Content uploaded by Waiching Sun

Author content

All content in this area was uploaded by Waiching Sun on Nov 07, 2016

Content may be subject to copyright.

International Journal for Multiscale Computational Engineering, 14 (4): 389–413 (2016)

IDENTIFYING MATERIAL PARAMETERS FOR A

MICRO-POLAR PLASTICITY MODEL VIA X-RAY

MICRO-COMPUTED TOMOGRAPHIC (CT) IMAGES:

LESSONS LEARNED FROM THE CURVE-FITTING

EXERCISES

Kun Wang,1WaiChing Sun,1,∗Simon Salager,2SeonHong Na,1&

Ghonwa Khaddour2

1Department of Civil Engineering and Engineering Mechanics, Columbia University, 614 SW

Mudd, Mail Code: 4709, New York, New York 10027, USA

2UJF-Grenoble 1, Grenoble-INP, CNRS UMR 5521, 3SR Laboratory, 38041 Grenoble, France

∗Address all correspondence to: WaiChing Sun, E-mail: wsun@columbia.edu

Unlike a conventional ﬁrst-order continuum model, the material parameters of which can be identiﬁed via an inverse

problem conducted at material point that exhibits homogeneous deformation, a higher-order continuum model requires

information from the derivative of the deformation gradient. This study concerns an integrated experimental-numerical

procedure designed to identify material parameters for higher-order continuum models. Using a combination of micro-

CT images and macroscopic stress–strain curves as the database, we construct a new ﬁnite element inverse problem

which identiﬁes the optimal value of material parameters that matches both the macroscopic constitutive responses and

the meso-scale micropolar kinematics. Our results indicate that the optimal characteristic length predicted by the con-

strained optimization procedure is highly sensitive to the types and weights of constraints used to deﬁne the objective

function of the inverse problems. This sensitivity may in return affect the resultant failure modes (localized vs. dif-

fuse), and the coupled stress responses. This result signals that using the mean grain diameter alone to calibrate the

characteristic length may not be sufﬁcient to yield reliable forward predictions.

KEY WORDS: micro-CT imaging, micro-polar plasticity, critical state, higher-order continuum, Hostun

Sand

1. INTRODUCTION

Granular matters are some of the most commonly encountered materials in our daily lives. Ranging from natural

materials, such as sand, slit, gravel, fault gauges to the man-made materials, such as pills, sugar powder, and candy,

granular matters may be composed of grains with various shapes, sizes, and forms. The mechanical behavior of

granular assemblies are therefore not only depending on the material properties of the grains that form the assemblies,

but also the evolution of the grain contact topology or fabrics (Kuhn et al., 2015; Lenoir et al., 2010; Li and Dafalias,

2011; O’Sullivan, 2011; Satake, 1993; Subhash et al., 1991; Sun et al., 2013; Walker et al., 2016; Wang and Sun,

2015). In the past decades, theoretical and numerical studies have achieved great success to describe and predict the

granular materials that exhibit solid-like behavior when subjected to sufﬁcient conﬁning pressure. In those cases,

granular materials are often treated as ﬁrst-order Boltzmann continua which possess no internal microstructures or

length scale. For instance, constitutive models that employ the critical state soil mechanics concept are now able

to replicate the monotonic and cyclic responses of sand of different relative densities, conﬁning pressure and initial

states (Been and Jefferies, 1985; Borja and Sun, 2007; Dafalias and Manzari, 2004; Lade and Duncan, 1975; Manzari

1543–1649/16/$35.00 c

⃝2016 by Begell House, Inc. 389

390 Wang et al.

and Dafalias, 1997; Pestana and Whittle, 1999; Schoﬁeld and Wroth, 1968; Sun, 2013; Wu et al., 1996). In these

macroscopic models, microstructural attributes, such as pore size distribution, grain shapes, and size distribution,

grain-scale heterogeneity are not explicitly taken into account at the grain scale. Instead, they are represented by a set

of internal variables that represents the loading history. The evolution of these internal variables are then governed

by phenomenological evolution laws that require material parameters calibrated from experiments. Nevertheless, the

success on replicating constitutive responses at a material point level often does not automatically lead to realistic

simulated responses. For instance, the classical ﬁrst-order continuum model does not possess physical length scale and

this may cause ﬁnite element or ﬁnite difference models to exhibit spurious mesh dependence, unless a regularization

limiter is applied (Bazant et al., 1984; Belytschko et al., 1988; Fish et al., 2012; Lasry and Belytschko, 1988; Liu

et al., 2015; Na and Sun, 2016; Needleman, 1988; Sun and Mota, 2014). This regularization limiter can be introduced

via various means, ranging from incorporating rate dependence, inserting embedded strong discontinuity to adapting

nonlocal and gradient-based constitutive relations.

For instance, classical bifurcation analysis by Rudnicki and Rice (1975) and Issen and Rudnicki (2000) can be used

to predict the onset and orientiation of a deformation band. Beyond the bifurcation point, additional degree of freedom

that represent an embedded strong discontinuity can be introduced via assumed strain [e.g., Borja (2000)], extended

ﬁnite element [e.g., Song et al. (2006)] or localization or bifurcated elements [e.g., Belytschko et al. (1988); Ortiz

et al. (1987); Yang et al. (2005)]. Nevertheless, recent experimental and theoretical studies have both indicated that

granular materials may initially form multiple deformation bands at the bifurcation point, but the interaction among

deformation band may lead to some of the deformation band vanishing. Meanwhile, a dominated persistent shear band

may emerge among them in the post-bifurcation regime (Borja et al., 2013; Gajo et al., 2004; Rechenmacher, 2006).

This discovery indicates that the adaptive insertion of the enhanced mode(s) may lead to mesh bias unless there is a

coarsen mechanism in place to prevent the ﬁxture of shear band localization (Lin et al., 2015).

Opposite to the enrichment approach, higher-order continuum theory is another attractive option to circumvent the

pathological mesh dependence exhibited in the Boltzmann continua. The history of generalized continuum theories

can be tracked back to the work of the Cosserat brothers who published a book titled “Th´

eorie des corps d´

eformables

in 1909” (Cosserat and Cosserat, 1909). As pointed out by Maugin and Metrikine (2010), there exist multiple inter-

pretations and generalizations of Cosserats’ work. Nevertheless, the necessary ingredients of the higher-order con-

tinuum theory are the incorporation of higher-order kinematics (micro-rotation, micro-stretch, micro-torsion, etc.)

and additional internal degree of freedom that breaks the symmetry of the Cauchy stress (Eringen, 1999; Maugin

and Metrikine, 2010; Mindlin, 1964). The micropolar continuum theory is a sub-class of the generalized continuum

theory in which the material point is associated with micro-structures exhibiting not only deformation but changing

orientations that are independent of the macroscopic deformation. These changes of orientations can be represented

by Euler’s angle, quaternions, and spinors, among other mathematical tools (Duhem, 1893). By incorporating the

additional rotational degrees of freedom at the microscopic level, micropolar effect of granular materials, especially

those associated with the onset of strain localization as observed by Desrues and Viggiani (2004); Hall et al. (2010),

can be replicated properly in numerical simulations without exhibiting the spurious mesh dependence.

Nevertheless, incorporating the micropolar effect to improve the accuracy and physical underpinning of the nu-

merical models also comes with a price. For instance, the identiﬁcation of the the micropolar material parameters in

addition to the non-polar counterparts can be difﬁcult for material testing procedures that are originally designed for

homogeneous specimen subjected to ﬁrst-order boundary conditions. The assumption that the strain ﬁeld developed

inside the specimen remains approximately homogeneous in convectional triaxial and shear apparatus is unavoidably

violated when high-order kinematics and couple stress are considered. These difﬁculties are partly responsible for the

limited applications of higher-order continuum theory in engineering practices.

Our objective is to overcome this technical barrier by introducing a simple experimental-numerical method de-

signed speciﬁcally for identifying material parameters for higher-order continua. In particular, we introduce the meso-

scale kinematics data from X-ray tomographic images obtained from drained triaxial compression tests to construct

a multiscale objective function. The tomographic images at selected axial strain levels are combined with the con-

ventional macroscopic stress–strain curve and the state path to constitute the measured data set. A new objective

function is then deﬁned to simultaneously minimize the discrepancy between the simulated and measured macro-

scopic responses while enforcing the simulated grain-scale responses to be consistent with the kinematics information

International Journal for Multiscale Computational Engineering

Identifying Micropolar Material Parameters via Micro-CT Images 391

provided by the tomographic images. Our results show that the incorporation of microstructural information from

the grain scale may profoundly change the predicted length scale and failure mode of the numerical specimen. More

importantly, the predicted coupled stress response is also highly sensitive to the way microstructural information is

incorporated in the inverse problem.

To the best knowledge of the authors, this work is the ﬁrst contribution to employ X-ray tomographic images

to identify material parameters via a multiscale inverse problem for granular materials. In addition, the proposed

method provides a new method to quantify the sensitivity of characteristic length predicted by material identiﬁcation

procedure. The ﬁndings from the Hostun sand experiment indicate that a multiscale model capable of replicating

the macroscopic responses may nevertheless produce incorrect bifurcation modes and completely different micro-

scale responses. This result has important implications for material modeling, as it clearly shows that calibrating

macroscopic responses alone is insufﬁcient to provide reliable forward prediction for multiscale models. In other

words, a seemingly good match between experimental and simulated responses at the macroscopic level may be a

consequence of high-quality grain-scale simulations or a product manufactured by excess tuning and manipulations

of meso-scale material parameters.

The rest of this paper is organized as follows. In Section 2, we ﬁrst describe the experimental procedure used to

obtain the tomographic images. Then, the ingredients that constitute the inverse problem for a micropolar critical state

plasticity model are discussed in Section 3. The calibration of a single unit cell and the ﬁnite element inverse problem

are described in Sections 4 and 5, followed by a discussion of major ﬁndings. Finally, concluding remarks are given

in Section 6.

As for notations and symbols, bold-faced letters denote tensors; the symbol “·” denotes a single contraction of

adjacent indices of two tensors (e.g., a·b=aibior c·d=cij djk ); the symbol “:” denotes a double contraction of

adjacent indices of tensor of rank two or higher (e.g., C:ϵe=Cijkl ϵe

kl); the symbol “⊗” denotes a juxtaposition

of two vectors (e.g., a⊗b=aibj) or two symmetric second order tensors [e.g., (α⊗β) = αijβkl]. As for sign

conventions, we consider the direction of the tensile stress and dilative pressure as positive.

2. EXPERIMENTS AND MICRO-CT IMAGES FOR VALIDATIONS

This section introduces the triaxial test performed for this research at 3SR Lab and details the experimental proce-

dures followed. The objective of performing the triaxial tests is to provide a multiscale benchmark database for model

calibration and material identiﬁcation. The bemchmark data consist of the macroscopic stress–strain curve and the

grains position at different stress levels during the shearing, characterizing the grains-scale behavior of the granular

media. These data form the basis for a multiscale objective function to search the optimal material parameters to pre-

dict the most consistent mechanical behavior measured by a weighted norm that minimizes the discrepancy between

simulations and experimental observations at both grain- and specimen scales.

2.1 X-Ray CT

X-ray micro-CT is a non-destructive, 3D high-resolution imaging method that allows the internal structure of the

scanned objects to be investigated (And`

o et al., 2012; Hall et al., 2010; Lenoir et al., 2010; Sun et al., 2011). Micro-

CT technology yields images that map the variation of x-ray attenuation within the objects, based on the composition

of the object. Because sand and water have different x-ray attenuation coefﬁcients, a signiﬁcant contrast for the two

phases is observed in an x-ray transmission image, Fig. 1. A quantitative analysis that provides detailed information

on the arrangement and distribution of particles can be built from the processed 3D images.

The imaging process can be summarized as follows: the cell ﬁrst is placed on a turntable stage whose rotation

can be accurately controlled. An x-ray source generates a continuous x-ray beam; the beam passes through the object

and casts an x-ray shadow onto the detector, see Fig. 2. The radiation that hits the detector is converted into an

electronic charge, that is subsequently passed to a computer to create a radiograph (i.e., digital image). A series

of images (radiographies) is acquired while rotating the object step by step through 360◦at a pre-deﬁned angular

increment. These radiographies record projections that contain cumulative information on the position and density of

Volume 14, Issue 4, 2016

392 Wang et al.

FIG. 1: Reconstructed x-ray image shows the clear contrast among grain, water, and air. The experiments in this study

are performed using x-ray setup of Laboratory 3SR at Grenoble University.

FIG. 2: X-ray CT scanner at Laboratory 3SR (left) and the whole arrangement of triaxial setup inside x-ray cabinet

(right)

the absorbing features within the specimen. The data obtained can be used to perform the numerical reconstruction of

the ﬁnal 3D image (And`

o et al., 2012; Hall et al., 2010).

2.2 Triaxial Cell

The triaxial cell used in this work is designed to allow x-ray scanning and to ﬁt to the mechanical loading system

available in 3SR Laboratory. The cell is made of PMMA, which is transparent to X-rays, and quite resistant to cell

pressure (conﬁnement) applied in triaxial tests. Figure 3 shows the PMMA cell in front of X-ray source, with the

specimen installed at its place inside the cell. The loading system applies and measures the axial force compressing

the specimen from below, and the vertical displacement of the lower piston. The maximum axial force that can be

measured by the force meter, which is in contact with the bottom of the lower piston is 0.5 kN. The measurement of

the axial displacement is made by a linear variable differential transformer (LVDT) which is attached to a tie bar and

measures the vertical displacement of the loading head. These measurements allow plotting the relation between the

axial stress and axial shortening during the triaxial test. The speed range for the loading head is from 0.1 µm/min to

100 µm/min, which corresponds to a strain rate from 0.0005 % to 0.5% per minute (for a specimen of height 20 mm).

The speed range considered in this work for the three triaxial tests is 20 µm/min (strain rate is 0.1% for a specimen of

height 20 mm). The motor is driven remotely from a laptop used for data acquisition. The system of data acquisition

registers the force, the conﬁnement pressure, and LVDT measurements which were calibrated at the beginning of the

test.

International Journal for Multiscale Computational Engineering

Identifying Micropolar Material Parameters via Micro-CT Images 393

FIG. 3: PMMA cell in front of x-ray source with the specimen installed at its place inside the cell

2.3 Material Used

The experimental program is conducted on Hostun sand (HN31). Hostun sand is used as a reference material in

different laboratories [cf. Amat (2008); Desrues and And`

o (2015); Sadek et al. (2007)]. Its chemical components

principally consist of silica (SiO2>98%). The grain shape is angular as shown in the scanning electron microscope

image (SEM) of a few grains of Hostun sand, taken from Flavigny et al. (1990), in Fig. 4(a). A particle size distribution

analysis for Hostun sand is shown in Fig. 4(b).

2.4 Specimen Preparation

The specimen prepared for the triaxial tests is a cylindrical specimen of 1 cm diameter and 2 cm height (slenderness

ratio 2). The technique used for specimen preparation in the three triaxial tests is water pluviation, and thus all the

FIG. 4: (a) SEM image of Hostun sand from Flavigny et al. (Reprinted with permission from Jacques Desrues,

Copyright 1990), (b) Grain size distribution of Hostun sand from Amat (2008)

.

Volume 14, Issue 4, 2016

394 Wang et al.

pores in the specimen are completely ﬁlled with de-mineralized water in the initial state. This procedure ensures that

the specimen is fully saturated at the beginning of the mechanical tests.

2.5 Macroscopic Response

The axial stress is calculated by the axial force measured over the cross section of the specimen at the initial state.

The axial force comes from the raw measurement of force. The axial strain (in %) is obtained from the shortening

applied by axial compression of the specimen, with respect to its initial height. The specimen behaves as expected for

a dense granular material: there is a peak in the specimen’s axial stress response, followed by strain softening until a

plateau of residual stress is reached. The peak is reached at axial strain ϵa= 5.5%, and at axial stress σ= 620 kPa.

The residual stress is σ= 440 kPa. Figure 5 shows the cross section of the deformed specimen at different axial strain

level. While the initial deformation remains relatively homogeneous at low axial strain, the specimen subsequently

deforms into a barrel shape while large strain (ϵa= 21%) develops toward the end of the test.

2.6 Segmentation

Measurements of the spatial distribution of grains require a separation of the solid and water phases. The volume has

been “binarized” using thresholds, one between air and water peaks, and the second one between water and grain

peaks. This last one is chosen in such a way that volume of voxels identiﬁed as grain phase correspond to volume of

grains that has been estimated from the weight of the sample (in a dry state). The threshold between air and water

is chosen to correspond to the local minimum in the gray level distribution. Figure 6 shows the cross section of the

processed images of which the spatial distribution of the solid and water phases are ideniﬁed based on the threshold

values.

FIG. 5: Micro-CT images of the cross section of the specimen at different axial strain level

International Journal for Multiscale Computational Engineering

Identifying Micropolar Material Parameters via Micro-CT Images 395

FIG. 6: Cross sections of the “binarized” two-phase images at axial strain = 0, 8, and 21%.

2.7 Grains Position

Water phase is separated from the whole 3D volume using Visilog by setting the threshold to a gray value equal to 128.

In order to remove the noise inside the images, represented by water volumes of one voxel, a morphological process

of erode and dilate by one voxel is applied. Later, water clusters are labeled. Labeling process can be summarized as

giving each individual water cluster an Identiﬁcation (Id). This Id is represented in the images by colors (i.e., each color

represents a speciﬁc water cluster). The process of labeling is very useful to be able to refer to any water cluster directly

by its Id number. These images where each water cluster has a speciﬁc Id and color can be considered as a mask applied

to the reconstructed image (i.e., the gray-scale information is always kept). At this point, more information about

water clusters is needed as the number of clusters, the center of mass, 3D volume, 3D area, ...etc. This information

is obtained using another tool of Visilog “analysis individual.” The measurements of clusters properties are saved in a

DAT ﬁle type associated primarily with data which can be read by a text editor.

3. INVERSE PROBLEMS FOR MICROPOLAR PLASTICITY MODEL

This section presents the calibration procedure for the micropolar hypoplasticity model via the micro-CT images.

The macroscopic constitutive model and the parameter identiﬁcation procedure conducted via the open source toolkit

Dakota [cf. Adams et al. (2009)] are brieﬂy described. We will ﬁrst treat the entire specimen as a single unit cell and

we will use the gradient based approach to obtain the material parameters. The open-source optimization code Dakota

is employed. Then we make use of the micro-CT images to construct the geometry of the Hostun sand sample at the

beginning of the triaxial compression loading. The initial distribution of void ratio is also inferred from micro-CT

information. They serve as the starting point of the model simulations. We calibrate the material parameters by two

sets of benchmark data. One set consists of the macroscopic stress ratio and volumetric strain vs. the axial strain

curves. The other set contains the geometry and void ratio distribution when the axial stress reaches the peak and at

the residual stage.

Volume 14, Issue 4, 2016

396 Wang et al.

3.1 Micropolar Finite Element Model

The inverse problem introduced in this paper is applicable to a wide spectrum of constitutive laws that considers

high-order kinematics. In this study, we extend the previous two-dimensional implementation of the micropolar hy-

poplasticity model in Lin and Wu (2015), Lin et al. (2015) to a 3D formulation and use the 3D model to calibrate

with experimental data. This model can be considered as an extension of the non-polar hypoplasticity model in Wu

et al. (1996) and is suitable to characterize the Houstun sand specimen obtained for this study. For completeness, we

include a brief outline of the 3D ﬁnite element formulation of the micropolar hypoplasticity model. Readers interested

at the details of the micropolar hypoplasticity constitutive model may refer to Wu et al. (1996), Lin and Wu (2015),

Lin et al. (2015).

The non-symmetric Cauchy stress sand the coupled stress µare given in nonlinear path-dependent rate form in

terms of the current state of the stresses, the strain rate ˙e, the curvature rate ˙

κ, and the void ratio e. The micropolar

constitutive equations writes

◦

s=C1tr(s)˙e+C2tr(˙e)s+C3ψs+C4fd|| ˙e||2+l2|| ˙

κ||2(s+s′),

◦

µ=C1l2tr(s)˙

κ+C2tr(˙e)µ+C3ψµ+ 2C4fd|| ˙e||2+l2|| ˙

κ||2µ,(1)

where ◦

sand ◦

µare the Jaumann rates of sand µ,s′is the deviatoric part of s, and

ψ=tr(s·˙e)−tr(µ·˙

κ)

tr(s),(2)

whereas

fd= (1 −a)e−emin

ec−emin +a(3)

is a linear scalar function that accounts for the effect of critical state on the constitutive behavior. The material pa-

rameters to be determined are C1,C2,C3,C4,a, the critical void ratio ec, the minimum void ratio emin, and the

characteristic length l, which is the only additional parameter in micropolar framework compared to classical con-

tinuum hypoplasticity model. Note that this micropolar hypoplasticity model is valid under the assumption that ˙

κ,µ,

and ◦

µare all skew-symmetric. An explicit numerical integration scheme (Forward-Euler) is adopted to integrate the

highly nonlinear constitutive equations Eq. (1), i.e.,

st1≈st0+˙st0(st0,µt0,˙e,˙

κ,et0)∆t, µt1≈µt0+˙

µt0(st0,µt0,˙e,˙

κ,et0)∆t. (4)

The ﬁnite element implementation for 2D small strain problems was previously described in great detail (Lin

et al., 2015). However, the 3D ﬁnite element counterpart has not yet been established. Since the tomographic images

are three-dimensional and the bifurcation modes can be either symmetric or non-symmetric, an extension to three-

dimensional space is necessary for comparing experimental data and simulated responses at the meso-scale level.

As a result, the micropolar hypoplasticity model is implemented with a 3D micropolar ﬁnite element model. Con-

sequently, the degrees of freedom for each node consist of three translational DOFs and three rotational DOFs, i.e.,

u= [u1u2u3ϕ1ϕ2ϕ3]T. Adopting the Voigt notation, the vectors that store the components of the generalized

strain and stress tensors read

{s}= [s11 s22 s33 s12 s21 s23 s32 s13 s31 µ21 µ32 µ31]T,

{e}= [e11 e22 e33 e12 e21 e23 e32 e13 e31 κ21 κ32 κ31]T.(5)

International Journal for Multiscale Computational Engineering

Identifying Micropolar Material Parameters via Micro-CT Images 397

The element strain-displacement matrix Beis modiﬁed accordingly and takes the form:

Be= [B1,B2, . . . Bnen ],Ba=

Na,10 0 0 0 0

0Na,20 0 0 0

0 0 Na,30 0 0

Na,20 0 0 0 Na

0Na,10 0 0 −Na

0Na,30Na0 0

0 0 Na,2−Na0 0

Na,30 0 0 Na0

0 0 Na,10−Na0

0 0 0 0 Na,10

0 0 0 0 0 Na,2

0 0 0 0 0 Na,1

,(6)

where nen is the number of element nodes, Nais the shape function of element node a,Na,1,Na,2,Na,3are the

derivatives of the shape function Nawith respect to x1,x2,x3directions in a Cartesian coordinate system. Consider

a body Bsubjected to traction tand body force bpartitioned into element Be. The matrix form of the balance of linear

and angular momentum therefore reads

FINT =FEXT,(7)

where

FINT =Anel

e=1 Be

BeT·sdV ;FEXT =Anel

e=1 Be

NeT·bdV +∂Be

NeT·tdV,(8)

where Ais the assembly operator [cf. Hughes (2012)] and Ne= [N1,N2,N3, ...Nnen ].

3.2 Optimization Algorithms for Material Parameter Identiﬁcation

Material parameter identiﬁcation of soil samples can be done via solving inverse problems. Previous works, such as

Herle and Gudehus (1999), use multiple simple experiments, such as angle of repose, minimum and maximum index

density, shear test, and oedometric compression test to identify material parameters in an almost one-by-one fashion

for a non-polar hypolasticity model parameter for granular materials. Meanwhile, Ehlers and Scholz (2007) introduce

a two-step staggered procedure for micropolar constitutive model for granular materials in which the identiﬁcation

algorithm ﬁrst seeks the standard non-polar elasto-plastic material parameters from a series of cyclic triaxial tests that

keeps the specimen in homogeneous states, then searches for the optimal values for the micropolar counterpart with

another set of biaxial compression tests. In both approaches, the material identiﬁcation procedure requires multiple

mechanical tests on multiple specimens for the same materials. Nevertheless, due to the particulate nature of the

granular materials, it is almost impossible to reconstruct specimens that have identical microstructural attributes.

In this work, we propose a different approach that only requires one mechanical test to identify all material pa-

rameters. Instead of using multiple tests to constitute sufﬁcient constraints to identify all material parameters, the new

approach uses tomographic imaging techniques to track the spatial variability of the void ratio and deformation and

uses this kinematics information as the additional constraints for the objective function. One key assumption we made

here is that the spatial relative density or porosity variation is the dominating factorfor the post-bifurcation responses,

but the material parameter variation is negligible. This approximated approach has been found to be successful in

yielding consistent macroscopic constitutive responses and bifurcation modes for sand specimen that evolves with

complex bifurcation modes. In particular, Borja et al. (2013) found that by taking account only the relative density

Volume 14, Issue 4, 2016

398 Wang et al.

variation, one may correctly predict the onset of multiple shear bands at the bifurcation points, the sequential interac-

tion mechanisms of these shear bands, and the persistent shear band that lasts till the end of the loading program.

The calibration of the micropolar hypoplasticity model is completed by ﬁnding the optimal values of the material

parameters that minimize the discrepancy between the simulated results and the experimental data via iterations.

How “optimal” the values of the material parameters are is determined by the speciﬁc objective function used for the

inverse problem. If a global minimum of the objective function exists, then an optimization algorithm may start from

differential initial guess but ﬁnd the same set of optimal material parameters. In this work, we use a gradient-based

method to ﬁnd the optimal values of the material parameters. As a result, multiple boundary value problems are run

and the results and their corresponding gradient with respect to the parametric space are computed. The objective

function we used takes the following form:

f(x) =

N

i=1

wiri[si(x), di]2,(9)

where xis the vector of all material parameters, Nis the number of available experimental data for calibration, diis

the ith data point, wiis its weight, and si(x)is the corresponding result obtained from model simulation using the

parameter set x.riis the residual between the simulation data si(x)and the experimental data di. In this work, we

normalize the residuals by deﬁning them as the relative errors, i.e.,

ri[si(x), di] = si(x)−di

di

,(10)

since the laboratory data we employed for parameter calibration include the stress ratio, volumetric strain, and void

ratio, which are not in the same order.

In this work, we employ the Dakota analysis toolkit which provides ﬂexible interface between simulation codes

and robust optimization algorithms (Adams et al., 2009). Among the available methods we choose the gradient-based

least-squares algorithm (NL2SOL) which minimizes f(x)using its gradient:

∇f(x) = J(x)r(x), Jij =∂ri

∂xj(11)

and approximates the Hessian matrix with Gauss–Newton method:

∇2f(x) = J(x)2.(12)

The parameter calibration procedure requires an input ﬁle specifying the chosen optimization algorithm, the pa-

rameters to be determined, their initial values, upper and lower limits, and the experimental data set to be compared.

At the beginning of each iteration, Dakota passes the current guess of the material parameters to the ﬁnite element

simulation code developed for this study. The focus of this research is not the development of state-of-the-art opti-

mization algorithm to identify material parameters. Interested readers may refer to, for instance, Mahnken and Stein

(1996), Cooreman et al. (2007), Adams et al. (2009), Hu and Fish (2015) for details.

The ﬁnite element code then runs the simulations using the material parameters provided by Dakota and returns

the simulation results back to Dakota to compute the discrepancy measured by the objective function, and the cor-

responding gradient. Based on these inputs, Dakota then determines the next iteration of material parameters and

repeats the gradient estimation procedure and the discrepancy is again measured after another run of the simulation

code (Adams et al., 2009). This procedure repeats until the discrepancy is below a pre-deﬁned tolerance or when the

number of iterations reaches its maximum.

4. UNIT CELL CALIBRATION

The Dakota calibration scheme is ﬁrst applied to calibrate a unit cell represented by a tri-linear 8-node quadrilateral

element representing the entire granular sample. This test is conducted under the additional assumptions that (1)

International Journal for Multiscale Computational Engineering

Identifying Micropolar Material Parameters via Micro-CT Images 399

the specimen is composed of an effective medium of homogeneous properties inferred from the real heterogeneous

specimen and (2) the deformation remains homogeneous. With these additional assumptions, the inverse problem

is simpliﬁed such that only macroscopic responses from the experiment are utilized in the objective function [as

demonstrated previously in Liu et al. (2016)]. The target data for curve-ﬁtting are the stress ratio s (σ1/σ3)lab and

volumetric strain elab

valong the triaxial compression test. As a result, the objective function reads

f(x) =

Nσ1/σ3

i=1

wiriσ1

σ3model

i

(x),σ1

σ3lab

i2

+

Nev

j=1

wjrjemodel

vj(x), elab

vj2

,(13)

where Nσ1/σ3and Nevare the number of stress ratio data and volumetric strain data, respectively. Both data have 30

data points and the weights are equal to 1. The subscript idenotes the ith data point. The unit cell and the boundary

conditions are presented in Fig. 7. The initial void ratio and the minimum void ratio are set to be 0.6, which is the

average value of the sample. Note that there is no micropolar effect in this single element example, thus the material

length ldoes not affect the macroscopic response. It is set to be equal to the element size 10 mm. The parameters to

be identiﬁed via Dakota are C1,C2,C3,C4,ec, and a.

The Dakota calibration procedure takes in total 92 evaluations, of which 72 evaluations are performed for deter-

mination of the gradient of the 6 material parameters, while the remaining 20 evaluations are making guesses based

on the gradients. To demonstrate the convergence of the material parameters, the trial material parameters and the cor-

responding values of the objective function are presented in Table 1. The macroscopic responses obtained by the trial

material parameters compared with the experimental data are shown in Fig. 8. Despite the large discrepancy between

the initial guess and the laboratory results, f(x)decreases rapidly: it is reduced by about 95.7% after 50 evaluations.

This example demonstrates the robustness of the NL2SOL scheme in Dakota for nonlinear models and least-squares

problems for which the residuals do not tend to vanish.

FIG. 7: Domain and boundary condition for single unit cell calibration

TABLE 1: Evolution of the material parameters during the Dakota calibration procedure

C1C2C3C4eca f(x)

Initial guess −33.33 −104.61 −336.44 −105.90 0.800 0.900 143.77

Evaluation 20 −31.63 −62.44 −484.71 −107.15 0.650 0.905 97.60

Evaluation 50 −40.60 −1054.35 −1565.66 −138.86 0.651 0.923 6.18

Calibration result −63.14 −1831.49 −2563.92 −237.52 0.732 0.848 2.85

Volume 14, Issue 4, 2016

400 Wang et al.

FIG. 8: Macroscopic stress ratio (left) and volumetric strain (right) responses from selected value sets of material

parameters in unit cell calibration

5. MICROPOLAR HYPOPLASTIC MODEL CALIBRATION WITH MICRO-CT IMAGES FROM TRIAXIAL

COMPRESSION TEST

To analyze how the spatial variability of porosity (or relative density) affects the macroscopic responses, we recon-

struct a detailed 3D numerical specimen with the exact geometry and porosity distribution of the laboratory specimen

using the data extracted from micro-CT image analysis. The process of converting the micro-CT experimental data

into numerical specimen is illustrated in Fig. 9. After obtaining the images from the X-ray CT scan, the position and

effective diameter of each grain are recorded. Three micro-CT images taken at initial (0% axial strain), peak (6%

axial strain) and residual (15% axial strain) stages are used. The boundary particles are identiﬁed and thus the outer

boundary of the 3D specimen can be extrapolated from the position of these particles. Following this step, the domain

of the specimen is discertized by ﬁnite element and the void ratio of each ﬁnite element is calculated by established

the total solid volume using the positions and effective diameters of the particles, as shown in Fig. 9.

All micro-CT based ﬁnite element simulations are performed on the numerical specimen with identiﬁed initial

geometry and initial void ratio distribution. Here we adopt the hypothesis that the dominating factor that governs the

transition from compressive to dilatant behavior of granular materials is the relative density or porosity, the same

simpliﬁcation used in Borja et al. (2013). As a result, the material parameters C1,C2,C3,C4,ec,a, and lare assumed

to be homogeneous within the specimen, while the spatial variation of the void ratio inferred from micro-CT image

of the initial conﬁguration is incorporated to study the effect of the spatial variation of voidratio. Note that the results

of the inverse problem depend on how the boundary conditions are applied in model simulations. In this study, these

conditions are deﬁned based on the experimental setup with assumptions and simpliﬁcations. For example, because

of the complexity of the interaction between the Hostun sand sample and the loading pistons of the triaxial cell, the

top and bottom surfaces of the specimen are not fully constrained, neither in terms of the transnational nor rotational

degrees of freedom (as shown in Fig. 4). In particular, we observed that the loading plates placed on the top of the

specimen has slid. In this study, the authors assume that the nodal displacements on the top surface of the specimen

are totally constrained, while the bottom surface is compressed under a constant strain rate in the Z direction and all

nodes at the bottom boundary have the same vertical displacement in the XY plane in order that the surface area does

International Journal for Multiscale Computational Engineering

Identifying Micropolar Material Parameters via Micro-CT Images 401

FIG. 9: Construction of the numerical sample and void ratio distribution from micro-CT experimental data at the

initial stage (0% axial strain) of the triaxial compression test

not change. This constraint is applied via the Lagrange multipliers. Meanwhile, the rotations on both surfaces are

prohibited. The lateral surface of the FEM model is under constant conﬁning pressure of 100 kPa and is free to rotate,

neglecting the effect of the rubber membrane in the testing apparatus.

The Dakota calibration procedure of the material parameters is carried out for two inverse problems. In the ﬁrst

problem (Case A), only the macroscopic responses serve as the constraints for material parameters. The second prob-

lem (Case B) takes into account, in additional to the macroscopic constitutive responses, the local void ratio developed

at the peak and residual stages, thus adopting information from micro-CT images as additional target. The weight of

the void ratio data in Case B is intentionally set larger than the macroscopic data such that the microstructural evolu-

tion can be compatible. These two extreme cases are studied to separate the inﬂuence of either macroscopic behavior

or meso-scale behavior on the parameter calibration. In the third numerical experiments, we introduce a new mul-

tiscale objective function that takes account of both macroscopic and meso-data in a more balanced way. We then

reuse the calibrated material parameter sets from Case A and Case B as the initial guesses for the restarted material

parameter identiﬁcation procedure to study the sensitivity of the calibration procedure.

5.1 Case A: Results from Macroscopic Objective Function

The objective function for Case A is the same as Eq. (13), which only consists of macroscopic stress ratio and volu-

metric strain data. Both data types contain 30 data points, thus their weights are identical. Unlike elasto-plastic models

for granular materials, the micropolar hypoplastic constitive model does not separate elastic and plastic parameters.

Thus the material parameters are calibrated simultaneously, not in a stepwise manner (Ehlers and Scholz, 2007). The

initial guess of the material parameters and the calibrated results in Case A is presented in the Table 2. The macro-

scopic responses obtained by the parameter sets from initial guess, the 20th evaluation, the 50th evaluation, and the

ﬁnal calibration result are compared in Fig. 10. The evolution of the curves shows that the iterations converge to the

ﬁnal solution that minimizes the objective function. However, the local void ratio distribution does not converge to the

actual experiment data. This is shown in Fig. 11 where the experimental data of void ratio map in cross-section YZ

and the relative error map [deﬁned in Eq. (10)] computed from simulations are presented. Since the sample geometry

and void ratio distribution are not included in the objective function of Case A, the calibration procedure does not

correct the void ratio discrepancy with the initial guess and leads to a numerical solution that a dominant shear band

Volume 14, Issue 4, 2016

402 Wang et al.

TABLE 2: Calibration of material parameters of entire sample using Case A: only macroscopic responses; Case

A1: equal weights of stress ratio, volumetric strain and local void ratio data, starts from results of Case A; Case B:

macroscopic responses and local void ratio distributions; Case B1: equal weights of stress ratio, volumetric strain and

local void ratio data, starts from results of Case B

Number of C1C2C3C4eca l

iterations

Initial guess — −68.00 −767.60 −2742.70 −257.50 0.650 0.980 0.200 mm

Case A 74 −70.57 −832.40 −2524.10 −261.70 0.637 0.960 0.468 mm

Case A1 30 −69.67 −1372.67 −2075.62 −251.90 0.641 0.976 0.364 mm

Case B 117 −67.21 −920.08 −2312.79 −259.45 0.636 0.971 0.977 mm

Case B1 30 −67.91 −1229.29 −2244.82 −262.41 0.6358 0.970 0.979 mm

FIG. 10: Stress ratio and volumetric strain responses of full sample simulation during the calibration procedure using

only macroscopic responses (Case A)

is formed inside the sample, while the actual specimen developed a “barrel” shape that exhibit diffusive bands (Ikeda

et al., 2003).

In the correction step (Case A1), we modify the objective function used as previous calibrated material parameters

set as an initial guess. This modiﬁed objective function incorporates additional terms to constrain the material param-

eter set such that the numerical specimen also exhibits the same peak and residual shear strength due to the last two

terms in (14), which reads

f2(x) = 1

Nσ1/σ3

Nσ1/σ3

i=1

wiriσ1

σ3i

(x),σ1

σ3lab

i2

+1

Nev

Nev

j=1

wjrjevj(x), elab

vj2

+1

Nelement

Nelement

k=1

wkrkepeak

k(x), epeaklab

k2

+1

Nelement

Nelement

m=1

wmrmeresidual

m(x), eresiduallab

m2

,

(14)

International Journal for Multiscale Computational Engineering

Identifying Micropolar Material Parameters via Micro-CT Images 403

(a) Void ratio distribution from Micro-CT images

(b) Distribution of residuals [deﬁned in Eq. (10)] at selected steps of the Dakota calibration in Case A

FIG. 11: Relative error of local void ratio distribution between full sample simulation and micro-CT data (shown in

cross section in plane YZ) during Dakota calibration procedure using Case A: only macroscopic responses

where eiare void ratio in element i,Nelement is the number of elements in the FEM model. The contributions of stress

ratio, volumetric strain, local void ratio data are balanced by the number of data points of each type. The weights of

each data type equal to 1.

5.2 Case B: Results from Multiscale Objective Function

The objective function for Case B takes the form

Volume 14, Issue 4, 2016

404 Wang et al.

f2(x) =

Nσ1/σ3

i=1

riσ1

σ3i

(x),σ1

σ3lab

i2

+

Nev

i=1

rievi(x), elab

vi2

+

Nelement

i=1

riepeak

i(x), epeaklab

i2

+

Nelement

i=1

rieresidual

i(x), eresiduallab

i2

.

(15)

In this function, the residuals of all available experimental data point are treated equally in the objective function.

Since the number of elements is much larger than the number of stress ratio and volumetric strain data, the objective

function Eq. (15) has much larger weight on micro-CT data than the macroscopic responses. The calibration procedure

in Case B employs the same initial guess of the material parameters previously used in Case A. The calibrated results

are also shown in Table 2 to show comparisons. Compared to other hypoplastic material parameters, the material

length parameter l, which accounts for the micropolar effect in the model, varies more signiﬁcantly when experimental

data of local void ratio are included in the least square problem. The macroscopic responses obtained by the parameters

sets from initial guess, the 20th evaluation, the 50th evaluation, and the ﬁnal calibration result are compared in Fig. 12.

Although the volumetric strain response approaches the experiment data along the iterations, the macroscopic stress

ratio response deviates from the experimental response in the sense that the peak stress and residual stress do not

coincide and the softening phenomenon is not apparent. As for the meso-scale data shown in Fig. 13, the calibrated

parameters lead to a deformed conﬁguration much closer predication to actual specimen geometry than that of the

Case A results shown in Fig. 11. In particular, the Case B calibrated simulation correctly predicts the increased of

porosity at the middle of specimen, which is consistent to the observation in the laboratory. The Case A calibrated

simulation leads to porosity increase highly concentrated in a single persistent anti-symmetric shear band that was not

triggered in the actual experiment. This observation indicates the necessity of including micro-structural information

in material parameter identiﬁcation procedures. Note that the relative errors of void ratio near the top and bottom

surfaces, unlike the central areas, are not signiﬁcantly reduced during the Dakota calibrations. This is because the

boundary conditions in model simulation does not perfectly represent the experimental setup.

The objective function Eq. (14) is again adopted to perform the correcrtion step from the calibrated results

(Case B1). The correction is made by using the equilibrium weights of different types of experimental data. The

FIG. 12: Stress ratio and volumetric strain responses of full sample simulation during the calibration procedure using

macroscopic responses and local void ratio distribution (Case B)

International Journal for Multiscale Computational Engineering

Identifying Micropolar Material Parameters via Micro-CT Images 405

Distribution of residuals at selected steps of the Dakota calibration in Case B

FIG. 13: Relative error of local void ratio distribution between full sample simulation and micro-CT data (shown in

cross section in plane YZ) during Dakota calibration procedure using Case B: macroscopic responses and local void

ratio distribution

results are recapitulated in Table 2 and the macroscopic responses in different cases are compared in Fig. 14. Recall

that in the correction step Case A1, the prediction of peak stress has been improved. In this case, the discrepancy

between the model response and experimental data in Case B1 has not been improved or even changed signiﬁcantly

in the correction step, which indicates that both objective functions Eqs. (15) and (14) havesimilar local minimizers.

5.3 Discussion

Two calibration strategies are employed in this study. The resultant material parameters and calibrated simulations

are analyzed. In the ﬁrst strategy, we ﬁnd material parameter that allows the ﬁnite element simulations to replicate

the macroscopic responses as close as possible, but neglect all meso-scale information provided by the micro-CT

images. This optimized material parameter set (optimized in terms of macroscopic responses only), are then used as

the initial guess of the next inverse problem. Following this predictor step, another inverse problem is deﬁned by a

new multiscale objective function that takes account of both the macroscopic data and local void ratio properties used

for calibration. This approach mimics the idea in Ehlers and Scholz (2007) for determining material parameters for

micropolar constitutive laws. The major departure here is the usage of micro-CT image and the elimination of the need

to use multiple experimental tests to generate constraints for the objective function. To analyze the importance of the

initial guess and whether a global optimal value for the material parameter set exists, we employ another alternative

strategy in which the meso-scale information is used right at the predictor step. Then, the same multiscale objective

function used in the corrector step [i.e., Eq. (14)] is used to balance the weights of global and local data.

5.3.1 Comparisons of Results

At the ﬁrst look, the approach that starts with calibrating macroscopic parameter seems to be better in terms of replicat-

ing compatible shear stress history as shown in Fig. 14(a), even though both calibrated ﬁnite element simulations yield

Volume 14, Issue 4, 2016

406 Wang et al.

FIG. 14: Stress ratio and volumetric strain responses of full sample simulation during the calibration procedure using

Case A: only macroscopic responses; Case A1: equal weights of stress ratio, volumetric strain, and local void ratio

data, starts from results of Case A; Case B: macroscopic responses and local void ratio distribution; Case B1: equal

weights of stress ratio, volumetric strain, and local void ratio data, starts from results of Case B

similar volumetric responses. In particular, the second approach is unable to capture the macroscopic peak and resid-

ual shear stresses at the predictor step, and again fails to make any signiﬁcant improvement in capturing the peak and

residual shear strength after switching to the multiscale objective function, as shown in Fig. 14. As a result, evidence

provided in the macroscopic responses seems to favor the staggered approach similar to the one proposed in Ehlers

and Scholz (2007) in which the calibration process begins with an inverse problem that ﬁrst curve-ﬁt macroscopic

behaviors, followed by a correction step that uses multiscale objective function to enforce consistency of kinematics.

However, a closer look at the deformed conﬁguration and the meso-scale responses may lead to an opposite

conclusion. In particular, we ﬁnd that the weight modiﬁcation approach that begins with calibrating meso-scale in-

formation from micro-CT images actually yields the experimentally observed bifurcation mode, while the macro-

then-microscopic approach does not. In Case A, l= 0.468 mm, which is close to the mean particle diameter d50 =

0.338 mm, a persistent shear band has developed in the specimen. This persistent shear band is non-symmetric and has

not been observed in the micro-CT images captured during the drained triaxial compression test. In Case B, however,

l= 0.977 mm, which is 2.5 times larger than the mean particle diameter and the barrel-shaped deformed specimen

develops a barrel deformed conﬁguration followed by the development of an X-shaped shear localization zone. These

kinematic features are consistent with what is observed in the physical experiment, even though the Case B simulation

does not replicate the shear stress responses as close as the Case A simulation does.

In other words, the set of material parameters that leads to the best replica of macroscopic responses observed in

laboratory does not necessarily yield the correct bifurcation mode. As a result, the reasonable strategy is to design

an objective function that acts as a compromise between matching macroscopic responses and maintaining consistent

kinematics at meso-scale level. Furthermore, the notable difference in the macroscopic and microscopic responses

predicted by the staggered approach and weight modiﬁcation approach indicates that the calibration exercise is highly

path-dependent and multiple local minima are likely to exist, thus making it difﬁcult to ﬁnd the global minimum point

for the multiscale objective function in the parametric space.

International Journal for Multiscale Computational Engineering

Identifying Micropolar Material Parameters via Micro-CT Images 407

5.3.2 Source of Error

The discrepancy between the experimental data and simulation results may also be attributed to the idealized bound-

ary condition applied on numerical specimen. In particular, the specimen-loading-plate interaction between the sample

and the loading pistons as well as the elastic membrane are not accurately and explicitly modeled (Albert and Rud-

nicki, 2001). Moreover, the assumption that material parameters are homogeneous throughout the specimen may also

oversimplify the spatial variability of the physical specimens. Finally, the micropolar ﬁnite element model is formu-

lated using the Jaumann rate of the Cauchy stress. The Jaumann rate is an objective rate which is suitable for materials

in the geometrical nonlinear regime with small strain and large rotation. However, some previous works, such as

Molenkamp (1986), have pointed out that the Jaumann rate should be avoided for problems with large deviatoric

strain. These limitations will be considered in future studies.

5.3.3 Length Scale, Higher-Order Kinematics and Bifurcation Modes

The higher-order quantities, namely the curvature κand the coupled stress µ, can be computed from the micropolar

model simulation, while this information is not available from micro-CT images. In this study, two types of shear bands

are encountered with different material parameters in the ﬁnite element simulations even though both simulations

began with the same initial geometry and void ratio distribution. In Case A and B, the calibrated material length l

varies signiﬁcantly, compared to other material parameters.

In both cases, the distribution of the norm of the couple stress tensor |µ|=µ2

21 +µ2

32 +µ2

31 and the norm of the

curvature tensor |κ|=κ2

21 +κ2

32 +κ2

31 are shown in Figs. 15 and 16, respectively. The coupled stress and curvature

localize in the transition zone between the shear localization region and homogeneously deformed region, suggesting

that the micropolar effect becomes very important when high-gradient deformation occurs in granular materials. This

observation is consistent with numerical experiments conducted with discrete element method (Ehlers et al., 2003;

Oda and Iwashita, 2000; Wang and Sun, 2016a,b).

The evolution of shear bands simulated in Case A and Case B is illustrated in Fig. 17 using color map of |µ|, as

well as in Fig. 18 using color map of |κ|. At the beginning of the triaxial loading, two-axes symmetric localization

patterns emerge for both simulations (Ikeda et al., 2003). After the peak stress of Case A (at 3% axial strain), the

pattern bifurcates to an asymmetry pattern: one shear band becomes stronger than the other. Upon further loading, the

dominant shear band becomes persistent and the other weak band gradually dies out. As for Case B, the two bands

compete with each other along the deformation but neither predominates. The initial pattern bifurcates to bilateral

symmetry so that the symmetry with respect to the horizontal axis is broken. The diffuse mode is preserved to the end

(a) Case A (b) Case B

FIG. 15: Norm of the coupled stress tensor |µ|=µ2

21 +µ2

32 +µ2

31 for results of inverse problems A and B at the

residual stage

Volume 14, Issue 4, 2016

408 Wang et al.

(a) Case A (b) Case B

FIG. 16: Norm of the curvature tensor |κ|=κ2

21 +κ2

32 +κ2

31 for results of inverse problems A and B at residual

stage

FIG. 17: Evolution of shear localization in cross section YZ of calibraion results of Case A and Case B, illustrated in

the norm of coupled stress

of the loading. This analysis shows that, even though the spatial variability of the material parameters is neglected,

the numerical values of the material parameters, particularly the material length l, still impose strong effects on the

failure mode of numerical specimen.

International Journal for Multiscale Computational Engineering

Identifying Micropolar Material Parameters via Micro-CT Images 409

FIG. 18: Evolution of shear localization in cross section YZ of calibraion results of Case A and Case B, illustrated in

the norm of curvature tensor

5.3.4 Critical State

The anisotropy density function fdof Eq. (3) is an indicator of the critical state, a condition at which particulate

materials keep deforming in shear at constant void ratio and stress (Casagrande, 1936; Roscoe et al., 1958). In the

micropolar hypoplasticity adopted in this study, the local void ratio approaches the critical void ratio ec,fdapproaches

1. The distribution of fdof the two numerical specimens captured at residual stage of Cases A and B are presented in

Fig. 19. Since the different calibrated material parameters are used in Case A and B, the failure modes as well as the

locations where the Hostun sand ﬁrst reaches the critical state are also different. In Case A, the elements residing in

the persistent shear band are close to the critical state, while the elements outside the zone are not close to it. In Case

B, the pattern of fdalso coincides with the diffuse failure mode presented in Fig. 17, showing that the unit cells inside

the shear band are also closer to critical state than the host matrix, but the difference between the numerical value

of fdwithin the host matrix and inside the shear band are less signiﬁcant than Case A. Both ﬁndings are consistent

with Tejchman and Niemunis (2006); Wang and Sun (2016b) where the strain localization triggered by the material

bifurcation in dense assemblies tends to have the void ratio approaching its critical value locally, but the specimen

itself does not necessarily reach critical state globally. Comparing this observation with the coupled stress norm shown

in Fig. 17, we observed that the shear band is not only much closer to the critical state, but also has signiﬁcantly lower

coupled stress magnitude. This result is consistency with the norm of the curvature tensor shown in Fig. 18 where

the specimen has signiﬁcant amount of micro-rotation at the boundary of the shear band and the host matrix but the

micro-polar kinematics is not signiﬁcant inside the shear band and in the host matrix.

6. CONCLUSIONS

In this work, we incorporate information obtained from both macroscopic measurement and meso-scale kinematics to

analyze the sensitivity of the predicted characteristic length and mechanical responses of a 3D micropolar hypoplastic-

ity ﬁnite element model. To the best knowledge of the authors, this is the ﬁrst contribution that incorporates micro-CT

Volume 14, Issue 4, 2016

410 Wang et al.

FIG. 19: Distribution of fdfor results of inverse problems A and B at the residual stage. fd= 1 indicates that the

material is in critical state.

images and multiscale objective function into the material parameter identiﬁcation procedure for micropolar plasticity

for granular materials.

The results show that the incorporation of meso-scale information may signiﬁcantly change the predicted length

scale obtained from the inverse problems and leads to different bifurcation modes in the ﬁnite element simulations.

Even though similar macroscopic responses are observed from simulations conducted with single-scale and multi-

scale objective functions, the macroscopic responses in the former case may yield meso-scale responses that are not

consistent with those of the real specimen. As a result, the apparently good curve-ﬁtting of macroscopic response

is not a good indicator of forward prediction capacity. This result has important implications for the validations of

grain-scale simulation tools (such as discrete element, lattice-beam, and lattice spring models) in which macroscopic

stress–strain responses are often the only experimental data available for benchmark and validations. The numerical

results, particularly the difference of the failure modes obtained from different objective functions, indicate that using

macroscopic stress–strain curve alone to evaluate or validate grain-scale simulations is neither productive nor reliable.

Although the higher-order quantities, the curvature κand the coupled stress µ, are not available from micro-CT

images, they are computed from the micropolar simulations with the material parameters optimized for different ob-

jective functions. Depending on which set of material parameters is employed, the numerical specimen may either

develop a persistent shear band, which is not observed in the experiment, or a diffuse failure mode, which is observed

from micro-CT images. Comparisons between the simulation results with micro-CT images suggest that a staggered

predictor–corrector procedure that ﬁrst employs the macroscopic objective function to curve-ﬁt the macroscopic re-

sponses, then use the multiscale objective function to enforce kinematic constraints seem to yield a more compatible

macroscopic constitutive response with the experimental counterpart, butthe material parameters that lead to the best

curve-ﬁtting macroscopic responses also lead to an incorrect bifurcation mode. This ﬁnding is alerting, as the material

parameters that lead to the wrong bifurcation mode in the backward calibration exercise is also likely to generate

even more unrealistic forward prediction. The apparently good match in the macroscopic curve can be misleading and

generate a false sense of conﬁdence for the numerical model. This is a noteworthy concern, as there is an alerting

trend in which the forward-predictive capacity of grain-scale models are often incorrectly measured by how well they

curve-ﬁt the macroscopic stress–strain curve, rather than how well they are able to generate compatible and consistent

mechanical behaviors across length scales. The issue associated with this calibration approach is not apparent when

calibration is conducted at the unit cell level in which only homogeneous deformation is considered. However, when

macroscopic stress–strain curve is used to calibrate meso-scale or grain-scale models, the dimensions of the paramet-

ric space can be larger than the number of constraints provided by the macroscopic responses. The insufﬁciency of

constraints then makes it possible to generate simulations that apparently match the macroscopic calibration with a

International Journal for Multiscale Computational Engineering

Identifying Micropolar Material Parameters via Micro-CT Images 411

completely inconsistent microstructure. By incorporating microscopic information from micro-CT images to calibrate

material parameters, this research provides important evidence to suggest that constraining micro-mechanical model

to match macroscopic responses is not sufﬁcient. Nor is it a meaningful way to measure the quality of numerical

predictions. These lessons are important for the calibration and validation of high-order and multiscale ﬁnite element

models.

ACKNOWLEDGMENTS

This research is supported by the Earth Materials and Processes program at the US Army Research Ofﬁce under

grant contract W911NF-14-1-0658 and W911NF-15-1-0581, the Mechanics of Material program at National Science

Foundation under grant contracts CMMI-1462760 and EAR-1516300, and Provost’s Grants Program for Junior Fac-

ulty who Contribute to the Diversity Goals of the University at Columbia University. These supports are gratefully

acknowledged.

REFERENCES

Adams, B. M., Bohnhoff, W. J., Dalbey, K. R., Eddy, J. P., Eldred, M. S., Gay, D. M., Haskell, K., Hough, P. D., and Swiler, L. P.,

Dakota, a multilevel parallel object-oriented framework for design optimization, parameter estimation, uncertainty quantiﬁca-

tion, and sensitivity analysis: Version 5.0 users manual, Sandia National Laboratories, Tech. Rep. SAND2010-2183, 2009.

Albert, R. A. and Rudnicki, J. W., Finite element simulations of tennessee marble under plane strain laboratory testing: Effects of

sample–platen friction on shear band onset, Mech. Mater., vol. 33, no. 1, pp. 47–60, 2001.

Amat, A. S., Elastic Stiffness Moduli of Hostun Sand, PhD thesis, Universitat Polit`

ecnica de Catalunya, Escola, T`

ecnica Superior

d’Enginyers de Camins, Canals i Ports de Barcelona, Departament d’Enginyeria del Terreny, Cartogr`

aﬁca i Geof´

ısica, (Enginy-

eria Geol`

ogica), 2008.

And`

o, E., Hall, S. A., Viggiani, G., Desrues, J., and B´

esuelle, P., Grain-scale experimental investigation of localised deformation

in, sand: a discrete particle tracking approach, Acta Geotech., vol. 7, no. 1, pp. 1–13, 2012.

Bazant, Z. P., Belytschko, T. B., and Chang, T.-P., Continuum theory for strain-softening, J. Eng. Mech., vol. 110, no. 12, pp.

1666–1692, 1984.

Been, K. and Jefferies, M. G., A state parameter for sands, G´

eotechnique, vol. 35, no. 2, pp. 99–112, 1985.

Belytschko, T., Fish, J., and Engelmann, B. E., A ﬁnite element with embedded localization zones, Comput. Methods Appl. Mech.

Eng., vol. 70, no. 1, pp. 59–89, 1988.

Borja, R. I., A ﬁnite element model for strain localization analysis of strongly discontinuous ﬁelds based on standard galerkin

approximation, Comput. Methods Appl. Mech. Eng., vol. 190, no. 11, pp. 1529–1549, 2000.

Borja, R. I. and Sun, W. C., Estimating inelastic sediment deformation from local site response simulations, Acta Geotech., vol. 2,

no. 3, pp. 183–195, 2007.

Borja, R. I., Song, X., Rechenmacher, A. L., Abedi, S., and Wu, W., Shear band in sand with spatially varying density, J. Mech.

Phys. Solids, vol. 61, no. 1, pp. 219–234, 2013.

Casagrande, A., Characteristics of cohesionless soils affecting the stability of slopes and earth ﬁlls, Contrib. Soils Mech., 1925-

1940, vol. 23, no. 1, pp. 13–32, 1936.

Cooreman, S., Lecompte, D., Sol, H., Vantomme, J., and Debruyne, D., Elasto-plastic material parameter identiﬁcation by inverse

methods: Calculation of the sensitivity matrix, Int. J. Solids Struct., vol. 44, no. 13, pp. 4329–4341, 2007.

Cosserat, E. and Cosserat, F., Th´

eorie des Corps D´

eformables, Paris, 1909.

Dafalias, Y. F. and Manzari, M. T., Simple plasticity sand model accounting for fabric change effects, J. Eng. Mech., vol. 130, no.

6, pp. 622–634, 2004.

Desrues, J. and And`

o, E., Strain localisation in granular media, Comptes Rendus Physique, vol. 16, no. 1, pp. 26–36, 2015.

Desrues, J. and Viggiani, G., Strain localization in sand: An overview of the experimental results obtained in Grenoble using

stereophotogrammetry, Int. J. Numer. Anal. Methods Geomech., vol. 28, no. 4, pp. 279–321, 2004.

Duhem, P., Le potentiel thermodynamique et la pression hydrostatique, In Annales Scientiﬁques de l’ ´

Ecole Normale Sup´

erieure,

vol. 10, pp. 183–230, 1893.

Volume 14, Issue 4, 2016

412 Wang et al.

Ehlers, W., Ramm, E., Diebels, S., and d’Addetta, G. A., From particle ensembles to cosserat continua: Homogenization of contact

forces towards stresses and couple stresses, Int. J. Solids Struct., vol. 40, no. 24, pp. 6681–6702, 2003.

Ehlers, W. and Scholz, B., An inverse algorithm for the identiﬁcation and the sensitivity analysis of the parameters governing

micropolar elasto-plastic granular material, Arch. Appl. Mech., vol. 77,no. 12, pp. 911–931, 2007.

Eringen, A. C., Theory of micropolar elasticity, In Microcontinuum Field Theories, pp. 101–248, Springer, Berlin, 1999.

Fish, J., Jiang, T., and Yuan, Z., A staggered nonlocal multiscale model for a heterogeneous medium, Int. J. Numer. Methods Eng.,

vol. 91, no. 2, pp. 142–157, 2012.

Flavigny, E., Desrues, J., and Palayer, B., Note technique: Le sable d’Hostun RF, Revue. Fran. C G´

eotech., vol. 53, pp. 67–70,

1990.

Gajo, A., Bigoni, D., and Wood, D. M., Multiple shear band development and related instabilities in granular materials, J. Mech.

Phys. Solids, vol. 52, no. 12, pp. 2683–2724, 2004.

Hall, S. A., Bornert, M., Desrues, J., Pannier, Y., Lenoir, N., Viggiani, G., and B´

esuelle, P., Discrete and continuum analysis

of localised deformation in sand using x-ray µct and volumetric digital image correlation, Geotechnique, vol. 60, no. 5, pp.

315–322, 2010.

Herle, I. and Gudehus, G., Determination of parameters of a hypoplastic constitutive model from properties of grain assemblies,

Mech. Cohesive-Frictional Mater., vol. 4, no. 5, pp. 461–486, 1999.

Hu, N. and Fish, J., Enhanced ant colony optimization for multiscale problems, Comput. Mech., vol. 57, no. 3, pp. 447–463, 2015.

Hughes, T. J. R., The Finite Element Method: Linear Static and Dynamic Finite Element Analysis, Courier Corporation, 2012.

Ikeda, K., Yamakawa, Y., and Tsutsumi, S., Simulation and interpretation of diffuse mode bifurcation of elastoplastic solids, J.

Mech. Phys. Solids, vol. 51, no. 9, pp. 1649–1673, 2003.

Issen, K. A. and Rudnicki, J. W., Conditions for compaction bands in porous rock, J. Geophys. Res.: Solid Earth (1978–2012), vol.

105, no. B9, pp. 21529–21536, 2000.

Kuhn, M. R., Sun, W., and Wang, Q., Stress-induced anisotropy in granular materials: Fabric, stiffness, and permeability, Acta

Geotech., vol. 10, no. 4, pp. 399–419, 2015.

Lade, P. V. and Duncan, J. M., Elastoplastic stress-strain theory for cohesionless soil, J. Geotech. Eng. Div., vol. 101, no. 10, pp.

1037–1053, 1975.

Lasry, D. and Belytschko, T., Localization limiters in transient problems, Int. J. Solids Struct., vol. 24, no. 6, pp. 581–597, 1988.

Lenoir, N., Andrade, J. E., Sun, W. C., and Rudnicki, J. W., In situ permeability measurements inside compaction bands using x-ray

CT and lattice Boltzmann calculations, Adv. Comput. Tomography Geomater.: GeoX2010, pp. 279–286, 2010.

Li, X. S. and Dafalias, Y. F., Anisotropic critical state theory: role of fabric, J. Eng. Mech., vol. 138, no. 3, pp. 263–275, 2011.

Lin, J. and Wu, W., A comparative study between DEM and micropolar hypoplasticity, Powder Technol., vol. 293, pp. 121–129,

2015.

Lin, J., Wu, W., and Borja, R. I., Micropolar hypoplasticity for persistent shear band in heterogeneous granular materials, Comput.

Methods Appl. Mech. Eng., vol. 289, pp. 24–43, 2015.

Liu, Y., Sun, W., Yuan, Z., and Fish, J., A nonlocal multiscale discrete-continuum model for predicting mechanical behavior of

granular materials, Int. J. Numer. Methods Eng., vol. vol. 206, no. 2, pp. 129–160, 2015.

Liu, Y., Sun, W., and Fish, J., Determining material parameters for critical state plasticity models based on multilevel extended

digital database, J. Appl. Mech., vol. 88, no. 1, pp. 011003-1–16, 2016.

Mahnken, R. and Stein, E., A uniﬁed approach for parameter identiﬁcation of inelastic material models in the frame of the ﬁnite

element method, Comput. Methods Appl. Mech. Eng., vol. 136, no. 3, pp. 225–258, 1996.

Manzari, M. T. and Dafalias, Y. F., A critical state two-surface plasticity model for sands, Geotechnique, vol. 47, no. 2, pp. 255–272,

1997.

Maugin, G. A. and Metrikine, A. V., Mechanics of generalized continua, Adv. Mech. Math., vol. 21, 2010.

Mindlin, R. D., Micro-structure in linear elasticity, Arch. Rational Mech. Anal., vol. 16, no. 1, pp. 51–78, 1964.

Molenkamp, F., Limits to the jaumann stress rate, Int. J. Numer.Anal. Methods Geomech., vol. 10, no. 2, pp. 151–176, 1986.

Na, S. and Sun, W., Wave propagation and strain localization in a fully saturated softening porous medium under the non-isothermal

conditions, Int. J. Numer. Anal. Methods Geomech., 2016.

International Journal for Multiscale Computational Engineering

Identifying Micropolar Material Parameters via Micro-CT Images 413

Needleman, A., Material rate dependence and mesh sensitivity in localization problems, Comput. Methods Appl. Mech. Eng., vol.

67, no. 1, pp. 69–85, 1988.

Oda, M. and Iwashita, K., Study on couple stress and shear band development in granular media based on numerical simulation

analyses, Int. J. Eng. Sci., vol. 38, no. 15, pp. 1713–1740, 2000.

Ortiz, M., Leroy, Y., and Needleman, A., A ﬁnite element method for localized failure analysis, Comput. Methods Appl. Mech.

Eng., vol. 61, no. 2, pp. 189–214, 1987.

O’Sullivan, C., Particulate Discrete Element Modelling, Taylor & Francis, 2011.

Pestana, J. M. and Whittle, A. J., Formulation of a uniﬁed constitutive model for clays and sands, Int. J. Numer. Anal. Methods

Geomech., vol. 23, no. 12, pp. 1215–1243, 1999.

Rechenmacher, A. L., Grain-scale processes governing shear band initiation and evolution in sands, J. Mech. Phys. Solids, vol. 54,

no. 1, pp. 22–45, 2006.

Roscoe, K. H., Schoﬁeld, A. N., and Wroth, C. P., On the yielding of soils, Geotechnique, vol. 8, no. 1, pp. 22–53, 1958.

Rudnicki, J. W. and Rice, J. R., Conditions for the localization of deformation in pressure-sensitive dilatant materials, J. Mech.

Phys. Solids, vol. 23, no. 6, pp. 371–394, 1975.

Sadek, T., Lings, M., Dihoru, L., and Wood, D. M., Wave transmission in hostun sand: Multiaxial experiments, Rivista Italiana

Geotecnica, vol. 41, no. 2, pp. 69–84, 2007.

Satake, M., New formulation of graph-theoretical approach in the mechanics of granular materials, Mech. Mater., vol. 16, no. 1,

pp. 65–72, 1993.

Schoﬁeld, A. and Wroth, P., Critical State Soil Mechanics, McGraw-Hill, p. 310, 1968.

Song, J.-H., Areias, P. M. A., and Belytschko, T., A method for dynamic crack and shear band propagation with phantom nodes,

Int. J. Numer. Methods Eng., vol. 67, no. 6, pp. 868–893, 2006.

Subhash, G., Nemat-Nasser, S., Mehrabadi, M. M., and Shodj, H. M., Experimental investigation of fabric-stress relations in

granular materials, Mech. Mater., vol. 11, no. 2, pp. 87–106, 1991.

Sun, W., A uniﬁed method to predict diffuse and localized instabilities in sands, Geomech. Geoeng., vol. 8, no. 2, pp. 65–75, 2013.

Sun, W. and Mota, A., A multiscale overlapped coupling formulation for large-deformation strain localization, Comput. Mech.,

vol. 54, no. 3, pp. 803–820, 2014.

Sun, W., Andrade, J. E., Rudnicki, J. W. and Eichhubl, P., Connecting microstructural attributes and permeability from 3D tomo-

graphic images of in situ shear-enhanced compaction bands using multiscale computations, Geophys. Res. Lett., vol. 38, no. 10,

2011.

Sun, W., Kuhn, M. R., and Rudnicki, J. W., A multiscale DEM-LBM analysis on permeability evolutions inside a dilatant shear

band, Acta Geotech., vol. 8, no. 5, pp. 465–480, 2013.

Tejchman, J. and Niemunis, A., Fe-studies on shear localization in an anistropic micro-polar hypoplastic granular material, Gran-

ular Matter, vol. 8, no. 3-4, pp. 205–220, 2006.

Walker, D. M., Tordesillas, A., and Kuhn, M. R., Spatial connectivity of force chains in a simple shear 3D simulation exhibiting

shear bands, J. Eng. Mech., p. C4016009, 2016.

Wang, K. and Sun, W., Anisotropy of a tensorial bishops coefﬁcient for wetted granular materials, J. Eng. Mech., p. B4015004,

2015.

Wang, K. and Sun, W., A semi-implicit micropolar discrete-to-continuum method for granular materials, In VII European Congress

on Computational Methods in Applied Science and Engineering, 2016a.

Wang, K. and Sun, W., A semi-implicit discrete-continuum coupling method for porous media based on the effective stress principle

at ﬁnite strain, Comput. Methods Appl. Mech. Eng., vol. 304, pp. 546–583, 2016b.

Wu, W., Bauer, E., and Kolymbas, D., Hypoplastic constitutive model with critical state for granular materials, Mech. Mater., vol.

23, no. 1, pp. 45–69, 1996.

Yang, Q., Mota, A., and Ortiz, M., A class of variational strain-localization ﬁnite elements, Int. J. Numer. Methods Eng., vol. 62,

no. 8, pp. 1013–1037, 2005.

Volume 14, Issue 4, 2016