) nice scaling is obtained8 but otherwise no physical meaning can be attributed to this scaling type. With S = Kt and f 0 one ends up with something very similar to the external work constraint of Bathe and Dvorkin (1983). A rather general equation (7.4) can be further specialized to load () control with u 0; f 1 and state control9 with u 1; f 0 and S = I. In the nite element literature, the term displacement control has been traditionally associated with the case in which only one of the components of the displacement vector u10 is specied. This may be regarded either as a variant of state control, in which a norm that singles out the ith component is used, or as a variant of the control if the control parameter is taken as ui . It is, of course, possible to make the previous parameters variable, functions of dierent unknowns. For example if one denes uref = uT Su then close to the limit point u 0

2 u u2 ref

2 f that makes our constraint from equation (7.4) behave like state

control. One important aspect of scaling constraint equations by using S = diag (Kt ) or S = Kt is the possibility of nonpositive denite stiness matrix Kt . It usually happens that after the limit point is passed, at least one of the eigenvalues of Kt is nonpositive, thus rendering the constraint hypersurface nonconvex. In order to get better control of the solution to the system of equations (7.10) one may directly introduce the constraint from equation (7.6) by following the approach proposed by Batoz and Dhatt (1979), as described by Criseld (1991) and Felippa (1993).

6 Or 7 It

augmented Jacobian. actually reduces to the original work of fRiks (1972), Riks (1979) and Wempner (1971). 8 For example if FEM model includes both translational and rotational DOFs. 9 That is the cylindrical constraint, or general displacement control. 10 Say u . i

Jeremi et al.

Lecture Notes

180

The iterative displacement u is split into two parts, and with the Newton change at the new unknown load level:

new = old +

(7.11)

becomes:

1 is an iterative where ut = K t fext is the displacement vector corresponding to the xed load vector fext , and u

change that would stem from the standard load-controlled Newton-Raphson, at a xed load level old . With the solution11 for the u from (7.12), the new incremental displacements are:

(7.13)

where is the only unknown. The constraint from equation (7.4) can be used here, and by rewriting it as:

= (l)

(7.14)

then by substituting unew from equation (7.13) into equation (7.14) and by recalling that new = old + one ends up with the following quadratic scalar equation:

2 u + ut uold + u u2 ref

+ ut + old + S uold + u

2 f

= (l)

(7.15)

11 But

Jeremi et al.

Lecture Notes

181

2 u 2 uT S ut + f 2 uref t

+2

2 u uold + u 2 uref T

+ (7.16)

l2 + old S uold + u

2 f

=0

or:

a1 2 + 2a2 + a3 = 0 where:

(7.17)

a1 2a2 a3

= = =

2 u 2 uT S ut + f 2 uref t

2 u uold + u 2 uref

l2 + old S uold + u

2 f

a2 +

a2 2 a1 a3 a1

= 2 =

a2

a2 2 a1 a3 a1

(7.18)

(7.19)

+ ut unew = uold + u

(7.20)

An ambiguity is introduced in the solution for from (7.18). The tangent at the regular point on the equilibrium path has two possible directions, which generally intersect the constraint hypersurface at two points. However, some exceptions from that rule are possible, so the solutions from (7.18) can be categorized as:

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

182

Real roots of opposite signs. This occurs when the iteration process converges normally and there is no limit or turning point enclosed by the constraint hypersurface. The root is chosen by applying one of the schemes proposed below. Real roots of equal sign opposite to that of old . This usually happens when going over a "at" limit point. Real roots of equal sign same as that of old . This is an unusual case. It may signal a turning point or be triggered by erratic iteration behavior. Complex roots. This is an unusual case too. It may signal a sharp turning point, a bifurcation point, erratic or divergent iterates. For the rst two cases, the correct can be chosen by applying one of the following schemes.

7.6.1

7.6.1.1

Positive External Work

The simplest rule requires that the external work expenditure over the predictor step be positive:

1 T T W = fext u = fext K t fext > 0

(7.21)

1 T The simple conclusion is that should have the sign of fext K t fext . This condition is particularly eective at 1 limit points. However, it fails when fext and K t fext are orthogonal: 1 T fext K t fext = 0

(7.22)

This can happen at: Bifurcation points, Turning points, The treatment of bifurcation points is of a rather special nature and is left for the near future. Turning points12 can be traversed by a modication of a previous rule, as described in the next section.

7.6.1.2

Angle Criterion

Near a turning point application of the positive work rule (7.21) causes the path to double back upon itself. When it crosses the turning point it reverses so the turning point becomes impassable. Physically, a positive work rule

12 One

might ask "why treating turning points in a material nonlinear analysis?". The answer is rather simple: "try to prevent all

unnecessary surprises". For a good account of some of surprises in material nonlinear analysis one might take a look at some examples in Criseld (1991) pp. 270.

Jeremi et al.

Lecture Notes

183

Load f

Limit Point

Primary Equilibrium Path Turning Point Bifurcation Point Secondary Equilibrium Path

Displacement u

Figure 7.3: Simple illustration of Bifurcation and Turning point. is incorrect because in passing a turning point the structure releases external work until another turning point is

encountered. To pass a turning point imposing a condition on the angle of the prediction vector proves more eective. The idea is to compute both solutions 1 and 2 and then both pnew and unew : 1 1

+ 1 ut unew = uold + u 1

(7.23)

+ 2 ut unew = uold + u 2

(7.24)

The one that lies closest to the old incremental step direction uold is the one sought. This should prevent the solution from double backing. The procedure can be implemented by nding the solution with the minimum angle between uold and pnew , and hence the maximum cosine of the angle:

cos =

uold uold

+ ut uold + u old + ut u + u

(7.25)

1 old 1 = K where u and ut = K t r t fext . Once the turning point has been crossed, the work criterion should

be reversed so the external work is negative. By directly introducing the constraint from equation (7.6) and following the method through equations (7.12) to (7.25) a limitation is introduced. Precisely at the limit point, Kt is singular and the equations cannot be solved.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

184

However, Batoz and Dhatt (1979) and Criseld (1991) report that no such problem has occurred, because one appears never to arrive precisely at limit point.

7.6.2

Predictor step

The predictor solution is achieved by applying one forward Euler, explicit step from the last obtained equilibrium point:

1 1 up = K t qe = p Kt fext = p ut

(7.26)

where Kt is the tangent stiness matrix at the beginning of increment. Substituting equation (7.26) into the constraint equation (7.14) one obtains:

=

2

2 p

2 u 2 uT S ut + f 2 uref t

(l)

(7.27)

p =

2 u u2 ref

l

2 | uT t S ut | + f

(7.28)

where l > 0 is the given increment length. The absolute value of | uT t S ut | is needed if the stiness matrix is chosen as a scaling matrix, i.e. S = Kt , since, after passing limit point, the stiness matrix is nonpositive denite so uT t S ut 0. The question of choosing the right sign + or in (7.28) is still a vigorous research topic. In the simplied procedure13 the negative sign is chosen with respect to the occurrence of one negative pivot during factorization of the tangent stiness matrix Kt . If more than one pivot happens to be negative, one is advised14 to stop the analysis and try to restart from previously converged solution with smaller step size.

13 Which 14 For

is not guaranteed to work if one takes into account bifurcation phenomena. more details see Criseld (1991).

Jeremi et al.

Lecture Notes

185

7.6.3

Automatic Increments

A number of workers have advocated dierent strategies for controlling the step length size. In this work we will follow the strategy advocated by Criseld (1991). The idea is to nd the new incremental length by applying:

lnew = lold

Idesired Iold

(7.29)

where lold is the old incremental factor for which Iold iterations were required, Idesired is the input, desired number of iterations15 and the parameter n is set to

1 2

7.6.4

Convergence Criteria

Introduction of an iterative scheme calls for the introduction of an iteration termination test. There are several convergence criteria that can be applied. Displacement Convergence Criteria. The change in the last correction u of the state vector u, measured

in an appropriate norm, should not exceed a given tolerance the termination criteria can be written as: u

scaled

u.

( u) S ( u)

(7.30)

Scaling matrix S is used in order to ensure that for a problem involving mixed variables17 , all parameters

1 have the same dimension. Here, an obvious choice for the scaling matrix is S = diag (K t ). If, on the

other hand we dont have mixed variables in state vector u then the simplest choice for scaling matrix is S = I. Residual Convergence Criteria. Since the residual r(u, ) measures the departure from the equilibrium path, an appropriate convergence test would be to compare Euclidean norm of residual with some predened tolerance: r(u, )

scaled

(r) S (r)

(7.31)

Here, an obvious choice for scaling matrix is S = diag (Kt ) Energy Based Convergence Criteria. The two previous convergence criteria can be combined in a single work change criterion: ( u) (r) =

T

( u) (r)

u r

(7.32)

A word of caution is appropriate at this point. As pointed out by Criseld (1991), it follows that: ( u) (r) = ( u)

15 Say 16 The

1 K t Kt (r) = ( u) Kt ( u)

u r

(7.33)

Idesired 3. so called 2norm. 17 For example, if rotations and displacements are involved.

Jeremi et al.

Lecture Notes

186

1 where the iterative change was ( u) = K t r. It should be noted that equations (7.33) give a measure of

the stiness of Kt . This merely implies that a stationary energy position has been reached in the current iterative direction, u. This can occur when the solution is still far away from equilibrium. Since u and r usually have physical dimensions, so do necessarily work with ratios that render the norm of total displacement u and and

r.

of NewtonRaphson iteration this dependency on physical units is undesirable and it is more convenient to

u r

scaled

scaled :

u scaled u scaled

(7.34)

(7.35)

Another important thing to be considered is Divergence. The NewtonRaphson scheme is not guaranteed to converge. Some sort of divergence detection scheme is therefor necessary in order to interrupt an erroneous iteration cycle. Divergence can be diagnosed if either of following inequalities occur: u scaled gu u scaled (7.36)

r scaled gr rpredictor scaled where gu and gr are dangerous growth factors that can be set to, for example gu = gr = 100.

(7.37)

In some cases, the NewtonRaphson iteration scheme will neither diverge nor converge, but rather exhibit oscillatory behavior. To avoid excessive bouncing around, a good practice is to put upper limit to the number of iterations performed in one iteration cycle. Typical limits to the iteration number are 20 to 50.

7.6.5

The progress of the scheme will be briey described, in relation with the Figure (7.1). The procedure starts from a previously converged solution (u0 , 0 fext ). An incremental, tangential predictor step u1 , 1 is obtained18 and the next point obtained is (u1 , 1 fext ). The rst iteration would then use quadratic equation 7.17 where constants a1 , a2 and a3 should be computed with uold = u1 and old = 1 , to calculate 1 and19

18 As 19 From

Jeremi et al.

Lecture Notes

187

1 1 20 u1 = K would lead to: t r (u1 , 1 ) + 1 Kt fext . After these values are calculated, the updating procedure

2 = 1 + 1

and

u2 = u1 + u1

(7.38)

When added to the displacements u0 and load level 0 , at the end of the previous increment this process would lead to the next point (u2 , 2 fext ). The next iteration would then again use quadratic equation 7.17 where constants a1 , a2 and a3 should be computed + 2 ut . After these values are calculated, with uold = u2 and old = 2 , to calculate 3 and u3 = u the updating procedure would lead to:

3 = 2 + 2

and

u3 = u2 + u2

(7.39)

When added to the displacements u0 and load level 0 , at the end of the previous increment this process would lead to the next point (u3 , 3 fext ).

20 See

Jeremi et al.

Lecture Notes

188

Jeremi et al.

Chapter 8

(1989-2006-)

189

Lecture Notes

190

8.1 8.2

(see section 2.2 on page 30). Great reading on this subject is also a book by Argyris and Mlejnek (1991b).

8.3

8.3.1

Newmark Integrator

The Newmark time integration method (Newmark, 1959a) uses two parameters, and , and is dened by the following two equations:

n+1 n+1

x = x =

1 x + t n x + 2 t [( ) n x + 2 n n n+1 x + t [(1 ) x + x ]

n+1

x ]

(8.1) (8.2)

These equatons give relationship between knowns variables at time step n to the unknown variables at next time step n + 1. Method is in general an implicit one, except when = 0 and = 1/2. There are several possible implementation methods for Newmark Integrator. One of them, the predictor-corrector method is dened through a predictors:

n+1

= =

n+1

1 x + t n x + t2 ( ) n x 2 n x + t (1 ) n x

n

(8.3) (8.4)

n+1 n+1

x = x =

n+1 n+1

x + t2

n+1

n+1

x + t

n+1

n+1

n+1

x,

n+1

R=M

n+1

x +C

n+1

x +K

n+1

(8.7)

Jeremi et al. University of California, Davis

(8.8)

Version: March 21, 2012, 17:15

Lecture Notes

191

Equation 8.3 to 8.8 constitute an iterative solving procedure. An alternative Newmark integration approach is to use displacement as the basic unknowns, and the following dierence relations are used to relate

n+1 n+1

x and

n+1

x to

n+1

n n

x x

= =

n+1

t 1 t 2

n+1

x nx + 1 x nx

x + 1

(8.9) (8.10)

n+1

1 n 1 x + 1 t 2

n+1

x x

n+1

n n x+ 1 x + 1 t 2 1 n 1 n 1 n = + 1 x 2 x t x 2 t =

(8.11) (8.12)

n+1

x x

n+1

n+1

x +

(8.13) (8.14)

The Newton integration method becomes M C + + K x = n+1 R 2 t t Equation 8.11 to 8.15 constitute an iterative solving procedure (Argyris and Mlejnek, 1991b). If the parameters and satisfy 1 , 2 1 1 ( + )2 4 2 (8.16) (8.15)

the procedureit is unconditionally stable and second-order accurate. Any value greater than 0.5 will introduce numerical damping. Well-known members of the Newmark time integration method family include: trapezoidal rule or average acceleration method for = 1/4 and = 1/2, linear acceleration method for = 1/6 and = 1/2, and (explicit) central dierence method for = 0 and = 1/2. If and only if = 1/2, the accuracy is second-order Hughes (1987).

8.3.2

HHT Integrator

Numerical damping introduced in the Newmark time integration method will degrade the order of accuracy. The Hilber-Hughes-Tailor (HHT) time integration -method (Hilber et al., 1977a), (Hughes and Liu, 1978a) and (Hughes and Liu, 1978b) using an alternative residual form by introducing an addition parameter to improve the performance:

n+1

R=M

n+1

x + (1 + )F (n+1 x, n+1 x) F (n x, n x)

n+1

(8.17)

Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

192

but retaining the Newmark nite-dierence formulas 8.1 and 8.2 or 8.9 and 8.10. If = 0, Equation 8.17 is reduced to Equation 8.7, and this special case of the HHT time integration method is exactly the Newmark time integration method. Decreasing value increase numerical dissipation (Hughes, 1987). The iteration method for HHT time integration is similar to that of Newmark time integration. Due to the change of Equation 8.17, Equations 8.8 and 8.15 respectively become M + (1 + ) tC + (1 + ) t2 K x = n+1 R for acceleration iteration and M (1 + )C + + (1 + ) x = n+1 R t2 t for displacement iteration. If the parameters , and satisfy 1/3 0, = 1 (1 2), 2 = 1 (1 )2 4 (8.20) (8.19) (8.18)

it is unconditionally stable and second-order accurate (Hughes, 1987). There are dierent denotation meaning of the parameter , e.g. the parameter in the HHT integrator codes of OpenSees equals to the conventional plus one.

Jeremi et al.

Chapter 9

(1999-2005-)

(In collaboration with Dr. Zhao Cheng and Prof. Mahdi Taiebat)

193

Lecture Notes

194

9.1 9.2

9.2.1

Background

For single-phase material encountered in structural mechanics, the response under ultimate load (failure) can be predicted using very simple calculations, at least for static problems. But for soil mechanics, simple, limit-load calculations can not be fully justied under static situation. However, for problems of soil dynamics, the use of simplied methods is almost never admissible. As the strength of the soil can be determined once the pore water pressures are known, it is possible to reduce the soil mechanics problem to that of the behavior a single phase. Then we can use again the simple, single-phase analysis approaches. Now we will introduce the concept of eective stress. The relationship between eective stress, total stress and pore pressure is (assume tensile components of stress as positive and compressive pressure, p is positive) (Zienkiewicz et al., 1999a)

ij = ij + ij p

(9.1)

where ij is eective stress tensor, ij is total stress tensor, ij is Kronecker delta. ij = 1, when i=j, and ij = 0, when i = j . For isotropic materials, = 1 KT /KS . KT is the total bulk modulus of the solid matrix, KS is the bulk modulus of the solid particle. For most of the soil mechanics problems, as the bulk modulus KS of the solid particles is much larger that that of the whole material, 1 can be assumed. Equation (9.1) becomes ij = ij + ij p (9.2)

9.2.2

Before proceeding with the general equations, we introduce the associated notation. In this we have ij ui wi p the total Cauchy stress in the mixture at any instant the displacement of the solid skeleton the pseudo-displacement of the uid phase relative to the skeleton of solid the pore water pressure the strain increment of the solid phase the rotation increment of the solid phase

ij = 1 2 (dui,j + duj,i )

1 ij = 2 (dui,j duj,i )

, s , f

Jeremi et al.

the densities of the mixture, and the solid phase and water respectively

University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

195

the porosity the rate at which volume of water changes per unit total volume of mixture

= w i,i

With these denitions, and others introduced as necessary, we can write the governing equation of the coupled system.

9.2.2.1

The overall equilibrium or momentum balance equation for the soil-uid mixture can be written as (Zienkiewicz et al., 1999a) ij,j u i f [w i + w jw i,j ] + bi = 0 (9.3)

Where u i is the acceleration of the solid part, bi is the body force per unit mass, w i + w jw i,j is the uid acceleration relative to the solid part, w i is local acceleration, w jw i,j is convective acceleration. The underlined terms in the above equation represent the uid acceleration relative to the solid and convective

terms of this acceleration. Generally this acceleration is so small that we shall frequently omit it. And for static problems, equation(9.3) only consists of the rst and last terms. For fully saturated porous media (no air inside), from denition = = = = = Mt Vt Ms + Mf Vt Vs s + Vf f Vt Vf Vt Vf f + s Vt Vt nf + (1 n)s (9.4)

= nf + (1 n)s

where Mt , Ms and Mf are the mass of total, solid part and uid part respectively. Vt , Vs and Vf are the volume of total, solid part and uid part respectively.

9.2.2.2

For the pore uid, the equation of momentum balance can be written as (Zienkiewicz et al., 1999a) p,i Ri f u i f [w i + w jw i,j ]/n + f bi = 0 (9.5)

where R is the viscous drag forces. It should be noted that the underlined terms in (9.5) represent again the convective uid acceleration and are generally small. Also note that, throughout report, the permeability k is

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

196

used with dimensions of [length]3 [time]/[mass], which is dierent from the usual soil mechanics convention, the permeability has the same dimension of velocity, i.e. [length]/[time]. Their values are related by k = K/f g , where g is the gravitational acceleration at which the is measured. Assuming the Darcy seepage law: nw = Ki, here i is the head gradient. Seepage force is then R = f gi. R can be written as

1 Ri = kij w j

or

Ri = k 1 w i

(9.6)

kij or k are Darcy permeability coecients for anisotropic and isotropic conditions respectively.

9.2.2.3

The nal equation is supplied by the mass conservation of the uid ow (Zienkiewicz et al., 1999a) w i,i + ii + p f +n +s 0 = 0 Q f (9.7)

The rst term of equation(9.7) is the ow divergence of a unit volume of mixture. The second term is the volume change of the mixture. In the third term, Q is relative to the compressibility of the solid and uid. The underlined

terms represent change of density and rate of volume expansion of the solid in case of thermal changes. They are generally negligible. n n n 1n 1 + + = Q Kf Ks Kf Ks where Ks and Kf are the bulk moduli of the solid and uid phases respectively. Thus, we got the total mixture equilibrium equation (9.3), uid equilibrium equation(9.5) and the ow conservation equation (9.7) for saturated soil. By omitting the convective acceleration (the underline terms in (9.3) and (9.5)), density variation and the volume expansion due to the thermal change (the underline terms in (9.7)), the equations of the total coupled system can be further simplied, they are summarized as below ij,j u i f w i + bi = 0 f w i + f bi = 0 n (9.9) (9.8)

p,i Ri f u i

(9.10)

w i,i + ii +

p =0 Q

(9.11)

9.2.3

9.2.3.1

Solid part equilibrium equation

First of all, a new variable Ui is introduced in place of the relative pseudo-displacement wi , it is dened as Ui = ui + UiR = ui +

Jeremi et al.

wi n

University of California, Davis

(9.12)

Version: March 21, 2012, 17:15

Lecture Notes

197

Figure 9.1: Fluid mechanics of Darcys ow (wi ) versus real ow (Ui = wi /n). Insertion of denition (9.12) into equation (9.9)(9.10) and subtraction of [n (9.10)] from equation (9.9) leads to the equation of skeleton equilibrium ij,j u i + bi + np,i + nRi + nf u i nf bi = 0 By substituting = (1 n)s + nf ij,j (1 n)s u i nf u i + (1 n)s bi + nf bi + np,i + nRi + nf u i nf bi = 0 ij,j + np,i + (1 n)s bi (1 n)s u i + nRi = 0 By using the denition of eective stress (9.1),(9.13) becomes ij,j ( n)p,i + (1 n)s bi (1 n)s u i + nRi = 0 (9.15) (9.14) (9.13)

9.2.3.2

The uid part equilibrium equation can be obtained simply by [n (9.10],i.e. np,i nRi nf u i f w i + nf bi = 0 w i np,i nRi nf ( ui + ) + nf bi = 0 n From equation (9.12), we have w i i = u U i + n so that equation (9.16) becomes: i nRi = 0 np,i + nf bi nf U (9.18) (9.17)

(9.16)

9.2.3.3

Jeremi et al. University of California, Davis

(9.19)

Version: March 21, 2012, 17:15

Lecture Notes

198

Notice that ii = u i,i , so that equation (9.19) becomes i,i n w i,i = nU ii By substituting (9.20) to (9.11), we obtain i,i n nU ii + ii + or: 1 i,i = ( n) nU ii + p Q (9.22) p =0 Q (9.21) (9.20)

Thus we got the whole set of modied governing equations (9.15),(9.18) and (9.22). They are summarized as below ij,j ( n)p,i + (1 n)s bi (1 n)s u i + nRi = 0 (9.23)

(9.24)

(9.25)

From the modied equation set (9.23),(9.24) and (9.25), we can see that only u i occurs in the rst equation,and only Ui in the second, thus leading to a convenient diagonal form in discretization. Now we have a complete equation system given by (9.23),(9.24) and (9.25). With the basic denitions introduced earlier, there are three essential 1. solid displacement u 2. pore pressure p 3. absolute uid displacement U The boundary conditions imposed on these variables will complete the problem. These boundary conditions are: For the momentum balance part, on boundary t , traction ti (t)(or ij nj ), where ni is the i-th component of the normal to the boundary is specied. On boundary u , the displacement ui is given. For the uid part, again the boundary is divided into two parts. On p , the pressure p is specied, on w , the normal outow w n is specied. For impermeable boundary a zero value for the outow should be specied. The boundary conditions can be summarized below ti ui = t u on = u

University of California, Davis

= ij nj = ti = ui on

= w (9.26)

Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

199

and p nT w = = = p w p wn on on = p = w (9.27)

9.3

The solutions to the problems governed by the modied governing equation set (9.23),(9.24) and (9.25) can be found by solving partial dierential equations, which can be written as + B + L() = 0 A (9.28)

where A, B are matrices of constant, and L is an operator involving spatial dierentials. The dot notation represents the time dierentiation. is a vector of dependent variables, in this problem, we say it represents the displacement u or the pressure p. The nite element solution of a problem will always proceed as the following procedures: 1. Discretize or approximate the unknown functions by a nite set of parameters k and shape function Nk . They are specied in space dimensions. Thus, we can write

n

= h =

k=1

Nk k

(9.29)

into the dierential equations to obtain a residual, then 2. Insert the value of the approximating function we can write a set of weighted residual equations in the form

h + B h + L(h ))d = 0 WjT (A In nite element method, the weighting functions Wj are usually identical to the shape functions.

(9.30)

The solid displacement ui , the pore pressure p, and the absolute uid displacement Ui can be approximated using shape functions and nodal values. ui p Ui

u uKi = NK p = NK pK U = NK U Ki

(9.31)

p u U where NK , NK and NK are shape functions for solid displacement, pore pressure and uid displacement respec-

tively, uKi ,pK ,U Ki are nodal values of solid displacement, pore pressure and uid displacement respectively.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

200

9.3.1

u To obtain the numerical solution of the rst equation, we premultiply (9.23) by NK and integrate over the domain.

u NK ij,j d

=

t

u NK nj ij d u NK (ti t u (f 1 )Ki

u NK,j ij d u NK,j ij d

= = = = = =

+ ni p)d

u (f 1 )Ki [ u (f 1 )Ki

(9.32)

where K EP is the stiness matrix of the solid part,ni is the direction of the normal on the boundary. Second term of (9.23) becomes

u NK ( n)p,i d

=

p

u NK ( n)ni pd + u NK ( n)ni pd + [ p

(9.33)

(9.34)

u NK (1 n)s ij u j d

= = =

u u Lj NK (1 n)s ij NL d]u

where Ms is the mass matrix of solid part. By substituting equations (9.6) and (9.12),last term of (9.23) (Damping

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

201

Matrix) becomes

u NK Ri d

j[ u 2 1 U NK n kij NL d]U L

C u u u u + C2 U + G1 p + f5 f4 Ms u K EP u + f1 1 =0 or +C u K EP u G1 p C2 U 1 + Ms u = f s

(9.37)

(9.38)

where

u u u f s = f1 f4 + f5

(9.39)

in index form (C ) EP Lj + (Ms )KijL u Ki = (f s )Ki KKijL (G1 )KiL pL + (C2 )KijL U Lj 1 KijL u Where K EP = (K EP )KimP G1 = (G1 )KiM C2 = (C2 )KijL C1 = (C1 )KijL Ms = (Ms )KijL f = (f s )Ki =

u u NK,j Dijml NP,l d p u NK,i ( n)NM d u 2 1 U NK n kij NL d u 2 1 u NK n kij NL d u u NK (1 n)s ij NL d u u u (f 1 )Ki (f4 )Ki + (f5 )Ki

(9.40)

= = = =

(9.41)

9.3.2

1 Ri = n2 kij (Uj u j)

(9.42)

University of California, Davis Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

202

U By premultiplying (9.43) by NK and integrating over the domain

(9.43)

U nNK p,i d

=

p

U nNK ni pd +

U nNK,i pd

= (f1 )Ki + [

p U nNK,i NM d]pM

= (f1 )Ki + (G2 )KiM pM = (f1 )Ki + (G2 )p Second term of (9.43) is then

U NK f bi d = (f2 )Ki

(9.44)

(9.45)

U j d NK nf ij U

= [

U U NK nf ij NL d]U Lj

U 2 1 NK n kij Uj d + U 2 1 NK n kij u j d

(9.46)

U 2 1 U NK n kij NL d]U Lj

(9.47)

+[

U 2 1 u NK n kij NL d]uLj

+ CT u C3 U 2

(9.48)

C U + CT u =0 f1 + G2 p + f2 Mf U 3 2 or +M U =f T G 2 p C2 u + C3 U f f where f f = f2 f1

Jeremi et al. University of California, Davis

(9.49)

(9.50)

(9.51)

Version: March 21, 2012, 17:15

Lecture Notes

203

in index form (G2 )KiM pM (C2 )T LjiK uLj + (C3 )KijL U Lj + (Mf )KijL ULj = (ff )Ki where (ff )Ki G2

T C2

(9.52)

= = = = =

T (C2 )KijL = p U nNK,i NM d U 2 1 u NK n kij NL d U 2 1 U NK n kij NL d U U NK nf ij NL d

C3 Mf

(9.53)

9.3.3

By integrating (9.25) in time and noticing ii = ui,i , we can obtain nUi,i = ( n)ii + 1 p Q (9.54)

p By multiplying (9.54) by NM and integrating over the domain, rst term of (9.54) becomes p U NM nNL,j d]U Lj = (G2 )M Lj U Li = GT 2U

(9.55)

p NM ( n)ui,i d

= = =

p u NM ( n)NL,j d]uLj

p NN

1 p N d]pN Q M

= PN M pM = P p

(9.57)

T GT 2 U + G1 u + P p = 0

(9.58)

Jeremi et al. University of California, Davis

(9.59)

Version: March 21, 2012, 17:15

Lecture Notes

204

9.3.4

The numerical forms of governing equations (9.38),(9.50) and (9.58) can be written together in the matrix form as Ms 0 u C1 0 0 p + 0 T 0 Mf U C2 0 0 0 0 C2 0 C3 u K EP p + GT 1 0 U G1 P G2 0 u

0 0

GT 2 p 0 U fs = 0 ff (9.60)

or in index form (Ms )KijL 0 0 0 0 0 0 0 Lj u (C1 )KijL p N + 0 U Lj (C2 )LjiK (G1 )KiM PM N (G2 )KiL 0 0 (C2 )KijL 0 Lj u p N U (9.61) where Ms Mf C1 C2 C3 K EP G1 G2 P = = = = = = = = = (Ms )KijL =

u u NK (1 n)s ij NL d U U NK nf ij NL d u 2 1 u NK n kij NL d u 2 1 U NK n kij NL d U 2 1 U NK n kij NL d u u NK,m Dimjn NL,n d p u NK,i ( n)NM d p U nNK,i NM d

(Mf )KijL = (C1 )KijL = (C2 )KijL = (C3 )KijL = (K EP )KijL = (G1 )KiM =

(G2 )KiM =

PN M =

p NN

1 p N d Q M (9.62)

Jeremi et al.

Lecture Notes

205

f Ki f Ki

solid

= = =

u u u (f 1 )Ki (f4 )Ki + (f5 )Ki U U (f1 )Ki + (f2 )Ki u NK nj ij d t u NK ( n)ni pd p u NK (1 n)s bi d U nNK ni pd p U nNK f bi d

f luid

u (f1 )Ki u (f4 )Ki u (f5 )Ki U (f1 )Ki U (f2 )Ki

= = = =

(9.63)

Here we have N u ,N p ,N U as the shape functions of skeleton, pore pressure and uid, ,s ,f are the density of the total, the solid and the uid phases, respectively, n is the porosity, and by its denition = (1 n)s + nf , the symbol ni is the direction of the normal on the boundary, u is the displacement of the solid part, p is pore pressure and U is the absolute displacement of the uid part. Equation (9.60) represents the general form (u p U ) for coupled system which can be written in a familiar form as Mx + Cx + Kx = f (9.64)

where x represents the generalized unknown variable. Solution of this equation for each time step will render unknown eld for given initial and boundary conditions.

9.3.5

Isoparametric elements are used in previous sections, where the coordinates are interpolated using the same shape functions as for the unknown. This mapping allows the use of elements of more arbitrary shape than simple forms such as rectangles and triangles. But in static or dynamic undrained analysis the permeability (and compressibility) matrices are zero, i.e.(Q ,and P 0),resulting in a zero-matrix diagonal term in the equation(9.61). The matrix to be solved is the same as that in the solutions of problems of incompressible elasticity or uid mechanics. Actually a wide choice of shape functions is available if the limiting(undrained) condition is never imposed. Due to the presence of rst derivatives in space in all the equations it is necessary to use "Co -continuous" interpolation functions and the suitable element forms are shown in Fig.9.2.

9.4

9.4.1

u-p Formulation

Governing Equations of Porous Media

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

206

Figure 9.2: Shape functions used for coupled analysis, displacement u and pore pressure p formulation The rst governing equation of porous media is total momentum balance equation:

ij,j u i + bi = 0 where ij = ij pij and = (1 n)s + nf . The second governing equation is the uid mass balance equation: (kij (p,j + f bj )),i + u i,i + where kij = kij kij = gf f p =0 Qsf

(9.65)

(9.66)

(9.67)

and kij is the permeability in Darcys law with the same unit as velocity. Qsf = Ks Kf Ks + Kf (9.68)

is the total compression modulus, Ks and Kf are the solid and uid compression modulus respectively. The boundary conditions are i ij nj = t ui = u i ni wi = ni kij (p,j + f bj ) = w = q p=p where w is the outow and q is the inux.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

on on on on

= t = u = w = p

Lecture Notes

207

9.4.2

The solid displacement ui and the pore pressure p can be approximated using shape functions and nodal values: ui p

u = NK u Ki p = NL p L

(9.73) (9.74)

9.4.2.1

u NK (ij,j u i + bi ) d = 0

(9.75)

u NK ij,j d

=

t

u NK,j ij d

= = = = =

u NK,j ij d +

u NK,j Dijml ml d + [

+[

u (f1 )Ki

ep (KKimP ) uP m

p u NK,i NN d] pN

+ (QKiN ) pN (9.76)

u = f1 (K ep ) u + Qp

u NK u i d

= = = =

[ [ij

(MKijL )u Lj M u (9.77)

u NK bi d u = f2

u (f2 )Ki

(9.78)

University of California, Davis Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

208

ep u u (MKijL )u Lj (QKiN ) pN + (KKijL ) uLj = (f1 )Ki + (f2 )Ki = (f u )Ki

(9.79)

or

u u + f2 = fu Mu Qp + (K ep ) u = f1

(9.80)

9.4.2.2

p NM

p Qsf

d = 0

(9.81)

p NM (kij (p,j + f bj )),i d p NM wi ni d p NM w d p NM,i kij (p,j + f bj )d p NM,i kij f bj d

(9.82)

=

w

= = =

p (f1 )M p (f1 )M

+[

p p = f1 + Hp f2

(9.83)

p NM u i,i d

= =

p u Lj NM NL,j d]u

Lj (QLjM )u (9.84)

= QT u

p NM

p d Qsf

= =

1 p N NN d]p Q sf N (SM N )p

p NM

= Sp

(9.85)

p p Lj + (SM N )p N = (f1 (HM N ) pN + (QLjM )u )M + (f2 )M = (f p )M

(9.86)

Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

209

or

p p + Sp = f1 = fp + f2 Hp + QT u

(9.87)

9.4.2.3

(K ep )KiLj 0

(9.88)

where fu fp and fu 1 fu 2 fp 1 fp 2 M

u (f1 )Ki = t u (f2 )Ki = p (f1 )M = w p )M = (f2 p NM w d u NK bi d u NK ti d u u u fKi = (f1 )Ki + (f2 )Ki p p p fM = (f1 )M + (f2 )M

(9.90) (9.91)

MKiLj = ij

Q QKiN =

p u NK,i NN d

S H

SM N =

p NM

1 N p d Qsf N

HM N =

Jeremi et al.

Lecture Notes

210

Jeremi et al.

Chapter 10

(1998-2000-2005-)

211

Lecture Notes

212

10.1 10.2

10.2.1

Introduction

Domain Decomposition approach is the most popular and eective method to implement parallel nite element method. The underlying idea is to physically divide the problem domain into subdomains and nite element calculations will be performed on each individual domain in parallel. Domain Decomposition can be overlapping or non-overlapping. The overlapping domain decomposition method divides the problem domain into several slightly overlapping subdomains. Non-overlapping domain decomposition is extensively used in continuum nite element modeling due to the relative ease to program and organize computations and is the one that will be examined in this chapter. In general, a good non-overlapping decomposition algorithm should be able to handle irregular mesh of arbitrarily shaped domain. minimize the interface problem size by delivering minimum boundary connectivity, which will help reducing the communication overheads. The well-known idea of domain decomposition method can be found in a 1870 paper by the father of domain decomposition, H.A. Schwarz (Rixena and Magouls, 2007). Domain decomposition method is also the underlying paradigm of substructuring methods developed in the sixties, which aim at reducing the dimension of models in structural analysis by applying static condensation-type techniques to subdomains. Other than static condensation, Farhat and Roux (1991); Farhat (1991); Farhat and Geradin (1992) proposed FETI (Finite Element Tearing and Interconnecting) method for domain decomposition analysis. In FETI method, Lagrange multipliers are introduced to enforce compatibility at the interface nodes. Rigid body modes are eliminated in parallel from each local problem and a direct scheme is applied concurrently to all subdomains in order to recover each partial local solution. The contributions of these modes are then related to the Lagrange multipliers through an orthogonality condition. This FETI method has been shown that it can deliver high eciency for parallel implicit transient simulations in structural mechanics (Crivelli and Farhat, 1993). Domain decomposition itself has become a active topic as parallel processing techniques receive much more attention in mathematics and engineering world during recent years. Domain decomposition was revived as a natural paradigm for parallel solvers (Rixena and Magouls, 2007). Many papers have discussed two algorithms that are currently receiving much research eort, namely the FETI-DP (or Dual Primal Finite Element Tearing and Interconnecting) method and the even more recent BDDC (or Balancing Domain Decomposition by Constraints). FETI-DP is the third generation FETI method (Bavestrello et al., 2007) developed for the fast, scalable, and domain-decomposition-based iterative solution of symmetric systems of equations arising from the nite element (FE) discretization of static, dynamic, structural and acoustic problems (Farhat et al., 2001, 2000).

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

213

BDDC, on the other hand, derives its formulation from substructuring method by enforcing constraints associated with disjoint sets of nodes on substructure boundaries using constrained energy minimization concepts (Dohrmann, 2003; Mandel and Dohrmann, 2003). An early endeavor on dynamic computational load balancing was presented by McKenna (1997). Limited number of examples show that run time, dynamic computational load balancing can indeed improve parallel program performance in some cases, particularly when nonlinearities are involved. Although many works have been presented on domain decomposition methods, the most popular methods such as FETI-type and BDDC all stem from the root of subdomain interface constraints handling. The merging of iterative solving with domain decomposition-type preconditioning is promising as shown by many researchers (Pavarino, 2007; Li and Widlund, 2007). Schwartz-type preconditioners for parallel domain decomposition system solving have also shared part of the spotlight (Hwang and Cai, 2007; Sarkis and Szyld, 2007). In solid nite element methods, it has been assumed that the equation solving is the most computational expensive part so it is totally reasonable that all focus has been set on equation solver during the past decades. Work presented in this chapter, however, has originated from the observation that for highly nonlinear materials,

the constitutive level computation can be at least equally costly as equation solving, if not more expensive. The novelty of this chapter is to break out of the existing substructuring or FETI frameworks to further address the fundamental load balance issue of parallel computing. Namely, in order to achieve better parallel performance, we want to keep all processors equally busy. Load imbalance issue resulted from nonlinear constitutive level computations is too important to be neglected. This chapter proposes the Plastic Domain Decomposition algorithm which focuses on adaptive load balancing operation for nonlinear nite elements. From the implementation point of view, for mesh-based scientic computations, domain decomposition corresponds to the problem of mapping a mesh onto a set of processors, which is well dened as a graph partitioning problem (Schloegel et al., 1999). Formally, the graph partitioning problem is as follows. Given a weighted, undirected graph G = (V ; E ) for which each vertex and edge has an associated weight, the k -way graph partitioning problem is to split the vertices of V into k disjoint subsets (or subdomains) such that each subdomain has roughly an equal amount of vertex weight (referred to as the balance constraint), while minimizing the sum of the weights of the edges whose incident vertices belong to dierent subdomains (i. e., the edge-cut). In computational solid mechanics, the element graph is naturally used in parallel nite element method due to the fact that elemental operation forms the basis of nite element method. On the other hand, for material nonlinearity simulations, the element calculations represent the most computationally expensive part. In order to facilitate consistent interfaces for computational load measuring and data migration, element graph has been utilized as fundamental graph structure in this chapter, although it has been shown that the node-graph can be used as well for structure dynamics problem and the element-cut partitioning can make certain algorithms simpler (Krysl and Bittnar, 2001). The graph partitioning problem is known to be NP-complete 1 . Therefore, generally it is not possible to compute

1 The

complexity class NP is the set of decision problems that can be solved by a non-deterministic Turing machine in polynomial

Jeremi et al.

Lecture Notes

214

optimal partitioning for graphs of interesting size in a reasonable amount of time. Various heuristic approaches have been developed, which can be classied as either geometric, combinatorial, spectral, combinatorial optimization techniques, or multilevel methods (Dongarra et al., 2003). In nite element simulations involving nonlinear material response, static graph partitioning mentioned above does not guarantee even load distribution among processors. Plastication introduces work load that is much heavier than pure elastic computation. So for this kind of multiphase simulation, adaptive computational load balancing scheme has to be considered to keep all processing units equally busy as much as possible. Traditional static graph partitioning algorithm is not adequate to do multiphase partition/repartitioning. A parallel multilevel graph partitioner has been introduced in this research to achieve dynamic load balancing for inelastic nite element simulations. In this chapter, the algorithm of Plastic Domain Decomposition (PDD) is proposed. The adaptive multi-level graph partitioning kernel of the PDD algorithm is implemented through the ParMETIS interface. Studies are performed to extract optimal algorithmic parameters for our specic applications.

10.2.2

The distinct feature of inelastic (elastic-plastic) nite element computations is the presence of two iteration levels. In a standard displacement based nite element implementation, constitutive driver at each Gauss point iterates in stress and internal variable space, computes the updated stress state, constitutive stiness tensor and delivers them to the nite element functions. Finite element functions then use the updated stresses and stiness tensors to integrate new (internal) nodal forces and element stiness matrix. Then, on global level, nonlinear equations are iterated on until equilibrium between internal and external forces is satised within some tolerance. Elastic Computations In the case of elastic computations constitutive driver has a simple task of computing increment in stresses (ij ) for a given deformation increment (

kl ),

kl )

It is

important to note that in this case the amount of work per Gauss point is known in advance. The amount of computational work is the same for every integration point. If we assume the same number of integration points per element, it follows that the amount of computational work is the same for each element and it is known in advance. Elastic-Plastic Computations On the other hand, for elastic-plastic problems, for a given incremental deformation the constitutive driver is iterating in stress and internal variable space until consistency condition is satised (F = 0). The number of iterations is not known in advance. Initially, all Gauss points are in elastic range, but as we incrementally apply loads, the plastic zones develop. For Gauss points still in elastic range, there are no iterations, the constitutive driver just computes incremental stresses from closed form solution. Computational load will

time. the NP-complete problems are the most dicult problems in NP ("non-deterministic polynomial time") in the sense that they are the smallest subclass of NP that could conceivably remain outside of P, the class of deterministic polynomial-time problems. (http://en.wikipedia.org/wiki/NP-complete)

Jeremi et al.

Lecture Notes

215

increase signicantly for integration of constitutive equations in plastic range. In particular, constitutive level integration algorithms for soils, concrete, rocks, foams and other granular materials are very computationally demanding. More than 70% of wall clock time during an elastic-plastic nite element analysis is spent in constitutive level iterations. This is in sharp contrast with elastic computations where the dominant part is solving the system of equations which consumes about 80% of run time. The extent of additional, constitutive level iterations is not known before the actual computations are over. In other words, the extent of elastic-plastic domain is not known ahead of time. The traditional preprocessing type of Domain Decomposition method (also known as topological DD) splits domain based on the initial geometry and assigns roughly the same number of elements to every computational node and minimizes the size of subdomain boundaries. This approach might result in serious computational load imbalance for elastic-plastic problems. For example one domain might be assigned all of the elastic-plastic elements and spend large amount of time in constitutive level iterations. The other domains will have elements in elastic state and thus spend far less computational time in computing stress increments. This results in program having to wait for the slowest domain (the one with large number of elastic-plastic nite elements) to complete constitutive level iterations and only proceed with global system iterations after that.

This illustrates a two-fold challenge with computational load balancing for inelastic simulations in mechanics. These two challenges is described below in some more detail.

10.2.2.1

Adaptive Computation

First, these computations are dynamic in nature. That is, the structure of elastic and elastic-plastic domains changes dynamically and unpredictably during the course of the computation. For this reason, a static decomposition computed as a pre-processing step is not sucient to ensure the computational load-balance of the entire computation. Instead, periodic computational load-balancing is required during the course of the computation. The problem of computing a dynamic decomposition shares the same requirements as that of computing the initial decomposition (i.e., balance the mesh elements and minimize the inter-processor communications), while also requiring that the cost associated with redistributing the data in order to balance the computational load is minimized. This last requirement prevents us from simply computing a whole new static partitioning from scratch each time computational load-balancing is required. Often, the objective of minimizing the data redistribution cost is at odds with the objective of minimizing the inter-processor communications. For applications in which the computational requirements of dierent regions of the domain change rapidly, or the amount of state associated with each element is relatively high, minimizing the data redistribution cost is preferred over minimizing the communications incurred during parallel processing. For applications in which computational load-balancing occurs very infrequently, the key objective of a loadbalancing algorithm is in obtaining the minimal inter-processor communications. For many application domains, it is straightforward to select a primary objective to minimize (i.e., minimize whichever cost dominates). However, one of the key issues concerning the elastic-plastic computation is that the number of iterations between compuJeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

216

tational load-balancing phases is both unpredictable and dynamic. For example, in the case of static problems, zones in the 3D solid may become plastic and then unload to elastic (during increments of loading) so that the extent of plastic zone is changing. The change can be both slow and rapid. Slow change usually occurs during initial loading phases, while the later deformation tends to localize in narrow zones rapidly and the rest of the solid unloads rapidly (becomes elastic again). The narrow, localized zone has heavy computational load on the constitutive level (in each integration point within elements). Similar phenomena is observed in seismic soil-structure interaction computations where sti structure interacts with soft soil and elastic and elasto-plastic zones change signicantly during loading cycles. In this type of computation, it is extremely dicult to select the type of computational load-balancing algorithm to employ. Furthermore, the preferred computational loadbalancing algorithm is liable to change during the course of the computation, and so the selection must be made dynamically.

10.2.2.2

Multiphase Computation

The second challenge associated with computational load-balancing elastic-plastic computations in geomechanics is that these are two-phase computations. That is, elastic-plastic computations follow up the elastic computations.

There is a synchronization phase between the computations, as only after the elastic computation is nished is it possible to check if the elastic-plastic computation is required for a given integration (Gauss) point within an element. For regions of the mesh in which this check indicates that the elastic-plastic computation is necessary, lengthy elastic-plastic computations are then performed. The existence of the synchronization step between the two phases of the computation requires that each phase be individually load balanced. That is, it is not sucient to simply sum up the relative times required for each phase and to compute a decomposition based on this sum. Doing so may lead to some processors having too much work during the elastic computation (and so, these may still be working after other processors are idle), and not enough work during the elastic-plastic computation, (and so these may be idle while other processors are still working), and vice versa. Instead, it is critical that every processor have an equal amount of work from both of the phases of the computation.

10.2.2.3

Elastic-plastic FE computation can be understood as a two-phase calculation, which is also dynamic in nature. Traditional graph partitioning formulations are not adequate to ensure its ecient execution on high performance parallel computers. In this chapter very recent progresses from the graph partitioning algorithm research will be investigated. We need new adaptive graph partitioning formulations, which can compute adaptive partitioningrepartitionings that can satisfy an arbitrary number of balance constraints. Static Graph Partitioning Given a weighted, undirected graph G = (V, E ), for which each vertex and edge has an associated weight, the k -way graph partitioning problem is to split the vertices of V into k disjoint subsets (or subdomains) such that each subdomain has roughly an equal amount of vertex weight (referred to as the balance constraint), while minimizing the sum of the weights of the edges whose incident vertices belong to dierent subdomains

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

217

(i.e., the edge-cut). 1. Geometric Techniques Compute partitioning based solely on the coordinate information of the mesh nodes, without considering edge-cut. Popular methods include, Coordinate Nested Dissection (CND or Recursive Coordinate Bisection), Recursive Inertial Bisection (RIB), Space-Filling Curve techniques and Sphere-Cutting approach. 2. Combinatorial Techniques Attempt to group together highly connected vertices whether or not these are near each other in space. That is combinatorial partitioning schemes compute a partitioning based only on the adjacency information of the graph; they do not consider the coordinates of the vertices. They tend to have lower edge-cuts but generally slower. Popular methods include, Levelized Nested Dissection (LND) and Kernighan-Lin/Fiduccia-Mattheyses (KL/FM) partitioning renement algorithm, which needs an initial partition input to do swapping renement. 3. Multilevel Schemes The multilevel paradigm consists of three phases: graph coarsening, initial partitioning, and multilevel

renement. Firstly, we form coarse graph by collapsing together selected vertices of the input graph. After rounds of coarsening, we get coarsest graph, on which an initial bisection will be performed. Then the KL/FM algorithm can be used to rene the partition back to the nest graph. The multilevel paradigm works well for two reasons. First, a good coarsening scheme can hide a large number of edges on the coarsest graph, which makes the task of computing high-quality partitioning easier. Second reason, incremental renement schemes such as KL/FM become much more powerful in the multilevel context. Popular algorithms include Multilevel Recursive Bisection and Multilevel k -Way Partitioning. Adaptive Graph Partitioning For large scale elasto-plastic FE simulations, it is necessary to dynamically load-balance the computations as the analysis progresses due to unpredictable plastication inside the domain. This dynamic load balancing can be achieved by using a graph partitioning algorithm. Adaptive graph partitioning shares most of the requirements and characteristics of static graph partitioning but also adds an additional objective. That is, the amount of data that needs to be redistributed among the processors in order to balance the load should be minimized. If the vertex weight represents the computational cost of the work carried by the vertex, another metric, size of the vertex needs to be considered as well, which reects distribution cost of the vertex. Thus, the repartitioner should attempt to balance the partitioning with respect to vertex weight while minimizing vertex migration with respect to vertex size. Dierent approaches are available. One can simply compute a new graph from scratch, so called ScratchRemap Repartitioner, which expectedly introduces more data redistribution than necessary. Diusion-Based Repartitioner attempt to minimize the dierence between the original partitioning and the nal repartitioning

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

218

by making incremental changes in the partitioning to restore balance. This method has been an very active topic during recent years, Dongarra et al. (2003) gives up-to-date review. Multiconstraint Graph Partitioning We can see traditional graph partitioning typically balances only a single constraint (i.e., the vertex weight) and minimizes only a single objective (i.e., the edge-cut). If we replace the vertex weight, which is a single number, with a weight vector of size m, then the problem becomes that of nding a partitioning that minimizes the edge-cuts subject to the constraints that each of the m weights is balanced across subdomains. Multilevel graph partitioning algorithms for solving multiconstraint/multiobjective problems have been very successful Schloegel et al. (1999). The software libraries METIS and ParMETIS are widely used in computational mechanics research.

10.2.2.4

In this chapter, the Plastic Domain Decomposition (PDD) has been developed using multi-level, multi-objective

graph partitioning algorithm. This algorithm automatically monitors load balancing condition and updates element graph structure accordingly as the simulation progresses. Element redistribution will be triggered to achieve load balance when nonlinearity of materials brings down the parallel performance.

10.2.3

Kaypis and Kumar (1998) present a k -way multilevel partitioning algorithm whose run time is linear in the number of edges |E | (i.e., O(|E |)); whereas the run time of multilevel recursive bisection schemes is O(|E |logk ) for k way partitioning. Kaypis and Kumar (1998) show that the proposed multilevel partitioning scheme produces partitioning that are of comparable or better quality than those produced by multilevel recursive bisection, while requiring substantially less time. This paradigm consists of three phases: graph coarsening, initial partitioning, and multilevel renement. In the graph coarsening phase, a series of graphs is constructed by collapsing together selected vertices of the input graph in order to form a related coarser graph. This newly constructed graph then acts as the input graph for another round of graph coarsening, and so on, until a suciently small graph is obtained. Computation of the initial bisection is performed on the coarsest (and hence smallest) of these graphs, and so is very fast. Finally, partition renement is performed on each level graph, from the coarsest to the nest (i.e., original graph) using a KL/FM-type algorithm Dongarra et al. (2003). Figure 10.1 illustrates the multilevel paradigm. This algorithm is available in METIS Karypis and Kumar (1998b) which is used in this research to provide initial static partitioning. Adaptive graph repartitioning algorithm can be used to achieve dynamic load balancing of multiphase elastic-plastic nite element simulations. Adaptive graph partitioning diers from static graph partitioning algorithm in the sense that one additional objective has to be targeted. That is, the amount of data the needs to be redistributed among the processors in order to balance the load should be minimized. In order to measure this redistribution cost,

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

219

not only does the weight of a vertex, but also its size have to be considered. In our implementation for the purpose of this research, the vertex weight represents the computational load of each nite element, while the size reects its redistribution cost. Thus, the application of adaptive graph partitioning algorithm aims at balancing the partitioning with respect to vertex weight while minimizing vertex migration with respect to vertex size. A repartitioning of a graph can be obtained simply by partitioning a new graph from a scratch, which tends to bring much more unnecessary communications because the old distribution has not been taken into account. Diusion-based Repartitioner is more popular in which one attempts to minimize the dierence between the original partitioning and the nal repartitioning by making incremental changes in the partitioning to restore balance. Dongarra et al. (2003) gives a comprehensive review on this subject. Adaptive repartitioning is available in ParMETIS Karypis et al. (2003) and Jostle Warshaw (1998). The former is chosen in this research considering the fact that ParMETIS provides seamless interface for METIS 4.0 which makes the comparison between static and adaptive partitioning schemes more consistent. PARMETIS is an MPI-based parallel library that implements a variety of algorithms for partitioning and repartitioning unstructured graphs and for computing ll-reducing orderings of sparse matrices Karypis et al. (2003). PARMETIS is particularly suited for parallel numerical simulations involving large unstructured meshes. In this type of computation, PARMETIS dramatically reduces the time spent in communication by computing mesh decompositions such that the numbers of interface elements are minimized. The algorithms in PARMETIS are based on the multilevel partitioning and ll-reducing ordering algorithms that are implemented in the widely-used

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

220

serial package METIS Karypis and Kumar (1998a). However, PARMETIS extends the functionality provided by METIS and includes routines that are especially suited for parallel computations and large-scale numerical simulations. In particular, PARMETIS provides the following functionality Karypis et al. (2003): Partition unstructured graphs and meshes. Repartition graphs that correspond to adaptively rened meshes. Partition graphs for multi-phase and multi-physics simulations. Improve the quality of existing partitioning. Compute ll-reducing orderings for sparse direct factorization. Construct the dual graphs of meshes. Both METIS and PARMETIS are used in this research. METIS routines are called to construct static partitioning for commonly used one-step static domain decomposition, while adaptive load-balancing is achieved by calling PARMETIS routines regularly during the progress of nonlinear nite element simulations.

Adaptive load-balancing through domain repartitioning is a multi-objective optimization problem, in which repartitionings should minimize both the inter-processor communications incurred in the iterative mesh-based computation and the data redistribution costs required to balance the load. PARMETIS provides the routine ParMETIS_V3_AdaptiveRepart for repartitioning the previous unbalanced computational domain. This routine assumes that the existing decomposition is well distributed among the processors, but that (due to plastication of certain nonlinear elements) this distribution is poorly load balanced. Figure 10.2 Schloegel et al. (2000) shows common steps involved in the execution of adaptive mesh-based simulations on parallel computers. Initially, the mesh is equally distributed on dierent processors. As all elements are elastic at the very beginning (carrying the same amount of elemental calculation work), computation load balance can be guaranteed with a even distribution. A number of iterations of the simulation are performed in parallel, after which plasticity occurs in certain nonlinear elements thus introducing some amount of load imbalance. A new partitioning based on the unbalanced domain is computed to re-balance the load, and then the mesh is redistributed among the processors, respectively. The simulation can then continue for another number of iterations until either more mesh adaptation is required or the simulation terminates. If we consider each round of executing a number of iterations of the simulation, mesh adaptation, and loadbalancing to be an epoch, then the run time of an epoch can be described by, Schloegel et al. (2000) (tcomp + f (|Ecut |))n + trepart + g (|Vmove |) (10.1)

where n is the number of iterations executed, tcomp is the time to perform the computation for a single iteration of the simulation, f (|Ecut |) is the time to perform the communications required for a single iteration of the simulation, and trepart and g (|Vmove |) represent the times required to compute the new partitioning and to redistribute the data. Here, the inter-processor communication time is described as a function of the edge-cut

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

221

Figure 10.2: A diagram illustrating the execution of adaptive scientic simulations on high performance parallel computers Schloegel et al. (1999)

of the partitioning and the data redistribution time is described as a function of the total amount of data that is required to be moved in order to realize the new partitioning. Adaptive repartitioning aects all of terms in Equation 10.1. How well the new partitioning is balanced inuences tcomp . The inter-processor communications time is dependent on the edge-cut of the new partitioning. The data redistribution time is dependent on the total amount of data that is required to be moved in order to realize the new partitioning. It is critical for adaptive partitioning schemes to minimize both the edge-cut and the data redistribution when computing the new partitioning. Viewed in this way, adaptive graph partitioning is a multi-objective optimization problem. There are various approaches how to handle this dual-objective problem. In general, two approaches have primarily been taken when designing adaptive partitioners. Schloegel et al. (2000) gives a comprehensive review on this topic. The rst approach is to attempt to focus on minimizing the edge-cut and to minimize the data redistribution only as a secondary objective. This family of methods can be called scratch-remap repartitioner. These use some type of state-of-the-art graph partitioner to compute a new partitioning from scratch and then attempt to intelligently remap the subdomain labels to those of the original partitioning in order to minimize the data redistribution costs. Since a state-of-the-art graph partitioner is used to compute the partitioning, the resulting edge-cut tends to be extremely good. However, since there is no guarantee as to how similar the new partitioning will be to the original partitioning, data redistribution costs can be high, even after remapping. The second approach is to focus on minimizing the data redistribution cost and to minimize the edge-cut as a secondary objective, or so-called diusion-based repartitioner. These schemes attempt to perturb the original partitioning just enough so as to balance it. This strategy usually leads to low data redistribution costs, especially when the partitioning is only

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

222

slightly imbalanced. However, it can result in higher edge-cuts than scratch-remap methods because perturbing a partitioning in order to balance it also tends to adversely aect its quality. These two types of repartitioner allow the user to compute partitioning that focus on minimizing either the edgecut or the data redistribution costs, but give the user only a limited ability to control the tradeos among these objectives. This control of the tradeos is sucient if the number of iterations that a simulation performs between load-balancing phases (i.e. the value of n in Equation 10.1 ) is either very high or very low. However, when n is neither very high nor very low, neither type of scheme precisely minimizes the combined costs of f (Ecut |)n and g (|Vmove |). Another disadvantage exists for applications in which n is dicult to predict or those in which n can change dynamically throughout the course of the computation. As an example, one of the key issues concerning the elastic-plastic soil-structure interaction computations required for earthquake simulation is that the number of iterations between load-balancing phases is both unpredictable and dynamic. Here, zones in the 3D solid may become plastic and then unload (during increments of loading) so that the extent of the plastic zone is changing. The change can be both slow and rapid. Slow change usually occurs during initial loading phases, while the later deformation tends to localize in narrow zones rapidly and the rest of the solid unloads rapidly (becomes elastic again) Jeremi and Xenophontos (1999).

Schloegel et al. (2000) presents a parallel adaptive repartitioning scheme (called the Unied Repartitioning Algorithm) for the dynamic load-balancing of scientic simulations that attempts to solve the precise multi-objective optimization problem. By directly minimizing the combined costs of f (Ecut |)n and g (|Vmove |), the proposed scheme is able to gracefully tradeo one objective for the other as required by the specic application. The paper shows that when inter-processor communication costs are much greater in scale than data redistribution costs, the proposed scheme obtains results that are similar to those obtained by an optimized scratch-remap repartitioner and better than those obtained by an optimized diusion-based repartitioner. When these two costs are of similar scale, the scheme obtains results that are similar to the diusive repartitioner and better than the scratch-remap repartitioner. When the cost to perform data redistribution is much greater than the cost to perform inter-processor communication, the scheme obtains better results than the diusive scheme and much better results than the scratch-remap scheme. They also show in the paper that the Unied Repartitioning Algorithm is fast and scalable to very large problems.

10.2.3.1

A key parameter used in Unied Repartitioning Algorithm (URA) is the Relative Cost Factor (RCF). This parameter describes the relative times required for performing the inter-processor communications incurred during parallel processing and to perform the data redistribution associated with balancing the load. Using this parameter, it is possible to unify the two minimization objectives of the adaptive graph partitioning problem into the unied cost function |Ecut | + |Vmove | (10.2)

where is the Relative Cost Factor, |Ecut | is the edge-cut of the partitioning, and |Vmove | is the total amount of data redistribution. The Unied Repartitioning Algorithm attempts to compute a repartitioning while directly

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

223

minimizing this cost function. The Unied Repartitioning Algorithm is based upon the multilevel paradigm that is illustrated in Figure 10.1, which can be described as three phases: graph coarsening, initial partitioning, and uncoarsening/renement Schloegel et al. (2000). In the graph coarsening phase, coarsening is performed using a purely local variant of heavy-edge matching. That is, vertices may be matched together only if they are in the same subdomain on the original partitioning. This matching scheme has been shown to be very eective at helping to minimize both the edge-cut and data redistribution costs and is also inherently more scalable than global matching schemes.

10.2.3.2

The RCF in the URA implementation controls the tradeo between two objectives, minimizing data redistribution cost or edge-cut. In our application, ParMETIS library has been linked to a Modied OpenSees Services (MOSS) analysis model to facilitate the partitioning/adaptive repartitioning scheme. The RCF is dened as a single parameter ITR in ParMETIS Karypis et al. (2003). This parameter describes the ratio between the time required for performing the inter-processor communications incurred during parallel processing compared to the time to perform the data redistribution associated with balancing the load. As such, it allows us to compute a single

metric that describes the quality of the repartitioning, even though adaptive repartitioning is a multi-objective optimization problem. As recommended by Karypis et al. (2003), appropriate values to pass for the ITR Factor parameter can be determined depending on the times required to perform 1. all inter-processor communications that have occurred since the last repartitioning, and 2. the data redistribution associated with the last repartitioning/load balancing phase. Simply divide the rst time measurement by the second time measurement. The result is the correct ITR Factor. In case these times cannot be ascertained (e.g., for the rst repartitioning/load balancing phase), Karypis et al. (2003) suggests that values between 100 and 1000 work well for a variety of situations. By default ITR is between 0.001 and 1000000. If ITR is set high, a repartitioning with a low edge-cut will be computed. If it is set low, a repartitioning that requires little data redistribution will be computed.

10.3

10.3.1

Introduction

In this chapter, parallel performance of the proposed PDD algorithm is thoroughly investigated. There are two major focuses for the timing analysis. Firstly we want to see how much performance gain we can have by introducing the PDD algorithm into inelastic nite element calculations. Secondly, we also want to show how scalable the proposed PDD algorithm is. As our nal objective is to apply PDD in large scale SFSI nite element simulations, nite element models of SFSI have been set up to study the parallel performance of the PDD based parallel program. Implicit constitutive

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

224

integration scheme Jeremi and Sture (1997) has been used to expose the load imbalance by plasticity calculation. Only continuum element has been studied due to the fact they can be easily visualized to obtain partition and/or repartition gures. Distributed memory Linux/Unix clusters are major platforms used in this chapter for speed up analysis.

10.3.2

Parallel Computers

Performance measurement has been carried out on two SMP-based clusters. IBM eServer p655 The DataStar IBM eServer p655 cluster consists of 176 8-way P655+ nodes at San Diego Supercomputer Center. System conguration is shown in Fig 10.3. The network benchmark is shown in Table 10.1. Table 10.1: Latency and Bandwidth Comparison (as of August 2004) MPI Latencies (sec) Intra-node 3.9 7.65 Bandwidth (MBs) 3120.4 1379.1

Inter-node

TeraGrid IA-64 Intel-Based Linux Cluster The TeraGrid project was launched by the the National Science Foundation with $53 million in funding to four sites: the National Center for Supercomputing Applications (NCSA) at the University of Illinois, Urbana-Champaign, the San Diego Supercomputer Center (SDSC) at the University of California, San

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

225

Diego, Argonne National Laboratory in Argonne, IL, and Center for Advanced Computing Research (CACR) at the California Institute of Technology in Pasadena. SDSCs TeraGrid cluster currently consists of 256 IBM cluster nodes, each with dual 1.5 GHz Intel R Itanium R 2 processors, for a peak performance of 3.1 teraops. The nodes are equipped with four gigabytes (GBs) of physical memory per node. The cluster is running SuSE Linux and is using Myricoms Myrinet cluster interconnect network. Table 10.2 shows the technical conguration of the IA64 cluster, on which the second part of the performance study has been done. Table 10.2: Technical Information of IA64 TeraGrid Cluster at SDSC IA-64 Cluster (tg-login.sdsc.teragrid.org) COMPONENT Architecture Access Nodes DESCRIPTION Linux Cluster quad-processor ECC SDRAM memory: 8 GB 2 nodes (8 processors) dual-processor

Compute Nodes

ECC SDRAM memory: 4 GB 262 nodes (524 processors) Intel R Itanium R 2, 1.5 GHz

Integrated 6 MB L3 cache Peak performance 3.1 Tops Myrinet 2000, Gigabit Ethernet, Fiber Channel 1.7 TB of NFS, 50 TB of GPFS (Parallel File System) Linux 2.4-SMP (SuSE SLES 8.0) Intel: Fortran77/90/95 C C++ GNU: Fortran77 C C++ Portable Batch System (PBS) with Catalina Scheduler

10.3.3

A soil-shallow-foundation interaction model as shown in Figure 10.4 has been set up to study the parallel performance. 3D brick element with 8 integration (Gaussian) points is used. The soil is modeled by Template3D elasto-plastic material model (Drucker-Prager model with Armstrong Frederick nonlinear kinematic hardening rule) and linear elasticity is assumed for the foundation. More advanced constitutive laws can be applied through Template3D model although the model used here suces the purpose of this research to show repartitioning triggered by plastication. It is shown in this research that the speedup by adaptive load balancing is signicant even for seemingly simple constitutive model. The material properties are shown in Table 10.3 and the vertical loading is applied at 5.0kN increments. The performance analysis has been carried out on DataStar supercomputer

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

226

at San Diego Supercomputing Center (P655+ 8-way nodes). Table 10.3: Material Constants for Soil-Foundation Interaction Model Soil Elastic modulus Poisson ratio Friction angle Cohesion Isotropic Hardening Kinematic Hardening E = 17400kPa = 0.35 = 37.1 c=0 Linear A/F nonlinear (ha = 116.0, Cr = 80.0) Foundation Elastic modulus Poisson ratio E = 21GPa = 0.2

10.3.4

As described in Section 10.2.3.2, the parameter ITR in ParMETIS describes the ratio between the time required for performing the inter-processor communications incurred during parallel processing compared to the time to perform the data redistribution associated with balancing the load. It acts like a switch on algorithmic approaches of ParMETIS repartitioning kernel. With ITR factor being very small, the ParMETIS tends to do that repartitioning which can minimize data redistribution cost. If the ITR factor is set to be very large, ParMETIS tends to minimize edge-cut of the nal repartition. In parallel design of PDD, if repartitioning is necessary to achieve load balance after each load increment, the whole AnalysisModel McKenna (1997) has to be wiped o thus a new analysis container can be dened to reload subsequent analysis steps. The data redistribution cost can be much higher than communication overhead only. In order to determine an adequate ITR value for our application, preliminary study needs to be performed to investigate the eectiveness of the URA. In this research, two extreme values of the ITR (0.001 and 1, 000, 000) are prescribed and then parallel analysis is carried out on 2, 4 and 8 processors to see how the partition/repartition algorithm behaves. Two soil-structure interaction models as shown in Figure 10.6 have been used in this parametric study. Timing data and partition gures have been collected to investigate the performance of dierent approaches. The one that tends to bring better performance will be adopted in subsequent parallel analysis for prototype 3D soil structure interaction problems. Figure 10.7 to Figure 10.12 shows the initial partition and nal repartition gures for two dierent types of algorithms. With ITR factor to be very small, the URA tends to present results that minimize data redistribution cost, in which diusive repartitioning approach is used. On the other hand, if the ITR factor is set to be very large, then the URA algorithm tends to give repartitioning with lowest edge cut but with considerably higher data redistribution cost.

Jeremi et al.

Lecture Notes

227

Figure 10.4: Example Finite Element Model of Soil-Foundation Interaction (Indication Only, Real Model Shown in Each Individual Section)

Jeremi et al.

Lecture Notes

228

Figure 10.5: FE Models (1,968 Elements, 7,500 DOFs) for Studying Soil-Foundation Interaction Problems

Figure 10.6: FE Models (4,938 Elements, 17,604 DOFs) for Studying Soil-Foundation Interaction Problems

Jeremi et al.

Lecture Notes

229

Figure 10.7: Partition and Repartition on 2 CPUs (ITR=1e-3, Imbal. Tol. 5%), FE Model (1,968 Elements,

7,500 DOFs)

Figure 10.8: Partition and Repartition on 2 CPUs (ITR=1e6, Imbal. Tol. 5%), FE Model (1,968 Elements, 7,500 DOFs)

Jeremi et al.

Lecture Notes

230

Figure 10.9: Partition and Repartition on 4 CPUs (ITR=1e-3, Imbal. Tol. 5%), FE Model (1,968 Elements,

7,500 DOFs)

Figure 10.10: Partition and Repartition on 4 CPUs (ITR=1e6, Imbal. Tol. 5%), FE Model (1,968 Elements, 7,500 DOFs)

Jeremi et al.

Lecture Notes

231

Figure 10.11: Partition and Repartition on 7 CPUs (ITR=1e-3, Imbal. Tol. 5%), FE Model (1,968 Elements,

7,500 DOFs)

Figure 10.12: Partition and Repartition on 7 CPUs (ITR=1e6, Imbal. Tol. 5%), FE Model (1,968 Elements, 7,500 DOFs)

Jeremi et al.

Lecture Notes

232

Figure 10.13: Partition and Repartition on 7 CPUs (ITR=1e-3, Imbal. Tol. 5%), FE Model (4,938 Elements, 17,604 DOFs)

Figure 10.14: Partition and Repartition on 7 CPUs (ITR=1e6, Imbal. Tol. 5%), FE Model (4,938 Elements, 17,604 DOFs)

Jeremi et al.

Lecture Notes

233

Figure 10.15: Timing Data of ITR Parametric Studies (1,968 Elements, 7,500 DOFs, Imbal. Tol. 5%)

Jeremi et al.

Lecture Notes

234

Figure 10.16: Relative Speedup of ITR=1e-3 over ITR=1e6 (1,968 Elements, 7,500 DOFs, Imbal. Tol. 5%)

Figure 10.17: Timing Data of ITR Parametric Studies (4,938 Elements, 17,604 DOFs, Imbal. Tol. 5%)

Jeremi et al.

Lecture Notes

235

Figures 10.15, 10.16 and 10.17 show the speedup data of parametric study on ITR factors. The purpose is to expose the the more ecient approach to do repartitioning for our specic parallel SFSI simulations, either scratch-remap approach (IT R = 1e6) or diusive approach (IT R = 1e 3). Through the study of this chapter, some conclusions can be drawn. 1. Smaller value of ITR (1e-3) outperforms larger value (1e6). The performance gain is up to 22.1% for 7 processors. As the model gets larger, the speedup tends to get better. 2. With small ITR value, the URA algorithm tends to give results for diusive partition/repartitioning scheme, which is good for performance for our application in overall due to the fact that the overhead associated with data redistribution in this research is very high. Diusive approach minimizes possible data movement thus delivers better performance. The drawback is the diusive approach typically gives very bad or even disconnected graphs with very high edge-cut as shown in Figures 10.7, 10.9 and 10.11. So careful attention must be paid to these graph structures when programing the nite element calculation. In this sense, the diusive algorithm is not as robust as scratch/remapping. One very important observation was that repetitive repartitionings tend to yield totally ill-connected graph. 3. With large ITR value, the URA algorithm adopts the scratch/remapping scheme which inevitably introduce

huge data redistribution cost. But this approach gives high quality graph and the integrity of original graph is well preserved as shown Figures 10.8, 10.10 and 10.12. This will be of great meaning for parallel nite element method based on substructure-type methods. Another important observation was, the scratch/remapping approach performed much more repartitionings than diusive approach for same analysis. Repetitive repartitionings by scratch/remapping method tends to totally migrate all elements out of their initial partitioning and repartitioning never stops even though the computation is stabilized (in the sense of formation of plastic zones). This also explains in part why the diusive approach can substantially outperform scratch/remapping. 4. Based on the timing analysis performed in this chapter, ITR=1e-3 is the best choice that brings substantially better performance over large ITR values. With the increase of number of processing units or the model size, the performance gain is more signicant as shown in Figures 10.15, 10.16 and 10.17. Robustness of the diusive approach has not caused much trouble in our application.

10.3.5

Timing routines have been implemented in PDD (and parts of OpenSees framework and other used libraries, such as Template3DEP/NewTemplate3Dep) to study the parallel performance. The preprocessing unit, like reading model data from le, has not been timed so the speed up here reects only algorithmic gain by graph partitioning. In the current phase of this research, the equation solving problem has not been addressed yet. More meaningful perspective would be to consider performance gains by simply switching from plain graph partitioning to adaptive graph partitioning, which is also the basic aim of this research. As we can see from the results below, adaptive graph partitioning improves the overall performance of elasto-plastic nite element computations. The partitioning/repartitioning overhead has been minimized by using parallel partitioner.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

236

As stated in previous sections of this chapter, there are a couple of key parameters that control performance of the adaptive load balancing algorithm. One is the ITR factor, and the other is the computational load imbalance tolerance. 1. ITR is the key parameter which determines the algorithmic approach of the adaptive load balancing scheme. Depending on dierent applications and network interconnections, this value can be set to very small (0.001) or very large (up to 1,000,000) and algorithm focus will be set to minimizing data redistribution or edge-cut respectively as explained in previous sections. 2. Computational Imbalance Load Tolerance is the other key factor aecting greatly the overall performance of the whole application codes. Basically speaking, with larger nite element model, the tolerance should be set higher due to the fact that data redistribution and subsequent analysis-restarting overhead can be substantially higher as the nite element model size increases. The performance tunings on ITR factor tend to yield consistent results as stated previously that smaller ITR (0.001) brings better performance over large ITR values. Diusive repartitioning algorithm outperforms scratch/remapping in our application.

While on the other hand, tuning on load imbalance studies has been more illusive. The rst conclusion is that load imbalance tolerance larger than 5% was not able to work robustly as the size of nite element model increases in the application study of this chapter. Detailed parametric studies have been performed on DataStar IBM Power4 and IA64 Intel clusters to indicate the eectiveness of the proposed adaptive PDD algorithm. Models with dierent sizes have been tested on various number of processors to show the scalability of computational performance. All results will be compared with static one-step Domain Decomposition approach to investigate the advantage of proposed PDD algorithm in nonlinear elastic-plastic nite element calculations. Table 10.4: Test Cases of Performance Studies Model Sizes (DOF) 4,035, 17,604, 32,091, 68,451 # of CPUs ITR Factors Imbalance Tolerance 3, 5, 7, 16, 32, 64 0.001, 1,000,000 5%, 10%, 20%

In the following sections, timing data and partition/repartition gures will be presented and results will be discussed at the end of this chapter.

10.3.5.1

The partition/repartition gures by PDD have been shown in Figure 10.18, 10.19, 10.20.

Jeremi et al.

Lecture Notes

237

Figure 10.18: 4,035 DOFs Model, 2 CPUs, ITR=1e-3, Imbal Tol 5%, PDD Partition/Repartition

Figure 10.19: 4,035 DOFs Model, 4 CPUs, ITR=1e-3, Imbal Tol 5%, PDD Partition/Repartition

Jeremi et al.

Lecture Notes

238

Figure 10.20: 4,035 DOFs Model, 8 CPUs, ITR=1e-3, Imbal Tol 5%, PDD Partition/Repartition

Figure 10.21: Timing Data of Parallel Runs on 4,035 DOFs Model, ITR=1e-3, Imbal Tol 5%

Jeremi et al.

Lecture Notes

239

Figure 10.22: Absolute Speedup Data of Parallel Runs on 4,035 DOFs Model, ITR=1e-3, Imbal Tol 5%

Figure 10.23: Relative Speedup of PDD over Static DD on 4,035 DOFs Model, ITR=1e-3, Imbal Tol 5%

Jeremi et al.

Lecture Notes

240

10.3.5.2

This is the same model as described before but with more elements as shown in 10.24. Timing data has been collected to indicate performance gains by adaptive load balancing Partition and repartition gures are shown from Figure 10.28 to 10.30. The partition/repartition gures by PDD have been shown in Figure 10.28, 10.29, 10.30.

Jeremi et al.

Lecture Notes

241

Figure 10.24: Finite Element Model of Soil-Foundation Interaction (4,938 Elements, 17,604 DOFs)

Jeremi et al.

Lecture Notes

242

Figure 10.25: Timing Data of Parallel Runs on 4,938 Elements, 17,604 DOFs Model, ITR=1e-3, Imbal Tol 5%

Figure 10.26: Absolute Speedup Data of Parallel Runs on 4,938 Elements, 17,604 DOFs Model, ITR=1e-3, Imbal Tol 5%

Jeremi et al.

Lecture Notes

243

Figure 10.27: Relative Speedup of PDD over Static DD on 4,938 Elements, 17,604 DOFs Model, ITR=1e-3, Imbal Tol 5%

Figure 10.28: 4,938 Elements, 17,604 DOFs Model, 2 CPUs, PDD Partition/Repartition, ITR=1e-3, Imbal Tol 5%

Jeremi et al.

Lecture Notes

244

Figure 10.29: 4,938 Elements, 17,604 DOFs Model, 4 CPUs, PDD Partition/Repartition, ITR=1e-3, Imbal Tol 5%

Figure 10.30: 4,938 Elements, 17,604 DOFs Model, 8 CPUs, PDD Partition/Repartition, ITR=1e-3, Imbal Tol 5%

Jeremi et al.

Lecture Notes

245

10.3.5.3

The mesh is shown in Figure 10.31. Speed up results are shown from Figure 10.32 to Figure 10.34. Partition and repartition gures are shown from Figure 10.35 to Figure 10.39.

Figure 10.31: Finite Element Model of Soil-Foundation Interaction (9,297 Elements, 32,091 DOFs)

Jeremi et al.

Lecture Notes

246

Figure 10.32: Timing Data of Parallel Runs on 9,297 Elements, 32,091 DOFs Model, ITR=1e-3, Imbal Tol 5%

Figure 10.33: Absolute Speedup Data of Parallel Runs on 9,297 Elements, 32,091 DOFs Model, ITR=1e-3, Imbal Tol 5%

Jeremi et al.

Lecture Notes

247

Figure 10.34: Relative Speedup of PDD over Static DD on 9,297 Elements, 32,091 DOFs Model, ITR=1e-3, Imbal Tol 5%

Figure 10.35: 9,297 Elements, 32,091 DOFs Model, 3 CPUs, PDD Partition/Repartition, ITR=1e-3, Imbal Tol 5%

Jeremi et al.

Lecture Notes

248

Figure 10.36: 9,297 Elements, 32,091 DOFs Model, 5 CPUs, PDD Partition/Repartition, ITR=1e-3, Imbal Tol 5%

Figure 10.37: 9,297 Elements, 32,091 DOFs Model, 7 CPUs, PDD Partition/Repartition, ITR=1e-3, Imbal Tol 5%

Jeremi et al.

Lecture Notes

249

Figure 10.38: 9,297 Elements, 32,091 DOFs Model, 16 CPUs, PDD Partition/Repartition, ITR=1e-3, Imbal Tol 5%

Figure 10.39: 9,297 Elements, 32,091 DOFs Model, 32 CPUs, PDD Partition/Repartition, ITR=1e-3, Imbal Tol 5%

Jeremi et al.

Lecture Notes

250

10.3.6

Algorithm Fine-Tuning

From performance analysis results in previous sections, it has been shown that adaptive graph partitioning algorithm based on element graph can improve overall load balance for nonlinear elastic-plastic nite element calculations. Speed up has been observed on example problems. While on the other hand, we can also see as the model size increases, the eciency of proposed PDD algorithm dropped sharply as shown in Figures 10.33 and 10.34. So the naive implementation of PDD does not work as expected. With load balancing, one expects that the performance of PDD should not be worse than the DD case. It otherwise implies that the PDD does not bring performance gain that can completely oset its own extra load balancing operations-related overheads. In this chapter, more detailed algorithm ne-tuning has been performed to address the problems we had in previous sections of the naive PDD implementation. In order to improve the overall eciency of proposed PDD algorithm. we have to consider two levels of costs when one wishes to balance the computational load among processing units. One is the data communication cost, and the other one is nite element model regeneration overhead associated with specic application problems.

Currently the adaptive graph partitioning algorithm does not consider the fact that the network communication patterns might dier much among processing nodes. The single IT R value indicates the algorithmic approach of the graph partitioning algorithm, but the real communication performance has not been addressed in the implementation. On the other hand, certain applications impose extra problem-dependent overhead to repartitioning operations. For example, whenever data communications happen, the nite element model has to be wiped o and regenerated. This is not inherent with the graph partitioning algorithm but still needs to be addressed in order to get the best performance. As observed in this chapter, model regeneration overhead increases when the nite element model becomes bigger. In order to improve the overall performance of our application, we hope to consider both data communication and model regeneration cost and create a new strategy through which we can adaptively monitor the extra overheads to assure that load balancing operation can oset both costs. This chapter will rst investigate the eect of load balance tolerance on performance and then a new globally adaptive strategy will be proposed to handle both communication and model regeneration overhead. Speedup analysis have been done to show performance gains.

10.3.7

If one nds out that the application-associated overhead (say, model regeneration cost) overwhelms when repartitioning happens, the most natural way to improve performance is to increase the load imbalance tolerance of the adaptive repartition routine. In this way, one hopes to increase the critical load imbalance that can trigger the balancing routine and so that the repartition counts can be reduced. As a result, model regeneration cost can

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

251

do less harm to the overall performance. This should rather viewed as a work-around and has not been eective in our application. The tuning approach aims at improving eciency of previous runs that failed showing speedup over static domain decomposition method. Shallow foundation model with 9,297 Elements, 32,091 DOFs has been chosen to study the eect of imbalance tolerance on parallel performance. Model setup has been the same as in previous sections. Speedup analysis results have been shown in Figures 10.40, 10.41 and 10.42.

Figure 10.40: Timing Data of Parallel Runs on 9,297 Elements, 32,091 DOFs Model, ITR=1e-3, Imbal Tol 20%

Jeremi et al.

Lecture Notes

252

Figure 10.41: Absolute Speedup Data of Parallel Runs on 9,297 Elements, 32,091 DOFs Model, ITR=1e-3, Imbal Tol 20%

Figure 10.42: Relative Speedup of PDD over Static DD on 9,297 Elements, 32,091 DOFs Model, ITR=1e-3, Imbal Tol 20%

Jeremi et al.

Lecture Notes

253

From the performance results, we can see that increasing load imbalance tolerance does not lead to eciency for our application. As the number of processing units increases, the whole performance of application codes deteriorates. It is also important to note that the adaptive graph partitioning/repartitioning kernel in ParMETIS has not been capable of producing adequate partitions for nite element calculations when the load imbalance tolerance is larger than the recommended 5% Karypis et al. (2003). The application crushed with 20% imbalance tolerance for same models tested in previous sections. The conclusion reached for the application in this chapter is that load imbalance tolerance larger than 5% has not been proved more ecient. This can also be explained in more details. In the implementation of ParMETIS, load imbalance tolerance is one of the most important parameters in the sense that this value determines whether repartition will be switched on. The other equally signicant implication of this value comes from the fact that it also establishes target load imbalance residual to be achieved after adaptive load balancing. That means for each repartition, the ParMETIS will only reduce the load imbalance to the provided tolerance. In current implementation, the load imbalance tolerance is set to be the same for both switch-on and target values, which is not capable of bringing the best performance into our application due to the fact that aside

from data redistribution cost, analysis model reconstruction is equally expensive. The dilemma is described by numerical example as shown in Table 10.5.

Table 10.5: Observation on Load Imbalance Tolerance %5 Model 20,476 Elements, 68,451 DOFs CPUs Imbalance Before Imbalance After Model Regeneration Total Step Time 32 7.018% 4.9% 57.2934 seconds 140.961 seconds

We can easily see that tiny portion of data movement to balance out 7.018 4.9 = 2.228% loads still invoked analysis model regeneration, which accounts for extra overhead that is about 40.6% of total step time. Because the load balance tolerance is also the target value that the repartitioning operation hopes to achieve. The implication is that after repartitioning, the load distribution among processing units is barely under this acceptable tolerance. The performance study conducted so far showed that continuous plastication can easily creates load imbalance over this tolerance so another round of repartitioning would be launched again. It greatly brings down the performance of the whole application when the huge data redistribution overhead is taken just to overcome a tiny imbalance. This explains why changing the tolerance was not able to bring better performance in our application. In order to improve performance while still minimizing load imbalance, we hope to maximize the eciency of model regeneration routine in our application. This is a two-fold statement, rstly, we dont want to blindly

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

254

increase the load imbalance because it basically claims we fail our adaptive PDD algorithm by not switching on repartitioning (5% is suggested by the author of ParMETIS Karypis et al. (2003) and has been proved to be the most stable value in this chapter), secondly, with each repartitioning, we hope to achieve "perfect balance" as much as possible and in this way, the huge model regeneration cost can be oset by performance gain. What was proposed as future extension of this chapter is the idea of dual load imbalance tolerances. Load balancing triggering tolerance and the target tolerance can be dened separately. We can set higher triggering tolerance to reduce the number of repartition counts, while on the other hand a strict target tolerance can be set close to 1.0 to get better load distribution out of the balancing routine. With proposed approach, our application in this chapter will be able to fully take advantage of the repartition routines without sacricing too much on model regenerations.

10.3.8

One signicant drawback of current implementation is that neither network communication nor model regeneration cost has been considered in element-graph-based type domain decomposition algorithm. Element graph only records computational load carried by each element. Only one IT R factor characterizes algorithmic approach of the load balancing operation and this is apparently too crude for complicated network/hardware congurations. The ignorance of the repartitioning-associated overheads inherent with application codes can lead to serious performance drop of the proposed PDD algorithm as shown in Figure 10.43.

Figure 10.43: Absolute Speedup Data of Parallel Runs on 9,297 Elements, 32,091 DOFs Model, ITR=1e-3, Imbal Tol 5%

Jeremi et al.

Lecture Notes

255

This drawback can harm the overall performance of the whole application code more seriously when the simulation is to be run on heterogeneous networks, which means we can have dierent network connections and nodes with varied computational power. The dilemma is, without exact monitoring of network communication and local model regeneration costs, we can easily sacrice the performance gain by load balancing operations. A second approach proposed in this chapter was the idea of modied Globally Adaptive PDD algorithm. The novelty comes from the fact that both data redistribution and analysis model regeneration costs will be monitored during execution. Load balancing will be triggered only when the performance gain necessarily oset the extra cost associated with the whole program. Domain graph structures will be kept intact till successful repartitioning happens. Meanwhile all elemental calculations will be timed to provide graph vertex weights. Data will be accumulated till algorithm restart happens, when all analysis model and vertex weights will be nullied. This improvement aims at handling network communication and any specic application-associated overheads automatically at the global level in order to remedy the drawback that the element graph repartitioning kernel currently supported by ParMETIS is not capable of directly reecting this application level overheads. The new strategy is to automatically monitor network communication and local model regeneration timings which will be integrated to the entry of load balancing routines to act as additional triggers of the operation along with the load imbalance tolerance. Performance study shows that PDD algorithm with the new additions signicantly improve performance even when the number of processing units is large. This modication xes the drawback shown in previous sections that the performance of PDD was beaten by static domain decomposition when the number of processors increases. This strategy is called to be globally adaptive because both data communication and model regeneration costs are monitored at the application level, which tells best how the real application performs on all kinds of networks. Whatever the network/hardware congurations might be, real application runs always deliver the most accurate performance counters. This information can be applied on top of graph partitioning algorithm as a supplement to account for the drawback that the algorithm kernel is not capable of integrating global data communication costs.

10.3.8.1

Implementations

We can dene the global overhead associated with load balancing operation as two parts, data communication cost Tcomm and nite element model regeneration cost Tregen , Toverhead := Tcomm + Tregen Performance counters have been setup to study both. Tcomm Data communication patterns characterizing the network conguration can be readily measured as the program runs the initial partitioning. As described in previous sections, initial domain decomposition needs to be done to send elements over to processing nodes. This step is necessary for parallel nite element

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

(10.3)

Lecture Notes

256

processing and it provides perfect initial estimate how the communication pattern of the application performs on specic networks. Timing routines have been added to automatically measure the communication cost. This cost is inherently changing as the network condition might vary as simulation progresses, so whenever data redistribution happens, this metric will be automatically updated to reect the network conditions. Tregen Model regeneration cost basically comes from the fact that if data redistribution happens, the analysis model needs to be regenerated to reect changes of nodes and elements inside the domain. Detailed operations include renumbering DOFs and rehandling constraints. This part of cost is application-dependent. In current implementation of PDD, eorts have been made to set up timing stop at the entry and exit of model regeneration routines to get the accurate data for the extra overhead. It is also important to note that model regeneration happens when the initial data distribution nishes, again the initial domain decomposition phase provides perfect initial estimate of the model regeneration cost on any specic hardware congurations. Naturally, for the load balancing operations to pay o, the Toverhead has to be oset by the performance gain Tgain . This chapter also creates a strategy to estimate the performance gain Tgain even before the load balancing

operation happens and this metric provides global control on top of the existing graph repartitioning algorithm. As implemented in previous sections, the computational load on each element is represented by the associated vertex weight vwgt[i]. If the SU M operation is applied on every single processing node, the exact computational distribution among processors can be obtained as total wall clock time for each CPU as shown in Equation 10.4,

n

Tj :=

i=1

vwgt[i], j = 1, 2, . . . , np

(10.4)

in which n is the number of elements on each processing domain and np is the number of CPUs. If we dene, Tsum := sum(Tj ), Tmax := max(Tj ), and Tmin := min(Tj ), j = 1, 2, . . . , np (10.5)

one always hope to minimize Tmax because in parallel processing, Tmax controls the total wall clock time. By load balancing operations, we mean to deliver evenly distributed computational loads among processors. So theoretically, the best execution time is, Tbest := Tsum /np, and Tj Tbest , j = 1, 2, . . . , np if the perfect load balance is to be achieved. Based on denitions above, the best performance gain Tgain one can obtain from load balancing operations can be calculated as, Tgain := Tmax Tbest Finally, the load balancing operation will be benecial IF AND ONLY IF Tgain Toverhead = Tcomm + Tregen

Jeremi et al. University of California, Davis

(10.6)

(10.7)

(10.8)

Version: March 21, 2012, 17:15

Lecture Notes

257

10.3.8.2

Performance Results

The newly improved design has been compared to the old design to see the eectiveness of the globally adaptive switch of PDD algorithm.

Figure 10.44: Performance of Globally Adaptive PDD on 9,297 Elements, 32,091 DOFs Model, ITR=1e-3, Imbal Tol 5%

Jeremi et al.

Lecture Notes

258

Performance of Globally Adaptive PDD on 20,476 Elements, 68,451 DOFs Model, ITR=1e-3,

Jeremi et al.

Lecture Notes

259

From Figures 10.44 and 10.45, advantage of the improved globally adaptive PDD algorithm have clearly been shown. After considering the eect of both data communication and model regeneration costs, the adaptive PPD algorithm necessarily outperforms the static Domain Decomposition approach as expected. This new design also signicantly improves the overall scalability of the proposed PDD algorithm as shown in Figure 10.46 and 10.47.

Figure 10.46: Scalability Study on 4,938 Elements, 17,604 DOFs Model, ITR=1e-3, Imbal Tol 5%

Jeremi et al.

Lecture Notes

260

Figure 10.47: Scalability Study on 9,297 Elements, 32,091 DOFs Model, ITR=1e-3, Imbal Tol 5%

Jeremi et al.

Lecture Notes

261

10.3.9

The ultimate purpose of this chapter is to develop an ecient parallel simulation tool for large scale earthquake analysis on prototype SFSI system. After in-depth development-rening process conducted in previous sections, real 3-bent production models have been set up to study the parallel performance of the proposed PDD algorithm using real world earthquake ground motions.

10.3.9.1

As described in later sections, various sizes of a 3 bent bridge SFSI system has been developed to study dynamic behaviors of the whole system in dierent frequency domain. These models provide perfect test cases for parallel scalability study of our proposed PDD algorithm. Detailed model description will be presented in later chapters of this chapter and only model size and mesh pictures are shown here to indicate the range of model sizes we have covered.

Figure 10.48: Finite Element Model - 3 Bent SFSI, 56,481 DOFs, 13,220 Elements, Frequency Cuto > 3Hz, Element Size 0.9m, Minimum G/Gmax 0.08, Maximum Shear Strain 1%

Jeremi et al.

Lecture Notes

262

Figure 10.49: Finite Element Model - 3 Bent SFSI, 484,104 DOFs, 151,264 Elements, Frequency Cuto 10Hz, Element Size 0.3m, Minimum G/Gmax 0.08, Maximum Shear Strain 1%

Figure 10.50: Finite Element Model - 3 Bent SFSI, 1,655,559 DOFs, 528,799 Elements, Frequency Cuto 10Hz, Element Size 0.15m, Minimum G/Gmax 0.02, Maximum Shear Strain 5%

Jeremi et al.

Lecture Notes

263

10.3.9.2

Scalability Runs

The models with dierent detail levels have been subject to 1997 Northridge earthquake respectively for certain time steps and total wall clock time has been recorded to analyze the parallel scalability of our proposed PDD. The result is presented in Figure 10.51.

1024

Parallel Speedup

512

256

128 64 32 16 8 1

1 8 16 32 64

128

256

512

1024

Number of CPUs

Figure 10.51: Scalability Study on 3 Bent SFSI Models, DRM Earthquake Loading, Transient Analysis, ITR=1e-3, Imbal Tol 5%, Performance Downgrade Due to Increasing Network Overhead

Jeremi et al.

Lecture Notes

264

10.3.10

Conclusions

Through detailed performance studies as presented in previous sections, some conclusions can be drawn and future directions can be noted. Plastic Domain Decomposition (PDD) algorithm based on adaptive multilevel graph partitioning kernels has been shown to be eective for elastic-plastic parallel nite element calculations. PDD algorithm consistently outperforms classical Domain Decomposition method for models tested so far in this chapter as shown in Figures 10.52 and 10.54.

Figure 10.52: Relative Performance of PDD over DD, Shallow Foundation Model, Static Loading, ITR=1e-3, Imbal Tol 5%

Jeremi et al.

Lecture Notes

265

There are some parameters that can be calibrated in the current implementation. As indicated by results of thorough numerical tests, ITR=0.001 and load imbalance tolerance ubvec=1.05 (5%) should be adopted and studies on our application in this chapter have shown they are adequate and able to bring performance not worse than the commonly used domain decomposition method in parallel nite element analysis. For the parameters suggested in the chapter, we can see a general trend that the eciency of PDD will drop as the number of processors increases. This can be explained. The implication of increasing processing units is that the subdomain problem size will decrease. It is naturally evident that the repartition load balancing wont be able to recover the overhead by balancing o small size local calculations. The improved design of globally adaptive PDD algorithm has been implemented in this chapter and both data communication and model regeneration costs associated with graph repartitioning have been integrated into the new globally adaptive strategy. With the new design, it has also been shown that the PDD algorithm consistently outperforms classic one step domain decomposition algorithm and better scalability can be obtained as shown in Figure 10.53. It has been shown that even for large number of processors, the current implementation can always guarantee that the performance of PDD is not worse than static DD method as shown in Figure 10.54. (the repartition routine has less than 5% overhead of the total wall clock time).

Jeremi et al.

Parallel Speedup

Jeremi et al.

128

120

112

e pe du p

104

a tic lS

nts

96

e Th or e

88

DO

2 s (1

5,2

e lem E 5

80

721 98,

72

64

56

118,2

75 DO

lemen 037 E , 6 3 ( s

ts)

Lecture Notes

48

40

32

1 68,45 1 32,09

17,6 s (4,938 04 DOF ts) Elemen

DO

ents)

24

16 8 0 0 8 16 24 32 40

DO

297 E Fs (9,

48

56 64 72 80 Number of CPUs

88

96

Figure 10.53: Scalability of PDD, Static Loading, Shallow Foundation Model, ITR=1e-3, Imbal Tol 5%

266

Computational Geomechanics Group

Jeremi et al.

Lecture Notes

Figure 10.54: Relative Speedup of PDD over DD, Static Loading, Shallow Foundation Model, ITR=1e-3, Imbal Tol 5%

267

Lecture Notes

268

If the problem size is xed, there exists an optimum number of processors that can bring the best performance of the proposed load balancing algorithm. As the number of processing units increases after this number, the eciency of proposed algorithm drops, which is understandable because the local load imbalance is so small overall that balancing gain wont oset the extra cost associated with repartitioning. But still the bottom line of proposed adaptive PDD algorithm is that it can run as fast as static one-step domain decomposition approach with less than 5% overhead of repartitioning routine calls. On the other hand, if the number of processing units is xed, bigger nite element model will exhibit better performance. The conclusion is shown clearly in 3D in Figure 10.54. It is also worthwhile to point out that even without comparing with classical DD, PDD itself exhibits deteriorating performance as the number of processing units increases. Here the reproduction of Figure 10.53 is presented with some downside performance noted as shown in Figure 10.55.

Jeremi et al.

Computational Geomechanics Group

Parallel Speedup

Jeremi et al.

256

Lecture Notes

128

64

32

1 16 32

Number of CPUs

269

Figure 10.55: Full Range Scalability of PDD, Static Loading, Shallow Foundation Model, ITR=1e-3, Imbal Tol 5%, Performance Downgrade Due to Increasing

Network Overhead

Lecture Notes

270

The implication is explained as follows: The performance drop partly is due to the communication overhead gets bigger and bigger so parallel processing will not be able to oset the communication loss. It is also noted that as the number of processing units increases, the elemental level calculation drops very scalably with the number of CPUs. This is inherently advantage of the proposed PDD algorithm. PDD through domain decomposition is very scalable for local level calculations because inherently local comp is element-based. when elements are distributed, loads are spread out evenly (during initial and redistribution). So as the number of CPU increases, the equation solving becomes more expensive. For the case of 56,481 DOFs prototype model with DRM earthquake loading, it has been observed that for sequential case (1 CPU), elemental computation takes 70% of time. As for parallel case (8 CPUs), we optimized parallel elemental computations through PDD, elemental computation only accounts for about 40%. As the number of CPU increases, parallel case (32 CPUs), the local level computation will only take less than 10% of total wall clock time. In other words, as the number of CPUs increases, PDD loses scalability because of the equation solving

now dominates. As being discussed in Chapter 10.4, the parallel direct solver itself is not scalable up to large number of CPUs Demmel et al. (1999a). Parallel iterative solver is much more scalable but dicult to guarantee convergence. This is now also the most important topic in the whole scientic computing community. For one set of xed algorithm parameters, such as ITR and load imbalance tolerance, basic conclusion is there exists an optimal number of processors that can bring best performance and as nite element model size increases, this number increases as listed in Table 10.6. Table 10.6: Best Performance Observed for ITR=0.001, Load Imbalance Tolerance %5 # of DOFs Speedup # of CPUs 4,035 17,604 32,091 68,451 1.553 1.992 1.334 1.068 4 7 7 16

The second point is related to the implementation of the multilevel graph partitioning algorithm. In current implementation of ParMETIS used in this chapter, vertex weight can only be specied as an int. That means in order to get timing data from local level calculation for each element, double data returned by MPI timing routine has to be converted to int. Signicant digit loss can happen depending on what accuracy the system clock can carry. We can also adjust the vertex weight by amplifying the timing by scale factors in order to save eective digits. 10 millisecond has been used in this chapter to represent the eective timing digits when converting from double to int.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

271

10.4

10.4.1

Introduction

Finite element method has been the most extensively used numerical method in computational mechanics. Equation solver is the numerical kernel of any nite element package. Gauss elimination type direct solver has dominated due to its robustness and predictability in performance. As modern computer becomes more and more powerful, more advanced and detailed models need to be analyzed by numerical simulation. Direct solver is not the favorite choice for large scale nite element calculations because of high memory requirements and the inherent lack of parallelism of the method itself. The motivation for presented work on iterative solvers stems from the need to expand the toolset of parallel iterative solvers for large scale simulation problems related to Earthquake-Soil-Structure interaction problems In this section, the eectiveness of Krylov iterative methods has been tested in solving soil-structure interaction

problems. Preconditioning techniques have been introduced. Robustness of iterative solvers has been investigated on equation systems from real soil-structure interaction problems. Several popular parallel algorithms and tools have been collected and implemented on PETSc platform to solve the SFSI problems. Performance study has been carried out using IA64 super computers at San Diego Supercomputing Center. A complete implementation has been developed within our computational system, within MOSS libraries, with extensive use of ParMETIS, and other material and numerical libraries.

10.4.2

Projection techniques are dened as methods to nd approximate solutions x for Ax = b (A Rnn ) in a subspace W of dimension m. Then in order to determine x , we need m independent conditions. One way to obtain these is by requiring the residual b Ax is orthogonal to a subspace V of dimension m, i.e., x W , b Ax V The conditions shown in Equation 10.9 are known as Petrov-Galerkin conditions (Bai). There are two key questions to answer if one wants to use projection techniques in solving large scale linear systems. Dierent answers lead to many variants of the projection method. Choice of Subspaces Krylov subspaces have been the favorite of most researchers and a large family of methods have been developed based on Krylov subspaces. Typically people choose either V = W or V = AW with V and W both Krylov subspaces. Enforcement of Petrov-Galerkin Conditions

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

(10.9)

Lecture Notes

272

Arnoldis procedure and Lanczos algorithm are two choices for building orthogonal or biorthogonal sequence to enforce the projection conditions. The iterative methods discussed in this section are generally split into two categories, one based on Arnoldis procedure and the other on Lanczos biorthoganalization. The most popular for the rst family are Conjugate Gradient and General Minimum Residual methods, while Bi-Conjugate Gradient and Quasi-Minimum Residual methods represent the Lanczos family.

10.4.2.1

The conjugate gradient (CG) algorithm is one of the best known iterative techniques for solving sparse symmetric positive denite (SPD) linear systems. This method is a realization of an orthogonal projection technique onto the Krylov subspace Km (A, r0 ), where r0 is the initial residual. Because A is symmetry, some simplications resulting from the three-term Lanczos recurrence will lead to more elegant algorithms (Demmel, 1997). ALGORITHM CG (Saad, 2003)

1. Compute r0 := b Ax0 , p0 := r0 2. For j = 0, 1, , until convergence, Do 3. 4. 5. 6. 7. 8. EndDo j := (rj , rj )/(Apj , pj ) xj +1 := xj + j pj rj +1 := rj j Apj j := (rj +1 , rj +1 )/(rj , rj ) pj +1 := rj +1 + j pj

Applicability Matrix A is SPD. Subspaces Choose W = V = Km (A, r0 ), in which initial residual r0 = b Ax0 . Symmetric Lanczos Procedure This procedure can be viewed as a simplication of the Arnoldis procedure when A is symmetric. Great three-term Lanczos recurrence is discovered when the symmetry of A is considered (Demmel, 1997). Optimality If A is SPD and one chooses W = V , enforcing Petrov-Galerkin conditions minimizes the A-norm of the error over all vectors x W , i.e., x solves the problem, min x x

A, x

xW

= A 1 b

University of California, Davis

(10.10)

Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

273

From the lemma above, one can derive global minimization property of the Conjugate Gradient method. The vector xk in the Conjugate Gradient method solves the minimization problem min (x) =

x

1 x x 2

2 A, x

x0 Kk (A, r0 )

(10.11)

Convergence In exact arithmetic, the Conjugate Gradient method will produce the exact solution to the linear system Ax = b in at most n steps and it owns the superlinear convergence rate. The behavior of Conjugate Gradient algorithm in nite precision is much more complex. Due to rounding errors, orthogonality is lost quickly and nite termination does not hold anymore. What is more meaningful in application problems would be to use CG method for solving large, sparse, well-conditioned linear systems in far fewer than n iterations.

10.4.2.2

GMRES

The Generalized Minimum Residual method is able to deal with more general type of matrices. ALGORITHM GMRES (Saad, 2003) 1. Compute r0 := b Ax0 , := r0 2 , and v1 := r0 / 2. For j = 1, 2, , m, Do 3. 4. 5. 6. 7. 8. 9. Compute j := Avj For i = 1, , j , Do hij := (j , vi ) j := j hij vi EndDo hj +1,j = j

2.

vj +1 = j /hj +1,j

10. EndDo m = {hij }1im+1,1j m 11. Dene the (m + 1) m Hessenberg matrix H m y 2 , and xm = x0 + Vm ym 12. Compute ym , the minimizer of i e1 H

Applicability Matrix A is nonsingular. Subspaces Choose W = Km (A, r0 ) and V = AW = AKm (A, r0 ), in which initial residual r0 = b Ax0 . Arnoldis Procedure Classic Arnoldis procedure (modied Gram-Schmidt) is followed in GMRES (Bai).

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

274

Optimality If one chooses V = AW , enforcing Petrov-Galerkin conditions solves the least square problem b Ax Convergence It has been shown that in exact arithmetic, GMRES can not breakdown and will give exact solutions in at most n steps. In practice, the maximum steps GMRES can run depends on the memory due to the fact it needs to store all Arnoldi vectors. Restarting schemes have been proposed for a xed m, which is denoted by GMRES(m). Typical value for m can be m [5, 20]. GMRES(m) can not breakdown in exact arithmetic before the exact solution has been reached. But it may never converge for m < n (Bai).

2

= min b Ax

xW

(10.12)

10.4.2.3

These two methods are based on nonsymmetric Lanczos procedure, which is quite dierent from Arnoldis in the sense that it formulates biorthogonal instead of orthogonal sequence. They are counterparts of CG and GMRES method, which follows similar derivation procedure except the Lanczos biorthogonalization is used instead of Arnoldis procedure (Bai).

10.4.3

Preconditioning Techniques

Lack of robustness is a widely recognized weakness of iterative solvers relative to direct solvers. Using preconditioning techniques can greatly improve the eciency and robustness of iterative methods. Preconditioning is simply a means of transforming the original linear system into one with the same solution but easier to solve with an iterative solver. Generally speaking, the reliability of iterative techniques, when dealing with various applications, depends much more on the quality of the preconditioner than on the particular Krylov subspace accelerator used. The rst step in preconditioning is to nd a preconditioning matrix M . The matrix M can be dened in many dierent ways but there are a few minimal requirements the M is supposed to satisfy (Benzi, 2002). 1. From practical point of view, the most important requirement of M is that it should be inexpensive to solve linear system M x = b. This is because the preconditioned algorithm will all require a linear system solution with the matrix M at each step. 2. The matrix M should be somehow close to A and it should not be singular. We can see that actually most powerful preconditioners are constructed directly from A. 3. The preconditioned M 1 A should be well-conditioned or has very few extreme eigenvalues thus M can accelerate convergence dramatically. Once a preconditioner M is available, there are three ways to apply it.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

275

1. Left Preconditioning

M 1 Ax = M 1 b 2. Right Preconditioning

(10.13)

AM 1 u = b, x M 1 u 3. Split Preconditioning

(10.14)

It is a very common situation that M is available in factored form M = ML MR , in which, typically, ML and MR are triangular matrices. Then the preconditioning can be split,

1 1 1 ML AMR u = b, x MR u

(10.15)

It is imperative to preserve symmetry when the original matrix A is symmetric, so the split preconditioner seems mandatory in this case.

Consider that a matrix A that is symmetric and positive denite and assume that a preconditioner M is available. The preconditioner M is a matrix that approximates A in some yet-undened sense. We normally require that the M is also symmetric positive denite. In order to preserve the nice SPD property, in the case when M is available in the form of an incomplete Cholesky factorization, M = LLT , people can simply just use the split preconditioning, which yields the SPD matrix L1 ALT u = L1 b, x LT u (10.16)

However, it is not necessary to split the preconditioner in this manner in order to preserve symmetry. Observe that M 1 A is self-adjoint for the M inner product (x, y )M (M x, y ) = (x, M y ) since (M 1 Ax, y )M = (Ax, y ) = (x, Ay ) = (x, M (M 1 A)y ) = (x, M 1 Ay )M (10.18) (10.17)

Therefore, an alternative is to replace the usual Euclidean inner product in the CG algorithm with the M inner product (Saad, 2003). If the CG algorithm is rewritten for this new inner product, denoting by rj = b Axj the original residual and by zj = M 1 rj the residual for the preconditioned system, the following sequence of operations is obtained, ignoring the initial step: 1. j := (zj , zj )M /(M 1 Apj , pj )M , 2. xj +1 := xj + j pj ,

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

276

3. rj +1 := rj j Apj and zj +1 := M 1 rj +1 , 4. j := (zj +1 , zj +1 )M /(zj , zj )M , 5. pj +1 := zj +1 + j pj . Since (zj , zj )M = (rj , zj ) and (M 1 Apj , pj )M = (Apj , pj ), the M inner products do not have to be computed explicitly. With this observation, the following algorithm is obtained. ALGORITHM Preconditioned CG (Saad, 2003) 1. Compute r0 := b Ax0 , z0 := M 1 r0 , p0 := z0 2. For j = 0, 1, , until convergence, Do 3. 4. 5. 6. 7. 8. 9. EndDo j := (rj , zj )/(Apj , pj ) xj +1 := xj + j pj rj +1 := rj j Apj zj +1 := M 1 rj +1 j := (rj +1 , zj +1 )/(rj , zj ) pj +1 := zj +1 + j pj

10.4.4

Preconditioners

Finding a good preconditioner to solve a given sparse linear system is often viewed as a combination of art and science. Theoretical results are rare and some methods work surprisingly well, often despite expectations. As it is mentioned before, the preconditioner M is always close to A in some undened-yet sense. Some popular preconditioners will be introduced in this section.

10.4.4.1

Jacobi Preconditioner

This might be the simplest preconditioner people can think of. If A has widely varying diagonal entries, we may just use diagonal preconditioner M = diag(a11 , , ann ). One can show that among all possible diagonal preconditoners, this choice reduces the condition number of M 1 A to within a factor of n of its minimum value.

10.4.4.2

Another simple way of dening a preconditioner that is close to A is to perform an incomplete Cholesky L T , but with less or no ll-ins factorization of A. Incomplete factorization formulates an approximation of A L relative to the complete factorization A = LLT (Demmel, 1997).

Jeremi et al.

Lecture Notes

277

ALGORITHM Incomplete Cholesky Factorization (Saad, 2003) 1. For j = 1, 2, , n, Do 3. 4. 5. 6. 7. 8. EndDo ljj := ajj

j 1 2 k=1 ljk

j 1 k=1 lik ljk )/ljj

There are many ways to control the number of ll-ins in IC factorization. No ll-in version of incomplete Cholesky factorization IC(0) is rather easy and inexpensive to compute. On the other hand, it often leads to a very crude approximation of A, which may result in the Krylov subspace accelerator requiring too many iterations to converge. To remedy this, several alternative incomplete factorizations have been developed by researchers by allowing more ll-in in L, such as incomplete Cholesky factorization with dropping threshold IC( ). In general, more accurate IC factorizations require fewer iterations to converge, but the preprocessing cost to compute the factors is higher.

10.4.4.3

Incomplete factorization preconditioners are quite eective for many application problems but special care must be taken in order to avoid breakdowns due to the occurrence of non-positive pivots during the incomplete factorization process. L T has been established for certain classes of matrices. For The existence of an incomplete factorization A L the class of M -matrices, the existence of incomplete Cholesky factorization was proved for arbitrary choices of the sparsity pattern (Meijerink and van der Vorst, 1977). The existence result was extended shortly thereafter to a somewhat larger class (that of H -matrices with positive diagonal entries) (Manteuel, 1980; Varga et al., 1980; Robert, 1982). Benzi and Tma (2003) presents reviews on the topic of searching for robust incomplete factorization algorithms and an robust algorithm based on A-Orthogonalization has been proposed. In order to construct triangular factorization of A, the well-known is not the only choice. (2003) shows how the factorization A = LDL

T

orthogonalization process applied to the unit basis vectors e1 , e2 , , en . This is simply the Gram-Schmidt process with respect to the inner product generated by the SPD matrix A. This idea is not new and as a matter of fact, it was originally proposed at as early as 1940s in Fox et al. (1948). It has been observed in Hestenes and Stiefel (1952) that A-orthogonalization of the unit basis vectors is closely related to Gaussian elimination but this algorithm costs twice as much as the Cholesky factorization in the dense case.

10.4.4.3.1

In reference Benzi et al. (1996) A-orthogonalization has been exploited to construct factored sparse approximate inverse preconditioners noting the fact that A-orthogonalization also produces the inverse factorization A1 =

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

278

ZD1 Z T (with Z unit upper triangular and D diagonal). Because the A-orthogonaliza-tion, even when performed incompletely, is not subject to pivot breakdowns, these preconditioners are reliable (Benzi et al., 2000). However, they are often less eective than incomplete Cholesky preconditioning at reducing the number of PCG iterations and their main interest stems from the fact that the preconditioning operation can be applied easily in parallel because triangular solve is not necessary in approximate inverse preconditioning. Reference Benzi and Tma (2003) investigates the use of A-orthogonalization as a way to compute an incomplete factorization of A rather than A1 thus a reliable preconditioning algorithm can be developed. The basic A-orthogonalization procedure can be written as follows (Benzi et al., 2000). ALGORITHM Incomplete Factored Approximate Inverse (Benzi et al., 1996) 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. EndDo 11. Let zi := zi

(i1)

Let zi

(0)

= ei , for i = 1, 2, , n

(i1) (i) (i1) (i1)

:= aT i zj

(i1)

pj pi

(i1)

(i1)

Apply dropping to

(i) zj

, for i = 1, 2, , n.

The basic algorithm described above can suer a breakdown when a negative or zero value of a pivot pi . When

T no dropping is applied, pi = zi Azi > 0. The incomplete procedure is well dened, i.e., no breakdown can occur,

if A is an H -matrix (in the absence of round-o). In the general case, breakdowns can occur. Breakdowns have a crippling eect on the quality of the preconditioner. A negative pi would result in an approximate inverse which is not positive denite; a zero pivot would force termination of the procedure, since step (7) cannot be carried out. The way proposed to avoid non-positive pivots is simply to recall that in the exact A-orthogonalization process, the pi s are the diagonal entries of matrix D which satises the matrix equation Z T AZ = D hence for 1 < i < n

T pi = zi Azi > 0

(10.19)

(10.20)

since A is SPD and zi = 0. In the exact process, the following equality holds

T pi = zi Azi = aT i zi T and pj = zi Azj = aT i zj

(10.21)

Version: March 21, 2012, 17:15

Jeremi et al.

Lecture Notes

279

Clearly it is more economical to compute the pivots using just inner product aT i zi rather than the middle expression involving matrix-vector multiply. However, because of dropping and the resulting loss of A-orthogonality in the approximate z -vectors, such identities no longer hold in the inexact process and for some matrices one can have aT i i z

T z i Az i

(10.22)

T The robust algorithm requires that the incomplete pivots p i s be computed using the quadratic form z i Az i

throughout the AINV process, for i = 1, 2, , n. ALGORITHM Stabilized Incomplete Approximate Inverse (Benzi et al., 2000) 1. 2. 3. 4. 5. 6. 7. Let zi

(0)

= ei , for i = 1, 2, , n

(i1)

(i1) (i) (i1)

For j = i, i + 1, , n, Do

(i1) T := vi zj (i1)

pj pi

(i1) (i1)

Apply dropping to zj

( i)

and pi := pi

(i1)

, for i = 1, 2, , n.

Obviously, the robust (referred to as SAINV) and plain algorithm are mathematically equivalent. However, the incomplete process obtained by dropping in the z -vectors in step (9) of the robust algorithm leads to a reliable approximate inverse. This algorithm, in exact arithmetic, is applicable to any SPD matrix without breakdowns. The computational cost of SAINV is higher than basic AINV and special care has to be taken to do sparse-sparse matrix-vector multiply.

10.4.4.3.2

Consider now the exact algorithm (with no dropping) and write A = LDLT with L unit lower triangular and D diagonal. Observe that L in the LDLT factorization of A and the inverse factor satisfy AZ = LD or L = AZD1 (10.23)

where D is the diagonal matrix containing the pivots. This easily follows from Z T AZ = D

Jeremi et al.

and Z T = L1

University of California, Davis

(10.24)

Version: March 21, 2012, 17:15

Lecture Notes

280

T If we recall that pivot dj = pj = zj Azj = Azj , zj , then by equating corresponding entries of AZD1 and

L = [lij ] we nd that (Benzi and Tma, 2003; Bollhfer and Saad, 2001) lij = Azj , zi Azj , zj ij (10.25)

Hence, the L factor of A can be obtained as a by-product of the A-orthogonal-ization, at no extra cost. In the implementation of SAINV, the quantities lij in Equation 10.25 are the multipliers that are used in updating the columns of Z . Once the update is computed, they are no longer needed and are discarded. To obtain an incomplete factorization of A, we do just the opposite; we save the multipliers lij , and discard the column vectors zj as soon as they have been computed and operated with. Hence, the incomplete L factor is computed by columns; these columns can be stored in place of the zj vectors, with minimal modications to the code. Here, we are assuming that the right-looking form of SAINV is being used. If the left-looking one is being used, then L would be computed by rows. Please refer to Benzi and Tma (2003) for more implementation details.

10.4.5

Numerical Experiments

Matrices from soil-structure interaction nite element analysis have been extracted from simulation system to study the performance of dierent preconditioning techniques on PCG method. The prototype of soil structure model has been shown in Figures 10.56 and 10.57. In order to introduce both of the nonlinear theories for soil and structures, we use continuum elements to model the soil and beam elements for the structures. Matrices from static pushover analysis and dynamic ground motion analysis have been collected for this research. Table 10.7: Matrices in FEM Models Continuum Model (Static) Matrix m1188 m1968 m11952 Property SPD SPD SPD Dimension 3336 5373 33081 # Nonzeros 220283 364877 2424543

Soil-Beam Model (Static and Dynamic) Matrix SoilBeam SoilBeamDyn Property SPD SPD Dimension 8442 8442 # Nonzeros 547678 547671

Jeremi et al.

Lecture Notes

281

Jeremi et al.

Lecture Notes

282

Jeremi et al.

Lecture Notes

283

Jeremi et al.

Lecture Notes

284

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

285

Jeremi et al.

Lecture Notes

286

SPD matrices have been studied using Conjugate Gradient method with or without preconditioning. Performance has been summarized in Table 10.8. Table 10.8: Performance of CG and PCG Method (Continuum FEM) 3336 DOFs FEM (Static) Preconditioner Jacobi IC(0) IC(1e-6) RIF2(1e-2) RIF3(1e-2) # Iter 4376 1612 413 5 571 541 Pre Time(s) 0.01 2.19 5.90 9.37 6.80 Iter Time(s) 54.82 20.18 11.07 0.47 14.94 14.13 Total Time(s) 54.82 20.19 13.26 6.37 24.31 20.93 Density2 1.00 5.88 0.94 0.94

5373 DOFs FEM (Static) Preconditioner # Iter 4941 1711 437 6 599 566 Pre Time(s) 0.01 6.5 19.81 25.71 21.31 Iter Time(s) 103.78 36.61 20.38 1.3 26.55 25.23 Total Time(s) 103.78 36.62 26.88 21.11 52.26 46.54 Density 1.00 8.10 0.96 0.96

3

Density 1.00

RIF2(1e-2)

2

694

1172.7

211.88

1384.58

0.99

RIF3(1e-2) 664 1245.4 202.67 1448.07 0.99 Density is dened as the number of non-zeros of the incomplete factor divided by the Could not continue because memory requirement larger than 1.4GB.

3

10.4.6

1. For the soil-structure interaction problems investigated in this section, Conjugate Gradient method works ne and the convergence is acceptable for most cases. 2. Incomplete Cholesky factorization preconditioner has been shown to be very powerful in static pushover

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

287

Jeremi et al.

Lecture Notes

288

8842 DOFs Soil-Beam FEM (Static) Preconditioner Jacobi IC(0) IC(1e-6) RIF2(1e-6)5 RIF3(1e-6)5 # Iter 3274 1687 26 6 23 31 Pre Time(s) 0.01 15.77 110.17 3364.8 34541 Iter Time(s) 102.5 54.56 1.95 2.79 3.44 9.26 Total Time(s) 102.5 54.57 17.72 112.96 3368.24 34550.26 Density4 1.00 15.11 4.32 16.37

8842 DOFs Soil-Beam FEM (Dynamic) Preconditioner Jacobi IC(0) IC(1e-6) RIF2(1e-2) RIF3(1e-2)

4 Density

Pre Time(s) -

Density -

is dened as the number of non-zeros of the incomplete factor divided by the number of nonwith tolerance 1e-2 failed to converge.

5 Iteration

Jeremi et al.

Lecture Notes

289

Jeremi et al.

Lecture Notes

290

problems. 3. Dynamic problems formulated by Newmark integration scheme have not been extensively tested. But according to the data available so far, neither IC nor RIF preconditioners performed well and further testing is necessary to reach a more persuasive conclusion. The diculty in dynamic analysis results from the fact that consistent mass and damping matrices used in continuum nite element formulations signicantly degrade the conditioning number of the nal system. This situation deteriorates when penalty handler is used to apply multiple point constraints, which introduces huge o-diagonal numbers to stiness, mass and damping matrices (Cook et al., 2002). 4. Robust incomplete factorization preconditioning based on A-orthogonalization has not been shown competitive with IC preconditioners in this research. It is also worth noting that all timings are taken in MATLAB. There are much more improvement can be achieved with a carefully coded FORTRAN program. 5. Static analysis has been extensively studied and it can be safely concluded that IC(0) and Jacobi preconditioners are good choices for the nonlinear soil-beam interaction simulations. 6. Dynamic analysis has also been studied but more work is needed to draw any detailed conclusion. Generally

speaking, one should be alert if iterative solver is to be used for dynamic analysis. This partially comes from the fact that mass and damping matrices undoubtedly alter the structures of the coecient matrix. This situation becomes more complicated if penalty handler is used to introduce o-diagonal numbers when handling multi-point constraints. So direct solver would be a more stable option for solving dynamic equations.

10.5

The motivation of this section is to introduce a robust and ecient parallel equation solver into our parallel nite element analysis framework. Aside from sparsity, which has been well known as the result of compact support that is inherent with nite element method, there exist some other special considerations that make the equation solving in nite element simulation a more involved problem. In nonlinear nite element simulations, handling of constraints signicantly aects the condition number of assembled equation systems. In SFSI simulations, multiple-point constraint is necessary to enforce the connection between soil and pile elements. In this research, penalty handler has been adopted to impose multiple point constraints on the assembled equation systems. Transformation and Lagrange multipliers are among those popular methods as well (Belytschko et al., 2000; Cook et al., 2002). The method of Lagrange multipliers adds extra constraints to the system and the resulted coecient matrix will lose symmetric positive deniteness. Transformation is favorable especially in the sense that it reduces the order of the equation systems by condensing out slave/constrained DOFs. But the transformation is the most dicult to code and the situation of one single master/retained node with multiple slave/constrained nodes further complicates the problem. Penalty method is chosen in this research due to the fact that it well preserves the symmetric positive deniteness

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

291

of the system if the nice property is observed. Another consideration comes from the easiness with which the penalty methods can handle the single master/retained multiple slave/constrained situations. This is proven to be extremely valuable when data redistribution is required in adaptive parallel processing because the DOF_Graph object can be clearly tracked during partition and repartition phases. The incapability of handling constraints accurately has been long known as the weakness of penalty method. The choice of the key penalty number seems arbitrary and largely depends on experience. The dilemma is with larger penalty number, the system can handle constraints more accurately while the coecient matrix can become very ill-conditioned. This can lead to serious convergence problem for iterative solvers. The majority of coecient matrices resulted from nite element analysis are inherently symmetric positive denite, for which lots of numerical algorithms have been proposed and solving SPD, symmetric or closely symmetric systems has been relatively maturer than more common unsymmetric cases. Unfortunately, in geotechnical nite element simulations, unassociated constitutive models lead to unsymmetric stiness matrices (Jeremi, 2004). More general parallel solvers must be coded to solve the problem. In this section, both iterative and direct solvers are coded using the consistent PETSc interface (Balay et al., 2001, 2004, 1997). Popular direct solvers for general unsymmetric systems such as MUMPS, SPOOLES, SuperLU,

PLAPACK have been introduced and performance study has been carried out to investigate the eciency of dierent solvers on large scale SFSI simulations with penalty-handled unsymmetric equation systems. GMRES is always the rst choice of iterative method when general unsymmetric systems are concerned. Preconditioning techniques have been thoroughly studied in this research to explore possible advantage of preconditioned iterative solver over direct solving. Jacobi, incomplete LU decomposition and approximate inverse preconditioners represent the most popular choices for Krylov methods and they are chosen in this performance survey. All numerical algorithms have been implemented through interface of PETSc, which provides a consistent platform on which implementation issues can be avoided to expose individual algorithmic performance.

10.5.1

The methods that we consider for the solution of sparse linear equations can be grouped into four main categories: general techniques, frontal methods, multifrontal approaches and supernodal algorithms (Dongarra et al., 1996).

10.5.1.1

The so-called general approach can be viewed as parallel versions of sparse LU decomposition. Special cares must be taken to handle the sparse data structures. Sparsity ordering is crucial in parallel sparse equation solving in order to reduce ll-in and discover large-grain parallelism (Demmel et al., 1993). Freely available package SPOOLES provides minimum degree (multiple external minimum degree (Liu, 1985)), generalized nested dissection and multisection ordering schemes for matrix sparsity ordering. Fundamental supernode tree built on top of vertex elimination tree is used to explore granularity in parallel (Ashcraft, 1999; Ashcraft et al., 1999).

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

292

10.5.1.2

Frontal methods have their origins in the solution of nite element problems from structural analysis. The usual way to describe the frontal method is to view its application to nite element problems where the matrix A is expected as a sum of contributions from the elements of a nite element discretization (Dongarra et al., 1996). That is,

m

A=

l=1

A[l] ,

(10.26)

where A[l] is nonzero only in those rows and columns that correspond to variables in the lth element. If aij and aij denote the (i, j )th entry of A and A[l] , respectively, the basic assembly operation when forming A is of the form aij aij + aij . It is evident that the basic operation in Gaussian elimination aij aij + aip [app ]1 apj . (10.28)

[l] [l]

(10.27)

may be performed as soon as all the the terms in the triple product 10.28 are fully summed (that is, are involved in no more sums of the form 10.27). The assembly and Gaussian elimination processes can therefore be interleaved and the matrix A is never assembled explicitly. This allows all intermediate working to be performed in a dense matrix, termed frontal matrix, whose rows and columns correspond to variables that have not yet been eliminated but occur in at least one of the elements that have been assembled. For general problems other than nite element, the rows of A (equations) are added into the frontal matrix one at a time. A variable is regarded as fully summed whenever the equation in which it last appears is assembled. The frontal matrix will, in this case, be rectangular. The idea of multifrontal method is to couple a sparsity ordering with the eciency of a frontal matrix kernel so allowing good exploitation of high performance computers. The basic approach is to develop separate fronts simultaneously which can be chosen using a sparsity preserving ordering such as minimum degree. Elimination tree, again is the most important notion in the factorization process and also utilized to discover the potential of parallelism. An elimination tree denes the a precedence order within the factorization. The factorization commences at the leaves of of the tree and data is passed towards the root along the edges in the tree. To complete the work associated with a node, all the data must have been obtained from the children of the node, otherwise work at dierent nodes is independent. Freely available package MUMPS (MUltifrontal Massively Parallel sparse direct Solver) has been used in this research to investigate the performance of multifrontal methods (http://graal.ens-lyon.fr/MUMPS/, 2006). MUMPS is a package for solving systems of linear equations of the form Ax = b, where A is a square sparse matrix that can be either unsymmetric, symmetric positive denite, or general symmetric. MUMPS uses a multifrontal technique which is a direct method based on either the LU or the LDLT factorization of the matrix. MUMPS exploits both parallelism arising from sparsity in the matrix A and from dense factorizations kernels.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

293

The main features of the MUMPS package include the solution of the transposed system, input of the matrix in assembled format (distributed or centralized) or elemental format, error analysis, iterative renement, scaling of the original matrix, and return of a Schur complement matrix. MUMPS oers several built-in ordering algorithms, a tight interface to some external ordering packages such as METIS and PORD, and the possibility for the user to input a given ordering. Finally, MUMPS is available in various arithmetics (real or complex, single or double precision). The software is written in Fortran 90 although a C interface is available. The parallel version of MUMPS requires MPI for message passing and makes use of the BLAS, BLACS, and ScaLAPACK libraries. The sequential version only relies on BLAS. MUMPS distributes the work tasks among the processors, but an identied processor (the host) is required to perform most of the analysis phase, to distribute the incoming matrix to the other processors (slaves) in the case where the matrix is centralized, and to collect the solution. The system Ax = b is solved in three main steps: 1. Analysis. The host performs an ordering based on the symmetrized pattern A + AT , and carries out symbolic factorization. A mapping of the multifrontal computational graph is then computed, and symbolic information is transferred from the host to the other processors. Using this information, the processors estimate the memory necessary for factorization and solution. 2. Factorization. The original matrix is rst distributed to processors that will participate in the numerical factorization. The numerical factorization on each frontal matrix is conducted by a master processor (determined by the analysis phase) and one or more slave processors (determined dynamically). Each processor allocates an array for contribution blocks and factors; the factors must be kept for the solution phase. 3. Solution. The right-hand side b is broadcast from the host to the other processors. These processors compute the solution x using the (distributed) factors computed during Step 2, and the solution is either assembled on the host or kept distributed on the processors. Each of these phases can be called separately and several instances of MUMPS can be handled simultaneously. MUMPS allows the host processor to participate in computations during the factorization and solve phases, just like any other processor. For both the symmetric and the unsymmetric algorithms used in the code, a fully asynchronous approach with dynamic scheduling of the computational tasks has been chosen. Asynchronous communication is used to enable overlapping between communication and computation. Dynamic scheduling was initially chosen to accommodate numerical pivoting in the factorization. The other important reason for this choice was that, with dynamic scheduling, the algorithm can adapt itself at execution time to remap work and data to more appropriate processors. In fact, the main features of static and dynamic approaches have been combined and the estimation obtained during the analysis to map some of the main computational tasks has been used; the other tasks are dynamically scheduled at execution time. The main data structures (the original matrix and the factors) are similarly partially mapped according to the analysis phase.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

294

10.5.1.3

The left-looking or column Cholesky algorithm can be implemented for sparse system and can be blocked by using a supernodal formulation. The idea of a supernode is to group together columns with the same nonzero structure, so they can be treated as a dense matrix for storage and computation. Supernodes were originally used for (symmetric) sparse Cholesky factorization (Demmel et al., 1999a). In the factorization A = LLT (or A = LDLT ), a supernode is a range (r : s) of columns of L with the same nonzero structure below the diagonal; that is, L(r : s; r : s) is full lower triangular and every row of L(r : s; r : s) is either full or zero. Then in left-looking Cholesky algorithm, all the updates from columns of a supernode are summed into a dense vector before the sparse update is performed. This reduces indirect addressing and allows the inner loops to be unrolled. In eect, a sequence of col-col updates is replaced by a supernode-column (sup-col) update. The sup-col update can be implemented using a call to a standard dense Level 2 BLAS matrix-vector multiplication kernel. This idea can be further extended to supernode-supernode (sup-sup) updates, which can be implemented using a Level 3 BLAS dense matrix-matrix kernel. This can reduce memory trac by an order of magnitude, because a supernode in the cache can participate in multiple column updates (Demmel et al., 1999a). It has been reported in (Ng and Peyton, 1993) that a sparse Cholesky algorithm based on sup-sup updates typically runs 2.5

to 4.5 times as fast as a col-col algorithm. Indeed, supernodes have become a standard tool in sparse Cholesky factorization. To sum up, supernodes as the source of updates help because of the following (Demmel et al., 1999a): 1. The inner loop (over rows) has no indirect addressing. (Sparse Level 1 BLAS is replaced by dense Level 1 BLAS.) 2. The outer loop (over columns in the supernode) can be unrolled to save memory references. (Level 1 BLAS is replaced by Level 2 BLAS.) Supernodes as the destination of updates help because of the following: 3. Elements of the source supernode can be reused in multiple columns of the destination supernode to reduce cache misses. (Level 2 BLAS is replaced by Level 3 BLAS.) Supernodes in sparse Cholesky can be determined during symbolic factorization, before the numeric factorization begins. However, in sparse LU, the nonzero structure cannot be predicted before numeric factorization, so supernodes must be dened dynamically. Furthermore, since the factors L and U are no longer transposes of each other, the denition of a supernode must be generalized. Freely available package SuperLU proposed a couple of ways to generalize the symmetric denition of supernodes to unsymmetric factorization (Demmel et al., 1999a). It is now not possible to use Level 3 BLAS eciently for unsymmetric systems. The implementation in SuperLU performs a dense matrix multiplication of a block of vectors and, although these can not be written as another dense matrix, it has been shown that this Level 2.5 BLAS has most of the performance characteristics of Level 3 BLAS since the repeated use of the same dense matrix allows good use of cache and memory hierarchy.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

295

There are three versions of libraries collectively referred as SuperLU (Demmel et al., 2003), Sequential SuperLU is designed for sequential processors with one or more layers of memory hierarchy (caches). Multithreaded SuperLU (SuperLU_MT) is designed for shared memory multiprocessors (SMPs), and can eectively use up to 16 or 32 parallel processors on suciently large matrices in order to speed up the computation (Demmel et al., 1999b). Distributed SuperLU SuperLU_DIST is designed for distributed memory parallel processors, using MPI for interprocess communication. It can eectively use hundreds of parallel processors on suciently large matrices (Li and Demmel, 2003). Parallelizing sparse direct solver for unsymmetric systems is more complicated than parallel sparse Cholesky case. The advantage of sparse Cholesky over the unsymmetric case is that pivots can be chosen in any order from the main diagonal while guaranteeing stability. This lets us perform pivot choice before numerical factorization begins, in order to minimize ll-in, maximize parallelism. precompute the nonzero structure of the Cholesky factor, and optimize the (2D) distributed data structures and communication pattern (Li and Demmel, 2003). In contrast, for unsymmetric or indenite systems, distributed memory codes can be much more complicated for at least two reasons. First and foremost, some kind of numerical pivoting is necessary for stability. Classical partial pivoting or the sparse variant of threshold pivoting typically cause the ll-ins and workload to be generated dynamically during factorization. Therefore, we must either design dynamic data structures and algorithms to accommodate these ll-ins, or else use static data structures which can grossly overestimate the true ll-in. The second complication is the need to handle two factored matrices L and U, which are structurally dierent yet closely related to each other in the lled pattern. Unlike the Cholesky factor whose minimum graph representation is a tree (elimination tree), the minimum graph representations of the L and U factors are directed acyclic graphs (elimination DAGs). In SuperLU_DIST, a static pivoting approach, called GESP (Gaussian Elimination with Static Pivoting) (Li and Demmel, 1998) is used. In order to parallelize the GESP algorithm, a 2D block-cyclic mapping of a sparse matrix to the processors is used. An ecient pipelined algorithm is also designed to perform parallel factorization. With GESP, the parallel algorithm and code are much simpler than dynamic pivoting. The main algorithmic features of SuperLU_DIST solver are summarized as follows (Li and Demmel, 2003): supernodal fan-out (right-looking) based on elimination DAGs, static pivoting with possible half-precision perturbations on the diagonal, use of an iterative algorithm using the LU factors as a preconditioner, in order to guarantee stability, static 2D irregular block-cyclic mapping using supernodal structure, and loosely synchronous scheduling with pipelining.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

296

In particular, static pivoting can be performed before numerical numerical factorization, allowing us to use all the techniques in good sparse Cholesky codes: choice of a (symmetric) permutation to minimize ll-in and maximize parallelism, precomputation of the ll pattern and optimization of 2D distributed data structures and communication patterns. Users are referred to Li and Demmel (2003) for algorithm details.

10.5.2

In this section, performance study on popular parallel direct and iterative solvers has been conducted. The purpose is to provide some guidelines on appropriate use of dierent solvers with the parallel nite simulation framework. Matrix systems from SFSI analysis are used as test cases. The performance investigation uses IA64 Intel-based cluster at SDSC.

10.5.2.1

Equation System

Jeremi et al.

Lecture Notes

297

10.5.2.2

Performance Results

Table 10.10: Performance Study on SFSI Systems (N =33081) Direct Solvers Solvers Num of CPUs 4 MUMPS 8 16 4 SuperLU_DIST 8 16 4 SPOOLES 8 16 Iterative Solvers (GMRES) Time (s) 6.0312 7.0534 5.3472 20.358 13.803 13.755 10.696 7.5338 6.2448

Preconditioner Jacobi

Num of CPUs 16 4

Parallel ILU(0)

8 16

10.5.3

Conclusion

This section presents the parallel solvers implemented in parallel nite element framework. Table 10.10, draws several conclusions about appropriate use of solvers: Direct solvers outperforms the iterative solver signicantly for general cases. It is worthwhile to note that nonsymmetric solvers are used here due to their generality. For special cases such as SPD system, preconditioned CG will show much better performance. The Conjugate Gradient method applies only to Symmetric Positive Denite (SPD) system. This puts restriction on the material models we can use in our simulations. Generally speaking, elastic material will yield a SPD stiness matrix. Plastic material with associative ow rule also satises this category. Plastic material with non-associative ow rule has non-symmetric element stiness matrix and so will be the global coecient matrix of the equation system. Another category of matrix that deserves attention is the stiness matrix from softening materials, which possesses at least one negative eigenvalue so the SPD property will be broken. For advanced geo-materials

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

298

subject to complicated loadings, as the material develops nonlinearity, the condition of stiness matrix might vary greatly from SPD (elastic phase), to singular (elastic-perfectly-plastic), and non-symmetric nonpositive-denite (elastic-non-associative-plastic-softening) cases. This poses another challenge when one tries to use iterative solver for production runs. The unpredictability of stiness matrix will disable the application of powerful solvers such as Conjugate Gradient for iterative case and Cholesky for direct case. The reason why iterative solver exhibits poor eciency is partly due to the problem size. We can also see from the Table 10.10 that parallel direct solver is not scalable in general. Iterative solver, on the other hand, is more scalable and it is safe to project that when the size of matrix increases, iterative solver has the advantage from the memory requirement point of view. Parallel equation solving itself is a complicated topic in numerical computing community. In this section, the main purpose is to introduce a robust and generally ecient parallel solver for nite element simulations. So in this sense, parallel direct solvers such as MUMPS and SPOOLES are recommended.

Jeremi et al.

Part II

299

Part III

301

Part IV

303

Chapter 11

(1999-2002-)

305

Lecture Notes

306

11.1 11.2

Chapter Summary and Highlights Numerical Analysis of Pile Behavior under Lateral Loads in Layered ElasticPlastic Soils

Material presented here has been previously published in our paper Yang and Jeremi (2005).

11.2.1

Introduction

The p y approach (Reese et al. Reese et al. (2000a)) has been widely used to design piles subjected to lateral loading. Based on the Winkler foundation theory, the method models the lateral soil-structure interaction with empirically derived nonlinear springs. The advancement of computer technology has made it possible to study this problem using more rigorous Finite Element Method (FEM). Mentioned are a few representative nite element applications. Maqtadir and Desai Muqtadir and Desai (1986)

studied the behavior of a pile-group using a three dimensional program with nonlinear elastic soil model. An axisymmetric model with elastic-perfectly plastic soil was used by Pressley and Poulos Pressley and Poulos (1986) to study group eects. Brown and Shie Brown and Shie (1990a) Brown and Shie (1990b) Brown and Shie (1991) and Trochanis Trochanis et al. (1991) conducted a series of 3D FEM studies on the behavior of single pile and pile group with elastic-plastic soil model. In particular, interface element was used to account for pile-soil separation and slippage. Moreover, Brown and Shie derived p y curves from FEM data, which provide some comparison of the FEM results with the empirical design procedures in use. A number of model tests of free- or xed-headed pile groups under lateral loading has been simulated by Kimura et al. Kimura et al. (1995) and Wakai et al. Wakai et al. (1999) using 3D elasto-plastic FEM. A good correlation between the experiments and the analysis has been observed in these studies. All these results demonstrated that FEM can capture the essential aspects of the nonlinear problem. It is noted that there is not much literature reporting on FEM studies of pile behavior under lateral loading in layered soil system. In addition to that, there is a very small number of studies on the eects of layering system on the commonly used p y curve approach. This paper describes four 3D nite element models of a laterally loaded pile embedded in uniform and layered soil proles with the dimensions and soil parameters similar to those used in the centrifuge study by McVay et al. McVay et al. (1998) and Zhang et al. Limin Zhang and Lai (1999). The bending moments derived by integrating vertical stresses from FEM are numerically dierentiated once and twice to compute the shear force and pressure diagrams, respectively. Particularly, p y curves are generated and cross compared to illustrate the eects of soft clay (sand) layer on the p y curves of the overlaid sand (soft clay) layer. The results from FEM are also compared with those from centrifuge test and LPILE. In addition, a limited parametric study of pressure redistribution is conducted by changing the undrained shear strength of the soft clay layer and the friction angle of the sand layer to further investigate the layering eects. An early version of OpenSees OpenSees Development Team (Open Source Project) (1998-2005) nite element program was used in presented computations. Developed models are now available within our new framework F E I. Soil modeling was performed using Template ElasticPlastic approach

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

307

11.2.2

Constitutive Models

Two simple models were used in this numerical study. Specically, clay was modeled by a simple von Mises material model which is completely dened with the undrained shear strength. Sand was simulated by a DruckerPrager material model with non-associated ow rule. The reason for using such simple models is that the experimental results used to compare our simulations against did specify only those two material properties for sands and clays. Figure 11.1 presents yield surfaces for both models. In both material models, the Youngs moduli vary with conning pressure, as shown in Eqn. (11.1). p pa

a

E = Eo

(11.1)

where Eo is Youngs Modulus at atmospheric pressure, p is the eective mean normal stresses, pa is the atmospheric

pressure, and a is constant for a given void ratio. In this work, 0.5 was used.

q Yield Surface e pl

a2

1 a1 1

Const

(a)

(b)

Figure 11.1: Elastic plastic models used in this study: (a) DruckerPrager model specied with friction angle and dilation angle, and (b) von Mises model specied with undrained shear strength Cu .

The following parameters were used for medium dense sand: friction angle of 37.1o , Shear modulus at a depth of 13.7 m of 8960 kPa (Eo = 17400 kPa), Poissons ratio of 0.35 and unit weight of 14.50 kN/m3 . These parameters were given by Zhang et al. Limin Zhang and Lai (1999). A dilation angle of 0o is used in this work (Brown and Shie Brown and Shie (1990a)). The undrained shear strength, Youngs modulus, Poissons ratio and unit weight of clay were chosen to be 21.7 kPa, 11000 kPa, 0.45, 13.7 kN/m3 , respectively. It should be noted that the above material models are available within the Template ElasticPlastic Material Modeling paradigm (Jeremi and Yang Jeremi and Yang (2002)). It should also be noted that the use of simple DruckerPrager model can over-predicted the friction angle to triaxial extension stress path. However this inuence is limited to the zone behind the pile, within the interface zone and thus this drawback of the DruckerPrager model was neglected.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

308

11.2.3

Simulation Results

Presented in this subsection are representative results related to the behavior of piles in uniform and layered soil systems. Presented results are compared with those from the centrifuge study (McVay et al. McVay et al. (1998)), and with results obtained using LPILE program (Reese et al. Reese et al. (2000a,b).

11.2.3.1

Pile Models

A number of static pushover tests for single pile models were simulated using uniform soil and layered soil setups. Figure 11.2 shows the model setups. There are four main setups. Two of these are dealing with uniform sand

1 0 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 1 0 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1

1.72

1.72

1.72

1.72

7.86

7.86

(a)

(b)

Figure 11.2: (a) Single pile models, dimensions and layers of case #1 and #2. (b) Single pile models, dimensions and layers of case #3 and #4.

and clay soils, while two others are featuring layered soils. In particular, the case # 1 is a uniform soft clay soil, case # 2 includes top and bottom layers of soft clay with an inbetween layer of medium dense sand. On the other hand, case # 3 features uniform medium dense sand soil, while case # 4 features top and bottom layers of medium dense sand with an inbetween layer of soft clay. Detailed layering setup is given in Figure 11.2. Figure 11.3 shows the nite element mesh for all four cases. Based on symmetry, only half of the model is meshed. Twenty node brick elements are used for both soil, pile and interface. It should be noted that these quadratic elements exhibit high accuracy even for high aspect ratios and can model accurately bending of solid piles with two layers of elements. During mesh design stage, a study was performed to decide on appropriate (balanced) mesh size. That study showed that a much larger mesh, with many more elements (with lower aspect ratios) would account for a fairly small change in results, so it was decided that the current mesh is sucient for our

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

309

analysis. The square pile, with a width of 0.429 m, consist of four elements (per cross subsection) with the elastic property of aluminum. The ne mesh in the upper part of the model is to provide data points for the computation of shear forces and p y curves of sucient reliability as well as for the investigation of layering eects. The sides and bottom of the model are xed with the exception of the symmetric boundary, which is only supported in Y direction. The interface layer between aluminum pile and surrounding soil is represented by one thin layer of elements. The purpose of this layer is to mimic the installation eects on piles (drilled or driven). It also serves a purpose of a simplied interface which allows for tension cut-o (gaping) and controlled, coupled horizontal and vertical stiness. All interface elements were simulated by DruckerPrager model with a friction angle of 25o , and a dilation angle of 0o .

Figure 11.3:

Mesh of single pile model, side view, top eight layers of nite elements are either clay or sand

(depending on the cases), middle eight layers of nite elements are sand or clay (again depending on the cases) and the bottom is all uniform clay or sand, interface zone around the aluminum pile is also present.

11.2.3.2

Plastic Zones

The static pushover test were conducted using load control at pile head. The nal plastic zones are depicted in Figures 11.4, 11.5. Plastic zones are actually presented by plastied Gauss points. In particular, Figure 11.4(a) shows developed plastic zones for the uniform clay soil (case # 1). It is interesting to note that the plastic zone propagates fairly deep while it does not extend far from the pile in clay. Moreover, compression side (right side)

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

310

features much larger plastic zone while the plastic zone for the extension side (left side) is conned to the interface layer and a few Gauss points outside the interface layer. The case with clay and sand layer inbetween is shown in Figure 11.4(b). The main dierence is that the plastic zone is even smaller than for uniform clay layer. It is worth mentioning that this case, which includes sand layer, is stier than the uniform clay case, thus displacements are smaller in clay and the plastic zone does not propagate as much as in uniform clay soil.

a)

b) Figure 11.4: The plastic zones for (a) case # 1, and (b) case # 2 at lateral loading of 400kN.

Figure 11.5(a)(b) shows plastic zones at the end of loading process for sand and sand and clay soils. In particular, Figure 11.5(a) shows the plastic zone for uniform sand. It is interesting to note that the plastic zone propagates toward the surface with the collapse mechanics similar to the active and passive failure. In this case of course the system is 3D and so the failure propagation angles do not match the active and passive failure angles, however the dierence between active and passive zones propagation angles is almost exactly /2. Figure 11.5(b) shows plastic zone for the case # 4 which includes a layer of clay between 1.72m and 3.44m (Z coordinate, origin is in the pile center at the ground surface) . It is noted that the plastic zone is deeper, but not as nicely dened as in the previous case.

11.2.3.3

p y Curves

Results from static pushover tests on piles were used to generate p y curves. The bending moments derived by integrating vertical stresses are numerically dierentiated once and twice to compute the shear force and pressure diagrams, respectively. Direct integration of shear stresses was also performed to check results and it was found that shear forces were within 5% accuracy. The combination of calculated pressures (p) and displacements obtained from the nite element solution, allowed for generation of p y curves at various depths along the pile.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

311

a)

b) Figure 11.5: The plastic zones of case 3 and 4 at lateral loading of 400kN.

In what follows, presented are generated p y curves for both uniform soils (sand and clay) as well as for layered systems. It is noted that the graphical presentation of results for bending moments, shear forces and lateral pressures (load) on a pile beam are shown with 10 lines, each one representing results for one increment (1/10) of the total load.

11.2.3.3.1

Figure 11.6 shows bending moments, shear forces and pressures along the depth of a pile in clay soil. It should be noted that the maximum bending moment, as well as the switching of sign for shear force, moves quite a bit from the depth of approximately 1.7m all the way to the depth of 3.4m. Pressure distribution shows that the top layers are already at the ultimate values of pressures and thus the pressure diagram propagates downward. There is a slight uctuation of pressures at the depths of 4 5m, which is attributed to the small numerical problems while doing double dierentiations. Figure 11.7 shows generated p y curves for uniform clay layer. It is obvious that most of the clay (at least until the depth of 2.6m) has reached its peak resistance.

11.2.3.3.2

Figure 11.8 shows bending moments, shear forces and pressures for a pile in a uniform sand soil. In this case it is interesting to note that the maximum bending moment, as well as the change of sign for the shear force is moving only between the depths 1.8m and 2.0m. Moreover, the pressure diagram shows steady increase (with

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

312

2

SOFT CLAY Cu = 21.7 kPa 3.436 SOFT CLAY Cu = 21.7 kPa

Depth (m)

10 1000

1000

2000

10 400 200

10

200

400

Pressure (kN/m)

Figure 11.6: Bending moment, shear force and pressure distributions for the uniform clay prole.

350 Depth 0.322 Depth 0.537 Depth 0.752 Depth 0.966 Depth 1.181 Depth 1.396 Depth 1.611 Depth 1.825 Depth 2.040 Depth 2.255 Depth 2.470 Depth 2.684

300

250

200

150

100

50

10

12

Jeremi et al.

Lecture Notes

313

top layers reaching ultimate pressures) until the depth of 1.7m and then steadily decreases, and changes sign at greater depths (below 4.0m).

2

SAND = 37.1 3.436 SAND = 37.1o

o

Depth (m)

10 1000

1000

2000

10 400 200

10

100

100

200

300

Pressure (kN/m)

Figure 11.8: Bending moment, shear force and pressure distributions for the uniform sand prole.

Figure 11.9 shows generated p y curves for the uniform sand case. It is interesting to note that only the top layer at the depth of about 0.3m will reach the ultimate pressure. All the other sand material is far away from corresponding ultimate pressures. It is also worth noting that the displacements in the case of uniform sand are much smaller (almost twice as small) than what has been observed in uniform clay case.

11.2.3.3.3

Figure 11.10 shows bending moments, shear forces and pressures for a layered soil case. In this case a layer of sand extends from 1.72m to 3.44m. The rest of soil is soft clay. It is interesting to note a large jump in pressures for the sand layer (as expected) and that the pressures in the top clay layer (from the surface to 1.7m) reaches ultimate values. Small nonuniform distribution of the pressures at the interface of sand and clay at 3.44m is attributed to the coarseness of the nite element mesh. In comparing Figure 11.10 with the results for uniform clay case (Figure 11.6) it is obvious that the sand layer arrests the propagation of deformation and forces in depth and xes the maximum moment to approx. 2.1m. Figure 11.11 shows generated p y curves for the layered case (single layer of sand in clay). The p y curves were generated only for the top layer of clay and middle layer of sand, to the depth of 2.7m. It is interesting to note that the p y curve for clay at the depth of 1.61m (close to the sand layer) exhibits strong hardening, unlike similar curve for the uniform clay soil, in Figure 11.7. The increase in pressure (transversal loading on the pile) between uniform clay (Fig. 11.7) and clay underlain by a medium dense sand layer (Fig. 11.11) at the

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

314

300 Depth 0.322 Depth 0.537 Depth 0.752 Depth 0.966 Depth 1.181 Depth 1.396 Depth 1.611 Depth 1.825 Depth 2.040 Depth 2.255 Depth 2.470 Depth 2.684

250

200

150

100

50

10

12

2

SAND = 37.1 3.436 SOFT CLAY Cu = 21.7 kPa

o

Depth (m)

10 1000

1000

2000

10 400 200

10

200

400

Pressure (kN/m)

Figure 11.10: Bending moment, shear force and pressure distributions for the clay soil with a sand layer.

Jeremi et al.

Lecture Notes

315

350 Depth 0.322 Depth 0.537 Depth 0.752 Depth 0.966 Depth 1.181 Depth 1.396 Depth 1.611 Depth 1.825 Depth 2.040 Depth 2.255 Depth 2.470 Depth 2.684

300

250

200

150

100

50

10

12

Figure 11.11: Calculated p y curves for the clay soil underlain by a medium dense sand layer.

11.2.3.3.4

Figure 11.12 shows bending moments, shear forces and transversal pressures for a case where a layer of soft clay is present within sand soil. Unlike the case of uniform sand soil (Figure 11.8) the presence of soft clay layer will change the depth of maximum moment by almost 1m (from 2.0m to 3.0m). In addition to that, the distribution of pressures on a pile is changed signicantly, as seen in the right plot of Figure 11.12. The reduction of pressures will extend into the sand layer and present signicant inuence of soft clay on pressures in sand. Figure 11.13 shows generated p y curves for the case of sand with a soft clay layer. It is noted that the p y curves for sand that is some distance away from the interface with clay are much the same as for the uniform sand case (refer to Fig. 11.9 and Fig. 11.19(a)). However, the p y curves in sand close to the interface are changed in some cases signicantly. For example, the p y curve at depth of 1.61m is showing pressure of approx. p = 265kN/m at the displacement of 0.042m for the uniform sand case, while the same p y curve, still in sand, has a drop in pressure at the same displacement to p = 140kN/m. Similar trend is observed for other p y curves close to the interface of sand with clay.

11.2.3.4

Comparison of pile behavior in uniform and layered soils can also be performed by looking at the displacement and bending moment distributions. For example, Figure 11.14 compares the distributions of displacements for

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

316

2

SOFT CLAY Cu = 21.7 kPa 3.436 SAND = 37.1o

Depth (m)

10 1000

1000

2000

10 400 200

10

100

100

200

300

Pressure (kN/m)

Figure 11.12: Bending moment, shear force and pressure distributions for the sand soil with a soft clay layer.

300 Depth 0.322 Depth 0.537 Depth 0.752 Depth 0.966 Depth 1.181 Depth 1.396 Depth 1.611 Depth 1.825 Depth 2.040 Depth 2.255 Depth 2.470 Depth 2.684

250

200

150

100

50

10

12

Figure 11.13: Calculated p y curves for the sand soil underlain by a soft clay layer.

Jeremi et al.

Lecture Notes

317

the uniform sand case with the sand and clay layer case. First observation is that the uniform sand layer allows smaller displacements of the pile head (0.12m) while the inclusion of clay layer raises those displacements to 0.22m. Second observation is that the point of rotation for the pile (point which does not move as the loading is applied) is pushed deeper, from 5m to approximately 6m. Moreover, the propagation of displacements along the depth of a pile is much greater for a layered case, the surface displacement is extended from 0.09m to almost 0.13m.

Depth (m)

Depth (m)

0 5 10 15 20 25

10 5

10 5

10

15

20

25

Figure 11.14: Pile displacement distributions along the depth in a uniform sand prole (left) and sand with clay layer prole (right).

Figure 11.15 shows similar results for uniform clay and clay with a layer of sand case. In this case, the inclusion of a sand layer will increase the stiness of the pile (as expected) and will also reduce propagation of displacements with depth. Figure 11.16 shows comparison of pile head displacements for all four cases. It is noted that the two layered cases exhibit similar behavior in terms of displacements, both at the pile head and in terms of displacement proles (compare right plot in Fig. 11.14 and left plot in Fig. 11.15). Figure 11.17 shows comparison of the maximum bending moment calculated for the pile for all four cases. It is interesting to note that the dierence between the two uniform soil cases (uniform sand and uniform clay) is not that pronounced. Of course one has to remember that the material for pile was assumed to be linear elastic, no yielding was allowed for the aluminum pile. The p y curves for uniform clay and clay with a layer of sand were plotted together in Figure 11.18 (a) for comparison. It can be seen that all the p y curves in clay except the one right next to the layer interface are almost identical. In order to measure the magnitude of the eects of sand layer on the pressure of soft clay layer,

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

318

Depth (m)

Depth (m)

10 5

10

15

20

25

30

35

10 5

10

15

20

25

30

35

Figure 11.15: Pile displacement distributions along the depth in a uniform clay prole (left) and clay with sand layer prole (right).

400

350

300

250

200

150

100

50

Sand Profile SandSoft Clay Profile Soft ClaySand Profile Soft Clay Profile 0 5 10 15 20 25 30 35

Displacement (cm)

Figure 11.16: Pile head displacement comparison for all the four cases.

Jeremi et al.

Lecture Notes

319

1800

1600

1400

1200

1000

800

600

400 Sand Profile SandSoft Clay Profile Soft ClaySand Profile Soft Clay Profile 0 5 10 15 20 25 30 35

200

Figure 11.17: Maximum bending moment comparison for all the four cases.

the ratio of pressures in clay layer for clay soils with a sand layer and uniform clay soils lateral displacement of 12%D, i.e. 5.15 cm, were computed and plotted against the distance in terms of times of pile width D in Figure 11.18. It is noted that the disturbance to the pressure eld is much more conned to the immediate vicinity (within 0.75D) of the layer interface. In addition, the results from two more analysis of the same model with dierent sands (friction angles = 25o , 30o respectively, other parameters remain the same.) were included in Figure 11.18. It is shown that the lateral pressure ratio is aected considerably when sand friction angle increases from 25o to 37o (from 1.5 times to 2.2 times more pressure). The p y curves for uniform sand and sand with a layer of soft clay were also plotted together for comparison purposes. It was found that the eect of soft clay on the pressures in sand propagates far away from the layer interface. Therefore, three cases of an additional model with a thicker sand layer (2.4m in thickness) underlain by a soft clay layer were analyzed by varying the undrained shear strength (Cu = 13.0 kP a, 21.7 kP a, 30.3 kP a) of the soft clay layer. Similarly, the pressure ratios at 6.5% D, i.e. 2.8 cm, were plotted in Figure 11.19. It is noted that the eects extends to as far as 4.75D from the layer interface and the reduction of pressures adjacent to the interface is about 0.6 in all three cases.

Jeremi et al.

Lecture Notes

320

300 Clay Profile: Depth 0.322 Clay Profile: Depth 0.537 Clay Profile: Depth 0.752 Clay Profile: Depth 0.966 Clay Profile: Depth 1.181 Clay Profile: Depth 1.396 Clay Profile: Depth 1.611 ClaySand Profile: Depth 0.322 ClaySand Profile: Depth 0.537 ClaySand Profile: Depth 0.752 ClaySand Profile: Depth 0.966 ClaySand Profile: Depth 1.181 ClaySand Profile: Depth 1.396 ClaySand Profile: Depth 1.611

250

200

150

100

50

10

12

a)

3.5

Sand: = 25o Sand: = 30o Sand: = 37o Deflection 0.120 xD 2.5

1.5

0.5

0 0.5

b)

Pressure Ratio P

1.5

claysand

/P

2.5

clay

Figure 11.18: (a) Comparison of p y curves for uniform clay versus clay with a layer of sand ( = 37o ). (b) Pressure ratio distributions in clay layer for sands with dierent friction angle ( = 25o , 30o , 37o ).

Jeremi et al.

Lecture Notes

321

300 Sand Profile: Depth 0.322 Sand Profile: Depth 0.537 Sand Profile: Depth 0.752 Sand Profile: Depth 0.966 Sand Profile: Depth 1.396 Sand Profile: Depth 1.825 Sand Profile: Depth 2.255 SandClay Profile: Depth 0.322 SandClay Profile: Depth 0.537 SandClay Profile: Depth 0.752 SandClay Profile: Depth 0.966 SandClay Profile: Depth 1.396 SandClay Profile: Depth 1.825 SandClay Profile: Depth 2.255

250

200

150

100

50

10

12

14

a)

5 Clay: Su = 13.0 kPa Clay: Su = 21.7 kPa Clay: Su = 30.3 kPa Deflection 0.065 xD 4

4.5

3.5

2.5

1.5

0.5

0 0.5

0.55

0.6

0.65

b) Figure 11.19:

0.7

0.75

0.8

0.85

0.9

0.95

(a) Comparison of p y curves for uniform sand versus sand with a layer of soft clay (Cu =

21.7 kP a). (b) Pressure ratio distributions in sand layer for clays with dierent undrained shear strength (Cu = 13.0 kP a, 21.7 kP a, 30.3 kP a).

Jeremi et al.

Lecture Notes

322

11.2.3.5

The pile head displacements for uniform sand prole from 3D FEM, LPILE (Reese et al. Reese et al. (2000a,b)), and centrifuge test (McVay et al. McVay et al. (1998)) were plotted against pile head load in Figure 11.20. It can be seen that they agree with each other fairly well. It should be noted that the material properties for our 3D nite element simulations were not in any particular way calibrated to improve the results. They were simply used as presented in the centrifuge study by McVay et al. McVay et al. (1998) and numerical simulation by Zhang et al. Limin Zhang and Lai (1999). Whereas, the results from LPILE were back-tted since the coecient of subgrade reaction h was back-calculated as 2714 kN/m3 (Zhang et al. Limin Zhang and Lai (1999)).

300

250

200

150

100

Displacement (cm)

The bending moments, shear forces and lateral pressures of uniform sand and clay proles from 3D FEM and LPILE were plotted against pile depth at several pile head loads in Figure 11.21 and 11.22. In general, there is a good agreement between the results from FEM and LPILE in uniform sand prole. In uniform soft clay prole, it is noted that the pressures at shallow depth from LPILE are smaller than those computed by FEM, which agrees with one of the ndings by the work of Steven and Audibert Stevens and Audibert (1979). For example, the pressures at lateral load of 120kN and 200kN from LPILE are only about half of those from FEM. Because the pressures at shallow depths are so small in LPILE that the pile head has to deform much more than in FEM and the passive pressure zone in LPILE extends to fairly large depth. Since LPILE currently uses the equivalent depth method developed by Geogiadis Georgiadis (1983) for layered soil proles, the LPILE output pressure distribution along pile depth, especially across the layer interface does not take into account of the layering eect, thus it is not that meaningful to compare pressure distributions of layered proles from LPILE versus FEM.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

323

FEM, Q= 40kN FEM, Q=160kN FEM, Q=280kN LPILE, Q= 40kN LPILE, Q=160kN LPILE, Q=280kN

2

SAND = 37.1 3.436 SAND = 37.1

o o

Depth (m)

10 1000

1000

10 400 200

200

400

10 100

100

200

300

Pressure (kN/m)

Figure 11.21: Comparison of bending moment, shear force and pressure computed by FEM and LPILE in uniform sand prole(case #3).

FEM, Q= 40kN FEM, Q=120kN FEM, Q=200kN LPILE, Q= 40kN LPILE, Q=120kN LPILE, Q=200kN

2

SOFT CLAY Cu = 21.7 kPa 3.436 SOFT CLAY Cu = 21.7 kPa

Depth (m)

10 1000

1000

10

200

200

10 100

100

200

Pressure (kN/m)

Figure 11.22: Comparison of bending moment, shear force and pressure computed by FEM and LPILE in uniform soft clay prole(case #1).

Jeremi et al.

Lecture Notes

324

It is also interesting to compare the p y curves derived from FEM with those used in LPILE. Figures 11.24 and 11.23 show FEM derived and LPILE used p y curves for uniform clay and sand proles, respectively. It should be noted that the coecient of subgrade reaction h was again back-calculated as 8969 kN/m3 in order to get a reasonable p y curves. From Figures 11.24 (a) and (b), it is clear that p y curves in sand prole from LPILE have lower resistance at depth close to ground surface. The p y curves for clay prole shown in Figures 11.24 (a) and (b) are seen to have much lower resistance at shallow depths.

400 Depth 0.322 Depth 0.537 Depth 0.752 Depth 0.966 Depth 1.181 Depth 1.396 Depth 1.611 Depth 1.825 Depth 2.255 Depth 2.684

350

300

250

200

150

100

50

10

a)

400

Depth 0.323 Depth 0.537 Depth 0.752 Depth 0.968 Depth 1.183 Depth 1.397 Depth 1.612 Depth 1.828 Depth 2.258 Depth 2.688

350

300

250

200

150

100

50

10

b)

Figure 11.23: p y curves from FEM (a) and LPILE (b) in uniform sand prole (h = 8969 kN/m3 , = 37.1o ).

Jeremi et al.

Lecture Notes

325

160

140

120

100

80

60 Depth 0.322 Depth 0.537 Depth 0.752 Depth 0.966 Depth 1.181 Depth 1.396 Depth 1.611 Depth 1.825 Depth 2.255 Depth 2.684 0 2 4 6 8 10 12 14 16 18 20

40

20

a)

160

Depth 0.322 Depth 0.537 Depth 0.752 Depth 0.966 Depth 1.181 Depth 1.396 Depth 1.611 Depth 1.825 Depth 2.255 Depth 2.684

140

120

100

80

60

40

20

10

12

14

16

18

20

b)

Figure 11.24: p y curves from FEM (a) and LPILE (b) in uniform clay prole (

50

Jeremi et al.

Lecture Notes

326

11.2.4

Summary

This paper presents results from a nite element study on the behavior of a single pile in elasticplastic soils. The analysis included single pile behavior in sand, clay and layered soils. Based on the results presented, it is concluded that three dimensional nite element analysis using very simple elastic-plastic soil models can predict the pile head deection with very good accuracy. The main ndings of this numerical study can be summarized as follows: When a sand layer is present within a clay deposit, the increase in lateral pressure in clay near the interface is conned to a narrow zone, up to two times of pile width, therefore the layering eect in this case is not prominent. When a clay layer is present within a sand deposit, the reduction in pressures spread well into the sand layer (up to four times of pile width). The layering eects are of more importance in this case since the disturbance zone is large and the pressure reduction is signicant. Reduction factors are given in terms of charts of pressure reduction versus the distance from the interface.

In addition, comparison with centrifuge data shows generally a good agreement between the bending moments, shear forces and lateral resistance. Moreover, a comparison with results from program LPILE, used in extensively in practice, show some discrepancies ultimate pressures in shallow soil layers.

11.3

Material presented here has been previously published in our paper Yang and Jeremi (2005).

11.3.1

Introduction

The theory of beams on a Winkler-type subgrade (Hartog (1952) ), also known as the py approach, has been widely used to design piles subjected to lateral loading. Based on that theory, the method models the lateral soilfoundation interaction with empirically derived nonlinear springs (py curves). The advancement of computer technology has made it possible to study this problem using more rigorous elasticplastic Finite Element Method (FEM). Here mentioned are a few representative examples of nite element studies of pile foundations. Muqtadir and Desai (1986) studied the behavior of a pilegroup using a three dimensional (3D) program with nonlinear elastic soil model. An axisymmetric model with elastic-perfectly plastic soil was used by Pressley and Poulos (1986) to study group eects. Brown and Shie (1990a), Brown and Shie (1990b), Brown and Shie (1991), and Trochanis et al. (1991) conducted a series of 3D FEM studies on the behavior of a single pile and a pile group with elastic-plastic soil model. These researchers used interface elements to account for pilesoil separation and slippage. Moreover, Brown and Shie derived py curves from FEM data, which provide some comparison of the FEM results with the

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

327

empirical design procedures in use. Kimura et al. (1995) conducted 3D FEM analysis of the ultimate behavior of laterally loaded pile groups in layered soil proles with the soil modeled by DruckerPrager model and pile modeled by nonlinear beam elements. A number of model tests of free or xedheaded pile groups under lateral loading in homogeneous soil proles have been simulated by Wakai et al. (1999) using 3D elasto-plastic FEM. Pan et al. (2002) studied the performance of single piles embedded in soft clay under lateral soil movements. A good correlation between the experiments and the analysis has been observed in these studies. All these results demonstrated that FEM can capture the essential aspects of the nonlinear problem. Information about the lateral behavior of piles in layered soil proles is very limited. Some analytical studies have been conducted by Davisson and Gill (1963) and Lee and Karunaratne (1987) to dene the inuence of pile length, the thickness of upper layer and the ratio of stiness ratio of adjacent layers on the pile response based on the assumption that the soil is elastic. Reese et al. (1981) conducted small scale laboratory tests on a 25 mm diameter pile and a eld test with 152 mm diameter pile in layered soils and found that there was a relatively good agreement between deections measured in the tests and deections computed using homogeneous py curves at small loads. Georgiadis (1983) proposed an approach which is currently used in the LPILE program (Reese et al. (2000a,b)). This method assumes the py curves of the rst layer are the same as those for homogeneous soils. The eects of upper layers on the py curves of the lower layers are accounted for by the equivalent depth of the overlying layers based on strength parameters. To the Authors knowledge, there is no literature reporting on FEM study of layering eects on the behavior of laterally loaded piles in layered proles. However, it is of great interest to investigate the layering eects since in practice, most of soil deposits are layered systems. In a predominantly clay site with a minor sand layer, the sand layer will still be counted on to provide most of the soil resistance. In this case, the layering eects (probably reduction of resistance in the sand layer) must be considered. Current practice is to make an educated guess to reduce the sand py curves to account for the soil layering eects (Lam and Law (1996)). Obviously, an educated guess might not result in optimal design. It is very important to nd out how these layers in the layered system aect each other in order to carry out a more accurate analysis of pile foundation and therefore provide a more eective way for the design of pile foundations in layered soil systems. This paper describes four 3D nite element models of a laterally loaded pile embedded in uniform and layered soil proles, with the dimensions and soil parameters similar to those used in the centrifuge studies by McVay et al. (1998) and Limin Zhang and Lai (1999). Visualization tool Joey3D (Yang (2002)) was used to compute the bending moment, shear force and lateral resistance diagrams along the pile. Model calibration, comparison of nite-element analysis results with those from centrifuge tests and the LPILE program, and comparison of nite-element generated py curves with traditional py curves are summarized in a separate paper (Yang and Jeremi (2003). In this paper, py curves from each model were cross compared to illustrate both the eects of an intermediate soft clay (or sand) layer on the py curves of the sand (or soft clay) layers and the eects of sand (or soft clay) layers on the intermediate soft clay (or sand) layer. In addition, a limited parametric study was conducted to further investigate the layering eects in terms of lateral resistance ratios. The OpenSees OpenSees Development Team (Open Source Project) (1998-2005) nite element framework was employed for all the computations. Soil modeling was performed using the Template ElastoPlastic Framework (Jeremi and Yang

Jeremi et al.

Lecture Notes

328

(2002)) and solid elements while the piles were modeled using linear elastic solid elements, all developed by the Authors.

11.3.2

Single pile nite element models with the dimensions similar to the prototype model described in the above centrifuge tests were developed and a number of static pushover tests were simulated with 3D FEM using uniform soil and layered soil cases. The models for all cases were illustrated in Figure 11.25 (a). There are four main

11 00 00 11 00 11 00 11 00 11 00 11 0.429m 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 Clay 11 00 11 00 11 00 Sand 11 00 11 00 11 00 00 11 Clay 11 00 11 00 Sand 11 00 11 00 11 00 11 00 11 00 11 00 11 00 Clay 11 00 11 00 00 11 Sand 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11

2.40m

Case 1&2: Case 3&4: Case 1&4: Case 2&3: Case 1&2: Case 3&4:

1.72m

1.72m

7.86m

Interface

a)

b)

Figure 11.25: (a) Single pile models, dimensions and layers for models #1, #2, #3 and #4, including pile-soil interfaces and (b) 3D mesh of the single pile model.

analysis models. Two of them are dealing with uniform sand and clay deposits, while the other two are featuring layered soil deposits. In particular, model # 1 has a uniform soft clay deposit, model # 2 includes top and bottom layers of soft clay with an interlayer of medium dense sand. Model # 3 features uniform medium dense sand deposit, while model # 4 features top and bottom layers of medium dense sand with an interlayer of soft clay. Figure 11.25 (b) shows the nite element mesh for all four models. Based on symmetry, only half of the model is meshed. Twentynode brick elements are used to mesh the soil, pile and pilesoil interface. The square pile, with a width of 0.429 m and length of 13.7 m1 , is divided into four elastic elements (per cross subsection) with the properties of aluminum. The mesh is rened at the upper part of the model in order to provide data points for the computation of shear force and lateral resistance of sucient reliability as well as for investigation of the layering eects. Additional nite element analysis of a cantilever beam using the same mesh as the pile was carried out and comparison of the beam displacement from FEM and beam theory solution indicated that the mesh was ne enough to capture the pile behavior. As to the boundaries, the sides and bottom of the model are xed with the exception of the symmetric boundary, which is only supported in Y direction. Since the sides are 13 times of the pile width away from the pile center, it is believed that the xed boundaries have very limited eects on the

1 All

Jeremi et al.

Lecture Notes

329

results. In addition to that the model size is closely following that of the physical, centrifuge model, which resided in a container of similar size. The pilesoil interface is represented by one thin layer of elements. The purpose of this layer is to mimic the installation eects on the pile (drilled or driven). It also serves a purpose of a simplied interface which allows for tension cut-o (gapping) and controlled, coupling of horizontal and vertical resistance according to Coulomb frictional laws.

11.3.3

Constitutive Models

Two simple models were used in this numerical study. Specically, clay was modeled by von Mises material model which is completely dened with the undrained shear strength. Sand was simulated by DruckerPrager material model with nonassociated ow rule, dened with the friction and dilation angles. The reason for using such simple models is that the experimental results used in comparison with simulations did specify only very limited number of material properties for sands. Furthermore, a small number of model parameters needed by simple models are convenient for parametric study. In both material models, the Youngs moduli vary with conning pressure, as shown in Eqn. (11.2) (cf. Janbu (1963), Duncan and Chang (1970)): p pa

a

E = Eo

(11.2)

where Eo is Youngs Modulus at atmospheric pressure, p is the eective mean normal stresses, pa is the atmospheric pressure, and a is constant for a given void ratio. In this work, 0.5 was used. The following parameters were used for medium dense sand: friction angle = 37.1o , Shear modulus G at a depth of 13.7 m = 8960 kPa (Eo = 17400 kPa), Poissons ratio = 0.35 and unit weight = 14.50 kN/m3 . These parameters were given by Limin Zhang and Lai (1999). A dilation angle of = 0o is used in this work (Brown and Shie (1990a)). The undrained shear strength, Youngs modulus, Poissons ratio and unit weight of clay were chosen to be Cu = 21.7 kPa, E0 = 11000 kPa, = 0.45, = 11.8 kN/m3 , respectively. The interface elements were simulated by DruckerPrager model with a friction angle = 25o , and a dilation angle = 0o . All material properties were summarized in Table 11.1.

Table 11.1: Material properties of sand, clay, pile and soilpile interface used in FEM analysis. Soil Medium dense sand Loose sand Clay Pile Soilpile interface Eo (kPa) 17400 16000 11000 69000000 Variable 0.35 0.35 0.47 0.33 Variable (kN/m3 ) 14.5 14.1 11.8 26.8 Variable (o ) 37.1 34.5 25 (o ) 0 0 0 Cu (kPa) 21.7

Jeremi et al.

Lecture Notes

330

11.3.4

This subsection presents representative results related to the behavior of piles in uniform and layered soil deposits. Specically the py response curves derived for 3D FEM results for homogeneous and layered soil deposits are compared with each other to investigate the layering eects.

11.3.4.1

The py curves of uniform clay deposit and clay deposit with a layer of sand were compared in Figure 11.26. It is clearly seen that the py curve (Z = 3.75D) close to the interface (Z = 4D) is signicantly dierent from that in uniform soil prole.

300 Z/D= 0.75 300 Z/D= 1.75 200 100 0 300 Z/D= 3.75 Z/D= 4.76 200 100 0 300 Z/D= 6.76 Z/D= 8.26 200 100 0 Uniform Clay ClaySandClay 200 100 0 200 100 0 300 Z/D= 5.76 300 Z/D= 2.75

p (kN/m)

10

10

10

p (kN/m)

10

10

10

y (cm)

p (kN/m)

200 100 0

10

10

y (cm)

y (cm)

Figure 11.26: Comparison of py curves of uniform clay deposit versus clay deposit with an interlayer of sand (Sand: = 37o ; Clay: Cu = 21.7 kPa).

In order to measure the magnitude of the eects of the intermediate sand layer on the lateral resistance of the soft clay layers and vice versa, the ratios of soil lateral resistances in the layered (p) and uniform models (phomog.

model )

at several lateral displacements (i.e. 0.5%, 1.0%, 2.0%, 2.5%, 8.0% and 10.0% of pile width D) were computed and plotted against vertical coordinate (Z) normalized by pile width D in Figures 11.27 and 11.282 . In addition,

2 The

lateral resistance ratio is only shown for the upper clay layer since the resistance corresponding to large y is not available at

larger depth due to the fact that the pile is loaded at the pile head and the deection decreases quickly as depth increases. Also due

Jeremi et al.

Lecture Notes

331

the results from two more analyses of the same model with dierent sands (friction angle were varied from 25o to 30o , while originally, the friction angle was set to 37o ) were also included in these gures. From Figure 11.27, it is observed that the lateral resistance ratios are independent of friction angle of sand at small lateral displacements ranging from 0.5% to 1.0% of pile width D. When the lateral displacement is greater than 1.0%, the variation in starts to aect the lateral resistance ratio, as shown in Figure 11.28.

0 0 1 =25o,Eo=17400kPa =30o,Eo=17400kPa =37o,Eo=17400kPa 1 =25o,Eo=17400kPa =30o,Eo=17400kPa =37o,Eo=17400kPa

CLAY =11.8kN/m3

CLAY =11.8kN/m3

Z/D

5 SAND =14.5kN/m3

5 SAND =14.5kN/m3

CLAY =11.8kN/m3

CLAY =11.8kN/m3

y/D = 0.005

1 1.5 9 0.5

y/D = 0.010

1 1.5

9 0.5

p/phomog. case

p/phomog. case

Figure 11.27:

Lateral resistance ratio distributions (Clay:Cu = 21.7 kP a, Eo = 11000 kP a) for sands with

Overall, the eect of the sand layer reduces to less than 10% at about one pile width above the upper sand interface and the lateral resistance ratio at a quarter pile width above the upper sand interface is 1.3 at lateral deection of 0.5% pile width. It may be noted that the two dashed vertical lines in the lateral resistance plots correspond to lateral resistance ratios of 0.9 and 1.1, indicating 10% change in lateral resistance. The 10% change will be used to judge the extent of inuence throughout the rest of the paper. The resistance ratio below the lower sand interface was not processed since the mesh is becoming coarse and the results are aected by mesh eects and numerical dierentiation, and the pile displacements are very small. Besides the eect the sand interlayer has on the clay layer, it is interesting to observe in Figure 11.27 that the soft clay layers also have signicant eect on the lateral resistance of the intermediate sand layer. The lateral resistance ratios are less than 0.9 throughout the interlayer of sand. Surprisingly, the eects are not symmetric at lateral deection of 0.5%. The resistance ratio is 0.85 at 0.25D below the upper sand interface, while that is 0.72 at 0.25D above the lower sand interface. This nonsymmetry is probably due to the nonsymmetric deformation3 mode in the pile. As the pile is loaded laterally at the pile head, the righthandside sand close to the pile below certain depth tends to move downward to the right, which can be observed in Figure 11.32 (b). Therefore, the sand close to the upper interface moves against sand, while that close to the lower interface moves against soft

to the limit of space, plots for 2.0%, 2.5% pile width are not shown in this paper. 3 Nonsymmetric with respect to the horizontal plane in between the interfaces (midway through the sand layer).

Jeremi et al.

Lecture Notes

332

0.5

0.5

y/D = 0.080

y/D = 0.100

1.5

1.5

Z/D

2.5 3

3.5

3.5

3

p/phomog. case

p/phomog. case

2.5

Figure 11.28: Lateral resistance ratio distributions in clay layer (Cu = 21.7 kP a, Eo = 11000 kP a) for sands

clay. This type of movement results in the larger reduction in resistance at the lower sand interface than at the upper sand interface. The decrease in lateral resistance is mainly due to the lower stiness in the adjacent soft clay layers. In addition, the smaller unit weight of the soft clay results in smaller mean eective normal stresses in the sand layer than the homogeneous model, which will reduce the stiness of the sand and therefore also contribute to the reduction in lateral resistance at the intermediate sand layer.

11.3.4.2

Uniform Sand Deposit and Sand Deposit with an Interlayer of Soft Clay.

By comparing the py curves of uniform sand deposit and sand deposit with an interlayer of soft clay, it was found that the eect of soft clay on the lateral resistance of sand propagates further away from the interface than Clay-Sand-Clay case, as described above in subsection 11.3.4.1. In addition to that, it was found that the heave in front of the pile will aect the lateral resistance of sand at shallow depth. Therefore, for sand deposit with an interlayer of soft clay, the thickness of upper sand layer was increased from 1.72 m to 2.36 m (the thickness of the soft clay layer was kept the same) to investigate the range of layering eects. Three models were analyzed by only varying the undrained shear strength Cu ( i.e. 13.0, 21.7 and 30.3 kPa ) of the soft clay layer. Similar to the previous analysis, the py curves from the uniform deposit and the re-congured layered deposit were compared in Figure 11.29 and the lateral resistance ratios at several lateral displacements (i.e. 0.5%, 1.0%, 2.0%, 2.5%, 5.0% and 6.5% of pile width D ) for all three models were computed and shown in Figures 11.30 and 11.31. It may be observed from Figure 11.29 that obvious dierence may be observed in several p y curves further away from the interface. From Figure 11.30, it is noted that the eects of the intermediate soft clay layer are also independent of its

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

333

300 Z/D= 1.75 200 100 0 300 Z/D= 3.75 Z/D= 4.76 200 100 0 300 Z/D= 6.76 Z/D= 8.26 200 100 0

300 Z/D= 2.75 200 100 0 300 Z/D= 5.76 200 100 0

p (kN/m)

10

10

10

p (kN/m)

10

10

10

y (cm)

p (kN/m)

200 100 0

10

10

y (cm)

y (cm)

Figure 11.29: Comparison of py curves for uniform sand deposit versus sand deposit with an interlayer of soft clay (Sand: = 37o ; Clay: Cu = 21.7 kP a).

SAND =14.5kN/m3

SAND =14.5kN/m3

Z/D

5

6 CLAY =11.8kN/m3

6 CLAY =11.8kN/m3

y/D = 0.005

9 0.5

y/D = 0.010

1.5 9 0.5

p/p

homog. case

p/p

1.5

homog. case

Figure 11.30: Lateral resistance ratio distributions (Sand: = 37o , Eo = 17400 kP a) for clays with various Cu at lateral displacements of 0.5% and 1.0% pile width.

Jeremi et al.

Lecture Notes

334

y/D = 0.050

y/D = 0.065

Z/D

3 SAND 3 =14.5kN/m

3 SAND 3 =14.5kN/m

p / phomog. case

0.7

0.8

p / phomog. case

0.7

0.8

0.9

Figure 11.31: Lateral resistance ratio distributions (Sand: = 37o , Eo = 17400 kP a) for clays with various Cu

undrained shear strength at small lateral displacements ranging from 0.5% D to 1.0% D. When the lateral displacement is greater than 1.0% D, the change in Cu starts to aect the lateral resistance ratio, as shown in Figure 11.31. Similar to the ClaySandClay model, the eect of the intermediate soft clay layer reduces to less than 10% at one pile width above the clay interface. The lateral resistance ratio at 0.25D above the clay interface is about 0.75. For large lateral displacements ranging from 5.0% D to 6.5% D, the 10% change in lateral resistance extends to 1.5 D - 2 D, as can be observed in Fig. 11.31. It may be noted that, at a lateral displacement of 6.5% D, the lateral resistance ratio at 0.25D above the clay interface changes from 0.58 to 0.67 when Cu increases from 13.0 kPa to 30.3 kPa. Figures 11.32 (a) and (b) show the details of displaced models around the interfaces for the Sand-Clay-Sand and Clay-Sand-Clay proles, respectively. The deformed model was overlapped with undeformed model for comparison. Ground heave can be easily observed in front of the pile from both gures. It is noted from Figure 11.32 (a) that the sand crosses the upper clay interface and moves into the intermediate soft clay layer. The movement slightly strengthens the soft clay soil and partially causes the slight increase of lateral resistance at the top of soft clay layer. Most importantly, the movement will soften the sand close to the upper layer interface, due to the reduction of connement to the sand. For the Clay-Sand-Clay prole, the stronger sand layer penetrates into the softer clay layers at both interfaces. This penetration softens the sand close to both interface, due to the same reason as above.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

335

(a)

(b)

Figure 11.32: Details of displaced model indicating ground heave and movement of soils across the layer interfaces at lateral load of 400 kN: (a) sand deposit with an interlayer of soft clay and (b) clay deposit with an interlayer of medium sand. The pile elements are removed so that the interface layer in the middle can be seen clearly.

11.3.5

Parametric Study for the Lateral Resistance Ratios in Terms of Stiness and Strength Parameters.

To further investigate the eects of soil stiness on the lateral resistance ratios at small displacement and/or large displacement, further analyses were carried out for the ClaySandClay and SandClaySand models by changing both stiness parameter (i.e. Eo ) and strength parameter ( Cu for clay, or for sand) using the same nite element models as above. The model congurations and intermediate layer soil parameters were summarized in Tables 11.2 and 11.3.

Table 11.2: Summary of model congurations and intermediate sand layer parameters for ClaySandClay model in the parametric study. Case 1 2 3 Soil Prole ClaySandClay ClaySandClay ClaySandClay Depth of Interfaces Upper -1.72 m -1.72 m -1.72 m Lower -3.43 m -3.43 m -3.43 m Intermediate Sand Layer Eo (kPa) 11500 13500 17400 25o 30o 37o

Jeremi et al.

Lecture Notes

336

Table 11.3: Summary of model congurations and intermediate clay layer parameters for SandClaySand model in the parametric study. Case 1 2 3 Soil Prole SandClaySand SandClaySand SandClaySand Depth of Interfaces Upper -2.36 m -2.36 m -2.36 m Lower -4.08 m -4.08 m -4.08 m Intermediate Clay Layer Eo (kPa) 8000 11000 12500 Cu (kPa) 13.0 21.7 30.3

Lateral resistance ratios were plotted in Figures 11.33 and 11.34 for the ClaySandClay model, and in Figures 11.35 and 11.36 for the SandClaySand model. By comparing Figures 11.28 and 11.34 for pile displacements of 8% D and 10% D, and Figures 11.31 and 11.36 for pile displacements of 5% D and 6.5% D, it is clear that the lateral resistance ratios are almost the same for the upper layer soil even if the stiness parameter Eo of intermediate layer soil was varied by more than 30%. However, the lateral resistance ratios at small displacement (0.5% D and 1.0% D) were obviously inuenced by the variation of Eo , as can be observed by comparing Figures 11.27 and 11.33 for the ClaySandClay model, and Figures 11.30 and 11.35 for the SandClaySand model. For medium pile displacements (e.g. 2% D and 2.5% D), both stiness and strength parameters have eects on the lateral resistance ratios.

0 =25o,Eo=11500kPa =30o,Eo=13500kPa =37o,Eo=17400kPa 0 =25o,Eo=11500kPa =30o,Eo=13500kPa =37o,Eo=17400kPa

CLAY =11.8kN/m3

CLAY =11.8kN/m3

Z/D

5 SAND =14.5kN/m3

5 SAND =14.5kN/m3

CLAY =11.8kN/m3

CLAY =11.8kN/m3

y/D = 0.005

1 1.5 9 0.5

y/D = 0.010

1 1.5

9 0.5

p/p

homog. case

p/p

homog. case

Figure 11.33:

Lateral resistance ratio distributions (Clay:Cu = 21.7 kP a, Eo = 11000 kP a) for sands with

It will be useful to relate the eects of (a) the relative stiness which controls the lateral resistance ratio at small lateral displacements and (b) the relative strength which determines the lateral resistance ratio at large lateral displacements with the lateral resistance ratio. To exclude the eects of unit weight, only the results above the

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

337

0.5

0.5

y/D = 0.080

y/D = 0.100

1.5

1.5

Z/D

2.5 3

3.5

3.5

3

p/phomog. case

p/phomog. case

2.5

Figure 11.34: Lateral resistance ratio distributions in clay layer (Cu = 21.7 kP a, Eo = 11000 kP a) for sands with various and Eo at lateral displacements of 8% and 10% pile width.

SAND =14.5kN/m

3

SAND =14.5kN/m3

Z/D

5

6 CLAY =11.8kN/m3

6 CLAY =11.8kN/m3

y/D = 0.005

9 0.5

y/D = 0.010

1.5 9 0.5

p/p

homog. case

p/p

1.5

homog. case

Figure 11.35: Lateral resistance ratio distributions (Sand: = 37o , Eo = 17400 kP a) for intermediate layer of clays with various Cu and Eo at lateral displacements of 0.5% and 1.0% pile width.

Jeremi et al.

Lecture Notes

338

y/D = 0.050

y/D = 0.065

Z/D

3 SAND =14.5kN/m3

3 SAND =14.5kN/m3

p / phomog. case

0.7

0.8

p / phomog. case

0.7

0.8

0.9

Figure 11.36: Lateral resistance ratio distributions (Sand: = 37o , Eo = 17400 kP a) for clays with various Cu

upper interface in the SandClaySand model are processed. The ratio of Youngs moduli of clay and sand soils was used to dene the relative stiness Rstif f ness of the two layers. On the other hand, the ratio of largest lateral resistances of uniform clay and sand4 at the upper interface (-2.36 m) was used to dene the relative strength RstrengthF EM , as described in Equations (11.3) and (11.4). Rstif f ness = Eoclay Eosand (11.3)

RstrengthF EM =

pclayF EM psandF EM

(11.4)

The lateral resistance ratios at lateral displacement of 6.5% D were plotted against Cu in Figure 11.37. For comparison, the relative stiness Rstif f ness and relative strength RstrengthF EM were also included in the same plot. As can be observed from this plot, the lateral resistance ratio decreases from 0.69 to 0.56 almost proportionally as Cu drops from 30 kPa to 13 kPa at 0.25 D above the upper interface, and the ratio is greater than the relative strength RstrengthF EM . Since the ultimate resistance of uniform sand will be larger than the computed largest value (which is still increasing , as can be observed from Figure 11.29 at Z=-3.75D) and that of uniform clay almost will almost remain the same (refer to Figure 11.26 at Z=-3.75D), this relative strength value will drop and the above statement still holds. There is certain correlation between the lateral resistance ratio close to the upper interface and RstrengthF EM at 6.5%D pile displacement. As the distance to the upper interface increases,

4 It

would be better to use the ultimate lateral resistances for both clay and sand to dene the relative strength but these values

are not available from the current numerical results since pile displacement y is not large enough.

Jeremi et al.

Lecture Notes

339

0.8

0.6

homog. case

0.4

Rstiffness RstrengthFEM Distance = 0.25D Distance = 1.25D Distance = 2.25D Distance = 3.25D Distance = 4.25D

0.2 12

14

16

18

20

22

24

26

28

30

32

Figure 11.37: Lateral resistance ratios in the upper sand layer ( = 37o ) at various distances from the interface

this correlation diminishes. The relative stiness curve intercepts with the lateral resistance ratio curves at 0.25D above the upper interface. This implies that the presence of the clay, which is softer than the sand, somehow caused the layered system to be softer than either of the homogeneous models. This seems illogical, and in fact previous discussions and comparisons showed that RstrengthF EM is more important than Rstif f ness at these large relative displacements. It is also interesting to examine the relationship between the lateral resistance ratio and the relative variables (i.e. strength and stiness) when lateral displacement increases, as presented in Figures 11.38 and 11.39. Figure 11.38 shows that the lateral resistance ratios at 0.25D above the interface decreases and come closer to the relative strength RstrengthF EM curve as the lateral displacement increases from 4.0%D to 6.5%D. The relative stiness Rstif f ness was also plotted in Figure 11.38 and it intercepts with the lateral resistance ratio curve, which has similar implications as the above discussion for Figure 11.37 and is illogical. On the other hand, as the lateral displacement decreases from 1.5%D to 0.5% D, the lateral resistance ratios keep decreasing and come closer to the relative stiness ratio Rstif f ness , as shown in Figure 11.39. There is almost a linear relationship between the lateral resistance ratio and the relative stiness at small displacements. From the above analysis, it is safe to say that the lateral resistance ratio is dominated by the relative stiness Rstif f ness at small displacement (i.e. 0.5%D), while that is controlled by the relative strength RstrengthF EM at large displacement ( i.e. 4.0%D). For small displacement, the smaller the displacement is, the closer the lateral resistance ratio is to the relative stiness; for large displacement, the larger the displacement, the closer the lateral resistance ratio is to the relative strength. Figures 11.40, 11.41 and 11.42 summarize observed lateral resistance ratios in layered proles. Figure 11.40 shows

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

340

0.8

0.75

0.7

0.65

0.6

homog. case

0.55

0.5

0.45

0.4

0.35 12

14

16

18

20

22

24

26

28

30

32

Figure 11.38: Lateral resistance ratio at a quarter pile width above the upper clay interface for various deections.

1.1

0.9

0.8

homog. case

0.7

0.6

0.5

0.4 12

14

16

18

20

22

24

26

28

30

32

Figure 11.39:

Lateral resistance ratio at one quarter pile width above the upper clay interface for clays with

various Cu and Eo at small pile displacements ranging from 0.5% to 1.5% pile width.

Jeremi et al.

Lecture Notes

341

the lateral resistance ratios in the intermediate sand layer corresponding to various relative stiness Rstif f ness at pile displacement of 0.5%D for the ClaySandClay model. Figures 11.41 and 11.42 show the lateral resistance ratios corresponding to various relative stiness Rstif f ness and relative strength RstrengthF EM at pile displacements of 0.5% D and 6.5% D for the SandClaySand model. The eects of the intermediate clay layer on the upper sand layer reduce to less than 10% at a distance of 0.5 to 1.5 D above the interface at small pile displacement (e.g. 0.5% D), while that eects reduce to less than 10% at a distance of 1.25 to 2.0 D above the interface at large pile displacement (e.g. 6.5%D). One may notice that the lateral resistance ratios corresponding to the relative stiness Rstif f ness = 0.63 in Figures 11.40 and 11.41 are not the same. The ratios close to the lower sand interface in the ClaySandClay model is slightly larger than that in the SandClaySand model. This dierence is due to the fact that the lateral resistance ratios in the intermediate sand layer also include the eects of smaller unit weight of upper layer clay.

stiffness

y/D = 0.005

3 CLAY 4 =11.8kN/m3

Z/D

5 SAND =14.5kN/m3 6 7 8

CLAY =11.8kN/m3

9 0.5

0.6

0.7

0.8

0.9

p/phomog. case

1.1

1.2

1.3

1.4

1.5

Figure 11.40: Summary of observed lateral resistance ratios from FEM analysis for the ClaySandClay prole at small deection (y/D=0.5%).

11.3.6

Summary

This subsection summarizes results from nite element analysis on the behavior of a single pile in elasticplastic layered soils. Based on the results presented, the following conclusions can be drawn. 1. The layering eects are twoway. Not only the lower layers are aected by the upper layers, but the upper layers are also aected by the lower layers. Furthermore, the layering eects are not symmetric. In the case of pile laterally loaded at the pile head, the eect of an interface extends further into the layer above the

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

342

stiffness

y/D=0.005

SAND =14.5kN/m3

Z/D

6 CLAY =11.8kN/m3

9 0.5

p / phomog. case

1.1

1.2

1.3

1.4

1.5

Figure 11.41: Summary of observed lateral resistance ratios from FEM analysis for the SandClaySand prole at small deection (y/D=0.5%).

y/D=0.065

SAND =14.5kN/m3

6 CLAY =11.8kN/m3

9 0.5

p / phomog. case

1.1

1.2

1.3

1.4

1.5

Figure 11.42: Summary of observed lateral resistance ratios from FEM analysis for the SandClaySand prole at large deection (y/D=6.5%).

Jeremi et al.

Lecture Notes

343

interface than it does into the layer below the interface at small displacements. 2. In the ClaySandClay model, the lateral resistance of soft clay increases by as much as 30% and the eect extends to one pile width above the upper sand interface for Rstif f ness = 0.63 at small pile displacement (0.5%D). Nonetheless, the increase of lateral resistance in the upper clay layer at large pile displacement (810%D) extends only one nite element above the upper sand interface. On the other hand, the clay layers also have signicant eects on the lateral resistance of sand throughout the intermediate layer. 3. In the SandClaySand model, the intermediate clay layer has considerable eects on the lateral resistance of the upper sand layer, and the sand layers also have signicant eects on the lateral resistance of the intermediate clay layer, causing 10 to 40% increase in its lateral resistance. 4. The lateral resistance ratio is dominated by the relative stiness at small displacements (i.e. 1.0%D), while that is controlled by the relative strength at large displacements (i.e. 5.0%D). It must be pointed out that the above observed lateral resistance ratios may only be applied to similar stratigraphies, pile deformation modes, and other conditions considered in this work. Further analyses are needed to investigate the eects of other stratigraphies, pile deformation modes, pile diameters, and other factors, in order

to draw more general guidelines. Future studies with a rened mesh around the interface will provide better resolution of the resistance ratio around the interface. Future studies of the eects of the interface layer on the layering eects will also be very interesting.

11.4

Material presented here has been previously published in our paper Yang and Jeremi (2005).

11.4.1

Introduction

Single pile foundations have been successfully modeled by the p y approach, as implemented in LPILE (Reese et al. Reese et al. (2000a) ). However, the behavior of a pile within a group may dier greatly from that of a single pile and vary depending on the position due to the interaction between the neighboring piles. To study this interaction eects, only a couple of eld tests have been carried out because of the large costs incurred. Brown et al. Brown et al. (1988) conducted cyclic loading tests on instrumented 3 3 steel pile group. The pmultiplier concept was presented based on the measured soil resistance data and specic pmultipliers were suggested for the three rows. Ruesta and Townsend Ruesta and Townsend (1997) reported an in-situ test on piles at Roosevelt Bridge. Rollins et al. Rollins et al. (1997) tested another fullscale pile group founded in clay and suggested a set of pmultipliers for corresponding pile groups. Ng et al. Ng et al. (2001) presented results on fullscale lateral load tests of one single pile and three pile groups with largediameter bored piles. Besides in-situ testing, many centrifuge tests were conducted to predict the behaviors of pile groups under static and dynamic loading. Recently, McVay et al. McVay et al. (1995) McVay et al. (1998) conducted a series of lateral load tests on large pile groups (3 3 to 7 3) founded in sands to study the interaction eects within a group.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

344

Based on these eld and centrifuge tests, it was found that: 1. All the test results have clearly shown that the lateral resistance of a pile within the group is strongly inuenced by its row position and the pmultiplier method was suggested by Brown et al. Brown et al. (1988) to account for this behavior. Specically, each row within the group is assigned a dierent p multiplier fm and the py curve for a single pile is multiplied by fm to produce py curves for all the piles in the same row. 2. The pmultipliers are independent of soil density and only depend on the pile geometry. And at sucient deformation and under static loading, they are constant for practical purposes (McVay et al. McVay et al. (1998)). 3. For 3diameter spacing, the suggested pmultiplier fm was 0.8, 0.4, 0.3 (front row to back row) for the 3 3 group (Brown et al. Brown et al. (1988)), 0.8, 0.4, 0.3, 0.3 for the 4 3 group, and 0.8, 0.4, 0.3, 0.2, 0.3 for the 5 3 group, and 0.8, 0.4, 0.3, 0.2, ..., 0.3 for all larger group size (McVay et al. McVay et al. (1998)). 4. For concentric loading (located at the geometrical center of the pile group), the dierence between the side

and middle piles within a row is small and may be neglected, while the moments in the side piles within a given row are slightly larger than that in the middle piles but may be represented by the average (McVay et al. McVay et al. (1998)). Together with the physical modeling, a few numerical simulations have also been performed. We mention a few representative nite element studies of pile groups. Maqtadir and Desai Muqtadir and Desai (1986) studied the behavior of a pile-group using a three dimensional program with nonlinear elastic soil model. An axisymmetric model with elastic-perfectly plastic soil was used by Pressley and Poulos Pressley and Poulos (1986) to study group eects. Brown and Shie Brown and Shie (1990a) Brown and Shie (1990b) Brown and Shie (1991) and Trochanis Trochanis et al. (1991) conducted a series of 3D Finite Element Method (FEM) studies on the behavior of single pile and pile group with elastic-plastic soil model. In particular, interface element was used to account for pile-soil separation and slippage. Moreover, several model and eld tests of free- or xed-head pile groups have been analyzed by Kimura et al. Kimura et al. (1995) and Wakai et al. Wakai et al. (1999) using 3D elasto-plastic FEM. This paper describes 3D elastic-plastic nite element modeling of two pile groups founded in sands with emphasis on the interaction eects within pile group. Specically, bending moment and load distribution in individual piles were examined and compared with centrifuge test data. Special attention was given to out-of-loading-plane bending moment and py behavior of individual piles in a group. The OpenSees OpenSees Development Team (Open Source Project) (1998-2005) nite element framework was employed to complete all the computations. Soil modeling was performed using the Template ElasticPlastic approach (Jeremi and Yang Jeremi and Yang (2002)). This paper is organized as follows. Section 11.4.2 summarizes the centrifuge tests and describes nite element models including the soil elasticplastic model used for 3 3 and 4 3 pile group simulations. Section 11.4.5

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

345

presents a number of results and discussion describing simulated behavior of analyzed pile groups. In particular, presented are developed plastic zones (Section 11.4.6), pile bending moments (Section 11.4.7), pile load distributions (Section 11.4.8), comparison of p y curves for individual piles (Section 11.4.9), and comparison with centrifuge tests (Section 11.4.10). Section 11.4.11 gives concluding remarks.

11.4.2 11.4.3

In the centrifuge tests reported by McVay et al. McVay et al. (1998), 3 3 to 7 3 pile groups embedded in homogeneous sands under lateral load were tested. The rectangularshaped sample container was fabricated from aluminum alloy with an inside dimensions of 0.254 m wide, 0.457 m long, and 0.305 m high. The piles were spaced by three times the pile width and the pile caps (made of aluminum) were rigidly connected with the piles. The model square piles and pile cap were fabricated from solid square aluminum (alloy 6061) bars. Each individual pile is 9.5 mm wide and 304.8 mm long. To simulate the installing eects of eld driven piles, the piles were driven in ight into sands by hydraulic equipment and tested at 45 g.

The sands (articially mixed by a number of dierent gradations) studied were at two dierent relative densities: a loose sand with relative density Dr=36%, unit weight = 14.05 kN/m3 and a medium dense sand with Dr = 55%, = 14.50 kN/m3 . The shear modulus G, Poissons ratio and friction angle are 8230 kN/m2 , 0.35, 34.5o for the loose sand and 8960 kN/m2 , 0.35, 37.1o for the medium dense sand. It is noted that the friction angles were determined from drained triaxial compression tests, and the shear moduli were backcomputed from instrumented vertical load tests ( Zhang et al. Zhang et al. (1998)) and were valued at a depth of 13.7 m.

11.4.4

Among these tested pile groups, 3 3 and 4 3 groups were chosen to be modeled in prototype scale using 3D elastoplastic nite element method to investigate pile group interaction eects. The typical layout of 4 3 pile group is shown in Fig. 11.43. The whole centrifuge model in prototype scale is 22.8 m wide, 20.6 m long and 13.2 m deep. Only half of each centrifuge model is meshed considering the symmetry . Figure 11.44 shows the nite element mesh for the 4 3 pile group. Additional nite element analysis of a cantilever beam using the same mesh as an individual pile in the group was carried out and comparison of the displacement at the top of the beam from FEM and beam theory solution indicated that the mesh was ne enough to capture the pile behavior. Soil, pile and soilpile interface are all modeled with twenty node brick elements. Each pile consists of four elements (per cross section) made of elastic material with properties corresponding to aluminum. There are 1268 and 1414 brick elements in the two models respectively. The sides and bottom of each model are xed in all three coordinate directions with the exception of the symmetric boundary, which is only supported in the direction perpendicular to the symmetry plane. This type of boundary conditions are fairly close to the actual friction boundary conditions in the centrifuge tests. All the parameters except the Youngs moduli for sands were the same as from the centrifuge studies. The Youngs

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

346

modulus is assumed to depend on the mean eective normal stress p (Manzari and Dafalias Manzari and Dafalias (1997) ) as: p pa

n

E = Eo

(11.5)

where Eo is Youngs modulus at the atmospheric pressure, p =ii /3 is the mean eective normal stress, pa is the atmospheric pressure, and n is constant for a given void ratio. Usually 0.5 is used for n. For the medium dense sand, the Youngs modulus at the atmospheric pressure computed by Eqn. (11.5) (Lateral pressure coecient Ko = 0.55 was used) from the backcomputed shear modulus was 200000 kPa and then adjusted to be 17400 kPa for the medium dense sand, so as to well simulate the loaddisplacement curve obtained from centrifuge tests. For the loose sand, the computed Youngs modulus at the atmospheric pressure from the backcomputed shear modulus was 18700 kPa and similarly adjusted to be 16000 kPa. Sand was simulated by DruckerPrager material model with nonassociated ow rules. Since the centrifuge studies we used to compare our simulations against did specify only the friction angle of test sands which were obtained from drained triaxial compression tests, the yield surface was chosen to agree with MohrCoulomb hexagon at

triaxial compression. Future study using MohrCoulomb material model will be useful to determine the eects of varying friction angle in DruckerPrager model on the results. Since there is no test data on the dilation angle of the tested sands, a dilation angle of 0o was used in this work, as similar dilation angle was also used in Brown and Shie Brown and Shie (1990a). The soilpile interface was represented by one thin layer of elements. The material of the interface element was also simulated by DruckerPrager model with a friction angle of 25o , a dilation angle of 0o , and the same Youngs modulus and Poissons ratio as corresponding sands. In the future, a realistic dilation angle needs to be used for sand to further investigate the eects of dilation angle on the pile group interaction behavior, especially the outofloadingplane bending moment.

Lateral Load

Y

3D

2.3 m

3D

Side Piles

X

D=0.429 m

Side Piles

a)

Jeremi et al.

10.9 m

Middle Piles

Lecture Notes

347

11.4.5

Simulation Results

In this section we present results related to the behavior of 3 3 and 4 4 pile groups in loose and medium dense sands. A number of static pushover tests were simulated using FEM. Specically, modied NewtonRaphson method were used to solve the system of equations in the nite element level and implicit algorithm (Jeremi and Yang Jeremi and Yang (2002)) was used in constitutive level integration. Results are also compared with those from the centrifuge studies by McVay et al. McVay et al. (1998) and Zhang et al. Limin Zhang and Lai (1999).

11.4.6

Plastic Zone

The static pushover tests were conducted using load control at the pile head with the loading applied in the X direction. The nal plastic zones (represented by plastied Gauss points) for two pile groups are depicted in Figures 11.45 and 11.46. In particular, Figure 11.45 shows dierent views of the 3D plastic zone developed in the 3 3 pile group at the lateral load of 2,200 kN. Figure 11.46 shows the dierent views of the 3D plastic zone developed in the 4 3 pile group at the lateral load of 2,970 kN. Both gures clearly show a wedge shaped plastic zone at the shallow depth. Also apparent is the propagation of the plastic zone (shear yielding) along the pilesoil interface, resulting from the rocking behavior of the group.

Jeremi et al.

Lecture Notes

348

a)

b) Figure 11.45: Plastic GaussPoints for the 3 3 pile group: (a) 3D view and (b) side view.

a)

b) Figure 11.46: Plastic GaussPoints for the 4 3 pile group: (a) 3D view and (b) side view.

Jeremi et al.

Lecture Notes

349

11.4.7

Bending Moment

The maximum bending moment in the piles often controls the design of pile groups and therefore has to be analyzed accurately. In order to generate the bending moments from the stress eld in piles vertical stresses at Gauss points from each pile element are integrated numerically. The moments with reference to the Y-axis (My ) of each individual pile in the 3 3 and 4 3 pile groups are plotted in Figure 11.47 and 11.48, respectively. Figure 11.47 shows the moment diagrams of the 3 3 group at a lateral load of 2200 kN. The calculated maximum moment for each pile occurs at the pile cap, which is consistent with the xed pile cap condition. The lead row piles, for both side and middle piles, in both 3 3 and 4 3 pile groups, carry the maximum bending moment. It is also interesting to note that in the 4 3 group, the moment diagrams for the piles in the third and fourth rows appear to be almost identical, which implies that they behave almost the same within the group. The variations of maximum bending moments in each pile of 3 3 and 4 3 groups are illustrated in Figures 11.49 and 11.50, respectively. The maximum moments develop in the leadrow side piles, while the smallest maximum moments occur in the trailrow middle piles in both groups. It is obvious that the maximum moments developed in the middle and side piles within the lead row are quite dierent for both pile groups, implying the load shared by each pile in the same row is dierent. For example, in the 3 3 group, the maximum moment in

the middle pile at the end load is 600 kN.m, while that of the side pile is 670 kN.m, the dierence is about 11%. For the 4 3 pile group, the maximum moment on the leadrow side pile was about 10% greater than that for the leadrow middle pile.

2 2

Depth (m)

8 Side Pile Trail Row Side Pile Middle Row Side Pile Lead Row 10 1000 500 My (kN.m) 0 500

8 Middle Pile Trail Row Middle Pile Middle Row Middle Pile Lead Row 10 1000 500 My (kN.m) 0 500

Figure 11.47: Comparison of bending moment diagram at lateral load of 2200 kN for piles in 3 3 group.

It is interesting to look at the moments with reference to the X-axis (out-of-loading-plane moment, Mx ). Figure 11.51 (a, b) shows Mx diagram for each pile as well as deformed piles for the 3 3 pile group. Similar plots for the 4 3 pile group are shown in Figure 11.53 (a, b). The maximum value of Mx in the 3 3 pile group reaches

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

350

Depth (m)

8 Side Pile Trail Row Side Pile Second Row Side Pile Third Row Side Pile Lead Row 10 1000 500 My (kN.m) 0 500

8 Middle Pile Trail Row Middle Pile Second Row Middle Pile Third Row Middle Pile Lead Row 10 1000 500 My (kN.m) 0 500

Figure 11.48: Comparison of bending moment diagram at lateral load of 2970 kN for piles in 4 3 group.

700 Trail Row, Side Pile Middle Row, Side Pile Lead Row, Side Pile Trail Row, Middle Pile Middle Row, Middle Pile Lead Row, Middle Pile

600

500

400

300

200

100

10

Jeremi et al.

Lecture Notes

351

900 Trail Row, Side Pile Third Row, Side Pile Second Row, Side Pile Lead Row, Side Pile Trail Row, Middle Pile Third Row, Middle Pile Second Row, Middle Pile Lead Row, Middle Pile

800

700

600

500

400

300

200

100

10

50 kN.m, which is about 8% of the maximum value of My . For the 4 3 pile group, the maximum value of Mx is about the same amount, which is about 6% of the maximum value of My . It is noted that the signs of the Mx moments in the lead and trail rows are dierent, indicating the bending directions are opposite to each other. This is further veried by looking at the deformed shape of the pile group shown in Figure 11.51(b) with only the displacement in Y direction shown. This kind of bending is caused by the the complex displacement eld of the soil surrounding the pile group, as illustrated by the horizontal displacement vector and contour plots of displacement in X direction in Figures 11.52 and 11.54 for the 3 3 and 4 3 pile groups, respectively. The soil in front of the lead row tends to squeeze into the group, while the soil outside of the trail row tends to come back toward the pile group when the pile group is moving forward, which consequently results in the fact that the lead row bends outward and the trail row bends inward.

Jeremi et al.

Lecture Notes

352

Depth (m)

8 Side Pile Trail Row Side Pile Middle Row Side Pile Lead Row 10 50

a)

0 Mx (kN.m)

50

b)

Figure 11.51: (a) Out-of-loading-plane bending moment diagram and (a) Out-of-loading-plane deformation for the 3 3 pile group.

4 3.5 3 1.07 2.5 2 Y (m) 1.5 1 0.5 0 0.5 1 5 4 o Pile Centers 3 2 1 Z = 0.0m Displacement (cm) 0 X (m) 1 2 3 4 5 2.5 1.43 5.36 5.36 5.01 3.93 2.5 1.79 2.86 0.358 0.715

Figure 11.52: Horizontal displacement vector and contour of displacement in X direction at ground surface for the 3 3 pile group.

Jeremi et al.

Lecture Notes

353

Depth (m)

8 Side Pile Trail Row Side Pile Second Row Side Pile Third Row Side Pile Lead Row

10 50

a)

0 My (kN.m)

50

b)

Figure 11.53: (a) Out-of-loading-plane bending moment diagram and (b) Out-of-loading-plane deformation for the 4 3 pile group.

4 3.5 3 2.5 2 Y (m) 6.59 1.5 1 2.35 0.5 0 0.5 1 5 4 o Pile Centers 3 2 1 Z = 0.0m Displacement (cm) 0 X (m) 1 2 3 4.71 7.06 7.06 6.12 2.35 0.471 1.41

0.942

4.24

Figure 11.54: Horizontal displacement vector and contour of displacement (cm) in X direction at ground surface for 4 3 pile group.

Jeremi et al.

Lecture Notes

354

11.4.8

Load Distribution

In order to compute the load taken by each pile, the values of bending moment at element centers along with the boundary condition at the bottom of pile (zero moment) were tted with a 5th order polynomial by least square technique. According to the theory of beam on a Winkler-type subgrade (Hartog Hartog (1952) ), the moment curve was dierentiated once to compute the shear force. Then, the shear forces at three sampling points between the ground surface and the pile cap were averaged to compute the load carried by each individual pile. The accuracy of the load measuring scheme has been veried by comparing the total load actually applied on the pile cap and the sum of all loads carried by each individual pile.

1200

1000

800

600

400

Figures 11.55 and 11.56 show the load and percentage of total load carried by each row of the 3 3 group. It is easily seen that not only the load but the percentage of total load taken by each row, especially the lead and trail rows, change steadily during loading process. It is observed in Figure 11.55 that the lead row and trail row share almost the same amount of load at small lateral displacement. However, as deection increases, the lead row picks up the load much faster that the trail row, although the load taken by the three rows all increase. More interestingly, the variation in the percentage of load carried by each row exhibits completely dierent trends. In the initial loading stage, the percentage shared by the lead row increases and that by the trail row drops quickly, while the percentage shared by the middle row almost remains constant. In the nal loading stage, however, the percentages carried by the lead and trail rows tend to stabilize. The lead row takes the most load, more than 46%, while the trail row takes the least, only around 25%. In addition, Figure 11.56 shows that the trail row takes a greater percentage of load than the lead row at small lateral deformation, which were attributed to some load measuring error and the fact that the denominator is relatively small. Similar plots for the 4 3 group are shown in Figures 11.57 and 11.58. While the lead row still carries much

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

355

45

40

35

30

25

20

10

Figure 11.56: Variations of percentage of total load taken by each row in the 3 3 group.

more load that the trail row, the third row and the trail row share almost the same amount of load at large lateral displacement, which is in agreement with the fact that the same pmultiplier was recommended for the third and fourth row by McVay et al. McVay et al. (1998). It should also be observed from Figures 11.57 and 11.58 that the lead row carries more than twice the load of the trail row.

1200

1000

800

600

400

It is worthwhile noting that the distribution of load in the same row can be quite dierent. Figures 11.59 and 11.60

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

356

35

30

25

20

15

10

11

Figure 11.58: Variations of percentage of total load taken by each row in the 4 3 group.

show the variations of load and the percentage of total load taken by each pile of the 3 3 group, respectively. Similar plots for the 4 3 group are shown in Figures 11.61 and 11.62. It is obvious that the piles at the sides take more load than the piles in the middle at the same row. For the 3 3 group, the side pile in lead row takes 350 kN or 16% of total load while the middle pile takes 325kN or 14.6% of total load at the end of loading.

300 Trail Row, Side Pile Middle Row, Side Pile Lead Row, Side Pile Trail Row, Middle Pile Middle Row, Middle Pile Lead Row, Middle Pile

250

200

150

100

50

10

Jeremi et al.

Lecture Notes

357

16 Trail Row, Side Pile Middle Row, Side Pile Lead Row, Side Pile Trail Row, Middle Pile Middle Row, Middle Pile Lead Row, Middle Pile

15

14

13

12

11

10

10

Figure 11.60: Variations of percentage of total load taken by each pile in the 3 3 group.

400 Trail Row, Side Pile Third Row, Side Pile Second Row, Side Pile Lead Row, Side Pile Trail Row, Middle Pile Third Row, Middle Pile Second Row, Middle Pile Lead Row, Middle Pile

350

300

250

200

150

100

50

10

Jeremi et al.

Lecture Notes

358

15 Trail Row, Side Pile Third Row, Side Pile Second Row, Side Pile Lead Row, Side Pile Trail Row, Middle Pile Third Row, Middle Pile Second Row, Middle Pile Lead Row, Middle Pile

14

13

12

11

10

10

11

Figure 11.62: Variations of percentage of total load taken by each pile in the 4 3 group.

Jeremi et al.

Lecture Notes

359

11.4.9

p y Curve

To further investigate the behavior of each pile, it is worthwhile to examine the p y behavior of each pile in the group. The tted moment curve and the resulting displacement from FEM were used to derive the p y curves according to beam theory. Figures 11.63, 11.64, 11.65 and 11.66 show the derived p y curves at given depths for all individual piles in the two groups. As is evident from p y plots for the two groups, the piles in the lead row exhibit much larger resistances than the piles in the middle and trail rows at large lateral displacement, due to the wellknown shadowing eect. Comparison of the p y curves at dierent depths in the same pile shows that the lateral resistance p increases as depth increases. For example, at deection of 4 cm in the 3D (3 pile diameters) spaced 3 3 group, the lateral resistance on the lead-row side pile is 120 kN/m at a depth of -1.54 m, while it is only 90 kN/m at a depth of -0.58 m. This is caused by the increases in vertical stress and Youngs modulus as depth increases. More interestingly, it is seen that the py curves of the leadrow and trailrow piles at the depth of -1.54 m are identical at small lateral displacement (y < 0.5 cm). Then the py curves of the piles within the trail row soften drastically as lateral displacement increases in the 3 3 group. Similarly for the 4 3 group, both the py curves of the leadrow and trailrow piles, and the py curves of the thirdrow and secondrow piles at the depth of

-2.04 m are almost identical at small lateral displacement (y < 0.7 cm). As deection increases, the py curves of the trailrow and thirdrow piles soften quickly. Obviously, each pile in the group exhibits quite dierent behavior than each other. It is believed that the dierent behavior of each pile is directly related to the yielding of soil in front of these piles. This observation can be veried by the fact that the softening behavior of the trail-row piles starts at larger lateral displacement as the depth increases, since the plastic zone rst develops at the ground surface and then extends downward as deection increases. These observations also imply that the pmultiplier approach might not be appropriate, especially at small deection, since it obtains the p y curves for piles within the group by simply scaling the singlepile p y curve. In addition, as a comparison with 3D (3 pile diameters) spaced pile group, also plotted in Figures 11.63 and 11.64 is the py curve of a leadrow side pile from a 6D spaced pile group. It is apparent that the lateral resistance increases with the increase of spacing.

11.4.10

The pile head displacements for the two pile groups from 3D FEM and centrifuge tests (McVay et al. McVay et al. (1998)) were plotted against pile head load in Figures 11.67 and 11.68. It can be seen that they agree with each other fairly well at the small lateral displacement and the FEM model is relatively stier at large lateral displacement. The accuracy of nite element modeling can also be examined in terms of maximum bending moments. The maximum moment developed in 4 3 group was compared with that from the centrifuge study ( Zhang et al. Limin Zhang and Lai (1999) ) in Figure 11.69. The results from the centrifuge study are slightly larger than that from FEM, which is partially due to the relatively simple elastic-plastic soil model used. Moreover, all the

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

360

160 140 Lead Row, 6D Trail Row, 3D Middle Row, 3D Lead Row, 3D Depth 0.583 m Side Piles

160 140 120 100 80 60 40 20 0 Lead Row, 6D Trail Row, 3D Middle Row, 3D Lead Row, 3D Depth 0.583 m Middle Piles

120 100 80 60 40 20 0

Figure 11.63: p-y curves for each individual pile in the 3 3 group at the depth -0.58m.

200 180 Lead Row, 6D Trail Row, 3D Middle Row, 3D Lead Row, 3D Depth 1.538 m Side Piles

200 180 160 140 120 100 80 60 40 20 0 1 2 3 4 5 0 0 1 2 3 4 5 Depth 1.538 m Middle Piles Lead Row, 6D Trail Row, 3D Middle Row, 3D Lead Row, 3D

Figure 11.64: p-y curves for individual piles in the 3 3 group at the depth of -1.54m.

above results can be further improved if the mesh is rened, since the rened pile group mesh will exhibit softer loaddisplacement response, and therefore develop larger bending moment in individual piles. The percentage of total lateral load taken by each row from 3D FEM and centrifuge tests (McVay et al. McVay et al. (1998)) at a lateral load of 1650 kN and 2300 kN for the 3 3 and 4 3 pile groups, respectively, were compared in Figures 11.70 and 11.71. Results for both loose and medium dense sand cases are included. Figures 11.70 and 11.71 show that the density of sand does not have much eects on the load distribution. It is evident that the load distributions to all rows as obtained from FEM and centrifuge tests in all the cases for both pile groups agree very well (the dierences are within 3%). Finally, the variation of distribution of total load to each row as obtained from 3D FEM and centrifuge tests for the 4 3 pile group was compared in Figure 11.72. At small lateral displacement, the FEM computed percentage of total load acting on the lead row was smaller than measured in the centrifuge tests and the load acting on the

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

361

140 120 Trail Row, 3D Third Row, 3D Second Row, 3D Lead Row, 3D Depth 0.734 m Side Piles

140 120 100 80 60 40 20 0 Trail Row, 3D Third Row, 3D Second Row, 3D Lead Row, 3D Depth 0.734 m Middle Piles

100 80 60 40 20 0

120 Trail Row, 3D Third Row, 3D Second Row, 3D Lead Row, 3D Depth 2.038 m Side Piles

120 Trail Row, 3D Third Row, 3D Second Row, 3D Lead Row, 3D Depth 2.038 m Middle Piles

100

100

80

80

60

60

40

40

20

20

trail row was larger than measured in the centrifuge tests. When lateral displacement is beyond 34 cm, however, the FEM computed load distribution tends to stabilize and agrees well with that from centrifuge tests.

Jeremi et al.

Lecture Notes

362

2000

1500

1000

500

10

12

Figure 11.67: Comparison of load displacement response for the 3 3 pile group.

2500

2000

1500

1000

500

10

Figure 11.68: Comparison of load displacement response for the 4 3 pile group.

Jeremi et al.

Lecture Notes

363

700

600

500

400

300

200

100

10

Figure 11.69: Comparison of maximum bending moment response for the 4 3 group.

50 Centrifuge Dense Sand Centrifuge Loose Sand FEM Dense FEM Loose

45

40

30

25

20

15

10

Figure 11.70: Comparison of percentage of total lateral load taken by each row in the 3 3 group

Jeremi et al.

Lecture Notes

364

40 Centrifuge Dense Sand Centrifuge Loose Sand FEM Dense FEM Loose

35

25

20

15

10

Figure 11.71: Comparison of percentage of total lateral load taken by each row in the 4 3 group

45

40

FEM Trail Row FEM Third Row FEM Second Row FEM Lead Row Centrifuge Trail Row Centrifuge Third Row Centrifuge Second Row Centrifuge Lead Row

35

30

25

20

15

10

Figure 11.72: Variation of percentage of total load taken by each row in the 4 3 pile group.

Jeremi et al.

Lecture Notes

365

11.4.11

Conclusions

This paper presents results from the nite element study on the interaction eects of pile groups founded in sands. Specically the 3 3 and 4 3 pile groups were analyzed in terms of plastic zone, bending moment and load distribution among individual piles. Comparison of results from FEM and centrifuge study shows that elastoplastic nite element analysis can predict the behavior of pile group with very good accuracy. Particularly, load distribution results from nite element analyses agree very well with that from centrifuge study. It was shown that not only the load taken by each row in the group is dierent, but the load shared by individual piles and maximum bending moment developed in individual piles within the same row vary quite a bit, as observed in the centrifuge tests. Although the dierence between the loads taken by leadrow middle and side piles is less than 2% of the total load, the dierence between the maximum bending moments developed in the leadrow middle and side piles reaches 1011% in the two pile groups. Interestingly, it was found that bending moment also occurs in the plane perpendicular to the loading direction. The numerically generated py curves were used to study the behavior of each pile in a group. It was found that individual piles in the group exhibit quite dierent py behavior at small deection, which means that the interaction in elastic range is dierent than that for loading in plastic range. Therefore, dierent interaction factors would be more appropriate depending on the loading range. Since FEM can capture the critical aspects of group eects, it could now be used to systematically study various pile group congurations at much smaller cost than actual load tests, and derive interaction factors for elastic and plastic loading levels that could be used in standard design practice.

Jeremi et al.

Lecture Notes

366

Jeremi et al.

Chapter 12

(1989-2002-2009-2010-2011-)

367

Lecture Notes

368

12.1 12.2

Jeremi (2010)

12.2.1

Earthquakes release large amounts energy at the source1 Part of released energy is radiated as mechanical waves ( 1.6 105 ) and part of that energy makes it to the surface where SFS system is located. Mechanical seismic wave energy enters the SFS system through a closed surface that encompasses (signicant) soil volume as well as foundation system and the structure (see Figure 12.7). Kinetic energy ux through closed surface includes both incoming and outgoing waves and can be calculated using Domain Reduction Method (Bielak et al., 2003a) as:

Ef lux =

+ 0 + 0 + 0 + 0 0; Mbe u e Kbe ue ; Meb u b + Keb ub i

ui

+ + + + where Mbe , Meb , Kbe , Keb are mass and stiness matrices, respectively for a single layer of elements just 0 outside of the boundary , while u 0 e and ue are accelerations and displacements from a free eld model for nodes

belonging to that layer of elements. Alternatively, energy ux can be calculated using (Aki and Richards, 2002):

t

Ef lux = Ac

0

u 2 i dt

Outgoing kinetic energy can be obtained from outgoing wave eld wi , (from DRM, Bielak et al. (2003a)), while the dierence then represents the incoming kinetic energy that needs to be dissipated with SFS region.

12.2.2

Seismic energy that enters the SFS system will be dissipated in a number of ways. part of the energy that enters SFS system can be reected back into domain outside by wave reection from impedance boundaries (free surface, soil/rock layers...). SFS system oscillation radiation. While the rest of seismic energy is dissipated through one of the following mechanisms within SFS domain:

1

for example, some of the recent large earthquake energy releases are listed: Northridge, 1994, MRichter = 6.7, Er = 6.8 1016 J ;

Loma Prieta, 1989, MRichter = 6.9, Er = 1.1 1017 J ; Sumatra-Andaman, 2004, MRichter = 9.3, Er = 4.8 1020 J ; Valdivia, Chile, 1960, MRichter = 9.5, Er = 7.5 1020 J ;

Jeremi et al.

Lecture Notes

369

Elasto-plasticity of soil Viscous coupling of porous solid with pore uid (air, water) Elasto-plasticity/damage of the foundation system Elasto-plasticity/damage of the structure viscous coupling of structure with surrounding uids (air, water) It is also important to note that in numerical simulations (advocated and used in this work), part of the energy can be dissipated or produced by purely numerical means. That is, numerical energy dissipation (damping) or production (negative damping) has to be carefully controlled (Argyris and Mlejnek, 1991b), (Hughes, 1987).

12.2.2.0.1

Elastic-plastic deformation of soil, foundation and structure is probably responsible for major part of the energy dissipation for large earthquakes. This, displacement proportional dissipation is a result of dissipation of plastic work (W = ij d

pl ij )

and is present in all three components of the system (soil, foundation and the structure).

Ideally, majority of the incoming energy would be dissipated in soil, before reaching foundation and structures. The possibility to direct energy dissipation to soil can be used in design by recognizing energy dissipation capacity for dierent soils. For example, simple elastic-plastic models of sti and soft clay as well as dense and loose send predict dierent energy dissipation capacities, as shown in Figure 12.1, for single loading-unloading-reloading cycle. While Figure 12.1 shows that sti clay and dense sand have much higher dissipation capacity, it is important to

Figure 12.1: Energy dissipation capacity for one cycle at various strains for four generic soils. note that soft/loose soils can undergo much larger deformation/strain, thus oering increased energy dissipation capacity through exibility.

12.2.2.0.2

Viscous coupling of pore uid (air, water...) and soil particles and/or foundation or structural components is responsible for velocity proportional energy dissipation. In particular, viscous coupling of porous solid and uid

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

370

i u results in Evc = n2 k 1 (U i )2 energy loss per unit volume. It is noted that this type of dissipation is realistically modeled using u p U formulation (Jeremi et al., 2008).

12.2.2.0.3

As noted above, numerical integration of nonlinear equations of motions aects calculated energy in various ways. Most common eect for nonlinear (elastic-plastic) systems is the positive (energy dissipation) and negative (energy production) damping. For example Newmark (N) (Newmark, 1959b) and HilberHughesTaylor (HHT) (Hilber et al., 1977b) are energy preserving for linear elastic system with proper choice of constants ( = 0.0; = 0.25, = 0.5). Both methods can also be used to dissipate higher frequency modes for linear elastic models 0, = by changing constants so that for N: 0.5, = 0.25( + 0.5)2 , while for HHT: 0.33 0.5(1 2), = 0.25(1 )2 . However, for nonlinear problems it is impossible to maintain energy of the system throughout computations (Argyris and Mlejnek, 1991b).

12.3

Seismic waves propagate from the source (rupturing fault) through the bedrock, surface rock and soil layers to the site of interest where a Nuclear Power Plant is located (or planned). Seismic (compressional and shear) waves do travel through various rock and soil domains, which can be represented by layers, which are sometimes horizontal, but mostly inclined. In addition to that, layers usually have variable thicknesses, creating a complex underground picture of soil/rock domains of dierent stiness, mass, energy dissipating characteristics (damping). Most commonly, the stiness of layers (horizontal and/or inclined) increases with stiness. This change in stiness results in seismic wave refraction, as shown in Figure 12.2.

se rupturing fault

is

a cw i m

p ve

rop

a ag

n tio

Figure 12.2: Propagation of seismic waves in nearly horizontal local geology, with stiness of soil/rock layers increasing with depth, and refraction of waves toward the vertical direction.

When soil/rock layers are mostly horizontal, inter-layer refraction causes seismic waves (both P and S) to tend toward vertical propagation. This case is shown in Figure 12.2. However, (near) horizontal layering of geologic layers is not very frequent, unless young deposits of soil in river valleys are of concern. More often the soil/rock

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

soil/rock stiffness

Lecture Notes

371

layers are inclined, thus creating conditions for variable directivity of seismic motions due to refraction. Figure 12.3 shows one such case where inclined soil/rock layers contribute to mainly horizontal propagation of seismic motions close to the surface, in the vicinity of an NPP.

is se

mi

a cw

ro ep

a ag

n tio

rupturing fault

Figure 12.3: Propagation of seismic waves in inclined (close to vertical) local geology, with stiness of soil/rock layers increasing through geologic layers, and refraction of waves away from the vertical direction.

A general conclusion can be made that seismic waves arriving at the particular site (surface) will be fully three dimensional (3D), uncorrelated, and incoming at an oblique angle. As noted by Zerva (2009), seismic motions will feature lack of correlations of motions between two monitored points at the surface. Lack of correlation is mainly due to (see more details in section 12.3.1): Attenuation eects, Wave passage eects, Scattering eects, Extended source eects Due to all the above mentioned 3D eects, realistic modeling of seismic wave propagation can only properly be done by fully and realistically incorporating: Body waves (P and S), Surface waves (Love and Rayleigh), seismic waves coming at an oblique angle (which will actually be done when the above two types of waves are properly modeled), Full three dimensional (3D) wave eld, Lack of correlation (incoherence) eects, Energy dissipation eects (damping).

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

372

12.3.1

Seismic motion incoherence (as it is called for frequency domain analysis, for time domain analysis it is called lack of correlation) is a phenomena that results in spatial variability of ground motions over small distances. Signicant work has been done in researching seismic motion incoherence over the last few decades. The main sources of lack of spatial correlation, according to Zerva (2009) are due to: Attenuation eects, Wave passage eects, Scattering eects, Extended source eects Figure 12.4 shows an illustration of main sources of lack of correlation.

Attenuation 1

Fault

Figure 12.4:

Wave Front(s)

Four main sources contributing to the lack of correlation of seismic waves as measured at two

observation points.

12.3.1.0.4

Attenuation eects

are responsible for change in amplitude and phase of seismic motions due to the distance between observation points and losses (damping, energy dissipation) that seismic wave experiences along that distance. This is a signicant source of lack of correlation for long structures (bridges), however for NPPSSS it is not of much signicance.

12.3.1.0.5

contribute to lack of correlation due to dierence in recorded wave eld at two location points as the (surface) wave travels, propagates from rst to second point.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

373

12.3.1.0.6

Scattering eects

are responsible to lack of correlation by creating a scattered wave eld. Scattering is due to (unknown or not known enough) subsurface geologic features that contribute to modication of wave eld.

12.3.1.0.7

contribute to lack of correlation by creating a complex wave source eld, as the fault ruptures, rupture propagates and generate seismic sources along the fault. Seismic energy is thus emitted from dierent points (along the rupturing fault) and will have dierent travel path and timing as it makes it observation points. Early studies concluded that the correlation of motions increases as the separation distance between observation points decreases. In addition to that, correlation increased for decrease in frequency of observed motions. Moreover, there is a strong probabilistic nature of this phenomena, as signicant uncertainty is present in relation to all four sources of lack of correlation, mentioned above. A number of excellent references are available on the subject of incoherent (or lacking correlation) seismic motions Abrahamson et al. (1991); Roblee et al. (1996); Abrahamson (1992a, 2005, 1992b); Zerva and Zervas (2002); Liao and Zerva (2006); Zerva (2009)

It is very important to note that all current models for modeling incoherent seismic motions make an ergodic assumption. That is all the models assume that a variability of seismic motions at a single site source combination will be the same as variability in the ground motions from a data set that was collected over dierent site and source locations Walling (2009). Unfortunately, there does not exist a large enough data set for east North American seismic seismic events that can be used to develop incoherence models. Rather, there are models that are used to model possible incoherent behavior for east North American seismic wave propagations.

12.3.2

Incorporation of lack of correlation eects in seismic motions can be done using the following methods: along 1D, in one direction, usually one of the horizontal directions, where all the points that are the same distance (in 1D) from a control point, plane) share a single lack of correlation, along 2D, in two directions, usually in a vertical plane or in a horizontal plane, where all the points in a set of vertical planes or a set of horizontal planes, share same lack of correlation, and in full 3D, where every point of interest (in a 3D volume of soil/rock) has it own, specied, lack of correlation. The method used here is using the so called seed motions, motions obtained through DRM in full 3D (inclined, body and surface waves), that are than enriched with appropriate uncorrelated (incoherent) components. It is important to note that only translational motions are used to model incoherence eects, while the rotational motions are not perturbed/made incoherent. This stems from the fact that currently vast majority of seismic recording stations only record translational motions and that only those translational motions are used to develop incoherent motions models. Code developed by Abrahamson (1992b) is used for this purpose

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

374

MORE TO COME

12.4

Both multi-directional and seismic input coming at an inclined angle is possible by using (and in fact it is inherent to) the DRM, described in previous section. It is important to note that both seismic body waves (P and S) and seismic surface waves (Rayleigh and Love) are present in all situations, and do contribute to multi-directional and inclined seismic input.

Figure 12.5: Illustration of seismic body waves, namely the P (primary) and S (secondary) waves (illustration from MTU web site).

12.4.0.0.8

It has long been recognized that surface waves are responsible for majority of destruction and the seismic energy at some distance from the epicenter during earthquakes (Kramer (1996a), Semblat and Pecker (2009)). In a perfectly linear elastic half-space, Rayleigh waves become signicant source of seismic motions at distance R = h/( (vp /vR )2 1) where h is hypocenter depth (for uniform half-space) and vp is the compressional (primary) wave velocity, while vR is the Rayleigh wave velocity (Kramer (1996a)). For example, if vR 0.93vS = 0.93 0.6vp = 0.56vp it follows that R 1.5h where vS is the shear wave velocity (Semblat and Pecker (2009)). Of course for a case with realistic geology, the pattern of occurrence of Rayleigh waves is much more complex.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

375

Figure 12.6: Illustration of seismic surface waves, namely the Rayleigh and Love waves (illustration from MTU web site).

However, since Rayleigh waves represent the result of interaction of body waves (P and S) with the free surface,

12.5

Current design practice for structures subject to earthquake loading regards dynamic SFSI to be mainly benecial to the behavior of structures (Jeremi and Preisig, 2005). Including the exibility of the foundation reduces the overall stiness of a system and therefore reduces peak loads caused by a given ground motion. Even if this is true in most cases there is the possibility of resonance occurring as a result of a shift of the natural frequencies of the SFS-system. This can lead to large inertial forces acting on a structure. As a result of these large inertial forces caused by the structure oscillating in its natural frequency the structure as well as the soil surrounding the foundation can undergo plastic deformations. This in turn further modies the overall stiness of the SFS-system and makes any prediction on the behavior very dicult. Dynamic SFSI also becomes important in the design of large infrastructure projects. As authorities and insurance companies try to introduce the concept of performance based design to the engineering community more sophisticated models are needed in order to obtain the engineering demand parameters (EDPs). A good numerical model of a soil-foundation-structure system can therefore not only prevent the collapse or damage of a structure but also help to save money by optimizing the design to withstand an earthquake with a certain return period. A variety of methods of dierent levels of complexity are currently being used by engineers. In the following an overview over the most important ones is presented. A more thorough discussion on methods and specic aspects of dynamic SFSI is available in Wolf (1985) and more recently in Wolf and Song (2002). No SFSI

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

376

The ground motion is applied directly to the base of the building. Alternatively, instead of applying the ground motion directly to the base of the structure, eective earthquake forces proportional to the base acceleration can be applied to the nodes. This procedure is reasonable only for exible structures on very sti soil or rock. In this case the displacement of the ground doesnt get modied by the presence of the structure. For stier structures on soil the ground motion has to be applied to the soil. The model has to incorporate propagation of the motion through the soil, its interaction with the structure and the radiation away from the structure. Direct methods Direct methods treat the SFS-system as a whole. The numerical model incorporates the spatial discretization of the structure and the soil. The analysis of the entire system is carried out in one step. This method provides most generality as it is capable of incorporating all nonlinear behavior of the structure, the soil and also the interface between those two (sliding, uplift). Substructure methods Substructure methods refer to the principle of superposition. The SFS-system is generally subdivided into

a structure part and a soil part. Both substructures can be analyzed separately and the total displacement can be obtained by adding the contributions at the nodes on the interface. This method reduces the size of the problem considerably. As the time needed for an analysis doesnt increase linearly with an increasing number of equations the substructure method is much faster than the direct method. The biggest drawback of the method however is the fact that linearity is assumed. For nonlinear systems the substructure method cannot be used. For the direct method dierent levels of sophistication are possible: Foundation stiness approach The behavior of the soil is accounted for by simple mechanical elements such as springs, masses and dash pots. Dierent congurations of the subsoil can be taken into account by connecting several springs, masses and dash pots whose parameters have been determined by a curve tting procedure Wolf (1994). This approach is very popular among structural engineers as it is relatively easy to be integrated in a commonly used nite element code. Other methods use frequency dependent springs and dash pots and therefore require an analysis in frequency domain. Relatively complex congurations of layered subsoil and embedded foundations can be modeled with good accuracy by replacing the (elastic) soil with a sequence of conical rods Wolf and Song (2002) and Wolf and Preisig (2003). p-y methods Attempts have been made to apply the static p-y approach for evaluating lateral loading on pile foundations to dynamic problems. E. and H. (2002) lists several references and provides a parametric study of single piles and pile groups in dierent soil types under simplied loading cases.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

377

Even if p-y curves are widely used for estimating lateral loading on piles they are rarely used in full dynamic soil-structure interaction analysis. Current work trying to implement these methods into nite element codes is likely to make them more popular with the engineering community. Full 3d Full nonlinear three-dimensional modeling of dynamic soil-foundation-structure interaction can be regarded as the brute force approach. Displacements and forces can be obtained not only for the structure as in the above mentioned methods but also for the soil. In spite of the computational resources and modeling eort required for an analysis it is the only method that remains valid for all kinds of problems involving material nonlinearities, contact problems, dierent loading cases and complex geometries.

12.6

Introduction

The Domain Reduction Method (DRM) was developed recently by Bielak et al. (2003a); Yoshimura et al. (2003a)). It is a modular, two-step dynamic procedure aimed at reducing the large computational domain to a more

manageable size. The method was developed with earthquake ground motions in mind, with the main idea to replace the force couples at the fault with their counterpart acting on a continuous surface surrounding local feature of interest. The local feature can be any geologic or man made object that constitutes a dierence from the simplied large domain for which displacements and accelerations are easier to obtain. The DRM is applicable to a much wider range of problems. It is essentially a variant of globallocal set of methods and as formulated can be used for any problems where the local feature can be bounded by a continuous surface (that can be closed or not). The local feature in general can represent a soilfoundationstructure system (bridge, building, dam, tunnel...), or it can be a crack in large domain, or some other type of inhomogeneity that is fairly small compared to the size of domain where it is found. In what follow, the DRM is developed in a somewhat dierent way than it was done in original papers by Bielak et al. (2003a); Yoshimura et al. (2003a)). The main features of the DRM are then analyzed and appropriate practical modeling issues addressed.

12.7

A large physical domain is to be analyzed for dynamic behavior. The source of disturbance is a known time history of a force eld Pe (t). That source of loading is far away from a local feature which is dynamically excited by Pe (t) (see Figure 12.7). The system can be quite large, for example earthquake hypocenter can be many kilometers away from the local feature of interest. Similarly, the small local feature in a machine part can be many centimeters away from the source of dynamic loading which inuences this local feature. In this sense the term large domain is relative to the size of the local feature and the distance to the dynamic forcing source.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

Local feature

378

ui

ub

P e (t)

ue

Figure 12.7:

Large physical domain with the source of load Pe (t) and the local feature (in this case a soil

foundationbuilding system.

It would be benecial not to analyze the complete system, as we are only interested in the behavior of the local feature and its immediate surrounding, and can almost neglect the domain outside of some relatively close boundaries. In order to do this, we need to somehow transfer the loading from the source to the immediate

vicinity of the local feature. For example we can try to reduce the size of the domain to a much smaller model bounded by surface as shown in Figure 12.7. In doing so we must ensure that the dynamic forces Pe (t) are appropriately propagated to the much smaller model boundaries .

12.7.1

Method Formulation

In order to appropriately propagate dynamic forces Pe (t) one actually has to solve the large scale problem which will include the eects of the local feature. Most of the time this is impossible as it involves all the complexities of large scale computations and relatively small local feature. Besides, the main goal of presented developments is to somehow reduce the large scale domain as to be able to analyze in details behavior of the local feature. In order to propagate consistently the dynamic forces Pe (t) we will make a simplication in that we will replace a local feature with a simpler domain that is much easer to be analyzed. That is, we replace the local feature (bridge, building. tunnel, crack) with a much simpler geometry and material. For example, Figure 12.8 shows a simplied model, without a foundationbuilding system. The idea is to simplify the model so that it is much easier to consistently propagate the dynamic forces to the boundary . The notion that it is much easier to propagate those dynamic forces is of course relative. This is still a very complex problem, but at least the inuence of local feature is temporarily taken out. It is convenient to name dierent parts of domain. For example, the domain inside the boundary is named 0 . The rest of the large scale domain, outside boundary , is then named + . The outside domain + is still the same as in the original model, while the change, simplication, is done on the domain inside boundary . The displacement elds for exterior, boundary and interior of the boundary are ue , ub and ui , on the original domain.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

379

u0 i

u0 b u0 e

P e (t)

Figure 12.8: Simplied large physical domain with the source of load Pe (t) and without the local feature (in this case a soilfoundationbuilding system. Instead of the local feature, the model is simplied so that it is possible to analyze it and simulate the dynamic response as to consistently propagate the dynamic forces Pe (t).

The equations of motions for the complete system can be written as M u + K u = Pe (12.1)

or if written for each domain (interior, boundary and exterior of ) separately, the equations obtain the following form:

Mii Mib + Mbb + Mbb + Meb

0

+ Mbe + Mee

M bi 0

u i K ii + b Kbi u 0 u e

0

+ Kbe + Kee

0 ui = ub 0 (12.2) Pe ue

In these equations, the matrices M and K denote mass and stiness matrices respectively; the subscripts i, e, and b refer to nodes in either the interior or exterior domain or on their common boundary; and the superscripts and + refer to the domains over which the various matrices are dened. The previous equation can be separated provided that we maintain the compatibility of displacements and equilibrium. The resulting two equations of motion are

Mii Mbi Mib Mbb

u i K + ii u b Kbi

Kib Kbb

u 0 i = , ub Pb

in

(12.3)

and

+ Mbb + Meb + Mbe + Mee

u K + b + bb + u e Keb

+ Kbe + Kee

u P b b , = ue Pe

in+

(12.4)

Compatibility of displacements is maintained automatically since both equations contain boundary displacements ub (on boundary ), while the equilibrium is maintained through actionreaction forces Pb . In order to simplify the problem, the local feature is removed from the interior domain. Thus, the interior domain is signicantly simplied. In other words, the exterior region and the material therein are identical to those of

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

380

the original problem as the dynamic force source. On the other hand, the interior domain (denoted as 0 ), is simplied, the localized features is removed (as seen in gure 12.8). For this simplied model, the displacement eld (interior, boundary and exterior, respectively) and actionreaction

0 0 0 + forces are denoted by u0 i , ub , ue and Pb . The entire simplied domain 0 and is now easier to analyze.

for the auxiliary problem can now be written as: + + u 0 P 0 Kbb Kbe b b + + = K K + u0 Pe

eb ee e

(12.5)

Since there was no change to the exterior domain + (material, geometry and the dynamic source are still the same) the mass and stiness matrices and the nodal force Pe are the same as in Equations (12.3) and (12.4). Second part of previous equation (12.5) can be used to obtain the dynamic force Pe as

+ 0 + 0 + 0 + 0 Pe = Meb u b + Mee u e + Keb ub + Kee ue

(12.6)

The total displacement, ue , can be expressed as the sum of the free eld u0 e (from the background, simplied

model) and the residual eld we (comming from the local feature) as following: ue = u0 e + we (12.7)

It is important to note that this is just a change of variables and not an application of the principle of superposition. The residual displacement eld, we is measured relative to the reference free eld ue 0. By substituting Equation (12.7) in Equation (12.2) one obtains:

Mii Mib + Mbb + Mbb + Meb

0

+ Mbe + Mee

M bi 0

u i u b u 0 e e +w

Kii

0

+ Kbe + Kee

+ Kbi 0

ui ub

u0 e + we 0 (12.8) 0 Pe

which, after moving the free eld motions u0 e to the right hand side, becomes

Mii Mib + Mbb + Mbb + Meb

0

+ Mbe + Mee

M bi 0

u i Kii + b u Kbi w e 0

+ 0 + 0 Mbe u e Kbe ue

u i + = Kbe ub + Kee we 0 0

+ 0 Kee ue

(12.9)

+ 0 Mee u e

+ Pe

By substituting Equation (12.6) in previous Equation (12.9), the right hand side can now be written as

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

381

Mii

0

+ Mbe + Mee

M bi 0

u i Kii + b u Kbi w e 0

0

+ Kbe + Kee

u i = u b we 0

(12.10)

The right hand side of equation (12.10) is the dynamically consistent replacement force (so called eective force, P ef f for the dynamic source forces Pe . In other words, the dynamic force Pe was consistently replaced by the eective force P ef f : ef f 0 P i ef f + 0 + 0 = = ue Mbe u e Kbe Pb ef f + 0 + 0 Meb u b + Keb ub Pe

P ef f

(12.11)

12.7.2

12.7.2.0.9

Method Discussion

Single Layer of Elements used for P ef f .

The Equation (12.11) shows that the eective nodal forces P ef f involve only the sub-matrices Mbe , Kbe , Meb , Keb . These matrices vanish everywhere except the single layer of nite elements in domain + adjacent to . The signicance of this is that the only waveeld (displacements and accelerations) needed to determine eective forces P ef f is that obtained from the simplied (auxiliary) problem at the nodes that lie on and between boundaries and e , as shown in Figure 12.9.

Local feature

e

e

u0 i u0 e

e

u0 b

P e (t)

+

u0 e

+

Figure 12.9: DRM: Single layer of elements between and e is used to create P ef f .

Jeremi et al.

Lecture Notes

382

12.7.2.0.10

Another interesting observation is that the solution to problem described in Equation (12.10) comprises full unknowns (displacements and accelerations) inside and on the boundary (ui and ub respectively). On the other hand, the solution for the domain outside single layer of nite elements (outside e ) is obtained for the residual unknown (displacement and accelerations) eld, we only. This residual unknown eld is measured relative to the reference free eld of unknowns (see comments on page 380). That eectively means that the solution to the equation Equation (12.10) outside the boundary e will only contain additional waves eld resulting from the presence of a local feature. This in turn means that if the interest is in behavior of local feature and the surrounding media (all within boundary ) one can neglect the behavior of the full model (outside e in + ) and provide appropriate supports (including xity and damping) at some distance from the boundary e into region + . This is signicant for a number of reasons:

large models can be reduced in size to encompass just a few layers of elements outside boundary e (signicant reduction for, say earthquake problems where the size of a local feature is orders of magnitudes smaller then the distance to the dynamic source force Pe (earthquake hypocenter).

the residual unknown eld can be monitored and analyzed for information about the dynamic characteristics of the local feature. Since the residual wave eld is we is measured relative to the reference free eld ue 0, the solution for we has all the characteristics of the additional wave eld stemming from the local feature.

12.7.2.0.11

In all the derivations in section 12.7 no restriction was made on the type of material inside the plastic bowl (inside e ). That is, the assumption that the material inside is linear elastic is not necessary as the DRM is not relying on principle of superposition. The Equation 12.7 was only describing the change of variables, and clearly there was no use of the principle of superposition, which is only valid for linear elastic solids and structures. It is therefor possible to assume that the derivations will still be valid with any type of material (linear or nonlinear, elastic or inelastic) inside e . With this in mind, the DRM becomes a very powerful method for analysis of soilfoundationstructure systems.

12.8

The accuracy of a numerical simulation of dynamic SFSI is controlled by two main parameters: a) the spacing of the nodes of the nite element model (h) and b) the length of the time step t. Assuming that the numerical method converges toward the exact solution as t and h go toward zero the desired accuracy of the solution can be obtained as long as sucient computational resources are available.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

383

12.8.1

Grid Spacing h

In order to represent a traveling wave of a given frequency accurately about 10 nodes per wavelength are required (Bathe and Wilson, 1976; Hughes, 1987; Argyris and Mlejnek, 1991b). Fewer than 10 nodes can lead to numerical damping as the discretization misses certain peaks of the wave. In order to determine the appropriate maximum grid spacing the highest relevant frequency fmax that is present in the model needs to be found by performing a Fourier analysis of the input motion. Typically, for seismic analysis fmax is about 10 Hz. By choosing the wavelength min = v/fmax , where v is the wave velocity, to be represented by 10 nodes the smallest wavelength that can still be captured partially is = 2h, corresponding to a frequency of 5 fmax . The maximum grid spacing should not exceed h v = 10 10 fmax (12.12)

where v is the lowest wave velocity that is of interest in the simulation. Generally this is the shear wave velocity.

12.8.2

The time step t used for numerically solving nonlinear vibration or wave propagation problems has to be limited for two reasons. The stability requirement depends on the numerical procedure in use and is usually formulated in the form t/Tn < value. Tn denotes the smallest fundamental period of the system. Similar to the spatial discretization Tn needs to be represented by about 10 time steps. While the accuracy requirement provides a measure on which higher modes of vibration are represented with sucient accuracy, the stability criterion needs to be satised for all modes. If the stability criterion is not satised for all modes of vibration, then the solution may diverge. In many cases it is necessary to provide an upper bound to the frequencies that are present in a system by including frequency dependent damping to the model. The second stability criterion results from the nature of the nite element method. As a wave front progresses in space it reaches one point after the other. If the time step in the nite element analysis is too large the wave front can reach two consecutive elements at the same moment. This would violate a fundamental property of wave propagation and can lead to instability. The time step therefore needs to be limited to t < h v (12.13)

12.8.3

If nonlinear material models are used the considerations for stability and accuracy as stated above dont necessarily remain valid. Especially modal considerations need to be examined further for these cases. It is however save to assume that the natural frequencies decrease as plastic deformations occur. The minimum time step required

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

384

to represent the natural frequencies of the dynamic system can therefore taken to be the same as in an elastic analysis.

10 8 6 Acceleration [m/s2] 4 2 0 2 4 6 8 Linear algorithm NewtonRaphson algorithm

A high frequency component is introduced due to plastic slip and counter balancing of the resulting displacement. This is especially true if a linear algorithm with no iterations within one time step is used. Figure 12.10 shows a part of an acceleration time history from an analysis involving elastic-plastic material. It can be seen that the out-of-balance forces at the end of a time step can be quite large if a linear algorithm is used. While the Newton-Raphson algorithm minimizes out-of-balance forces within one time step the linear algorithm requires several time steps to return to a stable equilibrium path. The frequencies corresponding to these peaks are typically of the order of 1 /(a few t). Normally the time step is small enough so that these frequencies dont interfere with the input motion. They can be prevented from propagating through the model by an appropriate choice of algorithmic or material damping. For stability the time step used in a nonlinear analysis needs to be smaller than in a linear elastic analysis. By how much it has to be reduced is dicult to predict as this depends on many factors such as the material model, the applied loading or the numerical method itself. Argyris and Mlejnek (1991a) suggest the time step to be reduced by 60% or more compared to the time step used in an elastic analysis. The best way to determine whether the time step is appropriate for a given analysis consists in running a second analysis with a reduced time step.

12.9

Domain Boundaries

One of the biggest problems in dynamic SFSI in innite media is related to the modeling of domain boundaries. Because of limited computational resources the computational domain needs to be kept small enough so that it can be analyzed in a reasonable amount of time. By limiting the domain however an articial boundary is introduced. As an accurate representation of the soil-structure system this boundary has to absorb all outgoing waves and reect no waves back into the computational domain. The most commonly used types of domain

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

385

boundaries are presented in the following: Fixed or free By xing all degrees of freedom on the domain boundaries any radiation of energy away from the structure is made impossible. Waves are fully reected and resonance frequencies can appear that dont exist in reality. The same happens if the degrees of freedom on a boundary are left free, as at the surface of the soil. A combination of free and fully xed boundaries should be chosen only if the entire model is large enough and if material damping of the soil prevents reected waves to propagate back to the structure. Absorbing Lysmer Boundaries A way to eliminate waves propagating outward from the structure is to use Lysmer boundaries. This method is relatively easy to implement in a nite element code as it consists of simply connecting dash pots to all degrees of freedom of the boundary nodes and xing them on the other end (Figure 12.11).

0 1 1 0 0 1 0 1

Cs Cp

Cs

1 0 0 1 0 1 0 1

Cp Cs

00 11 00 11 00 11 00 11 C

s

Cs

11 00 00 11 00 11 00 11

1 0 0 1 0 1

Cp

11 00 00 11 00 11 00 11 00 11 00 11

Cs

111 000 1 0 0 1 0 1

Cp Cs

Cs

11 00 00 11 00 11 00 11

Figure 12.11: Absorbing boundary consisting of dash pots connected to each degree of freedom of a boundary node

Lysmer boundaries are derived for an elastic wave propagation problem in a one-dimensional semi-innite bar. It can be shown that in this case a dash pot specied appropriately has the same dynamic properties as the bar extending to innity (Wolf, 1988). The damping coecient C of the dash pot equals C = Ac (12.14)

where A is the section of the bar, is the mass density and c the wave velocity that has to be selected according to the type of wave that has to be absorbed (shear wave velocity cs or compressional wave velocity cp ). In a 3d or 2d model the angle of incidence of a wave reaching a boundary can vary from almost 0 up to nearly 180 . The Lysmer boundary is able to absorb completely only those under an angle of incidence of 90 . Even with this type of absorbing boundary a large number of reected waves are still present in the

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

386

domain. By increasing the size of the computational domain the angles of incidence on the boundary can be brought closer to 90 and the amount of energy reected can be reduced. In Section ?? the Lysmer boundary is tested on a two-dimensional model. More sophisticated boundaries modeling wave propagation toward innity For a spherical cavity involving only waves propagating in radial direction a closed form solution for radiation toward innity, analogous to the Lysmer boundary for wave propagation in a prismatic rod, exists (Sections 3.1.2 and 3.1.3 in Wolf (1988)). Since this solution, in contrast to the Lysmer boundary, includes radiation damping it can be thought of as an ecient way of eliminating reections on a semi-spherical boundary surrounding the computational domain. More generality in terms of absorption properties and geometry of the boundary are provided by the various boundary element methods (BEM) available in the literature.

12.9.1

Problem Statement

The model used for verication of the DRM consists of 13 8-node brick elements (Figure 12.12). Each element

Model 1 (FreeField)

Interior Domain

Boundary Element

Boundary Element

Exterior Domain

is a cube of 1 meter side length. The four nodes in a horizontal plane are constrained to move together in all directions. Model 1 represents the free eld whereas Model 2 has a column with a lumped mass attached to its top. The translational degrees of freedom of the base of the beam-column are connected to the nodes at the top of the soil column, the rotational degrees of freedom are xed. The natural period of the pile, assuming it is xed at the soil surface, is 0.5 seconds. As input motion, a sinusoidal horizontal ground motion with a period of 1 second has been applied to the base of the soil column, resulting in an upward propagating shear wave.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

387

The following material properties have been used in the analysis: Youngs modulus E = 1944 kP a, Poissons ratio = 0.35 and specic density = 1.8 kg/m3 . The resulting shear wave velocity is 20m/s. The lowest natural frequencies are 3 Hz for model 1 and 0.31 Hz for model 2.

12.9.2

Results

As a rst test Model 2 is analyzed by prescribing the sinusoidal motion to the base of the model. The displacement and acceleration time histories of the nodes of the boundary layer are recorded. In a second analysis these time histories are used to calculate P ef f which then is applied to the nodes of the boundary layer. As can be seen in Figure 12.13 the results are identical in the interior domain. In the exterior domain the result from the DRManalysis shows no motion at all. This is the expected result since the large domain used to obtain u e and ue and the local domain are identical. The residual eld in the exterior domain we therefore vanishes. As a second test the same analysis as above is performed, but this time u e and ue from a free eld analysis have been used to calculate P ef f (Figure 12.14). In the interior domain the results from the DRM- and from the base-shaking analysis are identical at the beginning and start to diverge after about 2 seconds. Since the free-eld motions arent compatible with the soil-structure system, the motions in the exterior domain dont vanish. Those downward propagating waves, which result from interaction with the structure, will not be canceled out by P ef f in the boundary layer and will therefore leave the interior domain. At the base they will eventually be reected and sent upward into the interior domain, where undesired interference occurs. In order to reduce this undesired interference absorbing boundaries of the Lysmer type can be used (Section 12.9). The result of the previous analysis, this time with absorbing boundaries, is given in Figure 12.15. The Lysmer boundary reduces the waves in the exterior domain considerably. The displacement in the exterior domain is now entirely due to the outgoing wave caused by interaction of the free-eld motion with the beam-column at the surface. It can be seen that the residual displacement eld contains the frequency of the sinusoidal input motion and the natural frequency of the SFSI-model (0.31 Hz).

Jeremi et al.

Lecture Notes

388

Surface 4 Base shaking DRM 2 xDisp. [m] xDisp. [m] 1 0 1 2 4 0 2 Time [s] 4 6 3 0 2 Time [s] 4 6 3 2 Interior domain, z = 3 m

2 Time [s]

Base, z = 13 m

0.5

2 Time [s]

Jeremi et al.

Lecture Notes

389

Surface 4 Base shaking DRM 2 xDisp. [m] xDisp. [m] 1 0 1 2 4 0 2 Time [s] 4 6 3 0 2 Time [s] 4 6 3 2 Interior domain, z = 3 m

2 Time [s]

Base, z = 13 m

0.5

2 Time [s]

Jeremi et al.

Lecture Notes

390

Surface 4 Base shaking DRM with Absorbing Boundary 2 xDisp. [m] xDisp. [m] 1 0 1 2 4 0 2 Time [s] 4 6 3 0 2 Time [s] 4 6 3 2 Interior domain, z = 3 m

2 Time [s]

Base, z = 13 m

0.5

2 Time [s]

Figure 12.15: As in Figure 12.14 but with absorbing boundary at the base

Jeremi et al.

Chapter 13

(2003-2007-2011-)

391

Lecture Notes

392

13.1

Jeremi et al.

Lecture Notes

393

13.2

13.2.1

The nal objective this work is to improve current modeling techniques through the comparative study between numerical and experimental components. In this work, a whole prototype bridge model has been built using nite element techniques demonstrated in previous sections. Domain Reduction Method (Bielak et al., 2003b; Yoshimura et al., 2003b) (DRM) has been used to reduce the model size while still preserves the accuracy of the ground motion analysis.

13.2.1.1

Soil Model

Capitol Aggregates, a local quarry located in the south of Austin, has been selected to be the test site for this project. Site characterization has been preformed to collect information on the soil (Kurtulus et al., 2005). Based on the only triaxial test data available, a nonlinear soil constitutive model is developed in this work for prototype

nite element analysis. Undrained triaxial compression test has been carried out on one, 1.5in diameter triaxial test specimen trimmed from an undisturbed soil sample obtained from borehole at an approximate depth of 10.6f t. The initial size and index properties of the soil specimen are given in Table 13.1 (Kurtulus et al., 2005). Table 13.1: Index Properties of the Undisturbed Triaxial Test Specimen Soil Index Property Diameter D (inch) Height H (inch) Total Unit Weight t (pcf ) Water Content w (%) Dry Unit Weight d (pcf ) Void Ratio e

1 1

In the triaxial cell, the specimen was allowed to come into equilibrium (compress/consolidate with drainage lines open) under an isotropic pressure equal to the assumed in-situ mean total stress, which is about 5.6psi. Upon equilibrating, the specimen was sheared under undrained conditions with a strain rate of %1 per hour. No pore pressure readings were taken since the specimen was unsaturated. The resulting stress-strain curve is presented in Figure 13.1. An estimate of the undrained shear strength in terms of total stresses was measured as 13.41psi (1931psf ) at about 9% strain. The specimen failed in a bulging mode. The index properties of the specimen at failure are presented in Table 13.1.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

394

Figure 13.1:

Total Stress Strain Curve Determined from Undrained Triaxial Compression Test (Undisturbed

Based on the laboratory triaxial test data, a nonlinear elastic-plastic soil model has been developed to calibrate the nite element simulation. Associative Drucker-Prager plasticity model, combined with nonlinear ArmstrongFrederick kinematic hardening rule, yields good match between laboratory data and numerical results, as shown in Figure 13.2. The same model has been exposed to various connements to test robustness of the model for soils at dierent depths.

13.2.1.2

The accuracy of a numerical simulation of dynamic SFSI (Soil-Structure-Foundation-Interaction) problems is controlled by two main parameters (Preisig, 2005): 1. The spacing of the nodes of the nite element model h 2. The length of the time step t. Assuming that the numerical method converges toward the exact solution as t and h go toward zero the desired accuracy of the solution can be obtained as long as sucient computational resources are available.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

395

Figure 13.2: Total stress strain curve obtained using simulation of triaxial test (from depth 10.6f t)

As presented in Preisig (2005), in order to represent a traveling wave of a given frequency accurately about 10 nodes per wavelength are required. Fewer than 10 nodes can lead to numerical damping as the discretization misses certain peaks of the wave. In order to determine the appropriate maximum grid spacing the highest relevant frequency fmax that is present in the model needs to be found by performing a Fourier analysis of the input motion. Typically, for seismic analyses fmax is about 10Hz . By choosing the wavelength min = v/fmax , where v is the wave velocity, to be represented by 10 nodes the smallest wavelength that can still be captured partially is = 2h, corresponding to a frequency of 5fmax . The maximum grid spacing should not exceed h v min = 10 10fmax (13.1)

where v is the smallest wave velocity that is of interest in the simulation. In this work, the prototype site chosen is Capitol Aggregates, a local quarry located in the south of Austin. According to the site characterization report Kurtulus et al. (2005), we obtain Table 13.2 for element size determination.

Mechanical properties of soil changes with cyclic loadings. In order to predict more accurately the dynamic behaviors of soil subject to earthquake loadings, various laboratory and in situ tests have been performed to examine the degradation of dynamic soil properties. Equivalent linear model has been used extensively in practice (Kramer,

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

396

Table 13.2: Maximum Element Size Determination (fmax = 10HZ ) Depth (f t) 0 1 2.5 7 14 21.5 38.5 Thickness (f t) 1 1.5 4.5 7 7.5 17 half-space vshear (f ps) 320 420 540 660 700 750 2200 hmax (f t) 3.2 4.2 5.4 6.6 7.0 7.5 22.0 hmax (m) 0.98 1.28 1.65 2.01 2.13 2.29 6.7

1996b). Moduli reduction curve (G/Gmax ) and damping ratio relationship have been obtained for prototype soil at the site of Capitol Aggregates as shown in Figure 13.3.

Jeremi et al.

Lecture Notes

397

Figure 13.3: Comparison of the Variation in Normalized Shear Modulus and Damping Ratio with Shearing Strain from the Resonant Column Tests with Modulus Reduction Curves proposed by Seed et al. (1986) and Darendeli (2001) (Kurtulus et al., 2005)

Jeremi et al.

Lecture Notes

398

Table 13.3: Element Size Determination after Degradation of G(vs ) (fmax = 10HZ ) Depth (f t) 0 1 2.5 7 14 21.5

2

38.5 half-space 2200 0.36 1320 13.20 The value is obtained from the lab test data, which corresponds to 0.2% strain level.

The degradation of dynamic soil properties as observed in experiments has to be considered in nite element analysis in order to capture more accurate behaviors. As the shear wave velocity correlates with shear modulus by Equation 13.2,

vshear =

(13.2)

we can readily obtain the dynamic degradation of wave velocities. This leads to smaller element size required for detailed simulation of wave propagation. The newly calculated element sizes are listed in Table 13.3. A three bent prototype nite element model has been developed with element size h = 0.6m as shown in Figure 13.4

13.2.1.3

As stated in Preisig (2005), the time step t used for numerically solving nonlinear vibration or wave propagation problems has to be limited for two reasons. The stability requirement depends on the numerical procedure in use and is usually formulated in the form t = Tn < value. Tn denotes the smallest fundamental period of the system. Similar to the spatial discretization Tn needs to be represented by about 10 time steps. While the accuracy requirement provides a measure on which higher modes of vibration are represented with sucient accuracy, the stability criterion needs to be satised for all modes. If the stability criterion is not satised for all modes of vibration, then the solution may diverge. In many cases it is necessary to provide an upper bound to the frequencies that are present in a system by including frequency dependent damping to the model. The second stability criterion results from the nature of the nite element method. As a wave front progresses in space it reaches one point after the other. If the time step in the nite element analysis is too large the wave front can reach two consecutive elements at the same moment. This would violate a fundamental property of wave propagation and can lead to instability. The time step therefore needs to be limited to t < h v (13.3)

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

399

Figure 13.4: Detailed Three Bent Prototype SFSI Finite Element Model - 3 Bent SFSI, 484,104 DOFs, 151,264 Elements, Frequency Cuto 10Hz, Element Size 0.3m, Minimum G/Gmax 0.08, Maximum Shear Strain 1%

According to Table 13.3, the time step requirement can be obtained as t < in seconds. h = 0.00256 v (13.4)

13.2.1.4

Domain reduction method was originally proposed in Bielak et al. (2003b); Yoshimura et al. (2003b) The theory aims at reducing the size of simulation domain by means of variable interchange. This method features a two-stage strategy for complicated three dimensional earthquake engineering simulations. The rst is an auxiliary problem that simulates the earthquake source and propagation path eects with a model that encompasses the source and a background structure from which the localized feature has been removed. The second problem models local site eects. Its input is a set of equivalent localized forces derived from the rst step. These forces act

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

400

only within a single layer of elements adjacent to the interface between the exterior region and the geological feature of interest. The beauty of this theory comes from the fact that we can use established numerical and/or experimental approaches to solve the rst-stage wave propagation problem. With the outcome of the rst phase solution, we greatly reduce the size of the problem and then eorts can be focused on the second phase to deliver more accurate simulation on local responses. This approach can be successfully used in soil-foundation-structureinteraction nite element modeling without the need to incorporate unnecessary far-eld motion simulations.

13.2.1.5

Structural Model

The nonlinear structure model developed in this work is a joint eort of UCB and UCD. Experimental data has been collected from UNR shaking table tests to calibrate the structural models. The eort in this work has been focused on how to integrate advanced structure model with geotechnical model to enable full-scale prototype simulations. The assumption that the plastic hinge forms either on the top of column or at the xed bottom does not hold for SFSI problems. This restriction has been removed as the geotechnical and structural models are connected together.

Figure 13.5: Simplied Hinge Model Developed for SFSI Prototype Simulations (Dryden, 2005)

Jeremi et al.

Lecture Notes

401

Jeremi et al.

Lecture Notes

402

Jeremi et al.

Lecture Notes

403

13.2.1.6

Simulation Scenarios

In order to fully investigate how the relative strength of soil to the structure can aect the system behaviors, a thorough parametric study has been performed. The prototype model used in this work is a 4-span bridge structure with 3 bents. The supporting soil foundation can be varied according to dierent site conditions. The mesh of the prototype nite element model is shown in the Figure 13.8. The underlying soil of the bridge can be soft bay mud or sti sand. In order to fully investigate the SFSI response, various scenarios are simulated as shown in Table 13.4 and results are analyzed.

Jeremi et al.

Lecture Notes

404

Jeremi et al.

Figure 13.8: Finite Element Model for 3 Bent Prototype Bridge System

Lecture Notes

405

Table 13.4: Simulation Scenarios for Prototype SFSI Studies Simulation Cases Case 1 Case 2 Case 3 Case 4 Case 5 Case 6 Case 7 Case 8 Soil Block 1 Sti Sand Sti Sand Sti Sand Sti Sand Soft Clay Soft Clay Soft Clay Soft Clay Soil Block 2 Sti Sand Sti Sand Soft Clay Soft Clay Sti Sand Sti Sand Soft Clay Soft Clay Soil Block 3 Sti Sand Soft Clay Sti Sand Soft Clay Sti Sand Soft Clay Sti Sand Soft Clay

13.2.1.6.1

For the sti sand, the constitutive model developed in previous sections, as shown in Table 13.1 will be used.

13.2.1.6.2

This soil model aims at simulating in-situ undrained behavior of soft bay mud. Undrained shear strength can be easily determined and a simple von Mises model is used in this research.

13.2.2

Starting from this section, detailed numerical simulation results will be presented to show how the nite element simulation techniques can be used in prototype earthquake simulations. The results are presented here, and discussions will follow. Figure 13.9 shows the input motion recorded from 1994 Northridge which contains lots of high frequency contents. There are totally two motions are selected for this work, one with primary short period (high frequency) contents, and the other with primary long period (low frequency) contents. The purpose is to study every single component of the SFSI system trying to expose how each aects the SFSI system response.

13.2.2.1

Input Motion

Jeremi et al.

Lecture Notes

406

Jeremi et al.

Lecture Notes

407

13.2.2.2

Displacement Response

This section presents the displacement results from numerical simulations. There are a couple of interesting things that deserve attention. Firstly, in the 1994 Northridge earthquake which contains much high frequency content, the structural response from softer soil actually is smaller than those on top of sti soil. This is interesting because it basically controdicts the common notion that stier the soil, stabler the structure. That is the case for static design. But for earthquake design, we are presenting dierent stories. Secondly, we see the soil displacement near structure is largely aected by the SFSI. So the question if it is valid to apply outcrop motion directly to xity point to excite the structure, just as people commonly do, might need a revisit.

Figure 13.10: Simulated Displacement Time Series, Northridge 1994, Century City, Comparison of Eight Cases with Free Field Motions (Soil Block 1)

Jeremi et al.

Lecture Notes

408

Figure 13.11: Simulated Displacement Time Series, Northridge 1994, Century City, Comparison of Eight Cases

(Structure Bent 1)

Figure 13.12: Simulated Displacement Time Series, Northridge 1994, Century City, Comparison of Eight Cases with Free Field Motions (Soil Block 2)

Jeremi et al.

Lecture Notes

409

Figure 13.13: Simulated Displacement Time Series, Northridge 1994, Century City, Comparison of Eight Cases

(Structure Bent 2)

Figure 13.14: Simulated Displacement Time Series, Northridge 1994, Century City, Comparison of Eight Cases with Free Field Motions (Soil Block 3)

Jeremi et al.

Lecture Notes

410

Figure 13.15: Simulated Displacement Time Series, Northridge 1994, Century City, Comparison of Eight Cases (Structure Bent 3)

Jeremi et al.

Lecture Notes

411

13.2.2.3

Acceleration Response

The acceleration results are shown in this section, which also supports the observation that stier soil might not necessarily enhance the stability of the structure. Acceleration time series consistently show that the sti soil will excite larger amplication for structures.

Figure 13.16: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Eight Cases with Free Field Motions (Soil Block 1)

Figure 13.17: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Eight Cases (Structure Bent 1)

Jeremi et al.

Lecture Notes

412

Figure 13.18: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Eight Cases

Figure 13.19: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Eight Cases (Structure Bent 2)

Jeremi et al.

Lecture Notes

413

Figure 13.20: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Eight Cases

Figure 13.21: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Eight Cases (Structure Bent 3)

Jeremi et al.

Lecture Notes

414

13.2.2.4

It will be very interesting to look into the frequency domain what is going on. The structure is always stier than underlying soils. So if the underlying soil is stronger, it implies the natural frequency of the sti soil will be closer to the structure on top of it. If, the input motion contains much high frequency content, it will directly excite the stier soil so the structure on top will receive very large amplication. This conclusion is supported by following plots.

Figure 13.22: Simulated Displacement Response Spectra, Northridge 1994, Century City, Comparison of Eight Cases with Free Field Motions (Soil Block 1)

Figure 13.23: Simulated Displacement Response Spectra, Northridge 1994, Century City, Comparison of Eight Cases (Structure Bent 1)

Jeremi et al.

Lecture Notes

415

Figure 13.24: Simulated Displacement Response Spectra, Northridge 1994, Century City, Comparison of Eight

Figure 13.25: Simulated Displacement Response Spectra, Northridge 1994, Century City, Comparison of Eight Cases (Structure Bent 2)

Jeremi et al.

Lecture Notes

416

Figure 13.26: Simulated Displacement Response Spectra, Northridge 1994, Century City, Comparison of Eight

Figure 13.27: Simulated Displacement Response Spectra, Northridge 1994, Century City, Comparison of Eight Cases (Structure Bent 3)

Jeremi et al.

Lecture Notes

417

13.2.2.5

The acceleration spectra have also been plotted to support the observation we made before. The consistent discovery is that stier soil will have amplication concentrated to the lower period side. If the input motion also have lower period contents, those will amplify the response the structure can see.

Figure 13.28: Simulated Acceleration Response Spectra, Northridge 1994, Century City, Comparison of Eight Cases with Free Field Motions (Soil Block 1)

Figure 13.29: Simulated Acceleration Response Spectra, Northridge 1994, Century City, Comparison of Eight Cases (Structure Bent 1)

Jeremi et al.

Lecture Notes

418

Figure 13.30: Simulated Acceleration Response Spectra, Northridge 1994, Century City, Comparison of Eight

Figure 13.31: Simulated Acceleration Response Spectra, Northridge 1994, Century City, Comparison of Eight Cases (Structure Bent 2)

Jeremi et al.

Lecture Notes

419

Figure 13.32: Simulated Acceleration Response Spectra, Northridge 1994, Century City, Comparison of Eight

Figure 13.33: Simulated Acceleration Response Spectra, Northridge 1994, Century City, Comparison of Eight Cases (Structure Bent 3)

Jeremi et al.

Lecture Notes

420

13.2.2.6

Structural Response

Finally we come to the point that we can see exactly how the structure responds to excitation mechanically. The moment time series shown here tells some important stories during dynamic shaking. Firstly, the structure on top of stier soil will yield much faster than those in soft soils. This makes perfect sense after the observations we made in previous sections. The input motion contains very similar frequency content as the sti soil so sti soil and the structure on top of it are excited much more than the soft-soil-structure system. Secondly, The structure on top of soft soil will see larger residual response than the stier soil. This exactly tells the story that the soft soil will respond much largely to the long period content of input motions which is much closer to the natural period of the soft soil.

Figure 13.34: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Eight Cases (Structure Bent 1 Pile 1)

Jeremi et al.

Lecture Notes

421

Figure 13.35: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Eight

Figure 13.36: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Eight Cases (Structure Bent 2 Pile 1)

Jeremi et al.

Lecture Notes

422

Figure 13.37: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Eight

Figure 13.38: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Eight Cases (Structure Bent 3 Pile 1)

Jeremi et al.

Lecture Notes

423

Figure 13.39: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Eight Cases (Structure Bent 3 Pile 2)

Jeremi et al.

Lecture Notes

424

13.2.3

It has been well observed that the characteristics of input ground motion also aects the SFSI system response. From the perspective of soil dynamics, stier soil will have shorted natural period and thus higher natural frequency. One can argue that the conclusion that has been established in Section 13.2.4.1.1 might exactly reect the case that for earthquake input motions containing much high frequency content, the sti soil will always receive much stronger shaking. The 1994 Northridge earthquake has been known to contain very high frequency component as shown in Figure 13.40.

So it will be also legitimate to question that if the ground motion has much long frequency (long period) content, the structure supported by soft soil might be the one to be exposed. In this work, the question has been studied using the exact nite element models that we created in the previous sections. The 1999 Turkey Kocaeli earthquake motion recorded at station Yarimca (YPT330) has been used as the target long period motion to study the SFSI behavior with dierent soil proles.

Jeremi et al.

Lecture Notes

425

Jeremi et al.

Lecture Notes

426

It has been shown in this work that long period (low frequency) motion will excites stronger response from softer clay soil layers which has a lower natural frequency as opposed to the conclusion we draw before in Section 13.2.4.1.1.

13.2.3.1

Displacement Response

This section is designed specically to expose the eect of input motion on SFSI system responses. From the pictures shown below, you can see that now the structure on top of soft soil will show much larger response from the shaking. The story behind is obvious to explain. Now the input motion from Turkey Kocaeli earthquake contains primary long period content, which is similar to the natural frequency content of the soft soil. During shaking, this underlying resonance excites the response of the whole SFSI to a larger degree. While on the other hand, the sti soil now is further away from the primary frequency of the input motion. The consequence is that now the sti soil will not see much excitation, neither will the structure on top of it.

Figure 13.42: Simulated Displacement Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two Cases with Free Field Motions (Soil Block 1)

Jeremi et al.

Lecture Notes

427

Figure 13.43: Simulated Displacement Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two Cases

(Structure Bent 1)

Figure 13.44: Simulated Displacement Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two Cases with Free Field Motions (Soil Block 2)

Jeremi et al.

Lecture Notes

428

Figure 13.45: Simulated Displacement Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two Cases

(Structure Bent 2)

Figure 13.46: Simulated Displacement Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two Cases with Free Field Motions (Soil Block 3)

Jeremi et al.

Lecture Notes

429

Figure 13.47: Simulated Displacement Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two Cases (Structure Bent 3)

Jeremi et al.

Lecture Notes

430

13.2.3.2

Acceleration Response

Acceleration plots might not show as clearly as those in displacement plots due to the reason that acceleration is derivative of displacement, so acceleration is more responsive to higher frequency contents than lower frequency contents. This eect actually can be observed from any recorded displacement and acceleration spectra records. The acceleration spectra will shift to the low period or high frequency side. This observation will be further explained in later sections when we discuss the moment time series.

Figure 13.48: Simulated Acceleration Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two Cases with Free Field Motions (Soil Block 1)

Figure 13.49:

Simulated Acceleration Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two Cases

(Structure Bent 1)

Jeremi et al.

Lecture Notes

431

Figure 13.50: Simulated Acceleration Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two Cases with

Figure 13.51:

Simulated Acceleration Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two Cases

(Structure Bent 2)

Jeremi et al.

Lecture Notes

432

Figure 13.52: Simulated Acceleration Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two Cases with

Figure 13.53:

Simulated Acceleration Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two Cases

(Structure Bent 3)

Jeremi et al.

Lecture Notes

433

13.2.3.3

The response spectra are useful tools to look deep into the frequency domain. As we explained before, the Turkey Kocaeli shaking was picked for this study because it contains much longer period contents that were not present in the 1999 Northridge earthquake. In this section, we can clearly see how the soil foundation can aect the structures on top of it. Firstly, structures supported by sti soil will only see amplication from those low period, high frequency components. Correspondingly, structures supported by soft soil tend to respond much more to long period, low frequency components of the shaking motion. Secondly, this plots exactly show how important the SFSI analysis is. It will not make much sense if one wants to analyze the structure without acknowledging the characteristics of the underlying soil foudation and the input motion. How the structure behaves is the combination of answers to many questions such as the stiness of soil and the frequency contents of the motions.

Figure 13.54:

Simulated Displacement Response Spectra, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Jeremi et al.

Lecture Notes

434

Figure 13.55:

Simulated Displacement Response Spectra, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Figure 13.56:

Simulated Displacement Response Spectra, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Jeremi et al.

Lecture Notes

435

Figure 13.57:

Simulated Displacement Response Spectra, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Figure 13.58:

Simulated Displacement Response Spectra, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Jeremi et al.

Lecture Notes

436

Figure 13.59:

Simulated Displacement Response Spectra, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Jeremi et al.

Lecture Notes

437

13.2.3.4

As we discussed before, the acceleration response shows much focused content to the high frequency (low period) side as acceleration is the derivative of displacement. But it is still clearly shows the dierence between the structures supported by sti soil and those supported by soft soil. SFSI is crucial in the sense that the structure response must be determined by both short period and long period components. The overall response is determined by the primary periods of both underlying soil and the input motion.

Figure 13.60:

Simulated Acceleration Response Spectra, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Figure 13.61:

Simulated Acceleration Response Spectra, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Jeremi et al.

Lecture Notes

438

Figure 13.62:

Simulated Acceleration Response Spectra, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Figure 13.63:

Simulated Acceleration Response Spectra, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Jeremi et al.

Lecture Notes

439

Figure 13.64:

Simulated Acceleration Response Spectra, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Figure 13.65:

Simulated Acceleration Response Spectra, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Jeremi et al.

Lecture Notes

440

13.2.3.5

Structural Response

The structure response now directly shows how dierently the structure can react to the same input motion, given dierent soil conditions. We can again see the structure on top of soft soil exhibits much response to long period content especially for the fact that the structure on top of the soft soil reaches maximum moment for about 3 seconds, which is not present in the sti soil case. This is also consistent with the observation that the displacement of the structure supported by soft soil is much larger due to the much longer time for plastic slip when the plastic moment is reached.

Figure 13.66:

Simulated Maximum Moment Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Figure 13.67:

Simulated Maximum Moment Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Jeremi et al.

Lecture Notes

441

Figure 13.68:

Simulated Maximum Moment Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Figure 13.69:

Simulated Maximum Moment Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Jeremi et al.

Lecture Notes

442

Figure 13.70:

Simulated Maximum Moment Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Figure 13.71:

Simulated Maximum Moment Time Series, Turkey Kocaeli 1999, Yarimca, Comparison of Two

Jeremi et al.

Lecture Notes

443

Table 13.5: Simulation Scenarios for Prototype ESSI Studies Simulation Cases Case 1 Case 8 Soil Block 1 Sti Sand Soft Clay Soil Block 2 Sti Sand Soft Clay Soil Block 3 Sti Sand Soft Clay

13.2.4

In the following sections we analyze various Earthquake Soil Structure Interaction (ESSI) eects.

13.2.4.1

It has been well known that stier soil layer will provide higher bearing capacity of structures so site improvements will always be preferred in engineering practice when one talks about foundation design. While for dynamic cases, this widely-held impression will not be valid anymore. In order to see how stiness of soil can aect the response of the whole ESSI system during earthquake shaking, two distinct scenarios listed in

Table 13.5, Case 1 with all sti soil foundations, and Case 8 with all soft soil foundations have been extracted to show the dynamic system response.

13.2.4.1.1

Figure 13.72: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Two Cases First 25s (Structure Bent 1)

Jeremi et al.

Lecture Notes

444

Figure 13.73: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Two Cases -

Figure 13.74: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Two Cases First 25s (Structure Bent 3)

Jeremi et al.

Lecture Notes

445

Figure 13.75: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Two

Figure 13.76: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Two Cases - First 25s (Structure Bent 2)

Jeremi et al.

Lecture Notes

446

Figure 13.77: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Two Cases - First 25s (Structure Bent 3)

Jeremi et al.

Lecture Notes

447

It has been shown in this research that if the soil layer is stier, structures will see much more amplication from earthquake shaking. Figures 13.72 to 13.74 show the acceleration response from ESSI systems with sti sand and soft clay, which clearly shows that sti sand delivers much stronger energy input to superstructures. The superstructures on top of stier soil also yield sooner than softer soil layers as shown in maximum moment time series Figures 13.75 to 13.77. This observation proves one interesting point that in order to improve structural stability, site improvement is not necessarily improving the dynamic resistance of the ESSI system. In later section of this work, further observation will be made to correlate this conclusion with the characteristics of the input motion. As for the soil side, near-structure soil motion will be also aected by ESSI. So the traditional way of assuming that recorded ground motion can be used as input motion to ESSI analysis should be revisited. According to Figures 13.78 13.79 13.80, we can see that sti soil also shows stronger surface motion records during earthquake shaking.

Figure 13.78: Simulated Displacement Time Series, Northridge 1994, Century City, Comparison of Two Cases First 25s (Soil Block 1)

Jeremi et al.

Lecture Notes

448

Figure 13.79: Simulated Displacement Time Series, Northridge 1994, Century City, Comparison of Two Cases -

Figure 13.80: Simulated Displacement Time Series, Northridge 1994, Century City, Comparison of Two Cases First 25s (Soil Block 3)

Jeremi et al.

Lecture Notes

449

Table 13.6: Simulation Scenarios for Prototype ESSI Studies Simulation Cases Case 3 Soil Block 1 Sti Sand Soil Block 2 Soft Clay Soil Block 3 Sti Sand

13.2.4.2

Bridge is always constructed over gulfs or bays. It is a common case that for multi-span bridge, dierent bents will inevitably sit on soil foundations with totally dierent strength. This site nonuniformity complicates design because individual structure response might be largely varied. In order to study the eects of site non-uniformity on the dynamic response of ESSI system, Case 3 listed in Table 13.6 has been selected as the test bed of our simulation. This scenario corresponds to the case that a 4-span bridge sits on solid abutments but with much softer bay-mud type foundation in the middle of a bay.

13.2.4.2.1

Case 3

Figure 13.81: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases First 25s (Structure Bent 1)

Jeremi et al.

Lecture Notes

450

Figure 13.82: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases -

Figure 13.83: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases First 25s (Structure Bent 3)

Jeremi et al.

Lecture Notes

451

Figure 13.84: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three

Figure 13.85: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three Cases - First 25s (Structure Bent 2)

Jeremi et al.

Lecture Notes

452

Figure 13.86: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three Cases - First 25s (Structure Bent 3)

Jeremi et al.

Lecture Notes

453

Table 13.7: Simulation Scenarios for Prototype ESSI Studies Simulation Cases Case 2 Case 5 Soil Block 1 Sti Sand Soft Clay Soil Block 2 Sti Sand Sti Sand Soil Block 3 Soft Clay Sti Sand

From acceleration Figures 13.81, 13.82, 13.83, and moment Figures 13.84, 13.85 and 13.86, we can clearly see that due to the presence of a soft soil block, all structures show smaller magnitude in response. This is especially true for the superstructure that directly sits on top of the softer soil foundation. For Case 3 listed in Table 13.6, the superstructure (bent 2) gets much smaller response because it is right on top of clay (bay mud) foundation. As a matter of fact, the middle bent (bent 2) does not yield at all. We want to extend this observation to other cases as listed in Table 13.7. Case 2 shows the scenario that the soil foundation supporting bent 3 is soft bay mud, while Case 5 shows it for bent 1. The same reasoning can be applied to these similar cases. Figures 13.87 to 13.92 show the results of Case 2 and Figures 13.93 to 13.98 for Case 5.

13.2.4.2.2

Case 2

Figure 13.87: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases First 25s (Structure Bent 1)

Jeremi et al.

Lecture Notes

454

Figure 13.88: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases -

Figure 13.89: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases First 25s (Structure Bent 3)

Jeremi et al.

Lecture Notes

455

Figure 13.90: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three

Figure 13.91: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three Cases - First 25s (Structure Bent 2)

Jeremi et al.

Lecture Notes

456

Figure 13.92: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three Cases - First 25s (Structure Bent 3)

Jeremi et al.

Lecture Notes

457

13.2.4.2.3

Case 5

Figure 13.93: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases First 25s (Structure Bent 1)

Figure 13.94: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases First 25s (Structure Bent 2)

Jeremi et al.

Lecture Notes

458

Figure 13.95: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases -

Figure 13.96: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three Cases - First 25s (Structure Bent 1)

Jeremi et al.

Lecture Notes

459

Figure 13.97: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three

Figure 13.98: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three Cases - First 25s (Structure Bent 3)

Jeremi et al.

Lecture Notes

460

Table 13.8: Simulation Scenarios for Prototype ESSI Studies Simulation Cases Case 4 Case 6 Case 7 Soil Block 1 Sti Sand Soft Clay Soft Clay Soil Block 2 Soft Clay Sti Sand Soft Clay Soil Block 3 Soft Clay Soft Clay Sti Sand

As a nal conclusion, if there exists one substantially weaker soil layer for the ESSI system, the dynamic response for the whole ESSI system will be attenuated. The structure that directly sits on top of the soft soil block will receive signicantly smaller excitation during earthquake shakings. The other side of story can be also formulated using results obtained in this work. What if there is a substantially stronger soil foundation in the ESSI system? It would be worthwhile exploring the other side of reasoning. In this work, three other cases have been used to observe how the ESSI system will behave for the cases that one soil block is much stronger that the other blocks as listed in Table 13.8.

13.2.4.2.4

Case 4

Figure 13.99: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases First 25s (Structure Bent 1)

Jeremi et al.

Lecture Notes

461

Figure 13.100: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases

Figure 13.101: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases - First 25s (Structure Bent 3)

Jeremi et al.

Lecture Notes

462

Figure 13.102: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three

Figure 13.103: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three Cases - First 25s (Structure Bent 2)

Jeremi et al.

Lecture Notes

463

Figure 13.104: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three Cases - First 25s (Structure Bent 3)

Jeremi et al.

Lecture Notes

464

It can easily be seen from Figures 13.99 to 13.104 that the structure on top of the strong soil block will exhibit much bigger response and while the other structures sitting on weaker soil blocks will not get much shaking at all.

13.2.4.2.5

Case 6

Figure 13.105: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases - First 25s (Structure Bent 1)

Figure 13.106: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases - First 25s (Structure Bent 2)

Jeremi et al.

Lecture Notes

465

Figure 13.107: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases

Figure 13.108: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three Cases - First 25s (Structure Bent 1)

Jeremi et al.

Lecture Notes

466

Figure 13.109: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three

Figure 13.110: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three Cases - First 25s (Structure Bent 3)

Jeremi et al.

Lecture Notes

467

13.2.4.2.6

Case 7

Figure 13.111: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases - First 25s (Structure Bent 1)

Figure 13.112: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases - First 25s (Structure Bent 2)

Jeremi et al.

Lecture Notes

468

Figure 13.113: Simulated Acceleration Time Series, Northridge 1994, Century City, Comparison of Three Cases

Figure 13.114: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three Cases - First 25s (Structure Bent 1)

Jeremi et al.

Lecture Notes

469

Figure 13.115: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three

Figure 13.116: Simulated Maximum Moment Time Series, Northridge 1994, Century City, Comparison of Three Cases - First 25s (Structure Bent 3)

Jeremi et al.

Lecture Notes

470

As a conclusion, dynamic behaviors of the ESSI system on non-uniform sites can be much more complicated than the uniform case. Generally speaking, the existence of soft soil foundation will attenuate the dynamic response of the whole ESSI system. The structure supported by soft soil block will see far less shakings than the ones supported by sti soil. From the design perspective, the structure on top of strong soil foundations should be designed to higher safety with larger margin.

13.2.4.3

All the results have been shown in previous section, the purpose is to investigate how the whole ESSI system will respond to excitations with dierent predominant frequency. From Figure 13.117 and 13.118, we can clearly see the dierence in frequency contents of the two ground motions studied. THe 1999 Turkey Kocaeli earthquake contains more long period components that are not present at the 1994 Northridge site.

Jeremi et al.

Lecture Notes

471

Jeremi et al.

Lecture Notes

472

Depending on dierent site conditions, ESSI system will exhibit varied response to excitations with dierent predomiant frequencies. As we can see from Figure 13.42 to Figure 13.47, the structure sitting on softer soil foundations now exhibits much larger structural response when it is subject to long period excitations. This conclusion has been supported by every single plot that records the response spetra of structures such as Figure 13.61, Figure 13.63 and Figure 13.65. For the long perdominant periods present in the Kocaeli earthquake, the structures supported by soft clays exhibit much larger amplication than those supported by sti soils. It is also worthwhile to look deeper into the transfer function that tells some other aspects of the simulation scenarios. We can see from Figure 13.119 to Figure 13.130 that in terms of transfer function, which denes the amplication from soil surface to top of the structure, the structure is able to pick up much high frequency content that is not present within the soil layers. The visual representation of this is that the transfer function has much larger value at the high frequency end (low period end). This can be explained that typically people construct pile foundation inside the sti soil (or bedrock). In our simulation, we also have a very sti sustaining layer to provide pile tip resistance. The results shown in transfer functions make sense and also veries our numerical model. Acceleration results are also shown as following.

Jeremi et al.

Lecture Notes

473

Figure 13.119: Transfer Function of Simulated Displacement Time Series for Both Long Period and Short Period Motions (Structure Bent 1, from Soil Surface to Top of Bent)

Figure 13.120: Zoomed View: Transfer Function of Simulated Displacement Time Series for Both Long Period and Short Period Motions (Structure Bent 1, from Soil Surface to Top of Bent)

Jeremi et al.

Lecture Notes

474

Figure 13.121: Transfer Function of Simulated Displacement Time Series for Both Long Period and Short Period Motions (Structure Bent 2, from Soil Surface to Top of Bent)

Figure 13.122: Zoomed View: Transfer Function of Simulated Displacement Time Series for Both Long Period and Short Period Motions (Structure Bent 2, from Soil Surface to Top of Bent)

Jeremi et al.

Lecture Notes

475

Figure 13.123: Transfer Function of Simulated Displacement Time Series for Both Long Period and Short Period Motions (Structure Bent 3, from Soil Surface to Top of Bent)

Figure 13.124: Zoomed View: Transfer Function of Simulated Displacement Time Series for Both Long Period and Short Period Motions (Structure Bent 3, from Soil Surface to Top of Bent)

Jeremi et al.

Lecture Notes

476

Figure 13.125: Transfer Function of Simulated Acceleration Time Series for Both Long Period and Short Period Motions (Structure Bent 1, from Soil Surface to Top of Bent)

Figure 13.126: Zoomed View: Transfer Function of Simulated Acceleration Time Series for Both Long Period and Short Period Motions (Structure Bent 1, from Soil Surface to Top of Bent)

Jeremi et al.

Lecture Notes

477

Figure 13.127: Transfer Function of Simulated Acceleration Time Series for Both Long Period and Short Period Motions (Structure Bent 2, from Soil Surface to Top of Bent)

Figure 13.128: Zoomed View: Transfer Function of Simulated Acceleration Time Series for Both Long Period and Short Period Motions (Structure Bent 2, from Soil Surface to Top of Bent)

Jeremi et al.

Lecture Notes

478

Figure 13.129: Transfer Function of Simulated Acceleration Time Series for Both Long Period and Short Period Motions (Structure Bent 3, from Soil Surface to Top of Bent)

Figure 13.130: Zoomed View: Transfer Function of Simulated Acceleration Time Series for Both Long Period and Short Period Motions (Structure Bent 3, from Soil Surface to Top of Bent)

Jeremi et al.

Chapter 14

(2010-2011-2012)

479

Lecture Notes

480

14.1

Literature review

Almost in all the soil-structure systems, there are interfaces between dierent materials which are in contact. These contact areas can get detached during the loading so there will be gaps in the model. The focus of this section is on the contact of concrete and soil. In fact transferring the loads to structure is dened based on the contact areas which will aect eh response of the soil-structure systems (Kikuchi and Oden (1988)). The main objective at the contact areas is to dene normal and shear forces, and displacements at the common surfaces. The idea of contact element has become from elements so called joint elements which was initially developed for modeling of rock joints. Typically normal and tangential stiness were used to model the pressure and friction at the interface (Sheng et al. (2007b); Haraldsson and Wriggers (2000); Desai and Siriwardane (1984)). The study of two dimensional and axisymmetric benchmark examples have been done by Olukoko et al. (1993) for linear elastic and isotropic contact problems. Study was done considering Coulombs law for frictional behavior at the interface. In many cases the interaction of soil and structure is involved with frictional sliding of the contact surfaces,

separation, and reclosure of the surfaces. These cases depend on the loading procedure and frictional parameters. Sheng et al. (2007b) discussed how frictional contact is important for structural foundations under loading, pile foundations, soil anchors, and retaining walls. Khoei and Nikbakhti (2006) used the extended nite element method (X-FEM) for modeling of frictional contact problems. In order to be able to simulate discontinuities in model during the analysis without considering the boundary conditions of domain, special functions should be included within standard nite element framework. A nite element method for analysis of frictional contact problems was developed by Chandrasekaran et al. (1987). The nonlinear problem is solved by imposing geometric constraints on equilibrium conguration which compatibility conditions are violated. Dierent conditions happening at contact area, such as sticking and slipping, are determined from the relative magnitudes of the normal and tangential nodal forces. Developing contact nite element for analysis of connections between laminated composite plates has been done by Barbero et al. (1995). Such element is compatible with a three-dimensional plate element based on constant shear theory. The contact element formulation is based on a simple regularization of the unilateral contact with a Coulomb friction problem. Two-dimensional frictional polynomial to segment contact elements are developed by Haraldsson and Wriggers (2000) based on non-associated frictional law and elastic-plastic tangential slip decomposition. Several benchmarks are presented by Konter (2005) in order to verify the the results of the nite element analyses performed on 2D and 3D modelings. In all proposed benchmarks the results were approximated pretty well with a 2D or an axisymmetric solutions. In addition, 3D analyses were performed and the results were compared with the 2D solutions. Hjiaj et al. (2004) presented an algorithm for solving frictional contact problems where sliding rule is nonassociated. The algorithm is a combination of the classical unilateral contact law and an anisotropic friction

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

481

model with a non-associated slip rule. It has been observed that slip rule has a strong inuence on the frictional behavior. In following sections, the sliding behavior of earthquake-soil-foundation-structure systems is studied. There are zones exist at the interface of foundation and soil which are in contact. These zones should be modeled appropriately by using contact elements and friction law. This will let the interface zones to slide on each other based on the dened contact rules instead of modeling the contact zones to be in touch with each other during the whole analysis. In seismic modeling of soil-structure systems there are aspects such as the ow of input seismic energy through the system, propagating, and dissipation mechanisms to be considered. In soil-structure interaction modeling, the response of the system is controlled by interaction of dynamic properties of structure, foundation, soil or rock, as well as the ground motion. There are benets for designing process if there is an understanding in numerical simulations regarding where and when energy dissipation occurs in the system. Therefore the possible place of damage or unacceptable performance can be evaluated. In other words, knowing the location and timing of energy dissipation in soil-structure system can determine the damage happens during seismic event to structure. In case modeling the slipping behavior, parts of the foundation and soil layer might get separated during the

shaking and gap openings might be created. This can model the rocking of the foundations which will aect the movement of the structural elements. In addition, through the slipping behavior, part of the seismic energy can be dissipated due to friction. In following, nite element models used for simulations are shown. Seismic waves are propagated considering both cases of 1D wave propagation assumption and also the full 3D wave propagation. for seismic input motion, both earthquake ground motion and articial wavelet are used. Morgan Hill earthquake (1984) recorded at Coyote Lake Dam station is used along with Ricker wavelet as input motions for dierent cases. Using the Ricker wavelet helps since it is less complicated than the actual earthquake ground motion in terms of frequency content and they are shorter in time.

Jeremi et al.

Lecture Notes

482

14.2

The 3D nite element model is shown in Figure (14.1). The nite element model is created using both 8 node brick elements for modeling the soil part and 27 node brick elements for modeling the foundation part, and displacement beam element for structural components. The size of the model is 140m along the X and Y directions, and 50m along the Z direction (the height of the soil layer). Shear wave velocity of the soil is considered to be 700m/s with density of 2200kg/m3 . Foundation has the height of 5m embedded in the soil layer, with size of 90m in each horizontal direction. Concrete is chosen to have shear wave velocity of 2000m/s with density of 2400kg/m3 . The size of the elements are 5m in each direction for both foundation and soil elements. Structural beam is composed of 12 displacement beam elements attached to each other with dierent stiness and mass properties. Domain Reduction Method is used to apply the input eective forces. Both 1D and 3D wave propagation cases are using the same nite element model for second stage of DRM analysis. Frictional contact elements are placed at the interface of foundation and soil layer. The contact element used here has the same normal and tangential stiness with magnitude of 108 N/m. Friction ratio and cohesion of the contact element are = 0.4 and c = 0.0 respectively. Contact elements are oriented along the Z axis of global model. In general it is accepted that in order to represent a traveling wave of a given frequency, the size of the nite elements have to have about 10 nodes per wavelength (Hughes (1987); Argyris and Mlejnek (1991b)). Using fewer than 10 nodes per wave length specially for linear elements leads to numerical damping of higher frequencies as such element discretization misses certain frequencies of the wave. In order to satisfy this requirement, the size of the mesh should satisfy Equation (14.1) which is a function of the maximum frequency (fmax ) of the input motion and also the shear wave velocity of the media Vs = of the soil. G/ where G is the shear modulus and is density

(14.1)

In this model, the size of the elements are chosen to be 5m in each direction. Considering that the shear wave velocity of the soil is 700m/s, the maximum frequency able to be propagated through this model is 14Hz .

Jeremi et al.

Lecture Notes

483

Figure 14.1: 3D nite element SFSI model considering slipping behavior at the interface of foundation and soil layer

Jeremi et al.

Lecture Notes

484

14.3

In this section, slipping behavior of SFSI systems is studied under 1D seismic wave propagation assumption to come up with the input motion to be used for the model prepared for second stage of DRM. Morgan Hill earthquake and Ricker wavelets are used as the source of motion. Domain Reduction Method is used to apply the input eective forces. In order to obtain the displacement and acceleration time histories used for calculating the DRM layer eective forces, a nite element soil column with shear behavior is considered representing the 1D wave propagation. The motions are applied at the base of the 1D nite element model and propagated through the soil column using multiple support excitation pattern.

Jeremi et al.

Lecture Notes

485

14.3.1

Figures (14.2) to (14.3) show the acceleration and displacement time histories and FFT of the Morgan Hill earthquake at the base of the 1D model. Figure (14.4) shows the acceleration time histories recorded at the bottom and top of the structure in X and Z directions. As it is observed, the acceleration time history of slipping and no-slipping model has dierences in amplitude and the phase in X direction of the model. However, the dierences in phases of the recorded motions seems to be more along the Z direction. It shows that considering the slipping behavior has caused a lag in phase of the response. Comparison of displacement time histories of the same models is shown in Figure (14.5) along X and Z directions. As it is shown, displacement time histories along X direction are quite similar in this case while along Z direction they have dierent amplitudes due to initial settlement and dierent phases. FFT of the accelerations do not have much of a dierence in terms of frequency content (shown in Figure (14.6) but they are dierent in amplitude which has been observed in acceleration time histories as well. In general it is observed that in this earthquake the the magnitude of accelerations and displacements are less at top of the structure comparing to the ones recorded at the bottom of the structure for both slipping and no-slipping behavior cases.

Figure (14.7) shows the distribution of sliding at the interface of foundation and the soil layer at 9 dierent time steps of analysis from 0.5 to 1.3 seconds. As it is observed, location of maximum sliding is changed along the foundation in time while magnitude and direction of the applied motion changes. Slipping happens at the specic steps and parts of the interface zone in which the applied force is more than the resistant one.

Jeremi et al.

Lecture Notes

486

2 1.5

0.0025 0.002

Acceleration (m/s2)

Displacement (m)

Time (s)

Time (s)

Figure 14.2: Acceleration and displacement time histories of Morgan Hill earthquake

0.012 0.01

FFT Amplitude

Frequency [Hz]

Jeremi et al.

Lecture Notes

487

Slip Behavior No-slip Behavior

Slip Behavior No-slip Behavior

Acceleration (m/s2)

Acceleration (m/s2)

Time (s)

Time (s)

Slip Behavior No-slip Behavior

Slip Behavior No-slip Behavior

Acceleration (m/s2)

Acceleration (m/s2)

Time (s)

Time (s)

Figure 14.4: Comparison of acceleration time histories of the structure between slipping and no-slipping models for Morgan Hill earthquake

Jeremi et al.

Lecture Notes

488

Slip Behavior No-slip Behavior

Slip Behavior No-slip Behavior

0.03

0.03

Displacement (m)

Displacement (m)

0.02

Time (s)

Time (s)

Slip Behavior No-slip Behavior

Slip Behavior No-slip Behavior

0.03

0.03

Displacement (m)

Displacement (m)

0.02

Time (s)

Time (s)

Figure 14.5: Comparison of displacement time histories of the structure between slipping and no-slipping models for Morgan Hill earthquake

Jeremi et al.

Lecture Notes

489

0.8

Slip Behavior No-slip Behavior

0.8

Slip Behavior No-slip Behavior

0.7

Frequency (Hz)

Frequency (Hz)

0.8

Slip Behavior No-slip Behavior

0.8

Slip Behavior No-slip Behavior

0.7

Frequency (Hz)

Frequency (Hz)

Figure 14.6: Comparison of FFT of the acceleration of the structure between slipping and no-slipping models for Morgan Hill earthquake

Jeremi et al.

Lecture Notes

490

0.5s

110 100

0.002 0.0025

0.6s

110 100

0.002 0.0025

0.7s

110 100

0.002 0.0025

90

90

90

0.0015

Y (m)

Y (m)

0.0015

Y (m)

80 70 60 50 40 30

80 70 60 50 40

80 70 60 50 40

0.0015

0.001

0.001

0.001

0.0005

0.0005

0.0005

30 40 50 60 70 80 90 100 110

30

30 40 50 60 70 80 90 100 110

30

30 40 50 60 70 80 90 100 110

X (m)

X (m)

X (m)

0.8s

110

0.0025

0.9s

110 100

0.002 0.002 0.0025

1.0s

110 100

0.002 0.0025

100 90

90

90

0.0015

Y (m)

Y (m)

0.0015

Y (m)

80 70 60 50 40 30

80 70 60 50 40

80 70 60 50 40

0.0015

0.001

0.001

0.001

0.0005

0.0005

0.0005

30 40 50 60 70 80 90 100 110

30

30 40 50 60 70 80 90 100 110

30

30 40 50 60 70 80 90 100 110

X (m)

X (m)

X (m)

1.1s

110 100

0.002 0.0025

1.2s

110 100

0.002 0.0025

1.3s

110 100

0.002 0.0025

90

90

90

0.0015

Y (m)

Y (m)

0.0015

Y (m)

80 70 60 50 40 30

80 70 60 50 40

80 70 60 50 40

0.0015

0.001

0.001

0.001

0.0005

0.0005

0.0005

30 40 50 60 70 80 90 100 110

30

30 40 50 60 70 80 90 100 110

30

30 40 50 60 70 80 90 100 110

X (m)

X (m)

X (m)

Figure 14.7: Distribution of sliding along the contact interface for Morgan Hill earthquake (gray scale given in meters)

Jeremi et al.

Lecture Notes

491

14.3.2

Ricker wave

In order to investigate more, the Ricker wave with dominant frequency of 1Hz is used for analysis of the same model. The maximum of this function happens at 1 second. Figures (14.8) and (14.9) show the acceleration and displacement time histories as well as the FFT of the Ricker wave respectively. As shown in Figure (14.10), displacement time history of top of the structure along X direction is higher when slipping behavior is not considered, while along Z direction it is higher considering the slipping behavior. Since structural components have rotational degrees of freedom, the slipping behavior and gaps in the model will lead to have more rotational movement and rocking of the foundation. There is also a phase lag in response of the structure in slipping behavior model. The displacement time histories at the bottom of the structure are slightly dierent in X direction. It can also be observed that along Z direction, the response of the structure in slipping model is higher comparing to the no-slipping one. In this analysis slipping behavior make the movement to be less along the X direction. This might be due to the fact that while the foundation and structure are moving, there are gaps created at some parts of the interface which will aect the rocking movement of the foundation.

Figure (14.11) shows the FFT of the acceleration at the bottom and top of the structure. As it is observed, there is a slight shift in predominant frequency of the response between the slipping and no-slipping behavior. By considering the slipping behavior, the dominant frequency of the motion is decreased along the X direction while is increased along the Z direction. This means that the system gets softer along X direction and stier along Z direction. This is due to the gap openings and slidings at the foundation and soil layer interface. Since sliding happens along the X direction, and also considering the gap openings occur along the Z direction which can lead to the rocking of the foundation, it makes the system softer along the X direction. In general this shows the fact that the natural frequency of the system can be changed due to the slipping behavior. Figures (14.12) and (14.13) show the distribution of gap openings and slidings at the interface of foundation and the soil layer at 9 dierent time steps of analysis from 0.5 to 1.3 seconds. It can be observed from Figure (14.12) how the location of gap openings are changed on dierent parts of the contact interface as the dynamic motion is applied. The location of gap openings are changed from one side to the other while the magnitude and direction of the applied motion changes. The maximum of the gap openings at these time steps happen at 1.3 second with maximum amount of 0.1m on the right side of the interface. Maximum sliding at the shown range of time steps happens at 1.0 second with maximum value of 0.07m on the left side of the interface.

Jeremi et al.

Lecture Notes

492

6 5

0.06 0.04

Acceleration (m/s2)

Displacement (m)

0 0.5 1 1.5 2 2.5 3 3.5 4

4 3 2 1 0 -1 -2 -3 -4

0.02 0 -0.02 -0.04 -0.06 -0.08 -0.1 0 0.5 1 1.5 2 2.5 3 3.5 4

Time (s)

Time (s)

Figure 14.8: Acceleration and displacement time histories of Ricker wave with dominant frequency of 1Hz

0.00016

Frequency (Hz)

Jeremi et al.

Lecture Notes

493

0.6 0.4 0.2 0 -0.2 -0.4 -0.6 0 0.5 1 1.5 2 2.5 3 3.5 4

Slip Behavior No-slip Behavior

0.6 0.4 0.2 0 -0.2 -0.4 -0.6 0 0.5 1 1.5 2 2.5 3 3.5 4

Slip Behavior No-slip Behavior

Displacement (m)

Time (s)

Displacement (m)

Time (s)

0.6 0.4 0.2 0 -0.2 -0.4 -0.6 0 0.5 1 1.5 2 2.5 3 3.5 4

Slip Behavior No-slip Behavior

0.6 0.4 0.2 0 -0.2 -0.4 -0.6 0 0.5 1 1.5 2 2.5 3 3.5 4

Slip Behavior No-slip Behavior

Displacement (m)

Time (s)

Displacement (m)

Time (s)

Figure 14.10: Comparison of displacement time histories of the structure between slipping and no-slipping models for Ricker wave

Jeremi et al.

Lecture Notes

494

8

Slip Behavior No-slip Behavior

8

Slip Behavior No-slip Behavior

6 5 4 3 2 1 0 0 5 10 15 20

7 6 5 4 3 2 1 0 0 5 10 15 20

Frequency (Hz)

Frequency (Hz)

8

Slip Behavior No-slip Behavior

8

Slip Behavior No-slip Behavior

6 5 4 3 2 1 0 0 5 10 15 20

7 6 5 4 3 2 1 0 0 5 10 15 20

Frequency (Hz)

Frequency (Hz)

Figure 14.11: Comparison of FFT of the acceleration of the structure between slipping and no-slipping models for Ricker wave

Jeremi et al.

Lecture Notes

495

0.5s

110 100

0.08 0.1

0.6s

110 100

0.08 0.1

0.7s

110 100

0.08 0.1

90

90

90

0.06

Y (m)

Y (m)

0.06

Y (m)

80 70 60 50 40 30

80 70 60 50 40

80 70 60 50 40

0.06

0.04

0.04

0.04

0.02

0.02

0.02

30 40 50 60 70 80 90 100 110

30

30 40 50 60 70 80 90 100 110

30

30 40 50 60 70 80 90 100 110

X (m)

X (m)

X (m)

0.8s

110

0.1

0.9s

110 100

0.08 0.08 0.1

1.0s

110 100

0.08 0.1

100 90

90

90

0.06

Y (m)

Y (m)

0.06

Y (m)

80 70 60 50 40 30

80 70 60 50 40

80 70 60 50 40

0.06

0.04

0.04

0.04

0.02

0.02

0.02

30 40 50 60 70 80 90 100 110

30

30 40 50 60 70 80 90 100 110

30

30 40 50 60 70 80 90 100 110

X (m)

X (m)

X (m)

1.1s

110 100

0.08 0.1

1.2s

110 100

0.08 0.1

1.3s

110 100

0.08 0.1

90

90

90

0.06

Y (m)

Y (m)

0.06

Y (m)

80 70 60 50 40 30

80 70 60 50 40

80 70 60 50 40

0.06

0.04

0.04

0.04

0.02

0.02

0.02

30 40 50 60 70 80 90 100 110

30

30 40 50 60 70 80 90 100 110

30

30 40 50 60 70 80 90 100 110

X (m)

X (m)

X (m)

Figure 14.12: Distribution of gap openings along the contact interface for Ricker wave (gray scale given in meters)

Jeremi et al.

Lecture Notes

496

0.5s

110 100 90

0.07 0.06 0.05

0.6s

110 100 90

0.07 0.06 0.05 0.04

0.7s

110 100 90

0.07 0.06 0.05 0.04

Y (m)

Y (m)

0.04

70

0.03

70

0.03

Y (m)

80

80

80 70

0.03

60 50 40 30 30 40 50 60 70 80 90 100 110

0.02 0.01 0

60 50 40 30 30 40 50 60 70 80 90 100 110

0.02 0.01 0

60 50 40 30 30 40 50 60 70 80 90 100 110

0.02 0.01 0

X (m)

X (m)

X (m)

0.8s

110

0.07 0.06 0.05

0.9s

110 100 90

0.07 0.06 0.05 0.04

1.0s

110 100 90

0.07 0.06 0.05 0.04

100 90

Y (m)

Y (m)

0.04

70

0.03

70

0.03

Y (m)

80

80

80 70

0.03

60 50 40 30 30 40 50 60 70 80 90 100 110

0.02 0.01 0

60 50 40 30 30 40 50 60 70 80 90 100 110

0.02 0.01 0

60 50 40 30 30 40 50 60 70 80 90 100 110

0.02 0.01 0

X (m)

X (m)

X (m)

1.1s

110 100 90

0.07 0.06 0.05

1.2s

110 100 90

0.07 0.06 0.05 0.04

1.3s

110 100 90

0.07 0.06 0.05 0.04

Y (m)

Y (m)

0.04

70

0.03

70

0.03

Y (m)

80

80

80 70

0.03

60 50 40 30 30 40 50 60 70 80 90 100 110

0.02 0.01 0

60 50 40 30 30 40 50 60 70 80 90 100 110

0.02 0.01 0

60 50 40 30 30 40 50 60 70 80 90 100 110

0.02 0.01 0

X (m)

X (m)

X (m)

Figure 14.13: Distribution of sliding along the contact interface for Ricker wave (gray scale given in meters)

Jeremi et al.

Lecture Notes

497

14.4

In this section, the same nite element SFSI model as previous section is used with 3D wave propagated motions as input motions for simulations. The input motion used here is Ricker wave (as shown in Figure (14.8)). In order to study the slipping behavior of SFSI considering 3D wave propagation for rst stage of DRM simulation, a nite element model with dimensions of 10000m 50m 5000m is considered. Two cases are studied here with the source of motion (fault) to be located at (x = 3000m, y = 0, z = 3000m) and (x = 3000m, y = 0, z = 3000m). Figures (14.14) and (14.15) show these two cases respectively. The size of the elements is chosen to be 50m in all directions for both cases in order to reduce the computational time. The soil parameters are: shear wave velocity of 700m/s and density of 2200kg/m3 . Analyses for the fault slip model are done by applying the motion at the nodes of one element. This is done in order to represent the the wave propagation starting from the fault using multiple support excitation pattern. This is representing the rst stage of analysis of DRM in which a big model including the fault is considered for free eld case in order to obtain the required motions for DRM layer.

Jeremi et al.

Lecture Notes

498

0m 5000m 2000m

5000m

10000m

Fault

2000m

0m

Figure 14.14: Domain to be analyzed for the 1st stage of DRM with fault located at an angle of 45 with respect to the top middle point of the model

0m 5000m 2000m

5000m

10000m

Fault

3000m

0m

Figure 14.15: Domain to be analyzed for the 1st stage of DRM with fault located at an angle of 34 with respect to the top middle point of the model

Jeremi et al.

Lecture Notes

499

14.4.1

Ricker wave, with fault located at 45 towards the top middle point of the model

For the rst case to be studied here, Ricker wavelet is considered as an input motion with dominant frequency of 1Hz . The fault is located with angle of 45 toward the top middle point of the model (Figure (14.14)). Displacement time histories of the structure along X and Z directions are shown in Figure (14.16). It is observed that along the X direction, the amplitude of displacement is little less at the beginning of the shaking of the slipping behavior model and a slight phase lag can be observed as well. However, when the actual pick of the motion is gone, it is shown that the motion for no-slipping model will damp out sooner. So the magnitude of displacements are higher at this time range for slipping model which could be because of the gap/sliding at the interface zone. The displacement time histories do not have a signicant dierence at the bottom of the structure. Displacement time histories observed along the Z direction shows the fact that magnitude of displacements for slipping model is higher and will damp out later comparing to the no-slipping model. Acceleration time histories along X and Z directions at the bottom and top of the structure are shown in Figure (14.17). It is observed that along the X direction the amplitude of acceleration is less considering the slipping behavior and also there is a phase lag in the observed motion. However, the amplitude of the acceleration along Z direction is much higher in case of considering the slipping case. This is the same behavior observed in 1D wave propagation as well but with higher dierence in time histories amplitudes. As shown in Figure (14.18), the dominant frequency of the response in case of slipping is less than the one observed in no-slipping case along X direction while it is higher along Z direction. Figures (14.19) and (14.20) show the distribution of sliding at the interface of foundation and the soil layer at 9 dierent time steps of analysis from 4.5 to 5.3 seconds. It can be observed from Figure (14.19) that maximum gap opening of 0.12m is occured at 4.7 seconds while the location of the openings are changed during the analysis. In addition, maximum sliding at the interface zone in this case happens at 4.8 seconds with magnitude of 0.03m. In both gap and slide distribution plots, it can be observed that the place of maximum is close to the middle of the foundation which is where the structure is located. Distribution of cumulative dissipated energy due to sliding of the foundation and soil layer contact zone is shown in Figure (14.21). By modeling the slipping behavior at the interface zone, part of the seismic energy is dissipated through the sliding and rocking of the foundation and therefore, less amount will be transferred to the structural components. Figure (14.22) shows how energy can be dissipated during the analysis for the point at the middle of the interface zone (location of the structure). When sliding happens, some part of the energy is dissipated as shown while there will be no change in dissipated energy if the foundation and soil are sticking to each other.

Jeremi et al.

Lecture Notes

500

0.6 0.4 0.2 0 -0.2 -0.4 -0.6 0 1 2 3 4 5 6 7 8

Slip Behavior No-slip Behavior

0.6 0.4 0.2 0 -0.2 -0.4 -0.6 0 1 2 3 4 5 6 7 8

Slip Behavior No-slip Behavior

Displacement (m)

Time (s)

Displacement (m)

Time (s)

0.6 0.4 0.2 0 -0.2 -0.4 -0.6 0 1 2 3 4 5 6 7 8

Slip Behavior No-slip Behavior

0.6 0.4 0.2 0 -0.2 -0.4 -0.6 0 1 2 3 4 5 6 7 8

Slip Behavior No-slip Behavior

Displacement (m)

Time (s)

Displacement (m)

Time (s)

Figure 14.16: Comparison of displacement time histories of the structure between slipping and no-slipping models for Ricker wave

Jeremi et al.

Lecture Notes

501

Slip Behavior No-slip Behavior

Slip Behavior No-slip Behavior

40

40

Acceleration (m/s2)

20 0 -20 -40 0 1 2 3 4 5 6 7 8

Acceleration (m/s2)

20 0 -20 -40 0 1 2 3 4 5 6 7 8

Time (s)

Time (s)

Slip Behavior No-slip Behavior

Slip Behavior No-slip Behavior

40

40

Acceleration (m/s2)

20 0 -20 -40 0 1 2 3 4 5 6 7 8

Acceleration (m/s2)

20 0 -20 -40 0 1 2 3 4 5 6 7 8

Time (s)

Time (s)

Figure 14.17: Comparison of acceleration time histories of the structure between slipping and no-slipping models for Ricker wave

Jeremi et al.

Lecture Notes

502

6

Slip Behavior No-slip Behavior

6

Slip Behavior No-slip Behavior

4 3 2 1 0 0 5 10 15 20

5 4 3 2 1 0 0 5 10 15 20

Frequency (Hz)

Frequency (Hz)

6

Slip Behavior No-slip Behavior

6

Slip Behavior No-slip Behavior

4 3 2 1 0 0 5 10 15 20

5 4 3 2 1 0 0 5 10 15 20

Frequency (Hz)

Frequency (Hz)

Figure 14.18: Comparison of FFT of the acceleration of the structure between slipping and no-slipping models for Ricker wave

Jeremi et al.

Lecture Notes

503

4.5s

110 100 90

0.12

4.6s

110 100 90

0.08 0.12

4.7s

110 100 90

0.08 0.12 0.1 0.1

0.1

Y (m)

Y (m)

70 60 50 40 30 30 40 50 60 70 80 90 100 110

0.06

70 60 50

Y (m)

80

80

80 70 60 50

0.08

0.06

0.06

0.04

0.04

0.04

0.02

40 30 30 40 50 60 70 80 90 100 110

0.02

40 30 30 40 50 60 70 80 90 100 110

0.02

X (m)

X (m)

X (m)

4.8s

110

0.12

4.9s

110 100 90

0.08 0.12

5.0s

110 100 90

0.08 0.12 0.1 0.1

100 90

0.1

Y (m)

Y (m)

70 60 50 40 30 30 40 50 60 70 80 90 100 110

0.06

70 60 50

Y (m)

80

80

80 70 60 50

0.08

0.06

0.06

0.04

0.04

0.04

0.02

40 30 30 40 50 60 70 80 90 100 110

0.02

40 30 30 40 50 60 70 80 90 100 110

0.02

X (m)

X (m)

X (m)

5.1s

110 100 90

0.12

5.2s

110 100 90

0.08 0.12

5.3s

110 100 90

0.08 0.12 0.1 0.1

0.1

Y (m)

Y (m)

70 60 50 40 30 30 40 50 60 70 80 90 100 110

0.06

70 60 50

Y (m)

80

80

80 70 60 50

0.08

0.06

0.06

0.04

0.04

0.04

0.02

40 30 30 40 50 60 70 80 90 100 110

0.02

40 30 30 40 50 60 70 80 90 100 110

0.02

X (m)

X (m)

X (m)

Figure 14.19: Distribution of gap openings along the contact interface for Ricker wave (gray scale given in meters)

Jeremi et al.

Lecture Notes

504

4.5s

110 100 90

0.03

4.6s

110 100 90

0.02 0.03

4.7s

110 100 90

0.02 0.03 0.025 0.025

0.025

Y (m)

Y (m)

70 60 50 40 30 30 40 50 60 70 80 90 100 110

0.015

70 60 50

Y (m)

80

80

80 70 60 50

0.02

0.015

0.015

0.01

0.01

0.01

0.005

40 30 30 40 50 60 70 80 90 100 110

0.005

40 30 30 40 50 60 70 80 90 100 110

0.005

X (m)

X (m)

X (m)

4.8s

110

0.03

4.9s

110 100 90

0.02 0.03

5.0s

110 100 90

0.02 0.03 0.025 0.025

100 90

0.025

Y (m)

Y (m)

70 60 50 40 30 30 40 50 60 70 80 90 100 110

0.015

70 60 50

Y (m)

80

80

80 70 60 50

0.02

0.015

0.015

0.01

0.01

0.01

0.005

40 30 30 40 50 60 70 80 90 100 110

0.005

40 30 30 40 50 60 70 80 90 100 110

0.005

X (m)

X (m)

X (m)

5.1s

110 100 90

0.03

5.2s

110 100 90

0.02 0.03

5.3s

110 100 90

0.02 0.03 0.025 0.025

0.025

Y (m)

Y (m)

70 60 50 40 30 30 40 50 60 70 80 90 100 110

0.015

70 60 50

Y (m)

80

80

80 70 60 50

0.02

0.015

0.015

0.01

0.01

0.01

0.005

40 30 30 40 50 60 70 80 90 100 110

0.005

40 30 30 40 50 60 70 80 90 100 110

0.005

X (m)

X (m)

X (m)

Figure 14.20: Distribution of sliding along the contact interface for Ricker wave (gray scale given in meters)

Jeremi et al.

Lecture Notes

505

110 100 90

Y (m)

80 70

600

60 50 40 30 30 40 50 60 70 80 90 100 110

400 200 0

X (m)

Figure 14.21: Distribution of cumulative dissipated energy due to sliding along the contact interface for Ricker wave (gray scale given in kJ)

3500 3000

Energy (kJ)

Time (s)

Figure 14.22: Cumulative dissipated energy time history due to sliding at the mid-center of the contact interface for Ricker wave (gray scale given in kJ)

Jeremi et al.

Lecture Notes

506

14.4.2

Ricker wave, with fault located at 34 towards the top middle point of the model

The last simulation to be studied here is the case which the location of the fault has an angle of 34 with respect to the top middle point of the model (Figure (14.15)). Same Ricker wave is used as an input motion to be propagated through the model built for rst stage of DRM analysis. In this case, since the source of the motion is farther from the interested domain (comparing to previous case with the angle of 45 ), motions will arrive later at the interface zone, and structure start shaking later. Displacement time histories at the bottom and top of the structure are recorded and shown in Figure (14.23) along X and Z directions. Comparing to the previous case along X direction, at the beginning of the shaking the magnitude of the displacement for slipping behavior model is a bit smaller while it damps out sooner as well. Along the X direction, the trend of the displacement time histories are pretty much the same with a slight dierence in magnitudes in time. Acceleration time histories of the structure are shown in Figure (14.24) along X and Z directions. The same behavior is observed here as previous case such that the amplitude of the acceleration is signicantly higher for case of no-slipping along the X direction while it is less along the Z direction. The seismic wave damps out sooner along X direction in case of considering the slipping behavior. The frequency content change is shown in Figure (14.25) by comparing the FFT of the accelerations obtained from the model considering the slipping behavior and the one with no-slipping behavior. It can be observed that there is a slight change in frequency content and predominant frequencies along both X and Z directions. Figure (14.26) shows the distribution of gap openings along the soil-foundation contact zone for 9 time steps from 4.5 to 5.3 seconds. The maximum gap opening happens at 4.5 seconds with magnitude of 0.12m. On the other hand, the sliding at the contact zone is shown in Figure (14.27) with maximum magnitude of 0.0035m happening at 4.5 seconds. In this case, the maximum sliding happens around the location of the structure. Figures (14.28) and (14.29) show the distribution of the dissipated seismic energy at the interface zone and dissipated energy time history at the middle of the contact zone respectively.

Jeremi et al.

Lecture Notes

507

0.6 0.4 0.2 0 -0.2 -0.4 -0.6 0 1 2 3 4 5 6 7 8

Slip Behavior No-slip Behavior

0.6 0.4 0.2 0 -0.2 -0.4 -0.6 0 1 2 3 4 5 6 7 8

Slip Behavior No-slip Behavior

Displacement (m)

Time (s)

Displacement (m)

Time (s)

0.6 0.4 0.2 0 -0.2 -0.4 -0.6 0 1 2 3 4 5 6 7 8

Slip Behavior No-slip Behavior

0.6 0.4 0.2 0 -0.2 -0.4 -0.6 0 1 2 3 4 5 6 7 8

Slip Behavior No-slip Behavior

Displacement (m)

Time (s)

Displacement (m)

Time (s)

Figure 14.23: Comparison of acceleration time histories of the structure between slipping and no-slipping models for Ricker wave

Jeremi et al.

Lecture Notes

508

40 30

Slip Behavior No-slip Behavior

40 30

Slip Behavior No-slip Behavior

Acceleration (m/s2)

Acceleration (m/s2)

Time (s)

Time (s)

40 30

Slip Behavior No-slip Behavior

40 30

Slip Behavior No-slip Behavior

Acceleration (m/s2)

Acceleration (m/s2)

Time (s)

Time (s)

Figure 14.24: Comparison of acceleration time histories of the structure between slipping and no-slipping models for Ricker wave

Jeremi et al.

Lecture Notes

509

10

Slip Behavior No-slip Behavior

10

Slip Behavior No-slip Behavior

8 6 4 2 0

0 5 10 15 20

8 6 4 2 0

10

15

20

Frequency (Hz)

Frequency (Hz)

10

Slip Behavior No-slip Behavior

10

Slip Behavior No-slip Behavior

8 6 4 2 0

0 5 10 15 20

8 6 4 2 0

10

15

20

Frequency (Hz)

Frequency (Hz)

Figure 14.25: Comparison of FFT of the acceleration of the structure between slipping and no-slipping models for Ricker wave

Jeremi et al.

Lecture Notes

510

4.5s

110 100 90

0.12

4.6s

110 100 90

0.08 0.12

4.7s

110 100 90

0.08 0.12 0.1 0.1

0.1

Y (m)

Y (m)

70 60 50 40 30 30 40 50 60 70 80 90 100 110

0.06

70 60 50

Y (m)

80

80

80 70 60 50

0.08

0.06

0.06

0.04

0.04

0.04

0.02

40 30 30 40 50 60 70 80 90 100 110

0.02

40 30 30 40 50 60 70 80 90 100 110

0.02

X (m)

X (m)

X (m)

4.8s

110

0.12

4.9s

110 100 90

0.08 0.12

5.0s

110 100 90

0.08 0.12 0.1 0.1

100 90

0.1

Y (m)

Y (m)

70 60 50 40 30 30 40 50 60 70 80 90 100 110

0.06

70 60 50

Y (m)

80

80

80 70 60 50

0.08

0.06

0.06

0.04

0.04

0.04

0.02

40 30 30 40 50 60 70 80 90 100 110

0.02

40 30 30 40 50 60 70 80 90 100 110

0.02

X (m)

X (m)

X (m)

5.1s

110 100 90

0.12

5.2s

110 100 90

0.08 0.12

5.3s

110 100 90

0.08 0.12 0.1 0.1

0.1

Y (m)

Y (m)

70 60 50 40 30 30 40 50 60 70 80 90 100 110

0.06

70 60 50

Y (m)

80

80

80 70 60 50

0.08

0.06

0.06

0.04

0.04

0.04

0.02

40 30 30 40 50 60 70 80 90 100 110

0.02

40 30 30 40 50 60 70 80 90 100 110

0.02

X (m)

X (m)

X (m)

Figure 14.26: Distribution of gap openings along the contact interface for Ricker wave (gray scale given in meters)

Jeremi et al.

Lecture Notes

511

4.5s

110 100 90

0.035 0.03 0.025

4.6s

110 100 90

0.035 0.03 0.025 0.02

4.7s

110 100 90

0.035 0.03 0.025 0.02

Y (m)

Y (m)

0.02

70

0.015

70

0.015

Y (m)

80

80

80 70

0.015

60 50 40 30 30 40 50 60 70 80 90 100 110

0.01 0.005 0

60 50 40 30 30 40 50 60 70 80 90 100 110

0.01 0.005 0

60 50 40 30 30 40 50 60 70 80 90 100 110

0.01 0.005 0

X (m)

X (m)

X (m)

4.8s

110

0.035 0.03 0.025

4.9s

110 100 90

0.035 0.03 0.025 0.02

5.0s

110 100 90

0.035 0.03 0.025 0.02

100 90

Y (m)

Y (m)

0.02

70

0.015

70

0.015

Y (m)

80

80

80 70

0.015

60 50 40 30 30 40 50 60 70 80 90 100 110

0.01 0.005 0

60 50 40 30 30 40 50 60 70 80 90 100 110

0.01 0.005 0

60 50 40 30 30 40 50 60 70 80 90 100 110

0.01 0.005 0

X (m)

X (m)

X (m)

5.1s

110 100 90

0.035 0.03 0.025

5.2s

110 100 90

0.035 0.03 0.025 0.02

5.3s

110 100 90

0.035 0.03 0.025 0.02

Y (m)

Y (m)

0.02

70

0.015

70

0.015

Y (m)

80

80

80 70

0.015

60 50 40 30 30 40 50 60 70 80 90 100 110

0.01 0.005 0

60 50 40 30 30 40 50 60 70 80 90 100 110

0.01 0.005 0

60 50 40 30 30 40 50 60 70 80 90 100 110

0.01 0.005 0

X (m)

X (m)

X (m)

Figure 14.27: Distribution of sliding along the contact interface for Ricker wave (gray scale given in meters)

Jeremi et al.

Lecture Notes

512

110 100 90

Y (m)

80 70

600

60 50 40 30 30 40 50 60 70 80 90 100 110

400 200 0

X (m)

Figure 14.28: Distribution of cumulative dissipated energy due to sliding along the contact interface for Ricker wave (gray scale given in kJ)

3500 3000

Energy (kJ)

Time (s)

Figure 14.29: Cumulative dissipated energy time history due to sliding at the mid-center of the contact interface for Ricker wave (gray scale given in kJ)

Jeremi et al.

Lecture Notes

513

14.5

Summary

On previous sections, the sliding behavior of earthquake-soil-foundation-structure systems was studied. Sliding behavior is due to existence of two adjacent materials with dierent properties. Based on the results of dierent simulations, the importance of considering the slipping behavior in numerical modeling was shown. Dissipation of seismic energy due to sliding at the contact zone was studied and shown how energy could be dissipated at the interface zone when two contact areas are sliding on each other. There are benets for designing process if there is an understanding in numerical simulations regarding where and then energy dissipation occurs in the system. The contact zone is one of the possible places for dissipation of energy. The 3D nite element model for studying the slipping behavior were consisted of a soil layer and foundation modeled with 8 and 27 node brick elements, as well as a structure modeled by using displacement beam elements. Frictional contact elements were placed at the interface of foundation and soil layer considering appropriate coecients of friction and stiness. The slipping behavior of SFSI model under 1D seismic wave propagation was studied considering Morgan Hill earthquake and articial Ricker wavelet. It was shown that for case of using Morgan Hill earthquake, in general there was a deamplication of the motion from the bottom of the structure to top of it. The magnitude of displacements and accelerations were slightly higher for case of slipping model and signicant phase lag was observed along Z direction. For simulation with using Ricker wavelet as an input motion, it was observed that displacement time history of top of the structure along X direction is higher when slipping behavior is not considered, while along Z direction it is higher when the slipping behavior is considered. It was discussed that this might be due to the fact that structural response is aected by gap openings and sliding at the interface zone leading to have rocking of the foundation and structure. There was also a slight shift in predominant frequency and change in frequency content of the recorded motions for these two models. This change in frequency content or shifting the predominant frequency of the motion could change the seismic design of the structure. There might be a case in which considering the slipping behavior in modeling the SFSI system brings the dominant frequency of the input motion close to the natural frequency of the structure. In this case resonant may happen and structure might not have been designed to tolerate bigger magnitudes of motions. It was also observed that due to slipping behavior, the frequency content and predominant frequency of the response could be shifted. In case the predominant frequency of the motion gets further from the structural natural frequency, structure can be designed more economically. For the cases of SFSI slipping behavior under 3D wave propagation simulations, Ricker wavelet used as an input motion. A model with the width size of 10000m and height of 5000m was considered for free eld 3D wave propagation in order to come up with the displacement and acceleration time histories required for smaller model to be used for second stage of DRM. Location of the fault considered to be at the depth of 2000m. Two models were prepared with two dierent locations for the fault: 3000m and 2000m. Results of these simulations showed that there could be signicant amount of dierence in displacement or acceleration responses by properly modeling the slipping behavior of the interface zone.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

514

Due to the presence of the vertical ground motions in 3D wave propagation models, signicant dierence between the responses obtained from two models of slipping and no-slipping was observed comparing to the ones obtained from the case with 1D wave propagation assumption. It was also shown how dissipation of energy can be monitored in time due to sliding of the contact zones.

Jeremi et al.

Chapter 15

(2002-2006-2009-)

(In collaboration with Dr. Zhao Cheng, Prof. Mahdi Taiebat, Ms. Panagiota Tasiopoulou)

515

Lecture Notes

516

15.1 15.2

Liquefaction is one of the most complex phenomena in earthquake engineering. Liquefaction also represents one of the biggest contributors to damage of constructed facilities during earthquakes (Kramer, 1996a). Prediction of behaviors of liqueable soils is dicult but achievable. There are number of methods that can be utilized to predict such behaviors. Methods currently used can have varying prediction accuracy and certainty. Of particular interest in this paper is the description of veried and validated numerical simulation methodology based on rational mechanics that is used to model, simulate and predict behavior of a single pile in liqueable soil subjected to seismic loading. Both level and sloping ground pile systems are modeled and simulated. Detailed description of background theory, formulation and implementation were recently given by Cheng et al. (2007) and Jeremi et al. (2008). It should be noted that presented development does show great promise in analyzing a myriad of liquefaction related problems in geotechnical and structural engineering. The eectiveness and power of numerical simulation tools for analyzing liquefaction problems becomes even more important and prominent in the light of potential disadvantages of models used in experimental simulations. These disadvantages, related to proper scaling (Wood, 2004) and problems in maintaining appropriate similarities (Harris and Sabnis, 1999) for rst order important phenomena, can render scaled models ineective, when used for physical simulations (under one-step or multiplestep gravity loading). In what follows, a brief literature review is provided. The literature review comprises sections on observations of liquefaction behavior in case studies, noncontinuum modeling eorts, review of redistribution of voids and pore uid volume/pressures phenomena and continuum modeling eorts.

15.2.0.0.1

Observation of Behavior.

Liquefaction behavior was observed during a number of earthquakes in the past. During Alaskan Earthquake (1964), liquefaction was the main cause of severe damage to 92 highway bridges, moderate to light damage to another 49 highway bridges, and moderate to sever damage to 75 railroad bridges (Youd and Bartlett, 1989). During Niigata Earthquake (1964) liquefaction induced damage to foundation piles under Yachiya bridge (Hamada, 1992). During that same earthquake, girders of Showa Bridge toppled as the support structure and piles moved excessively due to liquefaction (Japanese Society of Civil Engineers, 1966). During Kobe Earthquake (1995), liquefaction was the primary cause of damage to many pile supported or caisson supported bridges and structures. For example, ShinShukugawa bridge was subjected to excessive pile foundation movement due to liquefaction (Yokoyama et al., 1997). Opposed to these failures and collapses, there were a number of bridges with pile foundations that did not suer much or even minor damage even though there was liquefaction around foundations. For example, pile foundations of the Landing Road Bridge in New Zealand performed quite well during Edgecumbe earthquake (1987) even with

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

517

a signicant liquefaction recorded (Berril et al., 1997; Dobry and Abdoun, 2001). In addition to that, Second Maya Bridge piles (large steel pipes) were not damaged during Kobe earthquake despite signicant liquefaction in surrounding soils (Yokoyama et al., 1997).

15.2.0.0.2

Modeling and simulation of piles in liqueed grounds has been focus of a number of recent studies. The simple approach, based on scaling of p-y springs has been suggested early by Japanese Road Association (1980), Architectural Institutive of Japan (1988), Liu and Dobry (1995), Miura et al. (1989) and ORourke (1991). However, large inconsistencies with material parameter selection are present when p-y spring approach is used for piles in liqueed soils. Since p-y methodology for liqueed soils is not based on rational mechanics, appropriate choice of material parameters is primarily based on empirical observations of behaviors of piles in liqueed soils in experimental studies. A number of experimental studies have carefully examined pile behaviors in liqueable soils. We mention Tokida et al. (1992), Liu and Dobry (1995), Abdoun et al. (1997), Horikoshi et al. (1998) and Boulanger

and Tokimatsu (2006). Studies using physical model can be used to obtain very high quality data on behavior of piles in liqueed soils, provided that similarity of important physical phenomena is maintained (Wood, 2004; Harris and Sabnis, 1999). Some of the recent papers that discussed use of these models and gave recommendations about parameter choices are listed for reference: Tokimatsu and Asaka (1998), Martin et al. (2002), Dobry et al. (2003), Liyanapathirana and Poulos (2005), Rollins et al. (2005), ubrinovski and Ishihara (2006), Brandenberg et al. (2007).

15.2.0.0.3

Mechanics of pile behavior in liqueable grounds is based on the concept of redistribution of voids and pore uid volume/pressures (RVPFVP). It should be emphasized that geomechanics phenomena of redistribution of voids pore uid volume/pressure is used here in purely mechanistic way. That is, RVPFVP is a phenomena that occurs in saturated soils and that phenomena is responsible for (is manifested in) liquefaction related soil behaviors with or without piles. This is noted as in some recent publications, RVPFVP terminology is explicitly used for problems of liquefaction induced failures of sloping grounds without piles. Our understanding of the RVPFVP phenomena is that RVFVP is responsible for many more facets of behavior of liqueed soils, rather than only failure of liqueed slopes. The early investigation of the RVPFVP phenomena was related to the behavior of innite slopes. For example, loss of shear strength in innite slopes is one of the early understood manifestations of RVPFVP (Whitman, 1985; National Research Council, 1985; Malvick et al., 2006). Laboratory investigation of sand was also used to observe the RVPFVP phenomena (Casagrande and Rendon, 1978; Gilbert, 1984)

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

518

15.2.0.0.4

Continuum based formulations for modeling liquefaction problems have been present for over two decades. In a landmark paper, Zienkiewicz and Shiomi (1984) presented three possible coupled formulations for modeling of soil skeleton pore uid problems. The most general and complete one is the so called u-p-U formulations while the other two, the u-p and the u-U have a number of restrictions on the domain of application. Here, the unknowns are the soil skeleton displacements u; the pore uid (water) pressure p; and the pore uid (water) displacements U . The u-p formulation captures the movements of the soil skeleton and the change of the pore pressure, and is the most simplistic one of the three mentioned above. This formulation neglects the dierential accelerations of the pore uid (it does account for acceleration of pore uid together with soil skeleton, but not the dierential one if it exists), and in one version neglects the compressibility of the uid (assuming complete incompressibility of the pore uid). In the case of incompressible pore uid, the formulation requires special treatment of the approximation function (shape function) for pore uid to prevent the volumetric locking (Zienkiewicz and Taylor, 2000). The majority of the currently available implementations are based on this formulation. For example Elgamal et al. (2002) and Elgamal et al. (2003) developed an implementation of the u-p formulation with the multi-surface plasticity model by Prevost (1985), while Chan (1988) and Zienkiewicz et al. (1999a) used generalized theory of

plasticity Pastor et al. (1990). The u-U formulations tracks the movements of both the soil skeleton and the pore uid. This formulation is complete in the sense of basic variables, but might still experience numerical problems (volumetric locking) if the dierence in volumetric compressibility of the pore uid and the solid skeleton is large. The u-p-U formulation resolves the issues of volumetric locking by including the displacements of both the solid skeleton and the pore uid, and the pore uid pressure as well. This formulation uses additional dependent unknown eld of pore uid pressures to stabilize the solution of the coupled system. The pore uid pressures are connected to (dependent on) displacements of pore uid. With known (given) volumetric compressibility of the pore uid, pore uid pressure can be calculated. Despite its power, the u-p-U formulation has rarely been implemented into nite element code, and has never (at least to our knowledge) been used to analyze pile liqueed soil interaction. This can be attributed in part to a sophistication of implementation that is required, and to a sizable increase in computational cost for u-p-U elements.

15.3

Material presented here is from Cheng et al. (2007); Jeremi et al. (2008). Liquefaction of level and sloping grounds represents a very common behavior during earthquakes. Of interest is to estimate settlement for level ground, and horizontal movements for sloping grounds. In next few sections, presented are results for a 1D, vertical (level ground) and sloping ground cases for dense and loose sand behavior during seismic shaking.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

519

15.3.1

Model Description

Vertical soil column consists of a multiple-elements subjected to an earthquake shaking. The soil is assumed to be Toyoura sand and the calibrated parameters are from Dafalias and Manzari (2004), and are given in the Table (15.1). Table 15.1: Material parameters of Dafalias-Manzari model. material parameter Elasticity Critical sate G0 v M c c er Yield surface m value 125 kPa 0.05 1.25 0.712 0.019 0.7 0.934 0.01 Fabric-dilatancy Dilatancy material parameter Plastic modulus h0 ch nb A0 nd zmax cz value 7.05 0.968 1.1 0.704 3.5 4.0 600.0

The other parameters, related to the boundary value problem are given in table (15.2). For tracking convenience, the mesh elements are labeled from E01 (bottom) to E10 (surface) and nodes at each layers are labeled from A (bottom) to K (surface). A static application of gravity analysis is performed before seismic excitation. The resulting uid hydrostatic pressures and soil stress states along the soil column serve as initial conditions for the subsequent dynamic analysis. It should be noted that the self weight loading is performed on an initially zero stress (unloaded) soil column and that the material model and numerical integration algorithms are powerful enough to follow through this early loading with proper evolution. The boundary conditions are such that the soil and water displacement

Table 15.2: Additional parameters used in boundary value problem simulations (other than material parameters from the Table (15.1)). Parameter Solid density Fluid density Solid particle bulk modulus Fluid bulk modulus permeability HHT parameter Symbol s f Ks Kf k Value 2700 kg/m3 1000 kg/m3 3.6 107 kN/m2 2.2 106 kN/m2 5.0 104 m/s -0.2

Jeremi et al.

Lecture Notes

520

degree of freedom (DOF) at the bottom surface are xed, while the pore pressure DOFs are free; the soil and water displacement DOFs at the upper surface are free upwards to simulate the upward drainage. The pore pressure DOFs are xed at surface thus setting pore pressure to zero. On the sides, soil skeleton and water are prevented from moving in horizontal directions while vertical movement of both is free. It is emphasized that those displacements (of soil skeleton and pore uid) are dierent. In order to simulate the 1D behavior, all DOFs at the same depth level are connected in a masterslave fashion. Modeling of sloping ground is done by creating a constant horizontal load, sine of inclination angle, multiplied by the self weight of soil column, to mimic sloping ground. In addition to that, for a sloping ground, there should be a constant ow (slow) downhill, however this is neglected in our modeling. The permeability is assumed to be isotropic k = 5.0104 m/s. The input acceleration time history (Figure (15.1)) is taken from the recorded horizontal acceleration of Model No.1 of VELACS project Arulanandan and Scott (1993) by Rensselaer Polytechnic Institute (http://geoinfo.usc.edu/gees/velacs/). The

0.3

0.2

0.1

0.1

0.2

0.3

10

12

14

16

18

20

Time (s)

Figure 15.1: Input earthquake ground motion for the soil column.

magnitude of the motion is close to 0.2 g, while main shaking lasts for about 12 seconds (from 1 s to 13 s). For the sloping ground model a slope of % 3 was considered. It should be emphasized that the soil parameters are related to Toyoura sand, not Nevada sand which is used in VELACS project. The purpose of presented simulation is to show the predictive performance using veried and validated formulation, algorithms, implementation and models.

15.3.2

Figure (15.2) describes the response of the sample with loose sand e0 = 0.85. This gure shows the typical mechanism of cyclic decrease in eective vertical stress due to pore pressure build up as expected for the looser than critical granular material. The lower layers show only the reduction of eective vertical stress from the

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

521

beginning. Once the eective vertical (and therefore conning) stress approaches the smaller values, signs of the so-called buttery shape can be observes in the stress path. Similar observation can made in the upper layers which have the smaller conning pressure comparing to the lower layers from the very beginning due to lower surcharge. The upper layers have lower conning pressure (lower surcharge) at the beginning of the shaking, hence less contractive response is expected in these layer; however, soon after the initiation of the shaking these top layers start showing the liquefaction state and that type of response continues even after the end of the shaking. The top section of the model has remained liqueed well past the end of shaking. This is explained by the large supply of pore uid from lower layers, for which the dissipation starts earlier. For example, for the lowest layer, the observable drop in excess pore pressure start as soon as the shaking ends, while, the upper layers then receive this dissipated pore uid from lower layers and do liquefy (or continue being liqueed) well past end of shaking (which happens at approximately 13 seconds). It is very important to note the signicance of this incoming pore water ux on the pore water pressure of the top layers. Despite the less contractive response of soil skeleton at the top elements, the transient pore water ux, that enters these elements from the bottom, forces those to a liquefaction state. In other words, the top elements have not liqueed only due to their loose state but also because of the water ow coming from the bottom layers. The maximum horizontal strains can be observed in the middle layers due to liquefaction and prevents upper layers from experiencing larger strains. The displacements of water and

soil are presented in the last column. It shows that in all layers the upward displacement of water is larger than the downward displacement of soil. This behavior reects soil densication during shaking. Figure (15.3) describes the response of the sample with dense sand (e0 = 0.75). This gure also shows the typical mechanism of cyclic decrease in eective vertical stress. However, in case of this dense sample the decreasing rate of the eective conning pressure is much smaller than what was observed in the loose sample. Signs of the partial buttery shape in the eective stress path can be observed from early stages of shaking. The buttery is more evident in the upper layers with the lower conning pressure, i.e. more dilative response. In later stages of the shaking, i.e. when the conning pressure reduces to smaller value the buttery shape of the stress path gets more pronounced due to having more dilative response in the lower conning pressure based on CSSM concept. In comparing this dense case to the case of shaking the loose sand column, the current case does not show any major sign of liquefaction (when stress ratio ru = 1). This is due to the less contractive (more dilative) response of the sand in this case, which is coming from the the denser state of the sample. Because of having partial segments of dilative response, the whole column of the sand has not loosed its strength to the extent that happened for the case of loose sand and therefore smaller values of horizontal strains has been observed in the results. The absolute values of soil and water vertical displacements are also smaller than the case of loose sand which can be again referred to the less overall contractive response in this case. Overall, it can be noted that the response in the case of loose sand (e0 = 0.85) is mainly below the dilatancy surface (phase transformation surface) while the denser sand sample with e0 = 0.75 shows partially dilative response referring to the denser than critical state.

Jeremi et al.

Lecture Notes

522

20

20

h (kPa)

h (kPa)

Dz (cm)

E10

0

E10 Ru

E10

0

K

0 Soil Water

20 20

0 1

20 20

4 4

h (kPa)

h (kPa)

Dz (cm)

E07 Ru

E07

E07

H

0

20 20

0 1

20 20

4 4

h (kPa)

h (kPa)

Dz (cm)

E04 Ru

E04

E04

E

0

20 20

0 1

20 20

4 4

h (kPa)

h (kPa)

Dz (cm)

E01 Ru

E01

E01

B

0

20

v (kPa)

50

100

20

40

60

80

20

Time (s)

h (%)

20

40

60

80

Time (s)

E10

E09

E08

E07

E06

E05

E04

E03

E02

4

Figure 15.2: Seismic results for (loose sand) soil column in level ground (e0 = 0.85).

20

20

h (kPa)

h (kPa)

Dz (cm)

E10

0

E10 Ru

E10

0

K

0 Soil Water

20 20

0 1

20 20

4 4

h (kPa)

h (kPa)

Dz (cm)

E07 Ru

E07

E07

H

0

20 20

0 1

20 20

4 4

h (kPa)

h (kPa)

Dz (cm)

E04 Ru

E04

E04

E

0

20 20

0 1

20 20

4 4

h (kPa)

h (kPa)

Dz (cm)

E01 Ru

E01

E01

B

0

20

v (kPa)

50

100

20

40

60

80

20

Time (s)

h (%)

20

E01

40 60

80

Time (s)

E10

E09

E08

E07

E06

E05

E04

E03

E02

Figure 15.3: Seismic results for (dense sand) soil column in level ground (e0 = 0.75).

Jeremi et al.

E01

Version: March 21, 2012, 17:15

Lecture Notes

523

15.3.3

Figures (15.4) and (15.5) present the result of the numerical simulations for shaking the inclined soil columns (toward right) with loose and dense sand samples, respectively. The inclination of the soil column results in presence of the oset shear stress to the right side. This essentially poses asymmetric horizontal shear stresses (toward the direction of inclination) during cycles of shaking. On one hand, this oset shear stress makes the sample more dilative in the parts of shaking toward the right side (think about the state distance from the phase transformation line or dilatancy line in the p q space). As a result asymmetric buttery loops will be induced causing the soil to regain its stiness and strength (p) in the dilative parts of the corresponding cycles, therefore only instantaneous spikes of ru = 1 can be observed in case of the sloped columns of soil. There is also a permanent liquefaction in terms of having stationary portions of ru = 1 in this case. On the other hand, the oset shear stress results generation of more horizontal strains in the portions of loading which are directed toward the right side than those which are directed back toward the left side. As a result the horizontal shear strains will accumulative toward the right side and create larger permanent horizontal displacement comparing to the case of level ground soil column. Since the overall dilative response of the dense sample, i.e. Figure (15.5), is larger than that of the loose sample, i.e. Figure (15.4), the dense sample shows stier response and therefore less accumulative horizontal shear strains than the loose sample. The dierence in predicted horizontal displacements is almost three times, that is, for the dense sample the nal, maximum horizontal displacement is approx. 0.5 m, while for the loose sand sample, it almost 1.5 m,

25 1 25 1.5

h (kPa)

h (kPa)

E10

0

E10 Ru

E10

0

K Dx (m)

0 1.5

25 25

0 1

25 25

h (kPa)

h (kPa)

E07

0

E07 Ru

E07

0

H Dx (m)

0 1.5

25 25

0 1

25 25

h (kPa)

h (kPa)

E04

0

E04 Ru

E04

0

E Dx (m)

0 1.5

25 25

0 1

20

40

60

80

25 25

h (kPa)

h (kPa)

E01

0

E01 Ru

E01

0

B Dx (m)

25

v (kPa)

50

100

20

40

60

80

25

Time (s)

h (%)

10

20

10

20

Time (s)

B E02

E10

E09

E08

E07

E06

E05

E04

E03

Figure 15.4: Seismic results for (loose sand) soil column in sloping ground (e0 = 0.85).

Jeremi et al.

E01

Version: March 21, 2012, 17:15

Lecture Notes

524

25

25

1.5

h (kPa)

h (kPa)

E10

0

E10 Ru

E10

0

K Dx (m)

0 1.5

25 25

0 1

25 25

h (kPa)

h (kPa)

E07

0

E07 Ru

E07

0

H Dx (m)

0 1.5

25 25

0 1

25 25

h (kPa)

h (kPa)

E04

0

E04 Ru

E04

0

E Dx (m)

0 1.5

25 25

0 1

20

40

60

80

25 25

h (kPa)

h (kPa)

E01

0

E01 Ru

E01

0

B Dx (m)

25

v (kPa)

50

100

20

40

60

80

25

Time (s)

h (%)

10

20

10

20

Time (s)

B E02

E10

E09

E08

E07

E06

E05

E04

E03

Figure 15.5: Seismic results for (dense sand) soil column in sloping ground (e0 = 0.75).

15.4

Material presented here is from Cheng and Jeremi (2009a,b) Model development for a pile in the liqueable soil follows physics (mechanics) of the problem as close as possible. Numerical simulation of such problems in geomechanics is usually based on stages of loading and increments within those stages. All load stages are applied to a series of nite element models, all of which share features of an initial soil model. This initial soil model consists of a soil block with dimension of 12 12 15 m (length width depth). Due to the symmetry of the model, only half of the block is modeled. Symmetry assumptions is based on assumption that all the loads, dynamic shaking and other inuences are symmetric with respect to the plane of symmetry. This specialization to symmetric model reduces model generality (for example this use of symmetry will preclude analysis of dynamic shaking perpendicular to sloping ground dip). However, as our goal is to present a methodology of analyzing behavior of piles in liquefying ground, this potential drawback is not deemed signicant in this study. Finite element mesh for the model is presented in Figure (15.6). The initial mesh consists of 160 eight node u-p-U elements. Each node of the mesh has 7 degrees of freedom, three for soil skeleton displacements (ui ), one for pore water pressure (p), and three for pore water displacement (Ui ). While it can be argued that the mesh is somewhat

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

E01

Lecture Notes

z Lumped Mass x

525

e1 B9 e2 C9 e3 12 m e4 e5 e6 E9 F9 G9 H9 e8 I9 12 m 15 m D9

e7 3m

Figure 15.6: Left: Three dimensional nite element mesh featuring initial soil setup, where all the soil elements are present. The gray region of elements is excavated (numerically) and replaced by a pile during later stages of loading; Right: Side view of the pile-soil model with some element and node annotation, used to visualize results.

coarse, it is well rened around the pile, yet to be installed, in place of gray region in the middle. A single set of parameters is used with the Dafalias-Manzari material model. Soil is modeled as Toyoura sand and material parameters (summarized in Table 15.4) are calibrated using tests by Verdugo and Ishihara (1996), while initial void ration was set to e0 = 0.80. It is very important to emphasize that the state of stress and internal variables from initial state (zero for stress and given value for void ratio and fabric) will evolve through all stages of loading by proper modeling and algorithms, by using single set of material parameters. Table 15.4 presents additional parameters, other than material parameters presented in Table 15.4, used for numerical simulations.

15.4.1

The initial stage of loading is represented by the application of self weight on soil, including both the soil skeleton and the pore water. Initial state in soil before application of self weight is of a zero stress and strain while void ratio and fabric are given initial values. The state of stress/strain, void ratio and fabric will evolve upon application

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

526

Table 15.3: Material parameters used for Dafalias-Manzari elasticplastic model. Material Parameter Elasticity Critical sate G0 v M c c er Yield surface m Value 125 kPa 0.05 1.25 0.8 0.019 0.7 0.934 0.02 Fabric-dilatancy Dilatancy Material Parameter Plastic modulus h0 ch nb A0 nd zmax cz Value 7.05 0.968 1.1 0.704 3.5 4.0 600.0

Table 15.4: Additional parameters used in FEM simulations. Parameter Solid density Fluid density s f Ks Kf k g Value 2800 kg/m3 1000 kg/m3 1.0 1012 kN/m2 2.2 106 kN/m2 1.0 104 m/s 10 m/s2

Solid particle bulk modulus Pore uid bulk modulus permeability Gravity

of self weight. At the end of self weight loading stage, soil is under appropriate state of stress (K0 stress), the void ratio corresponds to the void ratio after self weight (redistributed such that soil is denser at lower layers), while soil fabric has evolved with respect to stress induced anisotropy. All of these changes are modeled using DafaliasManzari material model and using constitutive and nite element level integration algorithms developed within UC Davis Computational Geomechanics group in recent years. Boundary conditions (BC) for self weight stage of loading are set in the following way: Soil skeleton displacements (ui ), are xed in all three directions at the bottom of the model. At the side planes, nodes move only vertically to mimic self-weight eect. All other nodes are free to move in any direction. Pore water pressures (p), are free to develop at the bottom plane and at all levels of the models except at the top level at soil surface where they are xed (set to zero, replicating drained condition), Pore water displacements (Ui ), are xed in all three directions at the bottom, are free to move only vertically at four sides of the model and are free to move in any direction at all other nodes. These boundary condition are consistent with initial self-weighting deformation condition for soil and pore water at the site.

Jeremi et al. University of California, Davis Version: March 21, 2012, 17:15

Lecture Notes

527

For the case of sloping ground, an additional load sub-stage is applied after self weight loading, in order to mimic self weight of inclined (sloping) ground. This is eectively achieved by applying a resultant of total self weight of the soil skeleton times the sine of the inclination angle at uphill side of the model. This load is applied only to the solid skeleton DOFs, and not on the water DOFs. Physically it would be correct to consider the sloping ground eects on the pore water as well. This will create

## Гораздо больше, чем просто документы.

Откройте для себя все, что может предложить Scribd, включая книги и аудиокниги от крупных издательств.

Отменить можно в любой момент.