Numerical Modeling in Micromechanics Via Particle Methods 2004 Proceedings of The 2nd International PFC Symposium Kyoto Japan 28 29 October 2004

Cover
Page i
NUMERICAL MODELING IN MICROMECHANICS VIA PARTICLE METHODS—
2004
Page ii
This page intentionally left blank.
Page iii
PROCEEDINGS OF THE 2ND INTERNATIONAL PFC SYMPOSIUM, 28–29 OCTOBER 2004, KYOTO, JAPAN
Numerical Modeling in Micromechanics via Particle Methods—2004
Edited by
Y.Shimizu, R.D.Hart & P.A.Cundall
Itasca Consulting Group, Inc., Minneapolis, MN, USA

A.A.BALKEMA PUBLISHERS LEIDEN/LONDON/NEW YORK/PHILADELPHIA/SINGAPORE
Page iv
Copyright © 2004 Taylor & Francis Group plc, London, UK
All rights reserved. No part of this publication or the information contained herein may be reproduced, stored
in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, by photocopying,
recording or otherwise, without written prior permission from the publisher.
Although all care is taken to ensure the integrity and quality of this publication and the information herein,
no responsibility is assumed by the publishers nor the author for any damage to property or persons as a result
of operation or use of this publication and/or the information contained herein.
Published by: A.A.Balkema Publishers, Leiden, The Netherlands, a member of Taylor & Francis Group plc
www.balkema.nl, http://balkema.tandf.co.uk and www.tandf.co.uk
This edition published in the Taylor & Francis eLibrary, 2005.
To purchase your own copy of this or any of Taylor & Francis or Routledge’s collection of thousands of eBooks please go to www.eBookstore.tandf.co.uk.
ISBN 0203024419 Master ebook ISBN
ISBN 90 5809 679 3 (Print Edition)
Page v
Table of Contents
Numerical Modeling in Micromechanics via Particle Methods—2004—Shimizu, Hart & Cundall (eds.)
© 2004 Taylor & Francis Group, London, ISBN 90 5809 679 3

Preface IX
Organization XI

Soil/structural support

Use of DEM to model the interlocking effect of geogrids under static and cyclic loading 3

H.Konietzky, L.te Kamp, T.Groeger & C.Jenner
DEM modeling of soil geocomposite material for rockfall impact 13

D.Bertrand, F.Nicot & Ph. Gotteland
Microstructure modeling of granular media and interactions between pile and granular ground by DEM 21

K.Maeda, T.Sakurai, A.Oomura & H.Hirabayashi

Slope and wall stability

Application of PFC3D for determination of soil properties and simulation of the excavation process in front of sheet pile wall constructions 35

M.Pohl, M.Pulsfort & B.Walz
Finite element and discrete element modeling for the solution of spatial active earth pressure problems 45

F.tom Woerden, M.Achmus & K.AbdelRahman
Numerical study on the cause of a slope failure at a gravel pit using PFC and FLAC 51

T.Sakakibara, M.Ujihira & K.Suzuki
Evaluation of castle stone wall stability with the particle flow code 57

M.Mizumoto & G.Q.Cao

Rock fractures

2D 63
Simulation of rock cutting using distinct element method—PFC
S.Lei, P.Kaitkay & X.Shen
Evaluation of behavior of EDZ around rock cavern by AE measurements and PFC simulation 73

K.Aoki, Y.Mito, T.Mori & T.Maejima
Numerical simulation of the mechanical behavior of discontinuous rock masses 85

E.S.Park, C.D.Martin & R.Christiansson

Industrial

Introduction of research and development on electromechanics of electromagnetic particles for imaging technology 95

H.Kawamoto
DEM simulation of bulk material flow in a feeding system of blast furnace 103

J.F.Ferellec, C.Mochizuki & A.Hirata
Using particle flow code to investigate energy dissipation in a rotary grinding mill 111

R.S.Sarracino, A.T.McBride & M.S.Powell
The modeling of the separation process in a ferrohydrostatic separator using discrete element method 119

V.Murariu, J.Svoboda & P.Sergeant
Page vi
A discrete particle contact model for sintering 127

S.Luding
Discrete element modeling for dynamics of vibroimpact system 135

M.Saeki

Shear flow

GEM simulation of shear flows in granular media 143

Y.Kishino
3D 149
PFC modeling of interparticle percolation in caved rock under draw
M.E.Pierce
Flow behavior of cohesive and frictional fine powders 157

S.Luding, R.Tykhoniuk, J.Tomas, L.Heim, M.Kappl & H.J.Butt
Mixtures of rolling and nonrolling circular disks studied at high shear strains 165

T.Vietor

Geology

Simulation of magmatic fabrics using PFC 175

Y.Park, C.Lee & J.H.Ree
Simulation of a twodimensional ‘thinskinned’ orogenic wedge using PFC 181

C.Lee, Y.Park & S.Kwon
Simulation of accretionary prisms by distinct element method 187

Y.Yamada, S.Ueda & T.Matsuoka
2D 191
Numerical simulation of gas hydrate behavior in marine sediments using PFC
S.Kreiter, B.Grupe & V.Feeser
2D 199
The particle flow code, PFC , applied in planetary studies to model the tectonic evolution of chasma walls on Mars
B.Imre

Theory & development

On a new numerical analysis method for fracture problems 209

M.Hori & K.Oguni
A selfsimilar bond model for PFC2D 217

P.A.Cundall
A grid computing architecture for applications in distributed computation 225

X.Lin & J.R.Williams
Ab initio emergent phenomena in PFC 235

B.E.Hobbs, A.Ord, K.RegenauerLieb, F.Boschetti, Y.Zhang & S.Durrlemann
Discrete computational methods as the tools for studying behavior of materials, structures and complex media on different scales 241

S.G.Psakhie
Application of an ellipsoid approximation to discrete/distinct element modeling 249

S.Johnson, J.R.Williams & B.K.Cook

Fluid coupling

3D 257
Fluid coupling in PFC using the measurement spheres
Y.Okura, H.Ochiai, P.A.Cundall & Y.Shimizu
Page vii
Flow deformation of a river dike due to seismic liquefaction 265

T.Honda & I.Towhata
A study on the calculation of particle volumetric deformation in a fluid coupled PFC model 273

L.Li & R.M.Holt
2D 3D 281
Fluid coupling in PFC and PFC
Y.Shimizu
Design of gas storage rock cavern by the hydromechanical coupled discrete model 289

K.Aoki, Y.Mito, T.Matsuoka & D.Kondoh
DEMfluid model development for nearwellbore mechanics 301

B.K.Cook, D.F.Boutt, O.E.Strack, J.R.Williams & S.M.Johnson

Lab/element tests

Micromechanical modeling of granular materials under triaxial and oedometric loading 313

R.Katzenbach & A.Schmitt
Particle shape influence on the fracture behavior of concrete 323

N.M.Azevedo & J.V.Lemos
The use of particle flow code for investigating the stress memory effect in rocks 331

V.Louchnikov, S.P.Hunt & A.G.Meyers
Numerically modeling the strength of ice using discrete elements 341

L.U.Arenson
Effects of intergranular adhesive force on behavior of granular material 347

S.Kato, T.Sakakibara & H.Yoshimori
Development of compressive failure analysis of concrete specimen by 3dimensional distinct element method 355

D.Haraki, H.Shiraishi & S.Katsuki
Experimental validation of modeling irregular particle shapes using DEM 363

A.M.Sallam, A.K.Ashmawy & B.D.Runkles
Threedimensional discrete element simulations of direct shear tests 373

C.O’Sullivan, L.Cui & J.D.Bray
Calibration of a distinct element model for rock failure envelope and tensile strength 383

A.Fakhimi & T.Villegas
Influence of stress history on in situ and core compaction 391

R.M.Holt, C.J.Kenter, B.Verboom & D.Doornhof
Numerical modeling of weak sand behavior 397

R.Narayanasamy
Macromechanical behavior for an assembly of grains modeled by bonding spheres 407

Y.Nakata, M.D.Bolton & Y.P.Cheng
Elastoviscoplastic constitutive model of sea ice with strain rate dependent ice strength 415

D.Zyryanov & R.Korsnes
3D imagebased discrete element modeling for irregularlyshaped grains 421

T.Matsushima
Comparison between numerical and laboratory tests of rapidly sheared granular materials 429

K.Iwashita, K.Ichiba & M.Oda

Author index 435
Page viii
Page ix
Preface
Numerical Modeling in Micromechanics via Particle Methods—2004—Shimizu, Hart & Cundall (eds.) © 2004 Taylor & Francis Group, London, ISBN 90
5809 679 3
PFC2D and PFC3D model the movement and interaction of assemblies of discrete particles using the distinct element numerical method. These programs have been
distributed by Itasca Consulting Group, Inc. since their first commercial release in 1994. Initial development was partially funded by a consortium of companies for
engineering applications ranging from largescale mine block caving to microscale powder compaction and sintering.
The variety of applications of PFC has continued to increase in the 10 years since the first release of these programs. This volume contains a collection of 52 papers
selected for presentation at the 2nd PFC Symposium, held 28–29, 2004, in Kyoto, Japan. These contributions cover a wide range of engineering applications and
theoretical developments using PFC, and discrete methods in general. Topics include applications in civil engineering, slope and wall stability, rock fracture, shear
flows, geology and industrial engineering. New developments are also described for contact bond models, fluid coupling and model calibration.
This proceeding illustrates the great variety of PFC applications in different engineering fields, and includes case studies and general applications as well as research
presentations. The volume should serve as a guide to help PFC users in their applications.
We would like to acknowledge the contribution of CRC Solutions, Corp., Tokyo, Japan, especially Toshio Iwasaki, in hosting the 2nd PFC conference and
assisting with the preparation of these proceedings. We also thank Michele Nelson who served as Technical Editor for the papers contained in this volume, and Chad
Sylvain, who prepared the graphics illustration for the cover. Their assistance has been invaluable.
Yoshi Shimizu
Roger Hart
Peter Cundall
Editors
Page x
Page xi
Organization
The Organization Committee
Dr. Roger Hart, Itasca Consulting Group, Inc., USA
Mr. Toshio Iwasaki, CRC Solutions, Inc., Japan
Ms. Michele Nelson, Itasca Consulting Group, Inc., USA
Dr. Yoshiyuki Shimizu, Itasca Consulting Group, Inc., USA
Technical Committee
Prof. Kenji Aoki, Kyoto University, Japan
Prof. Jon Bray, University of California—Berkeley, USA
Prof. Bernard Cambou, Ecole Centrale de Lyon, France
Dr. Guoqiang Cao, CRC Solution Corp., Japan
Dr. Peter Cundall, Itasca Consulting Group, Inc., USA
Dr. Ben Cook, Sandia National Laboratories, USA
Dr. Bruce Hobbs, CSIRO, Australia
Prof. Rune Holt, SINTEF Petroleum Research, Norway
Prof. Yuji Kishino, Tohoku University, Japan
Dr. Heinz Konietzky, Itasca Consultants GmbH, Germany
Prof. Tatsunori Matsumoto, Kanazawa University, Japan
Prof. Toshifumi Matsuoka, Kyoto University, Japan
Prof. Graham Mustoe, Colorado School of Mines, USA
Prof. Masanobu Oda, Saitama University, Japan
Dr. David Potyondy, University of Toronto, Canada
Dr. Joerg Theuerkauf, DOW Chemical Co., USA
Prof. Yutaka Tsuji, Osaka University, Japan
Prof. John Williams, Massachusetts Institute of Technology, USA
Page xii
Page 1
Soil/structural support
Page 2
Page 3
Use of DEM to model the interlocking effect of geogrids under static and cyclic
loading
H.Konietzky, L.te Kamp & T.Groeger
ITASCA Consultants, Gelsenkirchen, Germany
C.Jenner
TENSAR International, Blackburn, United Kingdom
ABSTRACT: Discrete element modeling (DEM) was used to investigate the interlocking behavior of geogrids under static and cyclic loading
conditions. Both the geogrid itself and the soil particles were represented by particles using PFC. The geogrid model was calibrated in terms of
stiffness and strength by performing different tensile and rotational tests. Pullout tests and calibrated triaxial tests with different load levels were used
to investigate the interlocking effect.
1 INTRODUCTION
Geogrids have seen an increased application in civil engineering over the past decades. They are used to stabilize steep slopes and avoid erosion, to reinforce soil
retaining walls and bridge abutments, to reduce settlements for railway track ballast, pavements, foundations and embankments. In contrast to geotextiles the
mechanism of geogrids is dominated by the interlocking effect. Although this effect was detected as the key mechanism and numerous successful applications have
demonstrated that this mechanism works, so far only little knowledge was available about the physical details. While any discontinuum mechanical approach fails due to
the discrete nature of the process and the penetration of materials, the use of the Discrete
Figure 1. Sketch of geogrid interlocking mechanism.
Element Method allows a detailed view into the interaction of the particles with the geogrid. Figure 1 shows a sketch, which demonstrates the mechanism of
interlocking.
2 GEOGRID MODEL
2.1 Geogrid model setup
The numerical modeling was performed using the Particle Flow Code (PFC). This code is based on the DEM using stiff spherical particles with deformable contacts
and breakable bonds in between (Itasca 2003, Konietzky 2003).
The geogrid was built up by chains of very small particles using the socalled parallel bonds (Itasca 2003), which reproduce tensile strength, bending resistance as
well as normal and shear stiffness. The shape of the geogrid was reproduced by the development of a specific arrangement of the particles and parallel bonds. Figure 2
shows a photograph of the Tensar biaxial geogrid SS30, which is underlying the numerical modeling.
The numerical geogrid with 9 cells (Fig. 3) consists of approximately 31,000 particles. Figures 4 & 5 show detailed views of a junction in form of the arranged
particles and the contacts between them, respectively.
2.2 Calibration
The calibration was performed in terms of stiffness and strength. Three different tests were performed: a single rib test, a single junction test and an inplane rotation
Page 4
Figure 2. Photo of Tensar geogrid SS30.
Figure 3. Numerical SS30 geogrid with 9 cells.
Figure 4. Detailed view of numerical SS30 geogrid model (particles).
Figure 5. Detailed view of numerical SS30 geogrid model (contacts).
Table 1. Comparison between lab test and adequate numerical simulation.
Maximum load [N] Strain [%] at maximum load
Lab. result Model result Lab. result Model result
Single rib test 1372 1450 10.5 11.8
Single junction test 1258 950 9.2 8.0
test. Results of tension tests are shown in Table 1. The numerical inplane rotational test has given a value of 0.92 Nm/degree compared to 0.79 Nm/degree obtained
by lab tests.
A better match could be achieved by further adjustment of the microproperties. Although this match is not perfect, the underlying parameter set was used to study
the interlocking mechanism. It is obvious, that a better match would not change the general behavior.
3 NUMERICAL MODEL FOR PULLOUT TEST
A typical grain size distribution as shown in Figure 6 was used to model the aggregate particles. The aggregate particles were placed into the modeling box using a
random procedure. Figures 7 & 8 show detailed views of the model setup with particles and the geogrid. Figure 9 shows a part of the whole model with geogrid and a
complete set of aggregate particles according to the
Page 5
Figure 6. Particle size distribution for aggregate particles.
Figure 7. Geogrid and aggregate particles with diameters between 0.6 and 4 mm.
Figure 8. Geogrid and aggregate particles with diameters between 14 and 20 mm.
particle size distribution given in Figure 6. Figure 10 shows a sketch of the whole model, which consists of a box with a height of 70 cm and a lateral extension of 18 by
18 cm. The geogrid is pulledout horizontally in the negative zdirection.
Figure 9. Geogrid with complete set of aggregate particles.
Figure 10. Sketch of numerical model for simulation of pullout test.
Three different loading situations were considered: pure gravitational load (load case a), 13.2 kPa vertical load in addition to the gravitational load (load case b) and
35.4 kPa vertical load in addition to the gravitational load (load case c). Figure 11 shows
Page 6
Figure 11. Contact forces between particles (left: pure gravitational load, right: gravitational load+surface load of 35.4 kPa). Thickness of
lines correspond to the magnitude.
the contact forces between the particles for the load cases a and c.
4 RESULTS OF NUMERICAL PULLOUT TEST
The geogrid was pulledout with low velocity to guarantee quasistatic conditions. Figure 12 shows the contact forces distribution for load case c. Figure 13 shows the
total pullout force as a function of the pullout length for all 3 considered load cases. As expected, higher vertical loads need higher pullout forces.
The curves look similar to stressstrain curves of triaxial tests with high confining pressure or shear tests with compacted frictional material. Figure 14 shows, that
with increasing vertical load the peak force is shifted to larger values of the pullout length and the difference between peak and residual values become more
pronounced.
The numerical model allows a detailed investigation in respect to the deformation and force distribution inside the geogrid. The modeling has revealed a quite
complex mixture of tensile, compressive, shear and bending forces inside the geogrid. But during the pullout test, the tensile forces dominate as expected. For example,
Figure 14 shows the axial deformation between two outer junctions as a function of the pullout length.
Figure 12. Contact forces (thickness of lines correspond to the magnitude) in a plane vertical plot view during the pullout test (left: at the
begin of the test, right: after pulling of 12 mm).
Figure 13. Total pullout force as a function of pullout length for the load cases a, b and c.
Page 7
Figure 14. Axial strain between measured in pullout direction between 2 junctions (marked with stars).
Figure 15. Maximum normal force before and during the pullout test for the load cases a, b and c.
Figure 16. Average normal force before and during the pullout test for the load cases a, b and c.
Figure 17. Maximum shear force before and during the pullout test for the load cases a, b and c.
The Figures 15 to 18 show the recorded shear and normal forces inside the model before the pullout test was started and at the end of test, where a pullout length of
app. 12 mm was reached. The geogrid is located at a depth of 0.0 m. As can be seen, the geogrid leads to a significant increase of contact forces in the area near the
geogrid. This mobilization of forces and stresses, respectively, is a clear indication of the interlocking effect. At the end of the tests, the peak values for both the shear
and normal forces reach app. 10 times the values at the start of the test. For the average forces this value is about 3. The interlocking
Page 8
Figure 18. Average shear force before and during the pullout test for the load cases a, b and c.
effect is restricted to a relatively small height at both sides of the geogrid and vanishes at values greater than about 20 cm.
5 NUMERICAL MODEL FOR CYCLIC LOADING
Triaxial laboratory tests were performed by SINTEF (2001) to investigate the behavior of geogrids under cyclic loading. These tests have revealed that geogrids have a
significant influence on the resistance against permanent deformations of the crushed rock material. The experiments have shown, that in the case of installed geogrids
considerably higher deviatoric stresses are necessary to produce the same deviatoric strains as for samples without geogrids.
This effect cannot be explained by tension forces inside the geogrid. Backcalculations have revealed, that the permanent deformations of the geogrid are only
approximately 10% of the value, which would correspond to the stiffness of the geogrid and the apparent confining pressure. It seems, that the interlocking effect
causes this apparent confinement. Therefore, the geogrid acts like an amplifier or corset: it enables the crushed particle assemble to sustain higher forces; it mobilizes the
intrinsic bearing capacity. The numerical modeling via PFC should give deeper insight into these effects.
5.1 One geogrid layer
According to the performed lab tests a numerical model was setup. The cylindrical sample was approximated by 8 vertical walls according to Figure 19. The
Figure 19. Numerical model setup.
model height is 60 cm and the diameter is 16 cm. The initial porosity was set to 29%. The contact friction angle was set to 60°, which corresponds to an overall friction
angle in the sense of continuum mechanics of approximately 33°. An automatic calibration procedure was used to specify particlediameter dependent normal and shear
contact stiffness values to represent an overall Young’s modulus of 300 MPa.
The measured grain size distribution of the granular material with particle radii between 2 and 11 mm was reproduced as documented by Figure 19. The crushed
Page 9
Figure 20. Deformed geogrid.
Figure 21. Crushed rock sample with geogrid.
rock is represented by approximately 50,000 particles. The geogrid itself was setup as described in section 2. Three different load levels (vertical load) were
considered: 20 kPa, 30 kPa and 40 kPa. Figure 20 shows a deformed geogrid.
Figure 22. Contact forces and geogrid.
Figure 23. Force ratio ß after 3 load cycles at 30% of maximum load as a function of model height (the arrow indicates the location of
the geogrid).
Figure 21 shows the crushed rock sample together with the installed geogrid. Figure 22 illustrates that the inner four intact geogrid meshes (cells) cause interlocking,
which is manifested by a dense network of stronger forces (force chains). To quantify the interlocking effect a force ratio ß (=ratio of average force inside a brick,
which covers in the horizontal direction the 4 intact meshes and a height of 2 cm to the average force in the same height interval but considering the full cross section)
was defined. Figures 23 & 24 show this force ratio as a function of model height.
Whereas the sample without a geogrid shows fluctuating values around 1.0 as expected (Fig. 24), the sample with geogrid shows a significant increase in the force
ratio of up to 1.2 near to the location of geogrid (Fig. 23). For all load levels the confining pressure was held constant at approximately 10 kPa by a servo algorithm.
Page 10
Figure 24. Force ratio ß after 3 load cycles at 30% of maximum load as a function of model height for a sample without geogrid.
Figure 25. Deformed sample with 3 geogrid layers.
As already revealed by static pullout tests (shown above) and conformed by these cyclic loading tests, the influence of the geogrid in terms of increased contact forces
between the crushed rock material is approximately 10 cm to both sides.
Figure 26. Force ratio ß during loading as a function of model height (the arrows indicate the location of 3 geogrid layers).
Figure 27. Force ratio ß after partially vertical and horizontal unloading (see also Fig. 28) as a function of model height (arrows indicate
the location of the geogrid layers).
5.2 Three geogrid layers
Further modeling was performed with 3 parallel geogrid layers installed at a distance of 15 cm (Fig. 25). The axial and confining pressure histories are the same as for
the model with only 1 geogrid, but both, the vertical and the radial displacements were reduced by approximately 50%. Figure 26 shows the distribution of the force
ratio ß (compare to Fig. 23). The interlocking effect is much more pronounced during unloading. Figure 27 shows the force ratio ß during partially vertical and
horizontal unloading. At the locations, where the geogrids are installed the ratio ß shows local minima. Local maxima are found in between. This again can be explained
by the interlocking effect. In the immediate vicinity of the geogrid the interlocking effect keeps the crushed rock material together, which results in relatively low ßvalues
in the order of 1.2. In between the geogrid layers the interlocking effect is restricted to the center part of the model (bridge effect between the geogrid layers), near to
the model boundaries the contact forces are strongly reduced, which lead to high ßvalues.
Page 11
Figure 28. Contact force distribution and geogrid after partially horizontal and vertical unloading.
6 CONCLUSIONS
The DEM approach has given valuable insight into the interlocking mechanism of geogrids under shear loading and cyclic normal loading.
The developed methodology can be used to answer several questions of practical importance, e.g. the optimum geogrid cell size in relation to the aggregate particle
size and the optimum distance between several geogrid layers. This information can be used for both, improved geogrid design and production as well as dimensioning.
Due to the interlocking effect the installation of geogrids works as an additional confinement. The use of several geogrid layers at short distances leads to a significant
reduction of displacements, e.g. 3 layers of geogrids can reduce the settlements by a factor of 2. Figure 28 shows the contact forces observed during partially unloading
in the horizontal direction using 3 geogrids. The concentration of forces near the geogrids and the contact chains between the geogrid layers are clearly recognizable.
The action of geogrids is strongest under low confinement or at high differential stresses.
REFERENCES

Itasca Consulting Group, Inc. 2003. PFC3D—Particle Flow Code in 3 Dimensions, Ver. 3.0 User’s Manual. Minneapolis, Minnesota: Itasca.
Konietzky, H. (ed.) 2003. Numerical Modeling in Micromechanics via Particle Methods, 321 p., Rotterdam: Balkema.
SINTEF 2001. Geogridreinforced base courses cyclic triaxial tests, SINTEF Civil and Environmental Engineering, Geotechnical Engineering, Norway, unpublished report.
Page 12
Page 13
DEM modeling of soil geocomposite material for rockfall impact
D.Bertrand
France Gabion Society, Lirigm & Cemagref de Grenoble, Domaine Universitaire, SaintMartind’Hères Cedex, France
F.Nicot
Cemagref de Grenoble, Domaine Universitaire, SaintMartind’Hères Cedex, France
Ph. Gotteland
Lirigm UJF Grenoble 1, Domaine Universitaire, Grenoble Cedex, France
ABSTRACT: One innovative development for rockfall barrier is to construct structures with assembly of cells, i.e. gabion technique. The proposed
presentation is focused on the numerical modeling of the cell. This modeling is based on the Discrete Element Method (DEM) analysis using PFC3D.
Cells are composed of two main systems: the wire netting and the filling granular material. To take into account the wire mesh, assemblies of spherical
particles were considered. The wire mesh is described by a set of particles located at the nodes of the mesh. The filling material was modeled with
unbreakable particles associations. It makes it possible to model different shapes and sizes of elements in the aim to reproduce the nonsmooth surface
and the grain size distribution. The numerical calculation was calibrated from experimental results obtained from laboratory tests carried out with single
cells. Numerical parametric analyses were finally performed to assess the dissipative capacity of the cells.
1 INTRODUCTION
Reinforced soil structures, including flexible elements such as geosynthetic layers or metallic mesh, can be used to protect infrastructures against rockfall events. The
rock energy is dissipated, largely due to the constituent material, and the mass of the structures. One innovative development is to construct the front face and possibly
the core and the back face of the structure using a cellular assembly, i.e. the gabion technique (Figs. 1 and 2). The gabion technique is commonly
Figure 1. Cellular assembly structure.
used for the construction of retaining walls, and noise and rockfall protection structures. A single cell element consists of a wire mesh cage, which is filled with materials
(e.g. site material such as gravel, stones, sand including an appropriate geosynthetic sack, or adapted material such as a composite material). The gabion cell may be
filled manually or using a mechanical method.
Figure 2. Example of gabion structure.
Page 14
Large deformations within cells and relative displacements between cells in contact may occur, making it possible to increase energy dissipation. Using such a technique
allows easy repair of the structure following rockfall events.
This research work consists of analyzing the mechanical behavior of such geocomposite structures subjected to a dynamic impact loading. The approach will be
based on a multiscale modeling (Bertrand et al. 2004). The presentation is focused on a cellular scale. Studies were performed to investigate the behavior of one cell
subjected to different types of loadings: quasistatic compressive tests and impact tests. The behavior of a single cell is modeled using the Discrete Element Method
(DEM).
2 MODELING OF THE CELL
2.1 Cell composition
The gabion cell is a cage composed of two systems; a double twisted hexagonal mesh and the filling material. In this study, the cell is filled with stones (Fig. 3). The wire
netting acts as a peculiar boundary condition that strongly influences the mechanical behavior of the whole cell. The function of the mesh is to maintain the assembly of
the stones during the construction of each cell, to confine the filling material and to give to the overall system a particular ability to be deformed.
2.2 Numerical modeling of the cell
2.2.1 The Distinct Element Method (D.E.M.)
The DEM used in this study was developed by Cundall and Strack (1979). Today, this numerical method is used in many geomechanical fields such as: granular flows,
response of concrete structures to dynamic loading, wave propagation in particle assemblies, mining studies, general geotechnical problems (Chareyre & Villard 2002,
Dong & Moys 2001, Tanaka et al. 2001, Masson & Martinez 1999).
Figure 3. Photograph of a cell filled with stones and encircle by the double twisted hexagonal mesh.
In this paper, all developments and simulations were carried out in three dimensions using PFC3D environment. The discrete elements are spherical particles, which are
considered as rigid bodies. All the particles can interact with one another by overlapping and their interactions are governed through the contact law of the particles.
The calculation is an iterative process involving two steps. The first step allows the contact forces in the assembly to be calculated using the laws of contact and
overlapping. The second step determines the new position of each particle using Newton’s law.
The contact behavior is characterized using a soft contact approach. In our case, the contact law is elastoplastic (within the elastic domain, the penetration between
two particles is proportional to the force and these two quantities are related by the normal stiffness, k n , and shear stiffness, k s, at the contact). The friction, f, between
two elements is taken into account in the model through the usual Coulomb law (Itasca 2003).
2.2.2 Modeling of the filling material
The physical properties of the filling material are the grain size distribution and the shape of the stones. Because of stones used to fill cells are not spherical, it is
important to take into account the shape of the stones in order to achieve an accurate physical description. Moreover, the grain size distribution must be controlled. The
numerical stones are modeled using the bonding of the particles (Fig. 4).
In this purpose, a cubiclike shape, which has the desired dimensions of a stone is created and filled with a fixed number of particles. The rotation of the cubic
element is performed to avoid, during the filling process, the formation of vertical columns. After this stage, the edges of each clump are cut to reduce angularity. At this
time, all the particles are clumped
Figure 4. Modeling of a stone (a) generation of a cubiclike box filled with clumped spherical particles, (b) rotation of the clump, (c) edges
cutting of the clump.
Page 15
together, which means that they cannot be dissociated during calculations. Numerical stones are now called clumps, like in the PFC terminology. The grain size
distribution is generated to obey to the French Standards (NFP 94–325–1). Actually, the dimensions requested are included between two extreme sizes, which are
calculated on the wire mesh geometry.
To fill the numerical cell, several clumps are generated just over the cell (Fig. 5) and are put into the cell by applying the gravity. Rigid boundary conditions (walls)
define the initial cell limits (500×500×500 mm3). Clumps are positioned without friction under increased gravity to optimize both the final porosity of the cell and the
calculation time. After positioning the clumps, all the particles located above the height of the cell are suppressed. Thus, the clump shape of the upper layer is slightly
modified.
2.2.3 Double twisted hexagonal mesh modeling
The hexagonal steel mesh is modeled by a set of particles located at the nodes of the mesh (Fig. 6). Each linear wire between two adjoining nodes is modeled by a
force existing between the two corresponding nodes. This force takes into account the mechanical behavior of the steel wire. This behavior is described by an elasto
plastic model with failure. After a given loading history, an unloading path gives rise to plastic strains
Figure 5. Filling process of a cell (a) generation of columns of clump over the cell, (b) cell filling by applying the gravity, (c) suppressing
of the upper particles.
(Fig. 7). The double twisted wires are modeled as two parallel wires. The calibration of the local interaction is performed thanks to experimental data coming from a
tensile test carried out on a wire specimen. The numerical curve is built with linear interpolation between points from the experimental curve (Fig. 7).
Equations of motion are solved using an explicit centered finitedifference scheme. As a consequence, the computed solution is reputed stable if the time step does
not exceed a critical value. The critical time step can be assessed by Equation (1) where m is the mass of the particle and k the stiffness of the particle.
(1)
The chosen critical time step corresponds to the smallest value of all time steps computed for each particle.
Two stiffness are considered in the modeling: the contact stiffness and the wire elastic stiffness. The interaction stiffness of the particles, which composed
Figure 6. DEM modeling of the hexagonal steel mesh.
Figure 7. Interaction model between wire mesh particles. The model take into account the unloading following the Young modulus of the
steel.
Page 16
the mesh, is characterized by the steel Young modulus (E), the section of the wire (S) and the initial distance between neighboring particles (Linitial); that allows to
define the stiffness k mesh of the mesh wires (k mesh=E.S/Linitial). Thus, Equation (2) needs also to be considered for the calculation of the global time step of the
system.
(2)
The time step determination is made as described above and is taken to be the minimum of all the computed critical time steps.
In order to validate the modeling of the wire mesh, numerical tensile tests are performed. The results obtained are compared with experimental data (Agostini 1988).
To carry out a numerical tensile strength test, a wire mesh sheet is generated. All the lower particles are fixed and a loading velocity is applied on the upper particles
(Fig. 8). The two vertical sides of the sheet are free to move during the mechanical test. From a qualitative point of view, the striction of the wire mesh is well
reproduced (Fig. 8). Moreover, the maximal tensile force obtained of the wire mesh sheet is in agreement with experimental data. However, it must be noted that the
numerical modeling overestimates the global deformation of the mesh before rupture (Fig. 9). This can be explained, at least in part, by the fact that the real mechanical
behavior of the double twist is not modeled in this presentation. Indeed, it is a strong hypothesis to consider the double twist as two parallel wires. Actually, during
manufacturing process, the twisting acts as a hardening operation and thus the steel becomes more brittle. As a consequence, the local failure of the mesh could occur
sooner as observed in this case. Taking into account this particular behavior of the double twist is now in progress.
3 NUMERICAL SIMULATION
3.1 Introduction
Two studies were carried out in order to investigate the ability of the numerical modeling. First, two kinds of tests were performed in quasistatic conditions: an
unconfined compression test and a confined compression test. Experimentally, a set of mechanical tests were performed on similar cells in the same conditions (Exp 1–
2 in unconfined condition, Exp 3–4 in confined condition). On the one hand, information is provided on the phenomena that occur during a quasistatic loading. One the
other hand, from a quantitative point of view, the experimental results give a set of data for the calibration of the model.
Second, the mechanical response of a cell under dynamic loading is analyzed. The influence of the
Figure 8. Evolution of the wire mesh shape during a numerical simulation of a tensile test. Note the representation of the loading velocity
on the first picture.
Figure 9. Mechanical tensile test carried out on a sheet of wire mesh. Comparison between numerical results and experimental data.
damping coefficient on the dissipative capacity of the cell is explored. Moreover, a cell impacted with the same boulder for different boundary conditions is also
presented.
3.2 Initial state
Reproducing the organization of the elements and the void ratio of a cell can be regarded as a major numerical problem. This problem is due to the difference between
the experimental and numerical ways of filling. The initial positioning of the elements is not the same. Indeed, experimentally, the stones are put into the cell one by one
to optimize the density of the cell. The numerical filling is quite different because the clumps are introduced into the cell only dropped under
Page 17
Table 1. Initial void ratio of the experimental and numerical aggregates.
Initial stat Num 1 Exp 1 Exp 2 Exp 3 Exp 4

Initial void ratio 46% 42% 43% 43% 43%
gravity. Furthermore, the numerical and experimental grain size distributions are similar but not exactly the same. Experimentally, the stones are slightly smaller than the
numerical ones.
Simulations were performed using an initial numerical state (Num 1). The numerical grain size distribution of the initial state Num 1 is calibrated on the French
Standards (NFP 94–325–1). Note that experimental initial void ratios are smaller than the numerical one (Table 1).
3.3 Energy balance
During a simulation, different kinds of energy can be tracked. In our case, we will focus our attention on the kinetic energy, the strain energy and the frictional work
dissipated into the system. The kinetic energy is computed as Equation (3).
(3)
(4)
Np is the number of particles, M(i) is the generalized mass and V(i) the generalized velocity of each particle. Product M(i) V(i) is defined as Equation (4) where ν (i) is
the velocity components of the particle, ω(i) the angular velocity components, m the mass, I the moment of inertia. This expression takes into account the energy due to
the translational and the rotational motion.
The elastic strain energy computed over all the contacts is given in Equation (5) where Nc is the number of contact and, are the magnitudes of the normal and
shear components of the contact force.
(5)
This energy can be considered as a potential energy because of the reversible feature of the elastic contact.
The frictional work dissipated during a time step is assessed by Equation (6) where are the shear force and the increment of slip displacement on the
contact c, over the current time step.
(6)
Figure 10. Comparison between strain energy and kinetic energy during a confined compression test realized on the initial state Num 1.
3.4 Compression tests
DEM numerical simulations are time consuming, especially in quasistatic condition. Indeed, velocities induced during an experimental loading are rather small. Thus, it
is expedient to introduce a local nonviscous damping to increase the numerical loading velocity and then to reduce the computation time. The effect of the local non
viscous damping on the particle motion is to decrease the resultant force applied by a fixed ratio on each particle. As a consequence, the acceleration of the particle is
reduced by the same order of magnitude.
The experimental loading velocity is of 0.08 mm∙s−1. For the numerical simulations, the loading velocity is of 100 mm∙s−1 for a nonviscous local damping of 70%. In
order to verify that inertial phenomena are not preponderant during simulation, we controlled (Fig. 10) that kinetic energy is quite small and can be neglected with
respect to the strain energy.
3.4.1 Confined test
In confined conditions, several simulations were performed. The influence of the microparameters, namely k n and k s, is highlighted in Figure 11. From a qualitative
point of view, the behavior of the cell is quite different from the experimental results. It can be observed that at the beginning of the loading (for small displacements),
numerical curve obtained with high values of k n and k s (5∙106 to 108 N∙m−1) match well experimental data. But when axial displacement increases, experimental curves
tend toward the numerical ones obtained using smaller local stiffness (under 3∙106 N∙m−1). This feature deserves some comments: stones cracking are clear
experimental evidence. During the loading, certain stones fail, which induces a reduction of the macroscopic stiffness. At present, this phenomenon is not taken
Page 18
Figure 11. Influence of local stiffness upon the macroscopic response of a cell subjected to a confined compression test.
into account by the numerical model. The part of the computed curve, corresponding to important axial displacements, is therefore wellfitted with smaller local
stiffness.
Moreover, it can be noted that the initial void ratio is not the same between experimental and numerical initial states (Table 1). As a first approximation, the higher
values of k n and k s are used thereafter.
3.4.2 Unconfined test
Unconfined compression tests are now the subject of analysis. The main difference with the mechanical test carried out in quasistatic conditions is the double twisted
hexagonal steel mesh incorporation. From a qualitative point of view, the numerical response in strain of the cell is in a good agreement with experimental data (Figs. 12
and 13).
Physical phenomena are well reproduced by the numerical modeling. The microparameters have a significant influence on the over all behavior of the cell. The
maximal force closely depends on the contact stiffness. It must be noted that the initial compactness of the cell and the initial positioning of the elements may have an
influence too.
3.5 Dynamic loading
3.5.1 Influence of the damping coefficient
For dynamic problems, the local nonviscous damping coefficient plays a very important role. Indeed, the effect of the local nonviscous damping on the particle motion
is to decrease the acceleration of the particle at each time step of a fixed ratio. Thus, for dynamic analysis, this parameter has to be very small.
In the aim to identify the influence of the local nonviscous damping coefficient on the system, two impacts of 4.6 KJ were performed in confined
Figure 12. Comparison between experimental photograph and numerical view for a compression test (Num 1 and Exp 1).
Figure 13. Comparison between experimental data and numerical results in the case of an unconfined compression test conditions (Exp 1,
Exp 2 and Num 1 for k =k =108 N∙m−1, 107 N∙m−1, and 5∙106 N∙m−1)
n s
condition (Fig. 14). The boulder generated is a spherical particle of 144 Kg in weight and a diameter of 50 cm. Its initial velocity is 8 m∙s−1. In the first case, the
damping coefficient is 0.01 and applied on all the particles. In the second one, the clumps are affected of a damping coefficient of 0.01 and the particles, which
composed the wire mesh, have a damping coefficient of 0.7. The boulder has a damping coefficient of 0.01 for the two cases.
Page 19
Figure 14. Deformation of the cell after impact at time 0.0186 s.
Figure 15. Influence of the damping coefficient. Time evolution of the contact force applied on the boulder and its velocity.
Figure 15 shows the differences concerning the evolution of the boulder velocity and the contact force applied to the boulder during the impact. Obviously, the
decrease of the boulder velocity is higher when the damping coefficient is 0.7 and the boulder rebound does not exist. The damping of the wire mesh particles has an
important effect on the overall system mechanical response. Indeed, it must be noted that not all the phenomena can be reproduced by the model if a high damped
numerical simulation is considered. Furthermore, from an energetic point of view, the damping causes a numerical energy dissipation, which is not taken into account in
the energy balance of the overall system. Figure 16 presents the evolution of the kinetic energy and of the frictional work during the impact. A priori, all the kinetic
energy has to be dissipated in frictional work and during plastification of the wire mesh. However, damping modifies the frictional work.
Therefore, the damping coefficient has to be very small for dynamic numerical simulations in order to
Figure 16. Influence of the damping coefficient. Time evolution of the kinetic energy of the boulder and of the frictional work in confined
condition.
Figure 17. Cell shape during the impact at time 0.1174 s. (Left=confined; Right=unconfined).
reproduce correctly the physical phenomena and to avoid numerical dissipation.
3.5.2 Influence of the boundary condition on the mechanical response of the cell
The purpose of this section is to give an idea of the abilities of the model to carry out dynamic parametric studies. The cell can be considered with different boundary
conditions. Two types of boundaries are considered: confined and unconfined conditions (Fig. 17). The boulder has the same characteristics as in § 3.5.1. The damping
coefficient is 0.01.
The main difference concerning the reaction of the cell is the generation of a rebound when the cell is confined (Fig. 18). The origin of the rebound is the elastic
energy storage during the compression of the clumps. The model likely overestimates the rebound. Physically, the cracking of the stones may dissipate a large part of
this energy. One way to improve the model is to consider fragmentation and thus reduce the
Page 20
Figure 18. Comparison of the mechanical response between confined and unconfined conditions. Evolution of the velocity and the force
acting on the boulder during the impact.
restitution of the elastic energy. In addition, the maximal magnitude is higher. The containment increases the cell stiffness. As a consequence, for the same impact
energy, the force reached is more important in confined conditions.
4 CONCLUSION
To establish the behavior of a composite material (cubic steel wire mesh cages filled with stones) both quasistatic and dynamic numerical simulations were performed.
In quasistatic conditions, parameters were calibrated from comparison between experimental and numerical results. As a first element of validation, a rather good
agreement can be noted between results concerning unconfined compression test. Simulation of an impact of a boulder on the cell allows us to investigate the
mechanical response in dynamic conditions. In such cases, the local nonviscous damping coefficient plays a very important role. It has to be very small in order to
reproduce correctly the phy sical phenomena and to avoid numerical dissipation of energy.
Nevertheless, complementary works are required to take into account phenomena such as fragmentation of elements to improve modeling relevance, especially in
confined conditions.
ACKNOWLEDGEMENTS
Work presented was completed within the framework of research network RNVO: Natural Hazards and Vulnerability of Structures. This network is accredited and
funded by the French Ministry of Research. It combines different laboratories in RhôneAlpes region and offers a wide spectrum of multidisciplinary competencies on
natural and technological hazards. This project is financed by France Gabion SA, part of the group Maccaferri. We hold has to thank particularly Mr. F.Derache, Mr.
M.Piolanti, Mr. B.Rossi and Mr. Ferralliolo in charge of the management of the research project, as all the members of the research group, especially S.Lambert, for
their contribution and interest they had in this work.
REFERENCES
Agostini R., Mazzalai P., Papetti A. 1988. Filets métalliques à mailles hexagonales pour la protection des falaises et talus, Officine Maccaferri S.p.a, Bologne, Italy
Chareyre B., Villard P. 2002. Discrete element modelling of cuved geosynthetic anchorages with known macroproperties. In H.Konietzky et al. (eds) Proc. of the First
International PFC Symposium, Gelsenkirchen, Germany, 197–203
Cundall P.A., Roger D.H. 1992. Numerical modelling of discontinua, Engineering Computations, (9):101–113.
Cundall P.A., Strack O.D.L. 1979. A discrete numerical modelling method for granular assemblies, In Géotechnique, 29:47–65.
Bertrand D., Gotteland Ph., Lambert S., Nicot F., Derache F. 2004. Multiscale mechanical modelling of unusual geocomposite material under the impact of rockfall. In
Proc. of Ninth international symposium on numerical Models in Geomechanics, Ottawa, Canada, 6pp.
Dong H., Moys M.H. 2001, Assessment of the discrete element method for one ball bouncing in a grinding mill. International Journal of Mineral Processing, 65: 213–
226.
Itasca Consulting Group, Inc. 2003. PFC3D—Particle Flow Code in Three Dimensions, Vers. 3.0 User’s Manuals, Minneapolis: Itasca.
Lambert S., Gotteland Ph., Plé O., Bertrand D., Nicot F., Modelling of the mechanical behaviour of confined material cells, Journées nationales de géotechnique et de
géologie, Lille, France, 8pp
Masson S., Martinez J. 1999. Effect of particles mechanical properties on silo flow and stresses from distinct element simulations. Powder Technology, 190:164–178.
Mustoe G.G.W., Huttelmaier H.P. 1993. Dynamic simulation of a rockfall fence by the discrete element method. Microcomputers in Civil Engineering, 8:423–437.
NFP 94–325–1. 2004. Norme Française: Exécution des travaux géotechnique spéciaux—Ouvrages en gabion—Partie 1: Hors sites aquatiques, AFNOR, St Denis la
plaine.
Tanaka K., Nishida M., Kunimochi T., Takagi T. 2001. Discrete element simulation and experiment for dynamic response of twodimensional granular matter to the impact
of a spherical projectile. Powder Technology, 124: 160–173.
Page 21
Microstructure modeling of granular media and interactions between pile and
granular ground by DEM
K.Maeda
Department of Systems Engineering and Management, Nagoya Institute of Technology, Nagoya, Japan
T.Sakurai, A.Oomura & H.Hirabayashi
Department of Civil Engineering and Management, Nagoya Institute of Technology, Nagoya, Japan
ABSTRACT: Recently, many kinds of piles with fairly rough skin are developed and used in geotechnical practice. However, the mechanism of
interaction between the pile and the ground made of granular media has not been clarified sufficiently, due primarily to the plasticity and nonlinearity of
soils and the localization of deformation. Granular media show nonlinear deformation behaviors depending on density and stress histories. Grain
properties (e.g. grain size and grain shape) and evolutioncollapse of microstructures formed in granular media control the macro behaviors. Firstly, the
density and stresslevel dependencies of deformationfailure behaviors were calculated by Distinct Element Method (DEM) in two dimensions, where
noncircular grains were employed. Fabric tensor, coordinate number and stability of microstructure were investigated in order to clarify the
relationships between macroscopic and microscopic phenomena. Secondly, the interaction behaviors between piles with different skin roughness and
grounds with different conditions were analyzed. It was found the rule of stressinduced anisotropy was presented as a function of stress ratio, and the
nonlinearity of interaction between skin friction and point resistance around pile increased with skin roughness of pile.
1 INTRODUCTION
First, granular material shows nonlinear deformation behavior that is dependent on density and stress history Moreover, even if they are under same condition, the
mechanical properties of soils with different grain sizes and grain shapes should be intrinsically different. Grain shape influences macroscopic deformation properties
thorough variations stability against slip and deformability at contact points. In a series of papers (Maeda et al. 1995, Miura et al. 1997, Miura & Maeda 1999a, b,
Miura et al. 1997, 1998), the relationships between physical and mechanical properties with variations in confining stress and relative density were investigated
extensively for granular samples with different primary properties (e.g. density, hardness, grain shape and grain size). From test results, it was revealed that the primary
and physical properties of sand significantly influence its mechanical behavior and some reliable mutual relationships of physical and mechanical properties were
identified. An estimation method of mechanical properties from some index physical properties was also discussed. Void ratio extent (emax−emin) which increases with
increasing angularity of grain shape and decreasing grain size, was selected as a promising parameter for evaluating the ductile deformability, the degree of confining
stress and relative density dependencies. In that paper series, it was indicated that grain crushability, which increases with angularity and grain size, also induces further
reduction of shear strength with increasing confining stress. Therefore, the effects of grain shape on the deformationfailure behaviors of granular material should be
taken into account when simulating the ductile and compressive deformation behavior of soils by the Discrete Element Method (DEM); noncircular particle elements
were employed in this paper.
Secondly, the macroscopic behavior of granular materials is controlled by the evolution and collapse of microstructures. Stressinduced anisotropy in granular
materials under shear has been observed in model tests and simulations (e.g. Oda et al. 1985), and elastoplastic models with a complete evolution model for anisotropy
have been proposed (e.g. Nakai & Mihara 1984). The mechanism and evolution of fabric in granular materials therefore represent an important subject in geo
mechanics. In this paper, the geometric parameters such as the fabric tensor of stressinduced anisotropy and the coordinate number are investigated.
Finally, in geotechnics, the vertical bearing capacity of a pile is thought to be able to be estimated by
Page 22
superposition of the ultimate skin friction resistance and point bearing capacity. Spring models described in terms of coefficient of subgrade reactions are commonly
used to express the relationships between pile resistance and settlement. The reactions of ground materials, however, are remarkably nonlinear, and the interaction
between ground and pile is complex. The occurrence of large deformation and deformation localization enhance these tendencies. To establish design code that
accounts for ground deformation, it is therefore necessary to consider the nonlinearity of the pileground system directly and then examine the interaction mechanism
(e.g. Nakai et al. 1991). Moreover, these discussions on bearing capacity must be valid for discussions on sounding of ground by penetration; the last is inverse
problem of the former.
In this paper, the dependence of vertical pile bearing behavior on the density, stress and skin roughness of the constituent granular material is analyzed by DEM. In
these analyses, both circular and noncircular particles are considered in order to simulate the compressive and ductile behavior of deep ground. The deformation
failure behavior of soil around pile, deformation localization, and displacement increment vectors are then discussed based on the analysis results. It is revealed that the
compressibility below the pile point and the interaction between end bearing and skin friction mobilization are important factors contributing to the bearing capacity of
pile.
2 PROCEDURES AND CONDITIONS
2.1 DEM using noncircular particle
The deformation behavior of granular materials was analyzed by the DEM (Cundall & Strack 1979). The equilibrium contact forces and displacements of grains in the
stressed assembly were calculated individually at small time steps on the basis of Newton’s second law. The interactions between elements were modeled by contact
elements between rigid disks (e.g. springs, dashpots, sliders and nonextensional elements), as shown in Figures 1a and b. Calculations were conducted using PFC2D
code.
Figure 1. Contact elements at contact points in DEM: (a) normal direction, (b) tangential direction.
Circular (Fig. 2a) and noncircular (Figs 2b–e) particle elements were considered. For example, the noncircular particles shown in Fig. 2b were prepared by
connecting three circular particles of the same radii as a “clump particle” in the code. Simulations were performed for a material composed entirely of circular particles,
and then all circular particles were replaced by noncircular particles, where the circular particles are shown as broken lines in Figures 2c, d and e and circumscribe
noncircular particles. The parameters for particle elements used are described in Table 1.
2.2 Biaxial compression test procedure
Figures 3a & b show the specimen models employed for simulations of biaxial compression tests under isotropic compression and shear failure states. The analysis
involved approximately 4000 disc particles and four wall boundary elements, the movement of which was employed for external stress and strain control, as shown in
Figure 3, where gray balls and
Figure 2. Particles types used in the DEM: (a) c101 (circle), (b) c103 (triangle), (c) c104 (quadrate), (d) c106 (hexagon), (e) c108
(octagon), where broken line is circumscribing circle.
Table 1. Analysis parameters of particle elements in DEM.
Parameters ParticleParticle
3 2700
Density (kg/m )
Diameters (mm) 5 to 10
Normal spring coefficient
5×108
kn (N/m/m)
Tangential spring coefficient
5×107
ks (N/m/m)
Damping Critical damping
Resistant friction angle (deg.)
Page 23
black lines denote particles and the network of contact force transmission. The biaxial compression tests were simulated under zero gravity conditions to investigate the
change in fabric due to the change in macro stresses.
Specimens were modeled as follows. First, circular particles with initial resistance friction coefficient tan (= 0.25 in this simulation), and the initial void ratio e0
was obtained after stabilization under σm0.
Figures 4a and b show the change in the void ratio e0 with tan and n0=0.10. Therefore, minimum and maximum mechanical void ratios emax and emin
can be obtained by simulating both end conditions. Stresses were then varied to the prescribed consolidation stress. The analysis cases are shown in Table 2.
2.3 Vertical loading test procedure for pile with rough skin surface
Figure 5 shows the configuration for the vertical penetration test of a pile with constant overburden pressure under a one gravity (1×g) condition. The analysis
Figure 3. Specimens model for biaxial compression test simulation by DEM: (a) isotropic stress, (b) shear failure.
was performed on a halfspace taking account of the symmetry of the problem. The initial depth of the pile point was 20 times the pile radius (D/2; pile diameter D) and
was about 30 times the mean grain size (D50) of the soil. The width and depth of the analysis region were 50 and 40 times D/2, respectively. The soil was composed of
approximately 70,000 disc particles, and the density of the soil was determined in the same manner as for the biaxial compression test. The wall elements were
employed as boundary and pile elements, where the boundary wall was frictionless. The overburden pressure σv was maintained at the required

Figure 4. Void ratio e0 after conversion from initial resistant friction angle tan (=0.25): (a) initial porosity n 0=0.40, (b) n 0=0.10.
Table 2. Analysis cases of biaxial compression tests.
Grain shape Constant confining pressure Constant volume

Circular e0=0.20 –
–
σc0=2kn×10−4
Noncircular Density
Dense: e0=0.20 Dense: e0=0.20
Medium: e0=0.23 Medium: e0=0.22
Loose: e0=0.27 Medium: e0=0.23
– Loose: e0=0.27
Consolidation stress

σc0=2kn×10−4 σc0=2kn×10−4
–
10 kn×10−4
–
40 kn×10−4
Page 24

value by adjusting the upper boundary wall. The void ratio at σv0=k n ×10−4 was defined as the initial void ratio e0.
The skin friction of the pile (tan δ) and the geometric roughness (R) were changeable as shown in Figure 6. The geometric roughness was modeled by fixing circular
particles to the skin of the pile, and the velocities of particles were tied closely to that of the pile. The roughness was defined as R=(Rmax−Rmin), where the maximum
roughness (Rmax) and minimum roughness (Rmin) correspond to the maximum and minimum diameters of the surface particles, respectively.
Figure 5. Configuration of vertical loading test of pile under overburden pressure.
The analysis cases are listed in Table 3. Four cases of skin roughness were considered. The “smooth” case (R1) assumed a geometric roughness of R=0 and skin
friction coefficient of tan δ=0. R2 was a slightly rougher case with R=0 and tan δ=0.25, while R3 and R4 had roughness R equal to one and two times D50,
respectively.
2.4 Estimations of micro stress and strain in soil mass
In DEM analysis by PFC2D, the contact forces and particle displacements are computed on a microvolume of soil mass V as shown in Figure 7, but these discrete
quantities cannot be involved directly in a continuum model. An averaging procedure is necessary to transfer the discrete quantities on the microscopic scale to
Figure 6. Definitions of skin friction and geometric roughness: smooth surface and rough surface with particles fixed to smooth surface.
Table 3. Analysis cases of pile penetration tests.
Case no. Grain shape Initial density (e0) Friction coefficient of surface, tan δ Relative surface roughness, R/(D50) Overburden pressure σv (MPa)

CaseS1 Circle 0.16 0.25 0 0.1
CaseS2 Circle 0.16 0.25 0 0.5
CaseS3 Circle 0.16 0.25 0 2.0
CaseD1 Noncircle Loose, 0.27 0.25 0 0.1
CaseD2 Noncircle Dense, 0.20 0.25 0 0.1
CaseLR1 Noncircle Loose, 0.27 0 0 (absolutely smooth) 0.1
CaseLR2 Noncircle Loose, 0.27 0.25 0 0.1
CaseLR3 Noncircle Loose, 0.27 0.25 1.0 0.1
CaseLR4 Noncircle Loose, 0.27 0.25 2.0 (fairly rough) 0.1
CaseDR1 Noncircle Dense, 0.20 0 0 (absolutely smooth) 0.1
CaseDR2 Noncircle Dense, 0.20 0.25 0 0.1
CaseDR3 Noncircle Dense, 0.20 0.25 1.0 0.1
CaseDR4 Noncircle Dense, 0.20 0.25 2.0 (fairly rough) 0.1
Page 25
macroscopic stresses and strains. To calculate average values and investigate the microstructures formed in materials, a program was developed to conduct the
following simulation for an arbitrary shape V.
Both stress and strain are represented by three mode components, where the subscripts p and s denote pure shear and simple shear modes, respectively,
(1)
(2)

where σm is the mean normal stress and εν is volumetric strain in two dimensions. The maximum shear stress τm and strain γm are given by,
(3)
The averaged stresses on a particle (p) are calculated on the basis of the assumption that the stresses in each particle over V are continuous. Since each particle
is subjected to point forces at discrete contacts as shown in Fig. 7, the stresses in V can be calculated by
(4)
where vectors present the location and the force acting at contact (c), respectively (e.g. Oda & Iwashita 1999, Itasca 2002). The stress tensor σij averaged in
V, can be computed by volumetric average for all particles (p) in V as follows;
(5)

where Np is the number of particles contained in V.
Figure 7. Micro volume element V including particles.
The average strain rates in V can be determined by comparison between the discrete velocity field of particle centers ν (p)(xi) and the continuous velocity field ν(x i)

predicted on the assumption that linear interpolation is valid for the velocity field in V(e.g. Oda & Iwashita 1999, Itasca 2002). The differences between the velocities
of two neighboring particles are given by,
(6)
(7)
where as follows;
(8)
(9)
The tensor can be computed by a leastsquares procedure to minimize the residual error χ between the velocity value ν (p)(xi) calculated by DEM and the

predicted velocity value ν(x i) of the particles in V as follows;
(10)
(11)
where
(12)
Here, are mean velocities. Equations (10) to (12) give the minimization operation of error χ, which is the sum of the square deviation between the relative velocity
calculated by DEM and the predicted relative velocity. All four components of the strain rate are obtained in this way. The strain tensor components can also be
computed in a similar fashion.
3 RESULTS AND DISCUSSIONS OF BIAXIAL COMPRESSION TESTS
3.1 Macro deformation—failure behaviors
Figure 8 shows the stressstraindilatancy behavior of the specimen with circular particles under constant confining stress. The dilative volume change is remarkable and
stress ratio at failure is small; these
Page 26
tendencies mean that the assembly of circular particles is brittle.
Figures 9 and 10 show the confining stress dependence and density dependence of deformation behavior in the case of an assembly of triangular particles (cl03).
The relationships between the void ratio and the mean normal stress under isotropic compression and shear for selected particle shapes (cl01, cl03 and
Figure 8. Deformationfailure behaviors under constant confining stress, for circular particles (cl01): circular particles.
Figure 9. Stress level dependence of deformationfailure behavior under constant confining stress, for densely packed noncircular
particles.
cl06) are presented in Figures 11a–c. As observed in experiments (e.g. Fukushima & Tatsuoka 1984), the specimens with lower density exhibit a lower stress ratio at
failure and higher contractive deformation under higher confining stress, indicating ductility.
As shown in Figures 11a–c, there seems to be a critical state region (indicated by a broken line) in which the shear strain is large. However, the difference between
the critical state lines for dense and loose specimens is significant. This indicates that since deformation localization tends to occur in dense specimens, it may be
possible to investigate the relationships between the void ratio e and the mean normal stress σm for both the specimen as a whole and locally in the shear band (Wong
1999).
Figures 12a–c show the density dependence of deformation behavior under the constant volume condition in terms of the stressstrain curve, the reduction in
effective mean normal stress, and the stress path. In the case of the lowest density (e0=0.27), it can be found that brittle deformation occurs, causing fluid flow and
reducing the effective mean normal stress remarkably. These tendencies indicate the occurrence of static liquefaction. This fluidity reduces with increasing density.
Figure 10. Density dependence of deformationfailure behavior under constant confining stress for noncircular particles.
Figure 11. Relationships between void ratio and mean normal stress under isotopic compression and shear, where IC means isotropic: (a)
c101 (circular particle), (b) c103 (noncircular particle), (c) c106 (noncircular particle).
Page 27
Figure 12. Density dependence of deformationfailure behavior under a constant volume condition, for noncircular particles (c103),
corresponding to ‘undrained test’: (a) stress ratio and strain, (b) reduction of effective mean normal stress, (c) stress path.
Figure 13. Decrease in coordinate number with increasing shear deformation under constant confining pressure.
These tendencies of deformation failure have also been observed in experiments (Ishihara 1996). The density and stress level dependencies can therefore be simulated
well through the use of noncircular particles in the DEM, demonstrating that grain shape represents an important factor determining the compressibility and ductility of
soils.
3.2 Evolution of fabric in prefailure specimen
Figures 13 and 14 show the changes in coordinate number under constant confining pressure and at constant volume, respectively. The coordinate number decreases
remarkably from 3–4 to about 2.5 with the progression of shear. When Nc is less than 2.5, the stiffness remains low and is not restored. As the coordinate number
appears to represent the stability and restorative ability of the contact paths, the specimen with high Nc has a highly stable fabric (Oda 1997). Thus, Nc indicates the
degree of percolation of microstructures in granular media.
The stressinduced anisotropy was investigated based on the fabric tensor (Satake 1982), as defined by,
(13)
Figure 14. Decrease in coordinate number with increasing shear deformation at constant volume.
Figure 15. Increase in principle fabric tensor ratio F1/F2 with principal stress ratio σ1/σ2 for the case of shear under constant confining
stress.
where θ(k ) is the angular direction of the kth contact plane from the x axis, and M is the total number of contact points. The major and minor principal values are
denoted by F1 and F2, respectively. Figures 15 and 16 show the relationships between the anisotropy index F1/F2 and the principal stress ratio σ1/σ2 for all cases
under prefailure deformation. The index F1/F2 increases monotonically with σ1/σ2. Over all the various cases with different grain shapes, densities,
Page 28

Figure 16. Increase in principal fabric tensor ratio F1/F2 with principal stress ratio σ1/σ2 for the case of shear at constant volume.
Figure 17. Distribution of velocity vectors of particles in a loose specimen composed of noncircular particles (c103): (a) initial shear, (b)
postpeak deformation.

confining stress levels and volume constraint conditions, the relationship between F1/F2 and σ1/σ2 can be approximated by the following unique relation,
(14)
Thus, it can be said that the evolution of stressinduced anisotropy is dependent on the square root of principal stress. This relationship is useful for developing a
constitutive model. The relationship was first proposed by Satake (1982), and has subsequently been utilized for development of the concept of modified stress (Nakai
& Mihara 1984; Tobita & Yanagisawa 1992).
3.3 Deformationfailure behaviors and evolution of fabric in micro volume element
Figures 17 and 18 show the velocity vectors of particles loose and dense specimens (c103), where the magnitude of velocity is normalized by the velocity
Figure 18. Distribution of velocity vectors of particles in a dense specimen composed of noncircular particles (c103): (a) initial shear, (b)
postpeak deformation.
Figure 19. Changes in micro void ratios outside and inside the shear band during macro deformation of a dense specimen (c103).
of the boundary wall. The deformation of very loose specimens is uniform under shear, even after failure, whereas in dense specimens, deformation localization (shear
band) occurs.
To investigate macro and micro deformations, the specimen was divided into 100 rectangular micro elements V, and the void ratio, stress and fabric in V were
calculated. Figure 19 shows the micro and macro void ratios against the macro strain for a dense specimen (c103), where open and closed circles denote the behavior
outside and inside the shear band, respectively. In the prefailure state (εyy<4%), the micro void ratios are the same as the macro void ratios, indicating that the
specimens are homogeneous. However, at postpeak (εyy>4%) deformation, the void ratio outside shear band decreases or does not change with macro deformation,
whereas the inside void ratio increases as to attach the value of macro void ratio on the critical state for loose specimen. A similar result was observed
Page 29
Figure 20. Relationship between micro void ratio e and micro stress.
Figure 21. Change in micro stress ratio.
in computed tomography experiments (Desrues et al. 1996). Figure 20 shows the relationship between the micro void ratio and the micro mean normal stress σm. The
micro void ratio decreases with decreasing σm in the shear band, whereas the micro stress ratio in the shear band increases remarkably and the direction of major
principal stress rotates, as shown in Figure 21.
Figure 22 shows the changes in micro fabric tensor. The relationship between F1/F2 and σ1/σ2 inside and outside the shear band can also be approximated by the
same manner as for the macro case (eq. (14)).
4 RESULTS AND DISCUSSIONS OF PILE PENETRATION TESTS
4.1 Effect of overburden pressure
In Figures 23a and b, the skin friction resistance following mobilization of the skin surface (tan δm) and the
Figure 22. Micro stressinduced anisotropy.
Figure 23. Resistance behavior of a pile at different overburden pressures for the case of circular particles: (a) skin friction resistance,
(b) point resistance.
point resistance normalized by the overburden pressure (qp /σv ) are plotted against the relative settlement S/D under different σv conditions. The skin friction resistances

reaches ultimate values of tan δ=0.25 with
Page 30

Figure 24. Distribution of particle velocity around a pile at S/D=0.5: (a) σv=0.1 MPa, (b) σv=2 MPa.

smaller settlement than for the point resistances. As shown in Fig. 23b, qp /σv , which represents the bearing capacity coefficient, decreases with increasing σv .
Figures 24a and b show the distributions of velocity. From Fig. 23, under low σv conditions (0.1 MPa), the velocity vectors are directed downward, laterally and
slightly upward, whereas at high σv (2 MPa), the vectors are uniformly directed downward. Therefore, the range of the velocity distribution at large deformation
decreases with increasing σv . Since the dilatancy properties shift from dilative to contractive with increasing σv (Maeda & Miura 1999a), large deformation seems to
have difficulty propagating outward at higher stress levels compared to lower stress levels.
4.2 Effect of density
Figures 25a and b show the resistances of pile without geometric roughness at different densities. The resistance qp /σv increases with density.
Figures 26a and b show the velocity distribution around the pile. The range of the distribution decreases with decreasing density. This tendency is similar to that due
to the stress level. Since the dilatancy property shifts from dilative to contractive with decreasing density (Maeda & Miura 1999b), it is more difficult under loose
conditions for the deformation to propagate extensively outward.
4.3 Effect of surface roughness
Figures 27a and b show the resistance of a loose pile with different surface roughnesses. Here, for qp /σv , a smaller range of S/D is considered. The skin resistances tan
δm reaches ultimate values at greater settlement in the cases of LR3 and LR4 compared to cases LR1 and LR2. The ultimate values increase with roughness, and for
cases LR1 and LR2, the ultimate values equal the surface friction coefficients tanδ employed. For cases LR3 and LR4, the ultimate values are larger than tan δ and the
failure strength of the soil element, as shown in Fig. 9. In particular,
Figure 25. Resistance behavior of a pile at different densities for the case of noncircular particles: (a) skin friction resistance, (b) point
resistance.
Figure 26. Distributions of particles velocity around pile at S/D=0.5: (a) loose, (b) dense.
tan δm is very high for case LR4, which has geometric roughness equal to twice the particle radius of the soil. The occurrence of extremely high friction seems to be
controlled by simple shear deformation localization around the skin surface as well as earth pressures.
Page 31
Figure 27. Resistance behavior of piles with different surface roughnesses for the case of noncircular particle and loose packing: (a) skin
friction resistance, (b) point resistance.
The point resistance qp /σv seems to decrease with increasing roughness at S/D<0.02, and the value of qp /σv fluctuates with skin friction mobilization. At S/D>0.04, the

point resistances are almost the same, increasing slightly with roughness, particularly in cases LR3 and LR4. For the case of high roughness and lesser settlement, skin
friction bears the majority of the load applied to the pile, while the point resistance is not active.
Figures 28a and b show the distribution of σm around the pile at S/D=0.5 for smooth and rough cases. A stressfocused region with a bulb shape can be seen, the
depth and width of which are higher in the rough case than in the smooth case. Figures 29a and b show the velocity vector distributions. For the rough case (LR4), the
vectors are distributed below the pile point and also around the skin surface and laterally upward. Therefore, the range of shear stress mobilization increases with
roughness. The particles on the rough surface were dragged down during large settlement. Thus, if particles are dragged down the pile by roughness come to bear on
the ground around the pile point, resulting in increased overburden pressure. A similar tendency was observed for the dense case.
Figure 28. Distibution of mean normal stress σm (kPa) around pile at S/D=0.5 for loose case: (a) absolutely smooth (LR1), (b) fairly
rough (LR4).
Figure 29. Distributions of particle velocity around pile at S/D=0.5 for loose case: (a) absolutely smooth (LR1), (b) fairly rough (LR4).
These nonlinear interactions between skin friction and point resistance indicate that the total resistance cannot be obtained by simply superimposing separately
estimated resistances (Meyerhof 1951, Terzaghi 1965).
5 CONCLUSIONS
Twodimensional DEM simulations were performed for piles of circular and noncircular particles. The behavior on microscopic to macroscopic scales was
investigated through simulations of biaxial compression and pile bearing behavior. Conclusions on the basis of analysis results are summarized as follows:
1. It was found that assemblies of circular particles are more brittle than assemblies of noncircular particles, the latter of which are ductile and exhibit considerable
dependencies on density and stress level. This particle shape effect is considered to be induced by particleinterlocking due to rotation resistance.
2. The coordinate number was found to increase under isotropic compression and to decrease under shear,
Page 32
representing the stability and restorative ability of the contact paths (related to percolation theory).
3. The principal fabric tensor ratio increases with the principal stress ratio. The relationship is independent of grain shape, density and confining pressure, and can be
approximated by the square root power relation given by eq. (14).
4. In the shear band of a dense specimen, the micro void ratio increases due to a decrease in the mean normal micro stress with increasing stress ratio and anisotropy
(eq. (14)) and with stress rotation.
5. Using noncircular particles, the dependence of pile resistance behavior on density and stress level can be reproduced. The size of the shear failure zone around the
pile decreases with increasing stress level and with decreasing density, controlled by the dilatancy properties of the soil.
6. Friction mobilization on the skin surface increases with the skin friction coefficient and geometric roughness. Particularly in the case of large geometric roughness, the
ultimate friction value is greater than both the skin surface friction and the shear strength of the soil mass, attributable to earth pressure generated by the roughness.
7. At small settlement, point resistances seem to decrease with increasing skin roughness, where the external load is taken up mainly by skin friction while point
resistance is relatively inactive. However, at relatively large settlement, the point resistances remain the same or even increase with increasing roughness. This
behavior is attributed to an overbearing effect due to particles dragged down by the roughness. These results indicate that the interaction between skin surface
friction and point resistance is nonlinear, implying that the total resistance cannot be determined by superimposing separately estimated resistances.
8. Next interesting subjects for piles; Piles with other shapes, such as bell and nodes, have also been investigated (e.g. ACI and BIS). However, the mechanism of
bearing behavior has yet to be revealed. In addition to the interaction between skin friction and point bearing, multiple interactions among the pile, the soilcement
interface and surrounding ground are also important (Hara 2003, Sakamoto 2004). If cracks and disruptions occur in the soilcement interface, interaction between
the dilatancy properties of the soilcement interface and the stiffness of the soil will increase the complexity of the problem. Therefore, discrete analyses will be
necessary to reveal the mechanism and facilitate the development of improved piles.
REFERENCES
American Concrete Institute Committee 336, 1998. Design and Construction of Drilled Piers.
Bureau of Indian Standards, National Building Code, Structural Design, Section2, Foundations.
Ishihara, K. 1996. Soil behavior in earthquake geotechnics. Oxford: Oxford Science Publications.
Cundall, P.A. 1971. A Computer Model for Simulation Progressive, Large Scale Movement in Blocky rock system. Symp. ISRM, Vol. 2:129–136.
Desrues, J. et al. 1996. Void ratio evolution inside shear bands in triaxial sand specimens studies by computed tomography, Geotechnique, Vol. 46, No. 3, pp. 529–546.
Fukushima, S. & Tatsuoka, F. 1984. Strength and deformation characteristics of saturated sand at extremely low pressure. Soils and Foundations, Vol. 24, No. 4:30–48.
Hara, Y. 2003. Bearing behavior of piles with different skin shapes, Master thesis, Nagoya Institute of Technology.
Maeda, K., Miura, K. & Toki, S. 1995. Mechanical properties of elliptic microstructure formed in granular materials. Soils and Foundations, Vol. 35, No. 2:1–13.
Maeda, K. & Miura, K. 1999a. Confining stress dependency of mechanical properties of sands. Soils and Foundations, Vol. 39, No. 1:53–68.
Maeda, K. & Miura, K. 1999b. Relative density dependency of mechanical properties of sands. Soils and Foundations, Vol. 39, No. 1:69–80.
Meyerhof, G.G., 1951. The ultimate bearing capacity of foundations, Geotechnique, Vol. 1:301–332.
Miura, K., Maeda, K. & Toki, S. 1997. Method of measurement for the angle of repose of sands, Soils and Foundations, Vol. 37, No. 2:89–96.
Miura, K., Maeda, K., Furukawa, M. & Toki, S. 1997. Physical characteristics of sands with different primary properties, Soils and Foundations, Vol. 37, No. 3: pp. 53–64.
Miura, K., Maeda, K., Furukawa, M. & Toki, S. 1998. Mechanical characteristics of sands with different primary properties. Soils and Foundations, Vol. 38, No. 4: 159–
172.
Nakai, T. & Mihara, Y. 1984. A new mechanical quantity for soils and its application to elastoplastic constitutive models. Soils And Foundations, Vol. 24, No. 2:82–94.
Nakai, T., Sakai, S., Sumi, H. & Yamada, Y. 1991. Analysis of vertically loaded pile in sand and clay. Proceedings of the international geoengineering for coastal
development Geocoast ’91, Yokohama, 3–6 September: 707–712.
Oda, M. 1977. Coordination number and its relation to shear strength of granular material. Soils and Foundations, Vol. 17, No. 2:29–42.
Oda, M., NematNasser, S. & Konishi, J. 1985. Stressinduced anisotropy in granular masses. Soils and Foundations, Vol. 25, No. 3:85–97.
Oda, M. & Iwashita, K. 1999. An Introduction mechanics of granular materials. Rotterdam: BALKEMA.
Sakamoto, K. 2004. Analytical study on bearing behaviors of nodular pile with account for interaction between pile and soilcement. Graduated thesis, Nagoya Institute of
Technology.
Satake, M. 1982. Fabric tensor in granular materials. IUTAMConference on Deformation and Failure of Granular Materials: 63–68.
Tobita, Y. & Yanagisawa, E. 1992. Modified stress tensors of anisotropic behavior of granular materials. Soils and Foundations, Vol. 32, No. 2:85–99.
Page 33
Slope and wall stability
Page 34
Page 35
Application of PFC3D for determination of soil properties and simulation of the
excavation process in front of sheet pile wall constructions
M.Pohl, M.Pulsfort & B.Walz
Institute for Geotechnical Engineering, University of Wuppertal, Wuppertal, Germany
ABSTRACT: Within PFC3D (Particle Flow Code in three dimensions) granular materials such as soils are simulated by balls. Depending on the ball’s
size and the geometry of the container stochastic effects regarding the pore content and the inner friction angle of the numerical soil occur. For specific
numerical soils, reproducible ball parameters are determined in dependence on different container geometries. The simulation of the excavation
process is conducted in front of unsupported sheet pile walls, single strutted sheet pile walls and single strutted sheet pile walls with the allowance for
the formation of a plastic hinge. The resulting displacements, bending moments and transverse forces within the wall and the resulting earth pressures
acting on the wall are determined as well as the behavior within the soil for subsequent excavation steps. The wall is simulated by parallel bonded balls.
1 INTRODUCTION
When generating balls randomly above a container and pouring them repeatedly into the container, different pore contents and varying inner friction angles of the
agglomeration of balls in the container can be observed, when taking the ball parameters as constant. It is expected that these stochastic effects depend on the ratio
“balls per container wall length”. This relation will be determined for satisfying reproducible agglomeration states.
For the simulation of a specific agglomeration of balls, the ball parameters have to be determined by every user anew. The necessary microproperties of the balls—
depending on different container geometries—are derived from the simulation of specific soils with inner friction angles φ=30° to 40° and pore contents n=0.40.
Mathematical functions are presented, which enable the user to estimate the necessary ball parameters with good accuracy.
The simulations of the excavation process in front of sheet pile wall constructions are performed with the predominant goal to determine the earth pressures acting on
the wall in dependence on different static systems. In this respect, the earth pressures resulting from sheet pile walls in the state of plastic deformation are studied in
detail. Aside from the earth pressures, the displacements, bending moments and transverse forces over both walls height and width are determined for all excavation
steps.
2 DETERMINATION OF REPRODUCIBLE NUMERICAL SOIL PROPERTIES
2.1 Pore content
A container (width=4.70 m, height=3.00 m, depth= 0.70 m) is filled with balls (diameter=20 cm). The balls are generated above the container and repeatedly poured
into the container under normal gravitational acceleration. The user defined ball parameters (normal stiffness, shear stiffness, surface friction) are kept constant for all
simulations. The simulated states are: pouring of balls into the container, balls at rest (earth pressure at rest) and balls exposed to the displacement of one front wall
(active earth pressure).
Simulations by Herten & Pulsfort (1999)—parallel displacement of one front wall—demonstrated, that the soil behavior can be represented accurately using uniform
particles. Simulations with particle size distributions yielded the same results, but were more time consuming. Thus, it can be deduced that no regular packing of the
randomly generated uniform particles occurred. Such crystalline structures are more likely to happen in twodimensional simulations. The effect of regularly packed
particles is presented in chapter 4.1.1, Figure 14 of the present paper.
The maximum density of randomly close packed identical spheres has been proved by Torquato et al. (2000) to a pore content n=0.36. Monte Carlo schemes even
yield n=0.32. A very dense graded
Page 36
granular soil has a pore content of about n=0.365. Thus, soils can be modeled in general by uniform particles. For the determination of a ratio “(uniform) balls per wall
length” it is applicable to use uniform particles anyhow. Different pore contents of numerical soils can be achieved by varying the surface friction of the balls and their
shear stiffness, respectively.
The occurred pore contents in the described slender container are shown in Figure 1. A varying between 0.388≤n≤0.413 is measured. A mean value of the pore
content nm=0.4011 with a variation coefficient vc=0.019 and a standard deviation sd =0.0076 results. When taking the agglomeration of balls as a simulated specific soil
with a grain bulk unit weight γs =26.5 kN/m3, a deviation of ∆γd =0.67 kN/m3 of the dry bulk unit weight follows, which might imply a significant influence on the design
of a building construction.
The same simulations are executed in a square container with identical volume as the slender container. Its dimensions are therefore: width=depth=height= 2.14 m.
For studying the stochastic influence of this configuration, a repeated pouring of the balls of 8 times is sufficiently. The simulated pore content lies in a range between
0.399≤n≤0.405 (Fig. 2). A mean value of the pore content nm=0.4016 with a
Figure 1. Occurrences of pore content (slender container).
Figure 2. Occurrences of pore content (square container).

variation coefficient vc=0.005 and a standard deviation sd =0.0020 results.
A good reproducibility of the pore content at the square container is thus given. This can be explained by a more favorable distribution of the balls in the square
container. Mathematically, 10.5 aligned balls can stress the front wall at its maximum in the slender container, where 22.9 balls charge the front wall in the square
container mathematically. It can be stated that 11 balls should at least fit mathematically aligned on one confining wall for a satisfying reproducibility of the pore content
without large stochastic effects. Hence, this amount exceeds the demanded number of ten aligned particles in DIN 18137 (1990).
2.2 Inner friction angle
The influence of stochastic effects on the inner friction angle at rest of the soil is examined in the state balls at rest. The agglomeration of balls stress the front walls at
both sides of the container. These forces are measured and then recalculated by:
(1)
where E0=measured force on one front wall, γ=bulk unit weight, h=height of the wall, t=depth of the wall. By the application of frictionless walls and the relation
k 0=1−sin φ, the inner friction angle φ at rest can be recalculated. These inner friction angles are determined for the measured pore contents for both containers and are
depicted in Figures 3 & 4.
The inner friction angle recalculated at the slender container (Fig. 3) varies between 22.2°≤φ≤32.7°. At the two opposite front walls deviating friction angles occur.
This reflects the antisymmetric earth pressure force due to stochastic influences. Even at the same wall and the same pore content different friction angles are
encountered. The mean value of the inner friction angle on the left wall is φm=26.3° and φm=21.1° on
Figure 3. Inner friction angle at rest (slender container).
Page 37
Figure 4. Inner friction angle at rest (square container).
Figure 5. Inner friction angle in active state (slender container).

the right wall. The variation coefficient at the left wall is vc=0.148, at the right wall vc=0.230.
The same basic observations apply to the square container (Fig. 4), but the variation of the friction angle (29.5°≤φ≤35.6°) is less distinct and the mean value is higher
at both front walls: left wall: φm=32.3°, right wall: φm=32.0°. The variation coefficient at the left wall is vc=0.053, at the right wall vc=0.059.
The less pronounced variation of the inner friction angle of the numerical soil in the square container is due to the enlarged front wall area in respect to the slender
container. Thus, more balls can stress the wall and the influence of each single ball is less distinct. This also explains the overall higher inner friction angle of the
numerical soil.
Studies with square containers of larger sizes reveal that at least 15 balls should fit mathematically on one wall for achieving a good reproducibility of the inner friction
angle φ of the simulated soil at rest (vc=0.03; i.e. φ≈φm±1, 0°).
Figure 5 shows the inner friction angle of the soil in the active state (one front wall is moved away from the container), which leads to comparable curves as for the
soil under rest. In addition to the fluctuation of the inner friction angle, the necessary displacement of the wall away from the container for achieving the minimum of the
earth pressure is subjected to a variation.
Using particle size distributions, the presented stochastic effects will get smaller, but the calculation time will increase significantly. Nevertheless, the determined ratio
“balls per wall length” should be met for a good reproducibility.
3 BALL PARAMETERS OF SPECIFIC SOILS
As a conclusion of the previous chapter it can be stated that the ball parameters have to be determined anew for every configuration of container geometry and ball
size. For the simulation of three specific soils (φ=30°, φ=35° and φ=40°; n=0.400) and different container geometries, the ball parameters are determined in the same
way as described before. The container geometries are chosen in dependence on the ball size. Balls with diameters d from 0.002 m, 0.020 m, 0.10 m to 0.20 m are
used. With square containers of 15 to 25 times the ball size, dimensions of 0.03 m to 5.00 m are covered. Even with the smallest wall length (15×d) a good
reproducibility should be possible. Beside the square container, which can be used for the simulation of local spatial geotechnical questions (piles etc.), the ball
parameters are also determined in slender containers with lengths and heights of 25 to 30 times the ball size, which can be applied e.g. for the simulation of walls and
shallow foundations.
Mathematical functions are derived from the simulations for the ball parameters, which can be applied for different configurations.
3.1 Square container
Depending on the container size the soil is represented by 3900 to 19,500 balls. The normal stiffness kn of the balls is chosen in such a way that an overlapping of two
balls by 2.5% occurs, if the total container mass (filled with balls) acts on one of these balls. The necessary ball parameters are presented in Figures 6 & 7 for an inner
friction angle φ=30° and a pore content n=0.400. These are—in order to derive the desired pore content—the shear stiffness ks_pour of the balls and the surface
friction sf_pour, which is set to 0.2. It should be mentioned that the balls are enlarged by 2‰ at the pouring state and reduced by the same amount at the state at rest to
delete possible locking effects.
The trend of the normal stiffness kn of the balls for different ball sizes is linear due to the constant ratio of container size to ball size, when displaying the results in a
logarithmic scale. The function of the necessary shear stiffness of the balls at the pouring state ks_pour is described by a potential function to

(where br=ball radius).
Page 38
Figure 6. Ball properties (pouring, rest 30°).
Figure 7. Ball properties (active 30°).
The necessary surface friction at rest sf_rest can be regarded as a constant with values between 0.30 and 0.35 (Fig. 6).
For the determination of necessary ball parameters in the active limit state (Fig. 7), one front wall is moved away from the soil. The shear stiffness ks_active is
determined under the criterion that the minimum of the earth pressure force acts at a displacement of 1‰ of the wall’s height. A constant relationship between the shear
stiffness ks_active and the ball diameter can be adopted. Fluctuations may be due to stochastic effects.
The surface friction at this state sf_active_move is constant too, with values between 0.40 and 0.45. The constant surface friction values are due to identical relative
stresses for each ball size, resulting from identical boundary conditions, which themselves depend on the ball size. The necessary surface friction—determined by means
of a triaxial test (confining stress=100 kPa)—sf_active_triax, yields different results. For intermediate stress states a surface friction between the two examined
parameters may be applied.
For soils with inner friction angles φ of 30°, 35° and 40° and pore contents n=0.400, the necessary ball parameters are presented in Table 1.
The shear stiffness ks_active is also used at the determination of sf_rest. An estimation of the
Table 1. Ball parameters.
φ (°) sf_rest (−) sf_active (−) ks_active (kN/m)

30° 0.3–0.35 0.4–0.45 8e4–5e5
35° 0.65–0.7 0.9–1.1 8e4–2e5
40° 5–6 3.5–4.5 8e3–2e5
Figure 8. Ball properties (pouring, rest 35°), container size.
surface friction at rest is possible by: sf_rest= 0.00009∙e0.2659φ. The necessary surface friction in the active state for a desired inner friction angle φ can be assessed by:
sf_active=0.0006∙e0.2185φ.
In the following simulations, the ball parameters are determined at different square container sizes. Due to the knowledge from the previous studies, it is sufficient to
use only one ball diameter (here: 0.02 m). In order to study the effect of the container size, a soil with an inner friction angle φ=35° is simulated. The surface friction at
the pouring process is again set to 0.2. The criterion for the balls normal stiffness is as before. The necessary ball parameters for the pouring process and the soil at rest
are shown in Figure 8.
The function of the shear stiffness of the balls at the pouring state ks_pour is described by a potential function: ks_pour=0.00005∙cs2.6511 (where cs=container size).
The surface friction of the balls at rest can be estimated by: sf_rest=0.1263∙e0.1099 cs. The surface friction at the active state can be assessed by:
sf_active=0.1651∙e0.1099 cs. The shear stiffness ks_active can be regarded as a constant with a value between 9e3 and 8e4 kN/m (Fig. 9).
3.2 Slender container
Slender containers with depths of 15 times the ball size, and heights and lengths of 25 and 30 times the ball size are subjected to the same examinations as the square
containers. The containers hold 10,500 to 17,000 balls. A single ball size d=0.02 m is used.
Page 39
Figure 9. Ball properties (active 35°), container size.
Figure 10. Preliminary ball properties (active; 25×d).
The first simulation is conducted with a container having a wall length and height of 25 times the ball diameter. Figure 10 depicts the preliminarily determined necessary
surface frictions for the desired inner friction angles φ of the simulated specific soils. The surface friction for a certain inner friction angle φ can be assessed by:
sf_active=0.00006∙e0.2931 φ. The shear stiffness for the achievement of the minimum of the earth pressure after 1 ‰ displacement of the height of the front wall is
regarded as constant (stochastic effects) with values between 5e3 and 1e5 kN/m. The function of the shear stiffness ks_pour of the balls at the pouring state is
described by a potential function to Simulations with larger wall lengths are presently underway.
3.3 Conclusions
The presented simulations enable the user of PFC3D to estimate the necessary ball parameters before starting an actual simulation. The user has at least an idea of the
necessary parameters, even if a soil with a different pore content, a stiffer behavior or a graded soil is applied. For container size and/or soil parameters deviating from
the examined ones, the derived functions can be applied for a first approximation.
4 SIMULATION OF THE EXCAVATION PROCESS IN FRONT OF SHEET PILE WALL CONSTRUCTIONS
For the simulation of the excavation process in front of sheet pile wall constructions, a numerical sheet pile wall is generated by deterministically positioned and parallel
bonded balls. By application of suitable parameters to the parallel bonds and balls, the behavior of the numerical wall is equivalent to the behavior of a real sheet pile
wall LARSSEN 601; as was examined in various bending tests.
The dimensions of the system are chosen in a way that a sliding plane in front of and behind the wall can emerge without any constraints. Therefore, a container with
the dimensions: height=10.78 m and length=16.95 m results, when simulating a 7.70 m high sheet pile wall that is positioned at 9.25 m length. The soil is simulated with
balls, which have a diameter of 0.20 m. To save time, their number is limited to 20,000. Hence, the container depth is set to 0.70 m. Thus, the ratio of wall depth to ball
size is far below the proposed value derived from the stochastic examinations, but more balls would end up in incompatibly long computing times. The sheet pile wall
depth is set to 0.60 m. In this way an unimpeded movement of the sheet pile wall is possible. The diameter of the sheet pile wall balls is set to 0.10 m. For the
simulation of a sheet pile wall friction angle that is one half of the inner friction angle of the soil, the wall balls receive half of the surface friction of the soil balls. The
sidewalls of the container are maintained frictionless.
After generating and bonding the sheet pile wall balls at their defined positions, the head of the sheet pile wall is fixed, and the soil balls are randomly generated
above the container. Then, the soil balls are poured into the container under normal gravitational acceleration. Having reached an equilibrium, the parameters are
assigned—for the state at rest and thereafter the active state—to the soil balls. A soil with an inner friction angle φ=29° and a pore content n=0.392 is simulated in this
way. The given pore content is the mean value of the pore content of the soil balls in the whole container. Larger deviations are encountered, when measuring this value
in sections.
For the simulation of this system, three P IV 2.4–2.7 GHz computers with 512 MB RAM are linked in parallel. Due to the unfavorable behavior of the system the
simulation of the unsupported wall takes 3 months, the simulation of the single supported wall another month.
4.1 Unsupported system
For the unsupported wall, the bending moments and transverse forces, which are acting within the sheet pile wall, the displacements of the sheet pile wall and the earth
pressures acting on the wall are determined.
Page 40
Figure 11. Bending moment (unsupported).
The curvature of the bending moment and the transverse force is established by the interrogation of the parallel bond information. The displacement curve is determined
by the position of the sheet pile wall balls. The earth pressure acting on the wall is established by using measurement circles in front of and behind the sheet pile wall,
varying in depth and height.
The excavation process is simulated by deletion of balls in excavation steps of 1.0 m or less. The state of equilibrium for each excavation depth is checked by the
development of the unbalanced forces and the subsequent displacement of the sheet pile wall head. After an excavation depth of 4.0 m the system fails and the wall
moves continuously.
For intermediate excavation depths, the bending moment curves are shown in Figure 11. Remarkable is the initial bending moment curve prior to the excavation
process. Obviously, the wall is already stressed due to the pouring of the soil balls into the container. This initial imprint is over pressed, not until the ultimate limit state.
When determining the earth pressures acting on the wall, an approximation of the earth pressure gradient at rest to the one at the ultimate limit state (active/passive)
can be observed with subsequent excavation depth (Fig. 12). The trend and the magnitude of the earth pressures are in accordance to Coulomb (Weissenbach 1975).
A replacement force at the toe of the wall develops in the final excavation step according to Blum (1931).
The peaks in the earth pressure distribution are due to force chains, which form between the soil balls and lead forces compressed into the wall. Despite the
averaging of the earth pressures by the measurement circles, this effect is still present. The developed force chains at the ultimate limit state are displayed in Figure 13.
They represent a visualization of the earth pressure distribution. Due to the rotation of the wall around its toe, the soil behind the wall head settles.
Simulations with smaller soil balls lead to less fluctuations in the earth pressure distribution, because
Figure 12. Earth pressure (unsupported).
Figure 13. Contact forces (unsupported).
more force chains can form and thus the forces are distributed to more balls.
4.1.1 Regularly packed particles
A simulation with soil balls, being positioned in the container in a tetragonal spatial centered shape is carried out (Fig. 14). The original pore content of such a packing is
0.260. For a better comparison with the previous simulation it is changed to 0.397 (by creating gaps at the sides). The simulation leads to larger excavation depths
accompanied by smaller displacements, bending moments and transverse forces, compared to
Page 41
Figure 14. Earth pressure (positioned, unsupported).
the previous simulation. This behavior is due to an interlocking effect of the soil balls and wall balls, which finally leads to a toe of the wall, that has no vertical contact
with the soil balls. The supporting earth pressures in front of the wall (in a range directly below the excavation level) exceed by far those according to Coulomb.
4.2 Single supported system
For the simulation of a single supported sheet pile wall, the simulated unsupported sheet pile wall is shifted back against the soil—by applying a force at the support
level (0.55 m)—at an intermediate excavation depth of 3.0 m. Then, the two structural members, waling U 300 and strut IPE 220, both simulated by positioned and
parallel bonded balls, are generated. The strut is fixed at its end in all three directions but allows for free rotation. The simulation of the excavation process is again
executed in steps of 1.0 m or less. The state of equilibrium for each excavation depth is additionally controlled by the incremental displacement of the wall in one meter
steps. At an excavation depth of 6.0 m the ultimate limit state is reached. A further excavation leads to a continuous movement of the wall.
Figure 15 shows the developed contact forces at the ultimate limit state. Due to larger contact forces in the upper area behind the wall, an earth pressure distribution
towards the strut level can be deduced. The displacement of the wall head against the soil recurs the soil to its initial level. In accordance with the static system, no
replacement force at the toe of the wall develops and the supporting contact forces in front of the wall act in an area above the toe of the wall. The determined earth
pressures emphasize this statement (Fig. 16).
The displacements—belonging to the certain excavation depths—display that the strut stiffness allows the sheet pile wall only a small displacement at the support
level, whereas the toe of the wall deflects more
Figure 15. Contact forces (single supported).
Figure 16. Earth pressure (single supported).
with increasing excavation depth (Fig. 17). This leads finally to the failure of the system.
Studying the bending moment curvature (Fig. 18), the change of the static system with progressing excavation depth can be seen. At excavation depths of 3.0 m to
5.0 m the toe of the wall is partially fixed, whereas in the ultimate limit state (at 6.0 m excavation depth) the toe is free. No hogging moment forms at the support level.
This might be due to the limited amount of balls above the support level and a shear stiffness that is too low to build up a sufficient resistance.
Page 42
Figure 17. Displacement (single supported).
Figure 18. Bending moment (single supported).
Figure 19. Transverse force (single supported).
The force of the strut at the different excavation levels can be read off Figure 19. At 3.0 m excavation depth the strut force is 59 kN/m and it increases to 103 kN/m at
6.0 m excavation depth.
4.3 Single supported system with a plastic hinge
According to Eurocode 3 Part 5 (2003), apart from the elastic design, the plastic design of sheet pile walls is
Figure 20. Moment rotation behavior (plastic hinge).
Figure 21. Displacement (elastic/plastic wall).
allowed, when taking the moment rotation behavior of the sheet pile wall into account. Therefore, a virtual (because unknown) plastic capacity of the simulated
profile—thus weaker than the elastic LARSSEN 601—is chosen. A four point bending test reveals that at a maximum bending moment of 95 kNm/m a plastic hinge
develops, i.e. at 4.0 m excavation depth of the elastic wall (Fig. 20).
Characteristic for a plastic wall is the large displacement in the plastic hinge (Fig. 21), which develops at the point and depending on the magnitude of the maximum
bending moment. After an excavation depth of 6.0 m the ultimate limit state of this system is exceeded and it fails by a continuous movement of the toe of the wall, i.e.
by geotechnical failure.
Due to the large displacement in the plastic hinge, the soil cannot stress the wall as in the simulation with the elastic wall. This leads to a reduced maximum bending
moment in the plastic wall (Fig. 22). The contact forces (Fig. 23) resemble those of the elastic wall. But at approximately 5.0 m below ground level the contact forces
are larger than those of the elastic wall (Fig. 15).
The different interaction behavior of both systems can be studied by comparing the earth pressures acting on the elastic and the plastic wall respectively (Fig. 24).
Page 43
Figure 22. Bending moment (elastic/plastic wall).
Figure 23. Contact forces (plastic wall).
Obviously, in the area of the plastic hinge (in the ULS at 4.0 m below ground level) the earth pressure is further reduced. Due to the deflection of the wall, the earth
pressure directly above the plastic hinge is reduced but it increases in the area below the plastic hinge (in comparison to the elastic wall). The soil directly above the
hinge has lost its support by the deflection of the wall and finds a new support in the wall area below the hinge.
The mobilized supporting earth pressure in front of the plastic wall exploits the possible earth pressure
Figure 24. Earth pressure (elastic/plastic wall).
Figure 25. Transverse force (elastic/plastic wall).
according to Coulomb up to its limit. The supporting earth pressure of the elastic wall is mobilized to a smaller amount.
The transverse forces of the elastic and plastic wall (Fig. 25) show that the strut of the wall in the plastic state is less stressed, which implies that the earth pressure
distribution towards the strut level is less intense than it is at the same excavation level of the elastic wall. Large fluctuations of the transverse force around the plastic
hinge occur. The reason might be due to temporary maximum displacements of the wall in this area caused by the varying position of the maximum bending moment
within the wall.
5 CONCLUSIONS
PFC3D can be regarded as a suitable tool for the simulation of the interaction of a building construction and the soil. Here, the simulations focus on interaction of
different sheet pile wall systems and the soil. By visualizing the contact forces, the behavior of the different systems can be studied and examined for a
Page 44
better understanding. This is supported by the determination of the earth pressures acting on the wall, the bending moments and transverse forces as well as the
displacements of the wall. At plastic design of sheet pile walls, the large wall displacements can be taken into account explicitly by the implementation of a predefined
momentrotationcurve of the sheet pile wall.
The simulations of the unsupported and single supported wall yield realistic results. The behavior of a single supported wall with a plastic hinge could be studied for
the first time without hydrostatic water pressure and without having a fixed position of an artificially manufactured plastic hinge (Bornarel et al. 2001).
From the simulations with different container geometries, ball sizes and granular soils, the determined ball parameters enable the user to estimate the necessary ball
parameters for the simulation of a specific soil under certain boundary conditions. For a good reproducibility, the ratio of balls per wall length is chosen to fifteen at its
minimum. For the simulation of soils that do not comply with the examined ones, derived mathematical functions can be applied for a first calculation.
ACKNOWLEDGEMENTS
We thank the DFG (Deutsche Forschungsgemeinschaft) for supporting this research project.
REFERENCES
pr EN 1993–5 2003. Eurocode 3 Part 5. Design of steel structures—Piling, Brussels.
Blum, H. 1931. Einspannungsverhältnisse bei Bohlwerken, Ernst & Sohn.
Bornarel, BourneWebb, Dercks, Heinemeyer, Kastner, König, Pohl, Potts, Pulsfort, Sedlacek 2001. Advanced Design Tools for Steel Sheet Piling—Final Report.
Aachen.
Cundall, P.A. & Strack, O.D.L. 1979. A discrete numerical model for granular assemblies. Géotechnique 29(1): 47–65.
DIN 18137 1990. Bestimmung der Scherfestigkeit, Beuth Verlag.
Herten, M. & Pulsfort, M. 1999. Determination of spatial earth pressure on circular shaft constructions, Granular Matter 2:1–7.
Itasca Consulting Group, Inc. 1999. PFC3D—Particle Flow Code in 3 Dimensions, Version 2.0 User’s Manuals. Minneapolis: Itasca.
Torquato, S., Truskett, T.M., Debenedetti, P.G. 2000. Is Random Close Packing of Spheres Well Defined?, Physical Review Letter 84(10):2064–2067.
Weissenbach, A. 1975. Baugruben, Teil I–III. Ernst & Sohn.
Page 45
Finite element and discrete element modeling for the solution of spatial active
earth pressure problems
F.tom Woerden, M.Achmus & K.AbdelRahman
Institute of Soil Mechanics, Foundation Engineering and Waterpower Engineering, University of Hannover, Hannover, Germany
ABSTRACT: Regarding the active earth pressure acting on walls of limited widths only few experimental findings exist. From analytical calculation
methods proposed in the literature generally different results are obtained. Using the particle flow code, a discrete element model for the consideration
of earth pressure phenomena in sand is established. Calibration of the model parameters is done by comparison with the results of triaxial tests. The
model is then validated by calculation of a 2D earth pressure case. For the simulation of spatial (3D) earth pressure forces the discrete element model
and a finite element model with a hypoplastic material law for sand are used. Both methods lead to similar and plausible results. It is shown that for the
case considered, the analytical methods typically used overestimate the spatial active earth pressure force. As well, a distinct dependence on the
spatial earth pressure on the type of wall movement is found.
1 INTRODUCTION
The calculation of earth pressures on building structures is a central problem of soil mechanics. The horizontal pressure acting on a rigid wall is termed the earth
pressure at rest. The earth pressure is decreasing when the wall is moving away from the earth. A minimum value, which is called the active earth pressure, is reached at
a certain displacement. The resultant earth pressure force on the wall is termed Ea .
For plane strain (2D) conditions, i.e. walls with very high widths in relation to the depths, an analytical calculation method exists, which gives experimentally
confirmed results. For noncohesive soils the earth pressure force is:
(1)
Herein h is the wall height, γ is the unit weight of the soil and k a is the active earth pressure coefficient, which in general depends on the angle of internal friction of

the soil, on the angle of wall friction δa and also on wall and surface inclinations.
The active earth pressure in a depth z below surface for the case of a rotation of the wall about its base is given by
(2)
The active earth pressure acting on a wall with limited width is a spatial problem. Spatial active earth pressure has to be considered particularly for the design of slurry
trenches, but also for the load calculation of shafts or small and deep excavation pits.
For spatial earth pressure calculations a variety of analytical methods have been developed. These methods are only partially confirmed by experiments, and in
general different methods lead to different results for the resultant earth pressure force.
Basically the spatial (3D) active earth pressure is lower than the earth pressure for the 2D case. This can be formulated via a reduction factor λ:
(3)
This reduction factor depends mainly on the spatiality ratio n=h/b (with b=wall width), with λ=1 for n 0 or b>>h, respectively.
Two methods for the estimation of spatial earth pressure forces are in use in Germany. The regulation DIN 4085 (1987) gives reduction factors dependent on n,
which were derived as average values from different analytical methods (Piaskowski & Kowalewski 1965, Walz & Prager 1978, Karstedt 1980). In the regulation
DIN 4126 (1986) a wedge model is recommended for the calculation of earth pressure forces on slurry trenches. An overview on analytical calculation methods is
given in MüllerKirchenbauer et al. (1979).
A generally accepted method for the calculation of spatial earth pressures does not exist. Also there is no reliable calculation method and very few experimental
investigations on pressure distributions dependent
Page 46
on the wall displacement mode. Basic knowledge can be obtained with numerical methods. In the presented paper a spatial earth pressure problem in sand soil is
modeled with the discrete element method and for comparison also with the finite element method.
2 NUMERICAL MODELS
2.1 Finite element model
Numerical calculations in ground engineering are mainly carried out by means of the finite element method (FEM). The soil is herein considered as a continuum, whose
stressstrain behaviour is described on a macroscopic level with a material law. The quality of the calculation results is of course decisively dependent on the quality of
the chosen material law. A generally valid material law for soils does not exist. The choice of an approach is to be made under consideration of the kind and the
boundary conditions of the problem to be modeled.
The FEM calculations were executed with an 8parametric hypoplastic material law, which describes the stressstrain behavior of sand very reliably. For details on
the hypoplastic material law please refer to Herle & Gudehus (1999) and AbdelRahman (1999).
Material parameters derived for “Karlsruhe sand” were used in the calculations. The parameters and the index properties of this poorly graded sand are given in
Table 1.
For the initial state, a void ratio e0=0.65 was realized. This void ratio corresponds to a relative density of Dr=0.61.
2.2 Discrete element model
In the discrete element method (DEM) individual particles of a material are simulated discretely. The interaction of the different particles is treated as a dynamic
process. An equilibrium state is reached when all the internal contact forces are balanced. The dynamic behavior is represented numerically by a time stepping algorithm
using explicit time difference scheme.
Here the DEMcode PFC3D Version 3.0 (Itasca 2003) is used. The sand particles are idealized as perfect spheres. The contact behavior of the particles is
described by three parameters:
1. Normal stiffness k n : The normal stiffness determines the overlap of two adjacent particles dependent on the normal contact force.
2. Tangential stiffness k s: The tangential stiffness determines the amount of tangential relative (shear) displacement of two spheres in contact dependent on the tangential
contact force.
Table 1. Index properties and hypoplastic parameters of Karlsruhe medium sand.
26.5
Unit weight of the grains, kN/m3
D , mm
10
0.240
D60, mm 0.443
Uniformity coefficient, Cu 1.85
Min. void ratio, emin 0.53
Max. void ratio, emax 0.84

Hypoplastic parameters: h s=5800 MN/m2, ed0=0.53, ec0=0.84, ei0 =1.0, n=0.25, α=0.13, β=1.05

Figure 1. Geometry of the PFC3D model.
3. Friction coefficient μ: The friction coefficient determines the maximum tangential contact force via a Coulomb type friction law, i.e. dependent on the normal contact
force.
The geometry of the model box used is shown in Figure 1. Its base area is 140 cm×30 cm and its height is 50 cm. A plane earth pressure case is simulated by
movement of the wall A–C and a spatial earth pressure case by movement of the wall a–b. Thus, with a=40 cm and b=20 cm a spatiality coefficient of n=2 is
considered. The model walls were simulated as totally smooth, i.e. without wall friction.
The sand was modeled by spheres with a uniform diameter of D=2 cm and a density of 2650 kg/m3. When related to the average grain size diameter of Karlsruhe
sand, this gives a scaling factor of 400. This scaling factor was used to limit the necessary calculation time.
3 GENERATION AND CALIBRATION OF THE DEM MODEL
In the first step of a DEM calculation, a sample with the desired relative density has to be generated. To achieve this, the particles have to be poured into the
Page 47

Figure 2. Void ratio in the PFC model dependent on the coefficient of friction μ e (k n=k s=105 N/m).

model box. The parameters k n , ks and particularly the friction coefficient μe, which are found in preliminary investigations, determine the resulting relative density.
In the next step the desired initial stress state in the sample is generated. For that the diameter of all the particles is decreased by a small amount. The degree of
reduction and the friction coefficient μv during this step are again determined in preliminary investigations.
Using this method, sphere packings or samples suitable for the consideration of earth pressure problems can be generated (Herten & Pulsfort 1999). In the last step
the movement of the model walls is carried out, leading to the shear stresses in the sample. For this step again a new set of contact parameters is defined, which is
determined by calibration with triaxial tests.
The normal and tangential stiffnesses k n and k s were held constant in all calculation steps. Via the calibration on a triaxial test they were determined to be k n =k s=105
N/m.
The calibration process for the determination of the friction coefficients and the degree of grain size reduction in the discrete element model is described below.
3.1 Generation of the sphere packing
The particles of the system considered were generated by using the “interspersing” method (Mörchen & Walz 2003). Herein the particles are generated without
overlapping in a random distribution inside a volume which is larger than the target volume. After activation of gravity, the particles fall down into the box and move in
mutual interaction in free positions. The parameters k n and k s were chosen to be k n =k s=105 N/m. The resulting void ratio of the sample is dependent only on the
friction coefficient μe.
The dependence of the void ratio on the friction coefficient μe is shown in Figure 2. The densest packing is achieved by use of μe=0 with emin=0.56 and the loosest
packing by very high μevalues with emax=0.84. For the generation of the desired relative density of Dr=0.61 (e=0.67) a μevalue of 0.17 was chosen.
3.2 Initial stress state
Generally, after the particles pour into the model box, the average horizontal stresses acting on the sample walls are different from the desired initial stress state. For the
“at rest” state, an earth pressure coefficient of k 0=0.48 was desired, which is the value derived by calculation with the hypoplastic material law in the finite element
model for Dr=0.61.
In the state after pouring, the earth pressure coefficient in the DEM model, derived from the contact forces acting on the model walls, was about k=0.86. A
reduction to the desired earth pressure coefficient was achieved by a decrease of all particle diameters of 0.36% and with a friction coefficient of μv =0.6 during this
step. Due to the very small change of the particle diameters no significant change in the relative density of the sample occurred.
3.3 Calibration with the results of a triaxial test
In the last step of earth pressure simulation, the fixation of relevant contact parameters was carried out by simulation of a triaxial test with the DEM model and
compared with the behavior of real sand. For the numerical simulation of the triaxial test the particles were generated using the fill and expand method (Itasca 1995)
inside a cylinder with a diameter of 20 cm and a height of 40 cm.
The model box for the simulation of earth pressure has a height of 50 cm (see Fig. 1), i.e. the stress level inside the sample is relatively low with stresses between 0
and about 8 kN/m2. Thus, a triaxial test with a confining stress of 5 kN/m2 was used for calibration. Instead of the real test results, the results of a finite element
calculation with the hypoplastic material law described above were used for comparison with the DEM results.
Optimum agreement of the results was found for the parameters k n =105 N/m, k s=105 N/m and μs=2.3. Results obtained with these parameters are given in Figures
3 & 4. The mobilized friction angle is shown in Figure 3 is defined by the equation
(4)
The FEM results represent the actual behaviour of Karlsruhe sand in the triaxial test. It appears that the PFC model achieves a good agreement concerning the
development of mobilized friction angle with strain and the ultimate shear strength. Volumetric strain is overestimated for high strain levels, but this is not of decisive
relevance for the earth pressure simulation considered here.
Page 48
Figure 3. Mobilized friction angle dependent on axial strain.
Figure 4. Volumetric strain dependent on axial strain.
3.4 Validation of the DEM model
The consideration of the 2D earth pressure problem can serve as validation of the DEM model, because for this case a reliable calculation method exists. For the
considered case of a smooth wall (wall friction angle δa =0) the active earth pressure coefficient amounts to

(see Fig. 3) one yields k a =0.27.
For the numerical simulation the 50 cm high wall A–C of the model box (Fig. 1) was moved. The resultant earth pressure force was derived as the sum of all contact
forces acting perpendicular to the wall. With that, using Equation (1), the earth pressure coefficient can be calculated.
Three deformation modes of the wall were investigated: A rotation of the wall about its base, a rotation about the wall top and a pure horizontal translation. In Figure
5 the development of the earth pressure coefficients are given dependent on the dimensionless displacement u/h. u is herein defined as the maximum horizontal
displacement, i.e. for instance the horizontal head displacement in the case of the base rotation.
It appears that the magnitude of the earth pressure coefficients depends on the wall deformation mode. This is not considered in the analytical method, but
Figure 5. Plane strain earth pressure coefficients obtained with the DEM model.
the effect is known from experimental results and also from numerical investigations (see Fang & Ishibashi 1986, AbdelRahman 1999).
The earth pressure coefficients calculated with the DEM model lie between k a =0.23 (translation) and k a =0.34 (rotation about base) and are thus of similar size as
the analytical value.
The minimum earth pressure forces were reached at dimensionless displacements u/h between about 0.8% (translation) and about 1.5% (rotation about base).
These limit values are larger than values known from experience, which for dense sand lie between 0.05 and 0.5%. But, limit values obtained from FEM calculations
are also much larger than the experience values.
At least for the plane active earth pressure force, i.e. the minimum force reached at a certain displacement, it can be stated that the DEM model gives reasonable
results and is thus suitable for the calculation of spatial active earth pressure forces.
4 SIMULATION OF THE SPATIAL EARTH PRESSURE PROBLEM
Spatial earth pressure calculations were carried out with the presented DEM model as well as with a FEM model. Herein a spatiality ratio of n=2 was chosen as
exemplary. For the DEM simulation the 40 cm high wall a–b of the model box (Fig. 1) was moved. Again the three deformation modes rotation about base, rotation
about top and pure translation were considered.
For the FEM simulation a wall of 10 m height and 5 m width was investigated. The discretized system had a base area of 55 m×10 m and a height of 12.5 m. A fine
mesh was used. The interaction between the wall and the soil elements was modeled via a contact algorithm. The calculations were carried out with the Abaqus
program (Abaqus 2003).
Starting from the earth pressure at rest E0 a reduction of the earth pressure force is induced with
Page 49
Figure 6. Reduction factors κ obtained with DEM and FEM for wall base rotation, wall top rotation and translation.
increasing wall displacement. This reduction is described here by a factor κ:
(5)
In Figure 6 the obtained reductions of the κvalues with increasing dimensionless displacement u/h are shown for the different wall deformation modes. The results of
the FEM calculations and the DEM calculations are given for comparison.
First it has to be stated that both numerical methods give results which are in relatively good agreement. This is valid for the deformation dependence of κ and also
for the minimum value κmin, which gives the active earth pressure force. The κmin value obtained from the DEM is a bit lower than that obtained from the FEM.
For the reduction factor λ introduced in Equation (3) the following is valid:
(6)
Table 2. Values for
λ in 1
Base rotation Top rotation Translation
DEM 0.43 (0.34)* 0.39 (0.35)* 0.27 (0.31)*
FEM 0.57 0.52 0.39
DIN 4085 0.76 (independent of deformation mode)
DIN 4126 0.53 (independent of deformation mode)
*The values in brackets were determined with the k a values from the 2D earth pressure DEM calculations (see Figure 5).
In the case under consideration k 0=0.48 and k a = 0.27. The reduction factors calculated correspondingly are given in Table 2 and compared with the values from the

analytical calculation methods usually used in Germany. The κmin values corresponding to these analytical methods are also shown in Figure 6.
From the numerical results it is evident that in the 3D case the wall deformation mode influences the active earth pressure force even more than in the 2D case. For a
pure wall translation, the active earth pressure force amounts to only about 63% (DEM) respectively 69% (FEM) of the force for the case of a wall base rotation.
Obviously the development of arching behind the wall is significantly affected by the deformation mode.
The reduction factors λ obtained numerically from the DEM are, even for the case of base rotation, significantly lower than the factors resulting from the analytical
methods according to DIN 4085 and DIN 4126. This confirms that distinct further research need exists concerning spatial earth pressure calculation.
5 CONCLUSIONS
For the investigations presented here a DEM model for the simulation of earth pressure problems in sand was established. The sand was idealized by perfect spheres
with a constant diameter of 2 cm. To keep the calculation times within acceptable limits, a model box with relatively small dimensions (50 cm height) was generated.
By comparison with results of analytical calculations for the 2D earth pressure case and with the results of finite element calculations for the spatial (3D) case, it is
proved that even this rather simple DEM model is capable of simulating problems of that kind.
The need for research regarding the spatial active earth pressure problem is shown by the numerical calculations. On one hand, the earth pressure forces calculated
are significantly smaller than the values coming from analytical methods usually used for estimation. On the other hand, the numerical calculations show a distinct
dependence of the spatial earth pressure on the
Page 50
deformation mode i.e. the type of wall displacement. For further clarification more numerical, as well as experimental, investigations are necessary.
REFERENCES
ABAQUS 2003. User’s manual, Version 6.4.
AbdelRahman, K. 1999. Numerische Untersuchungen von Maßstabseffekten beim Erddruck auf Sand. Schriftreihe des Lehrstuhls Baugrund Grundbau, Universität
Dortmund, Heft 23.
DIN 4085. 1987. Berechnung des Erddrucks. Deutsches Institut für Normung. Beuth Verlag.
DIN 4126. 1986. OrtbetonSchlitzwände. Deutsches Institut für Normung. Beuth Verlag.
Fang, Y.S. & Ishibashi, I. 1986. Static Earth Pressure with Various Wall Movements. Journal of Geotechnical Engineering, 112(3): 317–333.
Herle, I. & Gudehus, G. 1999. Determination of a hypoplastic constitutive model from properties of grain assemblies. Mechanics of Cohesivefrictional Materials,
4:461–486.
Herten, M. & Pulsfort, M. 1999. Determination of spatial earth pressure on circular shaft construction. Granular Matter, 2:1–7.
Itasca Consulting Group, Inc. 2003. PFC3D Ver. 3.0, User’s Manual, Minneapolis: Itasca Consulting Group.
Karstedt, J. 1980. Beiwerte für den räumlichen, aktiven Erddruck bei rolligen Böden. Bauingenieur, 55: pp. 31–34.
Mörchen, N. & Walz, B. 2003. Model generation and calibration for a pile loading in the particle flow model. In Konietzky et al. (eds), Numerical Modeling in
Micromechanics via Particle Methods, Proc. of the 1st International PFC Symposium, Gelsenkirchen, Germany, Nov. 6–8, 2002. Lisse: Swets & Zeitlinger, pp.
189–195.
MüllerKirchenbauer, H., Walz, B. & Kilchert, M. 1979. Vergleichende Untersuchung der Berechnungsverfahren zum Nachweis der Sicherheit gegen Gleitflächenbildung
bei suspensionsgestützten Erdwänden. Veröffentlichungen des Grundbauinstituts der Technischen Universität Berlin, Heft 5.
Piaskowski, A. & Kowalewski, Z. 1965. Application of Thixotropic Clay Suspensions for Stability of Vertical Sides of Deep Trenches without Strutting. Proc. 6th intern.
conf. on Soil Mechanics and Foundation Engineering, Montreal, Vol. 2, pp. 526–529.
Walz, B. & Prager, J. 1978. Der Nachweis der äußeren Standsicherheit suspensionsgestützter Erdwände nach der Elementscheibentheorie. Veröffentlichungen des
Grundbauinstituts der Technischen Universität Berlin, Heft 4.
Page 51
Numerical study on the cause of a slope failure at a gravel pit using PFC and
FLAC
T.Sakakibara
CRC Solutions Corp., Osaka, Japan
M.Ujihira
Div. of Environment & Resource Engineering, Hokkaido University, Sapporo, Japan
K.Suzuki
Nittoc Construction Co., Ltd., Sapporo, Japan
ABSTRACT: Discrete particle modeling with PFC2D was used to simulate a slope failure due to ground water pressure. Seepage force caused by
hydraulic pressure gradient should be applied as additional force to each particle of calculation model. Calculation of seepage force was conducted
using FLAC, which can be used for continuous material. This method was applied to find out the cause of a slope failure accident at a gravel pit in
Yunicho, Hokkaido, Japan in 1999. From this numerical study, it was concluded that the slope failure should have occurred mainly because the
excavation was carried out under the condition of high pressure existence of groundwater at the back of the slope.
1 INTRODUCTION
An accident of a mining face collapse of a short slope occurred in Yunicho, Hokkaido, Japan in 1999 and a machinery operator lost his life. This accident warns us that
it is not safety even when a slope is short, if the sufficient cares are not paid for the stability of soft layers. Figure 1 is a photograph just after the slope failure at the
gravel pit. Cross section of the slope is shown in Figure 2. Where, the aggregate was mined only in summer season and the mined out area was
Figure 1. Photograph of a slope failure.
reclaimed in winter season, because the demand of the aggregate material for construction works and others is a little. In the case of this accident, the slope collapsed
when the mining level reached from 10 m to 17 m depth from ground level.
The level of 10 m is the value that was permitted in the safety regulations considering the stability of soft layers. In the process of reclamation of the mined out area,
there was the possibility that a lot of snow was mixed with reclamation soils. It is considered that the hydraulic pressure of melted water in the reclaimed area was the
cause of the appearance of steep pressure gradient and resultant failure of the protection zone.
DEM (Distinct Element Method) is suitable for the simulation of various kinds of slope failures. In this method calculation conditions such as strength,
Figure 2. Crosssection of a gravel pit and protection zone.
Page 52
cohesion, angle of internal friction, slope angle, slope height, slope shape, groundwater head and others should be reflected for analysis. Groundwater pressure should
be applied certainly to the calculation model when the existence of groundwater is examined at the back of a slope from the field investigations. Namely, seepage force
caused by hydraulic pressure gradient should be applied as additional force to each particle of the calculation model.
In this study, continuous modeling with FLAC was used to calculate the groundwater gradient and determine hydraulic pressure acting to the protection zone.
Obtained seepage force was applied to the particle model of PFC2D and slope deformation was calculated.
2 HORIZONTAL LOADING TEST
The horizontal loading test was carried out at the two points near the failure site. Table 1 shows the measurement results of insitu horizontal loading test.
In order to determine the DEM parameter, the numerical tests that simulate the insitu horizontal loading tests were performed. Figure 3 shows PFC model reflecting
the insitu horizontal test. A series of tests were performed to determine elastic properties and bond properties, which would influence to failure behavior.
The size of particle radius is set almost uniformly from 0.20 to 0.25 m, which is the same distribution as the radius of the gravel pit model executed later.
Initially, elastic properties of the sample that could not be measured directly in the field test were determined indirectly by performing a calculation of horizontal
loading test under high bond strength. Three set of runs were carried out with different values of normal stiffness. Shear stiffness was simply assumed to be 1/4 of the
value of normal stiffness.
Figure 4 shows 3 kinds of relationships between horizontal displacement of the loading wall and the force acts on it. From these results, it was confirmed that the
relationship between displacement and force in the case of kn=5e8 N/m was corresponding to the measured data in Table 1.
Secondly, in order to determine the yield strength of gravel layer, a series of run was carried out with changing the bond strengths (Table 2). It is difficult to set the
contact bond parameter, because we have only the data on yield stress as the measured parameter concerning strength. Therefore it was assumed that the calculation
should be executed by changing the bond strength to find out the adequate bond strengths. Where, it is said that bond strength is the adequate value when the
calculated strength is converged to the measured value of 0.5 MPa. Figure 5 shows the relationships between horizontal displacement of the loading wall and the
horizontal stress acts on it.
Table 1. Result of insitu horizontal loading test.
N value Yield stress (MPa) Young Modulus (MPa)

Point1 15 0.49 19.0
Point2 15 0.52 22.3
Figure 3. PFC model for insitu horizontal loading test.
Figure 4. Relationship between displacement and force for the evaluation of stiffness.
Table 2. Contact bond strength used for calibrations.
Normal bond strength, (N) Shear bond strength, (N)
Case1 5.0e4 5.0e4
Case2 5.0e4 1.0e5
Case3 1.0e5 5.0e4
Case4 1.0e5 1.0e5
Case5 1.0e5 2.0e5
Case6 2.0e5 1.0e5
From these results, in case of the parameter of bond strength, n_b=1e5 N s_b=1e5 N (Case4) were determined. The contact force distribution under above bond
strengths conditions is shown Figure 6. The calibrated PFC parameters are summarized in Table 3.
Page 53
Figure 5. Relationship between displacement and stress for the evaluation of bond strength.
Figure 6. Contact force distribution at Disp.=0.032 m.
Table 3. Calibrated parameters in PFC sample.
Normal stiffness (N/m) 5e8
Shear stiffness (N/m) 1.25e7
Friction coefficient 0.5
Normal bond strength (N) 1e5
Shear bond strength (N) 1e5
3 NUMERICAL MODEL
Figure 7 shows the planned cross section (a) in the stage of mine planning and the cross section (b) created after mined out. Compared with these two figures, mainly
slope angles are different. In order to investigate the cause of a slope failure, numerical calculations on both models have been executed.
In this study, the effect of fluid interaction to a granular assembly in unit thickness is simply given by the following equation:
(1)
Figure 7. Cross sections both of planned and practically mined.
Figure 8. FLAC model and hydraulic condition to calculate the hydraulic pressure distribution.
Figure 9. PFC model for the calculation of slope stability.
where Fi is the seepage force, R is the particle radius and P is the pressure.
To obtain seepage force, FLAC was used to determine the fluid flow field in a model containing two kinds of layers as shown in Figure 8.
Figure 9 shows the PFC model that corresponds with the FLAC model. Particle radii were set almost uniformly from 0.20 to 0.25 m and the number of particles
was 10414.
The flow vectors, which should be applied as body forces to each particle, are shown in Figure 10. Because the coefficient of permeability is ten times different
Page 54
Figure 10. Seepage forces obtained from the calculation using FLAC.
Figure 11. Failure of the model. (a) 10,000 cycles, (b) 20,000 cycles, (c) 30,000 cycles, and (d) 40,000 cycles.
between two layers, it is understood that hydraulic pressure by a high hydraulic gradient has occurred in the gravel layer. Figure 11 shows the state of the PFC model
after every 10,000 cycles calculation. Bold
Figure 12. Total displacement contours in two cases.
line represents the boundary of gravel layer. A slope failure occurred initially at the toe of slope and progressed toward the up side of the protection zone. These failure
behaviors of PFC simulation agree with the observations.
To identify and clarify the effect of hydraulic pressure and the gradient of slope, contour plots of equidisplacement were drawn (Figure 12). Figure 12(a) shows a
result of total displacement calculated by using the model with high gradient of slope. Figure 12(b) shows a result calculated by using the model with low gradient of
slope. In the later case, it had no damage in the side of gravel pit and the maximum displacement has occurred in reclaimed area. The reason is because there was no
measured data on reclaimed area so that the stiffness of the area was assumed to be 1/10 of the gravel stiffness.
4 CONCLUSIONS
Discrete particle modeling with PFC was used to simulate a slope failure at a gravel pit. In order to take
Page 55
into account of the hydraulic pressure, seepage forces obtained by FLAC are simply coupled with PFC.
From a series of calculated results by PFC simulations, two kinds of major causes of slope failure became clear. One is because they had excavated the slope under
the angle condition of 64 degree, and the other is because mining was continued without any water drainage in the reclaimed area.
REFERENCES
Itasca Consulting Group, Inc. 2002, PFC2D—Particle Flow Code in 2 Dimensions, Version 3.0 User’s Manuals. Minneapolis: Itasca.
Itasca Consulting Group, Inc. 2003, PFC3D—Particle Flow Code in 3 Dimensions, Version 3.0 Users Manuals. Minneapolis: Itasca.
Page 56
Page 57
Evaluation of castle stone wall stability with the particle flow code
M.Mizumoto
Sumitomo Mitsui Construction, Tokyo, Japan
G.Q.Cao
CRC Solutions, Tokyo, Japan
ABSTRACT: In the ruins of Japanese castles, most of the original towers have been lost. Many of the stone walls, which are important structures
serving as both retaining walls and foundation support for the towers, are still remaining but have been damaged or deformed during the passage of
time. Recently plans for the restoration of lost towers have been made. It is necessary to analyze the stability of the stone walls in order to judge
whether they need to be repaired. In this study, we proposed a new approach to analyze the stability of castle stone walls using the particle flow code.
Modeling methods and analysis results are presented in this paper. Influence of shapes of stone wall surface, location of backfill and size of stones are
also discussed.
1 INTRODUCTION
Many ruins of ancient castles exist in Japan. Over time, most of the original towers (or turrets) in these castles, which were constructed by traditional techniques, had
been damaged and lost due to earthquakes and other reasons. The time is ripe for the restoration of lost towers throughout the country.
In Japanese ancient castle construction, the stone wall (in other word stone masonry) is an important structure of both the retaining wall and the foundation
supporting the tower (see Fig. 1). However, many of the stone walls have been damaged or deformed during passage of time. Some of them may be in danger of
collapse. It is necessary to analyze the stability of such
Figure 1. Typical Japanese castle.
stone walls in order to judge whether they need to be repaired or not. In the case of the restoration of a tower on such a stone wall, evaluation of the castle stone wall
stability becomes more important, because it becomes not only the retaining wall but also the foundation supporting the tower.
In Japan, the stone walls in ancient castles are constructed by dry masonry. They consist of blockstones (or ashlars) on the surface, cobblestones as backfill and soil
behind them (see Figs 2 & 3). They are a very important part of historical heritage. In order to hand down the precious technique, it is necessary when
Figure 2. Section of typical Japanese castle stone wall.
Page 58
restoring the walls to use the same materials and techniques, as much as possible, as was used when the original stone wall was built. No modern construction
techniques, such as concrete reinforcement, are allowed. It is desirable to restore the castle stone walls to their original state.
In the present design method, it is difficult to model blockstones, cobblestones and soil. In this study we present a modeling method using the particle flow code. To
analyze the stability of the castle stone wall, the influence of the shape of the stone wall surface, and the arrangement location and size of cobblestones are discussed. In
this paper we compare the results and discuss how the stability of stone walls change with these factors.
2 ANALYSIS METHOD
In this analysis, the particle flow code PFC2D, developed by Itasca Consulting Group, Inc., is used. The model is simulated by circle elements that are contacted
through points. Based on actual shapes of stones, the clumps of circles are used in modeling. Forces are calculated at contacting points. At each contacting point, two
springs of normal and shear direction
Figure 3. Inside of the stone wall.
are considered. The shear spring is responded as Coulomb slip model. There is also a limiting strength considered for the normal spring. If normal force exceeds the
limiting strength, the spring will break and the pointtopoint contact will be deleted.
3 ANALYSIS MODEL
An analysis model is shown in Figure 4. The material properties of the stones and soil are shown in Table 1. The height and the shape of the stone wall surface, and the
shape, size and location of the stones are referenced to actual castles. The clumps are used for stones while the soils are modeled by small circle elements. The spring
stiffness and friction of the left boundary and bottom are assumed to be the same as the soil. The time step of the analysis is about 3×10−5 sec and 5% damping is used
in the dynamic analysis.
4 ANALYSIS CASES
Eight cases are shown in Table 2 where the sizes and arrangement locations of the stones were changed are
Figure 4. Analysis model of basic case.
Table 1. Material properties of stones and soil.
Description Parameter Unit Blockstone Cobblestone Soil

Shape – – Rectangle Ellipse Circle
Density ρ 3 25.0 25.0 22.0
kN/m
Porosity n % – – 16
Major diameter d1 m 1.2–1.6 0.2–0.4 0.1
Minor diameter d2 m 0.5 0.2–0.4 0.1
Friction coefficient τ – 0.8 1.5 0.5
Normal spring stiffness kn MN/m 200 100 50
Shear spring stiffness ks MN/m 200 50 25
Normal bond strength nb kN 0 0 0
Shear bond strength sb kN 0 0 2
Page 59
considered under upper loading 8.0 kN/m2. Furthermore, Case 1 and Case 6 are considered as dynamic cases as an earthquake input. As an example, Case 1
includes 6180 balls.
5 ANALYSIS RESULTS AND OBSERVATION
The positions of the resultant force applied on the bottom of the foot blockstone are shown in Table 3.
Table 2. Analysis cases.

Page 60
Table 3. Analysis results.
Case Content Distance X of resultant force (mm) Displacement (mm)

Static analysis 1 Basic case 887 0.3
2 Small diameter cobblestone 1062 0.3
3 Equal diameter cobblestone 762 0.4
4 Loose soil slope 1160 0.1
5 Curved blockstone 417 0.1
6 Small thickness of backfill 1001 6.9
7 Large diameter of cobblestone on the Surface 463 0.3
8 Small diameter of cobblestone on the Surface 911 0.5
Dynamic analysis 1 Basic case 911 8.2
6 Small thickness of backfill 1034 10.8
Figure 5. Force position X.
The position of the resultant force is measured as the distance X from left side of the foot blockstone (see Fig. 5). If this value of the distance is large, then the stone
wall may fall ahead and it becomes unstable. So this distance X is taken as one basic parameter for estimation of stone wall stability. Moreover, the relative
displacement of the top blockstone can be considered in estimation of castle stability and all results are shown in Table 3 for cases of both static and dynamic analysis.
The results are summarized as follows.
1. Case 1 and Case 2 show that using large size cobblestones gives more stable results than using small cobblestones since the distance X of resultant force is large in
the case of the smaller cobblestones.
2. By comparing Case 1, Case 2 and Case 3, it is found that a range of optimal particle diameters of cobblestone exists. By using the optimal size the stone wall will
become more stable.
3. Case 1 and Case 4 show that if a loose soil slope is designed, the stone wall may become more unstable since the distance X of resultant force is large.
4. Case 1 and Case 5 show that the distance X of resultant force by using a curved stone wall is small and it provides a more stable result.
5. Case 1 and Case 6 show that the distance X of resultant force by using large thickness of backfill is small and provides a more stable result.
Figure 6. Collapse simulation analysis model.
6. Case 7 and Case 8 show that the distance X of resultant force by using large cobblestones on the surface layer near blockstones is small and provides a more
stable result.
7. For dynamic analysis, the case of using a small thickness of backfill causes a large earthquake displacement as well.
6 CONCLUSION
A new method for evaluating castle stone wall stability is proposed in this study. The analysis results are in agreement with actual tendency and show that it is possible
to analyze influences of stone size, stone shape, stone wall shape and arrangement location, etc. using particle flow codes. Based on these results, collapse simulations
of actual castle model by an earthquake input are carried out (see Fig. 6). It is possible to simulate the process of slip of soil slope, fracture of foundation, as well as the
collapse and fall of a stone wall. In order to apply this analysis method to the restoration design of an actual castle, it becomes necessary to establish an accurate
method to determine the properties of stone and soil, including both the experiment and simulation method.
Page 61
Rock fractures
Page 62
Page 63

Simulation of rock cutting using distinct element method—PFC2D
S.Lei, P.Kaitkay & X.Shen
Kansas State University, Manhattan, Kansas, USA
ABSTRACT: Distinct element method (PFC2D) was used to model a rock material, Carthage marble, and simulate orthogonal machining of this
material with and without hydrostatic pressure. A methodology was used to apply hydrostatic pressure to the rock specimen during simulation. The
average cutting forces were summarized for different depth of cut, rake angle, and hydrostatic pressure. A parametric study was conducted to
investigate the effects of various parameters on the cutting force. Those parameters include damping coefficient, friction coefficients for particleparticle
and particlewall contact pairs, particle spin, and particle size. It was found that damping coefficient and particle spin have the most significant effects
on the cutting force. Suggestions are proposed to treat those effects properly for more accurate modeling and simulation of rock cutting.
1 INTRODUCTION
Rock drilling in the petroleum industry is one of the application areas that involve interactions between a cutting tool and rock. There is significant demand for
advancement in rock drilling technology to achieve high penetration rate and long tool life, which depends to a large degree on the understanding of the rock cutting
process. Many experimental investigations have contributed to a better understanding for the rocktool interaction, tool wear mechanisms, and thermal aspects during
cutting using a PDC cutter (Karasawa & Misawa 1992, Appl et al. 1993, Kuru & Wojtanowicz 1993, Cooper et al. 1994, Richard 1999), since its development in
the early 1970s. Because experiments are usually costly to run and it is difficult to directly observe the material removal process, numerical simulations have been
performed to elucidate rocktool interactions during rock cutting. Two approaches are generally used to model rocks, depending on how the damage is represented:
either indirectly via constitutive relations or directly by formation and tracking of a large number of discrete cracks.
Many studies have used continuum mechanics based approach such as FEM to study toolrock interactions. For example, Wang & Lhenhoff (1976) studied bit
penetration into rock using FEM; Korinets et al. (1996) used another FEM based code (DIANA) to simulate rock breakage with a spherical/flatbottomed indenter
and disc roller cutter. But very few studies have been found to model rock cutting process. One such study was done by Kou et al. (1999), who modeled the
machining of rock using RFPA2D, a FEM package which has the capability to model inhomogeneous rocks. Using distinct element method (DEM) developed by
Cundall & Hart (1992), Huang (1999) modeled rock samples and conducted simulations of rock cutting on the rock specimens. It was found that material removal is
realized by a ductile flow of the particles at small depths of cut while large fractures are formed at larger depths of cut, which Kou et al. failed to capture in their study
More recently, Lei & Kaitkay (2002) used the same method to study rock cutting under hydrostatic pressure with a simple contact bond model. In those work, both
Huang (1999) and Lei & Kaitkay (2002) have demonstrated the unique capability of the DEM in reproducing the complex fracturing process involved in rock cutting.
In this study, improvements were made to Lei & Kaitkay’s work by using a parallel bond model. Cutting simulations of a common type of rock material, Carthage
marble, were conducted. Hydrostatic pressure was considered to resemble the pressure of drilling mud that is typical for oil drilling operations. A parametric study was
conducted to depict the effects of various parameters on the simulation results. Finally, the findings from this study are summarized and a new approach is proposed for
future improvements in rock cutting simulation. For clarification, the following two pairs of terms are used interchangeably in this paper: ‘particle and ball’, ‘hydrostatic
pressure and confining pressure’.
2 MODELING OF ROCK SPECIMEN
The rock specimen shown in Figure 1 was generated following a threephase specimen generation
Page 64

Figure 1. Specimen generated in PFC2D.
procedure using PFC2D (Itasca 1999). The main parameters used in the process are given in Table 1. The first phase is to generate a well connected, densely packed,
and bonded assembly of circular particles to approximate a solid material. There are four steps involved in this phase: (1) generating a compact initial assembly, (2)
installing specified isotropic stress, (3) eliminating floaters, and (4) forming bonded particle assembly. Parallel bonds are used in this study. In the second phase,
numerical biaxial tests are conducted to study the behavior of the rock specimen. From the biaxial test, the stressstrain curve is obtained and the following properties
are extracted: Young’s modulus of elasticity, Poisson’s ratio, and the compressive strength. This is an iterative process that involves the adjustment of the relevant
parameters in Table 1 and comparison of the properties between the PFC specimen and the actual rock material. The iteration ends when a satisfactory result is
obtained. Biaxial tests under hydrostatic pressure were also performed by installing and maintaining a specified confining pressure on the two sides of the specimen, so
that the compressive strength at the confining pressure was also considered. As shown in Table 2, the properties of the rock specimen match well with those of the
actual material (Terra Tek, Inc. 1987), with errors of less than 10%. Hence it can be stated that the rock specimen generated has properties close to those of Carthage
marble.
In the final phase, a specimen with a gradient of particles was created for cutting simulation. The model is 15 mm in height and 50 mm in width. It consists of 3566
circular disks of unit thickness, and the radii of the balls range from 0.1 to 0.52 mm. Thus a graded assembly is created with the ball size increasing downward from the
top surface. The reason for using a gradient of balls is to have a reasonably large rock specimen but without significantly increasing the number of balls. A balance is
needed between accuracy requirement and computational expense. Since biaxial test results have shown that the material properties remain nearly the same for the ball
radii ranging from 0.1 to 1 mm (Kaitkay, 2002), the differences in ball size within this range should not adversely affect the
Table 1. Parameters for specimen generation.
Parameters Symbol Value

Sample height (mm) H 140
Sample width (mm) W 70
Minimum ball radius (mm) Rmin 0.625
Ball size ratio, uniform distribution Rmax/Rmin 1.3

ρ 2650
Ball density (kg/m3)
Ballball contact modulus (GPa) Ec 35.5
Ball stiffness ratio k n/ks 2.6

Parallel bond radius multiplier 1

Parallelbond modulus (GPa) 35.5

Parallel bond stiffness ratio 2.6

Ball friction coefficient μ 0.577
Parallel bond normal strength, mean (MPa) 45.4

Parallel bond normal strength, std. dev. (MPa) 1
Parallel bond shear strength, mean (MPa) 105
Parallel bond shear strength, std. dev. (MPa) 1
Table 2. Results of numerical biaxial tests.
Property Carthage marble PFC specimen

Young’s modulus of elasticity, E (GPa) 44.8 43.9
Poisson’s ratio, ν 0.24 0.23
Compressive strength at zero pressure, σc (MPa) 103 111
Compressive strength at pressure of 34.4 MPa, σc (MPa) 186 180
properties of the rock specimen. The cutting tool was modeled using wall segments.
3 SIMULATION OF ROCK CUTTING
3.1 Simulation conditions
Simulations were conducted using the PFC specimen that consists of a gradient of particles as shown in Figure 1. The cutting tool is modeled as a rigid body by using
wall elements. The top wall of the specimen is deleted to allow chip formation. The origin of the coordinate system is at the center of the specimen with the Xaxis in the
horizontal direction. The two sides and the bottom of the specimen are constrained with walls to represent the surrounding rocks. Cutting simulation begins by moving
the cutting tool at a specified speed
Page 65
Table 3. Simulation conditions.
Parameter Values
Depth of cut, d, (mm) 0.4, 0.8
Rake angle, α, (deg) 15, 25
Confining pressure, p, (MPa) 3.44, 34.4
Velocity, V, (m/s) 1
in the horizontal direction. Forces acting on the cutter surfaces in both horizontal and vertical directions are extracted from the simulation. The simulation conditions are
listed in Table 3. For simulation under a hydrostatic pressure, a previously developed algorithm was implemented in PFC2D to find all the particles on the boundary
surface and to apply an external force to each particle in the right direction and magnitude that is derived from the confining pressure.
3.2 Simulation results
Figures 2 & 3 show the simulation results for cutting under zero and 34.4 MPa confining pressure, respectively. The depth of cut is 0.8 mm and the rake angle is 15
deg. The cutting tool has moved approximately 25 mm in Figures 2a and 2b. When no confining pressure is applied, as seen in Figure 2a, chip segments have formed in
front of the cutter and many particles have broken loose and are ejected from the bulk material. In contrast, when a confining pressure of 34.4 MPa is applied on the
top boundary, a dense chip and machined surface form. The boundary particles to which the pressure is applied are shown in gray.
The histories of the two force components are shown in Figure 3. Although the forces fluctuate significantly for both cases, cutting with a confining pressure exhibits a
larger force magnitude due to the application of the hydrostatic pressure. As given in Table 2, the compressive strength has an increase by more than 60% under a 34.4
MPa confining pressure. The large force fluctuations in both cases are caused by the breakage of the bonds between particles. Large variations in cutting force were
also noted by Huang (1999) when simulating cutting of a different rock material using PFC2D. Generally speaking, the force in the cutting direction (main force or Fx) is
greater than that perpendicular to it (thrust force or Fy ).
Figure 4 shows the variations of cutting forces with depth of cut for the two rake angles. Both force components increase with depth of cut because of the increased
amount of material to be removed. It also can be seen that the average forces increase as the rake angle increases. This is mainly due to the increased degree of
difficulty in chip flow for a large rake angle. The variations of the average cutting force with confining pressure and rake angle are plotted in Figure 5.
Figure 2. Chip formation after cutting for 25 mm.
Figure 3. Variations of cutting forces with length of cut.
Both force components increase with confining pressure. The main component, Fx, approximately doubles the thrust component, Fy . The average force appears to be
nonlinearly related to confining pressure.
Page 66
Figure 4. Variations of average cutting forces with depth of cut and rake angle.
Figure 5. Variations of average cutting forces with confining pressure and rake angle.
4 EFFECT OF VARIOUS PARAMETERS ON SIMULATION RESULTS
During the simulation process, it was found that there are some parameters that either are difficult to determine accurately or may not represent the actual material well.
Those parameters include damping coefficient, wall friction coefficient, friction coefficient between two particles, particle size, and particle rotation. In this section, these
factors and their effects on the cutting forces are investigated.
4.1 Damping coefficient
A damping coefficient of 0.7 was used for the simulations in Section 3, which is believed to be significantly larger than that of an intact crystalline rock material. The
damping coefficient for a rock is not easy to
Table 4. Parameters of specimens with different damping coefficients.
Damping coefficient 0.7 0.4 0.015

Young’ modulus of elasticity (GPa) 43.9 42.6 42.4
Poisson’s ratio 0.238 0.239 0.239
Compressive strength (MPa) 107 110 106
Parallelbond normal strength, mean (MPa) 53 65 88
Parallelbond shear strength, mean (MPa) 123 150 204
determine, especially when it undergoes a highly dynamic process such as rock cutting. Studies have shown that damages and cracks in brittle materials could
significantly increase material damping (Adams et al. 1973 and Mantena et al. 1986). With consideration of the effect of stress relief fractures on damping coefficient,
Hazzard et al. (2000) used a damping coefficient of 0.015 in their simulation of biaxial tests for Lac du Bonnet Granite. Because a good representative damping
coefficient for the rock material undergoing cutting is not known, the effects of damping coefficients on rock cutting are assessed in terms of average cutting forces for
three selected damping coefficients in this study: 0.7, 0.4, and 0.015.
In order to study how damping coefficient influences the cutting forces, three PFC specimens were created as given in Table 4 with approximately the same macro
properties. Only parallelbond normal and shear strength were adjusted from those in Table 1 according to the damping coefficient. The simulations in this subsection
were conducted using a newer version of PFC2D Ver. 3.0 (Itasca 2003). Cutting of each specimen was simulated at a depth of cut of 0.8 mm, a rake angle of 15 deg,
and zero confining pressure.
Figure 6 shows the simulation results for two different damping coefficients after the cutting tool has moved approximately 15 mm. It appears that a large volume of
small chip segments have accumulated in front of the cutting tool and less particles or clusters have been ejected for the highly damped specimen. In contrast, for the
damping coefficient of 0.015, a few large chip segments are formed in front of the tool and more particles “fly away” from the specimen because of less resistance to
their motion. The corresponding histories of the cutting forces are plotted in Figure 7. The difference is clearly seen in both the profile and magnitude of the two force
components. For the highly damped specimen, there exists a noticeable level of “background” force superimposed by large fluctuations throughout the entire length of
cut, which indicates a fairly stable chip load in front of the cutting tool. When damping is very small, on the other hand, only sparse force spikes are noted which are
separated by gaps where the force is virtually zero.
Page 67
Figure 6. Chip formation after the tool has traveled approximately 15 mm.
Figures 6 & 7 suggest that the chip formation process may be different for the two significantly different damping coefficients. We postulate that, for the highly damped
material, cracks develop relatively slowly and stably in the vicinity of the cutting tool and thus result in small chip segments which accumulate in front of the cutting tool.
When damping coefficient is much smaller, relatively large fractures develop quickly by fast creation of new cracks at the onset of an initial bond failure and thus
resulting in formation of larger chip segments. The ejection of a large chip segment leaves a virtually empty space in front of the cutting tool and little resistance is
encountered by the cutting tool until it passes the void zone. The postulation of fast crack propagation may be explained by the work of Hazzard et al. (2000), who
have shown that, in a dynamic simulation of a lowdamping rock, the waves emitted by cracking events could cause more cracks to form in a nearby region. In cutting,
when the cutting tool moves into the material, cutting force increases rapidly because of the resistance of the intact bonds until it peaks and a crack or a few cracks
initiate. Then the stress waves emitted by the cracks cause more cracks which coalesce to form large fractures, and hence cause segmentation of large chips. Less
Figure 7. Variations of cutting forces with length of cut.
Figure 8. Variations of average cutting forces with damping coefficient.
energy is required to remove materials when damping is small, because the bond breakage within the specimen is facilitated by the stress waves originated from the
microcracks formed due to the advancement of the cutting tool. In other words, cutting is assisted by the passing waves capable of causing bond failure within the
material.
Figure 8 shows the variations of the average cutting forces with damping coefficient. The relationship seems to be more linear for the thrust force. As the damping
coefficient increases, there is a significant increase in both force components. For example, the main cutting force for damping coefficient of 0.7 is approximately 24
times that of 0.015. Hence, knowledge of an appropriate value for damping
Page 68
coefficient is very important for simulating highly dynamic processes such as rock cutting.
4.2 Wall friction coefficient
The wall friction coefficient was set at 1.0 for the simulations. However, the actual friction coefficient between the cutter surface and Carthage marble was not known,
which could be as low as 0.1 or 0.2. To understand this effect, simulations were conducted using the graded particle assembly for the wall friction coefficients of 0.2
and 1.0. The depth of cut, rake angle and confining pressure were fixed at 0.8 mm, 15 deg and 3.44 MPa, respectively. Those values also remain the same for the
simulations hereafter. The average cutting forces were calculated for a tool travel between 8 and 20 mm for both simulations. Figure 9 shows the variations of the
average cutting forces with wall friction coefficient. As seen from the figure, there is a significant change in the thrust force but only about 10% change in the main force.
While cutting using a rake angle of 15 deg, the main material movement along the cutting tool is in the vertical direction. As the friction coefficient increases, the
resistance to the vertical movement of the particles increases and hence there is a corresponding increase in the thrust force.
4.3 Particle friction coefficient
Graded assemblies were created to study the effect of friction coefficient between particles on the cutting forces. PFC specimens were generated having similar micro
properties but different coefficients of friction between particles. The coefficient of friction between particles has been shown to have negligible effects on material
macro properties by Kaitkay (2002) and Huang (1999). This is understandable since the coefficient of friction will most likely only affect the properties once the
particles are free to move. Simulations were run for three different values of friction coefficients: 0.4, 0.6 and 0.8. The cutting forces
Figure 9. Variations of average cutting forces with wall friction coefficient.
generated from the various assemblies are plotted in Figure 10 against the friction coefficient. It is shown that as the particle friction coefficient increases, there is a
corresponding increase in the average cutting force. The increase for the main force is greater than for the thrust force. This is because that the increase in the coefficient
of friction between particles will increase the resistance to movement of the particles and hence, the cutting forces. However, while conducting biaxial tests, there were
no appreciable movements of the particles. Therefore, the macro properties are not affected considerably with an increase or decrease in friction coefficient.
4.4 Minimum particle size
To understand the effect of minimum particle size of a PFC specimen on the average cutting forces, simulations were conducted using different minimum particles sizes.
The range for the minimum particle size was selected to be from 0.05 to 0.2 mm, although the optimal range for this parameter is not known. Graded specimens having
the same micro properties but different minimum particle radius were generated. The cutting tool was moved to a distance of 15 mm for each simulation and the
average cutting force is found between 8 to 15 mm of the tool travel.
Figure 11 shows the variations of the average cutting force with minimum particle size. It was observed that the variations in cutting forces were as high as 30%. The
ratio between the main force and the thrust force does not change considerably between 0.05 to 0.125 mm of minimum particle size. The ratio reduces dramatically for
the 0.2 mm minimum particle radius. If the particle size is reduced considerably and the specimen dimensions are kept constant, there would be a corresponding
increase in the number of particles within the assembly and thus computational time. At the same time the particle size should be small enough to provide for enough
degrees of freedom and enough micro events to occur. Therefore, the ratio of depth of cut to particle radius, d/R, has to be definitely greater
Figure 10. Variations of average cutting forces with particle friction.
Page 69
than 1, but no optimal value for the d/R ratio has been found. Hence, it can be seen that the d/R ratio does affect the cutting forces.
4.5 Particle rotation
When conducting simulations, the particles were set to rotate freely about their own axes of rotation. Therefore, large degrees of rotations were observed for the
particles with broken bonds. However, for an actual crystalline rock material that consists of grains, such large rotation either does not exist or should be limited, due to
the varying shapes and sizes of the grains. Therefore, to study the effect of selfrotation or spin of the particles about its own axis on the average cutting forces,
simulations were performed by fixing the rotation of the particles in PFC2D. Figure 12 shows the variations of the average cutting forces with the rotational condition of
a particle.
From Figure 12 it can be seen that the rotation or spin of the particles affects the cutting forces significantly. The main force increases almost five times and the thrust
force has a twofold increase when rotation is disabled. The particles now have to slide along each other and the wall surface instead of rolling
Figure 11. Variations of average cutting forces with minimum ball radius.
Figure 12. Variations of average cutting forces with particle rotation.
and hence create greater resistance to movement. Figures 13a & 13b show the damage zone while cutting with and without free spin. When particle spin is not allowed,
the cutting takes place mostly as brittle failure with bonds breaking in normal as well as shear mode. When particle rotation is allowed within the assembly, the particle
bonds fail mainly in normal mode. Since the bond shear strength is approximately 2.3 times the normal strength as indicated in Table 1, a larger force is required to
break the bonds in shear, thus resulting in an increase in cutting force for the case of no spin. Another reason for the larger force would be the decrease in the degrees
of movement for the particles. Due to the lack of rotation, the particles are forced to slide against each other and the wall, thus increasing the difficulty for chip flow as
indicated by the short chips in Figure 13a. It is not clear why the fixation of rotation causes such a considerable change in the cutting forces and damage mode within
the specimen. Particle rotation is considered as a nonquantifiable parameter. Although there exist two
Figure 13. Damage zone with spin and without spin (Black: Shear failure, Grey: Normal failure).
Page 70
distinct conditions, the condition with free spin is considered closer to the experimental conditions. This is because even the particles with broken bonds are not allowed
to spin when rotation is fixed. In real conditions, the particles with bonds intact would definitely have no rotation but the particles with broken bonds would be able to
spin, although the extent of particle spin in an actual cutting process is not known.
5 CONCLUSIONS
Rock cutting simulation with and without hydrostatic pressure was carried out using distinct element method. With appropriate selection of particle scale properties for
the particle assembly, a rock specimen was created in PFC2D whose major mechanical properties closely match those of Carthage marble. A methodology was used
to apply a confining pressure to the boundary particles during the simulation. A confining pressure was found to assist chip growth and suppress particle ejection. The
simulation results show an increase in cutting force with confining pressure. This increase appears nonlinear and the main force is approximately two times that of the
thrust force. Loose particles were not observed from the machined surface when cutting with confining pressure. Cutting force increases with depth of cut and rake
angle as expected.
Various parameters affect the cutting force based on the simulation results. A parametric study was presented and it was found that although most of these
parameters do not affect the material macro properties, they do affect the average cutting forces to different degrees. Among those parameters, the damping coefficient
and particle spin were found to have the most significant effects on the forces. The lack of accurate data for these parameters is seen as a roadblock to accurate
simulations of rock cutting. To improve rock cutting simulation, one of the future research is to obtain realistic values for these parameters so that a PFC specimen can
represent a real material closely. Another future research is to model the microstructure of a crystalline rock in a PFC specimen, i.e., to use clusters of particles to
model the grains of various shapes and sizes based on the microstructure of a real material. This could effectively limit the ease of particle selfspin when only circular
particles are used to model a material. The results presented in this study also place an argument on the use of the biaxial test as the only means to define the properties
of a PFC specimen and the subsequent assumption that the material will behave similarly in machining. Although the biaxial test gives a reasonable approximation for
the material, there are various parameters that have little effect on the macro properties from the biaxial test but have a significant impact on the cutting forces. It is
suggested that perhaps the post peak behavior of a PFC specimen should be characterized and matched with that of a real material. Overall, it can be said that the
DEM provides a valuable approach in the simulation of the rocktool interactions during rock cutting. One application is in the petroleum industry where cutting
simulation can play an important role in the design of drill bits for efficient underground drilling.
ACKNOWLEDGEMENT
The authors wish to thank the National Science Foundation under Grant No. DMI0134579 for partial financial support of this work.
REFERENCES
Adams, R.D., Flitcroft, J.E., Hancox, N.L. & Reynolds, W.N. 1973. Effects of Shear Damage on the Torsional Behavior of Carbon Fiber Reinforcement Plastics. Journal
of Composite Materials 7:68–75.
Appl, F.C., Wilson, C. & Lakshman, I. 1993. Measurement of Forces, Temperatures and Wear of PDC Cutters in Rock Cutting. Wear 169(1):9–24.
Cooper, G., Liu, Z. & Yang, M. 1994. Wear of Thermally Stable Diamond during Rock Cutting. Proceedings of the EnergySources Technology Conference 56:181–
186.
Cundall, P.A. & Hart, R.D. 1992. Numerical Modeling of Discontinua. Engineering Computations 9:101–113.
Hazzard, J.F. & Young, R.P. 2000. Micromechanical Modeling of Cracking and Failure in Brittle Rocks. Journal of Geophysical Research 105(B7):16683–16697.
Huang H. 1999. Discrete Element Modeling of ToolRock Interaction. Ph.D. Thesis, University of Minnesota.
Itasca Consulting Group, Inc. 1999. PFC2D—Particle Flow Code in 2 Dimensions, Version 2.0. Minneapolis: Itasca.
Kaitkay, P. 2002. Modeling of Rock Cutting Using Distinct Element Method. M.S. Thesis, Kansas State University.
Karasawa, H. & Misawa, S. 1992. Development of New PDC Bits for Drilling of Geothermal Wells—Part 1: Laboratory Testing. Journal of Energy Resources
Technology, Transactions of the ASME 114(4):323–331.
Korinets, A.R., Alehossein, H., Chen, L., Lim, W. & Baker, G. 1996. DIANA Modeling of a Rolling Disc Cutter and Rock Indentation. In M.Aubertinin, F.Hassani &
H.Mitri (eds), Rock Mechanics, Tools and Techniques: 647–654. Rotterdam: Balkema.
Kou, S.Q., Lindqvist, P.A., Tang, C.A. & Xu, X.H. 1999. Numerical Simulation of the Cutting of Inhomogeneous Rocks. International Journal of Rock Mechanics and
Mining Sciences 36:711–717.
Kuru, E. & Wojtanowicz, A.K. 1993. Compound Effect of Cutting Depth and Bit Dull on Cutters’ Temperature for Polycrystalline Diamond Compact Bits. Journal of
Energy Resources Technology, Transactions of the ASME 115(2): 124–132.
Lei, S. & Kaitkay, P. 2002. Micromechanical Modeling of Rock Cutting under Pressure Boundary Conditions Using
Page 71
Distinct Element Methods. Transactions of the North American Manufacturing Research Institution of SME 30: 207–214.
Mantena, R., Gibson, R.F. & Place, T.A. 1986. Damping Capacity Measurements of Degradation in Advanced Materials. SAMPE Quarterly 17(3):20–31.
Richard, T. 1999. Determination of Rock Strength from Cutting Tests. M.S. Thesis, University of Minnesota.
Terra Tek, Inc. 1987. Terra Tek Rock List, March 1987.
Wang, J.K. & Lehnhoff., T.F. 1976. Bit Penetration into Rock: A Finite Element Study. International Journal of Rock Mechanics, Mining Science 13(1):11–16.
Page 72
Page 73
Evaluation of behavior of EDZ around rock cavern by AE measurements and PFC
simulation
K.Aoki & Y.Mito
Dept. of Urban and Environmental Engineering, Faculty of Engineering, Kyoto University, Kyoto, Japan
T.Mori
Kajima Technical Research Institute, Tokyo, Japan
T.Maejima
Tokyo Electric Power Company, Tokyo, Japan
ABSTRACT: In highly stressed rock, the stressinduced cracks are often developed by the excavation of caverns and EDZ (Excavation Disturbed
Zone) will be formed around the caverns. The monitoring and control of EDZ is the one of the most important required technology for the design of
caverns such as the highlevel radioactive waste repository. Thus the authors are developing the effective design and construction system for
underground caverns based on the monitoring of EDZ using AE (Acoustic Emission) that sensitively detects the generation of the stressinduced
cracks. In this study, the behaviors of several rock specimens at different stress levels of the loading path, through yield stress, peak stress to residual
stress, are examined from the results of laboratory triaxial compression tests using a stiff and servocontrolled testing system. As AE sensors are
attached to the both pedestals, they provide complete stressstrain curve and AE data. Each of the complete stressstrain curves obtained from the
triaxial compression tests is fitted by the stressstrain curve of twodimensional bonded particles model simulated by PFC with appropriate
microproperties determined so as to express the macroscopic behavior of the tested rock specimen. Numbers of tension and/or shear cracks
generated in the simulation are examined with the corresponding AE parameters such as AE counts, predominant AE frequency and mvalue
measured in the tests. The relationship between rock fracturing behavior and features of AE parameters is clarified. The field measurements of the
EDZ developed around the actual large underground cavern in highly stressed rock and AE behaviors are conducted in accordance with the progress
of the excavations. As the results, the spatial distribution of EDZ is determined by a number of measurements such as the initiated crack observation
and stress measurement. Furthermore, the prediction of the behavior of EDZ by PFC simulation of the bonded particles model is carried out using the
results of the insitu rock tests and initial rock stress measurement before the cavern excavation. The results of the simulation shows good agreement
with the actual measurement results, which provides high performance for the behavior analysis of underground excavation together with a consistent
interpretation of the behaviors of AE parameters.
1 INTRODUCTION
In highly stressed rock, the stressinduced cracks are often developed by the excavation of caverns and EDZ (Excavation Disturbed Zone) will be formed around the
caverns. The monitoring and control of EDZ is the one of the most important required technology for the design of caverns such as the highlevel radioactive waste
repository. Thus the authors are developing the effective design and construction system for underground caverns based on the monitoring of EDZ using AE (Acoustic
Emission) that sensitively detects the generation of the stressinduced cracks.
In this study, the behaviors of several rock specimens at different stress levels of the loading path, through yield stress, peak stress to residual stress, are examined
from the results of laboratory triaxial compression tests using a stiff and servocontrolled testing system. As AE sensors are attached to the both pedestals, they provide
complete stressstrain curve and AE data. Each of the complete stressstrain curves obtained from the triaxial compression tests is fitted by the stressstrain curve of
twodimensional bonded particles model simulated by DEM with appropriate microproperties determined so as to express the macroscopic behavior of the tested rock
specimen.
Page 74
Numbers of tension and/or shear cracks generated in the simulation are examined with the corresponding AE parameters such as AE counts, predominant AE
frequency and mvalue measured in the tests. The relationship between rock fracturing behavior and features of AE parameters is investigated.
The field measurements of the EDZ developed around the actual large underground cavern in highly stressed rock, and the AE behaviors, are also conducted in
accordance with the progress of the excavations. Furthermore, the prediction of the behavior of EDZ by DEM simulation of the bonded particles model is carried out
using the results of the insitu rock tests and initial rock stress measurement before the cavern excavation.
2 AE PARAMETERS
Acoustic Emission (AE) is a transient elastic wave, which is generated by the rapid release of strain energy within a solid material. In this study, the behaviors of the
following three AE parameters are fully discussed, comparing with the failure behaviors of rocks under the specific stress condition.
2.1 AE counts
AE counts denote the number of AE events with individual destruction of the solid structure, which is usually observed within a specific period of time.
AE counts is a well known parameter that directly represents microcrack initiation events, and its relation to rock failure has been often discussed in rock mechanics
field.
2.2 Predominant AE frequency
Predominant AE frequency varies with progress of rock failure dues to the phenomenon that higher frequency wave is generally attenuated more than lower frequency
wave.
It is inferred that the ratio for which high frequency AE accounts increases when larger compressive stress is applied to the rock, because rock structure becomes
denser and pores that intercept the high frequency AE decreases. On the other hand, it is inferred that the ratio for which low frequency AE accounts increases when
the rock is failed, because the initiated cracks intercept higher frequency AE. Thus, predominant AE frequency has the possibility to reflect both the change in rock
stress and the change in rock property.
In this study, dominant AE frequency is defined as the expectation of power spectrum density distribution of AE frequency, which can be obtained from FFT
analysis.
2.3 mvalue
It can be inferred that AE amplitude, which reflects the magnitude of the released strain energy, also varies with progress of rock failure. For instance, the gradual
microscopic failure, which the magnitude of the released strain energy is small, occurs before the rock stress exceeds the peak strength, while the explosive
macroscopic failure, which the magnitude of the released strain energy is small, occurs immediately after the rock stress exceeds the peak strength.
Ishimoto & Iida (1939) found out a power law concerning the amplitude distribution of seismic events.
N(A)dA=HA−mdA
(1)
where H and m (mvalue)=constant; and N(A)dA= the number of earthquakes with amplitude between A and A+dA. In this study, such mvalue is used as the
quantitative parameter that reflects the AE amplitude distribution.
3 LABORATORY TEST
A series of laboratory triaxial compression tests using a stiff and servocontrolled testing system with AE sensors attached to the both pedestals, which provides
complete stressstrain curve and AE data, is conducted in order to grasp the behaviors of AE parameters of several rock specimens at each stress level of the stress
path through yield stress, peak stress and residual stress.
3.1 Test apparatus
A stiff and servocontrolled triaxial compression testing system that gives the complete stressstrain curve is used in this study. Two AE sensors are attached to the both
pedestals as shown in Figure 1.
Broadband AE sensors are connected to pre and main amp system, and then the waveforms converted by ADC (analog to digital converter) are recorded on the
database. 100 kHz to 1 MHz band pass filter is used to prevent interferences such as electric noise.
3.2 Test cases
The rock samples are collected from highly stressed rock masses around the actual underground powerhouse located at approximately 500 m depths, which are
composed of Mesozoic sedimentary rocks. The cylindrical specimen with a height of 100 mm and diameter of 50 mm is subjected to the fixed confining pressure. The
displacementcontrolled loading with increasing displacement rates of 0.05 mm/min is performed until the vertical strain reaches an extent of 1.5 to 2.5%.
Page 75
Figure 1. Triaxial test machine and AE sensors.
Table 1. The description of the specimen and the applied confining pressure.
Case Rock type Remarkable weak plane Confining pressure

1A Mudstone Nothing particular 10 MPa
1B Sandstone 2 sharp quartz veins 10 MPa
2A Sandstone Nothing particular 0.5 MPa
2B Sandstone Nothing particular 6 MPa
2C Sandstone Nothing particular 15 MPa
Table 1 shows the description of the specimen and the applied confining pressure. The test series of 1A to 1B intend to examine the relationship between complete
stressstrain curve and behaviors of AE parameters, while the series of 2A, 2B and 2C intends to examine the effect of confining pressure on the behaviors of AE
parameters.
3.3 Relationship between complete stressstrain curve and behaviors of AE parameters
Figure 2 shows the complete stressstrain curve and behaviors of AE parameters, which are observed in the test cases 1A and 1B. The stressstrain curve in the case
1A shows ductile behavior, while that in the case 1B shows brittle behavior.
The following relationship between complete stressstrain curve and behaviors of AE parameters can be inferred from the figure.
1. AE counts starts to increase when the stress exceeds yield strength, shows the maximum value immediately after the stress reaches peak strength,
Figure 2. The complete stressstrain curve and behaviors of AE parameters (the cases 1A and 1B).
decreases until the stress reaches residual strength, and becomes close to zero.
2. Predominant AE frequency increases as stress level increases, drastically decreases when the stress reaches peak strength, continues to decrease until the stress
reaches residual strength, and becomes stable.
3. mvalue suddenly decreases and becomes stable when the stress reaches around yield strength, and suddenly increases and becomes stable when the stress reaches
residual strength.
4. The difference of the behaviors of AE parameters between the cases 1A and 1B is found in the behaviors from peak to residual, in which the magnitudes of negative
slope of complete stressstrain curve are completely different. As vertical strain increases after peak, AE parameters gradually change in the ductile case 1A, those
suddenly change in the brittle case 1B.
3.4 Effect of confining pressure on behaviors of AE parameters
Figure 3 shows the complete stressstrain curve and behaviors of AE parameters, which are observed in the
Page 76
Figure 3. The complete stressstrain curve and behaviors of AE parameters (the cases 2A, 2B and 2C).
test cases 2A, 2B and 2C with a confining pressure of 0.5 MPa, 6 MPa and 15 MPa, respectively. The peak strength and the residual strength with higher confining
pressure show higher values, while the elastic constants in the three cases show almost similar values. On the other hand, the negative slope of complete stressstrain
curve with higher confining pressure shows gentler.
The followings can be inferred from the figure.
1. The relationship between complete stressstrain curve and behaviors of AE parameters (AE counts, Predominant AE frequency and mvalue) is the same as
mentioned in 3.3.
2. The magnitude of variation of AE parameters is larger in the case of lower confining pressure.
3. The difference of the behaviors of AE parameters between the cases is mainly found in the behaviors after peak to residual, in which the magnitudes of negative
slope of complete stressstrain curve are different, as well as the cases in 3.3.
4 PFC SIMULATION OF LABORATORY TEST
4.1 PFC modeling
The contact bonded particles model on the basis of the twodimensional Particle Flow Code, PFC2D, (Itasca 2002) is used for the simulation of the laboratory tests. A
series of numerical biaxial tests are implemented by applying the servocontrolled confining pressure and the displacementcontrolled vertical load, as applied in the
triaxial tests. The values of microproperties of the bonded particles are specified such that the complete stressstrain behavior is similar to that of the tested specimen.
4.2 Curve fitting
The complete set of microparameters that characterizes a contactbonded material is given by the following:
1. Ballball contact modulus, EC
2. Ball stiffness ratio, k n /ks
3. Ball friction coefficient, μ
4. Contactbond normal strength, σC
5. Contactbond shear strength, τC
The macroscopic elastic behavior of the bonded particles, which is perceived before yield stress, is governed by the values of EC and k n /ks. Each of the contact bond
strengths, σC and τC is described with its mean and standard deviation. The mean value of the strength influences the magnitude of peak strength while the standard
deviation of the strength influences on the increment from yield strength to peak strength. The value of the friction coefficient, μ affects on the residual strength.
Considering such basic relationship between microparameters and macroscopic response, the microparameters are calibrated by try and error method such that the
mean deviation between those curves shows within 10% of the peak strength.
Table 2 shows the sets of microparameters, each of which is determined for each test case. Only in the case 2 A, two weak planes whose bond strength is lower
than the other parts are specified in the bonded particles, corresponding to the actual configuration of the quartz veins in the tested specimen. Further, for the cases 2B
and 3B, an identical bonded particles model, which can reproduce the complete stressstrain curve of the both cases with different confining pressure, is
Page 77
Table 2. Mechanical microproperties setting for the simulation.
Cases EC [GPa] k n /ks μ σC [MPa] τC [MPa]

m s/m m s/m
1A 52 10 0.3 83 0.88 83 0.88
1B* 75 10 0.01 280 0.2 280 0.2
2A 75 10 0.005 100 0 100 0
2B & 2C 75 10 0.2 100 0.1 100 0.1
* The bond strength along the weak plane is 0.14 MPa.
Figure 4. An illustration of (a) complete stressstrain curve that is reproduced by PFC simulation (Case 1A) with (b) sketch of the rock
specimen after the test and (c) initiated crack distribution of the bonded particles after the simulation.
specified. Figure 4 shows an illustration of the complete stressstrain curve (case 1A) that is reproduced by PFC simulation.
4.3 Relationship between complete stressstrain curve and microcrack initiation
Figure 5 shows the transition of crack distribution in the cases 1A and 1B as vertical strain increases. The first (left) two figures of each case show the microcrack
distributions when the stress level reaches yield strength and peak strength, respectively. The next (right) six figures show the crack distribution at every strain increment
of 0.2% after differential stress reaches peak strength. White circle and black circle express the newly initiated crack due to tensile stress (tensile crack) and shear
stress (shear crack), respectively. Gray circle also expresses previously initiated crack. The radius of the circle represents the increment of the distance between the
bondfailed balls (equivalent to increment of aperture), which is strongly related to the released strain energy when microcrack occurs.
The following relationship between complete stressstrain curve and microcrack initiation can be inferred from the simulation results.
1. Before differential stress reaches yield strength, only tensile cracks with small released energy are uniformly initiated in all the parts of specimen in the case 1A, while
most of microcracks are initiated along the weak planes in the case 1B.
2. As differential stress increases from yield to peak, tensile and shear cracks with larger released energy are initiated in particular next to the previous cracks due to
local stress concentration. The crack initiation event increases suddenly.
3. Immediately after differential stress reaches peak strength, microcracks with large released energy are initiated along the macroscopic conductive fractures, which
are formed previously. The number of crack initiation events and their released energy becomes the maximum.
4. As differential stress decreases from peak to residual, the number of crack initiation events and their released energy also decrease.
On the other hand, the difference between the cases 1A and 1B is mainly found in the spatial distribution of microcrack. The most of microcracks are initiated along
thin shear bands bounded by weak plane in the brittle case 1B, while microcracks are initiated in wide shear bands and the other parts in the ductile case 1A. The
number of crack initiation events decreases suddenly in the brittle case 1B, while the microcrack initiation with small released energy successively occurs in the ductile
case 1A.
4.4 Effect of confining pressure on microcrack initiation
Figure 6 shows the transition of crack distribution in the cases 2A, 2B and 2C in the same manner as Figure 5.
The following can be inferred from the simulation results.
1. The relationship between complete stressstrain curve and microcrack initiation is the same as mentioned in 4.3.
Page 78
Figure 5. The transition of spatial distribution of microcracks as vertical strain increases in the cases 1A and 1B.
Figure 6. The transition of spatial distribution of microcracks as vertical strain increases in the cases 2A, 2B and 2C.
Page 79
2. Once the macroscopic shear bands are formed immediately after the peak, a few tensile cracks are initiated only along them in the low confining pressure case 2A.
On the other hand, in the higher confining pressure cases 2B and 2C, different shear bands are successively formed by shear cracks even if the strain becomes large.
5 RELATIONSHIP BETWEEN AE PARAMETERS AND ROCK FAILURE
The relationship between AE parameters and rock failure is examined here based on the abovementioned experimental and numerical investigations. The measured
behavior of each AE parameter is compared to the computed crack initiation parameter, which represents crack initiation behavior, along an identical stress path.
5.1 AE counts and number of crack initiation events
Figure 7a shows the comparison between the measured AE counts and the computed number of crack initiation events in the case 1A. The behavior of AE counts is
consistent with the variation of crack initiation events along an identical stress path.
Both AE events and crack initiation events start to increase when the stress exceeds yield strength, show the maximum value immediately after the stress reaches
peak strength, decrease until the stress reaches residual strength, and become close to zero.
5.2 Predominant AE frequency and porosity
Figure 7b shows the comparison between the measured predominant AE frequency and the computed porosity of the bonded particles in the case 1A. The behavior of
predominant AE frequency can be interpreted using the variations of stress and porosity of the bonded particles along an identical stress path.
As mentioned in 2.2, when larger compressive stress is applied to the rock, the ratio for which high frequency AE accounts increases, because rock structure
becomes denser and pores that intercept the high frequency AE decreases. On the other hand, when the rock is failed, the ratio for which low frequency AE accounts
increases, because the initiated cracks intercept higher frequency AE.
It is clear that predominant AE frequency increases as the stress level increases, start to decrease as the crack initiation events increase, drastically decrease when
the porosity suddenly starts to increase (immediately after differential stress reaches peak strength), and continue to decrease as the porosity continues to increase.
5.3 mvalue and total increment of crack aperture
Figure 7c shows the comparison between the measured mvalue and the computed total increment of crack aperture in the case 1A. The behavior of mvalue can be
interpreted using the total increment of crack aperture along an identical stress path.
Both mvalue and total increment of crack aperture represent the magnitude of crack initiation. Low
Figure 7. Comparison between AE parameters and crack initiation parameters as vertical strain increases in the case 1A.
Page 80
Figure 8. AE parameters and crack initiation parameters in the cases 2A, 2B and 2C with different confining pressure.
mvalue and large increment of crack aperture imply the crack initiation with high released strain energy.
In general, gradual microcrack initiation with small released strain energy progresses in the beginning of loading, and then the explosive crack initiation with large
released strain energy occurs when differential stress varies around peak strength. After differential stress reaches residual strength, the released strain energy becomes
small again. The both behaviors of the measured mvalue and the computed total increment of crack aperture, clearly explain such a general rock behavior along the
stress path.
5.4 Effect of confining pressure
Figure 8 shows the comparison between AE parameters and crack initiation parameters in the cases 2A, 2B and 2C with different confining pressure. The both
variations of the measured AE parameters and the computed crack initiation parameters with different confining pressure, consistently explain the different processes of
rock failure in relation to the confining pressure.
6 FIELD MEASUREMENT IN LARGE ROCK CAVERN
6.1 The cavern for research
The field measurement data, which was obtained from a large rock cavern for underground power station,
Figure 9. Geological section of the cavern for research. M1=mudstone, S1=sandstone, CG1=conglomerate, CG2=coarse sandstone,
FS1=sandstone, and FS2=sandstone.
is discussed here. This rock cavern with a height of 52 m, span of 33 m, and length of 200 m was excavated by top headings and 12 bench cuts at the depth of 500 m
in hard Mesozoic sedimentary rocks with a few closed rock joints. Figure 9 shows the vertical geological section. In this study, the measured data, which was obtained
from the point at depth of 3 m from the cavern wall along the horizontal measuring line H10 at the level of the bench No. 4, is particularly examined in detail.
Page 81
Figure 10. The measurement result at depth of 3 m from the cavern wall along the measuring line H10 at the level of the bench No. 4.
6.2 Relationship between stress, strain and AE parameters
Figure 10 shows the time series data of change in stress, strain, AE counts, predominant AE frequency, and mvalue at depth of 3 m along H10.
The following relationship between stress/strain and behaviors of AE parameters can be inferred based on the conclusion of the laboratory test.
1. AE event was first observed when the bench No. 1 was excavated. A little change in stress was observed although strain could not be perceived at this time. This
fact shows that AE can capture a premonition of EDZ with high accuracy.
2. When excavating the benches No. 2 and No. 3, AE counts and predominant AE frequency increased and mvalue decreased, as stress level increased. Those
behaviors of AE parameters consistently show the fact that stress concentration progressed and stress induced cracks increased at this time.
3. When excavating the benches No. 4 and No. 5, AE counts showed maximum, predominant AE frequency decreased, and mvalue became constant as stress
showed its maximum level and strain increased. These behaviors of AE parameters consistently show the fact that the stress had already reached the peak strength
of the rock at this time.
4. When excavating the benches No. 6, AE counts became close to 0 and mvalue suddenly increased and became its initial level, as stress and strain level became
constant. These behaviors of AE parameters consistently show the fact that the stress had already reached the residual strength of the rock and the EDZ had
completely been formed at this time.
Figure 11. Global model and sector model for simulation.
7 PFC SIMULATION OF EXCAVATION OF ROCK CAVERN
7.1 PFC modeling
The contact bonded particles model on the basis of the twodimensional Particle Flow Code, PFC2D is also used for the simulation of the excavation of the rock
cavern as discussed in 6.2, in order to capture the formation of the EDZ.
The following procedure is implemented to grasp the detailed distributions of stress and initiated cracks in rock mass, as the bench cuts progresses.
1. The values of microproperties of the bonded particles are specified such that the macroscopic behavior of rock mass is corresponded to the result of a series of in
situ tests on rock deformability and rock strength.
2. The simulation using the global model is implemented. The inclined maximum and minimum principal stress, which is predetermined by initial rock stress
measurement, is applied to the global model. The top of the cavern is removed as shown in Figure 11 and PFC simulation is implemented.
Page 82
Figure 12. Distributions of differential stress and initiated crack, which are computed by PFC.
3. The sector model, which contains about 23000 balls with a diameter of 12–20 cm are constructed. Referring the stress distribution, which is calculated from the
global model simulation, the initial stress condition of sector model is specified. As each of the benches No. 1 to No. 7 is sequentially removed from the model, the
PFC simulation is implemented.
7.2 Relationship between stress, strain and rock fracturing
Figure 12 shows the computed stress distribution and initiated crack distribution as the bench cut excavation progresses.
Figure 13. Total crack aperture distribution along H10 line, which is measured by borehole television observation.
Figure 14. Comparison between the measured stress and the computed stress by PFC (and the computed stress by FEM) as the
excavation progresses (at depth of 3 m from the cavern wall along the measuring line H10).
When the benches No. 1 and No. 2 are excavated, stress concentration occurs only in the neighborhood of the excavated benches and tensile cracks are mainly
initiated there. When the benches No. 3 and No. 4 are excavated, macroscopic fracture is formed along the geological boundary. When the bench No.5 is excavated,
the cracks with large released strain energy are initiated along the macroscopic fracture, and the region of stress concentration moves to the deeper region from the
cavern wall. When the bench No. 6 is excavated, only a few tensile cracks with small released energy occurs, while shear cracks start to occur in the deeper parts
along the geological boundary.
Figure 13 shows total amount of initiated crack aperture along H10 line, which is measured by borehole television observation when the benches No. 3 and No. 6
are excavated, respectively. The computed crack distribution as shown in Figure 12 shows very good agreement with this figure.
Figure 14 shows the comparison between the measured stress and the computed stress by PFC as the excavation progresses. In this figure, the computed stress by
FEM considering strainsoftening behavior of rock mass is also plotted. The computed stress by PFC shows very good agreement with the measured value. The
computed stress by FEM also shows good agreement, however the stress value when the bench No. 5 is excavated still increases, although the measured stress
decreases. It is inferred that this dues to
Page 83
Figure 15. Comparison between the measured AE parameters and the computed crack initiation parameters as the excavation progresses
(at depth of 3 m from cavern wall along the measuring line H10). The broken line shows stressstrain curve.
the large element size in FEM with a dimension of 2 m and more, which sometimes causes the imreasonable stress concentration, while PFC uses small balls with a
diameter of 10 to 12 cm, by which gradual crack initiations can be represented.
Finally, the relationship between AE parameters and rock failure is examined here based on the abovementioned measured and computed results. The measured
behavior of each AE parameter is compared to the computed crack initiation parameter. Figure 15 shows the comparison between the measured AE parameters and
the computed crack initiation parameters by PFC as the excavation progresses in the same manner as the laboratory examination case in section 5, and all the relations
are consistent with the conclusions, which are provided in section 5.
8 CONCLUSION
In this study, the behaviors of several rock specimens at different stress levels of the loading path, are examined from the results of laboratory triaxial compression tests
using a stiff and servocontrolled testing system. Each of the complete stressstrain curves obtained from the triaxial tests is fitted by the stressstrain curve of two
dimensional bonded particles model simulated by PFC with appropriate microproperties determined so as to express the macroscopic behavior of the tested rock
specimen. Numbers of tension and/or shear cracks generated in the simulation are examined with the corresponding AE parameters such as AE counts, predominant
AE frequency and mvalue measured in the tests. The relationship between rock fracturing behavior and features of AE parameters is clarified. The field measurements
of the EDZ developed around the actual large underground cavern in highly stressed rock and AE behaviors are conducted in accordance with the progress of the
excavations. As the results, the spatial distribution of EDZ is determined by a number of measurements such as the initiated crack observation and stress measurement.
Furthermore, the prediction of the behavior of EDZ by PFC simulation of the bonded particles model is carried out. The results of the simulation shows good
agreement with the actual measurement results, which provides high performance for the behavior analysis of underground excavation together with a consistent
interpretation of the behaviors of AE parameters.
REFERENCES
Cundall, P.A. & Strack, O.D.L. 1979. A Discrete Numerical Model for Granular Assemblies, Géotechnique 29:47–65.
Ishimoto, M. & Iida, K. 1939. Observations of earthquakes registered with the microseismograph constructed recently, Bull. Earthquake Res. Inst., Tokyo Univ., 17:443.
Itasca Consulting Group, Inc. 2002. PFC2D—Particle Flow Code in 2 Dimensions, Ver. 3.0. Minneapolis: Itasca.
Mori, T., Aoki, K., Morioka, H., Iwano, K., Tanaka, M. & Kanagawa, T. 2003. Application of borehole seismic and AE monitoring technique in the rock cavern, 10th
ISRM Congress, pp. 845–848.
Page 84
Page 85
Numerical simulation of the mechanical behavior of discontinuous rock masses
E.S.Park
Korea Institute of Geoscience & Mineral Resources, Daejeon, Korea
C.Derek Martin
Geotechnical Engineering, University of Alberta, Edmonton, Alberta, Canada
R.Christiansson
Swedish Nuclear Fuel and Waste Management Co., Stockholm, Sweden
5809 679 3
ABSTRACT: Although the evaluation of the mechanical properties and behavior of discontinuous rock masses is very important for the design of
underground openings, it has always been considered the most difficult problem. One of the difficulties in describing the rock mass behavior is
assigning the appropriate constitutive model. This limitation may be overcome with the progress in discrete element software such as PFC, which does
not need the user to prescribe a constitutive model for rock mass. In this paper a 30 m×30 m×30 m jointed rock mass from the 400m level of SKB’s
Äspö Hard Rock Laboratory was analyzed. A discrete fracture network was developed from the joint geometry obtained from core logging and
tunnel wall mapping. Initially the rock mass PFC models were generated with the 2D trace sections extracted from the 3D DFN model. PFC
simulations were carried out, starting with the intact rock and systematically adding the joints and the stressstrain response of the rock mass including
the post peak region was recorded for each case. In the PFC model the behavior of the rock mass changed from brittle to perfectly plastic with
increase in the number of joints. As expected, the presence of joints had a pronounced effect on post peak response of the rock mass. More
importantly the postpeak response of the PFC model could be determined without a user specified constitutive model, while the UDEC results were
determined by the constitutive model applied to the joints.
1 INTRODUCTION
Although the evaluation of the mechanical properties and behavior of discontinuous rock masses is very important for the design of underground openings, it has always
been considered the most difficult problem. The reason is that it is often impossible to carry out largescale in situ tests and, although widely used, the correlations
between strength parameters and quality indexes (for instance GSI or Q index) are still affected by considerable uncertainties (Ribacchi 2000).
Experimental investigations have shown that the presence of joints reduces the strength of the rock mass relative to the intact rock material. The strength of this
reduction is governed by such factors as j oint geometry configuration, and intact and joint material properties (Kulatilake, Malama et al. 2001).
At present, the rock mass strength and deformation characteristics are evaluated using empirical methods such as the rock mass classification systems of
Bieniawski’s RMR (1976) and Barton’s QValues (1974). More recently Hoek et al. (1995) introduced the Geological Strength Index (GSI) to provide a system for
estimating the rock mass strength for different conditions.
Deformation properties and rock mass strength are not only dependent on the intact rock, but also on the fracture network (number and orientation of fracture sets,
intensity, mineralization, and so on) and the presence of deformation zones. Therefore, characterization of both the intact rock and of the fractures is required to define
the mechanical behavior of the rock mass.
The deformation modulus of a jointed rock mass may be estimated using the geometric data of the rock joints as well as the elastic parameters of the intact rocks
and the rock j oints if available. Analytical studies have been made on this subject by a number of researchers, for example Fossum (1985), Amadei (1993) and Huang
et al. (1995). Amadei (1993) presented a general solution to the deformability of regularly jointed rock masses. His work shows that the deformation modulus is
anisotropic in jointed rock
Page 86
masses. The structure and the stiffness of the rock joints predominately determine the magnitude of the deformation modulus of the rock mass (Li 2001).
Recent developments in numerical modeling that allow study of the overall response of a synthetic material containing discrete heterogeneities and discontinuities both
at the micro (particle) scale and at the larger scale of jointed rock masses can greatly aid the interpretation and application of laboratory test results on these materials
(Potyondy & Fairhurst 1999).
In this paper, the methodology developed in Sweden for the Rock Mechanics Descriptive Model is reviewed (Andersson et al. 2002). The results from the UDEC
rock mechanical descriptive model are compared to the results using the discrete element software PFC2D (Itasca 1999).
2 THE TEST CASE AREA
2.1 Presentation of the test case area
The Äspö Hard Rock Laboratory (HRL) is located on the Swedish east coast, about 350 km south of Stockholm. Site investigations started in 1986, and Construction
of the Äspö Hard Rock Laboratory (HRL) began in the early 1990’s. The HRL has been the site for many insitu experiments since its completion in 1995. The
deepest level of HRL is 450 m below ground surface.
A test case for Rock Mechanics Descriptive Model was established at SKB(Swedish Nuclear Fuel and Waste Management Co.)’s Äspö Hard Rock Laboratory at
a depth of approximately 420 m. The location of the test site and the location of the boreholes used in the analysis are shown in Figure 1. The test site was chosen as it
provided by boreholes and tunnel information so that a comparison could be made between both data sets.
2.2 Fractured rock mass model
The rock mass at Äspö HRL in the vicinity of the test site is typical of the granitic rock masses found in the Scandinavian Shield. It contains several joint sets that vary in
both trace length and fracture frequency.
A discrete fracture network (DFN) model was developed for the Äspö HRL in the vicinity of the test site from both borehole and tunnel fracture information. The
threedimensional discrete fracture network generated by FracMan®, which is the software used to generate fractures in three dimensions within a given rock volume,
must be transferred to twodimensional fracture trace sections to provide the ground data for simulations of fractures in the PFC model.
To assess the rock mechanical considerations for a deep repository, the tunnel scale with resolution
Figure 1. Location of the test case at SKB’s Äspö Hard Rock Laboratory in southern Sweden (from Staub 2002).
Figure 2. Fracture traces for a 2D section from a DFN model with fractures larger than 2 m, data from (Staub 2002).
of 30×30 m in the numerical model is most likely sufficient to cover rock mechanical aspects in “the repository scale” (Staub et al. 2002). So the size of the fractured
rock mass was set as 30 m×30 m. This 900 m2 contained a sufficient representative number of both large scale and small scale fractures. In addition, many tunnels in
the Swedish KBS3 disposal concept are 5m diameter or larger, and hence the 30 m×30 m rock mass is representative of the tunnel scale.
The fracture traces for a twodimensional section from the discrete fracture network model are shown in Figure 2.
3 PFC MODELING
3.1 Background
One of the limitations of conventional numerical models, such as UDEC, that rely on a user specified
Page 87
constitutive model is that the results are prescribed by the constitutive model assigned to both the intact rock and the fractures. An ideal numerical model is one where
the behavior of the rock mass is not prescribed a priori by the user.
The discrete element or discontinuous technique avoids the issue described above by representing the rock mass as an assemblage of particles similar to that
described by Cundall and Strack (1979). This assemblage must be bonded together and therefore the bonding at the micro level must be described. When particles
come into contact, forces F are generated as a function of the relative displacements u at the contacts. According to Cundall (2001) simple linear relations are often
adequate to describe the Fu contact behavior. By assigning a tensile bond strength and frictional bond strength, the micro strength properties can be expressed in
terms familiar to geotechnical engineers. However, the challenge is finding the appropriate microscale parameters to represent the macroscale response.
The particle flow code PFC2D, also available from Itasca Consulting Group Inc. was used to model the 30 m×30 m rock mass described previously. The PFC2D
model simulates the mechanical behavior of a collection of distinct, nonuniformly sized circular particles that displace independently and interact only at contacts. The
movement and interaction of the particles is modeled using the discrete element method, as described by Cundall and Strack (1979). Also PFC models are categorized
as direct, damagetype numerical models in which the deformation is not a function of prescribed relationships between stresses and strains, but of changing
microstructure (Wanne 2002).
One of the earliest applications of the discrete element method to rock was the modeling of brittle failure observed around AECL’s Mineby test tunnel (Martin et al.
1996). That effort, using PFC, began in 1995 and resulted in many refinements to the PFC code that is commercially available today. In this modeling effort the basic
“PFC Model for Rock” is used, which provides a numerical laboratory for simulating the full dynamic response of particlebased systems using the discreteelement
method.
The PFC model does not require assigning a priori constitutive models, instead the contact parameters are assigned.
In the present PFC modeling the parallel bond model was considered. The required microproperties are stiffness and strength parameters for the particles and the
bonds. The difficulty with PFC is in determining the micromechanics contact parameters to represent the rock mass response. Fortunately, Potyondy and Cundall
(2001) have investigated the relationship between micro and macro properties for Lac du Bonnet granite and their findings were used to establish the micro parameters
for Äspö diorite, the dominant rock
Table 1. Microproperties used to represent intact rock.
Items Values
Rmin (mm) 100.0
Rmin/Rmax 1.66
2630
Particle density, ρ (kg/m3)
Particle contact modulus, Ec (GPa) 62
Particle normal/shear stiffness, Kn/Ks 2.5
Particle friction coefficient, μ 0.50
1.0
Particle bond radius multiplier,
Parallel bond normal/shear stiffness, Kn /Ks 2.5
Parallel bond modulus, Ec (GPa) 62
Parallel bond normal strength, (MPa) 157±36
Parallel bond shear strength, (MPa) 157±36
type in the Rock Mechanics Descriptive model test site.
3.2 PFC model for discontinuous rock masses
First of all some numerical calibration tests were necessary to simulate the mechanical properties of intact rock. Calibration involves finding the micromechanical
parameters of the assembly from the macroscopic properties such as Young’s modulus and uniaxial compressive strength.
To perform this calibration, the micro mechanical parameters except the minimum radius of particles were estimated using the work of Itasca (1999).
The bestfit PFC2D material is an unclustered, parallel bonded material for which both tensile and shear micro failures are allowed. The shortterm micro properties
that were assigned are shown in Table 1.
The most basic mechanical properties of rock are those that are determined in the laboratory by standard testing methods, such as compressive strength, tensile
strength, Young’s modulus, Poisson’s ratio, friction angle, etc. As the rock mass is comprised of intact rock and discontinuities the accurate capture of intact rock
properties becomes critical and the first priority is creating a rock mass that truly represents reality (Wang, Tannant et al. 2003).
The model size of 30 m×30 m was used in PFC computations to consider the mechanical behavior of intact rock and fractures and fracture geometry. The PFC
model for discontinuous rock masses consists of approximately 13,000 particles.
The PFC model is built in the following steps. First, the intact rock is modeled as a single rock block of the model size. Then the fracture traces are integrated in the
model. The fracture model in PFC2D was generated from 2D trace sections extracted from the 3D DFN model developed for Äspö ZEDEX tunnel—see Figure 3(a),
(b) and (c).
In this PFC2D model, the fractures are modeled by generating bands of particles which consist of a
Page 88

Figure 3. The resulting rock mass model with fracture traces in PFC2D.
number of finite circular balls within the matrix. A joint set is generated by assigning the joint set ID number to all contacts between particles that lie on opposite sides of
each joint in the jointset.
After installation of fractures in PFC model, the microproperties of the joint set particles and bonds are modified. The given factors are applied to reduce the
particle and bond stiffnesses and strengths, as it is assumed that the joint is softer and weaker than the matrix.
3.3 PFC computations
The mechanical testing is simulated by a vertical loading applied on the top of the model by means of a constant velocity displacement obtained by a constant boundary
velocity applied during a specified number of computational cycles. And the lateral wall is maintained by a constant confining stress during each test.
The plate velocity is set to 0.05 m/s enough for avoiding the effect of loading rate (dynamic effect). The compression test simulation is performed until failure occurs.
Both loading and the test are continued until the postpeak stress is 80% of the peak strength.
The PFC simulations were carried out by starting with the intact rock and systematically adding the joint sets. For each test simulation, 5 repetitions of compression
tests were performed with same parameter and different packing arrangements. Also the stressstrain response of the rock mass including the post peak region was
recorded.
These results were averaged to obtain the mean and their standard deviation values for Young’s modulus, Poisson’s ratio, peak strength and crack initiation stress.
The crack initiation stress corresponds to stress at which 1% of the total number of cracks existing at peak load has been formed.
The effect of joints on the rock mass properties was studied by carrying simulations with different numbers of joint sets, which was varied from 0 to 3 sets. The
values of Young’s modulus, Poisson’s ratio, peak strength and crack initiation stress determined from each simulation are presented with respect to the numbers of joint
set in Figure 4(a), (b), (c) and (d) and Table 2.
As the number of joint sets increased, the values of Young’s modulus, peak strength and crack initiation stress decreased to about 50% of those values of rock mass
without joints, while that for Poisson’s ratio increased a little.
Especially in the PFC model the behavior of the rock mass changed from brittle to perfectly plastic and the residual strength of rock mass increased a little with
increase in the number of joints—see Figure 5. As expected the presence of joints has a pronounced effect on mechanical behavior and properties of rock mass.
3.4 Damage patterns in the PFC model
The damage plot provides information about the intensity, location and orientation of the microcracks and the possible coalescence into macrofractures that have
formed throughout the model, and corresponds with the postpeak state for which the axial stress has fallen to 80% of the peak strength. The crack distribution
Page 89
Figure 4. Simulated all values for Young’s modulus, Poisson’s ratio, peak strength and crack initiation stress with respect to number of
joint sets.
Table 2. Mean values for Young’s modulus, Poisson’s ratio, peak strength and crack initiation stress for the different fracture models confined at 22.4
MPa.
Number of joint sets Poisson’s ratio Young’s modulus (GPa) Peak strength (MPa) Crack initiation stress (MPa)

0 0.228 76.5 318.5 204.6
1 0.229 60.7 230.7 119.9
2 0.237 47.8 183.2 103.8
3 0.236 43.4 168.1 102.6
Figure 5. Simulated stressstrain curves of rock mass with regard to numbers of joint set (confined at 22.4 MPa).
in the plots (tension or shearinduced bond failure) is depicted as bicolored lines lying between the two previously bonded particles.
The damage patterns in the fracture model of PFC2D for discontinuous rock masses and the calculated stressstrain curve for a fracture model are shown in Figure 6
(a), (b), (c) and (d). As shown in Figure, the amount of damage that developed in PFC model increased and the damage patterns are changed by the presence of joint
sets. Also, the tensioninduced bond failure is predominant over shearinduced bond failure.
4 CONCLUSIONS
From the numerical tests of fractured rock mass model, it can be concluded that as the number of joint sets increased, the values of mechanical properties of rock
masses were decreased to about 50% of those values of rock mass without joints.
In the PFC model, the behavior of the rock mass changed from brittle to perfectly plastic with increase in the number of joints. As expected, the presence of
Page 90
Figure 6. Damage patterns occurred in PFC model confined at 22.4 MPa.
joints had a pronounced effect on post peak response of the rock mass.
More importantly the postpeak response of the PFC model could be determined without a user specified constitutive model, while the UDEC results were
determined by the constitutive model applied to the fractures.
The amount of damage that developed in PFC model was increased and the damage patterns were changed by the presence of joint sets. Also, the tensioninduced
bond failure that occurred in PFC model is predominant over shearinduced bond failure and the failures were mainly occurred along the fractures in PFC model.
Therefore, the PFC model without specified constitutive laws could be used further to simulate the mechanical behavior (including postpeak region) and properties
of discontinuous rock masses which is not yet possible by existing continuum theories.
Further studies are needed to continuously improve the particle model to better describe the reality, e.g. by developing various contact models and upscale the
modeling dimensions to simulate experiments at larger scale.
REFERENCES
Andersson, J., Christiansson, R. & Hudson, J.A. 2002. Strategy for Development of a Descriptive Rock Mechanics Model: Site Investigation. SKB Technical Report
TR02–01, Stockholm, Sweden.
Barton, N.R., Lien, R. & Lunde, J. 1974. Engineering classification of rock masses for the design of tunnel support. Rock Mech. 6(4):189–239.
Bieniawski, Z.T. 1976. Rock mass classification in rock engineering. In Exploration for rock engineering, Proceedings of the Symposium, eds. Z.T.Bieniawski, 97–
106. Cape Town: Balkema.
Cundall, P.A. & Strack, O.D.L. 1979. A discrete numerical model for granular assembles. Geotechnique. 29:47–65.
Cundall, P.A. 2001. A discontinuous future for numerical modelling in geomechanics? Proceedings of the Institute of Civil Engineers, Geotechnical Engineering,
January Issue 1, 41–47.
Hoek, E., Kaiser, P.K. & Bawden, W.F. 1995. Support of underground excavations in hard rock. Rotterdam: Balkema.
Itasca Consulting Group Inc. 1999. PFC2D—Particle Flow Code in 2 Dimensions, Minneapolis: Itasca.
Itasca Consulting Group Inc. 2000. UDEC—Universal Distinct Element Code in 2 Dimensions., Minneapolis: Itasca.
Li, C. 2001. Technical Note: A Method for Graphically Presenting the Deformation Modulus of Jointed Rock Masses. Rock Mechanics and Rock Engineering 34(1):
67–75.
Martin, C.D. & Read, R.S. 1996. AECL’s Mineby Experiment: A test tunnel in brittle rock. Proceedings of the 2nd N ARMS Symposium, Montreal. Edited by
M.Aubertin, F.Hassani & H.Mitri. Vol 1 pp. 13–24. A.A.Balkema, Rotterdam.
Page 91
Potyondy, D.O. & Cundall, P.A. 2001. The PFC Model for Rock: Predicting rockmass damage at the underground research laboratory, Itasca Consulting Group,
Inc. Report to Atomic Energy of Canada Limited (AECL), Report No. 06819REP01200–10061R00.
Potyondy, D.O. & Fairhurst, C.E. 1999. The Value of Numerical Modeling in Understanding the Complete Load/Deformation Behavior of CohesiveFrictional Materials.
Nondestructive and Automated Testing for Soil and Rock Properties, ASTM STP 1350, W.A.Marr and C.E.Fairhurst, Eds., American Society for Testing and
Materials: 290–299.
Ribacchi, R. 2000. Mechanical Tests on Pervasively Jointed Rock Material: Insight into Rock Mass Behaviour. Rock Mechanics and Rock Engineering 33(4):243–266.
Staub, I., Fredriksson, A. & Outters, N. 2002. Strategy for a Rock Mechanics Site Descriptive Model: Development and testing of the theoretical approach. SKB
Report R02–02, Stockholm, Sweden.
Wang, C., Tannant, D.D et al. 2003. Technical Note: Numerical analysis of the stability of heavily jointed rock slopes using PFC2D. International Journal of Rock
Mechanics & Mining Sciences 40:415–424.
Wanne, T. 2002. Rock strength and deformation dependence on schistosity: Simulation of rock with PFC3D. Itasca Consulting Group, Inc. Report to POSIVA Oy,
POSIVA 2002–05.
Page 92
Page 93
Industrial
Page 94
Page 95
Introduction of research and development on electromechanics of electromagnetic
particles for imaging technology
H.Kawamoto
Dept. of Mechanical Engineering, Waseda University, Tokyo, Japan
5809 679 3
ABSTRACT: The following basic researches are being carried out on electromechanics of particles for the printing technology: (1) Experimental,
numerical, and theoretical investigations have been conducted on statics and dynamics of magnetic bead chains in the magnetic field. Electrostatic pull
off force for the magnetic bead chain was then investigated to clarify the mechanism of beadcarryout phenomenon in the twocomponent brush
development system in electrophotography. (2) The influence of charge of toner particles on the formation of the magnetic chain was investigated to
apply for the magnetic singlecomponent development system in electrophotography. Numerical results were useful to clarify some image defects and
they were utilized for the improvement of the system. (3) Another investigation was carried out on transport of particles and particle size classification
in electrostatic traveling field. Numerical calculation was also conducted by the hard sphere model of the DEM.
1 INTRODUCTION
Because electromechanics of particles is a basis of electrophotography (Williams 1993, Schein 1996), we are conducting a fundamental study on electromechanical
transport of particles from three different aspects.
One is statics and dynamics of magnetic beads in the electromagnetic field. They are expected to be utilized for a magnetic brush development system in highspeed
and high image quality color laser printers. A typical schematic of this system is shown in Figure 1. Magnetic carrier beads with electrostatically attached toner particles
are introduced into the vicinity of a rotator sleeve with a stationary magnetic roller inside it. Diameter of carrier particle is several tens micrometer and that of toner
particle is about 10 micrometer.
Figure 1. Dualcomponent (magnetic brush) development system in electrophotography.
Carrier beads form chains as shown in Figure 1 on a sleeve by virtue of the magnetic field. Tips of chains touch the photoreceptor surface at the development area and
toner particles on chains move to electrostatic latent images on a photoreceptor to form real images. Carrier chains play an important role in this development step. In
order to realize high quality imaging, it is necessary to clarify the relationship between kinetic characteristics of formed chains and design parameters, such as magnetic
flux density and properties of carriers. In this study, experimental, numerical, and theoretical investigations have been carried out on statics and dynamics of the chain.
The second is dynamics of toner particles in the electromagnetic field. It is expected to be utilized for a magnetic singlecomponent development system in lowspeed
blackandwhite laser printers. Figure 2 shows a typical schematic of this system (Kawamoto & ElHage 1998). The photoreceptor drum and the sleeve are both
turning in opposite directions as the dualcomponent system. Magnetic toner particles are attracted to the stationary magnetic roller, and by the magnetic and friction
forces, they are transported around by the rotating sleeve. The thickness of the toner layer is controlled by the doctor blade. The doctoring process determines the
amount of toner on the sleeve in the development zone. As the sleeve rotates, the toner particles are charged by the blade, a process called static electrification. Once
in front of the photoreceptor, the toner particles jump towards the photoreceptor drum to form real images or return to
Page 96
Figure 2. Magnetic singlecomponent development system in electrophotography.
the sleeve near the development zone depending on the direction of force applied to the toner particles. It is also critically important to clarify toner dynamics in this
system to realize high image quality.
The third is a technique to transport dielectric particles utilizing traveling electrostatic field. Although a number of experimental and theoretical studies have been
reported on this technique (Masuda et al. 1987, Melcher et al. 1989a, b, Schmidlin 1991, Schmidlin 1995, Moesner & Higuchi 1999, Kober 2003), they are still
qualitative. We are conducting a basic research on this technology to utilize it for the transport of toner and/or carrier particles in laser printers.
2 NUMERICAL SIMULATION ON DYNAMICS OF PARTICLES IN ELECTROMAGNETIC FIELD
Numerical simulation was performed based on the Distinct Element Method (DEM) (Cundall & Strack 1979). In the calculation the following momentum equations are
solved for each particle.
(1)
where mj, uj, Ij, φj, Fj, Mj are mass, displacement vector, moment of inertia, rotation angle, applied force vector, and applied moment to a particle j, respectively. The
force and moment contain mechanical, magnetic, and electrical components.
The mechanical interaction force in normal direction at the contact point was estimated from Hertzian contact theory. The interaction force in tangential direction was
assumed proportional to the normal force with a proportional constant 0.25. Young’s modulus of the particles was assumed to be 10 GPa and that of the wall and the
floor was assumed to be 100 GPa. Poisson’s ratio and the friction coefficient were assumed to be 0.3 and 0.2, respectively. Other mechanical forces are air drag,
gravitational, centrifugal, and Van del Waals forces.
The magnetic force Fmj and rotational moment Mmj with the magnetic moment pj are given by the following expressions under the assumption that each particle
behaves as a magnetic dipole placed at the center of the magnetized particle.
(2)

The magnetic moment pj and magnetic flux density Bj at the position of the jth particle are
(3)
(4)
where N is the number of the particles, μ0 is the permeability of free space, μ is the relative permeability of particles, and aj is the diameter of the jth particle (Paranjpe
& Elrod 1986). The first term in the right hand side of Equation 3 is the applied magnetic field by the magnet and the second term is the field at the jth particle due to
dipoles of the other particles. Bkj is given by
(5)
where rkj is the position vector from the kth to the jth particle. The magnetic force is determined by solving Equations 3–5 simultaneously and by substituting the
calculated results, pj, and Bj, to Equation 2.
The electrostatic force consists of Coulomb force qiE and polarization force Fdipole. The electrostatic field E consists of the potential gradient due to the applied
voltage E0 and field created by other charged particles Eq .
(6)
where is potential applied by the power supply, ε0 is permittivity of free space, and q is charge of particle. The force applied to the polarized particle in the gradient
field is (Jones 1995)
(7)
where ε is a relative permittivity of the particle.
Page 97

Figure 3. Time variation of chain profile. (Averaged diameter 88 μm, surface loading 0.64 kg/m2, coil current 3 A).
3 CHAIN FORMATION IN MAGNETIC FIELD
Experimental and numerical investigations have been carried out on the formation and statics of magnetic bead chains in a magnetic field (Nakayama et al. 2002,
Nakayama 2003). Chains formed on a solenoid coil were observed by a digital microscope. Calculated chain forming process just after the magnetic field is applied is
shown in Figure 3 compared with experimental observations. In the chain forming process, deposited particles disperse upward first due to the repulsive force between
dipoles, then they form chain clusters to become magnetically stable state [0–20 ms]. The particles and clusters fall on the floor and grow up to treelike chains [20–40
ms]. Finally, stable cornshaped chains are formed by 40 ms. It is clearly recognized that the calculated chain forming process and the profiles are very similar to the
experimental observations.
Time variation of potential energy in the chain forming process of Figure 3 is plotted in Figure 4. Although the gravitational energy increases with the dispersion of
particles, magnetic and total potential energy decreases with time. The results support the hypothesis that the stable chain configuration is determined to minimize its
total potential energy that consists of gravitational and magnetic energy.
Chain profiles are deduced from calculated chain profiles and results were compared to experimental ones. Measured chain length increases with the increase in the
coil current, i.e. magnetic field, and the surface loading of particles as shown in Figure 5. It is shown that the calculated results give fairly good agreement with the
measured. Figure 6 shows an
Figure 4. Time variation of potential energy.
Figure 5. Comparison of calculated chain length with experimental results. (Averaged diameter 88 μm. Coil current 1A corresponds to
0.00616 T at the end of the coil.)
Figure 6. Formation of magnetic bead chains in magnetic field. (Left, experimental observation; Right, numerical result.)
example of experimental observation and numerical result at steady state.
An additional experiment was performed on a formation of magnetic bead chains in an extremely high magnetic field created by superconducting coil. Both the chain
length and interval increased with the increase in the magnetic field up to 0.1 T, while they decreased in the higher field. These results were also qualitatively confirmed
by the DEM calculation.
Page 98
4 DYNAMICS OF CHAIN IN MAGNETIC FIELD
It is also important to clarify resonant frequency and stiffness of the chain in the magnetic field, because hard chains might cause stripe image degradations, on the other
hand, “beadcarryout” might take place if chains are too soft (Williams 1993). In this study the following four independent methods were employed to evaluate
resonant frequency and stiffness of a magnetic bead chain in the magnetic field (Nakayama et al. 2003).
1. SineWave Excitation. Chains formed on a solenoid coil were vibrated by the sinusoidal wave to deduce resonant frequency.
2. Impact Test. An impact test method was employed to determine the resonant frequency. Free vibration response of the chains measured by a laser displacement
meter was introduced to an FFT analyzer then the resonant frequency was specified.
3. Static Inclination Method. Chains are not exactly directed parallel to the magnetic flux if the flux inclines to the gravity. The static stiffness of the chain was estimated
from chain inclinations in inclined magnetic fields. Because chains collapse when the inclination of the magnetic field exceeds a critical value, a critical angle was also
measured and calculated to evaluate the strength of chain.
Figure 7. Calculated impact response of magnetic chain.
Figure 8. Resonant frequency of magnetic chain.
4 Static Deformation. Chain was laterally pushed by a fine cantilever and lateral deformation was measured by a microscope to determine static stiffness of the chain.
The following were deduced from the investigations.
1. The resonant frequency of the chain was 20–60 Hz and it was almost irrelevant with the magnetic flux density, the bead diameter and surface loading of the beads.
2. The static stiffness of the chain was 10−4− 10−2 N/m. It was large in the high magnetic field and for large diameter beads. These characteristics were confirmed by a
theoretical consideration based on the assumption of potential energy minimization and the DEM calculation (refer to Fig. 7). Experimental and numerical results on
the resonant frequency are summarized in Figure 8.
5 BEADCARRYOUT PHENOMENON
It is well known as the beadcarryout (BCO) phenomenon that the electrostatic force acts on chains and bead(s) are separated from chains and moved to
photoreceptor surface (Williams 1993). The attached carrier bead(s) on the photoreceptor surface cause significant image degradations. Because clarification of the
BCO mechanism and requirements to prevent BCO is important to realize high image quality of the dualcomponent development system, experimental and numerical
investigations have been carried out on this phenomenon (Nakayama et al. 2003). In the experiment, a critical pulloff field intensity to separate bead(s) from chains
formed on a solenoid coil were measured as shown in Figure 9 and effects of bead diameter and coil current were evaluated. Amount of pulledoff carriers was also
measured in another experiment with a magnet roller used in a commercially available laser printer. In addition, numerical simulation of chain formation was performed
to estimate magnetic bonding force between beads in the chain with the DEM calculation. An electrostatic pulloff force was estimated by the electric field
Figure 9. Experimental setup to measure critical BCO conditions.
Page 99
Figure 10. Comparison of electrostatic and magnetic forces that BCO takes place.
calculation with the Finite Element Method including charged chains. These numerical results were compared to experimental results to evaluate adequacy of the
method and to study pulloff characteristics. In conclusion, it was clarified that the BCO phenomenon can be predicted qualitatively from the comparison of calculated
magnetic force and electrostatic force as shown in Figure 10.
6 TONER DYNAMICS IN MAGNETIC SINGLECOMPONENT DEVELOPMENT SYSTEM
A magnetic singlecomponent development process in electrophotography has interesting characteristics on a magnetic and electrostatic interaction of toner particles.
Magnetic force restrains toner adhesion to nonimage area and this feature enhances the development. However, long chains made by magnetic interaction force
between magnetic toner particles deteriorate tonerimage quality on the photoreceptor as shown in Figure 11. It is known from a past experimental study that the
magnetic chain length is influenced not only by the magnetic field but also by the electrostatic charge of toner particles. This study confirmed it by a numerical analysis
(Nakano et al. 2003). Simulation of toner dynamics was carried out with the Finite Difference Method for the electric field calculation and the DEM for the calculation
of charged magnetic toner motion in the magnetic and electrostatic field. The following were deduced from the investigation.
1. The chain length of toner particles was shortened by the toner charge.
2. In the case that particles does not include wrongsign toner, charge distribution does not change
Figure 11. An example of developed toner image of 6 pt Japanese KANJI font on the photoreceptor. Defects in white circles were
caused by magnetic chains.
Figure 12. Simulated development process on latent image.
the chain formation but wrongsign toner particles influence the chain length and number.
3. The magnetic chain was reconstructed under the application of AC biased potential at the development gap.
4. Toner particles with wrongsign toner caused a defect of chain images.
Figure 12 shows snapshots of developed image of 600 dpi 3line images when the development gap was smallest. Chains without wrongsign toner particles were
shortened by the reconstructed effect; on the other hand, toner particles with wrongsign toner caused the chain image as shown in Figure 11.
A microscopic measurement of the image on the photoconductor is also carried out to confirm the numerical model. Figure 13 shows an example of the measured
line image.
7 TRAVELING WAVE TRANSPORT OF PARTICLES AND PARTICLE SIZE CLASSIFICATION
An experimental research was carried out on transport of particles and particle size classification in electrostatic traveling field. Particle conveyers consisting of
Page 100
parallel electrodes were constructed and fourphase traveling electrostatic wave was applied to the electrodes as shown in Figure 14 to transport particles on the
conveyer (Kawamoto and Hasegawa in press). The following were clarified by the experiments and a hard
Figure 13. Measured toner profiles of line image on photoconductor.
sphere DEM model in which friction and repulsion between particleparticle and particleconveyer were included to replace initial conditions after mechanical contacts.
1. Coulomb force is the most predominant to drive particles.
2. Particles were transported almost linearly with time. A transport rate was also linear with the applied voltage but a threshold existed due to adhesion force.
3. The direction of particle transport did not always coincide with that of the traveling wave but partially changed depending on the frequency of the traveling wave, the
particle diameter, and the electric field as shown in Figure 15. Although some
Figure 14. Electrostatic particle conveyer and power supply.
Figure 15. Measured and calculated rates of forward, backward, and nontransported particles with respect to frequency of traveling
wave. (ACM235:35 μm, ACM255:55 μm, ACM2107:107 μm).
Page 101
Figure 16. Distributions of particle diameter in forward and backward transported particles and corresponding photographs of particles.
particles overtook the traveling wave at the very low frequency, the motion of particles at low frequency was synchronized with the traveling wave and that at
medium frequency was opposite and slow to the wave. Particles were vibrated but not transported under high frequency field.
4. Particles were efficiently transported under conditions of high electrostatic field of rectangular wave.
5. Particles moved essentially through the electric flux line but collision between particleparticle and particleconveyer made trajectories complex.
6. Particles were classified with size under the voltage application of appropriate frequency. Figure 16 shows evidence on the particle classification with particles whose
diameter was distributed in wide range. Under the condition of the optimum frequency, 140 Hz in this case, a significant classification was realized.
8 CONCLUDING REMARKS
The Distinct Element Method was applied to simulate kinetics of electromagnetic particles in the electromagnetic field. It was demonstrated that the method was useful
for the improvement of the electrophotography process.
The author would like to express his thanks to all members of his laboratory who contribute to this work. This work is supported by The 21st Century COE
Program of Japan Society for Promotion of Science.
REFERENCES
Cundall, P.A. & Strack, O.D.L. 1979. A Discrete Numerical Model for Granular Assemblies. Géotechnique. 29(1): 47–65.
Jones, T. 1995. Electromechanics of Particles. Cambridge: Cambridge University Press.
Kawamoto, H. & ElHage, W. 1998. Dynamics on Magnetic Single Component Development System in Electrophotography. Advances in Information Storage Systems.
10: 237–248.
Kawamoto, H. & Naoto, H. in press. Traveling Wave Transport of Particles and Particle Size Classification. J. Imaging Sci. Technol.
Kober, R. 2003. Transport geladener Partikel in elektrischen Wanderfeldern. Dissertation. Achen Univ. Technology.
Masuda, S, Washizu, M. & Iwadare, M. 1987. Separation of Small Particles Suspended in Liquid by Nonuniform Traveling Field. IEEE Trans. Ind. Appl. IA23(3):474–
480.
Melcher, J.R, Warren, E.P. & Kotwal, R.H. 1989a. Theory for FinitePhase TravelingWave BoundaryGuided Transport of Triboelectric Particles. IEEE Trans. Ind.
Appl. 25(5): 949–955.
Melcher, J.R, Warren, E.P. & Kotwal, R.H. 1989b. TravelingWave Delivery of SingleComponent Developer. IEEE Trans. Ind. Appl. 25(5):956–961.
Moesner, F.M. & Higuchi, T. 1999. Electrostatic Devices for Particle Microhandling, IEEE Trans. Ind. Appl. 35(3): 530–536.
Nakano, M, Kawamoto, H. & Itoh, I. 2003. Influence of Toner Charge on Magnetic Chain Formation in Magnetic SingleComponent Development System of
Electrophotography. IS&T’s NIP19: International Conference on Digital Printing Technologies. New Orleans. 74–78.
Nakayama, N, Kawamoto, H. & Yamaguchi, M. 2002. Statics of Magnetic Bead Chain in Magnetic Field. J. Imaging Sci. Technol. 46(5):422–428.
Nakayama, N, Kawamoto, H. & Yamada, S. 2003. Resonance Frequency and Stiffness of Magnetic Bead Chain in Magnetic Field. J. Imaging Sci. Technol. 47(5):408–
417.
Nakayama, N. 2003. Study on Dynamics of Electromagnetic particles in Electromagnetic Field and Application to Imaging Technology. Dissertation. Waseda University.
Nakayama, N, Watanabe, Y, Watanabe, Y. & Kawamoto, H. 2003. Experimental and Numerical Study on the BeadCarryOut in TwoComponent Development Process
in Electrophotography. IS&T’s NIP19: International Conference on Digital Printing Technologies. New Orleans. 69–73.
Paranjpe, R.S. & Elrod, H.G. 1986. Stability of chains of permeable spherical beads in an applied magnetic field. J. Appl. Phys. 60(1):418–422.
Schein, L.B. 1996. Electrophotography and Development Physics (Revised Second Edition). Morgan Hill, CA: Laplacian Press.
Schmidlin, F.W. 1991. A New Nonlevitated Mode of Traveling Wave Toner Transport. IEEE Trans. Ind. Appl. 27(3): 480–487.
Schmidlin, F.W. 1995. Modes of traveling wave particles transport and their applications. J. Electrostat. 34: 225–244.
Williams, E.M. 1993. The Physics and Technology of Xerographic Processes. Florida: Krieger Publishing.
Page 102
Page 103
DEM simulation of bulk material flow in a feeding system of blast furnace
J.F.Ferellec, C.Mochizuki & A.Hirata
IshikawqjimaHarima Heavy Industry Co. Ltd., Yokohama, Japan
5809 679 3
ABSTRACT: This study concerns the flow of coke in the feeding system of a blast furnace. As an experimental study of this phase of the flow is
difficult to carry out, a simulation using the distinct element method has been employed. The contact law model consists of linear springs and dashpots
for normal and tangential displacements associated with a MohrCoulomb type slider. The parameters of the models are determined so as to represent
coke material but also to maintain reasonable calculation duration at the same time. The flow is analyzed in terms of flow pattern and distribution of
particles at different locations in the pipes and hoppers of the system. The results show that the bulk material does not flow uniformly in the different
sections of the pipes and flows along preferential walls of the hoppers. These asymmetrical flow conditions affect the flow in the chute of the blast
furnace itself.
1 INTRODUCTION
The feeding process of blast furnace is a delicate phase as it conditions optimization of ironmaking process. A perfect knowledge of the flow of bulk materials during
the feeding process is then necessary to guaranty this optimization. However conducting experiment on such facility is difficult considering its size. This obstacle can be
overcome by the use of numerical tools. The numerical model, which considers coke bulk material as discrete, is well adapted to this kind of challenge. The cases using
such a model are indeed numerous and show that discrete models are powerful tools. Among these models, the most popular is probably the distinct element model.
This model has been successfully utilized in various situations like storage in silos (Ferellec et al. 2001, Masson & Martinez 2000), bulk material handling, mixing
process (Cleary 2000, Cleary & Sawley 2002), fluidized bed (Yuu et al. 2000). Blast furnace has also been the subject of study using such a method for the flow
inside the furnace itself (Zhang et al. 2002). In the present paper, our attention is focused on the distribution system of a bellless blast furnace, which goes from the
conveyor belt to the furnace itself. This system consists in various devices like duct, hopper, bunker and chute. The resulting complexity of the system, which is
necessary to guarantee a smooth functioning of the blast furnace, implies complicated flow path of the bulk material. The aim of this study is precisely to analyze this
path using the distinct element method. The flow has been analyzed in two steps corresponding to two locations in the distribution system of the blast furnace. The
parameters introduced in the model, essentially stiffness and friction of coke, are deduced from repose angle and compression tests. Diverse variables have been
analyzed in simulations: flow pattern, velocity distribution, particle distribution, wall pressure, etc. The paper is organized in two main parts. In the first, we describe the
distinct element model used here and the conditions of the simulations. In the second part, we present the different results of simulation by locations from the top to the
furnace. Finally a conclusion closes this paper by recalling the main results.
2 CONDITIONS OF SIMULATIONS
2.1 DEM model
The model used here (PFC3D) is based on the distinct element method or DEM (Cundall & Strack 1979). In this model, the granular material is considered as a set of
spheres interacting with each other by means of contact laws. The contact law is elasticplastic and represented by linear springs and dashpots for normal and tangential
displacements, associated with a slider (Fig. 1). Springs are characterized by normal and tangential stiffness coefficients kn and kt, dashpots by normal and tangential
damping coefficients Cn and Ct. A Coulomb type friction coefficient μ limits the tangential component of the contact force. A similar elasticplastic behavior is adopted
for particlewall contact. Particles motion resulting from unbalanced force is determined by application of the law of dynamics of Newton in accordance with an explicit
time difference scheme.
Page 104
2.2 Geometry of the system
Two independent simulations, corresponding to two different locations in the feeding system of the blast furnace, have been performed. The first one considers the flow
from the conveyor belt into the temporary storage bunker (Figs 2 & 3). For confidentiality, dimensions are not provided. From the conveyor belt, the bulk material falls
into a rectangular funnel, and then into a succession of vertical pipes to a bent pipe that redirects
Figure 1. Contact model.
Figure 2. Geometry of part 1 of the feeding system (two side views).
the flow in a hopper above the top inlet of the temporary storage bunker. Inside the bunker, the flow of bulk material is deviated by a threeplate device, with
adjustable inclination. The angle between the conveyor belt line and the main plan of the blast furnace, that is the plan of the bunker, is 57.25 degrees. The particles,
which represent coke, are generated continuously on the conveyor belt, forming a heap with 30 degrees slope angles. Particles, with density of 1200 kg/m3, are fed into
the system assuming a flow rate equal to 0.8 m3/s and porosity of the bulk material equal to 0.559. Particle size distribution is divided into four classes (Table 1). The
values of the mechanical parameters of the model are summarized in Table 2. They have been deduced from experimental data on coke. In order to better render the
behavior of real angular pieces of coke, the free rolling of the particles of the model has been prevented. Friction coefficient between particles has been adjusted so as
to get a similar repose angle of the bulk material.
Figure 3. Geometry of part 1 of the feeding system (top view).
Table 1. Particle size distribution.
Diameter (mm) 25.0 37.5 62.5 75.0

% 1.0 52.3 33.0 13.7
Table 2. Mechanical parameters of the model.
Parameter Unit p/p* p/w*

kn N/m 1.7×10 6
3.3×106

kt N/m 1.7×106

3.3×106

Cn N.s/m 5.4×102

7.5×102

Ct N.s/m 5.4×102

7.5×102

μ – 0.2 0.2
*p/p=particle/particle contact, p/w=particle/wall contact.
Page 105
The stiffness of the particles has been deduced from compression tests on coke. However, in order to keep reasonable calculation time, this stiffness has been
reduced hundred times. Preliminary simulations with the same geometry but using lower flow rate, that is fewer particles, have been performed to observe the possible
difference induced by the reduction of stiffness value. As flow pattern and flow rate at different locations were similar using either real stiffness of reduced one, the
simulation with nominal flow rate has been performed using the reduced stiffness.
The second simulation concerns the flow of coke from the outlet of the temporary storage bunker to the furnace itself (Fig. 4). It consists first of a hopper and two
vertical pipes: one upper eccentric, guiding the flow from the bunker outlet to the hopper and another one lower centered, going from the hopper outlet to a chute
located at the top of the furnace. A low circular
Figure 4. Geometry of part 2 of the feeding system (two side views).
Figure 5. Geometry of part 2 of the feeding system (top view).
wall surrounds the outlet inside the hopper. At the top of this ringshape wall, two vertical protuberances rise. The chute, which consists roughly of half a cylinder,
rotates along vertical at 8 rpm. Particles are generated continuously above the upper pipe as if they were discharged from the bunker using a flow rate equal to 1 m3/s
and porosity equal to 0.559. Other parameters are similar to the first simulation parameters.
3 RESULTS
3.1 From conveyor belt to storage bunker
The following results concern the flow when it has reached a stable state i.e. the flow rate measured at different locations in the feeding system is constant. Figure 6
shows two side views of the flow from the conveyor belt to the bunker (not represented on the figure). The flow of particles does not seem to occupy the entire volume
of the pipes. When the flow reaches the bunker, it becomes asymmetrical. A nonnegligible part of the particles do not hit the plate located in the bunker. This will
probably lead to an asymmetrical filling of the bunker and subsequently an asymmetrical loading on the bunker structure. A more precise observation of the
Figure 6. Side views of the flow of particles from the conveyor belt to the storage bunker.
Page 106
Figure 7. Horizontal sections of the flow of particles at different locations in the first simulation of the feeding system of the blast furnace
(top part).
flow in the different pipes and hoppers provides clearly the reason of this asymmetrical flow. Figures 7 & 8 show horizontal sections of the flow of bulk material at
different positions in the feeding system. Top and bottom of the system have been dissociated as they are in two different vertical plans. At top (Fig. 7) we can observe
that the flow rebounds on the slope wall opposite to the conveyor belt and then its velocity acquires a nonnegligible horizontal component.
Subsequently the particles flow only along a fraction of the section of the following lower pipes (Fig. 8). Moreover, as top and bottom of the system are in two
different plans separated by 57.25 degrees, the trajectory of the flow does not follow naturally the curve of the bent pipe but adopts a complex shape. As a result, the
particles hit the lateral wall of the hopper and enter the bunker asymmetrically. An additional simulation with free rolling particles presented a quite similar flow pattern.
A previous simulation in which particles occupy all the section of the upper pipe has also been made and presented this time a symmetrical flow in the bunker. All these
results tend to confirm that the position of the conveyor belt in a plan different from all other components of the blast furnace generates an asymmetrical flow in the
feeding system.
A verification of the distribution of the particles of different sizes at different locations in this part of the feeding system has been realized for this simulation. No
obvious segregation phenomenon occurs at any location.
Figure 8. Sections of the flow of particles at different locations in the first simulation of the feeding system of the blast furnace (bottom
part).
3.2 From bunker outlet to furnace
As the huge number of particles required to get a stable flow in the hopper up to the furnace implies unreasonable calculation time, the simulation in this part has been
conducted far enough to get input flow conditions for the chute below. Figure 9 illustrates the flow of particles in the hopper located up to the furnace. The different
shades correspond to different generation times. In order to make the reading of particle velocity easier, only one vector in ten is represented on the picture. The figure
shows clearly that initially particles hit the bottom of the hopper and flow around the outlet avoiding falling into it because of the presence of the low ring and the
protuberance mentioned above. When the area around the outlet is filled, particles accumulate under the feeding pipe creating a relatively stable heap, which will
probably remain if feeding stops. The velocity of the particles under the surface of the heap is much lower and produces a relatively stagnant zone. Particles later slide
along the slope of this heap and flow directly into the outlet. Figure 10 shows the corresponding flow in the pipes that extend after the hopper outlet. Particles do not
enter the pipes vertically as they flow along the slope of the heap in the hopper. They hit the side of the pipes facing the heap and then fall vertically along this side.
Subsequently the flow does not occupy all the section of the pipes but rather the part opposite to the heap
Page 107
Figure 9. Flow pattern and velocity in hopper up to furnace.
side. The components of total momentum Px, Py and Pz divided by total mass m of the bulk material at hopper outlet and pipes outlet have been determined (Fig. 11).
After 4.5 s these components are stable. At the end of the pipes, Px/m and Py /m components are essentially equal to zero.
Particles flow vertically as confirmed on the velocity field in Figure 10. Taking into account these components and theoretical flow rate, that is the input flow rate of
hopper, flow conditions and, more particularly, the flow section surface at the end of the pipes can be determined. Then, in the simulation of distribution of material in
the furnace using the rotating chute, particles are introduced with the following conditions. First, the initial velocity of the particles corresponds to Pz/m, from the
previous simulation after 4.5 s at the pipes outlet. Second, particles are generated in a section located on one side of pipes outlet, which is
Figure 10. Flow pattern and velocity in pipes under hopper outlet.

Figure 11. Components of momentum Px, Py and Pz per unit of mass versus time.
deduced from the theoretical flow rate of 1 m3/s. These particular conditions are important as they correspond to an asymmetrical feeding of the following rotating
chute.
Page 108
Figure 12. Flow pattern in chute and position of sections at t=2 s. (a) represents top view of chute, (b) side view and (c) side view of
sections.
Figure 13. Velocity pattern in chute at t=2 s. (a) represents top view of chute and (b) side view.
Figure 12 shows a top view (a) and a side view (b) of the flow in the chute at t=2 s. The outlet of the upper pipes can be seen on the left. The asymmetrical feeding is
also discernible. Particles are generated in a zone that is half of the section of the pipes outlet. Rotation of the chute along the vertical direction is clockwise from the
top.
The effect of this rotation on the flow of bulk material is clearly observable: the material is positioned along the inner side of the chute that faces the rotation direction.
In order to make flow analysis easier, the flow of bulk material has been cut at different positions in the chute (Fig. 12c). Simultaneous observation of the velocity field
at t=2 s (Fig. 13) and evolution of the different sections (Fig. 14) allows decomposing the different steps of the flow in the chute. Initially (t=1 s), the particles hit the
chute on its right side because of asymmetrical feeding. Then as they are along a vertical wall, they tend to move towards the middle of the chute and finally reach the
opposite wall. However because of the rotation of the chute, they come back in the middle of the chute and move
Figure 14. Evolution of flow pattern sections in chute viewed from the end of the chute. Numbers on left correspond to sections in the
chute.
again progressively towards the first wall (the inner wall that faces the rotation direction) until they reach the end of the chute. At the end of the chute, particles are close
to escape from the top.
Figure 15 shows the evolution of average wall pressure in the chute for the corresponding flow. Average pressure is calculated as the ratio of the summation of the
normal components of the contact force on an elementary wall to the surface area of this wall.
Pressure is given along the circumference of the chute for the first half (sections 1 to 5) and second half (sections 5 to 10) of the chute. The profiles of pressure
confirm the results above. Initially in the first half, pressure is higher at the bottom. But because of rotation, pressure becomes higher along the inner right side of the
chute. Later pressure increases for all walls because the amount of particles increases. On the other hand, in the second half of chute, high pressure occurs essentially
on the right side of the chute, because material is located on this side, and increases progressively.
Page 109
Figure 15. Evolution of pressure (kPa) in first (left) and second (right) half of chute.
Figure 16. Flow of bulk material from top after the chute.
These results show that most of the time, initial impact of the particles occurs on the inner side of the chute generating, in combination with chute rotation, a complex
flow pattern.
However this case presents only the beginning of the flow in the chute. The flow may present different patterns depending on the relative position of the asymmetrical
feeding and the angular position of the chute.
After the chute, the bulk material falls into the furnace itself. The furnace and its walls have not been simulated in this study. The complete filling of the furnace would
require unachievable calculation time regarding the number of particles necessary to fill it. Then in the present simulation, particles fall freely after chute. However a wall
reference of the top furnace is represented by a circle on Figure 16. This figure shows the flow of particles from the top successively at t=1, 2, 3 and 4 s. The flow
presents a curvilinear trajectory and will probably hit the walls of the furnace. This will possibly affect the distribution of the bulk material in the furnace. Further
simulation would be necessary to clarify this delicate point of bulk material distribution in the furnace.
4 CONCLUSION
The simulation of the flow of coke in the feeding system of a blast furnace using the distinct element method has been conducted. The results show that during flow, the
bulk material does not occupy the entire sections of the conducts, and flows along preferential sides in the different parts of the blast furnace. The relative angular
position of the conveyor belt and the blast furnace is strongly involved in this phenomenon. As a consequence, bulk material flows asymmetrically in the temporary
storage bunker. This asymmetrical feeding of the bunker will probably lead to an asymmetrical filling of the bunker. This may also affect the discharge of the bunker and
the loading of the structure of the bunker.
The results of the simulation also show an accumulation of bulk material on one side of the hopper under the bunker. This leads to an asymmetrical flow in the pipes
located before the inlet of the furnace itself. The resultant asymmetrical feeding of the chute, combined with its rotation, generates a complex flow in the chute. Material
is not discharged along all the section of the chute but only along one of the vertical sides because of the rotation. After being discharged from the chute, the flow
adopts a curvilinear trajectory that should hit the furnace wall.
Finally, the distinct element method proves to be again an efficient tool for the study of bulk material flow under gravity in complex arrangements of pipes, hoppers
and storage devices even more when experiment is hardly realizable. This method has shown here results that other methods could not apprehend so easily.
However, the present distinct element method remains perfectible. It concerns mainly the contact model. A model that could take into account the effect of the shape
of the particle should be more efficient like, for example, a model including rolling
Page 110
resistance (Iwashita & Oda 1998, Oda & Iwashita 2000, Sakaguchi et al. 1993).
REFERENCES
Cleary, P.W. 2000. DEM simulation of industrial particle flows: case studies of dragline excavators, mixing in tumblers and centrifugal mills. Powder Technology 109: 83–
104.
Cleary, P.W. & Sawley, M. 2002. DEM modeling of industrial granular flows: #D case studies and the effect of particle shape on hopper discharge. Applied
Mathematical Modelling 26(2):89–111.
Cundall, P.A. & Strack, O.D.L. 1979. A discrete numerical model for granular assemblies. Geotechnique 29(1): 47–65.
Ferellec, J.F., Iwashita, K., Martinez, J. & Masson, S. 2001. Influence of particle rolling resistance on silo flow in DEM simulations. In Y.Kishino (ed.), Powders &
Grains 2001; Proc. Fourth Intern. Conf. on Micromechanics of Granular Media, Sendai, Japan, 21–25 May 2001. Rotterdam: Balkema.
Iwashita, K. & Oda, M. 1998. Rolling resistance at contacts in simulation of shear band development by DEM. Journal of Engineering Mechanics 124(3):285–292.
Masson, S. & Martinez, J. 2000. Effect of particle mechanical properties on silo flow and stresses from distinct element simulations. Powder Technology 109:164–178.
Oda, M. & Iwashita, K. 2000. Study on couple stress and shear band development in granular media based on numerical analysis. Int. J. Eng. Sc. 38:1713–1740.
Sakaguchi, H., Ozaki, E. & Igarashi, T. 1993. Plugging of the flow of granular materials during the discharge from a silo. Int. J. Modern Physics B(7):1949–1963.
Yuu, S. Umekage, T. & Johno, Y. 2000. Numerical simulation of air and particle motions in bubbling fluidized bed of small particles. Powder Technology 110:158–168.
Zhang, S.J., Yu, B., Austin, P. & Zulli, P. 2002. Numerical simulation of the solids flow in a blast furnace. Applied Mathematical Modelling 26:141–154.
Page 111
Using particle flow code to investigate energy dissipation in a rotary grinding
mill
R.S.Sarracino, A.T.McBride & M.S.Powell
Department of Chemical Engineering, University of Cape Town, South Africa
5809 679 3
ABSTRACT: An overview of the procedure used to extract and visualize an impact energy spectrum from a numerical simulation of tumbling mills,
performed using the Discrete Element software Particle Flow Code (PFC), is presented. A linear viscous damping contact model is developed with
the User Defined Contact Model in PFC to facilitate the extraction of the data pertinent to the construction of the impact spectra. Simple numerical
experiments are performed with PFC and an explicit embedded 4th–5th order RungeKutta time integration scheme to assess the validity of the linear
viscous damping model for comminution related applications. It is shown that the explicit time integration scheme used in PFC can be problematic
under certain conditions due to the stiff nature of the governing differential equations. Other contact models are assessed including one with nonstiff
governing equations and their relative merits discussed.
1 INTRODUCTION
The Discrete Element Method (DEM) has been widely utilized for simulating rotary grinding mills in the field of comminution with the ultimate aim of using DEM as a
design tool capable of predicting important milling parameters such as power draw, liner wear, impact energy spectrum and particle breakage. In order for DEM to
realize its potential as a design tool, accurate measures of the energy dissipated during contact events within the mill and of impact energy spectra need to be extracted
from the DEM simulations.
This paper presents an overview of the procedure implemented within the DEM simulation software Particle Flow Code in 3D (PFC3D) developed by Itasca
(2002) to record the energy dissipated during contact events. The widely implemented viscous damping contact model was developed in conjunction with Itasca using
the framework of the user defined contact model (UDCM). The UDCM allows the user to develop a contact model using the programming language C++. It also
allows the user direct access to the contact information that is calculated during the contact force resolution process by PFC3D. Significant simulation time reductions
result from using the UDCM to extract data necessary to produce an impact energy spectrum (a histogram plot of the number of contact events versus the magnitude of
the event) as the contact information does not need to be recalculated using FISH, the internal scripting language of PFC3D.
Due to the large number of contacts that occur within a mill, efficient numerical schemes were developed to record the data for subsequent postprocessing. The
impact energy spectra of collision events for each particle size within a distribution is calculated and can be related to experimental breakage data via experiments. The
duration of each contact event is also recorded for potential incorporation into more complex breakage models.
Three different techniques were employed to determine the energy spectrum of the mill, and a comparison of each technique is presented. This paper also takes a
preliminary look at several possible contact models. A combined experimental and numerical program is underway at the University of Cape Town to develop contact
models suitable for modeling the inelastic collisions between the steel balls, rocks and liners used in mills. A further combined program to develop and implement
breakage models is being planned. Extensions of the Hertz theory of impact to oblique impact have been developed (see, e.g., Maw et al. 1976, 1981). However, to
implement a full Hertzian model in DEM modeling of mills would be prohibitively expensive; consequently, cheaper models which reproduce the essential behavior must
be developed.
We perform a simple numerical experiment in a beginning assessment of the standard springanddashpot model, two other viscous damping models, and a particular
hysteretic model, the bilinear model (Sarracino et al. 2003).
Page 112
2 OVERVIEW OF THE DISCRETE ELEMENT METHOD
Contact between particles is modeled using the soft contact formulation, wherein individual particles are permitted to overlap one another at the point of contact. A
forcedisplacement law relates the overlap between the particles to a restoring force, as depicted schematically in Figure 1. A numerical spring, the normal contact
spring, is applied in the direction of the vector linking the centroids of the contacting objects to “push” the objects apart. A second numerical spring, the shear contact
spring, is applied in a plane whose normal is in the direction of the normal contact spring to simulate the frictional forces at the contact. The maximum frictional force is
limited by the MohrCoulomb law. The stiffness of the numerical springs is determined by a contactstiffness law. Energy dissipation mechanisms can be accounted for
by a combination of asymmetric stiffnesses and symmetric or asymmetric numerical dashpots.
One of the most widely used contactstiffness laws for simulating comminution devices is the linear viscous damping model for interaction between unbonded
particles (the wellknown springanddashpot model),
(1)
where U is the particle overlap, b is the damping constant and k is the contact stiffness. Including the effect of gravity one obtains the inhomogeneous equation
(2)
where g is the acceleration of gravity. Here it is assumed that gravity acts to increase overlap, as would be the case for a particle falling on a liner. However, due to the
weakness of the gravitational force relative to the spring force, one would expect the solution of Equation 2 to be only a minor perturbation to that of Equation 1. This
model will be discussed later in the paper.
The use of normal and tangential contact springs is common to the majority of forcedisplacement laws in DEM codes, with the particular values and mix of
stiffnesses, as well as values of the friction and damping coefficients, used to distinguish various materials.
The time domain is discretized into intervals whose duration is sufficiently small to ensure that the disturbance caused to the system by a contact event does not
propagate beyond the contacting objects within the
Figure 1. Schematic of the viscous damping contact model.
timestep. The net force and moment acting relative to the centroids of the individual particles due to contact forces and gravity are resolved and the acceleration of the
particles calculated using Newton’s second law. An explicit finite differencing scheme is then used to approximate the velocity and displacement of the particles from
the acceleration. The positions of all the particles are updated and a contact search algorithm is used to determine the contacting objects and the relative contact
overlap for the next timestep. The DEM is characterized by the repeated implementation of a forcedisplacement law at each contact within the system and Newton’s
second law for each particle.
3 USE OF THE USER DEFINED CONTACT MODEL
The UDCM allows the user to create virtually any contact model. The contact model is developed within the object oriented environment of C++ and communicates
with PFC via the protocols defined in the contact model’s class header file. This model is then compiled to a dynamic link library (dll) and the PFC simulation made
aware of its presence by a standard set of FISH commands in the input deck. The UDCM overrides the standard contact model. The user controls which variables are
to be passed to and from the contact model in the header file. A number of contact models, initially developed as UDCMs, are now present within the standard
PFC3D software package, including the linear viscous damping models (Equations 1 & 2).
The internal scripting language of FISH allows the user access to a substantial amount of data calculated by PFC3D during particle interaction events (e.g. the
identities of the contact partners, the relative contact velocity, etc). However, not all data is accessible. Energy dissipated by the numerical dashpots in the
Page 113
normal and shear directions and by friction in the shear direction—information necessary for the construction of an impact energy spectrum—although calculated by the
contact model in PFC3D, is not available to the user via FISH. This data can be recalculated using FISH, or an external process, but this adds significant computational
overhead and is not feasible for simulations involving a large number of particles. With the UDCM, however, since, as mentioned above, the user can control which
variables are passed to and from the model, obtaining the necessary data to generate the impact energy spectrum becomes straightforward.
4 GENERATION OF AN IMPACT ENERGY SPECTRUM
In this section we illustrate the process used to extract and present the impact data generated by PFC3D. An impact energy spectrum is a histogram plot of the
frequency of contact events versus the magnitude of the event. We are interested specifically in the amount of energy dissipated to motion normal and tangential to the
contact surfaces. In the case of the linear viscous damping model, energy dissipation normal to the contact surface would be determined by the normal damping
constant, and energy dissipated tangentially would be determined by both the (dynamic) coefficient of friction and the shear damping constant, and would depend in a
complicated fashion on the direction of the impulse vector. Experimental procedures such as drop weight tests used to assess the rock progeny formed from a single
impact event of a certain magnitude or an accumulation of impact events could then be used to estimate breakage from the impact energy spectrum provided by the
simulation. While this approach is by no means novel within the DEM literature (Mishra 1991), this is, to the authors’ knowledge, the first published work that makes
use of PFC3D to investigate impact energy distributions within a mill.
To provide statistically meaningful means of describing the charge motion within a mill, a probability distribution function of particle position is generated using a
recursive binning algorithm (Govender et al. 2001). The binning algorithm can be used to display the spatial variance in the magnitude and frequency of the impact
events, as shown in Figure 2 for an experimental mill 142 mm in diameter, rotating at 71 rpm. The cross section of the mill is uniformly divided into a fine Cartesian grid
(50×50 cells) where each cell represents a bin. The mill system is assumed planar for purposes of generating the bin plots: events occurring along the length of the mill
are projected onto a plane whose normal is parallel to the axis of rotation.
The position of the measured variable determines to which bin the variable is to be associated. The
Figure 2. Spatial distribution of the frequency and magnitude of contact events in an experimental mill.
Page 114
normalized count of experimental points within each bin represents the probability distribution function of particle position within the cross section of the mill. The bin
algorithm allows the frequency of any measured variable to be expressed as a function of position within the mill.
Three PFC data arrays store the accumulated energy dissipated to the normal, shear and frictional slip mechanisms. A fourth array stores the number of contact
events occurring within each bin. The arrays can be visualized as a Cartesian grid superimposed upon the cross section of the mill. The array positions correspond to
the number and position of the bins (see Fig. 2).
A recursive binning algorithm was implemented within PFC3D to ensure that the allocation of the energy data to the correct location within the arrays takes at worst
seven iterations per dimension, or in this case, at most fourteen iterations for the twodimensional plot. A defined particle size distribution is imposed upon the mill
charge when it is created, as illustrated in the two dimensional simulation of a large mill in Figure 3. The charge particles are then classified by dividing the size
distribution range into intervals. The four PFC data arrays described have a third dimension allowing the spatial distribution of the impact events to be categorized by
the size of the impacting media. An energy distribution plot per particle size class can therefore be generated (see Fig. 3).
The impact energy spectrum can be generated in a variety of ways. Three approaches were implemented in the current work. Firstly, each particle within the charge
was tracked and the accumulated amount of energy dissipated during all contact events that occurred within a mill revolution, as well as the number of contact events,
recorded using extra ball variables (a PFC function for allocating additional data to particles, walls and contacts; see Fig. 4). The duration of each contact event was
also determined. Although the duration of the event is not currently used to predict particle breakage it is envisaged that more complex breakage models will takes this
into account. If one assumes that each particle experiences only one dominant form of contact event during each mill revolution, then the average energy dissipated to
the various components can be calculated per particle per revolution. This energy can then be further averaged by classifying the charge by size and averaging the
accumulated energy dissipated and the number of impacts. An impact energy spectrum can be generated from this data.
A second approach is to divide the three PFC data arrays containing the accumulated impact energy dissipated in a component wise manner with the contact count
array (Fig. 3). The elements of the resulting array can be plotted in a histogram that is an approximation to the impact energy spectrum.
Figure 3. Various impact energy spectra for each particle size class within a two dimensional mill.
Page 115
Figure 4. Duration of individual contact events and the energy dissipated for each contact event per mill revolution.
The third approach is the standard approach adopted in the literature. An energy range for the magnitude of the possible impact events is assumed a priori. This range
is divided into intervals based upon a logarithmic scale (there are orders of magnitude more impacts in the low energy range than in the high energy range). A number of
particles are tracked and each impact event allocated to an energy interval. The number of impact events is also tracked. In this manner an impact energy spectrum is
produced. One problem with applying this method, however, is that the a priori estimation of the energy range for shear and frictional energy dissipation is not obvious
and may require some initial scoping simulations.
All of the approaches suggested are equally valid. The authors found that a combination of the first and second approaches was preferable, as this allowed the spatial
variation of the energy distribution to be assessed as well as providing data at the level of the individual particle. All methods are computationally expensive and to this
end the use of the parallel processing option in PFC3D was investigated. For relatively simple problems, those for which visualization of the bulk charge motion was
more important than tracking data associated with individual particles, the parallel processing option showed reasonable increases in performance. For more complex
problems, the procedure of geometrically dividing the space into regions and then giving the user the responsibility of tracking data as it flowed between regions was
found to be cumbersome and error prone.
5 THE LINEAR VISCOUS DAMPING MODEL
The choice of a contact model is critical if an accurate energy spectrum is to be generated, as it is from this spectrum that the breakage of particles is to be quantified.
The linear viscous damping model is currently in wide use, probably because it is a linear model, with a long history, and so well understood. However, the model fails
on two accounts.
Firstly, this model predicts that energy dissipation is highest when velocity is highest—that is, at the first instance of contact, before any significant deformation and
when strain should still be in the elastic regime, and at the last instance of contact, when the balls are about to separate. This is surely unphysical.
Secondly, the differential equations are stiff. As is well known, stiff differential equations present problems for explicit schemes, in that the region of stability may be
so small as to increase the solution time inordinately. Implicit schemes are unconditionally stable, but, as is also well known, they are not inherently more accurate than
explicit schemes operating within their region of stability, and may require greater overhead in iterating to a solution of the same accuracy.
We conducted a basic numerical assessment of the linear viscous damping model in PFC and in an explicit RungeKutta code, in order to check for accuracy and
stability for different values of the damping constant and to see just how serious the stiffness problem would be. A single particle of radius 0.1 m, density 7860 kg/m3
and stiffness 3e10 (the density and stiffness of steel) is dropped from a height of 1 m onto a rigid boundary. The energy balance is monitored and the coefficient of
restitution εn , defined as the absolute value of the ratio of rebound velocity, or velocity of the particle at last contact with the surface, to the initial velocity, or velocity at
first contact, is measured. Shear and tangential effects are not present in the current configuration.
There are two very different time scales in this problem. The time of fall from 1 m is on the order of 0.5 s, whereas the duration of contact with the rigid surface is on
the order of 100 μs. The automatic timestep
Page 116
mechanism in PFC is designed for quasistatic solutions (Itasca 2002, 2004), and hence with this scenario, typical for milling, we had to impose a restriction on the
maximum timestep to ensure energy balance.
The solution to Equation 1 for a system in free fall (gravity turned off) is
(3)
where
(4)
where U=0 at t=0, and where V0 is the velocity upon initial contact. The solution with gravity, Equation 2, in this case representing a particle in free fall hitting a
stationary wall, is
(5)
As one can see from Equation 5, gravity perturbs the solution to Equation 3 so as to result in a longer contact duration and a lower coefficient of restitution, without
affecting the critical value of the damping constant
(6)
The solutions to Equations 1 and 2 contain terms of the form exp(λt), where λ is a complex number with negative real part—i.e. they are stiff differential equations. For
such equations explicit schemes remain stable only if the product of λ and the stepsize is in the region of stability of the method, which can place an inordinate restriction
on stepsize (see, e.g. Burden et al. 1981). An implicit solver would guarantee stability, although this guarantee of stability could come at the cost of significantly greater
computational effort to obtain an acceptable level of accuracy. Explicit time integration schemes are the standard in DEM codes for good reason. PFC was not initially
designed for the simulation of milling, however, and we recommend that consideration be given to creating an alternate implicit scheme.
To test the accuracy of the numerical solutions the system was modeled from the point at which the ball, having acquired kinetic energy through free fall, had just
made contact with the wall, through to the point at which it just left the wall upon rebound. The results were compared with solutions to Equations 3 & 5, respectively,
by performing the simulation with gravity off and gravity on. Each simulation was conducted with PFC and with an embedded 4th–5th order RungeKutta, an explicit
solver with an adaptive stepsize, utilizing a computationally efficient error checking mechanism (Press et al. 1992). The solution obtained from the 4th order method is
compared to that obtained from the 5th order method. If the two solutions differ by more than a user specified error tolerance the stepsize is reduced and the solutions
recalculated.
When we took care to restrict the stepsize (with PFC) or to set a strict error tolerance (with the RungeKutta solver) the solutions were generally robust. However,
with gravity on and with high values of relative damping, greater than 0.80 with PFC and 0.85 with the RungeKutta solver, in this case, the stepsize dropped so
precipitously as to effectively prevent a solution. This value of the damping constant corresponds to a coefficient of restitution of 0.08 with gravity off, that is, for two
particles colliding while in freefall—a coefficient one might expect when simulating the impact of rock on rock. With gravity on, however—a rock impacting on a rock
resting on the liner, for instance—the coefficient of restitution becomes much lower and the solution process freezes. The problem of stability is worse for Equation 2
than for Equation 1 because of the constant mg/k term in the solution—despite the fact that in general the solution to Equation 5 represents only a perturbation of the
solution to Equation 3. The problem becomes worse when one lowers the stiffness, as is often done, since overall charge behaviour in milling simulations seems to be
relatively insensitive to stiffness, and simulation times can be significantly reduced.
Internal error checking has the potential to deliver more accurate solutions, although it requires extra overhead. When one is concerned with the overall flow of an
assembly of particles the extra accuracy may not be that important. In simulations of milling, however, obtaining an accurate energy spectrum is surely vital, and the
benefits of adding stricter error control may far outweigh the extra cost. These findings could have farreaching repercussions with regards to the simulation of milling
using DEM, for several reasons. Explicit time integration schemes are the standard in the use of DEM for milling, there has been significant investment in the linear
viscous damping model and it is in wide use, and, perhaps because of the focus on overall charge shape and charge motion, little consideration seems to have been
given to modeling at the level of accuracy that would yield an accurate and detailed energy spectrum.
6 ALTERNATIVE CONTACT MODELS
Several alternative contact models have been simulated, and will be tested more extensively in future
Page 117
numerical mill simulations. Two are nonlinear viscous damping models, and a third, a simple hysteretic model, is the bilinear model.
One may set the viscosity equal to the product of the velocity and the overlap, obtaining
(7)
This corrects the unphysical feature of the linear model. We also experiment with a variation of Equation 7,
(8)
where is the sign of the velocity. The bilinear model is given by
(9)

where k2<k1. The coefficient of restitution for Equation 9 is
(10)
In this model, during relaxation the stressstrain curve lies on a line of slope k2 which intersects the origin, and during penetration the stressstrain curve follows a line
with slope k1 which may or may not intersect the origin. A sample interaction is illustrated in the stressstrain diagram of Figure 5. Upon initial contact the stress rises
along line A from the origin, point 1, to point 2. As the particles begin to move apart the stress drops to point 3 on line B and moves along the line to point 4, say, when
collision with a third particle forces the initial two particles to move together again. The stress then rises along line C, whose slope is equal to that of line A, to point 5.
Upon relaxation the stress again drops to line B, point 6, and moves back to the origin, point 1, which represents the point of separation.
One can solve Equations 7 without the spring term to find that as in the case of the linear damping model, velocity decays exponentially. However, the full contact
models Equations 7 & 8 do not have a known
Figure 5. Schematic of the bilinear contact model.
exact solution. Consequently there is no independent check of the accuracy of numerical results, and the value of the damping constant for critical damping is not
known. Numerical solutions have the same character as the solutions of Equations 1 & 2. There does appear to be critical damping, and with a gravity term included
the number of time steps rises precipitously for large values of the damping constant.
The solution to Equations 9 modified to include gravity is
(11)

where V0 is the initial velocity, the frequencies are
(12)

the time t l is
(13)

and t f is the value of t for which U returns to zero. The arguments ω1t 1≥π/2 and ω2(t f−t l)≥π, with
Page 118
the equality signs holding when gravity is ignored. The coefficient of restitution from Equation 11 is more complicated than that from Equation 10, and is given by
(14)
This model does not have the potential instabilities of viscous damping models, although there is an unphysical discontinuity in acceleration at the turn around point. With
the RungeKutta solver, simulations utilizing the bilinear model with gravity on and with very low coefficients of restitution were more accurate than results from the
linear viscous damping model, readily iterating to an accurate solution in the same time as with higher coefficients of restitution.
7 SUMMARY AND CONCLUSION
By utilizing the User Defined Contact Model in PFC a procedure has been developed to extract an impact energy spectrum from a numerical mill. Simple numerical
experiments have been conducted in PFC and with an embedded RungeKutta solver to test the linear viscous damping model, two nonlinear viscous damping models
and a bilinear model. When one includes the effects of gravity the solutions to the viscous damping models can become problematic for high values of the damping
constant, in that the stepsize becomes inordinately small. The problem is exacerbated for low particle stiffness and high density. The bilinear contact model does not
produce these problems.
An hysteretic contact model has physical and numerical advantages over a viscoustype damping model, and we recommend that the former be adopted for milling
simulations. We also recommend that because of the widely varying time scales in milling and the accuracy required to generate a correct impact spectrum, an adaptive
stepsize method with error checking be developed in PFC, and possibly an alternate implicit solver.
ACKNOWLEDGEMENTS
Dr. Yoshiyuki Shimizu from Itasca developed the user defined contact models. His help and contribution to aspects of this work is greatly appreciated.
REFERENCES
Burden, R.L., Faires, J.D. & Reynolds, A.C. 1981. Numerical Analysis, 2nd Edition, Prindle, Weber & Schmidt, Publishers, pp 253–257.
Cundall, P.A. & Strack, O.D.L. 1979. A discrete numerical model for granular assemblies. Geotechnique, 29(1): 47–65.
Govender, I., Balden, V., Powell, M. & Nurick, G. 2001. Validated DEMPotential major improvements to SAG mill modelling. In Barratt et al. (eds.) Proceedings
International autogenous and semiautogenous grinding technology, Published CIM, vol. IV, pp. 101–114, September 30–October 3.
Itasca Consulting Group, Inc. 2002. PFC3D—Particle Flow Code in 3 Dimensions, Ver. 3.0, Minneapolis: Itasca.
Itasca, private communication, 2004.
Maw, N., Barber, J.R. & Fawcett, J.N. 1976. The Oblique Impact of Elastic Spheres. Wear. 38(1):101–114.
Maw, N., Barber, J.R. & Fawcett, J.N. 1981. The Role of Elastic Tangential Compliance in Oblique Impact. Transactions of the ASME, Vol 103:74–80.
Mishra, B.K. 1991. Study of media mechanics in tumbling mills by the discrete element method. PhD thesis, Dept. Metallurgical Eng., University of Utah.
Press, W.H., Teukolsky, S.A., Vetterling, W.T. & Flannery, B.P. 1992. Numerical Recipes in C, Cambridge University Press, pp 710–722.
Sarracino, R.S., Dong, H.J. & Moys, M.H. 2002. An experimental and theoretical study of ballwall impact for Discrete Element Method (DEM) modeling of tumbling
mills. In Discrete Element Methods, Proceedings of the Third International Conference, Santa Fe, N.M., USA, Sept 23–25 2002, pp 229–233.
Page 119
The modeling of the separation process in a ferrohydrostatic separator using
discrete element method
V.Murariu, J.Svoboda & P.Sergeant
TSS Technology, De Beers Consolidated Mines Ltd., Johannesburg, South Africa
5809 679 3
ABSTRACT: This paper presents a model of the separation process in a ferrohydrostatic separator, which has been designed and developed at TSS
Technology, De Beers, South Africa. The model was developed using special Discrete Element Method software package called PFC2D (Particle
Flow Code in 2D). The software models in 2D the movement and interaction of circular particles by the discrete element method. The simulation that
was carried out using particles of different densities and sizes has proved that interactions among particles are the most important factor leading to the
misplacement of particles. The theoretical results are then compared with the experimental results obtained in the De Beers FHS machine. The
comparison has shown that the DEM model can predict the experimental results very well and can also accurately describe mechanisms of
misplacements of the particles.
1 INTRODUCTION
A ferrohydrostatic separator (FHS) features a stationary magnetic fluid, housed in a separation chamber placed between polepieces of a magnet. Magnetic fluidbased
separations were pioneered in the early 1960s (Andres 1965). More recently, numerous ferrohydrostatic separators have been designed for separation of nonferrous
metals (Khalafalla et al. 1973 and Shimoiizaka et al. 1980), scrap metal (Mir et al. 1973), coal preparation (Fujita et al. 1990), for recovery of gold, platinumgroup
metals and diamonds.
A single nonmagnetic particle suspended in a ferrofluid is acted upon by force of gravity and by two buoyancy forces. The first is the classical Archimedes gravity
related force, and the other is the magnetically induced buoyancy force due to the magnetic “weight” of the magnetic fluid as introduced by Rosensweig (1966). These
forces acting on a nonmagnetic particle will force the particle to move along a certain path and the trajectory of the particle can be computed as it was described in
Murariu & Svoboda (2002).
A particle of radius b (small enough to assume that the magnetic field gradient is uniform within the particle), volume Vp , and density ρp immersed in a magnetic fluid
of density ρf, dynamic viscosity η and saturation magnetization Mf in the presence of an external magnetic field gradient will experience the following forces (see
Fig. 1):
(i) Force of gravity:
(1)
Figure 1. Section through the magnetic fluid bridge of depth δ, inclined with an angle α.
(ii) Buoyancy force:
(2)
(iii) The magnetically derived buoyancy force:
(3)

where μ0=4π×10−7 Tm/A is the magnetic permeability of vacuum.
Page 120
Figure 2. Trajectories of particles of different densities for h=3 mm. The dashed straight lines represent the trajectories of particle of
ρp=3500 and 3550 kg/m3 for h=0.
(iv) Drag force (written for the regime where Stokes Law is applicable and with the magnetic fluid at rest):
(4)
Equations (1)–(4) are written in SI units and this system of units is used throughout.
In equilibrium the total force of the particle is equal to zero and the socalled apparent density of the ferrofluid can be written as:
(5)
α being the angle between the direction of the magnetic field gradient and the horizontal direction. In Equation (5) the magnetization and the magnetic field intensity are
expressed in A/m.
Figure 2 depicts trajectories of nonmagnetic particles (of diameter 3 mm) of different densities (given in kg/m3) as they were computed in Murariu & Svoboda
(2002). The values of the parameters used in calculation are: the apparent density of the ferrofluid ρapp=3500 kg/m3, the ferrofluid density ρf=990 kg/m3, the ferrofluid
viscosity η=1.65× 10−3 kg m−1 s−1, the angle α=2° and drop height h=3 mm. It can be observed that a particle having the density greater than the apparent density will
follow a sink trajectory while a particle of density lower than the apparent density will float. Besides, a particle having the same density as the apparent density of the
fluid is sinking.
It is obvious that this model based on a single particle is not suitable for an assembly of particles interacting with one another. The interactions can deflect the particle
from its correct path and thus can lead to the misplacement of the particles. A single particle model can be used, eventually, for the calibration process, where single
particles are introduced into the fluid.
A software package, which can model the interaction between particles, can be successfully used to model the separation process. Such software is PFC2D and
PFC3D (Particle Flow Code in 2D and 3D) developed by Itasca Consulting Group Inc., USA.
Discrete Element Method (DEM) has been already used to model different processes occurring in mineral processing equipments such as tumbling mills (Mishra
2003, Clearly 1998, Clearly 2000), jigs (Mishra 2001), sieves (Li 2003) and crushers (Djordjevic 2003). There is no attempt yet to model the separation process in a
ferrohydrostatic separator by DEM.
This paper presents the results obtained by simulating, only in 2D, the separation process in the FHS machine using this software. In order to find out if the model
can predict the experimental results, the theoretical results were then compared to the experimental data. Theoretical partition curves were generated and compared to
experimental partition curves at the same feed rate. The theoretical curves are very close to the experimental ones, proving that the simulation can predict the
experimental results.

2D
2 GENERAL FEATURES OF THE PFC SOFTWARE
PFC2D models, in 2D, the dynamic behavior of assemblies of arbitrarily sized circular particles. A particle generator allows the statistical generation of specified
distributions of particles to be done automatically. Particle radii may distribute uniformly or according to a Gaussian distribution. Proprieties are associated with
individual particles (like initial velocity) or contacts. Any type of force can act on the particles, e.g. drag force, gravity force, friction force etc. Any number of arbitrarily
oriented line segments may be specified as walls, each with its own set of contact proprieties. These walls represent usually boundary conditions which force the
particles to move only inside a certain area (in 2D version of the software) or volume (in 3D version). The most important feature of the software is the contact model.
PFC2D allows assigning a contact model to each particle or wall, which describes the physical behavior occurring at a contact.
Looking at the features of the PFC2D software it can be easily observed that this software is very flexible and allows us to simulate the behavior of the particles in a
ferrohydrostatic separator.
3 THE MODELING STEPS
The simulation process has been developed in several steps, each step being described in the following sections.
Page 121
Figure 3. The walls defined in the model.
Figure 4. The electromagnet design for FHS machine and the plane in which the simulation takes place.
3.1 The definition of the walls based on the design of the chamber
Based on the chamber design and the position of the ferrofluid pool inside the chamber, the walls used in the modeling are the following: the ferrofluid surface, bottom
of the chamber, the splitter plate and the baffle plate. The walls can be also visualized in Fig. 3. It can be observed that the right hand side of the chamber is open. The
reason was to allow the software to delete the particles leaving the chamber and thus, to reduce the computational time.
3.2 The modeling of the magnetic field in the electromagnet gap and calculation of the apparent density of the fluid
In order to calculate the apparent density of the ferrofluid it was necessary firstly to model the electromagnet and calculate the magnetic field from the model. The
electromagnet was modeled using PC OPERA software for electromagnetic designs. The FHS electromagnet design is presented in Figure 4, as it was modeled in
OPERA 3D (Vector Fields).
Figure 5. The profile of apparent density along the vertical line in the middle of the gap.
After the modeling, a feature of the software, called commands input file was used to calculate the apparent density of the fluid along the vertical line of a plane situated
in the middle of the gap (see Fig. 4). This plane represents the plane in which the simulation takes place.
The profile of the apparent density along the vertical line is shown in the plot presented in Figure 5.
The modeling and the calculation of the apparent density prove that the apparent density is constant along the vertical only in a limited region (it was already
observed experimentally by Svoboda et al. 1996). Especially at the bottom of the chamber the apparent density drops steeply and becomes even negative. This means
that at the bottom of the chamber the particles are pushed downward instead of being lifted upward. The main reason for installing the baffle plate is thus to prevent the
particles introduced in ferrofluid from reaching the bottom of the chamber. Without the baffle plate all the particles introduced into the chamber, from a certain height,
would sink. The points on the plot are fitted using an equation of this form:
ρapp=y0+ay+by2
(6)

The values of constants y0, a and b, for which the fit is almost perfect (R2=0.93) are:
y0=−224073.5, a=6659.3 and b=−48.6
(7)
Equation (6) with the values from Equation (7) was used to calculate the magnetic buoyancy force.
The values of the apparent density inside the electromagnet gap are changing when the current density in the coils is changing. Using different values of the current
density in the coils it is possible to change the cutpoint density for the ferrofluid. The cutpoint density represents that density for which, ideally, all
Page 122
Table 1. The gray shades of particles corresponding to each density range.
the particles of density greater than the cutpoint density are sinking and all the particles of density lower than the cutpoint density are floating. By changing the current
density in the coils the apparent density will still be given by Equation (6), but the values of the parameters in Equation (7) will be different for different current densities.
This procedure will be used to set the cutpoint density in the model.
3.3 The particles used in DEM simulation
The particles were modeled as disks of uniform size and density distributions. The ranges of sizes and densities of the particles were the following: the size between 4
and 8 mm and the densities between 2900 and 3900 kg/m3. In the model particles belonging to a certain density range were colorcoded in different gray shades
(Table 1). All the particles had the same properties required by the DEM, such as stiffness and friction.
The shape of the particles was not considered here as being an important factor that can affect the final results. As long as the separation is based mainly on the
density of the particles, the contacts among particles can influence only the agglomeration process on the surface of the baffle plate.
3.4 The forces used in the model
The forces used to model the separation process were presented already in the Introduction section and they have the following expressions:
– Gravity force: G=ρp gVp
(8)

Figure 6. Calibration using virtual tracers of (a) density 3400 kg/m3 and (b) density 3530 kg/m3.
– Buoyancy force: FA=ρfgVp
(9)
– Magnetic buoyancy force: FM=ρappgVp
(10)
– Drag force: FD=6πηbν p
(11)
In the above equations ρp , ρf, ρapp represent the density of the particle, of the ferrofluid and apparent density of the fluid, respectively. All the other parameters have
been presented already above. Regarding the interactions between particles, the software can provide different contact models. The contact model chosen in the
present simulation was the HertzMindlin Contact Model (Itasca Consulting Group, Inc. 2003). With the other contact model, called Linear Contact Model, the results
were not so close to the experimental ones as the results obtained using HertzMindlin contact model.
3.5 Setting the cutpoint (CP) density
Before each simulation the cutpoint was checked using particles of certain density (called virtual tracers). For instance, to check if the desired cutpoint is 3250 kg/m3,
ten virtual tracers of density 3200 kg/m3 and ten of density 3300 kg/m3 are introduced into the fluid. The density of these virtual tracers was not constant. A range of
values was used according to the manufacturing error. Two static images from the animated simulation are presented in Figure 6a, b for virtual tracers of density 3400
kg/m3 and 3530 kg/m3, respectively. The images show that the virtual tracers of density 3400 kg/m3 are floating and those of density 3530 kg/m3 are sinking.
If the cutpoint density had not been the desired one, the current density in the electromagnet coils was changed and the new apparent density was calculated.
Page 123

Figure 7. Calibration using virtual tracers of density 3437 kg/m3.
Then, the calibration was checked again until the desired cutpoint density was achieved.
In contrast to experimental investigations in which the densities of the tracers have fixed values, any density of the virtual tracers can be set. For instance, by changing
the density of the virtual tracers, the actual cutpoint for the apparent density profile presented in Figure 6 is 3437 kg/m3, somewhat different from the average value
between 3400 and 3530 kg/m3, namely 3465 kg/m3 as determined by real tracers. Two static images from the simulation are presented in Figure 7. It can be observed
that the virtual tracers are colliding with the splitter plate, the fact that proves that the density of the virtual tracers is very close to the cutpoint density.
The modeling process is based on introduction of 2000 circular particles of different radii and densities into the ferrofluid. The initial velocity of the particles was
taken as the velocity of a particle falling freely in the air from a height of 4 cm. Particles were introduced one after another at a certain time interval. By modifying the
time interval between two successively particles introduced into the fluid, it was possible to change the feed rate. The shorter the time interval, the higher the feed rate.
In order to compare the model with the experimental results a certain feed rate had to be achieved. In the model, the feed rate was computed by adding the masses of
all particles and then the total mass was divided by the real time of the simulations (this time is given by the software during the simulation). Thus, the theoretical feed
rate was calculated using the following equation:
(12)
The time of simulation, in seconds, provided by the software after each simulation, can be input in Equation (12) and the result will represent the theoretical feed rate in
kg/h.
Figure 8. Snapshot taken during the simulation.
Figure 9. Theoretical partition curve obtained after simulation.
3.6 Generation of the theoretical partition curve
During the simulation, the particles leaving the chamber are deleted. This process is taking place only after the particle is counted and recorded in the right fraction (sink
or float) and the right density range. Thus, at the end of the simulation the percentage of each density range found into the sink fraction is recorded and written in a file.
A snapshot from the simulation is presented in Figure 8.
The theoretical partition curve obtained at a certain feed rate can be then compared with the experimental one at the same feed rate. For instance, in the case of the
simulation where the above image is taken from, the corresponding theoretical partition curve is presented in Figure 9. Two parameters are calculated from the curve,
namely: actual cutpoint (CP50) and the Ep value. These parameters are defined as follows (see Fig. 9):
– the actual cutpoint represents the value of the density for which the percentage in the sink fraction is 50%;
– Ep defines the sharpness of the separation and can be calculated using the equation:
(13)

In the graph the values of CP50 and Ep are also shown. Both parameters are measured in kg/m3.
Page 124
Figure 10. Experimental and theoretical partition curves at CP=3400 kg/m3. (a) Feed rate=33 kg/h; (b) Feed rate=83 kg/h; (c) Feed
rate=180 kg/h.
Figure 11. Experimental and theoretical partition curves at CP=3250 kg/m3. (a) Feed rate=33 kg/h; (b) Feed rate=83 kg/h; (c) Feed
rate=180 kg/h.
Page 125

Table 2. Comparison between the experimental and theoretical values of CP50 and Ep.
Set cutpoint (kg/m3) Feed rate (kg/h) CP50 (kg/m3) Ep (kg/m3)

Experimental Theoretical Experimental Theoretical
3400 33 3430 3440 38 37
83 3420 3430 53 58
180 3480 3510 146 110
3250 33 3270 3270 46 66
83 3330 3330 55 80
180 3330 3340 68 57
4 MODEL VERSUS EXPERIMENT
In order to compare the model with the experimental results, theoretical partition curves have been generated at the following cutpoints and feed rates:
– CP=3400 kg/m3 and the feed rates 33, 83 and 180 kg/h;
– CP=3250 kg/m3 and the feed rates 33, 83 and 180 kg/h.
Each theoretical partition curve obtained by simulation is then compared with an experimental curve. The experimental work was carried out on the FHS De Beers
separator using gravel of size range −8+4 mm and density tracers of 6 mm size. The gravel were mixed with the density tracers and then passed through the separator.
After separation, the tracers in each fraction were counted and experimental partition curves were drawn. The same feed rates, namely 33, 83 and 180 kg/h were used.
Both curves at the same feed rate are plotted on the same graph for the first cutpoint (3400 kg/m3) and second cutpoint (3250 kg/m3) in Figures 10 & 11,
respectively.
The experimental and theoretical values for CP50 and Ep are presented in Table 2.
It can be observed that the theoretical curves are close to the experimental ones, a fact which proves that the model can predict well the experimental results. As
seen in Table 2, the experimental and theoretical values of the actual cutpoint are very close. On the same table can be observed differences in the Ep values,
especially for the cutpoint equals to 3250 kg/m3. The reason could be the less number of particles (2000) used to simulate the separation process compare to the
mass of the gravel used in experiment (20 kg). Also, the model demonstrates that interactions between particles are the main reason for the misplacement of the
particles in this specific design of De Beers FHS machine.
It can be observed from these figures that the separation at the cutpoint of 3250 kg/m3 is better than the separation at 3400 kg/m3 for the same feed rate. The
reason is that at CP=3250 kg/m3 the particle mass split is around 50/50 compared to the case of CP=3400 kg/m3 in which most particles move
Figure 12. Two misplaced particles.
upward. The same results have been obtained during the test work, fact that proves once again the efficiency of the model.
During the experimental work has been observed a strange behavior of the density tracers of 3600 kg/m3 at cutpoint set 3400 kg/m3 and feed rate equals to 83
kg/h. In this particular case approximately 30% of these tracers have been reported into float fraction. It can be observed on the same plot that the same result has
been obtained in the simulation.
5 THE MECHANISM OF MISPLACEMENT
The misplacement of particles is due mainly to the interactions among particles. However, the particles can be also entrapped by other particles and carried in the
wrong fraction. Two static images in Figure 12 show two misplaced particles (using ellipses). The first image shows a light particle of density between 3000 and 3100
kg/m3 surrounded by heavy particles and, thus, forced to move downward. This particle will report, finally, into the sink fraction even if its density is lower than the cut
point density. The second image shows a heavy particle of density between 3700 and 3800 kg/m3 pushed upward by two light particles. This particle will report into
the float fraction.
Page 126
6 CONCLUSIONS
Comparative experimental and theoretical results on the separation process in De Beers ferrohydrostatic separator were presented. The theoretical results were
obtained by DEM modeling using a software package Particle Flow Code while the code for simulation was written by the authors.
By comparing the experimental partition curves with the theoretical curves obtained by simulations it was observed that the model could predict the experimental
results well. The theoretical values of the actual cutpoint CP50 were very close to the experimental ones for both cutpoints and all the feed rates used in experimental
tests. The theoretical values for Ep were closer to the experimental ones only in the case of the cutpoint equals to 3250 kg/m3. Thus, even if the model has used only
2000 particles for simulation, the prediction of the actual cutpoint was very good. It seems that this number was not enough for a well prediction of the Ep .
The model proved that the interactions among particles play the main role in the separation process in the case of De Beers FHS machine.
The model also proved that at higher feed rates the efficiency of separation decreased because of agglomerations of particles above the baffle plate.
The model explained the mechanism of misplacement of particles. Most of the misplaced particles were due to the entrapments and interactions among particles.
The simulations showed that when the cutpoint was chosen such a way that the particle mass split was around 50/50 the efficiency of separation was increased. This
is due to the fact that the number of entrapments in this case was reduced compared to the case of the cutpoint equals to 3400 kg/m3 when a predominant flux of
particles moving upward was present.
The circular shape of the particles seems do not affect the results of the simulation. The main reason was the fact that the separation process is based mainly on the
density of the particles. The shape could probably affect the agglomeration process that takes place on the surface of the baffle plate, but this fact has no major impact
on the separation process itself.
The 3D simulation of the separation process has not been considered here. The third dimension complicates the model due to the presence of a magnetic force that
pushes the particles towards the middle of the separation chamber. The authors plan to develop the 3D model as the subject of another paper.
An empirical equation for the separation process has not been derived here from the partition curves. This will be done after the 3D model of the FHS unit will be
developed. The PFC package proved to be a very powerful tool in simulating motion and interaction of particles.
REFERENCES
Andres, U.T. 1965. New Gravitational Methods of Mineral Dressing, Itogi Nauki Tekhniki, Ser., Moskow (Russian).
Clearly, P.W. 1998. Predicting charge motion, power draw, segregation, wear and particle breakage in ball mills using discrete element methods. Miner. Eng., 11(11):
1061–1080.
Clearly, P.W. & Hoyer, D. 2000. Centrifugal mill charge motion and power draw: comparison of DEM predictions with experiment. Int. J. Miner. Process., 59:131–148.
Djordjevic, N., Shi, F.N. & Morrison, R.D. 2003. Applying discrete element modelling to vertical and horizontal shaft impact crushers. Miner. Eng., 16:983–991.
Fujita, T., Mori, S., Mamiya, M. & Shimoiizaka, J. 1990. An improved sinkfloat testing apparatus for coal preparation using water based magnetic fluid. 11th Int. Coal
Prep. Congress, Tokyo, pp. 109–114.
Itasca Consulting Group, Inc., 2003. PFC2D—Particle Flow Code in 2 Dimensions, Ver. 3.0, Theory and Background Manual. Minneapolis: Itasca.
Khalafalla, S.E. & Reimers, G.W. 1973. Separation nonferrous metals in incinerator residue using magnetic fluids. Separation Science, 8:161–178.
Li, J., Webb, C., Pandiella, S.S. & Campbell, G.M. 2003. Discrete particle motion on sieves—a numerical study using the DEM simulation. Powder Technol., 133:190–
202.
Mir, L., Simard, C. & Grana, D. Recovery of nonferrous metals from scrap automobiles by magnetic fluid levitation. 3rd Urban Technology Conference and Technical
Display, Boston, AIAA Paper, No. 73–959.
Mishra, B.K. 2003. A review of computer simulation of tumbling mills by the discrete element method: Part I—contact mechanics and Part II—Practical applications. Int.
J. Miner. Process, 71:73–112.
Mishra, B.K. & Mehrotra, S.P. 2000. A jig model based on the discrete element method and its experimental validation. Int. J. Miner. Process, 63:177–189.
Murariu, V. & Svoboda, J. 2002. The effect of magnetic susceptibility on the motion of particles in a ferrohydrostatic separator. Magnetic and Electrical Separation, 11
(1–2):51–61.
Rosensweig, R.E. 1966. Fluidmagnetic Buoyancy. AIAA Journal, 4(10):1751–1758.
Shimoiizaka, J., Nakatsuka, K., Fujita, T. & Kounosu, A. 1980. Sinkfloat separators using permanent magnets and water based magnetic fluid. IEEE Trans. Magn., 16(2):
368–371.
Svoboda, J., Felix, J., Krüger, L., Dumbu, S. & Phakati, Z. 1996. Experimental investigation of ferrohydrostatic separator and development of its control system. Internal
De Beers Report.
Svoboda, J. 2000. Separation in magnetic fluids: time to meet technological needs. In: Proc. MINPREX 2000 Congress, Melbourne, Australia, p. 297.
Vector Fields Inc., PCOpera, Software for Electromagnetic Design.
Page 127
A discrete particle contact model for sintering
S.Luding
Particle Technology, DelftChemTech, TU Delft, Netherlands
5809 679 3
ABSTRACT: A discrete model for the sintering of discrete particles in two dimensions is presented with empirical contact forcelaws, taking into
account plastic deformations, cohesion, temperature dependence (melting and freezing), and also longtime relaxation effects. Samples are prepared
under constant isotropic load, and are sintered for different sintering times. During sintering, the sample shrinks, i.e. becomes denser with increasing
sintering time. Both increasing external load and sintering time lead to a stronger, stiffer sample after cooldown. With a compression test, the sample is
destroyed via vertical cracks opening.
1 INTRODUCTION
There are two ways to describe and model a particulate, inhomogeneous material like powder, ceramics or concrete. The first approach is based on continuum theory
and relies on empirical assumptions about the macroscopic material behavior (Schwedes 1968; Olevsky 1998; Vermeer, Diebels, Ehlers, Herrmann, Luding, and
Ramm 2001; Riedel and Blug 2001; Luding, Manetsberger, and Müllers 2004). This macroscopic approach is complemented by a “microscopic” description of the
material on the particle or grain level where the particles and their interactions are modeled one by one using socalled discrete element methods (DEM) (Herrmann,
Hovi, and Luding 1998; DeLo, Dutton, and Semiatin 1999; Vermeer, Diebels, Ehlers, Herrmann, Luding, and Ramm 2001). The former involves stress, strain and
plastic yield conditions (Redanz and Fleck 2001), whereas the latter deals with local forcedeformation laws for each contact (Jagota and Dawson 1990; Luding,
Manetsberger, and Müllers 2004).
The microscopic interaction model for plastic deformations, friction, and cohesion is discussed and time, temperature, and history dependent behavior is introduced
and rationalized. After definition of the contact force laws, a model system in a twodimensional box filled with disks of different sizes, is examined by means of DEM.
2 MODEL SYSTEM
The numerical “experiment” chosen is a biaxial box setup, see [S.Luding et al., in this volume], where the
Figure 1. (Left) Preparation of the sample with pressure on both sides. (Right) After the system is relaxed the evolution in time of the
sample with time is examined.
left and bottom walls are fixed, and a stress or straincontrolled deformation is applied to the other walls. In the future, quantitative agreement between experiment and
simulation has to be achieved, however, this issue is far from the scope of this paper where, mainly, the new contact model is proposed.
Inside the system, N disks with radii ai (i= 1,…, N) and height h are placed. The radii are drawn from a homogeneous distribution with mean a0 and relative width

w0 so that The particleparticle interactions and the parameters involved are discussed in the next section.
3 DISCRETE PARTICLE MODEL
The elementary units of granular materials are mesoscopic grains which deform under stress, possibly yield, change their properties with time, and can behave different
for different temperatures. Since the
Page 128
realistic modeling of the deformations of the particles is much too complicated to allow for a subsequent manyparticle simulation, we relate the interaction force to the
overlap δ of two particles. In the absence of longrange forces, an interaction takes place only if particles are in contact and thus δ>0. In that case, the forces are split
into a normal and a tangential component denoted by n and t, respectively.
If all forces fi acting on the particle i, either from other particles, from boundaries or from external forces, are known, the problem is reduced to the integration of
Newton’s equations of motion for the translational and rotational degrees of freedom. The integration of the equations of motion is performed with the simplest
molecular dynamics Verlet algorithm together with Verlettable neighborhood search (Allen and Tildesley 1987)—this is done in order to prove that the contact model
is stable (given a small enough integration timestep) and does not rely on advanced algorithms.
3.1 Normal contact model

Two particles i and j at positions ri and rj, with radii ai and aj, interact only if they are in contact so that their overlap
δ=(ai+aj)−(ri−rj)∙n
(1)
is positive, δ>0, with the unit vector n=nij=(ri− rj)/|ri−rj| pointing from j to i. The force on particle i, from particle j can be written as with n
perpendicular to t. In this subsection, the normal forces are discussed.
3.1.1 Short time contact model
As first step, we discuss the time and temperatureindependent behavior of the contact forces between a pair of particles. For this, we modify and extend the linear
hysteretic spring model (Walton and Braun 1986; Luding 1998; Tomas 2000). It is the simplest version of some more complicated nonlinearhysteretic force laws
(Walton and Braun 1986; Zhu, Shukla, and Sadd 1991; Sadd, Tai, and Shukla 1993), which reflects the fact that at the contact point, plastic (permanent) deformations
may take place. The repulsive (hysteretic) force can be written as
(2)

with k 1≤k 2 see Fig. 2.
During the initial loading the force increases linearly with the overlap δ, until the maximum overlap δmax is reached, which is kept in memory as a history parameter.
The line with slope k 1 thus defines the
Figure 2. Force law for two springs with stiffness k 1 and k 2 for initial loading and subsequent un/reloading, respectively. Attractive forces

are possible due to the cohesion strength k c.

maximum force possible for a given δ. During unloading the force drops, on the line with slope k 2, from its value at δmax down to zero at the forcefree overlap
δ0=(1−k 1/k 2)δmax.
Reloading at any instant leads to an increase of the force along this line, until the maximum force is reached; for still increasing δ, the force follows again the line with
slope k 1 and δmax has to be adjusted accordingly. Unloading below δ0 leads to negative, attractive forces until at the overlap
the minimum force f min=−k cδmin, i.e. the maximum attractive force, is obtained as a function of the model parameters k 1, k 2, k c, and the history parameter δmax.

Further unloading leads to attractive forces on the branch f=−k cδ. The highest possible attractive force, for given k 1 and k 2, is reached for k c ∞, so that f min=−
(k 2−k 1)δmax. This would lead to a discontinuity at δ=0 that is avoided by using finite k c.
The cone formed by the lines with slope k 1 and −k c defines the range of possible force values. If a force would fall outside the cone, it is forced to remain on the limit
lines. Departure from these lines into the cone takes place in the case of unloading and reloading, respectively. Between these two extremes, unloading and reloading
follow the same line with slope k 2. Possible equilibrium states are indicated as circles in Fig. 2, where the upper and lower circle correspond to a prestressed and
stressfree state, respectively.
3.1.2 Viscous dissipation
In the case of collisions of particles and for large deformations, dissipation takes place due to the hysteretic nature of the forcelaw. For small displacements around
some equilibrium state, the model does not
Page 129
contain strong dissipation. Therefore, in order to allow for stronger dissipation and thus faster relaxation, a viscous, velocity dependent dissipative force in normal
direction,
(3)
is assumed with some damping coefficient γ0. The halfperiod of a vibration around the equilibrium position, see Fig. 2, can be computed for arbitrary values of k 1 and
k c, as long as the overlap fulfills the condition δmin<δ<δmax. In this case, k 2 determines the stiffness and one obtains a typical response time on the contact level (Luding
1997),
(4)
the eigenfrequency of the contact, the rescaled damping coefficient η0=γ0/(2m12), and the reduced mass m12=m1m2/(m1+m2). The timestep of the simulation t MD has
to be chosen such that
t MD≈t c/50

for a proper integration of the equations of motion, so that we chose a typical timestep t MD=π/(50ω) after the model parameters k 2 and γ0 are specified.
3.1.3 Stiffness increase with contact area
In order to account for the fact that a larger contact surface leads to a larger contact stiffness, the coefficient k 2 is made dependent on the maximum overlap history
parameter δmax (and thus on the forcefree overlap δ0), as long as the overlap is below the threshold δfluid fluid that corresponds to the “complete melting” of the
particles (Luding, Manetsberger, and Müllers 2004).

The stiffness is maximal in the fluid limit for δ0=δfluid, which corresponds to and varies between k 1 and k 2 for smaller overlaps, so that
(5)
For large overlaps (in the fluid regime), the stiffness and the force is thus only dependent on k 2, independent of k 1. For smaller overlaps both k 1 and k 2 affect the force

together with the history of this contact.
The hysteretic stiffness model takes into account an increasing stiffness with increasing overlap. The first loading is plastic with low stiffness, and subsequent un and
reloading are stiffer because the material was initially compressed. Thus, also the maximum cohesive force depends on the maximum compression which was
experienced by the contact during its history.

Figure 3. Force law with varying stiffness k 2(δmax). If δ0 becomes larger than δfluid, the stiffness remains equal to k 2.
3.2 Density temperature dependence
If a solid (we assume a simple material here), is heated, in general, its volume increases so that its density decreases. Therefore, we assume a temperature dependent
density of the single particles (disks with radius a and height h):
(6)

with the density change per unit temperature δρT. This corresponds (in linear approximation) to a change of the particle radius
a(T)=a(Tmelt) [1−δaT (Tmelt−T)],
(7)
with the relative change of the radius per unit temperature δaT. This approximation can be used if the range of temperatures is rather narrow and the changes per unit
temperature are very small.
3.3 Contact temperature dependence
For the temperature dependence, we focus on an inhomogeneous material with a melting temperature Tmelt and assume that the material behaves static, as described
above, if the temperature T is much smaller than the melting temperature. The behavior of the stiffness k 1 is schematically shown in Fig. 4 as a function of the
temperature.
The material becomes softer when T −T becomes small and will lose all stiffness in the limit
melt

Page 130

Figure 4. Schematic plot of the stiffness k 1 as a function of the temperature.

and in the hot and the liquid regime, Tvar, one has k 1 0 and the particles should behave “liquid” with considerable overlap.
3.3.1 Increasing temperature
When the temperature is increased to a rather large value, close to the melting point, two particles under stress and in equilibrium due to compressive forces will lose
stiffness and thus will deform stronger so that their overlap becomes larger. Therefore, we assume for the stiffness coefficient
(8)
with the dimensionless temperature difference τ. When k 1 is reduced due to an increase in temperature (+), we assume that
Thus the material volume shrinks due to sintering (increasing overlap) at the contact level. Note that k 2 is not changed
directly when T is increased.
If, as an example, the material has a melting point Tmelt=120°C with a range of softening of Tvar=10°C, for a temperature of 120°C, the stiffness is only 0.4 of the
cold limit.
3.3.2 Decreasing temperature
If, later in time, the temperature is decreased again, k 1(T) is adjusted according to Eq. (8), but since the melted (sintered) area around the contact point will not return
to its previous state, we now assume that see the right panel in Fig. 5.
The temperature dependence thus can lead to changes of the stiffness and of the overlap (deformation) of the particles.
3.4 Temperature dependence with time
In addition to the direct effect of a temperature change on the particle size and the stiffness, the material
Figure 5. Force laws for varying stiffness k 1, according to Eq. (8). (Left) Temperature increase: k 1 is reduced while remains

constant (solid line with slope k 2).
may change its internal, atomistic structure such that, e.g., defects heal and disappear. This effect will occur mainly in the regime of higher temperatures. Therefore, a
time dependence is introduced that leads to an explicit change of the material stiffness k 1 with time. The change takes place extremely slowly with an algebraic time
dependence, so that k 1(T, t) lags behind when varying from its actual value to the desired, final value k 1(T), as defined in the above Eq. (8).
When the temperature is increased from a small value T0 to T, then k 1(T0) changes to k 1(T) with the rate of change
(9)
with the time scale t 0 on which a typical change takes place. Note that k 1, k 1(T), and k 1(T, t) are different, in general, and correspond to the maximum k 1, the

temperature dependent k 1(T)<k 1, and the time dependent k 1(T, t) that tends towards k 1(T). The sign in Eq. (9) is chosen according to the sign of [k 1(T)−k 1(T, t)].
Assume T0=20°C and T=118°C, so that k 1(T0)=1 corresponds to k 1(T)=0.6. The stiffness as a function of time is plotted in Fig. 6 for the time constant t 0=10 s.
Note that, due to the factor k 1(T) in the denominator of Eq. (9), the change of stiffness is faster for higher temperatures. In the hot limit, changes take place very
rapidly, whereas in the cold limit changes are extremely slow.
Finally, we note that the adaption/relaxation of k 1(T, t) to the desired k 1(T) value follows the above equations in all cases except when the temperature is going
down and k 1(T)>k 1(T, t). In that situation the contacts freeze rapidly and thus have to become strong as fast as the system cools down.
3.5 Cohesion dependence on stiffness and friction
The cohesive properties of a particle contact depend on the temperature, in so far that a melted contact
Page 131
Figure 6. Variation of the stiffness k 1 with time. The inset shows the same function with a logarithmic time axis—only after about 103 s
the final stiffness is reached.
should have weak tensile and compressive strength. Therefore, we couple the cohesive parameter k c to the magnitude of k 1(T, t), which decreases with temperature

increasing. In addition, in order to take into account a reduced tensile strength of a soft contact with weak deformation and thus small overlap, the cohesion is directly
related to the stiffness k 2(δmax):
(10)
This is an arbitrary choice for the temperature dependence of the cohesive force strength k c, but as long as no detailed experimental results are available, we stick to
this empirical law.
3.6 Tangential contact model
The force in tangential direction is implemented in the spirit of Cundall and Strack (Cundall and Strack 1979; Luding 2004), who introduced a tangential spring in order
to account for static friction. Various authors have used this idea and numerous variants were implemented, see Brendel and Dippel 1998; Luding 2004 for a summary
and discussion, and the paper by Luding et al. in this volume for a description of the tangential spring implementation.
In brief, the tangential force is coupled to the normal force via Coulomb’s law, i.e. f t≤μf n , where for the limit case one has sliding friction and for the case of small
forces f t, one has static Coulomb friction. As a consequence of the cohesion force in normal direction, attractive forces are possible so that Coulomb’s law has to be
modified
f t≤μ(f n −f min),
(11)
with the minimum (maximum attractive) force
(12)
Thus the tangential force is related to the normal force relative to the point of cohesionfailure. Note that this changed force law can lead to a stable equilibrium of the
solid also at f n ≈0 and activated tangential static friction.
3.7 Temperature dependence in tangential direction
In parallel to the change of normal stiffness, the tangential stiffness is always kept in a constant ratio α to k 2 so that
k t=αk 2(δmax),
(13)
since the stiffness in tangential direction is based on the same arguments as the material stiffness in normal direction.
The friction is coupled to the temperature dependent value of the stiffness k 1(T, t), because friction should not be present in a liquid at large enough temperatures, so
that
(14)
Thus friction is modified together with the changes in normal direction. No new ideas are introduced for the tangential forces.
4 RESULTS
Now, the sintering model can be applied to the sintering process of a particulate material sample (Luding, Manetsberger, and Müllers 2004). The material is initially a
loose powder and first has to be prepared at low temperature from time t=0 to time t heat, see Fig. 7. The preparation takes place with isotropic external pressure
p=100, where pressure is measured in units of Nm−1 due to the twodimensional nature of the model, and a pressure of p=100 corresponds to f n ≈2ap, as rough
estimate. The other system parameters are summarized in (Luding, Manetsberger, and Müllers 2004).
4.1 Sample preparation
The preparation of the sample consists of an initial relaxation period at constant temperature T=80° C until time t heat, when the system is heated up to T=140°C
between time t heat and Tsinter. During the sintering time t s, the system is allowed to sinter with a much shorter relaxation time, t 0=0.1 instead of
Page 132
Figure 7. Schematic plot of the temperature variation during simulation in a preparationheatingsinteringcoolingrelaxation cycle.
Figure 8. Material density ν as function of time for simulations with N=300 particles and side stress p=100.
t 0=103 as used during the rest of the process. This “trick” allows for a long time sintering simulation while keeping the simulation time small. During sintering, the time
axis should be stretched by a factor of 104 in order to obtain the realtime behavior. At the end of the sintering process, at time t cool, the sample is slowly cooled down
and, at time t relax allowed to relax at constant temperature until time t f. Here, the times t heat=0.2, t sinter−t heat=0.1, different sintering times t s:=t cool−t sinter,
t relax−t cool=0.1, and t f−t relax=0.1 were Used.
4.1.1 Density
The longer the sintering time t s, the lower the value of k 1 gets. For short sintering time, the lowest values are never reached, because the system is cooled down before
the sintering is finished. At the end of the sintering time, k 1 is increasing during the cooling process of the sample and reaches its inital value. However, the melting and
sintering of the contacts is not reversed, as becomes evident when plotting the volume fraction ν in Fig. 8.
The initial preparation step leads to a rather low density of ν ≈0.8. At time t heat, the relaxed sample is heated up and, at the same time, the wall friction is switched
off. The latter has an immediate effect,

Figure 9. Material density as function of the vertical strain for samples prepared with different sintering times ts and pressure p w=100.
a slight increase in density is possible due to reorganizations. The increased temperature becomes only effective at time t sinter, when the relaxation time t 0 is decreased

in order to accelerate the evolution of the system. Note that the increase in density due to the sintering seemingly appears quite rapid—due to the quenched time axis
during sintering.
In summary, longer sintering and larger confining pressures (data not shown) lead to higher densities of the sintered sample. Through test simulations with different
parameters, we verified that the increase in density is only partially due to the contact model, but also is caused by reorganizations in the sample. A small thermal
expansion of the particles, thus, does not affect the results.
4.2 Material properties
The samples prepared via the procedure described above are now tested via a compression test, where the vertical confining pressure is slowly increased. More data
and also a vibration test are provided in Luding, Manetsberger, and Müllers 2004.
The compression test is a variant of the biaxial compression frequently used in soil or powder mechanics. A given compressive stress in one direction (p=σxx) is
kept constant by moving the right, vertical wall, if neccessary. The other confining stress σzz is increased by moving the vertical wall down in a defined way, where
compression is defined positive, εzz(t)>0. The density evolution of a sample with N=300 particles is plotted against the strain for various sintering times in Fig. 9.
The vertical stress is plotted against the strain in Fig. 10. With increasing strain, the vertical stress in the sample increases and the material fails at a
Page 133

Figure 10. Vertical stress as function of the vertical strain for samples prepared with different sintering times ts and pressure p w=100.
Figure 11. Snapshot of a compression test with N=300 and p=10, at deformation εzz=0.04 and ts=1.0. (Top) The circles are the particles
with the greyscale coding the average stress; dark and light correspond to small and large stresses. (Bottom) The lines are
contacts; cracks are where no lines are drawn anymore. The small solid circles denote the tangential forces, with their size
proportional to its magnitude.
certain maximal stress and critical strain (Luding, Manetsberger, and Müllers 2004). The material stiffness (dashed line in Fig. 10) is increased by about a factor of two
when the confining stress is increased by a factor of ten. In summary, the critical strain and stress where the material fails increase with increasing confining pressure,
sintering time and system size.
One snapshot from a compression test for a long sintering time, N=300 and p=10 is presented in Fig. 11. Gaps (cracks) between the parts of the sample have
opened and fragments are evident. This is only a representative example for the compression test; a more detailed study of the fracture behavior, samplesizeand
sinteringtimedependence is far from the scope of this study.
5 CONCLUSION
In summary, a discrete contact model for the sintering of particulate materials was introduced and some material samples were sintered for different times. Then they
were tested with respect to their anisotropic load strength: Longer sintering and stronger confining pressure systematically increases the density and the strength of the
material. Depending on the sintering duration, either isolated particles (short sintering), fragments or a single solid block of material (long sintering at large pressure)
could be produced. Further research includes more model tests via a comparison with experimental data, extension of the model to three dimensions—the model was
designed such that this should be straightforward. The last missing ingredients in the model, like rolling and torsionresistance will make the model even more realistic.
Note, however, that already the model presented here increased the amount of computation necessary for each contact by some factor.
ACKNOWLEDGEMENTS
The author thanks M.Lätzel, K.Manetsberger, Johannes Müllers, J.Tomas, S.Diebels, H.Besserer, and G.A.D’Addetta for discussions and acknowledges financial
support by the Deutsche Forschungsgemeinschaft (DFG) and by the DaimlerChrysler research division.
REFERENCES
Allen, M.P. and D.J.Tildesley (1987). Computer Simulation of Liquids. Oxford: Oxford University Press.
Brendel, L. and S.Dippel (1998). Lasting contacts in molecular dynamics simulations. In H.J.Herrmann, J.P.Hovi, and S.Luding (Eds.), Physics of Dry Granular Media,
Dordrecht, pp. 313. Kluwer Academic Publishers.
Cundall, P.A. and O.D.L.Strack (1979). A discrete numerical model for granular assemblies. Géotechnique 29(1), 47–65.
Page 134
DeLo, D.P., R.E.Dutton, and S.L.Semiatin (1999). A comparison of discrete element model prediction to observation of metal powder consolidation. Scripta Materialia
40(10), 1103–1109.
Herrmann, H.J., J.P.Hovi, and S.Luding (Eds.) (1998). Physics of dry granular media—NATO ASI Series E 350, Dordrecht. Kluwer Academic Publishers.
Jagota, A. and P.R.Dawson (1990). Simulation of the viscous sintering of 2 particles. Journal of the American Ceramic Society 73(1), 173–177.
Luding, S. (1997). Surface waves and pattern formation in vibrated granular media. In Powders & Grains 97, Amsterdam, pp. 373–376. Balkema.
Luding, S. (1998). Collisions & contacts between two particles. In H.J.Herrmann, J.P.Hovi, and S.Luding (Eds.), Physics of dry granular media—NATO ASI Series
E350, Dordrecht, pp. 285. Kluwer Academic Publishers.
Luding, S. (2004). Micromacro transition for anisotropic, periodic, elastic solids. in preparation.
Luding, S., K.Manetsberger, and J.Müllers (2004). A discrete model for long time sintering. J. Mech. Phys. Solids, in press.
Olevsky, E.A. (1998). Theory of sintering: from discrete to continuum. Materials Science and Engineering R23, 41–100.
Redanz, P. and N.A.Fleck (2001). The compaction of a random distribution of metal cylinders by the discrete element method. Acta Materialia 49, 4325–4335.
Riedel, H. and B.Blug (2001). A comprehensive model for solid state sintering and its application to silicon carbide. In T.J.Chuang and J.W.Rudnicki (Eds.), Multiscale
deformation and fracture in materials and structures, the James R.Rice 60th anniversary volume, Solid mechanics and its application—Vol. 84, Dordrecht, pp.
49–70. Kluwer Academic Publishers.
Sadd, M.H., Q.M.Tai, and A.Shukla (1993). Contact law effects on wave propagation in particulate materials using distinct element modeling. Int. J. NonLinear
Mechanics 28(2), 251.
Schwedes, J. (1968). Flieβνerhalten von Schüttgütern in Bunkern. Weinheim: Verlag Chemie.
Tomas, J. (2000). Particle adhesion fundamentals and bulk powder consolidation. KONA 18, 157–169.
Vermeer, P.A., S.Diebels, W.Ehlers, H.J.Herrmann, S.Luding, and E.Ramm (Eds.) (2001). Continuous and Discontinuous Modelling of Cohesive Frictional
Materials, Berlin. Springer. Lecture Notes in Physics 568.
Walton, O.R. and R.L.Braun (1986). Viscosity, granulartemperature, and stress calculations for shearing assemblies of inelastic, frictional disks. Journal of Rheology 30
(5), 949–980.
Zhu, C.Y., A.Shukla, and M.H.Sadd (1991). Prediction of dynamic contact loads in granular assemblies. J. of Applied Mechanics 58, 341.
Page 135
Discrete element modeling for dynamics of vibroimpact system
M.Saeki
Niigata Institute of Technology, Fujihashi, Kashiwazaki, Niigata, Japan
5809 679 3
ABSTRACT: Granular materials are used to improve the damping of vibrating structures. This is known as particle impact damping. Presented here is
a numerical method for calculating the performance of a particle impact damper equipped with a number of boxlike cavities. This method is based on
the discrete element method (DEM). The validity of this method is examined by comparison with experimental results. The effects of system
parameters on the performance of the damper are considered.
1 INTRODUCTION
Granular materials have proved useful for vibration suppression. This is known as particle impact damping or granular damping (Araki et al. 1985, Panossian 1992,
Sato et al. 1995, Yokomichi et al. 1996, Papalou & Masri 1996 and Friend & Kinra 2000). Particle impact damping involves granular materials constrained to move
in the cavity of a primary vibrating system. The damping effect results from the exchange of momentum during the impact of granular materials against the wall of the
cavity. This damping has the advantage of being independent of the environment. Therefore, particle impact dampers can be applied in extreme temperature
environments, where most conventional dampers would fail. In order to use the particle impact dampers effectively, it is necessary to solve the equations of motion for
each particle, considering the granularity. The discrete element method (DEM) (Cundall 1979) has been found to be effective for revealing the dynamics of particle
impact damping (Saeki 2002). The important parameters of the dampers are the cavity dimensions and mass ratio, which is the total mass of granular materials divided
by the primary mass. When the mass of the primary system is large, a large cavity or a number of small cavities must be attached to the primary system. Many
researchers have been studying particle impact dampers. However, most previous studies have focused on particle impact dampers equipped with a single cavity.
This paper investigates the damping efficiency of a particle impact damper equipped with a number of boxlike cavities (termed the multiunit particle damper in the
following). The objective of this work is to develop an analytical method for the damper, using DEM. The validity of the method is confirmed by a comparison of
experimental results. It is shown that the particle size and number of unit influence the damping performance.
2 MULTIUNIT PARTICLE DAMPING
Figure 1 shows the model of the multiunit particle damper. The primary system is supported with a spring and dashpot, and acts as a singledegreeoffreedom system.
A number of cavities are attached to the primary system. Particles of uniform size and number are placed in each cavity. When the primary system is excited by the
motion of the support point U, particles may contact each other or the wall of the cavity. The collision of the particles with the wall of the cavity results in a reduction of
the primary system
Figure 1. Model of the multiunit particle damper.
Page 136
amplitude through momentum transfer. The equation of the motion for the primary system is given by
(1)
where u is the harmonic displacement of the support point U, and a and ω are the amplitude and the angular frequency of the vibrating support point U, respectively. M
is the mass of the primary system, and C and K are the damping constant and spring stiffness of the vibratory system, respectively. Fp is the contact force between the
primary system and particles in the x direction. The dot denotes a time derivative.
3 DISCRETE ELEMENT MODELING
The damping performance depends on the behavior of all particles. Therefore, in order to evaluate the entire system in detail, the motion for each particle must be taken
into consideration. In this study, an analytical method based on the discrete element method (DEM) (Cundall 1979) is provided to estimate the damping performance.
The equation of motion for a particle i is given by
(2)
where m is the particle mass, I is the moment of inertia of the particle, g is the acceleration vector due to gravity, p is the position vector of the center of gravity of the
particle, φ is the angular velocity vector, and F is the summation of the contact forces acting on the particle. T is the summation of the torque caused by the contact
forces.
Next, consider the contact state between the two particles i and j. The contact force acting on a particle i can be divided into the normal component f N and
tangential components f T of the contact force. According to Tsuji et al. (1992), the normal component f N of the contact force is modeled by the sum of the spring force
and viscous damping force, and is expressed by Eq. (3). The tangential components f T of the contact force are given in Eq. (4), considering Coulomb’s law of friction.
(3)
(4)
(5)
In these expressions, δ and are the displacement and velocity of particle i relative to particle j, respectively. k N is the spring constant, cN is the damping coefficient, r

is the particle radius, and μ is the coefficient of friction between two particles or between a particle and the wall of cavity. In this study, the adhesive force between two
particles is neglected.
By applying the Herzian theory of elastic contact for spheres of the same size, the spring constant k N in Eq. (3) is expressed as
(6)

where Ep and ν p are the modulus of elasticity and Poisson’s ratio of particles, respectively. In the case of contact between a sphere and the wall, k N is expressed as
(7)
where E0 and ν 0 are the modulus of elasticity and Poisson’s ratio of the wall, respectively. When the damping coefficient cN is defined by the following equation, cN is a
function of the coefficient of restitution e (Tsuji et al. 1992).
(8)
where α is determined by the coefficient of restitution e. Each particle may be in contact with other particles and the wall at the same time. However, it is assumed that
the contact force can be defined by Eqs. (3)–(8).
4 EXPERIMENT
Figure 2 shows the experimental apparatus. The primary system, which is fabricated from acrylic resin, is excited by the horizontally vibrating shaker through the leaf
springs, and acts as a singledegreeoffreedom system. Five cavities are attached to the primary system, as shown in Figure 1.
The dimensions of each cavity are length L= 60 mm, height H=40 mm and width W=60 mm. The motion of the primary system was measured with an
accelerometer. The impacting granular spherical particles in each cavity used in this study are made of acrylic resin, and are of uniform size (r=3 mm). The equivalent
properties were found to be M=1.153 kg, C=0.423 Ns/m, K=3795.1 N/m. Table 1 lists the particle properties. To investigate the effect of the number of cavities on
the damping efficiency, granular materials are placed as shown in Figure 3.
Page 137
Figure 2. Experimental apparatus.
Table 1. Particle properties.
Total number 425–855
Density
1190 kg/m3
Coefficient of friction μ 0.52
Damping constant α 0.077
Spring constant k N (particle/particle) 1.0×107 Nm−3/2

Spring constant k N (particle/wall) 1.3×107 Nm−3/2

Figure 3. Cavity arrangement.
5 EXPERIMENTAL AND ANALYTICAL RESULTS
Figures 4–6 show comparisons between experimental and analytical results. All the figures show the relationship between the root mean square (r.m.s.) value of the
primary system amplitude and the frequency of the vibrating support point. In Figures 4 & 5, the numbers of cavities are 3 and 5, respectively. The amplitude is smaller
in Figure 6 than it is in Figures 4 & 5. From these figures, it is evident that the calculated result indicates the experimental trend with reasonable accuracy. The approach
in the present work is therefore effective for estimating the damping performance of the multiunit particle damper.
Next, the effect of system parameters on the damper performance is discussed in terms of the analytical results.
Figure 7 shows the time history of the behavior of the granular materials in a cavity. The broken line shows the cavity when the primary system is not
Figure 4. Comparison between experimental and calculated results (a=1.0 mm, λ=0.10).
Figure 5. Comparison between experimental and calculated results (a=1.0 mm, 5 cavities).
excited. It is clear that the granular materials repeatedly collide with the wall of the cavity. It is also found that granular materials collide with the wall in the direction
opposite to the primary system’s motion.
Figures 8 & 9 show the effect of the particle radius on the damper performance. The cavity length L is smaller in Figure 8 than in Figure 9. In the case of Figure 8,
the maximum amplitude of the primary system occurs at a lower frequency for r=5.55 mm than it does for r=3 mm. This trend is different from that in Figure 9. In the
case of Figure 9, the maximum amplitude occurs at almost the same frequency and the resonance amplitude of the primary system
Page 138
Figure 6. Comparison between experimental and calculated results (a=0.5 mm, 5 cavities).
Figure 7. Behavior of granular materials (a=1.0 mm, f=9.0 Hz, 3 cavities, λ=0.10).
Figure 8. Effect of particle radius (L=53 mm, 5 cavities, a=1.0 mm, λ=0.10).
decreases as the particle radius increases. The reason for the difference seems to be that the difference in the apparent volume results in a difference in the response
amplitude.
Figure 9. Effect of particle radius (L=73 mm, 5 cavities, a=1.0 mm, λ=0.10).
Figure 10. Effect of cavity length (a=1.0 mm, 5 cavities, λ=0.10).
Figure 10 shows the effect of the cavity length on the maximum r.m.s. value of the primary system amplitude. It is clear that the optimum cavity length decreases with
decreasing the particle radius. It is also found that the maximum r.m.s. values of the primary system amplitude at the optimum cavity length are almost the same.
6 CONCLUSIONS
The performance of the particle impact damper equipped with a number of boxlike cavities was analyzed by means of the DEM. The validity of the method was
confirmed by a comparison with experimental results. It is found that the damping efficiency depends on the number of the cavities, particle radius and cavity
dimensions. The behavior of granular materials obtained by the analytical method appears to be realistic.
Page 139
ACKNOWLEDGEMENT
I wish to acknowledge the assistance of T.Aiba in the experimental phase of the investigation.
REFERENCES
Araki, Y., Yokomichi, I. & Inoue, J. 1985, Impact dampers with granular materials. JSME, 28(241):1466–1472.
Panossian, H.V. 1992, Structural damping enhancement via nonobstructive particle damping technique. ASME J. of Vibration and Acoustics, 114:101–105.
Sato, T., Tanaka, K., Aida, S. & Mouri, Y. 1995, Vibration isolation in a system using granular medium. JSME Int. J., Series C, 38(3):434–440.
Yokomichi, I., Araki, Y., Jinnouchi, Y. & Inoue J. 1996, Impact damper with granular materials for multibody system. ASME J. of Pressure Vessel Techn., 118:95–103.
Papalou, A. & Masri, S.F. 1996, Performance of particle dampers under random excitation. ASME J. of Vibration and Acoustics, 118:614–621.
Friend, R.D. & Kinra, V.K. 2000, Particle Impact Damping. J. of Sound and Vibration, 233(1):93–118.
Cundall, P.A. & Struck, O.D. 1979, A discrete numerical model for granular assemblies. Geotechnique, 29:47–65.
Saeki, M. 2002, Impact Damping with granular materials in a horizontally vibrating system. J. of Sound and Vibration, 251(1):153–161.
Tsuji, Y., Tanaka, T. & Ishida, T. 1992, Lagrangian numerical simulation of plug flow of cohesionless particles in a horizontal pipe. Powder Technology, 71:239–250.
Page 140
Page 141
Shear flow
Page 142
Page 143
GEM simulation of shear flows in granular media
Y.Kishino
Department of Civil Engineering, Tohoku University, Tohoku, Japan
5809 679 3
ABSTRACT: Granular media behave solidlikely when they are put in a quasistatic environment and fluidlikely when they lose their static
equilibrium. Their widerange mechanisms are difficult to be investigated in the current continuummechanical framework. In this paper, the transitional
behavior between solid and fluid is studied in terms of a numerical method called the granular element method (GEM). The granular element method
was originally developed by the author to study quasistatic problems of granular media. Recently, the static code was extended to a dynamic version.
The application of the new code to shear flow problems showed that no drastic change in granular arrangements was observed for a quasistatic case
and that the initial fabric was totally destroyed in an inertiadominant flow where the confining stress was equilibrated by the momentum transfer from
colliding particles.
1 INTRODUCTION
As is observed in debris flows or any other granular flows, granular media behave solidlikely when they are put in a quasistatic environment and fluidlikely when they
lose their static equilibrium. Our final goal is to formulate a universal constitutive model that covers widerange behaviors of granular media. However, due to the
complexity of behaviors, we should say that the goal is still far.
Through a discrete particle computer simulation of the twodimensional Couette flow, Zhang & Campbell (1992) studied the phase change from fluidlike behavior
to solidlike behavior and found that the transition was a yieldlike phenomenon. Louge et al. (2001) proposed a shear cell apparatus to investigate the transitional
behavior in the space laboratory. To support such studies, three dimensional simulations will be inevitable.
The granular element method (GEM) presented by Kishino (1988, 2001) may be one of the most suitable methods to study the transitional behaviors. This method is
similar to the distinct element method proposed by Cundall & Strack (1979). However, it adopts an iterative algorithm to reduce numerical errors. The original quasi
static method, which is basically a stiffness method, was recently extended to a dynamic version and utilized to shear cell simulations. This paper explains the algorithm
of dynamic version of GEM and demonstrates the variety of shear flow behaviors observed in shear cell simulations under wide range of shear rates.
2 GRANULAR ELEMENT METHOD
2.1 Static algorithm
The granular element method was initially proposed for quasistatic analyses of granular media. The essential part of the original method is briefly reviewed first.
The movement of each particle in a granular assembly is characterized by its translation and rotation. Let x be the configuration vector of the granular assembly that
consists of the Cartesian coordinates of particle centers and the rotations of particles, and F be the force and moment vector associated with the configuration vector.
In static problems, the forces acting on each particle are a set of contact forces and the body force. The normal component of a contact force can be directly
calculated as the product of the normal stiffness k n and the amount of overlapping between particles, while the tangential component should be calculated by summing
the past incremental forces that are determined by the product of the tangential stiffness k t and the tangential component of incremental relative displacement. The
tangential component should also be modified according to the Coulomb’s law characterized by the friction angle
Let FS and FB be the resultant force vector of contact forces and the body force vector. Then the static equilibrium equation is simply expressed as
FS +FB=0
(1)
Page 144
The granular element method adopts an iterative process to solve the above equation. In an incremental loading step starting from an equilibrated configuration x0, the
Nth approximation of the configuration is expressed as
(2)
where un is the displacement vector that consists of incremental translations and rotations in the nth iteration step. The resultant force vector FS changes in the (N+1)th

iteration step as
(3)
where K is the stiffness matrix. The stiffness matrix gives the incremental relationship between x and FS at the current state, and it is defined by assembling the normal
and tangential contact stiffnesses with suitable coordinate transformations between local systems at contact points and the global system.
The GEM algorithm, which adopts relaxation procedure, does not use the full stiffness matrix. Instead, it uses a simplified stiffness matrix; an assembly of a set of
particle stiffness matrices with 6×6 components that we can derive by letting neighboring particles fixed. As the summation of FB and is expected to be zero
from Equation 1, we get the following iteration equation for static problems.
(4)
It should be noted that the unbalanced forces on the righthand side of the above equation is calculated in terms of total forces instead of incremental forces. It may help
to avoid an accumulation error. The iteration process is repeated until the absolute value of vector quantity on the righthand side of Equation 4 becomes sufficiently
small.
2.2 Dynamic algorithm
The dynamic GEM algorithm is basically the same as the static one. Dynamic forces added to the equilibrium are the inertia and damping forces.
(5)
(6)
The inertia vector FI consists of the inertia forces and rotational inertias. The mass matrix M is diagonal and the components are the masses and inertia moments of
particles. The damping force vector FD consists of resultant forces and moments of a set of damping forces at contact points. The physical model of damping
mechanism in GEM is the normal and tangential dashpots that are placed in parallel with the normal and tangential springs at each contact point. Thus the procedure to
get the damping matrix C in terms of the damping coefficients ηn and ηt is similar to that of the stiffness matrix K.
Adding these force vectors to Equation 1, we have the following dynamic equilibrium equation.
FI+FD+FS +FB=0
(7)
By assuming that the acceleration changes linearly during a time step ∆t, we can express the increments of velocity and acceleration by means of the increment of
configuration ∆x as
(8)
(9)
where are the velocity and acceleration vectors at the beginning of the time step.
Then the Nth approximations of velocity and acceleration are given by
(10)
(11)
Then the inertia and damping force vectors for (N+1)th iteration step are expressed as
(12)
(13)
Therefore, we have the following iterative equation for dynamic problems.
(14)
The right hand side of the above equation represents the resultant of current total forces acting on particles. Thus the accumulation error may be restricted. It is
observed from this equation that the inertia term in the left hand side becomes dominant when the time
Page 145
interval becomes sufficiently small. In such extreme cases, we need no matrix inversion to solve Equation 14. After getting the incremental configuration ∆x for the time
step ∆t, we can calculate the new velocity and acceleration in terms of Equations 8 & 9, respectively. However, to avoid the integration error, we may adopt the direct
calculation of velocity and acceleration as follows:
(15)
(16)
These new velocity and acceleration are substituted into Equations 8 & 9 as respectively, for the next time step.
Boundaries are also controlled according to Equation 14. In the kinematic control, the incremental displacement and the velocity are prescribed for a boundary
element and we do not need to obtain the boundary displacement itself. The prescribed values of incremental displacement and velocity are utilized to calculate the
amounts of spring force and damping force appearing in the right hand side of Equation 14 for each neighboring particle.
On the other hand, in the kinetic control, we have to obtain the stiffness matrix in the left hand side of Equation 14 first for a boundary element. We have to take into
account the contribution of all the contacting inner elements. At the beginning of each time step, we eliminate the gap between prescribed stress and current stress by
moving the boundary element according to Equation 14. For strict stress control we do not need to include the acceleration and damping terms in the left hand side of
this equation for a boundary element.
3 SHEAR FLOW SIMULATIONS
The dynamic version of granular element method was applied to numerical experiments of shear flows under the nongravitational field. Figure 1 shows the initial
packing of granular assembly used in the numerical experiments. A pair of shearwalls are located in parallel at both sides of the granular assembly and each of them
consists of regular arrangements of particles with a uniform radius of 1. The distance between shearwalls is always kept constant. In the region between these walls,
439 particles with radii ranging from 0.5 to 1.5 are packed randomly.
We assume that the upper and lower boundaries are smooth planes, so that no damping and friction forces exist between each of these planes and contacting
particles. The lower plane is assumed to be fixed and the upper plane is controlled such that it moves vertically without rotation and that a constant confining stress is
maintained.
Figure 1. Initial state of granular assembly.
The magnitudes of confining stress assumed in the following simulation are 1, 10 and 100. In x direction, we assume periodic boundaries with a constant length of
periodicity. These boundary conditions simulate a shear cell that has two coaxial cylinders with different radii that can rotate in opposite directions. This type of shear
cell has been proposed by a research group of Professor J.T.Jenkins (Louge et al. 2001).
In Figure 1, the upper parts of shearwalls are not drawn for convenience. The initial particle concentration was 0.61. The assumed values of density of particle (ρ)
and other material constants are as follows:
Shear flows are activated in the granular assembly by the opposite movements of shearwalls as shown in Figure 1. The shear rates adopted in the following simulations
range from 1%/sec to 100%/sec. The period of time step is 10−3 sec for the slower two cases and 10−4 sec for the faster three cases.
Figure 2 shows a typical relationship between shear stress and shear strain. The shear rate in this example is 1%/sec, and the motion is quasistatic. As the particle
movements in the usual direction of dilatancy are restricted because of the structure of shear cell model, the rigidity for shear deformation is rather high and the shear
stress takes its peak value at rather small strain.
Figure 3 shows contact planes where the slipping takes place in the final state of the above simulation. Although the number of particles in shear cell is small, we
notice that slipping has a tendency to take place in a restricted zone.
Figures 4 & 5 show the changes in the height of granular assembly (H) normalized by the initial height (H0). In Figure 4 the confining pressure at the upper
Page 146
Figure 2. Typical stressstrain relationship (shear rate: 1%/sec, confining pressure: 1).
Figure 3. Slipping contact planes (shear rate: 1%/sec, shear strain: 5%, confining pressure: 1).
boundary is fixed to 1, and in Figure 5 the shear rate is fixed to 100%/sec. The maximum height becomes higher as the shear rate increases or the confining pressure
decreases. For the fastest shearing under the smallest confining pressure, the maximum height is 13 times higher than the initial height. From Figure 5, we notice the
definite role of the confining pressure in shear flows of granular assembly.
Figures 6–8 show snapshots of particle motions after transient responses for shearwall movements of three different shear rates.
Figure 4. Change in the height of granular assembly (confining pressure: 1).
Figure 5. Change in the height of granular assembly (shear rate: 100%/sec).
The quasistatic flow, where the constitutive relation can be assumed to be rate independent, is represented by Figure 6. As observed in Figure 4 no drastic change in
dilatancy takes place in such cases. Thus, some part of initial contacts may have chances to be preserved.
On the other hand, Figure 8 is a typical example of the inertiadominant flow, and the confining stress at the upper boundary is equilibrated by the momentum transfer
from colliding particles.
The above examples are two extremes of shear flows of granular assemblies. Between two of these, there may be many other types of transitional granular flows.
For example, Figure 7 may be looked upon as quasistatic. However, from Figure 4, the assembly has once experienced huge dilatancy and the initial fabric has been
totally destroyed.
Page 147
Figure 6. Snapshots of fully developed particle motions (shear rate: 1%/sec, shear strain: 5%).
4 CONCLUSION
The dynamic version of the granular element method, which was recently extended from the original static version, was applied to a study of the transitional behaviors of
granular assembly under dynamic shearing. The most important advantage of GEM is to secure equilibrated state through an iterative process. Simulations of shear cell
experiments under nongravitational field show the efficiency of this method. The quasistatic flow, where the constitutive relation is assumed to be rate independent, is
one of the extreme cases of shear flows. In this case, no drastic change in dilatancy takes place and the initial fabric is partially preserved. On the other hand, in the
inertiadominant flow the confining stress at the upper boundary is equilibrated
Page 148
by the momentum transfer from colliding particles and the initial fabric is totally destroyed.
To formulate universal constitutive laws for granular media, such transitional behaviors should be fully investigated.
REFERENCES
Cundall, P.A. & Strack, O.D.L. 1979. A discrete numerical model for granular assemblies. Géotechnique, 29(1): 47–65.
Kishino, Y. 1988. Disc model analysis of granular media. In M.Satake & J.T.Jenkins (eds), Micromechanics of Granular Materials: 143–152. Amsterdam: Elsevier.
Kishino, Y., Akaizawa, H. & Kaneko, K. 2001. On the plastic flow of granular materials. In Y.Kishino (ed.), Powders & Grains 2001:199–202. Rotterdam: Balkema.
Louge, M.Y., Xu, H. & Jenkins, J.T. 2001. Studies of gasparticle interactions in a microgravity flow cell. In Y.Kishino (ed.), Powders & Grains 2001:557–560.
Rotterdam: Balkema.
Zhang, Y. & Campbell, C.S. 1992. The interface between fluidlike and solidlike behaviour in twodimensional granular flows. J. Fluid Mech., 237:541–568.
Page 149

PFC3D modeling of interparticle percolation in caved rock under draw
M.E.Pierce
5809 679 3
ABSTRACT: A study was undertaken to examine the role of shearinduced interparticle percolation in preferential downward migration of fine
fragments in caved rock under draw. Following a demonstration of the phenomenon in a largescale simulation of draw in PFC3D, a series of simple
shear tests was carried out in PFC3D on dilute mixtures of coarse and fine spherical particles to replicate the results of physical testing carried out by
Bridgwater et al. (1978). Very similar percolation rates were observed, validating PFC3D as a tool for quantifying percolation rate. The test work was
extended in PFC3D to examine the impact of particle shape and shearing direction on percolation rate and to quantify percolation rates for materials
more closely resembling caved rock, which has a distribution of sizes, nonspherical shapes and increased density and modulus. The results indicate
that the percolation rate in materials with a Gaussian distribution of particle sizes is lower than in dilute mixtures and that it can be related to shear strain
and the ratio of particle diameter to mean particle diameter. A percolation rate equation was developed for caved rock based on these findings.
1 INTRODUCTION
The preferential movement of small particles through an assembly of large ones under the force of gravity is termed interparticle percolation. The phenomenon
contributes to the mixing and demixing of many types bulk solids during flow and transport (Wood 1986). Interparticle percolation may also contribute to preferential
downward migration of fines in cave mining operations. Caved rock generally contains a wide range of fragment sizes and so when a volume of such rock is drawn
from its base, the opportunity exists for finer fragments to move more rapidly downward and be drawn earlier than coarser fragments. Because fine and coarse
fragments may have significantly different ore grades, the preferential movement of fines can have a significant impact on caving operations. The main objectives of the
research presented in this paper were to explore the mechanisms controlling these relative movements, quantify the relative movements through controlled experiments
in PFC3D (Itasca 2003), and develop equations for incorporation of interparticle percolation mechanisms into REBOP, a gravity flow simulator developed by Itasca
for rapid analysis of draw in cave mining operations (Lorig & Cundall 2000, Pierce 2004).
2 MECHANISM OF PERCOLATION
For particles that are smaller than the minimum interstices between larger particles, percolation can occur spontaneously. Particles that are larger than the minimum
interstices may only percolate if the bed is disturbed in some way (e.g. through shearing). Shearinduced percolation has been studied extensively by Bridgwater et al.
(1978) who describe it as a releasecapture process in which the percolating particle travels downward through layers of larger bed particles in discrete steps. The
particle is released from a layer when the motion associated with shearing causes a gap of sufficient size to open up within the bed particles below it. After falling
through the gap, it may be captured again in a lower layer. If the particle is close to the size at which spontaneous percolation occurs, it may travel through several
layers before being captured again.
Bridgwater et al. carried out a large number of physical experiments to measure shearinduced percolation rates. They used a simple shear cell to deform a
rectangular bed of spheres to the shape of a parallelogram. By altering the direction of movement an unlimited amount of horizontal shearing could be applied. Individual
fine particles were placed in the top of the bed and the time required for them to pass to the base measured.
Percolation rates were examined by Bridgwater et al. for a number of different spherical particles with varying diameter, density and elastic modulus. Of the particle
properties, the ratio of percolating particle diameter to bed particle diameter (diameter ratio) was found to have the largest effect on percolation rate. Based on a large
number of tests, they were able
Page 150

to develop the following relationship between mean percolation distance, L, shear strain, γ, and diameter ratio, dp /db :
(1)

k 1 and k 2 are two arbitrary constants that were found to be equal to 20 and 8 respectively.
The term on the left is referred to as the dimensionless percolation rate. It is a measure of the most probable distance in bed particle diameters that an isolated
percolating particle is expected to drop when subjected to a shear strain of 1.0. The negative exponential relation with diameter ratio indicates that as the size of the fine
particle approaches that of the coarse particle, it takes an increasingly greater amount of shear strain for percolation to occur. Percolation rates were less sensitive to
the material properties and bed variables. In general, materials with greater density and lower modulus displayed a higher percolation rate. Rates were also found to
increase at strain rates below 0.4 s−1 and at lower levels of normal stresses applied to the top of the bed during shearing.
3 INTERPARTICLE PERCOLATION IN CAVED ROCK UNDER DRAW
The results of PFC3D simulations (Lorig 2000, Pierce et al. 2002) indicate that particles above a drawpoint are primarily subject to shear deformation when mobilized
(Fig. 1). Based on the work of Bridgwater et al. it was hypothesized that this shearing could be a primary driver of interparticle percolation in caving operations. In
order to test this hypothesis, a draw simulation was conducted on a similar PFC3D model in which a 0.75 m thick horizontal layer of fine particles was added 20
meters above the drawpoint The particle properties used to represent the main assembly of caved rock within PFC3D are outlined in Table 1. The fine particles had
similar properties to the main assembly except they were spherical with a uniform diameter equal to 20% of the mean diameter within the surrounding particle assembly.
The methodology used to simulate draw of caved rock is described in Pierce et al. (2002). A number of functions were added to the model to permit tracking of the
fine vs. coarse particles for this particular investigation.
The fines in the layer, and the coarser clumps surrounding them, were divided into equal width circumferential cells to permit tracking of average displacements as a
function of horizontal distance from the drawpoint. Figure 2 is a plan section through the model at the elevation of the fines layer showing the locations of the different
cells.
Figure 1. Quartersymmetry model used to simulate draw from a column of caved rock. The dimensions of the model are 19.25 m×19.25
m×42 m high and the drawpoint radius is 7 m. The altered positions of the colored layers (originally horizontal) reveal that the
material above the drawpoint is deformed in shear. (The front half of the model is hidden.)
Table 1. Base case particle properties used in simulation of caved rock under draw.
Shape 2ball clumps (1.5:1 aspect ratio)
Mean radius (of equivalentvolume sphere), Rmean 0.75 m
Standard deviation, σd 0.1 m
Cutoff radius, Rmin 0.01 m
Density, ρ
3333 kg/m3
Contact Young’s Modulus, E 0.125 GPa
Normal/shear stiffness, k /k
n s
1.0
Friction coefficient, μ 0.4
Initial porosity, n 0.4
Draw from the assembly was simulated by opening a circular hole at the bottom of the model and imposing a linear velocity profile on particles exiting the hole. Figure 3
illustrates the locations of the fines and clumps from each cell at several stages of draw (the remaining assembly has been hidden). The average radial and
circumferential deformation of the particle assembly is plotted on a grid to indicate the overall sense of movement resulting from draw. Examination
Page 151
Figure 2. Plan section through the fines layer with colors indicating the cells used to track movements of fines within the layer and the
coarse particles surrounding them. Cells are numbered 1 above the drawpoint through to 6 at the outer edge of the model. The
black line indicates the outer edge of the drawpoint 20 meters below.
of the particle positions indicates that in the early stages of draw, some of the fines percolate downward a short distance under slight disturbances to the assembly. As
draw continues, the layer begins to move more rapidly and deform in shear immediately above the drawpoint (near Cell #1). At this point, the average vertical
displacement of the fines increases relative to the coarse material. Percolation continues at a much slower rate farther from the drawpoint (near Cell #6) where
significant movement and shearing have yet to occur. The differences in vertical displacement and percolation are illustrated in Figure 4, which plots the average vertical
displacement of fine versus coarse particles in Cell #2 vs. Cell #6.
The results of the draw simulation indicate that percolation can occur in caved rock under draw and that is likely induced by the shearing that accompanies
movement toward the drawpoint. This suggests that if an equation of the form derived by Bridgwater et al. could be developed for caved rock under draw then it may
be possible to predict interparticle percolation outside of PFC3D within a draw simulator. The fact that both fragment size distribution and shear strain are traceable
quantities within REBOP makes this approach feasible.

3D
4 MEASUREMENT OF PERCOLATION RATES IN PFC
There are some key differences between the materials and bed conditions used in Bridgwater’s percolation
Figure 3. Position of coarse and fine particles within a horizontal layer after drawing 0.25%, 2.9% and 4.9% of the total assembly
volume. The remainder of coarse particles within the assembly has been hidden. The average circumferential and radial
deformation of the particle assembly is plotted as a grid. The black line represents the limit of all coarse particles that have
moved 1 m vertically.
rate experiments (horizontal shearing of isolated fines within a bed of uniformly sized spheres) and those existing within a column of caved rock under draw:
1. Caved rock exhibits a range of particle sizes with nonspherical shapes.
Page 152
Figure 4. Vertical displacement of coarse versus fine particles after 4.2% draw in cell 2 (close to the drawpoint center) and cell 6 (far
from the drawpoint center).
2. Caved rock has a greater proportion of finer fragments.
3. The shearing direction is vertical (or near vertical) in caved rock under draw rather than horizontal.
In order to obtain percolation rates under conditions more typical of cave draw, a number of simple shear experiments were carried out in PFC3D. As a first step, the
shear tests of Bridgwater et al. were replicated to validate PFC3D as a tool for quantifying percolation rates. A number of experiments were then carried out to
examine the effects of shearing direction and particle shape on the percolation rates predicted by Bridgwater. All subsequent tests were aimed at determining
percolation rates for a material with properties more closely simulating caved rock.
4.1 Replication of Bridgwater experiments
A simplified model of the experimental apparatus of Bridgwater et al. was constructed in PFC3D (Fig. 5). The model consisted of a rectangular box (354 mm× 355
mm×246 mm high) comprised of six planar frictionless walls filled with spherical particles of uniform size. The shearing conditions were equivalent to those used in the
standard tests of Bridgwater et al. A rotational velocity was applied to two opposing side walls to achieve simple shear at a rate of 0.4 s−1, the rotation being reversed
after a shear strain of 1.0 from vertical in order to achieve continuous shearing. The top wall was controlled with a servo function to maintain a normal stress of 1.4 kPa
on the bed of particles.
With the exception of contact modulus, the properties of the particles used in the simple shear model (Table 2) are equivalent to those of the acrylic resin spheres
most commonly used in Bridgwater’s experiments. In order to increase the timestep, the contact Young’s Modulus was set to a low value of 12.5 MPa.
Figure 5. PFC3D model of simple shear test used to examine percolation rates. The sidewalls are rotated back and forth in tandem to
achieve continuous shearing at a set rate. The percolation distances of fine particles are tracked as a function of total shear
strain applied to the assembly (visible in section here).
The fine particles have equivalent properties to the coarse particles except their diameter is smaller. A hysteretic damping model (Itasca 2002) was used to permit
realistic energy damping upon contact and
Page 153
Table 2. Base case particle properties used in simulation of acrylic resin spheres subject to simple shear.
Diameter, d b 11 mm
Density, ρ
1190 kg/m3
Contact Young’s Modulus, E 12.5 MPa
Normal/shear stiffness, k /k
n s
1.0
Friction coefficient, μ 0.25
Initial porosity, n 0.4
settlement of particles while maintaining realistic freefall accelerations. This is an important consideration since particles are expected to experience both settlement and
freefall within the discrete capturerelease process of shearinduced percolation. The damping coefficients were chosen to replicate the coefficient of restitution of 0.69
obtained from drop tests carried out on resin beads by Bridgwater et al. (1969).
The particle tracking procedure used in the PFC3D simple shear experiments differed from that of Bridgwater et al. in some respects. Rather than examine the
percolation of single particles one at a time passing through the entire bed, thirty fine particles were placed at the top of the coarse assembly and tracked
simultaneously. The particles were placed far enough apart so as to not interfere with one another. In addition, only particles in the upper half of the shear box at some
distance from the walls were tracked for percolation distance. This was meant to avoid interaction effects with the side and base walls, a problem identified in early
Bridgwater experiments and subsequently corrected through improvements to the shearing apparatus. These changes in tracking procedure allowed percolation rates to
be determined from fewer PFC3D experiments in a more simplified model apparatus.
Three main tests were carried out on three different particle diameter ratios: 0.25, 0.375 and 0.5. The percolation distance of each fine particle was recorded and
normalized to bed particle diameter and shear strain to permit direct comparison with Bridgwater’s dimensionless percolation rate. The match between PFC3D and the
experimental results is very good (Fig. 6) and indicates that PFC3D is capable of capturing the essential mechanisms involved in interparticle percolation.
4.2 Effect of shearing direction and particle shape
Additional shear tests were carried out in PFC3D to examine the effects of shearing direction and particle shape on the percolation rates predicted by Bridgwater.
Since vertical shearing is predicted to exist in a column of caved rock under draw, the shearing direction in the PFC3D model was rotated 90 degrees. Particle
diameter ratios of 0.25 and 0.375 were examined
Figure 6. Comparison of dimensionless percolation rate versus particle diameter ratio observed in the simple shear tests of Bridgwater et
al. and PFC3D simulations.
Figure 7. Dimensionless percolation rates from PFC3D for horizontal shearing, vertical shearing and nonspherical clumps versus
Bridgwater’s relation for horizontal shearing of spheres.
to determine the potential effect on percolation rate. As shown in Figure 7, the percolation rate increased only slightly, indicating that shearing direction (relative to
gravity) does not influence percolation rate significantly.
The vertical shear tests with particle diameter ratios of 0.25 and 0.375 were repeated with twoparticle clumps (1:1.5 aspect ratio), rather than spheres, to examine
the influence of particle shape on percolation rate. As shown in Figure 7, this did not have a significant influence on percolation rate. Several other shapes should be
examined before final conclusions are drawn regarding the effect of shape.
4.3 Examination of Gaussian distributed material representing caved rock
In order to obtain percolation rates for material more closely representing caved rock, a series of larger scale simple shear tests were carried out on twoparticle
Page 154

Figure 8. Size distributions examined in PFC3D shearinduced percolation rate tests.
Figure 9. Plot of tracked PFC3D particles after 400% shear strain. The surrounding particles have been hidden to show how finer
particles have percolated down preferentially.
clump assemblies with a distribution of sizes. In addition, the friction coefficient, density and contact Young’s Modulus of the clumps were increased to 0.4, 2700
kg/m3 and 0.125 GPa respectively. The normal stress applied to the top of the assembly was also increased from 1.4 kPa to 200 kPa based on measurements of stress
obtained from the PFC3D draw simulation reported earlier in this paper. The size of the shear box itself was increased to a 20 m cube. Shear tests were simulated on
four different twoparticle clump assemblies, each with a Gaussian distribution of particle diameters but varying standard deviations of 0.3, 0.4, 0.5 and 0.6 meters (Fig.
8).
For tracking purposes, the size distribution was divided into 12 classes with a width of half the standard deviation. All of the particles generally fell within classes 2
through 11 (−2.5 to +2.5 standard deviations from the mean). The mean percolation distance for each class was calculated at regular intervals of shearing.
Figure 9 shows the clumped assembly with a standard deviation of 0.5 meters after 400% shear strain.
Figure 10. Percolation distance versus shear strain for a range of particle sizes from an assembly with a Gaussian distribution of particle
diameter and a standard deviation of 0.5 meters.
Only the tracked clumps are shown in this plot to illustrate how some of the finer particles have percolated down toward the bottom of the shear box. The mean
percolation distances for particles less than the mean diameter are plotted as a function of shear strain in Figure 10. The plot indicates that particles with diameters more
than 0.5 standard deviations below the mean are undergoing preferential downward migration with the smallest particles percolating down as much as 6.5 meters.
Particles larger than this do not percolate down significant distances.
As discussed previously, Bridgwater observed that in monosized assemblies, percolation distance is a function of the particle diameter ratio (i.e. diameter of
percolating particle over bed particle diameter). It was hypothesized that a similar relation may hold for distributed assemblies, with the mean particle diameter in place
of bed particle diameter. To test this hypothesis, the mean percolation distances from all PFC3D shear box tests were normalized by shear strain and the mean
assembly diameter (to achieve Bridgwater’s dimensionless percolation rate) and plotted versus the particle diameter ratio (i.e. percolating particle diameter to mean
particle diameter). The results, shown in Figure 11, suggest that the hypothesis is valid, i.e. that the percolation rate for a particular sized fragment is controlled by the
ratio of its diameter to the mean diameter of the assembly. As with Bridgwater’s relation for isolated fines percolating through a bed of monosized spheres, the relation
between dimensionless percolation rate and particle diameter ratio developed here is negative exponential in form. The percolation rates are much lower, however,
since the wide distribution of sizes restricts the ability of fines to percolate down through the column.
The fact that percolation rate can be related to the mean particle diameter of a Gaussian distributed
Page 155
Figure 11. Dimensionless percolation rates for Gaussian distributed PFC3D material compared with rates from Bridgwater’s lab tests on
monosized spheres.
material makes sense since it represents the most common diameter. In reality, it is likely that the mode of the fragmentation distribution that will control percolation,
since this represents the most common size. For the Gaussian distributions examined here, these two values (the mean and the mode) are the same. Additional shear
tests could be carried out on nonsymmetrical size distributions to determine if the mode of the distribution is more generally valid as a controlling parameter.
4.4 Effect of strain rate
A base case strain rate of 0.4 s−1 was used in the preceding simple shear tests in order to replicate the standard mean rate used by Bridgwater in his lab tests.
Bridgwater et al. (1978) carried out a series of experiments to examine the effect of strain rate and found that percolation rates were faster when the strain rate was
reduced below 0.4 s−1. Strain rates of 0.1 s−1 and 0.2 s−1 were modeled in PFC3D to determine the influence on the percolation rates determined for a distributed
material. The results (plotted in Fig. 12) show that percolation rates increase significantly when the strain rate is reduced. This makes sense since the gaps that form
between coarse particles as a result of shearing will be open for more time as the shear strain rate is reduced. The probability of a fine particle making it through an
underlying gap would be expected to increase in response.
By fitting negative exponential curves to the data for each strain rate, some relations were established between the two constants in the percolation rate equation and
the strain rate, as shown in Figure 13. By extrapolating back to zero strain rate, we can obtain the constants for the percolation rate equation that represents an upper
limit to percolation rate. This curve is plotted in Figure 12.
Since the process of drawing material from a column of caved rock is normally discrete rather than

Figure 12. Dimensionless percolation rates for a Gaussian distributed PFC3D material as a function of strain rate.
Figure 13. Values of constants in dimensionless percolation rate equation as a function of shear strain rate.
continuous, the overall shear deformation resulting from multiple draw events over a specific time period is not suitable for estimating shear strain rate. Rather, one must
estimate the mean shear strain rate experienced during a single increment of draw and associated shear. Such strain rates can be estimated directly from PFC3D
simulations of intermittent draw from caved rock material of appropriate size distribution and density.
5 CONCLUSIONS
Shearinduced interparticle percolation has been identified as a significant mechanism leading to preferential migration of fines in caved rock under draw. Comparisons
between the results of physical simple shear tests on monosized bed particles and isolated percolating particles carried out by Bridgwater
Page 156
et al. (1978) and similar tests conducted in PFC3D show a very good match, indicating that PFC3D is able to capture the essential mechanisms involved in shear
induced percolation.
The work of Bridgwater et al. was extended to examine the impact of particle shape and shear direction on percolation rate and obtain percolation rates for materials
more closely resembling caved rock, which has a distribution of sizes, nonspherical shapes and increased density and modulus. The impact of particle shape (spherical
versus clumped) and direction of shearing was found to be negligible in tests on monosize assemblies with isolated percolating particles. PFC3D testing of simulated
nondilute mixtures more typical of caved rock indicate that percolation rates are lower for materials containing a Gaussian distribution of particle sizes and that
percolation rate can be related to shear strain and the ratio of particle diameter to mean particle diameter. A percolation rate equation was developed for caved rock
based on these findings and will be incorporated into REBOP, a gravity flow simulator developed by Itasca for analyzing draw in caving mines.
It was also demonstrated that percolation rate for such materials is sensitive to strain rate and that measures of instantaneous strain rate from fullscale flow
simulations should be used to determine the appropriate constants in the percolation rate equation developed from this research.
REFERENCES
Bridgwater, J., Sharpe, N.W. & Stocker, D.C. 1969. Particle Mixing by Percolation. Trans. Inst. Chem. Eng., 47:114.
Bridgwater, J., Cooke, M.H. & Scott, A.M. 1978. Interparticle Percolation: Equipment Development and Mean Percolation Velocities. Trans. Inst. Chem. Eng., 56:157.
Itasca Consulting Group, Inc. 2002. Hysteretic Damping Model Minneapolis: Itasca.
Lorig, L.J. & Cundall P.A. 2000. A rapid gravity flow simulator. Final Report, International Caving Study, (Ed: E T Brown). Brisbane: JKMRC and Itasca Consulting
Group, Inc.
Pierce, M.E. 2004. Final Report, International Caving Study II. Brisbane: JKMRC and Itasca Consulting Group, Inc.
Pierce, M.E., P.A.Cundall, G.J.van Hout and L.Lorig. 2003. PFC3D Modeling of Caved Rock Under Draw. In H.Konietzky (ed), Numerical Modeling in
Micromechanics via Particle Methods, Proceedings of the 1st International PFC Symposium, Gelsenkirchen, Germany, November 2002, pp. 211–217. Lisse:
Balkema.
Wood, P. 1986. Fundamentals of Bulk Solids Flow, IEA Coal Research Report, ICTIS/TR31, London, 1986.
Page 157
Flow behavior of cohesive and frictional fine powders
S.Luding
Particle Technology, DelftChemTech, TU Delft, Delft, Netherlands
R.Tykhoniuk & J.Tomas
Mechanische Verfahrenstechnik, OttovonGuerickeUniversität Magdeburg, Magdeburg, Germany
L.Heim, M.Kappl & H.J.Butt
MaxPlanckInstitut für Polymerforschung, Mainz, Germany
5809 679 3
ABSTRACT: Continuum mechanical models and appropriate measuring methods are applied to describe the flow behavior of cohesive powders and,
for example, for practical design of process apparatus as silos. Studies of the particle mechanics can provide a better physical “microscopic”
understanding of the essential constitutive functions of a powder “continuum”. With the discrete element method (DEM), a tool is available that allows
to consider the details of the contact and adhesion forces for each particle contact. Based on both experimental and empirical observations, a general
law is introduced, involving an irreversible (plastic), contact flattening (that leads to an increase of adhesion force for increasing flattening). The DEM
results show an elastic regime, yielding and steadystate flow of a twodimensional model powder.
1 INTRODUCTION
Given the particle geometry and their interaction forces, the flow behavior of powders under large deformations can be studied, using the discrete element method
(DEM), a convenient tool to gain insight into the evolution of, e.g., shear bands. Powders are typically inhomogeneous, disordered, and anisotropic on a “microscopic”
scale (Herrmann et al. 1998, Vermeer et al. 2001 and Kishino 2001), where the typical microscopic size is the particle size. Irregular random arrays respond to
deformations via inhomogeneous and anisotropic rearrangements and stressresponse. An initially isotropic contact network becomes anisotropic before the structure
of the network reaches its limit of stability, i.e., the yield stress. Before the peak, one has softening, and beyond weakening is obtained (Vermeer et al. 2001, Kishino
2001, Thornton & Antony 2000 and D’Addetta et al. 2002), which is typical for overconsolidated powders. Our work complements recent studies on shear band
formation in frictionalcohesive granular media (Thornton & Antony 2000, D’Addetta et al. 2002, Oda & Iwashita 2000, Kruyt & Rothenburg 2001 and Luding et al.
2001), for micro and macromodeling (Tomas 2001a and Tomas 2001b), and in various systems (Tykhoniuk & Tomas 2003, Luding & Herrmann 2001, Luding et
al. 2003, and Lätzel et al. 2003) for different materials.
Here, only spherical particles are used, as opposed to nonspherical objects like polygons (Tomas 2001), where roughness can also be mimicked by additional
torques (Oda 1993). The recently developed micromacro transition procedures (Oda & Iwashita 2000, Kruyt & Rothenburg 2001, Luding et al. 2001, Tomas
2001a, Tomas 2001b, Tykhoniuk & Tomas 2003, Luding & Herrmann 2001 and Luding et al. 2003) aim at a better understand of the macroscopic powder flow
behavior on microscopic foundations. Besides the experimental verification of the simulation results (Lätzel et al. 2003), the formulation of constitutive relations in the
framework of continuum theory is the great challenge. One promising material model, especially for sand is the hypoplastic theory (Tejchman & Wu 1993, Bauer &
Huang 1999, Bauer 1996, Tejchman 2002 and Nübel 2002), for which the material parameters can be determined experimentally, or from DEM simulations, as shown
in this study for some cases.
2 MODEL
2.1 DEM and the contact laws
The elementary units of granular materials, the “mesoscopic” particles, locally—at the contact point—deform under stress. The realistic modeling of the
Page 158
deformations inside is too much effort, so that we relate the interaction force only to the overlap δ of two particles, see Figure 1. As a further simplification, these two
particles interact only if they are in contact (short range forces), and the force between them is decomposed into a normal and a tangential part.
The normal force is, in the simplest case, a linear spring that takes care of repulsion, and a linear dashpot that accounts for dissipation during contact. Here, we
propose a new model that takes into account plastic contact deformation and cohesion (attractive forces). The force displacement scheme is shown in Figure 2.
Figure 1. Two particle contact with overlap δ.
Figure 2. Force displacement law for the DEM simulations.
For initial loading of the contact, the force increases linearly with stiffness k 1, which takes care of perfectplastic repulsion (Luding et al. 2001, Tomas 2001a and
Luding & Herrmann 2001). In addition, a linear dashpot accounts for dissipation during contact. Elasticity at the contact level is added by a spring, with a larger
stiffness, k 2, for un and reloading, so that the stiffness increases due to the irreversible, plastic contact deformation. Cohesion (or an attractive adhesion force)
between the contacts comes into the model by a “cohesive stiffness” k c, which allows for attractive forces (here negative) up to a minimal (maximal attractive) force
f min. Cast into an equation, the normal force on particle i is:

with the normal direction unit vector (as defined below) and the viscous dissipation parameter γn .
The tangential force involves dissipation due to Coulomb friction, but also some tangential elasticity that allows for stickslip behavior on the contact level (Thornton
& Antony 2000, Tomas 2001a, Tomas 2001b, Luding et al. 2003 and Lätzel et al. 2003). The implementation used here is slightly different from those presented in the
literature, in so far that an implementation in two and three dimensions is equally simple and also static and dynamic friction can be used with different values for the
respective coefficients.
In the static case, the tangential force is coupled to the normal force via Coulombs law, i.e. f t≤μsf n , where for the limit sliding case one has the dynamic friction with
f t=μ f n .
d
The dynamic and the static friction coefficients follow, in general, the relation μd ≤μs. However, for the following simulations, we will apply μ=μd =μs. The static case
requires an elastic spring in order to allow for a restoring force, i.e. a nonzero remaining tangential force in static equilibrium due to activated Coulomb friction.
If a contact exists with nonzero normal force, the tangential force is active too, and we project the tangential spring into the actual tangential plane. This is necessary,
since the frame of reference of the contact may have slightly rotated since the last timestep. with the total relative
Page 159
velocity of the contact surfaces of the two particles i and j:
(1)
Next, we calculate the tangential testforce as the sum of the tangential spring and a tangential viscous force (in analogy to the normal viscous force)
(2)
with the tangential spring stiffness k t and a tangential dissipation parameter γt. As long as is reached again, static friction is activated.
In the former, static case, the tangential spring is incremented is to be used in the next iteration and the tangential force as defined above is used.
In the latter, sliding case, the tangential spring is adjusted to a length, which is consistent with Coulombs condition
(3)
with the tangential unit vector,
Note that are not necessarily parallel in three dimensions. However, the mapping works always, rotating the new spring such that the direction of the frictional
force is unchanged and, at the same time, limiting the spring in length according to Coulombs law.
In short notation the tangential force on particle i reads for sticking contacts, as defined above.
Note that the tangential force described above is identical to the classical CundallStrack spring only in the limits μ=μd =μs and γt=0. The sequence of computations
and the definitions and mappings into the tangential direction, however, is new to our knowledge in so far that it can be easily generalized to three dimensions.
If all forces acting on a selected spherical particle (either from other particles, from boundaries or from external forces) are known, the problem is reduced to the
integration of Newton’s equations of motion for the translational and rotational degrees of freedom:
(4)
with the gravitational acceleration
2.2 Model system
The simulations with the discrete element model (Thornton & Antony 2000, D’Addetta et al. 2002, Oda & Iwashita 2000, Kruyt & Rothenburg 2001, Luding et al.
2001, Tomas 2001a and Tomas 2001b) use a twodimensional biaxial box, see Figure 3, where the left and bottom walls are fixed. Stress or straincontrolled
deformation is applied to the sideand topwalls, respectively. In a typical simulation, the top wall is slowly shifted downwards, while the right wall moves, controlled by
a constant stress px, responding on the forces exerted on it by the material in the box. The motion of the topwall follows a cosine function, in order to allow for a
smooth startup and finish of the motion so that shocks and inertia effects are reduced, however, the shape of the function is arbitrary as long as it is smooth.
2.3 Initial conditions
Initially, the particles are randomly distributed in a huge box, with rather low overall density. Then the box is compressed by defining an external pressure p, in order to
achieve an isotropic initial condition with kinetic energy much smaller than the potential energy stored in the contacts. Starting from this relaxed,
Figure 3. (Left) Schematic drawing of the model system. (Right) Position of the topwall as function of time for the straincontrolled
situation.
Page 160

isotropic configuration, the strain εzz is applied to the top wall and the response of the system is examined.
3 RESULTS
The system examined in the following contains N=1950 particles with radii randomly drawn from a homogeneous distribution with minimum 0.5 mm and maximum 1.5
mm. The friction coefficient used in the twodimensional simulations is μ=0.5. The total mass of the particles in the system is about 0.02 kg. If not explicitly mentioned,
the material stiffness parameters are k 2=105 N/m, k 1/k 2=1/2, and the contactviscosity is 0.1 kg/s. The eigenfrequency of the particle contact is hence typically 10−5 s
so that an integration timestep of 2∙10−7 s is used, in order to allow for a “safe” integration (Herrmann et al. 1998 and Luding & Herrmann 2001).
Figure 4. Volumetric strain (top) and stresses (bottom) for a biaxial box simulation with side pressure p=500.
3.1 Biaxial test with varying cohesion
In Figure 4 (top), the volume change of a typical simulation with cohesion, but without friction, μ=0, shows first compression, then dilatancy. The stronger the cohesion,
the more the material can be compressed, while the dilatancy is almost unaffected by cohesion. At the same time, the stress response (the index zz denotes the vertical
stress) becomes more and more anisotropic, i.e., the vertical stress increases until it reaches a maximum, while the horizontal stress remains constant. After the peak,
softening behavior and large fluctuations are evidenced. The peak stress value increases with the microscopic cohesion or adhesion force fmin .
From the macroscopic point of view, the flow behavior of the system can be examined by plotting Mohrcircles for different confining pressures (leftmost point on
the circle) and for the maximum stress (rightmost point), see Figure 5. The tangent to these circles can be seen as the flow function for the peak stress, which
corresponds to a yield locus of an overconsolidated packing. It is linear for the examined parameters with a slope slightly larger than expected from the microscopic
friction at the contacts alone. If no microscopic friction is active, a friction angle of about 13° is obtained (0° expected).
Figure 5. (top) Mohr circle representation of the flow function at maximum stress for cohesion and no friction (bottom). Macroscopic
cohesion as function of the microscopic cohesive strength. The points are taken from the flow functions; the line corresponds to
the analytical expression for the maximal attractive force as function of k c/k 2.
Page 161
3.2 Biaxial test with varying pressure
In Figure 6 (top), the volume change of a typical simulation shows first compression, then dilatancy, and eventually a very weak change at very large deformations, up
to 20 percent. At the same time, the stress response, in Figure 6 (bottom) (where the indices xx and zz denote horizontal and vertical stresses, respectively), shows
elastic, softening, and critical state flow behavior.
First, the vertical stress increases linearly; then the slope gradually decreases (softening), until the stress reaches its maximum (peak yield stress). After the peak,
further softening/weakening behavior (with negative slope) is followed by a constant, strongly fluctuating stress for larger deformations.
Figure 6. Volumetric strain (top) and stresses (bottom) during large deformations, both plotted against vertical strain, for different side
pressure, as indicated in the inset. The peak yield stress is marked by arrows.
It is possible to also examine the flow behavior of the system by plotting Mohrcircles for the maximum stress (rightmost point on the circle) for different confining
pressures (leftmost point), see Figure 7. The eigendirections of the system are parallel to the walls, because there is no friction active between particles and walls, so
that the left and rightmost points on the circles are indeed corresponding to the wall stresses; note that in an arbitrary geometry, it is not necessarily that simple. The
tangent to the circles (slope 0.588) can be seen as the flow function for peak stress. The corresponding friction angle is about 30.5° (26.6° expected from micro
friction). It is linear for the examined parameters, with its slope only slightly larger than expected from the microscopic friction at the contacts alone. Since we have not
used cohesive forces, the macroscopic cohesion c is nonexistent, i.e., the flow function hits the origin.
3.3 Macroscopic material parameters
From the simulation data presented in Figures 4–7, it is possible to obtain the following material parameters, as based on an isotropy assumption:
1. The initial slope (−0.59) of the volumetric strain allows to determine the Poisson ratio.
2. The slope of the volumetric strain in the dilatancy regime (+0.19 without friction, +0.80 with friction) is related to the dilatancy angle.
3. The initial slope of the stress is related to the bulk modulus (results not discussed further).
4. The peak (yield) stress is related to the flow function of the material. Interesting is a macroscopic friction angle that is always larger than the microscopic one.
Figure 7. Mohr circle representation of the flow function at peak stress, see the arrows in Figure 6 (bottom).
Page 162
Figure 8. Pore number plotted against side stress (top), and deviatoric stress ratio plotted against pore number (bottom)—for different
confining pressures.
3.4 Some material parameters for hypoplasticity
Some of the material parameters involved in a hypoplastic material theory (Tejchman & Wu 1993, Bauer & Huang 1999, Tejchman 2002 and Nübel 2002) can also
be extracted from the simulation data. An essential ingredient of the theory is the functional behavior of the pore number:
(5)
as a function of the pressure. The empirical model parameters for this function (based on experimental findings) involve the pore number at vanishing stress, the
socalled granular hardness, hs, and an empirical power n.
The function in Equation (5) is astonishingly close to the fitfunction (6) for the initial and the critical state porenumber envelope, see Figure 8:
e(e0, n)=e0−[p/hs] n
(6)
where the parameters e0 and n can be read off from the inset, and the granular hardness was set equal to the spring stiffness hs=k 2=105 in the DEM contact model.
Equations (5) & (6) can be related to each other via a series expansion in the small variable p/hs.
The representation of the deviatoric stress ratio:
(7)
in Figure 8 (bottom) is another way to extract the macroscopic friction angle. For peak stress, the simulations are almost in agreement with the microscopic friction
coefficient μ=0.5, whereas for the critical state flow stressvalue (besides strong scatter due to fluctuations), the macroscopic friction angle decreases with increasing
pressure. For the largest pressure used here, the friction coefficient is smaller than μcs=0.4.
4 SUMMARY AND CONCLUSIONS
In summary, a set of DEM simulations was presented, and several macroscopic material parameters like, e.g., the friction angle, were extracted from the simulation
data with cohesion (no friction) and with friction (no cohesion). Also the behavior of density (porenumber) and friction angle as function of the confining pressure were
discussed and related to a hypoplastic material law (Tejchman & Wu 1993, Bauer & Huang 1999, Bauer 1996, Tejchman 2002 and Nübel 2002).
The present results are a first step of a micro modeling approach for cohesive frictional powders. Further material parameters have to be identified, and also the
dependence of cohesion has to be examined more closely, not only for frictionless (Tykhoniuk & Tomas 2003, Luding & Herrmann 2001 and Luding et al. 2003), but
also for frictional materials.
Also the role of particle rotations is an open issue, as related to micropolar constitutive models. In both simulation and experiment, rotations are active in the shear
band—like in micropolar hypoplastic material models, where the rotational degree of freedom is activated in the shear band too. The corresponding parameter
identification and the micromacrotransition is another task for the future, like the implementation and simulation of experimentally determined forcelaws (Kappl &
Butt 2002) in threedimensional systems.
Page 163
ACKNOWLEDGEMENTS
This work was funded by the Deutsche Forschungsgemeinschaft (DFG) in the framework of the researchgroup: “Verhalten Granularer Medien”; we acknowledge
helpful and inspiring discussions with M.K. Müller, K.Nübel, R.Pitchumani, and J.Tejchman.
REFERENCES
Bauer, E. 1996. Calibration of a comprehensive hypoplastic model for granular materials, Soils and Foundations 36: 13–26.
Bauer, E. & Huang, W., 1999. Numerical study of polar effects in shear zones. In G.N.Pande, S.Pietruszczak, & H.F.Schweiger (eds.) Numerical Models in
Geomechanics, Rotterdam: Balkema, pp. 133–141.
D’Addetta, G.A., Kun, F. & Ramm, E. 2002. On the application of a discrete model to the fracture process of cohesive granular materials, Granular Matter 4(2):77–90.
Herrmann, H.J., Hovi, J.P. & Luding, S. eds. 1998. Physics of dry granular media, NATO ASI Series E 350 Dordrecht: Kluwer Academic Publishers.
Kappl, M. & Butt, H.J. 2002. The colloidal probe technique and its application to adhesion force measurements, Particle & Part. Sys. Charact. 19(3):129–143.
Kishino, Y. ed. 2001. Powders & Grains 2001, Rotterdam: Balkema.
Kruyt, N.P. & Rothenburg, L. 2001. Statistics of the elastic behavior of granular materials. Int. J. of Solids and Structures 38:4879–4899.
Lätzel, M., Luding, S., Herrmann, H.J., Howell, D.W. & Behringer, R.P. 2003. Comparing Simulation and Experiment of a 2D Granular Couette Shear Device, Eur. Phys.
J. E11:325–333.
Luding, S. & Herrmann, H.J. 2001. MicroMacro Transition for Cohesive Granular Media. In S.Diebels (ed.), Zur Beschreibung komplexen Materialverhaltens,
Stuttgart: Institut für Mechanik, pp. 121–134.
Luding, S., Lätzel, M., Volk, W., Diebels, S. & Herrmann, H.J. 2001. From discrete element simulations to a continuum model, Comp. Meth. Appl. Mech. Engng. 191:21–
28.
Luding, S., Tykhoniuk, R. & Tomas, J. 2003. Anisotropic material behavior in dense, cohesive powders, Chem. Eng. Tech. 26(12):1229–1232.
Nübel, K. 2002. Experimental and Numerical Investigation of shear localization in granular material, PhD Thesis, Karlsruhe.
Oda, M. 1993. Microfabric and couple stress in shear bands of granular materials. In C.Thornton (ed.) Powders and Grains, Rotterdam: Balkema, pp. 161–167.
Oda, M. & Iwashita, K. 2000. Study on couple stress and shear band development in granular media based on numerical simulation analyses, Int. J. of Engineering
Science 38: 1713–1740.
Tejchman, J. & Wu, W. 1993. Numerical study on patterning of shear bands in a Cosserat continuum, Acta Mechanica 99:61–74.
Tejchman, J. 2002. Patterns of shear zones in granular materials within a polar hypoplastic continuum. Acta Mechanica 155:71–95.
Thornton C. & Antony, S.J. 2000. Quasistatic deformation of a soft particle system, Powder Technology 109(1–3): 179–191.
Tomas, J. 2001a. Assessment of mechanical properties of cohesive particulate solids—part 1: particle contact constitutive model, Particulate Sci. Technol. 19:95–110.
Tomas, J. 2001b. Assessment of mechanical properties of cohesive particulate solids—part 2: powder flow criteria, Particulate Sci. Technol. 19:111–129.
Tykhoniuk, R. & Tomas, J. 2003. Simulation der Scherdynamik von kohäsiven Pulvern, Chem. Ing. Techn. 75, (in press).
Vermeer, P.A., Diebels, S., Ehlers, W., Herrmann, H.J., Luding, S. & Ramm, E., eds., 2001. Continuous and Discontinuous Modelling of Cohesive Frictional Materials,
Lecture Notes in Physics 568, Berlin: Springer.
Page 164
Page 165
Mixtures of rolling and nonrolling circular disks studied at high shear strains
T.Vietor
Geoforschungszentrum, Potsdam, Germany
5809 679 3
ABSTRACT: Grain rolling has been suggested to be an important mechanism for weakening of fault rocks at high strains. This study presents
numerical experiments on fault rock analogs i.e. mixtures of twodimensional rolling and nonrolling disks using a shear box with periodic boundaries.
In 18 experiments three different particle frictions and six different fractions of rolling disks were used. The tests indicate that grainrolling is a first
order controlling factor for the micromechanics and the bulk shear strength evolution of particle aggregates with high and intermediate particle friction.
These aggregates show large variations of their peak and residual strength with the fraction of nonrolling particles. In low concentrations rolling
particles interspersed in a nonrolling matrix obstruct the formation of long stresssupporting particle columns and thus drastically reduce the strength of
nonrolling aggregates. In higher concentrations rolling particles decouple the opposite shear zone walls and stabilize deformation localization. If the
rolling disks constitute more than half of the particles, the aggregates evolve towards a rollerbearing configuration that progressively reduces its bulk
friction by internal restructuring in order to reduce the slip along particle contacts.
1 INTRODUCTION
Seismically active faults commonly contain crushed rock material which has distinctly different mechanical and physical properties than the surrounding rock. These fault
gouges play an important role for the strength and the sliding stability of faults (Scholz 2002) and hence influence the frequency and intensity of earthquakes. Numerical
studies of simulated fault rocks using the Distinct Element Method (DEM) have improved the understanding of their kinematic and dynamic evolution (Morgan 1999,
Morgan & Boettcher 1999, Mora & Place 1999). However, some discrepancies between numerical and physical experiments remain. For example, particle rolling
plays an important role for fault weakening in numerical studies (Morgan & Boettcher 1999, Mora & Place, 1999) while physical experiments with simulated gouge
material indicate that particle rolling is uncommon at higher confining pressures where grain fracturing occurs (Mair & Marone 1999, 2002). This suggests that rolling
circular disks may be inappropriate to represent fault gouges.
This study investigates the influence of particle rolling on the kinematics and dynamics of numerically simulated aggregates of circular disks at high strains. For that
purpose a densely packed and noncohesive assemblage of disks is deformed in a shear box with periodic boundaries where the maximum shear strain is only limited
by computation time. In order to detect the changes in bulk friction and micromechanics induced by particle rolling we investigate three aggregates with different particle
frictions (μd =0.2, 0.5, 0.8) and vary the fraction of rolling particles in six steps from 0 to 1.
2 METHOD
2.1 Computational basis
The numerical experiments presented here apply the twodimensional version of the Particle Flow Code PFC2D (Itasca 2002). The details of the DEM technique as
implemented in PFC2D have been described by Cundall & Strack (1979) and in Itasca (2002) and only the general concept and the contact laws used in the
subsequent experiments will be outlined here. In the DEM implementation of PFC2D geological materials are simulated by aggregates of circular disks. The particles
are independent and interact via forcedisplacement laws installed at their mutual contacts. These contacts may form and break repeatedly during the deformation of the
aggregate and the strain of the
Page 166
bulk material is not limited by the modeling procedure. During the experiment the interparticle forces and the particle positions are updated in a twostage explicit time
stepping scheme. First, the contact forces are calculated from the actual positions of the disks and summed up for every element. Second, the total forces (including
body forces) and moments are integrated using a fixed timestep to calculate incremental displacements and rotations. These are added to the positions of the particles
and the calculation process is restarted.
The constitutive law at the contacts used in this study consists of a stiffness model and a slip model. The stiffness model provides an elastic relation between the
relative displacement and the contact force. The elastic contact force from the stiffness model can be decomposed into normal and shear components which are parallel
and orthogonal to the line joining the centers of two disks in contact. The normal force Fn is calculated as the product of the total normal displacement Un (the overlap
of the disks) and the normal stiffness kn of the contact. The shear force is computed incrementally. Shear motions at a newly formed contact give rise to increments of
shear force ∆Fs which are progressively added up to the present shear force value. The shear force increments are calculated as the product of the displacement
increments ∆Us and the shear stiffness ks . The slip model allows frictional slip by limiting the shear force to Fs max=μd ×Fn , where μd is the friction coefficient of the
disks.
2.2 Experimental setup and testing program
In order to study the evolution of highstrain shear zones in aggregates of circular disks, a twodimensional shear box with periodic boundaries was constructed (Fig.
1a). The periodic boundaries were implemented by copying particles to the opposite side of the box whenever they touched the periodic boundaries. Details of the
procedure can be found in the work of Jensen et al. (1999). Contact forces acting on one of the boundary particles are copied to the partner particle at the opposite
end of the box in every timestep.
The construction of the samples started with the creation of four walls which served as a temporary container for the particles. Approximately 1250 circular disks
were generated within this box at random locations, without initial overlap and without surface friction. Compaction was performed by stepwise enlargement of the
disks. After each enlargement step the repulsive forces between overlapping disks lead to particle rearrangements during a subsequent equilibration phase. After five to
ten enlargement steps the particles had formed a dense aggregate supported exclusively by normal forces. Then, the lateral walls were removed, periodic boundary
conditions were activated and the aggregate was compacted to 150 MPa confining pressure by moving the top and
Figure 1. a) Shear box with periodic boundaries. In order to avoid crystallization effects, the disks have a linear radius distribution (by
number) between 1 and 3 cm. They are 1 cm thick and their density is 2800 kg/m2. The normal stiffness and shear stiffness of
the contacts are 10 GPa. The tests are run at 150 MPa constant confining pressure and constant shear velocity of 5 cm/s.
Friction coefficients of the disks are either 0.2, 0.5 or 0.8 (“low, intermediate and high friction”). The gray disks have fixed x
velocities and drive the deformation. Tests were terminated at a shear strain γ=dx/h0=3.5. The experiment shown uses
intermediate friction particles (μ d=0.5) and 68% of the grains are not allowed to rotate (xf=0.68). The black disks were initially
arranged in stripes normal to the shear direction and indicate localized deformation in the upper half of the shear box. b)
Example result for the experiment shown in a) that illustrates the determination of bulk peak friction μ bmax and residual friction
μ bres (determined as maximum friction between 2.0<γ<2.5).
bottom walls. This mechanism also kept the confining pressure constant (within<3%) during the shearing experiment. At this stage the rotational degree of freedom of a
fraction xf of the disks varying from 0.0 to 1.0 was locked. The nonrotating disks were chosen randomly from the assemblage. Finally, the top and bottom rows of the
balls were assigned fixed velocities to start the experiment at constant shear strain rate and constant confining pressure.
During the tests the bulk shear strength and the dilatation of the aggregate were recorded in steps of 0.05% shear strain. The complete model was stored for post
processing in steps of 0.5% shear strain. The dimensions and the mechanical properties of the
Page 167
particles, as well as the dimensions of the shear box are depicted in Figure 1a. Figure 1b shows an example strength evolution of one of the aggregates and the
determination of peak and residual strength in terms of bulk friction coefficient μb .
The experiments presented subsequently will be used to demonstrate the influence of the particle friction and the fraction of nonrotating particles on the deformation
of twodimensional circular disk aggregates. The aggregates have either low, intermediate or high particle friction (μd =0.2, 0.5 and 0.8). For each of the friction values
we varied the fraction of nonrotating disks xf in six discrete steps (1.0, 0.86, 0.69, 0.45, 0.18, 0.0). Experiments that study the influence of the particlesize
distribution and the confining pressure are underway and will be presented at a later stage.
3 RESULTS
3.1 Bulk shear strength
The Figures 2 and 3 summarize the peak and residual strengths of the low, the intermediate and the highfriction series in terms of the bulk friction μb . In the three
series the peak strength μbmax and residual strength μbres of the aggregates increase with particle friction and with the fraction of nonrotating particles. Hence, the
lowest bulk strength values were observed when both the particle friction and the fraction of nonrotating particles were low (μd =0.2, xf=0.0). On the other hand, the
highest values for peak and residual strength were observed in the test with high particle friction (μd =0.8) and no rotating disks (xf=1.0).
The increase of μbmax and μbres with the fraction of nonrolling particles is strongest in the highfriction series. From xf=0 to xf=1.0 the bulk peak strength μbmax of
the highfriction series increases by a factor of 3.7 and the bulk residual strength μbres increases by a factor of 6.8. Both the increase of μbmax and μbres are nonlinear
for the highfriction series and steepen for a nonrolling fraction xf>0.5. The strongest increase is observed from xf=0.86 to xf=1.0 where the peak strength μbmax rises
from 1.26 to 2.11 and the residual strength from 1.08 to 2.43. This indicates that a few rolling particles substantially weaken the highfriction nonrolling aggregate.
The intermediate and lowfriction series show the same overall trends with increasing xf, however, the nonlinear increase of the bulk friction is weaker. The
intermediate and lowfriction aggregates increase their bulk peak strength (bulk residual strength) only by a factor of 2.6 and 1.7 (3.6 and 1.6), respectively. This
suggests that the influence of the nonrolling particle fraction decreases with decreasing particle friction.
Figure 2. Peak friction μ bmax. Measured in all but one test as maximum friction. For μ b=0.8 and xf=1.0: peak friction was defined as the

maximum friction in 0.0<γ<0.1).

Figure 3. Residual friction μ bres, measured as the maximum bulk friction during 2.0<γ< 2.5. Symbols as in Figure 2.
The different sensitivity of the three aggregates to the varying fraction of rolling particles xf is also illustrated by the strength of the nonrolling and allrolling aggregates.
Among the nonrolling aggregates (xf=1.0) the peak (residual) strength of the highfriction assembly is higher by a factor of 3.2 (4.1) than the values for the lowfriction
aggregates. In contrast, in the allrolling particle aggregates (xf=0.0) the influence of particle friction on the bulk strength is much smaller. The peak strength of the high
friction aggregate is only 1.6 times stronger than the lowfriction aggregate. And finally the residual strength of the allrolling aggregates appears to be completely
independent of the particle friction. In the residual shear phase the highfriction disks are even weaker than the lowfriction disks. Apparently, particle rolling decouples
bulk shear strength of the allrolling aggregate and surface friction of the particles.
As a measure for postfailure strain softening the relative residual strength κ=μbres/μbmax was calculated. Figure 4 shows κ for the different values of
Page 168
Table 1. Experimental results for the low, high and intermediatefriction series (μ d: particle friction, xf: fraction of nonrolling disks, μ bmax: bulk peak

friction, μ bres: bulk redisual friction, κ=μ bres/μ bmax.
μd xf μbmax μbre κ

Low friction 0.2 0.00 0.38 0.37 0.96
0.2 0.18 0.40 0.35 0.86
0.2 0.45 0.45 0.38 0.85
0.2 0.68 0.49 0.45 0.91
0.2 0.86 0.57 0.52 0.91
0.2 1.00 0.64 0.59 0.92
Int. friction 0.5 0.00 0.49 0.35 0.71
0.5 0.18 0.57 0.40 0.70
0.5 0.45 0.69 0.53 0.77
0.5 0.68 0.83 0.68 0.82
0.5 0.86 1.05 0.91 0.87
0.5 1.00 1.26 1.25 0.99
High friction 0.8 0.00 0.57 0.36 0.64
0.8 0.18 0.68 0.42 0.62
0.8 0.45 0.84 0.48 0.57
0.8 0.68 1.04 0.65 0.63
0.8 0.86 1.26 1.08 0.86
0.8 1.00 2.11 2.43 1.16
Figure 4. Relative residual strength κ=μ bres/μ bmax. Symbols as in Figure 2. Filling indicates localization behavior. Open symbols: strong

localization. Halffilled: weak localization. Filled: homogenously distributed shear. Insets show examples for the distribution of
shear in the aggregates with strong, weak and no localization at the end of the experiment (γ=3.5).
xf. For the lowfriction particle assembly (μd =0.2), κ does not drop below 0.8 indicating a low tendency to produce localized deformation bands. In the intermediate
iriction series (μd =0.5), κ is approximately 0.7 for xf=0.0 and 0.18. For xf>0.3 κ increases slowly to 1.0 at xf=1.0.
The highfriction series (μd =0.8) shows the strongest changes in strain softening behavior with the variation of xf. In the range 0.0<xf<0.5 we observe values for κ
around 0.6 with the minimum of κ=0.55 at xf=0.45. For xf>0.8 κ increases strongly reaching κ=1.15 at xf=1.0. The postfailure shearstrength increase indicated by
κ>1.0 in the highfriction nonrolling test is unique among the experiments.
Strain softening is generally thought to be the reason for deformation localization. If the relative residual strength κ can be interpreted in that sense, the aggregates
with high surface friction and a fraction of nonrotating particles around 0.5 should have the strongest tendency to localize the shear deformation. The localization
behavior of the samples was checked using average shear displacement profiles normal to the shear direction. Inspection of the profiles reveals three categories of
experiments with strong, weak and no localization (see insets in Fig. 4 and symbol filling). The tests showing localization (open symbols) are grouped in the lower half of
the diagram indicating that localization indeed is favored by strain softening in the aggregates. Weak or no localization is observed in all experiments with low particle
friction. Experiments with high and intermediate particle friction generally show strong localization, but only if both rotating and nonrotating particles are present.
3.2 Grainscale mechanics
Peak and residual shear strength in the intermediate and highfriction series are strongly dependent on the fraction of rolling particles. As the bulk strength of the DE
model is controlled by the interactions at grain scale, the correlation suggests that the variation of the fraction of rolling particles induces changes in the micromechanics
of the aggregates.
Figure 5 shows the contact force networks of five example experiments during prefailure loading (shear strain γ=3%) and in the postfailure and residual shear
phases (γ=50% and γ=250%). Figure 5a, which depicts the experiment with high friction and norotating particles (xf=1.0, μd =0.8) shows that during prefailure, post
failure and in residual shear, long columns of highly stressed particles are formed. The columns are stable even with minimal lateral support by adjacent particles. This is
also indicated by the sharpening of the contact orientation distribution around the principle stress direction at high strains (insets in Fig. 5a) and the simultaneous increase
of the average shear strength. The stability of the particle columns results from the inability of the individual particles to rotate. Failure of the stresssupporting particle
columns occurs only by frictional slip along contacts. Moreover, the increase of the average bulk shear strength in the highfriction nonrotation test indicates that the
internal restructuring into particle columns becomes increasingly effective during the test. The restructuring process allows stresses to become
Page 169
Figure 5. Contact force network in five experiments discussed in the text. Insets show histograms of contact orientations in 10° bins. Left
to right: 3% (prefailure), 50% (postfailure), 250% shear strain (residual shear). a: high particle friction (μ d=0.8), fraction of
nonrotating particles xf=1.0; b: μ d=0.8, xf=0.0; c: μ d=0.2, xf=1.0; d: μ d=0.8, xf=0.86; e: μ d=0.8, xf=0.45.
concentrated along only a few particle columns (Fig. 5a, right). This results in a sharp division of the particles into a nonstressed matrix and a stresssupporting
network. The strong dilatation of the experiment and the formation of large voids are also a result of the long stable particle bridges.
In contrast, the aggregates with either low particle friction (Fig. 5b: xf=1.0, μd =0.2) or a high amount of rolling particles particle (Fig. 5c: xf=0.0, μd =0.8) do not
form long stresssupporting particle columns. Here, the forces between the moving shear zone margins are transmitted by a stressed particle network (Fig. 5b and 5c,
and insets).
Both the rolling highfriction and the nonrolling lowfriction experiments show lower strength values than the highfriction nonrolling aggregate. Hence, we
tentatively conclude that the strong aggregates are able to form particle columns due to the higher particle friction and a low content of rotating particles. The addition of
rolling particles or the decrease of the particle friction obstruct the formation of the particle columns which results in weaker aggregates.
This view is confirmed by the test with only a few rolling particles and high particle friction. Figure 5d shows the contact forces of the high particlefriction experiment
with 14% of rolling disks (xf=0.86). Both the branching structures in the contact force image in Figure 5d and the less sharp distribution of the contact orientations (Fig.
5d, inset) indicate that the interspersed rolling particles obstruct the restrengthening within the failed region. The experiment shows a decrease of the residual strength
(measured at 2.0<γ<2.5) by 60% with respect to the experiment with xf=1.0.
A further increase of the rollingparticle fraction (xf<0.68) leads to the localization of the permanent shear deformation in a shear band. In the cases with xf=0.45
(Fig. 5e) and xf=0.32 residual strength factors κ lower than 0.55 indicate considerable strain softening. This agrees with the observation that in the residual shear phase
of these two tests the slope of the dilatation curve is considerably smaller than in the tests with lower rolling fractions. Fewer particles per shear strain increment are
broken out of the shear zone walls and incorporated into the dilated shear band. Apparently an increased fraction of rolling disks in these tests enables a lowstress
motion between the opposite sides of the shear band. This is supported by the particle rotations in the high friction experiments with xf=0.68 and 0.45 which show that
particle rolling is most active within the shear band. Hence, localization of deformation within a shear zone is stabilized if sufficient rolling particles are present that
prevent the formation of interlocked particle groups or columns within the weakened region.
A further reduction of the fraction of nonrotating disks to xf=0.18 and xf=0 again reduces the tendency for strain localization. Relative residual strength κ of the two
experiments is approximately 0.62 and almost as low as with xf=0.68. However, despite the substantial strain weakening the initial shear zones are unstable. From initial
failure to high strains we observe that the fraction of slipping contacts continuously decreases. In contrast, rolling of particles is observed throughout the whole assembly
in the postfailure and residual shear phases. This indicates that the aggregate progressively optimizes its internal structure for particle rolling in order to reduce frictional
slip. Finally, the shear deformation is absorbed by a maximum of rotation and a minimum of frictional slip. Apparently, high amounts of rolling particles are able to
substantially weaken the shear zones by rolling around each other.
The results of the intermediate friction series are very similar to the high friction series. Bulk stresses are lower, however, the localization characteristics are similar to
the highfriction series. Only mixtures of rolling and nonrolling particles exhibit strain localization, while aggregates containing either only rolling or nonrolling particles
do not.
Page 170
This differs from the experiments from the low friction series which never exhibit localization of deformation. Here, the relative residual strength κ is generally higher
than in the intermediate and highfriction series. Hence, the lowfriction aggregates appear to be unable to produce shear zones which are substantially weaker than the
unsheared aggregate.
4 DISCUSSION AND CONCLUSIONS
Previous numerical experiments as well as physical analog modeling have shown that simulated fault rocks consisting only of circular grains are unable to reproduce
natural fault strength (Ng & Dobry 1992, Morgan & Boettcher 1999). The low strength of aggregates of circular shapes has been attributed to excessive rolling of the
particles. Moreover, freerolling particles do not exhibit stable deformation localization (Morgan 1999). In contrast, experiments which use noncircular shapes or
proxy mechanisms to mimic grain interlocking achieve strain localization and show higher shear strengths (Oda et al. 1997, Thomas & Bray 1999). The observations
made in the shearbox experiments on mixtures of rolling and nonrolling circular disks described in this study underline the importance of grain rolling for the bulk
strength and the micromechanics.
The increase of the rolling fraction leads to characteristic changes in the micromechanics that are reflected by changes of the bulk strength of the aggregates. A non
rolling consolidated assembly (xf=1.0) of highfriction disks subjected to shear deformation under constant confining pressure forms long columns of particles that
support large differential stresses. Restructuring at grain scale continuously increases the shear strength of the aggregate during the experiment. However, a few rolling
particles interspersed in the nonrolling aggregate (xf=0.86) effectively obstruct the formation of the stresssupporting particle columns and thus severely reduce the
bulk strength. This behavior resembles the shear failure in sandstones observed by Oda & Kazama (1998). In their biaxial experiments shear failure resulted from
buckling of stress supporting particle columns enabled by rotation of particles. A higher concentration of the rolling particles (xf=0.68 and xf=0.45) enables deformation
localization. Within the shear bands rolling particles limit the formation of interlocking particle groups that lock the shear zone and act as stress bridges. Aggregates with
only low contents of non rolling particles (xf=0.18 and xf=0.0) show even lower bulkstrength values. Remarkably, the residual strength of these aggregates keeps
declining throughout the experiments. In the allrollingcase the residual strength even drops below the residual strength of the low friction aggregate. This suggests that
highfriction aggregates, in which all particles are free to rotate, undergo permanent restructuring at the grain scale in order to reduce frictional slip. These aggregates
hence possibly develop towards a “roller bearing” structure in 2D with minimal slip of the grain contacts and a low bulk strength as proposed by Scott (1996). One
may speculate that with progressive grain fracturing that results in the fractal grain size distributions observed in fault rocks these highfriction rollerbearing structures
may even approach Appolonian packings that enable rolling with only minimal frictional slip (Baram et al. 2004).
Deformation localization is not observed in low friction aggregates, independent of the fraction of nonrolling particles. Deformation localization hence requires (1)
intermediate to high particle friction and (2) the presence of nonrolling grains which both enable the formation of stable interlocked stresssupporting contact networks
in the lowstrain regions and (3) the presence of rolling particles which impede restrengthening within shear zones.
REFERENCES
Baram, M.R., Herrmann, H.J. & Rivier, N. 2004. Spacefilling bearings in three dimensions. Physical Review Letters 92:044301.
Cundall, P.A. & Strack, O.D.L. 1979. A discrete numerical model for granular assemblies. Geotechnique 29: 47–65.
Itasca. 2002. PFC2D. Particle Flow Code in 2 Dimensions Version 3.0. Theory and Background. 2nd edition.
Jensen, R.P., Bosschner, P.J. Plesha, M.E. & Edil, T.B. 1999. DEM simulation of granular media—structure interface: Effects of surface roughness and particle shape.
International Journal for Numerical and Analytical Methods in Geomechanics 23:531–547.
Mair, K. & Marone, C. 1999. Friction of simulated fault gouge for a wide range of velocities and normal stresses. Journal of Geophysical Research B104(12):28899–
28914.
Mair, K., Frye, K.M. & Marone, C. 2002. Influence of grain characteristics on the friction of granular shear zones. Journal of Geophysical Research B 107(10):ECV 4–
1–4–9.
Mora, P. & Place, D. 1999. The weakness of earthquake faults, Geophysical Research Letters 26(1):123–126.
Morgan, J.K. & Boettcher, M.S. 1999. Numerical simulations of granular shear zones using the distinct element method; 1. Shear zone kinematics and the micromechanics
of localization. Journal of Geophysical Research B 104(2): 2703–2719.
Morgan, J.K. 1999. Numerical simulations of granular shear zones using the distinct element method; 2. Effects of particle size distribution and interparticle friction on
mechanical behavior. Journal of Geophysical Research. B104(2):2721–2732.
Ng, T.T. & Dobry, R. 1992. A nonlinear numerical model for soil mechanics. International Journal for Numerical and Analytical Methods in Geomechanics 16:247–
263.
Page 171
Oda, M. & Kazama, H. 1998. Microstructure of shear bands and its relation to the mechanisms of dilatancy and failure of dense granular soils. Geotechnique 48(4):465–
481.
Oda, M., Ivashita, K. & Kazama, H. 1997. Microstructure developed in shear band of dense granular soils and its computer simulation: mechanisms of dilatancy and
failure. In N.A.Fleck & A.C.F.Cocks (eds.), Mechanics of Granular and Porous Materials, Proceedings of the IUTAM Symposium: 353–364. Dordrecht: Kluwer.
Scholz, C.H. 2002. Mechanics of Earthquakes and Faulting, 2nd edition. Cambridge: Cambridge University Press.
Scott, D.R. 1996. Seismicity and stress rotation in granular model of the brittle crust. Nature 381:592–595.
Thomas, P.A. & Bray, J.D. 1999. Capturing noncircular shape of granular media with disk clusters. Journal of Geotechnical and Geoenvironmental Engineering 125
(3): 169–178.
Page 172
Page 173
Geology
Page 174
Page 175
Simulation of magmatic fabrics using PFC
Y.Park, C.Lee & J.H.Ree
Department of Earth and Environmental Sciences, Korea University, Seoul, Korea
5809 679 3
ABSTRACT: Magmatic foliation, defined by preferred alignment of phenocry sts in igneous rocks, is developed by flow processes in a magma
chamber. Internal convection and tectonic stress field are considered as important for the magmatic flow. For the modeling of magmatic foliation, we
used Particle Flow Code, PFC (Itasca 1999), which is based on the distinct element method. The phenocryst model in our experiments consists of 65
attached balls (5×13 in dimensions), and varying numbers of phenocrysts are placed in a viscous matrix. The initial orientations of the longer axes of
phenocrysts are randomly chosen. Wall displacement and wall rotation are used for the boundary conditions for pure shearing flow and simple
shearing flow, respectively. The observations made from numerical experiments agree well with those from physical experiments. We expect that this
type of approach will lead to better understanding of local development of microstructures (e.g. crystal plasticity at high stress contacts) and the
phenocrystfractiondependent rheology in crystal mush.
1 INTRODUCTION
Magmatic foliation is defined by preferred alignment of phenocrysts in igneous rocks. It is commonly found in a variety of igneous compositions from gabbro to granite,
and also in a variety of igneous textures reflecting deeper to shallowlevel emplacements. Flow processes related to ascent and emplacement of magma are known to
be responsible for forming magmatic foliation (Cruden 1988, Paterson et al. 1989). Magmatic foliation in granitic rocks also has fundamental importance in interpreting
regional deformation. This is because granites are closely associated with formation of orogenic belts in space and time (Paterson 1991). Thus, correct interpretation of
magmatic foliation is critical in constraining timing and mode of pluton emplacement (Paterson 1989).
Magmatic foliation is formed by rotation and translation of relatively stronger phases (e.g. Kfeldspar phenocrysts in granite) in much weaker liquidbearing phases.
Jeffery (1922) provided the theoretical background for rigidbody rotation in a flowing matrix. Equations in Jeffery (1922) relate rotation rate to the orientation of a
rigid particle with respect to the kinematic reference frame. However, Jeffery’s solutions are limited to situations where a single rigid particle is present in the system.
Therefore, for systems where multiple particles are present and mechanical interaction occurs, Jeffery’s solutions are not suitable, although the solutions are still valuable
as a reference.
When multiple rigid particles are present, the particles start to interact with each other as the volume fraction of the particles increases. A remarkable strength change
has been reported for partial melts. The strength of partial melts abruptly increases when particles start to touch each other and particletoparticle interaction starts to
occur. The volume fraction of such change is known as RCMP in literatures (Arzi 1978, Van der Molen & Paterson 1979). For simulating these situations, physical
experiments have been performed for materials with low to high phenocryst fraction (Ildefonse et al. 1992, Park & Means 1997). Numerical models are also
developed for systems with relatively high particle fractions (e.g. Tikoff 1994), but these are kinematic models without mechanical consideration. Models based on
FEM or FDM can also be used (e.g. Treagus & Lan 2003) to simulate development of magmatic foliation. However, there are two major limitations when we adopt
these models:
1 It is known that large strain is necessary for the development of magmatic foliation, but it is technically difficult to sustain large strain in these models.
2 Extreme differences in material properties between rigid particles and flowing matrix also make the FEM or FDM approach difficult because maintaining mesh
structures becomes difficult. DEM codes such as PFC are ideal to overcome these technical difficulties. Here we explore the possibility of using DEM codes to
simulate rotation of rigid particles in a much weaker ductile matrix.
Page 176
Figure 1. Sketch of magmatic foliation in the Cheongsan granite, Korea (Kwon 1997). The magmatic foliation is defined by preferred
orientation of the longer axes of Kfeldspar phenocrysts (white phases). The dotted area represents the matrix phase.
2 NUMERCIAL MODEL
The magmatic foliation in granitic rocks (Fig. 1) represents the structure that we want to model using PFC. The magmatic foliation is defined by preferred alignment of
Kfeldspar phenocrysts. The matrix phase consists of quartz, Kfeldspar, plagioclase and biotite. Crystallinity in the matrix at the time of formation of magmatic foliation
is not clearly known for the granitic rock in Figure 1. Possible ranges of crystallinity can be from very low (near zero) to ~70% where loadbearing framework starts to
form. Very low crystallinity may represent situations where magmatic foliation is formed by growing phenocrysts during suspensiontype flow (Paterson 1989). If
magmatic deformation occurred above ~70% crystallinity (RCMP, rheologically critical melt percent) (Arzi 1978, Van der Molen & Paterson 1979), we expect
presence of microstructures indicating solidstate deformation. However, we do not observe the evidence of solidstate deformation in the granitic rock. Thus, we think
that crystallinity of ~70% may be maximum phenocryst percent for the granitic rock. Although we do not have strong evidence for the crystallinity in the granitic rocks
(Fig. 1), we modeled development of magmatic foliation with data structures of rigid phenocrysts and weaker matrix because development of magmatic foliation in
general involves rotation and translation of rigid phenocrysts in flowing matrix.
Figure 2 shows the initial and boundary conditions for the model. For pure shear deformation (Fig. 2a), we used the standard biaxial testing routines included in
Figure 2. Initial and boundary conditions for the model. (a) Pure shear deformation. (b) Simple shear deformation.
PFC software. For simple shear deformation (Fig. 2b), we used two sets of parallel walls. One set of walls (the upper and lower walls in Figure 2a) is fixed while the
other set of walls (the left and right walls) rotates with the rotation center fixed at the centers of the vertical walls.
The model phenocrysts are prepared using an external PASCAL program. The PASCAL program randomly chooses the position and orientation of a phenocryst
consisting of a group of balls (5 by 13 dimensions). The program also performs duplicate checking to avoid the presence of overlapped crystals. Then, the individual
positions of each ball that constitutes a phenocryst are written in FISH language in order to group them into a single rigid body by applying a high value of bond
strength. After placing the rigid phenocryst, the matrix phase is made using the GENERATE command in PFC.
3 RESULTS
Figure 3 shows the development of preferred orientations of the model phenocrysts. The rose diagrams show the degree of preferred orientation of the longer axes.
The measurement of the longer axis for each phenocryst was carried out by image analysis using NIH Image (Rasband 1994).
In pure shear experiments, linear strain up to 0.6 (ε1) is imposed. Figure 3a, b represent the experimental results with 110 phenocrysts (~28% volume) and
Page 177
Figure 3. Experimental results. (a) Pure shear deformation with 110 phenocrysts. (b) Pure shear deformation with 44 phenocrysts. (c)
Simple shear deformation with 44 phenocrysts. The rose diagrams represent the orientation of the longer axes of phenocrysts.
44 phenocrysts (~11% volume), respectively. In both experiments, it is observed that magmatic foliation is developed by synthetic rotation of phenocrysts. The high
phenocryst volume fraction experiment (Fig. 3a) shows stronger preferred orientation than the low phenocryst volume fraction experiment (Fig. 3b). The differences in
the degree of preferred orientation can be explained by the degree of phenocrysttophenocryst interactions. When rigid phenocrysts start to touch each other as
phenocryst volume fraction increases, stress transmission among phenocrysts becomes more effective.
Simple shear experiments with different volume fraction of phenocrysts were carried out. Figure 3c shows the results for the 110 phenocryst experiments with shear
strain (γ) up to 1.95. Development of preferred orientation is also observed. Most of phenocrysts rotate toward the shear direction, but some of them rotate backward.
The observed back rotation may be interpreted in three ways:
1 Backrotation of phenocrysts to approach the stable orientation (e.g. Ghosh & Ramberg 1976).
2 Backrotation of phenocrysts by local disturbance in flow field by neighboring phenocrysts.
3 Backrotation of phenocrysts near the rotating walls because of the velocity field imposed by the boundary condition.
Figure 4. Contact force and velocity field in deformed sample by pure shear deformation. (a) Deformed state by pure shear deformation.
Box represents the region for (b) and (c). (b) Contact force within phenocrysts. (c) Velocity field.
4 DISCUSSION
As we have shown in the previous section, the development of magmatic foliation by preferred alignment of rigid phenocrysts can successfully be modeled using PFC.
The processes that lead to formation of magmatic foliation include rotation and translation of phenocrysts. In high phenocryst volume fraction experiments, we have also
shown that particletoparticle interaction is an important process in forming magmatic foliation. In addition to these geometric changes related to magmatic foliation,
PFC model also provides velocity field and force diagrams that are useful in inferring processes that may operate in deforming magma. For example, when we combine
the velocity field and contact force diagrams (Fig. 4), we can infer magmatic processes such as crystal plasticity and migration of liquid phase. From Figure 4, we can
expect crystal plastic deformation will occur at sites where contact force is high. Migration of softer crystalline matrix along with liquid phase can also be inferred from
the maps of velocity field and contact force.
The bulk strength of crystalmelt mixtures is known to be a function of volume fraction of crystalline phases (Arzi 1978, Van der Molen & Paterson 1989). In
general, spherical shapes for rigid body are assumed. Therefore, direct application of these relations to magma can be problematic because the shapes of phenocrysts
are not spherical. In the PFC model, the bulk strength of crystalmelt mixture can be measured for varying contents of phenocryst. Experiments can also be designed
to study the effect of preferred orientation of phenocrysts. Since the mechanical properties of phenocrystbearing magma are one of the fundamentally important
parameters for magma intrusion
Page 178
Figure 5. Problems related to wall geometry for simple shear deformation. (a) Velocity field near the rotating wall. (b) & (c) Improved
wall geometries for simple shear deformation. Sliding walls (b), and ringshear type boundary condition for large strain simple
shear deformation (c).
(Marsh 1995), the results from PFC experiments will be helpful in understanding the physics of magma chamber (e.g. Ord 2003).
Although it is not attempted in this study, another important but unexplored area is simulation of crystallization. Crystallization of phenocrysts in magma chamber may
be modeled by volumetovolume replacement of matrix phase by growing phenocryst. The development of algorithms for the replacement of stressed balls remains to
be done but this type of replacement will eventually lead to model general crystallization process in magma chamber. Textural development in stressed cumulate pile
may also be modeled.
The geometry of boundary condition for simple shear deformation in our experiments turned out to be problematic because of the vortical velocity field imposed by
rotating walls. Figure 5a represents the velocity field of the experiment used in Figure 3c. Near the lower left, a vortex is formed due to spatial constraints by the
rotating walls. This problem may be solved by employing a different geometry of boundary conditions as illustrated in Figure 5b. Large (or infinite) shear strain
experiments can also be done using ringshear rigs (e.g. Fig. 5c).
5 CONCLUSION
From the numerical experiments described above, it can be seen that fabric development in magma chamber can successfully be modeled using PFC. We have studied
the magmatic flow only in the context of fabric development, but PFC experiments can be used to understand important issues related to magmatic flow. They include
(1) studies of crystal plastic deformation in crystalmelt mixtures, (2) studies of mechanical properties of crystalmelt mixtures as a function of (i) volume fraction of
crystalline phases, (ii) shape of crystalline phases, and (iii) degree of preferred orientation of crystalline phases, and (3) studies of more complicated processes such as
crystal growth in deforming cumulate pile by developing crystalgrowth algorithm in PFC. We feel that this approach will be helpful for geologists interested in
quantitative understanding of the physics of magmatic flow.
ACKNOWLEDGEMENTS
We thank Michele Nelson for editorial assistance and Taewook Ha (Basis SoftKorea, Inc.) for technical support. This work was supported by KOSEF grant R02–
2003–000–10039–0.
REFERENCES
Arzi, A.A. 1978. Critical phenomenon in the rheology of partially melted rocks. Tectonophysics 44:173–184.
Cruden, A.R. 1988. Deformation around a rising diapir modeled by creeping flow past a sphere. Tectonophysics 7: 1091–1101.
Ghosh, S.K. & Ramberg, H. 1976. Reorientation of inclusions by combination of pure shear and simple shear. Tectonophysics 34:1–70.
Ildefonse, B., Launeau, O., Bouchez, J.L. & Fernandez, A. 1992. Effect of mechanical interactions on the development of shape preferred orientations: a two dimensional
experimental approach. Journal of Structural Geology 14: 73–84.
Itasca Consulting Group, Inc. 1999. PFC2D—Particle Flow Code in Two Dimensions, Ver. 2.0 User’s Manual. Minneapolis: Itasca.
Jeffery, G.B. 1922. The motion of ellipsoidal particles immersed in a viscous fluid. Proceedings of the Royal Society London A102:161–179.
Kwon, S. 1997. Structural evolution of the Cheongsan area, centralsouthern part of the Ogcheon belt, South Korea (unpublished M.S. thesis). Seoul: Korea University.
Marsh, B.D. 1995. Solidification fronts and magmatic evolution. Mineralogical Magazine 60:5–40.
Ord, A. 2003. Modeling magma ascent, In Heinz Konietzky (ed.), Numerical modeling in micromechanics via particle methods: 291–294. Lisse: Balkema.
Park, Y. & Means, W.D. 1997. Crystal rotation and growth during grain flow in a deforming crystal mush. In Sudipta Sengupta (ed.), Evolution of geological structures
in micro to macroscales: 245–258. London: Chapman & Hall.
Paterson, S.R., Vernon, R.H. & Tobisch, O.T. 1989. A review of criteria for the identification of magmatic and tectonic
Page 179
foliations in granitoids. Journal of Structural Geology 11: 349–363.
Paterson, S.R., Vernon, R.H. & Fowler, T.K. 1991. Aureole tectonics. In Derrill M.Kerrick (ed.), Contact Metamorphism, Reviews in Mineralogy Volume 26:673–722.
Chelsea: Mineralogical Society of America.
Rasband, W. 1994. NIH Image, National Institutes of Health.
Tikoff, B. & Teyssier, C. 1994. Strain and fabric analyses based on porphyroblast interaction. Journal of Structural Geology 16:477–491.
Treagus, S.H. & Lan, L. 2003. Simple shear of deformable square objects. Journal of Structural Geology 25:1993–2003.
Van der Molen, I. & Paterson, M.S. 1979. Experimental deformation of partiallymelted granite. Contributions to Mineralogy and Petrology 70:299–318.
Page 180
Page 181
Simulation of a twodimensional ‘thinskinned’ orogenic wedge using PFC
C.Lee & Y.Park
Department of Earth and Environmental Sciences, Korea University, Seoul, Korea
S.Kwon
Heesong Geotek Co., Ltd., Seoul, Korea
5809 679 3
ABSTRACT: Orogenic belts are major parts of the upper crustal deformation and the largescale evolution of orogenic wedges has broadly been
explained by critical taper theory. We used particle flow code (PFC) for the modeling of an orogenic wedge. We model the wedge shape by changing
controlling factors such as backstop velocity and basal slope and compare it with natural examples. Our results show that experiments with higher
backstop velocity (higher strain rate) behave like the noncohesive Coulomb wedge, while those with lower backstop velocity (lower strain rate)
behave differently. Our experiments suggest that the use of proper material properties is an important factor for developing more realistic orogenic
wedge model. Using PFC, we have been able to simulate deformation within a more realistic orogenic wedge; in particular, we have been able to
model the initiation of discrete thrust fault surfaces and the displacements along these faults.
1 INTRODUCTION
Foldthrust belts (FTBs) and accretionary prisms are a major component of most orogenic belts and have been the subject of study in many orogenic settings (e.g.
Alps, Appalachians, Jura Mountains and Sevier FTBs). Chapple (1978) first discussed the concept of an ideally plastic wedge that has shortened and thickened from
the back to the front of the wedge creating a surface slope (critical taper), which allows the whole wedge to move by sliding on a weak basal layer. Davis et al. (1983)
and Dahlen (1990) discussed a noncohesive frictional Coulomb wedge, which might be considered as frictional plastic wedge (Mandl 1988). These theoretical models
of critically tapered wedges (Chapple 1978, Davis et al. 1983) provide a basic framework for understanding the largescale evolution of FTB wedges (e.g. Mitra
1997, Kwon & Mitra in press). Analog experiments (e.g. Davis et al. 1983, Marshak & Wilkerson 1992, Mulugeta & Koyi 1987, Koyi 1995, Mandal et al. 1997)
have been used to compare with theoretical model predictions (e.g. Dahlen 1984). These theoretical FTB wedge models have also provided useful information in
explaining largescale features observed in active FTBs (e.g. Taiwan FTBs) and accretionary prisms (e.g. Barbados accretionary complex) (Davis et al. 1983, Dahlen
1990).
Even though theoretical wedge models (Chapple 1978, Davis et al. 1983, Dahlen 1990) and analog experiments (e.g. Davis et al. 1983, Marshak & Wilkerson
1992, Mulugeta & Koyi 1987, Koyi 1995, Mandal et al. 1997) have greatly advanced our understanding of the mechanism of largescale evolution of orogenic belts,
they have inevitable limitations. For instance, these models could not provide quantitative data such as stress trajectories and material displacement during the evolution
of a wedge.
Numerical modeling has been attempted in two dimensions (e.g. Strayer & Hudleston 1997, Allmendinger 1998) and three dimensions (e.g. Kwon & Mitra 2004)
to further our understanding of orogenic belts from kinematics to dynamics by taking advantage of both graphical and quantitative interpretation of these numerical
models. For understanding the mechanical aspects of structures, various numerical methods have been developed, e.g., such as the finite difference method, the finite
element method (e.g. Barthe 1996) and the particle flow code (PFC) (Itasca 1999). Among these methods, PFC has the advantage of modeling the initiation of
discrete boundaries such as thrust faults within an orogenic wedge, which allows the wedge to deform in a more realistic manner.
By using this method, we show how PFC can be applied to explain critical wedge theory by deriving
Page 182
known relations between surface and basal slopes for the evolution of a critically tapered wedge. We also discuss the effects of velocity on the development of the
orogenic wedge.
2 WEDGE THEORY
Chapple (1978) discussed the characteristic features of thinskinned FTBs:
(1) the belts contain particular stratigraphic horizons that experience thrusting and thrustrelated folding above the crystalline basement.
(2) the basal layer that bounds the thinskinned belt is generally composed of weaker rocks such as evaporite or shale.
(3) the initial sedimentary prism is deformed to have wedgeshaped crosssectional geometry, which is thicker at the back (where the thrusts first developed) and
thinner at the front toward the direction of propagation of faults.
(4) the rear of the wedge is strongly shortened and thickened, while the frontal part of wedge shows little internal deformation.
Figure 1 shows simplified geometry of a model wedge. The basal (β) and surface (α) slopes are defined by an angle between the undeformable basement and the
horizontal surface, and an angle between the topographic and horizontal surfaces, respectively; these angles form the critical taper angle (α+β).
Davis et al. (1983) derived an equation for noncohesive frictional Coulomb wedge having small α and β (thus, smallangle approximation for trigonometric functions
becomes possible, i.e., sin α≈α and sin β≈β). According to their equation, the wedge shape reflects material properties as well as gravityrelated terms (e.g. α and β);
these include internal friction angle pore fluid pressure (Pf) and the water overburden. By neglecting the effects of pore fluid pressure and the water overburden,
they derived a simplified equation
Figure 1. Boundary conditions used in the model. The wedge containing balls is pushed by a backstop with the indicatedranges of velocity
(g: gravity, α: surface slope, β: basal slope).
that can be applied to noncohesive Coulomb wedge as shown in the following,
(1)

where K is the dimensionless quantity related to and μb . K is defined as the following equation,
(2)
By measuring the coefficient of basal friction (μb ) based on experiment using noncohesive Coulomb material (i.e. dry sand), they obtained a value of 0.30 for μb ,
which yields a value of 1.92 for K. By using these experimental data, equation (1) is reduced to a linear relationship between surface and basal slopes.
α=5.9°−0.66β
(3)
Even though they confirmed these theoretical predictions of noncohesive Coulomb wedge by comparing them with the observed field data from Taiwan, as mentioned
by Davis et al. (1983), we have to caution that this type of wedge can only be applied to the upper crust above the ‘brittleductile’ transition zone (thickness less than
10 to 15 km). Materials below the ‘brittleductile’ transition zone do not follow Coulomb behavior. In addition, parameters such as basal friction, dip of decollement
layer, pore fluid pressure etc. may vary during wedge evolution. This suggests that the evolution of orogenic wedges may experience much more complicated
deformation histories than those predicted by the theoretical noncohesive Coulomb wedge model. However, this theory still provides a basic framework for
understanding the largescale evolution of orogenic wedges, and can be broadly applied to the quantitative analysis of orogenic wedges using numerical models.
3 MODELING SETUP
The data structure of the model wedge consists of layers of balls and walls. The layers with balls represent deformable noncohesive Coulomb wedge, while the walls
represent boundary conditions for undeformable basement and backstop (Fig. 1). We assume the boundary between the undeformable basement and the deformable
wedge as a basal decollement of the model wedge (Fig. 1). We also setup a backstop at the rear of the wedge, so that we can give uniform velocities from the back
(Fig. 1). This uniform loading condition causes the initial sedimentary prism to be deformed to have wedgeshaped geometry before it starts to move along the basal
decollement.
Page 183
In general, there are two types of boundary conditions for this type analysis. One is moving basal decollement (e.g. Davis et al. 1983) and the other is moving
backstop (e.g. Strayer & Suppe 2002). We used the moving backstop for our analysis because walltoball friction is difficult to control in PFC.
Experimental procedures follow two stages: (1) preparation of the model setup and (2) deforming the wedge by applying velocity boundary conditions at the back.
During the preparation stage, we prepared an open box that consists of the left, bottom, and right walls. Then, we created balls above the topside of the open box. The
x coordinates for the ball positions range from the coordinate of the left wall to the coordinate of the right wall. Then gravitational force is introduced, so that the
generated balls fall down and pile up within the open box. The reason we chose this type of procedure is to make the material property of our model similar to that of
noncohesive Coulomb material (e.g. dry loose sand), which is commonly used in sandbox experiments (e.g. Davis et al. 1983, Marshak & Wilkerson 1992, Mulugeta
& Koyi 1987, Koyi 1995, Mandal et al. 1997). Since the pile of the balls may not be in static equilibrium, we allow sufficient time until it reaches to static equilibrium.
After this step, balls above a certain elevation within the pile are removed so that the initial layers are parallel to decollement. The wedgedeforming stage involves the
loading (applying velocity to the left wall that is used as a backstop during the analysis) from the back after removing the right wall. The layers of balls then start to form
the wedgeshaped geometry and the deformed shape remains constant once it starts to move along the basal decollement (critical taper).
The length of the basal decollement (or the bottom wall) is 400 m with friction coefficient of 0.25, normal and shear stiffness. The heights for the left (backstop) and
right walls are 50 m. We did not impose friction on the backstop because low friction materials (e.g. graphite) are commonly used near the backstops in sandbox
experiments (Davis et al. 1983, Mandal et al. 1997). The radii of the balls range from 0.1 m to 0.3 m with Gaussian size distribution to prevent predetermined faults
within the initial wedge due to hexagonal packing (Strayer & Suppe 2002). The ball aggregates (n=1200 with the average radius of 0.2 m) consist of layers with
different colors for representing internal deformation during the analysis. Since the sediments in the upper crust are considered to have low cohesion (e.g. 5 to 20 MPa,
Hoshino et al. 1972), we do not impose bonds on balls.
We test two controlling parameters to see how they affect the deformed wedge shape: (1) velocity boundary conditions that are given to the backstop (referred to
backstop velocity hereafter), and (2) basal slopes. For this, we give 5 different velocities to the backstop: 0.001 m/s, 0.0005 m/s, 0.0025 m/s, 0.000125 m/s and
0.0000625 m/s. Instead of changing basal slope, we change the direction of gravity to have similar effects of varying basal slope with the dip angles of 0, 2, 4, 6, 8 and
10 degree.
4 RESULTS
Figure 2 shows the evolution of our model wedge. The rear of the wedge is shortened and thickened, by development of internal faults and faultrelated folds, before it
reaches to the critical stage. Once critical taper is obtained, the entire wedge passively moves forward without further internal deformation. Note the detachment folding
that is formed along the basal decollement.
Figure 3 shows the effects of velocity (strain rate) on basal slope (β) vs. surface slope (α) space. In general, higher velocity causes higher surface slope of the wedge.
In higher velocity experiments (0.001 m/s and 0.0005 m/s), surface slope decreases as basal slope increases. For experiments with lower velocities (0.00025 m/s,
0.000125 m/s, 0.0000625 m/s), changing basal slopes has little effect on surface slopes.
Figure 4 shows the relationship between backstop velocity and critical taper (α+β). Each group of data points is composed of 6 different decollement slopes. In each
group, the lowest critical taper value represents a decollement slope of 0 degree and the highest critical taper value represents a decollement slope of 10 degree. The
dispersal of the values in each group becomes smaller as the backstop velocity increases (Fig. 4). The critical taper angle (α+β) also increases as the backstop velocity
increases (Fig. 4).
Figure 2. Evolution of the model wedge.
Page 184
Figure 3. Plots of basal slope vs. surface slope for different backstop velocities.
Figure 4. Plots of backstop velocity vs. critical taper. Each group of plots (indicated with the same symbols) represents experimental
results with the same backstop velocities. The upper symbols within a group indicate higher β (basal slope).
5 DISCUSSION
Our orogenic wedge model demonstrates that the surface slope of a constant basal slope experiment increases as backstop velocity increases (Fig. 3). Because we do
not change any material properties during the experiments, the effects of backstop velocity for the critical taper angle need to be explained. In ideal Coulomb material in
which deformation occurs
Figure 5. Velocity field from one of the lower backstop velocity experiments. See text for detail.
by frictional sliding of rocks, the material behavior should be rateindependent as expressed in the Equation 1 by Davis et al. (1983). However, critical taper angles
varying with backstop velocity (and, therefore, strain rates caused by loading from the back) from our experiments (Fig. 4) suggest that the material properties of our
model wedge (aggregates of balls) have ratedependent stressstrain relation. Our experiments show that our model wedge behaves like a Coulomb wedge with the
backstop velocities of higher than 0.001 m/s (Figs 3 & 4). However, the lower backstop velocity experiments reflect strainrate dependent behavior, such that
applications of the Coulomb wedge theory are not valid.
Our model results show relatively low surface slope in the cases of lower backstop velocity experiments (lower strain rate) (Fig. 3). It is also interesting to note that
the surface slope remains nearly constant regardless of the basal slope for these lower backstop velocity experiments. Velocity field (Fig. 5) from one of the lower
backstop velocity experiments (backstop velocity=0.00025 m/s) shows ball movement when decollement slope is 10 degree. Displacement vectors show that particles
moves toward the backstop (or downward) like fluid. This may be due to insufficient internal frictions between the particles (balls). This type of downward fluidlike
behavior explains the near zero surface slopes observed in our experiments. In summary, at the low strainrate regimes, our experimental material does not show
Coulomb behavior.
The higher backstop velocity experiments reproduced the results that are compatible with Coulomb wedge theory (Fig. 4), although the critical taper angle is
considerably higher than those from natural orogenic belts (Davis et al. 1983, Dahlen 1990). For developing more realistic models close to natural orogenic belts, we
need to better understand parameters such as internal friction between the balls and basal friction along the basal decollement. We will conduct biaxial compression
type experiments for calculating material properties (e.g. internal friction, cohesion and volumetric strain) that best reproduce the conditions within natural orogenic
belts. Instead of using constant rheology throughout the analysis, we will, in future models, consider varying rheology (e.g. Koyi 1995, Mandal et al. 1997) where the
particles at rear will undergo
Page 185
Figure 6. Velocity field from one of the higher backstop velocity experiments. (a) Geometry of ball aggregates during wedge formation.
(b) Velocity field and inferred structural development of the boxed region in (a).
changes in mechanical properties, which will already have an effect on the overall taper of the wedge.
We could not successfully simulate the sandbox model because we failed to produce ‘synthetic sand’ which possessed the properties of Coulomb materials.
Nevertheless, our experimental results show similarities to sandbox experiments. Figure 6a shows a snapshot (backstop velocity=0.00025 m/s) of a deforming wedge,
showing internal deformation such as faults and detachment folds. Velocity fields of parts of wedge (box shown in Fig. 6a) show velocity vector of each particle (Fig.
6b); from this we can see the evolution of internal structures such as successive thrusting and thrustrelated folding (Fig. 6b). These types of data (e.g. velocity field) will
help more quantitative analysis of structures during wedge evolution.
6 CONCLUSIONS
From the numerical experiments of ‘thinskinned’ orogenic wedge evolution, we made the following observations and suggestions for future experiments:
(1) The critical taper angle of our model wedge is closely related to the backstop velocity. The faster backstop velocity experiments produce the model wedge with
higher critical taper. Our results reveal that experiments with higher backstop velocity (higher strain rate) show the behavior of the noncohesive Coulomb wedge.
However, experiments with lower backstop velocity (lower strain rate) do not follow the simple theoretical prediction of noncohesive Coulomb wedge.
(2) Although our higher backstop velocity experiments reproduced the results that are compatible with Coulomb wedge theory, the critical taper angles are
considerably higher than those from natural orogenic belts. This is closely related to the material properties (ball aggregates) of our model wedge and suggests
proper choice of material properties is an important factor for developing more realistic orogenic wedge models. More experiments (e.g. biaxial compression test)
are required to choose material properties that can produce a realistic orogenic wedge model.
(3) We have demonstrated that the simulation of orogenic wedge deformation using PFC is useful for quantitative understanding of the natural wedges. In particular,
initiation of thrust faults, displacements along fault surfaces, interactions of structural elements, and analysis of stress field can be studied with PFC modeling.
ACKNOWLEDGEMENTS
We thank anonymous reviewers for clarifying the paper, Michele Nelson for editorial assistance, and Taewook Ha (Basis SoftKorea, Inc.) for technical support.
C.Lee and Y.Park thank J.H.Ree for financial support (KOSEF grant: R022003–000–10039–0).
REFERENCES
Allmendinger, R.W. 1998. Inverse and forward numerical modeling of trishear faultpropagation folds. Tectonics 17: 640–656.
Barthe, C.J. 1996. Finite element procedures. New Jersey: Prentice Hall.
Chapple, W.M. 1978. Mechanics of thinskinned foldandthrust belt. Geological Society of America Bulletin 89: 1189–1198.
Dahlen, F.A. 1984. Noncohesive critical Coulomb wedges: An exact solution. Journal of Geophysical Research 89: 10125–10133.
Dahlen, F.A. 1990. Critical taper model of foldandthrust belts and accretionary wedges. Annual Review of Earth and Planetart Sciences 18:55–99.
Davis, D., Suppe, J. & Dahlen, F.A. 1983. Mechanics of foldandthrust belts and accretionary wedges. Journal of Geophysical Research 88:1153–1172.
Itasca Consulting Group, Inc. 1999. PFC2D—Particle Flow Code in 2 Dimensions. Minneapolis: Itasca.
Hoshino, K., Koide, H., Inami, K., Iwamura, S. & Mitsui, S. 1972. Mechanical properties of Japanese Tertiary sedimentary rocks under high confining pressures.
Geological survey of Japan Report 244. Kawasaki: Geological survey of Japan.
Koyi, H. 1995. Mode of internal defomation in sand wedges. Journal of Structural Geology 17:293–300.
Kwon, S. & Mitra, G. 2004, Threedimensional finiteelement modeling of a thinskinned foldthrust belt wedge: Provo salient, Sevier belt, Utah. Geology 32: 561–564.
Kwon, S. & Mitra, G. in press, Strain distribution, strain history and kinematic evolution associated with the formation of arcuate salients in foldthrust belts: the example of
the Provo salient, Sevier orogen, Utah. Geological Society of America Special Publication.
Page 186
Mandal, N., Chattopadhyay, A. & Bose, S. 1997. Imbricate thrust spacing: experimental and theoritical analysis. In Sudipta Sengupta (ed.), Evolution of geological
structures in micro to macroscales: 143–165. London: Chapman & Hall.
Mandl, G. 1988. Mechanics of tectonic faulting: models and basic concepts. Amsterdam: Elsevier.
Marshak, S. & Wilkerson, M.S. 1992. Effect of overburden thickness on thrust belt geometry and development. Tectonics 11:560–566.
Mitra, G. 1997. Evolution of salients in a foldandthrust belt: the effects of sedimentary basin geometry, strain distribution and critical taper. In Sudipta Sengupta (ed.),
Evolution of geological structures in micro to macroscales: 59–90. London: Chapman & Hall.
Mulugeta, G. & Koyi, H. 1987. Threedimensional geometry and kinetics of experimental piggyback thrusting. Geology 15:1052–1056.
Strayer, L.M. & Hudleston, P.J. 1997. Numerical modeling of fold initiation at thrust ramps. Journal of Structural Geology 19:551–566.
Strayer, L.M. & Suppe, J. 2002. Outofplane motion of a thrust sheet during alongstrike propagation of a thrust ramp: a distinctelement approach. Journal of Structural
Geology 24:637–650.
Page 187
Simulation of accretionary prisms by distinct element method
Y.Yamada, S.Ueda & T.Matsuoka
Kyoto University, Kyoto, Japan
5809 679 3
ABSTRACT: Deformations at plate subduction zones, in particular accretionary prisms, have been the key subject for a number of geoscientists not
only because of the methane hydrate distribution but also possible major earthquakes. Sandbox experiments have successfully reproduced typical
deformation architecture of accretionary prisms, such as a sequence of faulting that a new thrust generates in the footwall of the previous one. In order
to examine the deformation in detail, we have applied the Distinct Element Method (DEM) and extracted velocity and stress data from each particle.
These results suggest that the development process of faults is intermittent that strongly suggests stickslip motions along fault surfaces. Such
information is extremely useful to evaluate the timing of faulting, which is vital information for methane hydrate exploration and earthquake prediction.
1 INTRODUCTION
At plate subduction zones, a sequence of sediments is striped from the ocean plate by thrusting and forms accretionary prisms at the front of the overriding plate. The
accretionary prisms generally have wedge geometry that is cut by a series of thrust faults. Such accretionary prisms well develop in SW Japan, where Cretaceous to
modern sediments have been accreted from the Philippine Sea Plate. At the deformation front of the accretionary prism, the plate bounding fault (called ‘decollement’
surface) does not coincide with a lithologic change but is characteristically within a thick mudstone. The deformation process of accretionary prisms has been analyzed
by physical experiments using dry sand as its material (i.e. sandbox experiments; Yamada et al. 2003 and others) for methane hydrate exploration and for the research
of major earthquake prediction. By the experiments, however, it is generally difficult to examine stress state inside the accretionary prisms. In order to extract such
internal information, we employed the Distinct Element Method (DEM) (Cundall & Strack 1979, Hakuno 1997) and simulated the accretionary process numerically.
2 MODEL SIMULATIONS
The models in this study aim to reproduce the twodimensional deformation of accretionary prisms. With the knowledge of the sandbox models, the physical properties
and input parameters of the simulations were determined to make the results similar to the
Figure 1. Initial arrangement.
Table 1. Parameters in this simulation.
Radius (m) 0.08×0.12
3 2500
Density (kg/m )
Particle 8314
kn (N/m)
2×107
ks (N/m)
8.13×106
sandbox experiments. The size of the particles was randomized to avoid formation of preferential weak planes at the initial arrangement of the particle assembly (Fig.
1). In order to model the decollement surface in sediments, we have employed a threelayered model with an intercalation of a layer of lowfriction coefficient. The
deformation was produced by moving the left vertical wall to the right. The parameters used in this study are shown in Table 1.
3 SIMULATION RESULTS
The simulation results are shown in Figure 2. The compression by the moving wall generated deformation of
Page 188
Figure 2. Simulation results.
the particle assembly in front of the wall. The overall geometry of the deformation is a wedge shape that increased the size as the displacement of the wall proceeded.
The deformation also includes a lot of discontinuity surfaces, which can be identified by displacement at the marker layers. They are, however, generally not obvious
because of their minor displacements.
The velocity distribution diagrams show that velocity heterogeneity frequently appeared within the particle assembly. A group of particles being the same velocity
vector form a boundary to an adjacent particle group. This boundary corresponds to a surface of two different velocity groups, thus can be interpreted as a fault. The
velocity diagram shows that faulting within the particle assembly is quite unstable and frequently changes the position and the character. In particular, the development of
faults is intermittent that strongly suggests stickslip motions along fault surfaces. The lowfriction layer seems to affect the development of the faults and act as a
decollement surface within the particle assembly.
The contact force diagram shows the stress state within the particles. It is understood that the moving wall pushes the particles and the stress chains initiated from the
wall. The stress chains can generally be traced toward the base (downtotheright direction), but sometimes show subhorizontal to even uptotheright directions in
some stages. These clearly correspond to the fault activity when compared with the velocity diagram. From this stress chains, it can also be observed that the low
friction layer behaved as a decollement surface.
4 COMPARISONS TO THE REAL WEDGE AND ANALOGUE MODELS
The simulation results can be correlated with real prism geometry and with analogue experiments (Fig. 3). A seismic profile (Fig. 3a) illustrates a characteristic structural
style at the toe of the Nankai accretionary prism off Muroto, where the Philippine Sea Plate is subducting underneath the Eurasia Plate.
Page 189
Figure 3. Comparison of a seismic profile, the simulation result and experimental results. a) A part of a seismic profile at the toe of the
Nankai accretionary wedge offshore Muroto (Bangs et al. 2004). Note that the detachment horizon is at the middle of the mud
stone (low reflective sediments) and not at the uppermost horizon of the basaltic basement (the strong reflector). b) Simulation
with a low friction layer (cf. Fig. 2). c) ExperimentI with pure dry sand (from Yamada et al. 2003). A typical sandwedge
experiment with a piggyback sequence of thrusts. d) ExperimentII with dry sand and a layer of micro glass beads (Yamada
et al. 2004). The beads layer behaved as the detachment horizon, from which the thrust sequence developed. Note that the
geometry of the wedge is strongly controlled by the physical property of the detachment.
Page 190
The profile has no vertical exaggeration (i.e. the vertical scale equals to the horizontal scale), thus can be directly compared with the simulation results. The decollement
surface is within the mudstone and the ocean floor surface has a gently inclined geometry toward the trough. This top surface geometry is similar to that observed in the
simulation (Fig. 3b).
There are several analogue experiments conducted under the boundary conditions similar to those of the numerical simulation. One typical experiment with a
homogeneous dry sand reproduced structures similar to the real wedge; such as the piggyback sequence of thrusting, the landward tilting thrust geometry and formation
of the wedge geometry (Fig. 3c). The model succeeded to make the decollement surface within the sand pack, but the angle of the prism surface (line linking the tips of
the thrusts) is much steeper than that of the real case. In addition, the distance between the thrusts is significantly closer than that seen offshore Muroto.
Another experiment intercalated a micro beads layer within the sand to change the physical property of the detachment (Fig. 3d). The results (ExperimentII)
succeeded to reproduce the horizontal geometry of the decollement surface. In the upper sand layers, a series of thrusts were formed and converged to the
decollement. The characteristics of the structures formed in the ExperimentII coincided with that of a real prism toe. In addition, the angle of the prism surface (line
linking the tips of the thrusts) was gentler and the distances between the fault planes were larger than that in the ExperimentI. These suggest that the ExperimentII
reproduced a geometry closely similar to that observed at the toe of the Nankai accretionary wedge.
In our simulation results, the development of the faults was quite limited and no major thrust system was observed. This suggests that strain and stress localization
was not well developed in the simulation. This might suggest probable refinements in the input parameters. The simulations, however, brought dynamic process of the
prism deformation by extracting velocity and stress state within the particle assembly, which showed that the deformation is extremely dynamic and frequently changed
the state in the particle velocities and force chain patterns due to the possible stickslip motions. This information cannot be easily acquired by the experiments.
5 CONCLUSIONS
The simulation results reproduced some key features seen in the accretionary prism formation. The results also showed that the development process of faults is
intermittent, which strongly suggests stickslip motions along fault surfaces. Such information is extremely useful to evaluate the timing of faulting, which is vital information
for methane hydrate exploration and earthquake prediction at plate subduction margins.
REFERENCES
Bangs, N.L., Shipley, T., Moore, G., Gulick, S. & Kuramoto, S. 2004. Evolution of the Decollement from the Trench into the Seismogenic Zone Inferred from Mapping
Nankai Trough Decollement Seismic Reflections in 3D, Muroto Transect. Geology, in press.
Cundall, P.A. & Strack, O.D.L. 1979. A discrete numerical model for granular assemblies, Geotechnique, 29, 47–65.
Hakuno, M. 1997. Simulation of destruction (in Japanese), Morikitasyuppan, ISBN: 4627915713.
Yamada, Y., Matsuoka, T. & Baba, K. 2003. Sandbox Modeling: A Tool to Analyze Geometry and Development Process of Geologic Structures. Proceedings of the
EITJAPANAIT joint Workshop in Modern Technology on Computer in Rock Engineering. Bangkok, Thailand, p. 21–26.
Yamada, Y, Kaneda, K., Baba, K. & Matsuoka, T. 2004. Analogue Modeling of Accretionary Prisms With Granular Materials. Proceedings of the International
Symposium in Methane Hydrates and Fluid Flow in Upper Accretionary Prisms (PrismFluid2004). Kyoto, Japan, p. 14–17.
Page 191

Numerical simulation of gas hydrate behavior in marine sediments using PFC2D
S.Kreiter & B.Grupe
AG. WUM, Technische Universität, Berlin, Germany
V.Feeser
Institut für Geowissenschaften, Universität Kiel, Germany
5809 679 3
ABSTRACT: The occurrence of gas hydrate at continental margins is considered to be one cause for submarine slope failure and tsunamis in coastal
regions. In order to provide the base for a stability prediction of submarine slopes, the growth and the decay of gas hydrate was simulated with the
two dimensional distinct element method (DEM) of PFC2D. Growing gas hydrate crystals are simulated by a surface tensed ensemble of balls. The
influence of surface tension on gas hydrate growth and the influence of gas hydrate growth on the stress state were modeled. Furthermore the decay of
gas hydrate located near the sediment surface was simulated. The method used for modeling the surface tensed gas hydrate crystals is supposed to be
applicable for the simulation of other polyphase processes, too.
1 INTRODUCTION
1.1 General
Gas hydrate occurring in the sediments of continental margins is a hazard for coastal regions. The decay of gas hydrate was the cause for several submarine slides,
some of which created huge tsunamis (Bondevik 2001, Brückmann 2003). This study is part of a joint project with the aim to study continental slope stability. As gas
hydrates exist only at high pressure, they are very difficult to sample and to study in nature. They are also very difficult to be handled in soil mechanical experiments in
the laboratory, and there the process of gas hydrate growth in fine sediment is very time consuming. During simulations with the distinct element method (DEM) gas
hydrate growth can be speeded up and so the soil mechanical important interactions between sediment and gas hydrate can be examined faster.
Gas hydrate is an ice like solid consisting of water and small gas molecules. Natural gas hydrates contain mainly methane molecules. They form when the constituents
come in contact at low temperature and high pressure (Fig. 1). Methane hydrate crystals have the composition 8 CH4 46 H2O. They are crystalline solids denser than
typical hydrocarbons, so the gas molecules they contain are effectively compressed. Under normal conditions the melting of methane hydrate releases the 164 fold
volume as methane gas and the 0.9 fold
Figure 1. Stability field of gas hydrate and typical temperature pattern in the ocean and the continental slope sediment, BGHZ=Base of
Gas Hydrate stability Zone (after Kvenvolden 1987).
volume as water. The volume increases and the high concentration of hydrocarbon in gas hydrates is the cause for soil mechanical impact and for climate effects (Maslin
2004). Furthermore the high hydrocarbon concentration is the base of the potential use of gas hydrate as an energy source (Sloan 2003).
The continental margins contain vast quantities of methane hydrates and store more organic carbon than the deposits of oil, gas and coal together (Sloan 1998). They
are located in the upper part of the sediment
Page 192
Figure 2. Gas hydrate bearing sediment from the core M54–109, sampled by the R.V.Meteor during the cruise 54 of the SFB 574. (Photo
by Tobias Mörz).
column in water depths of more than 300 m. In greater sediment depth gas hydrate becomes unstable because of the heat transfer from the earth’s interior (Fig. 1). The
base of gas hydrate stability zone (BGHZ) is not fixed in room or time, it is highly variable depending on pressure and heat flow. The processes at the BGHZ are until
now not well understood, but nevertheless essential for the forecast of submarine slope failure.
Figure 2 shows a sample of gas hydrates in marine sediment. The white lenses within the sediment are gas hydrate. At the time the photograph was taken, the gas
hydrate was decaying and transforming the sediment into a mud slurry. This mud slurry is seen at the rim of the core.
Hydrate crystals grow and decompose in the sediment pores depending on changes of the ambient pressure and temperature. These processes change the overall
volume in the pore space, and modify the fabric of the grain skeleton. Therefore methane hydrate alters the compaction process and the strength of the sediment. The
strength alteration of gas hydrate bearing sediment is important, because the BGHZ (Fig. 1) is found to be the failure plane of several, partly huge submarine slides
(Brückmann 2003, Mienert 1998, Mienert 2001).
1.2 Processes related to gas hydrates
In the drilling campaign “leg 164” of the Ocean Drilling Project no gas hydrate was found under a certain sediment depth even though gas hydrate was predicted by the
thermodynamic conditions (Rupel 1997). The thermodynamic conditions for the occurrence of gas hydrate were derived from bulk experiments, so the influence of the
sediment had to be examined. Natural gas hydrates are growing in the tiny pores of the fine grained sediment and the crystals are therefore forced to form very high
curvatures. This led to the assumption that the surface energy reduces the thermodynamic stability of gas hydrates in tiny pores (Clennell 1995, 1999).
Except for coarse sandy sediments all natural gas hydrates found were distributed very heterogeneously in form of lenses or layers (c.f. Fig. 2). So it was postulated
that the surface energy of growing gas hydrate crystals in tiny pores is high enough to push the surrounding particles away (Feeser 1997). The importance of the surface
tension was confirmed by laboratory experiments which proved that gas hydrate crystals tend to form in the center of the pores (Spangenberg 2001, Tohidi 2001). But
the substrate of these experiments being glass beads their results might not apply to the conditions in natural fine grained sediment. Up to now the formation of gas
hydrate lenses could not be observed either in soil specimens or in numerical experiments. To assess how gas hydrates grow in the sediment is one aim of this study.
The growth of gas hydrate does not only depend on thermodynamic conditions, but also on the kinetics. The water molecules in the fluid phase are organized in big,
relatively stable clusters. These clusters have to undergo severe reorganizations to fit in the crystal lattice of gas hydrates (Sloan 1998).
Under stable conditions of continental margins, steady state upward fluid convection is common. By this convection the fluids are transported deeper and deeper into
the stability field of gas hydrate (c.f. Fig. 1) and this could be expected to lead to a homogeneous distribution of gas hydrate in the sediment, but for gas hydrate crystals
it is difficult to nucleate. The convection and the hindered kinetics together will cause a slight oversaturation of methane with regard to gas hydrate in the bulk sediment.
In contrast once a crystal nucleated will keep on growing, fed by the oversaturated water, depended on the curvature and the surface energy, reducing the
oversaturation. The kinetics will then prevent the nucleation of new crystals and is therefore the cause for the inhomogeneous distribution of gas hydrate in the sediment.
Very few crystals suffice to create steady state concentration gradients in the continental margin sediment. So it is reasonable to model the gas hydrate—sediment
interaction with the main focus on few crystals and their surface tension.
2 DEM SIMULATION OF GAS HYDRATE

The distinct element model (DEM) PFC2D was used as a base for the simulation (Cundall 1978, Itasca 2002). The gas hydrate related features were added by FISH
Page 193
and C++ routines. The basic idea was to model the growing gas hydrate crystals by growing PFC “balls” and to measure the reaction of the sediment in numerical
compression experiments. The simulations in this study were made parallel to the construction of the Gas Hydrate Test System (GTS) laboratory. The GTS is in
principle an oedometric compression test but has multiple additional sensors. The numerical experiments performed are adapted to the planned experiments of the
GTS.
2.1 Physical concept
The very base of surface tension is the physical interaction between multiple molecules. The behavior of surfaces has been successfully modeled by molecular dynamics
simulations (MD). The approach taken here is to simplify the MD models and to scale them up. The upscaled models then only match the macroscopic behavior of
gas hydrates crystals.
The molecular interaction potentials between water molecules and between water and gas molecules are complicated quantum physical phenomenon (Israelachvili
1991), but for practical applications a number of empirical interaction potentials can be used. Because in this simulation only the postulated surface tension has to be
simulated, the Mie potential was chosen, which describes the isotropic attraction between hard spheres:
w(r)=Ar−n −Br−m
(1)
where w=potential; r=distance between the midpoints of the spheres; and A and B=adjustable constants. The surface energy of a medium can be calculated by half the
energy needed to separate two planar surfaces of this medium. Because the Mie potential is negligible in greater distances, only the nearest neighbors have to be
considered. In a dense packing of two dimensional disks (PFC balls) a disk looses two of its six nearest neighbors during the separation of two planar surfaces, so the
surface energy γ can be calculated by:
(2)
with γ=surface energy; w∞ =the energy needed to move a neighbor disk from the touching distance to infinity; and 2 r=length of the surface, which is covered by this
disk. The energy w∞ to move an adjacent disk from the distance σ to infinity in a two dimensions model is calculated by:
(3)
Figure 3. Potential and force over the distance of the two midpoints of the interacting circles (balls), form 10% overlap to the double of
the contact distance.
where w(r)=potential at distance r; σ=the touching distance of the two disk midpoints; and r=actual distance of the two disk midpoints. For different disk radii the
constants A and B have to be adjusted in such a way, that the potential minimum of the Mie potential is at the distance in which the disks come in contact. This yields
for the chosen exponents n=3 and m=6 to:
(4)
(5)
with γ=surface tension; σ=the sum of the two disks radii or in other words the distance between the midpoints of two touching disks. The force originated by the Mie
potential is:
(6)
where F=force; w=potential; r=actual distance of the disks midpoints. This force is only applied when the disks do not overlap. In the case of overlapping disks the
standard spring potential of PFC is used.
As shown in Figure 3 the Mieforce and the springforces are zero at the distance σ, where the disks touch each other. The potential minimum is at the same distance.
The growing crystal starts as one disk (PFC ball), but this disk decays if the forces acting on it become too big. The detection of the neighboring disks was
performed by the implementation of a PFC “user defined contact law” with a lesser effective radius than the standard contact law. The pictures therefore show no PFC
balls but user defined plot objects, the PFC balls are actually strongly overlapping. The overlap makes the neighbor disks accessible by the use of the contact list of the
disk (PFC ball).
Page 194
As the aim of the study was the simulation of gas hydrate growth, the sediment used in the experiments were simple disks with the standard spring contact model and
an internal friction.
2.2 Results
2.2.1 Different grain sizes
The virtual oedometric experiment is a box constructed for consolidation experiments, the walls and the bottom of it are fixed, and the top can be moved to apply a
certain load or to conduct a constant strain experiment. First experiments were made to check the influence of the surface tension on the way gas hydrate crystals grow
in the sediment. The nucleation of gas hydrate crystal was implemented by placing a small hydrate disk (PFC ball) into the mid of the pore. The growth of the hydrate
crystals was then set to a constant volume per time.
Figure 4 shows the state after gas hydrate growth with the application of surface tension in coarse grained sediment. Figure 5 shows the state after gas hydrate
growth with the application of surface tension in fine grained sediment with the same surface tension and the same geometry, but the geometry is five times smaller.
The hydrate crystal in the fine grained sediment stays in round shape and pushes away the surrounding sediment grains. It does not intrude into the edges of the
pores. The hydrate crystal in the coarse grained sediment intrudes into the edges of the pores and does not push the sediment grains away until the connected pore
space is completely filled up. This clearly illustrates the influence of surface tension when gas hydrate interacts with the sediment grains.
2.2.2 Gas hydrate growth during consolidation
To get a first impression of the influence of gas hydrate on the state and the history of a sediment sample, the interaction of sediment and gas hydrate growth was
simulated. The horizontal force was measured to assess the stress conditions in a continental margin sediment in which gas hydrate is growing. The nucleation of the gas
hydrate crystals was simulated by a random distribution of several small hydrate balls in the virtual sediment during the preparation procedure, which was done by
radius multiplication (Itasca 2002). The gas hydrate was then growing at a constant rate. In a first experiment the gas hydrate was growing before consolidation and in a
second experiment after consolidation (Fig. 6). During the experiment the void ratio, the normal vertical stress and the normal horizontal stress were registered. Figure 7
shows the void ratio over the normal vertical stress. Growing gas hydrate extends the pore space and lets the void ratio rise. Without much vertical stress the rise of
void ratio is much higher, and the consolidation starts with a
Figure 4. Hydrate growth in coarse grained sediment; dark gray balls are hydrate; light gray balls are sediment. Lines in hydrate balls
mark the direction of the applied force. Scale has the same absolute length as in Figure 5.
Figure 5. Hydrate growth in fine grained sediment; dark gray disks are hydrate; light gray disks are sediment. Five times smaller than
Figure 4, but same relative geometry and same relative hydrate volume as Figure 4. Lines in hydrate disks mark the direction of
the applied force. Scale has the same absolute length as in Figure 4.
collapse (until Fig. 7 point c). Although the history of the samples is quite different, in the void ratio vs. effective stress diagram graphs of the two experiments end
nearly at the same point.
In Figure 8 the deviatoric stress is plotted over the effective isotropic stress. The deviatoric stress which
Page 195
Figure 6. Virtual compression test with gas hydrate, gas hydrate growth after compression. Dark gray balls are gas hydrate; light gray
balls are sediment.
Figure 7. Void ration over vertical stress during gas hydrate growth and consolidation.
Figure 8. Deviatoric stress over isotropic stress during gas hydrate growth and consolidation.
is generated by the consolidation of the first consolidating experiment is reversed by the growth of gas hydrate. In the case of the early growing gas hydrate deviatoric
strain can not build up as it would do without a surface tensed crystal in the pore space. The stress state of gas hydrate bearing sediment is shown to be
Figure 9. Simulation of gas hydrate decay during the decay Light gray balls=methane; medium gray balls=sediment; dark gray balls=gas
hydrate.
heavily dependent on the history of load application and the history of gas hydrate growth in the sediment.
2.3 Simulation of surface near gas hydrate decay
Most of the gas hydrate research is on surface near gas hydrates, because they are accessible with standard research ships with no possibility to drill and because they
are related to interesting structures like mud volcanoes or active faults. In such active environments methane bubbles escape into the free water probably being released
from gas hydrate by the warmth of an ascending fluid (Wood 2002). A melting procedure was developed to simulate the surface near decay of gas hydrate to methane.
The methane bubble nucleates a random ball of the gas hydrate and growths then in mass balance with the decaying gas hydrate. The methane bubbles were simulated
like the gas hydrate crystals as a surface tensed medium out of PFC balls. The surface tension, the mass and the stiffness were adjusted to fit the properties of methane
in 1000 m water depth, a depth common for mud volcanoes and other fluid escape structures (Söding 2003). The buoyancy of the methane balls in water is calculated
and is applied as an additional force to the methane representing balls. Figure 9 shows the virtual sediment during the process of hydrate decay. Figure 10 shows the
resulting methane bubbles in a later stage of the experiment. Some of the methane bubbles are moving through the sediment, some have escaped in the meantime.
3 DISCUSSION
Experiments have shown that the simulation of surface tension in the DEM is possible with methods of the molecular dynamic simulation. The influence of gas
Page 196
Figure 10. Simulation of surface near gas hydrate decay, after the decay during the escape of the methane. Light gray balls=methane;
medium gray balls=sediment.
hydrate growth on the sediment state was modeled as well as the escape of methane bubbles into the water. It is possible to extend the application of surface tensed
ball—ensembles to other processes like the freezing of soil or other processes in polyphase mixtures.
The above mentioned oedometric GTS gives the opportunity to test sediments under the conditions of the BGHZ. This experimental data is needed to validate the
way the gas hydrate related processes are simulated. This data will also be used to calibrate the sediment and hydrate crystal properties and to get a deeper
understanding of the processes in nature. The calibrated sedimenthydrate ensemble can then be used for other numerical experiments, which measure the shear
strength of the sediment like biaxial tests or shear—box tests.
A severe shortcoming of the two dimensional simulation appears during the simulation of gas hydrate growth in coarse grained sediment. The growth of gas hydrate
pushes the surrounding particles away because the pore space is not connected. There is no possibility in two dimensions to transfer material to a neighboring pore, so
always a sort of overpressure develops and the sediment grains are pushed away. In Figure 4 the uppermost ball is lifted by this process. In three dimensions where the
pore space is connected, hydrate crystals would grow around the grains without moving them.
In this state of the model, the pores of the sediment are empty, in real deep sea sediment they are filled with fluids. The presented simulation is therefore only valid for
slow changes of stress and at least a medium permeability of the sediment, where the fluid can always escape and no overpressure develops. For faster processes the
mechanical influence of the pore fluid has to be taken into account. Also the mass transport within the fluid has not been modeled so far, which is needed to fully
understand the formation of the gas hydrate lenses.
4 OUTLOOK
To overcome the known shortcomings of the simulation, it is intended to move to a three dimensional model, with a coupled fluid model and the consideration of mass
transport, as detailed the computational costs admit. The properties of the sediment will also be refined to match the properties of a cohesive material.
ACKNOWLEDGMENTS
The underlying project was funded by German Federal Ministry for Education and Research (BMBF) under the no. 03G0560A. The authors are solely responsible for
the contents of this paper.
REFERENCES
Bondevik, S., Harbitz, C.B., et al. 2001. The Storegga Slide tsunami along the Norwegian coast—from the geological record to numerical simulations (internet
publication). 2003.
Brückmann, W., Müller, C., et al. 2003. Submarine mass wasting associated with the upper limit of the gas hydrate stability zone at the Costa Rica continental margin.
EGSAGUEUG Joint Assembly, Nice, France.
Clennell, M.B., Hovland, M., et al. 1999. Formation of natural gas hydrates in marine sediments 1. Conceptual model of gas hydrate growth conditioned by host sediment
properties. Journal of Geophysical Research 104(B10): 22985–23003.
Clennell, M.B., Hovland, M., et al. 1995. Role of capillary forces, coupled flows and sedimentwater depletion in the habit of gas hydrate. EOS 76 (17, supplement):164–
165.
Cundall, P.A. & Strack, O.D.L. 1979. A Discrete Numerical Model for Granular Assemblies. Géotechnique 29:47–65.
Feeser, V. 1997. Gashydratbildung in Tiefseesedimenten—Zur Rolle der sedimentmechanischen ProzessSteuereung. DGMKTagungsbericht 9706:51–60.
Israelachvili, J.N. 1991. Intermolecular and surface forces. London, Academic Press Limited.
Itasca Consulting Group, Inc. 2002. PFC2DParticle Flow Code in 2 Dimensions, Version 3.0 User’s Manual. Minneapolis: Itasca.
Kvenvolden, K.A. 1987. Methane hydrate; a major reservoir of carbon in the shallow geosphere. Geological Society of America, 1987 annual meeting and exposition.
Geological Society of America. 19; 7:736.
Maslin, M., Owen, M., et al. 2004. Linking continentalslope failures and climate change: Testing the clathrate gun hypothesis. Geology 32(1):53–56.
Mienert, J., Posewang, J., et al. 2001. Changes in the Hydrate Stability Zone on the Norwegian Margin and their Consequence for Methane and Carbon Releases into the
Oceanosphere. The Northern North Atlantic: A Changing Environment. P.Schäfer, W.Ritzrau, M.Schlüter and J. Tiede. Berlin, Springer.
Mienert, J., Posewang, J., et al. 1998. Gas hydrates along the northeastern Atlantic Margin; possible hydratebound margin instabilities and possible release of methane.
Gas
Page 197
hydrates; relevance to world margin stability and climate change. J.P.Henriet and J.Mienert. London, United Kingdom, Geological Society of London. 137:275–291.
Ruppel, C. 1997. Anomalously cold temperatures observed at the base of the gas hydrate stability zone on the U.S. Atlantic passive margin. Geology (Boulder) 25(8):
699–702.
Sloan, E.D. 1998. Clathrate hydrates of natural gases. New York, Basel, Marcel Dekker, Inc.
Sloan, E.D. 2003. Fundamental principles and applications of natural gas hydrates. Nature 426:353–359.
Söding, E., Wallmann, K., et al. 2003. RVMeteor Cruise Report M54/2+3. Fluids and Subduction Costa Rica 2002. Kiel, GEOMAR.
Spangenberg, E. 2001. Modeling of the influence of gas hydrate content on the electrical properties of porous sediments. Journal of Geophysical Research 106(B4):
6535–6548.
Tohidi, B., Anderson, R., et al. 2001. Visual observation of gashydrate formation and dissociation in synthetic porous media by means of glass micromodels. Geology
(Boulder) 29(9): 867–870.
Wood, W.T., Gettrust, J.F., et al. 2002. Decreased stability of methane hydrates in marine sediments owing to phaseboundary roughness. Nature 420:656–660.
Page 198
Page 199
The particle flow code, PFC2D, applied in planetary studies to model the tectonic
evolution of chasma walls on Mars
B.Imre
Department for Planetary Research, German Aerospace Center (DLR), Berlin, Germany, now Division of Geotechnical Engineering, Swiss
Federal Institute of Technology (ETH), Zurich, Switzerland
Numerical Modeling in Micromechanics via Particle Methods—2004—Shimizu, Hart & Cundall (eds.) © 2004 Taylor & Francis Group, London, ISBN 90 5809
679 3
ABSTRACT: Valles Marineris offers a deep natural insight into the upper crust of Mars. Its wall slope morphology reflects properties of its wall
materials, thereby constraining models of structure, composition, and evolution of the Martian crust. This study successfully investigates mechanical
rock mass parameter of the northern wall of eastern Candor Chasma by inferring them from simulations of its present day wall morphology and its
tectonic history by applying the distinct element code PFC2D within a parameterstudy. The best fit model yields a homogenous distribution of low
level mechanical rock mass properties distributed over the entire stratigraphic column of that wall. The values are 5.0 (±0.7) MPa for the uniaxial
compressive strength, 1.6 (±0.2) MPa for the Brazilian tensile strength, 4.7 (±1.5) GPa for the Young’s modulus, 0.20 (±0.15) for the Poisson’s ratio,
22 (±2)° for the internal friction angle, 1.6 (±0.2) MPa for the cohesion and 2200 (±500) kg m−3 for the density.
1 INTRODUCTION
Valles Marineris (Mariner valleys, named in honor of the achievements of the Mariner 9 mission (Kieffer et al. 1992) which encountered Mars in 1971) represents a
vast, about 4000 km long system of troughs in the Tharsis region of Mars just south of the Martian equator (Fig. 1).
The majority of the troughs exhibit a relief up to 8 km. (Beside the term “trough” also the terms “chasma” and “canyon” are widely used to describe these
depressions.) The very high wall slopes of Valles Marineris offer a deep natural insight into the upper crust of Mars. The wall morphology reflects properties of the wall
materials, thereby constraining models of structure, composition, and evolution of the upper layers of the Martian crust (Carr 1981). Or in other words: The angle of a
slope is a function of the strength of the slope forming material against failure, the forces acting on this material and the time until the slope reaches stable state. The way
around is also valid: The angle of a stable slope allows inferring mechanical rock mass parameters. That leads to the scope of the present study: To gain information on
mechanical wall rock mass parameters by modeling the tectonic history and mechanical slope stability of a wall within the Valles Marineris System. An accurate analysis
of a slope stability problem needs to consider large evolving strains and path dependency by a constitutive model what exceeds the capabilities of conventional limit
equilibrium methods (e.g. Laouafa & Darve 2002).
Therefore the present study investigates mechanical rock mass strength and deformability properties by numerical nonlinear analysis based on the distinct element
code PFC2D version 2.00 (Itasca 2002).
Naturally the results will never be unique what is inherent to open systems like the present one. But it is possible to validate model results on observed data (Oreskes
et al. 1994).
Finally the present study uses only tools which have been already successfully applied to a wide range of applications (e.g. Konietzky 2003) and which are originally
implemented in PFC2D 2.00 and its Augmented Fishtank.
2 SITE SELECTION AND 2D WALL SLOPE PROFILE FOR MODEL CALIBRATION
For model calibration and interpretation it is crucial to choose a Valles Marineris wall slope which is, on the one hand, representative and, on the other hand,
structurally as simple as possible. That is to reduce the number of unknown variables which can lead to complex and unwieldy models of very questionable means (e.g.
Oreskes et al. 1994, Starfield & Cundall 1988). Additionally a large number of high quality remotesensing data of this wall section has to be
Page 200
Figure 1. Digital elevation model of central Valles Marineris modified after DLR—RIPF (2002), merged with an orthophoto mosaic
based on Viking orbiter images (DLR—RIPF 2002, MSSS 2004). The study area at the northern wall of eastern Candor
Chasma is marked by a white square. Nearby geographic features are indicated with numbers: Candor Chasma (1), Ophir
Chasma (2), Melas Chasma (3), Coprates Chasma (4), Ophir Planum (5), Lunae Planum (6). Geoid height in meters with zero
elevation as described in Smith, et al. 2001.
Figure 2. Nonexaggerated arithmetic mean profile of the northern wall of eastern Candor Chasma. It clearly displays the steep rim and
the terrace.
available. These requirements are good fulfilled by the northern wall of eastern Candor Chasma on which this study will focus on (Fig. 1). This wall segment is
described as straight aligned and free of visible signs of tectonic ruptures except a basal fault scarp (e.g. Peulvast et al. 2001). It displays a homogenous spur and gully
morphology and its wall rocks are only partially mantled by eolian or other materials.
For calibration of a 2D numerical model a simple but still representative 2D profile of that 3D chasma wall is created by calculating a mean slope profile (Fig. 2) out
of elevation data collected by the Mars Orbiter Laser Altimeter (MOLA) on the Mars Global Surveyor Mission which encountered Mars in 1997 (Smith et al. 2001).

3 SETUP OF THE PFC2D MODELS
The present PFC2D models of this chasma wall assume that the wall slope evolves naturally from an initially entirely flat plane to its final height due to normal faulting
alone, along preexistent basement faults under Martian gravity conditions (g=3.71 ms−2, Archinal 2003). The assumption that normal faulting is the exclusive agent
forming the height of the slope is
Page 201
Figure 3. Simplified flow diagram showing the bipartite modeling process of the present study to perform a parameterstudy on a solid
material within PFC. The tectonic chasma wall modeling environment simulates the morphology and evolution of the northern
wall or eastern Candor Chasma. The strength test environment calibrates microball properties on given mechanical rock mass
macrostrength properties. Or in the way around it determines mechanical rock mass macrostrength properties from an
assembly of balls with given microproperties.
based on the discussion on which the northern wall of eastern Candor Chasma shows signs of diastrophic processes that have led or at least have contributed to its
present height (e.g. Peulvast et al. 2001). Values for the dip of the assumed basement faults are extracted from literature without any preference to test their influence
on the wall slope development (45°: Schultz & Lin 2001, 60° and 90°: Mège & Masson 1996). The basic model consists of only one fault plane. A modified model
with a set of two faults is applied to simulate the hypothesis of the origin of a wallterrace induced by two parallel faults. To choose a proper faulting rate for a dynamic,
pathdependent model solution (Itasca 1999) which fulfills both requirements of low unbalanced forces which otherwise could cause unnatural damage and short
calculation time, numerous tests are performed addressing this issue. It turns out that a faulting rate of 0.90 ms 1 is the best compromise in this question. These models
are true scale models (initial width/height: 12500/37000 m) to prevent scale effects.
PFC simulates the mechanical behavior of a system comprised of a collection of discs of unit width (also referred as “particles” or “balls”). Its distinct particles can
displace independently from one another and interact only at contacts between them. Newton’s laws of motion provide the fundamental relationship between particle
motion and the forces causing that motion. In this study particles display diameters from 60 to 90 m following a Gaussian normal distribution. The behavior of massive
rock is modeled by bonding the particles together at their contact points to create an artificial material. A physical problem like the behavior of loose sand that is
concerned with the movement and interaction of particles can be modeled correctly in PFC just by defining microproperties of balls and its contacts. But so far it is
not possible to define directly bulk properties of an artificial material like solid rock within PFC.
Therefore the macromechanical (bulk) properties of a ball assemble representing a solid material have to be determined in an additional step by performing a series
of tests within a uniaxial/biaxial compressiveand Brazilian tensile strength test environment (Fig. 3). The uniaxial and biaxial compressive strengths, the static Young’s
modulus and Poisson’s ratio are derived
Page 202
from those compressive tests. Material tensile strength properties are directly derived from Brazilian tensile strength tests.
To derive the important material strength properties internal friction angle φ and cohesion c from the failure envelope of an artificial material obtained from series of
biaxial compressive tests, the failure criterion after MohrCoulomb is applied (modified after Mandl 1988):
τ=c′+σ′ tan φ′
(1)
where τ=shear stress, c′=effective cohesion, σ′=effective normal stress and φ′=effective internal friction angle. The advantages of this failure criterion are that it is
simple, and very often successfully applied. It can be adopted that the intermediate principal stress has no influence on failure as is modeled here in PCF2D. Hence it is
purely 2 dimensional as it is PFC2D. The initial width of tested samples within the strength test environment amounts 1800 m and is therefore 20 times greater than the
radius of their largest balls. Test samples intended for compressive strength tests display a width to height ratio of 1:2 (1800/3600 m). Both ratios follow
recommendations on strength tests by Vutukuri et al. (1974).
On account of the size of both the tectonic chasma wall model and the test samples within the strength test environment the macro (bulk) strength properties of the
tested artificial materials can be treated as strength properties of a rock mass. No efforts are made to evaluate intact rock strength properties out of rock mass strength
values because this would require numerous assumptions which can hardly be made seriously for Valles Marineris wall rocks based on actual knowledge. The pore
pressure is implemented in the failure criterion via the effective normal stress σ′. It is mentioned for the sake of completeness because its acting in the past within the wall
rock of Candor Chasma can not be excluded. But the effects of enhanced pore pressure are not modeled explicitly. Mars Orbiter Camera (MOC) narrow angle
images of the Mars Global Surveyor Mission which encountered Mars in 1997 (Malin et al. 1998) analyzed in this study, show no evidence of present acting of
volatiles like seeps on the wall slope (e.g. Hartmann 2001). Furthermore enhanced pore pressure is not required to explain the results of the tectonic chasma wall
model. In compressive strength tests the artificial materials dilate slightly during failure. The amount of dilatation is not recorded in the present study. Therefore no exact
statements can be made but it is estimated that internal friction angles φ may be increased by 10% due to dilatation. Finally it must be stated that the values of Poisson’s
ratio obtained in PFC2D are not strictly comparable to the Poisson’s ratio of a real material obtained from triaxial tests, because conditions in a twodimensional biaxial
test within PFC2D are neither plane strain nor plane stress (Itasca 1999).
4 SIMULATING THE TECTONIC EVOLUTION OF THE CHASMA WALL
4.1 Chasma wall models with two parallel basement faults
In a first attempt it is assumed that normal faulting is the only wall forming agent and care has to be taken that the resulting slope not only matches the real wall in its
slope angle but also in its overall profile (e.g. Fig. 2). To find answers on the development of the terrace within the wall slope a parameterstudy is performed on double
fault chasma wall models (Fig. 4). Those models indeed develop a terrace what would imply that the steeper wall segments bounding the terrace are the scarps of two
parallel normal faults. Although this is a reasonable attempt, one major observation excludes this slope forming process and also casts a different light on the evolution
of the real chasma wall itself: The best fit model (in terms of the resulting slope angle) forms two fault zones with tilted strata within instead of distinct fault planes. But
observations on MOC narrow angle images reveal horizontal layering within the wall rock to depth of at least 6 km. Based on this finding this type of model is not
further pursued.
4.2 Chasma wall models with a single basement fault
Because a purely structurally controlled chasma wall development alone does not coincide with flat laying strata, an additional process has to be taken into account:
Erosional widening of chasmata. This process is discussed in literature in great number (e.g. Peulvast et al. 2001). In the attempt to simulate and to discuss erosion as
an agent of wall slope development, a mathematical denudation model after Scheidegger (1991) is utilized. A combination of both processes is modeled by manually
combining a PFC2D single fault chasma wall model to simulate the endogenic process and a denudation model to simulate the exogenic process. Although possible the
denudation model is yet not implemented directly within the program code of the tectonic chasma wall models. This might be done in future studies. Therefore this
report focuses on the PFC2D single fault tectonic chasma wall models alone.
4.2.1 Initial situation
At the beginning a 12.5 km thick block of the upper Martian crust at the position of present day Candor Chasma is assumed. This block will form the future southern
rim of the Lunae planum and the northern wall of eastern Candor Chasma. It is also assumed that
Page 203
Figure 4. Tectonic chasma wall model with two predefined parallel basement faults. Fault “1” is activated first. After it has reached a
throw of about −4200 m it stops and fault “2” is activated. The cumulative throw amounts about −6200 m. Clearly visible is that
both basement faults induce the formation of a terrace under intense deformation of strata within the two emerging surface
fault zones.
Figure 5. Best fit tectonic chasma wall model showing the first phase of wall slope development, after normal faulting has created a slope
with an “ancient” (Peulvast et al. 2001) relief of about 4.5 km. The relief is higher than the indicated fault throw because the
hangingwall block experiences horizontal extension. It can be clearly seen that faulting induces within the relatively
incompetent best fit material a fault zone with tilted strata (indicated by white marker balls).
this block consists of horizontally layered material—the future wall rock.
4.2.2 The first “ancient” faulting period
In a first phase it is assumed that normal faulting creates a slope with an “ancient” (Peulvast et al. 2001) relief of about 4.5 km, the approximated height of the slope
section above the terrace. This process is now simulated within the tectonic chasma wall modeling environment by a single basement fault model (Fig. 5). Numerous
simulations with varying strength and density properties of the wall rock mass are performed in a parameterstudy. The modeling focuses on the evaluation of rock mass
strength values necessary to support an about 20° dipping slope like the upper section of the mean profile. Keeping the properties of the rock mass
Page 204
constant the influence of different initial basement fault dips on the development of the wall slope and surface fault geometry is tested as well. As expected it turned out
that basement fault dips of 45°, 60° or 90° have only slight influence on the morphology of wall slopes higher 3 km. The dip of such high slopes is in the last
consequences primarily controlled by the lithostatic stresses behind the slope which are a function of the weight of the overburden. A different story tells the
development of surface faults. For the case of basement faults with dips of 45° and 60° a steep, about 70° dipping planar synthetic surface fault evolves reaching finally
a dip of about 45°. Hence the inclination δ of the synthetic fault zone (zone of maximum shear displacement) decreases with time, depending on the throw of the normal
fault. This behavior can be described mathematically such that the inclination δ of the fault with respect to the orientation of the maximum principal stress is given by
the following equation:
(2)
where φ′ is the internal friction angle of the rock mass (modified after Mandl 1988). Initially the maximum lithostatic principle stress is vertically oriented. While the
wall slope evolves the orientation of the principle stresses rotates under control of gravity and the morphology of the overburden what causes the fault to rotate and a
broad fault zone to develop (e.g. Dresen et al. 1991). This behavior is clearly shown in the present model where faulting induces within the relatively incompetent best
fit material rather a fault zone than a distinct fault plane (e.g. Figs 4 & 5). Additionally a planar antithetic surface fault evolves with a constant dip of about 55°. In the
cases of a basement fault dipping 90°, surface faults develop, as expected as steeply inclined “palm tree structures”. But this behavior does not coincide with
observations made on Mars.
The best fit result of the parameterstudy to evaluate rock mass strength values necessary to support an about 20° dipping slope is shown in Fig. 5. It turns out that
the stratigraphic column consists mechanically of only one, massive wall rock unit.
As mentioned previously the mechanical properties of an artificial material are defined by the microproperties of its balls and balltoball bonds. The best fit micro
properties of the wall rock are displayed in Table 1. The bulkproperties or macroproperties of this unit are determined as described above by performing a series of
strength tests on that particular artificial material. Their results are summarized in Table 2.
4.2.3 The second “recent” faulting period
At the beginning of this phase it is assumed that denudation sets in at high rates causing the slope
Table 1. Best fit microproperties for the artificial material “wall rock”.
Microproperty Value Unit

“Wall rock” ball properties
Ball density 1900–3300
kg m−3
Static Young’s Modulus 3.9e9 Pa
Ball normal to shear stiffness ratio 0.4 –
Friction coefficient 0.5 –
“Wall rock” parallel bond properties
Static Young’s Modulus 3.6e9 Pa
Normal to shear stiffness ratio 0.4 –
Normal strength 4.8e6 Pa
Normal strength standard deviation 1.6e6 Pa
Shear strength 4.8e6 Pa
Shear strength standard deviation 1.6e6 Pa
Multiplier of bond radius 1.0 –
Table 2. Best fit rock mass properties for the material “wall rock”.
Macro properties (rock mass properties) Value Unit

Uniaxial compressive strength 5.0 (±0.7) MPa
Brazilian tensile strength 1.6 (±0.2) MPa
Static Young’s modulus 4.7 (±1.5) GPa
Poisson’s ratio 0.20 (±0.15)* –
Internal friction angle 22 (±2)** degrees
Cohesion 1.6 (±0.2) MPa
Density 2200 (±500)
kg m−3
2D
* Values of Poisson’s ratio obtained in PFC are not strictly comparable to the Poisson’s ratio of a real material obtained from triaxial tests (Itasca 1999).
** The internal friction angle may be increased by 10% due to dilation.
to recess forming a faultline scarp. To implement the results of the denudation model after Scheidegger (1991) the corresponding balls are deleted manually within the
best fit tectonic chasma wall model of phase one.
This second phase is then simulated by reactivating the existing basement fault. Within this phase the “recent” (Peulvast et al. 2001) fault scarp and a corresponding
“recent” level of the chasma floor evolve. In addition a second, planar antithetic fault evolves with a constant dip of about 55° (Fig. 6).
4.2.4 Discussion
This combined model is capable to explain the overall profile of the investigated wall section, the undisturbed horizontal layering in the upper wall section and the dip of
the fault zone at the base of the wall
Page 205
Figure 6. Sketch of the combined, best fit models at the end of the second faulting phase. This result matches the mean profile very well.
Furthermore it displays undisturbed layering in the upper slope sections and a debris slope with incorporated tilted blocks in the
lower slope section as it is seen on MOC narrow angle images.
(Fig. 6). This model reveals that the stratigraphic column may consist mechanically of only one, massive wall rock unit. The main reason may be found in the fact that
the present tectonic chasma wall models show a maximum resolution of about 1000 m caused by their particle diameters. The assumption, that at such a scale the
chasma walls are mechanically homogeneous, can also be implied from their vertically fairly uniform erosional patterns (e.g. Schultz, 2002). The rock mass strength
results of the present study compare remarkably well with the studies by Schultz, 2002 or Clow & Moore, 1988.
5 CONCLUSION
This study successfully investigates mechanical rock mass parameters of the northern wall of eastern Candor Chasma by inferring them from its present day morphology
and its slope forming history. It applies the distinct element code PFC2D to simulate the tectonic element of the slope evolution, and to perform a parameterstudy on
mechanical rock mass parameters consistent with past and present wall slope morphology. Additionally a mathematical denudation model is applied to simulate the
acting of denudation to complete the picture of the slope forming history. It is shown that two periods of normal faulting together with a massive interim denudational
scarp recess are a valid model for the evolution of the northern wall of eastern Candor Chasma. The best fit tectonic model yields a homogenous distribution of low
level wall rock mass strength and deformability properties distributed over the entire stratigraphic column of the northern wall of eastern Candor Chasma.
6 FUTURE SCOPES
So far the denudation model after Scheidegger (1991) has been only manually combined with the PFC2D chasma wall models. The next step will be to implement this
denudation model directly within their program codes. The presence of ground ice at equatorial regions in past and present is discussed intensively and is of great
interest (e.g. Carr 1996, Clifford 1993, Fanale et al. 1986). The results of the currently ongoing radar mapping campaign of the European “Mars Express” orbiter
mission are expected to be very valuable for answering this question. PFC2D would allow implementing the influence of volatiles on slope stabilities directly. It is also
expected that the High Resolution Stereo Camera (HRSC) on “Mars Express” will provide numerous high quality remote sensing data of the present study area as well.
High resolution stereo images and improved digital terrain models generated from these data will allow investigating Valles Marineris wall slopes in even greater detail.
That is of vivid interest for future attempts to model questions related to these wall slopes or other morphologic features on Mars.
ACKNOWLEDGEMENTS
This study was performed at the German Aerospace Centre (DLR), Berlin, Germany. I am very grateful to Tim Vietor and Erik Rybacki both from the
Geoforschungszentrum (GFZ) where I was able to perform simulations on PFC2D. I am also grateful to Ernst Hauber and Marita Waehlisch (DLRAdlershof), Marcel
Naumann (GFZ Potsdam), Heinz Konietzky (ITASCA Germany) and Sarah Springman (ETH Zurich) for thorough comments.
REFERENCES
Archinal, B. 2003. A Proposed Standard for Mars Surface Representation. Submitted to the Mars Geodesy and Cartography Working Group (unpublished).
Carr, M.H. 1996. Water on Mars. New York: Oxford Univ. Press.
Carr, M.H. 1981. The Surface of Mars. New York: Oxford Univ. Press.
Page 206
Clifford, S.M. 1993. A Model for the Hydrologic and Climatic Behavior of Water on Mars. J. Geophys. Res. 98(E6): 10,973–11,016.
Clow, G.D., Moore, H.J. 1988. Stability of Chasma Walls in the Valles Marineris, Mars. In Lunar Planet. Sci. XIX, 201–202. Lunar and Planetary Institute, Houston, TX.
DLR—RIPF German Aerospace Center, last update August 30, 2002. Regional Planetary Image Facility. Pieth, S., Keller, C. (supervisor). BerlinAdlershof, Germany:
German Aerospace Center. URL: http://solarsystem.dlr.de/RPIF/
Dresen, G., Gwildis, U., Kluegel, TH. 1991. Numerical and Analogue Modelling of Normal Fault Geometry. In Roberts, A.M., Yielding, G., Freeman, B. (eds), The
Geometry of Normal Faults. Geological Society Special Publication No. 56:207–217. London, UK: The Geological Society.
Fanale, F.P., Salvail, J.R., Zent, A.P., Postawko, S.E. 1996. Global Distribution and Migration of Subsurface Ice on Mars. Icarus 67:1–18.
Hartmann, W.K. 2001. Martian Seeps and their Relation to Youthful Geothermal Activity. Space Science Reviews 96: 405–410.
Itasca Consulting Group Inc. 2002. PFC2D—Particle Flow Code in 2 Dimensions, Version 2.00–103 (April 2, 2002), Minneapolis: Itasca.
Itasca Consulting Group Inc. 1999. PFC2D—Particle Flow Code in 2 Dimensions, Version 2.00 User’s Manual, Minneapolis: Itasca.
Kieffer, H.H., Jakosky, B.M., Snyder, C.W. 1992. The Planet Mars: From Antiquity to the Present. In Kieffer, H.H., Jakosky, B.M., Snyder, C.W., Matthews, M.S. (eds).
Mars: 1–33. Tucson AZ: University of Arizona Press.
Konietzky, H. (ed) 2003. Numerical Modeling in Micromechanics via Particle Methods. Proc. 1 st International PFC Symp., Gelsenkirchen/Germany, 6–8 November
2002. Rotterdam: Balkema.
Laouafa, F., Darve, F. 2002. Modelling of Slope Failure by a Material Instability Mechanism. Computers and Geotechnics 29:301–325.
Malin, M.C., Carr, M.H., Danielson, G.E., Davies, M.E., Hartmann, W.K., Ingersoll, A.P., James, P.B., Masursky, H., McEwen, A.S., Soderblom, L.A., Thomas, P.,
Veverka, J., Caplinger, M.A., Ravine, M.A., Soulanille, T.A., Warren, J.L. 1998. Early Views of the Martian surface from the Mars Orbiter Camera of Mars Global
Surveyor. Science 279:1681–1685.
Mandl, G. 1988. Mechanics of Tectonic Faulting: Models and Basic Concepts. Zwart, H.J. (series ed), Developments in Structural Geology, 1. Amsterdam, the
Netherlands: Elsevier Science Publishers B.V.
Mège D., Masson, P. 1996. Amounts of Crustal Stretching in Valles Marineris, Mars. Planet. Space Sci. 44(8): 749–782.
MSSS—Malin Space Science Systems, Inc., last update 2004. Viking Image Archive. San Diego California: Malin Space Science Systems, Inc. URL:
http://barsoom.msss.com/http/vikingdb.html
Oreskes, N., ShraderFrechette, K., Belitz, K. 1994. Verification, Validation, and Confirmation of Numerical Models in the Earth Sciences. Science 263:641–646.
Peulvast, J.P, Mege D., Chiciak, J., Costard, F., Masson, P.L. 2001. Morphology, Evolution and Tectonics of Valles Marineris Wallslopes (Mars). Geomorphology
37:329–352.
Vutukuri, V.S., Lama, R.D., Saluja, S.S. 1974. Handbook on Mechanical Properties of Rocks: Testing Techniques and Results, Vol. I. Series on Rock and Soil
Mechanics. Clausthal, Germany: Trans Tech Publications.
Scheidegger, A.E. 1991. Theoretical Geomorphology (3rd edition). Berlin Heidelberg, Germany: Springer Verlag.
Schultz, R.A., 2002. Stability of Rock Slopes in Valles Marineris, Mars. Geophys. Res. Lett. 29(19), 1932, doi:10.1029/2002GL015728.
Schultz, R.A., Lin, J. 2001. ThreeDimensional Normal Faulting Models of the Valles Marineris, Mars, and Geodynamic Implications. J. Geophys. Res. 106(B8): 16,549–
16,566.
Smith, D.E., Zuber, M.T., Frey, H.V., Garvin, J.B., Head, J.W., Muhleman, D.O., Pettengill, G.H., Phillips, R.J., Solomon, S.C., Zwally, H.J., Banerdt, W.B., Duxbury,
T.C., Golombek, M.P., Lemoine, F.G., Neumann, G.A., Rowlands, D.D., Aharonson, O., Ford, P.G., Ivanov, A.B., Johnson, C.L., McGovern, P.J., Abshire, J.B., Afzal,
R.S., Sun, X. 2001. Mars Orbiter Laser Altimeter: Experiment Summary after the first Year of Global Mapping of Mars. J. Geophys. Res. 106(E10):23,689–23,722.
Starfield, A.M., Cundall, P.A. 1988. Towards a Methodology for Rock Mechanics Modelling. Int. J. Rock Mech. Min. Sci. & Geomech. Abstr. 25(3):99–106.
Page 207
Theory & development
Page 208
Page 209
On a new numerical analysis method for fracture problems
M.Hori & K.Oguni
Earthquake Research Institute, University of Tokyo, Japan
5809 679 3
ABSTRACT: A new numerical analysis method, called FEMβ, has been proposed. While it is to solve a boundary value problem of continuum
mechanics, FEMβ discretizes a displacement function in terms of nonoverlapping characteristic functions and appears as if it is a method similar to
DEM, which solves problems of rigidbodyspring model. Due to the rigorousness of discretization, however, spring properties are determined in
terms of material properties, and brittle fracture or cracking is easily expressed as breakage of spring. This paper presents the basic formulation of
FEMβ, paying attention to the discretization scheme that uses a set of nonoverlapping characteristic functions. As an example, explicit expression of
spring constants in terms of linear elasticity is shown.
1 INTRODUCTION
The advancement of computers and analysis methods enables us to solve various problems of solid mechanics. In particular, techniques of solving solid mechanics
problems that are related to structures have been matured. However, the analysis of brittle failure remains a challenging research topic. This is mainly due to difficulty in
computing the growth of crack; it requires numerous calculations to determine a crack path even though the constitutive relations of the crack surfaces are known.
In reality, the crack path is not straight. It bents, kinks or bifurcates, forming complicated failure surfaces. The presence of slight material heterogeneity often
contributes the formation of such complicated surfaces. However, measuring heterogeneity distribution is difficult, and hence the reproduction of the crack path
becomes a task practically impossible. A technique similar to that of solving ductile failure problems is needed to evaluate the effects of local heterogeneity upon crack
growth.
At this moment, there are a variety of analysis methods for brittle failure. Since analytic solutions are available for some crack problems of linear elasticity, most
analysis methods are based on FEM for continuum mechanics problems; see Arnold (1982); see also recent works of Bardenhagen et al. (2000), Duarte et al. (2000),
and Beltyschko et al. (2001). Special treatment is made in discretizing field variables, except for the use of fine discretization to obtain solution with singularity that is
inherent to fields near the crack tip. Thus, analyzing an ideally homogeneous case is computationally heavy for FEMbased methods, and effects of local heterogeneity
upon crack growth are often ignored.
There is another class of analysis methods of brittle failure, which is based on different modeling of a deformable body; see Cundall (1971); see also Sakaguchi &
Muhlhaus (1997) and Oliver (2000). This model is rigidbodyspring model (RBSM); a body is regarded as assembly of rigidbodies, which are mutually connected
through springs, and cracking or brittle failure is represented as the breakage of spring. A numerical analysis method for RBSM is called distinct element method
(DEM) in this paper. It is easy for DEM to reproduce nonsmooth crack growth path since springs break along the boundaries of distinct elements. Thus, if the element
size is of the order of heterogeneity length scale, the effects of local heterogeneity upon crack growth can be evaluated. However, DEM has a serious problem in
determining the properties of springs. No rigorous methodology has been established to relate material properties to spring constants. Table 1 summarizes the
comparison of FEM
Table 1. Comparison of FEM and DEM as analysis methods of ductile failure problems.
Input of material property Computation of crack path Implementation of heterogeneity

FEM Rigorous Difficult Difficult
DEM Ambiguous Easy Easy
Page 210
and DEM as a numerical analysis methods of ductile failure problems.
The authors have been proposing a new numerical analysis method, called FEMβ, to solve ductile failure problems; see Oguni et al. (2004). The key characteristic
of FEMβ is discretization of displacement in terms of nonoverlapping characteristic functions. While FEMβ solves a boundary value problem of continuum
mechanics, this discretization transforms a continuum model to an equivalent RBSM in the sense that the spring properties are rigorously determined in terms of material
properties. It is certainly true that computing strain or stress is not straightforward when displacement is discretized in terms of characteristic functions that do not have
smooth derivatives. FEMβ avoids this problem by using a different set of characteristic functions for strain and stress.
In this paper, we present the basic formulation of FEMβ. The discretization scheme that uses nonoverlapping characteristic functions and the treatment of calculating
derivatives of discretized functions are explained in Section 2. Making use of the discretization scheme, we derive the governing equation of FEMβ and show the
formulation of FEMβ in Section 3. The derivation is rigorous such that spring constants are determined in terms of material properties. In Section 4, we present an
alternative derivation of the governing equation, which is less mathematical and more based on mechanical consideration.
This paper uses Cartesian coordinates and index notation, i.e. x i stands for the ith coordinate. The summation convention is employed, and an index following a
comma stands for the partial differentiation with respect to the corresponding coordinate.
2 DISCRETIZATION SCHEME THAT USES NONOVERLAPPING CHARACTERISTIC FUNCTIONS
Discretization of a function is essential in numerically solving a boundary value problem. Discretization of a derivative is also important, although the derivative is usually
discretized in the same form as the function. The authors have been proposing a new discretization scheme, which uses sets of nonoverlapping characteristic functions;
function and derivative are discretized by using a different set. The scheme is suitable to represent function with discontinuity. This is because the discretized function has
many discontinuities across the boundary of the domain for the characteristic function.
For simplicity, we consider a twodimensional function, f, to explain the discretization scheme that uses nonoverlapping characteristic functions. We denote by V a
domain in which f is defined, and decompose V into blocks, {Φα}, applying the Voronoi diagram; for a given set of mother points, {xα}, a block Φα is formed as a set

of points whose closest mother point is xα. The characteristic function of Φα is denoted by is
(1)

We determine the coefficients {fα} by minimizing a suitably defined error of discretization. The following square error is used for this end:
(2)

Then, {fα} that minimize Ef are given as the average of f taken over Φα, i.e.,
(3)

where Φα stands for the Voronoi block area.
We consider discretization of derivative of f d . Since the characteristic function is used, differentiation of Eq. (1) does not yield a bounded function; the derivative of
in the following form:
(4)

As f d is determined by minimizing Ef given by Eq. (2), we define the following error of discretizing
(5)

By minimizing Eg , we obtain
(6)
with being the Delaunay triangle area.
In Figure 1, we compare discretized function in one Delaunay triangle formed by the three mother points, x1, x2 and x3. According to an ordinary discretization
scheme, which is used in FEM, a discretized function is expressed as a plane over the triangle. Three linear shape functions, which overlap each other, are
Page 211
Figure 1. Discretization of function in triangular domain.
used to express this plane. The present discretization scheme discretizes a function as three prisms, each of which is produced by a characteristic function. As is seen,
discretizing derivative is easy for the ordinary discretization scheme, but not easy for the present discretization scheme. Thus, it computes the average of derivative
taken over the triangle.
In closing this section, we emphasize that discretization scheme that uses nonoverlapping characteristic functions may not be mathematically accepted; discretized
function does not belong to a class of functions which are admissible for ordinary physical problems. However, as will be shown later, the governing equation for the
discretization coefficients can be derived from the physical conditions, and it is similar to the governing equation for the coefficients of the ordinary discretization.
3 FORMULATION OF NEW NUMERICAL ANALYSIS METHOD
We now formulate FEMβ, applying the discretization scheme shown in the preceding section to solve a boundary value problem for continuum mechanics. As the
simplest problem, we consider linear elasticity, assuming quasistatic state and infinitesimally small deformation.
Suppose that body force bi and displacement ν i are prescribed in a body V and its boundary ∂V, respectively. Then, the following boundary value problem is posed
for displacement:
(7)
Here, cijkl is the homogeneous linear elasticity tensor of the body. The boundary value problem (7) is transformed to an equivalent variational problem. Since function is
discretized in terms of characteristic functions whose derivative behaves like delta function, it is not admissible to put such function into an ordinary functional, which
uses a quadratic form of strain. Instead, the following functional for displacement and stress is used:
(8)

where −1 on cijkl stands for the inverse and εij is strain given as εij=1/2(ui,j+uj,i). Note that ui is required to satisfy the boundary conditions. The variation of I is
(9)
from which the governing equation is derived. Thus, the variational problem of I is equivalent to the boundary value problem (7).
We decompose the domain V according to the Voronoi diagram and the conjugate Delaunay tessellation, suitably distributing mother points within V. Applying the

discretization scheme shown in the preceding section, we discretize ui and σij in terms of as
(10a)
(10b)
Substituting Eq. (10) into Eq. (8), we can analytically compute as
(11)
Page 212
where is
(12)

Note that this integration is carried out just by using the geometry of a Voronoi block Φα and a Delaunay triangle
(13)
This is the governing equation of FEMβ. Thus, FEMβ solves Eq. (13) with discretized boundary conditions and obtains a solution of discretized displacement
function, just like ordinary FEM.
The matrix for corresponds to concentrated body force of FEM.
Since FEMβ uses a set of characteristic functions to discretize displacement, it appears that each Voronoi block moves as if it is a rigidbody, even though FEMβ
is formulated to solve a boundary value problem of continuum mechanics. In other words, the FEMβ provides an RBSM which is equivalent to a continuum model for
a deformable body. The matrix in Eq. (13) represents springs, which connect rigid bodies and the matrix components, which are rigorously determined in terms of
material properties, give the spring constants. Indeed, for a spring connecting the αth and βth Voronoi blocks, spring constant for the ith and jth displacement
component is explicitly expressed in terms of elasticity cijkl as
(14)
This formula of the spring constant is due to the discretization scheme that uses nonoverlapping characteristic functions. As shown in Figure 2, DEM seeks to
determine spring constants analyzing relative displacement of two neighboring elements. However, the spring constants cannot be determined in terms of
Figure 2. Schematic view of spring connecting rigid bodies.
material properties, since they relate strain to stress; three strain components cannot be computed by analyzing relative displacement of two elements. Applying the
discretization scheme shown in Section 2, FEMβ uses three Voronoi blocks to compute strain in a Delaunay triangle, and succeeds to rigorously determine spring
constants in terms of material properties.
While FEMβ provides an RBSM that is equivalent to a continuum model, it is formulated to solve a boundary value problem of continuum. Hence, some care must
be taken in breaking springs to express cracking. We should recall that the matrix is derived from the functional I which integrates strain energy density over V; see Eq.
(8). In FEMβ, therefore, fracture is regarded as domain, which has zero strain energy, and hence the contribution to I from such domain is excluded. This does not
lead to making the
Page 213
matrix component zero, but the component is reduced to a certain value; see Oguni et al. (2004) for detailed explanation.
We should mention that it is straightforward to include the rotational degreeoffreedom in FEMβ. Springs, which are related to this degreeoffreedom, are
determined in terms of material properties, as well. When infinitesimally small deformation is assumed, the rotational degreeoffreedom can be expressed in terms of
three rotational angles around the coordinate axis, and displacement is discretized in the following form:
(15)

with εijk being the permutation symbol. The coefficients as
(16)
with
(17)
The matrix equations for
(18a)
(18b)
with
Here, the origin of the coordinate is taken such that the integration of x i for each Φα vanishes. It should be mentioned that due to local rotation of Φα, and the
matrix of Eq. (18) give spring constants related to the rotationaldegreesoffreedom. For instance,
(19)
is the spring constant for the rotation (around the 3rd axis) of the αth Voronoi block to force in the ith direction acting to the βth Voronoi block; compare Eq. (19)
with Eq. (14).
For dynamic conditions, we need to add inertia to the governing equation of Eq. (7). The boundary value problem becomes an initial boundary value problem.
Instead of using an equivalent variational problem, we consider the weak form of the initial boundary value problem and derive the governing equation for the
discretization coefficients. Again, in view of characteristic functions being used in discretization, we write the equation of motion in the following form:
ρüj(x,t)=σij,i(x,t),
(20a)
σij,i(x,t)=cijklεk l(x,t).
(20b)
Here, for simplicity, the body force is assumed to be zero, and ρ is the homogeneous density. The weak form of Eq. (20) is
(21)
for each instance t. Here, pi and qij are weight functions; pi satisfies pi=0 on ∂V. Using integration by parts for the product of stress and weight function, we rewrite
Eq. (21) as
(22)

We discretize pi and qij in the same form as ui and σij, respectively. If the rotational degreeoffreedom is included, ui and qij are discretized as
(23a)
(23b)

and σij and qij are discretized as
(23c)
(23d)

Substituting Eq. (23) into Eq. (22), we can analytically compute to make coefficients of the discretized weight function qij vanish, as
(24)
Page 214

We derive the matrix equations for the discretized coefficients making coefficients of the two weight functions qij and ri vanish, as follows:
(25a)
(25b)
Here, is the moment of inertia for the Voronoi block, defined as
(26)
Equation (25) is the dynamic governing equation of FEMβ. With discretized initial and boundary conditions, FEMβ solves Eq. (26) using numerical time integration
methods; either implicit or explicit algorithm is applicable.
In Eq. (25a), the matrix in the summation coincides with the stiffness matrix of FEM with triangle elements. The first term corresponds to the concentrated mass
matrix of FEM, just like body forces for the static state being expressed as concentrated forces; see Eq. (13). On the other hand, since is rewritten as
(27)
the first term of Eq. (25b) is different from the moment of inertia for the Voronoi block.
4 ALTERNATIVE DERIVATION OF GOVERNING EQUATIONS
The governing equation of FEMβ is derived in a purely mathematical manner, i.e., just by substituting discretized functions into the functional or the weak form. There
is an alternative derivation of the governing equation, by considering the equilibrium of each Voronoi block. Thus, this derivation is intuitively appealing. In this section,
the alternative governing equation for dynamic state is presented as the most general case.
We start from writing the local equation of motion in the following form:
ρüj(x,t)=σij,i(x,t),
(28a)
εijkσjk(x,t)=0.
(28b)
Here, “local” means that Eq. (28) holds for any x in V. Note that Eq. (28b) implies the symmetry of the stress tensor, i.e. σij=σji. We take the volume average of Eq.
(28) with the aid of the Gauss theorem. The volume average of Eq. (28a) taken over an arbitrary domain D is
(29a)
where ni is the outer unit normal of the boundary ∂D. Using the identity σjk=σjlx k,l, we rewrite Eq. (28b) as εijk((σjlx k ),l−ρüjx k )=0, and calculate the volume average

over D as
(29b)
These two averaged equations hold for any arbitrary domain D. However, they are only necessary conditions for the local equation of motion, Eq. (27), to hold when
D is fixed.
Now, we substitute discretized displacement and stress, Eqs. (23a) and (23c), into Eq. (29). Replacing D with the Voronoi block Φα, we have
(30a)
(30b)
Note that the derivation of these equations requires some manipulation, in particular, to compute the integration of the characteristic functions.
The alternative governing equations for the coefficients we immediately have the same equation as Eq. (25), i.e.,
(31a)
(31b)
This is the alternative derivation of the governing equation.
For DEM, the governing equation is derived in a manner similar to derive Eq. (31b), i.e., considering
Page 215
the equilibrium of motion, translational and rotational, for each element. In the current derivation, the equilibrium of motion for element is derived as the average of the
local equation of motion; see Eq. (29), which is derived from Eq. (29). The major difference is the evaluation of external force and moment acting on the element or the
Voronoi block. With some mechanical consideration, we can derive the equation of motion for an element of DEM. However, such mechanical consideration cannot
help determine external forces and moments if only interaction between two elements is considered. We need to make mathematical analysis in computing strain from
displacement; it is easily seen that ≥3 elements are needed to calculate four displacement gradient components since three elements have two relative displacements and
they have four independent components.
On the other hand, the alternative derivation of the governing equation presented here gives FEMβ a physical interpretation. As mentioned, the formulation of FEM
β is purely mathematical, just by substituting function discretized in terms of nonoverlapping shape functions into the weak form. In the current derivation, FEMβ can
be derived from the equilibrium of each Voronoi block, with stress being evaluated in Delaunay elements, which are partially included in the Voronoi block.
5 CONCLUDING REMARKS
On the view of point of constructing RBSM, the rigorous determination of spring properties of FEMβ is worth being carefully examined. The determination can be
extended to a case when material is nonlinear plastic or a case when deformation is large and strain is related to displacement in nonlinear manner.
We should emphasize that the essence of FEMβ is the discretization scheme that uses a set of nonoverlapping characteristic functions. The scheme can be applied
to other boundary value problems of classical mathematical physics. As an RBSM that is equivalent to a continuum model is created, it renders a particlephysicslike
perspective to various problems.
In this paper, the Voronoi diagram and the conjugate Delaunay tessellation are used to determine sets of nonoverlapping characteristic functions. However, there is
no necessity to use them, i.e., other pair of the domain decomposition are admissible. We need further study to find the domain decomposition that gives the most
suitable discretization of function and derivative.
REFERENCES
Arnold, D.N. 1982. An interior penalty finite element method with discontinuous elements, SIAM, J. Numer. Anal. 19: 742–760.
Bardenhagen, S.G., Brackbill, J.U. & Sulsky, D. 2000. The materialpoint method for granular materials, Comput. Methods. Appl Mech. Engng, 187:529–541.
Belytschko, T., Moes, N., Usui, S. & Parimi, C. 2001. Arbitrary discontinuities in finite element, Int. J. Numer. Meth. Engng. 50:993–1013.
Cundall, P. 1971. A computer model for simulating progressive, large scale movements in blocky rock systems, in: Proc. International Symposium on Rock Fractures,
Nancy, France.
Duarte, A.V.C., Dutra do Carmo, E.G. & Rochinha, F.A. 2000. Consistent discontinuous finite elements in elastodynamics, Comput. Methods. Appl. Mech. Engng. 190:
193–223.
Oguni, K., Hori, M. & Sakaguchi, H. 2004. New analysis method of failure problem—proposal of FEMβ, J. of Struct. Eng./Earthquake Eng., JSCE (in print).
Oliver, J. 2000. On the discrete constitutive models induced by strong discontinuity kinematics and continuum constitutive equations, Int. J. Solids and Structures, 37:
7207–7229.
Sakaguchi, H. & Muhlhaus, H.S. 1997. Mesh free modelling of failure and localization in brittle rock, in Deformation and Progressive Failure in Geomechanics, (ed. by
Asaoka, A., Adachi, T. & Oka, F.), Pergamon, New York, 15–21.
Page 216
Page 217

A selfsimilar bond model for PFC2D
P.A.Cundall
Itasca Consulting Group, Inc., Minneapolis, Minnesota, USA
5809 679 3
ABSTRACT: The default bonding laws in PFC are quite simple and the behavior of an assembly of particles, each using a default bond model, is quite
different from that of the model at each contact. A more consistent law is developed by testing a model of the bond region itself, using many PFC2D
particles to represent the cementlike material constituting the bond. By observing the damage, and recording the relations between forces/moments
and displacements/rotations for tests on the macromodel of a bond, a damage formulation is developed that accounts for the observed behavior. By
adjusting various free parameters, the prescribed response of each contact and the measured response of the “bond object” are matched. The contact
model is then said to be selfconsistent. Several interesting features of a selfconsistent contact model are observed. First, there is a definite ratio
between shear and normal contact stiffness. Second, tensile behavior is seen to be brittle, but shear and rotation responses are more plastic. Third,
dilatancy is observed during shear and rotation. It may be noted that the new contact model provides a moment reaction to relative rotations of the
attached particles; it is thus equivalent to a model with a distributed contact area (or multiple contact points). The model is offered as a basis for
research into new ways to represent bonds, not as a fullydeveloped formulation. Only a two dimensional implementation and formulation is presented.
1 INTRODUCTION
There are two forms of bonding built into PFC (Itasca 2004)—the contact bond and the parallel bond. These supply simplified representations of cement between two
particles that resist relative movement between the particles and allow for breakage. When an assembly of such bonded particles is constructed and tested, the
macroscopic behavior exhibits many of the characteristics of real material, such as strength dependence on mean stress, plasticity and coupling between components of
strain (e.g. dilation), even though the bond formulation does not represent such effects directly. This emergent behavior is commonly observed in systems composed of
many simple elements: the macro behavior is more complex than the micro behavior.
It is possible to use PFC itself to investigate the constitutive response of cementlike material that forms a bond. Thus, a multiparticle model of an individual bond
may be exercised under various modes of deformation—tensile, shear and bending—and the response noted. A constitutive contact model may then be developed
from the observed response and used, in turn, to represent the bond behavior within a new multiparticle of an individual bond. It is interesting to inquire if a constitutive
law for the contact model (expressed in nondimensional form) can be found that is the same as the observed behavior of a multiparticle assembly representing a single
bonded contact. If so, the resulting contact model may be said to be selfsimilar, in that the same behavior occurs at all scales. Further, the exercise of matching may
reveal certain relations between micro parameters that are required in order to achieve selfsimilarity. The question of whether selfsimilarity exists in nature is not
considered; the present exercise is intended to provide insight into the formulation of a bond model, not to reproduce any particular material or form.
In this paper, the two dimensional form of the selfsimilar bond is developed; there should be no difficulty in extending the approach to three dimensions. In order to
avoid confusion, the multipleparticle model of a single bond will be referred to as a “bond object,” to distinguish it from a regular PFC contact between two particles.
The selfsimilar bond model is not intended as a fullydeveloped model and is simply offered as a basis or source of ideas for developing such models. The topic of
matching to lab data is not addressed.
2 INITIAL NUMERICAL TESTS
Figure 1 shows the conceptual model for a bond object, in which a thin region of cement (shown in black) joins two disks. Since the bond object represents the bond
Page 218
Figure 1. Conceptual model of a bond object (shown in black).
Figure 2. Compacted assembly used in numerical testing of bond object (bonds on left and particles on right).
between two PFC disks, the vertical boundaries are regarded as rigid and are moved in extension, shear and rotation, to reproduce the three modes of relative particle
motion in two dimensions. An assembly of 10:1 aspect ratio with 2000 disks is compacted and used in all subsequent tests. The assembly is shown in Figure 2. The
lockedin forces are very low.
Initial tests of the bond object are made under conditions of tension, rotation and shear. The tests
Figure 3. Crack patterns developed under three modes of deformation: (a) tension, (b) bending, and (c) shear.
are actually performed with the contact model to be described later, but the results would be similar if PFC’s parallel bond model were used. At this stage, we are
concerned only with qualitative behavior; the topic of quantitative matching will be addressed later. Figure 3 shows crack patterns that result from three modes of
imposed boundary displacement. In (a) the gray boundary particles are moved rigidly apart at constant velocity. In (b) the set of left boundary particles is given a
clockwise rigidbody rotation, and the right set is given a counterclockwise rotation. In (c), the lefthand boundary particles are moved downwards at constant
velocity, and the righthand boundary particles are moved upwards.
In tensile loading, cracks form parallel to the long axis of the bond object, while cracks progress from one end of the object when boundary rotations are imposed
(corresponding to opposite spins of the two particles connected by the bond object). Shear deformation causes a series of tensile cracks to develop at approximately
45° to the long axis of the object.
3 BASIC FORMULATION
The results described in Section 2 indicate that—for tension and bending—the behavior may be described
Page 219
Figure 4. Bond object with cracks extending from two ends.
by the evolution of cracks parallel to the bond object For the purpose of analysis, a set of independent elastic/brittle fibers is assumed to extend from one side of the
bond object to the other, perpendicular to the long axis of the object. Failure occurs by progressive “deletion” of fibers from either end of the bond object. A similar
scheme was devised by Trent (1987), but he assumed the crack extended from the center of the bonded region, not from the outer fibers. Two variables, ip and in ,
record the proportion of intact fibers in the “positive” and “negative” directions from the center point: see Figure 4. These variables start with values of unity (denoting a
full population of fibers) and progressively (and independently) evolve towards zero as complete failure is approached. It is assumed that broken fibers contribute no
load.
Considering extension of the bond object perpendicular to its long axis, the relation between the normal force, Fn , and the accumulated normal displacement, μn , is:
(1)
where t is the outofplane thickness of the object, is the “radius” of the object (i.e., its halfheight) and k n is the tangent stiffness of the layer of fibers (stress per

unit normal displacement). If relative rotation, θ, of the long edges of the object is superimposed, the additional normal force is found by integrating along the height, y,
of intact fibers:
(2)
Similarly, the moment, M, generated in response to both normal displacement and relative rotation is found by integration, so that the composite relations for normal
force and moment become:
(3)
Note that there is a crosscoupling term of magnitude
(4)
where k s is the tangent shear stiffness of a fiber layer and ∆us is the incremental shear displacement. A further plastic coupling between shear force and normal force is
found to be necessary for selfsimilarity, but this will be introduced later.
The evolution of the damage parameters, ip and in , occurs when any tensile fiber stress, σL—at distance L from the centerline—exceeds the tensile strength, σt.
Assuming that positive un and positive θ each cause tensile stress in the nsection of the bond object, the condition for failure in this section is:
(5)
where
(6)
Thus, the condition for further fiber breakage in the nsection is:
(7)
In this event, the new value of the nsection damage, is:
(8)
noting that for the special case of θ=0. Similarly the condition for failure in the psection (θ<0) is:
(9)
Page 220
and the new psection damage is:
(10)
Since the shear force is calculated incrementally, a reduction in intact fiber length must produce a corresponding reduction in shear force, as follows:
(11)
The formulation described above is embodied in a DLL (Dynamic Link Library) model, available to the user as ssbwrv.dll, with the model name pfb. Note that a shear
strength, σs, is also assigned to fibers, but shear failure does not affect damage. When fully damaged, the bond model reverts to a frictional contact model, governed by
k n , ks and friction,
4 CALIBRATION AND COMPARISON
The objective is to match the response from a single contact containing the pfb model with the response of the 2000disk bond object, in which all individual contacts
contain the pfb model.
4.1 Stiffness and strengthratio similarity

First, tests are performed to determine the elastic properties of the bond object for various ratios of contact normal stiffness to shear stiffness, k n /ks. Figure 5
Figure 5. Relation between the (input) normal/shear stiffness ratio of contacts and the observed stiffness ratio of the bond object (solid
line). The dashed line denotes equal values.
shows the relation between macro stiffness ratio and the micro stiffness ratio (solid line).
Clearly there is a point (k n /k s=2.55) where the two stiffness ratios are equal. This particular value of stiffness ratio is used for the remaining tests.
In the next series of tests, the ratio of normal strength to shear strength (σt/σs) of bonds is varied and the corresponding ratio noted for the bond object. Figure 6
records the results. The flat part of the curve for low normal/shear ratios reflects the fact that failure in both modes of deformation is dominated by micro failure in
tension; i.e. the shear strength of individual bonds plays no part in the shear strength of the complete bond object (c.f., the tension cracks shown in Fig. 3c). For this
realization of a bond object, the selfsimilar strength ratio is 1.35.
4.2 Strength and postfailure similarity
We now seek to obtain selfsimilarity in terms of strength and postfailure curve shapes, including the coupling effects (e.g. effect of rotation on normal stress, etc.). The
result of a direct tensile test on a single bond is shown in Figure 7, plotting the nondimensional tensile stress (σ/σt) against nondimensional displacement (dkn /σt),
where σ is the measured stress and d the measured displacement, respectively. The expected elastic/brittle behavior is seen.
The mapping of individual bond strength to strength of the complete bond object depends on the packing of particles in the bond object, so normalization is done by
dividing by the observed tensile strength for the bond object, Using identical bond properties as for the singlebond test, the results for the bond object
are shown in Figure 8. The response is also brittle, but with a finite softening slope.
Figure 6. Relation between (input) normal/shear strength ratio and observed ratio for the bond object (solid line). The dashed line denote
equal values.
Page 221
For the bending test (opposite rotations applied to the two bond boundaries), the normalizing factors may be derived from Equations (3), (5) & (6). Thus, the non
dimensional rotation, are given by
(12)
(13)

There is a normal force, Fn , induced by the rotation, and this is presented as a nondimensional stress, given by:
(14)
Figure 7. Tensile response of single bond; nondimensional tensile stress versus nondimensional displacement.
Figure 8. Tensile response of bond object; nondimensional stress versus displacement. The induced shear stress & moment are very
close to zero, and are therefore not shown.
For a single bond with applied rotation, the results are shown in Figure 9. Note that the induced normal stress is compressive; i.e. dilation would occur if the particles
were not constrained in the normal direction.
A similar bending test is performed with the bond object (2000disk assembly). The normalizing factors are similar to those presented in Equations (12), (13) &
(14), with the strength and stiffness values being those derived from the tensile test on the bond object—i.e. no further calibration is performed in order to obtain the
bendingtest results. Figures 9 & 10 show good agreement, both in the form of the response and in the numerical values.
Finally, shear tests are performed both on individual bonds and on the bond object. An almost plastic response is seen after yield for the bond object, even when the
individual bonds behave in a brittle manner.
Figure 9. Results of bending test on single contact; moment (black) and normal stress (gray) versus rotation (all normalized, as
described in the text).
Figure 10. Results of bending test on bond object; moment (black) and normal stress (gray) versus rotation (all normalized as described
in the text).
Page 222
Figure 11. Results of shear test on single contact; shear stress is in black and normal stress is in gray, normalized as described in the text.
Thus, the bond model is given a perfectly plastic yield response: i.e.
(15)
where is the new shear force. Shear yielding is also observed to give rise to a normal force (under normal confinement). This behavior is reproduced by modifying
the accumulated normal displacement, which then causes a change in normal force, according to Equation (1):
(16)
where Φ is a dilation coefficient (0≤Φ<1). A shear test is performed with the measured shear force converted to a normalized shear stress as follows:
(17)

and the shear displacement normalized by dividing by σs/ks. The results are shown in Figure 11. The nondimensional normal stress is given by:
(18)
A similar shear test is performed on the contact object. The displacement to failure and the yield stress both depend on packing, so the normalization uses the observed
values of these quantities. However, the normal stress is made nondimensional according to Equation (18), using the independent measurement of bondobject tensile
strength, derived previously
Figure 12. Results of shear test on contact object; shear stress is in black and normal stress is in gray, normalized as described in the text.
Table 1. Parameters required for selfsimilar bond model.
k n Normal stiffness (stress/displacement)
k s Shear stiffness (stress/displacement)
σt Tensile strength (stress)
σs Shear strength (stress)
Bond “radius”—halfwidth (length)
Φ Dilation parameters (dimensionless)
Friction coefficient (when bond broken)
for the tensile test on the bond object. Comparing Figures 11 & 12 show that satisfactory selfsimilarity is achieved for shear behavior. However, it should be noted
that the match for shear behavior is largely cosmetic, because most microfailure in the bond object under shear loading occurs when the tensile limit of individual bonds
is exceeded. Thus, almost the same response as shown in Figure 12 is seen if the bond shear strength is set to infinity.
4.3 Model parameters
The necessary material parameters, introduced in previous sections, are summarized in Table 1. Apart from Φ and these parameters are similar to those needed for
PFC’s parallel bond model. However, it is not necessary to supply additional contact properties (as with the parallelbond model) because the behavior of the self
similar model reverts to a linearstiffness/frictional model when damage is complete.
5 CONCLUSIONS
A relation between forces and moments at particle contacts has been developed that represents bonding in such a way that the nondimensional response of a
Page 223
single bond is almost identical to the response of a “bond object” composed of many particles attached together with such single bonds. During the process of
establishing the equivalence of the micro and macro versions of the bond, certain observations were made.
1. The ratio of normal stiffness to shear stiffness of a bond must have a specific value in order to obtain the required equivalence. For the specific packing used, the
value is 2.55.
2. Similarly, there is a specific ratio of bond tensile strength to shear strength to obtain equivalence (1.35, for the example packing, and for the stiffness ratio given).
3. Using the given stiffness ratio and strength ratio for individual bonds, almost all microfailure is by tension, even for shear loading of the macroscopic object.
4. Coupling between various components (e.g. between bending displacement and normal stress) is necessary for equivalence. No such coupling exists in PFC’s
parallelbond model.
The new selfsimilar bond resembles the parallel bond (built into PFC) in that it supplies bending resistance (and associated failure), but when broken, the selfsimilar
bond reverts to an elastic/frictional contact. Thus, only one set of parameters is needed for the selfsimilar bond; for example it is not necessary to supply contact
stiffnesses, which are usually needed in addition to parallel bond stiffnesses when parallel bonds are used. Further, the selfsimilar bond incorporates damage
parameters that evolve, to give the bond a progressive weakening behavior.
The selfsimilar bond was developed out of a notion that material behavior might be similar at all levels, apart from molecular levels. We have shown that it is
possible to achieve this scaleindependence, but have not attempted to show that this behavior actually occurs in real materials. Very few tests have been performed
with the selfsimilar bond model, so it is too early to say if some undesirable behavior will be observed when the model is exercised more fully. The model is offered as
a basis for research, not as a fullydeveloped formulation.
REFERENCES

Itasca Consulting Group, Inc., 2004. PFC2D—Particle Flow Code in 2 Dimensions, Version 3.1, User’s manual. Minneapolis: Itasca.
Trent, B.C. 1987. The effects of microstructure on the macroscopic behavior of cemented granular material. Ph.D. Thesis, University of Minnesota.
Page 224
Page 225
A grid computing architecture for applications in distributed computation
X.Lin & J.R.Williams
Massachusetts Institute of Technology, Cambridge, Massachusetts, USA
5809 679 3
ABSTRACT: Grid computing has been proposed as a framework for distributing computations over a network of computers. This paper reviews the
ideas of Grid Computing in the context of numerical simulations and presents a solution based on Microsoft’s Common Language Runtime (CLR) and
recent standards such as XML, SOAP messaging and Web Services. The goal is to provide a middleware layer for managing computing cycles over a
network of machines and to provide a good programming model for this new framework. To ease the programming burden, an alternative to the
Message Passing Interface (MPI) is proposed that has higher level abstractions for routing and controlling messages between worker threads on
different machines. The solution extends the common scope of “Parallel Computing” and “Distributed Computing” by incorporating remote service
lookup and management of computational processes over the Internet using Web Services that are accessible from a wide range of devices, including
PDAs and smart phones. Comparisons with MPI and with the Open Grid Services Architecture are provided. Finally, the potential of Grid Computing
in discrete computational mechanics is discussed and examples of programming LatticeBoltzmann fluid flow and multibody particle dynamics are
shown.
1 INTRODUCTION
This paper addresses the problem of distributing numerical simulations and other computations across both computer clusters and more loosely coupled machines. This
style of computing is often referred to as Grid Computing (Foster et al., 2004a, 2004b). In Grid Computing, the assumptions that are usually made in cluster computing
are not valid, including:
– Dedicated machines running a single operating system with all resources available
– All machines reside on the same local network and security is not a concern
– Reliable message passing exists.
In Grid Computing we envisage computers that “belong” to many different owners in different administrative domains that are willing to provide compute cycles but
perhaps with restrictions on resources, such as disk access.
In general a user distributes code to be executed across a network of machines and coordinates the messaging between the machines so that some computational
goal is achieved. The coordination of the machines can be achieved by using a master machine (Fig. 1). Here we assume that the Master is a “special” trusted machine
that has coordination and other responsibilities not shared by the Workers.
To protect the worker machines from malicious code, each worker machine loads the code to be
Figure 1. Master and worker machines that form a simple Computational Grid.
executed into a “sandbox” that protects other resources on the machine from attack or infection by limiting the resources available to threads within the sandbox.
A major problem in distributed computing concerns messaging. Messages must be transferred between machines. These messages may be as simple as a packet of
data or more complex, such as executable code or streaming data. In a simulation environment the passing of these messages must be coordinated so that Worker
machines remain synchronized. “Wiring” together the Workers and coordinating the messages between them is a major issue for programmers, especially those writing
time stepping simulations where synchronization at the end of every time step is required. Indeed, we contend that simplifying this
Page 226
programming burden should be a major goal of any Grid environment.
2 GRID COMPUTING
Grid Computing extends the concept of cluster computing to a scale that crosses organizational boundaries. The Grid refers to an infrastructure that enables the
integrated, collaborative use of highend computers, networks, databases, and other resources, such as scientific instruments, that may be owned and managed by
multiple organizations. Grid applications often involve large volume data (>Terabyte) and often require secure resource sharing across organizational boundaries that are
not easily handled by today’s Internet and Web infrastructures (Foster et al., 2004a, 2004b). Examples of where Grid Computing is used include the following:
– Collaborative engineering Highbandwidth access to shared virtual spaces that support interactive manipulation of shared datasets and steering of sophisticated
simulations, for collaborative design simulations or for collaborative design of complex systems (typical Data Grid.) e.g. Sloan Digital SkyServer
http://skyserver.sdss.org/dr2
– Distributed supercomputing: Ultralarge virtual supercomputers constructed to solve problems too large for any single computer (typical Computational Grid), e.g.
Discrete particle analysis
– Parameter studies: Rapid, largescale parametric studies, in which a single program is run many times in order to explore a multidimensional parameter space, e.g.
Monte Carlo analysis.
2.1 Current state of Grid Computing
Although the Grid Computing community generally agrees with the definition above, there is no common concept about the detailed form of the Grid. As a result of this
confusion quite a few standards and implementations are available.
The Globus Alliance (http://www.globus.org) has published their own standard, the Open Grid Services Architecture (OGSA), and distributes a toolkit called the
Globus Toolkit (GT1–3). The standard, as implemented in GT1 and GT2, was originally formulated without reference to the industry standard of Web Services.
However, the most recent release (GT3) has moved fairly aggressively towards Web Services as the underlying architecture. Although GT3 incorporates many recent
standards, such as XML and Web Services, it has a complicated structure that contains much legacy GT2 code. OGSA to some extent overextended its reach by
proposing broad standards before any implementation experience was available. This detailed standard is vulnerable to the risk of deviating from the real needs of users
and also to the quickly evolving technologies and standards associated with Web Services, such as WSSecurity, WSRouting, WSCoordination…etc. One
consistent set of standards is being developed by Microsoft under the title of Global Web Services Architecture (GXA). However, OGSA is proposing others, such as
WSResources that are not compatible with GXA and therefore there is some confusion as to which standards will prevail.
Sun’s N1 is another Grid Computing framework that Sun Inc. promotes. However, to date neither OGSA nor N1 has not clearly established a leadership position in
either industry or academia. To complicate matters even further, industry has just formed its own organization called Grid Computing Association to promote its own
standards.
Other parallel computing frameworks are MPI and Matlab*P. However, they have limited functionality for resource control, have no security infrastructure and they
do not comply with recent open standards, such as XML.
After studying these technologies and their problems, we proposed a new Grid Computing framework, called GridGarden. NET with the goal of having a small
footprint and addressing the needs of computational software developers.
The Microsoft .NET Framework provides an excellent foundation for this effort. The .NET Framework combines recent technologies, such as XML, SOAP and
Web Services. .NET provides numerous utilities that comply with these standards and makes them easy to utilize. Furthermore, .NET provides an object oriented
programming language, C#, which is powerful and efficient. The compiled code, based on the Common Language Runtime (CLR), is similar to Java byte code. It is
platform independent, efficient, highly reusable and perhaps most importantly is “managed” code. Managed code means that the computer rather than the programmer
manages the garbage collection of unused objects. It avoids the memory leaks and “dangerous” pointers of C++ and C. Lastly, although .NET Framework is not open
source and only runs on Windows computers, there are several Open Source implementations of .NET Framework on other system, such as Mono (which is now a
software product of Novell, Inc.) on Linux and Rotor on FreeBSD (Rotor is provided by Microsoft under its “Shared Source” policy).
In the next sections we describe the programming and messaging issues associated with developing a Grid Computing environment and detail our solutions to some,
but by no means all, of the problems that arise.
3 MESSAGING AND COORDINATION
There are several models of parallel computing that can exist on the Grid topology described above. One simple
Page 227
model is where every Worker runs the same code but operates on different data, as would be the case in say Monte Carlo parametric studies or in the SETI search for
extraterrestrial life. Each worker thread computes independently and only communicates with the Master. These kinds of problems are sometimes called
“Embarrassingly Parallel”. In the case of discrete particle computations, and indeed for most mechanics computations of interest, this is not the case.
In DEM computations, communication between Worker machines occurs at the end of every time step, when particles that have moved across domain boundaries
are moved from one Worker to another. This kind of problem requires communication between the Worker threads and also coordination of those communications,
i.e. a thread handling one domain must not proceed until all the necessary data from other domains has been received.
The design of the messaging system is critical to Grid Computing (and to Service Oriented Architectures) and there are several associated problems associated. The
next section describes in more detail how modern messaging is achieved at the application level.
3.1 Remote procedure calls
There are many layers involved in the messaging stack but from the programmers perspective messages in computations are passed from one machine to another as
either an argument or return value of a function call. The heart of CORBA, COM+ and today’s Web Services is the remote procedure call or a call to a method on a
remote service (Web Method call). Below we show a call made from the Client to a remote services (Service).
ResultType result;
result = (ResultType) Service.Doit
(Type1 object);
Here we call the function Doit that is a function on the Service object that exists on the remote machine. We pass to the remote machine an object of Type 1. This
object can contain an array of other objects and can transport particle data, say, to the remote machine. An important concept is the “Proxy” object on the Client
machine. This Proxy allows us to program on the Client as if there is a local copy of the remote object. The Web Service Description Language (WSDL) allows the
Remote Service to publish its interface, as a WSDL file, and the programmer of the Client can use this to instantiate a Proxy object that matches that interface.
The .NET messaging “plumbing” then arranges for a message to be correctly packed using the SOAP protocol and sent to the Remote Service. At the remote machine
the SOAP message is unpacked and sent to the correct application that is hosting the actual
Figure 2. Message passing using the SOAP protocol and the Proxy object on the client to fire a method on the Remote Service.
Service object, fire the correct function (Doit) with its associated argument (Fig. 2).
In this example, the Client thread waits for the response object from the Remote Service that is again sent as a SOAP message. The message is unpacked and an
object of type ResultType created. Alternatively, we could arrange for our call to the remote service to be made asynchronously.
We believe that this style of programming is far easier than the MPI style that is predominant today. In our GridGarden.NET environment we leverage the concept of
Proxy objects to provide a clean and simple style of message passing.
4 DESIGN AND IMPLEMENTATION
The scope of a computational Grid architecture can be very broad. Therefore, the feature list was restricted to the following:
– Provide a framework that has the basic functions of MPI, such as process control, communication among processes and data management.
– Provide Grid capabilities such as code distribution, Internet ID addressing of the distributed application threads, and good process control and security utilities.
– Provide messaging coordination and control for a wide range of messages size (approximately 8 orders of magnitude).
– Provide good programming abstractions for numerical simulation and general message passing.
4.1 Overview
Some terms are useful to understand the design:
– Grid Application—A group of computer processes whose purpose is to achieve a single computing task that are not necessarily running on one computer but on a
grid of computers over the network.
Page 228
Figure 3. Components of GridGarden.NET.
– Net Application Domain (NAD)This is the only identity that is used to run or stop a Grid Application, or access Global Data.
– NAD Global Data—A collection of data that can be accessed over the Internet using Web Services.
– Seed—A .NET CLR class that has necessary entries required to be loaded by the GridGarden.NET system onto a Worker machine (Fig. 3).
This is a masterslave architecture in which we call the master computer a “Proxy” and the slave computer a “Worker”. We use Proxy for the Master because it
provides users with “proxy” objects for manipulating the Workers. For example, a user can, over the Internet, command one or more of the Worker threads to
“Start/Resume”, “Pause” or “Kill”.
The computing is done through a number of “worker processes”. Workers, the computers that hold the worker processes, are distributed across the network by
URLs. For the current design, two communication channels are provided to communicate with the Proxy, namely, .NET TCP Remoting channel and HTTP Web
Service channel. The TCP Remoting channel is used for the local area network (typically the Workers) due to its high performance, while the HTTP channel is used for
communication with Internet clients, since the SOAP based Web Service is an open standard.
Programs running as worker processes are coded in the Common Language Runtime (CLR), which is platform independent (though presently limited to Windows,
Linux and BSD). A grid program is normally a CLR dynamic loaded library (DLL) that contains the minimum functionality needed by GridGarden.NET system to
Start/Resume, Pause and Kill it. We call the DLL a “Seed”. The GridGarden.NET framework stores the DLLs in the Seed Pool for later distribution to the Workers.
A number of abstract seed classes are provided as templates for the programmer. Typically, the programmer will inherit and extend the abstract classes to their specific
needs, such as particle simulation.
One important idea to be introduced is the “Net Application Domain” (NAD). Similar to MPI’s “Communication Group”, the NAD facilitates identification and
communication of worker processes for a Grid Application. One difference between NAD and MPI Communication Group is that the NAD is not just for
communication. It also provides a shared memory block, similar to a global COMMON, for worker processes, and its data can be exported over the Internet. The
share memory block is called “Global”. It is critical for Grid Computing so that results from all the Workers can be aggregated and fed on to other systems, such as a
graphics server or a database server. Grid users accessing information about a simulation or controlling the application can do this by Read or Write to the Global data.
Data stored in Global can be arbitrary serializable objects and can represent any data type.
Controlling the Grid Application is currently achieved through a Windows GUI that allows the Grid Application administrator to Run, Pause, Kill etc, which forms a
mini distributed operating environment that is quite similar to UNIX. One difference is that GridGarden.NET commands are managing distributed the application
through Web Services.
4.2 Process (computing threads) management
The process management scheme proposed in this design has the following features:
– Each individual Grid application is identified by NAD ID, which is unique across the network.
– Each individual computing thread is identified by an integer starting from 0. This is similar to MPI.
– The SandBox is a client that listens to commands through a TCP .NET Remoting Channel on a specific port.
– Computing threads run under the control of the “SandBoxes”. The idea of the SandBox is borrowed from the Java Applet. Introducing the SandBox not only makes
the managing of individual thread much easier but it also provides a security boundary between the Worker’s system and the “foreign” executable it is hosting.
– Each worker machine listens (on port 8075) for commands from the Master Proxy. The commands are usually for creation and dumping of worker processes
(computing threads).
Upon receiving a job request from the Proxy on port 8075, the worker machine fetches the proper DLL (Seed) from the Seed Pool via a Web Service.
When commands or data are sent to the SandBox it passes the information to computing threads. This is done by activating correct functions of the Seed
Page 229
Figure 4. Thread creation illustration.
Figure 5. Run time control components.
objecton the Proxy machine. SandBoxes (in .NET these run in separate AppDomains) on the Worker are created if successful are recorded on the Proxy as available
resources. Computing threads are then created inside the SandBoxes, following the security policies imposed by SandBoxes (Fig. 4).
To allow users to issue instructions to a grid application, there must be an interface between the computing threads and user. To achieve run time control, the Proxy
is used for hosting the interface (Fig. 5). In fact, as shown in Figure 3, the Proxy has two sets of process control services. The external services are exposed through an
HTTP channel as Web Services. This allows users to control the Grid Application from the Internet using any tools provided they follow our standard. The internal
services are mainly for synchronization among computing threads. They are exposed as Remoting service and only used inside the local network. Specifically, the “Sync
Clock” is used in time step based simulation; the “Locker” is for more general consensus based synchronization.
Figure 6. Thread creation illustration.
4.3 Intercommunication between computing threads using the Event Subsystem
Most parallel computing algorithms are not “embarrassingly parallel”, which means they require intercommunication among computing nodes. There are a number of
issues in messaging between Worker threads, namely:
– Routing the message
– Buffering messages
– Dealing with both small and large messages
– Dealing with unreliable messaging so that a “conversation” can be reliable guaranteed.
Figure 6 shows three methods for a thread to send data to another thread. The first method is “quick messaging”. Small volume data can be sent to the other thread via
the Proxy. The message is “pushed” from the Worker to the Proxy. The Proxy then “pushes” the message into an “InBox” on the correct Worker. This mechanism is
not suitable for large volume of data since the Proxy may be overloaded by dealing with Worker communication.
The second method deals with large volume data and is called “big object messaging”. When sending large volume data, the initial thread merely sends a notification
message, via “quick messaging” to the other Worker informing it that there is data ready to be pickedup. It is then the responsibility of the Worker receiving the
notification to “pull” the data from the other Worker in peertopeer mode. This avoids congestion at the Proxy both in terms of network traffic and demand on CPU
cycles.
The last method is direct peertopeer connection. This is allowed in the framework but is not recommended because worker machines workload is typically high
because they are running the simulation, while the Proxy acts to direct small data traffic. If a worker fails, other workers sending messages to it may
Page 230
block or even crash. The Proxy, however, is not take part directly in the simulation and also has mechanisms to detect failure of Workers. This is done presently by
pinging the service of each worker machine at regular intervals. Other mechanisms, such as “lease managers” are used in distributed computation (.NET uses these for
managing remote objects) and may be implemented in the future.
If a Worker fails then the Proxy assures that the other computing threads will be notified of this failure if they attempt to send messages to the failed machine. The first
two messaging methods described above involve short messages to and from the Proxy and this technique is very similar to event system used in many operating
systems. For this reason, this scheme was named “Event Subsystem”. Communication reliability is guaranteed by a reference number attached to each message and a
resending loop on the sender side. On the receiver side, special handling of the received message ensures the uniqueness of a single message as well as the order
among messages.
5 PROGRAMMING MODEL
The targeted users of this Grid Computing framework are engineers who are not necessarily professional programmers. Therefore, programming a Seed should be as
straightforward as possible, in particular there should be good abstractions that hide the system internals.
For example, sending an object to worker process “2” from “1” is as simple as:
…
// Identify myself.
if ( this.id = = 1)
// Send object “obj” to “21”.
this.SendObj( 2, Obj);
…
Receiving an object is very different in our system, compared with MPI. A data type, “InBox”, is provided as a container for received objects. Upon receiving objects,
the system put them into the correct InBox. The InBox acts as a message buffer with queuing characteristics that can be easily set by changing the InBox’s instance
fields. For example, InBox.order specifies how arrived objects line up:
...
//First in first out.
inBox1.order=InBox.FIFO;
//Only preserve the last object.
inBox2.order=InBox.LAST;
...
//Oldest object.
obj1=inBox1.pop();
//Last object.
Obj2=inBox2.pop();
...
We can achieve blocking receiving, as in MPI, by modifying InBox.mode:
...
//Blocking receiving.
inBox1.mode =InBox.BLOCKING;
//Nonblocking
inBox2.mode=InBox.NONBLOCKING;
...
//Blocked here if no object.
obj1=inBox1.pop();
//Always return immediately,
//null if no object.
Obj2=inBox2.pop();
...
Event hooks are built into the framework to allow processuponreceiving using the event handling mechanism in C#. In this case the inbox.Ereceiving delegate will be
fired upon message receipt. In the example below we bind our own function “receiving” to the event handler list so that it will be fired:
{
...
//Add event handler.
inBox1.Ereceiving +=
System.EventHandler(receiving);
...
}
// user defined function
private void receiving(...,...)
{
//handle the event.
...
}
Programming in the Computational Grid framework we developed is efficient. Some programmers pay more attention to the reusability and extensibility of the
framework. To address this concern, we provide an extensible layer between the system kernel and user side program—the GridLib package. It has a tree design
presented in Figure 7.
At the root level, programming is in direct contact with GridGarden.NET’s kernel systems, such as the thread control mechanism, the communication system and the
network topology. While most programmers do not need to program at this level, access is provided to give programmers full control of the system. At the leaf
Page 231
Figure 7. Structure of GridLib package.
level, programming is easier with a lot of functionality available. For example, there is a Seed type used as an alternative to MPI that appears in the third level. It is
named “MultiStore” because the Seed uses multiple storage entities for incoming objects. GridLib is open in the sense that any programmer can add his own functions
into it.
The coding of the framework is mostly done in C#. However, C++ and VB for .NET can be used since they support and can be compiled into CLR code. There
are some other language bindings for CLR that are mostly Open Sourced, such as Perl, Python and even PHP. These are also potential language for GridGarden.NET
developers but at present have not been tested.
High performance programming libraries, such as Lapack, can be easily embedded into GridGarden.NET by inter operation. The current system contains a matrix
manipulation class “Matrix” in GridLib as a wrapper for a native code function written in C. This Matrix class is a good illustration of how to write a wrapper for legacy
libraries.
6 EXAMPLES
Two computational mechanics examples deployed in GridGarden.NET are shown below. The first is a LatticeBoltzmann simulation (Qi et al., 1998, Noble et al.,
1998, Cook et al., 2004) of fluid flow across a rectangular box (Fig. 8).
The calculation was separated into 8 vertical domains with each domain running on a separate Worker. A fixed horizontal velocity boundary was placed across the
top of the box. The thin lines denote the calculation domain boundaries across which the computing threads must exchange data.
The results from the Worker machines were collected in the Global data store for integrated graphics rendering. Alternatively, the results could have been separately
rendered on each Worker so that each worker displays its own results.
Figure 8. LatticeBoltazmann fluid flow in a box with the top boundary having a fixed velocity.
The second example, shown in Figure 9 illustrates rigid particle dynamics. The particles are originally allocated with a random velocity in a looped pipe so that each
domain has a given color. As particles move from one domain to another they must be moved across the Workers.
7 DISCUSSION
One concern in programming using highlevel abstractions is that the computational speed of execution is degraded. In particular the efficiency of .NET and the CLR
have rarely been tested in high performance computing. The CLR code is typically partially compiled to Intermediate Language code (IL) similar to Java byte code.
This means that we can distribute IL assemblies to various computer architectures because it is then compiled locally into executable code. However, there is a one
time penalty for this compilation. Even this can be avoided by forcing seed code to be fully compiled and storing multiple versions of the seed to be distributed
depending on the Worker architecture. We have found the one time compilation penalty (<10 seconds) to be minimal in terms of large simulations that may last hours.
To address the efficiency of C# code in numerical computations a scheme similar to the one Matlab uses was introduced. Manipulation of large volume of data was
moved to routines developed in C code.
This was tested in the simple scenario of matrix product of two square matrices. This function was wrapped into the “Matrix” data type. This data type
Page 232
Figure 9. Rigid particle dynamics in a looped pipe.
Figure 10. Comparison of native C, C# with primitive types, and C# with complex types for matrix multiplication.
is seen on the third level of GridLib. The usage of this function is shown below:
...
Matrix m1, m2, m3;
...
m3=Matrix.product(m1, m2);
...
An experiment was developed to measure the efficiency of different code.
Three types of code were studied, namely: natively compiled code, pure CLR code using primitive types (arrays of double precision values) and CLR code with
advanced data types (linkedlist of objects). The time cost for different size of matrix and different programming methods are plotted in Figure 10. Time measured is
only the time for product calculation.
Not surprisingly, the native code based matrix manipulation is the most efficient one. The difference between CLR code using primitive data types and native code is
approximately doubled in CLR code, but still within the same order of magnitude. The CLR code can be regarded as “moderately efficient”. The idea of wrapping
native functions for large volume data manipulation can be useful for some memoryintensive or computationintensive job. However, there is another efficiency issue.
Since the communication channels, both Remoting and Web Service, are based on XML/SOAP, which tags all data entities, the actual data volume transferred are
larger than absolutely necessary. For large XML documents, parsing takes significant time and memory resources. Unfortunately, in scientific computing, large data sets
are common. These introduce efficiency challenges in communication. Although we present here only a theoretically evaluation, communication efficiency can
undoubtedly be a bottleneck and should be explored in future studies.
7.1 Security
In a Grid network that crosses security boundaries, the likelihood of being attacked is greatly increased. Security however comes at a high cost, especially in
computational efficiency. For example, there is a considerable overhead in checking an X509 certificate.
MPI was developed for tightly controlled clusters of machines and not for cross domain applications where security may be an issue. This is not a problem within a
local area network where each machine is trusted. However, the Computational Grid is an open target over the Internet for potential attack and we should address
these security considerations. Security of the connection can be addressed because the .NET Framework supports Web Services using HTTPS. For security of the
runtime, the SandBox design greatly improves safety when running executables on worker machines.
8 CONCLUSIONS
A new Computational Grid framework has been developed that is capable of solving some parallel
Page 233
computing problems. This new framework has features that are attractive for engineers that are experts in computational algorithms and want to take advantage of high
level messaging abstractions and advanced programming techniques. A communication system has been proposed that addresses some problems that arise in Grid
Computing, not addressed by MPI. There are questions about the communication efficiency in Grids that cross domain authorities and where security and message loss
may become problematic. These issues are largely absent in a tightly coupled cluster of machines on a local highspeed network.
REFERENCES
Foster, I., Kesselman, C., Nick, J. & Tuecke, S. 2004. The Physiology of the Grid: An Open Grid Services Architecture for Distributed Systems Integration.
(http://www.globus.org/research/papers/ogsa.pdf)
Foster, I., Kesselman, C. & Tuecke, S. 2004. The Anatomy of the Grid—Enabling Scalable Virtual Organizations, 2001.
Noble D.R. & Torczynski, J. 1998. A lattice Boltzmann method for partially saturated computational cells. International Journal of Modern Physics C 9(8):1189–1201.
Qi, D. 1998. LatticeBoltzmann simulations of particles in nonzeroReynoldsnumber flows. Journal of Fluid Mechanics 385:41–62.
Cook, B.K., Noble, D.R. & Williams, J.R. 2004. A Direct Simulation Method for ParticleFluid Systems. Engineering Computations: Int J for ComputerAided
Engineering 21(2–4):151–168.
Page 234
Page 235
Ab initio emergent phenomena in PFC
B.E.Hobbs, A.Ord, K.RegenauerLieb, F.Boschetti, Y.Zhang & S.Durrlemann
CSIRO Exploration & Mining, Predictive Mineral Discovery Cooperative Research Centre, Perth, WA, Australia
5809 679 3
ABSTRACT: PFC is capable of producing selforganized or emergent behavior with the initial (or ab initio) incorporation of only the simplest of
contact relations between particles (Cundall 2001). In this paper we explore the proposal that emergent behavior within PFC arises because of long
range feedback interactions. Specifically we explore the nucleation and growth of a shear band within PFC and map out the nature of the nonlocal
feedback interactions.
1 INTRODUCTION
The term ab initio is used here in the same sense as it is used in molecular dynamic (MD) simulations. There, an assembly of atoms, represented as spheres, is modeled
with physically realistic interactions between the spheres; in the case of molecular dynamics, these interactions are quantum mechanical interactions. With this premise it
is now possible to derive most of the physical and chemical properties of the bulk solid, such as viscosity, bulk modulus, thermal expansion and Gibbs Free Energy of
formation, from a set of relatively simple, but computer intensive, calculations (see for example Colina et al. 2003). Moreover, even with these ground rules, new
unpredicted crystallographic structures and phase relations emerge spontaneously during the calculations without any a priori rules being applied. With similar scientific
rigor, we use an upscaled version of the MD approach, PFC, in which Newton’s Laws of Motion plus simple elastic, viscous and frictional contact laws take the place
of the quantum mechanical interactions. As Sethna (2003) has described so succinctly for condensed matter physics, we are trying to understand why ‘messy
combinations of zillions of’ basic building blocks ‘do such interesting simple things’.
2 EMERGENCE
Emergent phenomena are characteristic of large systems, in which energy is being dissipated, and which are being driven far from equilibrium. In such systems,
wonderful patterning and order can be generated, both in time and space, in systems that otherwise one would expect to behave in uniform, boring manners
(Kondepudi & Prigogine 1998). This can commonly be represented as a symmetry breaking phase transition (see Sethna 2003). A characteristic of such systems is
that there are first order feedback relationships between a number of processes. Another important characteristic is that the resultant patterning occurs on a number of
scales so that something of a fractal nature is developed.
As Cundall (2001) has pointed out, there is a basic difference between continuum codes and codes such as PFC2D (Itasca 2002) and PFC3D (Itasca 2003). In
continuum codes the (commonly imprecisely known, but complicated) physics has to be incorporated explicitly in the form of partial differential equations or empirical
data whereas relatively simple interactions between particles in PFC commonly suffice to produce quite complicated and emergent phenomena. Such emergent
phenomena do not arise in continuum codes unless the basic physics behind the phenomena is incorporated a priori into the calculations, or the evolution of the
dissipative structure is followed in a fully coupled way. Thus, as an example that we will explore in detail here, localization arises in continuum codes only if non
associated plasticity, corners on a yield surface or strain softening are specifically incorporated in the constitutive relation. An alternative continuum solution, where
emergent phenomena are resolved by full consideration of the thermodynamic feedback loop is numerically very expensive. By contrast, localization occurs in PFC
with the assumptions only of Newton’s Laws of Motion, simple elastic and frictional interactions between particles and an appropriate size distribution of particles.
Other similar examples of the emergent development of microstructure and patterning using a particle code for colloidal dynamics are given by Dzwinel et al. (2003).
The continuum approach to simulation is a classical Aristotelian one: the problem is first broken down
Page 236
into its constituent parts and then the relationships between the parts are defined by means of mathematical expressions of physical and chemical laws such as
constitutive relations, Darcy’s Law or the Law of Mass Action. The assumption is then that the behavior of the total system can be arrived at by combining the
constitutive parts in simple or, sometimes, quite complicated nonlinear ways. Normally this is done by writing a coupled set of partial differential equations that are then
solved by some differential equation solving engine such as FLAC (Itasca 2000).
The ultimate result of such a process is deterministic and arises solely from the imposed constitutive behavior together with assumed initial conditions and boundary
conditions. This is the classical Aristotelian—Galilean—Newtonian way of doing science. It is deterministic and the results follow once the physical and chemical laws
are prescribed along with initial and boundary conditions. As indicated above we now know that in these kinds of systems, emergent behavior will not appear unless
bifurcation phenomena are explicitly built into the mathematics that describe the system behavior. PFC on the other hand is nondeterministic and constitutive relations
capable of exhibiting bifurcation behavior do not need to be incorporated.
One important class of constitutive relations is the socalled nonlocal constitutive relations in which the mechanical behavior at a point depends on the behavior at
points far removed from that point. The gradient class (see for instance, Muhlhaus & Aifantis 1991 a, b), where the spatial gradients of the deformation rate or stress
rate are included within the constitutive relation, are a special case of nonlocal constitutive behavior. Such constitutive relations automatically include a length scale and,
in addition to having the advantage that mesh sensitivity is no longer an issue at bifiircation points, these constitutive relations also do not pose mathematical problems
such as loss of ellipticity at bifurcation points. These materials exhibit a large range of emergent behavior in continuum calculations. The physical significance of non
local constitutive relations is that since each node in a continuum mesh is influenced by the evolution of nodes more than once removed from itself, longrange
feedback interactions are incorporated into the mechanics of the system.
However, this type of behavior is an intrinsic feature of PFC in that commonly, the motion of an individual particle is strongly influenced by the motion of a particle
many times removed from itself. Thus, a form of nonlocal interaction is an intrinsic feature automatically incorporated within PFC.
3 APPLICATION
The development of fracture systems, shear zones, zoned mineral alteration systems (including ore bodies), and fluid percolation networks can all be thought of as
emergent phenomena in geomaterials. However, prediction of the initiation, development, and evolution of fracturing and shear zone formation, and of associated
properties such as dilatancy, is remarkably difficult. This is a function partly of the variability of rock properties, and partly because we cannot see into the specimen to
see how it behaves during physical deformation experiments. This is where computer experiments can provide such valuable information. In this case we look at the
opportunities provided by a particle code for exploring the emergence of microstructures during deformation.
4 EXPERIMENTS
We initiated a suite of experiments for validating a particle flow code, PFC, (Cundall 1988, Cundall, 2001) against physical deformation experiments, in this case the
triaxial experiments of Edmond and Paterson (1972; see also Ord, 1991), for various materials, including Gosford sandstone. The initial experiments were sufficiently
encouraging for us to attempt validation of the code in a plane strain biaxial experiment against the experiments of Ord et al. (1991). The same phenomenon,
localization of the deformation into a shear zone, emerged from a numerical specimen undergoing uniform loading (Figs 1 & 2).
This behavior is not promoted in any way through special conditions within the experiments; no notches were cut, there were no seeds of particularly weak zones. In
Figure 3 we see the result for shear band formation of loading the material to about 10% shortening, far beyond localization at about 0.6%.
Figure 1. (a) Displacement vectors displayed for plane in centre of the numerical models, highlighting the location of a shear zone. (b)
Square root of the horizontal displacement contour (dark grey—minimum; light grey—maximum).
Page 237
This is an excellent example of the emergence of patterned shear zones. Conjugate sets of shear bands are well developed.
5 HOW TO DETECT EMERGENCE
Coming to applications, we are first faced with a very practical task, how to detect emergence. This depends on a much deeper problem, that is, how to define
emergence. Here we follow the approach of Crutchfield (1994) who distinguishes two kinds of emergence: pattern formation and intrinsic emergence:
– Pattern formation refers to an external observer who is able to recognize how certain unexpected features (patterns) ‘emerge’ or ‘selforganize’ during a process
(e.g., spiral waves and Turing patterns in oscillating chemical reactions and convective rolls in fluid flow). The patterns do not have specific meaning within the
system, they obtain a special meaning to the observer once (and if) he/she is able to detect them. When this happens, the patterns become part of the toolbox the
observer can employ to describe and study the process.
– Intrinsic emergence refers to the emergent features which are important within the system, since they confer additional functionality to the system itself, like supporting
global coordinationcomputationbehavior (e.g., the emergence of coordinated behavior in flock of birds or in stock market pricing allows efficient global information
processing which benefits the individual components of the system).
The two types of emergence described above are not in antithesis and can be found at various degrees in different natural systems. It is tempting to suspect that, at least
in certain particle modeling exercises, the two views of emergence outline above may coincide. We will adopt this as a working conjecture and test it in our
experiments. Our justification for the conjecture goes like this: in a compression simulation particles need to arrange themselves spatially in response to the load. The
initial loose particle packing allows a local configuration to appear by more or less uncoordinated movements of individual particles. Once a tight packing is achieved,
the particles reach a configuration which is optimal locally, but not globally. At this stage, jumps to a more efficient spatial configuration require the coordinated
movements of large groups of particles. This coordination highlights particle aggregations; these aggregations not only have a meaning to an external observer who sees
such patterns arising from the initial disorganized medium, but also to the medium itself, as a means for global information processing.
Figure 3. Development of shear bands in numerical rock sample of dimensions 20×20 cm. Shortening is in the vertical direction and
confining pressure is 10 MPa. Bulk shortening of the model is about 10%.
Figure 2. Comparison of numerical results with physical experimental results. (a) Volumetric strain versus axial strain. (b) Axial
deviatoric stress versus axial strain.
Page 238
Computational Mechanics tools like the epsilonmachine (Shalizi et al. 2002), and its initial extension to 2D patterns (Feldman & Crutchfield 2003) suggest
themselves for this kind of analysis. In Ord et al. (2004) we reported on some first attempts of applications to 2D rock patterns. This is an active area of research,
which currently stumbles against a huge computational cost. These initial results demonstrate an alternative approach which is showing promising results at a fraction of
that computational cost. The idea is based on the work of Desolneaux et al. (2003). They describe a technique to detect meaningful features in images and meaningful
events in histograms. A feature or an event is described as meaningful when the ratio between its occurrence and the expectation of its occurrence is very high (in other
words, when we expect something to happen rarely but instead it happens frequently). Compared to similar approaches available in the literature, this simple algorithm
has four particularly nice features:
– It depends only weakly on a numerical parameter defining the threshold for meaningfulness, which allows the algorithm to work in an ‘unsupervised’ manner and with
minimal a priori information;
– It defines not only which patterns are meaningful, but also how many there are (a value often required a priori in standard image processing tools);
– It is supported by a convincing theory describing the asymptotic behavior of the algorithm;
– Meaningful classes are proved to be non overlapping, a feature which offers itself immediately to image segmentation (partitioning).
How does this simple approach lend itself to detection of emergence? Rather than applying the algorithm to features (like linear segments in an image), or events (like in
histogram analysis), we apply the algorithm to behaviors: we look for significant partitioning in the motion of the particles.
6 EXAMPLE
An example helps to describe the approach. In Figure 4 we see a frame taken from a PFC simulation.
Each particle is shaded according to the main direction of motion within an interval of approximately 8000 frames. In Figure 5a we can see the histogram of the
number of particles moving in different directions.
Based on the null hypothesis that the histogram should be evenly distributed, Desolneaux et al.’s algorithm detects 4 significant modes in the histogram. These are
shown in Figure 5b.
The algorithm tells us that the motion within these four directions happens so unusually often to deserve our attention. Following the conjecture mentioned above, we
suspect the medium (particles) has
Figure 4. Snapshot of the particle movements in the first 8000 steps of the simulation. Shades of grey map the direction of the vector
describing the particle total movement from step 1 to step 8000.
Figure 5. (a) Histogram of the number of particle having movement of a certain direction. Directions between −π and π are binned into
100 bins. (b) The four significant intervals as detected by Desolneux et al.’s algorithm (notice that because of circular
symmetry the last and first interval coincide).
Figure 6. Segmentation of the image according to the results in Figure 5b. Basic particle aggregations are shown.
Page 239
organized itself by producing 4 (emergent) “superparticles” characterized by coordinated motion.
More work is required before being able to give more sound foundation to the above statements but the results seem promising. In particular the following avenues
of research are worth pursuing:
– Generalize the algorithm so that more complicated motion than simple linear translation can be detected (rotations, oscillations, lack of motion, etc.)
– Apply the algorithm to the 3D cube resulting from the frames moving in time.
– Apply the algorithm to the spatial arrangement of the particle.
7 CONCLUSIONS
The simplicity of particle simulations lies in that they only use elastic couplings to explore longrange interactions without going into a full implementation of the potential
functions. The term “ab initio” usually is understood for more comprehensive approaches at atomistic level, which are fully coupled. Regardless of the complexity of
the model “ab initio” is here understood as describing a whole new class of approaches described by investigating nondeterministic dissipative structures that lead
to the phenomenon of emergence.
Such simple models generate very complex patterns. Similar to recent advances in biology, e.g. in observing DNA microarrays (Tilstone 2003), we need new
statistical methods to describe the system complexity. We have explored here a pattern analysis that allows us to identify domains of basic particle motions as one form
of emergence, possibly premonitoring the appearance of a shear band.
While we have presented a first approach only, the future of “ab initio” modeling lies in linking nature to the numerical analyses, to perform a comparative
implementation of statistical methods on both natural and numerical samples. In more general terms the general strategy has the potential to provide a new, common,
computationallyenhanced basis for all natural sciences.
REFERENCES
Colina, C.M., OliveraFuentes, C.G., Siperstein, F.R., Lisal, M. & Gubbins, K.E. 2003. Thermal properties of supercritical carbon dioxide by Monte Carlo simulations.
Molecular Simulation 29:405–412.
Crutchfield, J. 1994. The calculi of emergence: Computation, dynamics, and induction. Physica D 75:11–54.
Cundall, P.A. 1988. Computer simulations of dense sphere assemblies. In M.Satake and J.T.Jenkins (eds) Micromechanics of Granular Materials. 113–123.
Amsterdam: Elsevier Science Publishers B.V.
Cundall, P.A. 2001. A discontinuous future for modelling in geomechanics? Proc. Inst. Civil Eng. Geotechnical Engineering 149:41–47.
Desolneux, A., Moisan, L. & Morel, J.M. 2003. Maximal meaningful events and applications to image analysis. The Annals of Statistics 31(6):1822–1851.
Dzwinel, W., Yuen, D.A. & Boryczko, K. 2003. Bridging diverse physical scales with the discreteparticle paradigm in modelling colloidal dynamics with mesoscopic
features. Winter 2003 Research Bulletin of the Supercomputing Institute, University of Minnesota, 19(2):13.
Edmond, J.M. & Paterson, M.S. 1972. Volume changes during the deformation of rocks at high pressures. Int. J. Rock Mech. and Mining Sci. 9:161–182.
Feldman, D.P. & Crutchfield, J.P. 2003. Structural information in twodimensional patterns: Entropy convergence and excess entropy. Physical Review E67, 051104.
Itasca Consulting Group, Inc. 2000. FLAC—Fast Lagrangian Analysis of Continua, Ver. 4.0 User’s Guide. Minneapolis: Itasca.
Itasca Consulting Group, Inc. 2002. PFC2D—Particle Flow Code in 2 Dimensions, Ver. 3.0 User’s Guide. Minneapolis: Itasca.
Itasca Consulting Group, Inc. 2003. PFC3D—Particle Flow Code in 3 Dimensions, Ver. 3.0 User’s Guide. Minneapolis: Itasca.
Kondepudi, D. & Prigogine, I. 1998. Modern Thermodynamics: From Heat Engines to Dissipative Structures. Chichester: John Wiley and Sons.
Mühlhaus, H.B. & Aifantis, E.C. 1991a. A variational principle for gradient plasticity. Int. J. Solids and Structures 28:845–857.
Mühlhaus, H.B. & Aifantis, E.C. 1991b. The influence of microstructureinduced gradients on the localization of deformation in viscoplastic materials. Acta Mechanica
89: 217–231.
Ord, A. 1991. Deformation of rock: a pressuresensitive, dilatant material. Pure and Applied Geophysics 137: 337–366.
Ord, A., Boschetti, F. & Hobbs, B. 2004. 3D Imaging of jointed rock masses. In D.Kolymbas (ed), Fractals in Geotechnical Engineering. Berlin: Logos.
Ord, A., Vardoulakis, I. & Kajewski, R. 1991. Shear band formation in Gosford sandstone. Int. Jnl. Rock Mech. Min. Sci. and Geomechanics Abst., 28:397–409.
Sethna, J.P. 1992. Order parameters, broken symmetry, and topology. In L.Nadel & D.Stein (eds) Lectures in the Sciences of Complexity: 243–288. New York:
AddisonWesley.
Shalizi, C.R., Shalizi, K.L. & Crutchfield, J. 2002. An algorithm for pattern discovery in time series. Technical Report, Santa Fe Institute, 2002–10–60, cs.LG/02100025.
Tilstone, C. 2003. Vital statistics. Nature 424(6949): 610–612.
Page 240
Page 241
Discrete computational methods as the tools for studying behavior of materials,
structures and complex media on different scales
S.G.Psakhie
Institute of Strength Physics and Materials Science, Russian Academy of Sciences, Tomsk, Russia
5809 679 3
ABSTRACT: The discrete approach is developed on the basis of computational methods at the micro and mesolevels to simulate behavior of
materials under various types of loading. At the microscale level the molecular dynamics (MD) method is used. A promising method for the meso
level is the movable cellular automata (MCA) method. This method provides new possibilities to simulate complex behavior of materials like mass
mixing, penetration, etc. The equations of motion of the MCA method are transformed into equations of the molecular dynamics method on the nano
scale level. On this basis processes at various levels from atomic one to geomedium were investigated. Modeling of a contact problem at the atomic
and mesoscale levels was carried out as an example. The results obtained show the velocity gradient in nearsurface regions leads to the formation of
loosened and fragmented zones. This effect is accompanied by the appearance of tensile stresses in the loaded region and intensive massmixing
processes. It allows us to explain many experimental findings of mechanochemistry, such as phase formation at friction surfaces and alloying under
pressure plus shear loading. Computeraided investigation of the behavior of feebly bonded soil media under dynamic loading was carried out also. It
was shown that an analogue of “lifting force” in such a medium appears in the presence of velocity gradient. This effect is key to understanding the
liquefaction phenomenon during earthquakes. It should be noted that the MCA and MD methods can be successfully used to study different problems
in various areas such as materials science, rock and soil mechanics, civil engineering, earth science, and biomechanics.
1 INTRODUCTION
Recent development of technologies for design of new materials and coatings with gradient and nonequilibrium structures is of the utmost practical importance. The
technologies, based on highenergy contact interaction, are used to produce various types of gradient materials and coatings. In Bridgeman experiments, as far back as
30 years ago, the quasiviscous streaming of solids under “pressure+shear” loading conditions was discovered (Panin et al. 1985). The extraordinary behavior of solids
was found. The impressive results were obtained at the Institute of Chemical Physics (Larionov & Enikolopjan 1993). In these papers an anomalous masstransfer rate
was described. The masstransfer rate is 15 orders of magnitude greater than that at classical diffusion, the dielectric and semiconductor conductance was abruptly
changed, chemical reaction rate was profoundly increased and, at times, chemical reactions even changed their directions. These phenomena cannot be explained in the
frameworks of approaches based on the classical description of diffusion process. It allows us to conclude that a new branch of chemistry, mechanochemistry, has
emerged (Boldyrev et al. 1979, Urakaev & Boldyrev 2000, Grigorjeva et al. 2000). The nonequilibrium state of a material in contact areas can initiate peculiar
chemical reactions, which allows significant economic, ecological, technological, and other advantages in comparison with conventional approaches. So, using
mechanochemical solidphase synthesis it is possible not only to reduce the time of fabrication but also to produce materials with properties impossible for conventional
methods. This is a result of nonequilibrium processes of alloying under “pressure+shear” loading (Larionov & Enikolopjan 1993, Grigorjeva et al. 2000).
The present paper is devoted to the theoretical investigation of this problem using the present numerical techniques. In studies of contact highenergy interaction it is
necessary to take into account intensive processes of discontinuity formation, material fragmentation and mass mixing in contact areas. The numerical methods based on
the particle approach (discrete methods) have been developed during recent years. In the framework of these methods processes of damage formation, fragmentation
and mass mixing can be taken into account correctly. Moreover, in
Page 242
the framework of the discrete approach it is possible to simulate material response at various scale levels from the micro/atomic level (molecular dynamic method
(Heerman 1986)) to the macrolevel (rigid body dynamics and quasimolecular dynamics (Heerman 1986, Greenspan 1989)). It should be noted that the discrete
numerical methods are oriented mainly for simulation of granular media at the macro and mesoscale levels and cannot be used for description of behavior of
continuum media and bounded systems. Therefore, in the present paper the movable cellular automaton (MCA) method is used for simulation at the meso and macro
scale level (Psakhie et al. 1995, 1997, 2001). Although the MCAmethod is based on the discrete approach, the equations of motion, which are used in the MCA
method, differ from the classical ones in the particle approach. The MCAmethod allows simulating, in explicit form at the mesolevel, both the formation of
discontinuities of various types (from single damage generation to main crack propagation) and the effects of mass mixing (especially during highenergy contact
loading). Another discrete method—the molecular dynamic method—was used to investigate special features of material behavior in a contact area at the atomic level.
Note that the molecular dynamic method can follow from the movable cellular automat on method in the limit case when automat on size tends to a point and the
function of automaton response is changed by the interatomic interaction potential.
2 FORMULATION OF MOVABLE CELLULAR AUTOMATA METHOD
Within the framework of the MCA method, a system under simulation is considered as an ensemble of interacting automata (elements) of finite size. The concept of the
MCA method is based on the introduction of a new type of state (viz., the state of a pair of automata) into that of classical cellular automata. This has made it possible
to arrive at a space variable and use it as a parameter of switching. The overlapping of a pair of automata was chosen as such a parameter:
(1)

where where dj characterizes the automaton size. In the simplest case there are two states of the pair:
(2a)
(2b)

In general, the state switching criterion is a function of hij and angle of relative rotation of automata θij:
(3)
The linked state indicates chemical bonds between elements and the unlinked state indicates that there is no chemical bond between automata. It should be emphasized
that this parameter is of space dimension. This leads to a new property of automata, i.e. the capability of changing their positions in space and, consequently, changing
the spatial arrangement of the whole system. Thus, the change in the state of automata pairs is governed by relative displacements of automata forming a pair. A
medium constituted by such pairs can be considered a bistable medium.
2.1 Equation of motion of movable cellular automata
The system consisting of movable cellular automata is described by the translational equations of motion based on the WinerRosenbluth model for bistable cellular
automata:
(4)
Here pi(j)k(l) is the force of interelement interaction. It is defined by the function of the interautomata response. The coefficient C(ij,ik(jl)) is associated with the
transfer of the parameter h from the pair ik (or jl) to the pair ij. The term is determined by the Poisson’s ratio and, consequently, is related to the mutual
arrangement of the pairs of elements ij and ik.
The equation of motion for rotation can be written in the following form:
(5)
Here θij is the angle of the relative rotation of elements; qij is the distance from the center of the automaton i to the point of contact with the automaton j (moment arm);
τij is the tangential interaction of pairs; S(ij,ik(jl)) are coefficients associated with the transfer of the parameter θ from the pair ik (or jl) to the pair ij.
It can be seen that if C(ij,ik(jl))=C(ij,ik(jl))=1, Equations (4) & (5) are totally equivalent to the
Page 243
NewtonEuler equations of motion for manyparticle interaction.
The time lag is defined by the automata size, longitudinal and transverse sound velocities, and mutual arrangement of the pairs ij and ik (or jl). It is obvious that the
effects associated with generation and propagation of shock waves require particular investigation.
2.2 Response function in the MCA method
There are four basic types of response function in the MCA method (Fig. 1). In the simplest case interautomaton interaction is assumed to be elastic and linear. To this
approximation the response function can be considered as a linear function of the overlapping parameter as shown in Figure 1a. It should be noted that in this case,
loading and unloading follows the same curve. To take account of damage generation at a scale level lower than the automaton size, the linear response function should
be modified (Fig. 1b). It should be taken into account that Young’s modulus decreases due to material degradation starting with a certain loading value (higher than the
degradation point). The response functions mentioned above can be used for simulating fracture of brittle materials such as ceramics, concrete, etc.
The response functions for irreversible behavior of materials, like plastic deformation, should be selected particularly for each material with regard to loading
conditions. Examples of response functions describing irreversible behavior of materials are shown in Figures 1c & 1d. The case presented in Figure 1c corresponds to
plastic deformation and the response
Figure 1. Examples of response functions in the MCA method.
function shown in Figure 1d is a combination of plastic flow and degradation processes (damage generation) in the material. It should be emphasized that the examples
shown in Fig. 1c and Fig. 1d are illustrative in character. The determination of the response function for describing the irreversible behavior (particularly plastic
deformation) is a challenging problem and requires intimate knowledge of the mechanisms of plastic flow initiation and evolution.
3 CONTACT AREA MODELING AT THE MESOSCALE LEVEL
The initial structure of a simulated fragment is presented in Figure 2. The simulated region consists of two blocks 1 and 2, which belong to different interacting elements.
Automata of the upper layer of block 1 were subjected to an external force oriented along OYaxis and corresponding to a compressive pressure 4 MPa. Horizontal
loading was applied after the establishment of equilibrium. It was a constant velocity (5, 10, 20, and 40 m/s) oriented along OXaxis in one series of calculations and a
constant force in another one. The value of constant force was three times greater than the value of vertical loading. In all calculations, location of automata of the lower
layer of block 2 was fixed, in order to simulate elastic response of the substrate. Such loading conditions correspond to highconstrained conditions, which can be
easily realized during highenergy contact interaction. Along OXaxis the periodic boundary conditions are used to avoid the influence of boundary effects.
Figure 2. Structure of the simulated area and loading scheme at initial time moment.
Page 244
Table 1. Mechanical properties of the simulated material.
Young’s modulus (GPa) 200
Poisson’s ratio 0.3
Yield point (MPa) 100
Strength (MPa) 160
Ultimate strain (%) 2
2100
Density (kg/m3)
The longitudinal size of the calculated area is 0.5 μm and the automaton size is 10 nm. A number of automata in the simulated fragment is more than 5000. The physical
parameters of automata of both blocks are presented in the Table 1. The contact surfaces have initial roughness, which was preset in explicit form.
Loading was realized in four stages: (i) linear growth of vertical loading component up to the fixed value corresponding to pressure 4 MPa; (ii) the establishment of
equilibrium in the system under fixed vertical loading; (iii) linear growth of horizontal loading component (force/velocity) under fixed vertical loading; (iv) application of
constant vertical and horizontal components of loading. The results obtained have shown that a loosely fragmented layer is formed in the contact area. Material in this
layer is simultaneously subjected to pressure and shear. Intensive processes of fragmentation at the mesoII scale level and material mixing of blocks 1 and 2 occur.
3.1 Loading with constant velocity
The structure of the calculated area is presented in Figure 3a to demonstrate formation of a loosening layer. The velocity of the automata in the upper layer of block 1
was 20 m/s. It is seen that a loosely fragmented layer is formed under applied compressive stress.
The time dependence of Ycoordinate of the upper layer of block 1 at various loading velocities is presented in Figure 3b. The presented results conclude that due to
formation of a loosely fragmented layer, the value of Ycoordinate of the upper block grows oscillating in the vicinity of a certain average value. With loading velocity
growth the average value of the Ycoordinate increases. It should be noted that fragment formation and evolution occur in two stages. In the first stage, discrete
misorientation of adjacent fragments takes place and in the second stage their separate movement with formation of the loosely fragmented layer occurs. The calculation
results have shown that a certain threshold value of loading velocity should be reached to initiate loose layer formation. So, for loading velocity 5 m/s Ycoordinate of
the upper layer of block 1 gradually takes its initial value after the oscillation stage (Fig. 3b).
Figure 3. Loading with constant tangential velocity: structure of the simulated fragment at loading velocity 20 m/s (a); time dependence
of Ycoordinate of the upper layer of block 1at various loading velocity (b): 5 (1); 10 (2); 20 (3); 40 m/s (4); d is the automata
diameter.
3.2 Loading with constant force
The formation of the loosely fragmented layer induces a decrease in effective resistance to relative motion of the contact surfaces. In the case of constant tangential
force, sufficient for initiation of the fragmentation process, relative velocity of motion grows and the loose layer becomes thicker. The structure of the calculated area is
presented in Figure 4a.
It is clear from Figure 4b that the thickness of the fragmented layer increases, involving new subsurface layers in the mixing process. The maximum changes of Y
coordinate observed at the initial stage appear to be governed by intensive changes of the contact surface profiles. Then the normal pressure starts to dominate, which
leads to a decrease in Ycoordinate. Further, Ycoordinate of the upper layer of block 1
Page 245
Figure 4. Loading with constant tangential force: structure of the simulated fragment (a); time dependence of Ycoordinate of the upper
layer of block 1 (b); d is the automata diameter.
constantly grows. This effect is equivalent to the action of a “lifting force”, which induces formation of a loose layer and intensive mixing within this layer. Maximum
thickness of this loose layer is determined by parity between “lifting force” and normal pressure. Automaton velocity fields at different time periods are presented in
Figure 5. It is seen that a velocity gradient occurs in the loose layer. In doing so, it is clear that with velocity gradient growth, the thickness of the loose layer increases
according to the growth of the “lifting force” value.
A similar effect was described by Goldin (2001), where behavior of feebly bonded media (soil in the vicinity of a massive vibrator) was analyzed. It was shown that
“lifting force” initiation is governed by velocity gradient. In so doing, a “lifting force” value
Figure 5. Velocity fields of the simulated fragment at different time instants.
(taking into account the special features of medium motion) can be estimated using the Euler equation of motion in the GromekaLamb form. As it was shown by
Goldin (2001).
(6)
where ν=velocity; ρ=density; p=pressure; α= certain structuresensitive coefficient which effectively represents medium heterogeneity.
Page 246
Assuming that the velocity gradient along the vertical direction is linear and velocities of automata of the lower layer are equal to zero, Equation (6) can be written as:
(7)
where ν 1=automata velocity in the upper layer. As shown by Goldin (2001), Equation (7) provides good agreement with the results of numerical experiments. As noted
above, intensive processes of mass mixing take place in the fragmented layer.
4 LOCAL LOOSENING PHENOMENON ON THE ATOMIC SCALE LEVEL
It is of interest to consider, at the atomic level, the special features of behavior of an individual mesofragment during loosening of the surface layer when numerous
regions with velocity gradient are formed in contact areas. The molecular dynamic method was used for simulation. The simulated crystallite was parallelepiped in
shape. The manybody interatomic potential of interaction derived in the framework of the embedded atom approximation (Daw 1984) was used. The periodic
boundary conditions were preset in OX and OYdirections. Free surfaces were simulated along the OZaxis. In the central crystallite part (loading region), velocity
gradient along OXaxis was given. Length of the loading region was 1/3 of the total length. Y and Zcomponents of atom velocities in the loading region were set to
zero. Calculations were carried out for Vmax=50, 75, 100, 125, and 150 m/s (Vmax=the maximum value of atomic velocity in the center of the loading region). Atomic
velocities in the loading area during the first 5000 integration steps were linearly increased. The integration step was equal
Figure 6. Initial structure of the simulated crystallite and loading scheme.
to 100 atomic units. The distribution of applied loading is shown in Figure 6.
The calculation results obtained show that, due to the velocity gradient in the loading region, tensile stresses appear. The stresses are oriented from the center of the
loading region to its boundaries along the OZdirection. Appearance of tensile stresses at the atomic level induce an increase in effective atomic volume of the loading
region. Changes of the specimen volume with time (a number of integration steps) for various loading values are presented in Figure 7a. The time dependence of volume
changes when stationary state is reached is shown in Figure 7b. The velocity value in the figure peaks. It is clear that as the maximum velocity value of atoms in the
loading region grow, the mesofragment tends to increase the volume. It should be noted that the loosening process is mainly
Figure 7. Volume changes of simulated crystallite depending on steps of integration under various loading value (a); volume changes of
loading region depending on maximum value of loading velocities (b).
Page 247
(~95%) realized owing to the increase in volume of the loading region.
5 CONCLUSIONS AND DISCUSSION
In Figures 3 & 4 special attention must be given to the pronounced selforganization of mesofragments along the directions of maximum tangential stresses τmax within
the loose layer. This effect is manifested both at the mesoII scale level and at the atomic level. It indicates that the field of maximum tangential stresses governs
structure rearrangement in a contact area of blocks 1 and 2. The results obtained allow us to conclude that mass mixing effects in contact areas occur at various scale
levels. It should be also noted that the loss of crystal lattice stability in contact areas leads to the generation of an abnormally high defect density and to a decrease of
activation energy threshold of various chemical processes. For a fundamental understanding of mass transfer processes it is necessary to carry out additional theoretical
and experimental investigations at various scale levels.
Furthermore, the computeraided investigation of the behavior of feebly bonded soil media under dynamic loading was also carried out using the movable cellular
automaton method. Soil behavior under a 40t vibrator, which is used for generation of acoustic waves in geological media, was simulated. It was shown that an
analogue of “lifting force” in a feebly bonded granular medium appeared in the presence of a velocity gradient. The “lifting force” leads to local loosening of the medium.
The revealed effect is in agreement with experimental data on investigations of a massive soil by highfrequency seismic waves. The tentative dependence of the “lifting
force” value on such factors as medium density and velocity gradient was proposed. Knowledge of such processes is very important for human safety, therefore, the
significance of revealed dynamic phenomena, which accompany the process of local material loosening, extends the initial aim of the investigation.
ACKNOWLEDGEMENTS
This study was supported in part by a grant of the Ministry of Education (RF PD021.5–425), a CRDF grant (no TO016–02) and the “Russian Science Support
Foundation”.
REFERENCES
Boldyrev, V.V., Boulens, M. & Delmon, B. 1979. The Control of the Reactivity of Solids. Amsterdam: Elsevier Sci. Publ.
Daw, M.S. & Baskes, M.I. 1984. Embedded atom method: Derivation and application to impurities, surfaces, and other defects in metals. Phys. Rev. B29(12):6443–6453.
Goldin, S.V., Psakhie, S.G., Dmitriev, A.I. & Yushin, V.I. 2001. Structure rearrangement and “lifting” force phenomenon in granular soil under dynamic loading. Physical
Mesomechanics. 4(3):91–97.
Greenspan, D. 1989. Supercomputer simulation of cracks and fracture by quasimolecular dynamics J. Phys. Chem. Sol. 50(12):1245–1249.
Grigorjeva, T.F., Korchagin, M.A., Barinova, A.P. & Ljakhov, N.Z. 2000. Selfpropagating high temperature synthesis and mechanical alloying during formation of
monophase fine intermetallides. Khimia ν interesakh ustoichivogo razvitia. 8:685–691.
Heerman, D.J. 1986. Computer simulations methods in theoretical physics. SpringerVerlag.
Larionov, L.V. & Enikolopjan, N.S. 1993. Nitrocompound decomposition under isothermic compression. Doklady AN SSSR, 328(2):209–211.
Panin, V.E., Likhachev, V.A. & Grinjaev, Yu.V. 1985. Structural levels of deformation of solids. Novosibirsk: Nauka.
Psakhie, S.G., Horie, Y., Korostelev, S.Yu., Smolin, A.Yu., Dmitriev, A.I., Shilko, E.V. & Alekseev S.V. 1995. Method of movable cellular automata as a tool for
simulation within framework of physical mesomechanics. Izv. Vuzov. Fizika. 38(11):58–69.
Psakhie, S.G., Smolin, A.Yu., Shilko, E.V., Korostelev, S.Yu., Dmitriev, A.I. & Alekseev, S.V. 1997. About the features of transient to steady state deformation of solids.
Journal of Materials Science & Technology. 13(1):69–72.
Psakhie, S.G., Horie, Y., Ostermeyer, G.P., Korostelev, S.Yu., Smolin, A.Yu., Shilko, E.V., Dmitriev, A.I., Blatnic, S., Spegel, M. & Zavsek, S. 2001. Movable cellular
automaton method for simulating materials with mesostructure. Theoretical and Applied Fracture Mechanics. 37(1): 311–334.
Urakaev, F.Kh. & Boldyrev, V.V. 2000. Mechanism and kinetics of mechanochemical processes in comminuting devices. I. Theory. Powder Technology. 107(1–2):93–
107.
Page 248
Page 249
Application of an ellipsoid approximation to discrete/distinct element modeling
S.Johnson & J.R.Williams
B.K.Cook
Sandia National Laboratories, Albuquerque, New Mexico, USA
5809 679 3
ABSTRACT: The practical limit for the number of particles that can be simulated by the Discrete Element Method depends critically on the particle
shape, the data structures and the algorithms used for contact and force generation. Simple shapes, such as spheres, though optimal in the above
aspects, are not always able to reproduce phenomena of interest, and nonspherical shape needs to be included. The use of particles of nonspherical
shape requires both good data structures and good algorithms, especially for contact. The focus of this paper is on representing a particle’s geometry
as a general ellipsoid in a computationally efficient manner. A new contact resolution algorithm based on the Method of Equivalent Spheres is
presented, which takes advantage of the convexity and smoothly varying curvature of the ellipsoidal geometry in combination with an efficient iterative,
vectorbased search algorithm. Special attention is given to the software implementation of the algorithm, and a discussion of the computational
efficiency of the algorithm is also provided.
1 INTRODUCTION
The computational load of discrete particle simulations places a limit on the number of particles that can be handled on today’s computers. The processing and memory
resources required are determined by the number of particles, the complexity of the particles and the algorithms and data structures used to represent them. For this
reason there is a great benefit if simple shapes, such as spheres, can be used. However, nonspherical particle shape has found to be important in representing many
physical processes, such as the mixing of powders and the strength of granular materials. This paper seeks to develop algorithms for ellipsoids that can be competitive
with spheres.
The contact detection part of the pipeline is particular computationally intensive. In this paper, contact detection is split into two phases, neighbor sorting and contact
resolution. Since neighbor sorting algorithms typically use a bounding sphere or bounding box representation of objects the same algorithms can be adapted for use with
arbitrary shapes. For details of neighbor sorting see, Williams & O’Connor (1999), O’Connor (1996), Munjiza, Bicanic & Owen (1997); Wensel & Bicanic (1993);
Munjiza & Andrews (1998); and Perkins & Williams (2001), specifically as it relates to particle geometry representation in DEM models. Therefore, this paper focuses
on the remaining problem of contact resolution.
Geometry can be a firstorder effect in determining the behavior of many systems as shown in DEM simulations for angle of repose experiments by Johnson. Physical
experiments, such as Favier (2001), also have shown that angularity and sphericity of grains are firstorder effects in the angle of repose.
The sphere has several attributes that, while attractive for computational implementation, can hinder its ability to capture key behaviors of real particle systems that
contain nonspherical particles. Spheres tend to exhibit excessively rolling when subjected to small perturbations, because they cannot exert a resistive couple (i.e.
friction is the only means by which a resistive couple can be applied), they are unable to represent the particle interlocking observed in many systems. Some numerical
implementations address this by introducing couples artificially by either perturbing the normal force direction so that it does not pass through the centroid or by adding
an adjustable “rolling friction” factor (e.g. VuQuoc 1999). Spheres also tend to organize into dense, ordered packings as the coefficient of friction decreases (Grest
2001), which is not representative of many naturally occurring materials. Despite these drawbacks, spherebased DEM implementations remain the most common for
Page 250
3D modeling, since they are simple and compact (only 4 floating point numbers are required to describe the geometry, and contact resolution between spheres is trivial)
and hence are fast computationally. Indeed, spheres are the basis for several popular codes, including TRUBAL (Cundall 1987) PFC3D (Cundall 1996) and DMC
(Taylor & Preece 1992) as well as the application developed in Cleary (1998).
The ellipsoid is the next obvious choice after a sphere as a geometric primitive to represent a particle’s topology. It is part of the same family of quadrics as the
sphere, but it offers additional favorable properties, including the capability of providing geometric interlocking, resistance to pathological rolling, and a more flexible
and accurate convex hull for many particles of interest. The geometry is also fairly general and the C∞ smoothness of its surface allows for ease of implementation in
fluid coupling codes. However, because the additional computational requirements required for contact resolution beyond those for spheres is typically large, the
ellipsoid is rarely used in largescale DEM models.
Several schemes have been developed to represent ellipsoids computationally. However, there are restrictions in each of these that limit the contact resolution
abilities. Favier (2001), Mustoe (2001), and VuQuoc (2000), investigate the use of sphere aggregates to approximate ellipsoids. Favier (2001) investigates the use of
multisphere aggregates to approximate ellipsoids, and notes the undesirable effects of the resultant C0 continuous surface on the behavior of the system being
modeled. For Wang (1999) and Johnson (2003), the “ellipsoids” are simply approximates using the common draftsmen’s 3arc ellipse approximation rotated about the
principal axis, which yields a C1 continuous surface. Ouadfel (1999), Lin (1997), and Ting (1993) all approach the problem of actual ellipsoids and provide either an
analytical solution or an iterative solution to the resolution of the contact. These algorithms, however, have been too computationally intensive to use in largescale
systems and have only included the identification of contact between two ellipsoidal surfaces (e.g. no method is shown for contact with triangular facets). For many
applications, any of these restrictions is unacceptable.
In this paper we seek to take advantage of the simplicity of spheres while overcoming some of the problems inherent in the approaches just discussed. In particular
this paper provides a new contact resolution algorithm for the general ellipsoid that improves on the Method of Equivalent Spheres developed in Johnson (2001) for
use in the rotated 3arc ellipsoidal approximation. This method takes advantage of the convexity and smoothly varying curvature of the ellipsoidal geometry in
combination with an efficient iterative, vectorbased search algorithm to yield an effective algorithm that largely avoids the evaluation of computationally expensive
transcendental functions. Special attention is given to the software implementation and computational efficiency.
2 METHOD OF EQUIVALENT SPHERES
For two smooth convex bodies in contact, there is a single point where the normals of the two bodies are equal but opposite or a surface of contact where the regions
of deformation create equal magnitude but opposite direction tractions. The threedimensional case of contact has often frustrated those attempting to intersect such
general bodies.
In the most basic terms, the principle of the Method of Equivalent Spheres is to decouple the dimensions, and thus reduce the search space by a dimension, by
slicing the body into twodimensional pieces that can then be used to select better twodimensional slices until the correct pair of slices is identified. The condition for
convergence in this case is that the angle between the identified normal in each of the bodies at the identified contact point is arbitrarily close to π. By parameterizing the
convex body in terms of a locally defined cylindrical coordinate system, the search space can be constrained to twodimensional slices passing through the principal axis
(the local zaxis) identified uniquely by the angle component. The robustness and effectiveness of the method comes from the second search, which involves an iterative
refining of the contact point estimate, which will be shown in the following sections.
2.1 Development of vectorbased search
The development of the Method of Equivalent spheres began with the case of axially symmetric systems of rotated circular arcs. In this case, it was quickly visible that
given the correctly positioned sphere, one could, for the purposes of determining the contact point and normal, replace the contacting bodies with spheres that were
correctly placed.
Determining this correct placement is the nontrivial contribution of the algorithm, especially without incurring unreasonable computational costs. To achieve this,
several aspects of the geometry were leveraged, including the property that the center of curvature for the surface always lies on a circular path orthogonal to the axis of
rotation. This constraint specifies two onedimensional equations for the identification of the unique arc of contact for each rotated arc body. It is notable that the
convexity of the body guarantees convergence of the solution. Given the correctly identified arcs of contact, determining the contact point becomes trivial and is exactly
equivalent to the condition for contact of two spheres. Because
Page 251
Figure 1. Primary vectorbased search: (a) cast vector between the centroids of the bodies and project onto plane orthogonal to the
principal axis of each body then multiply by −1, (b) cast a vector from the end of the second projected vector to the centroid of
the first body and project this vector onto the plane orthogonal to the principal axis of the first body (c) repeat step b for the
other body.
of this, the method has been named the Method of Equivalent Spheres.
To further elucidate this concept, the algorithm is illustrated (Fig. 1) and explained in summary below and for further detail see Johnson (2003).
3 EXTENSION TO GENERAL ELLIPSOIDS
Extending this concept from the application of a body consisting of rotated arcs requires some further development. It quickly becomes clear that some of the
assumptions made for the case of rotated arcs are not applicable for the case of general ellipsoids. For instance, the center of curvature for every point on the body’s
surface is no longer constrained to a circular path about the principal axis, increasing the dimensionality of the search space from 1 to 2.
There are, however, other constraints imposed by the geometry that can aid the development of a vectorbased search algorithm, such as, the relationship of normal
to surface point.
3.1 Initial estimate for the vectorbased search
The initial estimate for beginning the search is based on the vector between the centroids of the two bodies:
(1)
where Ri is the rotation matrix for body i, CoGi is the geometrical center of body i, CoGj is the geometrical center of body j, and is the unit vector in the direction
of the principal axis in local coordinate of body i.
Using the angular estimate can be found as:
(2)
(3)
Let a, b, and c be the axis lengths for the local is given by the implicit equation:
(4)
(5)
Evaluating at z=0 gives the first point of the estimator, which is also at the point of the minimum curvature for the ellipse in Equation 4 given by the equation:
(6)
3.2 First iteration
The vector can now be updated as:
(7)
Project the line axis. To find the point of intersection of the line segment and the ellipse defined in Equation 4, we first define line segment as:
(8)
To determine the initial estimate for the contact point, we use Equation 4 and Equation 8:
Page 252
which yields the quadratic equation for r as:

By definition the r2 term and r are both nonnegative, yielding the solution as:
(9)
(10)
where
To calculate the curvature at the point the ellipse is reparameterized as:
Evaluating the equation for curvature,
yields the following equation:
To calculate the normal at point we first differentiate Equation 4 with respect to z:
(11)
(12)
which leads to the tangent at the point as:
(13)
and the normal at the point as:
(14)
4 SUCCESSIVE ITERATIONS
Successive iterations follow the template presented in the first iteration, except the terminal position of the along the normal found in Equation 14:
(15)
The vector can now be updated as:
is projected onto body i according to Equation 8. However, the known point is altered somewhat as:
Then the updated contact point estimate, the ellipse is reparameterized as specified above and the curvature, normal, and tangent is found as in Equations 11–14.
5 IMPLEMENTATION CONSIDERATIONS
Though the development appears complicated, the code required for implementation is reasonably compact. Each evaluation requires an initialization overhead of
approximately 80 flops as well as the evaluation of 6 square root functions. An iteration also requires approximately another 250 flops and the evaluation of 8 square
root functions. To reach within 1 degree of skew between the normals, at the determined contact point on each surface, approximately 7 iterations are required. This
leads to an estimate of 2000 flops per
Page 253
Figure 2. Ellipsoidal particles in contact.
detailed contact check. Assuming the required memory can be stored in cache, which is reasonable given the numbers that must be stored, on a 2.5 GHz processor
each detailed contact detection step should require on the order of 3 μs (or 3 seconds per million pairs).
6 CONCLUSIONS
An algorithm has been presented to offer a different approach for calculating the contact point between two generalized, axially asymmetric ellipsoids (Fig. 2) using the
Method of Equivalent Spheres. The algorithm relies on the constraints imposed by the ellipsoidal geometry to reduce the surface point search from a coupled three
dimensional search space to a coupled twodimensional search space.
Because of the almost exclusive use of vector operations, the evaluation of transcendental functions is minimized, which significantly reduces the computational
requirements. In terms of theoretical floating point operation count the algorithm is approximately 2000 floating point operations (flops), or approximately 3 μs per pair.
Because of the methodology behind the Method of Equivalent Spheres, extension to contact between ellipsoids and spheres or ellipsoids and triangular facets is
straightforward.
REFERENCES
O’Connor, R., 1996. A Distributed Discrete Element Modeling Environment—Algorithms, Implementation and Applications. PhD Thesis MIT, February 1996.
Williams, J.R., & O’Connor, R., 1999. Discrete Element Simulation and the Contact Problem. Archives of Computational Methods in Engineering, 6(4):279–304.
Munjiza, A., Bicanic, N., & Owen, D.R.J., 1993. BSD Contact Detection Algorithm for Discrete Elements in 2D. In Proceedings of the 2nd International Conference
on Discrete Element Methods, Massachusetts Institute of Technology, March 18–19, IESL Publications.
Wensel, O., & Bicanic, N., 1993. A quad tree based contact detection algorithm. In Proceedings of the 2nd International Conference on Discrete Element Methods,
Massachusetts Institute of Technology, March 18–19, IESL Publications.
Munjiza, A., & Andrews, K.R.F., 1998. NBS Contact Detection Algorithm for Bodies of Similar Size. International Journal for Numerical Methods in Engineering,
43: 131–149.
Perkins, E., & Williams, J.R., 2001. A fast contact detection algorithm insensitive to object sizes. International Journal for ComputerAided Engineering, 18(1):48–62.
Cook, B.K., & Jensen, R.P., 2002. Proceedings of the 3rd International Conference on Discrete Element Methods Numerical Modeling of Discontinua, Sept. 23–
25, ASCE Geotechnical Special Publication No. 117.
Favier, J.F., AbbaspourFard, M.H., & Kremmer, 2001. Modeling nonspherical particles using multisphere discrete elements. Journal of Engineering Mechanics, 127
(10): 971–977.
VuQuoc, L., & Zhang, X., 1999. An accurate and efficient tangential forcedisplacement model for elastic frictional contact in particleflow simulations. Mechanics of
Materials, 31:235–269.
Grest, G., Landry, J., Silbert, L., & Plimpton, S., 2001. Rheology of Granular Flow, Lecture at the Kavli Institute for Theoretical Physics, Santa Barbara, CA, 4. June.
Cundall, P.A., 1987. Computer Simulations of Dense Sphere Assemblies, In Proceedings of the U.S./Japan Seminar on the Micromechanics of Granular Materials,
SendaiZao, Japan, October 26–30.
Cundall, P.A., Konietzky, H., & Potyondy, D.O., 1996. PFC Ein Neues Werkzeug für Numerische Modellierungen, Bautechnik, 73(8):492–498.
Taylor, L.M., & Preece, D.S., 1992. Simulation of blasting induced rock motion using spherical element models. Engineering Computations, 9(2).
Cleary, P.W., 1998. Discrete Element Modeling of Industrial Granular Flow Applications, TASK. Quarterly—Scientific Bulletin, 2:385–416.
Mustoe, G.G.W., 2001. Cluster 2D Discrete Element Computer Code—Users Manual and 2D Discrete Element Workshop Notes. Retrieved from
http://egweb.mines.edu/gmustoe/eg535/main_EGES535.htm.
VuQuoc, L., Zhang, X., & Walton, O.R., 2000. A 3D discrete element method for dry granular flows of ellipsoidal particles. Computational Methods in Applied
Mechanical Engineering, 187:483–528.
Wang, CY., Wang, CF., & Sheng, J., 1999. A packing generation scheme for the granular assemblies with 3D ellipsoidal particles. International Journal for Analytical
Methods in Geomechanics, 25:815–828.
Johnson, S.M., Williams, J.R., & Cook, B.K., 2003. Contact resolution algorithm for an ellipsoid approximation for discrete element modeling. Engineering Computations
(in press).
Ouadfel, H., & Rothenburg, L., 1999. An algorithm for detecting interellipsoid contacts. Computers and Geotechnics, 24:245–263.
Lin, X., & Ng, TT., 1997. A threedimensional discrete element model using arrays of ellipsoids. Géotechnique, 47(2):319–329.
Ting, J.M., Khwaja, M., Meachum, L.R., & Rowell, J.D., 1993. An EllipseBased Discrete Element Model for Granular Materials. International Journal for Numerical
and Analytical Methods in Geomechanics, 17:603–623.
Page 254
Page 255
Fluid coupling
Page 256
Page 257

Fluid coupling in PFC3D using the measurement spheres
Y.Okura & H.Ochiai
Forestry and Forest Products Research Institute, Tsukuba, Japan
P.A.Cundall & Y.Shimizu
Itasca Consulting Group, Inc., Minneapolis, USA
5809 679 3
ABSTRACT: Fluid coupling into PFC3D is proposed to reproduce solidfluid interactions in soil masses, and eventually to predict debris flow motion
and the hazardous conditions. In the model, fluid is assigned to measurement spheres that surround and coincide with each solid particle. Fluid
pressure perturbations and fluid transmission are governed by volumetric strain and Darcy’s law, respectively, for each measurement sphere. The
newly developed model is verified by comparing simulation results to laboratory CU (consolidated undrained) triaxial compression tests, in which
liquefaction is approached in loose soil specimen, while pore pressure decreases, and compressive strength increases, are observed in dense soil. The
model is able to demonstrate similar phenomena with both types of soil, and Darcy flow is also revealed in the model when simulating pressure
diffusion tests.
1 DEDICATION
Since this paper was written, Dr. Yoichi Okura died in a tragic accident. Dr. Okura was a fine scientist, and he was strongly motivated to understand and model the
basic mechanisms that control debris flow, and thereby mitigate the hazards. His coauthors dedicate this paper to Dr. Okura’s memory, and hope that his goal will one
day be achieved.
2 OBJECTIVES
Debris flow hazard prediction is one of the critical issues for protecting residential areas and infrastructures from natural disasters in hillside regions. Debris flow consists
of a complex mixture of solid particles and liquid water, which flows down slopes or channels at high speed. The internal friction is decreased by excess porewater
pressure, which exceeds the hydrostatic pressure, and is related to high rates of volumetric change in a debris flow mass during rapid shearing (Iverson & LaHusen
1989).
Discrete numerical models (Cundall & Strack 1979) are recognized as effective tools for analyzing the stressstrain relations of particulate systems in various fields,
such as granular flow (Walton 1983) or soil mechanics (Ting et al. 1989). This study proposes a new conceptual model for fluid coupling in a discrete numerical model,
to reproduce solidfluid interactions and volumetric strain in soil masses; eventually it will be applied to predict debris flow motion and the hazardous areas affected by
the passage and sedimentation of the material.
3 METHODS
3.1 Numerical simulation model
PFC3D (Particle Flow Code in 3 Dimensions, Itasca 2003) is used for this study because it can perform both steadystate and dynamic simulations, and offers a
flexible interface language, which enables users to make their own constitutive models for the solidfluid system. Figure 1 shows the basic system of force transmission
between two solid particles through collision or contact in PFC3D.
Elastic springs exist in both the normal and tangential directions, and viscous dampers act to oppose relative motion between pairs of particles.
Several possible methods for incorporating fluid into PFC3D are listed below.
1. Measurement spheres—Fluid is assigned to measurement spheres, which surround each solid particle and coincide with it. Fluid pressure fluctuations are calculated
in the spheres according to the volumetric strains generated by relative displacements of particles and boundary walls intersected by the
Page 258

Figure 1. Force transmission system between particles in PFC3D. Kn=elastic spring, η=viscous damper, and μ=friction.
spheres. Fluid flow occurs between adjacent measurement spheres, using Darcy’s law and a mean permeability. The approach is appropriate for highly dense,
complex flow consisting of solid and liquid and having free surface boundaries, such as nonmarine debris flow, because pore water is discretized by measurement
spheres and moves with the associated solid particles. This model may have difficulty representing hydrostatic pressure.
2. Fluid particles—Fluid is represented by many small particles, and the interaction calculation is simpler than other methods because the same calculation process
applies to both solid and fluid particles. It can also reproduce hydrostatic pressure, but requires longer processing time because it needs a large number of fluid
particles, each having a diameter of less than 1/5 of the solid ones to allow fluid particles to pass freely between the solid particles.
3. Fluid contained in voids between solid particles—Fluid is assigned to actual voids formed with four adjacent solid particles in threedimensional analysis, and
flows between neighboring voids according to the hydraulic gradient. Fluid pressure fluctuations are calculated from volumetric strains in voids, and pressure forces
act on solid particles facing each void (Hakuno & Tarumi 1988). The apparent area of a pipe, which is formed between neighboring voids, is variable according to
contact conditions between the particles forming the pipe. This model is applicable to stressstrain analysis of slope stability, liquefaction etc., during fluid
penetration. However, it is difficult to trace major changes of void spaces during large deformations and difficult to maintain continuity conditions for the fluid.
4. NavierStokes fluid—Solid particles are set in a continuum fluid field within a closed boundary. Both liquid and gas may be used as the fluid. The entire field is
divided into rectangular cells, one side of which is larger than the mean particle diameter, and porosity is calculated from the number of particles contained in each
cell. Particles take drag force from fluid passing through a grid (Tsuji et al. 1993), and opposite forces are applied to fluid elements. This model is the only method
that treats fluid as a continuum and can effectively represent dynamically moving particles, such as suspended sediment, in a fluid field. But it is difficult to represent a
freesurface boundary, which changes dynamically with particle movements such as debris flow.
This study focuses on Method 1; Measurement spheres scheme for developing a representative model for debris flow dynamics, for the reasons mentioned above.
Figure 2 illustrates measurement spheres, in which pore fluid is assigned, and the application of pressure to neighboring particles and boundary walls. The diameter of
each measurement sphere is twice that of solid particles, and the center of each sphere is coincident with its associated solid particle. The pressure fluctuation in a
measurement sphere is calculated from mean volumetric strain in the sphere, taking account of relative displacements of particles and walls included within the
measurement sphere.
Volumetric strain during the increment of time ∆t given by
(1)

where έV=volumetric strain rate in a measurement sphere, and έV is calculated from Equation (2)
(2)

where ά11, ά22, ά33=strain rate components in the x, y, z directions respectively. άmn is computed using the following leastsquares procedure.
The strain rate tensor computed for a given measurement sphere represents the best fit to the N measured relative velocity of the particles and walls intersecting
the measurement sphere. The mean velocity and position of the particles and walls is given by
(3)
(4)
where ν m, xm=translational velocity and centroid location of particles respectively; m=axial direction index; and N=total number of particles and walls contained in a
measurement sphere. The measured relative velocity and position are given by
(5)
Page 259
Figure 2. Measurement spheres representing fluid field, and section areas for applying the fluctuations of fluid pressure to the adjacent
particles or walls.
(6)

For a given άmn, the predicted relative velocity can be written as
(7)
The error in these predicted values is given by
(8)
The condition for minimum z is
(9)
Substituting Equation (7) into Equation (8) and differentiating it as Equation (9), the following set of nine equations is obtained
(10)
where m, n=axial direction index (m, n=1, 2, 3). These nine equations are solved by performing a single LUdecomposition and back substitutions (Press et al. 2001).
Fluid pressure fluctuations are calculated by using εv as Equation (11), and applied as forces to the centers of particles which are within a measurement sphere
(referring to Fig. 2)
F=(P+∆P)s=(P+Evεv)πR2
(11)
Figure 3. Visualization for fluid gradient applied in Darcy’s law.
where ∆P=pressure change within ∆t; s=the area on which pressure acts; EV=bulk modulus of elasticity for water; and R=radius of particles. Pore water pressure is
applied to walls as expressed in Equation (12) when they are within the measurement sphere
F=(P+∆P)s=(P+Evεv)π{(2R)2−L2}
(12)
where L=distance between center of the measurement sphere and wall.
Fluid flow is also generated between neighboring measurement spheres according to Darcy’s law (refer to Fig. 3) and described as
∆qi,j=ksIi,j∆t
(13)

where k=mean permeability; s=crosssectional area of penetration; Ii,j=hydraulic gradient; and s and Ii,j are expressed as Equations (14) & (15) respectively
s=πR2
(14)
Page 260
(15)
where Pi, Pj=fluid pressure in measurement spheres corresponding to particles i and j respectively; ρ=density of water; G=gravitational acceleration; and Li,j=distance
between particle i and j. Pressure change derived by ∆qi,j is expressed as
(16)
where Vi=volume of water in a measurement sphere. For stability, the pressure response must be less than the original pressure perturbation, i.e. ∆Pi,j≤(Pj− Pi).
Hence, combining (13) to (16) gives us
(17)

The sum of ∆Pi,j for all particles and walls adjacent to particle i is calculated as
(18)
3.2 Triaxial compression tests
The model expressed above is verified by comparing simulations with CU (consolidated undrained) triaxial compression tests, using a specially designed apparatus able
to perform both stress and strain controlled compression tests (Photo1).
Test specimens are prepared in cylindrical samples of 10 cm diameter and 25 cm height (Photo 2) with various void ratios, using sandy soil having a mean
Photo 1. Triaxial compression test apparatus.
diameter of 0.5 mm and a uniformity coefficient of 3.5. Specimens are covered with a rubber sleeve on their side faces, and consolidated to 20 kPa while saturated
with water. They are then sheared through undrained compression under both stress and strain control.
These tests reveal a critical void ratio of 0.67 above which the specimens liquefy with volumetric contraction, and below which they dilate and display increased
compressive strength.
Figure 4 shows a cylindrical specimen generated by PFC3D in which the diameter is 5.0 cm and height
Photo 2. Specimen set on the pedestal of the loading unit.

Figure 4. Initial arrangement of particles in a specimen generated by PFC3D.
Page 261
was 10 cm. Solid particles having a diameter of 1.0cm are packed in the specimen with void ratios of 0.98 and 0.8 for loose and dense packed soil layers,
respectively. Particle number for the loose and

Table 1. Physical and calculation properties for the triaxial test simulations in PFC3D.
Setting Numerical value
Height of the specimen 0.1 m
Diameter of the specimen 0.05 m
Particle diameter 0.01 m
Particle density
2573 kg/m3
Kn for particles
5.0×106 N/m
Ks for particles
5.0×106 N/m
Kinetic friction 0.59
Kn for walls
5.0×106 N/m
Confining pressure
2.0×104 Pa
Back pressure
1.0×105 Pa
Modulus of elasticity
2.3×106 Pa
Axial loading speed 0.01 m/s
Permeability
1.6×10−3 m/s
dense specimens is 170 and 234, respectively. Table 1 records the physical and calculation conditions for the simulations.
4 RESULTS AND DISCUSSIONS
4.1 Triaxial compression tests
Figure 5 shows the results of loose and dense soil specimens in the triaxial compression tests (Okura et al. 2002). Back pressure of 100 kPa has been subtracted from
the porewater pressure axes. Compressive strength (principal stress difference) decreased and pore pressure increased in loose soil during compression, and pore
pressure coefficient A dramatically increases, which indicates the onset of liquefaction. Shear stress on the stress pass does not increase in the loose soil. However in
dense soil, compressive strength increases and pore pressure decreases through volumetric dilation, reading to an increase of shear stress along the failure line in a
stress pass.
Figure 5. Left side graphs show the results for CU triaxial compression tests with loose soil specimen prepared in 0.70 of void ratio. Right
side graphs are for the dense soil with 0.64 of void ratio.
Page 262
4.2 Numerical simulations
Figure 6 shows simulation results for loose and dense soil specimens. The principal stress difference shows little compressive strength increase in loose soil, whereas the
pore pressure increases to around 20kPa, almost the same as the confining pressure, indicating liquefaction is produced during compression. The stress pass does not
show any shear resistance.
In dense soil, on the other hand, the principal stress difference increases and pore pressure decreases with shear. Shear resistance increases along with the failure
line, the inclination of which is around 0.8—larger than the kinetic friction, because of interlocking between particles.
These tendencies in the stressstrain relations with loose and dense soil specimens in Figure 6 are similar to those observed in the experiments shown in Figure 5.
Figure 7 shows line segments expressing normal compressive forces between particles after the compression simulations for loose and dense soil specimens. The
forces disappear around the central
Figure 7. Normal force transmission between contacting particles after the triaxial compression simulations with loose (left) and dense
(right) soil specimens.
Figure 6. Left side graphs show the simulation results for CU triaxial compression tests with loose soil specimen of 0.98 void ratio. Right
side graphs are for the dense soil, with a void ratio of 0.80.
Page 263
area due to pore pressure increase of loose specimen, whereas in the dense specimen, stable chains of particles are developed (mainly in the axial direction), with many
small lateral supporting forces connected to the circumferential wall.
This model uses Darcy’s law for the fluid flow between measurement spheres. Figure 8 illustrates simulation results for verifications of Darcy’s flow. Particle
movement is restricted and boundary conditions at the bottom and top are set to constant pressures of 1.0 kPa and 0.0 kPa, respectively. Pressure diffusion from the
bottom to the top, caused by fluid flow in the specimen, are represented by the shaded particles.
The curves in Figure 9 show pressure distribution along the axis of the specimen from the initial stage to 800 steps after of PFC3D calculation cycles. The pressure
distribution in the final equilibrium stage is
Figure 8. Darcy’s flow simulation in a specimen with boundary conditions of 1.0 kPa at the bottom and 0.0kPa at the top. Shade
represents fluid pressure magnitudes.
Figure 9. Pressure distributions in several calculation steps along the axis of the specimen.
almost linear with axial distance, verifying that Darcy’s law was adequately represented in the full specimen.
There is no obvious difference in the results, represented in Figure 9, between simulations using the automatically estimated calculation time steps of PFC3D and
those using the restricted time step satisfying Equation (12), with the former time step being almost twice as large as the later one. Since Equation (12) is derived from
the least pressure difference between two measurement spheres, the time step estimated by Equation (12) may be too conservative when applied to a complete
assemble.
5 CONCLUSION
A new scheme for fluid coupling into PFC3D is proposed in this study, to reproduce stressstrain relations during large deformation of saturated soil layers, with
potential application to subaerial (nonmarine) debris flow dynamics and hazard prediction. Fluid is discretized and assigned to measurement spheres, containing a
solid particle at each center. Fluid pressure fluctuations are calculated from volumetric strains arising from relative displacements of solid particles and boundary walls
intersecting the measurement spheres. Fluid flows between measurement spheres is governed by Darcy’s law. The model is verified by comparing to experiment results
of CU triaxial compression tests and numerical tests for pressure diffusion.
Several main findings for the new model are listed below:
1. In a loose soil, liquefaction through undrained shear during a triaxial compression simulation is generated, porewater pressure increased and effective stress
decreases in the specimen.
2. In a dense soil simulation, porewater pressure decreases and effective stress increases, leading to the increase of shear stress.
3. Fluid pressure distribution along the axis of the specimen is linear for a flow simulation, which indicates that Darcy’s law is adequately represented in the model.
The newly proposed model is able to adequately reproduce the fluid pressure fluctuations related to volumetric strain of soil layers and fluid flow governed by Darcy’s
law. Therefore, it should be possible to apply it to debris flow dynamics and predictions of the associated hazards.
REFERENCES
Cundall, P.A. & Strack, O.D.L. 1979. A discrete numerical model for granular assemblies. Geotechnique 29(1): 47–65.
Page 264
Hakuno, M. & Tarumi, Y. 1988. A granular assembly simulation for the seismic liquefaction of sand. Structural Eng./ Eartquake Eng. 5(2):333s–342s. (Proc. of JSCE
398).
Iverson, R.M. & LaHusen, R.G. 1989. Dynamic porepressure fluctuations in rapidly shearing granular materials. Science 246:796–798.
Okura, Y., Ochiai, H. & Sammori, T. 2002. The effect of void ratio on flow failure generation caused by monotonic liquefaction. Proc. Interprevent 2002 in the Pacific
Rim., Matsumoto, 14–18 October 2002: 537–545. Tokyo: Nissei Eblo.
Press, W.H., Teukolsky, S.A., Vetterling W.T. & Flannery, B.P. 2001. Numerical Recipes in C++ 2nd Edition: 46–49. NY: Cambridge.
Ting, J.M., Corkum, B.T., Kauffman, C.R. & Greco, C. 1989. Discrete numerical model for soil mechanics. Geotechnical Engineering 115(3):379–398.
Tsuji, Y., Kawaguchi, T. & Tanaka, T. 1993. Discrete particle simulation of twodimensional fluidized bed. Powder Technology 77:79–87.
Walton, O.R. 1983. Particledynamics calculations of shear flow. In J.T.Jenkins & M.Satake (eds), Mechanics of Granular Materials: 327–338. Amsterdam: Elsevier.
Page 265
Flow deformation of a river dike due to seismic liquefaction
T.Honda & I.Towhata
The University of Tokyo, Tokyo, Japan
5809 679 3
ABSTRACT: Ground deformation of a river dike due to seismic liquefaction was simulated by using distinct element method with pore water model.
Simulation of another model with sheet piles as mitigation measure to minimize subsidence of a river dike was also carried out to evaluate the effect of
underground walls in restricting lateral movement of the soil under a river dike. For the simulation of the model with mitigation measure, a new
approach of modeling flexible boundaries such as sheet piles is introduced. In this approach, line elements with particle ends and with Young’s
modulus and viscosity to reproduce boundary bending stiffness are used. In the simulation, the sheet piles were represented with truss structures
composed of these line elements. Based on the stresses and strains observed in the simulations, the deformation mechanism and behavior of liquefied
soil and the effects of sheet piles were also discussed.
1 INTRODUCTION
Many river dikes were constructed on loose and soft soil. Such a soil could easily liquefy during earthquake and could cause damages such as cracks and subsidence
along a river dike. In the 1995 Kobe earthquake many damages of dikes were reported (Matsuo 1995). After the earthquake, many researchers have studied on
deformation mechanism of a dike with liquefiable subsoil and some mitigation techniques. For example, Mizutani et al. (2001) carried out shaking table tests of a river
dike and summarized the deformation mechanism as follows (Fig. 1): when the subsoil under a river dike liquefies during an earthquake, the subsoil moves laterally due
to the weight of a dike, resulting in subsidence of a dike. They also demonstrated the effect of underground walls in minimizing lateral movements.
The present study aims to investigate deformation mechanism of a river dike and contribute in understanding the behavior of liquefied soil. In this study, ground
deformation of a river dike during earthquake was simulated by using distinct element method with pore water model. Simulation of another model with
Figure 1. Deformation mechanism of a river dike due to seismic liquefaction.
sheet piles as mitigation measure was carried out to investigate the effectiveness of such mitigation measure. In the simulation with mitigation measure, new line elements
were introduced to model the sheet piles. Based on the stresses and strains observed in the simulations, deformation mechanism, the behavior of liquefied soil and the
effects of sheet piles were discussed.
2 NUMERICAL MODEL
2.1 Distinct element method with pore water model
Distinct element method with pore water model was adopted in order to simulate liquefaction of granular material. The original model was proposed by Nakase et al.
(1999). To minimize calculation time, the analyzed area is divided into squares, which are called cells (see Fig. 2). Pore water pressure in each cell is calculated from
injected volume and bulk modulus of the cell. For accuracy, it is important to determine the bulk modulus of the cell from that of water and particles. However, it is
reasonable to use the bulk modulus of water as that of the cell since the bulk modulus of water is much smaller than that of particles. Assuming that the bulk modulus of
the cell is the same as that of the water, pore water pressure is obtained by solving the following simple equation,
(1)
Page 266
Figure 2. Volume change in the cell by movement of particles and seepage.
where DV(I, J) is the injected volume of particles in the cell (I, J) from the adjoining cells and is calculated from average displacement of particles in the adjoining cells.
Q(I,J) is the injected water flow rate by the seepage based on Darcy’s rule. ∆t is the time interval of calculation step. Kw(I,J) is the bulk modulus of water in the cell
(I,J) and is the pore water pressure in cell (I,J) at time t. Once the pore water pressures are obtained, the seepage force on the particle is calculated from the
hydraulic gradient between adjacent cells on both sides. Additionally buoyancy and viscous resistance are calculated. These forces are added as total force acting on
individual particles. The remaining routines are the same as those proposed by Cundall (1971).
2.2 Line elements
To simulate sheet pilesoil interaction, a new approach for modeling flexible boundaries, such as sheet piles, is introduced. In this approach, line elements are used to
model these boundaries as illustrated in Figure 3. The displacement and rotation of line elements are determined by the displacement of two particles having masses at
the end of the line elements, in proportion to the distances where contact forces are acting. The sum of contact forces for each particle of a line element can be
calculated using the following equations:
(2)
where Fi1 and Fi2 are the sums of contact forces at the ends of Ei1, Ei2, respectively, in the ith boundary, Oj is the position of jth contact point and Fj is contact
force at the point of Oj. The boundary forces and stresses are also divided to the two particles of the line element. The displacement of the two particles is obtained
from the equation of taking into account the particle masses
Figure 3. Reproduction of a flexible boundary.
Figure 4. Modeling of sheet piles by using line elements.
and total force acting on them. Additionally, axial force on line elements is considered by using a coefficient of modulus and viscosity.
Figure 4 shows the modeling of a sheet pile with a truss structure composed of the line elements. Coefficients of modulus and viscosity of the line elements were
determined from calibrations in a free oscillation.
3 TEST PROGRAM AND PARAMETERS
Figure 5 shows the schematic diagrams of the two cases simulated herein. The first case is a model with no mitigation (unmitigated) and the second, with mitigation using
sheet piles. In both cases, no slip was assumed at the bottom boundary. Each model consists of a saturated soil layer with the width of 6400 mm and the depth of 800
mm and a dike with 400 mm in height. The soil above the water level was assumed to be unsaturated and constant meniscus forces are given between particles. Note
that the meniscus force was not intended to simulate the behavior of unsaturated soil but to add a certain stress to the particles near the water surface. Otherwise, the
particles would collide
Page 267
Figure 5. Schematic diagram of analyzed models.
Table 1. Parameters in the analysis of infinite slope.
Number of particles 14,000
Average diameter 20 mm
Minimum diameter 2 mm
Maximum diameter 40 mm
Density of particle
1.7E+03kg/m3
Density of water
1.0E+03 kg/m3
Normal stiffness (kn) 9.5E+06 N/m
Shear stiffness (ks) 2.4E+06N/m
Normal damping (ηn) 3.4E+04 Ns/m
Shear damping (ηs) 8.4×102 Ns/m

Cohesion (c′) 0.0 N
30.0 degrees
Friction angle
Coefficient of permeability 2.0E–04 m/s
Bulk modulus of water, Kw
4.5E+08 N/m2
Bending stiffness of sheet pile
1.6E+02 Nm2
Interval of a calculation step 5.0E–06 s
Calculation steps 800,000
Figure 6. Input acceleration of shaking wave.
with each other at low stress level during an earthquake and would hardly create static shear strength. In the case of sheet piles, they were inserted at the toes of a dike
to minimize lateral displacement. The bending stiffness of these sheet piles is the same as that of aluminum sheet with the thickness of 3 mm. The numerical values of the
parameters that were used in the simulation are listed in Table 1.
Liquefaction was induced by shaking with acceleration time history shown in Figure 6. The acceleration
Figure 7. The final ground deformation.
Figure 8. Locus of ground movement.
in Figure 6 acted on all the particle and fluid media as inertia force.
4 RESULTS AND DISCUSSION
4.1 Ground deformation and settlement
The final ground deformation and locus of ground movement are shown in Figures 7 & 8. It was observed in Case 1 (unmitigated model) that lateral movement was
largest at the middle depth of the saturated soil layer under the dike. The reason for this is that the friction at the bottom and the boundary with the dike restricted lateral
deformation. This deformation is similar to that of model tests. Close to side walls, direction of ground movements changed from lateral to upward (Fig. 8), and the
deformation became smaller because they were restricted by the side walls. It can be said from Figure 8 that the influence of side wall ranges at least to the same
distance as the depth. In Case 2 (model with sheet piles), lateral displacement at the inner side of sheet piles was minimized. Deformation at the middle depth of the
saturated soil layer was markedly reduced.
The settlement at the bottom of a dike for each of the two cases is shown in Figure 9. From this figure, the settlement in Case 1 started at about 1 sec, and then
increased up to 140 mm at an almost constant rate. On the other hand, the settlement in Case 2 increased with
Page 268
Figure 9. Settlement at the bottom of a dike.
Figure 10. The number of contact point in Case 1.
Figure 11. The number of contacts point in Case 2.
a repeating upanddown motion after about 1 sec. The upanddown motion was though to be due to sloshing between the two sheet piles.
Figures 10 & 11 illustrate particletoparticle contact at 0.5 sec time interval. This interval is the same as the period of input acceleration (Fig. 6). In these figures,
particles which are colored black have more than three points of contact with other particles. Particles colored white do not have any contact. i.e., these particles are
floating in pore water, a condition that can be called “perfect liquefaction”.
From Figures 10 & 11, it was observed that white particles gradually increased and decreased after the time of 3.5 sec. Comparing the distribution of white particles
in the case of the unmitigated model and that of the model with sheet piles, it can be seen that the perfectly liquefied region was limited in model with sheet piles and that
the soil around sheet piles hardly lost any contact points. The sheet piles apparently minimized movement of particles and shear deformation.
Figure 12. Distributions of mean effective stress at the time 1, 2, 3 and 4 sec. in Case 1.
Figure 13. Distributions of mean effective stress at the time of 1, 2, 3 and 4 sec. in Case 2.
4.2 Relationship of stress and strain
Figures 12 & 13 show the distribution of the mean effective stress. This stress was obtained from contact forces between particles by using the following equation
(Cundall et al. 1982):
(3)

where Np is the number of particles in Volume V, is the cth
Page 269
Figure 14. Time histories of the mean effective principal stress under a dike in Case 1 (The position of these points is shown in Fig. 5).
Figure 15. Relationship between shear stress and shear strain under a dike in Case 1 (The position of these points is shown in Fig. 5).

contact force, rk is the radius of the particle and mk is the number of contact points of kth particle.
From Figures 12 & 13, it was observed that the recovery of mean effective stress due to shear deformation was induced from the bottom boundary. However, the
degree of recovery was gradually decreased, and then the effective stress was lost at the end of shaking. In the model with sheet piles, the recovery of effective stress
was observed around the sheet piles.
Figures 14 & 15 show time histories of the mean effective stress and relationships between shear stress and shear strain at the several points under the dike in Case
1. The position of these points is shown in Figure 5. Strain was obtained from the following equations (Bardet & Proubet 1991):
(4)
(5)
Page 270
Figure 16. Time histories of mean effective principal stress around the sheet pile and the side wall in Case 2 (The position of these points
is shown in Fig. 5).
Figure 17. Relationship between shear stress and shear strain around the sheet pile and the side wall in Case 2 (The position of these
points is shown in Fig. 5).
where Fij is the average displacement gradient, ui is the displacement of particles, NI is the number of intercepts by the perimeter S of the averaging area V, and ν j is
the unit vector normal to S.
In Figure 14, the mean effective stress at the bottom under the center of a dike (P69) decreased from the beginning of the shake, became zero at about 1.2 sec and
hardly increased afterward. In contrast, the mean effective stress at the points apart from the center of a dike increased and decreased repeatedly. It was observed
from Figure 15 that shear stresses near the surface and under the center of a dike hardly changed, because the state of liquefaction lasted and the recovery of the
effective stress did not occur. In contrast, the soil elements in the area where the recovery of the effective stress was induced showed the tendency as follows; shear
modulus softened with shear stress increasing and shear strain largely increased with a low stress state after the perfect liquefaction occurred. This behavior is similar to
that of laboratory shear tests.
To examine the sheet pilesoil interaction, the mean effective stresses and relationships between shear
Page 271
stress and strain around the sheet pile and the side wall are shown in Figures 16 & 17. The position of these points is also shown in Figure 5. In the points of P2,
P1258 and P2357 near the side wall, shear stress was hardly observed though the mean effective stress increased and decreased repeatedly. The reason for this is that
soil beats against the side wall and induced deviator stress. Shear deformation near the surface is thought to be due to the sloshing. In the outside of sheet piles, the
large recovery of the mean effective stress was observed since passive earth pressure acted on the sheet piles. It can be said that the sheet piles and the outside soil of
them became in a body. In the inner side of sheet piles, the recovery of the mean effective stress was also observed except for near the surface. The reason why the
recovery of the effective stress hardly increased near the surface is thought to be due to sloshing.
5 CONCLUSIONS
The following conclusions were drawn from the simulations of a river dike with no mitigation and with sheet piles as mitigation measure.
1. In the unmitigated model, ground deformation obtained from the simulations was similar to that of model tests. In the model with sheet piles, sheet pile could be
reproduced by a truss structure composed of line elements.
2. The recovery of the effective stress due to shear deformation was induced from the bottom boundary. Ground deformation of a river dike was strongly influenced by
the boundary conditions. Additionally, the influence of the side walls ranges to at least to the same distance as the depth.
3. The sheet piles restricted the lateral displacements at the middle depth between them and limited the liquefied region. However, sloshing occurred in the inner side of
sheet piles and the lateral deformation near the surface were not reduced at all.
4. The sheet piles and the surrounding soil became in a body and resisted lateral displacement of soil in the inner side of the sheet piles.
REFERENCES
Bardet, J.P. & Proubet, J. 1991. A numerical investigation of the structure of persistent shear bands in granular media. Geotechnique 41(4):599–613.
Cundall, P.A. 1971. A computer model for simulating progressive large scale movements in blocky rock systems, Symp. ISRM, Nancy, France, Proc. Vol. 2, 129–136.
Cundall, P.A., Drescher, A. & Strack, O.D.L. 1982. Numerical experiments on granular assemblies: Measurement and observations. IUTAM Conf. on Deformation and
Failure of Granular Materials, Delft, 355–370.
Matsuo, O. 1996. Damages to river dikes. Special issue of Soils and Foundations, 235–240: Japanese Geotechnical Society.
Mizutani, T., Towhata, I., Shinkawa, N., Ibi, S., Komatsu, T. & Nagai, T. 2001, Shaking table tests on mitigation of liquefactioninduced subsidence of river dikes,
Proceedings of the 15th international conference on soil mechanics and geotechnical engineering, Istanbul, Turkey, 1207–1210.
Nakase, H., Takeda, T. & Oda, M. 1999. A simulation study on liquefaction using DEM. Proceedings of the Second International Conference on Earthquake
Geotechnical Engineering, Lisbon, Portugal, 637–642.
Page 272
Page 273
A study on the calculation of particle volumetric deformation in a fluid coupled
PFC model
L.Li
SINTEF Petroleum Research, Norway
R.M.Holt
Norwegian University of Science and Technology and SINTEF Petroleum Research, Norway
5809 679 3
ABSTRACT: A fluid coupled discrete particle model (based on PFC) is developed to simulate the poromechanical behavior of fluidsaturated porous
(granular) rock. A key issue is to obtain good accuracy in the calculation of the particle volumetric deformation resulting from pore pressure changes,
and from intergranular stresses. In PFC, the particles are assumed rigid, but they are permitted to overlap in compression. It was found through
detailed simulation of grain compression that the particle overlap may be used as a simple but practical estimate of particle volumetric deformation due
to particleparticle interaction. Moreover, a simple way to account for particle volumetric deformation due to pore pressure variation by altering the
particle radius has been implemented in the discrete particle modeling scheme. Numerical verification experiments have been performed with
unbonded particles, demonstrating good correspondence between modeled response and what is expected from poroelastic theory.
1 INTRODUCTION
Reservoir rocks are porous, made up of solid frame and pores filled with isolated or freely moving fluid (gas and/or liquid). Both the solid frame and the fluid contribute
to the mechanical response of the rocks. The wellknown thermodynamically based Biot theory (Biot 1941) is widely used to describe macroscopic mechanical
properties of the porous rocks. The Biot theory considers the interaction between the skeleton and the pore fluid. Pore pressure change induces rock deformation,
while rock deformation causes pore pressure alteration if the pore fluid cannot be drained out or injected instantaneously. This is the source of timedependent
deformation known as consolidation.
A PFC (Itasca 2002, 2003) model is a good numerical analogue of a porous rock, especially a granular rock like sandstone where model particles may be
associated with sand grains, and pores inherently exist in the numerical model. We have developed a scheme to couple fluid flow in a PFC model, starting from basic
ideas originally developed by Cundall (1999). A PFC model simulates an assembly of disks (2D) or spheres (3D). A flow network can be constructed based on the
pores in a PFC model (Li & Holt 2001). The interaction between the fluid and the solid frame is directly simulated. The deformation of the solid frame (particle
assembly) affects the geometry of the pores and the flow paths. The fluid pressure around a particle is integrated and applied to the particle. This scheme allows hydro
mechanical simulation at pore/grain scale. It can be applied to study the fluid transport in porous media, for instance by analyzing effects of stress on permeability, as
was done by Li & Holt (2001).
The preliminary version of the fluidcoupling scheme above needs to be improved in order to become fully consistent with poromechanical theory. The particles in
PFC are assumed rigid to start out with. In poromechanics, however, grain deformation needs to be accounted for. Two main refinements are required:
1. The fluid pressure acts on the particles. The pressure variation induces volumetric deformation of the particles.
2. The volume deformation of the particles as a result of intergranular stress needs to be accounted for.
In a PFC model, even if the particles are assumed rigid, grain contact deformation is mimicked by allowing overlap (so called soft contact). The overlap is used to
calculate the interaction between the particles according to the constitutive contact law. This is a proper approximation in normal mechanical calculations, even though
the real particles should not
Page 274
overlap and should not be rigid. However, in a fluid coupling simulation, the solid deformation affects pressure change in the pores. It is important to study how much
error is induced due to the assumption of rigid particles, in order to correct this error when it is necessary.
In the next section, we will briefly review Biot’s poroelastic theory. The coupled PFC scheme has not applied Biot’s formulations. However, the concept of the
mutual interaction of the solid and fluid in the numerical scheme is to some extent comparable to Biot’s theory. The coupled PFC model should be consistent with
Biot’s theory as long as both macroscopic and grain deformations obey linear elasticity. Beyond elasticity, the discrete particle approach should be more general than
Biot’s simple theory in the sense of its applicability, since complex features such as anisotropy, nonlinearity and plasticity all emerge from simple contact laws. The
basics of the coupling scheme will be presented in the third section. A study on particle deformation due to pore pressure loading and due to particleparticle interaction
will be presented in sections 4 and 5. As examples of application, simulations of simple poromechanical experiments are presented in section 6. Finally, we summarize
this paper.
2 BASICS OF BIOT’S POROELASTIC THEORY
Biot (1941) assumed some basic properties of the medium, such as homogeneity, isotropy, linearity of stressstrain relations (small strains), and reversibility of stress
strain relations under final equilibrium conditions. It should be emphasized here that the fluid coupling scheme presented in this paper is independent of Biot’s theory.
Biot’s theory is reviewed because of its wide application in the mechanics of fluid saturated porous media, and because it is a benchmark for the resulting numerical
approach.
According to Biot’s theory, the constitutive equations of a porous medium can be written as (Detournay & Cheng 1993):
σii=λεv+2Gεii−Cς,
p=Cεν−Mς (1)
where λ and G are (undrained) elastic parameters, while C and M are additional poroelastic parameters, required to describe a twophase solid+fluid medium. The
ratio C/M equals to the Biot coefficient α:
(2)
Here Kfr and Ks are drained rock and solid grain bulk moduli. σii and εii in Equation (1) are normal stress increments and normal strain increments, p is fluid pressure
increment, while ζ is the exchange of fluid volume per unit volume of the porous medium:
(3)

where φ is porosity, Vp is pore volume, and Kf is fluid bulk modulus. The Biot α controls the effective stress for deformation of the porous material:
σ′=σ−αp=Kfrεv.
(4)
The volumetric strain εv=∆V/V. If a drained test is performed (no pore pressure change), then the pore volume and bulk volume changes are related (from Equations
(1) & (2)) as
(5)
If another test is performed by injecting fluid, with no volume deformation, then the total stress increment from Equation (4) is
∆σ=αp.
(6)
Biot’s theory describes the macroscopically mechanical response of porous rocks, while the coupled PFC model mimics the porous rocks at grain scale. The coupled
model should be valid in a wider range. When the conditions for Biot’s theory are satisfied, the coupled PFC model should be consistent with Biot’s theory.
3 BASICS OF THE FLUID COUPLING SCHEME
The flow network model is constructed based on the voids in the particle model. It consists of pores connected by pipes. Here, the wide spaces in the pore system are
called pores. The throats between the pores are called pipes. Taking the 3dimensional case as an example, 4 neighboring particles are firstly selected in a model. They
are used to create a pore. The pore space is the tetrahedron subtracted the part occupied by the solid (Fig. 1a). The tetrahedron has three faces. Each forms a pipe.
Note that each pipe that is correspondent to a triangular face has two sides. Searching the nearest particle on one side of a pipe, this particle and the three particles that
form that pipe can be used to create a new pore. Going on in this way, a flow network that covers the whole particle packing can be created. Figure 1b shows a part
of the mesh of tetrahedrons. The flow resistance or conductance of a pipe is calculated according to the area of the triangular face subtracted the part occupied by the
solid. In order to improve the
Page 275
Figure 1. In a 3D model, a tetrahedral pore is created with four particles. The pore has four pipes connected to its neighbors. Each pipe
corresponds to a face of the tetrahedron. (a) A schematic pore; (b) Part of the tetrahedron mesh of a model.
computational efficiency, some wellconnected pores can be merged.
The fluid coupling schemes in a 2dimensional and a 3dimensional model are similar. However, the 2dimensional model has to be treated a bit more artificially
because of its geometrical limitation: Strictly speaking, there is no flow path in a 2D particle model if the particles are in contact or overlap. In order to introduce it, a
fictitious layer of particles is placed adjacent to the 2D model, and the fluid flow network is then constructed in the same way as for the 3D model.
The parameters of the flow network are calculated from the geometry of the particle model. They are updated with the deformation of the particle model
(microscopic failures can be mimicked using simplified approaches as well). The fluid pressures surrounding each particle are integrated and the resulted force is applied
to the particle. The fluid flow is solved in an explicit scheme. In a flow calculation cycle, the fluid exchange through a pipe is:
∆Vft=C∆p∆t,
(7)
where C is the conductance of the pipe, ∆p is the pore pressure difference between two pores connected by the pipe, ∆t the time step for fluid calculation. This
computation goes through all the flow pipes in a cycle.
Then the pore pressures are updated according to:
(8)
where pi,t and Pi,t+∆t are previous and updated pore pressures of the pore, ∆VfT is the total fluid taken in the pore in a time step by summing up ∆Vfi in Equation (7)
through all pipes connected to the pore, Vpore is the pore volume. This computation goes through all the pores in a cycle.
Figure 2. Schematic illustration of the fluidcoupling scheme in a 2D particle model. A particle model combines with a flow network that
consists of pores (the disks) linked by pipes (the lines) to make a coupled model on the right (with the arrows indicate fluid
flow). The principle also applies to fluid coupling in a 3D particle model.
Since the time step for flow computation may be different from the time step for particle interaction and motion computation, a cycle of flow computation may be
performed after many cycles of particle dynamic computation, or many cycles of flow computation may be performed after every particle dynamic computation. Figure
2 schematically shows the coupling approach.
4 REFINEMENT OF PARTICLE DEFORMATION DUE TO PORE PRESSURE ALTERATION
According to Biot mechanics as outlined in section 2, the poroelastic response is related to bulk volume (framework) and solid grain deformability. This means that the
rigid particle assumption underlying PFC needs to be relaxed in order to implement poromechanical coupling.
In the simulation of transient flow, the pore pressure in the model changes with time. When the fluid is injected in or when part of the particle model is compressed,
the pore pressure will increase. This pore pressure change results in deformation of the particles. Assuming this effect can be approximated as the effect of an average
pore pressure surrounding a particle, we calculate the volumetric deformation of the particles due to pore pressure change as below:
(9)
where p is the average pore pressure surrounding a particle, and Ks is the bulk modulus of the solid material. Practically, in the numerical model, the radii of the
particles are dynamically changed according to an equation as below:
(10)

where r0 is the particle radius when the pore pressure in zero.
Page 276
In a bonded model, the particle size modification may change the gap or overlap between the particles. This should be treated carefully. One can apply an additional
force to the bonds. This force will push or pull the particle pair in order to avoid changes of the gap or overlap between the particle pair due to the particle size
adjustment.
5 PARTICLE VOLUMETRIC DEFORMATION
Solid particle deformation as a result of intergranular stresses will affect the calculation of pore volume to be entered in Equation (8), and further also the pore pressure
in the model. One may in principle apply a similar approach as done in the previous section for pore pressure loading, but since the intergranular stresses are acting
through the contacts with neighboring particles, it is not trivial to calculate the resulting particle deformation.
PFC simulations were performed in order to study the particle deformation due to particleparticle interaction. One of the simulations was set up as in Figure 3. A
sphere made from PFC material (diameter 1 mm) was compressed by two semispheres. The middle sphere and the two semispheres used the same microscopic
parameters. In fact, the two semispheres were selected and copied from the middle sphere. Contact bonds were applied to all actual contacts within the middle sphere
and semispheres (There is no bonds between the simulated sphere and semispheres). The purpose of using contact bonds is to ensure the simulated material has the
same mechanical properties in tension and compression. The bonding strengths were set high enough to avoid any bond breakages during the simulation. In order to
obtain a smooth contact, the spheres were generated using a particularly designed spherical “wall”. 22485 particles were used for one whole sphere.
In this study, only the overall elastic properties of the sphere and semispheres are of interest. A cylindrical sample was cut from the sphere. With it, a simulation of a
triaxial test was performed. The Young’s modulus and the Poisson’s ratio of the simulated grain material were
Figure 3. Setup of a model used to simulate grain deformation.
obtained from the simulation results as: E=63 GPa and v=0.13.
As the middle grain was compressed by the semispheres, the displacements and total contact forces between them were monitored. Meanwhile, the volumetric
deformation of the middle grain was monitored as well. In order to do this, the middle grain was discretized by many tetrahedrons with the centers of every 4
neighboring particles as their vertexes. The volume of the grain did not account for the rough surface. It is an approximate value. However, the volumetric strain
obtained in this way is accurate enough for the analysis.
The stiffnesscontact force relation obtained from this simulation is compared with Hertzian theory in Figure 4. The grain elastic parameters used in the HertzMindlin
equations are obtained from the simulation of triaxial test with the simulated grain material, as mentioned previously. It is seen from the results that the PFC simulation is
consistent with Hertzian theory when the contact force is low. However, the simulated grain demonstrates higher stiffness when the contact force is high.
These results are not surprising. Since Hertzian theory is derived from a solution of a half space with Hertz
Figure 4. Comparison of contact behavior of the PFC simulated grains and Hertzian theory.
Figure 5. Comparison of volumetric deformation between PFC grain simulation and the assumed particle overlap.
Page 277
Figure 6. Setup of a model used to simulate grain deformation with coordination number=6.
Figure 7. Comparison of contact stiffness calculated by Hertzian theory and by the simulated sphere compressed in the cases of
coordination 2 and 6.
Figure 8. Comparison of volumetric deformation between PFC grain simulation (when the grain is compressed by 6 neighbors) and the
assumed overlap of rigid spheres.
proposed pressure acting on a circular area on the surface plane, it is valid only when the grain diameter is much larger than the diameter of the contact area (Johnson
1985). The simulation results are qualitatively consistent with some experimental observations (e.g. Liu et al. 1998). In experimental study, it is usually difficult to
conclude if the mismatch with Hertzian theory is due to plastification at the contact or due to any violation of the required condition.
Figure 5 shows the comparison of the volumetric deformation obtained from the grain compression simulation. The results are compared to the volume change
estimated from particle overlap in a soft contact model applied to the large spheres, treating them as rigid bodies. The displacement between the spheres is calculated
from the contact force according to Hertzian theory. The results show that the volumetric deformation for the overlapped rigid spheres is smaller than the volumetric
deformation of the simulated grain when the contact force is small. The discrepancy between them is not large in the simulated range when the contact force is low.
Since the grain diameter is 1 mm, the contact force 1N between the grains is approximately correspondent to stress 1 MPa in the porous rock. The discrepancy tends
to increase with increasing contact force up to contact forces of 100 N or so.
The set up in Figure 3 shows a case when the coordination number is 2. Usually grains in a rock may be compressed by more than two neighbors. Another
simulation was performed with a case when the coordination number is 6 (Fig. 6). In this case, the contacts around the middle grain affect each other. The simulation
results are shown in Figures 7 & 8. It is interesting to notice that for a coordination number of 6, the volumetric deformation calculated by rigid sphere overlap fits well
with the volumetric deformation of the simulated grain up to contact forces of 250 N or so (Fig. 8). The average number of loadbearing contacts in a 3D model of
densely packed spheres is usually is around 6. The simulation results suggest that it may be acceptable to assume the overlap as the particle volume change in many
cases.
6 NUMERICAL POROELASTIC EXPERIMENTS
Some numerical tests were performed using PFC3D. The refinement of particle volume change due to pore pressure change was implemented in the coupled models,
and the particle overlap was used as estimate of grain deformation caused by intergranular stress changes.
A drained test with constant pore pressure was performed with an unbonded model. The HertzMindlin contact law was applied in that model. The model simulated
a sample under hydrostatic loading. Fluid was allowed to freely move in the numerical sample.
Page 278
Figure 9. A drained test with constant pore pressure. The Biot coefficient calculated using different approaches but from the same
simulation are compared. In approach 1, the Biot coefficient was calculated from the total volume of the displaced fluid and the
volume change of the numerical sample (secant value). In Approach 2, the Biot coefficient was calculated from the increment
of the displaced fluid volume and the increment of total frame volume (tangent value). Approach 3 and approach 4 use the bulk
moduli of the frame and the particle material. The secant value (in approach 3) and tangent value (in approach 4) were
calculated respectively.
Therefore, the pore pressure in the numerical sample was constant. The Biot coefficient of the porous material (the numerical material) was calculated, even though
Biot’s theory does not apply to nonlinear elasticity. In order to verify the implemented numerical scheme, the Biot coefficients were calculated from the modeling
results in two ways: by displaced fluid volume (Equation (5) (α=ς/εv)), and from the ratio of the bulk moduli of the frame and the solid material (Equation (2)
(α=1−Kframe/Ks)).
Since this is a nonlinear material, the Biot coefficients are not constant. Both a secant value and a tangent value for α were calculated (as described in the caption of
Fig. 9). It can be seen that the results from different equations follow the similar trend. However, there is also an obvious difference between two approaches, possibly
indicating errors in the calculation of displaced fluid volume.
With the same numerical sample, keeping the hydrostatic stress at 18 MPa, an injection test with 0 frame deformation was performed. The results are shown in
Figure 10. In the model, the pore pressure increased during fluid injection. The particle radii therefore reduced due to the compression of the fluid. The total stress
(which is the summation of the fluid pressure and the stress in the solid frame) increment ∆σt was smaller than the pore pressure increment. The Biot coefficient for this
case can be calculated according to Equation (6).
The Biot coefficient was constant in this case because the bulk modulus of the frame did not change. Actually, when the particle shrank due to the increasing pore
pressure the contact stiffness of the particles may change. This nonlinearity may result in a nonconstant Biot coefficient. However, this effect was neglected in
Figure 10. The Biot coefficient in an injection test with 0 frame deformation.
Figure 11. The stress and pore pressure changes in an injection test with constant external stress.
this scheme. It could be acceptable if the pore pressure change is not crucial. Otherwise this scheme could be modified by changing the contact stiffness according to
the particle radii. Obviously, this will increase the complexity of the numerical model and reduce the computational efficiency.
Using the same numerical sample again and keeping the total external stress constant, another injection test was performed. In this case, the stress in the solid frame
σc decreased, while the pore pressure increased (Fig. 11). Therefore the total stress could be maintained constant. The Biot coefficient in this case was calculated
according to:
(11)
This is derived from Equation (4).
Because the stress in the solid frame changed when the pore pressure increased, the Biot coefficient was not constant (Fig. 12) due to the nonlinear behavior
Page 279
Figure 12. The Biot coefficient calculated for the numerical material under test as described in Figure 11.
of the model. It is observed that the Biot coefficients obtained in different tests and from different approaches are quantitatively consistent.
7 SUMMARY
Based on the results of the simulations, the simple refinement with which the effects of pore pressure variation on particle volumetric deformation seems proper. This
refinement improves the previous version of the fluid coupling scheme.
In many cases when the grains are compressed at many contacts (e.g., with a coordination number around 6), the particle volumetric deformation due to particle
particle interaction can be approximated by the overlap.
Further work is required to extend this study to cemented (bonded) granular media.
ACKNOWLEDGEMENT
This work has been supported by the Norwegian Research Council, ConocoPhillips, Norsk Hydro, Petrobrãs, Shell and Statoil through the project “Petrophysics
under stress”. The authors acknowledge the permission from these companies for publishing the results.
REFERENCES
Biot, M.A. 1941. General theory of threedimensional consolidation. Journal of Applied physics, 12(2):155–164.
Cundall, P.A. 1999. The incorporation of fluid coupling into PFC. Unpublished.
Detournay, E. & Cheng, A.H.D. 1993. Fundamentals of poroelasticity (Chapter 5). In J.A.Hudson (ed.), Comprehensive rock engineering, Vol II: 113–169. Oxford:
Pergamon Press.
Johnson, K.L. 1985. Contact mechanics. Cambridge: Cambridge University Press.
Li, L. & Holt, R.M. 2001. Simulation of flow in sandstone with fluid coupled particle model. In D.Elsworth, J.P.Tinucci and K.Heasley (eds), Rock mechanics in the
national interest: 165–172. Rotterdam: Balkema.
Liu, K.K., Williams, D.R. & Briscoe, B.J. 1998. The large deformation of a single microelastomeric sphere. Journal of physics, D: Applied physics, 31:294–303.
Page 280
Page 281

Fluid coupling in PFC2D and PFC3D
Y.Shimizu
5809 679 3
ABSTRACT: A particlefluid coupling scheme with a mixed LagrangianEulerian approach is implemented in PFC2D and PFC3D Ver. 3.1. The
scheme solves the continuity and NavierStokes (NS) equations numerically in Eulerian Cartesian coordinates, then derives pressure and velocity
vectors for each fixed grid (cell) by considering the existence of particles within the cell. Driving forces from the fluid flow are applied to the particles
as external body forces. The scheme is described in terms of the momentum interaction term between particles and fluid. It is also compared with
another scheme, used to solve underground flow coupling problems with Darcy’s law.
1 INTRODUCTION
Numerical analysis of particlefluid coupling to solve geoengineering problems has focused on fluid flow through soil, rock or porous medium. One approach makes
use of a method (“scheme A”) that is commonly used for solving groundwater flow problems, in which the equation of continuity is used for the fluid phase as the
governing equation to calculate pressure and velocity on an Eulerian grid. Darcy’s equation is used to provide a linear relationship between fluid flowrate and pressure.
The movement of the solid phase (particles) may be traced by the Discrete Element Method (DEM) as a Lagrangian method. Local porosity change by particle re
arrangement causes pressure change, resulting in a change in body forces applied to particles. Application of this mixed LagrangianEulerian approach is limited to a
small range of fluid velocity and porosity. Particles moving with the fluid flow are not handled by the formulation.
A new mixed LagrangianEulerian approach (“scheme B”) to the coupling problem is becoming popular, especially in the chemical and mechanical engineering fields
(Kawaguchi 1992 & 2003, Tsuji et al. 1993 & Hoomans et al. 1995). In this scheme, the fluid phase is described by using the continuity and momentum (Navier
Stokes) equations to calculate pressure and velocity of fluid, while the movement of the solid phase (particles) is traced by the DEM as a Lagrangian method.
Momentum exchange between particles and fluid is considered in the momentum equation for the fluid phase and in the forcedisplacement law for the solid phase.
This approach has many potential applications in the geoengineering field because it is applicable for a wide range of fluid velocity and for a high porosity range, in
which particle dynamics (for instance in particle migration processes) is as important as fluid flow through soil, rock or porous medium. Some expected application
examples are sand sedimentation, slurry flow (transportation) or fluid flow in aquifers in the vicinity of boreholes.
Scheme B is implemented in PFC2D Ver. 3.1 (Itasca 2004) and PFC3D Ver. 3.1 (Itasca 2004). In this paper, the scheme is described, with particular emphasis on
the interaction term between particles and fluid. Also, a comparison between Schemes A and B is made by performing a simple simulation of flow through a particle
assembly.
2 INTERACTION BETWEEN PARTICLES AND FLUID FLOW: SCHEME B
Figure 1 shows a fixed control volume ∆x∆y∆z in the model domain, in which several spherical particles exist within a flowing fluid.
Figure 1. Fluid flow through particle assembly.
Page 282
Assuming that the direction of fluid flow is in the xdirection, a pressure gradient dp/dx occurs only in the xdirection. By considering force balance in the xdirection
for particles within a cell, the sum of driving forces, is given by Equation (1).
(1)
where is the interaction force per unit volume between particles and fluid, dpi(i=1, np ) is the diameter at particle i and np is number of particles in the cell. The

minus sign of the first term means that the force applied to the fluid is positive. The second term represents the force acting on the particles by pressure gradient dp/dx.
Note that the minus sign of this term means that pressure decreases with positive flow in the xdirection.
By definition of porosity n,
(2)
By substituting Equation (2) into Equation (1),
(3)

Hence, the driving force applied to a single particle (i=1−np ) is given by Equation (4).
(4)
In general form, Equation (4) becomes,
(5)

where, i=1−np , j=x, y, z.
The continuity and NavierStokes (NS) equations for the fluid phase in the fluidsolid, twophase flow model, assuming incompressible fluid with constant density,
are given by Equations (6) & (7). In this formulation, both the fluid and solid phases contribute to the pressure gradient of the system (Bouillard et al. 1989).
(6)
(7)
where u is fluid velocity vector, τ is the viscous stress tensor, g is the gravity force vector, and fint is the total interaction force between particles and fluid, per unit
volume.
If the pressure gradient in Equation (7) arises only by interaction between particles and fluid (i.e., for steady, nondiverging flow),
(8)
By substituting Equation (8) into Equation (5), this case is given by Equation (9) (Asgian et al. 1995).
(9)
There are some empirical equations, relating to the interaction term, for pressure gradient through a lowporosity porous medium, based on Equation (8). One of them is
Darcy’s equation, widely used for liquid flow through porous media, in which the pressure gradient is proportional to the apparent velocity.
(10)

where, v f and ρf are kinematic viscosity and density of fluid, ux0 is apparent velocity, K is hydraulic conductivity and k is permeability.
The permeability may be related to porosity by the KozenyCarman equation (Bear 1979), for example.
(11)
where, c=0.003 to 0.0055 and is the average particle diameter.
Hence,
(12)
Darcy’s equation is valid for a range of Reynolds numbers, in which particle diameter is taken as the length dimension of 1 to 10 in the laminar flow range. For large
Reynolds numbers, nonlinear equations exist to evaluate the pressure gradient. Equation (13) records Ergun’s equation (Ergun 1952):
(13)
Comparing Equations (12) & (13), Darcy’s equation using the KozenyCarman relation is identical to the first term of Ergun’s equation except for the
Page 283
Figure 2. Comparison of pressure gradient between KozenyCarman and Ergun’s equation; data points show Reynolds number; the solid
line indicates equal values for both axes.
coefficient, 1/c and 150. The first term in Ergun’s equation, which is proportional to apparent velocity, plays a role in the laminar flow range: i.e. Re≤10. On the other
hand, the second term, which is proportional to the square of apparent velocity, is effective in the turbulent flow range.
Figure 2 shows a comparison of pressure gradient between Darcy’s equation with hydraulic conductivity by KozenyCarman relation and Ergun’s equation. As the
velocity increases, the discrepancy between the equations becomes large because of the dominance of the second term in Ergun’s equation in the high velocity range.
At low velocities, when the term proportional to apparent velocity is dominant, the pressure gradient by Ergun’s equation is in good agreement with that of Darcy’s
equation.
By substituting ux0=ux×n into Equation (13),
(14)
In numerical simulations, the relative velocity,
(where is the average velocity of particles in the domain) should be used instead of the absolute fluid velocity, since the driving force arises from relative fluid flow.
Then,
(15)
In generalized form,
(16)
In addition, for the highporosity range, the pressure gradient is calculated from the Wen & Yu (1966) equation, which is related to hindered
settling; Equation (17).
(17)
where the drag coefficient for a spherical particle is given by Equation (18) as a function of the particle Reynolds number,
(18)
where,
(19)
The interaction force per unit volume between particles and fluid is also given by Equation (20).
(20)
where, (j=x,y,z) is fluidparticle friction coefficient (Bouillard et al. 1989).
From Equations (8) & (20) the fluidparticle friction coefficient is given by Equations (21) & (22).
(21)
(22)
Figure 3 shows the flowchart of the entire PFC2D/3D calculation with the fluid scheme. The scheme is invoked at each fluid time step, interposed between the Law of
Motion and forcedisplacement law in mechanical (DEM) calculation, with the interaction forces on particles being added as external body forces in the mechanical
calculation. If the accumulated mechanical time, t m equals or exceeds the predicted time for
Page 284

the next fluid step (t f+dtf), then the fluid scheme is executed.
Figure 4 shows the flowchart of scheme B. The pressure and velocity vector of fluid in each cell are
Figure 3. Flow chart of entire PFC calculation with fluid scheme.
Figure 4. Flow chart of scheme B.
calculated by using the SemiImplicit Method for Pressure Linked Equations (SIMPLE) algorithm (Patanker 1980), which is widely used as a solver for incompressible
fluid. The interaction force between particles and fluid is calculated during the iteration.
3 SCHEME A
Assuming that the range of particle Reynolds number is 1–10 within the particle assembly, the fluidparticle friction coefficient is given by Equation (23) using the
KozenyCarmah equation with Darcy’s law.
(23)
If the substantive derivative of fluid phase in the NS equation is negligible, the NS equation deduces to Darcy’s law with the KozenyCarman equation.
(24)
By substituting Equation (24) into the continuity equation Equation (6),
(25)
where is pressure head.
Equation (25) is the equation that is used for solving groundwater flow problems. In scheme A, the governing equation with Darcy’s law is given by Equation (26) by
considering the compressibility of fluid in the time derivative term (Bear 1979).
(26)

The compressibility of fluid Bf and the coefficient of compressibility of assembly α are given by Equations (27) & (28) respectively.
(27)
Page 285
(28)
By substituting Equations (27) & (28) into Equation (26),
(29)
In scheme A, Crank Nicholson’s implicit scheme for the time derivative is used in the numerical formulation.
Figure 5 shows the flowchart of scheme A. The way the scheme is invoked in the entire PFC2D/3D calculation is similar to that of scheme B as shown in Figure 3.
The successive over relaxation (SOR) method is used for iterative solution of the algebraic equations. Local porosity in Equation (29) is traced in the DEM, and the
driving force applied to particles is given by Equation (9) in this case.
Figure 5. Flow chart of scheme A.
4 SIMULATION
A simple simulation is conducted for both schemes A and B, in which water flows through particle assembly by setting a pressure difference between each end (x
direction). The discharge flowrate is monitored during the simulation.
4.1 Model
Figure 6 shows the model, the initial state of the particle assembly and the fluid cells placed in the model. The model consists of particles and surrounding walls, 0.1
m×0.03 m×0.03 m (width× depth×height). 3108 spherical particles are created with a uniform size distribution 1.6–5 mm diameter. At x=0.1 m, line walls are set to
represent a filter, from which water is discharged. The space between these line walls is set to three times of the minimum diameter of particle in the assembly. 250 fluid
cells (10×5×5: width×depth×height) are created inside the rectangular parallelepiped space, with a cell size of 10×6×6 mm. A pressure boundary is specified for each
side end: 2×102 Pa for scheme A, 2×102 and 2×103 Pa for scheme B at x=0.0 m, and 0.0 Pa at x=0.1 m. Slipwall boundaries, in which the fluid velocity parallel to a
wall surface is nonzero at the wall surface, is specified for the other walls. The time step for the fluid calculation is 5.0×10−4 s.
Table 1 shows the material properties. The linear contact model was used, in which normal and shear stiffnesses are both 1.0×106 N/m. Particle density is 2650
kg/m3. The friction coefficients between particleparticle and particlewall contacts are 0.5 and 0.3, respectively. Viscous damping with no tensile force is used for
energy dissipation at contacts. The ratio of the damping constant to the critical damping constant in the normal and shear direction are 0.1. Density and viscosity of
water are 1000 kg/m3 and 1.0×10−3 Pa.s, respectively. The compressibility of fluid B in scheme A is 1.0×107 Pa.
f
Figure 6. Model: assembly.
Page 286
Table 1. Material properties.
Particle
Number 3108 [–]
Diameter 1.6–5 [mm]
Density 2650
[kg/m3]
Normal stiffness 6 [N/m]
1.0×10
Shear stiffness 6 [N/m]
1.0×10
Friction coefficient 0.5 [−]
Coeff icient of compressibility of assembly (used in scheme A) −4
1.0×10 [Pa−1]
Wall
Normal stiffness 6 [N/m]
1.0×10
Shear stiffness 6 [N/m]
1.0×10
Friction coefficient 0.3 [−]
Water
Density 1000
[kg/m3]
Viscosity [Pa.s]
1.0×10−3
Compressibility (used in scheme A) [Pa]
1.0×107

Figure 7. Result of scheme A, discharge flowrate at x=0.1 m, pressure 2.0×102 Pa applied at x=0.0 m.
The following simulation procedure is used. First, particles are created randomly within a box surrounded by six walls, at a porosity of 0.35, and then released under
gravity (xdirection). After reaching a steady state condition, the side wall at x=0.1 m is replaced by line walls to make a filter. The fluid scheme is invoked for 0.3
seconds after the steady state. During the simulation, discharge flowrate is recorded. The particle Reynolds number calculated by the Darcy’s law in the case of
pressure difference of 2×102 and 2×103 Pa are 4×101 and 4×102, respectively.
4.2 Results
4.2.1 Scheme A
Figure 7 shows the time history of discharge flowrate by scheme A. The pressure difference between

Figure 8. Result of scheme B after 0.3 s, fluid velocities, pressure 2.0×102 Pa applied at x=0.0 m.

Figure 9. Result of scheme B, discharge flowrate at x=0.1 m, pressure 2.0×102 Pa applied at x=0.0 m.

each end in the xdirection is 2×102 Pa. The discharge flowrate is calculated by summing the flowrate for fluid cells on the side boundary at x=0.1 m.
The flowrate converges to a constant value with very small fluctuation, which comes from the change of porosity by interaction between particles and fluid. The
flowrate calculated by Darcy’s law (Equation (10)) and the KozenyCarman equation (Equation (11)) with 1/150 as c and 0.35 as the porosity in the equation is
1.09×10−5 m3/s. The flowrate by the simulation is in good agreement with that by Equations (10) & (11).
4.2.2 Scheme B
Figure 8 shows the result of scheme B after 0.3 second in the case of 2×102 Pa, in which fluid velocity vectors within the model are displayed. The flow is almost
paralleled to the xdirection, but localized depending on the local porosity.
Page 287

Figure 10. Result of scheme B, discharge flowrate at x=0.1 m, pressure 2.0×103 Pa applied at x=0.0 m.
Figure 9 shows the time history of discharge flowrate by scheme B in the case of 2×102 Pa. The flowrate reaches about 7×10−6 m3/s with fluctuations following the
transient condition. The value is smaller than that of scheme A with Darcy’s law.
Figure 10 shows the time history of discharge flowrate by scheme B in the case of 2×103 Pa driving pressure. The flowrate after reaching steady state is about
3.6×10−5 m3/s. Compared with the case of 2×102 Pa (Fig. 9), this case reaches steady state within a short time. Also, there is no linear relationship between steady
state flowrate and pressure, because of the nonlinear equation, i.e. Ergun’s equation.
5 CONCLUSION
A newly implemented particlefluid coupling scheme with a mixed LagrangianEulerian approach is described regarding the particlefluid interaction term. It is also
compared with the scheme using Darcy’s law to make a simple simulation of flow through particle assembly. The results indicate that the scheme is able to simulate
coupling problems with large particle Reynolds number, in which Darcy’s law is not applicable to solve the fluid field. It is also expected that the scheme may be
applied in many geoengineering problems that include dynamics of particles.
ACKNOWLEDGEMENT
I sincerely thank Dr. Peter Cundall and Dr. Christine Detournay at Itasca Consulting Group, Inc. and Dr. Toshihiro Kawaguchi at Osaka University for useful
discussion and advice during developing the scheme in PFC2D/3D.
REFERENCES
Asgian, M.I., Cundall, P.A. & Brady, B.H.G. 1995. The Mechanical Stability of Propped Hydraulic Fractures: A Numerical Study. In Proceedings of the SPE 69th
Annual Technical Conference and Exhibition, vol. 1, pp. 475–489.
Bear, J. 1979. Hydraulics of Groundwater. McGrawHill, pp. 60–69 and pp. 84–94.
Bouillard, J.X., Lyczkowski, R.W. & Gidaspow, D. 1989. Porosity Distributions in a Fluidized Bed with an Immersed Obstacte. AIChE Journal 35(6):908–922.
Ergun, S. 1952. Fluid Flow through Packed Columns. Chemical Engineering Progress 48(2):89–94.
Hoomans, B.P.B, et al. 1995. Discrete Particle Simulation of Bubble and Slug Formation in a TwoDimensional GasFluidised Bed: A HardSphere Approach. Chemical
Engineering Science 51(1):99–118.
Itasca Consulting Group, Inc. 2004. PFC2D—Particle Flow code in 2 Dimensions, Version 3.1 (to be released). Minneapolis: Itasca.
Itasca Consulting Group, Inc. 2004. PFC3D—Particle Flow code in 3 Dimensions, Version 3.1 (to be released). Minneapolis: Itasca.
Kawaguchi, T., Tanata, T., & Tsuji, Y. 1992. Numerical Simulation of Fluidized Bed Using the Discrete Element Method (the Case of Spouting Bed). JSME (B), 58(551):
79–85.
Kawaguchi, T. 2003. PhD thesis, Osaka University.
Patanker, S.V. 1980. Numerical Heat Transfer and Fluid Flow. Hemisphere.
Tsuji, Y, Kawaguchi, T. & Tanata, T. 1993. Direct Particle Simulation of TwoDimensional Fluidized Bed. Powder Technology 77:79–87.
Wen, C.Y. & Yu, Y.H. 1966. Mechanics of Fluidization. Chemical Engineering Progress Symposium Series, 62: 100–111.
Page 288
Page 289
Design of gas storage rock cavern by the hydromechanical coupled discrete
model
K.Aoki, Y.Mito, T.Matsuoka & D.Kondoh
Dept. of Urban and Environmental Engineering, Faculty of Engineering, Kyoto University, Kyoto, Japan
5809 679 3
ABSTRACT: Gas storage plants in rock cavern with high internal pressure, such as LPG storage and CAES (Compressed Air Energy Storage) have
the risk of gas leakage through the conductive fracture networks around the cavern, which is formed by fracture initiation, propagation and coalescing
phenomenon due to the internal pressure. Therefore, for the planning and designing of gas storage caverns with high internal pressure, it is essential to
predict the behavior of fractures around the cavern subjected to internal gas pressure and evaluate the amount of gas leakage through the fracture
network considering the effect of groundwater flow that governs the gas penetration into rock mass. In this study, first, hydromechanical coupled
analysis based on twodimensional bonded particles model that explicitly simulate interparticle mechanics is developed in order to perform a realistic
simulation that can express rock deformation and fracturing coupled with fluid flow of gas and liquid. Then the appropriateness of this numerical
method is verified through the case study of the actual ACSC (Air Cushion Surge Chamber) site that had an experience of not only leakage of high
pressured air from the rock cavern but also air leakagecontrol with water curtain. Further, a planning and designing of the LPG storage cavern is
conducted, and the design procedure for the determination of rational design parameters is examined.
1 INTRODUCTION
The underground LPG storage plant and CAES (Compressed Air Energy Storage) plant are new style energy storage facilities, which have large rock caverns with high
internal gas pressure. The gas storage design basically relies on the principle of water pressure in surrounding rock being more than the ‘escape’ pressure of the stored
gas, while those have the risk of gas leakage through the conductive fracture networks around the cavern, which is formed by fracture initiation, propagation and
coalescing phenomenon due to the fluid pressure. Therefore, for the planning and designing of gas storage caverns with high internal pressure, it must be required to
predict the generation of fracture network due to internal gas pressure and evaluate the amount of gas leakage through the fracture network with taking the effect of
groundwater flow that governs the gas penetration to rock mass into account.
Particle flow code (PFC) described by Cundall & Strack (1979) is convenient to treat such process of fracturing as the progressive breaking of discrete bonds.
Potyondy & Cundall (1999) suggested the hydromechanical coupled formulation that utilizes DEM and channeldomain network concept. However, this formulation
has never been extended to the rock deformation and fracturing issue caused by the interaction of gas and liquid.
Thus, in this study, the authors develop the hydromechanical coupled analysis based on twodimensional PFC model that explicitly simulate interparticle mechanics in
order to perform a realistic simulation that can express rock deformation and fracturing coupled with fluid flow of gas and liquid.
The appropriateness of this numerical method is verified through the case study of the actual ACSC (Air Cushion Surge Chamber) site that had an experience of not
only leakage of high pressured air from the rock cavern but also air leakagecontrol with water curtain.
Further, a planning and designing of the LPG storage cavern is conducted, and the design procedure for the determination of rational design parameters is examined.
2 FORMULATION
2.1 Basic concept
The concept of the hydromechanical coupled model formulated by Potyondy & Cundall (1999) is shown as in Figure 1. Rock mass is represented by the bonded
particle assemble. A closed loop surrounded by the
Page 290
Figure 1. The concept of the hydromechanical coupled model formulated by Potyondy & Cundall (1999).
bonded particles, is called a “domain” that acts as a notational void in rock mass. A segment that links two neighboring domains, is called a “channel” that acts as a
notational crack with a specific aperture. The geometry of domain and channel are changed according to the relative displacement of the corresponded bonded
particles. It is assumed that this causes the following coupled phenomena in rock mass.
1. The relative displacement of the bonded particles or contact force causes the change in the geometry of the adjacent channel and then the hydraulic properties of the
channel are changed.
2. Mechanical changes in domain volumes due to fluid flow through the adjacent channel causes changes in domain pressure.
3. Pressure changes at the domains cause displacements of the bonded particles around the domains, or destruct the bond of the particles and then create a newer
channeldomain structure.
2.2 Flow formulation by Potyondy & Cundall
The channel that connects domains in bonded particles is assumed as the crack with specific aperture, which is hydraulically equivalent to the discontinuities in rock
mass. Simplify by assuming the parallel plate model, the flow rate through the crack is given by
(1)
where g=gravity; v=kinematic viscosity of fluid; a=aperture; (h2−h1)=head difference between the two adjacent domains (The sign convention is such that a positive
pressure difference produces a positive flow from domain 2 to domain 1.); and l=channel length which is taken to be the sum of the radii of the particles adjacent to the
contact.
The aperture for a crack is usually taken to be the opening of the crack. However, in order to account for the “background” permeability of an uncracked material,
each contact must be assumed to have a residual aperture when it is bonded but taking no load. It is also reasonable to assume that the addition of normal load may
decrease the apparent aperture (although the parameters may be chosen such that the coupling between normal load and aperture vanishes).
The following relation is empirical and simple: it ensures that the aperture is equal to the residual aperture for zero normal force, and then decreases asymptotically to
zero as the force tends to infinity. Note that compressive forces are taken as positive, and that the equation only applies for compressive normal forces.
(2)
where a0=residual aperture; and F0=the value of normal force, F, at which the aperture decreases to a0/2. The aperture remains at its residual value if F0 is large
compared to the expected contact forces.
When the normal force is tensile (intact bond) or it is zero, then the aperture is simply equal to the sum of the residual aperture and the normal distance between the
surfaces of the two particles, dC:
a=a0+ma dC
(3)

where ma=dimensionless multiplier which can be set different from unity for calibration purposes.
2.3 Pressure formulation
Each domain receives flows from the surrounding pipes, Σq. In one time step, ∆t, the increase in fluid volume is given by the following equation, assuming that inflow is
taken as positive.
(4)

where ∆Vd is increase in the apparent volume of the domain.
The different equation must be required to calculate the pressure change for 2 different fluid phases. For gas phase, assuming isothermal change, we can obtain
PiVGi=Pi−1VGi−1
(5)

where Pi=fluid pressure at time step, i; and VGi=gas volume at time step, i. For liquid phase, using fluid bulk modulus, k f,
(6)

where VLi=liquid volume at time step, i.
Page 291
2.4 Mechanical coupling
Of three forms of coupling between the fluid and the solid particles shown in section 2.1, the first item has already been covered by Equations (2) and (3) in Section
2.2. The second item is taken into account in the formulation in Section 2.3. The calculation method for the change in domain volume is as follows.
Domain volume, can be defined as product of apparent domain volume and porosity of assembles. Apparent domain volume is volume of a closed loop of the
centers of the adjacent particles and is given by
(7)
where w=unit width; (xk , yk ) is a coordinate of center of the adjacent particles that forms a closed loop. Thus, the change in domain volume at time step, i is calculated

from
Vdi+1−Vdi=md (Vdi+1'−Vdi')
(8)

using dimensionless multiplier, md .
For third item, to simplify the problem of applying pressure forces to particles, assume that the pressuredrop due to flow through a pipe is localized at the
corresponding contact. Then, the pressure in the region of the domain is uniform, and the tractions are independent of the path around the domain. If we choose the
polygonal path that joins the contacts that surround a domain, the force vector on a typical particle is
Fk =Pnk s
(9)

where nk =the unit normal vector of the line joining the two contact points on the particle, k; and s=the length of the line.
2.5 Solution scheme
Using an explicit solution scheme, the calculation alternates between applying the flow equation to all pipes and applying the pressure equation to all domains as
following procedure.
1. Calculation for flow rate through the channel between the 2 adjacent domains from their pressure difference.
2. Calculation for domain pressure from change in fluid volume in the domain.
3. Calculation for mechanical behaviors of particles, which body force generated by domain pressure is applied to.
4. Calculation for change in domain pressure due to change in domain volume.
It is assumed that the flow rates of gas and liquid, qG and qL, in the above calculation (1) can be obtained from the following equations, respectively.
(10)
(11)
where VG, VL and V=gas volume, liquid volume and domain volume; ν G and ν L=gas viscosity and liquid viscosity; g=gravity vector; and nC=the unit normal vector of

the channel. |source denotes the domain on the inlet side.
The calculation (2) can be applied only after the fluid volume governed by the calculation (1) is completely moved between domains. Therefore, it requires not only
taking the flux between domains but also the actual fluid velocity into consideration in order to perform the significant calculations, and then the increment of time step,
∆t should be set to the specific value that is given by
(12)
where lC=the channel length; and ν C=fluid velocity through the channel. However, it cannot be set a certain value that is strictly effective for all the channels in the
particles, because the channel length and the fluid velocity differ by each channel. Thus, the following equation is assumed in this study.
(13)

where E[*] denotes the expectation operator. E[lC] and E[ν C] are obtained from the following approximations;
E[lC]=lM
(14)
(15)

where lM=mean channel length; aM=mean channel aperture; and Φ=the macroscopic hydraulic gradient, initially observed in the analyzed region.
Page 292
3 VERIFICATION OF THE PROPOSED METHOD
3.1 Outline of verification
In this section, a case study concerning an actual hydromechanical phenomenon around the rock cavern with high internal gas pressure is presented to verify the
practicality of the proposed hydromechanical coupled formulation.
The case discussed in this study concerns the aircushion surge chamber (ACSC) in Kvilldal power plant, the largest Norwegian hydropumped power plant.
ACSC is an economical and operational measure for replacing the conventional surge shaft. Because of this replacement, the position of the powerhouse can be
chosen with greater freedom, and waterway tunnels, especially the expensive ones such as steellined valley crossings can be replaced by generally much less costly
and shorter deepseated headrace/ pressure tunnels.
Similarly to the underground gas storage cavern, the air storage design of the unlined ACSC relies on the principle of water pressure in surrounding rock being more
than the ‘escape’ pressure of the stored air, and if the sufficient water pressure cannot be secured, a series of overhead water curtain gallery is excavated in order to
maintain air storage pressure.
The ACSC in Kvilldal has experienced the air leakage and the countermeasure by the appropriate water curtain facility. Therefore, it is convenient to verify the
performance of the proposed method. The both the air leakage and the air sealing by water curtain are simulated using the proposed method, and the simulation results
are compared with the actual field data.
3.2 The Kvilldal ACSC investigation data
The location of the Kvilldal ACSC is shown in Figure 2. The host rock is gray to dark gray, hard amphibolitic gneiss of Precambrian age with pockets of quartz and
feldspar (mimetic gneiss).
Testing of cylindrical specimens of the cores showed (Kjorholt et al. 1989) 160 MPa for uniaxial compressive strength, σc; 15 MPa for tensile strength, σt; 35 GPa
for Young’s modulus, E; and 0.11 for Poisson’s ratio, v.
The volumetric joint count (joints/m3) was described by Broch et al. (1987) as “low to moderate” for Kvilldal. A total of up to 4 joint sets were registered when
mapping the ACSC walls and arch, but the jointing was spars.
Rock mass permeability measurements were carried out using exploratory boreholes in cavern vicinity, and showed 2×10−16m2 for the average permeability, K.
Figure 2. Layout of the Kvilldal ACSC, Kjorholt et al. (1989).
Figure 3. Dimension of the Kvilldal ACSC, Barton et al. (1987).
Page 293

Hydraulic fracturing stress measurements were also carried out (HFRAC performed by NGI), and showed 5.2 MPa for horizontal stress, σHmin.
The Kvilldal ACSC as shown in Figure 3, measures 16 m for span, B; 17–24 m for height, Z; 386 m for length, L; 260–370 m2 for vertical sectional area, AV;
5200 m2 for water surface area, AH; 110,000 m3 for cavern volume, Vt; and 70–80,000 m3 for air volume, Va .
No reinforcement or grouting was performed in this excellent quality, low permeability rock, although grouting was suggested prior to operations.
3.3 Air losses at the Kvilldal ACSC
At design stage of the Kvilldal ACSC, favorable high water pressure had been expected to prevent the gas leakage since the rocks around the ACSC are extremely
good and low permeable. However, when the highpressure air compression was carried out in September 1982, the impermissible air losses from the ACSC were
observed, and then the compressors were stopped. The air loss rate calculated to be 250 Nm3/hr. The air leakage rate into the surrounding rocks is estimated 240
Nm3/hr.
Figure 4 shows the temporal variation of PV product of the Kvilldal ACSC. The steady decline in the PV product is shown in Figure 4(a) for a 25day period in late
September through to mid October 1982. On the other hand, the stable result is shown in Figure 4(b)
Figure 4. The temporal variation of PV of the Kvilldal ACSC, Pleym & Stokkebo (1985).
for a 25day period in November through to December 1983, when the water curtain that was established as a countermeasure for the air losses, was fully worked.
The air loss rate was totaling 10 Nm3/hr, an amount equivalent only to air absorption, dissolution, or diffusion into the waterbed.
At a later data, piezometers placed within the ACSC to try to understand the leakage problem registered only 1.5–1.7 MPa water pressure. The Bjorndal zone,
which dips at about 45 degree above the ACSC and intersects the water tunnel of the plant, was suspected to have caused the local drainage and the drawdown of the
water table.
3.4 PFC modeling
As shown in Figure 2, the Kvilldal ACSC has the form of a rectangular loop. In this study the left side of the section B–B shown in Figure 3(a) is selected as the typical
twodimensional vertical section for the simulation as shown in Figure 5. The cavern with a span of 16 m and height of 24 m, which has a waterbed with a depth of 10
m, is established in a 100 m×100 m square area.
As the mechanical boundary condition, the measured horizontal stress of 5.2 MPa and the calculated vertical stress of 13.7 MPa from the overburden of 522 m and
the density of 2630 kg/m3 are applied to the bonded particle models, and then the ACSC cavern is excavated by removing the particles within the region of the cavern.
Until the change in domain volume becomes sufficiently small, only the mechanical calculation is carried out with a fixed domain pressure in order to prevent drastic
change in domain geometry, which causes the unstable solution.
Figure 5. The bonded particles model and the boundary condition for the simulation of the Kvilldal ACSC.
Page 294
The right side of the analyzed region is defined as noflow boundary (impermeable boundary) and the other 3 sides are defined specified head boundaries, assuming
the axisymmetric flow. The both the portion of the air in the cavern and the waterbed are the domains with the specified initial pressure.
The verification must be required to use relevant input microproperties, which is set to reproduce the relevant behaviors of the rocks. The appropriate
microproperties are determined by a calibration process, in which the response of the synthetic material is compared directly with the measured response of the rocks.
A series of biaxial tests using the 5 contactbonded particles model (particle radius, r=0.6–1.0 m: uniform distribution; density, ρ=2630 kg/m3) with a height of 100
m and width of 50 m, each of which has the same microproperties and the different layout of particles generated by different random number sequence, are carried out
as the calibration for the determination of the mechanical microproperties. In order to determine the contactbond strength, each of the 4 different confining pressures (0
MPa, 5 MPa, 10 MPa, and 15 MPa) is applied to the bonded particles, and the strength values are compared with the MohrCoulomb envelope plotted based on the
uniaxial compressive strength value and tensile strength value described in 3.2.
As the result of the calibration, the mechanical microproperties can be obtained as shown in Table 1. Table 2 shows the comparison of macroresponse related to
deformability between the synthetic material that has microproperties shown in Table 1 and the
Table 1. Mechanical microproperties setting for the simulation of the Kvilldal ACSC.
Microproperties Notation Unit Value

Ballball contact modulus EC GPa 43
Ball stiffness ratio k n/ks 1.2
Ball friction coeff icient μ 0.5
Contactbond normal strength σC MPa 137
Contactbond shear strength τ C MPa 137
Table 2. Comparison of macroresponse related to deformability between the synthetic material and the actual rock.
Specimen Young’s modulus [GPa] Poisson’s ratio

Model1 36.5 0.115
Model2 36.7 0.107
Model3 37.0 0.117
Model4 36.4 0.124
Model5 37.4 0.110
Model (Average) 36.8 0.115
Measured 35 0.11
actual rock. Similarly, Figure 6 shows the comparison of macroresponse related to strength between the synthetic material and the actual rock.
On the other hand, hydraulic macro response values, which are required with the channeldomain network analysis model, are porosity and permeability. The
porosity data is absent in the literature, therefore, a value of 0.1%, which is typical for hard granitic rocks, is assumed. Two dimensionless multipliers used in equations
(3) and (8), ma and md are set to 1×10−3 corresponding to the assumed porosity value. The normal force for calibration in equation (2), F0 of 1×10−12 is selected
such that the channel aperture is rather stable irrespective of magnitude of the applied stress value, as taking the extremely good rock mechanical properties into
consideration. The residual channel aperture, a0, which denotes the channel aperture under zero stress, is estimated 1.58×10−5 m by equations (1), (2) and (3) based
on the measured permeability of 2×10−16 m2. The abovementioned hydraulic microproperties are summarized in Table 3.
3.5 Simulation results on the case without water curtain
Figures 7 and 8 show a series of the simulation results on the case without water curtain, which the
Figure 6. Comparison of macroresponse related to strength between the synthetic material and the actual rock.
Table 3. Hydraulic microproperties setting for the simulation of the Kvilldal ACSC.
Microproperties Notation Unit Value

Multiplier for domain volume md 1×10−3

Residual channel aperture a 0 m 1.58 ×10−5

Normal force for calibration F0 N 1×1012

Multiplier for channel aperture ma 1×10−3

Page 295
Figure 7. The computed head distribution in the case that the apparent air leakage into the surrounding rocks was observed.
Figure 8. The computed water saturation distribution in the air leakage case (without the water curtain).
Page 296
Figure 9. Comparison between the computed PV and the measured PV in the air leakage case (without water curtain).
apparent air leakage into the surrounding rock mass was observed in 1982. Figure 7(a), (b) and (c) show the computed head distributions of fluid on September 1,
1982 (initial state), September 20, 1982 (20days later), and December 10, 1982 (100days later), respectively, while Figure 8(a), (b) and (c) similarly show the
computed water saturation distributions.
The internal pressure of the cavern is transmitted to the surrounding rock mass, and then the air is penetrated sequentially as time elapsed.
Figure 9 shows the comparison between the computed and the measured PV product as time elapsed. It is clear that the computed PV gives very good agreement
with the measured PV. The average of the computed air leakage rate into the surrounding rock mass for a 25 day period in late September through to mid October
1982 is estimated 245 Nm3/hr, which is very close to the measured air leakage rate of 240 Nm3/hr.
3.6 Simulation results on the case with water curtain
Figure 10 shows a series of the simulation results on the case with water curtain, which the apparent air leakage into the surrounding rock mass was observed in 1982.
Figure 10(a), (b) and (c) show the computed head distributions of fluid on May 1, 1983 (initial state), June 1, 1983 (1 month later), and January 1, 1984 (8 months
later), respectively.
The water pressure, Which is supplied from the water curtain, is transmitted to the underneath. When 8 months have passed since water curtain was operated, higher
head values than the internal pressure 4.1 MPa can be observed in the surrounding rocks.
Figure 11 shows the comparison between the computed and the measured PV product as time elapsed. It is clear that the computed PV gives good agreement with
the measured PV as well as the case without water curtain. The average of the computed air leakage rate is estimated 3.2 Nm3/hr, a sufficient small value as well as the
measured value.
Figure 10. The computed head distribution in the case that the water curtain is established.
Page 297
Figure 11. Comparison between the computed PV and the measured PV in the countermeasure case (with water curtain).
4 A TRIAL OF PLANNING AND DESIGN FOR LPG STORAGE CAVERN
As the results of the above simulations, it is clear that the actual behaviors of the air leakage from the rock cavern with high internal gas pressure and the air sealing by
means of the water curtain can be precisely predicted by the proposed hydromechanical coupled formulation, which can treat the gasliquid twophase flow, and its
practicality is verified.
In this section, a trial of the planning and design of LPG gas storage cavern is conducted with more probable parameters.
4.1 Design parameters for LPG storage cavern
The principal design parameters for LPG gas storage cavern are concerned with storage gas pressure, specifications of cavern, presumable deepest groundwater table,
and specifications of the water curtain. This paper mainly discusses the specifications of the water curtain, with fixed specifications of the cavern, which are
predetermined by the preliminary simulations, under the assumed rock conditions.
The principal design parameters for water curtain are distance above the cavern, length of the borehole, and longitudinal spacing of the borehole. The critical value of
each of those parameters is investigated under the condition that the values of any other parameters are fixed as shown in Table 4.
4.2 PFC modeling
The standard section and the longitudinal section are modeled as shown in Figure 12. The cavern with a span of 18 m, height of 22 m, and length of 60 m, which has a
waterbed with a depth of 5 m, is established in the analyzed region. The length of cavern in the longitudinal section, 60 m is selected for the interest of computation time,
although it is more or less shorter than the value in the ordinary case.
Assuming the axisymmetric flow, the right side of the analyzed region in the standard section is defined
Table 4. Design parameters setting for the simulation of the LPG storage cavern.
Design parameter Value (m)
Specifications of the cavern
Height of the cavern 22
Span of the cavern 18
Length of the cavern 60
Depth of the cavern EL120
Spacing of the cavern 60
Specifications of the water curtain (standard)
Distance above the cavern 25
Length of the borehole 60
Longitudinal spacing of the borehole 10
Figure 12. The bonded particles model and the boundary condition for the simulation of the LPG storage cavern.
as noflow boundary and the other 3 sides, specified head boundaries. In the case of longitudinal section, all the sides are specified head boundaries.
The values of rock properties are selected assuming the massive hard rock as shown in Table 5. For the rock properties, which are undefined in Table 5, the values
used in the Kvilldal case are selected.
Page 298
Table 5. Rock parameters setting for the simulation of the LPG storage cavern.
Rock parameter Value
Rock properties
Young’s modulus 8 GPa
Poisson’s ratio 0.25
Hydraulic conductivity
3.5×10−8 m/s
Rock condition
Vertical rock stress 4.1 MPa
Horizontal rock stress 4.5 MPa
Presumable deepest groundwater table EL30 m
Figure 13. The simulation results with different distance values above the cavern in the standard section.
4.3 Simulation results on the design parameters ofthe water curtain
Figures 13–16 show a series of simulation results on the specifications of the water curtain. Figures 13 and
Figure 14. The simulation results with different distance values above the cavern in the longitudinal section.
14 show the simulation results with different distance values above the cavern in the standard section and in the longitudinal section, respectively. Both results indicate
the critical distance above the cavern=25 m in the case with borehole length=60 m and borehole spacing=10 m.
Similarly, Figures 15 and 16 show the simulation results with different borehole length values in the standard section, and with different borehole spacing values in the
longitudinal section, respectively. Those show the critical borehole length=60 m and the critical borehole spacing=10 m.
Such systematic parametric study allows us to find the more reasonable design parameters of the LPG storage cavern, explicitly considering the hydromechanical
coupled phenomenon, which the conventional continuum analytical method cannot treat due to lack of degreeoffreedom (less elements) to reproduce realistic
predominant pathways that propagate in the specific direction.
5 CONCLUSION
In this study, hydromechanical coupled analysis based on twodimensional bonded particles model that explicitly simulate interparticle mechanics is examined in order
to perform a realistic simulation that can express rock deformation and fracturing coupled with fluid flow of gas and liquid. Then the appropriateness
Page 299
Figure 15. The simulation results with different borehole length values in the standard section.
of this numerical method is verified through the case study of the actual ACSC (Air Cushion Surge Chamber) site that had an experience of not only leakage of high
pressured air from the rock cavern but also air leakagecontrol with water curtain. Finally a planning and designing of the actual LPG storage cavern is conducted, and
the design procedure for the determination of rational design parameters is examined.
Figure 16. The simulation results with different borehole spacing values in the longitudinal section.
REFERENCES
Barton, N. et al. 1987. Rock mechanics investigations for unlined pressure tunnels and air cushion surge chambers. Proc. Int. Conf. On hydro power in Oslo, pp.641–
656.
Cundall, P.A. & Strack, O.D.L. 1979. A Discrete Numerical Model for Granular Assemblies,” Géotechnique, 29: 47–65.
Kjorholt, H. et al. 1989. Water curtain performance at the Kvilldal air cushion. Proc. Int. Conf. On storage of gases in rock caverns, pp. 26–28.
Pleym, A.S. & Stokkebo, O. 1985. Operational experience with the air cushion surge chamber at Kvilldal power station. Norwegian hydropower tunneling. Norwegian
Soil and Rock Engineering Association, No. 3.
Potyondy, D.O. & Cundall, P.A. 1999. Modeling of Notch Formation in the URL MineBy Tunnel: Phase IV—Enhancements to the PFC Model of Rock, Itasca
Consulting Group, Inc., Report to Atomic Energy of Canada Limited, Ontario Hydro Nuclear Waste Management Division Report No. 06819REP01200–10002R00.
Page 300
Page 301
DEMfluid model development for nearwellbore mechanics
B.K.Cook, D.F.Boutt & O.E.Strack
Sandia National Laboratories, Albuquerque, New Mexico, USA
J.R.Williams & S.M.Johnson
5809 679 3
ABSTRACT: Numerous important natural and industrial mechanical processes, such as sand production in the petroleum industry, are dominated by
the interaction of fluid and solids. This paper summarizes results from a joint MITSandia research effort to develop two and threedimensional
coupled models for the direct simulation of particlefluid systems, with a focus on nearwellbore applications of interest to the petroleum industry. Our
model resolves the interaction of individual solid particles (and cemented particle assemblies) with other solid particles and the surrounding fluid. The
numerical formulation of the model couples the discreteelement and latticeBoltzmann methods to achieve linear scaling in computational effort with
the number of simulated particles. The numerical methods have been implemented into a userfriendly modeling application to facilitate the simulation of
a wide variety of problems. Model results have been validated through comparison with several analytical and benchmark simulation problems,
including the drafting, kissing, tumbling interaction of two spheres settling in a fluid column. Applications include fluidinduced sand production from an
idealized wellbore perforation.
1 INTRODUCTION
Numerous important natural and industrial mechanical processes are dominated by the interaction of fluid and solids. For example, the petroleum industry is plagued by
the erosion of sand during oil production induced by elevated fluid and in situ stresses; the mitigation of this sanding problem alone costs the industry in excess of $1
billion annually. Existing continuumbased models are poor predictors of these phenomena as they do not resolve the critical underlying micromechanical interactions
between the fluid, saturated solid matrix, and individual solid particles. Meaningful experimental studies are difficult to perform because of system complexity and the
associated large parameter space. We believe that rapid increases in computational resources combined with the development of efficient discrete modeling schemes
will allow micromechanicsbased approaches to be successful in elucidating the dominant physical mechanisms and processes in many particlefluid systems.
The objective of this joint MITSandia project has been to develop two and threedimensional coupled models for the direct and highfidelity simulation of particle
fluid systems, with a focus on application to nearwellbore mechanics. The models approximate the interaction of individual solid particles (and cemented particle
assemblies) with other solid particles and the surrounding fluid. Our research efforts have combined the formulation of new discrete numerical methods with their
implementation into generalized twoand threedimensional simulators. These models were developed for use by researchers in petroleum engineering and other areas
to simulate physical processes that were heretofore unobservable with traditional modeling approaches. We have also begun to apply the models to investigate a suite
of petroleumspecific problems, such as sanding, which influence the well completion and production techniques employed by that industry.
This paper reviews our modeling approach and provides some representative examples of validation and application problems that have been performed with the
code. Our twodimensional code, SandFlow2D, is an extension of an earlier discrete element code known as MIMES, which was developed at the Massachusetts
Institute of Technology (Rege 1996) with support from Sandia National Laboratories. More recently, we have been working on a threedimensional code called
SandFlow3D. The numerical formulation of these models is based on the coupling of the discreteelement method (DEM) and the latticeBoltzmann method (LBM).
We sketch highlights of the numerical method, and then present a twodimensional sanding simulation to demonstrate the largescale application of the coupled model.
We conclude with a summary
Page 302
of our current implementation efforts on SandFlow3D, outlining the architecture of the code and presenting a sample simulation of the drafting, kissing, tumbling
interaction of two settling solid spheres in a fluid column.
2 COUPLED NUMERICAL METHOD
This section provides an overview of the coupled DEMLBM approach, which has been extensively documented in previous publications (Cook et al. 2000, Cook
2001, Cook et al. 2004a). DEM is a suite of numerical algorithms that allows for the efficient modeling of granular materials as a collection of arbitrarily shaped, rigid
bodies that interact through surface contacts and bonds. A collection of recent DEM papers can be found in Cook & Jensen (2002). LBM is a naturally parallel
numerical method for simulating the incompressible Navier Stokes equations (Chen & Doolen 1998), which govern fluid flow in the problems of interest. Forcing
mechanisms represented to at least the first order in the coupled model include fluidinduced forces on particles, intergranular bonding between particles, intergranular
forces from particle collisions and static formation stresses, and buoyancy forces. Given the expected familiarity of this readership with DEM, we will focus the
remainder of our discussion in this section on LBM and the coupling scheme.
Fluid flow at the particle scale can be fully resolved by LBM; consequently, in addition to the in situ formation stresses, the coupled DEMLBM method accurately
reproduces the local fluidinduced forces that can drive the disaggregation and transport of solids from a solid matrix (e.g., piping erosion in porous media). This is an
important benefit of direct simulation techniques like ours over other coarsergrained fluidDEM coupling techniques, such as DEM coupling to Darcy’s (Klosek 1997,
Jensen et al. 1999) and to pipe flow networks (Bruno 1994). These coarsegrained formulations are typically limited to quasistatic simulations of dense solid systems
(i.e. porous media) and consequently cannot be applied to many problems of interest that are characterized by either a sparse distribution of solids or by energetic (i.e.
inertial) flow where the Reynolds number is not small. Furthermore, because the coarsegrained techniques like Darcy’s generally only resolve the normal pressure
component of the fluid upon the solids and do not capture the shear component associated with viscous flows, they do not accurately model the detailed force and
associated torque that is generated by the fluid at the particle scale. The limitations of the coarsegrained coupling approaches are unacceptable for many problems,
such as sanding where the flow regime ranges from creeping in the far field to inertial near the wellbore, while particle concentrations are dense in the farfield,
undisturbed porous media and relatively dilute in the wellbore.
2.1 LatticeBoltzmann method
In the latticeBoltzmann method, the fluid domain is subdivided into a uniformspaced, regular computational grid called a lattice. At each computational node in the
lattice, the fluid is represented discretely as fluid particle packets. These fluid packets have a discrete set of velocities such that packets will move or stream from one
node to its neighboring nodes in a time step. At each node fluid packets collide and are redistributed according to specific relationships that conserve mass and
momentum and recover the NavierStokes equation. Formally, the linearized method (Chen & Doolen 1998) used in our implementation is written as
(1)
where f i(x,t) is the fluid particle density distribution with velocity e i at a node located at position x for a given time t; τ is the relaxation time; and is the

equilibrium density distribution for the fluid. The particle distribution f i can be thought of as the probable amount of fluid mass moving in the direction i (Noble &
Torczynski 1998). Equation 1 comprises a collision phase (shown in the second line), and a streaming phase that moves the distribution calculated on the righthand
side to the neighboring node indicated on the left, x+e i∆t. The macroscopic fluid variables at a node, density ρ and velocity v, are found from the moments of the fluid
particle density distribution:
(2)
Our twodimensional code uses a 9velocity lattice referred to in the literature as the 2DQ9 (Cook et al. 2000, Cook 2001, Cook et al. 2004a), while our three
dimensional code uses a 19velocity lattice known as the 3DQ19 (Strack & Cook in preparation). The 2DQ9 model is based upon the regular discretization of the
fluid domain into square cells; at the center of each cell is located a lattice Boltzmann node. At each node there are eight density distributions with the nonzero
velocities shown in Figure 1. There is also a zerovelocity or rest distribution at each node, f 0. The equilibrium distribution used in our twodimensional code is given by
Gallivan et al. (1997).
(3)
Page 303
Figure 1. Closeup of twodimensional lattice with nine nodes shown as filled squares. Velocity vectors, e i, for center node are also
shown as labeled. The area extent of the lattice cell associated with the center node is shown in gray.
for i=1, 3, 5, and 7
for i=2, 4, 6, and 8
As can be seen from an inspection of Equations 1 and 3, the majority of computational cost for each timestep update is associated with the calculation of the
equilibrium distribution: a local computation that requires no additional information from neighboring nodes. Data is only exchanged between lattice nodes during the
streaming of the distributions. These algorithmic characteristics of LBM result in the linear scaling of computational effort with the number of lattice nodes and enable
efficient parallel implementations.
The interaction between solids, which are represented by discrete elements, and fluid is resolved with a moving boundary condition that enforces the noslip
boundary condition along solidfluid interfaces. We use an immersedboundary scheme proposed by Noble & Torczynski (1998). In their method the fluid domain is
continuous, overlapping the solid particles suspended in the fluid. Unlike other immersedboundary conditions (Ladd 1994, Qi 1999) that use a discretized
representation of the solid computed from the lattice cells covered by the solid, Noble & Torczynski’s approach employs a modified collision operator in the lattice
Boltzmann equation that accounts for the varying amount of fluid mass scattered by solids partially or fully covering a lattice cell. For solidcovered lattice cells,
Equation 1 becomes
(4)
where B is a weighting function that depends on the cell solid fraction ε, defined here as the fraction of the lattice cell area covered by a solid(s), and the dimensionless
relaxation time as
(5)
where τ* is the dimensionless relaxation time equal to τ/∆t. ΩS is an additional collision term that bounces back the nonequilibrium part of the distribution and is given
by
(6)
where vs is the solid’s velocity at time t and −i is used to denote the distribution component having the opposite direction to i (e.g. f 5 for f 1, see Fig. 1). A modification

to Equation 6 was proposed by Holdych (2003) to use only the equilibrium distribution computed from the solid velocity as
(7)
This modification reduces the computational effort required to enforce the noslip condition by eliminating the calculation of the nodal equilibrium distribution at solid
covered cells. We have recently implemented this change with mixed results: it improves the accuracy of some simulations in our threedimensional code (Strack &
Cook in preparation), but also was found to introduce numerical instability in our twodimensional code. Additional work is underway to assess these issues.
Equation 4 modifies the density distributions at the nodes covered by a solid, effectively setting the fluid velocity to match the solid’s rigid body motion. The non
equilibrium part of the distribution calculated in Equations 6 and 7 represents a momentum imbalance that contributes to the fluidinduced force on a solid particle. The
total force of the fluid on a particle can be found by summing up the momentum transfer that occurs over the n nodes covered by the solid as
(8)
where h is the lattice node spacing. Fluidinduced torque can be easily found through a similar computation,
(9)

where xs is the centroid of the solid particle at time t.
Page 304
Figure 2. Model coupling between DEM and LBM computational modules. The method is an explicit, staggered coupling approach
whereby the DEM and LBM solvers are updated sequentially at each timestep based on the previous results of the other
solver.
Fluidparticle coupling is realized through the alternating update of particle positions with the DEM solver and the fluid state with the LBM solver as shown in Figure 2.
We should note that since new spatial sorting schemes allow the discrete element method to scale linearly with the number of simulated particles (Williams et al. 2004),
the coupled DEMLBM method can also scale linearly in the number of discrete elements with a constant of proportionality equal to the number of lattice nodes
associated with a discrete element.
3 VALIDATION AND APPLICATION
SandFlow2D was originally validated through a series of fundamental particlescale simulations including comparisons to the analytical solution for cylindrical Coutte
flow (Cook 2001, Cook et al. 2004a) and a number of particle sedimentation problems such as drafting, kissing, tumbling (Cook et al. 2000, Cook 2001). It was
subsequently applied by Boutt (2004) through a series of simulations at the specimen scale with O(1000) discrete elements to test its ability to reproduce poroelastic
phenomena such as Terzaghi’s consolidation problem (Boutt et al. 2003, Boutt et al. in review) and naturally induced hydraulic fracturing. In this section we present
results from a recent largescale simulation of sanding to illustrate the flexibility of the coupled model to analyze a widerange of particlefluid problems.
3.1 Simulation of sanding
The primary application focus of the code has been in the area of nearwellbore mechanics, including DEM simulations of borehole breakout and hydraulic fracturing
(Cook et al. 2004b), and more recently the DEMLBM simulation of sand production. Sanding occurs during the production of hydrocarbons
Figure 3. Visualization of sanding simulation with approximately 40,000 elements showing flow of fluid and discrete elements (colored by
speed) through opening at right. The square elements along the right boundary are fixed and represent the well casing, while
the gap represents a casing perforation. The other simulation boundaries are outside of the area shown in the image. The large
mushroomshaped region indicates elements fluidized by high drag forces.
from unconsolidated or poorly consolidated formations such as sandstones. Often, because of the uncertainty as to whether a given well will produce sand, the industry
uses unnecessary and expensive sand control measures like gravel pack filters. To better predict the need for and to design these control measures, the industry has
sought an improved understanding of the factors leading to the onset, subsequent rate, and occasional stabilization of sanding. The DEMLBM method is appealing
because it should capture the discontinuous failure and stabilization processes that are difficult to isolate in laboratory experiments and that are not resolved by
continuum methods.
In collaboration with a petroleum industry consortium, we have begun applying SandFlow2D to determine whether a discrete simulation approach can provide an
improved understanding of the parameters influencing sanding. In particular, we are investigating the role of microstructural properties, such as particle shape, in sanding
and stabilization mechanisms. The simulations are based on the twodimensional idealization of a vertical crosssection through an oilbearing formation adjacent to the
wellbore; the wellbore casing is perforated by an opening to allow flow. The simulated region is of O(10 cm) in length with either unbonded or bonded discrete
elements of O(1 mm) in diameter used to represent a unconsolidated sand or loosely consolidated sandstone. A confining stress is
Page 305
Figure 4. Closeup of perforation region showing the formation of a particle arch that stabilizes the discrete element matrix and allows it
to temporarily resist fluidinduced forcing. Compared to the model shown in Figure 3, this simulation has more elliptical
elements and a smaller perforation opening relative to the average element size. The arch collapses later in the simulation
resulting in episodic sanding much like that observed in the field.
applied at the top and bottom of the simulation domain to simulate the overburden stress found in field conditions. The fluid has the properties of oil and is forced across
the simulation domain to flow through the perforation; depending on the balance of the stabilizing interparticle contact forces and the destabilizing forcing from the fluid,
the discrete elements near the perforation may erode and be transported out of the formation into the wellbore. Preliminary results were described in Boutt et al.
(2003). Figures 3 & 4 show some recent results from this study. Sanding through the perforation opening is driven by a fixed pressure drop between the wellbore (right
side of model in Figs 3 & 4) and farfield boundary (left side of model).
4 3D MODEL DEVELOPMENT
The validation and application results obtained with the twodimensional code for idealized problems have been encouraging. However, most discrete systems of
practical interest are threedimensional, and it is unclear whether twodimensional models of such systems are appropriate. To address this issue, a recent focus of our
research program at Sandia has been to develop a generalized threedimensional modeling application, SandFlow3D, based on the coupled DEMLBM method
outlined earlier in this paper. The beta version of SandFlow3D implements spherical discrete elements coupled to the 3DQ19 LBM model; the LBM implementation
and moving boundary validation is described in Strack & Cook (in preparation). A related thrust of the MITSandia research program has been to develop improved
nonspherical element representations, which has led to the formulation of a pseudoellipsoid element by Johnson et al. (2004). In addition to the algorithmic and
computational challenges associated with a threedimensional code development, from a software engineering perspective, the appropriate application framework for a
useful threedimensional simulation tool is not well established. The next section presents the software design considerations and resulting architecture for SandFlow3D.
4.1 Software architecture
Based in part on feedback from beta testers of SandFlow2D and a general desire to facilitate the simulation and analysis of a wide variety of particlefluid problems, the
architecture of SandFlow3D was developed to satisfy several software design guidelines, including easeofuse, extensibility, high performance, platform independence,
and user programmability. For easeofuse, we wanted to create a modeling environment like MIMES (Rege 1996) that could serve as a virtual laboratory in which
numerical experiments can be easily and seamlessly defined, run, and visually analyzed. In keeping with this goal, the target platform for the application prototype was a
workstationclass PC. As a code that will inevitably evolve in time, extensibility is a necessary architectural property to expedite future code changes, such as the
addition of new discrete element representations. From a computational standpoint, some degradation in performance can be expected in a generalized implementation;
however, we wanted SandFlow3D to be competitive with optimized, “hardwired” implementations of the underlying numerical methods. With the emergence of Linux
as a scientific computing platform and the continued popularity of Windows, we also wanted a platformindependent application that could be built from the same or
slightly modified source code. Finally, we have found that the efficient execution of simulation tasks such as model generation requires user programmability, whereby
the application is controllable through a set of applicationspecific commands that can be executed individually or collectively through scripts.
The resulting architecture strives for balance between the application requirements: the need for extensibility suggested an objectoriented design; performance
demands required a compiled objectoriented language (C++); easeofuse necessitated a graphical user interface (GUI) to the application kernel; the desire for
platformindependence precluded
Page 306
Figure 5. SandFlow3D architecture.
tying the GUI directly to a specific windowing system (e.g. X or Windows); and the need for programmability suggested leveraging a scripting language to create an
application command language. This design approach was inspired in part by the MIMES architecture described in Rege (1996).
The resulting application architecture for SandFlow3D is presented in Figure 5. The code consists of (1) an objectoriented kernel programmed in C++; (2) a set of
Tcl scripts, which build the Tk user interface and console and also define a number of applicationspecific Tcl procedures; (3) a visualization window implemented with
the VTK class library; and (4) the VTK and Tcl/Tk dynamically linked libraries. The code is currently single threaded; however, performance considerations will likely
necessitate multithreading, which is easily accommodated by its modular architecture.
Users interact with the application through either a GUI, which has an integrated script editor as well as a limited menubased set of model geometry and parameter
control options, or through a console, which supports direct execution of application commands. Both the GUI and console are script built from the platform
independent Tk widget set. An interactive VTK window provides the user with a threedimensional visualization of the current simulation state. Commands issued
through the GUI and console are parsed by the embedded Tcl interpreter, which executes builtin Tcl commands directly or calls the application kernel for execution of
applicationdefined commands. The kernel in turn either passes back data through the interpreter for display in the Tk user interface, or passes the results directly
through memory to the VTK visualization module. The Tcl/Tk and VTK components are described further in the sections below.
4.1.1 Tcl/Tk
Tcl/Tk is a software package developed by John Ousterhout that provides a comprehensive programming system for developing and controlling graphical user interface
applications. It has been opensourced and is freely available online for a variety of common operating systems (www.scriptics.com). By itself, Tcl offers a powerful
scripting language, which is parsed and executed by the Tcl interpreter. Tk extends Tcl with a suite of platformindependent widgets that can be scripted to create
highly functional GUIs. The Tcl interpreter is embedded into the C++ kernel to provide a scripted or eventdriven (through the Tk GUI) frontend to the application.
The Tcl language can be extended to include both userdefined scripted procedures or compiled C++ commands, which are registered with the interpreter at compile
time. We should note that there are several other programming environments—including Java and Python—that are viable alternatives to Tcl/Tk (see, e.g.,
Komodromos & Williams (2004) for a discussion of a Javabased DEM simulation environment).
4.1.2 VTK
Visualization of DEM and fluid results in three dimensions poses significant coding and computational challenges. Many commercial 3D graphics applications use
OpenGL, a relatively lowlevel, industrystandard graphics application programming interface (API) that incorporates a broad set of 2D and 3D drawing functions. In
SandFlow3D, we wanted to provide the user with an interactive and realtime visualization of the simulation state, which could display both the DEM model along with
the fluid velocity or pressure fields. Because an OpenGL implementation of such a sophisticated scientific visualization interface was beyond the scope of this project,
we turned to a higherlevel visualization API called VTK to streamline the development process.
VTK is an objectbased C++ visualization API layered over OpenGL. VTK is also open source and is available online for most operating systems (www.vtk.org).
VTK provides a large number of powerful C++ classes that can be integrated by a programmer to create a data processing and rendering pipeline. To implement our
visualization display, we use the VTK interactive rendering window along with several VTK classes that render the DEM and fluid data. The threedimensional fluid
velocity and pressure datasets can be interactively sliced to display results along interpolative cutting planes as shown in Figure 6. More advanced visualization options,
such as fluid streamlines, are also available in VTK. The rendering widget supports a wide range of user interactions, including object picking and view manipulation
operations such as zooming, panning, and rotation. SandFlow3D datasets are saved in the VTK file format to allow for additional postprocessing visualization using
the VTKbased ParaView, an open source visualization application available online (www.paraview.org).
Page 307
Figure 6. VTK visualization in SandFlow3D of flow through an array of spherical discrete elements. The fluid cutting plane is shown at
the end of the element array and displays the fluid speed in the pore throats between the spheres.
4.2 Validation
We are currently in the process of validating the 3D implementation of DEMLBM in SandFlow3D. These tests, which include the detailed comparison of model
results against analytical and other numerical solutions for flow around fixed and moving spheres, will be reported in Strack & Cook (in preparation). Here we highlight
the qualitative validation of the code through the simulation of the drafting, kissing, tumbling phenomenon first observed by Joseph et al. (1987), and later simulated
numerically with several codes including SandFlow2D (Cook et al. 2000, Cook 2001).
Drafting, kissing, tumbling results from SandFlow3D are shown in Figure 7. The results are for a particle Reynolds number of O(10). In this coupled particlefluid
problem, two spheres are allowed to settle in an initially quiescent fluid under the influence of gravity. Under inertial conditions, where the particle Reynolds number
exceeds 1, a lowpressure wake develops behind the leading particle that allows the trailing particle to draft behind the lead particle. The drafting particle quickly
accelerates to catch up with, or “kiss”, the lead particle, at which point they effectively form an elongated body, which is forced broadside to the flow by a fluid
induced turning couple. With this motion, the trailing particle “tumbles” by the lead particle, at which point the drafting, kissing, tumbling cycle may begin again. The
numerical simulation of this phenomenon is a good test for the code because it requires the accurate resolution of inertial flow and moving solid interfaces to reproduce
the complex fluidsolid dynamics.
Figure 7. Visualization of drafting, kissing, tumbling simulation performed with SandFlow3D. Top panel from left to right shows a time
sequence of earlytime interactions of the spheres as the trailing sphere drafts the leading sphere and is accelerated
downwards. Lower panel shows a latertime sequence as the formerly trailing sphere tumbles by other sphere to assume the
lead position. Background coloration shows the dynamic fluid pressure.
5 DISCUSSION
We have achieved the coupling of DEM with a NavierStokes fluid solver by using LBM with an immersed moving boundary condition. The coupled method has been
implemented in both two and threedimensions to create two numerical laboratories, SandFlow2D
Page 308
and SandFlow3D, for the exploration of particlefluid physics. Much fundamental work remains, however, to mature DEMfluid modeling techniques into predictive
capabilities with widespread engineering value. In outlining the efforts associated with this research, we have skirted many significant and outstanding research questions
regarding the underlying mathematical modeling and numerical formulation of the coupled codes, as well as the challenges that arise in their application to simulate actual
physical systems. For example, in DEM, the importance of representing the detailed shape of real grains is unclear, as is the more general issue of the characterization
of real systems (e.g., a sandstone specimen) and the appropriate numerical specification of DEM material properties and geometry. On the fluid side, many relevant
problems (including sanding) involve multicomponent mixtures of fluids, such as oil and water, or oil, water, and air, introducing wetting or capillary effects not
obviously resolved by the current scheme. Compressibility may be important, as may be the Reynolds lubrication pressures that develop between adjacent solid grains.
Of course, even if the relevant physics for the problem of interest are correctly represented, the computational timetosolution, the time required to develop, execute,
and analyze a simulation, must be reasonable for a modeling technique to be of engineering value. These issues will likely offer fruitful research topics for the years
ahead.
ACKNOWLEDGEMENTS
Funding for this research was provided by the U.S. Department of Energy through the Natural Gas and Oil Technology Partnership, the Sandia Laboratory Directed
Research Program, and a Sandia CRADA with ChevronTexaco, Halliburton, Schlumberger, and Shell. Sandia is a multiprogram laboratory operated by the Sandia
Corporation, a Lockheed Martin Company, for the U.S. Department of Energy’s National Nuclear Security Administration under Contract DEAC0494AL85000.
REFERENCES
Boutt, D.F. 2004. Discrete analysis of the role of pore fluids in the genesis of opening mode fractures in the shallow crust (Ph.D. thesis). Socorro: New Mexico
Tech.
Boutt, D.F., Cook, B.K., McPherson, B.J.O.L. & Williams, J.R. 2003. Application of a directly coupled numerical model of fluidsolid mechanics. Soil and Rock America
2003. Rotterdam: Balkema.
Boutt, D.F., Cook, B.K., McPherson, B.J.O.L. & Williams, J.R. In review. Direct simulation of fluidsolid mechanics in porous media using the discrete element and
latticeBoltzmann methods.
Bruno, M. 1994. Micromechanics of stressinduced permeability anisotropy and damage in sedimentary rock. Mechanics of Materials 18:31–48.
Chen, S. & Doolen, G. 1998. Lattice Boltzmann method for fluidflows. Annual Review of Fluid Mechanics 30: 329–364.
Cook, B.K. 2001. A Numerical Framework for the Direct Simulation of SolidFluid Systems (Ph.D. thesis). Cambridge: Massachusetts Institute of Technology.
Cook, B.K., Noble, D.R., Preece, D.S. & Williams, J.R. 2000. Direct simulation of particleladen fluids. Pacific Rocks 2000. Ed. Girard, Liebman, Breeds, and Doe.
Rotterdam: Balkema.
Cook, B.K., Noble, D.R. & Williams, J.R. 2004a. A Direct Simulation Method for ParticleFluid Systems. Engineering Computations: International Journal for
ComputerAided Engineering 21(2–4):151–168.
Cook, B.K., Lee, M.Y., DiGiovanni, A.A., Bronowski, D.R., Perkins, E.D. & Williams, J.R. 2004b. Discrete Element Modeling Applied to Laboratory Simulation of Near
Wellbore Mechanics. International Journal of Geomechanics 4(1):19–27.
Cook, B.K. & Jensen, R.P., Eds. 2002. Discrete Element Methods: Numerical Modeling of Discontinua, Proceedings of the Third International Conference on
Discrete Element Methods, Santa Fe, NM. Reston: ASCE.
Gallivan, M., Noble, D., Georgiadis, J. & Buckius, R. 1997. An evaluation of the bounceback boundary condition for lattice Boltzmann simulations. International Journal
for Numerical Methods in Fluids 25:249–263.
Holdych, D. 2003. LatticeBoltzmann methods for diffuse and mobile interfaces (Ph.D. thesis). UrbanaChampagne: University of Illinois.
Jensen, R.P., Preece, D.S., Cook, B.K., Perkins, E.D. & Williams, J.R. 1999. Coupled fluidflow/particlemotion computer modeling of sand production in oil wells, SPE
No. 56630. Society of Petroleum Engineers Annual Conference, Houston, TX.
Johnson, S., Williams, J.R. & Cook, B.K. 2004. Contact detection algorithm for an ellipsoid approximation for discrete element modeling. Engineering Computations:
International Journal for ComputerAided Engineering 21(2–4):215–234.
Joseph, D., Fortes, A.F., Lundgren, T.S. & Singh, P. 1987. Nonlinear mechanics of fluidization of beds of spheres, cylinders, and discs in water. In G.Papanicolau (ed.),
Advances in Multiphase Flow and Related Problems, SIAM.
Klosek, J. 1997. The integration of fluid dynamics with a discreteelement modeling system: Algorithms, implementation, and applications (Masters thesis).
Cambridge: Massachusetts Institute of Technology.
Komodromos, P.I. & Williams, J.R. 2004. Dynamic simulation of multiple deformable bodies using combined discrete and finite element methods. Engineering
Computations: International Journal for ComputerAided Engineering 21(2–4):431–448.
Ladd, A. 1994. Numerical simulations of particulate suspensions via a discretized Boltzmann equation, Part 1. Journal of Fluid Mechanics 271:285–309.
Noble, D.R. & Torczynski, J. 1998. A lattice Boltzmann method for partially saturated computational cells. International Journal of Modern Physics C 9(8): 1189–1201.
Page 309
Qi, D. 1998. LatticeBoltzmann simulations of particles in nonzeroReynoldsnumber flows. Journal of Fluid Mechanics 385:41–62.
Rege, N. 1996. Computational Modeling of Granular Materials (Ph.D. thesis). Cambridge: Massachusetts Institute of Technology.
Strack, O.E. & Cook, B.K. In preparation. An evaluation of an immersed moving boundary condition for the latticeBoltzmann method.
Williams, J.R., Perkins, E.D. & Cook, B.K. 2004. A Contact Algorithm for Partitioning N Arbitrary Sized Objects. Engineering Computations: International Journal
for ComputerAided Engineering 21(2–4):235–248.
Page 310
Page 311
Lab/element tests
Page 312
Page 313
Micromechanical modeling of granular materials under triaxial and oedometric
loading
R.Katzenbach & A.Schmitt
Institute and Laboratory of Geotechnics, Technische Universität Darmstadt, Germany
5809 679 3
ABSTRACT: The efficiency of pile drilling machinery generally results from the interaction between the performance of machinery tools and the
modifications created to the soil characteristics. From the theoretical point of view the process of pile installation can be modeled more accurate by
using the 3D Distinct Element Method (DEM) than by approaches based on the continuum theory. This contribution presents results on
micromechanical particle simulations using PFC3D to assess the range of parameters influencing the stressstrain behavior of an assembly of rigid
particles. The stressstrain response of spherical and nonspherical particles mainly under triaxial but also under oedometric loading in comparison to
laboratory tests is investigated. Based on this experience a preliminary model is presented which is aiming to answer the question on how soil
displacement piles create modifications to the surrounding soil characteristics.
1 INTRODUCTION
The idea of using the DEM for simulating different aspects of pile installation came up during a research project on screw piles in coarsegrained soils with an adjustable
soil displacement during the installation phase. Although the bearing capacities of these piles could be predicted successfully by relying on a continuum approach using
the Finite Element Method (FEM), certain simplifications are indispensable to keep the model as simple as possible and to allow predictions of the loadsettlement
response in a tolerable time frame. However when a more detailed approach is required which accounts for detailed geometric shapes (such as auger stems and auger
starters) and the strong dependency of soil parameters on density changes in coarsegrained soils, classical constitutive models, as implemented in most commercial
FEcodes, rapidly reach their limits or remain unsatisfactory for the intended investigations. This shortage comes together with the often cumbersome troubleshooting
while carrying out large displacement analysis where single elements or complete FEmesh parts are subjected to extremely high shear deformations.
In order to derive a set of simulation parameters for further application, several triaxial and oedometric tests have been backanalyzed within this study. The results
and a model resulting from the experience gained, is presented in the following.
2 NUMERICAL MODEL
2.1 The particle model
The constitutive model utilized within this study to describe the particle interaction is an elastoplastic contact model that is characterized by normal and shear springs at
the particle contacts. The Particle Flow Code, PFC (Itasca 1999), that is used in this study was developed by Cundall (1971) and is based on the DEM but restricted
to modeling rigid spherical particles (Cundall & Strack 1979).
Numerical simulations are therefore based on the following assumptions:
– All particles are treated as undeformable rigid spheres (Fig. 1)
– Contacts only occur at points
– The behavior at contacts is described by a softcontact approach in which the rigid particles are allowed to overlap one another at their contact points (Itasca 1999)
– The magnitude of overlap is related to the contact forces by a force displacement law
– The overlap is small compared to the particle size (Fig. 2)
– All particles are simulated by spheres, overlapping spheres which form a permanent unbreakable clump or by connected (glued) spheres which form a breakable
cluster
Page 314
Figure 1. Simplified particle shape and contact behavior (based on Iwashita & Oda 2003).
Figure 2. Interaction between particles.
For a contact formed e.g. by the particles A and B with the unit normal ni (the suffix i denotes the ith contact), the contact force vector Fi can be resolved into two

parts, one acting in the direction of the normal vector ni (normal component

The normal force is related to the amount of overlap Un by the average secant normal contact stiffness of particle A and B:
(1)
When a contact is formed the total shear contact force is initialized to zero. All subsequent relative sheardisplacements are resulting in an additional increment of elastic
shear force that is added to the current value. From the contact displacement increment vector the shear component over a time step ∆t with the contact velocity
(shear component) is given with:
(2)
The corresponding shear force is related to the tangential shear stiffness by:
(3)
The shear contact force is calculated by summing the existing shear force vector at the beginning of
Figure 3. The calculation cycle of DEM (neglecting damping).
the timestep with the actual shear forceincrement vector:
(4)
In addition to the linear contact model a slip model has been considered. By introducing this model the shear forces transmitted between two contacting entities can be
limited by:
(5)
Slip or plastic deformation occurs when the shear force exceeds, in comparison to the normal force, a certain level, which depends on the dimensionless friction
coefficient μ. Every contact is checked for slip conditions by calculating the maximum allowable shear contact force In case of
(6)
slip will occur during the next calculation cycle and
The equation of motion, which subsequently is used to calculate the new particle positions, consists of two vector equations that relate the resultant force to the
translational motion and the resultant moment to the rotational motion as depicted in Figure 3. Fi is the resultant force as a sum of all externally applied forces acting on
the particle, Mi the corresponding resultant moment; m is the total mass of the particle and gi the body force acceleration vector (gravity loading).
The equations of motion Fi and Mi are integrated using a centered finite difference scheme procedure (Itasca 1999).
3 SAND INVESTIGATED
Several preliminary numerical simulations were performed on triaxial tests with a modified grain size distribution of a sand from Berlin (Fig. 4). In addition
Page 315
Figure 4. Grain size distribution/upscaled distribution (USC).
Figure 5. Type of numerical triaxial testing samples.
more comprehensive numerical simulations were conducted with soil samples corresponding to the grain size distribution of Karlsruhe sand. Similar tests and soil
properties of Karlsruhe sand are reported in Achmus & AbdelRahman (2003). The computational time is governed by the number of particles and the size of the
smallest particle for which the time increment is computed. To reduce simulation time all grain diameters of an existing grain size distribution are usually enlarged for the
numerical simulation by multiplying the actual diameter with a user defined factor (upscaling—USC). For Berlin sand an upscaling factor of 100 was used. A steeper
slope of the grain size distribution of Karlsruhe sand (Fig. 4) and a smaller sample size (Fig. 5) enabled reduction of the upscaling factor to about 40 for Karlsruhe sand.
In addition the original grain size distribution of Karlsruhe sand has slightly been modified by reducing the amount of fine and coarse grains, before applying any
upscaling to further reduce the simulation time.
4 TRIAXIAL TESTING
4.1 Generating specimen
The effect of particle size distribution and particle friction coefficient on the macroscopic behavior,
Table 1. Simulation parameters for triaxial testing (DEM).
Parameter/property Selected value
Number of particles max. 40000
Radii of particles 3–22 mm
Gravity
0.0 m/s2
Porosity of the numerical sample 0.38
Friction coefficient 0.1–10.0
Normal spring constant of particles (kn) 1.0×10(6 to 8) N/m

Stiffness ratio (shear/normal) 0.33–2.0
Normal spring constant of walls
1.0×10(5 to 8) N/m
Shear spring constant of walls 0.0
Friction coef. (spheres/wall) 0.0
especially on the macroscopic friction angle φ, has been investigated by performing triaxial tests on particle assemblies containing spheres as well as clumped particles,
representing a more realistic shape of sand grains. To create a packing of spheres and clumps with a specified porosity, an initial assembly of spheres is created in a test
box with spheres of an artificially low radius. The spheres are generated randomly within a threedimensional rectangular box (Fig. 5) comprising of 6 rigid planar,
frictionless boundary walls. The starting radius of each sphere is calculated by the final radius required for a specific porosity divided by a predefined factor. After the
testing box is filled with an arbitrary placement of the particles, the radius of spheres is expanded back again until the desired density is reached. Every grain is now
supported by the others or by the edges of the domain so that a stable assembly is obtained.
Simulations were performed for an upscaled sample of Berlin sand with a square cross section having a side length of a and a sample height of 2a. The sample
contained only spherical particles. Numerical triaxial tests based on upscaled grains of Karlsruhe sand were carried out with a unique side length of b (Fig. 5),
containing only nonspherical particles. Both models are equilibrated under a predefined friction coefficient and zero gravity.
The properties of the particles are set already at the beginning of the simulation, the frictional properties of the boundary walls are set after the specimengenesis
procedure. In Table 1 a summary of the range of simulation parameters used in the numerical triaxial tests is given.
As most published simulations are restricted to circular/disk shapes in 2D, the choice of parameters is based on an initial triaxial testing series and simulation
parameters documented in Ting et al. (1986), Cundall et al. (1989) and Oda et al. (1997). In a second step the particles within the testing box are subjected to a
confining pressure and equilibrated by adjusting the wall velocities. During the third step the lateral walls are replaced by stress controlled boundary walls.
Page 316
Figure 6. Different packing of spheres for numerical simulation of triaxial testing, (a) constant radius, (b) normal distribution, (c) Gauss
distribution, (d) upscaled grainsize distribution of Berlin sand.
The specimens were then subjected to a prescribed displacement representing a compression test by moving the upper boundary downward. The lateral walls are
permanently adjusted to provide a constant confining pressure on the specimen. Simulations were performed without gravity loading for several confining pressures
between 50 kN/m2 and 400 kN/m2.
4.2 Grain size distributions
In general every given grain size distribution can be simulated. The grain size distribution of the first test series satisfy different forms of distribution as presented in
Figure 6. Sample (d) represents the upscaled grain size distribution of Berlin sand as depicted in Figure 4.
From the results of the first triaxial testing series as depicted in Figure 7 it is obvious that the type of grainsize distribution has a strong effect on the stressstrain
behavior. Obviously numerical test specimen (a), (b) and (c) strongly overestimate the contractant behavior of the sample compared to sample (d).
Although the peak values of deviatoric stresses from sample (a) and (b) are more closely reaching the level of the Berlin sand specimen, the stress level decreases
significantly after reaching a peak value. Sample (c) and (d) do not reach the comparable stress level of the real soil sample but their postpeak behavior does not show
the remarkable decrease as observed for the first two numerical samples. Although also sample (c) overestimates the volumetric straining at the beginning of the test the
dilatant behavior—in terms of the maximum dilatation angle—is comparable to Berlin sand. The best approach regarding the volumetric straining obviously is given with
sample (d), which was used for further studies.
4.3 Upscaling
At the present stage for the simulation of large scale problems a number of obstacles does exist from which the most significant one apparently is the number of
Figure 7. Stressstrain response of triaxial testing on spherical particle specimen with different grain size distributions.
particles required to simulate a geotechnical problem in its original dimensions. Practically it is therefore also not possible at present to model geotechnical laboratory
tests and engineering problems by particle sizes that correspond to grain size dimensions in reality. For this reason an upscaling is used. The diameter of grains from the
initial grain size distribution is multiplied by a constant factor. Considering DIN 18137 for triaxial laboratory testing this leads to a triaxial testing container of about 10
times the largest grain diameter for uniform grain size distributions and 5 times the biggest grain diameter for nonuniform distributions. In proportion to the upscaling
applied within this study, the testing environment and boundary conditions were adjusted by using a cross section of 25 cm×25 cm for triaxial and oedometric testing
devices.
In Figure 8 the results from numerical triaxial tests with different numbers of spherical particles, representing different upscaling factors, are given. The grain size
distribution used in these tests corresponds to Figure 6d. The scattering of stressstrain curves is however less pronounced than for tests carried out with differently
shaped grain size distributions. It can clearly be seen that for a lower number of simulated grains (corresponding to a higher upscaling factor) the contractant behavior is
overestimated whereas the dilatant behavior in respect to its maximum angle is less influenced and more comparable among the tested samples.
Page 317
Figure 8. Stressstrain response of triaxial testing on spherical particle specimen with different amount of grains but equally shaped grain
size distributions.
4.4 Particle friction coefficient
The shearing resistance of granular materials such as sand strongly depends on the interaction between sliding particles, the shearing strain, the initial density of packing
and the particle geometry.
One of the most important micro parameters among them is the internal friction between spheres (Eq. 5). This parameter determines the transition between elastic
and plastic deformation on the micro level. The interaction forces between each actual contacting particle are determined according to a constitutive relationship or
interaction law as presented before. The appropriate modeling of friction certainly plays a significant role in the correct capture of the true features of most problems
that are considered within this work.
Again the grain size distribution of Figure 6d has been used to assess the influence of the particle friction coefficient Figure 9 shows that a nonlinear stressstrain
response is produced in which the initial stiffness modulus continually decreases until a peak axial load is reached. With Figure 9 it becomes obvious that the initial
stiffness modulus does also strongly depend on the coefficient of friction μ applied to the spheres. Less influence was observed when the normal or shear stiffness of the
spheres was altered or the ratio between shear and normal stiffness was changed. Also the peak
Figure 9. Influence of particle friction coefficients.
value reached in the stressstrain diagram is a function of the friction coefficient applied to the particles.
As shown in Figure 9 the coefficient of friction does have a significant influence on the dilatant behavior of the material. The experiments also show that by
considerably increasing the interparticle friction, the shear strength does increase less pronounced and obviously tends to reach an upper margin. If friction on the other
side is very low at the contact points, the material cannot sustain applied shear forces as they are developed during the triaxial tests.
In reality particles show irregular surface texture and are not spherical in shape. Particles may therefore contact their neighbors not only in points but also through
surfaces and contact lines.
In such cases rolling resistance can occur to some extent and even contact moments can be transmitted in addition to contact forces as shown in Figure 1. When
assuming spherical particle shapes the rolling resistance provided by arbitrary shaped particles on the microscopic scale is generally neglected. As shown in Figure 9
this rolling resistance, which leads to a higher resistance of arbitrary shaped grains during shearing, can only partly be compensated by increasing the interparticle friction
for spherical particles.
However, the side effect of increasing the friction coefficient is a rapid decrease of the postpeak stress level. This stressstrain behavior was also observed when
spherical particle simulations were carried out with contact bonds between the single particles to prevent particles from heavily rotating and sliding (Fig. 10). Contact
bonds are breaking if the shear and
Page 318
Figure 10. Influence of contact bonds.
Figure 11. Overlapping spheres forming different clumps.
normal forces exceed the predefined bondstrength of the material.
As these effects should vanish for nonspherical particles, which provide a more realistic grain shape and therefore a better resistance against the rolling of spheres,
further triaxial testing was performed with clumps.
4.5 Particle shape, nonspherical particles
Nonspherical particles (clumps) were generated by two overlapping spheres (Fig. 11). Considering an existing assembly of spheres, every sphere is replaced by a
clump with the same volume and an adjusted new density. The unbreakable clump is rotated randomly in space by a transformation matrix at the coordinates of the
replaced spherecenter to ensure an arbitrary orientation of the new clumped particle.
Figure 12. Stressstrain response of clumps with different degrees of overlapping.
Figure 13. Stressstrain response of clumps with different interparticle friction coefficients (μ).
Page 319

Figure 14. Stressstrain response of clumps with different stiffnesses (k n=k s).
In Figures 12–14 the results are depicted from simulations carried out with clumped particles in a cubic specimen representing an upscaled grain size distribution of
Karlsruhe sand (Fig. 4).
Simulations are again aiming to investigate the dependency of the stressstrain response from
– the degree of overlapping of spheres forming a clump,
– the coefficient of internal friction (μ) and
– the stiffness k n and k s.
In comparison to simulations carried out with spherical particles, the results of nonspherical particle simulations tend to oscillate considerably due to particles
overrunning each other but apparently can provide a much better fit to the stressstrain response of laboratory tests. For results presented in Figure 15 one set of
parameters was selected which is fitting the laboratory testing on Karlsruhe sand and which will be used in the following to simulate oedometric loading.
4.6 Macroscopic friction angle φ
A direct relationship between the macroscopic friction angle of sand, derived from numerical simulations, and the particle friction coefficient μ in the numerical model is
shown in Figure 16. Simulations were
Figure 15. Final sets of parameters for simulating Karlsruhe sand with clumped particles.

Figure 16. Relation between macroscopic friction angle and friction coefficient of particles (porosity: 43%; σ2=σ3=100 kN/m2).
carried out with 9800 spherical particles and clumps and a Gaussian distribution of particles. Figure 16 shows that specimens composed of clumps are leading to higher
macroscopic friction angles φ, even for very small friction coefficients μ, compared to simulations performed with spheres. For higher friction coefficients the
macroscopic angle of friction measured in the numerical simulations does not exceed a value of more than φ≈49° for clumps and φ≈36° for spherical particles.
Page 320
5 OEDOMETRIC TESTING
Further simulations were carried out to simulate oedometric loading with the objective to investigate the behavior of clumps under compression loading.
5.1 Test setup
First numerical tests were performed on spherical particles in a testing box as shown in Figure 17, which is identical to the geometry of a laboratory device.
More detailed simulations were later performed with clumps in a cubic specimen (side length 25 cm) taken from triaxial tests with properties depicted in Figure 20.
The outer boundaries (side walls) were fixed in horizontal direction and the bottom in vertical direction. The top boundary was loaded in 50–100 kN/m2 steps up to
400 kN/m2 with an unloading path starting at 300 kN/m2 (Fig. 20).
Simulations with samples that initially were created for triaxial testing by radius expansion (section 4.1), exhibited a very stiff response for lower vertical loading of up
to 200 kN/m2 (Fig. 20). For a higher loading and also for reloading paths the stiffness derived from the test results of the numerical model were comparable to the
stiffness derived from test results of laboratory testing.
Figure 17. First model for oedometric testing with spherical particles.
Figure 18. Second model for oedometric testing derived from triaxial testing with clumps, black lines show contact forces due to gravity
loading before oedometric loading of the sample started.
To investigate the influence of sample generation, the specimen was prepared alternatively according to Figure 19. In a 75 cm high sample with a quadratic cross
section (25 cm side length) clumps were created at arbitrary locations without contacting each other. All particles were then accelerated by gravity loading. At the same
time a friction coefficient of μ=0.25 was applied to all particles. After the sample reached equilibrium the upper part of particles was deleted to ensure a plane loading
level. The sample was than loaded in the same manner as aforementioned. Figure 20 shows that the sample created by gravity loading behaves less
Figure 19. Creating a sample by gravity loading.
Figure 20. Results of oedometric loading from a laboratory test, a sample taken from triaxial testing and a sample created by gravity
loading.
Page 321
stiff for lower vertical stresses and thus more accurately corresponds to results obtained from laboratory tests than a numerical sample that is created by the radius
expansion method.
However when compared to laboratory testing the numerical sample obviously still is packed too densely and exhibits a much stiffer response than the real soil
sample for low compressive stresses.
6 USE IN PILING TECHNOLOGY
6.1 Consequences from modeling triaxial and oedometric testing
Based on the experience gained with the numerical simulation of triaxial and oedometric loading on particle specimens, following conclusions are drawn in respect of
using the DEM for detailed investigations of the drilling processes during pile installation or the pile testing:
– Using nonspherical particles allows a realistic simulation of the stressstrain behavior of coarsegrained soils under triaxial loading.
– For the same set of parameters a realistic modeling of oedometric loading is only appropriate for higher vertical stress levels. A certain enhancement is achievable by
a modified assembly generation. However this requires comparatively high computational efforts.
– Simulation of loadsettlement behavior of piles will result in a strong overestimation of pile base resistance.
– Numerical particle models that focus on the simulation of skin friction effects and shear deformation seem to be most promising.
In the following a simple model is generated to investigate the soilstructure interaction in the transition zone of displacement piles and surrounding soil.
6.2 Soil displacement piles
A small section of a pile shaft has been modeled to learn more about the effect of soil displacement during pile installation, the development of skin friction, the change
in densities around the pile, and the changes in the stress distribution within the sample. The model represents a quarter of a pile (Fig. 21a) and is based on a sample
generated for oedometric testing. The initial specimen was generated by gravity loaded clumps as described in Section 5.
The top wall needed to simulate oedometric loading has been replaced by a layer of small spheres. Every single sphere has been loaded in vertical direction to ensure
an average overburden stress of about 100 kN/m2 to simulate a sample originally located within a deeper soil layer. After achieving equilibrium a small hollow cylinder
as depicted in Figure 21b is slowly expanded up the final shaft diameter. The bottom of the cylinder is closed by a plate, which is changing diameter according to the
expansion process to avoid clumps escaping upwards through the expanding cylinder. After reaching the final shaft diameter clumps below the expanded cylinder are
deleted and an outer casing is installed (Fig. 21c). This casing has a slightly larger diameter and is fixed in space. After equilibrium has been achieved again, the upper
pile shaft segment is moved downward by a prescribed constant speed and slides into the outer casing but without touching it. This modeling allows investigation of the
shaft resistance of a pile separately from any base resistance. The influence of a lower boundary condition is, although not entirely eliminated, considerably reduced.
The forces acting on the upper shaft segment are recorded and converted into shaft friction by dividing the actual vertical resistance through the actual area of the shaft
segment. The distribution of displacements and the contact forces after the expansion
Figure 21. Sketch of the model, cross section (a) and elevation (b, c).
Figure 22. Particle model to simulate the dependency of skin friction and density on the amount of soil displacement.
Page 322
Figure 23. Crosssection with displacement vectors (left) and distribution of contact forces (right).
Figure 24. Skin resistance in dependency of vertical displacement of the shaft segment.
process within the sample are exemplarily shown in Figure 23.
Qualitatively the measured skin resistance is shown in Figure 24 in dependency of the vertical downward movement of the expanded shaft segment. Although these
results are in good agreement with the shape of measured skin resistance in field tests, the amount of skin resistance strongly depends on the shaft stiffness and friction
assumed for the shaft surface within the numerical model. To evaluate this dependency detailed simulations are presently carried out. These studies do also include the
effect of different degrees of soil displacement on the final shaft resistance.
7 CONCLUSIONS
The stressstrain behavior of an assembly of spherical and nonspherical particles has been investigated under triaxial and oedometric loading. It was shown that non
spherical particles can be used to backanalyze existing results from laboratory triaxial tests with very good agreement. By using identical model parameters oedometric
loading can be simulated properly for higher vertical stress levels. For lower vertical stress levels the dependency on the specimen generation procedure becomes vital.
A preliminary model has been presented on how these results and PFC in general can be used to clarify questions related to the efficiency of soil displacement piles.
ACKNOWLEDGMENTS
The authors wish to acknowledge the financial support of the European Commission in the present work that was part of the Fifth Framework Program and the TOPIC
project.
REFERENCES
Achmus, N., AbdelRahman, K. 2003. The influence of “upscaling” on results of particle method calculations of noncohesive soils. In Konietzky et al. (eds.) Proc. of the
1st International PFC Symposium on Numerical Modeling in Micromechanics via Particle Methods, Gelsenkirchen, Germany, Nov. 6–8, 2002, Balkema ISBN
90 5809 532 0.
Cundall, P.A. 1971. A computer model for simulating progressive, largescale movements in blocky rock systems. In Proc. Symp. Int. Soc. Rock Mech., Nancy 2, art.8
Cundall, P.A., Strack, O.D.L. 1979. A discrete numerical model for granular assemblies. Géotechnique 29:47–65.
Cundall, P.A., Jenkis, J.T., Ishibashi, I., 1989. Evolution of elastic moduli in a deforming granular assembly. Powders and Grains. Balkema, ISBN 90 6191 984 3.
Itasca Consulting Group, Inc., 1999. PFC Ver. 2.0 Theory and Background. Minneapolis: Itasca.
Iwashita, K., Oda, M. 2003. Numerical simulation of granular materials flow in silo by modified distinct element method. In Konietzky et al. (eds.) Proc. of the 1st
International PFC Symposium on Numerical Modeling in Micromechanics via Particle Methods, Gelsenkirchen, Germany, Nov. 6–8, 2002. Balkema ISBN 90
5809 532 0.
Oda, M., Iwashita, K., Takiuchi, T., 1997. Importance of particle rotation in the mechanics of granular materials. Powder & Grains 97, ISBN 905410 884 3.
Ting, J.M., Corkum, B.T., Greco, C. 1986. Application of the distinct element method in geotechnical engineering. 2nd Symposium on Numerical Models in Geomechanics,
Ghent.
Page 323
Particle shape influence on the fracture behavior of concrete
N.Monteiro Azevedo
CAeMD, Lisboa, Portugal
J.V.Lemos
LNEC, Lisboa, Portugal
5809 679 3
ABSTRACT: The Discrete Element Method was initially applied to the analysis of discontinuous media, e.g. in rock mechanics and soil mechanics.
Recently the DEM has been used in fracture studies of nonhomogeneous continuous media, for example concrete and rock. A 2D circular rigid
discrete element formulation based on the Discrete Element Method that has been further developed to model concrete has been adopted. The
performance of the DEM based model enhanced with new particle generation algorithms has been evaluated for uniaxial tension and uniaxial
compression tests. Tests performed with circular and noncircular aggregates, with a given aspect ratio, have shown similar values of fracture
toughness.
1 INTRODUCTION
The Discrete Element Method DEM was initially introduced for the study of discontinuous rock. Rigid circular particle models have also been developed for the study
of the micromechanic behavior of soils Cundall & Strack (1979). Recently, circular particle models have also been applied in the study of fracture analysis in rock,
Potyondy & Cundall (1996), and concrete, Meguro & Hakuno (1989) and Takada & Hassani (1996). This has been accomplished through the development of
constitutive models for the interparticle contacts enabling the geomaterials modeling. This type of models have a great potential in the fracture study of large concrete
structures such as concrete dams and their foundations, namely the fracture analysis of the interface connecting the concrete to the rock. In this paper only the concrete
modeling is discussed.
Particle models, by taking directly into consideration physical contact mechanisms and the influence of the concrete mesostructure, are able to explicitly capture the
fracture mechanisms present in concrete. By simulating the concrete mesostructure the particle method prevents the localization of damage into regions, not sufficiently
large, when compared to the inhomogeneity size, Bazant (1986).
In the approach here adopted concrete is regarded as a skeleton of aggregate particles of various sizes, almost in direct contact with each other. The cement matrix
is intended to act as filler enabling the structure to be able to carry tensile stresses.
A 2D rigid circular particle model based on the DEM that has been developed to model concrete fracture has been adopted (Monteiro Azevedo 2003). This
particle model was shown to give a good agreement with concrete experimental data, namely the peak stress values, the crack initiation and crack localization
(Monteiro Azevedo et al. 2003).
Recently, particle generation algorithms have been proposed that enable the generation of an aggregate particle structure that follows the shape, the size, the aspect
ratio and the distribution of the aggregate particles present in concrete (Monteiro Azevedo & Lemos, 2003). The particle algorithms proposed include both the
generation of allcircular particle assemblies or hybrid assemblies with circular and polygonal particles. Here only allcircular particle assemblies are discussed.
The behavior of the adopted discrete model enhanced with the new particle generation schemes has been assessed in uniaxial tension and compression tests. The
influence of the particle geometry and the particle elongation ratio of the particles representing the concrete aggregates have been evaluated.
2 FORMULATION
In the DEM the domain is replaced with an assembly of discrete entities that interact with each other through contact points or contact interfaces. As proposed by
Cundall & Strack (1992) in order for an algorithm to
Page 324
be defined as DEM it has to allow finite displacements and rotations of the discrete bodies, including complete detachment. It also has to be able to recognize new
contacts as the calculation progresses.
In the DEM the set of forces acting on each block/particle are related to the relative displacement of the block/particle to its neighbors. Given the applied forces,
Newton’s 2nd law of motion is applied in order to define the new position of the block/ particle.
The relation between the generalized forces transmitted through the contact and applied at the center of gravity of the particles to the generalized displacements at the
center of gravity of the particles, see Figure 1, can be expressed by:
{f}=[Ke]{d}
(1)

where, {f} is the generalized force vector, {d} is the generalized displacement vector and [Ke] is the contact stiffness matrix in terms of particle displacements.
The generalized force and displacement vectors are given, in accordance with Figure 1, by:
{f}={f 1f 2f 3f 4f 5f 6}t, {d}={d1d2d3d4d5d6}t
(2)
The single contact element stiffness matrix, [Ke], in terms of particle displacements for a reference plane passing through the normal to the point contact ni=(n1, n2) is
defined by:
(3)
where:
k n and k s—contact normal and shear stiffness;
—Euclidean norm of the corresponding vectors;
—centers of gravity of particles A and B, respectively;
—contact point coordinates;
An explicit time marching calculation scheme based on the centereddifference algorithm is adopted. When it is only intended to obtain a steady state solution a mass
scaling algorithm is adopted in order to reduce the number of time steps necessary to reach the desired solution, Underwood (1983). The particle scaled mass
Figure 1. Forces and displacements for a given contact.
and scaled inertia used in the calculations are then given by:
mscaled=0.25 k t; Iscaled=0.25 k θ
(4)

The translation stiffness k t and the rotational stiffness k θ of each particle must include the contribution of all the particles in contact:
(5)
(6)

where Nc is the number of particles in contact.
The contact stiffnesses are defined using:
(7)
where, L is the interparticle distance, t is the thickness of the particle assembly, h is defined as the contact height being equal to the smallest diameter of the particles
involved, and for plane stress:
(8)
where, E and ν are the Young’s modulus and the Poisson’s ratio of the continuum material.
In the determination of the contact strength an analogy to the method used in the definition of the contact stiffness is used. The contact strength can then be
approximated by:
Fn,t=σn,t 2.0 Rmin t
(9)
where, Fn,t is the generalized ultimate contact force, σn,t is the ultimate value of the stress in the desired direction, normal or tangent to the contact plane and Rmin is the
minimum radius of the particles in contact.
For the interparticle contacts an extended MohrCoulomb model with bilinear softening is adopted, Figure 2. The bilinear softening diagram is based on the model
proposed by Rokugo (1989). The values of the maximum resistant tensile force and maximum cohesion force are reduced in function of the current value of the contact
damage.
Figure 2. Contact constitutive laws.
Page 325
3 PARTICLE GENERATION
The modeling of a given structure with the particle method requires the previous definition of the particle distribution based on a given sieve analysis. In Monteiro
Azevedo & Lemos (2003) particle generation algorithms are proposed which allow the generation of compact particles assemblies with the shape, aspect ratio and
distribution of the aggregate particles present in concrete. The random aggregate procedure, Bazant et al. (1990), is based on the following steps:
1. The volume ratio of each particle size is specified in advance setting the area of aggregate for each adopted grading segment, Dmin and Dmax, which are,
respectively, the maximum and the minimum diameter of a given grading segment.
2. A given particle diameter is defined using, D=Dmin+ηiDmax, in which ηi is a random number uniformly distributed between 0 and 1.
3. A uniform probability distribution is assumed for the particles center of gravity. Under this assumption a random number generator from a standard computer library
is used to generate the pairs of coordinates (xi, yi) of the particle centers: x i=Xmin+η1 (Xmax−Xmin) and yi=Ymin+η2 (Ymax−Ymin). Wnere: Xmin, Xmax, Ymin, Ymax
are the minimum and the maximum X−Y coordinates of the rectangular area in consideration, being η1 and η2 two independent random numbers uniformly
distributed between 0 and 1. Nonrectangular specimens need to be previously discretized with rectangular elements.
4. For each generated coordinate pair the new particle is checked for possible overlaps with the previously placed particles and with the specimen boundaries. The
coordinate pair is rejected if an overlap occurs. The minimum gap between two particles is taken as a function of the minimum particle diameter of the particles in
contact, γ min (DA , DB). A value of γ=0.30 is suggested in Wang et al. (1999). If necessary, the value of γ can be further reduced so all particles can be placed into
the domain. A minimum gap of γDA is adopted for the intersection between the particle and the boundary segment.
5. The random generation of the coordinate pairs for a given grading segment proceeds until the area of aggregate left to be generated is less than π The
remaining aggregate area to be generated is then transferred to the next grading segment. Particles are first inserted from the grading segment with the highest
maximum diameter to the grading segment with the smallest maximum diameter. The procedure is stopped after the last particle of the smallest grading segment is
inserted.
For the polygonal particle it is further required to define the particle shape, particle elongation ratio and particle orientation before step 4. The number of vertices of the
polygonal particle is a random variable varying uniformly from 4 to 10 vertices. Given the grading segment diameter D the radius of each vertex Ri is given by Ri=(D/2)
+ηjδ(D/2), where ηj is a random number uniformly distributed between 0 and 1, δ(D/2) is the aggregate radius amplitude, a value of 0.25 of the grading segment radius
is adopted.
Figures 6a, b present the aggregate structure adopting, respectively, circular shape and polygonal shape with a variable 1 to 2 aspect ratio. The random aggregate
particle assemblies have a high porosity without real particle contacts. In order to apply the DEM it is necessary to replace the domain of analysis with a particle
assembly with a low porosity.
A compact particle system is generated according to a void elimination scheme proposed by Monteiro Azevedo (2003) that adopts the DEM in the modeling of the
discontinuous behavior proper of a granular media. The void elimination procedure combines the radius expansion procedure Itasca (1995), the random macroparticle
growing procedure Sakaguchi & Mühlhaus (1997), the gap particle insertion mechanism, Monteiro Azevedo (2003) and the DEM cycle, Figure 3.
The radius expansion procedure places particles based on a given uniform radius grading with half their radius. After the particle placement, the real particle radius is
set and the DEM cycle is applied in order to obtain a particle assembly with smaller particle overlaps.
In the random macroparticle growing RMPG procedure algorithm seed particles, SP, positioned on a given domain have six directions of propagation into which
new particles can be placed. A seed particle can be any particle present in the assembly resulting from a radius expansion procedure. A specific direction of
propagation is ruled out if the corresponding position has already been occupied by a given particle. The generated particles are not allowed to overlap and they all
Figure 3. Void elimination procedure diagram.
Page 326
Figure 4. RMPG six propagation directions.
Figure 5. Gap particle insertion attempt.
have the same radius. The directions of propagation are randomly selected.
In a DEM solution cycle a contact detection procedure is applied in order to identify the particles in contact. In order to reduce the contact detection computational
costs an enlarged bounding box is usually adopted, Monteiro Azevedo (2003). All the particles that intersect the particle enlarged bounding box are recognized as
potential contactors. In the gap particle insertion mechanism, Figure 4, attempts are made to place particles in the contact location of particles with a gap higher than a
certain threshold value taken advantage of the information stored in a DEM program.
The interior domain of each aggregate particle is first discretized adopting the void elimination procedure. The same procedure is applied in the discretization of the
void space corresponding to the outer aggregate domain. A particle radius of 1.0 mm is adopted in the void elimination procedure.
Table 1 defines the aggregate content that is taken into consideration when generating the particle
Table 1. Aggregate grading.
Aggregate 1879 kg/m3
0.00–0.25 mm 8%
0.25–0.50 mm 12%
0.50–1. 00 mm 12%
1. 00–2.00 mm 10%
2.00–4.00 mm 14%
4.00–8.00 mm 20%
8.00–16.00 mm 24%
Figure 6. Particle assemblies—from the aggregate structure to the final compact assembly.
assemblies discussed. The particle assemblies are generated for an area of 100 by 50 mm. On all particle systems a minimum value for the aggregate diameter of 2.00
mm is adopted.
As shown in Figures 6a & 6b, the concrete aggregate structure is first generated. For the particle assemblies here discussed a minimum distance equal to 0.10 of the
minimum radius of the particles in contact is adopted.
The interior domain of each aggregate particle is first discretized, Figures 6c and 6d. Figures 6e and 6f show the final compact assemblies.
A superparticle algorithm, Itasca (1995), which enables the use of particle clusters is adopted for the particles belonging to the same aggregate particle reducing the
number of contacts present in the particle assemblies. An initial particle overlap is adopted on all the particle assemblies.
4 FRACTURE TESTS
Table 2 presents the microproprieties that need to be defined in the particle model. The elastic constants
Page 327
Table 2. Micro elastic and strength properties.
Ec [GPa] v c Cc [MPa] μc

25.0 0.40 3.45 11.50 0.20 6.5 900.0
Table 3. Macro elastic and strength properties/ numerical.
E [GPa] σt [MPa] σc [MPa] Gf [N/m]

Circ. 32.6 3.47 35.62 45.9
Pol. 1 32.6 3.40 37.02 37.6
Pol. 1–2 35.1 3.43 35.64 48.9
Pol. 2–3 37.5 3.30 35.62 40.4
are defined based on the Young’s modulus Ec and Poisson’s ratio nc of the contact equivalent continuum material. The contact stiffnesses are defined based on
Equation 7.
The macroscopic values of the Young’s modulus and Poisson’s ratio are numerically defined in uniaxial tension and compression tests in particle assemblies with a
rectangular geometry of 100 by 100 mm assuming plane stress constitutive equations.
The contact strength properties need to be previously calibrated in uniaxial tension and compression tests. A given value for the softening energy is adopted for the
normal direction of the contact Gft and for the contact shear direction Gfc. In Table 2 are presented the strength properties that give the best agreement in terms of the
peak values of stress and fracture toughness. Table 3 presents the macroscopic numerical values that were obtained for the several particle systems here discussed in
uniaxial tension and uniaxial compression.
The uniaxial tension tests are performed on particle assemblies with a rectangular geometry of 50 by 100 mm, while the uniaxial compression tests are performed for
particle assemblies with a rectangular geometry of 100 by 100 mm. The load is applied to the particle assembly by setting the displacements of a rigid wall, Itasca
(1995). In the fracture tests here discussed there is no transmission of shear forces at the interface connecting the particle assembly to the rigid wall modeling the
loading plate.
Four types of particle systems are analyzed:
– Circ.=circular particle assemblies,
– Pol. 1=polygonal particle assemblies with an aspect ratio of 1,
– Pol. 1–2=polygonal particle assemblies with a variable 1 to 2 aspect ratio, and finally,
– Pol. 2–3=polygonal particle assemblies with a variable 1 to 3 aspect ratio.
Figure 7. Tensile tests—final crack pattern.
Table 3 presents the macroproprieties that have been numerically calculated with the particle model, namely the Young’s modulus, the Poisson’s coefficient ν, the
ultimate tensile stress, the ultimate compression stress and the fracture energy.
From Table 3 it can be seen that the four particle sys stems here studied give similar results. The four particle assembly types have in common the same aggregate
quantity and the same microproperties.
Figure 7 shows the final crack patterns for each type of particle assembly. In general, the initial crack was occurring on the lateral zones of the particle assembly, the
initial cracks were then propagating towards each other to the interior. At the end of the tensile test a
Page 328
Figure 8. Compression tests—final crack pattern.
Figure 9. Stress/deformation relationships—tensile tests.
final crack crossing the particle assembly is formed. In some tests the final crack patterns have a given overlap, Figures 7a & c. A similar behavior is known to occur in
concrete.
In the compression tests the particle assemblies were divided in a variable number of triangular/conical pieces that can shear off, Figure 8. The shear cracks
Page 329
Figure 10. Stress/deformation relationships—compression tests.
are formed by an array of an “en echelon” splitting cracks. Even though the particle model includes a perfect friction contact model, pure friction forces on cracked
contacts either do not occur or when occurring transmit a low value of contact force when compared with the contact forces present in the assembly.
In the particle model here adopted the cracks can only propagate along the boundary of the rigid particle explaining a certain tortuosity of the final crack pattern. A
similar mechanism is observed in concrete. As shown in Figures 7 and 8, the particle model has the ability to follow the distinct fracture mechanisms known to occur in
concrete for both uniaxial tension and uniaxial compression.
Figure 9 shows that the tensile stress/deformation diagrams have a long tail. This phenomenon occurs even in the cases where an almost stress free macrocrack is
formed. This long tail can be explained by the fact that active bridging contacts can still transmit the load through the particle assembly. From Figure 9 it is possible to
verify that the geometry and the aspect ratio of the aggregate particle do not have a significant influence in the stress/deformation relationship in tension. The several
types of particles assemblies analyzed gave a similar response.
From the analyses of Figures 9 & 10 it is possible to conclude that the obtained numerical values are more brittle when compared to the numerical results obtained
by Vonk (1993), namely the compression tests. This can be partially explained by the fact that Vonk (1993) adopted an assembly of deformable polygonal particles for
both the particles representing the aggregate and for the particles representing the cement matrix with a higher variation of contact strength.
From Figure 10 it is possible to conclude that the four types of particles assemblies here studied have also a similar response in compression. From Figures 7–10
and Table 3, it is verified that it is possible to calibrate the particle model parameters in order to obtain a macroresponse similar to the obtained in concrete. It can also
be concluded that particle assemblies with noncircular aggregate shape have a softening behavior similar to the softening occurring in particle assemblies with circular
aggregates for both uniaxial tension and compression. The aspect ratio of the aggregate particle is also shown not to have a significant influence in the fracture behavior
of the particle assemblies.
5 CONCLUSIONS
A particle model is presented that captures the fracture mechanisms by taking into consideration the concrete aggregate structure and the physical mechanisms related
to the contact interaction. The particle model is shown to give figures of rupture similar to the obtained in concrete for similar fracture tests.
It is possible to calibrate the microproperties of the particle model in order to obtain a macroresponse similar to the observed for concrete. The calibration
Page 330
procedure needs to be executed for traditional fracture tests prior to the application of the model to other structural systems. The particle model is shown to capture the
distinct fracture mechanisms known to occur in concrete under uniaxial tension and compression.
The particles representing the concrete aggregate should have a shape and an aspect ratio closer to the concrete aggregate shape and the concrete aggregate aspect
ratio of the concrete to be modeled. For this reason it is important to devise flexible particle generation schemes.
When compared to the particle model adopted by Vonk (1993) the particle model here adopted is shown to be more brittle. Kotsovos (1983) through uniaxial
compression tests in concrete cylinders verified that when there is no shear force transfer at the interface between the loading plates and concrete, the concrete has an
immediate loss of loading capability as soon as the peak load is reached. The numerical results here obtained are in agreement with the experimental results obtained by
Kotsovos (1983).
Particle assemblies with noncircular particles representing the aggregate and particle assemblies with circular particles representing the aggregate have a similar
behavior. It should be noted that all the particle assemblies types here tested have a similar aggregate content and similar microproperties.
In order to increase the fracture toughness of the particle assemblies it is possible to generate aggregate particles with a higher surface roughness. In Monteiro
Azevedo (2003) it is shown that particle assemblies with aggregates having a higher surface roughness lead to an increase in the fracture toughness. In this work the
aggregate surface roughness of the four types of particle assemblies is similar and is given by the average particle size used to discretize the interior of each aggregate
particle.
It would also be relevant to develop particle generation schemes that allowed the definition of polygonal particles in zone corresponding to the cement paste. In this
study, particles placed outside the aggregate interior are circular. This would create a higher variation in the particle assembly maybe leading to higher fracture
toughness. Another issue would be to verify if the aggregate shape and aspect ratio does influence the concrete response if the material heterogeneity is taken into
consideration by assuming different stiffnesses for the aggregate particles and the cement paste particles and also assuming weaker interfaces at the interface cement
paste/aggregate.
ACKNOWLEDGEMENT
The financial support from the Foundation for Science and Technology of Portugal to the first author that allowed the development of the particle model here adopted is
gratefully acknowledged (Praxis XXIBD18018–98).
REFERENCES
Bazant, Z., 1986. Mechanics of distributed cracking. Appl. Mech. Rev., ASME, 4(5):675–705.
Bazant, Z., Tabbara, M., Kazemi, M. & Cabot, G., 1990. Random particle model for the fracture of aggregate or fiber composites. Journal of Engineering Mechanics,
ASCE, 116(8):1686–1705.
Cundall, P. & Strack, O., 1979. A discrete numerical model for granular assemblies. Geotechnique, 29(1):47–65.
Cundall, P.A. & Hart, R.D., 1992. Numerical Modelling of Discontinua, Engineering Computations, 9:101–113.
Itasca Consulting Group, Inc. 1995. PFC2D—Particle Flow Code in 2 Dimensions, Vers. 1.1. Minneapolis: Itasca.
Kotsovos, M., 1983. Effect of testing techniques on the postultimate behaviour of concrete on compression. Material and structures, RILEM, 16(91):3–12.
Meguro, K. & Hakuno, M., 1989. Fracture analysis of concrete structures by the modified distinct element method. Struc. Eng., Earthquake Eng., 6(2):283–294.
Monteiro Azevedo, N., May, I. & Lemos, V., 2003. Numerical simulations of plain concrete under shear loading condtions, In H.Konietzky et al. (Eds), Proceedings of
the 1st International Symposium on Numerical Modeling in Micromechanics via Particle Methods, Gelsenkirchen, Germany, Nov. 6–8, 2002. pp. 79–86.
Rotterdam: Balkema.
Monteiro Azevedo, N., 2003. A rigid particle discrete element model for the fracture analysis of plain and reinforced concrete. Ph.D. thesis, HeriotWatt University,
Scotland.
Monteiro Azevedo, N. & Lemos, V., 2003. Generation of random particle assemblies for fracture studies of concrete, pp. 1–50, LNEC, 2003 (to be published).
Potyondy, D. & Cundall, P.A., 1996. Modeling rock using bonded assemblies of circular particles, In Aubertin et al. (Eds.) Proc. 2nd North American Rock Mechanics
Symposium, pp. 1937–1944.
Rokugo, K., 1989. Testing method to determine tensile softening curve and fracture energy of concrete. In Fracture Toughness and Fracture Energy, pp. 153–163.
Rotterdam: Balkema.
Sakaguchi, H. & Mühlhaus, H., 1997. Mesh free modelling of failure and localization in brittle materials. In T.A.Asaoka & F.Oka (Eds.), Deformation and Progressive
Failure in Geomechanics, pp. 15–21. Pergamon.
Takada, S. & Hassani, N., 1996. Analysis of compression failure of reinforced concrete by the modified distinct element method. In C.B.GD Manolis & D.E.Beskos
(Eds.), Advances in earthquake engineering, earthquake resistant engineering structures. Comp. Mech. Publications.
Vonk, R., 1993. A micromechanical investigation of softening of concrete loaded in compression. Heron, 38(3): 1–94.
Wang, Z., Kwan, A. & Chan, H., 1999. Mesoscopy study of concrete I: generation of random aggregate structure and finite element mesh. Computers and Structures,
70: 533–544.
Underwood, P., 1983. Dynamic Relaxation, in Computational Methods for Transient Analysis, New Work: NorthHolland, Eds. T.Belytschko and T.Hughes, pp. 246–
265.
Page 331
The use of particle flow code for investigating the stress memory effect in rocks
V.Louchnikov & S.P.Hunt
Australian School of Petroleum, The University of Adelaide, SA, Australia
A.G.Meyers
Rocktest, Adelaide, SA, Australia
5809 679 3
ABSTRACT: Deformation Rate Analysis (DRA) is a nondestructive method for estimating insitu stress. Recent laboratory studies (Meyers et al.
2002, Hunt et al. 2003a) provided data suitable for use as input into a numerical model that replicates the physical tests. The discrete element
modeling code PFC2D was used to reproduce the damage that occurs to rock at the micro level, and to model dynamic failure propagation. The
software was suitable for this purpose due to its ability to explicitly model mechanical damage in rocks. Of particular interest is the algorithm allowing
contact bond breakage, which replicates the generation and propagation of microcracks. Earlier work (Hunt et al. 2003b) showed a strong link
between the DRA memory effect and the breakage of interparticle contact bonds. This paper highlights the link and extends the current understanding
of the phenomenon. The ability of the numerical model to reproduce the DRA effect was validated. Although the bulk of the numerical tests were
conducted on models of intact rock, it was found that changes in the lithology and introduction of discontinuities did not hinder the manifestation of the
DRA effect. Application of particle flow code for investigating the stress memory in rocks provided a valuable insight into this assumption. The results
of the study encourage further use of the code for understanding the micromechanical behavior of rocks under loading.
1 INTRODUCTION
Knowledge of an insitu stress tensor is a key element in the process of designing subsurface excavations. No matter what mode of failure occurs, it is ultimately related
to the stress conditions or, more correctly, to the stress changes that occur as a result of excavating the rockmass. Mining engineers need to understand the
characteristics of the insitu stress and its derivative, the mining induced stress, at many stages of mine planning e.g. when determining a layout for development
openings, designing efficient ground support, establishing correct stoping sequence. Petroleum engineers use stress data for understanding the natural and induced fluid
flow, reservoir drainage pattern, well integrity and flooding strategies for enhanced recovery Selection of the optimal direction for directional drilling to avoid wellbore
instability depends primarily on the magnitude and orientation of the insitu stresses. Civil engineers need stress data when determining the dimensions, orientation and
shape of tunnels and underground caverns. Knowledge of the initial state of stress is also a legislative requirement; proposed underground excavations will not be
approved unless a risk assessment had been carried out. The risk of rockmass failure is generally determined by relating the rockmass strength to the maximum insitu
stress. Stress measurement in rock has been made since the 1930s. All methods can be broadly allocated into three principal groups, indicative, fieldbased and core
based.
The first category involves techniques that do not come under the strict definition of ‘stress measurement’ but rather ‘stress estimation’. They involve analysis of
breakouts in boreholes and discs created in core. These methods still require further validation before they can be accepted as reliable methods for stress estimation.
The second category involves fieldbased methods that disturb the insitu rock conditions by inducing strains or by opening preexisting cracks. This group includes
wellestablished and widely accepted techniques involving borehole relief (e.g. overcoring and borehole slotting) and hydraulic fracturing, both of which allow stress to
be directly measured. These methods have tended to have a success rate
Page 332
of approximately 60% and, being laborintensive and involving complex equipment, they can be expensive to implement in large numbers over a site.
The third category encompasses various tests carried out on drill core. They are referred to as ‘indirect stress measurement techniques’ as they do not measure
stress directly in the field but rather in a laboratory. They provide data on what the insitu stresses were at the time the cores were recovered. The methods include
techniques involving strain recovery (anelastic strain recovery and differential strain curve analysis) and the Kaiser effect (acoustic emission AE and deformation rate
analysis DRA). Among these indirect methods, AE and DRA are preferred because they allow the stress to be measured directly as opposed to indirectly as a function
of strain as done with the strain recovery methods.
Currently the methods are used for research purposes rather than for practical applieations; however if they can be proven reliable, they have the potential to address
some of the concerns with fieldbased methods as multiple tests are significantly cheaper to carry out than field techniques. This paper presents the results from a study
involving the application of the numerical modeling code PFC2D for exploring the DRA technique.
2 STRESS MEMORY IN ROCKS
Discovery of the stress memory phenomenon in solid materials is attributed to Josef Kaiser, a German engineer and scientist. Kaiser (1950), while studying acoustic
emissions (AE) emanating from metals under compressive loading, noticed a sudden onset of AE from a specimen once the loading surpassed the preexisting level of
stress applied during a previous test. This phenomenon has since been referred to as the Kaiser effect (KE). Goodman (1963) pioneered experiments on the KE in
geomaterials by observing the onset of AE in rocks under compressive loading. For some time after that study, there was limited notable research into the KE until
Kanagawa et al. (1976) realized that the effect had the potential to be applied to the evaluation of insitu stress. Subsequently, many practitioners (e.g. Kurita et al.
1979, Holcomb & Costin 1986, Hughson & Crawford 1987, Friedel & Thill 1990, Momayez & Hassani 1992, Pestman et al. 2002) have researched the effect.
AE methods involve uniaxially compressing a core while monitoring for acoustic emissions emanating from the rock. The intensiveness of the emissions is low until the
loading exceeds the previously experienced maximum insitu stress. At this stress, the emissions increase significantly as defined by the AE cumulative counts. Figure 1
highlights the resulting distinctive inflection point, which is often referred to
Figure 1. Typical graph of AE cumulative counts vs. axial stress.
Figure 2. LRR test principle.
as the ‘takeoff point’. The axial stress applying is the same as the historical maximum stress previously applied to the core.
In the DRA method, instead of recording acoustic emissions, strain is monitored with strain gauges as the core undergoes cyclic uniaxial loading. A change in the
stressstrain behavior can be observed once the applied stress exceeds the previously applied maximum stress. The process involves loading the specimen above the
anticipated maximum insitu stress and then unloading it and then repeating the cycle as indicated in Figure 2. This procedure will hereafter be referred to as the ‘load
reloadreload’ or LRR test. The strain difference between the two loading curves, correlated with the same value of the stress, is plotted in terms of the axial stress. The
basic formulation of the DRA method is expressed by the strain difference function:
∆εi,j(σ)=εj(σ)−εi(σ)
(1)
Page 333
where εi(σ) is the axial strain in the core in the first loading and εj(σ) is the axial strain in the last loading at the applied stress σ. It was experimentally demonstrated by
Yamamoto, et al. 1990 that the curve changes its gradient when it passes the load that the core had been previously subjected to.
Most researchers (e.g. Holcomb & Costin 1986, Stevens & Holcomb 1980, Stuart et al. 1993, Li & Nordlund 1993, Barr & Hunt 1999) of the KE agree that the
application of stress (in the inelastic region) to a rock in the field results in the formation of the microcracks. The subsequent restressing of the rock in the lab does not
produce any further notable damage to it until the previously applied stress level is exceeded as no further microcracks develop until this stress has been reached. With
continuing application of the stress, past the preload value, the specimen again begins to experience microcrack nucleation and growth. Thus it can be said the damage
in the rock sample caused by the preload has been memorized in the form of a network of cracks, which subsequently can be interrogated using the DRA method.
3 PHYSICAL PROTOTYPE
A 61 mm diameter core of coarsegrained sandstone was selected as a physical prototype for the numerical model used in this study. The core comprised
approximately 40% quartz, 40% feldspar with minor opaques and finegrained cement. The specimen was wellcemented and compacted with the average grain size
being ≈0.25 mm. The bulk density of this rock was approximately 2.4 t/m3, uniaxial compressive strength of 82 MPa and Young’s modulus of 14 GPa. A detailed
description of DRA tests on this core can be found in Meyers et al. (2002). Strain vs. stress curves resulting from these tests are shown in Figure 3. These tests
produced reasonable estimates of the preload stress the core had been subjected to as summarized by the Felicity ratio in Table 1. This ratio is the preload stress
expressed in terms of the stress determined using DRA.
4 MODELING THE KAISER EFFECT
The twodimensional Particle Flow Code, PFC2D, (Itasca 1999) was used to replicate the physical tests on the sandstone core. A detailed description of the
procedure used for creating the numerical model can be found in Hunt et al. (2003a, 2003b). The parameters used in the model are listed in Table 2. PFC2D employs
a discrete element modeling (DEM) algorithm and as such has an advantage over a continuum modeling approach because it allows the user to explicitly model
Figure 3. KE determined under LRR test carried out at preloads: a) 15.4 MPa; b) 35.9 MPa; c) 54.7 MPa.
Table 1. Results of the LRR tests on physical specimen.
Test
1 2 3
Preload stress, MPa 15.4 35.9 54.7
Estimated stress using DRA, MPa 16.4 36.6 55.4
Felicity ratio 0.94 0.98 0.99
fracture damage directly allowing micromechanical damage to evolve. A rock model is represented as an assembly of circular particles bonded together. The bonds are
assigned shear and tensile strength values. When the tensile strength of a bond is exceeded, the bond breaks and its tensile strength is set to zero. When the shear
strength is exceeded, the shear force is set to a residual value that depends on the compressive normal force acting across the contact and the coefficient of
Page 334
Table 2. Microparameters characterizing the synthetic material.
PFC2D microparameter Value
Ball density, ρ
2397 kg/m3
Ball contact modulus, Ec 20.0 GPa
Ball stiffness ratio (normal stiffness over shear stiffness), k n/ks 1.8

Ball friction coefficient, μ 0.5
Mean contactbond normal strength, σc (mean) 82.0 MPa
Std. deviation of the mean contactbond normal strength, σc (std. dev.) 30.0 MPa
Mean contactbond shear strength, τ c (mean) 82.0 MPa
Std. deviation of the mean contactbond shear strength, τ c (std. dev.) 30.0 MPa
friction, which results in frictional sliding. Each broken bond is analogous to the initiation of a potential microcrack in a physical rock. After a bond breaks, stress is
redistributed away from it, which may cause more cracks to form nearby. If the rock model continues to be stressed, the cumulative localized bond breakages may
develop into an inclined microfracture. This action is analogous to the development of an inclined crack in a physical rock, which can be the precursor of failure.
There are various results that can be obtained from the output produced by PFC2D. One parameter of particular interest is crack initiation stress σci. By definition,
this stress is that at which the crack volumetric strain begins to grow. This parameter cannot however be confidently determined in a 2dimensional code. The crack
initiation stress is therefore defined as the axial stress, at which a specified number of microcracks have formed (as defined by the user as a percentage of the total
number of cracks in the failed specimen). In this study, this value was selected to be 1%. Based on this value, the crack initiation stress determined by running a
simulation of the LRR simulations was found to be 11.5 MPa. It was assumed that if the stress applied in a simulation was below this value, the deformations in the
specimen would be perfectly elastic, i.e. the strains would be fully recoverable, and no cracks would be developed. However, if the stress was above this value a
number of cracks would be permanently created in the model. This assumption was tested by running a serried of LRR simulations in the regions 0<σ<σci and σ>σci.
The axial load on the model was progressively increased until one preload had been applied in the elastic region and three subsequent loads were applied in the inelastic
region. The stressstrain diagrams from the simulations in the inelastic region are shown in Figure 4.
Figure 5a shows the differential strain vs. axial stress relationship obtained from an LRR simulation for a preload of 6 MPa. The differential strain is constant
Figure 4. Stressstrain response of the numerical model in the LRR tests at three preload values: a) 15.4 MPa; b) 35.9 MPa; c) 54.7
MPa.
(=0) throughout the entire range as expected for a simulation conducted in the elastic region. This behavior is different to that observed in the following three tests (Figs.
5b–5d), which were carried out in the
Page 335
Figure 5. DRA plots and microcrack snapshots for different preload values: a) 6 MPa; b) 15.4 MPa; c) 35.9 MPa; d) 54.7 MPa.
Page 336
Table 3. Results of the LRR tests on the numerical model.
Simulation
1 2 3
Preload stress, (MPa) 15.4 35.9 54.7
Stress observed from the KE, (MPa) 15.3 36.5 54.7
Table 4. Felicity ratios determined from plots of Histories 1 and 2.
Simulation
1 2 3
Preload stress, (MPa) 15.4 35.9 54.7
KE, history 1, (MPa) 17.5 36.9 54.3
KE, history 2, (MPa) 17.5 36.7 54.1
Figure 6. History 1. Cumulative number of microcracks in the model vs. axial stress: a) 15.4 MPa; b) 35.9 MPa; c) 54.7 MPa.
Figure 7. History 2: Frictional energy in the model vs. axial stress: a) 15.4 MPa; b) 35.9 MPa; c) 54.7 MPa.
inelastic region at the preloads listed in Table 3. In each simulation a distinctive ‘memory kick’ can be observed corresponding to the stress imposed during pre
loading. Trend lines have been added to enable accurate determination of the inflection points. In general, the inflection points obtained are similar to those obtained in
Figure 3, but the curve shapes are dissimilar. In Figure 5b the differential strain increases past the inflection point at 15.3 MPa, whereas in Figure 3a the differential
strain decreases. The same is true for the initial part of the curve after the inflection point (at 36.5 MPa) in Figure 5c. Generally a decrease would be
Page 337
Figure 8. Varying lithology sample: a) model assembly; b) microcracks at failure; c) displacement vectors at failure.
expected and it is thought this would be observed if the number of particles and hence number of bonds was increased. Alongside each plot is a microcrack equivalent’
snapshot’ recorded at the peak of each preloading cycle.
The KE/DRA can also be detected in the simulation by monitoring two histories: 1) cumulative number of particle bond breakages assumed to directly related to
microcrack generation and 2) frictional energy in the model. These histories have been recorded at for the first reloading cycle and are shown in Figures 6 & 7. The
Felicity ratios in Table 4, calculated from these histories, indicate that the preload values were estimated acceptably.
5 INFLUENCE OF LITHOLOGY AND DISCONTINUITIES ON KE
It is well known rocks are heterogeneous material, consisting of minerals of various mechanical properties, having variable porosity and intersected by discontinuities
(i.e. joints, faults, bedding planes). A series of simulations were carried out in which these features were included in the models. At this stage of the study, it was not
intended to create models that replicated actual heterogeneous rocks. Instead the aim was to observe how well the code could simulate the introduction of various
heterogeneities. The first simulations involved models that consisted of two different materials, strong bonds at the ends of the specimen and bonds 100 times weaker in
the middle. The development of shear planes during the simulation can clearly be seen in Figure 8. As with the previous simulations, the ability of the code to model the
DRA affect was also observed in these simulations. Figure 9 shows clear evidence of the occurrence of this effect.
The next series of tests involved the creation of discontinuities within the model. To do so, the codes built in JSET algorithm was used to create 2 sets of joints as
shown in Figure 10. As shown in Figure 11, the simulation again showed clear evidence of the KE.
6 CONCLUSIONS
PFC2D was found to simulate well the Kaiser effect (KE) observed in a natural rock subjected to axial compression. The stresses at which the effect occurred were
consistent with preload stresses applied to the models. The development of microcracks within a model, the action that forms the basis of the KE, was clearly evident
in the simulations through the breakage of as interparticle bonds. The presence of the KE was also observed in more complex heterogeneous models
Page 338
Figure 9. DRA on varying lithology sample. Preload=34.2 MPa, KE detected=35.5 MPa, Felicity ratio=0.96.
Figure 10. Jointed sample: a) model assembly; b) microcracks at failure; c) displacement vectors at failure.
Page 339
Figure 11. DRA on jointed sample. Preload=25.3 MPa, KE detected=25.6 MPa, Felicity ratio=0.99.
containing variable lithologies and joints. As a result of this investigation, it can be concluded that PFC2D is a useful aid in the study of fracture mechanics in rock.
Ongoing work is investigating how the addition of a confining stress will influence the development of the KE. Future work will apply the techniques to investigate the
influence of polyaxial loading on crack development, a phenomenon difficult to study using physical methods.
ACKNOWLEDGEMENTS
The authors would like to thank the Mining Engineering Research Group at the University of South Australia for funding this research and the School of Petroleum
Engineering at the University of Adelaide for providing resources and support.
REFERENCES
Barr, S.P. & Hunt, D.P. 1999. An elastic strain recovery and the Kaiser Effect retention span in the Carnmenellis granite, UK. Rock Mech Rock Eng 32:169–193.
Friedel, M.J. & Thill, R.E. 1990. Stress determination in rock using the Kaiser effect. US Dept. of the Interior, Bureau of Mines, Report of Investigations RI9286.
Goodman, R.E. 1963. Subaudible noise during compression in rocks. Geol. Soc. of Amer. Bull., 74:487–490.
Holcomb, D.J. & Costin, L.S. 1986. Detecting damage surfaces in brittle materials using acoustic emissions. J. App. Mech., 108:536–544.
Hunt, S.P., Meyers, A.G., Louchnikov, V. & Oliver, K.J. 2003a. Use of the DRA technique, porosimetry and numerical modeling for estimating the maximum insitu stress
in anisotropic rock core. In Proc. of 10th Int. Congress on Rock Mechanics, Johannesburg, South Africa, September 2003.
Hughson, D.R. & Crawford, A.M. 1987. Kaiser effect gauging: the influence of confining stress on its response. In Proc. 6th Int. Cong. Int. Soc. Rock Mech.,
Montreal, vol.2, pp. 981–985.
Hunt, S.P., Meyers, A.G. & Louchnikov, V. 2003b. Modelling the Kaiser effect and deformation rate analysis in sandstone using the discrete element method. Computers
and Geotechnics J., 30(7):611–621.
Kaiser, J. 1950. An investigation into the occurrence of noises in tensile tests or a study of acoustic phenomena in tensile tests. Ph.D. Thesis, Tech. Hochsch.
München, Munich.
Kanagawa, T., Hayashi, M. & Kitahara, Y. 1976. Estimation of the spatial geostress components in rock samples using the Kaiser effect of acoustic emissions. In Proc.
3rd Acoustic Emission Symp., Tokyo, pp. 229–248.
Kurita, K. & Fujii, N. 1979. Stress memory of crystalline rocks in acoustic emissions. Geophys. Res. Lett., 6:9–12.
Li, C. & Nordhund, E. 1993. Assessment of damage in rock using the Kaiser effect of acoustic emissions. Int. J. Rock Mech. Min. Sci. & Geomech. Abstr., 30(7):943–
946.
Meyers, A.G., Hunt, S.P. & Oliver, K.J. 2002. The use of the DRA technique and porosimetry for estimating the maximum insitu stress in rock from core. Australian
Geomechanics J., 37(1):17–22.
Momayez, M. & Hassani, F.P. 1992. Application of Kaiser effect to measure insitu stresses in underground mines. In Proc. 33rd US Symposium on Rock Mechanics,
Eds. Tillerson JR & Wawersik WR, Santa Fe, New Mexico. Rotterdam: A.A.Balkema. pp. 979–987.
Pestman, B.J., Holt, R.M., Kenter, C.J. & van Munster, J.G. 2002. Application of a novel corebased insitu stress estimation technique. SPE/ISRM 78158. SPE/ISRM
Rock Mechanics Conference OilRock2002.
Page 340
Stevens, J.L. & Holcomb, D.J. 1980. A theoretical investigation of the sliding crack model of dilatancy. J. Geophys. Res., 85(B12):7091–7100.
Stuart, C.E., Meredith, P.G., Murrell, S.A.F. & van Muster, J.G. 1993. Anisotropic crack damage and stressmemory effects in rocks under triaxial loading. Int. J. Rock
Mech. Min. Sci. & Geomech. Abstr., 30(7):937–941.
Yamamoto, K., Kuwahara, Y., Kato, N. & Hirasawa, T. 1990. Deformation rate analysis: a new method for insitu stress estimation from inelastic deformation of rock
samples under uniaxial compression. Tohoku Geophysical J., 33:1–13.
Page 341
Numerically modeling the strength of ice using discrete elements
L.U.Arenson
UofA Geotechnical Centre, University of Alberta, Edmonton, AB, Canada
5809 679 3
ABSTRACT: Numerical modeling of the mechanical properties of frozen soils is challenging. The strength of these materials depends on various
factors, such as temperature, applied strain rate, ice content, particle size and salinity. Instead of using one constitutive model that represents the frozen
soil as a continuum, discrete element modeling was selected to model the interaction between the different phases within a frozen geomaterial on a
micro level. This paper describes the use of the clump logic to model the strength of polycrystalline ice at different temperatures and compressive strain
rates, which is the first step in modeling frozen soils. It was possible to model the strength accurately. At varying applied strain rate the model showed
the initial yield point. The subsequent strain hardening—strain softening characteristic of ice could not be reproduced for all test conditions. However,
relationships between temperature and peak strength, as well as between applied strain rate and peak strength were able to show the effect of the
initial model parameters on sample response. The future research is focused on the implementation of this information into a more comprehensive
contact model.
1 INTRODUCTION
Frozen soils are complex geomaterials. They consist of solid grains, ice crystals, unfrozen water and gas. Their mechanical response is therefore highly affected by the
response of each constituent and their interactions. The distribution of the constituents is further affected by environmental parameters, such as temperature or the
salinity of the pore fluid. At temperatures close to the melting point of ice, the unfrozen water content, for example, can be very significant (Anderson et al. 1973). The
strength of a frozen soil can vary extensively (several order of magnitudes) depending on factors such as temperature, solid grain size distribution, specific surface of
solid grains, salinity or loading conditions (cf. Andersland & Ladanyi 1994: Chapter 5). Hivon and Sego (1995), for example, showed that the strength of a frozen soil
is primarily affected by the unfrozen water content, which again depends on the above mentioned factors. As a result various constitutive approaches for ice and frozen
soils under different boundary conditions have been presented in literature (Vyalov 1963, Adachi et al. 1990, Sunder & Wu 1990, Fish 1991, Fish & Zaretsky 1997,
Schapery 1997, He et al. 2000, DerradjiAouat & Evgin 2001, Neaupane & Yamabe 2001). Such approaches were then used in numerical codes to model the
thermomechanical behavior of frozen soils (e.g. Soo et al. 1985, Neaupane & Yamabe 2001).
Constitutive models that consider the frozen soil as continuum have the disadvantage that the changes in composition and the crucial interaction between the
constituents cannot be simulated. In a frozen soil, where the strength of the solid grains is significantly greater than the strength of the ice crystal, new approaches that
deal with the microstructural behavior have promise. Not only is it possible to model the different materials, but the ice can be modeled to reflect its brittle or ductile
behavior, depending on the loading conditions.
Discrete element modeling can account for interactions between the particles. A model has been presented on the destruction of sea ice (Zyryanov et al. 2002) or
crack formation in an Arctic sea ice pack (Hopkins 2002). The viscous behavior of the ice can be modeled using advanced contact models between the particles
(Zyryanov 2004). With this model, the uniaxial compression tests presented by Sinha (1988) have been reproduced accurately.
This paper presents a different approach to model polycrystalline ice. Since this research was initiated to develop a model for frozen soils, its heterogeneous
composition was kept in mind. The presented results are only the initial step that forms the basis for ongoing development. The current research focuses on the
dependency of the parameters used within the contact model as a function of the applied boundary conditions, such as the temperature and strain rate.
Page 342
Table 1. Parameters for parallelbonded contacts.
Model: Polycrystalline ice
Parameter Description Value
Density Ice density
917 kg/m3
Ec Ballball contact modulus

kn/ks Ball stiffness ratio 1.1
Parallelbond stiffness Parallelbond stiffness ratio 1.1
μ Ball friction coefficient 0.01
Parallelbond mean normal strength Parallelbond mean shear strength

2 NUMERICAL APPROACH
The discrete element code, PFC2D (Itasca 2000), was used to model triaxial compression tests on polycrystalline ice. An extensive laboratory investigation presented
by Rist and Murrell (1994) was utilized to calibrate the model parameters. The tests on polycrystalline ice were performed at temperatures between −20°C and −40°
C, and over a wide range of strain rates (10−2 s−1 to 10−5 s−1). In addition, confining pressures up to 46 MPa were applied. This vast variety in boundary conditions
resulted in different mechanical behavior and failure modes, i.e. brittle to ductile failure with dense cracking, intermediate cracking and sparse cracking was observed.
The size of the triaxial sample was chosen to represent a typical triaxial test with a height of 150 mm and a diameter of 74 mm. A minimum ball radius of 0.5 mm and
a ratio between the maximum and the minimum radius of 1.66 results in the generation of about 7000 balls. The clump logic (Cho et al. 2004) was further used to
simulate the ice crystals. This approach is not important at this stage of the project. It is, however, crucial with respect to the solid particles that will be incorporated
within the frozen matrix during the future modeling of frozen soils. In order to represent somewhat natural conditions a clump size of 5 balls was selected.
A parallelbonded contact model was chosen with the parameters shown in Table 1.
Most parameters are a function of temperature T or the axial strain rate The tests, however, showed that the stiffness ratios can be maintained constant for all test
conditions. The value for the angle of friction was also chosen to be constant and small. Triaxial compression tests on polycrystalline ice showed that the confining
pressure has a small effect on the shear strength (Jones 1982), in particular at low strain rates.
3 RESULTS
The data presented in Rist and Murrell (1994) were used for the calibration of the PFC parallelbonded
Figure 1. Stressstrain relationship for various axial strain rates at sample temperature of −20° C and confining pressure of 100 kPa
(q=σ1−σ3).
contact parameters. In this section the contact parameters will be discussed as a function of the applied strain rate and the temperature.
3.1 Stressstrain behavior
Figure 1 shows typical stress strain curves for triaxial compression at a temperature of −20°C, a confining pressure of 100 kPa and various strain rates. Unfortunately,
only the peak values are available for the laboratory tests.
The curves represent the post peak behavior of the laboratory tests on polycrystalline ice satisfactorily. At a low strain rate, an initial yield point (e.g. Andersen et al.
1995) can be observed (Fig. 1), followed by slight strain strengthening. However, modeling the post peak behavior has problems. Nearly no strain softening was
observed at the lowest strain rate, which is in accordance with the laboratory results. At higher strain rates, a more significant drop in the deviatoric stress was
expected. In addition, the curves are not smooth, which might be due to the coarse sample geometry. Increasing the number of balls would assist in smoothing the
curve, but significantly increasing the computational time. The coarse sample configuration might further be responsible for the nearly constant maximum deviatoric
stress at a strain rate of 10−2 s−1 between 0.3% and 1.0% axial strain. During this strain interval the development of the shear surface takes place. Only after its full
development was strain softening recorded.
The samples for two separate strain rates after accumulated axial strains of 1.9% and 3.5% are shown in Figure 2. A close look shows a number of fractures in the
sample that was strained at the lower strain rate (Fig. 2a), whereas the other sample shows the growth of a distinct shear zone (Fig. 2b). This behavior was also
documented by Rist and Murrell (1994) and
Page 343
Figure 2. Samples after an axial compression of 3.5% (a) and 1.9% (b) axial strain, respectively. (a)
Figure 3. Parallelbond parameters at various axial strain rates for two different temperatures.
represent the difference between a more ductile and a brittle behavior. The first sample showed dense cracking, whereas the latter displayed axial splitting with an
accompanying shear zone.
3.2 Strain rate effect
As shown in Figure 2, the samples behave differently at different applied strain rates. The parallelcontact parameters (Table 1) as shown in Figure 3 for the two
temperatures under investigation were found to best
Figure 4. Effect of temperature on the parameters for the parallelbond contact model at various axial strain rates.
represent the laboratory tests and account for changing sample response.
The stiffness and the strength parameters increase as the strain rate increases for both temperatures. As a result, the strength at higher strain rates is higher and the
peak resistance is achieved at a smaller axial strain. Both stiffness parameters Ec and seem to increase linearly with the logarithm of the strain rate. The increase is,
however, less pronounced at colder temperatures. This trend is in accordance with the change in the crack development described in Rist and Murrell (1994). The
authors showed that at colder temperatures the development of dense cracking prevails even at slow strain rates. On the other hand the cracking behavior changes for
warmer temperatures, where only sparse cracking occurred at low strain rates. This change in the crack development results in different changes in the stiffness of the
samples for the two temperatures.
The strength parameters show a slightly different trend. Both parameters increase parallel with increasing strain rate. In addition, the increase between 10−3 and
10−2 s−1 seems to be lower than the increase between 10−4 and 10−3 s−1. This trend would imply that a constant value might be reached at very high strain rates.
Unfortunately no data was available to confirm this trend. Nevertheless, it seems reasonable that there might be no further change in strength at very high strain rates,
since the sample fails in a brittle manner at small strains.
3.3 Temperature effect
The data presented in Figure 3 can also be shown as a function of the temperature (Fig. 4). The trends are
Page 344
similar to the ones described previously. The stiffness parameters change differently for the different strain rates. The increase in the strength parameters, on the other
hand, is independent of the strain rate. These trends need to be used with caution, since data are only available at two temperatures.
To some extent the trends can be extended to a temperature of 0°C, where the strength should be zero at the melting point of the ice. However, at temperatures
close to the melting point of the ice the behavior changes significantly, something that cannot be accounted for with the current model.
3.4 Confinement
The effect of the confinement could only be modeled for the lowest strain rate. The failure mechanism and the crack development at low strain rates is not influenced by
the confining pressure (e.g. Jones 1982). Therefore, only a slight increase in strength was noted. The PFC model showed an increase from 6.9 MPa at a confining
stress σ3 of 100 kPa to 13 MPa at 4.7 MPa confinement. Murrell and Rist (1994) report an increase from 6.5 MPa to 11 MPa for the same stress range at a strain
rate of 10−4 s−1 and temperature of −20°C. The increase in the test data for the same change in σ3 at a strain rate of 10−2 s−1 is much higher: 18 MPa to 32 MPa. The
numerical model only predicted an increase from 18 MPa to 24 MPa.
The discrepancy in the results can mainly be attributed to the change in sample behavior at higher strain rates. The numerical model used herein is not capable of
considering the change in behavior since the contact parameters are constant and independent of the changing confining stress. The laboratory test, however, showed
that the crack intensity changes as the confining stress increases at the same strain rate and temperature. New model parameters would therefore be necessary to better
represent the higher confinement.
4 CONCLUSION AND OUTLOOK
This paper presents the first steps in the use of discrete element modeling for ice and frozen soils. It was shown that the clump logic within PFC2D provided good
agreement of triaxial compression tests of polycrystalline ice. The parameters used to describe the parallelbonded material, however, varied with temperature and axial
strain rate.
The change in the crack density, as observed in laboratory investigations, could be represented by assigning different parameters for the changing boundary
conditions, such as axial strain rate and temperature. Samples after shearing also showed similar crack patterns to the original polycrystalline ice samples. However,
changes associated with confinement that result in a change of the mechanical response, in particular at higher strain rates, could not be captured.
The results presented herein are only the initial step towards the more complex modeling of frozen soils. The changes in the model parameter as a function of strain
rate are currently incorporated into the contact model. The stiffness and strength parameter will therefore change as a function of the local strain rate between the ice
particles. It is thought that constantly changing the model parameters will capture the change in the mechanical behavior that is observed during the different tests.
The next step will then be to include the solid particles and assigning different contact parameters. Further steps would include an extension into PFC3D and
temperature coupling. However, this modeling demands significant computer power that currently requires large computational times. Though it might not last very long
and the required CPU times are available to solve complex problems with constantly changing contact parameters (Munjiza & Latham 2002).
ACKNOWLEDGEMENT
The author wishes to acknowledge Namkak Cho from the UofA Geotechnical Centre for his help during the initial modeling steps. In addition the financial support from
the Kilam Trusts for the author’s current position is highly appreciated. Comments on the manuscript were kindly provided by Professor Dave Sego and Professor
Derek Martin.
REFERENCES
Adachi, T., Oka, F. & Poorooshasb, H.B. 1990. A constitutive model for frozen sand. Journal of energy resources technology—transaction of the ASME 112(3):208–
212.
Andersen, G.R., Swan, C.W., Ladd, C.C. & Germaine, J.T. 1995. Smallstrain behavior of frozen sand in triaxial compression. Canadian Geotechnical Journal 32(3):
428–451.
Andersland, O.B. & Ladanyi, B. 1994. An introduction to frozen ground engineering. New York: Chapman & Hall.
Anderson, D.M., Tice, A.R. & McKim, H.L. 1973. The unfrozen water and the apparent specific heat capacity of frozen soils. Proceedings of the Second International
Conference on Permafrost; North American Contribution: 289–295: Natl. Acad. Sci., Washington, D.C., United States.
Cho, N., Martin, C.D., Sego, D.C. & Christiansson, R. 2004. Modelling dilation in brittle rocks. In D.P.Yale, S.M.Willson & A.S.AbouSayed (eds), Proceedings of the
Sixth North American Rock Mechanics Symposium (NARMS). Gulf Rocks 2004, Houston, Texas, United States June 5–10, 2004: CDROM.
DerradjiAouat, A. & Evgin, E. 2001. A constitutive model for isotropic freshwater ice. Canadian Geotechnical Journal 38:818–827.
Page 345
Fish, A.M. 1991. Strength of frozen soil under a combined stress state. Ground Freezing 91, Beijing: 135–145. Rotterdam: A.A.Balkema.
Fish, A.M. & Zaretsky, Y.K. 1997. Temperature effect on strength of ice under triaxial compression. Proceedings of the Seventh International Offshore and Polar
Engineering Conference, Honolulu, Hawaii, United States, May 25–30, 1997: 415–422.
He, P., Zhu, Y. & Cheng, G. 2000. Constitutive models of frozen soils. Canadian Geotechnical Journal 37(4): 811–816.
Hivon, E.G. & Sego, D.C. 1995. Strength of frozen soils. Canadian Geotechnical Journal 32(2):336–354.
Hopkins, M.A. 2002. Using DEM to model Arctic sea ice. In B.K.Cook & R.P.Jensen (eds), Proceedings of the Third International Conference; Discrete Element
Methods: Numerical Modeling of Discontinua, Santa Fe, New Mexico, USA, September 23–25, 2002: 374–379: ASCE Geotechnical Special Publication No. 117.
Itasca Consulting Group, Inc. 2000. PFC2D—Particle Flow Code in Two Dimensions, Ver. 2.0. Minneapolis: Itasca.
Jones, S.J. 1982. The confined compressive strength of polycrystalline ice. Journal of Glaciology 28(98):171–177.
Munjiza, A. & Latham, J.P. 2002. Computational challenge of large scale discontinua analysis. In B.K.Cook & R.P.Jensen (eds), Proceedings of the Third International
Conference; Discrete Element Methods: Numerical Modeling of Discontinua, Santa Fe, New Mexico, USA, September 23–25, 2002: 5–10: ASCE Geotechnical
Special Publication No. 117.
Neaupane, K.M. & Yamabe, T. 2001. A fully coupled thermohydromechanical nonlinear model for a frozen medium. Computers and Geotechnics 28(8):613–637.
Rist, M.A. & Murrell, S.A.F. 1994. Ice triaxial deformation and fracture. Journal of Glaciology 40(135):305–318.
Schapery, R.A. 1997. Thermoviscoelastic constitutive equations for polycrystalline ice. Journal of Cold Regions Engineering 11(2):146–157.
Sinha, N.K. 1988. Crackenhanced creep in polycrystalline material—strainrate sensitive strength and deformation of ice. Journal of Materials Science 23(12):4415–
4428.
Soo, S., Wen, R.K. & Andersland, O.B. 1985. Finite element models for structural creep problems in frozen ground. In S.Kinosita & M.Fukuda (eds), Fourth
International Symposium on Ground Freezing, Sapporo, Japan, August 5–7, 1985: 23–28. Rotterdam: A.A.Balkema.
Sunder, S.S. & Wu, M.S. 1990. On the constitutive modeling of transient creep in polycrystalline ice. Cold Regions Science and Technology 18(3):267–294.
Vyalov, S.S. 1963. Rheology of frozen soil. Proceedings of the International Conference on Permafrost, Lafayette, Indiana, USA, November 11–15, 1963: 332–
339. Washington D.C., USA: National Academy of Sciences: National Research Council, Publication 1287.
Zyryanov, D., Divine, D. & Korsnes, R. 2002. Mechanical simulation of shore fast ice breakup: Kara and Laptev seas. International Workshop on SmallScale Sea
IceOcean Modeling (SIOM) for Nearshore Beaufort and Chukchi Seas, International Arctic Research Center Fairbanks, Alaska, USA, August 7–9, 2002.
Zyryanov, D.V. 2004. Columnargrained S2 ice contact model with failure. http://www.itascaudm.com/pages/contact.html (May 5, 2004).
Page 346
Page 347
Effects of intergranular adhesive force on behavior of granular material
S.Kato
Research Center of Urban Safety and Security, Kobe University, Kobe, Japan
T.Sakakibara
CRC Solutions, Ltd., Tokyo, Japan
H.Yoshimori
Kobe City Government, Kobe, Japan
5809 679 3
ABSTRACT: In this study, simulations of the true triaxial test for spherical granular material are carried out using the distinct element method analysis
in threedimensional state. In this analysis, we introduced a constant intergranular adhesive force that acts normal to the contact plane. The influence of
the intergranular adhesive force on the failure criterion and deformation are examined. From the analytical results, it is found that not only the failure
criterion but also deformation characteristic in shear process obey the Lade failure criterion.
1 INTRODUCTION
Unsaturated soil is an example of granular material in which intergranular adhesive force acts between particles. In unsaturated soil, soil water exits as meniscus at
contact points between particles. According to the meniscus, pressure difference, known as suction, is induced between pore air pressure and pore water pressure.
Suction causes intergranular adhesive force between particles perpendicular to the contact plane as shown in Figure 1, and it has influence on the complicated behavior
of unsaturated soil.
To study the behavior of unsaturated soil, many triaxial tests have been carried out. Based on the triaxial test results, constitutive models have been presented, but
some fundamental characteristics of unsaturated
Figure 1. Intergranular adhesive force induced by meniscus between two spherical particles.
soil are still not confirmed. For the compressibility of unsaturated soils, Alonso et al. 1990 proposed one equation in their constitutive model, in which the
compressibility index is treated as a variable for the change of suction. But there have been reports that the compressibility index is treated as a constant value
regardless of suction value if the additional induced confining pressure by suction is accounted for total confining pressure (Karube et al. 1986). For the strength
property, Fredlund et al. 1978 proposed one equation to represent shear strength for unsaturated soil as follows.
(1)
where τf=shear stress at failure on the failure plane, σ=normal total stress at failure on the failure plane, csat=cohesion under saturated state, =parameter concerning
to increase of shear strength with suction increase and ua, uw=pore air and water pressure respectively
In this equation, the internal friction angle is treated as a constant value from saturated state to unsaturated state. But there have been reported a triaxial test result for
unsaturated soil under constant suction, in which the internal friction angle increases with the increase of applied suction (Karube et al. 1986). The reason why these
different views are proposed is because there are some limitation and problems in the triaxial test for unsaturated soil.
Page 348
In many cases, the pressure plate method is used in triaxial test for unsaturated soil, in which a ceramic disk is used instead of the porous stone, to apply air pressure
to specimen. The ceramic disk has its own air entry value that limits the maximum applied air pressure, and it leads to the limitation of suction condition. Further, the
volume change of the specimen should be measured separately from measuring displacement of the drainage from specimen.
According to this condition, the double cell method is used in many cases. But in the double cell method, the volume of the inner cell will change for the lateral
pressure applied. We must correct the measurement value of the volume change of specimen for the lateral pressure change, and the obtained results contain some
errors. These limitation and problem have influence on the accuracy of the triaxial test for unsaturated soil and the comprehension for behavior of unsaturated soil.
We will not encounter in these limitation and problem mentioned above in simulation with DEM analysis. When we simulate true triaxial test for granular material with
adding the intergranular adhesive force, we are able to apply the arbitrary value of the intergranular adhesive force, which corresponds to the applied suction in the
triaxial test, and to know the volume of the specimen clearly.
In this study, by the distinct element method analysis in threedimensional state, simulations of the true triaxial test for spherical granular material is carried out. In this
analysis, the influence of meniscus water, which is mainly cause of the complicated behavior of unsaturated soil, is expressed by introducing a constant intergranular
adhesive force that acts normal to contact plane. The influences of the intergranular adhesive force on the failure criterion and shear deformation are examined to
understand the mechanical characteristics of unsaturated soil and the other granular materials with adhesion.
2 FAILURE CRITERIA FOR SOIL
Many failure criteria for soil have been presented, In this chapter, the failure criteria, which concern to our study, are briefly explained.
MohrCoulomb failure criterion is widely used in soil mechanics, and it has a shape of deformed hexagonal on the π plane as shown in Figure 2. Figure 3 shows
“Spatially Mobilized Plane (SMP)”, the direction of which is defined based on the principal values of stress tensor. Based on this plane, Matsuoka & Nakai 1974
proposed a failure criterion as follows. The normal and tangential components of the principal stress vector in principal stress space to the SMP are given as follows.
(2)
Figure 2. MohrCoulomb failure criterion plotted on the π plane.
Figure 3. Definition of the Spatially Mobilized Plane (SMP).
(3)

where a1, a2 and a3 are the direction cosines for the SMP given as follows.
(4)

Here I1, I2 and I3 are the invariants of principal stress defined as I1=σ1+σ2+σ3, I2=σ1σ2+σ2σ3+ σ3σ1, I3=σ1σ2σ3.
Page 349
Figure 4. The SMP failure criterion and the MohrCoulomb failure criterion plotted on the π plane.
Equation 5 is obtained, by taking the condition that the ratio of the normal component to the shear component becomes a constant value as the SMP failure criterion.
(5)
Eliminating the effect of confining pressure from the above equation derives the next equation.
(6)
Figure 4 shows the SMP failure criterion of Equation (6) and the MohrCoulomb failure criterion plotted on the π plane. The SMP failure criterion shows the same
shear strength at the triaxial compression and the triaxial extension stress state as those shown by the MohrCoulomb failure criterion, but in the other conditions, it
shows stronger shear strength.
Lade & Duncan 1975 carried out true triaxial tests for sand specimens under different stress paths on the π plane. Based on the obtained true triaxial test results,
they proposed the failure criterion shown as the next equation.
(7)
Figure 5. Lade failure criterion plotted on the π plane.
Figure 6. Comparison of Lade failure criterion with MohrCoulomb failure criterion.

where k3 is a constant that represents the shear failure state.
Figure 5 shows the Lade failure criterion of Equation (7) plotted on the π plane. As shown in this figure, the Lade failure criterion shows slightly different shape for
the different internal friction angles. Figure 6 compares the Lade failure criterion with the MohrCoulomb failure criterion on the condition that they have the same
internal friction angle at the triaxial stress state. From this figure, it is understood that the Lade failure criterion has bigger internal friction angle than the MohrCoulomb
failure criterion at the triaxial extension stress state.
Page 350
3 DEM ANALYSIS AND ANALYTICAL CONDITION
3.1 Outline of DEM analysis with intergranular adhesive force
The distinct element method (DEM) is one of the discontinuous corpora analysis method proposed by Cundall 1971. Unlike continuum analysis of the finite element
method or the boundary element method, it is suitable for analyzing the dynamic behavior of the granular material. In the DEM analysis, simple dynamic models (they
are generally the Voigt model and Coulomb’s friction rule) are introduced for normal and tangential directions to the contact plane at contact points and contacting
surfaces between the particles with the assumption that the particles are rigid. The independent equations of motion for every element are solved forwardly in the time
domain, and the interactions between particles and the deformation of particle aggregates are traced. This method has the merit that the necessary output data are easily
obtained, such as stress, strain, and rotation angle of some particles, and the setting of boundary conditions. In this study, we carried out the analysis with the distinct
element analysis program, PFC3D (Itasca 2003).
In the two dimensional analysis, the effect of meniscus water observed in unsaturated soil is expressed by introducing an intergranular adhesive force between
particles as follows. The intergranular force acting between contacting particle i and j is denoted as Pij. The x and y direction components, and the moment component
of the resultant force of the intergranular adhesive force at time t are given by the following equations considering that the direction of Pij is normal for the line direction
of the particle tangent plane.
(8)
(9)
(10)
where, are the x and y direction components, and the moment component of the resultant force of the intergranular adhesive force and αij
is the angle between the normal direction of the contact plane and the xaxis.
It is possible to introduce an intergranular adhesive force by deducting these components from each component of the total intergranular force except for the
intergranular adhesive force. By integrating these equations of motion for each particle by the Euler method with respect to time, the analysis was carried out.
Consequently, the solution becomes stabilized conditionally on the integral time increment. In the
Table 1. Parameters and material properties.
Particleparticle Particlewall
Normal stiffness kn (N/m) 9.0×10 9
1.8×1010

Shear stiffness ks (N/m) 3.8×108

6.0×108

Coefficient of friction 0.287 0.0

Density of particles 2700

ρ (kg/m3)
Figure 7. Rectangle specimen size used for analytical model.
three dimensional analysis, the similar procedure is applied.
Cundall 1971 proposed the following equation for the integral time increment.
(11)
where, m is mass of the disk particle, and k is the spring constant.
It has been found experientially that sufficient stability and accuracy can be ensured in the quasistatic problem at about 1/10∆t c, though Equation (11) is deduced
from the equation of motion for a singledegreeoffreedom system. But taking this equation as a standard, the integral time increment must be decided by trial and error.
In this study, ∆t=8/10∆t c is used. The parameters and the material properties necessary for the analysis are listed in Table 1.
3.2 Analytical con dition
A rectangle specimen of 4 meters height, and 2 meters width and deep was used for the analytical model as shown in Figure 7. The particles are drawn from a
Page 351
Figure 8. Stress paths on the π plane under confining pressures of 500 and 1000 kPa.
uniform distribution of radii with the specified upper limit of 10 m and lower limit of 0.05 m, and placed randomly inside the specimen. In total, 5438 particles are
placed inside the specimen, the void ratio of which is about 0.4. The six sides of the rectangle specimen are surrounded by the rigid wall elements that correspond to
the loading platens.
In this study, firstly, without the intergranular adhesive force, the specimen was compressed isotropically under confining pressures of 500 and 1000 kPa.
Then, the intergranular adhesive force of 5000 N or 10000 N was introduced between particle contact points. The same specimens are used for different stress
paths on the π plane, and the void ratio of the specimen and the state of particle distribution before the shear process are the same state for the each series of different
stress paths under constant mean principal stresses on the π plane. Afterwards, the shear process of the true triaxial test was simulated under the constant mean
principal stress. In the shear process, two wall elements in height direction were moved perpendicularly at a constant rate of 10 cm/s in order to apply the maximum
principal stress for the height direction. And the other pairs of lateral wall elements were controlled to keep the decided stress path and constant mean principal stress
condition.
Figure 8 shows the stress paths on the π plane under confining pressures of 500 and 1000 kPa. In one stress path, one kind of Lode parameter θ was kept constant
at 60, 75, 90, 105 and 120 degrees.
4 TEST RESULTS AND DISCUSSIONS
4.1 Effects of intergranular adhesive force on failure criterion
Figures 9 & 10 show the stressstrain relations for triaxial stress state and triaxial extension state under
Figure 9. Stressstrain relations for triaxial compression stress state under different intergranular adhesive forces.
Figure 10. Stressstrain relation for triaxial extension stress state under different intergranular adhesive forces.
confining pressure of 1000 kPa. In these results, the principal stress difference showed bigger peak value with the increase of the intergranular adhesive force, and then
decreased. And the more expansive behavior
Page 352
Figure 11. Relation between maximum principal stress differences and means principal stress.
Table 2. Adhesion and internal friction angle obtained.
Adhesion (N) c (kPa)
0 4.9 20.37
5000 −31 40.70
10000 −140 42.18
was observed with the increase of the intergranular adhesive force. These results are similar with that observed in the triaxial compression test for unsaturated clay
specimen under different constant suctions (Karube et al. 1986). The intergranular adhesive force resists for the slippage at contact points between particles, and affects
on these observed behaviors. The increase of shear strength is deduced, but the increase of volume expansion is not deduced simply from the effect of the intergranular
force.
Figure 11 shows the relation between maximum principal stress differences and mean principal stress. The dots show simulation results and the solid lines show the
failure lines for each intergranular adhesive force. Table 2 shows the adhesion and the internal friction angle obtained for the each failure line. From these
Figure 12. Maximum shear stress states on the π plane.
results, we find that the increase of intergranular force has influence not only on the adhesion but also on the internal friction angle. As mentioned in Section 1, in the
triaxial test results for unsaturated soil, the influence of the suction was not clear. But based on these results, we are able to state that the increase of the intergranular
adhesive force causes the increase of the internal friction angle.
Figure 12 shows the maximum shear stress states on the π plane. In this figure, the solid lines shows the Mises failure criterion, and the broken line and the one
dotted lines show the Lade and the MohrCoulomb failure criterion, which correspond to the internal friction angles at triaxial stress state, respectively. Figure 13
shows the normalized data shown in Figure 12 by changing the scale that the results of 5000 N and 10000 N of internal adhesive forces have the same σz intersect as
that for the result of no internal adhesive force. In this figure, The solid line and the one dot line show the Mises and the MohrCoulomb failure criterion, and the broken
line and the two kinds of two dotted line show the Lade failure criterion for different internal friction angle. From these results, we find that the Lade failure criterion is
able to explain the failure stress state, regardless of the value of the intergranular adhesive force.
4.2 Effects of intergranular adhesive force on residual stress state and shear deformation
Figures 14 & 15 show the residual stress states for the case of no intergranular adhesive force and 10000 N
Page 353
Figure 13. Comparison of normalized maximum shear stress states on the π plane.
Figure 14. Residual stress states for the case of no intergranular adhesive force.
of intergranular adhesive forces. As shown in Figures 9 & 10, the residual stress state is not clear in the stressstrain relations, but we decided it as the stress state when
the axial strain is 20%. In these figures,
Figure 15. Residual stress states for the case of 10000 N of intergranular adhesive forces.
the solid line, the broken line and the dotted line show the Mises, the SMP and the MohrCoulomb failure criterion respectively. From this figure, we find that the SMP
failure criterion explains the residual stress state well. This result corresponds to the result of true triaxial test for normally consolidated clay specimen (Shibata &
Karube 1965), in which the failure stress states are plotted on the π plane as a circumscribed curve for the MohrCoulomb failure criterion in the similar way as the
SMP failure criterion does as shown in Figure 4.
Figure 16 shows the counters of equivalent shear strain on the π plane. From the origin to outside direction, the each curve shows counter of shear strain from 0.003
to 0.01 by 0.01 division respectively. The dotted curve and the solid curve show the result of no intergranular adhesive force and 10000 N of intergranular adhesive
forces respectively. In the case of no intergranular adhesive force, the solid curves are near to concentric shape, but in the case of 10000 N of intergranular adhesive
forces, the dotted curves are deformed from the concentric shape.
Figure 17 compares the stress points at same shear strain near to the failure state with the Lade failure criterion. From this figure, we find that the stress points
distribute around the Lade failure criterion. From the results shown in Figures 16 & 17, we should conclude that the Lade failure criterion explains the counters of shear
deformation, and dominate the shear deformation of expansive granular materials.
Page 354
Figure 16. Counters of equivalent shear strain on the π plane.
Figure 17. Comparison of stress points at same shear strain near to the failure state with the Lade failure criterion.
5 CONCLUSIONS
By the distinct element method analysis in threedimensional state, simulations of the true triaxial test for spherical granular material was carried out. In this analysis, the
influence of meniscus water, which is mainly cause of the complicated behavior of unsaturated soil, is expressed by introducing a constant intergranular adhesive force
that acts perpendicular to the tangential plane at contact point. The influences of the intergranular adhesive force on the failure criterion and shear deformation are
examined. The obtained results are summarized as follows.
1 The internal friction angle and the adhesion for triaxial stress state increased with the intergranular adhesive force. This result corresponds to the triaxial test results
obtained for unsaturated soil, and shows the effectiveness of introducing the intergranular adhesive force for DEM analysis to study the mechanism of unsaturated
soil and the other granular materials with cohesion.
2 The Lade failure criterion on the π plane explained well the failure stress state regardless of the intergranular adhesive force. This result means that the Lade failure
criterion is able to use the failure criterion for unsaturated soil and the other granular materials with cohesion.
3 The residual stress points on the π plane distribute around the SMP failure criterion. This result means that the SMP failure criterion is able to use the contractive
material like normally consolidated clay.
4 The counters of shear strain show the similar shape with the Lade failure criterion at near to failure state. This result shows that the Lade failure criterion should
dominate the shear deformation of expansive granular materials with cohesion.
REFERENCES
Alonso, E.E., Gens, A. & Josa, A. 1990. A Constitutive Model for Partially Saturated Soils. Geotechnique 40(3): 405–430.
Cundall, P.A. 1971. A computer model for simulation progressive, large scale movement in blocky rock system. In Proc. of the ISRM Symp., Nancy, France, Proc.2, pp.
129–136.
Fredlund, D.G., Morgenstern, N.R. & Widger, R.A. 1978. The Shear Strength of Unsaturated Soils. Can. Geotech. J. 15(3):313–321.
Karube, D., Kato, S. & Katsuyama, J. 1986. Effective stress and soil constants in unsaturated kaolin. Japanese society of civil engineering 370/III5: pp. 179–188 (In
Japanese).
Lade, P.V. & Duncan, J.M. 1975. Elastoplastic StressStrain Theory for Cohesionless Soil. ASCE GE div. pp. 1037–1153.
Matsuoka, H. & Nakai, T. 1974. Stressdeformation and strength characteristics of soil under three deferent principal stresses. Japanese society of civil engineering
232: 59–70.
Shibata, T. & Karube, D. 1965. Influence of the variation of the intermediate principal stress on the mechanical properties of normally consolidated clays. 6th
ICSMFE, Vol 1, pp. 359–363.
Page 355
Development of compressive failure analysis of concrete specimen by 3
dimensional Distinct Element Method
D.Haraki, H.Shiraishi & S.Katsuki
Dept. of Civil and Environmental Engineering, National Defense Academy, Japan
5809 679 3
ABSTRACT: This paper discusses about the effect of MohrCoulomb’s parameters on failure behavior of concrete material using the Distinct Element
Method. The failure behavior of plane concrete, which is model by the particle elements arranged in a couple of parallel plane, is analyzed using
several constitutive law conditions. Two types MohrCoulomb condition are used, and various cohesion parameter and friction angles are combined in
some simulation condition. It is pointed out that 5 types failure modes appear in a global viewpoint depending on the MohrCoulomb condition.
1 INTRODUCTION
The Distinct Element Method, so called DEM, simulates well the initiation and/or development of crack of concrete material (Nanagawa et al. 1995), because the
method is originally based on the aggregation of separated pieces. Especially, in case of simulation of concrete structures subjected to an impact loading, the method
may simulate well the behavior of fragments, which cause secondary damage to the human or important devices near the structure.
Many research works, in which DEM is applied in failure of the concrete or the reinforced concrete structures, have been carried out. For instance, Meguro 1993
proposed an extended Distinct Element Method and applied it to failure analysis of concrete structure as a large deformation analysis.
Shimoda and others (Shimoda et al. 1992, 1993 & 1997, Suzuki et al. 1997a & 1997b) applied it to failure analysis of concrete check dam subjected to an impact
load of huge rock contained by the debris flow. Nakata et al. 1998, traced an impact experiment of RC beam by using DEM. Morikawa et al. 1995, discussed the
parameters used in the DEM in case of RC structure failure analysis. Beppu et al. 2004, analyzed the failure of concrete specimen subjected to highspeed load in
conjunction with a confined pressure.
According to those research works, it was clarified that the constitutive law of local springs affects very delicately to the global failure of concretes in the DEM. This
paper discusses about the effect of MohrCoulomb’s law parameter on the failure of concrete specimen fundamentally.
2 ANALYSIS METHOD
2.1 Formulations
Concrete is modeled as the aggregation of fine particles and contact springs. The equations of motion are solved by using the finite central difference (JSCE 1993),
which is more stable than Euler’s method, however, backward difference is using in a term of velocity.
(1a)
(1b)
where, u=displacement vector, (∙)=differentiation with respect to time, and t=time.
This form is solved simultaneously equation of motion. Then, the displacement vector of next step is given by
(2)
where, M=mass matrix, D=dumping matrix, K=stiffness matrix, F=external force vector.
Relation between load and deformation is modeled by Popovics’ form (Popovics 1971 & Katsuki et al. 2002) as shown in Figure 1.
(3)
Page 356
Figure 1. Constitutive law of normal spring.
Figure 2. Critical condition.

where, P=normal spring force, δ=deformation of normal spring, Pmax=strength of normal spring, δmax=maximum deformation of normal spring, n=nonlinear parameter.
The contact spring against tensile force resist until one tenth of compressive strength. Tangential spring model is linear, but has critical condition of MohrCoulomb’s
law.
Figure 3. Arrangement of elements.
Two types MohrCoulomb’s law are compared with each other. The first is onestep model as shown in Figure 2a. The second is twostep model as shown in Figure
2b in which the zero friction angle is given in the region that spring force exceeds half of its compressive strength.
Formulation for onestep model are,
Ps=k sδs
(4a)
(4b)
For twostep model are,
Ps=k sδs
(5a)
(5b)
(5c)
where, Ps=tangential force works on elements, k s=tangential spring constant, δs=deformation of tangential spring, C=cohesion, =friction angle, Pn =normal force

works on elements, Pmax=compressive strength of normal spring.
2.2 Analysis model
To emphasize the effect of constitutive law, particle elements are arranged during two parallel plane as shown in Figure 3a, b, and top plane moves downward
constantly.
Page 357
Table 1. Analysis condition.
Spring constant Damping ratio Critical compressive Critical tensile Pmin Nonlinear parameter n Cohesion C (N) Friction angle

h Pmax (N) (N)
Normal k n Tangential k s
(N/mm) (N/mm)
4 0.05 5.89×10 Before peak 1.522 After 15, 30, 45, 60, 75, 15, 20, 25, 30, 35
7.9×10 1.6×104 5.89 ×102
peak 0.95 90
Analysis parameters (Abe 1996) are shown in Table 1. Five types of friction angle, i.e. 20°, 25°, 30° and 35°, and six types of cohesion, i.e. C= 15 N, 30
N, 45 N, 60 N, 75 N, and 90 N, are combined each other.
3 ANALYSIS RESULT
3.1 Onestep model of MohrCoulomb’s law
3.1.1 Relation between load and deformation
Relations between load and deformation are shown in Figure 4a, b. Figure 4a shows friction angle effect on the elastic plastic curve with respect to constant cohesion
C=60 N. Figure 4b shows cohesion effect on the elastic plastic curve with respect to constant friction angle
The strength increases as cohesion and/or friction angle increase. These elastic plastic behaviors can be classified into 3 different shapes in broad view point, such
that Mode1: the resistance descend gradually after yield strength, Mode2: the resistance descend rapidly after yield strength, Mode3: the resistance asymptotically
increase until δ=0.5 mm.
All analysis results are categorized with respect to combinations of cohesion and friction angle as shown in Table 2. Elastic plastic Mode1 appears in case of low
friction angles and 20°. On the other hand Mode3 appears in case of combination of large friction angle and cohesion. The rest combinations are related
with Mode2.
3.1.2 Progress of failure condition
Figure 5 shows progress of failure conditions in representative combinations of cohesion and friction angle. Figure 5a shows analysis result associated with C=60 N,
(Mode1). In the progress of limit state of contact springs, at δ=0.04 mm, there are continuous diagonal line of ×marks which means that normal spring of
horizontal contact reaches critical tensile condition and its force is zero. This condition relates to descendent of stiffness as shown in Figure 4a. At δ=0.06 mm, the
region with ×marks spread widely. At δ=0.13 mm, a couple of diagonal lines, marked by •, appear. The symbol • means the
Figure 4. Relation between load and deformation.
Table 2. Combination of cohesion and friction angle and categorized Mode (Onestep model).
Cohesion C Friction angle
15° 20° 25° 30° 35°
15 N Mode1 Mode1 Mode2 Mode2 Mode2
Page 358
Figure 5. Progress of failure condition. •=Coulomb’s limit state; ×=Tensile limit state of horizontal connect; ◊=Tensile limit state of
diagonal connect.
Page 359
tangential spring force reaches MohrCoulomb’s condition in diagonal contact. This deformation δ=0.13 mm is associated with the maximum resistance of elastic plastic
curve as shown in Figure 4a. This failure lines maintains constantly until δ=0.37 mm. This fact shows that forming of cross line of local failure of MohrCoulomb
condition initiates the failure of this specimen. The figure of most right bottom of Figure 5a shows the displacement vectors of all elements. It is clearly found that the
specimens are divided into four regions bounded by the failure lines marked by symbol •. Therefore, the line of slight gap along the local failure line can be observed in
the figure of elements arrangement in Figure 5a at δ=0.5 mm.
Figures 5b, c are related with Mode2. It is very clear from the elements arrangements of Figure 5b that the specimen is cleaved vertically. In case of C=60 N,
(see Fig. 5b), until δ=0.06 mm the local spring limit state conditions are very similar to those of Figure 5a. However, at δ=0.23 mm some diagonal local
contact of the left part aggregation reached the tensile limit state which is marked by symbol ◊. Then the left part aggregation starts to leave horizontally from the rest
part aggregation. This deformation, i.e. δ=0.23 mm, is exactly corresponding to the maximum strength of elastic plastic curve shown in Figure 4a. This cleave behavior
is also very clear from the displacement vector of Figure 5b. Additionally, a similar progress is found in Figure 5c.
Figure 5d shows the results associated with C=60 N, (Mode3). In this case, at δ=0.04 mm and δ=0.06 mm, local spring limit state condition is also
very similar to Figure 5a. But, any limit state between diagonal contact does not appear at all until δ=0.50 mm. The horizontal contact limit state marked by × only
spread whole area of specimen homogenously. Whereas, the local contact limit state distribution is homogenous, the element movements are separated into four regions
clearly.
3.1.3 Progress of contact force at local connection
Figure 6 shows the progress of contact force at representative locations as shown in Figure 6e. Numbers 1–5 correspond to the typical points of elastic plastic
behavior defined in the Figure 4.
In case of Mode1, i.e., C=60 N, (see Fig. 6a), when the normal spring force at connection
Figure 6. Progress of contact forces at local connections.
Page 360
B reached on tensile limit at elastic plastic behavior step 2 the situations of connections A and D are in the elastic region. After losing the tensile force of normal spring
of B, at elastic plastic behavior step 3, the force combinations of A and D move toward the MohrCoulomb’s limit state on almost the same line. And after getting to the
MohrCoulomb’s limit state line, the force of D stays on the MohrCoulomb’s line, but force of connection A backs to the elastic region. In case of Mode2 (see Figs
6b, c), forces of A and D starts toward the limit state line along same way from step 3. But on the way to step 4, force combination of D apart from A and changes its
direction rapidly to the MohrCoulomb’s limit line.
Once the local force combination of D arrived at the MohrCoulomb’s limit line at step 4, the forces decrease suddenly and get to the tensile limit line at step 5. At
the same time, the forces of A decrease very quickly along its incremental process. It is found from this fact that the bifurcation of local limit state between connection,
e.g., connections A and D, induce the brittle elastic plastic behavior of Mode2.
In case of Mode3 (see Fig. 6d), forces of connections A and D stay near the compressive limit state line from step 4 to step 5 constantly and no force reached at
any limit states.
3.2 Twostep model of MohrCoulomb s law
3.2.1 Relation between load and deformation
Figures 7a, b show relation between load and deformation associated with the twostep model of MohrCoulomb’s law. Figure 7a shows the effect of friction angle on
the elastic plastic curves and Figure 7b shows effect of cohesion in comparison with the onestep model.
There is no difference between onestep and twostep mode of MohrCoulomb’s law in case of Those curves are different from any shapes of Mode1,
2 or 3 in Figure 4. Therefore, Mode4 and Mode5 are given to those curves, respectively. In Figure 7b, Mode4 and Mode5 are also found.
Table 3 shows the category of these modes in terms of friction angle and cohesion combination. When the twostep model of MohrCoulomb’s law is adopted,
most are classified into Mode5, which are related with Mode3 or Mode2 in the case of the onestep model. Some classified into Mode4, which are relate to
Mode2 in case of the onestep model.
Figure 8 shows the elastic plastic behavior of specitnen in case of Mode4 and Mode5. Figure 9 shows the progress of local contact forces of Mode4 and Mode
5.
From the figure of local limit state distribution associated with δ=0.39 mm in Figure 8a, it is found that
Figure 7. Effect of twostep model on relation between load and deformation.
Table 3. Combination of cohesion and friction angle and categorized Mode (Twostep model).
Cohesion C Friction angle
15° 20° 25° 30° 35°
Mode4 has similar failure pattern to Mode2 as shown in Figure 6c. However, the spring force of both connections A and D are on different limit state lines, i.e. D is
on first step MohrCoulomb line and A is on the second line.
In case of Mode5 (see Figs 8b, c), diagonal connection local limit state are concentrated at the center of specimen, and specimen is separated into four
Page 361
Figure 8. Effect of twostep model on progress of failure condition. •=Coulomb’s limit state; ×=Tensile limit state of horizontal connect;
◊=Tensile limit state of diagonal connect.
Page 362
parts very clearly from the view point of element displacement behaviors. In this case (see Fig. 9c), the local spring forces of connections A, D and E are all on the limit
line of second step MohrCoulomb’s law. In other words, no local connection is on or near the compressive limit.
4 CONCLUSION
This paper investigated an effect of MohrCoulomb’s parameters on failure behavior of concrete specimen using Distinct Element Method as fundamental study of
DEM. As a result, it is shown that slight difference on combinations of cohesion parameters and friction angles have much sensitively on elastic plastic behavior in
concrete specimen. Especially, elastic plastic behave much ductile in case of twostep model and lower friction angle.
REFERENCES
Abe, K. 1996. Spring Constants of distinct element in continuum analysis. Proc. of JSCE, No. 543/I36, pp. 83–90 (in Japanese).
Beppu, M., Sonoda, Y. & Ohno, T. 2004. A study on the capability of distinct element method to elastic plastic analysis of static and rapid speed triaxial compressive test
for concrete specimens. Proc. of structural engineering, Vol. 50A, (in Japanese).
JSCE, 1993. Impact behavior and design of structures. pp. 25–33 (in Japanese).
Katsuki, S., Kuroki, Y., Ishikawa, N. & Ohta, S. 2002. Dynamic elastic plastic behavior of concrete filled steel tube beam subjected to high speed loading. Proc. of JSCE,
No. 696/I58:61–76 (in Japanese).
Meguro, K. 1993. Application of the distinct element method to fracture analysis of concrete structures, Concrete Journal, 31(5):5–17 (in Japanese).
Morikawa, H., Sawamoto, Y. & Koshika, N. 1995. Fracture analysis of concrete structures by the distinct element method, J. Struct. Constr. Eng., AIJ, No 437:127–135
(in Japanese).
Nakata, Y., Masuya, H., Kjikawa, Y. & Morishima, Y. 1998. Analysis of impact behavior of reinforced concrete beam by the distinct element method. Proc. of
Structural Engineering, Vol. 44A:1761–1771 (in Japanese).
Nanagawa, M., Jiang, Y. & Esaki, T. 1995. A new approach of modeling generation and progress of cracks in discontinuous rock masses by using distinct element method,
J. of JSECE, No. 543/I36:83–90 (in Japanese).
Popovics, S. 1971. Factors affecting the elastic deformations of concrete, In Proc. of the Int. Conf. on Mechanical Behavior of Materials, Kyoto, Japan, August
1971, pp. 172–183.
Shimoda, Y., Suzuki, S., Ishikawa, N. & Furukawa, K. 1992. Impact response analysis of mortar cantilever by the distinct element method, Proc. of Structural
Engineering, Vol. 38A:1467–1476 (in Japanese).
Shimoda, Y., Suzuki, S., Ishikawa, N. & Furukawa, K. 1993. Impact failure analysis of concrete check dam by distinct element method, Proc. of JSCE, No. 480/I21:97–
106 (in Japanese).
Suzuki, S., Ishikawa, N., Furukawa, K., Mizuyama, T. & Ishikawa, Y. 1997. Impact failure analysis of check dam wing reinforced by steel bar using distinct element
method, Proc. of Structural Engineering, Vol. 43A: 1555–1566 (in Japanese).
Suzuki, S., Ishikawa, N., Furukawa, K., Mizuyama, T. & Ishikawa, Y. 1997. Dynamic behavior of concrete check dam wing model with longitudinal reinforcement under
high speed loading and failure analysis by distinct element method, Proc. of JSCE, No. 567/VI35:141–160 (in Japanese).
Page 363
Experimental validation of modeling irregular particle shapes using DEM
A.M.Sallam, A.K.Ashmawy & B.D.Runkles
University of South Florida, Tampa, Florida, USA
679 3
ABSTRACT: This paper presents an experimental verification of the ability of DEM to model angular particles that have been created using
“overlapping rigid clusters.” The experimental setup consists of a piston that can smoothly slide into a box to disturb model sand grains with pre
specified initial positions. Snapshots were taken at different stages of the experimental test to monitor the rotation and translation of each grain. A
numerical simulation of the experimental test was carried out using PFC2D and the results were compared quantitatively and qualitatively. Translations
and rotations of the laboratory manufactured modelgrains were measured experimentally and compared with the numerical modeling results. The
study revealed that the DEM can successfully capture the mechanical behavior of angular assemblies at small displacement ranges. The difference
between experimental and numerical simulations increases as a function of displacement, but significant variability also exists among experimental tests
with identical initial conditions.
1 INTRODUCTION
The Discrete Element Method (DEM) is a numerical tool capable of describing the mechanical behavior of discrete assemblies. The method was introduced by Cundall
and Strack (1979) to model granular assemblies within the context of geotechnical engineering. DEM has the capability of modeling the material at the microscopic level
and capture all the phenomena that pertain to the particulate nature of granular materials. The DEM solves the dynamic equilibrium equations for each element
subjected to either body or boundary forces. The method is capable of analyzing multiple interacting bodies undergoing large displacements and rotations. The discrete
element scheme, proposed by Cundall and Strack (1979), is based on the idea that a small enough time step should be chosen to ensure that, during a single time step,
disturbances do not propagate from any disc further than its immediate neighbors (Cundall & Strack 1979).
Following the pioneering work of Cundall and Strack (1979), numerous researchers have evaluated and improved the technique. Some of the corresponding fields
of study are: (i) fundamental investigation and application of the DEM in granular soils, cohesive soils, and powders, (ii) rock mechanics, (iii) experimental validation of
the DEM, (iv) modeling different shapes rather than discs/spheres, (v) developing improved contact models, (vi) coupled modeling methods, and (vii) largescale and
industrial applications.
1.1 Modeling noncircular particles
In the DEM scheme proposed by Cundall and Strack (1979), the granular particles were modeled as discs in 2D simulations and as spheres in 3D simulations.
Although computationally very efficient, circular particles have an inherent tendency to roll, which reduces their interlocking resistance. Polygonshaped particles were
proposed by Barbosa and Ghaboussi (1992) and Matuttis et al (2000). A more realistic representation of the particle shape can be achieved using polygonshaped
particles; however the method is computationally intensive. Ting et al (1993) and Ng (1994) presented ellipseshaped particles for 2D simulations, while Lin and Ng
(1997) introduced ellipsoidshaped particles for 3D Simulations. Ellipses and ellipsoids have a smaller tendency to rotate, but it is not a close representation of actual
particle shapes. As a result, angularityinduced dilation and particle interlocking are suppressed. Potapov and Campbell (1998) proposed ovalshaped particles, which
are computationally more efficient than the ellipses. Favier et al (1999) presented an axisymmetricalshaped particle, which can model a variety of shapes. However,
highly angular particles cannot be easily generated.
1.1.1 The concept of “clustering”
Jensen et al (1999) introduced the clustering technique by combining a number of sphericalshaped particles in a semirigid configuration, which is a better
representation of natural soil particles. Any number
Page 364
Figure 1. Disk elements inscribed within a particle outline to capture the shape.
of particles can be linked together to form a cluster as long as they rotate and translate as a rigid body. The interparticle contacts within the cluster are constrained to be
linearelastic with a high stiffness. Thomas and Bray (1999) presented a different idea for clustering by imposing kinematic restrictions on discs within a cluster to
prevent relative translations and rotations. Although these techniques represent some improvement over earlier methods, the simulated particle outlines did not resemble
those of actual particles. In addition, a substantial increase in computational time was incurred due to the high contact stiffness required within the cluster.
1.1.2 Overlapping rigid clusters (ORC)
Ashmawy et al (2003) proposed the use of overlapping rigid clusters (ORC) to accurately simulate angular particle shapes. A set of subroutines was introduced to
PFC2D using Itasca’s softwarespecific programming language, FISH. The builtin clump logic command was instrumental in formulating the new method. The clump
logic allows several disk elements to move as a rigid body without detecting contacts or calculating contact forces, which increases computational efficiency. Further
description of the clump logic is given in Itasca (1999). The ORC method relies on “clumping” a number of overlapping disc elements, such that the resulting outline
coincides with, and is almost identical to the actual particle’s outline.
First, twodimensional outlines of a series of particles are obtained using a digital microscope or scanner. Overlapping circles are then inscribed within the outline to
capture the shape, as shown in Figure 1. The number of overlapping circles that can accurately represent the actual particles depends on: (i) the degree of non
uniformity in the original particle shape and angularity, (ii) the desired level of geometric accuracy, and (iii) the required computation time limit. Typically, ten to fifteen
disks will adequately capture the “true” shape of a particle.
Due to overlapping, the density of each disk must be scaled to ensure that the mass of the particle remains proportional to the area, regardless of the number of
Figure 2. Random assembly of eight circular particles (left) and the transformed equivalent particles (right).
discs or level of overlap. An approximate method was proposed to scale the density as follows:
(1)
where ρd is the density of the disk elements, Ap is the area of the particle, ΣAd is the sum of the areas of the disk elements, and ρp is the density of the soil particle.

Equation (1) ensures that the mass of the particle is proportional to the area but it neither guarantees that the moment of inertia nor the center of mass of the ORC
particle is identical to that of the actual particle. The ORC method is implemented within PFC2D by means of a series of FISH functions that convert a particle
assembly of discs into their corresponding angular particles as follows:
1. Particles are automatically generated as discrete circular elements using any of the builtin particles generation techniques
2. The shape conversion algorithm is invoked to transform each circular particle into an angular equivalent through replacement with a corresponding set of disc
elements, selected randomly from a library of available particle shapes
3. A random rotation between 0 and 360° is applied to each transformed particle to avoid preferred orientations and subsequent system anisotropy.
Figure 2 shows a random assembly of circular particles that have been transformed to the corresponding angular shapes using the ORC method.
2 OBJECTIVES
Experimental verification of the ORC method is essential in order to evaluate the accuracy and significance of the new procedure. An experimental validation program
was carried out by: (i) designing and building an experimental setup that allows tracking translations and rotations of modelgrains resulting from an external disturbance,
(ii) numerically simulating the
Page 365
experimental setup using PFC2D, with the same initial and boundary conditions and material properties, (iii) tracking the positions and rotations of modelgrains in both
the experimental and numerical tests, and (iv) qualitatively and quantitatively comparing the experimental and numerical results.
3 EXPERIMENTAL PROGRAM
Fraser River sand was chosen as a natural angular material for the experimental verification. Only twodimensional modelgrains were considered for this study in order
to accurately track the particle motions. The scanned outlines of ten different Fraser River sand grains were resized and plotted on top of 25 mm thick wood. A
thickness of 25 mm and an equivalent diameter of 50 mm were chosen as the average disc dimension for all particles. For a particle with an original area of Ap (in
mm2), the distance scaling factor, λr, was therefore calculated from:
(2)
3.1 Material
Modelgrains were manufactured from maple wood, which has reliable strength, density, and stiffness. Typical maple wood properties are summarized in Table 1. A
scroll saw with hard carbon steel blade was used to cut the modelgrains by following the shape outlines plotted on top of the maple wood sheet. The upper and lower
edges were smoothed using fine sand paper to prevent edge smearing or tearing. For each of the ten particle shapes, fifteen particles were manufactured for use in the
testing program. An alphanumeric labeling scheme was used in order to facilitate the recognition of different modelgrains during the snapshots. For each particle, the
center of mass and a direction marker were labeled in order to facilitate the tracking procedure during comparison, as shown in Figure 3.
3.2 Test setup
The experimental boundary conditions were defined by a box and a piston, where the modelgrains were assembled (Fig. 4). The piston was 25×50×625 mm in
dimensions and was covered with the same laminated plastic sheets on all its sides.
A stainless steel ruledscale was attached to the top of the piston to measure its displacement as it penetrated into the box. A cast acrylic sheet was attached to the
top of the box to prevent the particles from “popping up” during the test. The cast acrylic sheet
Table 1. Maple wood properties.
406.6
Density, Kg/m3
Specific gravity 0.659
Bending strength, MPa 115
Modulus of elasticity, MPa 14100
Hardness 7290
Figure 3. Modelgrains samples.
Figure 4. The box with the piston.
was thick enough to prevent any deflection that would have caused friction with the top of the modelgrains. A digital camera with a 0.79 Mega pixel resolution was
used to capture the snap shots. The camera was centered and leveled at approximately 1 m above the box, and was supported by means of a wooden frame. Two
ruled Lsquares were used simultaneously to locate the center point of each modelgrain for initial setup. The modelgrains were distributed randomly within the box
and the two Lsquares were used to determine the X and Y coordinates for each modelgrain. The initial direction of each modelgrain was then adjusted and verified
using the two Lsquares.
Page 366
Figure 5. The initial condition of the experimental setup.
Figure 6. Distortion calibration using graded transparent sheet.
The initial condition of the modelgrains is shown in Figure 5.
3.3 Image distortion calibration
In order to accurately measure the location of the particles at various stages of the test, it is necessary to calibrate the captured image. A 12.5×12.5 mm graded
transparent sheet was, therefore, used to evaluate image distortion. The sheet was placed under the cast acrylic sheet on top of a few modelgrains as a support during
calibration as shown in Figure 6. The image was analyzed by counting the number of pixels within the center square, corner squares, and middle edge squares. The
squares sides’ lengths were calculated and compared to the lengths of the center square. In addition, an image of the initial condition of the modelgrains in the
laboratory was digitized. The X
Figure 7. LoadDisplacement curves for particle # 8Z.
Table 2. Contact normal stiffness results.
Particle Kn , N/m
Type 1Z
9.11×105
Type 2Z 6
1.07×10
Type 3Z
8.38×105
Type 4Z
9.79×105
Type 5Z
1.00×106
Type 6Z
8.63×105
Type 7Z
1.36×106
Type 8Z
1.06×106
Type 9Z 5
9.16×10
Type 10Z
8.36×105
Maximum normal stiffness
1.36×106
Minimum normal stiffness
8.36×105
Average normal stiffness 5
9.93×10
and Y coordinates for the modelgrains centers were extracted from the digitized image and then compared to the experimental X and Y coordinates. The results
confirmed that the distortion effect was minimal and can safely be neglected within the scope of this study.
3.4 Contact normal stiffness
The contact normal stiffness of the wooden modelgrains was determined experimentally using a compression loading machine. The compression test was performed on
each of the 10 different shapes of the modelgrains. In order to study the effect of grain angularity on the contact normal stiffness, the compression tests were
performed in different directions around the same modelgrain. Load displacement curves for particle 8 (Fig. 7) show the effect of loading direction on the load
displacement results. The results for all the particles are summarized in Table 2. Based on the experimental results, an average value of 106 N/m for the contact normal
stiffness was used in the numerical model. The contact shear stiffness value was taken the same as the contact normal stiffness.
Page 367
Figure 8. Particle positions after imposing 50 mm displacement.
3.5 Experimental test procedures and results
The experimental setup was set to the initial conditions and the piston was fitted in place with zero displacement. The piston was then moved successively into the box,
thereby displacing the particles, as snapshots were taken at 50 mm intervals. The test was terminated when the piston could not be pushed further. The experimental
test was repeated five times in order to study the inherent variability of the procedure due to slight changes in initial conditions. The resulted images for tests 1, 4, and 5
for displacements 50, 150, and 250 mm are shown in Figures 8–10.
The initial conditions for all five tests were identical. At 50 mm piston displacement, an apparent “contact force chain,” which mostly connected the grains that
moved, developed and was almost identical for all five tests (Fig. 8). At 150 mm displacement, variations
Page 368
in the force chain among the various tests began to appear (Fig. 9).
At 250 mm displacement, the difference in force chains became more noticeable, with differences in particles positions and rotations throughout the system (Fig. 10).
As the displacement increased beyond this point, big differences were noticed. Although the same initial condition was ensured, reproducible results could not achieved.
The results indicate that minor changes in the initial conditions may dramatically affect the overall behavior of the modelgrains assembly at large displacements. The X
and Y coordinates, along with the rotations of the modelgrains, for the five experimental tests, were tracked by digitizing the experimental snapshots and were stored
for comparison with the numerical simulation.
4 NUMERICAL SIMULATION
The Particle Flow Code, PFC2D, was used to numerically simulate the experimental setup. The box was simulated as four fixed rigid walls, whereas the piston was
simulated as four rigid walls that moved vertically at constant velocity. The vertical movement was slow enough to ensure a stable solution. The clumps were generated
at the same positions as the experimental setup using the ORC method described earlier. However, the particles were prevented from rotating randomly and, instead,
were aligned in the same direction as the modelgrains. The initial condition for the numerical model, along with the corresponding experimental initial condition for Test
1, is shown in Figure 11.
The density of the simulated grains was set to the maple wood density and the friction coefficient was assumed to be 0.25. The contact normal/shear stiffness was set
to 106 N/m. The simulated state, including grain positions and rotations, was saved at 50 mm piston movement intervals for further comparison with the experimental
results. Because the particles are highly angular, and because the interparticle contact forces are low, the friction coefficient and contact
Figure 11. Initial conditions for the test #1 and the numerical simulation.
stiffness do not have a significant effect on the displacement and rotation of the particles. Instead, the displacement pattern is governed mostly by the shapes and
interlocking mechanism between the grains. This is consistent with the objective of this study, that is, to evaluate the ability of the ORC technique to model angular
particles within a DEM framework.
5 DISCUSSION
The numerical results along with the experimental results of Test 1 at displacements of 50, 150, and 250 mm are shown in Figures 12–14. From a qualitative
standpoint, it is clear that there is good agreement between the numerical results using the ORC method and the experimental results within the small displacement
range. With minor exceptions, the grain
Figure 12. Numerical and experimental test #1 after imposing 50 mm displacement.
Figure 13. Numerical and experimental results after imposing 150 mm displacement.
Figure 14. Numerical and experimental results after imposing 250 mm displacement.
Page 369
chains are almost identical (Fig. 12). The grain chain begins to deviate as the piston displacement increases, and the difference may be considered significant at large
displacement, as seen in Figure 14. However, it is important to note that large differences were noticed not only between numerical modeling and the experimental tests,
but also among the five “identical” experimental tests. It is also essential to note that the previous qualitative comparison was performed between the numerical
simulation and experimental Test 1 only.
In order to evaluate the ability of the numerical model to simulate the tests, it is necessary to compare the results in a quantitative manner. Specifically, if certain
quantitative measures, such as particle rotation and displacement, are compared between the experimental and numerical results, the numerical values should fall within
the range of the experimental data. To this end, a statistical comparison between the experimental results and the numerical results was conducted. Histograms were
generated for cumulative vertical (Y) displacements and rotations of the modelgrains at 50 mm piston displacement intervals (Figures 15–20). The results of the five
experimental tests and the numerical model are plotted in these figures. The histograms for particles rotations at piston displacements of 50, 150, and 250 mm, shown in
Figure 15–17, demonstrate that the numerical modeling results, using ORC method, lie within the range of experimental results except at large displacements. Similarly,
the results for vertical displacement shown in Figures 18–20 indicate that the numerical modeling results closely simulate the experimental results. Because the horizontal
displacements of particles are very small under the given boundary conditions, a similar comparison would not be of much value and is therefore not reported within the
context of the paper.
As mentioned earlier, the scaling procedure described by Equation (1) only ensures that the mass of the particle is proportional to the area. However, it
Figure 15. Comparison between experimental and numerical particles rotations after 50 mm displacement.
Page 370
Figure 18. Comparison between experimental and numerical particles vertical movements after 50 mm displacement.
Page 371
neither guarantees that the moment of inertia nor the center of mass of the ORC particle coincides with that of the actual particle. While this may contribute to the
difference between the numerical model and the experiments, it is believed that the effect is minor because the particle accelerations are small. In a “fully dynamic”
system, where accelerations are more significant, it is important to ensure that the center of mass and the moment of inertia of the numerical model particles match the
experimental values.
Another important aspect is the loose packing of the particle assemblies described in this paper. The main reason behind avoiding a dense packing was to allow the
particles to move without imposing a high boundary force through the piston, which could potentially cause the wood to crack. If a densely packed system was used,
the system would have been much more stable and the chaotic element in the system response would have been reduced, thereby minimizing the difference numerical
and experimental results.
6 CONCLUSION
An experimental validation of the ability of DEM to model angular particles created using overlapping rigid clusters (ORC) was presented. In the ORC technique,
clumps that contain circular particles are created in such a way that the outline of each clump coincides with that of the actual particle. An experimental setup was built
and model particles were manufactured in the laboratory. The experimental tests were modeled numerically using PFC2D. Experimental test results from five identical
tests were compared qualitatively and quantitatively with the corresponding numerical simulation. The ORC technique was found to be effective in modeling the
behavior of angular particle assemblies. While particle displacements and rotations of both the numerical and experimental systems were almost identical at small
displacements, larger differences were observed at high displacements, even among the five experimental tests that were performed under identical initial and boundary
conditions. This variability is attributed to minor changes in the initial conditions, as well as other inherent uncertainties in the system, and cannot be practically avoided.
However, the numerical model revealed that the ORC simulated system response falls mostly within the range of the five experimental tests. The experimental validation
will be extended to different types of sands in the next research stage. The experimental setup will be altered to induce different modes of disturbance to the model
grains. Photostress analysis methods are currently being considered to evaluate and compare the interparticle contact forces obtained from the experimental tests and
the numerical simulation.
REFERENCES
Ashmawy, A.K., Sukumaran, B., & Hoang, A.V. 2003. Evaluating the influence of particle shape on liquefaction behavior using Discrete Element Method. In Proceedings
of the Thirteenth International Offshore and Polar Engineering Conference (ISOPE 2003) Honolulu, Hawaii, May 2003.
Barbosa, R., & Ghaboussi, J. 1992. Discrete finite element method. Engineering Computations, Vol. 9:253–266.
Cundall, P.A., & Strack O.D.L. 1979. A discrete numerical model for granular assemblies. Geotechnique, Vol. 29: 47–65.
Favier, J.F., AbbaspourFard, M.H., & Kremmer, M. 1999. Modeling nonspherical particles using multisphere discrete element. Journal of Engineering Mechanics, 127
(10):971–976.
Jensen, R., Bosscher, P., Plesha, M., & Edil, T. 1999. DEM simulation of granular mediastructure interface: effect
Page 372
of surface roughness and particle shape. International Journal for Numerical and Analytical Methods in Geomechanics, Vol. 23:531–547.
Itasca Consulting Group (1999). Particle Flow Code in Two Dimensions. Users Manual and Software Version 2.0.
Lin, X., & NG, T.T. 1997. A threedimensional discrete element model using arrays of ellipsoids. Geotechnique, 47(2):319–329.
Matuttis, H.G., Luding, S., & Herrmann, H.J. 2000. Discrete element simulations of dense packing and heaps made of spherical and nonspherical particles. Powder
Technology, 109(1):278–292.
Ng, T.T. 1994. Numerical simulation of granular soil using elliptical particles. Computers and Geotechnics, 16(2): 153–169.
Potapov, A., & Campbell, C. 1998. A fast model for the simulation of nonround particles. Granular Matter, Vol. 1:9–14.
Thomas, P.A., & Bray, J.D. 1999. Capturing nonspherical shape of granular media with disk clusters. Journal of Geotechnical & Geoenvironmental Engineering, 125
(3): 169–178.
Ting, J., Khawaja, M., Meachum, L., & Rowell, J. 1993. An ellipsebased discrete element model for granular materials. International Journal for Numerical and
Analytical Methods in Geomechanics, Vol. 17:603–623.
Page 373
Threedimensional discrete element simulations of direct shear tests
C.O’Sullivan & L.Cui
Department of Civil Engineering, University College Dublin, Ireland
J.D.Bray
Department of Civil and Environmental Engineering, University of California, Berkeley, USA
679 3
ABSTRACT: Using discrete element simulations, one can monitor the micromechanisms driving the macroresponse of granular materials and
quantify the evolution of local stress and strain values. However, it is important to couple these simulations with carefully controlled physical tests for
validation and insight. Only then can findings about the micromechanics of the material response be made with confidence. Moreover, the sensitivity
of the observed response to the test boundary conditions can be analyzed in some detail. The results of threedimensional discrete element simulations
of direct shear tests and as well as complementary physical tests on specimens of steel balls are presented in this paper. Previous discrete element
analyses of the direct shear test have been restricted to twodimensional simulations. For the simulations presented here, an analysis of the internal
stresses and contact forces illustrates the threedimensional nature of the material response. The distribution of contact forces in the specimen at larger
strain values, however, was found to be qualitatively similar to the twodimensional results of Zhang & Thornton (2002). Similarities were also
observed between the distribution of local strain values and the distribution of strains obtained by Potts et al. (1987) in a finite element analysis of the
direct shear test. The simulation results indicated that the material response is the stress dependent. However, the response observed in the simulations
was found to be significantly stiffer than that observed in the physical tests. The angle of internal friction for the simulations was also about 3° lower
than that measured in the laboratory tests. Further laboratory tests and simulations are required to establish the source of the observed discrepancies.
1 INTRODUCTION
To date, much of our understanding of soil response has developed using standard laboratory tests. Much can be learned by revisiting these tests and examining the
micromechanics underlying the observed macroscale response using discrete element method (DEM) analyses of these tests. This paper presents some preliminary
results of an ongoing analysis of the direct shear test that couples physical tests and DEM simulations. The physical tests on stainless steel spheres are described firstly
Following a detailed description of the details of the DEM simulation approach, the numerical results are compared with the physical tests results. Significant differences
in the material responses were observed. The next phase of this research will attempt to identify the source of these differences. Analysis of the simulation results
examined the local stresses, the particle displacements, the contact forces and the local strains. These results are compared with the earlier findings of Zhang &
Thornton (2002) and Potts et al. (1987).
2 BACKGROUND
Although there appears to be a slight difference in opinion on the origin of the direct shear test, it is unquestionably one of the oldest and most commonly used
laboratory tests in geotechnical engineering. Holtz and Kovacs (1981) state that the tests was originally used by Coulomb in the 18th century, while Potts et al. (1987)
cite Skempton (1949) who found that the test was first used in the 19th century by Alexandre Collin. In a preface to a Geotechnique symposium in print on “The
engineering application of direct and simple shear testing,” Toolan (1987) observed that the direct shear test declined in popularity with the emergence of triaxial testing.
Toolan argues that testing with a direct shear test provides essential data for applications with low factors of safety where moderate displacements of the soil can be
expected upon failure. A recent publication by Lings and Dietz (2004) proposes modifications to the apparatus to facilitate testing of sand, illustrating continued interest
in this apparatus. Other more recent research includes the
Page 374
experimental studies by Shibuya et al. (1997) and the finite element analyses described by Dounias et al. (1993) and Potts et al. (1987).
The standard test procedure for the direct shear test is documented in ASTM D3080 (1990). Both Potts et al. (1987) and Shibuya et al. (1997) note the limitations
of the test, including the nonuniformity of the stresses and strains induced in the soil and difficulties associated with interpreting the test results to define a failure
criterion. As noted by both Potts et al. and Shibuya et al. the popularity of the test can largely be attributed to its simplicity.
Shibuya et al. (1997) examined the influence of boundary effects, such as wall friction, on the specimen response. The distribution of normal and shear stresses in the
specimens tested were measured using twocomponent load cells. Considering direct shear tests on specimens of dense Toyoura sand, the effect of wall friction on the
specimen response were studied. It was found that as a result of wall friction, vertical stresses on the potential shear plane were found to be substantially lower than the
conventional measurement of vertical stress and that the value of at peak was overestimated when the stress was calculated from the applied normal force directly.
Potts et al. (1987) and Dounias et al. (1993) performed continuumbased numerical analyses of the test using the finite element method. Potts et al. found that the
reference stress state for the direct shear box is that of simple shear. The study also considered the influence of K0 on the specimen response. When a constitutive
model without strain softening was used to describe the material response, the distribution of stresses and strains within the material were highly nonuniform. However
when a model incorporating significant strain softening was used, the stresses and strains were more uniform.
Some earlier research has explored the applicability of DEM to simulate the direct shear test. Morgan & Boettcher (1999) and Morgan (1999) used the DEM code
TRUBAL to simulate direct shear tests with the objective of understanding the nature of deformation of granular fault gouge. Morgan & Boettcher calculated the
directional derivative of the horizontal displacement surface as a means to highlight discontinuities in the specimen. They observed that at lower strain values (about 2%)
the deformation was broadly distributed across multiple subhorizontal slip planes and progressively localized onto a single shear plane by 8–10% strain.
Masson & Martinez (2001) used the twodimensional code PFC2D (Itasca 2002) to simulate twodimensional direct shear tests using 1050 disks of diameter 2, 3
and 4 mm. The particle scale responses analyzed by Masson & Martinez included the particle instantaneous velocity field, the distribution of particle horizontal and
vertical displacements as a function of depth and the distribution of particle rotation as a function of depth. As other studies have shown (e.g. Iwashita & Oda 1998) the
particle rotations were found to be a significant indicator of strain localization. Masson and Martinez recognized the limitations of their twodimensional study and
highlighted the potential sensitivity of their results to the particleboundary friction value.
In another twodimensional study, using a single layer thickness of 5000 polydisperse spherical particles, Zhang & Thornton (2002) also analyzed the direct shear
test using the TRUBAL code. Comparing the findings of Zhang & Thornton and Masson & Martinez, the orientation of the contact forces at the peak strength exhibit a
more definite trend in the simulation of Zhang & Thornton, most likely as a consequence of the larger number of particles used.
All of these earlier studies of the direct shear test using DEM considered twodimensional analyses. The work of Thomas (1997) quantitatively demonstrates the
limitations of both twodimensional DEM simulations and physical tests using “Schneebli” rods to represent real soil response. The current study builds upon the
findings of these earlier researchers by extending the simulations to three dimensions. Furthermore, by coupling the numerical simulations with complementary physical
tests, further insight into the material response can be obtained.
3 PHYSICAL TESTS
3.1 Direct shear testing
A schematic diagram of the direct shear test apparatus used in the current study is given in Figure 1. The apparatus comprises a metal box of square cross section (60
mm wide), divided in two halves horizontally. The lower section of the box was moved forward at a constant velocity (0.015 mm/s) while the upper section of the box
remained stationary. The force required to maintain the upper section of the box in a stationary position was measured using a load cell and proving ring. The calibration
of the load cell was checked prior to testing. The vertical load was applied to the top of the shear box using a system of dead weights attached to a lever.
In the current study 0.9922 mm Grade 25 Chrome Steel Balls were used (manufacturer: Thompson Precision Ball). This material was used as the tolerances on the
ball diameter and sphericity were relatively tight (±7.5×10−4 mm). The ball material properties are summarized in Table 1. The coefficient of friction values obtained by
O’Sullivan (2002) are used in the current study, as the steel balls used are equivalent to the steel balls considered in the earlier study. For each test the balls were
placed in the shear box in three equal
Page 375
Figure 1. Schematic diagram of direct shear test setup.
Table 1. Material parameters of Grade 25 Chrome Steel Balls used in laboratory tests.
Parameter Value Unit

Radius 0.9922 mm
Density
7.8334×10−6 kg/mm3
Shear modulus
7.945×107 kg/(mm s2)
Poisson ratio 0.28
Friction coefficient* 5.5°
*As determined by O’Sullivan (2002) for equivalent balls.
Figure 2. Summary of physical test results. (a) Shear stress versus horizontal strain, (b) Shear stress versus normal stress.
Table 2. Summary of physical test results.
Test no. Normal stress (kPa) Peak shear stress (kPa) Void ratio

Dense 1–1 54.5 24.86 0.5841
Dense 1–2 54.5 27.23 0.5826
Dense 1–3 54.5 19.58 0.6089
Dense 2–1 109 45.90 0.5991
Dense 2–2 109 49.50 0.6097
Dense 2–3 109 42.75 0.5916
Dense 3–1 163.5 83.03 0.5818
Dense 3–2 163.5 71.55 0.6044
Dense 3–3 163.5 75.15 0.5969
layers. Each layer had a mass of 125 g (corresponding to 3900 balls). Following placement of the balls, five “taps” of a hammer were applied to each vertical side of
the box to increase the specimen density.
The test results are illustrated in Figure 2 and summarized in Table 2. The shear stress was calculated by dividing the horizontal load measured in the proving ring by
the specimen crosssectional area, while the vertical stress was calculated by dividing the applied vertical load by the specimen crosssectional area. As illustrated in
Figure 2(b), the friction angle for the material was found to be 24.4°. However, as illustrated in both Figure 2(a) and Figure 2(b), there was a degree of scatter in
the experimental results.
By reference to Table 2 it can be seen that the variation in specimen response appears to be largely a result of variation in the specimen void ratio values which were
difficult to control accurately. Considering the maximum shear stresses and the minimum shear stresses measured for each normal stress, the maximum friction angle
was found to be 26.2°, while the minimum friction angle was found to be 22.7°. To verify the experimental results the tests were repeated in the geotechnical laboratory
at Trinity College Dublin. For specimens with equivalent densities, the measured friction angles differed by less than 0.25°.
4 DEM SIMULATIONS
4.1 Implementation of direct shear testing in 3DDEM
The distinct element method program (3DDEM) used for the simulations described here is a modification of a code developed by Lin and Ng (1997) for three
dimensional ellipsoidal particles, which in turn is a modification of the Trubal DEM code (Cundall & Strack 1979). The program uses spherical particles and is further
described in O’Sullivan (2002). In the DEM code, the HertzMindlin contact model is used to model the contact between spheres. Details of the
Page 376
implementation of this model can be found in Itasca (2002) or Bardet (1998).
The method of simulating the direct shear test in the 3DDEM code was originally proposed by O’Sullivan (2002). A diagram illustrating the simulation procedure is
given in Figure 3. During the simulation, the specimen is enclosed by eight rigid boundaries, two horizontal boundaries at the top and bottom, two vertical boundaries at
the front and back, and four vertical boundaries at the left and right of the specimen. Prior to shearing, the two left boundaries are coplanar and the two right
boundaries are coplanar. During shearing, one set of vertical boundaries is moved laterally with a constant velocity. An imaginary horizontal plane is constructed
through the middle of the specimen. All of the particles with centroids located above this plane are checked for contact with the moving boundaries, while all of the
particles with centroids located below this plane are checked for contact with the stationary boundaries. Using this approach the computational cost associated with
contact detection between a sphere and the edge of the shear box is avoided.
The measurement sphere illustrated in Figure 3 is used to measure the average stresses in the specimen. Such a system to measure stress in discrete element
simulations has been described by a number of researchers including Bardet (1998). The average stress in the sphere is given by
(1)

where Nc is the number of contacts within the measurement sphere, V is the volume of the sphere, is the branch vector
Figure 3. Illustration of implementation of direct shear tests in code 3DDEM.
connecting two contacted particle, a and b. The branch vector is given by
(2)
where are the position vectors of the spheres a and b meeting at contact c.
Prior to shearing the measurement sphere is used in a servocontrolled system to bring the specimen into a prescribed initial stress configuration. If the stress
measured within the measurement sphere, the boundaries perpendicular to i direction are slowly moved so that the measured stress attains the required stress
values. A servocontrolled approach is also used during the test so that the vertical stress, as measured in the measurement sphere, remains constant during shearing. A
detailed description of the implementation of the servocontrolled system in 3DDEM is given by O’Sullivan (2002).
4.2 Analysis parameters
Table 3 lists the input parameters used in the simulation. A comparison of the parameters listed in Table 3 and Table 1 illustrates that the physical test and numerical
simulation are equivalent, apart from the differences in density value used. Density scaling is commonly used to reduce the computational cost associated with DEM
simulations for quasistatic analyzed (e.g. Morgan & Boettcher 1999, Thornton 2000, and O’Sullivan et al. 2004). The maximum stable time increment for explicit
DEM simulations is proportional to the square root of the particle mass, motivating the use of such scaling. In the current analyses attempts were made to run the
simulations without any density scaling. However, significant difficulties were encountered in bringing the system into equilibrium. These problems may have been a
consequence of the small particle masses used in comparison with the contact stiffnesses. Note that O’Sullivan (2002) performed a detailed analysis of the sensitivity of
simulations of quasistatic triaxial tests to the input parameters. The results of this analysis indicated that once the ratios σconfining/K and ρ/K remained constant there
was no variation in the simulation results (σconfining=confining pressure,
Table 3. Input parameters for DEM simulations.
Parameter Value Unit

Radius 0.9922 mm
Density
7.8334×103 kg/mm3
Shear modulus
7.945×107 kg/(mms2)
Poisson ratio 0.28
Friction coefficient 5.5°
Page 377
K=linear contact spring stiffness, ρ=particle density).
Previous analyses by O’Sullivan et al. (2004) that considered specimens with regular packing configurations confirmed that these precision steel ball particles can be
modeled using uniform spheres. While some variation in the coefficient of friction values for the particles was measured, such variation has a negligible influence on the
specimen response for specimens with irregular packing configurations (O’Sullivan et al. 2002). For the analyses discussed in this paper the ballboundary coefficient of
friction was assumed to be zero. In discrete element simulations of laboratory tests we can easily examine the influence of boundary friction values on the macroscale
response. Future analyses will explore this issue in more detail.
4.3 Specimen generation
The difficulties associated with generating a threedimensional specimen to a specific void ratio for use in a discrete element analysis cannot be overstated. A review of
some of the available approaches to specimen generation is provided by Cui & O’Sullivan (2003). As noted by O’Sullivan (2002), considerable effort has been
expended by mathematicians to develop algorithms to generate dense, random assemblies of spheres (e.g. Zong 1999). For the current analysis one such algorithm,
proposed by Jodrey & Tory (1985), was initially used to generate the specimen. The resultant specimen was however less dense than the specimens tested in the lab
(with a void ratio of 0.77 in comparison to void ratios of 0.58–0.61). To overcome this limitation and achieve the required density value, 11700 balls with a radius of
0.9393 mm were initially generated in a box of size 0.06×0.06×0.02 mm. This radius was then gradually expanded to 0.9922 mm, the box height was increased to 21
mm, and a DEM simulation was run until the system came into equilibrium. During this simulation the interparticle friction angle was assigned a value of zero. The
behavior of the particles during the equilibrium phase can be understood by reference to Figure 4. Initially upon initial radius expansion there was significant inter
particle overlap, with a maximum particle overlap of approximately 0.1059 mm (Fig. 4a). As the equilibrium phase of the analysis progressed the balls moved to
assume a dense packing configuration with a minimum amount of particle overlap (Fig. 4c). Prior to shearing the maximum overlap was approximately 1.73×10−3 mm.
4.4 Direct shear test simulations
Three DEM simulations were performed with initial isotropic stress conditions of 54.5 kPa, 109 kPa and 163.5 kPa. These initial stress conditions were attained using
the servocontrolled approach detailed above. The upper section of the box was moved at a velocity of 0.001 mm/sec. It is important in these servocontrolled tests to
verify firstly that the specified stress conditions are maintained for the duration of the simulation. A period of initial iteration was required to determine the appropriate
control parameters to meet this requirement. Figure 5 illustrates the vertical stress in the specimen during the simulations as a function of horizontal strain. Both the
vertical stress measured in the measurement sphere as well as the vertical stress calculated from the measured boundary forces are illustrated. The measured stresses
varied slightly during the simulations, with a maximum deviation of 5% from the specified value measured in the measurement sphere.
The macroscale test results are summarized in Table 4 and Figure 6. Figure 6a is a plot of shear stress versus horizontal strain, and Figure 6b is a plot of peak shear
stress versus normal stress. The results yield a value for the angle of internal friction of 19.6°, which compares with a value of 22.7° as the minimum value obtained in
the laboratory tests. The void ratios of the specimens in the DEM simulations were close to the minimum void ratio values of the test specimens.
Figure 4. Illustration of specimen generation phase of analysis.
Figure 5. Normal stresses (σz z) as calculated for measurement sphere and vertical stresses calculated from boundary forces for the
three simulations considered.
Page 378
Table 4. Summary of test simulation results.
Test no. Normal stress, kPa Peak shear stress, kPa Void ratio

1 54.5 15.9 0.587
2 109 37.5 0.587
3 163.5 60.4 0.587
Figure 6. Summary of DEM simulation results. (a) Shear stress versus horizontal strain. (b) Peak shear stress versus normal stress.
Hence, the DEM simulations were found to underestimate the strength of the ball assembly measured in the laboratory.
A comparison of Figures 2 & 6 indicates that the numerical simulations exhibit a significantly stiffer response in comparison to the physical tests. Considering the
volumetric strain response of the physical test specimens, some compression was observed at the start of shearing prior to dilation. However for the numerical
simulations the specimens dilated from the onset of shearing. These differences in volumetric strain response, suggest that there may be significant differences between
the fabric of the physical test specimens and the fabric of the specimens analyzed using DEM. In physical tests, O’Sullivan (2002) observed that the response of
specimens of uniform steel balls is very sensitive to the specimen preparation approach; specimens with similar void ratios can exhibit different responses depending
upon the specimen preparation approach used. Future research will address the source of these discrepancies considering the sensitivity of the response to the
specimen preparation method as well as any compliance of the laboratory apparatus. As a quality assurance procedure the tests will be replicated using an alternative
direct shear box located in another laboratory. Further analyses will explore the influence of ballboundary friction on the macroscale response as well as the contact
constitutive model used in the simulations.
Coupling physical tests and numerical simulations is nontrivial as it can be difficult to replicate the exact details of the physical test in the numerical simulation.
O’Sullivan (2002) describes some of the challenges associated with successful validation of discrete element codes using physical tests.
Notwithstanding the discrepancies between the stiffness observed in the laboratory tests and the numerical simulations, the numerical simulations clearly captured
some of the typical response characteristics of granular materials in direct shear tests. For example, the stressdependency of the specimen response can clearly be
seen with the ratio of the peak shear stress to the residual shear stress increasing as the normal stress increases. Such stress dependency, however, was not observed in
the physical tests. It is also interesting to observe in Figure 6a that the three specimens tested in the direct shear test appear to approach a similar value of shear stress
(about 20 kPa) as shearing progresses to large displacements.
5 ANALYSIS OF MICROSCALE PARAMETERS
5.1 Stress analysis
In a direct shear test the shear stress along the region of localization is inferred from the macroscale measurement of shear force. In the DEM analyses a series of
rectangular boxes were created and the average stresses within this box were calculated using Equation 1. The configuration of these boxes is illustrated in Figure 7; the
center of the stress measurement box was located at the center of the specimen and the thickness of the box, d, ranged from H/10 to 2H/5, where H is the specimen
height. The variation of the local stresses, σxz, as measured in the stress measurement box, with increasing strain is illustrated in Figure 8.
As illustrated in both Figure 8a & Figure 8b, the shear stress, σzx, measured increased as the thickness of the measurement box decreased. For an applied vertical
stress of 54 kPa (Fig. 8a), the σzx values consistently exceeded the global shear stress values
Page 379
Figure 7. Illustration of configuration of rectangular box used to calculate local stresses.
Figure 8. Variation of shear stress (σzx) values, measured in measurement boxes, as a function of horizontal strain. (a) Vertical
stress=54.5 kPa. (b) Vertical stress=109 kPa.
(calculated from the boundary forces). However, for an applied vertical stress of 109 kPa (Figure 8(b)) and an applied vertical stress of 163 kPa (not shown), the local
shear stress (σzx) values were less
Figure 9. Incremental displacement vectors for horizontal strain increment of 0.003 to 0.052, vertical stress 54.5 kPa. (a) Front view
(looking along yaxis) (b) Side view (looking along xaxis). Note: For ease of visualization only displacements exceeding
the mean displacement value are illustrated.
than the global shear stress. For the three vertical stresses considered, the ratio of the peak σzx value for d=1/5 H divided by the peak σzx value for d=2/5 H is
consistently 1.1, whereas the ratio of the peak σzx value for d=1/10 H divided by the peak σzx value for d=2/5 H is consistently 1.3.
5.2 Incremental displacements
A threedimensional plot of the incremental displacements for the simulation with a normal stress of 54 kPa is illustrated in Figure 9 for the horizontal strain increment
0.003 to 0.052. Figure 9a is a front view of the specimen and Figure 9b is a side view of the specimen. For ease of visualization only displacements exceeding the
average incremental displacement are illustrated. Referring firstly to Figure 9a, as would be expected, most of the displacement is concentrated in the upper half of the
shear box. There also is a concentration of displacement vectors close to the left and right boundaries. At the left hand side of the box there is significant downward
motion of the particles. The magnitude of the upward displacement of the particles in the upper
Page 380
Figure 10. Contact force vectors at a horizontal strain of 0.052, vertical stress 54.5 kPa. (a) Front view (looking along yaxis). (b) Side
view (looking along xaxis). Note: For ease of visualization only contacts where the forces exceeds the average contact
force +1 standard deviation are considered.
right section of the box is less than the magnitude of the downward motion of the particles close to the left boundary. Comparing Figures 9a & 9b, it is clear that the
components of the particle displacement vectors in the ydirection are small in comparison to the component of the vectors in the zdirection.
5.3 Contact forces
In order to appreciate the spatial variation in the contact forces, a threedimensional plot of the unit contact force vectors for the simulation with a normal stress of 54
kPa at a horizontal strain of 0.052 is illustrated in Figure 10. Figure 10a is a front view of the specimen and Figure 10b is a side view of the specimen. For ease of
visualization, only contacts where the magnitude of the contact force exceeds the average contact force +1 standard deviation are considered. The distribution of
contact forces illustrated in Figure 10a is qualitatively similar to the distribution of contact forces obtained by Zhang & Thornton (2002) in their twodimensional
discrete element analysis of the direct shear test. The forces are transmitted diagonally across the specimen. An orthogonal view of
Figure 11. Threedimensional plot of contact force vectors (a) Front view (looking along yaxis) (b) Side view (looking along xaxis).
the specimen, along the zaxis, is given in Figure 10b; however, it is difficult to identify any clear trend in the orientation of the contact force vectors from this
perspective.
Figure 11 is a threedimensional plot of the contact force vectors, giving an indication of both the magnitude and direction of the contact forces. As illustrated in
Figure 11a it is clear that the xcomponent of the contact force exceeds the zcomponent. Considering the distribution of the y and z components of the contact forces,
the plot is approximately circular, as illustrated in Figure 11b. The y and z components of the contact forces are therefore relatively evenly distributed. An analysis of the
local normal stresses as calculated from the contact forces indicated that σxx>σyy>σzz. The relatively large value in σyy is an indicator of the three dimensional nature of
the problem.
5.4 Strain localization analysis
For the simulation with a vertical stress of 163 kPa, the local strain values were calculated using the nonlinear homogenization technique proposed by O’Sullivan et al.
(2003). The strain values are plotted on a vertical plane through the center of the specimen,
Page 381
Figure 12. Strain contours for horizontal strain increment 0.003 to 0.017, plotted on a vertical plane through center of specimen with y=30.
Normal stress 163 kPa. Contour interval 0.025.
i.e. with y=30 mm. Figure 12 illustrates the local strain values for the horizontal strain increment from 0.003 to 0.017. While there are zones with high strain values at
mid height close to the left and right boundaries, no clear localization through the center of the specimen is apparent, even though the specimen is close to the peak
stress at this strain level. Considering the volumetric strain values (εvol), the localizations appear to be inclined. Similar inclined localizations were observed by Potts et
al. (1987) where a constitutive model without strain softening was used.
Figure 13 illustrates the distribution in strain values at a horizontal strain of 0.047. The distribution of shear strains is highly irregular and there is still no clear evidence
of a localization in the specimen. Similar strain distributions were observed for the specimens with vertical stresses of 54.5 kPa and 109 kPa. This contrasts with the
study of localizations in specimens with regular packing configuration described by O’Sullivan & Bray (2003) where a distinct localization was observed in three
dimensional simulations of plane strain tests.
Figure 13. Strain contours for horizontal strain increment 0.003 to 0.047, plotted on a vertical plane through center of specimen with y=30.
Normal stress 163 kPa. Contour interval 0.05.
6 DISCUSSION
The direct shear test has been studied using discrete element analysis in combination with a series of complementary physical tests. The agreement between the
laboratory tests and the numerical simulations was not wholly satisfactory. The angle of friction calculated using the results of the simulations was 3° lower than that
obtained in the physical tests. The response in the simulations was also stiffer than the response observed in the physical tests. Further tests and simulations are required
to resolve these discrepancies, although it is likely that test device friction may be a contributing factor.
An analysis of the results of the simulations considered the local stress values, the incremental displacements, the distribution of contact forces and the local strain
values. Whereas the particle displacements were predominantly restricted to the direction of shearing (i.e. parallel to the y=0 plane), significant contact forces
developed in the ydirection, illustrating the
Page 382
threedimensional nature of the material response at the particle scale. Within the specimen, the stresses are nonuniform with the shear stresses increasing closer to the
zone of shearing. The local strain values are nonuniform and the localization cannot be identified at horizontal strains of approximately 5%.
ACKNOWLEDGEMENTS
Funding for this research was provided by the Irish Research Council for Science, Engineering and Technology (IRCSET) under the Basic Research Grant Scheme.
Additional partial funding was provided by the Institution of Engineers of Ireland Geotechnical Trust Fund. The authors are grateful to Mr. George Cosgrave, University
College Dublin, for his assistance in performing the laboratory tests.
REFERENCES
ASTM, 1990. D3080: standard test method for direct shear test of soils under consolidated drained conditions. Annual book of ASTM standards: 417–422.
Cui, L. & O’Sullivan, C. 2003. Analysis of a Triangulation Based Approach for Specimen Generation for Discrete Element Simulations. Granular Matter, 5(3):135–145.
Dounias, G.T. & Potts, D.M. 1993. Numerical analysis of drained direct and simple shear test. Journal of Geotechnical Engineering, 119(12):1870–1891.
Holtz R.D. & Kovacs, W.D. 1981 .An Introduction to Geotechnical Engineering. Prentice Hall.
Itasca Consulting Group, Inc. 2002. PFC2D—Particle Flow Code in 2 Dimensions, Ver. 3.0 User’s Manual. Minneapolis: Itasca.
Iwashita, K. & Oda, M. 2000. Microdeformation mechanism of shear banding process based on modified distinct element method. Powder Technology 109:192–205.
Jodrey, W.S. & Tory, E.M. 1985. Computer simulation of close random packing of equal spheres. Physical Review A, 32(4):2347–2351.
Lings, M.L. & Dietz, M.S. 2004. An improved direct shear apparatus for sand. Geotechnique, 54(4):245–256.
Masson, S. & Martinez, J. 2001. Micromechanical analysis of the shear behavior of a granular material. ASCE Journal of Engineering Mechanics, 127(10):1007–1016.
Morgan, J. & Boettcher, M.S. 1999. Numerical simulations of granular shear zones using the distinct element method. 1. Shear zone kinematics and the micromechanics of
localization. Journal of Geophysical Research, 104(B2): 2703–2719.
Morgan, J. 1999. Numerical simulations of granular shear zones using the distinct element method. 2. Effects of particle size distribution and interparticle friction on
mechanical behavior. Journal of Geophysical Research, 104(B2):2721–2732.
O’Sullivan, C. 2002. The Application of Discrete Element Modelling to Finite Deformation Problems in Geomechanics. Ph.D. thesis, Dept. of Civil Engrg., Univ. of
California, Berkeley.
O’Sullivan, C., Bray, J.D. & Riemer, M.F. 2002. The Influence of Particle Shape and Surface Friction Variability on Macroscopic Frictional Strength of RodShaped
Particulate Media. ASCE Journal of Engineering Mechanics, 128(11):1182–1192.
O’Sullivan, C. & Bray, J.D. 2003. Evolution of Localizations in Idealized Granular Materials. Workshop on Quasistatic Deformations of Particulate Materials.
O’Sullivan, C., Bray, J.D. & Riemer, M.F. 2004. An examination of the response of regularly packed specimens of spherical particles using physical tests and discrete
element simulations. ASCE Journal of Engineering Mechanics, accepted.
O’Sullivan, C., Bray, J.D. & Li, S. 2003. A New Approach for Calculating Strain for Particulate Media. International Journal for Numerical and Analytical Methods
in Geomechanics, 27(10):859–877.
Potts, D.M., Dounias, G.T. & Vaughan, P.R. 1987. Finite element analysis of the direct shear box test. Geotechnique, 37(1):11–23.
Skempton, A.W. 1949. Alexandre Collin: a note on his pioneer work in soil mechanics. Geotechnique, 19(1): 75–86.
Thomas, P.A. 1997. Discontinuous deformation analysis of particulate media. Ph.D. thesis, Dept. of Civil Engrg., Univ. of California, Berkeley.
Toolan, F.E. 1987. Preface to “Symposium in print on the engineering application of direct and simple shear testing”, Geotechnique, 37(1):1–2.
Shibuya, S., Mitachi, T. & Tamate, S. 1997. Interpretation of direct shear box testing of sands as quasisimple shear. Geotechnique, 47(4):769–790.
Zhang, L & Thornton, C. 2002. DEM simulations of the direct shear test. In Proceedings of 15th ASCE Engineering Mechanics Conference, June 2–5, 2002.
Columbia University, New York, NY.
Zong, C. 1999. Sphere Packings. Springer.
Page 383
Calibration of a discrete element model for rock failure envelope and tensile
strength
A.Fakhimi
Departments of Mineral and Mechanical Engineering, New Mexico Tech, Socorro, NM, USA
Department of Civil Engineering, Tarbiat Modarres University, Tehran, Iran
T.Villegas
Department of Mineral Engineering, New Mexico Tech, Socorro, NM, USA
Department of Civil and Mining Engineering, University of Sonora, Hermosillo, Sonora, Mexico
679 3
ABSTRACT: A promising approach in modeling rock deformation and fracture is the implementation of a discrete element method in which rock is
modeled as an assembly of many circular or spherical particles that are glued to each other. The problem with this technique is the calibration of the
model in which appropriate micromechanical properties must be selected to result in correct macroscopic rock properties. So far, successful
calibration of rock modulus of elasticity, Poisson’s ratio and unconfined compressive strength has been achieved by appropriate selection of spring
constants and normal and shear bonds between particles. On the other hand, the effort to calibrate the numerical model for rock tensile strength,
cohesion and failure envelope has not been very successful. In this paper, a new technique is introduced to address the calibration of the model for the
above parameters using CA2 computer program developed by the first author. In this approach, slightly overlapping circular particles will be used.
The particles are overlapped by applying a hydrostatic stress in the process of sample genesis. In addition, a dimensional analysis is presented to
obtain calibration graphs for rock failure envelope and tensile strength. It is shown that by using slightly overlapped circular particles together with
these calibration charts, quick estimates of microscopic properties can be obtained that result in realistic simulated rock cohesion, tensile strength and
curvature of failure envelope.
1 INTRODUCTION
Rocks are made of different mineral constituents with directional properties. This fact together with other factors such as the presence of pores and fissures contribute
to the complexity of rock behavior under loading. Due to an applied deviatoric stress, damaged zones develop and extend inside a rock specimen. The joining and
coalescence of these damaged zones eventually break the rock sample through a macroscopic crack, known as a shear band. A shear band in a geomaterial can be
modeled as a bifurcation phenomenon, using closed form solutions in the plasticity framework for samples with simple geometries (Vermeer & De Borst 1984).
Numerical modeling of shear banding has been of interest as well. In this case, a nonlocal continuum approach is usually adapted to deal with mesh sensitivity involved
in the solution (Bazant & PijaudierCabot 1988).
A more recent technique in modeling rock crack propagation is the use of an assemblage of circular particles, which can interact through normal and shear springs.
This technique, for modeling geomaterials, was initially proposed by Cundall and Strack (1979) and later on extended and implemented in the PFC computer programs
(Itasca 1999). With this technique, development of new cracks, rock dilation and anisotropic rock behavior in the damaged zones can be modeled. But, this powerful
method has a few shortcomings. Firstly, it needs calibration. The synthetic material obtained by creating circular disks needs micromechanical properties, which cannot
be measured directly in the laboratory. For example, for a contact bond model used in this paper, five micromechanical constants, namely: normal spring stiffness (kn ),
shear spring stiffness (ks ), coefficient of friction (μ), normal bond (nb ) and shear bond (sb ) must be specified. One way to obtain these parameters is to assume some
initial values for them, create a synthetic model, and conduct some numerical experiments to obtain the macroscopic properties associated with the synthetic material.
These macroscopic properties
Page 384
could be Young’s modulus, Poisson’s ratio, compressive and tensile strength of rock. In this approach, several trials and errors are needed to acquire appropriate
microparameters to result in macroscopic properties similar to those of the real rock, i.e. to calibrate the numerical model. A more efficient method of calibration is to
use a dimensional analysis. This method was used by Huang (1999) for calibration of a circular particle interaction (CPI) model.
The second problem of using a discrete element approach is the difficulty of model calibration for rock tensile strength and failure envelope. A synthetic material
made of a CPI model is not usually capable of simulating realistic qu /σt (ratio of unconfined compressive strength to tensile strength) and failure envelope similar to
those of hard rocks (Potyondy & Cundall 2001). Recently, a new technique was introduced by Fakhimi (2004) to remedy the above problem. In this technique, the
disks are initially slightly overlapped by the socalled genesis stress (σ0). In this way a slightly overlapping circular particles interaction (SOCPI) model is developed
with strength properties similar to rocks.
In this paper, the SOCPI model is briefly introduced. In addition, dimensionless charts are presented for calibration of a SOCPI material. These calibration graphs
are different from those introduced by Huang (1999), in which the genesis pressure was zero. It is shown that a SOCPI model can be easily calibrated and it is capable
of simulating rock mechanical behavior closely. In addition, the ability of the model is examined by developing a synthetic material similar to Pennsylvania Blue
Sandstone. The details of calibration procedure for this sandstone are elaborated in this paper as well.
2 SAMPLE PREPARATION
The slightly overlapped circular particle interaction (SOCPI) model was selected to simulate mechanical behavior of rock using CA2 computer program. CA2
(continuum analysis, 2 dimensional) is a twodimensional computer code, developed by Fakhimi (1998). It is a hybrid explicit finite differencediscrete element code,
which can simulate the interaction of a continuum body and a circular particle domain.
To prepare the synthetic material, the surrounding flexible walls, which are made of a finite difference grid, must be created initially. The circular particles or disks are
then generated and randomly located within the confined domain between the walls. The initial radii of the disks are usually smaller to accommodate the particles inside
the desire domain. After the disk generation, they are inflated by multiplying their radii with a number. In this way, the disks may become partially overlapped. To obtain
an equilibrium configuration, the equations of motion together with a contact bond constitutive model for the disks interactions are
Figure 1. Finite difference grid together with the synthetic material in a numerical test.
solved iteratively. This is performed by implementing an explicit method, which consists of many numerical cycles. At this stage of analysis, the Kn and Ks (normal and

shear stiffness) parameters are used for elastic interaction of the disks with a friction coefficient of μ=0, i.e. the disks are allowed to slide with respect to each other and
with respect to the walls freely The interaction of the particles during the sample genesis is modeled using the following relationships:
Fn =Kn Un
(1a)
∆Fs =Ks ∆Us

(1b)
where Fn , ∆Fs , Un and ∆Us are the total normal force, the increment of shear force, the total overlap and the increment of relative shear displacement of the disks in

contact, respectively.
After equilibrium of the system and preparation of a dense compacted packing, the hydrostatic stress in the surrounding finite difference grid can be obtained which is
called the genesis stress (σ0) in this paper. The genesis stress is responsible for small overlap of the disks and plays an important role in deformation and strength
characteristics of the synthetic material as is explained later on. To complete the sample genesis, the locked in forces between the disks and stresses in the finite
difference grid are initialized to zero and the disks in contact are glued to each other by introducing the normal (nb ) and shear (sb ) bonds. This completes the sample
genesis and from this stage, the synthetic material is ready to be tested for example in biaxial loading. During the specimen loading, instead of Equation 1a, the normal
interaction between the disks is modeled with the following incremental law:
∆Fn =Kn ∆Un

(2)
This means that, in the numerical modeling, a double normal forcedisplacement law (Eqs 1a & 2) is
Page 385
used (Fakhimi 2004). With this, initial particles’ overlaps can be preserved and repulsive forces between the disks are avoided. Therefore, a synthetic sample with
slightly overlapped disks is created which does not include any initial locked in stress. A typical synthetic material together with the finite difference walls at the end of
sample genesis are shown in Figure 1.
3 DIMENSIONAL ANALYSIS
Dimensional analysis was used to obtain a series of graphs for fast and efficient calibration of the synthetic material. The microscopic parameters involved are Kn , Ks ,
μ, nb and sb . In addition, the genesis stress (σ0), average radius of disks (R) and the specimen size (D) are important factors, which could control the macroscopic
properties. The width of the specimen (in a biaxial test) or its diameter (in a Brazilian test) can be used to represent the size of a specimen. In the conducted dimensional
analysis, the inertia effects were considered negligible; the platen velocity for the specimen loading was very low compared to the speed of the compressive wave in the
sample such that the synthetic body was in a quasistatic situation during the numerical tests.
The parameters that control the elastic constants of the synthetic material, namely its Young’s modulus and Poisson’s ratio, were considered to be Kn , Ks , σ0, R and
D. These five parameters have two independent dimensions, i.e. force (F) and length (L). Therefore, based on π theorem in dimensional analysis (Sedov 1993), 5−2=3
nondimensional parameters are involved, i.e.
E′=Kn f1[R/D,Ks /Kn , σ0/Kn]

(3a)
v′= f2[R/D, Ks /Kn , σ0/Kn ]

(3b)

where f1 and f2 are dimensionless functions and E' and v ′ are apparent Young’s modulus and Poisson’s ratio of the synthetic material, defined in a uniaxial test as:
E′= ∆σaxial/∆εaxial
(4a)
v′ =−∆εlateral/∆εaxial
(4b)
The numerical model was considered as a plane strain problem. Therefore, Poisson’s ratio (ν) and Young’s modulus (E) of the synthetic material can be obtained from
the following equations:
ν=v′/(1+v′)
(5a)
E=E′(1−v2)
(5b)
Calibration charts for E′ and v′ for R/D=0.013 were developed by conducting several uniaxial tests with different Ks /Kn and σ0/Kn (Figs 2 & 3). It is evident from

Figure 2 that by increasing σ0 for a fixed Kn ,

Figure 2. Dimensionless apparent Young’s modulus as a function of Ks/Kn and σ0/Kn for R/D=0.013.

Figure 3. Apparent Poisson’s ratio as a function of Ks/Kn and σ0/Kn for R/D=0.013.
E′ increases which is due to the increase in the number of contact points in the sample. It is interesting to note that curves with σ0/Kn =0.007, in Figures 2 & 3, are
almost similar to those reported by Huang (1999) for σ0=0, which confirms the accuracy of the results. A few numerical biaxial tests and also the work by Potyondy
and Cundall (2001) show that E′ and v ′ are almost independent of R/D values provided that this ratio is kept small enough. Therefore, by introducing the macroscopic
parameters E′ and v ′ and deciding on the σ0/Kn value, Kn and Ks values can be estimated from Figures 2 & 3.
By conducting numerical tests, uniaxial strength (qu ) of the synthetic material can be obtained. The uniaxial strength is considered to be a function of eight parameters,
i.e. σ0, Kn , Ks , R, D, nb , sb and μ. Since there are two independent dimensions in this case, 8−2=6 dimensionless parameters are involved including: σ0/Kn , Ks /Kn ,
Kn R/nb , nb /sb , R/D and μ. Numerical analyses show that qu is a weak function of μ and Kn R/nb (Villegas 2004). In addition, it has been shown that a synthetic
material made of circular disks
Page 386
with a contact bond model does not show size effect with respect to its compressive strength, i.e. qu is not a function of R/D, provided that R/D is kept a small number
(Potyondy & Cundall 2001). Thus, only the three dimensionless parameters σo /Kn , Ks /Kn and nb /sb were needed to be considered. To simplify the problem, it was
decided to fix the following parameters: R/D=0.013 (Rmin =0.44 mm and Rmax=0.58 mm with a uniform random distribution of disks radii), μ=0.5, Kn =50 GPa and
nb =20,000 N/m. Thus, by conducting uniaxial tests on the synthetic material, four charts, each for a specific value of σ0/Kn (0.1, 0.08, .04, and 0.007) were obtained
(Villegas 2004). Figures 4 & 5 show two of these dimensionless charts. Comparison of these two figures shows that the qu R/nb increases by increasing the σ0/kn value.
This should be expected, because with a higher genesis stress (σ0), the number of contact points per unit area of the specimen increases which results in a stronger
specimen with higher unconfined compressive strength. It is interesting to observe the effect of Ks /Kn on the specimen strength. It is noticed that by reducing the Ks /Kn
to a value less than 0.5, qu R/nb decreases. The crack pattern in the synthetic material is also affected by the Ks to Kn ratio. In fact, it has been shown that for a small
Ks /Kn value, a diffuse crack pattern is obtained, in contrast to a shear banding failure of specimen when a higher Ks /Kn value is used (Villegas 2004).
Several numerical biaxial tests were conducted to evaluate the failure envelope of the synthetic material. The specimen strength (σ1) is a function of the confining
pressure (σ3), σ0, R, D, μ, Kn , Ks , nb and sb , i.e. σ1=g(σ3, σ0, R, D, μ, Kn , Ks , nb , sb ). Similar to the situation for qu , the biaxial strength of the specimen is a weak
function of μ, Kn R/nb and R/D. That means:
σ1R/nb =g1[σ3R/nb , Ks /Kn , σ0/Kn , nb /sb ]

(6)

Figure 4. Variation of dimensionless uniaxial compressive strength as a function of Ks/Kn and nb/sb (σ0/Kn=0.08, R/D=0.013, μ=0.5).
in which g1 is a dimensionless function. During the numerical biaxial tests, the following parameters were fixed: R/D=0.013 (with Rmin=0.44 mm and Rmax=0.58 mm),
μ=0.5, Kn =50 GPa, Ks = 27.5 GPa and nb =20,000 N/m. The results of numerical biaxial tests presented dimensionless failure envelopes for the values of σ0/Kn =0.1,
0.08, 0.06, 0.04, 0.02 and 0.007. In addition, different values of nb /sb (0.05, 0.1, 0.2, 0.3, 0.4, 0.5) were used in the analyses (Villegas 2004). Two sets of these non
dimensional failure envelopes for σ0/Kn =0.1 and σ0/Kn =0.08 are shown in Figures 6 & 7. From these figures, it is evident that increase in σ0 value, results in a stronger
specimen. In addition, it is noticed that with smaller nb /sb values, curvature of the failure envelope similar to that of a rock can be simulated. The slope of the failure
envelope in a σ1−σ3 space is equal to tan2(45+φ/2) in which φ is the rock friction angle. Considering the slope of these dimensionless failure envelopes for high values
of σ0 and nb /sb =0.05 to 0.1, it is evident that friction angles of about or more

Figure 5. Variation of dimensionless uniaxial compressive strength as a function of Ks/Kn and nb/sb(σ0/Kn=0.007, R/D=0.013, μ=0.5).

Figure 6. Dimensionless failure envelopes for different nb/sb values (σ0/Kn=0.1, R/D=0.013, μ=0.5).
Page 387
than 60° can be simulated. This has become feasible by using a nonzero genesis stress.
Two series of numerical Brazilian tests were carried out on the synthetic material to obtain its tensile strength. The dimensional analysis here is similar to that for
uniaxial tests. Thus, the important nondimensional parameters, which control the specimen tensile strength, are: σ0/Kn , Ks /Kn , nb /Sb and R/D.
Sample preparation for Brazilian test started by generating a circular wall of 40 mm in diameter (D). This wall was surrounded by an elastic finite difference grid (Fig.
8). Initially, the outside boundary of the finite difference grid was fixed in both x and y directions. The disks were created and placed randomly inside the area defined
by the circular wall subsequently. The number of disks varied between the samples depending on the porosity needed to reach the desired genesis stress (σ0). Finally,
after many numerical cycling, genesis stress

Figure 7. Dimensionless failure envelopes for different nb/sb values (σ0/Kn=0.08, R/D=0.013, μ=0.5).
Figure 8. Sample preparation for a Brazilian test.
was obtained from the finite difference grid. The rest of the procedure in the circular sample preparation is similar to that for a biaxial test, i.e. the normal and shear
contact forces between particles and stress components in the finite difference grid were initialized to zero to avoid locked in stresses, and normal bond, shear bond
and friction coefficient were introduced to the model. To apply load on the specimen, the original circular wall was removed and two new walls were generated. The
lower wall was fixed in the y direction while the upper one was given a constant velocity of 0.2×10−8 m/step in the downward direction (Fig. 9).
The first series of Brazilian tests were conducted on the synthetic material by varying the values of Ks /Kn and nb /Sb for specific values of σ0/Kn . The Ks /Kn values
used in the tests were 0.05, 0.25, 0.55, 1 and 1.5. Thus, four charts were generated, each for a value of σ0/Kn equal to 0.1, 0.08, 0.04, and 0.007. Figure 10
Figure 9. A numerical Brazilian test set up.

Figure 10. Variation of the dimensionless tensile strength as a function of Ks/Kn and nb/sb(σ0/Kn=0.1, R/D=0.013, μ=0.5).
Page 388
Figure 11. Variation of uniaxial compressive strength to Brazilian tensile strength ratio as a function of dimensionless sample size
(σ0/Kn=0.1, nb/sb=0.1, μ=0.5).
Figure 12. Variation of uniaxial strength to Brazilian tensile strength ratio as a function of dimensionless sample size (σ0/Kn=0.08,
nb/sb=0.2, μ=0.5).
shows the result for σ0/Kn =0.1. This figure shows the variation of dimensionless tensile strength of the specimen (σtR/nb ) with Ks /Kn . The general trend of the curves in

Figure 10 is similar to that for dimensionless unconfined compressive strength (Figs 4–5).
In the second series of Brazilian tests, the effect of R/D or size effect on the tensile strength of the synthetic material was investigated. Several graphs for different
values of σ0/Kn and nb /sb were obtained. Figures 11 & 12 show the typical results. In these figures, the variation of unconfined compressive strength to Brazilian tensile
strength ratio (qu /σt) is shown. Considering the fact that qu is not a strong function of R/D, it is noticed that σt reduces by decreasing the R/D value. In other words, by
increasing the specimen size or decreasing the disks radii, the synthetic sample becomes weaker in tension. This size effect of synthetic material has already been
reported by Potyondy & Cundall (2001) and Huang (1999) for a CPI model and it is confirmed here for a SOCPI material. An important observation here is that by
introducing a nonzero genesis stress, a qu /σt value of more that 25 can be obtained. This ratio is usually around 5 for a CPI model (Huang 1999), which is very low
compared to those for real rocks. Therefore, a SOCPI model is capable of simulating realistic compressive to tensile strength ratio similar to those of rocks.
4 CALIBRATION OF A SYNTHETIC MATERIAL FOR PENNSYLVANIA BLUE SANDSTONE
To observe the application of the calibration charts, presented in this paper, an effort was made to develop a synthetic material similar to Pennsylvania Blue Sandstone.
The mechanical properties of this rock as were obtained in the rock mechanics lab of Mineral Engineering Department, New Mexico Tech are as follow:
– Young’s modulus (E)=27 GPa
– Poisson’s ratio (v)=0.15
– Brazilian tensile strength (σt)=9.9 MPa
– Unconfined compressive strength (qu ) =122 MPa
The following four steps were followed to estimate the micromechanical parameters of the synthetic material similar to the sandstone.
4.1 Step 1: Selecting the nondimensional failure envelope
The first step was to search among nondimensional failure envelopes similar to those shown in Figures 6 & 7 to obtain the best match with that of the sandstone. In
Figure 7, the nondimensional failure envelope of sandstone together with those of synthetic material for σ0/Kn =0.08 is shown. From this figure, it was realized that a
curve with nb /sb =0.2 had a good match with the failure envelope of sandstone. Therefore, σ0/Kn =0.08 and nb /sb =0.2 were selected for the synthetic material.

4.2 Step 2: Finding the Ks, Kn and σ0 values
Using Equations 5a and 5b together with E=27 GPa and v=0.15 for sandstone, the apparent Young’s modulus and Poisson’s ratio could be obtained as: E′=28 GPa
and v ′=0.18. With v ′=0.18 and σ0/Kn =0.08, Ks /Kn =0.42 was found from Figure 3. The Ks /Kn =0.42 and σ0/Kn =0.08 together with Figure 2 resulted in E/Kn =0.66.
Therefore, the values of Kn =42.4 GPa and Ks =17.8 GPa were obtained. Finally, by considering σ0/Kn =0.08 and Kn =42.4 GPa, a genesis stress (σ0) of 3.4 GPa was
calculated.
4.3 Step 3: Obtaining the average radius of disks (R)
The qu /σt value for Pennsylvania Blue sandstone as measured in the laboratory is about 12. Using this
Page 389
Figure 13. Axial stressaxial strain and axial stresslateral strain curves of the synthetic material in a uniaxial numerical test.
value, nb /sb =0.2, σ0/Kn =0.08, Ks /Kn =0.42 and Figure 12, an R/D=0.013 was obtained which for a specimen width of D=40 mm resulted in the average disks radius

of 0.52 mm.

4.4 Step 4: Determining nb and sb values
An approximate value of R/nb was obtained through step 1, when the failure envelope was selected. In fact, R/nb was the scaling factor in plotting the sandstone failure
envelope in Figure 7. To obtain a better estimate of R/nb value, which incorporates the effect of Ks /Kn parameter, Figure 4 was used. For σ0/Kn =0.08, nb /sb =0.2 and
Ks /Kn =0.42, a qu R/nb equal to 5 was acquired from Figure 7. Inserting R=0.52 mm and qu =122 MPa in the qu R/nb =5, the nb =12,664 N/m was obtained which
together with nb /sb =0.2, resulted in sb =63, 320 N/m. Finally, a reasonable friction coefficient of μ=0.5 was assumed for the synthetic material. In summary, the
micromechanical parameters selected were as follow:
– σ0=3.4 GPa
– Kn =42.4 GPa
– Ks =17.8 GPa
– nb =12, 664 N/m
– sb =63, 320 N/m
– Rave=0.5 mm for D=40 mm
– μ=0.5
The synthetic material with the above parameters should have macroscopic properties similar to those of the sandstone. To verify this, numerical tests were carried out
on the synthetic material. The axial stressstrain and axial stresslateral strain curves of the synthetic material were obtained through a uniaxial test on a specimen with
D=0.04 m (Fig. 13). Curves in Figure 13 are in close agreement with those from the laboratory test (Fig. 14). The macroscopic parameters
Figure 14. Axial stressaxial strain and axial stressradial strain curves of Pennsylvania Blue Sandstone in a uniaxial laboratory test.
Figure 15. Peak and residual failure envelopes for the sandstone and the synthetic material.
found for the synthetic material through compressive and Brazilian tests are summarized as follow:
– E=27.2 GPa
– v=0.145
– σt=9.9 MPa
The above parameters are very close to those of the sandstone. The peak and residual failure envelopes of the synthetic material are shown in Figure 15, which
demonstrate excellent agreements with the results from laboratory tests. From this figure, it is realized that the curvature of the failure envelope has been simulated
accurately.
5 CONCLUSION
The calibration issue of a synthetic material, made of slightly overlapped circular particles was addressed in this paper. A dimensional analysis approach was suggested
and implemented by using many numerical experiments to obtain dimensionless graphs. These
Page 390
graphs can be used for quick and efficient calibration of a rock. The application of these graphs was demonstrated by developing a synthetic material similar to
Pennsylvania Blue sandstone. It was shown that the synthetic material obtained has many mechanical characteristics similar to the real sandstone. In addition, it was
demonstrated that the genesis stress is an important factor in dictating the deformation and strength characteristics of the synthetic material. By using high genesis
stresses which induce small overlap of the circular disks, more realistic friction angles, failure envelopes and tensile strengths for the synthetic models can be obtained
which was not possible previously by using circular disks with zero genesis stress.
REFERENCES
Bazant, Z.P. & PijaudierCabot, G. 1988. Nonlocal continuum damage, localization instability and convergence, Journal of Applied Mechanics, Vol. 55:287–293.
Cundall, P. & Strack, O.D.L. 1979. A discrete element model for granular assemblies. Geotechnique (29):47–65.
Fakhimi, A. 1998. Theory and user manual of CA2 computer program, Report no. 262 (in Farsi), Building and Housing Research Center, Tehran, Iran.
Fakhimi, A. 2004. Application of slightly overlapped circular particles assembly in numerical simulation of rocks with high friction angles. Engineering Geology 74, 129–
138.
Huang, H. 1999. Discrete element modeling of toolrock interaction. PhD dissertation, Department of Civil Engineering, University of Minnesota, USA.
Potyondy, D. & Cundall, P. 2001. The PFC model for rock: predicting rock mass damage at the underground research laboratory. Itasca Consulting Group, MN,
USA, Report no. 06819REP01200–10061R00.
Sedov, L.I. 1993. Similarity and dimensional methods in mechanics. 10th edition, CRC Press.
Vermeer, P.A. & De Borst, R. 1984. Nonassociated plasticity for soil, concrete and rock. Heron 29(3):1–65.
Villegas, T. 2004. Calibration of a discrete element model for rock: a dimensional analysis approach. MS dissertation, Department of Mineral Engineering, New
Mexico Tech, Socorro, New Mexico, USA.
Page 391
Influence of stress history on in situ and core compaction
R.M.Holt
Norwegian University of Science and Technology and SINTEF Petroleum Research, Norway
C.J.Kenter
Shell International Exploration & Production, Rijswijk, Netherlands
B.Verboom & D.Doornhof
Shell, Assen, Netherlands
5809 679 3
ABSTRACT: A set of PFC3D (Itasca 2003) simulations has been performed to provide understanding of how stress history associated with
geological processes such as uplift or overpressure may affect the compaction of a depleting reservoir. Also, because core data represent an essential
source of information for prediction of reservoir compaction, the combined effect of stress changes associated with uplift/overpressure and coring have
been simulated. The difference between core and reservoir compaction behavior is predicted to be reduced as a result of uplift/overpressure. The
predicted behavior can be understood in terms of stress changes as well as porosity changes and bond breakage occurring as part of the
uplift/overpressure process, and further damage during coring.
1 INTRODUCTION
Compaction and associated surface subsidence are important consequences of hydrocarbon reservoir depletion. Being able to predict the amount of compaction and
subsidence and the geomechanical impact on reservoir performance is becoming a more and more important part of reservoir management. In order to establish as
reliable predictions as possible, it is important to have a fundamental understanding of how reservoir rocks deform as a response to in situ stress conditions that are
changing during depletion. Since core data are the most important input to such predictions, it is also important to be able to correct core data for the effect of rock
alteration as a result of core drilling.
In past work (Holt et al. 2000), controlled laboratory experiments and discrete particle model simulations were performed to simulate in situ and core compaction.
Artificial sandstones were manufactured by cementing sand grains together under simulated in situ stress conditions, and then performing either a virgin compaction
experiment, or simulating coring by releasing the external stresses. The simulated core was then reloaded and compacted under the same stress conditions as the virgin
material. The data showed that the initial stiffness of the virgin material was significantly larger than the stiffness obtained during the core test. At high stress levels, the
virgin material would however soften, and the difference between virgin and core compaction curves would be reduced. In the numerical model (PFC3D, building on
the work by Cundall & Strack 1979), parallel bonds were inserted under in situ stress, and the simulations were otherwise following the same stress paths as the
physical experiments. Both laboratory tests and PFC3D simulations (Holt et al. 2002) showed the same kind of behavior.
In this work, we present further PFC3D simulations to study the effect of the in situ stress state being different from the forming stress. Such differences may for
instance be caused by uplift in the Earth, due to diagenesis taking place at a paleostress being larger than the present stress. It may also be caused by overpressure,
being generated prior to, or after, diagenesis. In this work, we mimic uplift and overpressure situations by either inserting parallel bonds between particles at a given
initial (“paleo”) stress state, followed by an unloading along a stress path representative for in situ conditions (here: Uniaxial strain), or the other way around: unloading
before parallel bonds are inserted. Virgin reservoir compaction is then simulated by loading in uniaxial strain conditions. We also simulate the behavior of a core taken
from the reservoir material, and compare the core compaction to the computed in situ behavior.
Page 392
2 MODELING APPROACH
A reference case for the PFC3D simulations was established by assuming an initial stress state typical for a normally pressured reservoir. The axial stress σz (mimicking
the effective vertical stress in the Earth) was chosen as 30 MPa, and the lateral (isotropic) stress σr (representing the effective horizontal stress) was set to 15 MPa.
This corresponds to a stress state at slightly less than 3 km depth. Starting from this initial reference state, three different scenarios have been simulated:
1. Cementation was performed under the reference stresses by inserting parallel bonds, followed by unloading in uniaxial strain conditions to a preset axial stress level
and a new lateral stress. This is then taken as equivalent of an initial (overpressured) reservoir stress state, corresponding to a case where diagenesis occurred prior
to pressurization.
2. The unbonded reference material was unloaded in uniaxial strain conditions to a preset axial stress level, and parallel bonds were then inserted at this state of stress.
This now represents an initial (overpressured) reservoir stress state, corresponding to a case where diagenesis occurred after pressurization.
3. The unbonded reference material was loaded to a new stress state σz & σr=30+∆σz & 15+∆σr, where parallel bonds were inserted. This material was then unloaded
in uniaxial strain conditions back to the reference stresses (30 & 15 MPa). This means that all compaction simulations start from exactly the sane state of stress, and
are therefore directly comparable. This situation is representative of uplift, where diagenesis occurred prior to stress reduction.
After the initial reservoir state has been reached in the three different scenarios, uniaxial (K0) virgin compaction simulations were performed. The materials were also
unloaded along an assumed coring stress path by reducing the vertical stress as far as possible before releasing the horizontal stress; reloaded back to the initial
reservoir stress state, and then the simulated cores were compacted (also along K0 paths).
The main part of the PFC3D simulations in this Paper was performed with input parameters as given in Table 1. The resulting macroscopic material properties when
the reference stress state (30 & 15 MPa) is used as forming stress (neither overpressure nor uplift) can be deduced from Figure 1. During virgin compaction, the initial
(first 10 MPa axial stress increase) uniaxial compaction modulus is 8.6 GPa and the initial K0=0.16, while yielding occurs around 25 MPa axial stress increase. During
core compaction, the initial compaction modulus is 4.8 GPa and the initial K0=0.22.

Table 1. Choice of parameters for the PFC3D simulations presented in this paper.
Parameter Value
Sample size 8×4×4 (5021 spherical particles)
Particle radius* 180±60 μm
Elastic modulus for loading of uncemented bonds 10 GPa
Ratio normal/shear stiffness for two uncemented particles 4.0
Interparticle friction coefficient 0.75
Number of parallel bonds in intact sample 11795
Elastic modulus for normal loading of cemented bonds 12 GPa
Ratio normal/shear stiffness for two uncemented particles 2.0
Parallel bond shear strength 20 MPa
Parallel bond tensile strength 40 MPa
Standard deviation of bond shear strength 20
Standard deviation of bond tensile strength 40
Radius of parallel bond/Particle radius 1.0
Figure 1. Axial stress increase vs. axial strain (top) and lateral vs. axial stress increase during PFC3D simulations of uniaxial virgin and
core compaction, inserting parallel bonds and starting loading from an axial stress of 30 and a lateral stress of 15 MPa.
3 SIMULATION OF OVERPRESSURE EFFECTS
Consider now the first two of the three different scenarios outlined above, namely the simulations of
Page 393
Figure 2. Stress evolution during PFC3D simulation of K0 unloading simulating pressurization. The black curve shows the stresses when
the particle assembly is cemented at the reference stress state (σz =30 & σr=15 MPa), the blue curve represents a particle
assembly which was cemented after unloading.
overpressuring. Figure 2 shows the stress evolution during K0 unloading for case 1 in the preceding Section, when diagenesis occurred prior to pressurization; i.e. when
the particle assembly was bonded at the reference stress and unloaded to reach the initial reservoir stress state, and for case 2 when the unbonded particle assembly
was unloaded and parallel bonds were inserted after the initial reservoir stress state was reached.
In both cases the axial stress is reduced far more rapidly than the lateral stress. This is a result of the Poisson effect, causing
In both cases, large overpressures (corresponding approximately to axial stress decrease) of the order 15–20 MPa or more cause the lateral ( horizontal) stress to
exceed the axial ( vertical) stress. Since Poisson’s ratio is higher for the unbonded particle assembly, the swap between the lateral and axial stresses as major and
minor principle stress occurs at larger axial stress release than when the particles were initially bonded.
Figure 3 shows the number of broken bonds during pressurization (K0 unloading) in case 1 (cementation before overpressure). It is seen that the number of broken
bonds increases with increasing overpressure. These broken bonds represent initial damage in the material prior to simulation of reservoir depletion. Also, as seen in the
same Figure, the volumetric expansion increases with increasing overpressure. This is likely to cause a reduced material stiffness (no healing of the broken bonds are
assumed). In case 2, where no bonds were inserted at the reference stress (cementation after overpressure), there are of course no bonds to break. The particle
assembly does however expand volumetrically, and more so than in the case of bonded particles.
Figure 3. Number of broken bonds and volumetric strain during PFC3D simulated K0 unloading of a material where parallel bonds were
inserted at the reference stress state (σz =30 & σr=15 MPa). Also shown is the volumetric strain in the case where K0
unloading was performed with unbonded particles.
Figure 4. Axial stress vs. axial strain during PFC3D simulated (uniaxial; K0) virgin (in bold) and core compaction for various initial
overpressures (0–25 MPa), for the case when parallel bonds were inserted at the reference stress state (σz =30 & σr=15 MPa)
prior to pressurization.
Figure 4 shows the axial stress vs. axial strain curves obtained during PFC simulations of uniaxial (K0) virgin (depletion induced) and core compaction for the material
where cementation took place prior to pressurization (scenario 1). Since the variation in overpressure generates a variation in initial reservoir stress conditions, the
curves start at different axial stress levels. Figure 5 shows the axial stress increase vs. axial strain during the same tests. The axial stress increase is a measure of the
simulated pore pressure depletion.
From these Figures, we make the following observations:
– The initial virgin compaction modulus is reduced with increasing overpressure, in particular when the overpressure becomes large relative to the interparticle bond
strength. As indicated above, this can be explained by bond breakage during pressurization, and by increasing porosity for increasing overpressure.
– The difference between virgin and core compaction decreases with increasing overpressure. As seen
Page 394
above, the damage in the virgin material increases with overpressure, so the additional damage due to coring contributes to reduce the difference between in situ
and core behavior. This can be documented by counting the number of additional broken bonds during coring simulation: This number is almost 6000 in the case of
no overpressure, but decreases to about 1500 for 25 MPa overpressure. In fact the total number of broken bonds is slightly higher (6010) in the reference case with
no overpressure than with 25 MPa overpressure (5670). This is because the coring stress path is more (but not much more) damaging than the K0 stress path
followed during simulation of pressurization. For high overpressures, the initial stress state when coring simulation starts is becoming more and more favorable for
core damage effects to be reduced, as the axial stress becomes closer to or even lower than the lateral stress. It is noticed that the initial stiffness of the cored
material is slightly lower at low than at high overpressures.
– Not shown in these figures, there are only small effects of overpressure on lateral stress response (K0). The samples that have been through a coring simulation show
slightly higher initial K0 values than the virgin samples.
– The axial stress level marking the transition between initial stiff and final soft behavior in the “in situ” (“virgin”) compaction tests appears to increase slightly with
increasing overpressure. This is primarily a result of softer behavior during initial compaction.
A similar set of PFC3D simulations has been performed on the basis of scenario 2; i.e. with parallel bonds being inserted after unloading the unbonded particle
assembly (cementation after pressurization). The axial stress vs. axial strain behavior in the virgin and core compaction tests is illustrated in Figure 6 and Figure 7. The
results are similar to those obtained for
Figure 5. Axial stress increase vs. axial strain during PFC3D simulated (uniaxial; K0) virgin (in bold) and core compaction for various
initial overpressures (0–25 MPa), for the case when parallel bonds were inserted at the reference stress state (σz =30 & σr=15
MPa) prior to pressurization.
scenario 1 above, and can be summarized as follows:
– The initial virgin compaction modulus decreases with increasing overpressure, but the effect is less pronounced than in case 1 above. In the present situation the only
source of stiffness reduction is the porosity enhancement caused by the overpressure.
– The difference between virgin and core compaction decreases with increasing compaction. In this case, the main reason is that the stress release during coring is less
damaging when the overpressure is high, because of low stress level, and also more isotropic stresses than with no overpressure. For example, the unloading—
reloading cycle associated with coring and core reloading creates only about 1200 broken bonds at 25 MPa overpressure, compared to 6010 n the reference case.
This is also the reason that the core stiffness increases with increasing overpressure.
– There is only negligible difference in the lateral vs. axial stress behavior (not shown in the Figures) as a result of varying overpressure.
Figure 6. Axial stress vs. axial strain during PFC3D simulated (uniaxial; K0) virgin (in bold) and core compaction for various initial
overpressures (0–25 MPa), for the case when no bonds were inserted at the reference stress state (σz =30 & σr=15 MPa) prior
to pressurization.
Figure 7. Axial stress increase vs. axial strain during PFC3D simulated (uniaxial; K0) virgin (in bold) and core compaction for various
initial overpressures (0–25 MPa), for the case when no bonds were inserted at the reference stress state (σz =30 & σr=15
MPa) prior to pressurization.
Page 395
– There is only a slight effect of overpressure on the apparent yield point in the virgin compaction curves.
4 SIMULATION OF UPLIFT
The third set of simulations was performed by bonding the particles together at an elevated stress state and then unloading until the reference stress state (σz=30 &
σr=15 MPa) is reached. This is a simulation of uplift (scenario 3). The stress evolution during the simulation is illustrated in Figure 8.
The resulting axial stress vs. axial strain curves from uniaxial virgin and core compaction simulations are shown in Figure 9, while the lateral vs. axial stress behavior is
shown in Figure 10. The main observations are:
– The initial virgin compaction modulus is reduced when uplift becomes large relative to interparticle bond strength. This is explained by an increasing
Figure 8. Axial and lateral stress in PFC3D simulation where parallel bonds were inserted at 45 MPa axial and 18.5 MPa lateral stress,
and uplift was simulated by unloading in uniaxial strain (K0) conditions until the reference stress state (σz =30 & σr=15 MPa)
was reached. After that, coring and core reloading was simulated, finally followed by a core test (also following a uniaxial
strain (K0) stress path).
Figure 9. Axial stress vs. axial strain during PFC3D simulations of uniaxial (K0) virgin (bold curves) and core compaction. The numbers
next to the virgin curves indicate the uplift in [MPa].
number of broken bonds as a result of increasing uplift (see Fig. 11). On the other hand, the more uplifted materials have been exposed to a higher stress during
cementation and are therefore more compacted. This lower porosity is reflected in an increase of virgin stiffness at high stress levels with increasing uplift.
– The simulated core stiffness decreases with increasing uplift. This is mainly a result of the increased amount of broken bonds present before stress release associated
with coring starts. The difference between virgin and core compaction appears to increase with increasing uplift.
– Virgin lateral stress response (K0) decreases while core K0 increases with increasing (and modest) uplift.
– Increasing uplift leads to an increase in the stress marking the transition between initially stiff and soft behavior in the virgin stress—strain curve. This may be
explained by an improved strength because of more compacted material when bonds are inserted.
The possible scenario when uplift occurred prior to cementation has not been simulated.
Figure 10. Lateral vs. axial stress during PFC3D simulations of uniaxial (K0) virgin (bold curves) and core compaction. The numbers next
to the virgin curves indicate the uplift in [MPa].
Figure 11. Development of broken bonds during uplift simulation for the cases of a 20 MPa uplift (starting from 50 MPa axial stress) and
5 MPa uplift (starting from 35 MPa axial stress).
Page 396
The results of the simulations above can all be understood in terms of two main mechanisms causing alteration of mechanical behavior: Porosity changes, either
enhancement associated with overpressure, or porosity reduction associated with the burial prior to uplift, have an impact on rock strength and stiffness, both if
cementation occurred prior to or after the geological event leading to pressurization/uplift. Second, but from the simulations even more important: Damage development
associated with stress release during uplift or pressurization of a cemented material leads to a softer behavior than in case of a reference material at normal stress
conditions. In reality, the question is to what extent such damage is healed or if creep over geological time will wipe out this damage. These simulations form a basis for
understanding the mechanics of uplifted or overpressured reservoirs, but the geological processes involved in the stress history of the reservoir may be much more
complex than simulated here. If these processes are known they need to be accounted for in field specific analyses. For instance, using a timedependent contact law in
PFC, one may simulate processes such as creep. No experimental data were collected in order to substantiate the findings from the numerical simulations. This should
be done in future work.
5 SUMMARY
The PFC3D simulations point to three important mechanisms being responsible for effects of geological processes such as overpressure and uplift on reservoir and core
compaction. These mechanisms are associated with stress evolution, porosity alteration, and damage associated with interparticle bond breakage during.
For an overpressure situation, in our case, we simulated rock being formed under normal (paleo) stress and then unloaded in uniaxial strain boundary conditions
before reaching the initial stress state of a depleting reservoir. If cementation occurred prior to overpressure, then bonds would break during the stress reduction
associated with the overpressure process. Also, the porosity would be reduced. The result is a softer material than that formed at initial paleostress conditions. If
cementation occurred after the overpressure situation was established, the overpressured material would still be softer because of the increased porosity during
pressurization. The difference between the compaction of cored material and reservoir material would be reduced in both cases, because the stress evolution during
pressurization led the material to a less anisotropic and hence less damaging in situ stress state.
If the reservoir was uplifted from a larger depth, and cementation occurred prior to uplift, then the stiffness was still found to be reduced as a result of bond breakage
during uplift, in particular if stress release during uplift becomes large relative to interparticle bond strength, This material would however have improved strength
because it is more compacted when bonds are inserted. The simulated core becomes softer with increasing degree of uplift, so that the difference between virgin and
core compaction increases with increasing uplift for modest degrees of uplift.
One should be aware that these simulations only describe simplistic assumptions about the geological processes involved. Further refinement and complexity needs to
be added if particular field situations are to be analyzed. We do feel however that the discrete particle simulations performed gives a basic insight into the mechanical
processes involved.
ACKNOWLEDGEMENT
This work has been supported by Shell through NAM (Nederlandse Aardolie Maatschappij).
REFERENCES
Cundall, P.A., Strack, O.D.L. 1979. A discrete numerical model for granular assemblies. Geotechnique 29:47–65.
Holt, R.M., Brignoli, M. & Kenter, C.J. 2000. Core Quality: Quantification of coring induced rock alteration. Int. J. Rock Mech. & Min. Sci. 37:889–907.
Holt, R.M., Doornhof, D., Kenter, C.J. 2002. Use of discrete particle modeling to understand stressrelease effects on mechanical and petrophysical behavior of granular
rocks. In Numerical Modeling in Micromechanics via Particle Methods, ed. H.Konietzky; Rotterdam; A.A.Balkema, pp. 269–276.
Page 397
Numerical modeling of weak sand behavior
R.Narayanasamy
HeriotWatt University, Edinburgh, UK
5809 679 3
ABSTRACT: Numerical simulation of the physical behavior of uncemented and weakly cemented sands was carried out using modeling software
based on the Discrete Element Method (DEM). A procedure was developed for triaxial and wave propagation tests using the numerical model.
Simulations were carried out to investigate how deformation was influenced by various material properties, such as density, grain stiffness, cement
stiffness and cement strength. The stiffness of the grains and cement controlled Young’s modulus while intergranular friction and the cement strength
affected the strength of the specimen. The amount of cement at contact points and timing of cementation also affected the stressstrain behavior of the
specimen. The numerical results compared well with the laboratory experiments. The density, stiffness of the material and cement, and the amount of
cement affected compression and shear wave velocities (Vp and Vs). A study on the frequency of a propagating wave suggested that specimens with
small values of L/D ratio might yield better waveforms and less discrepancy in wave velocities. Stress distribution around a hole in anisotropically
stressed specimen showed higher values of stress concentrations compared to analytical results. Stress concentrations near a hole were also affected
by bond strength.
1 INTRODUCTION
Several experimental studies had been conducted to study the behavior of artificial samples of sand, unconsolidated sand, weakly cemented sand and well cemented
sand by many researchers in the past. Samples in these studies were constructed using clean and round sand particles mixed with a measured amount of cementing
material. The advantage of using artificial samples is the ability to systematically vary critical properties independent of each other and study the effect of each property
The properties that are of interest are porosity, grain material, geometry, cement material, cement distribution and specimen size. Artificial specimens in previous studies
were made of sand grains or glass beads cemented with different materials. One such cementing material was sodium silicate (Holt et al. 1993, Nakagawa & Myer
2001).
Bernabè et al. (1992) measured the variations in strength, dilation and stressstrain behavior with cement content and observed the changes due to small amount of
cement deposited at graintograin contacts. Tests using glass beads cemented with ice (Liu et al. 1991, Tutuncu et al. 1997) and epoxy (Dvorkin et al. 1994) have
shown sharp increases in wave velocities due to small amounts of cement at grain contacts. Yale et al. (1995) associated the differences in static/dynamic moduli and
the dependence of Poisson’s ratio on porosity/cementation with the nonlinear mechanical behavior of the intergranular cement. The studies have led to a general
awareness that the influence of cement depends strongly on its distribution near the grain contacts.
Numerical modeling using Discrete Element Method (DEM) models the grains and cement in a physically realistic way. PFC3D (Itasca 1999) was used for
simulation of a numerical model to match the mechanical behavior of an artificially cemented sand sample made in the laboratory Wave propagation studies and the
effect of a hole in the model were analyzed using the calibrated model.
2 DESCRIPTION OF SAND SAMPLE
Test specimens used in the laboratory were created in a companion study (Camacho 2001) by artificially cementing commercially available sand known as “Hickory
sand” using sodium silicate. The sand was primarily quartz and the grains were well rounded with smooth surfaces, such that using spherical particles in PFC was an
adequate approximation. The grain size distribution (Fig. 1), initial porosity (39%), density (2650 kg/cu.m.) and confining
Page 398
Figure 1. Grain size distribution (from Camacho 2001).
Table 1. Values used in the numerical simulation of an uncemented specimen.
Description Numerical Model A Numerical Model B

Specimen size (cm) 1.5×1.5×3.0 (L×W×H) 3.0×3.0×3.0 (L×W×H)
Normal and shear stiffness of balls 4.5e6 N/m 4.5e6 N/m
Normal stiffness of lateral walls 4.5e6 N/m 4.5e6 N/m
Normal stiffness of confining walls 4.5e5 N/m 4.5e5 N/m
Friction coefficient of ball surface 0.7 0.7
Number of balls 16,422 65,678
Number of contacts (Pc=6.895 MPa) 69,280 277,708
pressure (Pc=6.895 MPa) applied on the physical specimen were obtained from a companion study (Camacho 2001).
The numerical model used the same parameters as the physical specimen except the 2.54 cm diam.×5.08 cm cylindrical size. Two numerical models were used in the
PFC3D analyses of sizes 1.5×1.5×3.0 cm (Model A) and 3.0×3.0×3.0 cm (Model B). The values used in the numerical simulations of both the models are shown in
Table 1.
3 NUMERICAL MODELING SEQUENCE
The minimum data required to create a numerical model that represents a physical specimen were: grain size distribution, specimen dimensions, density of balls and
initial porosity. The Young’s modulus (E) of the physical specimen was obtained from the stressstrain curve of a physical test. Using this value and the average ball
radius (ravg), the normal (Ball_kn) and shear (ks) stiffness of the balls were determined as Ball_kn=ks=4*E*ravg. Instead of using a constant stiffness value for all
balls as above, a different stiffness for each ball in the model can be computed using actual ball radii (r). Tests carried out using both the methods had shown little
difference in the physical behavior and, hence, constant ball stiffness was used.
The stiffness of axial walls was set same as the ball stiffness. A good initial guess for the stiffness of the lateral walls was 1/10th of the value of the walls at the ends.
This lateral wall stiffness can be modified to control the variation in porosity within the specimen during hydrostatic loading. The exact value of intergranular friction
coefficient was determined by comparison with physical tests, but a value between 0.50 and 0.70 was a good initial guess. Using the values assumed as described
above, numerical simulation of a triaxial test was carried out and matched with physical behavior by following the procedure given below.
1. Prepare the specimen based on parameters same as physical specimen and simulate triaxial behavior. Compare the simulated behavior with the laboratory results.
2. Increase the stiffness of balls if the slope in the elastic region is lower than that of physical test. Decrease the value otherwise. Rerun the triaxial simulation and iterate
until simulated and measured stressstrain curves match fairly well in the elastic region.
3. Increase the value of friction if the “peak stress” is lower than measured in the physical test. Decrease the value otherwise. Rerun the triaxial simulation and iterate
until simulated and measured “peak stress” values are matched.
4. Adjust the stiffness of lateral walls to obtain simulated porosity changes that match the physical test.
The numerical model was generated in PFC through a series of ball size manipulations and stabilizations. The physical specimen was created by pluviating the sand in
air (Camacho 2001). However, to create specimens with low values of porosity for a given grain size distribution, the model generation must simulate the actual
physical processes carried out in preparing the specimen.
After matching the stressstrain behavior of a numerical model with an uncemented laboratory specimen, the numerical model was used for other tests like wave
propagation, stress anisotropy and hole effects. Also, cement at intergranular contacts was modeled using parallel bond option in PFC to analyze the behavior of
cemented laboratory specimens.
Wave propagation tests can be carried out by different methods in PFC. Commands to control normal velocity, rotational velocity and displacement of balls or walls
provided the flexibility to simulate virtually all aspects of wave propagation. For this work, waves were generated by controlling the velocity of selected balls in the
desired direction at one end of specimen.
Page 399
The velocity in the desired direction of other selected balls at different locations within the specimen were monitored to interpret the propagation of a waveform. The
ability to monitor desired parameters at various locations within the numerical model helped our understanding of the physical processes during wave propagation within
the specimen.
4 STRESSSTRAIN BEHAVIOR
4.1 Uncemented model
The stressstrain behavior of numerical model was mainly affected by ball stiffness and friction (Figs 2 & 3). Secant Young’s modulus was measured from the stress
strain curve between a strain of 0% and 0.5%. A specimen with greater ball stiffness reached a stress level with lesser amount of strain (hence a rigid, stable framework
formed faster with less strain), whereas a specimen with lower ball stiffness experienced greater amount of ball reorientation before reaching the same stress level. The
porosity of the specimen measured at 1% strain was 37.27%, 37.63% and 37.86% for ball stiffness values of 2e6, 3e6 and 4.5e6 N/m respectively. The relatively
lesser reduction in porosity for greater ball stiffness showed lesser ball reorientation. An increase in friction between the balls resulted in a stronger specimen as
measured by the peak stress prior to yielding. At the same time, Young’s modulus of the specimen was relatively unaffected by changes in friction.
Using the procedure described earlier, the numerical model best matched the behavior of the uncemented physical specimen with normal and shear ball stiffness of
4.5e6 N/m and a friction coefficient of 0.7.
4.2 Cemented model
The stressstrain behavior of a model with parallel bond was affected by the bond stiffness, strength and size of the bond. The bond stiffness acts in parallel with the ball
stiffness. The strength of a model was affected by the failure of interparticle bonds, which in turn, depended on the bond strength and bond radius. The bond radius
was estimated to vary from 0.04 to 0.25 times the grain radius from the images of artificially cemented lab specimen.
The parallel bond stiffness (pb_kn, ks) acted similar to ball stiffness by increasing the Young’s modulus of the specimen (Fig. 4). The bond strength was high enough
so that bonds were not broken during the test. For constant values of ball and parallel bond stiffness, the peak sample stress increased almost linearly with increasing
bond strength (Fig. 5). It was evident that a bond strength (pb_n, s) of 1e7 Pa or less added little strength to the unbonded sample, indicating that all the
Figure 2. Effect of ball stiffness on deformation behavior of uncemented sand for Pc=6.895 MPa and friction μ=1.5.
Figure 3. Effect of friction on deformation behavior of uncemented sand for Pc=6.895 MPa.
Figure 4. Effect of parallel bond stiffness on deformation behavior of cemented sand for Pc=6.895 MPa.
Figure 5. Effect of parallel bond strength on deformation behavior of cemented sand for Pc=6.895 MPa.
Page 400
bonds broke as soon as a small strain was applied. Once the bonds were broken, sample strength came from the frictional characteristics of the graintograin contacts.
The parallel bond stiffness and strength in normal and shear directions were assumed to be same.
By choosing appropriate values of parallel bond strength and parallel bond stiffness/unit area for a given parallel bond radius, peak strength and the Young’s modulus
of a cemented model was matched with laboratory results of cemented specimens. Parallel bond radius (pb_r) in PFC is defined as the ratio of the radius of the bond
and the minimum radius of the two bonded balls. An increase in parallel bond radius increased the strength of the bond, by lowering the stress it must carry for a given
force carried between balls in contact. It also increased effective bond stiffness as compared to ball stiffness because bond stiffness in PFC is a function of bond size as
well (the units of bond stiffness is Pa/m, which is equivalent of Stiffness per unit area). Hence, an increase in bond radius (r) increased the bond stiffness by r2 since
effective bond stiffness=bond stiffness (per area)×area (πr2). The secant Young’s modulus measured between 0% and 0.4% strain showed an increase from 1.806
GPa to 2.493 GPa when parallel bond radius was increased from 0.1 to 0.6 in a cemented model with high bond strength value of 3e9 Pa (Fig. 6). The initial stress
strain behavior was controlled by both ball and bond stiffness but is dominated by ball stiffness. However, in the interval beyond 0.5% strain, the tangent Young’s
modulus measured between 1% and 3% strain was linearly dependant on the parallel bond radius, indicating that the behavior was controlled by the bond size.
The stressstrain behavior was different for samples cemented before and after applying confining pressure. The deformation behavior was affected by the timing of
the introduction of bond depending upon the stiffness of the bond (Fig. 7). The relative difference in the stress for the same amount of strain was small for lower values
of parallel bond stiffness, suggesting the ball stiffness was dominating the behavior. However, for high bond stiffness, the timing of bonding became important because
the bonds carried the load in parallel with the balls. Presumably, the increase in sample stiffness gained by cementation after hydrostatic loading was a reflection of an
increased number of contacts and shorter bond lengths, giving a stiffer geometric arrangement of balls. Natural and artificial specimens with similar properties were
expected to behave differently for the same reason. Increasing the number of contacts will also increase the number of bonds. However, the difference due to timing of
cementation for weakly cemented laboratory specimens was estimated to be insignificant for practical purposes.
The effect of ball density on the stressstrain behavior of the specimens was negligible as would be expected for quasistatic deformation. The density of
Figure 6. Effect of parallel bond radius on cemented sand for Pc=6.895 MPa.
Figure 7. Effect of cementation before and after loading on deformation behavior for bond strength of 3e9 Pa (CAL—Cementation After
Loading; CBL—Cementation Before Loading).
the balls was changed between 2050 and 3050 kg/cu.m and the difference in stressstrain behavior was insignificant. This may be due to the slow process of
deformation during a triaxial test, where acceleration of the balls was not a significant component in displacing the balls.
5 WAVE PROPAGATION
5.1 Effect of micromechanical properties
The focus on the numerical modeling of wave propagation characteristics in weak sand was to understand how micromechanical parameters such as ball stiffness,
friction, and bond stiffness/unit area, strength and bond radius affect measured wave velocities. In uncemented model, higher ball stiffness resulted in higher wave
velocities because they cause an increase in Young’s modulus (Fig. 8). Wave velocities from a numerical model with ball stiffness of 4.5e6 N/m and friction of 0.7
(obtained by matching stressstrain behavior of a laboratory specimen) were measured as Vp=1391 m/s and Vs=788 m/s. It showed approximately 4% error when
compared with wave velocities measured in the laboratory specimen as Vp=1448 m/s and Vs=823 m/s. The P and Swaves were simulated
Page 401
Figure 8. Effect of ball stiffness on wave velocities (uncemented, friction=1.5).
Figure 9. Effect of parallel bond stiffness on wave velocities.
as onehalf sinusoidal pulses at a frequency of 800 kHz in desired direction.
Parallel bonds were introduced in the calibrated uncemented model at a confining pressure of 6.895 MPa with bond strength of 3e9 Pa and parallel bond radius of
0.25 and the model was tested with onehalf sinusoidal pulses at a frequency of 800 kHz. Below a threshold parallel bond stiffness value of approximately 1e13 Pa/m,
the effect on wave velocities was insignificant (Fig. 9). Below this threshold value, the ball stiffness controlled the wave velocities and above which, the bond stiffness
controlled the wave velocities. Increasing the bond stiffness or the radius increased the velocity, but only when the effective bond stiffness was comparable to or greater
than the ball stiffness, otherwise ball stiffness dominated. In the same model, increasing the bond radius increased both Vp and Vs and resulted in a monotonic decrease
in Vp/Vs (Fig. 10). The bond stiffness was 4.5e12 Pa/m and bond strength was 3e9 Pa.
Both P and Swave velocities decreased when density of balls was increased (Fig. 11). The results agreed with the predicted values obtained using Equations 1 &
2.
(1)
Figure 10. Effect of parallel bond radius on wave velocities.
Figure 11. Effect of density of balls on uncemented specimen.
(2)
In both the uncemented and cemented models, an increase in confining pressure resulted in a minor increase in wave velocities and a decrease in Vp/Vs. Increasing the
confining pressure from atmospheric to 6.895 MPa resulted in an increase of 3.73% in Vp and 8.54% in Vs for an uncemented specimen. There was no effect of
friction on wave velocities for uncemented samples. Likewise, other strength parameters such as bond strength did not affect wave velocities.
5.2 Frequency dependant velocity dispersion
The wave velocities are frequency independent for dry rocks (Tutuncu et al. 1988). In fluid saturated rocks, however, compression and shear velocities are frequency
dependent even in the frequency range where scattering is negligible (Spencer 1981, Murphy 1982 & 1984, Winkler 1983 & 1985, Toksöz et al. 1979, Tutuncu et al.
1998). In general Vp, Vs and Vp/Vs increase with frequency (Moos & Zoback 1983, Murphy 1984, Goldberg & Zinsner 1989). Tutuncu et al. (1998) analyzed this
behavior and attributed the frequency dependence of wave velocity to the behavior at contact points.
PFC provides the flexibility to monitor the waveform at different locations within the specimen and velocities measured using such intermediate
Page 402
Figure 12. Effect of frequency variations on wave velocities for wave propagation along the length of the model.
Figure 13. Comparison between wave velocities measured in horizontal and vertical directions with different frequencies.
locations between transmitter and receiver can be used to analyze the propagation of waves. Also, an internal measurement of velocity avoids the interference of
external factors such as the bond between a specimen and end platen. Figure 12 shows the effect of frequency on the measured wave velocities in a numerical model of
uncemented specimen modeled using linear contact model and intergranular friction in PFC as described earlier. Similar behavior was noticed in parallel bonded
specimen as well. Most of this frequency dependence of velocity was attributed to specimen size effects at long wavelengths.
Instead of transmitting elastic waves along the vertical direction, which is typical of the laboratory procedure used to test a formation core, P and Swaves were
transmitted in a horizontal direction. Figure 13 shows the comparison between wave velocities measured in horizontal and vertical directions. The wave velocities
measured at frequencies less than 1 MHz are generally higher and tend to be similar to those measured at higher frequencies when measured in the horizontal direction
compared with vertical direction (Fig. 13). It appeared that boundary and size effects were significant while measuring wave velocities in laboratory samples. Hence,
specimens with smaller L/D ratios might yield better waveforms and less discrepancy in wave velocity measurements. Note that standard practice for measuring
strength in triaxial tests calls for the use of specimens with L/D ratios of at least 2, dimensions that may make it difficult to measure wave velocities accurately. Further
analysis indicated that higher area of excitation and reception in the laboratory experiment could be expected to produce fewer dispersion related anomalies. The effect
of wave travel in a simulated numerical model of an infinite medium is to be further analyzed.
6 STRESS DISTRIBUTION AROUND A HOLE
6.1 Comparison of stress distribution in numerical model and analytical results
The stress distribution around a hole in an anisotropically stressed weakly cemented sample was analyzed using PFC. The numerical model was a 3 cm cube with 0.25
cm radius cylindrical hole in the center along Z direction from face to face. The physical properties were matched with the behavior of uncemented laboratory sample
prior to applying anisotropic stress. Parallel bonds were added and a hole was created by removing the balls. The PFC model was compared with analytical solutions
to compare the stress distribution in the numerical model. For an infinitely large plate with applied stress of σxx and σyy and a hole of radius “a”, stresses in polar
components for plane strain are given by (Timoshenko & Goodier 1970):
(3)
(4)
(5)
where σrr is the radial stress, σθθ is the circumferential stress, σrθ is the shear stress, θ is the counterclockwise angle from the Xaxis, and r is the radial distance from
the origin. The stress distribution around a borehole of size 0.25 cm in an anisotropically stressed twodimensional body was calculated using the above equations and
converted into Cartesian coordinates as (Timoshenko & Goodier 1970):
σxx=σrr cos2 θ+σθθ sin2 θ−2σrθ sinθcosθ
(6)
Page 403
Figure 14. Stress distribution around borehole for σxx=σzz=6.9 MPa (1,000 psi) and σyy=10.3 MPa (1,500 psi), values in psi. a, b, c,—from
analytical solution; d, e, f—from numerical modeling.
σyy=σrrsin2 θ+σθθ cos2 θ+2σrθ sinθcosθ
(7)
σxy=(σrr−σθθ)sinθcosθ+σrθ(cos2 θ−sin2 θ)
(8)
The resultant normal and shear stresses are shown in Figure 14a, b and c for farfield applied stresses of σxx=6.9 MPa and σyy=10.3 MPa. σxx, σyy and σxy Of the
balls were obtained directly from PFC model.
Page 404
The constant stress boundary condition on the walls of the numerical model were σxx=σzz=6.9 MPa and σyy=10.3 MPa. The average values of σxx, σyy and σxy for
balls with their centers lying with in 3 ×3×2 mm grid blocks were computed for the PFC model in a layer parallel to Z direction in the middle of the cube. Figure 14d, e
and f shows the normal and shear stress distribution around the hole in PFC model. The above plots show the values of stress in psi.
Compare the nature of the stress in regions A (θ=0°) and B (θ=90°). For region A, the analytical model showed increased σxx (high ~2,400 psi) and decreased σyy
(low ~300 psi) due to the development of stress concentrations. In region B (θ=90°), the analytical model showed reduced σyy (low ~300 psi) and increased σxx (high
~1,300 psi). The PFC results showed the same trend but were different in detail because of the granular nature of the model. In region A, σyy was elevated (high
~3,400 psi) and σxx was reduced (low ~400 psi). In region B, σyy was reduced (low ~1,000 psi) and σxx was elevated (high ~2,000 psi). In both locations A and B,
the circumferential stress component was most compressive.
This point can be further illustrated by looking at plots of average radial and circumferential stress as a function of θ in near wellbore region (Fig. 15). For a numerical
model, σxx, σyy, σxy and the 3D position of each ball were obtained directly from PFC. The stresses from analytical solution were computed by averaging the stresses
in 3 mm grid blocks throughout the sampled layer using Equations 6–8 calculated at different points. Then the average normal and shear stresses for a region that lies
within a radius of 1 cm around the borehole in the 3 mm slice at the center was computed for both cases. The stresses were converted into polar coordinates to obtain
radial (σr) and hoop (σθ) stresses around the borehole using the relations (Timoshenko & Goodier 1970):
σr=σxx cos2 θ+σyy sin2 θ+2σxy sinθcosθ
(9)
σθ=σxx sin2 θ+σyy cos2 θ−2σxy sinθcosθ
(10)
Again, the results were highly variable due to the discrete nature of grains, but a bestfit curve to the results showed comparably similar results obtained from the
analytical solution for a uniform material (Fig. 16). The discrete nature of the numerical model showed high variations in ball stresses and almost twice the prediction
based on analytical solutions. However, it must be noted that the above comparison between analytical and numerical model is not appropriate since the numerical
model is not infinitely large. Analyzing a large numerical model is difficult due to computational limitations. Further analysis is required to confirm whether similar results
are noticed in such large numerical models.
Figure 15. Hoop stress in a numerical model with a borehole.
Figure 16. Radial stress in a numerical model with a borehole.
6.2 Stress distribution around borehole and bond strength
The stressstrain behavior of a numerical model was affected by the strength of parallel bonds as discussed previously. When all the bonds in the specimen failed, the
model behaved like an uncemented specimen. The study of the behavior of bonds in the vicinity of the borehole due to applied stress is useful to analyze the
characteristics of the nature of the failure. When the bond strength was low enough to result in bond failure, the borehole started to collapse and the differential stress
was reduced in the vicinity of the wellbore. Figures 17–20 shows the stress distribution obtained in a PFC2D simulation for low (3e6 Pa) and high (3e9 Pa) values of
strengths. For low values of bond strength, the normal stresses along X and Y direction showed uniform distribution around the borehole and stress concentrations were
absent in the minimum stress direction. For values of strength lower than this, the bonds broke and borehole collapse was noticed. Exact comparison of the values of
strength from the PFC2D simulation shown here with the 3D results could not be made because the input parameters in PFC2D simulation were different. The PFC2D
simulation was carried out using the following values: Length=10.6 cm, width=2.8 cm, minimum ball radius=0.075 mm, maximum ball radius=0.22 mm, normal stiffness
Page 405

Figure 17. Normal stress (psi) σxx for bond strength= 3e6 Pa.

Figure 18. Normal stress (psi) σyy for bond strength= 3e6 Pa.

Figure 19. Normal stress (psi) σxx for bond strength= 3e6 Pa.

Figure 20. Normal stress (psi) σyy for bond strength= 3e9 Pa.
of balls=5e9 N/m, shear stiffness of balls=0 N/m, porosity=16%, radius of the hole=4 mm, stress in X direction=6.9 MPa, stress in Y direction=13.8 MPa, parallel
bond stiffness=5e12 Pa/m and parallel bond radius (pb_r)=0.25.
7 CONCLUSIONS
The behavior of uncemented and weakly cemented sand was simulated numerically using PFC, a Discrete Element Method (DEM) based simulation software. The
effect of various micro parameters, deformation behavior and stress wave velocities in the numerical model was analyzed. The behavior of anisotropically stressed
material with a borehole was simulated and analyzed. The summary of the study and conclusions are presented below.
The ball stiffness affected the initial stressstrain behavior and, hence, the Young’s modulus of the specimen. Higher ball stiffness resulted in higher Young’s modulus,
Vp, Vs and Vp/Vs. Increasing interparticle friction increased the peak strength of the model. An increase in ball stiffness and density resulted in higher wave velocities.
Wave velocities were not affected by a change in the value of friction, when compared at the same confining pressure.
Parallel bonds under some conditions increased peak strength of a specimen. Strength of the model depended on the bonding material strength and the bond radius.
Parallel bonds also increased Young’s modulus of the model through bond material stiffness and bond radius. Deformation behavior, Young’s modulus and the strength
of the numerical model were affected by the timing of cementation. Cementing before applying confining pressure resulted in a lower stiffness material than when cement
was applied under stress. An increase in bond stiffness and bond radius resulted in higher wave velocities only when the bonds were dominant in altering the stress
strain behavior of the numerical model.
The difference between wave velocities measured at low frequencies (<1 kHz) and high frequencies (>1 MHz) decreased when they were measured along the width
compared with the length of the model. It showed that specimens with smaller L/D ratios yield better waveforms and less variation due to frequency of the propagating
wave.
The results indicated that strength of a specimen contributed by cementation alone was not predictable using wave velocities unless cement was stiffer than grain.
Also, in such cases, the wave velocities were affected by the bond size. The deformation behavior may be affected by many independent factors and, careful analysis at
laboratory is required to further refine our understanding of granular materials.
The stress distribution around a hole in an anisotropically stressed weakly cemented sample was analyzed using PFC. The results showed similar trend compared
with analytical solution for a uniform material. However, high variations in ball stresses were noticed and nearwellbore stresses were almost twice the prediction based
on analytical solutions. In anisotropically stressed numerical models with low values of bond strength, uniform stress distribution around the borehole was noticed and
stress concentrations were absent. However, models with high bond strength showed distinct presence of stress concentrations near the borehole.
Page 406
ACKNOWLEDGEMENTS
I acknowledge my past supervisors Dr. Jon E.Olson, Dr. Jon Holder, Dr. Alan F.Rauch and coresearcher Ms. Beatriz Camacho during my studies at The University
of Texas at Austin. I thank National Science Foundation, USA for financial assistance through grant no. CMS9978662. I acknowledge my current supervisors Dr.
David Davies and Dr. Jim Somerville for their encouragement and assistance to publish this paper. I thank Alumni fund, HeriotWatt University for financial assistance.
REFERENCES
Bernabè, Y., Fryer, D.T., Hayes, J.A. 1992. The effect of cement on the strength of granular rocks. Geophysical Research Letters, 19(14):1511–1514.
Camacho, B.I. 2001. Mechanical properties of granular material as measured from velocity of shear waves. Report, Austin: The University of Texas at Austin, USA.
Dvorkin, J., Nur, A., Yin, H. 1994. Effective properties of cemented granular materials. Mechanics of Materials, 18: 351–366.
Goldberg, D., Zinsner, B. 1989. Pwave attenuation measurements from laboratory resonance and sonic waveform data. Geophysics, 54:76–81.
Holt, R.M., Unander, T.E., Kenter, C.J. 1993. Constitutive mechanical behavior of synthetic sandstone formed under stress. Int. J. Rock Mech. & Min. Sci., 30(7):719–
722.
Liu, X., Dvorkin, J., Nur, A. 1991. The effect of intergranular cementation on rock properties: Experimental stady. AGU Annual Meeting, San Francisco; abstract.
Moos, D., & Zoback, M.D. 1983. In Situ studies of velocity in fractured crystalline rocks. Journal of Geophysical Research, 88(B3):2345–2358.
Murphy, W.F. III. 1982. Effects of partial water saturation on attenuation in Massilon sandstone and Vycor porous glass J. Acoustic Soc. Am., 71(6):1458–1468.
Murphy, W.F. III. 1984. Seismic to ultrasonic velocity drift: Intrinsic absorption and dispersion in crystalline rocks Geophysical Research Letters, 11(12):1239–1242.
Nakagawa, S., Myer, L.R. 2001. Mechanical and acoustic properties of weakly cemented granular rocks. In D.Elsworth, J.P.Tinucci, & K.A.Heasley, (eds), Rock
Mechanics in the National Interest: Proceedings of the 38th Annual U.S. Rock Mechanics Symposium. 1:3–10. Rotterdam: Balkema.
Narayanasamy, R., Olson, J., Holder, H., Rauch, A., Camacho, B. 2002. DEM study of wave propagation in weak sandstone. In B.K.Cook & R.P.Jensen (eds),
Proceedings of the Third International Conference on Discrete Element Methods, held in Santa Fe, New Mexico on September 23–25. ASCE.
Narayanasamy, R. 2002. The Effect of cementation on Granular Materials. MS Thesis. Austin: The University of Texas at Austin, USA.
Spencer, J.W. Jr. 1981. Stress relaxation at low frequencies in fluid saturated rocks: Attenuation and modulus dispersion. Journal of Geophysical Research, 86:1803–
1812.
Toksöz, M.N., Johnson, D.H., Timur, A. 1979. Attenuation of seismic waves in dry and saturated rocks: I. Laboratory experiments. Geophysics, 44(4):681–690.
Tutuncu, A.N., Dvorkin, J., Nur, A. 1997. Influence of cementation and permeability on wave velocities in poorly consolidated sands. Int. J. Rock Mech. & Min. Sci.
34:3–4, Paper No. 313.
Tutuncu, A.N., Podio, A.L., Gregory, A.R., Sharma, M.M. 1998. Nonlinear viscoelastic behavior of sedimentary rocks, Part I: Effect of frequency and strain amplitude.
Geophysics, 63(1):184–194.
Tutuncu, A.N., Podio, A.L., Sharma, M.M. 1998. Nonlinear viscoelastic behavior of sedimentary rocks, Part II: Hysteresis effects and influence of type of fluid on elastic
moduli. Geophysics, 63(1):195–203.
Winkler, K.W. 1983. Frequency dependent ultrasonic properties of high porosity sandstones. Journal of Geophysical Research, 88(B11):9493–9499.
Winkler, K.W. 1985. Dispersion analysis of velocity and attenuation in Berea sandstone. Journal of Geophysical Research, 90(B8):6793–6800.
Yale, D.P., Nieto, J.A., Austin, S.P. 1995. The effect of cementation on the static and dynamic mechanical properties of the Rotliegendes sandstone. Daemen & Schultz
(eds), Rock Mechanics, Rotterdam: Balkema.
Page 407
Macromechanical behavior for an assembly of grains modeled by bonding
spheres
Y.Nakata
Yamaguchi University, Yamaguchi, Japan
M.D.Bolton & Y.P.Cheng
Cambridge University, Cambridge, UK
5809 679 3
ABSTRACT: The use of Distinct Element Method (DEM) in modeling the behavior of crushable soil has aroused increasing attention since the
crushability of soil grains was included in the modeling procedure. The computer simulations were performed using the PFC3D computer code.
Numerical grains were made by bonding elementary spheres in “crystallographic” arrays, and by giving each sphere an existence probability of 0.8.
The bonding model which was adopted in the contact points between the spheres, serves to limit the total normal and shear forces that the contact can
carry by enforcing bondstrength limits. Triaxial tests on a cubical sample made of 389 grains were simulated. Isotropic compression and “drained”
triaxial compression under constant mean effective stress were calculated. The simulation provided not only the necessity in the soil modeling but also
the justification the use of the modeling in further geotechnical engineering applications.
1 INTRODUCTION
DEM simulation of perfectly elastic and infinitely strong grains provides many insights into the deformation of granular media (Thornton 2000). Qualitative agreement of
the mechanical behavior between the simulated results and real sand was found, yet the stress level dependency of granular behavior could not properly be represented
because the crushability of sand was ignored. The use of DEM in modeling the behavior of crusbable soil has aroused increasing attention since the crushability of soil
grains was included in the modeling procedure. Following Robertson (2000), numerical ‘grains’ (agglomerates) can be made by bonding elementary spheres in
probabilistically flawed ‘crystallographic’ arrays. McDowell and Harireche (2002) also validated the use of DEM in modeling soil particle fracture. Not only could
DEM simulate the crushing strength of a real sand grain, with diametral breakage of the bonded agglomerates between flat platens, it could also reproduce realistic
Weibull distributions of crushing strength in a batch of flawed agglomerates. Cheng et al. (2003) applied this DEM approach to simulate the compression and shearing
behavior of an element of crushable soil (silica sand) by reproducing the statistical crushing strength of a batch of uniformly sized ‘grains’ randomized by the removal of
20% of the microspheres. Reasonable agreement was found between the real data obtained from isotropically compressed silica sand and the DEM simulation when
stress was normalized by the characteristic crushing strength of the grains. Cheng et al. (2004) had the foregoing research with particular focus on the fundamentals of
yielding and plastic deformation.
The purpose of this paper is to introduce the simulation result by Cheng et al. (2003, 2004), and to justify the use of the modeled grains in further geotechnical
engineering applications.
2 MODELING OF GRAIN
2.1 Arraying and bonding of spheres
Grains were made from a regular assembly of spheres in hexagonal close packing (HCP), without initial overlap. As these agglomerate grains were intended to
represent solid particles, the main purpose of the regular packing was to minimize the space between the balls of the agglomerate. A stiffness model, a bonding model
and a slip model are included in the constitutive representation of contact points between the elementary spheres. The bonding model serves to limit the total normal and
shear forces that the contact can carry by enforcing bondstrength limits. The maximum tensile force that the bond can sustain in tension and the
Page 408
Figure 1. Examples of DEM agglomerate grains (Cheng et al. 2003).
maximum shear force it can withstand before breaking are specified when the bond is created. The bond breaks if either of these values is exceeded. Following
Robertson (2000), the parameter on the agglomerate grain, which models the properties of a typical sand grain, was given. The contact normal and shear stiffness and
the contact bond strength of each microsphere, 0.2 mm in diameter, were 4 MN/m and 4 N respectively. The frictional coefficient at the surface of the microsphere is
0.5, equivalent to In order to provide a statistical variability to strength and shape of the agglomerates, similar to that of a real sand, each elementary
sphere of an agglomerate was given a probability of existence of only 80% when it was created. As a result an average of 20% of the elementary spheres will not
appear in the final agglomerate used for subsequent testing, about 11 balls less than the maximum number of 57 as shown in Figure 1.
2.2 Mechanical behavior of modeled grains
Randomlyorientated agglomerates were then numerically crushed between two smooth and stiff platens under straincontrolled compression. This is similar to the
procedure described by Robertson & Bolton (2001). The initial separation of the platens was the same as the size of the agglomerate, which is 1.0 mm. With the
variability given to the agglomerates, different peak strengths were obtained from 20 tests. Figure 2 shows a typical result from the crushing tests; it is compared with
the crushing of a silica grain reported by Nakata et al (2001). The diameter of the tested silica sand is 1.4 to 1.7 mm, which is slightly larger than that of the computed
sand. The applied stress is calculated by normalizing the platen force by the initial diameter of the agglomerate or the sand grain. A similarity in the response is obtained,
in which both the computer simulation and the experimental single particle crushing test results produce a lower
Figure 2. Survival distribution curve on single agglomerate crushing (Cheng et al. 2003).
peak before the maximum peak strength. After the agglomerate is split at the maximum stress, the platens continue to approach one another with low contact stresses
until they find another good contact on the disintegrating agglomerate, to generate another split.
The survival probability of a batch of 20 agglomerates was calculated. The Weibull distribution can be used to describe the variability in tensile strengths of
apparently identical testpieces of a brittle material, in which the survival probability Ps is a function of normalized stress σ/σ0 given by
(1)

where σ0 is the stress at which 1/e or 37% of sample survive and m is the Weibull modulus. McDowell and Bolton (1998) explain that a similar approach can be
Page 409
used on the compressive normalized strength F/d2 in grain crushing tests treated as indirect tension tests, where F is the greatest force that was carried and d was the
initial separation of the platens. Figure 2 shows that a similar shape of the survival probability curve is obtained when the crushing peak stress is normalized by σ0,
although the silica sand has σ0=31 MPa and the simulated agglomerate has σ0=58 MPa. In addition, both of them give similar values of Weibull modulus of
approximately 3. Robertson (2000) showed that the Weibull modulus of a flawed agglomerate depends on the proportion of spheres was removed, or bonds
weakened.
3 ASSEMBLY OF MODELED GRAINS
3.1 Assembling for modeled grains
For each of the numerical tests reported here, an initial set of “exospheres” was first created at a size slightly smaller than the required agglomerates. They were placed
at random, but excluding overlaps. Then they were expanded to the required size, and cycled to equilibrium so as to reduce unwanted gaps. During this process shear
stiffness and friction were reduced to zero, while normal stiffness was increased 100fold. A linked list storing the coordinates of their centers was then created and the
exospheres deleted. Randomly rotated aggregates were then created in their place, centered at the coordinates in the list as shown in Figure 3, and the assembly
cycled to equilibrium again before commencing the tests. To reduce the likelihood of bonds between balls breaking during this stage, the strengths of the bonds were
initially set very high and reduced after a number of cycles. The final shear and normal stiffnesses of the balls of the agglomerates were set to their final values (4.0×106
Nm−1) and their coefficient of friction set to 0.5 (corresponding to a contact friction angle of 26.5°). Finally, bond strengths were fixed and variability was provided in
order to achieve the statistical distribution of breakage strength.
3.2 Mechanical behavior of assembly
Isotropic compression of a cubical arrangement of 389 agglomerates was begun by slow compression to 20 kPa, calculating the equilibrium at each 10 kPa increment
using a numerical iteration method recommended by the PFC3D manual and described in detail by Robertson (2000). Then, the sample was isotropically compressed
by 6 smooth stiff walls moving at a slow controlled rate. The voids ratio, calculated by using the solid volume as the total volume of the spheres, was then 2.08, shown
as the initial condition.
Figure 4 shows a comparison of isotropic compression curves between the silica sand and the DEM
Figure 3. Assembly of modeled grains (Cheng et al. 2003).
Figure 4. Normalized voids ratio plotted against mean effective stress normalized by grain tensile strength (Cheng et al. 2003).
simulation with a platen approach rate of 1 m/s. The effects of strain rate were investigated by Cheng, et al. (2003). There is a noticeable dynamic effect on the
compression behavior above 1 m/s. The silica sand has a more gentle transition into what we now recognize as clastic compression. One explanation is that real sand
particles have a greater variety of asperities, compared with the uniform microspheres in the DEM simulation. The opportunities for both elastic compression and
crushing at points of contact will be more variable. Similarly, the “normal compression line” in the DEM simulation begins to stiffen below e/ei of 0.6, in contrast to the
real sand. This is also tentatively attributed to the fact that agglomerates only have a limited number of component spheres.
Page 410
Figure 5. Validity of crushability in modeling.
3.3 Necessity of mo deled grains
Figure 5 shows the elog p curve for three particulate media having different properties of crushing calculating with 0.05 m/sec of walls moving. The thick black line
shows the compression result of the crushable assembly of agglomerates with 20% of the balls removed randomly from the regular array, as described earlier. The
agglomerates were supposed to have randomness of shape and flaws. The thin black line represents the elog p curve of the same assembly of agglomerates with the
same amount of elementary balls removed but the contact bonds that exist between the balls were nonbreakable in this case. The elog p curve of the breakable
agglomerates first deviated from that of the unbreakable ones at 6 MPa and started to curve at 17 MPa heading to the point known as the yield point. From a stress of
40 MPa onwards, the elog p plot looks almost linear with λ of 0.4. The amount of volumetric decrease for the thin black line is significantly small, and the micro
mechanism was restricted to the elasticity at contacts and rearrangement of the very rough agglomerates. Comparing the two lines shows that crushability is essential for
a realistic representation of plastic deformation with DEM.
The thick grey line of Figure 5 represents an assembly of agglomerates without removal of any elementary balls from the regular array, although the bonds are
crushable. Not only was the initial void ratio much less, there was also an apparent structural effect during the clastic yielding process. The sudden drop in volume at a
nearly constant mean stress of 60 MPa is due to the uniform crushing nature of these agglomerates without provision of randomness by ball removal, as was seen in the
compression of glass beads as reported by Nakata et al (2001). The variability in the crushing strength of agglomerates with no balls removed was low, with a Weibull
modulus of 9, compared with that of agglomerates with 20% of balls removed randomly, which had a Weibull modulus of 3. A similar tendency

Figure 6. Distribution for (volume of an agglomerate)1/3.
has been observed and discussed for undisturbed aged clay, sedimentary soft rock and cemented soils (e.g. Leroueil and Vaughan 1990). The structure of materials
having a sharp breakdown at yield is considered to be regular due to aging or adhesion processes.
The size distribution of the agglomerates was calculated and shown in Figure 6. The horizontal axis is the representative size, defined as the 1/3 power of the current
solidvolume of each individual agglomerate, while the vertical axis is the percentage of volume reduction. The curves correspond to the particle size distribution curves
in real sieving analysis. Initially, the agglomerates were nearly uniform in size. As mean stress increased, the agglomerates became welldistributed in size. It is similar to
the evolution of real particle size distribution. The smallest size, however, was restricted by the size of the single elementary sphere.
In order to further understand the significance of the introduction of crushability into the simulation, the constant mean stress triaxial compression shearing tests were
calculated for both breakable and unbreakable agglomerate assemblies after they were normallycompressed to various pressures, i.e. the thick and thin black lines
respectively in Figure 5. Figure 7 shows the results for the breakable agglomerate assembly compressed by rigid walls for: (a) the stressstrain curves and (b) the
volumetric and deviator strain curves.
The volumetric behavior with initial confining pressures of 1 and 5 MPa were dilative. As the initial confining pressure increased, the drained behavior changed from
dilative to contractive. In the cases with initial confining pressure of 20 and 40 MPa, the volumetric strain reached 15% contractively. The constant volumetric strain
condition at the final stage of the test appeared in the case of 1–10 MPa of mean stress. The results for the unbreakable agglomerate assembly are shown in Figure 8.
Shearing at all levels of initial confining pressure produced volume changes that were
Page 411
Figure 7. Behavior of breakable materials subjected to mean stress constant triaxial compression.
Figure 8. Behavior for unbreakable materials subjected to mean stress constant triaxial compression.
Figure 9. Critical state for DEM simulation (Cheng et al. 2003).
dilative although the degree of dilation became slightly weaker as the initial confining pressure increased.
3.4 Applications of modeled grains
3.4.1 Understanding of Critical State
The whole series of drained and undrained test paths is shown in Figure 9, together with indicative arrows if the final state of the simulation was continuing to change
when the computation was terminated. There is some ambiguity in the corresponding location of a critical state line on the e versus log p′ plot. In particular, the drained
test with was continuing to reduce in volume as bonds continued to break, even at 50% deviatoric strain. Nevertheless, there is some suggestion that
a critical state line exists about one unit of natural logarithms inside the fully developed normal compression line, just as Critical State Soil Mechanics suggests. The
simulations also suggest that the dissipation function used in Cam Clay would be improved if grain damage and grain rearrangement were recognized as distinct micro
mechanisms.
3.4.2 Understanding of plastic deformation and crushing
The yield point was determined in the conventional way as a change in the slope of a stressstrain curve for an overconsolidated DEM element. The element (with
ei=2.08) was first loaded to 40 MPa and then unloaded to 20 MPa, giving an overconsolidation ratio defined with respect to p′ of two (n=2) before stresspath tests
were simulated. Figure 10a shows the corresponding yield points of all stresspath directions in qp′ space from the results of the stresspath shearing tests. The amount
of bondbreakage could be quantified by the percentage of brokenbonds (pbb) counted from the beginning of shearing in each stress path.
Page 412
Figure 10. Plastic deformation and crushing (Cheng et al. 2004).
The numerical values of the pbb in a contour representation are shown in Figure 10b. The yield points obtained from the changeofslope method (Fig. 10a) can be
approximated by the fourpercent brokenbond contour (ppb=4%). It is clearly seen that bond breakage occurred before gross plastic yielding could be detected in
the stressstrain response. It is therefore clear that a significant proportion of bonds are broken as the stress state moves “inside” the observed yield surface. Significant
rearrangement (corresponding to macroscopic yielding) only occurs at about 4% pbb.
Plastic deformation was calculated from the measured total macroscopic strain of the element and the derived elastic strain using κ=0.04 but assuming the elastic
deviator strain to be negligible. The plastic strain increments were then plotted in Figure 10c. Both of them showed that the plastic flow followed a nonassociated flow
rule rather than the normality rule, similar to real soil.
3.4.3 Further applications
The following themes were considered as further applications for the modeled grains.
1. Compaction
2. Localization of deformation, e.g. slope stability with sliding surface
3. Cyclic degradation, e.g. liquefaction
4. Bearing capacity of pile and foundation, e.g. cone penetration
All simulations of processes that could crush grains should be undertaken with breakable agglomerates as described here. A simulation with crushable grains will give a
better physical understanding for macroscopic modeling and the selection and determination of meaningful material parameters.
4 CONCLUSION
The usefulness of Distinct Element Method in modeling the behavior of crushable soil was presented. Numerical grains were made by bonding elementary spheres in
“crystallographic” arrays, and by giving each sphere an existence probability of 0.8. The bonding model which was adopted in the contact points between the spheres
to limit the total normal and shear forces. Elementary tests, which were conventionally conducted in geotechnical engineering, on a cubical sample made of 389 grains
were simulated. The behaviors of isotropic compression and “drained” triaxial compression under constant mean effective stress showed a reasonable tendency
qualitatively. It indicated the use of the modeled grains to be justified in further geotechnical engineering applications.
REFERENCES
Cundall, P.A., and Strack, O.D.L. (1979). A discrete numerical model for granular assemblies’, Géotechnique 29(1):47–65.
Page 413
Cheng, Y.P., Nakata, Y. & Bolton, M.D. (2003). Distinct element simulation of crushable soil. Geotechnique 53(7): 633–641.
Cheng, Y.P., Bolton, M.D. & Nakata, Y. (2004). Crushing and plastic deformation of soils simulated using dem. Geotechnique 54(1).
McDowell G.R. & Harireche O. (2002). Discrete element modelling of soil particle fracture. Technical note. Geotechnique 52(2):131–135.
Nakata, Y., Kato, Y., Hyodo M., Hyde, A.L. & Murata, H. (2001). Onedimensional compression behaviour of uniformly graded sand related to single particle crushing
strength. Soils Found 41(2):39–51.
Robertson, D. (2000). Computer simulations of crushable aggregates. Ph.D. dissertation, Cambridge University.
Robertson, D. & Bolton, M.D. (2001). DEM simulations of crushable grains and soils. Powders and Grains 2001, Kishino (ed.), 623–626.
Thornton, C. (2000). Numerical simulations of deviatoric shear deformation of granular media. Géotechnique 50(1): 43–53.
Page 414
Page 415
Elastoviscoplastic constitutive model of sea ice with strain rate dependent ice
strength
D.Zyryanov
Institute of Water Problems, Russian Academy of Science, Moscow, Russia
R.Korsnes
Norwegian Defence Research Establishment, Division of electronics, Kjeller, Norway
5809 679 3
ABSTRACT: Recent remote sensing observations of the Arctic region by satellite imagery demonstrated sea ice dynamic challenges that cannot be
explained without deep consideration of underlying rheology of the ice media. Nonlinear stressstrain relationships coupled with strainrate dependent
ice strength are the factors defining Arctic ice mass behavior mostly observed in nature. This paper is directed to the sea ice rheology study and is
based on the authors proposed sea ice model linked with the PFC2D code. The simulation results revealed internal local interparticle forces observed
in lab experiments with rectangular ice specimens. Otherwise, mesoscale code running in narrow areas covered by ice demonstrated irregular areas,
where stress is significantly less than in coastal zones. The proposed code might be applied to the investigation of industrial structures interacting with
ice as well as for the study of pack ice dynamics in the Arctic.
1 ELASTOVISCOPLASTIC SEA ICE CONSTITUTIVE MODEL AND STRAIN RATE DEPENDENT ICE
STRENGTH
1.1 Introduction
Application of Discrete Element Models (DEM) to the challenges of sea ice dynamics (Hopkins et al. 1991 a, b, Hopkins 1993) revealed mechanisms of interparticle
interaction and underlying sea ice physics. The pack ice in the Arctic is forced by wind stress (Banke et al. 1973, Omstedt et al. 1996) and water current. Remote
sensing synthetic aperture radar (SAR) satellite images demonstrate a complex structure of polynyas and leads in the Arctic. This phenomenon occurs being joined with
sea rheology (Hopkins 1996, Loset 1994, Polyakov 2001). This paper is a further continuation of DEM approach based on PFC2D code (Itasca 2002). We first
apply the complex sea ice rheology coupled with strainrate dependent ice strength to numerical simulations of ice sheets destruction in mesoscale areas.
1.2 Basic equations
The total strain of columnargrained sea ice in tension can be superimposed into the following three terms according to Sinha (1988):
e=ee+ed +ev
(1)
where ee represents elastic strain, ed is delayed elastic recovery (exponentially dependent on time) and ev is a viscous strain. Sinha has proposed empiric formulas for
each of these terms in the form:
(2)
The relationship 2 is a polynomial relatively to stress σ. To explicitly describe numerical interparticle interaction, stress has to be a function of strain and elapsed
Table 1. The parameter values used in the nonlinear stressstrain relationship (Eq. 2).
Parameter* Description Value

E Stiffness in tension
5×109 Nm−2
σ0 Stress unit 1.0 Nm−2

c Material constant
9×10−3
b Time exp. for delayed elastic strain 0.34
a T Material constant 2.5×10−4 s−1

d An average grain diameter 1.0 mm
d 0 Grain diameter unit 1.0 mm
ev0 Viscous strain rate 1.76×10−7 s−1

*Empiric values.
Page 416
Figure 1. Normal (fn) and shear (fs) contact forces vs. total strain in according with (5) and (6). At the moment of the contact
appearance (t=0) stress in tension is a linear function of strain (the dash line). The dotdashed line denotes shear stress that is
throughout a linear function of e.
time from the moment a contact first appears in the system. Joining Eqs 1 & 2 and rearranging:
(3)
The discriminate of cubic Eq. 3 is:
(4)
which is positive for any t>0. It means that cubic Eq. 3 has one real root and two complex conjugate ones. Applying Cardano formula to Eq. 3 gives:
(5)
Compression as well as shear stress is considered to be pure elastic in the model:
(6)
where and k s denote normal stiffness in compression and shear stiffness respectively. Un /Us are particles overlap in normal/lateral direction. The relationships 5 &

6 are the core of the proposed model. Figure 1 shows normal force f n versus total strain e accumulated by the contact in a process of simulation.
If e>0 (compression), the normal stress is proportional to accumulated deformation e. Otherwise, if normal tension occurs in the contact (e<0) then stress becomes
nonlinear function of e (Eq. 5).
Figure 2. Normal (fn) and shear (fs) contact strength vs. strain rate according to Eq. 7.
Table 2. The parameter values used in the nonlinear strainrate sensitive ice strength in tension (Eq. 7).
Parameter Description Value

Unit strength in tension
1.0 Nm−2
(fs)max Shear strength 1.0×108 Nm−2

P* Proportional factor 212.0
p* Strainrate exp. 0.345
e
0
Unit strainrate 1.0 s−1

* Empiric values.
1.3 Strain rate dependent ice strength in tension
Experiments with ice specimens in tension (Cole et al. 1998, Sinha 1988) have demonstrated highly sensitive strain rate ice strength. Maximum stress in tension (i.e.
strength) occurs proportional to strain rate in degree of 0.345 (Sinha 1988):
(7)
Figure 2 shows a graph of versus strain rate for different P. The parameter values used in Eq. 7 are given in Table 2.

2 RUNNING PFC2D MODEL
2.1 Incrementally calculated bond forces
To avoid pure computation abnormalities we firstly applied a procedure of incrementally calculated bond forces between particles in the system. This approach is based
on expression of viscous normal velocity via normal forces before (.0) and after (./) one time step as:
(8)
Page 417
Figure 3a. The PFC2D simulation results of a ‘spring’ model with pure linear elasticity subjected to uniaxial gravity stress. The top edge
of a heavy specimen is fixed whereas the bottom (black) is free. Vertical lines show contact forces with thickness proportional
to their magnitudes.
Figure 3b. The comparison with the simulation results based on Eqs 6 & 8. Grey lines denote tension stress and black ones—
compression, where line thickness is proportional to the stress magnitude. Time scale is 12×104 s [corresponds to 1.5×103
computational steps].
Figure 4. Results of PFC2D model applied in Storfjorden area situated in the southeast of the Svalbard Archipelago between the islands
of Spitsbergen, Barentsoya and Edgeoya. Area of the interest is outlined by the white square. The line thickness is proportional
to stress magnitude occurred in balltoball contact.
Figure 3a shows the result of simple elastic material subjected to uniaxial gravity stress acted in vertical direction. Top edge of the heavy specimen is fixed whereas its
bottom is free.
Alternatively, Figure 3b demonstrates dynamic run of PFC2D code based on Eqs 6 & 8. In contrast with the linear model shown in the Figure 3a, in nonelastic
approach maximum becomes a function of elapsed time and model parameters (Table 1). Lines of local tension maximums (grey lines in the Fig. 3b) demonstrate a
common coincidence with cracks network observed in labs with columnargrained sea ice samples (Cole et al. 1998, Sinha 1988).
2.2 An explicit interparticle forces calculation. Application in mesoscale research areas
To reveal the role of elasticviscoplastic ice rheology in mesoscale applications we modeled ice sheet destruction of the Storfjorden gulf (Svalbard Archipelago)
caused by wind stress. Satellite remote sensing imaginary (SAR) demonstrated the persistence of latent heat polynyas in this region. Figure 4 shows the results of this
study. Balls in contact with a shore are fixed. The wind stress (drag force) is fixed and directed from the north. There is no any other external forcing in the system
caused by e.g. water
Page 418
currents. In contrast with elastic approach proposed earlier (Zyryanov et al. 2003a), elasticviscoplastic model (Eqs 5, 6 & 7) may lead to qualitatively different
simulation results.
In beginning of the fjord (highlighted by white rectangular) is observed area with dominant tension stress in numerical experiments. The investigation of the polynya
opening caused by winds of different directions (Zyryanov et al. 2003b) showed sensitivity of the icefree edge to wind forcing. Northern wind opens a larger polynya
(ca. 30%) then does a northeastern wind with the same speed.
The simulation results are in good agreement with the Storfjorden polynya outlines observations (Haarpaintner 1999). The more applied drag force the
Figure 5a. The result of Storfjorden ice sheet break up caused by tension stress directed from the North in nonlinear approach. Elapsed
time is 1.3×104 s.

Figure 5b. Continuation—Elapsed time is 9.2×105 s.

higher stress concentration difference in coastal zones of Spitsbergen and Edgeoya islands was observed in numerical experiments with PFC2D code.
3 CONCLUSIONS
The proposed nonlinear stressstrain approach coupled with strain rate dependent ice strength revealed a common coincidence with smallscale lab experiments with
the sea ice specimens. Thus, areas of local stress concentrations are in agreement with those found in nature.
Mesoscale modeling of Storfjorden polynya by means of the proposed ice rheology procedure showed a lensshaped local spot of low stress concentration, which
cannot be revealed e.g. by simple elastic models.
ACKNOWLEDGEMENTS
We wish to express our gratitude to Vladimir Smirnov from the Arctic and Antarctic Research Institute (AARI, department of sea ice physics) for his important
comments and directions. We also thank our colleagues from Water Problems Institute (WPI, RAN) for their technical assistance in preparing this manuscript.
APPENDIX

Table 3. Parameter values used in elasticviscoplastic simulation by PFC2D code shown in figures.
Setting Parameter Value

Figure 4 E
5×109 Nm−2
σ0 1 Nm−2

c
9×10−3
b 0.34
Time step 0.1 s
Time
4×102 s
Density
1×103 kg/m3
Drag force direction
00 (North)
Figure 5a E
5×109 Nm−2
σ0 1 Nm−2

c
9×10−3
b 0.34
Density
1×103 kg/m3
Time step 1.0 s
Time
1.3×104 s
00 (North)
Figure 5b E
5×109 Nm−2
σ0 1 Nm−2

c
9×10−3
b 0.34
Density
1×103 kg/m3
Time step 1.0 s
Time
9.2×105 s
00 (North)
Page 419
REFERENCES
Banke, E.G. & Smith, S.D. 1973. Wind stress on Arctic sea ice. J. Geophys. Res. 78(33):7871–7883.
Cole, D.M., Johnson, R.A. & Durell, G.D. 1998. Cyclic loading and creep response of aligned firstyear sea ice. J. Geophys. Res. 103(C10):21751–21758.
Haarpaintner, J. 1999. The Storfjorden polynya: ERS2 SAR observations and overview. Polar Res. 18, 175–182.
Hopkins, M.A. 1993. A mesoscale simulation of the Arctic ice pack. In J.P.Dempsey et al. (eds): Ice mechanics 1993. ASME AMD 163:85–96. New York: American
Society of Mechanical Engineers.
Hopkins, M.A. 1996. On the mesoscale interaction of lead ice and floes. J. Geophys. Res. 101(C8):18315–18326.
Hopkins, M.A. & Hibler, W.D. 1991a. Numerical simulations of a convergent system of ice floes. Ann. Glaciol. 15:26–30.
Hopkins, M.A., Hibler, W.D. III & Flato, G.M. 1991b. On the numerical simulation of the sea ice ridging process. J. Geophys. Res. 96(C3):4809–4820.
Løset, S. 1994. Discrete element modelling of a broken ice field. Part I: model development. Cold Reg. Sci. Technol. 22:339–347.
Omstedt, A., Nyberg, L. & Lepparanta, M. 1996. On the iceocean response to wind forcing. Dyn. Meteorol. Oceanogr. 48:593–606.
Polyakov, I. 2001. An eddy parameterization based on maximum entropy production with application to modeling of the Arctic ocean circulation. J. of Phys. Ocean. 31(8):
2255–2270.
Sinha, N.K. 1988. Crackenhanced creep in polycrystalline material: strainrate sensitive strength and deformation of ice. J. Mater. Sci. 23:4415–4428.
Zyryanov, D., Haarpintner, J. & Korsnes, R. 2003a. Storfjorden (Svalbard): modelling of the polynya development and sea ice ridging process. Modelling, Identification
and Control J. 24(1):37–48.
Zyryanov, D., Haarpintner, J. & Korsnes, R. 2003b. Numerical modelling of the Storfjorden (Svalbard) polynya development due to wind stress: role of the sea ice rheology
and damping forces. Polar Research J. 22(2):235–245.
Page 420
Page 421
3D imagebased discrete element modeling for irregularlyshaped grains
T.Matsushima
Institute of Engineering Mechanics and Systems, University of Tsukuba, Tsukuba, Japan
5809 679 3
ABSTRACT: Grains used in various engineering fields often have very irregular shape, and this grain shape greatly affects the mechanical behavior of
their assembly. In order to simulate shear behavior of such granular materials quantitatively an imagebased Discrete Element modeling has been
proposed (Matsushima & Saomoto 2002). This paper focuses on the 3D version of the modeling and its effectivity check. The material modeled is
Toyoura sand (a standard sand in Japan) whose 3D grain shape data were obtained with a micro Xray CT at SPring8 (Matsushima et al. 2004).
They are directly converted into Discrete Element models, and the accuracy of the model is discussed with two error indices. Then various “virtual”
specimens with different void ratios are prepared and a series of simple shear tests are performed under (1) constant confining pressure and (2)
constant volume. The results are basically in good agreement with experimental data, which proves the effectivity of the present grain shape modeling.
1 INTRODUCTION
A rapid increase of computer abilities in the recent past has drastically extended the availability of direct simulation of particulate system in various engineering fields.
Discrete Element Method (DEM) is the most popular method in geotechnical and powder engineering fields, and it has been applied into various boundaryvalue
problems including blasting and collision (Cook & Jensen 2003) that are difficult to be treated in continuum mechanics.
Recently such particle simulations are also used to explore constitutive relation of granular assembly. Since many particulate systems in engineering contain
tremendous number of particles, it is still quite important to simulate them as continua. In such cases it is more convincing to construct their constitutive relations based
on the behavior of constituent particles. Particle simulation is a great help for this purpose.
On the other hand, we still have encountered the difficulty to simulate with DEM the behavior of real engineering granular materials quantitatively. This is mainly due
to an excess simplification of grain modeling. In order to predict their behavior with DEM more accurately, it is necessary to develop more accurate model on
interparticle contacts, grain shape, grain crushing and so on. In particular, grain shape is of primary importance in plastic behavior and peak shear strength of the
granular materials. Natural sands or ‘topdown’ powders often have very complicated grain shapes that necessarily affect the macroscopic behavior of such assembly.
For instance, an assembly of angular grains has much higher shear strengths than that of roundish grains, but its mechanism is still not clear and its quantitative estimation
has not been successfully made yet.
There have been various DEM researches that deal with noncircular elements in 2D (Cundall 1971, Meguro et al. 1991, Rothenberg & Bathurst 1992,
Mirghasemi et al. 1997, Matsushima & Konagai 2001) or nonspherical elements in 3D (Ghaboussi & Barbosa 1990, Williams & O’Connor 1995, Lin & Ng 1997,
Munjiza & Latham 2002). Recently Matsushima & Saomoto (2002) proposed an efficient imagebased modeling method of irregularlyshaped grains and showed its
high adaptability into 2D and 3D DEM (Matsushima et al. 2003, Matsushima 2004). The present study aims to validate the modeling method in details through
comparison with physical element tests of Toyoura sand, a standard sand in Japan.
2 IMAGEBASED GRAIN SHAPE MODELING
2.1 Detection of 3D grain shapes of Toyoura sand
In order to obtain 3D grain shape data of Toyoura sand, Matsushima et al. (2003b) conducted a micro Xray CT experiment at SPring8, a third generation
Page 422
synchrotron radiation facility. Since the original CT images of the samples contain a lot of grains having contacts one another, an image processing algorithm has been
developed to distinguish every single grain from the others. The size distribution of the finally obtained grains is shown in Figure 1. It ranges from 0.1 to 0.3 mm, while
actual grain size of Toyoura sand ranges from 0.1 to 0.2 mm. This is partly because of the inaccuracy of the image processing, and partly because the diameter of an
irregular grain in the figure is calculated as that of the spheres of equivalent volume. Since the longest axis of the grain does not affect the sieving result, the size of
equivalent spheres tends to bigger than the one measured by sieving. Figure 2 shows some
Figure 1. Size distributions of the original grains and modeled grains.
Figure 2. Examples of the original 3D grain shapes obtained from CT data.
examples of 3D Toyoura sand grains obtained from the CT data.
2.2 Grain shape modeling
In the proposed modeling method, an irregular grain shape is represented with a certain number of primitive elements (circular elements for 2D modeling and spherical
elements for 3D modeling) mutually connected in a rigid way. Since such primitive elements only require the simplest contact detection algorithm, Discrete Element
simulation with such grains keeps high computational performance.
The key feature of the proposed method is that the sizes and the locations of the primitive elements are automatically computed so that the best model accuracy
possible is attained. The algorithm is based on a virtual timemarching scheme where a kind of virtual attraction is assumed between each surface point of a target
irregularlyshaped grain and the element closest to the point (Fig. 3).
As a result of this attraction, elements tend to move to reduce the distance from the surface point, and an optimized converged solution is obtained after a sufficient
number of calculation steps. The detailed explanation is available in Matsushima and Saomoto (2002). Figure 4 shows the resulting 3D models corresponding to the
grains in Figure 2, each of which is composed of 10 spheres.
In order to evaluate the accuracy of the model, the following surface error index is introduced:
(1)
where N is the number of surface points of the target grain, Req is the radius of the circle whose volume is equivalent to the one of the target grain, is the radius of
the element. In other words ES indicates the average distance between the original grain surface and that of modeled grain normalized by the average grain radius.
Figure 5 shows the distribution of ES for 400 grains used in Section 4. According to the figure the average of ES is about 0.058, which means that 5.8% of grain radius
is the average error of the model. Table 1 shows the ES ’s for the grains in Figure 4.
Figure 3. Concept of virtual force acting on the elements.
Page 423
Figure 4. Examples of the modeled grains each of which is composed of 10 spheres.

Figure 5. Distribution of surface error index ES.
Table 1. Error indices of the grain samples.
Grain No. ES E V
1 0.0483 −0.10905
2 0.0765 −0.13039
3 0.0582 −0.12712
4 0.0546 −0.15544
5 0.0484 −0.12114
6 0.0680 0.19149
7 0.0549 −0.14345
8 0.0605 −0.15263
9 0.0660 −0.18854

Figure 6. Distribution of volume error index EV.
Another index for the model evaluation is the volume error index, which is defined as:
(2)
where Vmodel and Vdata are the volumes of modeled grain and original CT data, respectively. Figure 6 shows the distribution of EV for the 400 grains, which indicates
that most of modeled grains have about 10 to 15% less volume than the grains from the CT data. This is because the modeled grains describe the original irregular grain
shape only by the combination of convex spherical shapes. There are also some grains whose volumes are much bigger than the original data. This is caused by the
inaccuracy of the modeling. However these errors do not greatly affect the resulting grain size distribution of the modeled grains as shown in Figure 1. EV for the grain in
Figure 4 is also listed in Table 1.
3 DEM USED IN THIS STUDY
3D DEM program used in this study was developed by the author, which allows combining a certain number of spherical elements in a rigid way to describe an
irregularlyshaped grain. Therefore, in the program, the contact judgment and the calculation of contact forces are conducted for each element, while the equation of
motion is solved for each grain. Since such grains are no more spheres, it is necessary to treat the moment of inertia as a tensor. Therefore, in the program, the three
principal values of the tensor and their orientations for each grain are computed in the beginning, and their rotations are calculated according to Euler’s equation of
rotational motion. It should be noted that this equation is nonlinear and has to be solved in an iterative way.
Page 424
4 SIMPLE SHEAR SIMULATION
Using 400 Toyoura sand grains model described in Section 2, a series of 3D DEM simple shear simulations were performed under (1) constant confining pressure and
(2) constant volume condition, that correspond drained and undrained torsional simple shear tests, respectively.
4.1 Preparation of specimens
DEM parameters used in this study are listed in Table 2. Stiffness of grains determined by the spring constant
Table 2. DEM parameters used in this study.
Density of grain
2.64 (g/cm2)
Spring constant (normal)
1.0e8 (g/s2)
Spring constant (shear)
2.5e7 (g/s2)
Damping coefficient (normal) 2.0e–2 (g/s)
Damping coefficient (shear) 1.0e–2 (g/s)
20, 27 (deg)
Interparticle friction angle
Time increment 1.0e6(s)
Figure 7. Snapshots of a specimen (Toyoura sand model, dense).
at the contact point is relatively small in comparison with that of real grains. Damping coefficient is sufficiently small so that the energy loss due to this damping can be
negligible. Interparticle friction angle of irregularlyshaped grain surface is difficult to determine from physical experiment. Therefore the simulations are performed with
two different values, 20 degrees and 27 degrees. The abovementioned 400 grains are consolidated under 100 kPa vertical stress. The resulting specimen is about 1.0
mm high, 1.5 mm wide and 1.5 mm deep (Fig. 7a). Smooth flat plates are set in front and back of the specimen, while periodic boundary is assumed on its left and right
sides.
Specimens with various void ratios are prepared by changing the interparticle friction angle during the consolidation process. The loosest specimen was prepared
with the interparticle friction angle of 27 degrees, while the densest specimen was prepared without the interparticle friction. After stabilization, the grains around top
and bottom edges are selected and connected, respectively, to make rigid frictional plates sandwiching the specimen.
4.2 Simple shear under constant confining pressure
Keeping the confining pressure acting on the top frictional plate constant (100 kPa), simple shear deformation is imposed to the specimen by moving the bottom
frictional plate laterally (Fig. 7b). Since the left and right sides of the specimen are periodic, the resulting shearbanding plane is almost horizontal. Figures 8 & 9
summarize the evolutions of stress ratio (horizontal shear stress divided by the vertical confining stress) and volumetric strain with respect to the shear strain for the
specimens with various initial void ratios, respectively. Interparticle friction angle is 20 degrees in this case. It is clear that the denser specimen exhibits the bigger peak
stress ratio. However, the residual stress ratio is almost identical for all specimens. It is noted that noticeable oscillation in
Figure 8. Evolution of stress ratio.
Page 425
stress ratio at residual state is caused by the insufficient number of grains in the specimens.
The figures include some experimental results on hollowcylindrical torsional simple shear (TSS) test of Toyoura sand (Pradhan et al. 1988). Simulation results are in
good agreement with the experimental results in a quantitative way. It is worth mentioning that the results shown in the figures are based on the macroscopic (overall)
strain measure. Therefore the responses after the yield state are strongly affected by the size of the specimen because of the strain localization. Since the height of the
DEM specimens are much smaller than those in experiments, the strain softening rate in the simulation becomes much milder. Also, the considerably large volumetric
strain in the simulations is due to the specimen size.
Figure 10 shows the internal friction angle with respect to the initial void ratio of the specimens. The figure includes the TSS test results by Pradhan et al. (1988).
Additionally, some DEM results of spheres are put together for reference. Note that the internal friction angles in the figures are computed by
The figure clearly indicates that
Figure 9. Evolution of volumetric strain.
Figure 10. Internal friction angle vs. initial void ratio.
the result in Toyoura sand DEM with the interparticle friction angle of 20 degrees is quite close to the experimental result. The difference in looser specimens is due to
the large oscillation of stress state at the residual state. According to Figure 8, the average residual stress rate is about 0.6 in all cases, which corresponds to of
31.0 degrees.
4.3 Simple shear under constant volume
Using the same specimens as the constantconfiningpressure tests, the top frictional plate is fixed and the shear deformation is applied from the bottom plate under
constant volume condition. This type of simulation is comparable to undrained monotonic loading test if the compressibility of grains and pore water is negligible. Figure
11 shows the stress paths for DEM simulation The experimental result (undrained TSS test) to be compared with is shown
Figure 11. Stress paths in DEM simulations (Toyoura sand model).
Figure 12. Stress paths in experiments (Toyoura sand, Yoshimine 1996).
Page 426
Figure 13. Evolution of mobilized friction angle (DEM Toyoura sand model).
Figure 14. Evolution of mobilized friction angle (Toyoura sand model, Yoshimine 1996).
in Figure 12 (data from Yoshimine 1996). Quantitative agreement is not sufficient in this case, mainly because this small strain behavior is quite sensitive to the initial
condition as well as the contact model related to the elastic behavior. On the other hand, the friction angle at steady state, which corresponds to the inclination of the
failure line, is about 31 degrees in DEM simulations (Fig. 13), which is almost identical to the experimental results (Fig. 14).
5 CONCLUSION
Toyoura sand model for Discrete Element Method was presented and its performance was investigated through a series of simple shear tests. Simple shear behavior
under constant confining pressure was in quantitative agreement with experiment when the interparticle friction angle was set to 20 degrees.
Constantvolume simple shear simulations result in rather different stress paths from those in experiment, though the internal friction angle at the steady state was
almost identical.
ACKNOWLEDGMENTS
The author gratefully acknowledges Dr. Yoshimine of the Tokyo Metropolitan University for providing his undrained TSS test data.
The experiment at the SPring8 was carried out with the approval of the Japan Synchrotron Radiation Research Institute (Proposal Number 2003A0127NDL2
np). This study was supported in part by a grantinaid for research of young scientists (No. 14750409) of the Ministry of Education, Culture, Sports, Science and
Technology (MEXT).
REFERENCES
Cook, B.K & Jensen R.P (eds). 2003. Discrete Element Methods, Proceedings of the third international conference on DEM, ASCE.
Cundall, P.A. 1971. A computer model for simulating progressive, largescale movements in blocky rock systems. In Symp. ISRM, Nancy, France. Proc., 2:129–136.
Ghaboussi, J. & Barbosa, R. 1990. Threedimensional discrete element method for granular materials, International Journal for Numerical and Analytical Methods in
Geomechanics., 14:451–472.
Lin, X. & Ng, T.T. 1997. A three dimensional discrete element model using arrays of ellipsoids, Geotechnique, 47(2):319–329.
Matsushima, T. & Konagai, K. 2001. Grainshape effect on Peak Strength of Granular Materials, Computer Methods and Advances in Geomechanics, Desai et al.
(eds). Balkema, 1, 361–366.
Matsushima, T. & Saomoto, H. 2002. Discrete element modeling for irregularlyshaped sand grains, Proc. NUMGE2002: Numerical Methods in Geotechnical
Engineering, Mestat (ed.), pp. 239–246.
Matsushima, T., Saomoto, H., Matsumoto, M., Toda, K. & Yamada, Y. 2003. Discrete element simulation of an assembly of irregularlyshaped grains: Quantitative
comparison with experiments, Electronic Proceedings of 16th Engineering Mechanics Conference, ASCE. 8p. (CDROM)
Matsushima, T., Saomoto, H., Uesugi, K., Tsuchiyama, A. & Nakano, T. 2004. Detection of 3D irregular grain shape of Toyoura sand at SPring8, Xray CT for
geomaterials. In Proc. International workshop on Xray CT for geomaterials. Otani & Obara (eds.), Balkema, pp. 121–126.
Matsushima, T. 2004. Virtual element test of sand, Computational Mechanics, WCCM VI in conjunction with APCOM’04, in press.
Meguro, K., Iwashita, K. & Hakuno, M. 1991. Fracture analysis of media composed of irregularlyshaped regions by the extended distinct element method, Structural
Eng./ Earthquake Eng., JSCE, 8(3):131s–142s.
Page 427
Mirghasemi, A.A., Rothenburg L. & Matyas, E.L. 1997. Numerical simulations of assemblies of twodimensional polygonshaped particles and effects of confining
pressure on shear strength, Soils and Foundations 37(3):43–52.
Munjiza, A. & Latham, J.P. 2003. Computational challenge of large scale discontinua analysis, In Proceedings of the third international conference on DEM, Cook,
B.K & Jensen R.P (eds.), ASCE, 5–10.
Pradhan, T.B.S., Tatsuoka, F. & Horii, N. 1988. Strength and deformation characteristics of sand in torsional simple shear, Soils and Foundations, 28(3):131–148.
Rothenberg & Bathurst 1992. Micromechanical features of granular assemblies with planner elliptical particles, Geotechnique, 42(1):79–95.
Williams, J.R. & O’Connor, R.M. 1995. A linear complexity intersection algorithm for discrete element simulation of arbitrary geometries. Engineering Computations, 12:
185–201.
Yoshimine, M. 1996. Undrained flow deformation of saturated sand under monotonic loading conditions, Doctoral dissertation, University of Tokyo.
Page 428
Page 429
Comparison between numerical and laboratory tests of rapidly sheared granular
materials
K.Iwashita, K.Ichiba & M.Oda
Department of Civil and Environmental Engineering, Saitama University, Saitama, Japan
5809 679 3
ABSTRACT: To study the micromechanics of rapid flow of granular material, apparatus was newly designed to generate twodimensional flow in a
laboratory. The characteristics of the apparatus are summarized as follows; it can deal with a wide range of shear strain rate and solid fraction, the
tangential stress can be measured (but normal stress cannot), and movement of individual particles can be traced using a highspeed camera with a
software for image analysis. At same time, numerical simulation based on Distinct Element Method (DEM) was also performed to get the detailed data
those are difficult to measure in the laboratory test. By comparison with previous results and experimental results, accuracy of the numerical simulation
are investigated.
1 INTRODUCTION
Some experimental works on rapid flow of granular materials have been performed with motivation to examine some results of kinetic theory of granular flow (e.g.
Bagnold 1954, Savage & Sayrd 1984). To do this, however, we must carefully take into account the effect of interstitial viscous fluid and gravity since most of the
theoretical works were discussed based on the assumption that dry, uniform sized spheres flow under the gravityfree condition. Earth’s gravity, for example, produces
a static pressure at the bottom of a 1 cm soil column equal to the dynamic normal stress produced by a shear strain rate at 5–10 (1/sec). In order to minimize the effect
of gravity, one can use the interstitial fluid of the same density as granular particles (Bagnold 1954). In this case, however, the influence of viscosity must be taken into
consideration. Possible ranges of solid fraction and shear strain rate are commonly so limited in the previous experiments that it is not easy to figure out the micro
mechanical behavior of entire range of granular flow. Moreover, it is difficult to observe directly micromechanical quantities such as individual particle velocity and local
solid fraction in a flow domain. In order to clarify the characteristic of granular flow from a micromechanical point of view, apparatus was newly designed to generate
twodimensional flow (Ichiba 2003, Iwashita et al. 2004). At same time, numerical simulation based on Distinct Element Method (DEM) was also performed to get the
detailed data those are difficult to measure in the laboratory test. By comparison with previous results and experimental results, accuracy of the numerical simulation are
investigated.
2 EXPERIMENTAL APPARATUS AND MEASUREMENT
2.1 Experimental apparatus
The photograph and the schematic view of the apparatus are shown in Figures 1 & 2, respectively.
The experimental apparatus consisting of an inner rotary cylinder and an outer fixed cylinder, which are placed on a bottom horizontal plate made of stainless. The
inner surfaces of the cylinders were roughened by
Figure 1. Experimental apparatus. A=shear stress measurement equipment, B=rotating velocity meter.
Page 430
Figure 2. Schematic view of the experimental apparatus.
Figure 3. Particles in the shear cell.
Table 1. Parameters of steel ball.
Radius r 0.50 cm
Mass 4.08 g
Density ρp 7.83 g/cm3

Volume
0.523 cm3
Rolling friction coefficient 0.0
Sliding friction coefficient 0.25
Repulsive coefficient 0.9–0.7
Number of particles 1–1124
means of gluing half circular rods with the same radius as the spherical steel balls (radius=5 mm), which were sheared between the inner and outer cylinders by rotating
the inner cylinder. The shear strain rate, U/H, was freely controlled such that the rotation speed U of the inner cylinder was changed (H=clearance between the inner
and outer cylinders). The solid fraction ν was controlled by means of changing the number of particles (balls). Since particles move on the
Figure 4. Distribution of particle velocity vector (v*=0.74, U/H=11.05 (1/sec)).
horizontal table twodimensionally, gravity does not take any effect on the movement of granular particles. Moreover, covering the upper surface by an acrylic
transparent plate, we could trace the movement of particles directly. (Spherical balls first slip immediately after each collision, and then start rolling when they travel,
say, about ten times particle diameter. Following the trajectory by a high speedcamera, we found that particle velocity after collision remains almost constant until next
collision takes place (Ichiba 2003)).
The concentric ring space surrounded by two cylinder walls is the shearing cell. In this study, radius of the outer cylinder was set to 40 cm, and radius of inner
cylinder was set to 35 cm. Therefore, the ring width was set to 5 cm, which was equivalent to 5 particles size. Figure 3 shows the particles in the shear cell. Upper
cylinder is the inner wall and lower cylinder is the outer wall. Those walls are covered with half circular rods with the same radius as the steel ball. The parameters of
the steel ball are shown in Table 1.
2.2 Measurement method
Figure 4 shows the distribution of the particle velocity vector measured by using highspeed camera. From
Page 431
Figure 5. Normalized particle velocity distribution in the shear cell (v*=0.74, U/H=11.05 (1/sec)).
Figure 6. Measured sheared stress waves.
this figure, it can be said that there is no strain localization and all the particles flow due to the inner cylinder rotation. Figure 5 shows the normalized particle velocity
distribution in the shear cell calculated from Figure 5. Because the tangential (shear directional) velocity is almost in proportional to the distance from the outer wall, the
shear strain rate, which is defined as a tangent of the graph, is almost constant in whole cell. According to this fact, the shear strain rate U/H is regarded as the
representative shear strain rate of the particle shear flow in the following part.
The outer cylindrical wall was subjected to rotational force arising from collision of particles against it, which was measured by means of a load cell attached at A in
Figure 1.
The shear stress τ was calculated by dividing the rotational force by the corresponding area of outer cylindrical wall. The typical shear stress waves are shown in
Figure 6. The shear stress, of course, oscillated during a test with a definite mean. The representative shear stress (e.g. Hanes & Inman 1985) as follows:
(1)
where D is the particle diameter, and ρp is the density of the individual particles, and is the shear strain rate equal to U/H. The solid fraction v is normalized by

maximum solid fraction v max
v*=v/vmax.
(2)
3 DISTINCT ELEMENT SIMULATION
3.1 Numerical model of apparatus
To simulate the experimental laboratory test, twodimensional Distinct Element Method (DEM) was used. Figure 7 shows the schematic view of the apparatus used in
the numerical simulation. By using Distinct Element particles, an inner rotary wall and an outer fixed wall, are simulated. The inner surfaces of the walls were roughened
with the same radius as the spherical steel balls. To simulate the experimental apparatus, particles were sheared between the inner and outer walls by rotating the inner
wall particles.
3.2 Determination of DEM Parameters
The DEM parameters are summarized in Table 2. Those parameters, e.g. size and stiffness, are determined from the steel ball used in the experimental study. Cylinder
particles are used as the distinct elements. Equivalent friction coefficient was applied to install both the rolling friction and sliding friction between the particle and the
bottom floor. Every particle has three degree of freedom, two horizontal movements and one rotation. Other two rotation components and vertical movement are
neglected in the simulation.
In the Distinct Element simulation, repulsive coefficient e on the collision is not used directly. Instead of e, damping coefficient is used, which depends on the
combination of stiffness and viscosity at the particle contact.
Figure 8 shows the relationship between damping coefficient h at the contact between Distinct Element and the repulsive coefficient e. One ball is released with some
initial velocities directly toward another ball that is at a standstill. From the velocity changes of two balls, e value is determined. According to Figure 8, it can be said
that there is a clear relationship between h and e, which is almost independent of initial velocity.
Page 432
Figure 7. Schematic view of the simulation.
Table 2. DEM parameters.
Radius r 0.50 cm
Mass 4.08 g
Density ρp 7.83 g/cm3

Volume
0.523 cm3
Rolling friction coefficient 0.0
Sliding friction coefficient 0.05
Friction coefficient between floors 0.05
Particle stiffness kn
1.14×107 N/m
Particle stiffness ks
1.14×106 N/m
Time increment
2×10−6 sec
Repulsive coefficient e 0.9
Figure 8. Relationship between damping coefficient h and repulsive coefficient e.
By using this graph, h value was determined as 0.05 which is equivalent to e=0.9.
3.3 Stress Measurement
In the simulation, the distribution of the particle velocity was also measured. It can be said that there is no strain localization and shear strain rate is almost constant in
whole cell.
In the simulation, the same stress measurement methods are used like as experimental work. In the numerical simulation not only τ but also normal stress σ are also
measured. As the representative stress, averaged stress are calculated by using Equation (1).
4 STRESS GENERATED IN THE FLOW
4.1 Comparison of measured stress
The relationship between measured averaged stress and shear strain rate are almost proportional to square of shear strain rate both in simulation and experiments, in
low strain rate range, stress seems independent of shear strain rate.
Figure 10(a) and (b) show the relationships between generated stress and fraction v*, of shear stress and normal stress respectively. Results of our experiments,
Bagnold (1954) and Hanes & Inman (1985) are also plotted. All the stress are nondimensionalized by using Equation (1). The results of numerical simulation seem to
agree well with the results of other previous studies. When fraction v* is smaller than 0.7, generated stress is small. In range of v* is greater than 0.7, non
dimensionalized stress suddenly increase. Shear stress of numerical simulation shows a good agreement with previous studies, however, normal stress of numerical
simulation is greater than results of previous studies. Because the generated stress has week dependence on shear strain rate when shear strain rate is slow, as seen in
Figure 9, some results of numerical simulation show larger nondimensionalized stress value.
4.2 Stress ratio
Figure 11 shows the relationship between fraction and stress ratio Other research data are also plotted. The results of numerical simulation show a good
agreement with other results. From Figure 11, it is difficult
Page 433
Figure 9. Measured stress vs. shear strain rate.
Figure 10. Nondimensional stress vs. normalized solid fraction.
Figure 11. Stress ratio vs. normalized solid fraction.
Page 434
Figure 12. Stress ratio vs. shear strain rate.
to find a clear relationship between stress ratio and fraction.
Figure 12 shows the relationship between stress ratio and shear strain rate. The range of stress ratio locates between the data obtained by Bagnold and Hanes &
Inman, while the tendency seems different. The stress ratio obtained in this study is almost independent of shear strain rate. However, other results show dependence,
both increasing and decreasing, on shear strain rate.
5 CONCLUSIONS
In this paper, twodimensional highspeed shear flow experimental apparatus was newly designed to clarify the characteristic of rapid granular flow. The range of the
shear strain rate is wide enough to cover the whole range of granular flow from quasistatic to inertiadominant; the movement of sphere particles can directly be
observed since particles move twodimensionally on the horizontal plate; the range of solid fraction is so wide that we can follow the wide range of microflowing
mechanism from streaming to multiple collision; and the shear stress generated by shear flow can be measured. Characteristics of this apparatus are as follows; (a) it
can deal wide range shear strain rate flows, (b) tangential stress can be measured; (c) movement of all particles can be observed by using a highspeed camera.
However, it is difficult to get the detailed data, such as contact condition of particles and distribution of particle rotation from laboratory test. To overcome this
difficulty, numerical simulation that emulates the apparatus is prepared. Twodimensional Distinct Element Method is used and ball elements are used as particles.
Boundary conditions are set to emulate the apparatus.
From the comparison between laboratory test and numerical simulation, the accuracy of the simulation is verified. By using those results, (a) effect of solid fraction on
shear stress generation, (b) effect of shear strain rate on the stress generation, (c) effect of solid fraction on the stress ratio and (d) effect of shear strain rate on the
stress ratio are studied.
REFERENCES
Bagnold, R.A. 1954. Experiments on a gravityfree dispersion of large solid spheres in a Newtonian fluid under shear. Proc. Royal Soc. Ser. A. 225:49–63.
Hanes, D.M. & Inman, D.L. 1985. Observations of rapidly flowing granularfluid materials. J. Fluid Mech. 50: 357–380.
Ichiba, K. 2003. Observations of rapidly sheared granular materials. Master Thesis, Saitama University.
Iwashita, K., Ichiba, K. & Oda, M. 2004. Observations of rapidly sheared granular materials (in Japanese), Journal of Geotechnical Engineering JSCE. (in press).
Savage, S.B. & Jeffrey, J.D. 1981. The stress tensor in a granular flow at high stress rates. J. Fluid Mech. 110:255–272.
Savage, S.B. & Sayed, M. 1984. Stresses developed by dry cohesionless granular materials in an annular shear cell. J. Fluid Mech. 142:391–430.
Page 435
Author index
5809 679 3
AbdelRahman, K. 45
Achmus, M. 45
Aoki, K. 73, 289
Arenson, L.U. 341
Ashmawy, A.K. 363
Azevedo, N.M. 323
Bertrand, D. 13
Bolton, M.D. 407
Boschetti, F. 235
Boutt, D.F. 301
Bray, J.D. 373
Butt, H.J. 157
Cao, G.Q. 57
Cheng, Y.P. 407
Christiansson, R. 85
Cook, B.K. 249, 301
Cui, L. 373
Cundall, P.A. 217, 257
Doornhof, D. 391
Durrlemann, S. 235
Fakhimi, A. 383
Feeser, V. 191
Ferellec, J.F. 103
Gotteland, Ph. 13
Groeger, T. 3
Grupe, B. 191
Haraki, D. 355
Heim, L. 157
Hirabayashi, H. 21
Hirata, A. 103
Hobbs, B.E. 235
Holt, R.M. 273, 391
Honda, T. 265
Hori, M. 209
Hunt, S.P. 331
Ichiba, K. 429
Imre, B. 199
Iwashita, K. 429
Jenner, C. 3
Johnson, S. 249
Johnson, S.M. 301
Kaitkay, P. 63
Kappl, M. 157
Kato, S. 347
Katsuki, S. 355
Katzenbach, R. 313
Kawamoto, H. 95
Kenter, C.J. 391
Kishino, Y. 143
Kondoh, D. 289
Konietzky, H. 3
Korsnes, R. 415
Kreiter, S. 191
Kwon, S. 181
Lee, C. 175, 181
Lei, S. 63
Lemos, J.V. 323
Li, L. 273
Lin, X. 225
Louchnikov, V. 331
Luding, S. 127, 157
Maeda, K. 21
Maejima, T. 73
Martin, C.D. 85
Matsuoka, T. 187, 289
Matsushima, T. 421
McBride, A.T. 111
Meyers, A.G. 331
Mito, Y. 73, 289
Mizumoto, M. 57
Mochizuki, C. 103
Mori, T. 73
Murariu, V. 119
Nakata, Y. 407
Narayanasamy, R. 397
Nicot, F. 13
O’Sullivan, C. 373
Ochiai, H. 257
Oda, M. 429
Oguni, K. 209
Okura, Y 257
Oomura, A. 21
Ord, A. 235
Park, E.S. 85
Park, Y 175, 181
Pierce, M.E. 149
Pohl, M. 35
Powell, M.S. 111
Psakhie, S.G. 241
Pulsfort, M. 35
Ree, J.H. 175
RegenauerLieb, K. 235
Runkles, B.D. 363
Saeki, M. 135
Sakakibara, T. 51, 347
Sakurai, T. 21
Sallam, A.M. 363
Sarracino, R.S. 111
Schmitt, A. 313
Sergeant, P. 119
Shen, X. 63
Shimizu, Y 257, 281
Shiraishi, H. 355
Strack, O.E. 301
Svoboda, J. 119
Suzuki, K. 51
te Kamp, L. 3
tom Woerden, F. 45
Tomas, J. 157
Towhata, I. 265
Tykhoniuk, R. 157
Ueda, S. 187
Ujihira, M. 51
Verboom, B. 391
Vietor, T. 165
Villegas, T. 383
Walz, B. 35
Williams, J.R. 225, 249, 301
Yamada, Y 187
Yoshimori, H. 347
Zhang, Y. 235
Zyryanov, D. 415

Numerical Modeling in Micromechanics Via Particle Methods 2004 Proceedings of The 2nd International PFC Symposium Kyoto Japan 28 29 October 2004

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Numerical Modeling in Micromechanics Via Particle Methods 2004 Proceedings of The 2nd International PFC Symposium Kyoto Japan 28 29 October 2004

Uploaded by

Copyright:

Available Formats

Cover

Initial stat Num 1 Exp 1 Exp 2 Exp 3 Exp 4

Grain shape Constant confining pressure Constant volume

Case no. Grain shape Initial density (e0) Friction coefficient of surface, tan δ Relative surface roughness, R/(D50) Overburden pressure σv (MPa)

The average strain rates in V can be determined by comparison between the discrete velocity field of particle centers ν (p)(xi) and the continuous velocity field ν(x i)

The tensor can be computed by a least­squares procedure to minimize the residual error χ between the velocity value ν (p)(xi) calculated by DEM and the

point resistance normalized by the overburden pressure (qp /σv ) are plotted against the relative settlement S/D under different σv conditions. The skin friction resistances

The point resistance qp /σv seems to decrease with increasing roughness at S/D<0.02, and the value of qp /σv fluctuates with skin friction mobilization. At S/D>0.04, the

φ (°) sf_rest (−) sf_active (−) ks_active (kN/m)

Herein h is the wall height, γ is the unit weight of the soil and k a is the active earth pressure coefficient, which in general depends on the angle of internal friction of

In the case under consideration k 0=0.48 and k a = 0.27. The reduction factors calculated correspondingly are given in Table 2 and compared with the values from the

N value Yield stress (MPa) Young Modulus (MPa)

Description Parameter Unit Block­stone Cobblestone Soil

Case Content Distance X of resultant force (mm) Displacement (mm)

Parameters Symbol Value

Ball size ratio, uniform distribution Rmax/Rmin 1.3

Ball stiffness ratio k n/ks 2.6

Property Carthage marble PFC specimen

Compressive strength at pressure of 34.4 MPa, σc (MPa) 186 180

Damping coefficient 0.7 0.4 0.015

Case Rock type Remarkable weak plane Confining pressure

Cases EC [GPa] k n /ks μ σC [MPa] τC [MPa]

Parallel bond normal strength, (MPa) 157±36

Parallel bond shear strength, (MPa) 157±36

Number of joint sets Poisson’s ratio Young’s modulus (GPa) Peak strength (MPa) Crack initiation stress (MPa)

Diameter (mm) 25.0 37.5 62.5 75.0

Parameter Unit p/p* p/w*

Cn N.s/m 5.4×102

Set cut­point (kg/m3) Feed rate (kg/h) CP50 (kg/m3) Ep (kg/m3)

Figure 2. Force law for two springs with stiffness k 1 and k 2 for initial loading and subsequent un/reloading, respectively. Attractive forces

δ0=(1−k 1/k 2)δmax.

the minimum force f min=−k cδmin, i.e. the maximum attractive force, is obtained as a function of the model parameters k 1, k 2, k c, and the history parameter δmax.

For large overlaps (in the fluid regime), the stiffness and the force is thus only dependent on k 2, independent of k 1. For smaller overlaps both k 1 and k 2 affect the force

Figure 5. Force laws for varying stiffness k 1, according to Eq. (8). (Left) Temperature increase: k 1 is reduced while remains

with the time scale t 0 on which a typical change takes place. Note that k 1, k 1(T), and k 1(T, t) are different, in general, and correspond to the maximum k 1, the

should have weak tensile and compressive strength. Therefore, we couple the cohesive parameter k c to the magnitude of k 1(T, t), which decreases with temperature

a slight increase in density is possible due to reorganizations. The increased temperature becomes only effective at time t sinter, when the relaxation time t 0 is decreased

In these expressions, δ and are the displacement and velocity of particle i relative to particle j, respectively. k N is the spring constant, cN is the damping coefficient, r

Spring constant k N (particle/wall) 1.3×107 Nm−3/2

where un is the displacement vector that consists of incremental translations and rotations in the nth iteration step. The resultant force vector FS changes in the (N+1)th

with the tangential spring stiffness k t and a tangential dissipation parameter γt. As long as is reached again, static friction is activated.

Figure 2. Peak friction μ bmax. Measured in all but one test as maximum friction. For μ b=0.8 and xf=1.0: peak friction was defined as the

Table 1. Experimental results for the low­, high­ and intermediate­friction series (μ d: particle friction, xf: fraction of non­rolling disks, μ bmax: bulk peak

μd xf μbmax μbre κ

Figure 4. Relative residual strength κ=μ bres/μ bmax. Symbols as in Figure 2. Filling indicates localization behavior. Open symbols: strong

Micro­property Value Unit

Macro properties (rock mass properties) Value Unit

Input of material property Computation of crack path Implementation of heterogeneity

where ni is the outer unit normal of the boundary ∂D. Using the identity σjk=σjlx k,l, we rewrite Eq. (28b) as εijk((σjlx k ),l−ρüjx k )=0, and calculate the volume average

where t is the out­of­plane thickness of the object, is the “radius” of the object (i.e., its half­height) and k n is the tangent stiffness of the layer of fibers (stress per

where is the interaction force per unit volume between particles and fluid, dpi(i=1, np ) is the diameter at particle i and np is number of particles in the cell. The

where w=unit width; (xk , yk ) is a coordinate of center of the adjacent particles that forms a closed loop. Thus, the change in domain volume at time step, i is calculated

where VG, VL and V=gas volume, liquid volume and domain volume; ν G and ν L=gas viscosity and liquid viscosity; g=gravity vector; and nC=the unit normal vector of

Microproperties Notation Unit Value

Ball stiffness ratio k n/ks 1.2

Contact­bond shear strength τ C MPa 137

Specimen Young’s modulus [GPa] Poisson’s ratio

Microproperties Notation Unit Value

Normal force for calibration F0 N 1×1012

Multiplier for channel aperture ma 1×10−3

where f i(x,t) is the fluid particle density distribution with velocity e i at a node located at position x for a given time t; τ is the relaxation time; and is the

where vs is the solid’s velocity at time t and −i is used to denote the distribution component having the opposite direction to i (e.g. f 5 for f 1, see Fig. 1). A modification

Ec [GPa] v c Cc [MPa] μc

The tensor can be computed by a leastsquares procedure to minimize the residual error χ between the velocity value ν (p)(xi) calculated by DEM and the

Description Parameter Unit Blockstone Cobblestone Soil

Set cutpoint (kg/m3) Feed rate (kg/h) CP50 (kg/m3) Ep (kg/m3)

Table 1. Experimental results for the low, high and intermediatefriction series (μ d: particle friction, xf: fraction of nonrolling disks, μ bmax: bulk peak

Microproperty Value Unit

where t is the outofplane thickness of the object, is the “radius” of the object (i.e., its halfheight) and k n is the tangent stiffness of the layer of fibers (stress per

Contactbond shear strength τ C MPa 137

Mean contactbond shear strength, τ c (mean) 82.0 MPa

Std. deviation of the mean contactbond shear strength, τ c (std. dev.) 30.0 MPa

Parallelbond stiffness Parallelbond stiffness ratio 1.1

Spring constant Damping ratio Critical compressive Critical tensile Pmin Nonlinear parameter n Cohesion C (N) Friction angle