Lecture Notes in Applied and Computational Mechanics 81

Kerstin Weinberg
Anna Pandolfi Editors

Innovative Numerical
Approaches for Multi-
Field and Multi-Scale
Problems
In Honor of Michael Ortiz’s 60th
Birthday
Lecture Notes in Applied and Computational
Mechanics

Volume 81

Series editors
Friedrich Pfeiffer, Technische Universität München, Garching, Germany
e-mail: [email protected]
Peter Wriggers, Universität Hannover, Hannover, Germany
e-mail: [email protected]
About this Series

This series aims to report new developments in applied and computational

mechanics—quickly, informally and at a high level. This includes the fields of fluid,
solid and structural mechanics, dynamics and control, and related disciplines. The
applied methods can be of analytical, numerical and computational nature.

More information about this series at http://www.springer.com/series/4623

Kerstin Weinberg Anna Pandolfi
•

Editors

Innovative Numerical
Approaches for Multi-Field
and Multi-Scale Problems
In Honor of Michael Ortiz’s 60th Birthday

123
Editors
Kerstin Weinberg Anna Pandolfi
Universität Siegen Politecnico di Milano
Siegen Milan
Germany Italy

ISSN 1613-7736 ISSN 1860-0816 (electronic)

Lecture Notes in Applied and Computational Mechanics
ISBN 978-3-319-39021-5 ISBN 978-3-319-39022-2 (eBook)
DOI 10.1007/978-3-319-39022-2

Library of Congress Control Number: 2016940007

© Springer International Publishing Switzerland 2016

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made.

Printed on acid-free paper

This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG Switzerland
Preface

During the past two decades research in the field of computational mechanics has
progressed remarkably, mainly because of the development of a sound mathe-
matical background and the introduction of new efficient computational strategies.
Beyond the classical finite element method, several innovative techniques and novel
approaches for the analysis of microstructural evolution, growth, damage, and
structural failure in multi-field and multi-scale problems have emerged vigorously.
With the aim to discuss different computational strategies for multi-field and
multi-scale problems, a remarkable group of scientists gathered in September 2014
to the IUTAM symposium “Innovative numerical approaches for materials and
structures in multi-field and multi-scale problems”. Hosted by the University of
Siegen, the venue of the symposium was the Castle Burg Schnellenberg, a mighty
fortress located in the green heart of Westphalia, Germany. There we discussed the
new horizons and perspectives of multi-field applied mechanics. The symposium
covered a large domain of recent research, from computational materials modeling,
crystal plasticity, micro-structured materials, and biomaterials to multi-scale sim-
ulations of multi-physics phenomena. The pioneering discretization methods for the
solution of coupled nonlinear problems at different length scales were particularly
emphasized.
The special occasion that motivated the organization of the symposium was the
60th birthday of Professor Michael Ortiz. Along his exceptional career, Michael
Ortiz has been at the forefront of computational mechanics, his work being a
constant source of inspiration for many. All participants of this symposium are
grateful to Michael Ortiz for being such an enthusiastic collaborator, a reliable
colleague, an illuminating scientist, and a valuable friend.
The friendship and fellow-feeling felt during the symposium inspired the idea to
collect the presentations of some of the convened researchers in a special book. Our
choice was to organize a book as part of the series ‘Lecture Notes in Applied and
Computational Mechanics’ (LNACM), which aims to document new high-level
developments in applied and computational mechanics. We are happy to present
here 13 high-quality contributions of current and past collaborators of Michael

v
vi Preface

Ortiz. All contributions have undergone full peer review, and we take the occasion
to thank the reviewers for their valuable comments.
It is our hope that the present volume will give the reader an insight into the
exciting new developments of computational solid mechanics which is still wide
open to discovery. The book attempts to provide a flavor of this challenging field
and to contribute to its popularity within the mechanics and physics communities.

Siegen Kerstin Weinberg

February 2016 Anna Pandolfi
Contents

Robust Numerical Schemes for an Efficient Implementation

of Tangent Matrices: Application to Hyperelasticity,
Inelastic Standard Dissipative Materials and
Thermo-Mechanics at Finite Strains . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Masato Tanaka, Daniel Balzani and Jörg Schröder
Folding Patterns in Partially Delaminated Thin Films . . . . . . . . . . . . . 25
David Bourne, Sergio Conti and Stefan Müller
Thermo-mechanical Behavior of Confined Granular Systems . . . . . . . . 41
Gülşad Küçük, Marcial Gonzalez and Alberto M. Cuitiño
Elastomeric Gels: A Model and First Results . . . . . . . . . . . . . . . . . . . . 59
Mariarita de Luca and Antonio DeSimone
A Tensegrity Paradigm for Minimal Mass Design of Roofs and
Bridges. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Gerardo Carpentieri, Fernando Fraternali and Robert E. Skelton
Universal Meshes for the Simulation of Brittle Fracture
and Moving Boundary Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Maurizio M. Chiaramonte, Evan S. Gawlik, Hardik Kabaria
and Adrian J. Lew
Free Energy, Free Entropy, and a Gradient Structure for
Thermoplasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
Alexander Mielke
Comparison of Isotropic Elasto-Plastic Models for the
Plastic Metric Tensor Cp ¼ FTp Fp . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Patrizio Neff and Ionel-Dumitrel Ghiba

vii
viii Contents

Quasi-Static Evolutions in Brittle Fracture Generated by Gradient

Flows: Sharp Crack and Phase-Field Approaches. . . . . . . . . . . . . . . . . 197
Matteo Negri
Improving the Material-Point Method . . . . . . . . . . . . . . . . . . . . . . . . . 217
Deborah Sulsky and Ming Gong
Meshfree Methods Applied to Consolidation Problems in Saturated
Soils . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
Pedro Navas, Susana López-Querol, Rena C. Yu and Bo Li
A Multiscale Microstructural Model of Permeability in Fractured
Solids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
Anna Pandolfi, Maria Laura De Bellis and Gabriele Della Vecchia
Thermal Diffusion in a Polymer Blend . . . . . . . . . . . . . . . . . . . . . . . . . 285
Kerstin Weinberg, Stefan Schuß and Denis Anders
Robust Numerical Schemes for an Efficient
Implementation of Tangent Matrices:
Application to Hyperelasticity, Inelastic
Standard Dissipative Materials and
Thermo-Mechanics at Finite Strains

Masato Tanaka, Daniel Balzani and Jörg Schröder

Abstract In this contribution robust numerical schemes for an efficient implemen-

tation of tangent matrices in finite strain problems are presented and their perfor-
mance is investigated through the analysis of hyperelastic materials, inelastic stan-
dard dissipative materials in the context of incremental variational formulations, and
thermo-mechanics. The schemes are based on highly accurate and robust numeri-
cal differentiation approaches which use non-real numbers, i.e., complex variables
and hyper-dual numbers. The main advantage of these approaches are that, contrary
to the classical finite difference scheme, no round-off errors in the perturbations
due to floating-point arithmetics exist within the calculation of the tangent matrices.
This results in a method which is independent of perturbation values (in case of
complex step derivative approximations if sufficiently small perturbations are cho-
sen). An efficient algorithmic treatment is presented which enables a straightforward
implementation of the method in any standard finite-element program. By means
of hyperelastic, finite strain elastoplastic, and thermo-elastoplastic boundary value
problems, the performance of the proposed approaches is analyzed.

1 Introduction

Many materials utilized in industry are characterized by micro-heterogeneous prop-

erties and a number of direct micro-macro transition methods have been proposed in
order to estimate the nonlinear stress-strain relationship at every macroscopic point

M. Tanaka (B)
Toyota Central R&D Labs., Inc., Nagakute, Japan
e-mail: [email protected]
D. Balzani
Institute of Mechanics and Shell Structures, TU Dresden, Dresden, Germany
e-mail: [email protected]
J. Schröder
Institute of Mechanics, University Duisburg-Essen, Essen, Germany
e-mail: [email protected]
© Springer International Publishing Switzerland 2016 1
K. Weinberg and A. Pandolfi (eds.), Innovative Numerical Approaches
for Multi-Field and Multi-Scale Problems, Lecture Notes in Applied
and Computational Mechanics 81, DOI 10.1007/978-3-319-39022-2_1
2 M. Tanaka et al.

by detailed modeling of the microstructure, i.e., the representative volume element

(RVE), see [2, 8, 19, 25, 30]. Especially when using the implicit finite element
method, the constitutive equations and consistent tangent moduli need to be explic-
itly specified and precisely implemented in computer programs. Whereas the exact
calculation of the stresses determines the accuracy of the numerical simulation, the
consistent algorithmic tangent moduli are required to achieve quadratic convergence
in a Newton iteration scheme as well as to detect material instabilities in localization
analysis, cf. e.g., [26]. However, for some complex material models their analytic
derivatives can be extremely elaborate and error-prone to be implemented. In those
cases numerical differentiation may be a useful alternative in particular to decrease
scientific development time, provided that robust methods are available.
Different robust numerical approaches have been proposed in the literature based
on either complex-step derivative approximation (CSDA) or hyper-dual-step deriv-
atives (HDSD). The CSDA scheme goes back to [14] and has been used since many
years in a different context than followed here, e.g. for the calculation of sensitivities
in structural optimization problems [16]. The concept of hyper-dual numbers has
been introduced in Fike [7] and also used for structural optimization in [6]. For inter-
ested readers, C++ or Matlab implementation packages can be downloaded from the
web site by J.A. Fike (http://adl.stanford.edu/hyperdual/#Code), see also a source
code example for the implementation of HDNs using operator overloading in Fortran
90/95 in [32]. In a different context, i.e. for the implementation of material models in
solid mechanics at finite strains, numerical methods based on such robust derivative
approximations have been proposed, cf. [3, 9–11, 23, 24, 31, 32].
These approaches mainly extend Miehe’s implementation technique [17] based
on finite differences (FD) to obtain more robust approximations of tangent matrices,
either for material tangent moduli or for tangent stiffness matrices in finite element
formulations. In this contribution we review some numerical schemes which are
suitable for the implementation of finite strain constitutive models, and present a new
framework for incremental variational formulations based on hyper-dual numbers.
It is organized as follows. Section 2 recapitulates some numerical differentiation
techniques, i.e., FD, CSDA and HDSD. Section 3 shows a formulation to numerically
compute tangent moduli from stress equations using the CSDA scheme and Sect. 4
shows a formulation to automatically compute both, stress and tangent moduli from
a strain energy function of hyperelastic material models using the HDSD scheme.
Section 5 extends the HDSD-based implementation method to an inelastic standard
dissipative material model by applying it to an incremental variational formulation.
In Sect. 6 the numerical calculation of tangent stiffness matrices in the context of
nonlinear thermo-mechanical finite element problems is shown.

2 Numerical Differentiation Techniques

For a summary of the individual numerical differentiation methods simple scalar-

valued functions of one single scalar quantity are considered. Given the function
f (x), most methods are based on the Taylor series expansion
Robust Numerical Schemes for an Efficient Implementation … 3

h 2  h 3 
f (x + h) = f (x) + h f  (x) + f (x) + f (x) + · · · . (1)
2! 3!
The often-used classical FD schemes consider a real-valued small perturbation h and
neglect the higher order terms O(h 2 ) and thus end up in the approximation

 f (x + h) − f (x)
f FD (x) ≈ , (2)
h

wherein, O(h 2 ) denotes Landau’s symbol to describe the asymptotic behavior of

the higher-order terms. This approach, however, is highly sensitive with respect to
the values of the perturbation h and of reduced accuracy even in the optimal range
of values which is typically small. For small h the explicit calculation of f (x + h)
leads to roundoff errors and for large h it is not acceptable to disregard higher order
terms in (1). To overcome this difficulty of the FD schemes, Lyness [14] devised
the CSDA scheme. In order to avoid the direct addition of perturbations along the
same (real) axis, the CSDA method uses perturbations along the imaginary axis of
complex numbers by replacing h by ih in (1) and by taking the imaginary part one
obtains the approximation

 [ f (x + ih)]
f CSDA (x) ≈ . (3)
h
 denotes the operation of taking the imaginary part of complex functions. This
expression provides a high accuracy for very small values of h being remarkably
close to the analytic ones. This is due to the fact that no roundoff errors occur in
the application of the perturbation itself and technically perturbation values up to
e.g. h = 10−99 are possible. The main drawback of this approach is that higher-
order derivatives can not be directly computed and combinations of FD and CSDA
have to be considered, see [13]. This however still suffers from the problems of the
FD method. Automatic differentiation (AD) techniques, which are typically derived
based on repeated application of the chain rule of differentiation, see e.g. [15], are
alternatives to compute highly accurate derivatives. However, these schemes are dif-
ficult to be derived and formulated, in particular for tensorial derivatives. An equiva-
lent with a forward type AD, but more practicable approach can be accomplished by
using dual numbers, which have been originally introduced by Clifford [5]. Fike [6]
developed a higher dimensional version of the dual numbers, the hyper-dual num-
bers (HDNs), which are mainly characterized by the consideration of more than one
imaginary axis. The HDNs of second-order have two non-real unit numbers ε1 and
ε2 with the properties

ε12 := 0, ε22 := 0, ε1 ε2 := ε2 ε1 , (4)

which implies that each of the numbers squared is 0 and products of them are com-
mutative. In order to better understand calculations with hyper-dual numbers we
4 M. Tanaka et al.

shortly recapitulate some basic characteristics and operations. Given the two HDNs
a = (a1 + a2 ε1 + a3 ε2 + a4 ε1 ε2 ) and b = (b1 + b2 ε1 + b3 ε2 + b4 ε1 ε2 ), a and b are
equal if and only if all of their real and non-real parts are equal, i.e., a1 = b1 , a2 = b2 ,
a3 = b3 , a4 = b4 . We define the symbols ε1 , ε2 and ε1 ε2 which denote the opera-
tions of taking the imaginary parts of HDNs such that ε1 [a] = a2 , ε2 [a] = a3 and
ε1 ε2 [a] = a4 . As proposed by Fike [6] the HDSD for a scalar function f := f (x)
of a single scalar-valued argument x is obtained by inserting hε1 + hε2 into h in (1)
with exactly vanishing higher order terms O(h 3 ). The first derivative f  (x) can then
be obtained by taking the coefficient of ε1 or ε2 , i.e.

 ε1 [ f (x + hε1 + hε2 )] ε [ f (x + hε1 + hε2 )]

f HDSD (x) = = 2 . (5)
h h

The second derivative f  (x) can be obtained by taking the coefficient, i.e.

 ε1 ε2 [ f (x + hε1 + hε2 )]

f HDSD (x) = . (6)
h2
Note that since expressions (5) and (6) do not have any subtraction operation the
roundoff errors do not arise. Furthermore, since O(h 3 ) is exactly 0, the expressions
do not have any truncation errors, too. This means that the value of h could be
completely arbitrary and h = 1 could be considered for simpler formulae.

3 Approximation of Material Tangents in Hyperelasticity

In this section, different methods for the numerical approximation of the derivatives
of stress tensors with respect to deformation tensors, known as tangent moduli, are
compared. For this purpose numerical differentiation schemes pointed out in the
previous section are extended to directional derivatives of tensor fields.

3.1 Numerical Approximation of Directional Derivatives

Let A be an arbitrary second-order tensor on R2 or R3 . Then, the directional derivative

of a second-order tensor function Z(X) of a second-order tensor argument X in
direction A is given by

Z(X + h A) − Z(X) ∂Z
DZ(X)[ A] := lim = : A. (7)
h→0 h ∂X

Herein, ∂ Z/∂ X is a fourth-order tensor that implies a gradient of Z with respect to

X. This expression directly yields the formula for the finite difference (FD) approach
by skipping the limit and treating h as the perturbation value:
Robust Numerical Schemes for an Efficient Implementation … 5

∂Z Z(X + h A) − Z(X)
: A≈ , (8)
∂X h
Second-order derivatives can be directly calculated by applying this formula also to
the second derivatives and using first-order FD for the individual first-order deriva-
tive function evaluations. By replacing the FD term by the complex step derivative
approximation, one obtains

∂Z [Z(X + ih A)]
: A≈ . (9)
∂X h
Unfortunately, the direct application of CSDA also to the second derivative yields
an expression which is not free from associated round-off errors and thus it may
also combined with the FD for the second derivative. The only approach which
enables second-order derivative approximation of high accuracy independent of the
choice of the perturbation values is based on hyper dual numbers. Such second-order
derivatives may be important in the context of hyperelasticity when considering the
strain energy function whose first derivative yields the stress and the second derivative
the tangent moduli. Thus, first and second directional derivatives of a scalar function
with respect to deformation tensors are considered now. Let A, B be arbitrary second-
order tensors on R2 or R3 . Then, the directional derivative of a scalar function z(X)
of a second-order tensor argument X in direction A is given by

z(X + h A) − z(X) ∂z
Dz(X)[ A] := lim = : A, (10)
h→0 h ∂X

with ∂z/∂ X denoting a second-order tensor. The second order directional derivative
of a scalar function z(X) of a second-order argument X in the two directions A, B
is
z(X + h A + k B) − z(X + h A) − z(X + k B) + z(X)
D2 z(X)[ A, B] := lim
h,k→0 hk
∂ z
2
= A: : B, (11)
∂ X∂ X

wherein ∂ 2 z/∂ X∂ X is a fourth-order tensor that implies a Hessian of z with respect

to X. This results in the formulae for the HDSD scheme
∂z ε [z(X + hε1 A)]
: A= 1
∂X h
∂2z ε1 ε2 [z(X + hε1 A + kε2 B)]
A: :B= , (12)
∂ X∂ X hk
6 M. Tanaka et al.

with small perturbation values h, k. Provided that appropriate energy, stress, strain
and directional tensors are substituted in z, Z, X and A, respectively, the required
tangent moduli can be automatically derived as shown in the following.

3.2 Application to Hyperelastic Constitutive Equations

Here the representations in oblique-angled coordinate systems are provided in order

to highlight the mappings required for a consistent perturbation calculation. However,
for the examples we focus on Cartesian coordinates and then we identify G I = E I ,
G I = E I , g i = ei , g i = ei . Herein, G I and G I denote the co- and contravariant general
base vectors in the reference configuration and g i and g i in the actual configuration;
the Cartesian base vectors in the reference and in the actual configuration are E I =
E I and ei = ei . Note that italic and non-italic indices represent components and
“name” indices, respectively. For an abbreviated representation here we describe
the constitutive equations in a material setting. Then the existence of a strain energy
function ψ := ψ(C) or ψ := ψ(E) with the right Cauchy-Green deformation tensor
C and the Green-Lagrange strain tensor E is postulated. The constitutive relation
for the second Piola-Kirchhoff stress tensor S then reads
∂ψ ∂ψ
S= =2 . (13)
∂E ∂C

The material tangent moduli C are defined as the derivative of S with respect to E:

∂S ∂ 2ψ ∂S ∂ 2ψ
C= = , or C = 2 =4 . (14)
∂E ∂ E∂ E ∂C ∂ C∂ C
∗ ∗
KL KL
Now, substitute S in Z, C in X, and C in A in (8) and (9), where C is defined
as
∗ 1 K
C KL
= (G ⊗ G L + G L ⊗ G K ). (15)
2
Then, one obtains for the finite difference and the complex-step-derivative approxi-
mation scheme
∗
S I J (C + h C KL
) − S I J (C)
CFD
I J KL
= , (16)
 h 
∗
 S I J (C + ih C KL
)
CCSDA
I J KL
=2 . (17)
h
Robust Numerical Schemes for an Efficient Implementation … 7

As mentioned above, the function evaluations for the second Piola-Kirchhoff stresses
could be replaced by numerical approximations in terms of the FD or the CSDA
approach. However neither of them would lead to a method without round-off errors.
This can be achieved by using hyper dual numbers. For this purpose consider Eq. (12)
∗ ∗ ∗ ∗
and substitute z by ψ, X by C, A by C IJ
and B by C KL
, where C IJ
and C KL
are
defined in (15), then one obtains
 
∗
ε1 ψ(C + hε1 C )
IJ
IJ
SHDSD =2 , (18)
 h 
∗ ∗
ε1 ε2 ψ(C + hε1 C IJ
+ kε2 C KL
)
CHDSD
IJKL
=4 . (19)
hk

3.3 Numerical Examples

The numerical calculation of stresses and tangent moduli of a hyperelastic material

model is analyzed. For that purpose the FD, CSDA and HDSD scheme are compared
with the implementation of the analytic stresses and moduli. In this study one of
the anisotropic polyconvex models proposed in [1] for fiber-reinforced materials is
considered where an isotropic part describes the ground substance and an anisotropic
part represents the embedded fibers as
nf 
  β2 
−1/3
ψ = α1 (I1 I3 − 3) + α2 (I3α3 + I3−α3 − 2) + β1 I1 J4(a) − J5(a) − 2 .
a=1
(20)

Herein, α1 > 0, α2 > 0, α3 > 0, β1 > 0 and β2 > 2 are material parameters, n f is
the number of fiber families and • denotes the Macaulay bracket. The invariants
are given by I1 = tr[C], I3 = det[C], J4(a) = tr[C M (a) ] and J5(a) = tr[C 2 M (a) ] with
the structural tensor M (a) = a0(a) ⊗ a0(a) , where a0(a) denotes the fiber direction.

3.3.1 Homogeneous Problem

As a first investigation, a homogeneous test is considered where a specific deforma-

tion gradient F is applied. This deformation gradient includes rotations Q as well
as deformations F 0 and is given by

F = Q F 0 with Q = Rθ1 Rθ2 Rθ3 . (21)

8 M. Tanaka et al.

Herein, Rθ1 , Rθ2 , Rθ3 are the individual rotation tensors and the deformations F 0
contain dilatation and shear deformation. They are chosen as
⎡ ⎤ ⎡ ⎤
cos π4 − sin π4 0 cos π3 0 sin π3
Rθ1 = ⎣ sin π4 cos π4 0 ⎦ , Rθ2 = ⎣ 0 1 0 ⎦,
0 0 1 − sin π3 cos π3
⎡ ⎤ ⎡ ⎤
1 0 0 1.1 γ γ
Rθ3 = ⎣ 0 cos π6 − sin π6 ⎦ , F 0 = ⎣ 0 0.9535 γ ⎦ , (22)
0 sin π6 cos π6 0 0 0.9535

γ is the amplitude of the shear deformation. The material parameters are chosen
as α1 = 1.0, α2 = 1.0, α3 = 0.1, β1 = 1.0 and β2 = 3.0. Two preferred directions
T
are considered, i.e. n f = 2, which are defined as a0(1) = 1/3 1 2 2 and a0(2) =
√ T
1/ 5 2 1 0 . In this example we focus on the numerical performance of computing
the tangent moduli by using different numerical schemes, i.e., the FD approach (16)
and the CSDA scheme (17) starting from the stress tensor, and the HDSD scheme (19)
starting from the strain energy function (20). We compare their results with the ones
of an analytically derived tangent modulus (see [31, 32] for the detailed expressions
of stress and tangent moduli of the model (20)). In order to compare the results the
relative error eC is defined as
 21  21
 
IJKL 2 IJKL 2
eC = (Canalyt ) I J K L − (Capprox ) / (Canalyt ) ,
I,J,K ,L I,J,K ,L
(23)

where (Canalyt ) I J K L denotes the coefficients of the analytic material tangent modulus
tensor in the Cartesian coordinate system, (Capprox ) I J K L denotes the values of the
approximated counterparts. This relative error eC is depicted for each numerical
approximation scheme in Fig. 1. As can be seen, the FD approach shows a quite
sensitive behavior having its optimal accuracy of eC ≈ 10−7 at perturbation value of

Fig. 1 Relative errors eC of 10 0

approximated tangent
moduli by using the FD, the 10 -4
Relative error e C

CSDA and the HDSD

scheme. Perturbation values 10 -8
FD
are varied from 1.0 × 10−20 CSDA
to 1.0 × 10−1 . In the case of
-12
10
HDSD
the HDSD, we fix h = 1.0
and instead vary the 10 -16
perturbation k in Eq. (19)
10 -20 -20
10 10-16 10-12 10-8 10-4
Perturbation
Robust Numerical Schemes for an Efficient Implementation … 9

≈ 10−7 . For increasing and decreasing perturbation values the error increases. The
CSDA approach already reaches an accuracy of eC ≈ 10−7 for perturbations smaller
than 10−4 and reaches an error at computer accuracy for perturbation values smaller
than approximately 10−9 . The HDSD is independent on the perturbation value and
always yields an error of eC ≈ 10−16 which is computer accuracy.

3.3.2 Cook-Type Problem

In order to analyze the influence of the numerical differentiation schemes onto

the convergence behavior of Newton-Raphson iterations required for the solu-
tion of nonlinear boundary value problems, the formulations described above are
implemented in the general-purpose finite-element program FEAP, developed by
R.L. Taylor (http://www.ce.berkeley.edu/projects/feap/). In this example, we ana-
lyze a Cook-type cantilever beam which is schematically illustrated in Fig. 2a.
Now, only one fiber family (n f = 1) is considered which is oriented in direction
√ T
a(1) = 1/ 3 1 1 1 , in order to induce some torsion around the x-axis accompa-
nied by some out of x-y-bending leading to a pronounced combined bending/torsion
deformation. The set of material parameters is chosen as α1 = 6.0, α2 = 100.0,
α3 = 5.0, β1 = 100.0 and β2 = 2.5. The maximum load p0 is increased in equidis-
tant load steps until the ultimate maximum load p0 = 5.0 is reached. In this example,
we compare different versions of numerical approximations. Starting from the strain
energy function (20) we apply the HDSD scheme (18), (19), in order to directly
compute both, stresses and moduli. Then, the FD scheme is applied twice to also

(a) p0 (b)

y 16

44
44

48 [mm]

Fig. 2 a Geometry with boundary conditions. The beam is fixed in all directions at x = 0 (dashed
area) and a distributed load is applied at the opposite side in vertical direction leading to a bending-
dominated deformation. b Deformed configuration with distribution of Kirchhoff stresses τ 11 in
x-direction
10 M. Tanaka et al.

compute stresses and moduli numerically based on the strain energy function. As
a third option we implement the analytic expression for the first derivative of the
strain energy and compute then numerically the tangent moduli using the CSDA
scheme (17). The Kirchhoff stress distributions τxx in the most deformed configu-
ration are depicted in Fig. 2b. For any numerical differentiation scheme, the stress
distribution is very similar to the analytic implementation of stress and moduli. The
Euclidean norm of the residual vector of the discretized weak form of equilibrium for
the calculations based on the analytic tangent, the FD and HDSD approach is plotted
in Table 1, where h S = h C = 1 is considered for the HDSD. Even the values of the
residuals are almost identical for the HDSD method compared with the ones resulting
from an analytic implementation of stress and moduli. The values for the FD method
differ by orders of magnitudes and thus, no quadratic convergence is observed. Next,
the computing time of the different approaches to calculate the stresses and tangent
moduli is compared in Fig. 3. As expected, the computing time of the FD method
depends on the perturbation values since a different number of iterations results
therefrom. When considering the best performing FD scheme with the perturbation
values h S = 10−4 and h S = 10−6 , the lowest computing time is obtained for coarse
meshes; the HDSD however, is only slightly slower. This behavior changes for an
increasing number of elements and at a reasonable discretization the HDSD becomes
even faster than the FD method with optimal values. This is due to the fact that with
an increasing number of elements the amount of time required for the solution of the
system of equations becomes dominant compared to the calculation of the stiffness
matrix, which is influenced by the numerical approximation of stress and moduli.
Then, for finer meshes the number of iterations and thus the advantage of the HDSD
and CSDA method becomes more important. It is emphasized that the optimal choice
of perturbation values is of course not known in general and thus, the FD method

Table 1 Euclidean norm of the residual of the Cook-type problem based on the FD and HDSD
scheme; red colors indicate that no quadratic convergence could be obtained
FD with hS = 10−4
Analytical HDSD
−4
hC = 10 hC = 10−6 hC = 10−8
5.4504 × 10−1 6.3167 × 10+1 6.3167 × 10+1 6.3167 × 10+1 5.4504 × 10−1
1.8568 × 10+2 9.3803 × 10+1 9.0099 × 10+1 9.0116 × 10+1 1.8568 × 10+2
1.5341 × 10+0 2.0612 × 10+0 2.4092 × 10+0 2.4096 × 10+0 1.5341 × 10+0
3.4458 × 10+0 3.2625 × 10+0 3.2628 × 10+0 3.2630 × 10+0 3.4458 × 10+0
8.4932 × 10−3 6.5687 × 10−3 1.4598 × 10−2 1.4692 × 10−2 8.4932 × 10−3
2.0186 × 10−4 4.2600 × 10−4 5.3153 × 10−4 5.7587 × 10−4 2.0187 × 10−4
5.0665 × 10−8 2.3998 × 10−5 2.0946 × 10−6 5.5211 × 10−6 5.0161 × 10−8
− 2.3392 × 10−6 3.3563 × 10−7 3.3765 × 10−7 −
− 4.2348 × 10−7 − − −
− 3.3564 × 10−7 − − −
Robust Numerical Schemes for an Efficient Implementation … 11

Fig. 3 Comparison of CPU

time for Cook-type problem, 1st, 2nd : HDSD
for the FD scheme the 1st: analytic, 2nd : CSDA

Normalized CPU time

perturbation value 1st, 2nd : FD with hC =10-6
1st, 2nd : FD with hC =10-4
h S = 10−4 is considered, for
the HDSD scheme we use
h S = h C = 1 and for the
CSDA scheme we consider
h = 10−30 . The CPU time is
normalized with respect to
the time required for the
calculation resulting from
the implementation of
analytic stresses and moduli 5000 10000
Number of elements

can typically be expected to be slower since not the optimal perturbation values will
be used. This is shown by the computing time resulting from the values h S = 10−4 ,
h C = 10−4 in Fig. 3 leading to a computing time being significantly higher as for
the HDSD and CSDA. Moreover, the HDSD method is still faster than the CSDA
scheme, although the first and second derivatives are calculated numerically in the
case of HDSD, while the CSDA scheme requires the analytic expressions for the
stresses.

4 HDSD Scheme for Incremental Variational Formulations

In this section a framework for the fully automatic calculation of internal vari-
ables, stresses and consistent tangent moduli for dissipative materials is proposed.
It mainly consists of applying the HDSD scheme to incremental variational formu-
lations (IVFs). Such formulations, see e.g. [18, 22], recast inelasticity theory as an
equivalent optimization problem where the incremental stress potential within a dis-
crete time interval is minimized in order to obtain the values of internal variables.
The IVFs provide a general framework for a broad range of standard dissipative
constitutive models in incremental stress potential regimes. Throughout this section,
isothermal conditions are considered.

4.1 Incremental Variational Formulation

Ortiz and Stainier [22] proposed an effective incremental stress potential W eff within
a discrete time interval Δt := tn+1 − tn as

W eff := inf W (F n+1 , q n+1 ). (24)

q n+1
12 M. Tanaka et al.

The deformation gradient tensor and the generalized vector of a collection of internal
variables at time tn+1 are denoted by F n+1 and q n+1 , respectively. The incremental
stress potential W comprises a Helmholtz free energy function ψ and a dissipation
potential φ by

Δq
W (F n+1 , q n+1 ) = ψ(F n+1 , q n+1 ) − ψ(F n , q n ) + Δtφ(q n+1 , ); (25)
Δt
Δq := q n+1 − q n . In many cases the minimization problem (24) is solved by using
a Newton-Raphson iteration and q n+1 is determined using the first and second deriv-
atives ∂q W , ∂qq
2
W by updating

−1
q (k+1) (k)
n+1 = q n+1 − ∂qq W
2 (k)
· ∂q W (k) , k = 0, 1, 2, . . . , (26)

until it converges; (k) denotes the iteration number. The stress response P n+1 at time
tn+1 is given via the effective incremental potential function W eff as

P n+1 = ∂ F W eff (F n+1 , q n+1 (F n+1 )), (27)

and the consistent tangent modulus An+1 at time tn+1 is given by

An+1 = ∂ F2 F W eff (F n+1 , q n+1 (F n+1 )), (28)

where P is the first Piola-Kirchhoff stress tensor and A is the corresponding nominal
tangent modulus tensor. Note that q n+1 is dependent on the current deformation
gradient F n+1 .

4.2 Implementation Using HDSD Scheme

At first, find a solution q n+1 to the minimization problem (24) by using the HDNs
ε1 and ε2 . Here, an asterisk (∗) is used as a superscript denoting the perturbed values
(k)
by means of the HDNs. The internal variables at k-th iteration q n+1 are perturbed by
using the HDN units ε1 and ε2 as
∗ (k) (k)
q n+1 = q n+1 + ε1 i r + ε2 i s , (29)

wherein i r denotes a unit vector of a set of internal variables q n+1 such that the j-th
component of i r is defined as i r j = 1 if j = r and 0 else. Then, the r -th components
of ∂q W (k) and accordingly of ∂qq 2
W (k) are obtained by taking the coefficients with
respect to ε1 and ε1 ε2 such as
∗  ∗ 
∂q W (k) r
= ε1 W (k)
, ∂qq
2
W (k) rs
= ε1 ε2 W (k)
, (30)
Robust Numerical Schemes for an Efficient Implementation … 13

and q n+1 is updated by Eq. (26). As a second step, the stresses P and tangent moduli
A are obtained by differentiating the minimized effective stress potential W eff with
respect to the deformation gradient F n+1 as

P n+1 = ∂ F W eff + ∂q W eff · ∂ F q n+1 , (31)

An+1 = ∂ F2 F W eff + ∂ Fq
2
W eff · ∂ F q n+1
+ ∂ F q n+1 · ∂q2 F W eff + ∂qq2
W eff · ∂ F q n+1 + ∂q W eff · ∂ F2 F q n+1 . (32)

Note that Eqs. (31) and (32) are indeterminate since ∂ F q n+1 and ∂ F2 F q n+1 cannot be
computed using only 2nd-order HDNs. In order to avoid the calculations of ∂ F q n+1
and ∂ F2 F q n+1 , we assume the strict stationary condition

∂q W eff = 0, (33)

and also

D F (∂q W eff ) = 0 or equivalently ∂q2 F W eff + ∂qq

2
W eff · ∂ F q n+1 = 0. (34)

Inserting (33) and (34) into (31) and (32), the stresses and tangent moduli in this
formulation are obtained by

P n+1 = ∂ F W eff , (35)

−1
An+1 = ∂ F2 F W eff − ∂ Fq
2
W eff · ∂qq
2
W eff · ∂q2 F W eff . (36)

The derivatives ∂ F W eff , ∂ F2 F W eff , ∂ Fq

2
W eff and ∂q2 F W eff can be computed using only
ε1 and ε2 in an analogous way as for the derivatives with respect to the internal
variables. Note that the main advantage of this proposed framework is that the users
are only required to implement the scalar energy functions ψ and φ.

4.3 Numerical Example: Elastoplastic Microstructure

In this numerical example, we analyze a finite strain elastoplastic model using a

von Mises yield function including exponential isotropic hardening, cf. [4, 20], see
also [27, 29]. The deformation gradient is multiplicatively decomposed into elastic
and isochoric plastic parts as F = F e · F p with det F p = 1 and the Helmholtz free
energy is additively decomposed into elastic and plastic parts as ψ = ψ e + ψ p . In
this model, the elastic response is captured by

λ e 2  
ψe = b1 + b2e + b3e + μ (b1e )2 + (b2e )2 + (b3e )2 . (37)
2
14 M. Tanaka et al.

λ and μ are material parameters and beA (A = 1, 2, 3) is the logarithm of each

eigenvalue λeA of the elastic left Cauchy-Green deformation tensor be = F e F eT as
beA = log(λeA ). The plastic response ψ p is captured by the exponential hardening

1 1
ψ p = y∞ α − (y0 − y∞ ) exp(−ηα) + hα 2 , (38)
η 2

with α being a strain-like isotropic hardening variable and the material parameters
y∞ , y0 , η and h representing the initial yield strength, the plastic yield strength at the
transition from exponential to linear hardening, the degree of exponential hardening,
and the slope of superimposed linear hardening, respectively. In our formulation, we
identify the internal variable q as
 T
q = Fp, α , (39)

wherein the square brackets denote an appropriate arrangement as column matrix.

The generalized internal force vector y dual to q is determined as
 T
∂ψ ∂ψ e ∂ψ p  T
y := − = − p, − = Σ F p−T , β , (40)
∂q ∂F ∂α

with Σ denoting the Mandel stress tensor and β denoting the thermodynamic conju-
gate force to α. The dissipation potential φ is obtained by the maximum-dissipation
principle as
 
φ= sup Σ n+1 : ΔL p + βn+1 · Δα , (41)
(Σ n+1 ,βn+1 )∈E

Herein, ΔL p is a constant plastic velocity gradient in the current time increment and
Δα := αn+1 − αn is a constant increment of hardening variable. The elastic domain
E is defined by a yield function f as
   

E := (Σ, β)  f (Σ, β) = devΣ − 2/3β ≤ 0 ; (42)

dev(•) is the deviatoric operator. Through the use of Karush-Kuhn-Tucker conditions,

the inf-sup problem (24) and (41) can be recast by one parameter minimization
problem using the Lagrange multiplier Δγ [21] as
 
W eff = inf ψ(F n+1 , Δγ ) − ψ(F n ) + Δγ f with Δγ f = 0, f ≤ 0,(43)
Δγ ≥0

with the updated variables as

 
p ∂f ∂f
F n+1 = exp Δγ · F pn and αn+1 = αn + Δγ . (44)
∂Σ ∂β
Robust Numerical Schemes for an Efficient Implementation … 15

23 [MPa]
(a) (b) (c)

300
275
l 250
225
200

l l

Fig. 4 a Initial configuration of SSRVE (the total number of elements is 10,889 and the number of
degrees of freedom is 43,578), b distribution of Kirchhoff stresses τyz in the deformed configuration
resulting from the classical standard return mapping method and c IVF with HDSD scheme

Table 2 Material parameters for individual phases of DP steel microstructure

λ (MPa) μ (MPa) y0 (MPa) y∞ (MPa) η (–) h (MPa)
Matrix 118846.2 79230.77 260.0 580.0 9.0 70.0
(Ferrite)
Inclusion 118846.2 79230.77 1000.0 2750.0 35.0 10.0
(Martensite)

The performance of the proposed implementation scheme for finite strain plas-
ticity is investigated by the finite element simulation of a statistically similar repre-
sentative volume element (SSRVE) of a dual-phase (DP) steel microstructure. The
SSRVE shown in Fig. 4a is obtained by constructing a simplified microstructure in
terms of statistical measures as similar as possible to the real random microstruc-
ture, cf. [2]. To somehow represent a typical DP steel microstructure consisting of
a ferritic matrix phase with an embedded martensitic inclusion phase the material
parameters in Table 2 are used. A macroscopic deformation gradient F̄ including the
shear F̄yz is applied by prescribing the homogeneous deformation field x = F̄ X + w̃
to the SSRVE and periodic deformation fluctuations w̃ and anti-periodic tractions

Fig. 5 Macroscopic 240

Macroscopic stress −τ23 [MPa]

stress-strain diagram for

simple shear deformation of
statistically similar RVE of 180
DP steel microstructure. The
diagram compares the results 120
of the HDSD scheme with Standard return-mapping
the results of the classical IVF with HDNs
standard return mapping 60
scheme

0
0 0.01 0.02 0.03 0.04
-
Macroscopic deformation gradient F23
16 M. Tanaka et al.

at the boundary of the SSRVE. The resulting macroscopic stress-strain response

is calculated as shown in Fig. 5. Figure 4b, c show deformed configurations of the
SSRVE with stress distributions τ yz as a result of two different schemes, i.e., the
proposed HDSD based implementation scheme and the classical standard return
mapping scheme following the implementation given in [12]. According to those
results, a good agreement is observed at both, the microscopic as well as the macro-
scopic scale. However, it is worth pointing out that the proposed scheme has the
important advantage of a straightforward implementation of any other complicated
constitutive model in the context of incremental variational formulations since users
are only required to modify scalar-valued quantities such as the functions ψ, φ and
the yield function f .

5 Robust Approach to Compute Tangent Stiffness Matrix

in Thermo-Mechanical Problems Based on CSDA

In this section a CSDA-based robust approximation scheme for the numerical calcu-
lation of tangent stiffness matrices is presented in the context of nonlinear thermo-
mechanical finite element problems and its performance is analyzed.

5.1 Formulations

The thermo-mechanical framework at large strains relies on the governing equations,

namely the balance of linear momentum, and the balance of energy, i.e.

− DivF S − f = 0, (45)
1
S · Ċ + ρ0 r − Divq 0 − ρ0 (ψ̇ + θ˙η) = 0. (46)
2
Herein, the Legendre transform ψ = U − θ η has been performed, where ψ, U , θ
and η denote the Helmholtz free energy, the specific internal energy, temperature
and the specific entropy. The internal dissipation is considered to consist of two
additive parts, i.e. Dint = Dmech + Dtherm , with the thermal part Dtherm = ρ0 θ η̇p and
a mechanical part Dmech . q 0 is the heat flux through the body in the reference con-
figuration, which is related to the Cauchy heat flux q = −kθ gradθ by q 0 = J F −1 q.
Herein, kθ is the heat conduction coefficient and J is the determinant of F. Note
that grad(•) denotes the gradient with respect to coordinates in the reference config-
uration, respectively, and Div(•) denotes the divergence with respect to coordinates
in the reference configuration. Also, f , r and ρ0 are the body force vector, internal
heat source and the reference density of the body, respectively. For the solution of
boundary value problems the standard Galerkin method is typically applied which
requires the weak form of the balance equations. In this context see e.g. [33] or [29]
for further details. Multiplying the balance Eqs. (45), (46) with test functions and
Robust Numerical Schemes for an Efficient Implementation … 17

integrating over the physical domain leads to the weak forms G u := G int
u − Gu = 0
ext

and G θ = G θ − G θ = 0, wherein the internal and external parts of the weak forms
int ext

are given by

1
u :=
G int S · δC dV, (47)
B0e 2

θ :=
G int ( q 0 · Gradδθ + ρ0 θ ∂θθ ψ θ̇ δθ + ρ0 θ ∂θα ψ α̇ δθ
2 2
B0e
e
+ ρ0 θ ∂θ2be ψ · ḃ δθ + Dmech δθ ) dV, (48)
 
G ext
u := t 0 · δu dA + f · δu dV, (49)
∂B0e B0e

G ext
θ := Q · Nδθ dA. (50)
∂B0e

Herein, B0e is the domain of the reference configuration, δu and δθ denote the test
functions associated with the displacement and temperature fields, respectively. Also
note that t 0 and Q are the surface traction and the heat flux vectors, respectively,
which are assumed here to be independent of the displacements or temperatures.
Exploiting the principle of maximum dissipation and applying the Karush-Kuhn-
Tucker optimality conditions, c.f. [28], theexplicit form for the mechanical dissipa-
tion for isotropic hardening is Dmech = λ 23 y(θ ), with the consistency parameter
λ and the initial yield stress y(θ ). Here, for the finite element implementation the
discretized weak forms of the balance equations using matrix notation read


n ele
int,h

n ele

u ≈ Gu
G int int,h
= (δd eu )T r int,e
u and G int
θ ≈ G θ = (δd eθ )T r θint,e , (51)
e=1 e=1

with n ele denoting the number of finite elements. The external parts of the discretized
weak forms G uext,h and G ext,h
θ are reformulated accordingly. The element vector of
mechanical degrees of freedom is denoted by d eu which can be represented for a
three-dimensional setting by d eu = [(d uI =1 )T (d uI =2 )T . . . (d nen
u ) ] with nen being
T T

the number of nodes per element and d u = [dux duy duz ] denoting the displacements
I T

at a particular node I .
In a three-dimensional physical space the number of degrees of freedom associ-
ated with the displacements is nd f u = 3, i.e. dux , du y , and duz . Then the total number
of mechanical degrees of freedom for one element is tdo f u = nen × nd f u . Analo-
gously, the number of thermal degrees of freedom is nd f θ = 1, i.e. dθ , such that the
total number of thermal degrees of freedom per element is tdo f θ = nen × nd f θ .
After inserting standard approximations for the nodal displacements and temper-
atures and shifting to the Voigt notation to represent tensorial quantities of second
order as vectors, the discretized internal element residual vectors are
18 M. Tanaka et al.

nen 

r e,int
u = (B uI )T S dV, (52)
I =1 B0e
nen 

r e,int
θ = (B θI )T q 0 + N I ρ0 θ ∂θθ
2
ψ θ̇ + N I ρ0 θ ∂θθ
2
ψ α̇
I =1 B0e
e 
+ N I ρ0 θ ∂θ2be ψ · ḃ + N I λ 2/3y(θ )) dV. (53)

Herein, B0e denotes the domain of the initial configuration of finite element e. The
external residual vectors r e,extu and r e,ext
θ are obtained analogously. Here, the same
nodal shape functions N are used for the displacements and the temperature; B uI are
I

the standard mechanical B-matrices associated with node I , c.f. [33], which consist
of the gradients of the nodal shape functions. The thermal B-matrix is given by B θI =
[N,1I N,2I N,3I ]T , where N,iI represents the derivative of the shape function with respect
to physical coordinate X i . Due to the material nonlinearities these equations have to
be solved numerically. For that purpose their linearizations LinG hu = G hu + ΔG hu and
LinG hθ = G hθ + ΔG hθ are required. Here ΔG hu and ΔG hθ are the linear increments of
the weak forms obtained as


n ele
ΔG int,h
u = (δd eu )T keuu Δd eu + keu θ Δd eθ , (54)
e=1

n ele
ΔG int,h
θ = (δd eθ )T keθu Δd eu + keθθ Δd eθ , (55)
e=1

with the individual derivatives of the residual vectors given by

∂ r e,int ∂ r e,int ∂ r e,int

θ ∂ r e,int
θ
keuu = u
, k e
= u
, k e
θu = , k e
θθ = . (56)
∂ d eu uθ
∂ d eθ ∂ d eu ∂ d eθ

Instead of implementing the analytic expressions for these derivatives the CSDA
scheme can be applied. Remember that the mechanical and thermal residual vectors
depend on all (mechanical and thermal) degrees of freedom, i.e. r eu := r eu (d eu , d eθ )
e
and r eθ := r eθ (d eu , d eθ ). Then the approximations of the k-th column vectors k̃uu(k) and
e e e
k̃θu(k) in keuu and keθu , respectively, and of the j-th column vectors k̃uθ( j) and k̃θθ( j)
in keuθ and keθθ , respectively, can alternatively be approximated by
 e

e ∂ r e,int  r eu (d eu + i h d̃ u(k) , d eθ )
k̃uu(k) := u  ≈ , (57)
∂ due k h
 e

e ∂ r e,int  r eu (d eu , d eθ + i h d̃ θ( j) )
k̃uθ( j) :=  u e  ≈ , (58)
∂ dθ j h
Robust Numerical Schemes for an Efficient Implementation … 19

 e

e ∂ r e,int  r eθ (d eu + i h d̃ u(k) , d eθ )
k̃θu(k) := θ  ≈ , (59)
∂ due k h
 e

e ∂ r e,int  r eθ (d eu , d eθ + i h d̃ θ( j) )
k̃θθ( j) := θ  ≈ , (60)
∂ dθe j h

where the indices k ∈ [1, tdo f u ] and j ∈ [1, tdo f θ ] on the left hand side of the
equations represent the column index. On the right hand side these indices correspond
to the individual perturbation vectors whose components with indices m ∈ [1, tdo f u ]
and q ∈ [1, tdo f θ ], respectively, are defined as

{d̃u(k)
e
}m = δ(k)m and {d̃θ(
e
j) }q = δ( j)q . (61)

The Kronecker symbol is defined as δab = 1 for a = b and δab = 0 otherwise.

5.2 Numerical Example: Thermo-Elastoplastic

Microstructure

Now the performance of the proposed approximation scheme is investigated in a

thermo-elastoplastic problem. A simplified RVE of DP steel is considered which
consists of a spherical inclusion embedded in a cubic matrix, cf. Fig. 6a. The dis-
placements at all surfaces are linked in normal direction such that the outer surfaces
remain planar. The surfaces at X = 0, Y = 0 and Z = 0 are fixed in the respec-
tive normal direction and a compressive normal stress is applied to the top surface
along the Z -direction. First, compression is increased up to 650 MPa and then it is

(a) (b)
(M) (T) (V)
-650 MPa 939 K 3%
Loads

0 1 2 3 4
Ferrite
Time (in s)
Martensite

Fig. 6 a RVE, discretized with 3623 quadratic tetrahedral elements, and b loading protocol; M, T
and V indicate the mechanical compressive load, the thermal change and the volumetric expansion
(of the inclusion), respectively
20 M. Tanaka et al.

Table 3 Material parameters of the phases

E ν αt (1/K) c k H ω y0
(MPa) (–) (mm2 /s2 K) (NK/s) (MPa) (MPa/K) (MPa)
Ferrite 206,000 0.3 1 × 10−5 0.46 × 10−9 49 5000 −0.295 260
Martensite 206,000 0.3 1 × 10−5 0.46 × 10−9 43 25000 −0.586 1000

unloaded again. Second, the temperature is prescribed over the entire microstruc-
ture, increasing first from 293 K up to 939 K and then decreasing back to 293 K,
to reflect some heat treatment. Third, during the cooling procedure additionally a
volumetric change is applied in the inclusion phase to characterize the mechanical
fields resulting from a 3 % phase transformation volume change. This loading pro-
tocol is depicted in Fig. 6b. It is remarked that this procedure is only an idealization
of the production process of DP steel and serves here only as a numerical exam-
ple showing the applicability of the approximation scheme. For the calculation the
strain energy function as proposed in [29] is used along with a linear hardening law.
The material parameters chosen for the two phases are listed in Table 3, wherein
the hardening modulus is denoted by H . For the initial yield stress the temperature
dependency y = ω(θ − θ0 ) + y0 − ỹ0 + ỹ0 is taken into account, where y0 is the
initial yield stress at room temperature θ0 = 293 K. The parameters ω and y0 for the
two phases are also given in Table 3. The resulting accumulation of plastic strains
in the microstructure is visualized in Fig. 7. At time t = 1 s, after the application of
compression, we see a development of plastic strains in the ferrite, which is however,
lower than at time t = 4 s. To demonstrate the performance of the CSDA scheme
the convergence patterns are analyzed at particular times of interest during the entire
simulation. Thus, the times t = 1 s where the total compressive load is active and
t = 4 s at the end of the volume jump and cooling process are considered. The norms
of the residual vectors versus the corresponding Newton iteration, obtained at each of
these times, are plotted in Fig. 8 for the perturbation values of h = 10−5 , h = 10−8 ,

Fig. 7 Values of internal variable α in the microstructure at (a) t = 1 s, after the application of
compressive loads, and at (b) t = 4 s, after the application of the volumetric expansion
Robust Numerical Schemes for an Efficient Implementation … 21

(a) (b)
100 10000
log(h) = -05 log(h) = -05

Absolute residual norm

log(h) = -08 log(h) = -08
Absolute residual norm

1 log(h) = -10 100 log(h) = -10

log(h) = -30 log(h) = -30
0.01 1

0.0001 0.01

1e-06 0.0001

1e-08 1e-06

1e-10 1e-08

1e-12 1e-10
1 10 1 10

Number of newton iterations Number of newton iterations

Fig. 8 Absolute values of the residual norms versus corresponding Newton iteration at times
(a) t = 1 s, (b) t = 4 s. Both axes are depicted in logarithmic scale

h = 10−10 and h = 10−30 . As can be seen, the perturbation h = 10−5 leads to too
inaccurate approximations and therefore gives unsatisfactory results, but smaller per-
turbation values lead to quadratically converging Newton iterations. Summarizing, a
stable convergence of the CSDA approach for the tangent stiffness matrix is demon-
strated in this example for a thermo-elastoplastic problem with a variety of loading
conditions.

6 Conclusion

This contribution presented robust numerical schemes for an efficient implemen-

tation of tangent matrices in finite strain problems. The schemes were based on
highly accurate numerical differentiation approaches which use non-real numbers,
i.e., complex-step derivative approximation and hyper-dual-step derivatives. Their
excellent performance was confirmed by analyzing different numerical problems,
a hyperelastic material, an inelastic standard dissipative material in the context of
incremental variational formulations and a thermo-mechanical calculation. The sim-
plicity of the algorithms enabled an uncomplicated implementation such that the
user is only required to program the scalar energy function for the approximation
of stresses and moduli for hyperelastic and standard dissipative materials, or the
residual vector for an approximation of the tangent stiffness matrix in finite element
environments. This advantage of saving time during the development process while
still attaining accurate approximations could be exploited for implementing complex
element formulations or material models for which the derivation and programming
of analytical derivatives would be difficult. In the context of incremental variational
formulations it has additionally the advantage of enabling a fully automatic formu-
lation which just requires the definition of strain energy and dissipation potential.
22 M. Tanaka et al.

Acknowledgments Financial funding by the DFG Priority Program 1648 (SPPEXA “Software
for Exascale Computing”), projects BA 2823/8-1 and SCHR 570/19-1 is greatly acknowledged
by D. Balzani and J. Schröder. Furthermore, assistance for the thermo-mechanical calculations by
Ashutosh Gandhi is appreciated.

References

1. Balzani, D., Neff, P., Schröder, J., & Holzapfel, G. A. (2006). A polyconvex framework for
soft biological tissues. Adjustment to experimental data. International Journal of Solids and
Structures, 43(20), 6052–6070.
2. Balzani, D., Scheunemann, L., Brands, D., & Schröder, J. (2014). Construction of two- and
three-dimensional statistically similar RVEs for coupled micro-macro simulations. Computa-
tional Mechanics, 54, 1269–1284.
3. Balzani, D., Gandhi, A., Tanaka, M., & Schröder, J. (2015). Numerical calculation of thermo-
mechanical problems at large strains based on robust approximations of tangent stiffness matri-
ces. Computational Mechanics, 55, 861–871.
4. Bleier, N., & Mosler, J. (2012). Efficient variational constitutive updates by means of a novel
parameterization of the flow rule. International Journal for Numerical Methods in Engineering,
89, 1120–1143.
5. Clifford, W. K. (1873). Preliminary sketch of biquaternions. Proceedings of the London Math-
ematical Society, 4(64), 381–395.
6. Fike, J. A. (2013). Multi-objective optimization using hyper-dual numbers. Ph.D. thesis, Stan-
ford university.
7. Fike, J. A., & Alonso, J. J. (2011). The development of hyper-dual numbers for exact second-
derivative calculations. In 49th AIAA Aerospace Sciences Meeting including the New Horizons
Forum and Aerospace Exposition.
8. Golanski, D., Terada, K., & Kikuchi, N. (1997). Macro and micro scale modeling of thermal
residual stresses in metal matrix composite surface layers by the homogenization method.
Computational Mechanics, 19, 188–201.
9. Kim, S., Ryu, J., & Cho, M. (2011). Numerically generated tangent stiffness matrices using
the complex variable derivative method for nonlinear structural analysis. Computer Methods
in Applied Mechanics and Engineering, 200, 403–413.
10. Kiran, R., & Khandelwal, K. (2015). Automatic implementation of finite strain anisotropic
hyperelastic models using hyper-dual numbers. Computational Mechanics, 55, 229–248.
11. Kiran, R., & Khandelwal, K. (2014). Complex step derivative approximation for numerical
evaluation of tangent moduli. Computers and Structures, 140, 1–13.
12. Klinkel, S. (2000). Theorie und Numerik eines Volumen-Schalen-Elementes bei finiten elastis-
chen und plastischen Verzerrungen. Dissertation thesis, Institut für Baustatik, Universität Karl-
sruhe.
13. Lai, K.-L., & Crassidis, J. L. (2008). Extensions of the first and second complex-step derivative
approximations. Journal of Computational and Applied Mathematics, 219, 276–293.
14. Lyness, J. N. (1968). Differentiation formulas for analytic functions. Mathematics of Compu-
tation, 352–362.
15. Martins, J. R. R. A., & Hwang, J. T. (2013). Review and unification of discrete methods for
computing derivatives of single- and multi-disciplinary computational models. AIAA Journal,
51(11), 2582–2599.
16. Martins, J. R. R. A., Sturdza, P., & Alonso, J. J. (2003). The complex-step derivative approxi-
mation. ACM Transactions on Mathematical Software, 29, 245–262.
17. Miehe, C. (1996). Numerical computation of algorithmic (consistent) tangent moduli in large-
strain computational inelasticity. Computer Methods in Applied Mechanics and Engineering,
134, 223–240.
Robust Numerical Schemes for an Efficient Implementation … 23

18. Miehe, C., & Lambrecht, M. (2003). Analysis of microstructure development in shearbands by
energy relaxation of incremental stress potentials: Large-strain theory for standard dissipative
solids. International Journal for Numerical Methods in Engineering, 58, 1–41.
19. Miehe, C., Schotte, J., & Schröder, J. (1999). Computational micro-macro-transitions and
overall moduli in the analysis of polycrystals at large strains. Computational Materials Science,
16, 372–382.
20. Mosler, J., & Bruhns, O. T. (2009). Towards variational constitutive updates for non-associative
plasticity models at finite strain: Models based on a volumetric-deviatoric split. International
Journal of Solids and Structures, 46, 1676–1684.
21. Mosler, J., & Bruhns, O. T. (2010). On the implementation of rate-independent standard dis-
sipative solids at finite strain—variational constitutive updates. Computer Methods in Applied
Mechanics and Engineering, 199, 417–429.
22. Ortiz, M., & Stainier, L. (1999). The variational formulation of viscoplastic constitutive updates.
Computer Methods in Applied Mechanics and Engineering, 171, 419–444.
23. Pérez-Foguet, A., Rodríguez-Ferran, A., & Huerta, A. (2000). Numerical differentiation for
local and global tangent operators in computational plasticity. Computer Methods in Applied
Mechanics and Engineering, 189, 277–296.
24. Pérez-Foguet, A., Rodríguez-Ferran, A., & Huerta, A. (2000). Numerical differentiation for
non-trivial consistent tangent matrices: An application to the mrs-lade model. International
Journal for Numerical Methods in Engineering, 48, 159–184.
25. Schröder, J. 2013. A numerical two-scale homogenization scheme: the FE2 -method. In Plas-
ticity and beyond—microstructures, chrystal-plasticity and phase transitions (CISM Lecture
Notes). Vienna: Springer.
26. Schröder, J., Neff, P., & Balzani, D. (2005). A variational approach for materially stable
anisotropic hyperelasticity. International Journal of Solids and Structures, 42(15), 4352–4371.
27. Simo, J. C. (1988). A framework for finite strain elastoplasticity based on maximum plastic
dissipation and the multiplicative decomposition: Part I. Continuum formulation. Computer
Methods in Applied Mechanics and Engineering, 66, 199–219.
28. Simo, J., & Hughes, T. J. R. (1998). Computational inelasticity. Berlin: Springer.
29. Simo, J., & Miehe, C. (1992). Associative coupled thermoplasticity at finite strains: Formu-
lation, numerical analysis and implementation. Computer Methods in Applied Mechanics and
Engineering, 98, 41–104.
30. Smit, R. J. M., Brekelmans, W. A. M., & Meijer, H. E. H. (1998). Prediction of the mechanical
behavior of nonlinear heterogeneous systems by multi-level finite element modeling. Computer
Methods in Applied Mechanics and Engineering, 155, 181–192.
31. Tanaka, M., Fujikawa, M., Balzani, D., & Schröder, J. (2014). Robust numerical calculation
of tangent moduli at finite strains based on complex-step derivative approximation and its
application to localization analysis. Computer Methods in Applied Mechanics and Engineering,
269, 454–470.
32. Tanaka, M., Sasagawa, T., Omote, R., Fujikawa, M., Balzani, D., & Schröder, J. (2015). A
highly accurate 1st- and 2nd-order differentiation scheme for hyperelastic material models
based on hyper-dual numbers. Computer Methods in Applied Mechanics and Engineering,
283, 22–45.
33. Zienkiewicz, O. C., & Taylor, R. L. (1967). The finite element method for solid and structural
mechanics. Oxford: Butterworth-Heinemann.
Folding Patterns in Partially Delaminated
Thin Films

David Bourne, Sergio Conti and Stefan Müller

Abstract Michael Ortiz and Gustavo Gioia showed in the 90s that the complex
patterns arising in compressed elastic films can be analyzed within the context of the
calculus of variations. Their initial work focused on films partially debonded from
the substrate, subject to isotropic compression arising from the difference in thermal
expansion coefficients between film and substrate. In the following two decades
different geometries have been studied, as for example anisotropic compression. We
review recent mathematical progress in this area, focusing on the rich phase diagram
of partially debonded films with a lateral boundary condition.

1 Introduction

Elastic films deposited on a substrate are often subject, after thermal expansion, to
compressive strains which are released by debonding and buckling, generating a
variety of microstructures. The work of Michael Ortiz and Gustavo Gioia in the 90s
[1, 2] opened the way for the use of the tools of calculus of variations in the study of
these structures. Their starting point was the Föppl-von Kármán plate theory, as given
in (4) below. One of their insights was that the key nonconvexity which gives rise
to the microstructure can be understood in terms of the out-of-plane displacement
alone, leading after some rescalings to the Eikonal functional, as given in (1) below.
This functional contains a term of the form (|Dw|2 − 1)2 , where w is the normal

D. Bourne
Department of Mathematical Sciences, Durham University, Durham, UK
e-mail: [email protected]
S. Conti (B) · S. Müller
Institut Für Angewandte Mathematik, Universität Bonn, Bonn, Germany
e-mail: [email protected]
S. Müller
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 25

K. Weinberg and A. Pandolfi (eds.), Innovative Numerical Approaches
for Multi-Field and Multi-Scale Problems, Lecture Notes in Applied
and Computational Mechanics 81, DOI 10.1007/978-3-319-39022-2_2
26 D. Bourne et al.

displacement, which favours deformations with the property that the gradient of w
is approximately a unit vector, independently of the orientation. Since the film is
still bound to the substrate at the boundary of the debonded region, the appropriate
boundary condition is w = 0, which prescribes that the average over Ω of the gradient
of w vanishes. Therefore the resulting low-energy deformations have gradient Dw
oscillating between different values. As in many nonconvex variational problems,
oscillations on very small scales may  be energetically convenient, see [3, 4]. Corre-
spondingly, the variational problem Ω (|Dw|2 − 1)2 d x is not lower semicontinuous,
and - depending on the boundary data and forcing - does not have minimizers. How-
ever, the curvature term σ 2 |D 2 w|2 penalizes oscillations on an exceedingly fine scale
and thereby ensures existence of minimizers. The solutions then have oscillations
on an intermediate scale, which is determined by the competition between the two
terms. The analysis of the specific functional proposed by Ortiz and Gioia is reviewed
in Sect. 2 below.
The approach of Ortiz an Gioia was later extended to the full vectorial Föppl-von
Kármán energy, and also to three-dimensional elasticity. These refinements explained
the appearance of oscillations on two different length scales, with coarse oscillations
in a direction normal to the boundary, and fine oscillations in the direction tangential
to the boundary, as discussed in Sect. 3 below.
Recently interest has been directed to controlling the microstructures by designing
the geometry of the debonded region appropriately [5, 6]. The key idea is to introduce
a sacrificial layer between the film and the substrate, and then to selectively etch away
a part of it, so that the boundary of the debonded region is straight. The film then
partially rebonds to the surface, leading to complex patterns of tunnels. A study
of these patterns within the Ortiz-Gioia framework, with a variational functional
containing the Föppl-von Kármán energy and a fracture term proportional to the
debonded area, is presented in Sect. 4. The mathematical analysis leading to the upper
bounds of Theorem 6 suggests the presence of different types of patterns in different
parameter ranges. The picture is rather easy in the two extreme cases in which the
bonding energy per unit area is very small or very large. Indeed, in the first one the
patterns observed for completely debonded films give the optimal energy scaling,
in the second one the optimal state corresponds to the film completely bound to the
substrate. In the intermediate regime we expect a richer picture, with bonded areas
separated by thin debonded tunnels. For a certain regime, depending on the relation
between the bonding energy per unit area, the film thickness and the compression
ratio, a construction in which the tunnels branch and refine close to the boundary has
a lower energy than the one with straight tunnels, see discussion in Sect. 4 below.
The microstructure formation in thin films can be understood at a qualitative level
as a form of Euler buckling instability. The relevant experiments, however, are well
beyond the stability threshold, as discussed in Sect. 5 below.
Folding Patterns in Partially Delaminated Thin Films 27

2 Scalar Modeling of Compressed Thin Films

Ortiz and Gioia showed that, if tangential displacements are neglected, the energy
of a compressed thin film can be characterized by the functional
  
 2
Iσ [w] = |Dw|2 − 1 + σ 2 |D 2 w|2 d x, (1)
Ω

subject to w = 0 on ∂Ω and w ≥ 0 in Ω. Here Ω ⊂ R2 represents the debonded

region, w : Ω → [0, ∞) the rescaled normal displacement, and σ is a small para-
meter related to the thickness of the film. This functional arises also naturally in the
study of liquid crystal configurations [7] and of magnetic structures in thin films [8].
Despite a large mathematical effort [7, 9–15] the problem (1) is not yet completely
understood; it has been shown that the minimal energy is proportional to σ but the
Γ -limit of σ −1 Iσ [w] has only been partially identified. The natural candidate is

1
I0 [w] = |[Dw]|3 dH 1 (2)
3 J Dw

restricted to functions w : Ω → R which solve the Eikonal equation |Dw| = 1 and

are sufficiently regular. Here, J Dw denotes the set of points (typically, a curve) where
the gradient Dw is not continuous, [Dw] denotes its jump across the interface, and
dH 1 the line integral along the interface. In particular, under the additional assump-
tion that Dw is a function of bounded variation, it has been shown that for σ → 0 the
scaled functionals σ −1 Iσ converge, in the sense of Γ -convergence, to I0 , see [11–13]
for the lower bound and [14–16] for the upper bound. However, it is also clear that
finiteness of the energy does not imply that Dw has bounded variation, but only that
w belongs to a larger space, called AG(Ω), see [10, 11]. Therefore the result is still
incomplete.
The Eikonal equation |Dw| = 1 with the boundary data w = 0 on ∂Ω is solved by
the distance to the boundary, w0 (x) = dist(x, ∂Ω). For example, if Ω is a square this
leads to the tent-shaped deformation illustrated in Fig. 1. The function w0 is however
only Lipschitz continuous, not twice differentiable, and makes the curvature term

Fig. 1 Sketch of a
deformation achieving the
optimal energy in (1). Here
the debonded region
Ω = (0, 1)2 is a square, and
the distance to the boundary
gives a “tent”-form. The
convolution in (3) makes the
folds have smooth transitions
on a small scale
28 D. Bourne et al.

w w

x2 x2

Fig. 2 Sketch of the effect of the mollification in (3) in a direction orthogonal to the fold. Left
panel: the distance from the boundary dist(x, ∂Ω) is a function with slope ±1 and sharp kinks.
Right panel: the mollification defined in (3) still has slope ±1 on large parts of the domain, but has
smooth transitions from one value to the other over a length of the order of σ , see Fig. 3


Ω σ |D w| d x infinite. Therefore Ortiz and Gioia [1, 2] proposed to use a smoothed
2 2 2

version of the distance function,


wσ (x) = dist(y, ∂Ω)ϕσ (x − y)dy (3)
Ω

where ϕσ is a mollifier on the scale σ , i.e., ϕσ ∈ Cc∞ (Bσ ) with R2 ϕσ d x = 1 and
|Dϕσ | ≤ c/σ 3 . Then the regularized gradient Dwσ has length close to 1 on most of the
domain Ω, but at boundaries between regions where Dw0 has different orientations
Dwσ changes smoothly over a length scale σ from one value to the other. The bending
energy is correspondingly localized in a stripe of thickness 2σ around the interfaces,
see Fig. 2. The prediction that minimizers of (1) are well represented by wσ is in good
agreement, at least for some geometries, with experimental observations [1, 2].
The work of Ortiz and Gioia was then extended to related problems, showing for
example that under anisotropic compression branching-type microstructures appear
close to the boundary [17, 18], or that in certain regimes telephone-cord blisters
develop [19–21] thanks to the interaction between the elastic deformation and the
fracture problem that determines the boundary of the debonded region.

3 Pattern Formation in Debonded Thin Films

A finer analysis of the nonlinear elasticity model that had led to (1) showed that, in
the case of isotropic compression, also the in-plane components exhibit fine-scale
oscillations which refine close to the boundary [22–25]. This analysis was based
on the Föppl-von Kármán model, which includes the tangential components of the
displacement u as well. After rescaling the energy takes the form (in the case of zero
Poisson’s ratio for simplicity)

 
E σ [u, w] = |Du + Du T + Dw ⊗ Dw − Id|2 + σ 2 |D 2 w|2 d x. (4)
Ω
Folding Patterns in Partially Delaminated Thin Films 29

Here Ω ⊂ R2 is, as above, the debonded region, and the displacements u and w
vanish at the boundary of Ω, corresponding to the fact that the rest of the film is still
bound to the substrate. The isotropic compressive strain has been scaled to 1, and
one can check that E σ [0, w] = 1 + Iσ [w]. The key result from [22, 23] was that the
minimum energy scales proportional to σ :
Theorem 1 (From [22, 23]) Let Ω ⊂ R2 be a bounded domain with piecewise
smooth boundary. Then there are two constants c L , cU > 0 such that

c L σ ≤ min{E σ [u, w] : u = 0, w = 0 on ∂Ω} ≤ cU σ. (5)

The argument used for proving the lower bound also proves that a finite fraction of
the energy is localized in a thin strip close to the boundary.
Similar statements hold if the plate theory in (4) is replaced by a fully three-
dimensional nonlinear elastic model. For v : Ω × (0, h) → R3 , h > 0, we define

1
E h [v] =
3D
W (Dv)d x (6)
h Ω×(0,h)

where W : R3×3 → [0, ∞) is the elastic stored energy density, which vanishes on
the set of proper rotations SO(3) and has quadratic growth, in the sense that

c dist2 (F, SO(3)) ≤ W (F) ≤ c dist2 (F, SO(3)) (7)

for some positive constants c and c . The factor 1/ h is included explicitly in (6) to
obtain an energy per unit thickness, corresponding to (4).
In the nonlinear case the thickness h of the film and the compression δ enter the
problem separately, however to leading order and after scaling the optimal energy
only depends on the combination σ = h/δ 1/2 . In order to understand this expression
it is instructive to recall the relation between the three-dimensional problem E h3D
and its two-dimensional counterpart E σ . In particular, a given pair (u, w) in (4)
corresponds to a three-dimensional deformation vδ of the form

vδ (x1 , x2 , x3 ) = (1 − δ) [ψ(x1 , x2 ) + x3 n(x1 , x2 )] (8)

where ⎛ ⎞
x1 + 2δu 1 (x1 , x2 )
ψ(x1 , x2 ) = ⎝x2 + 2δu 2 (x1 , x2 )⎠ (9)
(2δ)1/2 w(x1 , x2 )

represents the deformation of the x3 = 0 layer and

⎛ ⎞
−(2δ)1/2 ∂1 w(x1 , x2 )
n(x1 , x2 ) = ⎝−(2δ)1/2 ∂2 w(x1 , x2 )⎠ (10)
1
30 D. Bourne et al.

is, to leading order, the normal to the surface described by ψ and gives the out-of-
plane component of the strain. An expansion of E h3D [vδ ] for small δ shows that the
leading order contribution is proportional to δ 2 E σ [u, w] if the Poisson’s ratio of the
material vanishes. See for example [22, App. A and App. B] for a more detailed
discussion of this point. A rigorous relation between E σ and E h3D was derived in
[26, 27] by means of Γ -convergence, these results however are appropriate for a
different regime, with much smaller energy, and therefore do not apply directly to
the situation of interest here.
Theorem 2 (From [25]) Let Ω ⊂ R2 be a bounded domain with piecewise smooth
boundary, δ ∈ (0, 1), h ∈ (0, δ 1/2 ). Then there are two constants c L , cU > 0 such
that
1 3D
c L σ ≤ min{ E [u] : u(x) = (1 − δ)x for (x1 , x2 ) ∈ ∂Ω} ≤ cU σ (11)
δ2 h

where σ = h/δ 1/2 .

The significance of Theorems 1 and 2 is best understood by considering the key
ideas in the proofs. The upper bound in (5) and (11) is proven by explicitly construct-
ing a suitable deformation field (u, w). This is done in several steps. The first step is
the Ortiz-Gioia construction given in (3), which correctly describes the large-scale
behavior of the film and relaxes the compression in direction normal to the boundary,
as in Fig. 1. In the second step one adds fine-scale oscillations in the orthogonal direc-
tion, as illustrated in Fig. 3. This microstructure does not change the average shape
significantly but relaxes the strain component tangential to the boundary. Finally, one
realizes that optimal deformations have oscillations on a very fine scale close to the
boundary, to adequately match the boundary data, but much coarser oscillations in

Fig. 3 Sketch of a deformation achieving the optimal upper bound in (5). As in Fig. 1, the debonded
region Ω = (0, 1)2 is a square. The starting point, at a coarse scale, is the “tent”-form illustrated in
Fig. 1. At a finer scale, folds orthogonal to the boundary relax the tangential compression (left panel,
folds are only drawn in a small region). The period of the folds is of order h close to the boundary,
and via a sequence of period-doubling steps becomes coarser in the inside (right, blow-up of the
folds from the left panel)
Folding Patterns in Partially Delaminated Thin Films 31

the interior, to minimize the bending energy. Therefore a number of period-doubling

steps are inserted, as illustrated in Fig. 3. Analogous self-similar branched patterns
had previously appeared in the study of microstructures in shape-memory alloys [28,
29], where for a simplified model it had been possible to show that minimizers are
indeed asymptotically self-similar [30]. The scaling in the presence of finite elastic-
ity, both of the martensite and in the surrounding austenite, was then studied in [31,
32]; vectorial variants of the model were considered in [33, 34]. A similar approach
has been useful also for a variety of other problems, ranging from magnetic patterns
in ferromagnets [35–37] to field penetration in superconductors [38, 39], dislocation
structures in crystal plasticity [40] and coarsening in thin film growth [41].
This variational approach to microstructure formation in thin elastic sheets is
much more general, and indeed it can be applied to a number of related problems.
One example is paper crumpling [42, 43] in which a thin plate, completely detached
from the substrate, is confined to a small volume. In this case it has been possible to
construct deformations with much smaller energy per unit volume. In particular one
can obtain an energy per unit thickness proportional to h 5/3 [44, 45], and one can
approximate any compressive deformation with this energy.

Theorem 3 (From [45]) Let Ω ⊂ R2 be a bounded domain, r > 0. Then there is a

map v : Ω × (0, h) → Br (0) such that

E h3D [v] ≤ ch 5/3 . (12)

The constant c may depend on Ω and r but not on h. Further, if v0 : Ω → R3 is a

short map, i.e., a map which obeys |v0 (x) − v0 (y)| ≤ |x − y| for all x, y ∈ Ω, then
there is a sequence vh , converging to v0 , such that

1 3D
lim E [vh ] = 0 (13)
h→0 hα h

for any α < 5/3. Convergence of vh is understood as uniform convergence of the

vertical averages.

The proof of this is based on the combination of three ingredients. The first one is
an approximation of short maps with Origami maps:
Theorem 4 (From [45]) Let v0 : Ω → R3 be a short map, i.e., a map which obeys
|v0 (x) − v0 (y)| ≤ |x − y| for all x, y ∈ Ω. Then there is a sequence v j of Origami
maps converging uniformly to v0 .
Here we say that a map v : R2 → R3 is an Origami map if it is continuous and
piecewise isometric, i.e., if the domain can be subdivided into pieces such that v is
a linear isometry (a translation plus a rotation) in each piece. The number of pieces
is allowed to diverge only at infinity, in the sense that only finitely many pieces are
allowed in any bounded subset of R2 .
The second step is to approximate any Origami maps with low-energy maps:
32 D. Bourne et al.

Theorem 5 (From [45]) Let Ω ⊂ R2 be a bounded domain, v0 : Ω × (0, h) →

Br (0) be an Origami map. Then for any Origami map v0 there is a sequence of
maps vh : Ω × (0, h) → R3 , converging to v0 , such that

E h3D [vh ] ≤ Ch 5/3 . (14)

The constant may depend on Ω and v0 but not on h.

This is proven by an explicit construction around each fold.
Another related problem of high current interest is the study of wrinkling patterns
in graphene sheets [46, 47]. This has been addressed by a similar model, in which
the boundary conditions are replaced by a viscous term describing the interaction
with a substrate [48–50]. It would be interesting to see if the methods discussed here
can be useful also for this variant of the problem.

4 Pattern Formation in Rebonded Thin Films

The microstructures spontaneously developed by compressed thin films can be con-

trolled if the geometry of the debonded region is designed appropriately [5, 6]. One
possibility is to introduce a sacrificial layer between the film and the substrate, and
then to selectively etch away a part of it, so that the boundary of the debonded region
is straight, see sketch in Fig. 4. The film then partially rebonds to the surface, leading
to complex patterns of tunnels, which in some cases refine close to the boundary, see
Fig. 5.
These patterns can be studied by coupling the von-Kármán energy with a fracture
term proportional to the debonded area,
  
E σ,γ [u, w] = |Du + Du T + Dw ⊗ Dw − Id|2 + σ 2 |D 2 w|2 d x + γ |{w > 0}|.
Ω
(15)

x3

x2

x1

Fig. 4 Geometry of the partially delaminated film. The intermediate sacrificial layer is removed
chemically only for x1 > 0. The free-standing film is subject to compression at the Dirichlet bound-
ary x1 = 0 and may rebond to the substrate
Folding Patterns in Partially Delaminated Thin Films 33

Fig. 5 Experimental picture of tube branching in Si1−x Gex film on a thick SiO2 substrate. Left
AFM image of the network near the etching front. Right autocorrelation pattern. Reprinted from
[5, Fig. 2] with permission from Wiley

The three terms represent stretching, bending and bonding energies respectively.
Here u : Ω → R2 are the (scaled) tangential displacements and γ > 0 is the bonding
energy per unit area (related to Griffith’s fracture energy), |{w > 0}| represents the
area of the set where the vertical displacement w is nonzero. Equivalently one could
take the debonded state as reference and consider a negative term proportional to the
rebonded area, −γ |{w = 0}|; the two energies only differ by an additive constant. The
appropriate boundary conditions correspond to the film being bound to a substrate
on one side of the domain; for simplicity we shall focus on Ω = (0, 1)2 with u = 0
and w = 0 on the {x1 = 0} side of Ω. As above, we assume w ≥ 0 everywhere.
The mathematical analysis of the energy (15) leads to the rich phase diagram
sketched in Fig. 6, which contains four different regimes [51] that we now illustrate.
For large specific bonding energy γ the film is completely bound to the substrate.
In particular the film is flat, so that there is no bending energy, but the stretching
energy is not released. The total energy is then proportional to the area of Ω, and
one obtains E σ,γ [0, 0] = 2. This is regime A in Fig. 6 and Theorem 6.
The opposite case of very small bonding energy γ is also easy to understand
after the foregoing discussion: here the bonding term plays no significant role and

A
B

Fig. 6 Phase diagram for E σ,γ [u, w] in the (σ, γ ) plane

34 D. Bourne et al.

x2

Fig. 7 Sketch of the laminate regime (B)

the film is completely detached from the substrate. One recovers the result of the
blistering problem of Theorem 1, E σ,γ [u, w] E σ [u, w] ≤ cσ . The corresponding
deformations are those illustrated in Fig. 3. This is regime D in Fig. 6 and Theorem 6.
For intermediate values of γ the situation is more complex, in particular debonded
channels are formed, which separate wider bonded regions. In regime B the pattern
is periodic and, away from the Dirichlet boundary, depends only on the tangential
variable x2 . A large part of the film is bonded to the substrate, but bonded regions
are separated by thin tubes, see Fig. 7. Denoting by h the period of the oscillations,
and by δ the width of a tube, the total volume fraction of the tubes is δ/ h, therefore
the bonding energy is proportional to γ δ/ h. Each tube has to release a compression
of h over a width δ, therefore the term |Dw|2 is of order h/δ inside the tubes (the
stretching energy is then completely relaxed). This gives |Dw| ∼ (h/δ)1/2 in the
tubes, and hence |D 2 w| ∼ (h/δ)1/2 /δ. Therefore the total energy can be estimated
by
2
δ δ h 1/2 /δ 1/2 δ σ2
γ + σ2 =γ + 2 . (16)
h h δ h δ

Optimizing in δ we obtain δ ∼ σ 2/3 h 1/3 γ −1/3 (this is clearly only admissible if

δ ≤ h ≤ 1). The period h is fixed by the energetic cost of the interpolation region
close to the boundary. In this part of the domain there is no stretch-free construction,
and indeed an interpolation over a boundary layer of thickness ε results in a total
stretching energy of ε(1 + h 2 /ε2 ). Optimizing over ε we obtain ε ∼ h, and therefore
the total energy for the laminate construction is

δ σ2
h+γ + 2. (17)
h δ
Inserting the value of δ obtained above and minimizing in h we conclude that h
and E are proportional to (σ γ )2/5 . The width of each tube δ is then proportional to
σ 4/5 γ −1/5 . This is regime B in Fig. 6; a precise version of this construction proves
the second bound in Theorem 6.
If the bending term becomes more important, it is convenient to insert period-
doubling steps, just like in the discussion of the functional (4). The resulting pattern
is shown in Fig. 8. In comparison to the pattern of Fig. 3 the key difference is that the
bending is localized to a small region, whereas large parts of the film are bond to the
Folding Patterns in Partially Delaminated Thin Films 35

Fig. 8 Sketch of the tube branching regime (C)

substrate. The period-doubling steps are only possible at the expense of stretching
energy; balancing the different terms one finds [51] that the resulting energy is pro-
portional to σ 1/2 γ 5/8 . The result of the construction is summarized in the following
statement.

Theorem 6 (From [51]) Let γ > 0, σ ∈ (0, 1). There are u, w which obey the stated
boundary conditions and
⎧
⎪
⎪1 if σ γ > 1 (regime A),
⎨
(σ γ )2/5 if σ −4/9 ≤ γ ≤ σ −1 (regime B),
E σ,γ [u, w] ≤ c (18)
⎪
⎪σ 1/2 γ 5/8 if σ 4/5 ≤ γ ≤ σ −4/9 (regime C),
⎩
σ if γ < σ 4/5 (regime D).

The proof is based on making the constructions sketched above precise, details are
given in [22] for regime D and in [51] for regimes B and C. Regime A, as discussed
above, is immediate.
Optimality of the phase diagram just discussed can be at least partially proven
by providing matching lower bounds on the energy. In particular, one can show the
following.
Theorem 7 (From [51]) Let γ > 0, σ ∈ (0, 1). For any u, w which obey the stated
boundary conditions one has
⎧
⎨1 if σ γ > 1 (regime A),
E γ ,σ [u, w] ≥ c (σ γ )2/3 if σ 1/2 ≤ γ ≤ σ −1 (regime B ), (19)
⎩
σ if γ < σ 1/2 (regime D ).

Whereas the statement in regime D follows from [22], the other two bounds are
proven in [51] using the Korn-Poincaré inequality for SBD2 functions obtained in
[52].
36 D. Bourne et al.

Theorem 7 proves optimality in phases A and D. The bound in the intermediate

region does not, however, match the upper bounds stated in Theorem 6. Therefore it
is at this stage not clear if the branching patterns illustrated in Fig. 8 are optimal.

5 Linear Stability Analysis

The general form of the linearized Föppl-von Kármán plate theory under isotropic
compression is [53, 54]
  
1 h2  
E FvK [u, w] = Yh (1 − ν)|ε|2 + ν(Trε)2 + (1 − ν)|D 2 w|2 + ν(Δw)2 d x ,
2 Ω 12
(20)
see also [2, 22] for a discussion in the present context and [27] for a rigorous mathematical
derivation. Here ν ∈ [−1, 1/2] is the Poisson ratio, Y Young’s modulus, h the film thickness,
and the strain ε is defined by

ε = Du + (Du)T + Dw ⊗ Dw − 2δId , (21)

where δ is the eigenstrain (i.e., the compression enforced by the substrate). We recall
that we use |M|2 = TrM T M for the matrix norm. For ν = 0, after a rescaling (20)
reduces to (4). We recall that in [22, App. B] it was shown that the scaling behavior of
the functional E FvK is the same for all ν ∈ (−1, 1/2], hence our results hold also for
generic values of the Poisson ratio. Of course, the regime ν ≥ 0 is the most relevant.
For small δ one can linearize around the state u = 0, w = 0. After straightforward
computations this leads to
 
1
lin
E FvK [u, w] = Yh (1 − ν)|Du + Du T − 2δId|2 − 4(1 − ν)δ|Dw|2
2 Ω
h2  
+ ν(2divu − 4δ)2 − 8δν|Dw|2 + (1 − ν)|D 2 w|2 + ν(Δw)2 d x .
12

In this linearized functional u and w are decoupled. The dependence on u is convex,

hence u = 0 is the minimizer with the given boundary data. The dependence on w
is however not necessarily convex. Working for concreteness in a circle of radius R,
we can assume w to be radial, w(x) = ϕ(|x|), subject to ϕ(R) = 0, so that
x
Dw(x) = ϕ (|x|)
|x|

and
x x Id x⊗x
D 2 w(x) = ϕ (|x|) ⊗ + ϕ (|x|) − .
|x| |x| |x| |x|3
Folding Patterns in Partially Delaminated Thin Films 37

Inserting into the energy leads to the one-dimensional variational problem

 
R
1
Yh −4δ(1 + ν)(ϕ (r ))2
2 0
 
2
h2 ϕ (r ) ϕ (r )ϕ (r )
+ (ϕ (r ))2 + + 2ν r dr .
12 r r

This is positive definite if the first term, of order δ, is not larger then the second term,
of order h 2 /R 2 . Therefore the loss of stability, which corresponds to Euler buckling,
occurs at strains δ ∼ h 2 /R 2 . Inserting the experimental data from [5], namely, h ∼ 20
nm, R ∼ 10 µm, ν ∼ 0.277, leads to δcrit ∼ 4 × 10−6 , which corresponds to a strain
of 0.0004 %. This is over three orders of magnitude smaller than the experimentally
applied strain δExp ∼ 0.011 = 1.1 %. Therefore the experiments we discussed take
place well beyond the loss of linear stability, and a buckling-postbuckling analysis
does not seem appropriate to understand the deformations. Our variational approach
is instead constructed to deal with deformations and microstructures that appear in
the deeply nonlinear regime and is therefore more suitable to study the mentioned
experiments.

References

1. Ortiz, M., & Gioia, G. (1994). Dynamically propagating shear bands in impact-loaded
prenotched plates II. Numerical simulations. Journal of the Mechanics and Physics of Solids,
42, 531.
2. Gioia, G., & Ortiz, M. (1997). Delamination of compressed thin films. Advances in Applied
Mechanics, 33, 119.
3. Dacorogna, B. (1989). Direct methods in the calculus of variations. Applied Mathematical
Sciences, 78. Springer.
4. Müller, S.: F. Bethuel et al. (Ed.) Calculus of variations and geometric evolution problems.
Lecture Notes in Math (1713, pp. 85–210). Springer.
5. Mei, Y., Thurmer, D. J., Cavallo, F., Kiravittaya, S., & Schmidt, O. G. (2007). Semiconductor
sub-micro-/nanochannel networks by deterministic layer wrinkling. Advanced Materials, 19,
2124.
6. Cendula, P., Kiravittaya, S., Mei, Y. F., Deneke, C., & Schmidt, O. G. (2009). Bending and
wrinkling as competing relaxation pathways for strained free-hanging films. Physical Review
B, 79, 085429.
7. Aviles, P., & Giga, Y. (1987). A mathematical problem related to the physical theory of liq-
uid crystal configurations. Proceedings of the Centre for Mathematical Analysis, Australian
National University, 12, 1.
8. DeSimone, A., Kohn, R. V., Müller, S., & Otto, F. (2000). Magnetic microstructures : a paradigm
of multiscale problems. In J.M. Ball (ed.) ICIAM 99: proceedings of the Fourth International
Congress on Industrial and Applied Mathematics (pp. 175–190). Oxford: Oxford University
Press.
9. Aviles, P., & Giga, Y. (1996). The distance function and energy. Proceedings of the Royal
Society of Edinburgh Section A, 126, 923.
38 D. Bourne et al.

10. Aviles, P., & Giga, Y. (1999). On lower semicontinuity of a defect energy obtained by a singular
limit of the Ginzburg-Landau type energy for gradient fields. Proceedings of the Royal Society
of Edinburgh Section A, 129, 1.
11. Ambrosio, L., De Lellis, C., & Mantegazza, C. (1999). Line energies for gradient vector fields
in the plane. Calculus of Variations and Partial Differential Equations, 9, 327.
12. Jin, W., Kohn, R.V. (2000). Singular perturbation and the energy of folds. Journal of Nonlinear
Science, 10, 355.
13. DeSimone, A., Kohn, R. V., Müller, S., & Otto, F. (2001). A compactness result in the gradient
theory of phase transitions. Proceedings of the Royal Society of Edinburgh Section A, 131, 833.
14. Poliakovsky, A. (2007). Upper bounds for singular perturbation problems involving gradient
fields. Journal of the European Mathematical Society, 9(1), 1.
15. Conti, S., & De Lellis, C. (2007). Sharp upper bounds for a variational problem with singular
perturbation. Mathematische Annalen, 338, 119.
16. Poliakovsky, A. (2005). A method for establishing upper bounds for singular perturbation
problems. Comptes Rendus Mathematique Academy of Science Paris, 341, 97.
17. Gioia, G., DeSimone, A., Ortiz, M., & Cuitino, A. M. (2002). Folding energetics in thin-film
diaphragms. Proceedings of the Royal Society of London. Series A, 458, 1223.
18. Conti, S., DeSimone, A., & Müller, S. (2005). Self-similar folding patterns and energy scaling
in compressed elastic sheets. Computer Methods in Applied Mechanics and Engineering, 194,
2534.
19. Gioia, G., & Ortiz, M. (1998). Determination of thin-film debonding parameters from
telephone-cord measurements. Acta Materialia, 46(1), 169.
20. Audoly, B. (1999). Stability of straight delamination blisters. Physical Review Letters, 83,
4124.
21. Audoly, B. (2000). Mode-dependent toughness and the delamination of compressed thin films.
Journal of the Mechanics and Physics of Solids, 48, 2315.
22. Ben Belgacem, H., Conti, S., DeSimone, A., Müller, S. (2000). Rigorous bounds for the Föppl-
von Kármán theory of isotropically compressed plates. Journal of Nonlinear Science, 10, 661.
23. Jin, W., & Sternberg, P. (2001). Energy estimates for the von Kármán model of thin-film
blistering. Journal of Mathematical Physics, 42, 192.
24. Jin, W., & Sternberg, P. (2002). In-plane displacements in thin-film blistering. Proceedings of
the Royal Society of Edinburgh Section A, 132, 911.
25. Ben Belgacem, H., Conti, S., DeSimone, A., Müller, S. (2002). Energy scaling of compressed
elastic films. Archive for Rational Mechanics and Analysis, 164, 1.
26. Friesecke, G., James, R. D., & Müller, S. (2002). A theorem on geometric rigidity and the
derivation of nonlinear plate theory from 3d elasticity. Communications on Pure and Applied
Mathematics, 55, 1461.
27. Friesecke, G., James, R. D., & Müller, S. (2006). A hierarchy of plate models derived from
nonlinear elasticity by Gamma-convergence. Archive for Rational Mechanics and Analysis,
180, 183.
28. Kohn, R. V., & Müller, S. (1992). Branching of twins near an austenite-twinned-martensite
interface. Philosophical Magazine A, 66, 697.
29. Kohn, R. V., & Müller, S. (1994). Surface energy and microstructure in coherent phase transi-
tions. Communications on Pure and Applied Mathematics, 47, 405.
30. Conti, S. (2000). Branched microstructures: scaling and asymptotic self-similarity. Communi-
cations on Pure and Applied Mathematics, 53, 1448.
31. Conti, S. (2006). A lower bound for a variational model for pattern formation in shape-memory
alloys. Continuum Mechanics and Thermodynamics, 17, 469.
32. Zwicknagl, B. (2014). Microstructures in low-hysteresis shape memory alloys: scaling regimes
and optimal needle shapes. Archive for Rational Mechanics and Analysis, 213, 355.
33. Diermeier, J. (2013). Master’s thesis, Universität Bonn.
34. Chan, A., & Conti, S. (2015). Energy scaling and branched microstructures in a model for
shape-memory alloys with SO (2) invariance. Mathematical Models and Methods in Applied
Sciences, 25, 1091.
Folding Patterns in Partially Delaminated Thin Films 39

35. Choksi, R., & Kohn, R. V. (1998). Bounds on the micromagnetic energy of a uniaxial ferro-
magnet. Communications on Pure and Applied Mathematics, 51, 259.
36. Choksi, R., Kohn, R. V., & Otto, F. (1999). Domain branching in uniaxial ferromagnets: a
scaling law for the minimum energy. Communications on Pure and Applied Mathematics, 201,
61.
37. Viehmann, T. (2009). Uniaxial ferromagnets. Ph.D. thesis, Universität Bonn.
38. Choksi, R., Kohn, R. V., & Otto, F. (2004). Energy minimization and flux domain structure in
the intermediate state of a type-I superconductor. Journal of Nonlinear Science, 14, 119.
39. Choksi, R., Conti, S., Kohn, R. V., & Otto, F. (2008). Ground state energy scaling laws during
the onset and destruction of the intermediate state in a type-I superconductor. Communications
on Pure and Applied Mathematics, 61, 595.
40. Conti, S., & Ortiz, M. (2005). Dislocation microstructures and the effective behavior of single
crystals. Archive for Rational Mechanics and Analysis, 176, 103.
41. Conti, S., & Ortiz, M. (2008). Minimum principles for the trajectories of systems governed by
rate problems. Journal of the Mechanics and Physics of Solids, 56, 1885.
42. Lobkovsky, A. E., Gentges, S., Li, H., Morse, D., & Witten, T. A. (1995). Boundary layer
analysis of the ridge singularity in a thin plate. Science, 270, 1482.
43. Witten, T. A. (2007). Stress focusing in elastic sheets. Reviews of Modern Physics, 79, 643.
44. Venkataramani, S. C. (2004). Lower bounds for the energy in a crumpled elastic sheet a minimal
ridge. Nonlinearity, 17, 301.
45. Conti, S., & Maggi, F. (2008). Confining thin elastic sheets and folding paper. Archive for
Rational Mechanics and Analysis, 187, 1.
46. Li, X., Cai, W., An, J., Kim, S. et al. (2009). Large-area synthesis of high-quality and uniform
graphene films on copper foils. Science, 324(5932), 1312.
47. Zang, J., Ryu, S., Pugno, N., Wang, Q., Tu, Q., Buehler, M. J., et al. (2013). Multifunctionality
and control of the crumpling and unfolding of large-area graphene. Nature Materials, 12, 321.
48. Zhang, K., & Arroyo, M. (2013). Adhesion and friction control localized folding in supported
graphene. Journal of Applied Physics, 113, 193501.
49. Zhang, K., & Arroyo, M. (2014). Understanding and strain-engineering wrinkle networks in
supported graphene through simulations. Journal of Mechanics and Physics of Solids, 72, 61.
50. Arroyo, M., Zhang, K., & Rahimi, M. (2014). Mechanics of confined solid and fluid thin
films: Graphene and lipid bilayers. IUTAM Symposium on innovative numerical approaches
for materials and structures in multi-field and multi-scale problems.
51. Bourne, D., Conti, S., & Müller, S. (2015). Energy bounds for a compressed elastic film on a
substrate. Preprint arXiv:1512.07416
52. Chambolle, A., Conti, S., & Francfort, G. (2014). Korn-Poincaré inequalities for functions
with a small jump set. To appear in Indiana University Mathematics Journal. Preprint hal-
01091710v1.
53. Antman, S. S. (1995). Applied Mathematical Sciences. In Nonlinear problems in elasticity,
107. Springer.
54. Ciarlet, P. G. (1997). Theory of plates, Mathematical elasticity, vol. II. Elsevier.
Thermo-mechanical Behavior of Confined
Granular Systems

Gülşad Küçük, Marcial Gonzalez and Alberto M. Cuitiño

Abstract We present a mathematical formulation that integrates thermal contact

and Hertzian deformation models to understand the thermo-mechanical behavior
of consolidated granular systems. The model assumes quasi-static equilibrium and
quasi-steady heat conduction conditions that are appropriate for many thermally-
assisted manufacturing processes. We perform a parametric study that explores the
effect of applied thermal and mechanical loads, and of particles’ thermal expansion.
The nonlinearity of the multi-physics problem reveals that thermo-mechanical cou-
pling enhances the effective thermal conductivity and mechanical stiffness by directly
impacting the interrelation between contact conductance and overlapping between
the particles. Alterations in temperature profiles and displacements of particles are
significant for materials with higher thermal expansion coefficients. In this regards,
it is worth noting that the results of the proposed thermo-mechanical model depart
from those of conventional compaction models based on a continuum mechanics
description.

1 Introduction

Understanding the fundamental multi-physics behind the thermo-mechanically cou-

pled deformation of granular systems and its projections in macroscopic scale pro-
vides the essentials to fabricate particulate assemblies with specific functionalities.

G. Küçük (B) · A.M. Cuitiño

Department of Mechanical and Aerospace Engineering, Rutgers University,
Piscataway, NJ 08854, USA
e-mail: [email protected]
A.M. Cuitiño
e-mail: [email protected]
M. Gonzalez
School of Mechanical Engineering, Purdue University, 585 Purdue Mall,
West Lafayette, IN 47907-2088, USA
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 41

K. Weinberg and A. Pandolfi (eds.), Innovative Numerical Approaches
for Multi-Field and Multi-Scale Problems, Lecture Notes in Applied
and Computational Mechanics 81, DOI 10.1007/978-3-319-39022-2_3
42 G. Küçük et al.

A proper estimate of the mechanical strength, and of the thermal and electrical con-
ductivity of a compacted solid is contingent upon the knowledge of microstructure
formation during the deformation stage of the compression. Since thermally assisted
compaction of granular matter is of great importance for a wide set of manufacturing
processes, theoretical modeling and numerical simulations serve as significant tools
to forecast the macroscopic behavior of materials, essentially when experimental
techniques are also unfeasible.
At present, one of the most implemented methodologies to elucidate the col-
lective behavior of particulate materials is the continuum mechanics approach, in
which the granular material is assumed to be statistically homogenous [1]. This is
achieved by treating the system as units of ordered groups, simulating disordered
arrangements by statistical correlation functions or using empirical correlations. The
statistical averaging technique provides homogenized solutions of the highly hetero-
geneous granular media at the cost of imposing two assumptions: (i) affine motion
approximation, namely the motion of each grain follows the macroscopic strain, and
(ii) well-bonded structure, contact number and positioning do not change under the
applied load. Despite the fact that the effective medium theory particularly estimates
the effective elastic moduli of packed bed of spherical particles to a large extent, the
discrepancy between numerical and experimental results is remarkable. Makse and
co-workers questioned the relevance of force laws defined at single contact level,
where they pointed out that the simplification done in effective medium theory is the
misleading element in the formulation [2, 3]. Affine motion assumption demolishes
the ability of the approach to account for the relaxation and rearrangement of parti-
cles that are under shear deformation. Moreover concerning the variety of boundary
conditions and geometrical effects, experimentation techniques become insufficient
in providing sufficient information about the microstructure to feed empirical corre-
lations.
The second most adopted approach treats the particles as individual bodies. Orig-
inating from particle-particle interactions based on constitutive relations of contact
mechanics [4–6], the discrete element method has been widely used in the field of
particle scale research [7]. Pioneers of this approach, Cundall and Strack introduced
an explicit numerical scheme to practice the granular dynamics by defining particles’
interactions over the contact networks and solving for particles’ motion under the
state of force balance equilibrium [8]. The integration of particle motion and energy
to the macroscopic behavior of the assembly, provides the required understanding
of overall behavior of the confined material [9]. The main advantage of this method-
ology is the capability of presenting broad information about the micro-structural
arrangement of the granular media. Although there exists computational challenges
to model a large number of particles system with discrete elements methods, advances
in simulation techniques enhance the implementation of this approach into the field
of multi-physics problems of granular systems.
Recently researchers also focus on multi-scale approaches to describe the macro-
scopic behavior of granular systems. Zheng and Cuitino implemented a quasi-
continuum approach to bridge the gap between micro and meso scale description
by using a discrete-continuum formulation of elastic-inelastic deformations occur-
Thermo-mechanical Behavior of Confined Granular Systems 43

ring in the post-rearrangement regime of consolidation of inhomogeneous granular

beds [10]. Since this approach provides the flexibility of storing individual particle
interactions in a FEM scheme, it provides the overall behavior of the entire body
without loosing critical information specific to microstructure. Koynov et al. pre-
sented a notable adaptation of this approach on the topic of powder compactions
for pharmaceutical purposes [11]. In this study we present a new methodology to
explore the family of multi-physics problems such as thermo-mechanical coupling.
The method is an extension of discrete element method that accounts for the effec-
tive modeling of heat conduction, and similar in spirit to early studies of Vargas and
McCarthy [12] and Feng et al. [9].
Current study incorporates early mathematical models that are developed for
conforming thermal contact of elastic, spherical surfaces [13–15]. These theoret-
ical models are validated though experimental studies [16–18]. Also there exists
studies that aim to relax some of the assumptions by focusing on elasto-plastic con-
tacts [19], or rough surfaces of non-conforming contact [20–22]. Recently the field
of granular matter gained importance in the light of understanding the correlation
between geometry, loading conditions and anisotropic microstructural arrangements
that determine the macroscopic behavior of compacted particulate system [23]. Gon-
zalez and Cuitino introduced a new formulation that account for the interplay of
nonlocal mesoscopic deformations characteristic of confined granular systems. In
the absence of the classical restriction of independent contacts of Hertz law, the
extended theory of nonlocal contact formulation provides predictive models at mod-
erate levels of deformation and high confinement [24]. In their study on effects of
packing grains by thermal cycling, Chen et al. [25] showed that thermal expansion,
due to the imposed thermal gradient, has significant effect on the rearrangement
of particle bed. Vargas and McCarthy focused on the problem of how the forces
supporting the grains are distributed under the effect of thermal expansion [26].
It is the purpose of this study to suggest that insight into the nature of thermo-
mechanical behavior of confined granular materials. We aim to discover the effects
of thermal and mechanical coupling at the particle level and implement the required
amendments to continuum level models. We present the system of governing equa-
tions, which define prescribed state of the assembly under steady state conditions, in
terms of heat and force transfer between the contacting particle pairs. Owing to the
fact that the nature of the problem leads to highly non-linear coupled equations, regu-
lar packing simplifies the problem and makes it mathematically traceable. Moreover
we consider the analogous problem from the perspective of the conventional contin-
uum mechanics approach. Practicing a thermo-elastic continuum model to simulate
the system, we focus on the effective mechanical and transport properties to account
for the unique characteristics of granular materials.
44 G. Küçük et al.

2 Particle Mechanics Approach

Our point of departure for the particle-scale description of thermo-elastic contact of

spherical smooth particles is to integrate the well-known theory of Hertzian deforma-
tion, [4], and heat conduction through the common interface of deformed particles in
contact, [13, 14]. Under steady state conditions, the total heat transferred to individ-
ual particle m from neighboring particles n and the total of forces acting on particle
m are zero,

Qm = Q mn = 0 (1)
nεNm

F =m
F mn nmn = 0 (2)
nεNm
x − xn
m
nmn = . (3)
xm − xn 

where nmn is the unit normal vector defined from centers of particle n to particle m.
xm and xn are the position of the particles.
Johnson identifies the elastic deformation of locally spherical particles that are
subject to a compression load by contact mechanics considerations in his book [27].
Small-strain deformation of conforming surfaces results in a flat circle of contact
area. Collinear contact force at this elastic contact of the particles m and n is defined
through Young’s moduli, E m and E n ; Poisson’s ratios, ν m and ν n ; particle radii, R m
and R n of particle m and n; and overlap, γ mn , between these particles,

4 mn mn 1/2 mn 3/2
F mn = E (R ) (γ ) (4)
3
where
 −1
1 1
R mn
= + n (5)
Rm R
 −1
1 − (ν m )2 1 − (ν n )2
E mn = + (6)
Em En
γ mn = R m + R n − xm − xn  . (7)

One particular effect of applied thermal load on the system of particles is the
change in radii due to thermal expansion. Similar to previous studies in the literature
[26, 28], in the present study, linear thermal expansion formulation is taken into
consideration.
  
R m = Rrme f 1 + α m T m − Trme f (8)
Thermo-mechanical Behavior of Confined Granular Systems 45

Here α m is the thermal expansion coefficient, Tr e f is the reference temperature and

Rrme f is the radius of particle at the reference temperature. Due to the dependence
of contact geometry on the nature of thermo-mechanically coupled problem, it is
expected to capture a distribution of contact area formation throughout the compacted
medium.
There has been considerable research on thermal-contact models. The major heat
transfer mechanisms in compacted particle beds consist of conduction through solid,
conduction through the contact area between two touching particles, conduction
to/from interstitial fluid, heat transfer via convection, radiation between particle sur-
faces, radiation between neighboring voids [12]. For a system of granular media
where the thermal conductivity of the solid particles is much larger than the intersti-
tial medium, the driving mechanisms for the heat transfer are the first two. Concern-
ing the problem of thermally-assisted compaction of spherical particles in vacuum,
we focus on the thermal contact models that consider the conduction through solid
particle and through the contact area between two touching particles.
Analytical solution of the heat conduction through the solid phase of ordered
spherical particles has been proposed by Chan and Tien [14] and Kaganer [15].
Moreover the problem of heat transfer regarding the compaction of particles that
are in or nearly in contact is deeply investigated by Batchelor and O’Brien [13].
In an attempt to find the approximate effective thermal conductivity of ordered and
randomly packed granular beds, Batchelor and O’Brien discussed the heat flux across
the flat circle of contact between smooth, conforming, and elastic particles. In this
study we adopt Batchelor and O’Brien’s model for predicting the heat conductance,
which is the ability of two touching surfaces to transmit heat through their mutual
interface. Heat flux across the contact area of two spherical smooth particles is given
as
Q mn = 2a mn k mn (T m − T n ) (9)

where k mn is the arithmetic mean of the thermal conductivities of two conforming

particles, and a mn is the Hertzian contact area.
 
1 1 1 −1
k mn
= + n (10)
2 km k
a mn = γ mn R mn (11)

The total heat flow to an individual particle, Eq. (1), is calculated by adding the
heat flow at each contact of the particle between its neighboring particles Eq. 9. As
discussed by the thermal contact models introduced in the literature [13, 14], Eq. (1)
requires that at each contact of the individual particle, the temperature is equal to the
temperature calculated at the center of the particle. In other words, the temperature
does not very significantly within the particle, which also imposes that the contact
conductance at the interface of conforming particles is relatively smaller than the
heat conductance within the particle.
46 G. Küçük et al.

2k mn a mn
1 (12)
k mn A/R m

where A is the cross sectional area, A = π(R m )2 and Eq. (12) defines the state of
Biot number much less than 1. This assertion is applied by several authors in earlier
studies [12, 29]. The condition of a mn  R mn is also enforced by the assumption of
small-strain deformation of elastic bodies in contact.

2.1 Simulation Configuration

Referring to the previous experimental studies on regular and random packings of

granular media, Walton points out that although the regular packing models are
founded on extreme assumptions, they are capable of capturing vast majority of
the characteristics of a real granular media [30]. In the present study we consider
a simple cubic packing of identical elastic spheres, which are constrained between
parallel planes of infinite extent. Compression load, temperature gradient are applied
along the major and finite direction. Stress and heat flux are defined to only depend
on externally applied thermal and mechanical loads, and weight of the particles is
neglected. For such regular packings each layer of arrangement is isothermal normal
to the direction of applied load. Also, since these transversely oriented particles
are, at most, point contact, for each individual particle there is only one pair of
contact area aligned with the direction of applied thermal and mechanical load.
Due to the symmetry of the problem, it is sufficient to consider a single column of
square cross-section containing the longitudinally compressed spheres together. This
concept is similar, in spirit, to the work of Chan and Tien [14], who proposed the
effective thermal resistance, and to the work of Walton [30] who presented a method
to calculate the effective elastic moduli of such packings. The above description for
the specified granular media, which is under thermally-assisted compaction, can be
modeled as a chain of elastic particles, seen in Fig. 1.

Fig. 1 Sketch of initial configuration

Thermo-mechanical Behavior of Confined Granular Systems 47

2.2 Wall-Particle Interaction

Given such a setting that the chain of particles is compressed between two parallel
walls, which are maintained at different temperatures, the wall-particle interaction
is one of the important factors of this problem. In this study, analogous to ghost-cell
method, the contact between boundary particle and adjacent wall is simulated as the
contact between a boundary particle and a ghost particle. Based on the rigid wall
assumption, ghost particle and boundary particle are set to have the same material
properties and radius. The temperature difference between the ghost particle and the
wall surface is the same as the temperature deviation between the wall surface and the
boundary particle. The boundary wall is assumed to be located in the midst of these
symmetrically deformed particles. The temperature difference, overlap and contact
area are formulated in Eqs. (13)–(15), where T ws and T w refer to the temperature at
the wall surface and wall temperature, respectively. Subscript g is used to indicate
the ghost particle.

ΔT mg = 2(T m − T ws ) (13)
γ mg = 2(R m − ||xm − xws ||) (14)
R m 1/2
a mg = (γ mg ) (15)
2
At the boundary surfaces, heat transfer between the boundary particle to the wall
can be expressed by two main heat transfer mechanisms, (i) heat is conducted over
flat circle of contact between the particle and adjacent wall surface; (ii) convective
heat transfer, which is dependent on walls’ convection coefficient of h w , takes place
between the wall surface and the wall.

Q m−ws = k m a mg ΔT mg (16)
Q ws−w
= −h w π(a ) (T
mg 2 ws
−T )
w
(17)

The temperature at the wall surface can be obtained for the equilibrium of Eqs. (16)
and (17). The final set of equations, which define the wall-particle interaction, are
the following:

 4k m T m + h w πa mg T w 
Q = 4ka mg T m − (18)
4k + h w πa mg
m
E
F= R m γ mg . (19)
3(1 − (ν m )2 )
48 G. Küçük et al.

3 Conventional Continuum Mechanics Approach

While the particle mechanics approach aims to elucidate the formation and the evo-
lution of the microstructure of the granular media at particle-level, there has been
considerable research directed towards understanding macroscopic behavior of com-
pacted materials. Some of the early work on theoretical modeling of transport prop-
erties are devoted to the estimation of thermal, and electrical conductivity, elastic,
plastic mechanical properties of ordered and disordered arrangements. Originating
from the pair interactions between particles, the macroscopic properties are obtained
using various homogenization techniques and postulating continuum constitutive
laws [31]. In this study, we consider a continuum system that mimics the particle
level description for small strain thermoelasticity, which incorporates the proposed
effective mechanical and thermal properties for granular beds under compaction.
Governing field equations of motion and energy are the following

div (σ ) = 0 (20)
 
div k grad (T ) = 0 (21)

where the Cauchy’s stress, σ , is formulated as combination of classical linear elas-

ticity theory and simple linear thermal expansion.

σ = −λtr(ε)I − 2με + (3λ + 2μ)α(T − Tr e f )I (22)

For the basic problem of one dimensional steady state thermoelasticity of contin-
uum media, where the body forces are neglected, the solution depends linearly on
elastic constants (λ, μ), thermal expansion and conduction coefficients, compaction
strain and thermal gradient. Since ε22 = ε33 = 0 holds, ε11 is referred as ε. Equations
of motion and energy Eqs. (20) and (21) can be rewritten as

σ = −(λ + μ)ε + α(3λ + 2μ)(T − Tr e f ) , (23)

∂T
q=k . (24)
∂x
Effective mechanical properties of granular beds are of heavy interest in many
theoretical studies. Some of these include: calculation of the principal elastic modulus
of vertical compression of spherical particles without any lateral extension [30];
derivation of the finite and incremental elasticity of random packing of identical
particles using energy methods [32]; enhancement of the derived formulas based on
the pressure dependence of the elastic moduli of granular packings [2, 3].
In this study we extend the effective medium theory with the thermal contact
model principles by incorporating the particle interactions to account for the local
field effects. We re-formulate the effective elastic properties and effective thermal
conductivity accordingly, and implement these parameters in continuum mechanics
Thermo-mechanical Behavior of Confined Granular Systems 49

model. The effective mechanical properties can be expressed in terms of the applied
stress, σ , and bulk material properties [32],

μ E 1 2E
Cn = 4 =4 = (25)
1−ν 2(1 + ν) 1 − ν 1 − ν2
3  6π σ 1/3
λ̃ + 2μ̃ = Cn (φs Z )2/3 (26)
20π Cn
1  
2/3 6π σ 1/3
3λ̃ + 2μ̃ = Cn (φs Z ) (27)
4π Cn

where Cn is named as stiffness of the system, φs is the packing fraction, and Z

is the coordination number. Effective mechanical properties, λ̃, μ̃, and effective
thermal conductivity, k̃, are implemented in continuum mechanics model and listed as
conventional continuum solution. As a simple application of this theory, we consider
a case of particles’ chain that is compacted by a ratio of ε, under the effect of a thermal
gradient of T2w − T1w . The expression that defines the stress evaluation through the
chain is found as
 3 1/2  3 T w + T1w 3/2
σ = φs ZCn 2
|−ε − α( 2 − Tr e f )| . (28)
32π 5 2

Effective thermal conductivity, k̃, of the granular bed is substantially sensitive

to the thermal and the elastic properties of individual particle. Regarding ordered
cubic packing configuration, it is known that thermal contact models provide accu-
rate results in estimating steady and average temperature profiles [33]. Three major
analytical solutions in literature, by Batchelor and O’Brien [13], Chan and Tien [14],
Kaganer [15] and Siu and Lee [34], are proposed to determine the effective thermal
conductivity. Since Batchelor and O’Brien’s [13] solution stays in remarkable agree-
ment with the particle mechanics results in terms of heat transferred through the
chain, we adopted this solution for effective thermal conductivity coefficient in our
continuum mechanics approach. This comparison is shown in Fig. 2, where PMA and
CMA refer to particle mechanics approach, and conventional continuum mechanics
approach, respectively.

 6σ 1/3
k̃ B&O = k
Cn
 6σ 1/3
k̃ C&T = 0.9454k
Cn
 6σ 1/3
k̃ S&L = 0.8278k
Cn
50 G. Küçük et al.

0.4
PMA
CMA, Chan and Tien’s Sol’n.
0.35 CMA, Bathcelor and O’Brien’s Sol’n.
CMA, Siu and Lee’s Sol’n.

0.3

0.25
Q (W)

0.2

0.15

0.1

0.05
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
ε

Fig. 2 Comparison of continuum solutions adopting different thermal contact models with respect
to particle mechanics solution. Heat versus compaction strain, ε, is evaluated at T2w − T1w = 600 K

4 Results

According to the Hertz theory [35], the collinear contact force between the elastically
compressed particles is a nonlinear function of the overlap, which is generated under
the effect of the external load, between the particles. For the case of thermally-
assisted compaction of granular system of particles, this dependency is altered under
the effect of applied thermal gradient. Figure 3 shows the ratio of force, needed to
compress the system, in particle mechanics approach to the force in conventional
continuum mechanics approach. CMA significantly overestimates the thermal stress
within the chains system, particularly for the range of high thermal gradient and
low mechanical load. Moreover concerning the highly compacted systems CMA
underestimates PMA solution for the system of particles by 10 %.
Similar to compaction force comparison, conventional continuum solution pre-
dicts higher heat transferred values for analogous particles system. It is also shown
in Fig. 4 that as the packing density of the deformed particles system is increased,
conventional continuum mechanics solution becomes more effective in estimating
the particle-level solution.
Under three difference compaction strains, 2.5, 5, 10 %, the effect of wall-particle
interaction is examined through the chain of particles by imposing a thermal gradient
of 300 K between the two boundary walls of the system. For each case, wall heat
transfer coefficient, h w , is ranged from 1 to 107 W/m2 K. Figure 5 indicates the two
limiting cases of perfect insulating and perfect conducting walls.
Thermo-mechanical Behavior of Confined Granular Systems 51

1000 1.1

900 FPMA/ FCMA 1.05

800 1

0.95
700
0.9
T 2w − T 1 (K)

600
0.85
w

500
0.8
400
0.75
300
0.7
200
0.65
100
0.6

0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045

ε

Fig. 3 Comparison of force calculated in PMA and CMA under varying thermal and mechanical
loading conditions

1000 1.04
Q /Q
PMA CMA
900 1.02

800 1

700 0.98

0.96
T 2 − T1 (K)

600
0.94
w

500
0.92
w

400
0.9
300
0.88
200
0.86
100
0.84

0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045

ε

Fig. 4 Comparison of heat calculated in PMA and CMA under varying thermal and mechanical
loading conditions

4.1 Role of Thermal Expansion

Systems of granular materials with different thermal expansion properties respond in

various re-arrangements to a particular thermal and mechanical load. The following
52 G. Küçük et al.

0.014

0.012

Heat [Q / (k R0ΔT)] 0.01

Nondimensional

0.008

0.006

0.004

ε = 0.025
0.002 ε = 0.05
ε = 0.1

0
0 1 2 3 4 5 6
10 10 10 10 10 10 10
2
h (W/m K)
w

Fig. 5 Correlation between heat and wall conductance at different compaction strains

numerical experiment compiles the results of three different homogeneous system

of particles: SS304, Aluminum, and Teflon with different thermal expansion values,
17.3 10−6 1/K, 23.6 10−6 1/K, 250 10−6 1/K, respectively. In the above-mentioned
cases of comparison, the chain is compacted to 2.5 % of the initial length and a total
thermal gradient of 300 K is applied between the two boundary walls. Alterations
in displacement of each particle due to increase of thermal stress can be traced to
unveil the effect of thermal expansion coefficient on the system of particles under
thermally-assisted compaction. In Fig. 6 the displacement of each particle is divided
by the total mechanical deformation applied on the system. The non-dimensional
displacement of the particle in contact with the fixed boundary is listed as 0, whereas
the one in contact with the heated moving boundary wall is 1.
While reaching to equilibrium the two dominant mechanisms, thermal and
mechanical stresses, induce a nonlinear distribution of displacements, which is a
unique characteristic of particulate systems. This deviation from linear continuum
solution is enhanced for systems with high thermal expansion property.

4.2 Role of Applied Mechanical Load

Under the effect of a modest thermal gradient of 300 K, three different mechanical
loading conditions, 1, 2, 5–10 %, are compared in Fig. 7. In order to compare the
coupled effect of thermal gradient and mechanical deformation, two extreme cases for
wall-particle interactions are also considered. In case of perfect insulating walls, h w is
Thermo-mechanical Behavior of Confined Granular Systems 53

Displacement/Total Deformation [ ui / (ε L0) ]

SS304
Aluminum
Teflon
−0.2

−0.4

−0.6

−0.8

−1

−1.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Nondimensional Position [ xi /L0 ]

Fig. 6 Relative displacement of each particle within the chain at ε = 0.025 and T2w − T1w = 300 K

0
ε=0.01, hw=1
Displacement/Total Deformation [ ui / (ε L0 ) ]

−0.1 ε=0.01, hw=107 W/m2 K

ε=0.025, hw=1
−0.2
ε=0.025, hw=107 W/m2 K

−0.3 ε=0.1, hw=1

ε=0.1, hw=107 W/m2 K
−0.4

−0.5

−0.6

−0.7

−0.8

−0.9

−1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Nondimensional Position [ x i / L 0 ]

Fig. 7 Non-dimensional displacement versus initial position of the particles under different
mechanical loads, with varying heat convection coefficients of the boundary walls

assumed to be 1 W/m2 K. This particular condition is a simulation of pure mechanical

loading where we expect to have linear distribution of non-dimensionalized positions’
of particles within the chain system. On the other hand the case of h w = 107 W/m2 K
simulates the condition of perfectly conducting walls.
54 G. Küçük et al.

Regarding the system of SS304 spherical particles’ chain Fig. 7 indicates that
nonlinearity in distribution of displacements is more dominant for low mechanical
loading cases.

4.3 Role of Thermal Gradient

A recent experimental study on silos of spherical glass particles showed that thermal
cycling, and the difference in thermal expansion properties of the granular material
with respect to its container, significantly affect the packing fractions of granular
materials in the absence of mechanical compaction [25]. In the current study we
focus on the active interval where thermal gradient acts as a dominant mechanism
compared to mechanical deformation. A chain of spherical particles is gradually
consolidated up to a compaction strain of 5 % of their initial length, while thermal
gradient between the two boundary walls is increased to 1000 K.
The ratio of the displacements calculated in PMA to CMA indicates a discrep-
ancy between these two approaches. In Fig. 8 the maximum difference between par-
ticle mechanics approach and conventional continuum mechanics approach is traced.
Under the effect of low mechanical deformation, such as ε < 0.02, and high ther-
mal gradient conditions the continuum solution overestimates the actual position
of particles up to 40 % of the solution provided by particle mechanics approach.
The difference between these two solutions diminishes as the packing density of the
granular system increases.

0.95
Max (UPMA / UCMA)

0.9

0.85

0.8

0.75

0.7

0.65

0.06
1000
800 0.04
600
w w
400
200
0.02 ε
T 2 −T 1 0 0

Fig. 8 Maximum difference between the calculated displacements in particle mechanics and in
conventional continuum mechanics approaches
Thermo-mechanical Behavior of Confined Granular Systems 55

5 Conclusion

In this study we present a numerical model to describe the thermo-mechanical behav-

ior of a confined granular system by adopting a detailed description at the particle
level. We integrate thermal-contact and Hertzian deformation models to simulate the
temperature and displacement of consolidated granular medium. One-dimensional
model provides an opportunity to unveil the relation between two dominant mecha-
nisms affecting the thermal and mechanical equilibrium of the particulate systems. In
order to capture the actual physical conditions, we consider wall-particle interactions
ranging from perfect insulating to perfect conducting walls.
The numerical results indicate that integration of thermal deformation with the
elastic contact models induces the incongruity seen in mechanical deformation based
compaction models. The coupled phenomena introduce highly nonlinear system
of equations, and it imposes variation in contact areas and nonlinear temperature
distribution within the particulate material. This effect is enhanced for particles with
larger thermal expansion coefficient. It appears that the critical regime, where the
nonlinearity due to thermo-mechanical coupling becomes more dominant, is low
mechanical load and high thermal gradient conditions.
As a multi-physics problem, thermally-assisted compaction shows a significant
dependence on the thermal expansion of the particles. Discrete solution based on the
particle mechanics approach that adopts the thermal contact model, carries out this
dependence and the nonlinearity enhanced by thermal strains, successively. Despite
the fact that effective medium theory improves the continuum solution to a large
extend, it fails to capture the characteristics of multi-physics of the problem, partic-
ularly for the cases of low thermal gradient coupled with high mechanical load.
Looking toward to future, we are now in a position to address a variety of impor-
tant questions, such as; (i) what can be a further improvement in effective medium
theory that also account for an effective thermal expansion coefficient depending not
only on the bulk properties but also loading conditions of the compacted granular
assembly? (ii) what is the role of uneven distribution of contact areas and nonlinear
temperature distribution on formations of heterogeneous force and heat networks
within the concept of the micro-structural arrangement of granular system to macro-
scopic behavior of the thermally-assisted compacted end product?

Acknowledgments This work has been partially supported by U.S. Army ARDEC grant under
the project titled as: Multifunctional Nanomaterials: Processing, Properties, and Applications. The
authors would also like to acknowledge the support provided by the National Science Foundation
Engineering Research Center for Structured Organic Particle Systems (C-SOPS).
56 G. Küçük et al.

References

1. Vargas-Escobar, W. L. (2002). Discrete modeling of heat conduction in granular media. Ph.D.

thesis, University of Pittsburgh.
2. Makse, H. A., Gland, N., Johnson, D. L., & Schwartz, L. M. (1999). Why effective medium
theory fails in granular materials. Physical Review Letters, 83(24), 5070–5073.
3. Makse, H. A., Gland, N., Johnson, D. L., & Schwartz, L. (2001). The apparent failure of
effective medium theory in granular materials. Physics and Chemistry of the Earth, Part A:
Solid Earth and Geodesy, 26(1), 107–111.
4. Hertz, H. (1881). On the contact of elastic solids. Journal für die reine und angewandte Math-
ematik, 92(156–171), 110.
5. Mindlin, R. D. (1949). Compliance of elastic bodies in contact. Journal of Applied Mechanics,
16.
6. Mindlin, R. D., & Deresiewicz, H. (1953). Elastic spheres in contact under varying oblique
forces. Journal of Applied Mechanics, 20.
7. Zhu, H. P., Zhou, Z. Y., Yang, R. Y., & Yu, A. B. (2007). Discrete particle simulation of
particulate systems: Theoretical developments. Chemical Engineering Science, 62(13), 3378–
3396.
8. Cundall, P. A., & Strack, O. D. L. (1979). A discrete numerical model for granular assemblies.
Geotechnique, 29(1), 47–65.
9. Feng, Y. T., Han, K., Li, C. F., & Owen, D. R. J. (2008). Discrete thermal element modelling
of heat conduction in particle systems: Basic formulations. Journal of Computational Physics,
227(10), 5072–5089.
10. Zheng, S., & Cuitino, A. M. (2002). Consolidation behavior of inhomogeneous granular beds
of ductile particles using a mixed discrete-continuum approach. Kona, 20, 168–177.
11. Koynov, A., Akseli, I., & Cuitiño, A. M. (2011). Modeling and simulation of compact strength
due to particle bonding using a hybrid discrete-continuum approach. International Journal of
Pharmaceutics, 418(2), 273–285.
12. Vargas, W. L., & McCarthy, J. J. (2001). Heat conduction in granular materials. AIChE Journal,
47(5), 1052–1059.
13. Batchelor, G. K., & O’Brien, R. W. (1977). Thermal or electrical conduction through a granular
material. Proceedings of the Royal Society of London. A. Mathematical and Physical Sciences,
355(1682), 313–333.
14. Chan, C. K., & Tien, C. L. (1973). Conductance of packed spheres in vacuum. Journal of Heat
Transfer (United States), 95(3).
15. Kaganer, M. G. (1966). Contact heat transfer in granular material under vacuum. Journal of
Engineering Physics, 11(1), 19–22.
16. Hadley, G. R. (1986). Thermal conductivity of packed metal powders. International Journal
of Heat and Mass Transfer, 29(6), 909–920.
17. Nozad, I., Carbonell, R. G., & Whitaker, S. (1985). Heat conduction in multiphase systems—
II: Experimental method and results for three-phase systems. Chemical Engineering Science,
40(5), 857–863.
18. Shonnard, D. R., & Whitaker, S. (1989). The effective thermal conductivity for a pointcontact
porous medium: An experimental study. International Journal of Heat and Mass Transfer,
32(3), 503–512.
19. Sridhar, M. R., & Yovanovich, M. M. (1996). Elastoplastic contact conductance model for
isotropic conforming rough surfaces and comparison with experiments. Transactions-American
Society of Mechanical Engineers Journal of Heat Transfer, 118, 3–9.
20. Bahrami, M., Yovanovich, M. M., & Culham J. R., et al. (2005). A compact model for spherical
rough contacts. Transactions-American Society of Mechanical Engineers Journal of Tribology,
127(4), 884.
21. Fletcher, L. S. (1988). Recent developments in contact conductance heat transfer. ASME Trans-
actions Journal of Heat Transfer, 110, 1059–1070.
Thermo-mechanical Behavior of Confined Granular Systems 57

22. Majumdar, A., & Tien, C. L. (1991). Fractal network model for contact conductance. Journal of
Heat Transfer (Transactions of the ASME (American Society of Mechanical Engineers), Series
C;(United States), 113(3).
23. Majmudar, T. S., & Behringer, R. P. (2005). Contact force measurements and stress-induced
anisotropy in granular materials. Nature, 435(7045), 1079–1082.
24. Gonzalez, M., & Cuitiño, A. M. (2012). A nonlocal contact formulation for confined granular
systems. Journal of the Mechanics and Physics of Solids, 60(2), 333–350.
25. Chen, K., Cole, J., Conger, C., Draskovic, J., Lohr, M., Klein, K., et al. (2006). Granular
materials: Packing grains by thermal cycling. Nature, 442(7100), 257–257.
26. Vargas, W. L., & McCarthy J. J. (2007). Thermal expansion effects and heat conduction in
granular materials. Physical Review E, 76(4), 041301.
27. Johnson, K. L. (1987). Contact mechanics. Cambridge University press.
28. Lu, Z., Abdou, M., & Ying, A. (2001). 3d micromechanical modeling of packed beds. Journal
of nuclear materials, 299(2), 101–110.
29. Siu, W. W. M., & Lee, S. H.-K. (2004). Transient temperature computation of spheres in
three-dimensional random packings. International Journal of Heat and Mass Transfer, 47(5),
887–898.
30. Walton, K. (1975). The effective elastic moduli of model sediments. Geophysical Journal of
the Royal Astronomical Society, 43(2), 293–306.
31. Markov, K. Z. (2000). Elementary micromechanics of heterogeneous media. In Heterogeneous
media (pp. 1–162). Springer.
32. Norris, A. N., & Johnson, D. L. (1997). Nonlinear elasticity of granular media. Transactions-
American Society of Mechanical Engineers Journal of Applied Mechanics, 64, 39–49.
33. Vargas, W. L., & McCarthy, J. J. (2002). Stress effects on the conductivity of particulate beds.
Chemical Engineering Science, 57(15), 3119–3131.
34. Siu, W. W. M., & Lee, S. H.-K. (2000). Effective conductivity computation of a packed bed
using constriction resistance and contact angle effects. International Journal of Heat and Mass
Transfer, 43(21), 3917–3924.
35. Landau, L. D., & Lifshitz, E. M. (1959). Theory of elasticity. Pergamon Press.
Elastomeric Gels: A Model
and First Results

Mariarita de Luca and Antonio DeSimone

Abstract An elastomeric gel is a cross–linked polymer network swollen by a

solvent. Computational models of gels need to resolve the strong coupling between
the diffusion of the solvent and the deformation of the elastic network. We present
here a continuum mechanics model to describe the gel deformation and the coupled
fluid permeation in the polymeric network, and the first results we have obtained
with it. These consist of numerical simulations of two basic experiments: the free
swelling deformation of a dry specimen and an indentation test performed on a
swollen sample.

1 Introduction

An elastomeric gel is a cross–linked polymer network swollen by a solvent. Com-

putational models of gels need to resolve the strong coupling between the diffusion
of the solvent and the deformation of the elastic network.
In the literature there are many approaches to study the evolution of a system
composed by a polymeric structure plus solvent. Hong and coworkers [9] consider
the gel as a bulk soft material characterized by the free energy function proposed by
Flory and Rehner in 1943 [8]. Such free energy, derived on the basis of statistical
mechanics, takes into account the entropy change induced by stretching the polymer
network, and the enthalpy of mixing of polymer and solvent. Hong and others also
introduced a dynamic model for the diffusion of solvent molecules inside the gel
based on the assumption that the solvent molecules self diffuse inside the gel and
there is no macroscopic flow [9].
Following the same theory, Kang and Huang [12] first developed a finite element
model to compute the equilibrium swelling deformation of a gel in contact with a

M. de Luca · A. DeSimone (B)

SISSA, Via Bonomea 265, 34136 Trieste, Italy
e-mail: [email protected]
M. de Luca
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 59

K. Weinberg and A. Pandolfi (eds.), Innovative Numerical Approaches
for Multi-Field and Multi-Scale Problems, Lecture Notes in Applied
and Computational Mechanics 81, DOI 10.1007/978-3-319-39022-2_4
60 M. de Luca and A. DeSimone

solvent constrained to a fixed support. Subsequently, Zhang et al. [17] proposed a

full three dimensional model to simulate the diffusion of the solvent inside the gel
and the corresponding deformation.
Another approach is to consider the gel composed by a porous elastic material
and a fluid (solvent) flowing through its pores. This model combines the elasticity of
the porous structure with a mass transport model in the porous medium (Darcy Law),
obtaining the Biot model [2]. In its original formulation, the Biot model considers
only the pore pressure in a material element, defined as the pressure in an hypothetical
reservoir which is in equilibrium with this element, as the mechanism regulating the
flux of fluid between the reservoir and the material element [16]. For such reason it
is suitable to describe quasi-static processes, but it is not adequate for swelling gels,
then it needs to be extended in order to be able to correctly capture all the phenomena
that are interesting for gel applications. Following this idea, Murad et al. [14] have
extended the Biot model in the framework of hybrid mixture theory to develop a
theory for swelling porous media, in which the variation of the chemical potential is
the mechanism regulating the swelling and draining processes.
In this work we start from the idea proposed in [3], and we use a continuum
mechanics approach to describe the gel deformation and the coupled fluid per-
meation in the polymeric network. We present here the first preliminary results
obtained with our model, which has been implemented in the software package
Comsol Multiphysics® , and used to simulate two basic experiments. These are the
free swelling deformation of a dry specimen and an indentation test performed on a
swollen sample.

2 Governing Equations

In the following sections we present the governing equations of the model, the balance
of mass and the balance of linear momentum and we recall the kinematics needed to
describe the model.

2.1 Kinematics

Let Ω0 ⊂ R3 be the initial placement, at time t = 0, of the body. Points X ∈ Ω0 corre-

spond to both fluid and solid particles and it is not possible to distinguish among them,
hence the polymer-solvent mixture is treated as a single homogeneus continuum.
The vector function f : Ω0 × [0, T ] → R3 is the motion of the body, mapping points
X ∈ Ω0 at time t ∈ [0, T ] to points x ∈ Ωt ⊂ R3 , where Ωt = f(Ω0 , t). We write
x = f(X, t), F(X, t) = Grad f(X, t) = ∂f(X, t)/∂X for the deformation gradient,
subject to the orientation constraint J = det F > 0, v(X, t) = ḟ(X, t) = ∂f(X, t)/∂t
the velocity, and L = grad v = ∂v(f −1 (x, t), t)/∂x = ḞF−1 the spatial gradient of
the velocity.
Elastomeric Gels: A Model and First Results 61

As in [3], we consider a multiplicative decomposition of the deformation gradient

into a “swelling” and an “elastic” part:

F = Fe Fs , with Fs = λs I, λs > 0, (1)

where the swelling part Fs is assumed isotropic, with λs the swelling stretch. We refer
te reader to [5, 6] for further examples of using the multiplicative decomposition of
the deformation gradient in nonlinear models of polymers and of phase-changing
elastomers.
At a fixed time t ∗ , Fs (X, t ∗ ) represents the distortion of the body only due to
swelling, while Fe (X, t ∗ ) is the subsequent stretching and rotation of the corre-
sponding swollen network structure, representing the mechanical elastic distortion.
Then J e = det Fe , J s = det Fs , and J = J e J s . The right Cauchy–Green strain
tensor is
C = FT F = (Fe Fs )T Fe Fs = FsT FeT Fe Fs = FsT Ce Fs , (2)

where
Ce = FeT Fe (3)

is the elastic part.

The velocity gradient

L = Ḟ F−1 = (Fe˙Fs ) (Fe Fs )−1 = (Ḟe Fs + Fe Ḟs ) Fs−1 Fe−1

= Ḟe Fe−1 + Fe Ḟs Fs−1 Fe−1 = Le + Fe Ls Fe−1 , (4)

where

Le = Ḟe Fe−1 (5)

Ls = Ḟs Fs−1 . (6)

It is convenient to compute the time derivative of the determinants

J̇ = J F−T : Ḟ = J Ḟ F−1 : I = J trL (7)

J˙s = J s Fs−T : Ḟs = J s Ḟs Fs−1 : I = J s trLs , (8)

then
1 ˙s s−1
Ls = J J I. (9)
3
With reference to Fig. 1, F describes the deformation from the reference state to
the current state, while Fd is the deformation gradient from the dry state (when there
is no solvent) to the current state, F and Fd relate through F0 which is the spontaneous
62 M. de Luca and A. DeSimone

Fig. 1 Cartoon of dry, reference and current configurations

deformation that a dry body undergoes when immersed in a solvent (free swelling
deformation) [13]
Fd = F F0 , (10)

where
F0 = λ0 I. (11)

The free swelling deformation is isotropic, characterized by the free swelling stretch
λ0 . Then
Cd = Fd T Fd = (F F0 )T F F0 = F0 T FT F F0 = F0 T C F0 . (12)

2.2 Conservation of Mass

We call current concentration

n(x, t)
cc (x, t) = , (13)
V
Elastomeric Gels: A Model and First Results 63

the number of solvent moles n(x, t) absorbed by the solid network per unit current
volume V . The concentration field is a scalar function cc : Ωt × [0, T ] → R. The
mass of solvent in a region Rt ⊂ Ωt is

M(t) = cc (x, t) dV. (14)
Rt

The referential concentration, namely, the number of solvent moles absorbed by

the solid network per unit reference volume V0 is

n(X, t)
c(X, t) = , (15)
V0

where X = f −1 (x, t) and c : Ω0 × [0, T ] → R.

If we consider a fixed region R0 ⊂ Ω0 the conservation of mass implies
 
M(t) = c(X, t) dV0 = cc (x, t) dV, (16)
R0 Rt

using the volume transformation dV = J dV0 , then

 
c(X, t) dV0 = cc (x, t) J dV0 , (17)
R0 R0

and
c = J cc . (18)

Similarly,
n
cd = (19)
Vd

is the concentration referred to the dry volume of the body and

cd = J0 J cc = J0 c, (20)

where dV0 = J0 dVd .

If n0 is the number of solvent moles absorbed by the polymer network during the
spontaneous deformation (free swelling),
n0
c0 = (21)
V0
64 M. de Luca and A. DeSimone

is the concentration in the free swollen reference state per unit reference volume,
while n0
cd0 = (22)
Vd

is the concentration in the free swollen reference state per unit dry volume.
In our model, we impose the constraint of “molecular incompressibility”. This
means that the increase in volume of the body is only due to the solvent that enters
the polymer network. The increase in volume associated with the spontaneous free
swelling deformation described by F0 is

J0 = 1 + ν cd0 , (23)

where ν is the molar volume of the solvent.

Any deformation from the reference configuration induces a change in volume
described by J = J e J s , where J s accounts for the swelling part of the deformation.
Then the “molecular incompressibility” constraint reads

J s = J0 −1 + νc, or J s J0 = 1 + ν J0 c, (24)

because V0 = Vd + νn0 , V = V0 + ν(n − n0 ) = Vd + νn, and dividing all the terms

by Vd we obtain (24).
The global balance of liquid mass over a convected region Rt with boundary
measure dA in R3 is:
 
d
cc (x, t) dV = − jc (x, t) · n dA, (25)
dt Rt ∂Rt

where jc is the relative solvent flux through the boundary of Rt and n is the outward
unit normal to ∂Rt .
Using the formulas for surface and volume change

dV = J dV0 , n dA = JF−T n0 dA0 , (26)

we obtain  
d
J cc (x, t) dV0 = − jc (x, t) · J F−T n0 dA0 , (27)
dt R0 ∂R0

and  
d
J cc (x, t) dV0 = − JF−1 jc (x, t) · n0 dA0 . (28)
dt R0 ∂R0

Then, applying the divergence theorem and switching the derivative on the left hand
side with the integral sign we have:
 
d
c(X, t) dV0 = − Div j(X, t) dV0 , (29)
R0 dt R0
Elastomeric Gels: A Model and First Results 65

where
j(X, t) = JF−1 jc (x, t) (30)

is the Piola transformation of the relative solvent flux.

Finally, due to the arbitrariness of R0 , the local form of the balance of solvent
mass in the reference region Ω0 is

dc
= −Div j. (31)
dt

2.2.1 Polymer Fraction

We introduce the variable φ, the volume fraction occupied by the solid network in
the current configuration, defined as

Vd
φ= , (32)
V
then
V Vd + νn n
φ −1 = = =1+ν = 1 + νcd = 1 + ν J0 c = J0 J s . (33)
Vd Vd Vd

In the reference configuration Ω0 , the corresponding polymer volume fraction is

Vd
φ0 = = J0 −1 (34)
V0

and combining Eqs. (33) and (34) we have

φ = J s−1 φ0 . (35)

From (33) we can derive the concentration

1
c= (φ −1 − 1), (36)
ν J0

and its time derivative

1 φ̇ φ0 φ̇
ċ = − = (37)
ν J0 φ 2 ν φ2

Inserting relation (37) in Eq. (31) we have the mass balance with the polymer fraction
as independent variable:
φ0 φ̇
− = −Div j. (38)
ν φ2
66 M. de Luca and A. DeSimone

It is useful to define the fluid volume fraction to be able to compare our model
with the standard poro-elastic one

Vf
φf = 1 − φ = , (39)
V

where V f = V − Vd is the solvent volume. In the reference configuration Ω0

f
f V0
φ 0 = 1 − φ0 = , (40)
V0
f
where V0 = V0 − Vd is the fluid volume enclosed by the solid elastic network in the
reference configuration, then thanks to Eqs. (35), (39) and (40) we have

f
φ f = 1 − J s−1 (1 − φ0 ), (41)

and
f
Φ f = J s − 1 + φ0 , (42)

where Φ f = J s φ f = V f /V0 is the fluid fraction referred to the reference volume.

We remark that in classical poroelastic models, φ f is the “porosity” and it varies with
the deformation following the law (41) as described in [7], and

φf φf
e= = (43)
1 − φf φ

is the “void ratio”.

2.3 Balance of Linear Momentum

The balance of linear and angular momentum are unchanged with respect to the usual
expressions. Time scales associated with fluid diffusion are ususally considerably
larger then those associated with wave propagation, hence inertial effects can be
neglected. The body is subjected to volume forces b and surface forces t, where,
denoting by n the outward unit normal to ∂Ωt , we have

t(n) = T n, (44)

with T the Cauchy stress tensor.

Elastomeric Gels: A Model and First Results 67

The balance of linear momentum in a convected region Rt ⊂ Ωt reads

 
t(n) dA + b dV = 0, (45)
∂Rt Rt

or, using (44),  

T n dA + b dV = 0. (46)
∂Rt Rt

By using the area and volume transformation formulas (26) we obtain the balance
of linear momentum in a fixed region R0 ⊂ Ω0
 
J T F−T n0 dA0 + J b dV0 = 0. (47)
∂R0 R0

Thanks to the divergence theorem this becomes

 
Div(J T F−T ) dV0 + J b dV0 = 0, (48)
R0 R0

where
P := J T F−T (49)

is the First Piola—Kirchhoff stress tensor and b0 = J b are the reference body forces,
so that 
(DivP + b0 ) dV0 = 0. (50)
R0

Due to arbitrariness of R0 , we obtain the local form of the balance of linear momentum
in the reference configuration, which reads

DivP + b0 = 0 (51)

The balance of angular momentum implies

T = TT , (52)

and, in terms of the first Piola-Kirchhoff stress tensor,

P FT = F P T . (53)

We can also introduce the second Piola-Kirchhoff stress tensor

S := F−1 P = J F−1 T F−T , (54)

which is a symmetric second order tensor, i.e. S = ST .

68 M. de Luca and A. DeSimone

3 Thermodynamics

In this section we derive a consistent thermodynamic theory which involves chemi-

cal, thermal and mechanical processes. To do so, we recall the following quantities
defined per unit reference volume: ε0 is the internal energy, η0 is the entropy density,
Q0 is the external heat supply, and ρ0 is the mass density. Moreover, we consider
the heat flux per unit reference area q0 , the chemical potential μ, and the absolute
temperature θ .
We note that the mass density satisfies ρ = ρ s φ + ρ f φ f , where ρ s and ρ f are
the current mass densities (mass densities per unit current volume) corresponding
respectively to the solid and fluid part.
Given a region R0 ⊂ Ω0 with outward unit normal n0 we can define the corre-
sponding global energy terms

I (R0 ) = P : Ḟ dV0 : power of internal forces, (55)
R0
 
W (R0 ) = P n0 · ẋ dA0 + b0 · ẋ dV0 : power of external forces, (56)
∂R0 R0

E (R0 ) = ε0 dV0 : internal energy, (57)
R0

1
K (R0 ) = ρ0 |ẋ|2 dV0 : kinetic energy, (58)
R0 2
 
Q(R0 ) = − q0 · n0 dA0 + Q0 dV0 : heat flow, (59)
∂R0 R0

T (R0 ) = − μ j · n0 dA0 : energy flow due to fluid diffusion, (60)
∂R0

S (R0 ) = η0 dV0 : entropy, (61)
R0
 
1 Q0
J (R0 ) = − q0 · n0 dA0 + dV0 : entropy flow. (62)
∂R0 θ R0 θ

The conventional power balance is

W (R0 ) = K˙ (R0 ) + I (R0 ), (63)

and the first law of thermodynamics reads

E˙ (R0 ) + K˙ (R0 ) = W (R0 ) + Q(R0 ) + T (R0 ) (64)

= K˙ (R0 ) + I (R0 ) + Q(R0 ) + T (R0 ). (65)

Thus,
E˙ (R0 ) = I (R0 ) + Q(R0 ) + T (R0 ) (66)
Elastomeric Gels: A Model and First Results 69

By inserting Eqs. (57), (55), (59), and (60) in Eq. (66) we have
    
ε̇0 dV0 = P : Ḟ dV0 − q0 · n0 dA0 + Q0 dV0 − μ j · n0 dA0 .
R0 R0 ∂R0 R0 ∂R0
(67)
Then, by applying the divergence theorem
 
 
ε̇0 dV0 = P : Ḟ − Div q0 + Q0 − Div (μ j) dV0 , (68)
R0 R0

and finally, since the equation holds for each subset R0 ,

ε̇0 = P : Ḟ − Div q0 + Q0 − μ Div j − ∇μ · j. (69)

The second law of thermodynamics states that the net entropy production should
always be non negative:
S˙ (R0 ) − J (R0 ) ≥ 0. (70)

By inserting Eqs. (61) and (62) in Eq. (70) we have

  
1 Q0
η̇0 dV0 ≥ − q0 · n0 dA0 + dV0 . (71)
R0 ∂R0 θ R0 θ

Then, thanks to the divergence theorem

     
1 Q0
η̇0 dV0 ≥ − Div q0 + dV0 (72)
R0 R0 θ θ

and finally, since each integral holds ∀R0 ⊂ Ω0

 
1 1
η̇0 ≥ Q0 − Div q0 + q0 · ∇θ . (73)
θ θ

From Eq. (69) we can derive Q0 − Div q0 and inserting it in Eq. (73) to obtain
 
1 1
η̇0 ≥ ε̇0 − P : Ḟ + μ Div j + ∇μ · j + q0 · ∇θ . (74)
θ θ

Then, multiplying by θ each term and bringing all the terms on the right we have

1
ε̇0 − θ η̇0 − P : Ḟ + μ Div j + ∇μ · j + q0 · ∇θ ≤ 0. (75)
θ
We define the Helmotz free energy ψ0 as

ψ0 = ε0 − θ η0 , (76)
70 M. de Luca and A. DeSimone

then its time derivative is

ψ̇0 = ε˙0 − θ̇ η0 − θ η˙0 . (77)

We can replace ε˙0 − θ η˙0 = ψ̇0 + θ̇ η0 in Eq. (75) to have

1
ψ̇0 + θ̇ η0 − P : Ḟ + μ Div j + ∇μ · j + q0 · ∇θ ≤ 0, (78)
θ
that in isothermal condition reduces to

ψ̇0 − P : Ḟ + μ Div j + ∇μ · j ≤ 0. (79)

Thanks to the replacement of the mass balance (31) in Eq. (79) we finally have the
thermodynamic constraint

ψ̇0 − P : Ḟ − μ ċ + ∇μ · j ≤ 0. (80)

3.1 Stress Power

In this section we compute the P : Ḟ term in the thermodynamic inequality (80).

Given the kinematic decomposition of the deformation gradient F in Eq. (1), the
time derivative of F reads
Ḟ = Ḟe Fs + Fe Ḟs , (81)

then

P : Ḟ = P : Ḟe Fs + P : Fe Ḟs (82)

= PF sT
: Ḟ + F P : Ḟ
e eT s
(83)
= J T F−T FsT : Ḟe + J FeT T F−T : Ḟs (84)
s −T s −T
= J T (F F ) e
F sT
: Ḟ + J F
e eT
T (F F ) e
: Ḟ s
(85)
=J TF e−T
F s−T
F sT
: Ḟ + J F
e eT
TF e−T
F s−T
: Ḟ s
(86)
=J TF e−T
: Ḟ + J F
e eT
TF e−T
: Ḟ F
s s−1
(87)
=J TF e−T
: Ḟ + J F
e eT
TF e−T
:L s
(88)
= P : Ḟ + M : L ,
e e e s
(89)

in which we have used relations (49) and (6), and we have defined two new measures
of stress

Pe = J T Fe−T (90)
Me = J FeT T Fe−T . (91)
Elastomeric Gels: A Model and First Results 71

Recalling (9), and replacing it in Eq. (89) we obtain

1 s−1 ˙s
P : Ḟ = Pe : Ḟe + Me : J J I (92)
3
1 s−1 ˙s
= Pe : Ḟe + J J trMe (93)
3
= Pe : Ḟe − p̄ J˙s , (94)

where p̄ is the mean pressure

1
p̄ = − J s−1 trMe . (95)
3
We define another symmetric measure of stress Se as

Se = J Fe−1 T Fe−T , (96)

and we observe that

Pe = Fe Se , (97)

Thanks to Eqs. (91) and (96) we have

1 1
p̄ = − J e trT = − J s−1 Se : Ce . (98)
3 3
Then, recalling relation (3) we find
T
Se : Ċe = Se : (Ḟe Fe + FeT Ḟe ) (99)
T
=S : e
Ḟe F +S :F
e e eT
Ḟ e
(100)
= Se FeT : Ḟ eT + Fe Se : Ḟe (101)
= F S : Ḟ + F S : Ḟ
e e e e e e
(102)
= 2 P : Ḟ .
e e
(103)

3.2 Elastic Incompressibility Constraint

We introduce in the model the incompressibility constraint for the elastic polymer
network by imposing
J e = 1. (104)
72 M. de Luca and A. DeSimone

Then its time derivative is

J˙e = J e Ḟe Fe−1 : I = J e Ḟe : Fe−T = 0, (105)

that implies orthogonality between Fe−T and Ḟe . It follows that the introduction of
the incompressibility constraint in the form (105) does not have any effect if added
to Eq. (94) and, we can write

P : Ḟ = Pe : Ḟe − p̄ J˙s + p Fe−T : Ḟe , (106)

where p is a Lagrange multiplier associated with the imposed constraint.

Using Eq. (3), we can compute
 T

Fe−T : Ḟe = Fe−T : Fe−T Ċe − Fe−T Ḟe Fe (107)

= Fe−T : Fe−T Ċe − Fe−T : Fe−T Ḟ eT Fe (108)

= Fe−1 Fe−T : Ċe − Fe−1 Fe−T FeT : Ḟ eT (109)
=C e−1
: Ċ − F
e e−T
: Ḟe , (110)

then
Ce−1 : Ċe = 2 Fe−T : Ḟe . (111)

Introducing Eqs. (103) and (111) in Eq. (106) we have

1 e 
P : Ḟ = S + p Ce−1 : Ċe − p̄ J˙s . (112)
2
To conclude our computation we observe that

J˙s = ν ċ (113)

thank to Eqs. (24), (112) becomes

1 e 
P : Ḟ = S + p Ce−1 : Ċe − ν p̄ ċ. (114)
2

3.3 Dissipation Inequality

The introduction of Eq. (114) in the free energy inequality (80) gives:

1 e 
ψ̇0 − S + p Ce−1 : Ċe − (μ − ν p̄) ċ + ∇μ · j ≤ 0. (115)
2
Elastomeric Gels: A Model and First Results 73

We define

Seact = Se + p Ce−1 , (116)

μact = μ − ν p̄, (117)

where Seact is called “active stress”, and μact is called “active chemical potential”.
Introducing Eqs. (116) and (117) in Eq. (115), the free energy inequality becomes

1 e
ψ̇0 − S : Ċe − μact ċ + ∇μ · j ≤ 0 (118)
2 act
The independent variables of our model are the elastic deformation represented
by the right Cauchy—Green strain tensor Ce , and the concentration c of the solvent.
Hence we write

ψ0 = ψ̂0 (Ce , c), (119)

Seact= Ŝeact (Ce , c), (120)
μact = μ̂act (Ce , c), (121)
j = ĵ(C , c, ∇μ),
e
(122)
√
with Ce constrained to satisfy det Ce = J e = 1.

3.4 Thermodynamic Restrictions

In order to determine the thermodynamic restriction imposed by the free energy

inequality (118) we compute the time derivative of the free energy which, thanks to
Eq. (119), reads
∂ ψ̂0 ∂ ψ̂0
ψ̇0 = : Ċe + ċ. (123)
∂Ce ∂c
Substituting in Eq. (118) gives

∂ ψ̂0 1 ∂ ψ̂0
− Seact : Ċe + − μact ċ + ĵ(Ce , c, ∇μ) · ∇ μ ≤ 0. (124)
∂C e 2 ∂c

Inequality (124) must hold for all values of Ce , c, and ∇ μ. Since Ce and c appear
linearly, the only possibility is that their coefficients vanish. Then inequality (124)
implies

∂ ψ̂0
Seact = 2 , (125)
∂Ce
∂ ψ̂0
μact = , (126)
∂c
74 M. de Luca and A. DeSimone

and the dissipation inequality reduces to

ĵ(Ce , c, ∇μ) · ∇ μ ≤ 0. (127)

Recalling definitions (116) and (117), we finally have the expressions for the stress
tensor and the chemical potential:

∂ ψ̂0
Se = 2 − p Ce−1 , (128)
∂Ce
∂ ψ̂0
μ= + ν p̄. (129)
∂c
Combining Eqs. (128) and (98) we obtain

2 ∂ ψ̂0
p = J s p̄ − : Ce , (130)
3 ∂Ce
and the expression of the stress tensor Se depending on the mean (pore) pressure is:

8 ∂ ψ̂0
Se = − J s p̄ Ce−1 . (131)
3 ∂Ce

4 Constitutive Theory

The constitutive theory and the thermodynamic restriction that we propose are based
on ψ0 , the free energy density defined in the reference, free swollen, configuration.
In the literature the free energy density for polymeric swelling material is usually
defined per unit volume in the dry configuration, ψd , and is composed by a term due
to the mixing of the polymer and the solvent, ψd mix , and a term that accounts for the
energy change due to the stretching of the polymer molecules, ψd mech :

ψd (Cd , cd ) = μ0 cd + ψd mix (Cd , cd ) + ψd mech (Cd , cd ), (132)

where μ0 is the value of the chemical potential of pure solvent, and

    
ν cd 1
ψd mix (Cd , cd ) = R θ cd log +χ , (133)
1 + ν cd 1 + ν cd
G
ψd mech (Cd , cd ) = (trCd − 3). (134)
2
In Eq. (133), R = kB NA is the ideal gas constant, that is equal to the Boltzmann’s
constant kB times the Avogadro’s number NA , and θ is the absolute temperature,
while G in Eq. (134) is the shear modulus.
Elastomeric Gels: A Model and First Results 75

4.1 Free Energy Density per Unit Reference Volume

Recalling Eqs. (12) and (20) we obtain the free energy density per unit reference
volume as
1
ψ0 (C, c) = ψd (F0 T C F0 , J0 c), (135)
J0

where
    
ν J0 c 1
ψ0 mix (C, c) = R θ c log +χ , (136)
1 + ν J0 c 1 + ν J0 c
G  
ψ0 mech (C, c) = (tr F0 T C F0 − 3). (137)
2 J0

In the constitutive theory we chose as independent variables the concentration

c and the elastic deformation Ce , and since there is no dependence on the elastic
deformation
    
ν J0 c 1
ψ0 mix (Ce , c) = R θ c log +χ . (138)
1 + ν J0 c 1 + ν J0 c

To compute the mechanical part, thanks to Eq. (2), we have

   
tr FT 0 C F0 = tr F0 T FsT Ce Fs F0 , (139)

where
Fs = λs I = (J s )1/3 I = (J0 )−1/3 (1 + ν J0 c)1/3 I, (140)

then
G  
ψ0 mech (Ce , c) = ((J0 )−2/3 (1 + ν J0 c)2/3 tr F0 T Ce F0 − 3). (141)
2 J0

Having in mind future applications to biological tissues and intervertebral discs,

we want to also allow the presence of fibres in the gel, and we consider two families
of fibres that are stress-free in the swollen configuration. The energy density that
accounts for the fibre contribution is defined per unit reference volume, and it vanishes
when there is no elastic deformation Ce = I:
k1  
ψ0 fibre (Ce , c) = ΣAα exp k2 (Ce : Aα − 1)2 − 1 , (142)
2 k2

where α = 1, 2 identifies the first and the second family of fibres.

A1 = a1 ⊗ a1 , and A2 = a2 ⊗ a2 , (143)
76 M. de Luca and A. DeSimone

are the structural tensor that account for the material anisotropy and a1 and a2 are
the material unit vectors defining the direction of the two family of fibres.
To conclude, accounting for all contributions, the energy density for our material
reads

ψ0 (Ce , c) = μ0 c + ψ0 mix (Ce , c) + ψ0 mech (Ce , c) + ψ0 fibre (Ce , c). (144)

4.2 Partial Derivatives of the Free Energy ψ0

In this section we compute the partial derivatives of the energy density ψ0 that are
necessary to compute the stress and the chemical potential. We start with

∂ψ0 mix (Ce , c)

= 0, (145)
∂Ce
∂ψ0 mech (Ce , c) G
= (J0 )−2/3 (1 + ν J0 c)2/3 F0 F0 T , (146)
∂C e 2 J0

∂ψ0 fibre (Ce , c)  e 2 

= k1 (I e
− 1) exp k2 I − 1 A1 +
∂Ce A 1 A1

 2
k1 (IAe 2 − 1) exp k2 IAe 2 − 1 A2 , (147)

where
IAe α = Ce : Aα , with α = 1, 2. (148)

Then
∂ψ0 (Ce , c) ∂ψ0 mix (Ce , c) ∂ψ0 mech (Ce , c) ∂ψ0 fibre (Ce , c)
= + + (149)
∂C e ∂Ce ∂Ce ∂Ce
We compute the partial derivatives of ψ0 with respect to the concentration c:
   
∂ψ0 mix (Ce , c) ν J0 c 1 χ
= R θ log + + , (150)
∂c 1 + ν J0 c 1 + ν J0 c (1 + ν J0 c)2

∂ψ0 mech (Ce , c) 1 ν G  1/3  

= J0 (1 + ν J0 c)−1/3 tr F0 T Ce F0 , (151)
∂c 3 J0

and
∂ψ0 fibre (Ce , c)
= 0, (152)
∂c
Elastomeric Gels: A Model and First Results 77

then
∂ψ0 (Ce , c) ∂ψ0 mix (Ce , c) ∂ψ0 mech (Ce , c) ∂ψ0 fibre (Ce , c)
= μ0
+ + + . (153)
∂Ce ∂c ∂c ∂c

4.3 Stress Tensor

The insertion of Eq. (149) in Eq. (128) finally gives the expression of the elastic part
of the second Piola-Kirchhoff stress tensor
G s2/3
Se (Ce , c) = J F0 F0 T − p Ce−1
J0
 2 
+ 2 k1 (IAe 1 − 1) exp k2 IAe 1 − 1 A1
  
2
+ 2 k1 (IAe 2 − 1) exp k2 IAe 2 − 1 A2 , (154)

where
J s = J0 −1 (1 + ν J0 c) . (155)

Moreover, if we want to express the stress tensor Se as function of the deformation

and the polymer fraction, we have:

G −2/3 2/3
Se (Ce , φ) = φ φ0 F0 F0 T − p Ce−1
J0
 2 
+ 2 k1 (IAe 1 − 1) exp k2 IAe 1 − 1 A1
 2 
+ 2 k1 (IAe 2 − 1) exp k2 IAe 2 − 1 A2 , (156)

since
J s = φ −1 φ0 . (157)

Using Eqs. (96) and (54) we recover the expression of the second Piola-Kirchhoff
stress tensor
S = Fs−1 Se Fs−T = J s−2/3 Se = φ 2/3 φ0 −2/3 Se , (158)

and, thanks to Eq. (156) we have

G
S(Ce , φ) = F0 F0 T − p C−1
J0
 2 
+ 2 φ 2/3 φ0 −2/3 k1 (IAe 1 − 1) exp k2 IAe 1 − 1 A1
 2 
+ 2 φ 2/3 φ0 −2/3 k1 (IAe 2 − 1) exp k2 IAe 2 − 1 A2 , (159)
78 M. de Luca and A. DeSimone

where
C−1 = φ 2/3 φ0 −2/3 Ce−1 . (160)

4.4 Chemical Potential

The final expression of the chemical potential can be obtained by replacing Eq. (153)
in Eq. (129) to obtain
   
ν J0 c 1 χ
μ(Ce , c) = μ0 + R θ log + +
1 + ν J0 c 1 + ν J0 c (1 + ν J0 c)2
1 ν G  1/3  
+ J0 (1 + ν J0 c)−1/3 tr F0 T Ce F0 + ν p̄, (161)
3 J0

where p̄, defined in Eq. (98), can be explicitly computed once the expression of the
stress tensor Se is known, as in Eq. (154). In terms of the polymer fraction φ, the
chemical potential becomes

μ(Ce , φ) =μ0 + R θ log (1 − φ) + φ + χ φ 2

1 ν G  −1/3 1/3  T e 
+ φ0 φ tr F0 C F0 + ν p̄. (162)
3 J0

4.5 Solvent Flux

In the constitutive theory, we made the assumption that the solvent flux j depends on
the deformation, the concentration, and the gradient of the chemical potential, see
Eq. (122). Following [3] we suppose that the fluid flux obeys a Darcy-tipe relation,
hence it depends linearly on the gradient of the chemical potential

j = −M ∇μ, (163)

where M is the mobility tensor. To respect the constitutive restriction (127) the
mobility M must be a positive definite tensor.
The fluid flux j is the amount of solvent that flows through the boundary of the
body per unit surface and per unit time, then it is measured in mol/(s m2 ). If we
introduce Eq. (129) in Eq. (163) we have
 
∂ψ0
j = −M ∇ + ν p̄ , (164)
∂c
Elastomeric Gels: A Model and First Results 79

and, in turn,  
∂ψ0
ν j = −ν 2 M ∇ + p̄ . (165)
∂νc

We observe that ν j has the dimension of m3 /(m2 s),since the molar volume [ν] =
m3 /mol, that still represents the amount of solvent flowing through the boundary per
unit time and per unit surface, but simplifying we obtain that it has the dimension of
a velocity, i.e. m/s.
Recalling definition (34), (24), and (40), we have

Φ f = νc, (166)

and Eq. (165) becomes

 
Φf ∂ψ0
j = −ν 2 M ∇ + p̄ , (167)
c ∂Φ f

where j/c represents the relative velocity vfs of the fluid with respect to the solid
network that usually appears in the generalized Darcy’s Law
 
∂ψ0
Φ v = −K ∇
f fs
+ p̄ , (168)
∂Φ f

where K is the permeability tensor corresponding to

K = ν 2 M. (169)
∂ψ0
We remark that in Terzaghi’s consolidation theory the term ∂Φ f , which accounts for

adsorption and capillarity, is neglected. In our case this term is not negligible due to
the importance of swelling phenomena.
For isotropic materials the mobility, and also the permeability, can be represented
by scalar quantities, so that the corresponding tensors are

M = m I, K = k I, (170)

and thanks to Eq. (169) we have a scalar relation between mobility and permeability:

k = ν 2 m. (171)

In [4] a linear dependence of the mobility upon the polymer fraction is proposed as

D
m(φ) = γ (φ), (172)
ν Rθ
80 M. de Luca and A. DeSimone

Fig. 2 Comparison between 4

the two laws proposed to γ
model the dependence of the
3 k
permeability upon the
polymer fraction φ. The
linear law is the function γ 2
in Eq. (173), and the
exponential law is the
function k in Eq. (175) 1

0
0 0.2 0.4 0.6 0.8 1
φ

where D is a diffusion coefficient and γ (φ) is a positive function of φ

γ (φ) = (1 − γs ) φ φ0−1 + γs , γs > 1, (173)

such that in the reference state (φ = φ0 ), γ = 1, and m = D/(ν R θ ).

Having in mind a specific application to intervertebral discs, in [1] an exponential
law is proposed for the permeability
 2   
e (1 + e0 ) 1+e
k = k0 exp M −1 , (174)
e0 (1 + e) 1−e

where k0 is the permeability and e0 the void ratio in the reference state, and M is
an empirical positive coefficient. Using Eqs. (43) and (39) we obtain the relation
between the permeability and φ:
 2   
1−φ φ
k(φ) = k0 exp M −1 . (175)
1 − φ0 φ0

In the reference state φ = φ0 , and the permeability k assumes its reference value k0 .
In Fig. 2, a comparison between the two laws (173) and (175) is shown. Only for
this purpose arbitrary values of the constant are chosen, namely M = 10−6 , γs = 2,
φ0 = 0.5, k0 = 1. The important common feature of the proposed laws is the increase
of permeability with swelling, which causes a decrease in the polymer fraction.

5 Summary of the Equations of the Model

In this section we summarize the system of equations needed to solve a swelling-

deformation problem. We have a system of four equations, whose independent vari-
ables are the displacement u, the chemical potential μ, the polymer fraction φ and
the pressure p:
Elastomeric Gels: A Model and First Results 81
⎧
⎪
⎪DivP + b0 = 0
⎪
⎨− φ0 φ̇ = −Div (j)
ν φ2 (176)
⎪
⎪μ = μ0 + R θ log (1 − φ) + φ + χ φ 2 + ν p φ φ0−1
⎪
⎩ e
J =1

The first equation is the balance of forces, the second the balance of fluid mass, the
third is the relation between the chemical potential and the polymer fraction and
the last represents the constraint of elastic incompressibility. To close the problem
we need to consider the boundary condition for the displacement and the fluid mass
equations.

6 Spontaneous Swelling Deformations

The spontaneous swelling deformation is achieved when a dry sample is fully

immersed in solvent and undergoes free swelling. The computation of the spon-
taneous deformation is fundamental when we choose a stress free configuration as
reference configuration and we want to study and simulate a particular phenomenon
starting from that configuration.
The free energy function of the material in the dry configuration is given by
Eq. (132), and the spontaneous deformation is the deformation that minimizes such
energy. The free energy function ψd (Cd , cd ), in Eq. (132), is composed by the elastic
part ψd mech , in Eq. (134), and the mixing part ψd mix , in Eq. (133). For the computation
of the spontaneous deformation the free energy contribution due to the fibres is not
considered. We consider only the case of isotropic deformation, since free swelling
deformations are isotropic, such that

Fd = λd I, (177)

and the incompressibility constraint (105) becomes

Jd − 1
Jd = 1 + ν cd , and cd = , (178)
ν

with Jd = λ3d . When in Eq. (132) we replace Fd and cd with (177) and (178), we
obtain
    
Rθ  3  1 1 3  
ψd (λd ) = λd − 1 log 1 − 3 + χ 1 − 3 + G0 λ2 − 1 .
ν λd λd 2
(179)
82 M. de Luca and A. DeSimone

The minimum of free energy (179) is numerically computed using Newton’s method,
for which expressions of its first and second derivatives with respect to λd are needed:
   
dψd (λd ) Rθ 1 3 3χ
= 3 λd l log 1 − 3 +
2
+ 4 + 3 G0 λd , (180)
dλd ν λd λd λd

and
   
d2 ψd (λd ) Rθ 1 9 λd 3 12 χ
= 6 λd log 1 − 3 + 3 − 2 − 5 + 3 G0 . (181)
dλd2 ν λd λd − 1 λd λd

Using numerical data χ = 0.1, θ = 298 K, R = 8.314472 J K/mol, G0 = 1e6 Pa, we

find the value λ∗ = 1.6475 for which energy ψd (λ∗ ) is minimal, as shown in Fig. 3.
Recalling Eq. (11), we set λ0 = λ∗ as the spontaneous (free swelling) deformation,
such that in the reference state Ω0 (in Fig. 1), the free energy is minimum and the
body is stress free. Hence, thanks to Eq. (10), since Fd = F0 , in the reference free
swollen state the deformation gradient F = I.
Once the spontaneous deformation is determined, thank to Eq. (34), and recalling
that J0 = λ30 , we can compute the polymer fraction φ0 in the reference state. The
smaller φ0 is, the grater is the amount of solvent that entered the body. For λ0 =
1.6475 we have φ0 = 0.2236.
A key role in the determination of the spontaneous deformation is played by
the χ parameter. In Flory’s theory [8], χ is called enthalpy of mixing and it repre-
sents the energetic contribution due to the mixing of polymer and solvent. Values of
χ ≤ 0.5 promote swelling, while greater values prevent swelling due to some repul-
sion between the polymer and the solvent. Looking at Fig. 4, ψd mix has an horizontal
asymptote, as the deformation increases. While for χ ≤ 0.5, ψd mix strictly decreases
for deformations greater than 1, for χ > 0.5 the free energy ψd mix shows a change
of concavity and the presence of a minimum that becomes closer to the identity
deformation as χ increases. This can be attributed to growing repulsion between the
polymer and the solvent, until no solvent is allowed to enter the polymer. As an exam-

Fig. 3 The free energy ψd 10 8

in the dry configuration, 1
composed by a mechanical
part ψd mech and a mixing 0.5
contribution ψd mix . The
value λ∗ corresponds to the
deformation, greater than 1, 0
that minimizes ψd . The
value χ = 0.1 has been used -0.5
for the plot

-1
0 1 2 3 4 5
Elastomeric Gels: A Model and First Results 83

Fig. 4 Dependence of the 10 7

mixing free energy ψd mix 5
upon the entalphy of
mixing χ

-5
0 0.5 1 1.5 2 2.5 3

Fig. 5 The free energy ψd 10 8

in the dry configuration, 1
composed by a mechanical
part ψd mech and a mixing 0.5
contribution ψd mix . The
value λ∗ corresponds to the
deformation, greater that 1, 0
that minimizes ψd . The
value χ = 2 has been used in -0.5
the plot

-1
0 1 2 3 4 5

ple, in Fig. 5, the free energy ψd is computed using χ = 2, and the corresponding
minimizing spontaneous deformation is λ0 = λ∗ = 1.023. The corresponding poly-
mer fraction is φ0 = 0.934, i.e., a very small amount of solvent entered the polymer
network and the material stays almost dry.

7 Numerical Experiments

In this section we describe a series of numerical experiments that can be performed,

using the presented model, to simulate the physical behaviour of a hydrogel.

7.1 Free Swelling Deformation

A free swelling deformation, as previously pointed out, is an isotropic deformation

a body undergoes when, fully immersed in the solvent, it is free to swell without
any constraint. As described in the previous section, a freely swollen configuration
corresponds to a state in which the free energy is minimum and the body is stress
84 M. de Luca and A. DeSimone

free. In this section, we want to study the dynamics of the free swelling deformation,
by simulating the hydrogel that, starting from the dry configuration swells until it
reaches the free swelling state. Then, in this case, the initial configuration will be
the dry, not stress free, configuration Ω∞ in Fig. 1. The presented model can be used
to simulate such problem by setting F0 = I, so that F = Fd and ψd = ψ0 , and the
system of equations we are going to solve is the following:
⎧
⎪
⎪DivP + b0 = 0
⎪
⎨− 1 φ̇ = −Div (j)
ν φ2 (182)
⎪
⎪μ = μ0 + R θ log (1 − φ) + φ + χ φ 2 + ν p φ
⎪
⎩ e
J =1

Figure 6 shows the initial dry configuration of the body. It is a cube, whose edge
is 0.1 cm long. For symmetry reasons only one eighth of the cube is shown. The
boundary conditions for the problem are μ = μ0 for all the boundaries of the body,
because we suppose that the body is fully immersed in the pure solvent.
In Fig. 8 the body is fully swollen, the solvent inside the body is in equilibrium
with the solvent outside (they have the same chemical potential), and the polymer
fraction φ = 0.2236 is uniform throughout the body. The corresponding free swelling
deformations in the three directions are the same, confirming that the deformation
is isotropic and uniform, with values λx = λy = λz = 1.6475. The computed values
for the polymer fraction and the deformations coincide with the values for the spon-
taneous deformation computed in the previous section through the minimization of
the free energy.

Fig. 6 Initial dry

configuration for the free
swelling deformation. In the
figure the variable φ is
plotted and the legend shows
the assumed values
throughout the body. The
value φ = 1 corresponds to
the total absence of solvent
Elastomeric Gels: A Model and First Results 85

The numerical simulation of the swelling process allows us to observe the dynam-
ics of the solvent flowing inside the body. The intermediate states between the initial
and the final configuration are not characterized by an isotropic uniform deformation
as shown by Fig. 7. The solvent flows across the boundaries and migrates toward the
inner part of the body due to the gradient of the chemical potential. The diffusion
coefficient D = 5e − 12 m2 /s determines how fast the diffusion process happens.
After 36 h the dynamics of the solvent is over and equilibrium is reached (Fig. 8).

Fig. 7 Intermediate state for

the free swelling
deformation: the material is
swelling. The legend shows
the values of φ throughout
the body: at this stage the
polymer fraction is not
uniform in the volume. The
edges of the cube swell more
than the rest of the body, due
to a larger concentration of
solvent near the boundaries

Fig. 8 Final state for the

free swelling deformation.
The material is fully swollen,
the dymanics of the solvent
flowing through the body is
over and φ is again uniform
throughout the body
86 M. de Luca and A. DeSimone

7.2 Indentation Experiment

Indentation is one of the most common techniques for the mechanical characteri-
zation of materials. The test can be realized by pressing an indenter with different
possible shapes (spherical, conical, cylindrical flat-ended) on the upper surface of a
block of material. A control parameter can be the indentation depth or the pressure
exerted by the indenter and, from the material response it is possible to compute
material parameters which as the shear modulus and the Poisson coefficient.
When using sharp or spherical indenters the contact area varies during the inden-
tation, while with a flat-ended cylindrical indenter the contact area remains constant
[15]. In the following we will only consider the latter type of indenter, as shown in
Fig. 9.
If the deformation induced by the punch is sufficiently small, then the linear
theory of elasticity is applicable, and provided that the dimensions of the bodies are
large compared with the dimension of the contact area, the stresses in this region are
independent upon the shape of the bodies and the way they are supported far from
the contact area [11].
The displacement variable is u, the small strain tensor is

1 
ε= ∇u + ∇uT , (183)
2
and the stress tensor is
 
ν
σ = 2G ε + tr(ε)I , (184)
1 − 2ν

where G is the shear modulus, and ν is the Poisson coefficient.

Using the reference system depicted in Fig. 9, we consider the z axis along the
axis of the indenter and, due to the symmetry of the problem, a radial axis on the
surface of the material. The radius of the flat-ended indenter is a. The equilibrium
problem divσ = 0 with the boundary conditions

σzz (r, 0) = 0, r > a (185)

σrz (r, 0) = 0, 0 ≤ r ≤ a (186)
uz (r, 0) = h, 0 ≤ r ≤ a (187)

Fig. 9 Flat-ended punch

indentation scheme, adapted
from [15]
Elastomeric Gels: A Model and First Results 87

is solved in [10], and the results are reported in the following.

With reference to Fig. 9, corresponding to a penetration depth h, the distribution
of pressure σzz under the punch is

2Gh
σzz (r, 0) = − √ . (188)
π(1 − ν) a2 − r 2

When the material is incompressible the distribution of pressure under the punch is
obtained by replacing ν = 1/2 in Eq. (188):

4Gh
σzzinc (r, 0) = − √ . (189)
π a2 − r 2

The contact force exerted by the punch on the material can be obtained by integrating
the distribution of pressure over the area of the punch:
 2π  a  2π  a
F=− σzz r dr dα, and F inc = − σzzinc r dr dα, (190)
0 0 0 0

obtaining
4Gah
F= , and F inc = 8Gah, (191)
1−ν

which give the reaction forces that a material exerts against the indenter pressed with
indentation depth h. In Figure 11 we refer to F as F comp , because it is the value for a
compressible material.
To perform the numerical simulation of the indentation test we start from a cylinder
of hydro-gel fully swollen, in its stress-free reference state. The shear modulus we
are using in the numerical computation is G = 105 Pa, and the corresponding initial
polymer fraction φ0 = 0.0615. We suppose that the material is immersed in pure
solvent and it is in equilibrium, so that the chemical potential through all the body
is zero. From this condition a cylindrical flat indenter is pressed on the top center
of the cylinder, as shown in Fig. 10. The radius of the indenter is a = 0.1 cm, the
indentation depth is h = 0.03 cm.
When the indenter is pressed into the hydro-gel, the latter immediately deforms
elastically. The movement of the polymeric network causes a variation in the chemical
potential, that in order to find new equilibrium induces the discharge of the solvent
from the hydro-gel. Figure 10 shows a detail of the loaded area and the flux of the
solvent that is depicted by the arrows, which have the direction of the chemical
potential gradient. The solvent cannot flow through the contact area because the
indenter is considered impermeable. Figure 11 shows the graph of the reaction force
that the hydro-gel exerts against the indenter. It initially responds as an incompressible
elastic material and then the forces relax due to the migration of the solvent until the
new equilibrium is reached. The corresponding F comp is computed using the first of
Eq. (191), with a Poisson coefficient ν = 0.3.
88 M. de Luca and A. DeSimone

Fig. 10 Fully swollen hydro-gel loaded by a cylindrical flat indenter. The indentation causes a
change in the chemical potential, shown by the color bar, that induces a discharge of some solvent
from the body. A detail of the indentation zone is shown, where the arrows have the direction of
the chemical potential gradient. The tip of the indenter is considered impermeable

0.01
h [cm]
0 F gel [N]
F comp [N]
-0.01 F inc [N]
-0.02
-0.03
-0.04
-0.05
-0.06
0 0.5 1 1.5 2 2.5 3 3.5
t [s] 10 4

Fig. 11 Comparison of the reaction force Fgel exerted by the hydro-gel against the indenter with
the analytical values of the reaction forces for a compressible (Fcomp ) and an incompressible (Finc )
material. As expected, when the indentation starts the hydro-gel responds as an elastic (incompress-
ible) material, then after some time the material relaxes due to the discharge of some solvent. h is
the indentation depth

References

1. Argoubi, M., & Shirazi-Adl, A. (1996). Poroelastic creep response analysis of a lumbar motion
segment in compression. Journal of Biomechanics, 29(10), 1331–1339.
2. Biot, M. A. (1972). Theory of finite deformation of porous solids. Indiana University Mathe-
matics Journal, 21(7).
3. Chester, S. A., & Anand, L. (November 2010). A coupled theory of fluid permeation and large
deformations for elastomeric materials. Journal of the Mechanics and Physics of Solids, 58(11),
1879–1906.
Elastomeric Gels: A Model and First Results 89

4. Chester, S. A., & Anand, L. (2011). A thermo-mechanically coupled theory for fluid permeation
in elastomeric materials: Application to thermally responsive gels. Journal of the Mechanics
and Physics of Solids, 59(10), 1978–2006.
5. de Luca, M., & DeSimone, A. (2012). Mathematical and numerical modeling of liquid crystal
elastomer phase transition and deformation. In MRS Proceedings, 1403, 2012. doi:10.1557/
opl.2012.249. Copyright Materials Research Society 2012, Published online by Cambridge
University Press: 2012.
6. de Luca, M., DeSimone, A., Petelin, A., & Čopič, M. (2013). Sub-stripe pattern formation in
liquid crystal elastomers: Experimental observations and numerical simulations. Journal of the
Mechanics and Physics of Solids, 61(11), 2161–2177.
7. Ferguson, S. J., Ito, K., & Nolte, L.-P. (2004). Fluid flow and convective transport of solutes
within the intervertebral disc. Journal of Biomechanics, 37(2), 213–221.
8. Flory, P. J., & Rehner, J, Jr. (1943). Statistical mechanics of cross-linked polymer networks i.
rubberlike elasticity. The Journal of Chemical Physics, 11, 512.
9. Hong, W., Zhao, X., & Suo, Z. (2009). Formation of creases on the surfaces of elastomers and
gels. In Applied Physics Letters, 95
10. Ian, N. (1965). Sneddon. The relation between load and penetration in the axisymmetric boussi-
nesq problem for a punch of arbitrary profile. International Journal of Engineering Science,
3(1), 47–57.
11. Johnson, K. L. (1987). Contact mechanics. Cambridge: Cambridge university press.
12. Kang, M. K., & Huang, R. (2010). A variational approach and finite element implementation for
swelling of polymeric hydrogels under geometric constraints. Journal of Applied Mechanics,
77(6), 61004.
13. Lucantonio, A., Nardinocchi, P., & Teresi, L. (2012). Transient analysis of swelling-induced
large deformations in polymer gels. Journal of the Mechanics and Physics of Solids.
14. Murad, M. A., Bennethum, L. S., & Cushman, J. H. (1995). A multi-scale theory of swelling
porous media: I. application to one-dimensional consolidation. Transport in Porous Media,
19(2), 93–122.
15. Riccardi, B., & Montanari, R. (2004). Indentation of metals by a flat-ended cylindrical punch.
Materials Science and Engineering: A, 381(1), 281–291.
16. Rice, J. R., & Cleary, M. P. (1976). Some basic stress diffusion solutions for fluid-saturated
elastic porous media with compressible constituents. Reviews of Geophysics and Space Physics,
14(2), 227–241.
17. Zhang, J., Zhao, X., Suo, Z., & Jiang, H. (2009). A finite element method for transient analysis
of concurrent large deformation and mass transport in gels. Journal of Applied Physics, 105(9),
093522–093522.
A Tensegrity Paradigm for Minimal Mass
Design of Roofs and Bridges

Gerardo Carpentieri, Fernando Fraternali and Robert E. Skelton

Abstract This work presents a parametric design approach to simply-supported

structures, exhibiting minimal mass tensegrity architectures (axially-loaded pre-
stressible configurations of axially-loaded members) in two-dimensions. This pro-
vides minimal mass bridge structures in the plane. The mass minimization problem
considers a distributed loading condition, under buckling and yielding constraints.
The minimal mass structure is proved to be a tensegrity system with an optimal
complexity. This optimal complexity (number of structural elements) depends only
on material properties and the magnitude of the external load. The fact that the min-
imal mass structure is a Class 1 Tensegrity substructure has significant economic
advantage. Class 1 structures are less expensive to construct, and substructures are
easily deployable, offering portable applications for small spans. They can be easily
assembled for prefabricated component parts for large spans. This minimal mass the-
ory is then used to design a support structure for a solar panel cover of water canals,
stopping evaporative losses and generating power without requiring additional land.

1 Introduction

Tensegrity structures are very efficient, and tend to provide minimal mass solutions
to structure design under certain conditions. Some tensegrity papers have shown
minimal mass for tensile structures, subject to a stiffness constraint [1]. Some have

G. Carpentieri (B) · F. Fraternali

Department of Civil Engineering, University of Salerno, Via Giovanni Paolo II, 132,
84084 Fisciano, SA, Italy
e-mail: [email protected]
F. Fraternali
e-mail: [email protected]
R.E. Skelton
Mechanical and Aerospace Engineering, University of California San Diego,
9500 Gilman Drive MC 0411, La Jolla, CA 92093-0411, USA
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 91

K. Weinberg and A. Pandolfi (eds.), Innovative Numerical Approaches
for Multi-Field and Multi-Scale Problems, Lecture Notes in Applied
and Computational Mechanics 81, DOI 10.1007/978-3-319-39022-2_5
92 G. Carpentieri et al.

shown minimal mass for: compressive loads [2], cantilevered bending loads [3,
4], torsional loads [5], simply-supported bending loads [6], and distributed loads
on simply-supported spans, where significant structure is not allowed below the
roadway, [7]. The present work formulates a parametric design approach to simply-
supported (bridge-like) structures, which produces minimal mass shapes among all
possible tensegrity topologies (configurations of members).
The subject of form-finding of tensegrity structures continues to be an active
research area [8–13]. Particularly interesting is the use of fractal geometry as a form-
finding method for tensegrity structures, which is well described in [2, 3, 5, 14].
Such an optimization strategy exploits the use of fractal geometry to design tenseg-
rity structures, through a finite or infinite number of self-similar subdivisions of
basic modules. The strategy looks for the optimal number of self-similar iterations
to achieve minimal mass or other design criteria. This number is called the optimal
complexity, since this number fixes the total number of parts in the structure. The
‘fractal’ approach to tensegrity form-finding paves the way to an effective implemen-
tation of the tensegrity paradigm in parametric architectural design [9, 10, 15, 16].
Discrete to continuum approaches to trusses and tensegrity structures are available
in [17].
This paper finds the minimum mass design of tensegrity structures carrying simply
supported and distributed bending loads. In [7] numerical solutions where found for a
specified topology, without any theoretical guarantees that those topologies produced
minimal mass. This paper provides more fundamental proofs that provide necessary
and sufficient conditions for minimal mass.
The remainder of the paper is organized as follows. Section 2 describes the topol-
ogy of the tensegrity bridge under examination. For a simply-supported structure of
the simplest complexity, Sect. 3 describes the minimal mass bridge when the admis-
sible topology allows substructure or superstructure (that is, respectively, structure
below and above the roadbed). Section 4 defines deck mass and provides closed-form
solutions to the minimal mass bridge designs when only sub- or super-structure is
allowed. This finalizes the proof that the minimal mass bridge is indeed the substruc-
ture bridge. Section 4 also adds joint mass and shows that the optimal complexity is
finite. In Sect. 5, we describe an application of the above theory to the design of a
tensegirty bridge to be used for a solar panel covering of water canals. Conclusions
are offered at the end.

2 Planar Topologies of the Tensegrity Bridges Under Study

The tensegrity structures in this paper will be composed of rigid compressive mem-
bers called bars, and elastic tensile members called cables. We will assume that a
tensile member obeys Hooke’s law,

ts = k(s − s0 ), (1)
A Tensegrity Paradigm for Minimal Mass Design of Roofs and Bridges 93

where k is cable stiffness, ts is tension in the cable, s is the length of the cable,
and s0 < s is the rest length of the cable. The tension members cannot support
compressive loads. For our purposes, a compressive member is a solid cylinder,
called a bar. All results herein are trivially modified to accommodate pipes, tubes
of any material, but the concepts are more easily demonstrated and the presentation
is simplified by using the solid bar in our derivations. The minimal mass of a cable
with loaded length s, yielding strength σs , mass density ρs , and maximal tension ts is
ρs
ms = ts s. (2)
σs

The minimal mass to avoid yielding or buckling of a bar of length b, yielding

strength σb , mass density ρb , modulus of elasticity E b and with compression force
f b are, respectively

ρb fb
m b,Y = f b b, m b,B = 2ρb b 2
. (3)
σb π Eb

The planar bridge topology is considered here to elucidate the fundamental prop-
erties that are important in the vertical plane. We use the following nomenclature,
referring to Fig. 1: (i) a superstructure bridge has no structure below the deck level;
(ii) a substructure bridge has no structure above the deck level; (iii) a nominal bridge
contains both substructure and superstructure; (iv) Y means the design was con-
strained against yielding for both cables and bars; (v) B means the design was con-
strained against yielding for cables and buckling for bars; (vi) n means the number
of self-similar iterations involved in the design (n = 1 in Fig. 1); (vii) p means the
complexity of each iteration in the substructure ( p = 1 in Fig. 1c); (viii) q means
the complexity of each iteration in the superstructure (q = 1 in Fig. 1b); (ix) α is the
aspect angle of the superstructure measured from the horizontal; (x) β is the aspect
angle of the substructure measured from the horizontal.
We define the superstructure bridge of complexity (n, p = 0, q) where the sub-
structure below is deleted. We define the substructure bridge of complexity (n, p,
q = 0) where the superstructure above is deleted. It will be convenient to define the
following constants:

ρb /σb ρb L
ρ= , η= √ . (4)
ρs /σs (ρs /σs ) π E b F

Define a normalization of the system mass m by the dimensionless quantity μ:

m
μ= , (5)
(ρs /σs ) F L
94 G. Carpentieri et al.

(a)

(b)

(c)

Fig. 1 Basic modules of the tensegrity bridge with: a nominal bridge: n = q = p = 1; b super-
structure: n = q = 1; c substructure: n = p = 1. Compressive members (bars) are heavy black
lines, tensile members (cables) are thin red lines

where the mass m at the yielding condition is:

ρb  ρs 
nb ns
m= f i bi + ti si , (6)
σb i=1 σs i=1

where (bi , si ) is respectively the length of the ith bar or ith cable, and respectively
( f i , ti ) is the force in the ith bar or cable.
A Tensegrity Paradigm for Minimal Mass Design of Roofs and Bridges 95

3 Mass of Bridges of Complexity (n, p, q) = (1, p, q),

Under Yielding and Buckling Constraints

Now we consider structures by increasing p, q. This section finds the minimal mass
of substructure, and superstructure bridges with complexity (n, p, q) = (1, p, q), for
any p and q greater then 1.

3.1 Substructure Bridge with Complexity

(n, p, q) = (1, p > 1, 0)

Refer to Fig. 2 for the notation. The angle between the bars is:

2β
γ = . (7)
p−1

The lengths of the bars and cables are:

 
L L β L
s0 = , s1 = cos β, s2 = L sin β sin , b1 = b2 = sin β. (8)
2 2 p−1 2

From the equilibrium equations, we obtain the following relations for the forces:

F
f1 =      , f2 = 2 f1 , (9)
4 cos β + sin β(p−1
p−2) β
/ sin p−1
 
f2 β
t2 =   , t1 = t2 cos . (10)
2 sin β p−1
p−1

Fig. 2 Notations for forces and lengths of bars and cables for a substructure with complexity n = 1
and p > 1
96 G. Carpentieri et al.

The following theorems and corollaries can be obtained by minimizing the total
masses of the bridge in Fig. 2 (refer to [6] for an extended proof of the above theorem
and following theorems of this section).
Theorem 1 Consider a substructure bridge with topology described by (8), with
complexity (n, p, q) = (1, p, 0) (Fig. 2). At the yielding condition the dimensionless
total mass is:
⎡    ⎤
β β
t0 1 ⎣ ( p − 1) sin β sin p−1 + cos β cos p−1 ⎦
μY (β, p) = +    
F 4 cos β sin p−1 β
+ sin β(p−1
p−2)

( p − 1) sin β
+ρ      . (11)
4 cos β + sin β(p−1
p−2) β
/ sin p−1

Note that t0 is the force in deck cables s0 (see Fig. 2).

Corollary 1 The minimal mass in (11) is achieved at infinite complexity p → ∞
and t0 = 0. The minimal mass at yielding for a substructure bridge is:
 
1 √ 1
μ∗Y (βY∗ , ∗
p )= ρ + (1 + ρ) arctan √ , (12)
4 ρ

where p ∗ → ∞ and the optimal angle βY∗ is:

 
1
βY∗ = arctan √ . (13)
ρ

Theorem 2 Consider a substructure bridge with topology defined by (8), with com-
plexity (n, p, q) = (1, p, 0), See Fig. 2. At the buckling condition the dimensionless
total mass is minimized at p = 2 and t0 = 0, where:

1 + tan2 β η tan2 β
μ B (β, p = 2) = +   . (14)
4 tan β 2 1 + tan2 β 3/4

Corollary 2 The minimal mass substructure is achieved for p = 1.

3.2 Superstructure Bridge with Complexity

(n, p, q) = (1, 0, q > 1)

Refer to Fig. 3 for the notation. The angle between the bars is:

2α
γ = . (15)
q −1
A Tensegrity Paradigm for Minimal Mass Design of Roofs and Bridges 97

Fig. 3 Notations for forces and lengths of bars and cables for a superstructure with complexity
n = 1 and q > 1

The lengths of the bars and cables are:

 
L L L α
s0 = , s1 = s2 = sin α, b1 = cos α, b2 = L sin α sin . (16)
2 2 2 q −1

From the equilibrium equations, we obtain the following relations for the forces:

F t2
t2 =      , t1 = , (17)
α(q−2) α 2
2 cos α + sin q−1
/ sin q−1
 
t2 α
f2 =   , f 1 = f 2 cos . (18)
2 sin α q −1
q−1

The following theorems and corollaries can be obtained by minimizing the total
masses of the bridge in Fig. 3 (refer to [6] for an extended proof of the above theorem
and following theorems of this section).

Theorem 3 Consider a superstructure bridge, of total span L, topology defined by

(16), with complexity (n = 1, q > 1), Fig. 3. At the yielding condition under a vertical
load F the dimensionless total mass is:
t0 (q − 1) sin α
μY (α, q) = +     
F 4 cos α + sin α(q−2)
q−1
/ sin α
q−1
   
α α
ρ (q − 1) sin α sin q−1 + cos α cos q−1
+     . (19)
4 sin q−1α
cos α + sin α(q−2)q−1
98 G. Carpentieri et al.

Corollary 3 The minimal mass in (19) is achieved at infinite complexity q → ∞

and t0 = 0. Then the minimal mass at yielding for a superstructure bridge is:

1 √ √ 
μ∗Y (αY∗ , q ∗ ) = (1 + ρ) arctan ρ + ρ , (20)
4
where q ∗ → ∞ and the optimal angle αY∗ is:
√
αY∗ = arctan ρ. (21)

The left side of Fig. 4 illustrates superstructure bridges as q → ∞, where masses

are given for any q by (19).

Theorem 4 Consider a superstructure bridge with topology (16), and complexity

(n, p, q) = (1, 0, q > 1), see Fig. 3. At the buckling condition the dimensionless
total mass is:
t0 (q − 1) sin α
μ B (α, q) = +     
F 4 cos α + sin α(q−2)
q−1
α
/ sin q−1
⎡     ⎤
α α
⎢ cos 2
α cos q−1
+ 2 (q − 1) sin 2
α sin 2
q−1 ⎥
+ η⎢
⎣     
⎥ . (22)
⎦
α α(q−2)
2 sin p−1 cos α + sin q−1

Corollary 4 The minimal mass superstructure is achieved for q → ∞ and t0 = 0,

leading to the following mass:

α η cos2 α
μ B (α, q → ∞) = + √ . (23)
4 2 sin α

It is important to consider that, for the solution q → ∞, buckling is not the mode
of failure since the lengths of the bars approaches zero. Also note that at α = 90◦ ,
μ B = π/8.
The left side of Fig. 4 shows a sequence of superstructures under yielding con-
straints, as q increases. From (19) the mass is minimized at q → ∞ and αY∗ =
45◦ (ρ = 1). The right side of Fig. 4 shows a sequence of superstructures under
buckling constraints, as q increases. The mass is minimized at α = 90◦ for q = ∞
(η = 857.71, same steel/steel material as above).
Moreover, the left side of Fig. 5 shows a sequence of substructures under yield-
ing constraints, as p increases. From (11) the mass is minimized at p → ∞ and
βY∗ = 45◦ (ρ = 1). The right side of Fig. 5 shows a sequence of substructures under
buckling constraints, as p increases. The mass is minimized at β = 90◦ for p = 1
(η = 857.71, same steel/steel material as above).
A Tensegrity Paradigm for Minimal Mass Design of Roofs and Bridges 99

Fig. 4 Optimal topologies of superstructure bridges with complexity (n, p, q) = (1, 0, q → ∞)

under yielding constraints (left) and buckling constraints (right) for different q, (steel for bars and
cables, F = 1 N, L = 1 m)
100 G. Carpentieri et al.

Fig. 5 Optimal topologies of substructure bridges with n = 1 under yielding constraints (left) and
buckling constraints (right) for different p, (steel for bars and cables, F = 1 N, L = 1 m)

Theorem 5 A minimal mass superstructure constrained against yielding with hinge/roller

boundary conditions, has the same optimal topology as a minimal mass superstruc-
ture constrained against buckling and hinge/hinge boundary conditions.

Proof Michell [18] proved that the minimal mass structure constrained against yield-
ing with hinge/roller boundary conditions has the topology of the right side of Fig. 4
as q → ∞ and α → 90◦ . Theorem 4 provides the same topology for hinge/hinge
constraints. 
A Tensegrity Paradigm for Minimal Mass Design of Roofs and Bridges 101

Fig. 6 Minimal mass bridges under, a yielding constrained nominal bridges, b buckling constrained
superstructure bridge and c buckling constrained substructure bridge

Theorem 6 The minimal mass nominal bridge constrained against yielding is

obtained combining the optimal superstructure topology (Fig. 4, left side as q → ∞)
with the optimal substructure topology (Fig. 4, left side as p → ∞).

Proof Michell [18] obtained these same results by starting with a continuum and
optimizing the shape. 

Figure 6a illustrates the minimal mass nominal bridge under yielding constraints
(Theorem 6), leading to complexity (n, p, q) = (1, ∞, ∞). Figure 6b illustrates the
minimal mass superstructure bridge under buckling constraints, leading to complex-
ity (n, p, q) = (1, 0, q → ∞). Figure 6c illustrates the minimal mass substructure
bridge under buckling constraints, leading to complexity (n, p, q) = (1, 1, 0).

4 Introducing Deck and Joint Masses

In previous sections, complexity n was restricted to 1. This is appropriate only when

the external loads are all applied at the midspan. Real bridges cannot tolerate such
an assumption. So in this section we consider a distributed load. Part of the load is
the mass of the deck that must span the distance between adjacent support structures
(complexity n will add 2n − 1 supports). In the Sect. 4.4 we will consider adding
mass to make the joints, where high precision joints have less mass then rudely
constructed joints.

4.1 Including Deck Mass

The total load that the structure must support includes the mass of the deck, which
increases with the distance that must be spanned between support points of the
structure design (which is determined by the choice of complexity n). We therefore
consider bridges with increasing complexity n. We will show that the smallest n = 1
102 G. Carpentieri et al.

yields smallest structural mass and the largest deck mass. The required deck mass
obviously approaches zero as the required deck span approaches zero, which occurs
as n → ∞. We will show that the mass of the deck plus the mass of the structure is
minimized at a finite value of n.
The deck is composed by 2n simply supported beams connecting the nodes on
the deck.
Let the deck parameters be labeled as: mass m d , mass density ρd , yielding strength
σd , width wd . The mass of one deck section is equal to:

c1 c1 1
m d = 3n + 2n c2 + , (24)
2 2 22n

where:

3 wd g ρd2 L 3 16 σd F
c1 = , c2 = . (25)
8 σd 3 wd g 2 L 3 ρd2

Then, the normalized total mass of the deck structure is:

2n m d
μ∗d = . (26)
(ρs /σs ) F L

The total force acting on each internal node on the deck is the sum of the force due
to the external loads and the force due to the deck:

Ftot = F + 2n m d g. (27)

4.2 Adding Deck Mass for a Substructure Bridge

with Complexity (n, p, q) = (n, 1, 0)

In this case, we make use of the notation illustrated in Fig. 7 in which complexity p is
fixed to be one. Complexity n is defined to be the number of self-similar iterations of
the basic module of Fig. 1c. Each iteration n = 1, 2, . . . generates different lengths
of bars and cables. The lengths at the ith iteration are:

L L
bi = tan β, si = , i = 1 − n. (28)
2i 2i cos β

Observing the multiscale structure of Fig. 7 it’s clear that the number of bars and the
number of cables at the ith self-similar iteration are

n si = 2i , n bi = 2i−1 . (29)
A Tensegrity Paradigm for Minimal Mass Design of Roofs and Bridges 103

Fig. 7 Adopted notations for forces and lengths of bars and cables for a substructure with generic
complexity (n, p, q) = (n, 1, 0)

In this case the total force applied to the bridge structure is given by (27) and then
the forces in each member become:
F + 2n m d g F + 2n m d g
f bi = , t si = . (30)
2i 2(1+i) sin β

Theorem 7 Consider a substructure bridge with deck mass m d and topology defined
by (28), with complexity (n, p, q) = (n, 1, 0), see Fig. 7. The minimal mass design
under yielding constraints is given by:
 
1  md  
μ∗Y = 1 − n 1 + 2n g 1 + ρ, (31)
2 F

using the optimal angle:

 
1
βY∗ = arctan √ . (32)
1+ρ
√ √
Observe that (31) yields mass 1 + ρ/2 for complexity n = 1 and mass 1 + ρ
for complexity n = ∞. Note from (32), that the optimal angle βY∗ does not depend
upon the choice of n. Indeed, the minimal mass solution under yielding constraints
(31) depends on the material choice ρ (4), the complexity parameter n and the deck
properties. Note that, since the total external force F is a specified constant, the
mass is minimized by the complexity n = 1 if m d = 0. However since m d depends
upon n, the total vertical force including deck mass depends upon n, and the optimal
complexity will be shown to be n > 1 in that case.

Theorem 8 Consider a substructure bridge with topology defined by (28), with com-
plexity (n, p, q) = (n, 1, 0). The minimal mass design under yielding and buckling
constraints is given by:
 
1 + tan2 β B∗
μ∗B = β1 + ηβ2 tan2 β B∗ , (33)
2 tan β B∗
104 G. Carpentieri et al.

using the aspect angle:

   
1 β1
β B∗ = arctan β3 + β1 −1 , (34)
12β2 η β3

where:
 
1  md 
β1 = 1 − n 1 + 2n g , (35)
2 F

 √  
1+2 2 1 md
β2 = 1 − 3n/2 1 + 2n g , (36)
7 2 F

  1/3
β3 = 216β1 β2 η − β1 + 12 324β1 β2 η − 3β1 β2 η
2 2 3 2 4 4 4 2 2
. (37)

4.3 Adding Deck Mass for a Superstructure Bridge

with Complexity (n, p, q) = (n, 0, 1)

In this case, we make use of the notation illustrated in Fig. 8 in which complexity q
is fixed to be one. Complexity n is the number of self-similar iterations of the basic
module of Fig. 1b at different scales. After the ith self-similar iterations, the length
of the bars and cables for i ranging from 1 to n, are:

L L
bi = , si = i tan α. (38)
2i cos α 2

Fig. 8 Adopted notations for forces and lengths of bars and cables for a superstructure with
complexity (n, p, q) = (n, 0, 1)
A Tensegrity Paradigm for Minimal Mass Design of Roofs and Bridges 105

Observing the multiscale structure of Fig. 8 it’s clear that the number of bars and the
number of cables after the ith self-similar iterations are:

n si = 2i−1 , n bi = 2i . (39)

In this case the total force applied to the bridge structure is given by (27) and then
the forces in each member become:
F + 2n m d g F + 2n m d g
f bi = , t si = . (40)
2(i+1) sin α 2i
Theorem 9 Consider a superstructure bridge with topology defined by (38), with
complexity (n, p, q) = (n, 0, 1), Fig. 8. Under a given total vertical force (27), the
minimal mass design under yielding constraints is given by:
 
1  md  
μ∗Y = 1− n 1 + 2n g ρ (1 + ρ), (41)
2 F

using the aspect angle:

 
ρ
αY∗ = arctan . (42)
1+ρ

Theorem 10 Consider a superstructure bridge with topology defined by (38), and

complexity (n, p, q) = (n, 0, 1), see Fig. 8. The structure is loaded with a given
total vertical force (27) and the minimal bar mass, subject to yielding constraints is
given by:
δ1 55/4
μ∗B = + ηδ2 , (43)
2 4
using the aspect angle:

1
α ∗B = arctan , (44)
2
where:
 
1 md  1
δ1 = 1 + 2n g 1− n , (45)
2 F 2

 √   
√ 1+2 2 md 1
δ2 = 2 1 + 2n g 1 − 3n/2 . (46)
7 F 2

The proofs of Theorems 9 and 10 are similar to the proofs of Theorems 7 and 8
in Sect. 4.2 and can be founded in [6].
106 G. Carpentieri et al.

4.4 Penalizing Complexity with Cost Considerations: Adding

Joint Mass

Theorem 7, for m d = 0, leads

√ to an optimal complexity n = 1 which corresponds to
a minimal mass equal to 1 + ρ/2. As complexity n approaches √ infinity, instead,
the mass given in (31), for m d = 0, go to a limit equal to 1 + ρ. However, the
addition of the deck mass in Theorem 7 switches the optimal complexity from n = 1
to n = ∞, so small complexities n are penalized by massive √ decks. Also in this
latter case, the resulting optimal minimal mass is then 1 + ρ, as can be verified
looking the (31) or considering that as n goes to infinity the deck mass given in (24)
approaches zero. As a matter of fact, neither n = 1 or n = ∞ are believable solutions
due to practical reasons: the first solution leads only to a single force at the middle of
the span, the second solution leads to an infinite number of joints and connections.
The minimal masses obtained from (31) with or without deck correspond to perfect
massless joints. The addition of the joint masses to a tensegrity structure with n n
nodes, as illustrated in [5], leads to the following total normalized mass:

μ∗Y,tot = μ∗Y + μ∗d + Ωn n . (47)

Let $ j be the cost per kg of making joints and let $b be the cost per kg of making
bars. Then define Ω = $b /$ j . For perfect joints Ω = 0, for rudely made low cost
joints $ j is small and Ω is larger. Hence Ω is also approximatively the ratio of
material cost per joint divided by material cost per structural member being joined.
Consider the minimal masses of the substructure bridge (μ∗Y ) constrained against
yielding, for the cases with or without deck, see Eq. (31). Assume steel material for
cables, bars and deck beams and set F = 1 N, L = wd = 1 m. Without deck the
optimal aspect angle βY∗ (32) is 35.26◦ . For the case with neither deck nor joint√
mass,
∗
the optimum complexity n is 1, which corresponds to an optimal √ mass μY = 2/2.
As n approaches infinity the mass tends to a limit equal to 2, which is also the
optimal mass for the case with deck mass and perfectly manufactured joints, since
μ∗d approaches zero for n → ∞. Note that with the addition of joint masses as
illustrated in (47), the optimal complexity n ∗ can become a finite value. The above
procedure can be also used for the design under buckling constraints.
Figures 9 (for yielding) and Fig. 10 (for buckling) show the total minimal masses
obtained by using (47). In both Figs. 9 and 10 we also show with red curves the
minimal mass of substructures or superstructures only. In either case, the total mass
of the structure with deck (but no joint mass), is shown by black continuous lines in
Figs. 9 and 10, reaching minimum for an infinite complexity n. It is worth nothing
that, for infinite n, the mass of the deck is zero and the total minimum mass is just
the mass of the bridge structure. Then, with the dotted and dashed lines, we show
that a finite optimal complexity can be achieved if the joint’s masses are considered.
From Fig. 9 note that the minimal mass (μ ∼ = 21) bridge has complexity n = 11 for
Ω = 0.002, and has minimal mass μ ∼ = 15 with complexity n = 12 for Ω = 0.001.
Economic costs would decide if saving 25 % structural mass is worth the extra cost
of improving the joint precision by a factor of 2.
A Tensegrity Paradigm for Minimal Mass Design of Roofs and Bridges 107

Fig. 9 Optimal masses under yielding of the substructures and superstructure (red curve) and total
optimal mass with deck and different joint factors (dashed and dottled curves) for different values
of the complexity n (steel for bars, cables, deck, F = 1 N, L = wd = 1 m)

Fig. 10 Optimal masses under buckling of the substructures (left) and superstructure (right) (red
curves) and total optimal masses with deck and different joint factors (dashed and dotted curves)
for different values of the complexity n (steel for bars, cables, deck, F = 1 N, L = wd = 1 m)

5 Design of a Deployable Bridge for Solar Energy

Harvesting on Water Canals

From open canals that bring water to cities all over the world, water evaporation
represents a significant loss of water. This section designs a minimal mass cover for
water canals while using solar panels as the cover. This is not a new idea. Since 2012
India has built solar-panel-covered canals (Gujarata, India, 2012, [19]). The truss
structure used in Gujarata is massive, threatening the economic survivability of the
project. Here we design a minimal mass cover for such canals.
108 G. Carpentieri et al.

The optimal complexity of a deployable support structure for a solar panel cover-
ing of water canals is derived in the following, along with deployable schemes which
are useful for construction, repairs, for sun following, and for servicing. It is shown
that the minimal structure naturally has deployable features so that extra mass is
not needed to add the multifunctional features. The design of bridge structures with
tensegrity architecture will show an optimal complexity depending only on mate-
rial choices and external loads. The minimization problem considers a distributed
load (from weight of solar panels and wind loads), subject to buckling and yield-
ing constraints. The result is shown to be a Class 1 tensegrity substructure (support
structure only below the deck). These structures, composed of axially-loaded mem-
bers (tension and compressive elements), can be easily deployable and have many
portable applications for small spans, or they can be easily assembled for prefabri-
cated component parts for large spans. The focus of this section is an application
of these minimal mass tensegrity concepts to design shading devices to prevent or
reduce evaporation loss, while generating electric power with solar panels as the
cover. While the economics of the proposed designs are far from finalized, we show
a technical solution that uses the smallest material resources, and shows the technical
feasibility of the concept.

5.1 Description of the Model

In this example, we will assume buckling as a mode of failure of compressive mem-

bers since it has been shown in [6] that buckling is the mode of failure in most of the
practical cases.
In the previous sections we provided a theory to minimize the sum of deck mass,
structural mass, and joint mass. The solution is a Class 1 tensegrity structure (com-
pressive members do not make contact) with an optimal complexity (optimal number
of structural members) that is finite. That is, the optimal structure is not a contin-
uum (in contrast to the Michell truss, [18]) but a discrete structure with an optimal
number of elements. This optimal number depends on material choice, the span, and
the external load. This optimal bridge has no structure above the horizontal line (we
call this a substructure bridge). This example assures that the most efficient struc-
ture does not extend above horizontal, making it ideal for our proposed solar array
surface, since the surface is horizontal, and does not generate any shadows on the
solar panels.
For a water canal application, Fig. 11 shows a 3D deployable flat roof made of
repetitive 2D substructure bridges with multiscale topology defined in Fig. 7. Each
planar substructure bridge (Fig. 7) is constrained with two fixed hinges at both ends
(in practice these hinges might be pulleys that allow roll-up during construction or
repair). As illustrated in Fig. 11, this module can be replicated (along the longitudinal
direction) to build a deployable three-dimensional structure able to carry vertical
loads distributed on the horizontal plane of the solar array. Figure 12 shows a possible
application of this module to water canals.
A Tensegrity Paradigm for Minimal Mass Design of Roofs and Bridges 109

Fig. 11 Different configurations of a deployable solar roof for water canals: a open onfiguration,
b transition between open/closed configurations, c closed configuration

Fig. 12 Schematic of a deployable tensegrity system with solar panel

110 G. Carpentieri et al.

(a) (b)

Fig. 13 Details of the canal structure: a deck system, b deformed shape of the deck cross cables
subjected to the solar panel force

The bridge structures are stabilized out of the plane with a set of longitudinal cables
as illustrated in Fig. 12. In particular diagonal cables and horizontal longitudinal
cables (the magenta element showed in Section B-B of Fig. 12) are used to stabilize
out of plane vertical movement.
The deck is composed of different orders of cables (refer to Figs. 12 and 13):

• longitudinal cables: the elements connecting each tensegrity bridge unit along the
length of the canal;
• transversal cables: the elements of each tensegrity bridge lying on the transversal
direction;
• cross cables: the elements that directly carry the solar panel loads and transfer
their weight to the bridge structures.

The total vertical force Ftot can be computed designing the deck diagonal cables
represented in Fig. 13. These cables directly support two different solar panel modules
of sizes by wd /2 (see Fig. 13). We design these cables assuming that, at the fully-
deployed configuration of the structure, the deck diagonal cables are inclined at a
fixed angle αd with respect to the horizontal (Fig. 13). At this configuration the length
and the tensile force in each deck diagonal cable are:

fp wd2 + 2
td = , sd = . (48)
4 sin αd 2 cos αd
A Tensegrity Paradigm for Minimal Mass Design of Roofs and Bridges 111

We can compute the total mass of the deck diagonal cables as:

ρd ρd f p wd2 + 2
m d = 4 td s d = . (49)
σd σd 2 sin αd cos αd

The normalized total mass μ∗d and the total force Ftot can be computed with (26)
and (27), respectively.
The final total mass to be optimized is then the summation of the mass of the
bridge structure (33), the total mass of the deck (26) and the mass of the joints, Ωn n .

5.2 Numerical Results and Discussion

Let us now focus our attention on numerical results regarding the optimal design
of real-life roof structures featuring different complexities n. The examined struc-
tures show the following design data: L = 30.48 m, F = 12 kN, wd = 4.88 m,
αd = 1◦ , and the material properties of steel for bars (ρ = 7862 kg/m3 ; σ =
6.9 × 108 N/m2 ; E = 2.06 × 1011 N/m2 ) and Spectra® for cables (ρ = 970 kg/m3 ;
σ = 2.7 × 109 N/m2 ; E = 120 × 109 N/m2 ). We investigate on the optimal values
of the following parameters: μ∗B , μ∗tot and β B∗ , which respectively denote the dimen-
sionless minimal mass of a single bridge unit; the dimensionless minimal mass of the
overall system formed by the bridge and the deck; and the optimal aspect angle of
the bridge structure, under combined yielding and buckling constraints. The optimal
angle β B∗ of the examined structures can be computed from Eq. (34), and/or the plots
in Fig. 14. It is easy to verify that the minimal mas structure shows a markedly stream-
lined profile with β B∗ = 2.18◦ (Fig. 14-left). The global minimum of μ∗tot is attained
in correspondence with the complexity n ∗ = 3, both for Ω = 0 (μ∗tot ∼ = 23.07;
m ∗tot ∼
= 3.0318 kg, cf. Table 1), and for Ω > 0 (Fig. 14-right). The optimal design
leads to 0.10 kg mass per cables and 0.02 kg mass per bars per meter of the canal
lengthwise span (cf. Table 1).

Fig. 14 Left dimensionless masses μ B versus aspect angle β B ; right dimensionless total masses
μtot versus complexity n (η = 7569.04)
112 G. Carpentieri et al.

Table 1 Optimal masses μ∗B (33) and μ∗tot and optimal aspect angles β B∗ (34) of substructure
bridges with steel bars and Spectra® cables, under combined yielding and buckling constraints (B),
for different complexities n
n p Ftot (N) β B∗ (◦ ) μ∗B μ∗tot m ∗tot (kg)
1 1 12019.39 2.06 10.4357 25.4791 3.3480
2 1 12010.97 2.13 15.1186 23.6251 3.1044
3 1 12007.50 2.18 17.2522 23.0724 3.0318
4 1 12006.35 2.21 18.2414 23.1662 3.0441
5 1 12006.03 2.23 18.7080 23.3822 3.0725

6 Concluding Remarks

This paper provides closed form solutions (analytical expressions) for planar minimal
mass tensegrity bridge designs. The forces, locations, and number of members are
optimized to minimize mass subject to buckling and yielding constraints for a planar
structure with fixed-hinge/fixed-hinge boundary conditions.
We present the optimal complexity of the substructure bridge that minimizes the
sum of structural mass, deck mass and joint mass. Making better joints (less joint
mass) results in higher optimal complexity and less mass. So the economic tradeoff
between material cost of the truss structure and costs of making better joints will
lead to the proper trade between mass and labor costs.
We also define a 3D deployable tensegrity structure made of repetitive planar
substructure bridges (spanning the canal in the transversal direction) conveniently
stabilized out of plane with a set of cables, in both the transversal and the longitudinal
direction of the canal. Each planar structure has a self-similar fractal type of topology
generated by the complexity parameter n. The minimal mass solution yields com-
plexity n ∗ which depends upon material properties. Moreover, the topology of the 3D
structure is function of canal width (L), aspect angle (β) of the substructures bridges,
longitudinal aspect angle (αd ) governing the deploy-ability of the structure, the dis-
tance between consecutive repetitive structures in the longitudinal direction (wd ).
The design occupies much less volume and mass than the designs for the most
advanced attempts at energy production and shading over water canals (see Gujarata,
India, [19]). Formulas are given which will allow economic tradeoffs between mate-
rial costs of the structure, the labor cost (assuming price per joint is inversely propor-
tional to mass of the joint) of making more refined joints, and the choice of material
(steel, Spectra® , or other). Implicit in these tradeoffs, the optimized complexity n ∗
of the structure is derived to allow economic decisions on the number of components
(bars and cables) that will minimize mass for the given choice of material and joint
costs.
Numerical and experimental studies on the dynamics of these structures will
follow in subsequent work to impose further design constraints on stiffness issues
(vibrational frequencies, mode shapes [20, 21], displacements for high winds con-
A Tensegrity Paradigm for Minimal Mass Design of Roofs and Bridges 113

ditions, etc.), but the capability of all these choices and adjustments are within the
free parameters of the designs in this paper. The subsequent dynamics approach
will evaluate the value (economics and performance tradeoffs) the use of feedback
control for the deployable and service functions, or to adjust the stiffness of the
structure (varying the prestress of the cables) to modify stiffness or damping after
storm damage.

References

1. Skelton, R. E., & Nagase, K. (2012). Tensile tensegrity structures. International Journal of
Space Structures, 27, 131–137.
2. Skelton, R. E., & de Oliveira, M. C. (2010). Optimal complexity of deployable compressive
structures. Journal of the Franklin, I(347), 228–256.
3. Skelton, R. E., & de Oliveira, M. C. (2010). Optimal tensegrity structures in bending: the
discrete Michell truss. Journal of the Franklin, I(347), 257–283.
4. Nagase, K., & Skelton, R. E. (2014). Minimal mass tensegrity structures. Journal of the Inter-
national Association for Shell and Spatial Structures, 55(1), 37–48.
5. Skelton, R. E., & de Oliveira, M. C. (2010). Tensegrity systems. New York: Springer.
6. Carpentieri, G., Skelton, R. E., & Fraternali, F. (2015). On the minimum mass and optimal
complexity of planar tensegrity bridges, Internal Report of the University of California, San
Diego, Mechanical and Aerospace Engineering, No. 1-2014. www.fernandofraternaliresearch.
com/publications/arxiv_tensegrity_bridges_theory_2014.pdf. Accessed 23 June 2015.
7. Skelton, R. E., Fraternali, F., Carpentieri, G., & Micheletti, A. (2014). Minimum mass design of
tensegrity bridges with parametric architecture and multiscale complexity. Mechanics Research
Communications, 58, 124–132, ISSN 0093-6413, doi:10.1016/j.mechrescom.2013.10.017.
8. Koohestani, K. (2012). Form-finding of tensegrity structures via genetic algorithm. Interna-
tional Journal of Solids and Structures, 49, 739–747.
9. Rhode-Barbarigos, L., Jain, H., Kripakaran, P., & Smith, I. F. C. (2010). Design of tensegrity
structures using parametric analysis and stochastic search. Engineering and Computer, 26(2),
193–203.
10. Sakamoto , T., Ferrè, A., & Kubo, M. (Eds.). (2008). From control to design: Paramet-
ric/algorithmic architecture. Actar, Barcelona.
11. Sokóf, T., & Rozvany, G. I. N.: New analytical benchmarks for topology optimization and their
implications. Part I: bi-symmetric trusses with two point loads between supports. Structural
and Multidisciplinary Optimization, 46, 477–486 (2012).
12. Tilbert, A. G., & Pellegrino, S. (2011). Review of form-finding methods for tensegrity struc-
tures. International Journal of Space Structures, 18, 209–223.
13. Yamamoto, M., Gan, B. S., Fujita, K., & Kurokawa, J. (2011). A genetic algorithm based
form-finding for tensegrity structure. Procedia Engineering, 14, 2949–2956.
14. Fraternali, F., Marino, A., El Sayed, T., & Della Cioppa, A. (2011). On the structural shape
optimization through variational methods and evolutionary algorithms. Mechanics of Advanced
Materials and Structures, 18, 224–243.
15. Fraternali, F., Farina, I., & Carpentieri, G. (2014). A discrete-to-continuum approach to the cur-
vatures of membrane networks and parametric surfaces. Mechanics Research Communications,
56, 18–25.
16. Phocas, M. C., Kontovourkis, O., & Matheou, M. (2012). Kinetic hybrid structure development
and simulation. International Journal of Architectural Computing, 10(1), 67–86.
17. Fraternali, F., & Carpentieri, G. (2013). On the correspondence between 2D force networks
and polyhedral stress functions. International Journal of Space Structures, 29(3), 145–159.
114 G. Carpentieri et al.

18. Michell, A. G. M. (1904). The limits of economy of material in frame-structures. Philosophical

Magazine, 8, 589–597.
19. Kahn, M., & Longcore, T. (2014). A Feasibility Analysis of Installing Solar Photovoltaic Pan-
els Over California Water Canals. UCLA Institute of the Environment and Sustainability,
Los Angeles, CA. http://www.environment.ucla.edu/perch/resources/files/adeptfinalreport1.
pdf. Accessed 22 July 2014.
20. Carpentieri, G., Modano, M., Fabbrocino, F., & Fraternali, F. (2015). Optimal design and
dynamics of truss bridges. In COMPDYN 2015—5th ECCOMAS Thematic Conference on
Computational Methods in Structural Dynamics and Earthquake Engineering (pp. 1731–1740).
21. Modano, M., Fabbrocino, F., Gesualdo, A., Matrone, G., Farina, I., & Fraternali, F. (2015).
On the forced vibration test by vibrodyne. In COMPDYN 2015—5th ECCOMAS Thematic
Conference on Computational Methods in Structural Dynamics and Earthquake Engineering
(pp. 209–217).
Universal Meshes for the Simulation
of Brittle Fracture and Moving
Boundary Problems

Maurizio M. Chiaramonte, Evan S. Gawlik, Hardik Kabaria

and Adrian J. Lew

Abstract Universal meshes have recently appeared in the literature as a computa-

tionally efficient and robust paradigm for the generation of conforming simplicial
meshes for domains with evolving boundaries. The main idea behind a universal
mesh is to immerse the moving boundary in a background mesh (the universal mesh),
and to produce a mesh that conforms to the moving boundary at any given time by
adjusting a few elements of the background mesh. In this manuscript we present the
application of universal meshes to the simulation of brittle fracturing. To this extent,
we provide a high level description of a crack propagation algorithm and showcase
its capabilities. Alongside universal meshes for the simulation of brittle fracture,
we provide other examples for which universal meshes prove to be a powerful tool,
namely fluid flow past moving obstacles. Lastly, we conclude the manuscript with
some remarks on the current state of universal meshes and future directions.

1 Introduction

Predicting and understanding the behavior of a propagating fracture has applications

in a broad spectrum of disciplines. Perhaps the most renowned are applications in
civil, mechanical, and aerospace engineering for the safe design of structural and
mechanical components. More recently, a new wave of interest in understanding
fracture propagation has risen due to the insurgence of hydraulic fracturing for the

M.M. Chiaramonte (B) · E.S. Gawlik · H. Kabaria · A.J. Lew

Department of Mechanical Engineering and Institute for Computational
and Mathematical Engineering, Stanford University, Stanford, USA
e-mail: [email protected]
E.S. Gawlik
e-mail: [email protected]
H. Kabaria
e-mail: [email protected]
A.J. Lew
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 115

K. Weinberg and A. Pandolfi (eds.), Innovative Numerical Approaches
for Multi-Field and Multi-Scale Problems, Lecture Notes in Applied
and Computational Mechanics 81, DOI 10.1007/978-3-319-39022-2_6
116 M.M. Chiaramonte et al.

recovery of shale gas, as well as for engineering geothermal reservoirs. Alongside

hydraulic fracturing, the practice of abyssal sequestration [17] for the disposal of
radioactive waste also necessitates numerical tools capable of predicting the behav-
ior of fluid driven fractures. Beyond engineering, the modeling of fracturing finds
relevance in geophysics, for example for the prediction of ice-sheet separation and
its effect on global climate. Due to the pervasive nature of fracture mechanics in
many disciplines, there is a need for a deeper understanding of fracture evolution
accounting for the three-dimensionality of the fracturing process. These applica-
tions motivate the current efforts towards the creation of robust and computationally
efficient numerical methods to approximate the solutions of such fracture evolution
models.
From the numerical standpoint, one of the crucial challenges faced in this particu-
lar class of problems is the approximation of the evolving displacement discontinuity,
which is the focus of the work presented here. Several approaches have been pro-
posed in the literature to address this challenge. Albeit a comprehensive literature
review is beyond the scope of this manuscript, a very broad classification of the pre-
dominant classes of methods capable of handling the evolution of a few cracks can
be arguably categorized into basis-enriching methods or mesh-conforming methods.
Additionally it is worthwhile mentioning numerical methods to approximate solu-
tions of regularized theories of fracture. These theories, by assigning a finite width
to the fracture, circumvent the need to explicitly track the crack geometry. Some
examples are phase field methods [6], and Michael Ortiz’s own contributions on
eigenfracture [37] and eigenerosion [29, 31], to name a few. Also worth mention-
ing are methods for situations in which massive fragmentation appears, such as the
seminal contributions by Michael Ortiz based on cohesive elements [7, 27, 28, 30,
36].
Basis-enriching methods, such as the Extended (XFEM) [4, 24] and Generalized
(GFEM) [2, 22] finite element methods, endow the finite dimensional subspace
with discontinuous functions. These methods circumvent the need to accommodate
the evolving displacement discontinuity in the domain subdivision by implicitly
representing it through the discontinuous basis functions. Numerical integration can
be rather challenging, and, for problems such as hydraulic fracturing, when coupled
governing equations need to be solved on the crack faces, these methods fail to
provide a quality subdivision of the crack geometry. An example of the latter is
illustrated in Fig. 1.
Alternatively, conforming methods envision generating a subdivision which
accommodates exactly the evolving crack geometry. By ensuring that the mesh for
the domain always conforms to the crack path, any displacement discontinuity along
the crack is easily introduced. While the idea is simple it is nonetheless powerful.
The robustness of this class of methods is limited by the generation of a quality
conforming subdivision, a process which can be computationally demanding and
prone to failure. Some examples of this approach are locally re-meshing methods
as encountered in [3, 5, 32] as well as r -adaptive procedures as proposed in [1, 12,
23, 39]. Related to the latter are finite element spaces with embedded discontinuities
[18, 26].
Universal Meshes for the Simulation of Brittle Fracture … 117

Fig. 1 An arbitrary cut (in red) through a quality tetrahedralization, representing an imaginary
fracture, yields a poor discretization of the fracture faces

Herein we present a few ideas for the simulation of brittle fracture that fall in
the latter category of conforming mesh methods by taking advantage of universal
meshes. Universal meshes are a paradigm for mesh generation that envision the use
of a single “background” mesh (the universal mesh) whose vertices closest to the
crack geometry are perturbed to obtain a subdivision conforming to it. An example of
such a perturbation is illustrated in Fig. 2. Because the same mesh can be deformed to
conform to the geometry of a class of cracks, we say that the mesh is universal for such
a class. The salient features of the method are its robustness, computational efficiency,
and the mesh-independence of the solutions it provides (in fact, convergence).
We provide a description of the algorithm of universal meshes in Sect. 2 followed
by the presentation of the algorithm for the simulation of brittle fracture in Sect. 3.
Later, in Sect. 4, we highlight some applications of universal meshes beyond brittle
fracture. We conclude the manuscript on some recent developments of universal
meshes in three dimensions in Sect. 5.

2 Universal Meshes

We introduce the basic algorithmic ideas behind a universal mesh next. For concrete-
ness, we focus the description on the aspects relevant to crack propagation, bearing
in mind that similar ideas apply equally well to other classes of evolving domains,
such as those encountered in fluid-structure interaction, as discussed later in Sect. 4.

2.1 Algorithm

To illustrate the discretization of an evolving domain with a universal mesh, we

consider in this section the problem of triangulating a domain Ω(t) ⊂ R2 , 0 ≤ t ≤ T ,
which contains an evolving crack. In other words,
118 M.M. Chiaramonte et al.

Fig. 2 Using a universal mesh, a triangulation conforming to a crack (right) is constructed by

immersing the crack in a background triangulation Th (left) and adjusting a few of its elements.
This is accomplished by selecting a set of edges Γh in the background triangulation that lie near the
crack Γ , mapping them onto Γ via the closest point projection, and relaxing a few nearby vertices
to ensure the quality of the resulting triangulation

Ω(t) = D \ Γ (t)

where D ⊂ R2 is an open, bounded, polygonal domain and Γ (t) ⊂ D is a simple

open rectifiable curve.
Let Th be a triangulation of D, hereafter referred to as the universal mesh. We
use h to denote the maximum diameter of an element of Th . We do not assume that
the universal mesh conforms to Γ (t) at any given time; in general, Γ (t) may cut
through elements of Th arbitrarily, as in Fig. 2. Intuition would suggest, however,
that a conforming mesh can be constructed by adjusting a few elements of Th in a
neighborhood of Γ (t). This is the basic observation behind universal meshes.
To construct such a conforming mesh from Th , the following algorithm is adopted.
First, a subset of edges in Th lying near Γ (t) is identified. We denote the union of
these edges Γh (t). Next, these edges are mapped onto Γ (t) via the closest point
projection π : R2 → Γ (t), with a suitable modification that places nodes precisely
at the crack tips. Finally, the positions of nearby nodes are adjusted via a relaxation
step that ensures the quality of the resulting triangulation.
The precise choices for the edges constituting Γh (t) and the nodal adjustments
adopted during relaxation are detailed in [35]. Briefly, Γh (t) consists of positive
edges of positively cut triangles in Th . To define these notions, one designates an
orientation (positive or negative) for points in a neighborhood of Γ (t). A triangle
in Th is called positively cut if it has two nodes on the positive side of Γ (t) and
one on the negative side. An edge is then called a positive edge if it belongs to a
positively cut triangle and its endpoints both lie on the positive side of Γ (t). A minor
modification to Γh (t) is made if a triangle in Th has three nodes on Γh (t); see [35]
for details.
A key feature of the algorithm summarized above is its robustness. That is, the
algorithm returns a valid mesh, for both the crack and the domain, with the quality
of the elements bounded from below independently of the mesh size, provided that
Universal Meshes for the Simulation of Brittle Fracture … 119

three conditions are satisfied: (1) the background mesh is sufficiently refined in a
neighborhood of Γ (t), (2) all positively cut triangles in Th are acute, and (3) the
curve Γ (t) is sufficiently smooth. This statement was proved for a domain with C 2
boundary (no cracks) in [33, 34]. The numerical examples strongly suggest that this
should also be possible for domains with interfaces, such as cracks.

3 Simulating Brittle Fracture with Universal Meshes

The obvious way in which a Universal Mesh is useful for the simulation of a prop-
agating crack is by providing a mesh perfectly conforming to the crack at each step
of its evolution. However, there are advantages that are less evident: the conforming
mesh enables us to compute stress intensity factors to any order of accuracy, and
the few mesh changes from step to step make it possible to retain much of the data
structures in the computer implementation. The accuracy in the computation of the
stress intensity factors is a determinant factor in observing convergence of the crack
evolutions for “reasonable” mesh sizes.
In the following we present a numerical algorithm for the simulation of crack
evolution in a restricted set of problems, as introduced in [35]. The presentation of
the algorithm is followed by examples, including the formation of oscillatory crack
paths in quenched plates, which requires very accurate stress intensity factors to
converge.

3.1 Problem Statement

We consider the problem of an always propagating crack in an elastic medium, as

defined next. We parametrize the crack evolution by the crack length  ∈ [0 , max ]
and we denote by C () the crack tip position for the crack of length . Hence the crack
set is given by C ([0 , ]). The domain occupied by the cracked domain is denoted
by Ω() and its boundary is decomposed into a portion over which displacements
are prescribed, ∂d Ω(), and a portion over which boundary tractions are prescribed,
∂τ Ω(). Further we let C ([0 , ]) ⊆ ∂τ Ω().
The problem statement then reads: find the deformation u(·, ) : Ω() → R2 , the
load scaling factor C : [0 , max ] → R, and the crack set C ([0 , max ]) such that the
following holds for all  ∈ (0 , max ]:

−∇ · (C : ∇u) = b(), on Ω(),

(C : ∇u)n = t(), on ∂τ Ω(),
u = g(), on ∂d Ω(),
120 M.M. Chiaramonte et al.

K I [u] = K c ,
K I I [u] = 0,
C ([0 , − ]) ⊂ C ([0 , ]), ∀− < ,

where b() = C()b(), t() = C()t(), g() = C()g(). Here b(·, ) :

Ω() → R2 , t(·, ) : ∂τ Ω() → R2 , and g(·, ) : ∂d Ω() → R2 are arbitrary func-
tions representing the “shape” of the body forces and boundary conditions. Effec-
tively, for every crack length we know the “shape” of the applied body force (b),
boundary tractions (t) and displacements (g), and we must solve for the linearly scal-
ing coefficient C() such that the condition K I [u] = K c is always met, where K I [u]
and K I I [u] are the mode I and II stress intensity factors. The condition K I I [u] =
0 dictates the direction of crack propagation following the Principle of Local
Symmetry [19].
The “always propagating crack” problem circumvents some of the more delicate
issues in crack propagation, such as crack arrest and catastrophic crack propagation,
regularity of the crack path, and competition among multiple cracks. The algorithm
introduced next is applicable to this simpler class of problems.

3.2 Crack Propagation Algorithm

There are three critical steps in the computation of the evolution of brittle crack
paths: (1) the generation of a triangulation that conforms to the cracked domain, (2)
the calculation of the elasticity fields, and (3) the evaluation of the stress intensity
factors for curvilinear cracks. The steps are highlighted in Fig. 3.
We construct a triangulation that conforms to each cracked domains from a uni-
versal mesh, as described in Sect. 2. To ensure that the elasticity fields are sufficiently
resolved, we draw on a class of finite element methods for domains with corners or
cracks that retain optimal convergence rate for elements of any order in the face of
singular solutions [10], in contrast to standard methods or methods with enrichments.
Lastly, given that we count with higher order solutions of the elasticity fields, we
employ a family of interaction integrals specifically designed to handle curvilinear
cracks [9] which yield stress intensity factors that converge rapidly to the exact ones
(namely, they converge with twice the rate of convergence of the derivatives of the
solution of the elasticity fields; a motivation to the use higher-order finite element
methods).
We approximate the crack set C ([0 , ]) by a cubic spline interpolant Γh of a
(− )/Δ
finite set of crack tips A = {xn }n=0 0 , where 0 is the initial crack length and
Δ > 0 is a crack discretization parameter. For a discrete crack Γh ,  indicates the
chord length (the length along the polygonal line formed by points in A , plus the
initial crack length 0 ) instead of its length. At any value of  ≥ 0 , the algorithm
proceeds through the following three steps:
Universal Meshes for the Simulation of Brittle Fracture … 121

Fig. 3 The critical steps in the crack advancement algorithm

1. Generate a conforming triangulation to the crack Γh .

2. Find uh () ≈ u() and C h () ≈ C().
3. Advance the crack in the direction d(K IhI [uh ]/K Ih [uh ]) by Δ (namely A+Δ =
A ∪ {x(−0 )/Δ + d(K IhI [uh ]/K Ih [uh ]) Δ}).
Here the direction d : R → S 1 is chosen such that an infinitesimally short kink at
the chosen angle satisfies K I I = 0 up to first order in the kink angle itself. Evaluating
d in this way sidesteps the computationally intensive alternative of explicitly solving
for the direction d, but it likely restricts the order of convergence of the algorithm.

3.3 Examples

We next showcase the application of the algorithm for the simulation of brittle fracture
to two examples. The first example is a crack propagating along a circular arc,
which we compare against an exact solution, and the second is a crack propagating
in a perforated plate undergoing three-point bending, which we compare against
experimental results. We also show (preliminary) results on the application of a
slight modification of this algorithm to the computation of crack paths in a rapidly
quenched plate [8].
122 M.M. Chiaramonte et al.

3.3.1 A Crack Propagating Along a Circular Arc

The displacement and stress fields of an infinite medium that contains a crack shaped
as a circular arc subjected to far-field stresses and traction-free faces was computed
in [25]. The corresponding stress intensity factors can be found in [11]. We use this
solution to construct a loading history that, when is applied as Dirichlet boundary
conditions to a square-shaped domain, as illustrated in Fig. 4a, causes the crack to

(a) (b)

(c)

Fig. 4 The problem of a crack propagating along a circular arc. By imposing appropriate boundary
conditions g(), with the knowledge of the analytical solution as a function of the angle subdued
by the crack, and hence , we can guide the crack to propagate along a circular arc. This prob-
lem provides a benchmark to establish convergence of the computed crack paths. Four level of
refinements were used to perform the convergence of the computed crack path and convergence
rates were observed to be of the order O (h 1 ). a Modeled problem, b convergence of the solution,
c convergence of the crack path
Universal Meshes for the Simulation of Brittle Fracture … 123

propagate along a circular arc. For details on the construction of such a loading
history we refer the interested reader to [35].
Figure 4a shows a square-shaped domain Ω with a pre-existing crack of radius
R = 2 and angular span ϑ0 = π/8.The analyses were carried out with four pro-
gressively refined universal meshes. The coarsest background mesh as well as the
conformed mesh are shown in Fig. 5, and their refinements were constructed by
recursively subdividing each triangle of the background mesh into four similar
ones. The ratio Δ/ h ≈ 2 was kept constant over all simulations, where as usual
h denotes the maximum diameter of an element in a triangulation. As shown in
Fig. 4b, the crack path converges to a circular arc, and the convergence curves for the
L p ([0 , max ]), p = 2, ∞ and H 1 ([0 , max ]) norms are shown in Fig. 4c. Notably,
convergence of the tangents to the crack path is also obtained.
This simple, nonetheless illustrative example, suggests that the algorithm is indeed
convergent, and hence that the computed paths are largely independent of the chosen
mesh.

Fig. 5 In reference to the problem of a crack propagating along a circular arc, as presented in
Sect. 3.3.1, we showcase the universal mesh used for the entire crack propagation simulation (left).
Namely, the single background mesh (left) was used to generate a conforming triangulation to the
family of cracks {Γh } . On the right we showcase the conformed triangulation at the end of the
simulation with the background mesh in red
124 M.M. Chiaramonte et al.

3.3.2 Perforated Plate

We next present the problem of a perforated plate undergoing three-point-bending.

The problem setup is illustrated in Fig. 6. We performed the simulations for three con-
figurations of the initial crack position (d) and length (0 ). The values are tabulated in
Fig. 6. In Fig. 7 we illustrate a universal mesh employed for one of the three simula-
tions. It is worthwhile to note the adaptive nature of the background triangulation; in
fact universal meshes can be easily adopted in conjunction with adaptive refinement.
For each of the three simulations we generated a slightly different universal mesh to
comply with the varying location of the initial crack.
Experimental results for this test setup are available in [5, 20]. The experiments
were performed on polymethyl methacrylate (PMMA) plates. A comparison of the
computed crack paths with digitized points from [5, 20] show a good agreement
with experimental results. Further, relative convergence studies were performed on
the computed crack paths, and the results, that can be found in [35], show a similar
behavior to the one observed in Fig. 4.

Fig. 6 Geometry for a plate with holes

Fig. 7 Universal mesh for a plate with holes

Universal Meshes for the Simulation of Brittle Fracture … 125

3.3.3 Crack Path Instabilities in a Quenched Plate

Lastly we concisely present the problem of a thermally driven fracture in a quenched

plate. The problem consists of a plate of finite width cracked along its center line,
with low toughness (K c ), that, after being heated to temperature θ + , is immersed in an
ice bath at temperature θ − with a constant velocity (v). Refer to Fig. 9 to supplement
the above description (Fig. 8).
Depending on the material parameters, the presence of small deviations from the
idealized descriptions above, and the configuration of the experiment, the crack path
is expected to develop oscillations. Figure 10 showcases the results of experiments
performed by [38]. In Fig. 11 we showcase some snapshots of the computed crack
path for one set of inputs. These were computed through a modification to time-
dependent problems of the algorithm introduced here. Details will appear in [8].
Although not shown here, the crack paths are converged up to a small tolerance. In
our experience, this problem benefitted immensely from the high-order computation
of the stress intensity factors; our previous attempts with low-order methods required
excessively refined meshes to begin displaying some form of mesh-independent
results. We hope to use this platform to better understand the underlying physics.

Simulation
Experiment

Fig. 8 Comparison between experimental results and computed crack paths for the problem of
three-point-bending of a perforated plate. Experimental results were digitized from [5, 20]

Fig. 9 Quenched plate problem setup

126 M.M. Chiaramonte et al.

Fig. 10 Representative results of experiments of wavy crack patterns in rapidly quenched plates
[38]. In both cases shown above, crack propagation along a straight crack is unstable. These cases
correspond to different immersion speeds

Time

Fig. 11 Computed evolution of a thermally driven crack in a quenched plate. The contours show
the temperature profile along the crack with dark blue representing θ − and dark red representing
θ+

4 Beyond Brittle Fracture: Moving Boundary Problems

In addition to crack propagation, a variety of problems in science and engineering

involve partial differential equations posed on domains that change with time. Such
problems, collectively referred to as moving-boundary problems, appear in studies of
fluid-structure interaction, phase-transitions, free-surface flows, aeroelasticity, and
biolocomotion, to name a few. In this section, we demonstrate the applicability of
universal meshes to this broader class of problems.

4.1 Examples: Flow Past Moving Obstacles

In the setting of fluid-structure interaction, universal meshes provide a conform-

ing discretization of the fluid domain at all times, allowing finite element spaces of
Universal Meshes for the Simulation of Brittle Fracture … 127

any desired order of accuracy to be used to spatially discretize the Navier-Stokes

equations. This conforming discretization can be made to deform smoothly over
time intervals that are short in comparison to the mesh spacing, thereby allowing
standard numerical integrators to be used to solve the resulting system of ordinary
differential equations. A projector (such as the nodal interpolation operator) is then
used to transfer information between finite element spaces each time nodal positions
change discontinuously. Details of this procedure are given in [13, 16], and rigor-
ous theoretical bounds for a wide class of linear problems guarantee the high-order
nature of the resulting numerical scheme when high-order finite elements are adopted
[14, 15].
As an example, we consider in Fig. 12 the solution of incompressible, viscous
flow past a rotating propeller at Reynolds number Re = 290. We solved the problem
using a universal mesh having adaptive refinement in a neighborhood of the propeller,
together with Taylor-Hood finite elements. Figure 12 shows contours of the vorticity
at two instants in time. The robust nature of the method is patent in this example, as
traditional deforming-mesh methods could easily encounter difficulties with mesh
entanglement upon rotation of the propeller.
As a second example, we consider in Fig. 12 the solution of incompressible,
viscous flow past a pair of NACA0015 airfoils that change their pitch sinusoidally in
time. We solved the problem using a universal mesh together with Taylor-Hood finite
elements. For simplicity, the tips of the airfoils were blunted so that the algorithm
described in Sect. 2 (which applies to smooth geometries) could be applied in its
most basic form. Figure 12 shows contours of the vorticity at two instants in time
corresponding to the maximum and minimum pitch (17◦ and −17◦ , respectively) of
the airfoils.
Finally, we consider the solution of incompressible, viscous flow past an oscil-
lating disk with unit diameter at Reynolds number Re = 185. We solved the prob-
lem using a universal mesh having adaptive refinement near the disk (see Fig. 13a),
together with Taylor-Hood finite elements [16]. The disk’s motion was prescribed
using a sinusoidally varying vertical displacement with amplitude 0.2 and frequency
equal to 0.8 times the natural shedding frequency of a fixed disk of the same diameter.
Figure 13b shows a snapshot of the contours of the vorticity. Figure 13c shows the
observed convergence of the drag and lift coefficient time series under refinement of
the mesh, which were computed via direct integration over the boundary of the disk.

5 Outlook

Clearly for a universal mesh to be useful in engineering practice, it needs to be able

to handle evolving geometries in three-dimensions. We show next some incipient
results in this direction.
128 M.M. Chiaramonte et al.

Fig. 12 Vorticity contours for two representative examples of flow past a moving obstacle. The
simulations consist of incompressible viscous flow, computed using a universal mesh together with
Taylor-Hood finite elements. a Flow past a rotating propeller, b flow past pitching airfoils
Universal Meshes for the Simulation of Brittle Fracture … 129

(a) (b)

(c) 10
-1.5
CD
CL
-2
10

-2.5
10

0
10-0.5 10
h

Fig. 13 Numerical simulation of incompressible, viscous flow past an oscillating disk using a
universal mesh. In a, the background mesh adopted during the simulation is shown. In b, a snapshot
of vorticity contours are shown. In c, the convergence of the drag and lift coefficient time series
under mesh refinement is shown. The reported error E is the square root of the integrated squared
error (Ci (t) − C̄i (t))2 , i = L , D, over the time interval [0, 1], relative to a reference solution C̄i (t)
obtained from a fine mesh with h = 0.145. Nearly quadratic convergence is observed. See [16] for
details

5.1 Universal Meshes for Smooth Three-Dimensional

Domains

The construction of a universal mesh in three dimensions follows the steps described
in Sect. 2. Namely, given a smooth closed surface Γ ⊂ D ⊂ R3 immersed in a mesh
of tetrahedra Th , we first identify a set of faces Γh in Th that lie near Γ . These
faces are then mapped onto Γ via the closest point projection, and nearby nodes are
adjusted via a relaxation step that ensures the quality of the resulting mesh.
In analogy with the algorithm presented in Sect. 2, Γh is chosen as the union
of positive faces of positively cut tetrahedra in Th . A tetrahedron in Th is called
positively cut if it has three nodes on the non-negative side of Γ and one on the
negative side. A face is then called a positive face if it belongs to a positively cut
tetrahedron and all three of its vertices lie on the non-negative side of Γ . The closest
point projection defines a one-to-one mapping between a positive face and its image
on Γ provided that the mesh size is small compared to the local radius of curvature,
and more importantly, provided some dihedral angles in the mesh are acute [21].
130 M.M. Chiaramonte et al.

6
10

5
10
Number of Elements

4
10

3
10

2
10

1
10

0
10
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Quality

Fig. 14 The top row shows the tetrahedral meshes conforming to the geometry of an elephant in
two different postures. They were obtained from the same background universal mesh, discarding
exterior elements. In the middle row we show two cuts of the mesh displayed in the top right. In
the bottom row we showcase the distribution of the quality of the tetrahedra on a logarithmic scale

Some examples that illustrate the use of a universal mesh in three dimensions
are given in Figs. 14 and 15. In Fig. 14, two meshes of tetrahedra conforming to an
elephant undergoing changes in its posture were obtained from a single universal
mesh. In Fig. 15, the same procedure was used to construct conforming meshes of
tetrahedra of a human upper airway.
Universal Meshes for the Simulation of Brittle Fracture … 131

Fig. 15 The two figures in the middle show conforming meshes for two different configurations
of a human upper airway. The figures on the far left and the far right show a cut through each
tetrahedral mesh. We used a CT scan of a patient as an input for this example. The figure on the
far left shows the contours of the velocity field computed by solving the Navier-Stokes equations
inside. Coupling this tool to a solid mechanics analysis code for the upper airway would be useful
to study the collapse of the upper airway, quite often the area of interest for patients diagnosed with
obstructive sleep apnea

5.2 Universal Meshes for Evolving Curves on Surfaces

With an eye towards evolving crack fronts in three dimensions, we show next some
early results on how a universal mesh can conform to a smooth curve drawn over
a smooth surface, triangulating the interior of the curve over the surface, and con-
forming the tetrahedra to the surface and the mesh as well, see Fig. 16. To do so, a
planar parametrization of the surface was constructed, and a smooth approximation
of the given curve immersed in it. A conforming surface triangulation to the curve
was then achieved by using a modification of the algorithm in two dimensions, not
described here, and mapping the resulting planar triangulation back to the surface.
132 M.M. Chiaramonte et al.

10 7

10 6

Number of elements
10 5

10 4

10 3

10 2

10 1

10 0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Quality

Fig. 16 A conforming mesh of the given surface is generated from the background universal mesh,
as shown in the top left and top right figures. We then construct a planar parametrization from the
corresponding surface triangulation, and conform the mapped mesh to the surface in the parametric
planar domain. Mapping then back to the real space we obtain a surface triangulation that conforms
to the curve on the surface, and after a relaxation step of the nodes near the surface to ensure good
quality of the tetrahedra, we obtain the resulting mesh (bottom row). The distribution of the qualities
of tetrahedra in this mesh is shown at the bottom right

Acknowledgments This work was supported by the Office of Technology Licensing Stanford
Graduate Fellowship to Maurizio M. Chiaramonte, the National Science Foundation Graduate
Research Fellowship to Evan S. Gawlik, and the Franklin P. Johnson Jr. Stanford Graduate Fellow-
ship to Hardik Kabaria. Adrian J. Lew acknowledges the support of National Science Foundation;
contract/grant number CMMI-1301396.

References

1. Angelillo, M., Babilio, E., & Fortunato, A. (2012). Numerical solutions for crack growth based
on the variational theory of fracture. Computational Mechanics, 50(3), 285–301.
2. Armando Duarte, C., & Tinsley Oden, J. (1996). An h-p adaptive method using clouds. Com-
puter Methods in Applied Mechanics and Engineering, 139(1–4), 237–262.
3. Azócar, D., Elgueta, M., & Rivara, M. C. (2010). Automatic LEFM crack propagation method
based on local Lepp-Delaunay mesh refinement. Advances in Engineering Software, 41,
111–119.
Universal Meshes for the Simulation of Brittle Fracture … 133

4. Belytschko, T., & Black, T. (1999). Elastic crack growth in finite elements with minimal
remeshing. International Journal for Numerical Methods in Engineering, 45(5), 601–620.
5. Bittencourt, T. N., Wawrzynek, P. A., Ingraffea, A. R., & Sousa, J. L. (1996). Quasi-automatic
simulation of crack propagation for 2D LEFM problems. Engineering Fracture Mechanics,
55, 321–334.
6. Bourdin, B., Francfort, G. A., & Marigo, J.-J. (2000). Numerical experiments in revisited brittle
fracture. Journal of the Mechanics and Physics of Solids, 48(4), 797–826.
7. Camacho, G. T., & Ortiz, M. (1996). Computational modelling of impact damage in brittle
materials. International Journal of Solids and Structures, 33(20–22), 2899–2938.
8. Chiaramonte, M. M., Keer, L. M., & Lew, A. J. (2015). Crack path instabilities in thermoelas-
ticity.
9. Chiaramonte, M. M., Shen, Y., Keer, L. M., & Lew, A. J. (2015). Computing stress intensity
factors for curvilinear cracks. International Journal For Numerical Methods in Engineering.
10. Chiaramonte, M. M., Shen, Y., & Lew, A. J. (2015). The h-version of the method of auxiliary
mapping for higher order solutions of crack and re-entrant corner problem.
11. Cotterell, B., & Rice, J. R. (1965). Slightly curved or kinked cracks. International Journal of
Fracture Mechanics, 16, 155–168.
12. Fraternali, F. (2007). Free discontinuity finite element models in two-dimensions for in-plane
crack problems. Theoretical and Applied Fracture Mechanics, 47(3), 274–282.
13. Gawlik, E. S., & Lew, A. J. (2014). High-order finite element methods for moving boundary
problems with prescribed boundary evolution. Computer Methods in Applied Mechanics and
Engineering, 278, 314–346.
14. Gawlik, E. S., & Lew, A. J. (2015). Supercloseness of orthogonal projections onto nearby finite
element spaces. Mathematical Modelling and Numerical Analysis, 49, 559–576.
15. Gawlik, E. S. & Lew, A. J. (2015). Unified analysis of finite element methods for problems
with moving boundaries. SIAM Journal on Numerical Analysis.
16. Gawlik, E. S., Kabaria, H., & Lew, A. J. (2015). High-order methods for low reynolds number
flows around moving obstacles based on universal meshes. International Journal for Numerical
Methods in Engineering.
17. Germanovich, L., Murdoch, L. C., & Robinowitz, M. (2014). Abyssal sequestration of nuclear
waste and other types of hazardous waste.
18. Giacomini, A., & Ponsiglione, M. (2006). Discontinuous finite element approximation of qua-
sistatic crack growth in nonlinear elasticity. Mathematical Models and Methods in Applied
Sciences, 16(01), 77–118.
19. Gol’dstein, R. V., & Salganik, R. L. (1974). Brittle fracture of solids with arbitrary cracks.
International Journal of Fracture, 10(4), 507–523.
20. Ingraffea, A. R., & Grigoriu, M. (1990). Probabilistic fracture mechanics: A validation of
predictive capability. Technical report, Cornell University.
21. Kabaria, H., & Lew, A. J. (In preparation). Universal meshes for smooth surfaces with no
boundary in three dimensions. International Journal for Numerical Methods in Engineering.
22. Melenk, J. M., & Babuška, I. (1996). The partition of unity finite element method: Basic
theory and applications. Computer Methods in Applied Mechanics and Engineering, 139(1–4),
289–314.
23. Miehe, C., & Gürses, E. (2007). A robust algorithm for configurational-force-driven brittle
crack propagation with r-adaptive mesh alignment. International Journal for Numerical Meth-
ods in Engineering, 72(2), 127–155.
24. Moes, N., Dolbow, J., & Belytschko, T. (1999). A finite element method for crack growth
without remeshing. International Journal for Numerical Methods in Engineering, 46,
131–150.
25. Muskhelishvili, N. I. (Ed.). (1977). Some Basic Problems of the Mathematical Theory of Elas-
ticity: Fundamental Equations, Plane Theory of Elasticity, Torsion, and Bending (translated
from Russian) (2nd ed.). Leyden, The Netherlands: Noordhoff International Publishing.
26. Negri, M. (2005). A discontinuous finite element approach for the approximation of free
discontinuity problems. Advances in Mathematical Sciences and Applications, 15, 283–306.
134 M.M. Chiaramonte et al.

27. Ortiz, M. (1996). Computational micromechanics. Computational Mechanics, 18(5), 321–338.

28. Pandolfi, A., & Ortiz, M. (1998). Solid modeling aspects of three-dimensional fragmentation.
Engineering with Computers, 14(4), 287–308.
29. Pandolfi, A., & Ortiz, M. (2012). An eigenerosion approach to brittle fracture. International
Journal for Numerical Methods in Engineering, 92(8), 694–714.
30. Pandolfi, A., Krysl, P., & Ortiz, M. (1999). Finite element simulation of ring expansion and
fragmentation: The capturing of length and time scales through cohesive models of fracture.
International Journal of Fracture, 95(1–4), 279–297.
31. Pandolfi, A., Li, B., & Ortiz, M. (2012). Modeling fracture by material-point erosion. Interna-
tional Journal of Fracture, 184(1–2), 3–16.
32. Phongthanapanich, S., & Dechaumphai, P. (2004). Adaptive Delaunay triangulation with
object-oriented programming for crack propagation analysis. Finite Elements in Analysis and
Design, 40, 1753–1771.
33. Rangarajan, R. & Lew, A. J. (2013). Analysis of a method to parameterize planar curves
immersed in triangulations. SIAM Journal on Numerical Analysis, 51(3), 1392–1420.
34. Rangarajan, R., & Lew, A. J. (2014). Universal meshes: A method for triangulating planar
curved domains immersed in nonconforming triangulations. International Journal for Numer-
ical Methods in Engineering, 98(4), 236–264.
35. Rangarajan, R., Chiaramonte, M. M., Hunsweck, M. J., Shen, Y., & Lew, A. J. (2014). Sim-
ulating curvilinear crack propagation in two dimensions with universal meshes. International
Journal for Numerical Methods in Engineering.
36. Ruiz, G., Pandolfi, A., & Ortiz, M. (2001). Three dimensional cohesive modeling of dynamic
mixed-mode fracture. International Journal for Numerical Methods in Engineering, 52(12),
97–120.
37. Schmidt, B., Fraternali, F., & Ortiz, M. (2009). Eigenfracture: An eigendeformation approach
to variational fracture. Multiscale Modeling & Simulation, 7(3), 1237–1266.
38. Yuse, A., & Sano, M. (1997). Instabilities of quasi-static crack patterns in quenched glass
plates. Physica D: Nonlinear Phenomena, 108(4), 365–378.
39. Zielonka, M. G., Ortiz, M., & Marsden, J. E. (2008). Variational r-adaption in elastodynamics.
International Journal for Numerical Methods in Engineering, 74(7), 1162–1197.
Free Energy, Free Entropy, and a Gradient
Structure for Thermoplasticity

Alexander Mielke

Abstract In the modeling of solids the free energy, the energy, and the entropy play
a central role. We show that the free entropy, which is defined as the negative of
the free energy divided by the temperature, is similarly important. The derivatives
of the free energy are suitable thermodynamical driving forces for reversible (i.e.
Hamiltonian) parts of the dynamics, while for the dissipative parts the derivatives of
the free entropy are the correct driving forces. This difference does not matter for
isothermal cases nor for local materials, but it is relevant in the non-isothermal case
if the densities also depend on gradients, as is the case in gradient thermoplasticity.
Using the total entropy as a driving functional, we develop gradient structures for
quasistatic thermoplasticity, which again features the role of the free entropy. The
big advantage of the gradient structure is the possibility of deriving time-incremental
minimization procedures, where the entropy-production potential minus the total
entropy is minimized with respect to the internal variables and the temperature. We
also highlight that the usage of an auxiliary temperature as an integrating factor in
[30] serves exactly the purpose to transform the reversible driving forces, obtained
from the free energy, into the needed irreversible driving forces, which should have
been derived from the free entropy. This reconfirms the fact that only the usage of
the free entropy as driving functional for dissipative processes allows us to derive a
proper variational formulation.

1 Introduction

The mathematical theory of plasticity has its origin in the 1970 s based on the work
Moreau [22], Johnson [12], and Gröger [10], which treated the small-strain case with
quadrtic energies and fixed elastic domains. They developed a rich theory based on
convex analysis and monotone operators which allowed for significant generaliza-
tions, but still staying in the small-strain regime, see e.g. [1]. Finite-strain elastoplas-

A. Mielke (B)
WIAS, Mohrenstraße 39, 10117 Berlin, Germany
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 135

K. Weinberg and A. Pandolfi (eds.), Innovative Numerical Approaches
for Multi-Field and Multi-Scale Problems, Lecture Notes in Applied
and Computational Mechanics 81, DOI 10.1007/978-3-319-39022-2_7
136 A. Mielke

ticity also plays a fundamental role in engineering applications, and many algorithms
were derived starting in the 1980s, see e.g. [14, 29]. A major breakthrough was the
discovery in [24, 25] that incremental problems in finite-strain elastoplasticity can be
formulated as minimization problems jointly for the elastic and the plastic updates.
This means that elastoplasticity can be formulated in terms of a generalized gradient
system with a dissipation potential R and a free energy F such that it reads

Du F (t, u, z) = 0, 0 ∈ ∂ż R(u, z; ż) + Dz F (t, u, z),

where u is the displacement, and z contains all internal (dissipative) variables. This
theory even led to the first mathematical existence results for the rate-independent
case, see [9, 13, 20].
However, the whole theory is restricted to the isothermal case, and it remains
a challenging problem to find a corresponding mechanical and mathematical the-
ory for thermoplasticity. There are major differences in the mechanical modeling
between the isothermal and the non-isothermal case. In the former case there is one
free energy, and time-incremental minimization procedures can be formulated by
minimizing the sum of the free energy plus the dissipation in the time step. In the
non-isothermal case one has to take care of the mechanical forces still given by the
free energy, but instead of dissipation one now has to model entropy production.
A time-incremental minimization procedure should involve the entropy production
minus the total entropy. A first step in this direction was done in [30], and here we
connect our work [16, 17] with the latter.
The major observation is that one has to formulate thermoplasticity in a suitable
thermodynamically consistent way, in order to recast it in variational form. For this,
we start from the GENERIC framework (General Equations for Non-Equilibrium
Reversible Irreversible Coupling). For this we use the total energy E and the total
entropy S as functionals with densities E and S, respectively, depending on the
displacement u, the vector z of internal variables, and a thermodynamical variable
r. Using the entropy-production potential P and its Legendre dual P ∗ , we find the
form (cf. [16])
   
ρü = − Du E (q)−Θ∗Du S (q) = −Du F u, z, Θ(u, z, r) ,
 1   
ż = ∂ξz Pz∗ q, Dz S (q)− ∗Dz E (q) = ∂ξz Pz∗ q, Dz H (u, z, Θ(u, z, r)) ,
Θ
Δu S(q)[u̇] Δz E(q)[ż] 1  ∂r S 
ṙ = − − − div κ(q)∇ ,
∂r S(q) ∂r E(q) ∂r E ∂r E

where F and H are the total free energy and total free entropy expresses in terms
of the temperature θ = Θ(u, z, r), where r is an arbitrary scalar thermodynamic
E (u,z,r)
variable, such that Gibbs’ relation θ = Θ(u, z, r) = DDrrS (u,z,r)
holds.
Free Energy, Free Entropy, and a Gradient Structure for Thermoplasticity 137

The above form clearly shows the role of the free energy as the driving functional
for the reversible elastodynamics, while the free entropy

F(W, θ)
H (u, z, θ) = H(u, ∇u, z, ∇z, θ) dx with H(W, θ) = −
Ω θ

is the driving functional for the dissipative variables z (like the plastic tensor or the
hardening variables). Locally the free entropy is simply given as ‘minus the free
energy divided by the temperature’, but for functional derivatives, which involve
integration by parts, new terms appear and the naive relation Dz H (u, z, θ) =
− 1θ Dz F (u, z, θ) may be wrong. With F (u, z, θ) = Ω F(u, ∇u, z, ∇z, θ) dx we have

1   1
Dz H (u, z, θ) = − Dz F (u, z, θ) + ∂∇z F u, ∇u, z, ∇z, θ ∇ ,
θ θ
where the last term vanishes only in two important cases: (i) in the isothermal case
where ∇θ ≡ 0 and (ii) in the case “local case” where F does not depend on ∇z. In
these two cases, it is correct to use the derivative of the free energy and put the factor
θ into the dual entropy-production potential (thus turning it into a dual dissipation
potential). However, in all other cases, one has to distinguish the free energy and
the free entropy as driving functionals. Moreover, the inverse temperature 1/θ is the
driving force for heat transfer:

driving force for revers. dynamics :

Du F (u, z, θ) = Du E(u, z, r) − Θ(u, z, r)∗Du S(u, z, r),
driving force for dissip. dynamics :
Dz H (u, z, θ) = Dz S(u, z, r) − 1
Θ(u,z,r)
∗Dz E(u, z, r),
driving force for energy transport :
∂r S(u,z,r)
1/θ = 1/Θ(u, z, r) = ∂r E(u,z,r)
.

An important fact is that the terms on the right-hand side are independent of the choice
of the thermodynamical variable r, see Theorem 3, which gives a great flexibility in
the mathematical approaches.
Turning to the quasistatic case, we drop the interia term ρü and rewrite the remain-
ing system in the form
 
0 = Du E (q) − A(q)Dr E (q), ż = ∂ηz PZ∗ q; Dz S (q)−B(q)Dr S (q) ,
 
ṙ = −A(q)∗ u̇ − B(q)∗ ż − C(q)∗ div κ(q)∇(C(q)Dr S(q))
 ξr    (1)
with A(q)ξr = Dr S (q)
∗ Du S (q), B(q)ξr = Dr Eξr(q) ∗ Dz E (q),
ξr
and C(q)ξr = .
Dr E (q)
138 A. Mielke

Assuming that the first relation, which is static, can be solved in the form u = U(z, r),
Theorem 4 shows that this system can be formulated as a gradient system as follows

ż  
= ∂ξ P∗ z, r; DS(z, r) with S(z, r) = S (U(z, r), z, r),
ṙ

where P∗ is a suitably reduced dual entropy-production potential, and the reduced

energy E(z, r) = E (U(z, r), z, r) is conserved.
In general, this gradient structure is highly nonlocal, where P∗ involves the
derivatives Dz U(z, r) and Dr U(z, r), and thus less useful. However, in the case
A(u, z, r) ≡ 0, which occurs for a choice of r such that Du S (u, z, r) ≡ 0, one obtains
a system that allows for local approaches. Thus, the freedom of choosing r as freely
as possible is essential. For that case, we propose the time-incremental minimization
procedure
  
zk+1 1 z−zk 
∈ Arg min τ P qk ; − S (uk , z, r) ,
r k+1 (z,r) τ r−r k (2)
uk+1 ∈ Arg min E (u, zk+1 , r k+1 ),
u

where P is the primal entropy-production potential  obtained from the dual potential
P ∗ (u, z, r; ξz , ξr ) = PZ∗ (q; ξz −B(q)ξr ) + 21 Ω ∇(C(q)ξr )·κ(q)∇(C(q)ξr ) dx via
Legendre-Fenchel transform.
We discuss the abstract theory along specific thermomechanical examples. The
simplest is a spring-damper system, see Examples 5 and 7. Section 4.1 discusses the
Penrose-Fife model and shows how in [21] the gradient structure is exploited to do a
rigorous homogenization, where the effective entropy functional is obtained by aver-
aging of the free energy density. Section 4.2 treats a time-dependent thermoplastic
model where the gradient structure in terms of the entropy involves a time-dependent
entropy-production because of the elimination u(t) = U(z(t), (t)), where  is the
mechanical loading. Finally, a plastic model with thermal expansion is considered
in Sect. 4.3.
For all these models we need a specific and problem-dependent choice of the
thermodynamic variable r, which highlights the importance of a clear modeling in
terms of the free energy and free entropy giving the driving forces Du E(u, z, r) −
Θ(u, z, r)∗Du S(u, z, r), Dz S(u, z, r) − Θ(u,z,r)
1
∗Dz E(u, z, r), and 1/θ = ∂∂rr E(u,z,r)
S(u,z,r)
.

2 Heat Equation as a Starting Example

As a first example we treat the heat equation with energy density e and heat flux q:

ė + div q = 0 in Ω, q · n = 0 on ∂Ω.
Free Energy, Free Entropy, and a Gradient Structure for Thermoplasticity 139

Subsequently, we will drop all boundary conditions (like q · n = 0) and assume

that we have no-flux boundary conditions for all quantities, such that the system is
thermodynamically closed. We describe the energy density by an arbitrary scalar field
r, which may be the energy density e itself, the absolute temperature θ, the coldness
1/θ, or the entropy density s. This means that we have constitutive functions

θ = Θ(r), e = E(r), s = S(r).

Of course, by Gibbs’ relation θds = de we have the compatibility Θ(r) = E  (r)/S  (r).
Often this last relation is seen as the definition of the temperature. Note that already
here the inverse of the absolute temperature plays the role of an integrating factor
such that 1/θ de is the total differential ds, cf. [5].
In the classical form of the heat equation, the heat flux q is a linear function of
the temperature gradient, which is called Fourier’s law. In terms of r we arrive at
 
E  (r)ṙ − div k(r)∇Θ(r) = 0, (3)

where k ∈ Rd×d is the symmetric and positive definite heat conductivity matrix.
However, for a proper coupling to other mechanical effects, we want to have a
gradient flow formulation in terms of the total entropy S as a driving functional,
while the total energy E should be conserved
 
S (r) = s(r(x)) dx and E (r) = E(r(x)) dx.
Ω Ω

Hence, an entropic gradient structure must have the form

ṙ = K(r)DS (r), (4)

where K is a selfadjoint positive definite operator that maps the field ξr = DS (r)
to the rate ṙ, where ξr is the thermodynamical driving force associated with r. The
operator K will be called Onsager operator, since Onsager showed that such linear
operators should be symmetric. Indeed, in [23] the symmetry K = K∗ is called
“reciprocal relation”.
The positive semidefiniteness ξr , K(r)ξr  ≥ 0 guarantees that the second law
of thermodynamics is satisfied. Note that energy conservation needs the relation
K(r)DE (r) ≡ 0. Using the variational derivative DE (r) ≡ E  (r) we see that the
only Onsager operators which are compatible with the usual heat equation (3) and
the gradient structure (4) have the form

1   ξr 
K(r)ξr = − div κ(r)∇  ,
E  (r) E (r)

where κ(r) ∈ Rd×d

spd still can be chosen suitably.
140 A. Mielke

As a result we see that the heat equation takes the general structure

1  S  (r) 
ṙ = − div κ(r)∇ = K(r)DS (r),
E  (r) E  (r)

since DS (r) ≡ S  (r). In this general form we see that S  (r)/E  (r) = 1/Θ(r) is the
function under the spatial gradient, i.e. the heat flux has the form

q = κ(r)∇ 1/θ) = −k∇θ with κ(r) = k(r)Θ(r)2 .

Thus, we see that κ has to be chosen as k(r)Θ(r)2 . More importantly, we see that
the inverse temperature 1/θ is the driving force for energy flow, independently of the
choice of the scalar thermodynamical variable r.
To connect our theory to the work in [30] we introduce the dual entropy-production
potential (EPP), also called kinetic potential there, namely
  ξ   ξ 
1 1
P ∗ (r, ξ) := ξ, K(r)ξ = ∇ · κ(r)∇ dx.
2 2 Ω E  (r) E  (r)

By Legendre transform we can define the (primal) entropy-production potential via


  1
P(r, ṙ) = sup ξ, ṙ − P ∗ (r, ξ) = sup ξṙ − ∇ Eξ · κ∇ Eξ dx,
ξ ξ Ω 2

 ξ is obtained by solving the elliptic partial

which is nonlocal in ṙ, sincethe maximizer
differential equation − div κ∇(ξ/E  ) = ṙE  .
The gradient flow ṙ = K(r)DS (r) can be rewritten in the four fully equivalent
forms:

(i) ṙ = ∂ξ P ∗ (r, DS (r)), (ii) ṙ = Arg minv P(r, v) − DS (r), v,
(iii) ∂ṙ P(r, ṙ) = DS (r), (iv) DS (r) ∈ Arg minξ P ∗ (r, ξ) − ξ, ṙ.

Here the equivalence of (i) and (iii) is the Fenchel equivalence for the Legendre
transformation, while (ii) and (iv) are simply equivalent to (i) and (iii), respectively,
using the convexity of the EPPs P and P ∗ . To calculate the rate ṙ from the nonlo-
cal minimum principle (ii), the following local inf-sup formulation was introduced
in [30]:  
(
ξ, ṙ) solves inf sup ξ, v − P ∗ (r, ξ) − DS (r), v .
v ξ
Free Energy, Free Entropy, and a Gradient Structure for Thermoplasticity 141

3 Non-isothermal Dissipative Material Models

We now consider general elastic materials with internal parameters describing dis-
sipative effects such as plasticity, phase transformation, damage, magnetization, or
polarization, see e.g. [20]. We follow the approach presented in [16] but do not
emphasize the very useful framework GENERIC explicitly. Nevertheless we will
see remainders of the reversible (i.e. Hamiltonian) dynamics in the quasistatic elas-
tic force balance and of the irreversible dynamics in the flow rules for the internal
variable z and the heat equation.
We consider a body in the reference configuration Ω, which is a bounded domain
with Lipschitz boundary. The displacement is denoted by u : Ω ⊂ Rd → Rd , and
e(u) = 21 (∇u+∇uT ) is the linearized strain tensor. All the internal variables (also
called dissipative variables) are included in the variable z : Ω → Rm , which may
include plastic strains, phase indicators, or damage variables. By a general scalar
field r : Ω → R we describe the thermodynamical properties, e.g. r can be either
the temperature θ, the internal energy density e, or the entropy density s.
We consider a closed system, which means that we have no-flux boundary con-
ditions. The total energy and total entropy are given by

Ekin-pot (u, u̇, z, r) = Ekin (u̇) + E (u, z, r) and S (u, z, r) = S(∇u, z, ∇z, r) dx
Ω
 
ρ 2
where Ekin (u̇) := |u̇| dx and E (u, u̇, z, r) = E(u, ∇u, z, ∇z, r) dx,
Ω 2 Ω

∂r E(q)
where the consitutive laws E and S are related by Gibbs’ relation θ = Θ(q) := ∂r S(q)
.

3.1 Free Energy and Free Entropy as Driving Functionals

Before we discuss the equations of motions for such material models, we introduce
the free energy f and the free entropy s and discuss their role in continuum mechanics:

free energy f = e − θs (Gibbs 1873, Helmholtz 1882),

free entropy h = −f /θ = s − e/θ (Massieu 1869).

As is common for the free energy, we also consider the free entropy only as a function
of r = θ and use the densities (with W = (u, ∇u, z, ∇z))

f = F(W, θ) = E(W, θ) − θ S(W, θ),

F(W, θ) E(W, θ)
h = H(W, θ) = − = S(W, θ) −
θ θ
142 A. Mielke

as fields on Ω and define total free energy F and the total free entropy H via
 
F (u, z, θ) = F(u, ∇u, z, ∇z, θ) dx and H (u, z, θ) = H(u, ∇u, z, ∇z, θ) dx.
Ω Ω

The major point we want to address here is that F and H can serve as driving
functionals, since their partial derivatives with respect to any of the variables
W = (W1 , ..., Wk ) = (u, ∇u, z, ∇z) are independent of the particular choice of the
thermodynamic quantity r. The physical requirement for a driving force is that it takes
the same physical value, independent of the choice of the thermodynamic quantity.
The main observation is the following lemma which relies on the Gibbs relation.

Lemma 1 Consider smooth functions E : (W, r) → E(W, r) and S : (W, r) →

S(W, r) such that Θ(W, r) := ∂r E(W, r)/∂r S(W, r) > 0. Consider any transfor-
mation r = R(W, ρ) with ∂r R(W, r) = 0 and define

E(W, ρ) = E(W, R(W, ρ)) and 

 S(W, ρ) = S(W, R(W, ρ)).

Then, we have the identities

DW E(W, r) − Θ(W, r)DW S(W, r) = DW  

E(W, ρ) − Θ(W, ρ)DW  S(W, ρ)
∂ρ 
E(W, ρ)

and Θ(W, ρ) = = Θ(W, R(W, ρ)) if r = R(W, ρ).

∂ρ S(W, ρ)

In particular, for R(W, ρ) = Θ(W, θ) = θ and F(W, θ) = E(W, θ) − θS(W, θ) we

obtain
∂ρ E(W, ρ)
DW F(W, θ) = DW E(W, ρ) − θDW S(W, ρ), if θ = Θ(W, ρ) = .
∂ρ S(W, ρ)

Proof For the first result, we first establish the Gibbs relation using the chain rule:

∂ρ
E(W, ρ) ∂r E(W, R(W, r))∂ρ R(W, ρ) ∂r E(W, R(W, r))
= = = Θ(W, R(W, ρ)).

∂ρ S(W, ρ) ∂r S(W, R(W, r))∂ρ R(W, ρ) ∂r S(W, R(W, r))

For the driving forces for W we again use the chain rule to obtain

DW 
E(W, ρ) = DW E(W, R(W, ρ)) + ∂r E(W, R(W, ρ))DW R(W, ρ),
DW 
S(W, ρ) = DW S(W, R(W, ρ)) + ∂r S(W, R(W, ρ))DW R(W, ρ).

Thus, taking the linear combination DW 

E − ΘD W S and using Gibbs’ relation for Θ
we see that all terms involving DW R cancel and the result follows. Finally choosing
R(W, ρ) = Θ(W, ρ) =: θ and setting F(W, θ) = E(W, θ) − θS(W, θ) we obtain the
desired result since for Θ(W, θ) = θ we have DW Θ(W, θ) ≡ 0. 
Free Energy, Free Entropy, and a Gradient Structure for Thermoplasticity 143

To highlight the previous result we consider the following simple case.

Example 2 We consider E(z, θ) = 23 a(z)θ3/2 and S(z, θ) = a(z)θ1/2 , which gives

the free energy F(z, θ) = − 13 a(z)θ3/2 and the free entropy H(z, θ) = 13 a(z)θ1/2 .
Now consider r such that θ = R(z, r) = b(z)2 r 2 giving E(z, r) = 23 ab3 r 3 and
S(z, r) = abr. We can easily check the identity θ = b2 r 2 = Θ(z, r) = ∂r E/∂r S and
find
∂z E(z, r) − Θ(z, r)∂z S(z, r) = 13 a (z)b(z)3 r 3 = ∂z F(z, b(z)2 r 2 ),

i.e. the driving forces coincide as desired. However, for b (z) = 0 we have

∂z E(z, r) = ∂z E(z, b(z)2 r 2 ), ∂z S(z, r) = ∂z S(z, b(z)2 r 2 ), ∂z F(z, r) = ∂z F(z, b(z)2 r 2 )

where F(z, r) = E(z, r) − Θ(z, r)S(z, r) is the free energy expressed in r. 

As a consequence of the previous theorem we see that the only mechanically

relevant driving forces must be the derivative of the free energy DW F(W, θ) =
DW E(w, r) − Θ(W, r)DW S(W, r) or the temperature θ = ∂r E(W, r)/∂r S(W, r) or
any W -independent combination of these two. In fact, we will see that the following
three combinations are the most common:

driving force for reversible dynamics:

DW F(W, θ) = DW E(W, r) − Θ(W, r)DW S(W, r),
driving force for dissipative dynamics:
DW E(W, r)
DW H(W, θ) = DW S(W, r) − ,
Θ(W, r)
1 1 ∂r S(W, r)
driving force for energy transport: = = .
θ Θ(W, r) ∂r E(W, r)

However, there is still a major issue when considering fields over a body Ω and
considering the total free energy F and the total free entropy H . If we consider
variations of these functionals the variational derivatives involve integrations by part,
namely
 
Dz H (u, z, θ) = ∂z H(u, ∇u, z, ∇z, θ) − div ∂∇z H(u, ∇u, z, ∇z, θ) .

Now using the relation H = −F/θ we see that the differentials of F and H are not
simply related by multiplying with temperature, since we have

1 1
Dz H (u, z, θ) = − Dz F (u, z, θ) + ∂∇z F(u, ∇u, z, ∇z, θ)∇ .
θ θ

The last term, which destroys the naive relation Dz H (u, z, θ) = − 1θ Dz F (u, z, θ),
vanishes in two important cases: (i) in the isothermal case where ∇θ ≡ 0 and (ii) in
144 A. Mielke

the case “local case” where F does not depend on ∇z. In all other cases, we have to
be careful and distinguish the free energy and the free entropy as driving functionals.
In many situations it is helpful to use other thermodynamical fields r instead
of θ, in particular the internal-energy density e = E(W, θ) or the entropy density
s = S(W, θ) are often relevant. For these situations it is better to use the total energy
E and the total entropy S as function of (u, z, r). Hence, we need to adapt the nice
cancellation properties derived in Lemma 1 by introducing a multiplication “∗” for
scalar fields α and variational derivatives δz G as follows:
   
α ∗ δz G(z, ∇z) (x) := α(x)∂z G(z(x), ∇z(x)) − div α∂∇z G(z, ∇z) (x).

We also write α ∗ Dz G (u, z, r) for α ∗ δz G(u, ∇u, z, ∇z, r) and obtain the important
identities (5a, 5b) below. We should consider “α ∗ δz ” or “α ∗ Dz ” as one operator
acting on functions G or functionals G , respectively; see [20, Sect. 5.3] for a fully
abstract definition.

Theorem 3 Using the above definitions we have

Du F (u, z, Θ(u, z, r)) = Du E (u, z, r) − Θ(u, z, r) ∗ Du S (u, z, r), (5a)

1
Dz H (u, z, Θ(u, z, r)) = Dz S (u, z, r) − ∗ Dz E (u, z, r). (5b)
Θ(u, z, r)

Proof The right-hand side in the first line can be written in full detail as
 
RHS := ∂u E(W, r) − Θ(W, r)∂u S(W, r) − div ∂∇u E(W, r) − Θ(W, r)∂∇z S(W, r) ,

where W = (u, ∇u, z, ∇z). Using Lemma 1 we can apply the relation for ∂u and ∂∇u
independently and find
 
RHS = ∂u F(W, Θ(W, r)) − div ∂∇u F(W, Θ(W, r))
= δu F(W, θ))|θ=Θ(W,r) = Du F (u, z, Θ(u, z, r)).

This proves (5a), and the relation (5b) follows analogously. 

The importance of the formulas in Theorem 3 is that we are able to choose an

arbitrary thermodynamics field r of describing the heat effects in our material model.
This will allows us to find new mathematical formulations that cannot be accessed
by using the temperature θ, the energy density e, or the entropy density s, only.
We remark that in many papers and textbooks only the free energy is used as
driving functionals and that Dz F (u, z, r) is used as the driving force for the dissipa-
tive variable. This is correct for the cases of isothermal models or if F is local, i.e.
∂∇z F ≡ 0. In these two cases one has the relation Dz H = − 1θ Dz F , and the factor
−1/θ can be compensated in the dissipation potential, see [30] for the relevance of
the “integrating factor θ”.
Free Energy, Free Entropy, and a Gradient Structure for Thermoplasticity 145

However, in other cases the usage of Dz F leads to equations that are thermody-
namically inconsistent for the local balance laws, while the total energy conservation
and total entropy production may still be valid, see the discussion in [16, Remark 4.1].

3.2 The Balance Equations for Dissipative Material Models

Acoording to [11, 16] the GENERIC framework suggests to write the coupling of
elastodynamics for u, dissipative dynamics for z, and energy transport for r in the
form
 
ρü = − Du E (q)−Θ∗Du S (q) = −Du F (u, z, Θ(u, z, r)),
 1   
ż = ∂ξz Pz∗ q, Dz S (q)− ∗Dz E (q) = ∂ξz Pz∗ q, Dz H (u, z, Θ(u, z, r)) ,
Θ
Δu S(q)[u̇] Δz E(q)[ż] 1  ∂r S 
ṙ = − − − div κ(q)∇ ,
∂r S(q) ∂r E(q) ∂r E ∂r E

where the directional derivatives Δu S(q)[u̇] and Δz E(q)[ż] are defined via

Δw G(w)[v] := ∂w G(w, ∇w) · v + ∂∇w G(w, ∇w):∇v.

Here the first equation described elastodynamics and contains the Hamiltonian
part. In particular, we see that the reversible (i.e. Hamiltonian) part of the dynamics
is driven by the derivative Du F (u, z, Θ(u, z, r)) of the free energy F . In contrast,
the dissipative effects described by the internal variable z and the thermodynamical
field r are driven by Dz H (u, z, Θ(u, z, r)) and 1/Θ = ∂r S/∂r E, respectively. In
particular, we can define a joint dual entropy-production potential (EPP) P ∗ via

P ∗ (u, z, r; ξu , ξz , ξr ) = P0∗ (u, z, r; M(u, z, r)(ξz , ξr )T ) with


1
P0∗ (u, z, r; ηz , ηr ) = Pz∗ (u, z, r; ηz ) + ∇ηr · κ∇ηr dx and
2
  Ω
I ∂r− ∗Dz E (u, z, r)
M(u, z, r) = E(...)
 ,
0 ∂r E(...)

where  indicates the slot, where the corresponding argument (here ξr ) has to be
inserted.
We now discuss the two first terms on the right-hand side of the energy balance
for r, namely Δu S(u, z, r)[u̇]/∂r S and Δz E(u, z, r)[ż]/∂r E. The first term can be
seen as the latent-heat production term that is dual to the term ∂∂rr ES ∗ Du S in the
linear momentum balance and thus belongs to the reversible (=Hamiltonian) part of
dynamics. In particular it disappears completely if we choose r = s, which means
that it does not change the entropy. We refer to [16] for more details. In contrast, the
146 A. Mielke

second term Δz E(u, z, r)[ż]/∂r E is an entropy-production term that is dual to the

term ∂∂rr ES ∗ Dz E appearing in Pz∗ . We can now rewrite the system for (z, r) in the
form
   
ż 0 ∗ ∗ Dz S (u, z, r) 
= + M(u, z, r) ∂ ξ P u, z, r; M(u, z, r) (6)
ṙ − Δ∂u S[u̇]
rS
0
Dr S (u, z, r)

Before restricting to the quasistatic case with ρ = 0 we look at the total energy
balance and the total entropy production using the given abstract form. First we
observe that along solutions q(t) = (u(t), z(t), r(t)) we have

d ρ 2 
|u̇| dx + E (q(t))
dt Ω 2

= ρü · u̇ dx + Du E (q), u̇ + Dz E (q), ż + Dr E (q), ṙ
Ω
      
= −Θ ∗ Du S , u̇ + D∂rz EE , − Δ0u S[u̇] + M(q)∗ ∂η P0∗ q, M(q) D∂zrS
(1)
∂r S S
 Dz E  Dz S   (3)
(2) ∗
= 0 + 0 + M(q) Dr E , ∂η P0 q, M ∂r S = 0,

where we used the momentum balance and (6) in (1). Equality (2) follows from
Gibbs’ relation Θ = ∂r E/∂r S and the definition of “∗”, whereas (3) uses the
special form of M giving M(q)(Dz E , Dr E )T = (0, 1)T and the energy conserva-
tion (0, 1) , ∂η P ∗ (q, ξ) = 0, which follows from the translationally symmetry
P0∗ (q, η+λ(0, 1)T ) = P0∗ (q, η) for all λ ∈ R.
Similarly, the total entropy production can be calculated as follows:

d
S (q(t)) = Du S (q), u̇ + Dz S (q), ż + Dr S (q), ṙ
dt    0   Dz S 
= Du S (q), u̇ + D∂zrS
(i) ∗ ∗
S
, Δu S[u̇] + M(q) ∂η P0 q, M(q)
− ∂r S ∂r S
 Dz S      (ii)
= 0 + 0 + M(q) Dr S , ∂η P0∗ q, M D∂zrS S
≥ 0,

where we used (6) for (i) and the fact that P0∗ is a dual dissipation potential in (ii),
i.e. P0∗ (η) ≥ P0∗ (0) = 0 and convexity of P0∗ imply η, ∂P0∗ (η) ≥ 0.

3.3 A Gradient Structure for the Quasistatic Case

Subsequently we choose the quasistatic approximation and neglect the kinetic energy,
i.e. we set the density ρ = 0. It was shown already in [17] that, after elimination of the
displacement u, the remaining equation for (z, r) is a gradient system if one uses the
specific choice r = s (the density of the entropy). Here we follow [16] and show that
Free Energy, Free Entropy, and a Gradient Structure for Thermoplasticity 147

the result holds for any choice of r, which is extremely helpful, since traditionally
one prefers r = θ (the temperature) and more recently also the choice r = e (the
density of the internal energy), but general r gives more flexibility, see e.g. Sects. 4.2
and 4.3. To simplify the formulas we restrict our subsequent discussion to the simpler
case 
∗ ∗ 1
P0 (q; ηz , ηr ) = PZ (q; ηz ) + ∇ηr · κ(q)∇ηr dx.
Ω 2

The quasistatic thermomechanical system for q = (u, z, r) takes the form

0 = Du E (q) − Θ(q) ∗ Du S (q), (7a)

 1 
ż = ∂ηz PZ∗ q; Dz S (q)− ∗Dz E (q) , (7b)
Θ(q)
Δu S(q)[u̇] Δz E(q)[ż] 1  ∂r S 
ṙ = − − − div κ(q)∇ , (7c)
∂r S(q) ∂r E(q) ∂r E ∂r E

still displaying the driving forces in terms of free energy and free entropy. However,
the special GENERIC structure discussed in [16, Sect. 2.4] guides us to write the
system in the form

0 = Du E (q) − A(q)Dr E (q), (8a)

 
ż = ∂ηz PZ∗ q; Dz S (q)−B(q)Dr S (q) , (8b)
 
ṙ = −A(q)∗ u̇ − B(q)∗ ż − C(q)∗ div κ(q)∇(C(q)Dr S(q)) , (8c)

where the operators A(q), B(q), and C(q) are defined via
 ξr   ξr  ξr
Aξr = Dr S (q)
∗ Du S (q), Bξr = Dr E (q)
∗ Dz E (q), Cξr = Dr E (q)
. (8d)

By definition we have the following identities

ADr E = Θ ∗ Du S , BDr E = Dz E, CDr E = 1,

(9)
ADr S = Du S , BDr S = Θ1 ∗ Dz E , CDr S = 1/Θ.

For the following we assume that (7a) or (8a) can be solved uniquely in the form
u = U(z, r). As a shorthand, we also write q = Q(z, r) = (U(z, r), z, r).

Theorem 4 Assume that the mapping ξr → ξr + Dr U(z, r)∗ A(Q(z, r))ξr is invert-
ible and denote the inverse by J(z, r). Defining the functionals

S(z, r) = S (U(z, r), z, r), E(z, r) = E (U(z, r), z, r), and


     
P∗ (z, r; ξ) = PZ∗ Q(z, r); ξz −B(z, r)ξr + 21 ∇ C(z, r)ξr · κ(Q)∇ C(z, r)ξr dx,
 Ω
B(z, r) = B(Q(z, r))+Dz U(z, r)∗ A(Q(z, r)) J(z, r) and C(z, r) = C(Q(z, r))J(z, r),
148 A. Mielke

we obtain the following gradient structure:

  
d z ∗
 
(7a)−(7b) ⇐⇒ u = U(z, r) and = ∂ξ P z, r; DS(z, r) ,
dt r

and we have energy conservation via d

dλ
P∗ (z, r; ξ+λDE(z, r)) = 0.

Proof The last relation follows from the definition of P∗ and the identities

JDr E = Dr E (Q), CDr E(z, r) = C(Q)Dr E (Q) ≡ 1, and

Dz E−BDr E = Dz U∗ Du E + Dz E − (B+Dz U∗ A)Dr E )
= Dz U∗ (Du E −Θ∗Du S ) + Dz E − BDr E = 0,

where we used (7a) and (9)2 , respectively.

To see the equivalence between the evolution equations, first note

ż = ∂ξz P∗ (z, r; DS(z, r)) = ∂ηz PZ∗ (Q; Dz S − BDr S).

Proceeding as for Dz E−BDr E we obtain Dz S−BDr S = Dz S − BDr S |q=Q(z,r) ,

which is the physically correct driving force, namely the derivative of the free entropy.
Thus, the equation for z is identical to (8b).
For the r-equation we first observe CDr S = CDr S (Q) = 1/Θ(Q), which is the
correct driving force for heat conduction. Thus, the gradient-flow equation for r reads
    
ṙ = −B∗ ∂ξz P∗ (..) − C∗ div κ∇(1/Θ) = J∗ −(A∗ Dz U+B∗ )ż − C ∗ div κ∇(1/Θ) .

Now we use that by definition J∗ is the inverse of I + A∗ Dr U. Thus, we can rewrite

the last equation in the form
 
ṙ = −A(Q)∗ (U(z, r)ż+Dr U(z, r)ṙ) − B(Q)∗ ż − C(Q)∗ div κ(Q)∇(1/Θ(Q)) ,

which is the desired Eq. (8c), because of u̇ = Dz U(z, r)ż + Dr U(z, r)ṙ. 

Before going into more details, we present a simple finite-dimensional example,

where the reduction and the induced gradient structure can be calculated explicitly.

Example 5 We explain the derivation of the gradient structure by considering a

simple mass-spring-damper system, where we assume energy conservation, i.e. the
damping mechanics heats up the device, which additionally contains some thermal
expansion, see Fig. 1. To keep the model as simple as possible we choose the free
energy
1 1 4c
F(u, z, θ) = u2 + αuθ + (u−z)2 − θ3/2 ,
2 2 3
Free Energy, Free Entropy, and a Gradient Structure for Thermoplasticity 149

Fig. 1 A system with two

springs and one damper. The
upper spring undergoes
thermal expansion. The mass
at u(t) is neglected in our
quasistatic setting

z(t) u(t)

where α is the thermal expansion coefficient. The classical force balances are

0 = ∂u F(u, z, θ) = u + αθ + u − z and 0 = μż + z − u.

The evolution of θ will be determined by energy conservation.

For this we will transform the system into the above structure. First observe that

1 2 1 2c
E(u, z, θ) = u + (u−z)2 + θ3/2 and S(u, z, u) = 2cθ1/2 − αu.
2 2 3
The driving force for the damper is the derivative of the free entropy with respect to
z, which is

∂z H(u, z, θ) = ∂z S(u, z, θ) − θ∂z E(u, z, θ) = (u−z)/θ.

This is consistent with the choice of the EPP which differs from the dissipa-
tion potential R(ż) = μ2 ż2 by a factor of temperature, namely P(θ; ż) = 2θ
μ 2
ż and
θ 2
P ∗ (θ; ξ) = 2μ ξ . Together, the equations take the form (8), namely

0 = Du F(u, z, θ) = 2u−z+αθ,
θ 1
ż = ∂z H(u, z, θ) = (u−z),
μ μ
αθ1/2 z−u
θ̇ = −A(u, z, θ)∗ u̇ − B(u, z, θ)∗ ż, where A(q) = − and B(q) = 1/2 .
c cθ
There is no heat conduction term, since the temperature is the same in the whole
system.
For the reduction we immediately find u = U(z, θ) = 21 (z−αθ). Inserting this
relation into the above system we see that the reduced ODE is given by

1  α2  −z
ż = − (z+αθ) and 1+ θ1/2 θ̇ = ż, (10)
2μ 2c 2cθ1/2

where the last equation is equivalent to d

dt
E(z(t), θ(t)) = 0 after multiplication with
cθ1/2 .
150 A. Mielke

On the other hand, Theorem 4 provides a gradient structure via

1 2 α2 2 2c 3/2 α α2
E(z, θ) = z + θ + θ , S(z, θ) = − z + θ + 2cθ1/2 ,
4 4 3 2 2
θ  2 z
P∗ (θ; ξz , ξθ ) = ξz − B(z, θ)ξθ , B(z, r) = .
2μ 2cθ + α2 θ
1/2

It is easily checked that the gradient flow

 
ż = ∂ξz P ∗ θ; Dz S(z, r)−B(z, θ)Dθ S(z, θ) and ṙ = −B(z, θ)ż

is the same as (10), while the individual driving forces Dz S(z, θ) = −α/2 and
Dθ S(z, θ) = c/θ1/2 +α2 /2 are different from what one might naively expect. 

The above abstract result is a beautiful and mathematically clean reduction of the
quasistatically coupled system of elastostatics and dissipative material behavior to
a perfect gradient system driven by the physical entropy S. However, in practice it
is of limited use because of the involved nonlocal functionals. In particular, U(z, r)
depends nonlocally on (z, r), since it is obtained by solving an elliptic boundary
value problem. As a consequence, the operators J, B, and C are nonlocal as well.
Fortunately, there are cases, where the nonlocality disappears or is reduced to a
minimum. The most important case occurs if the entropy functional is independent
of u:
α ∗ Du S (u, z, r) = 0 for all α =⇒ A(q) ≡ 0.

As a consequence we obtain B(z, r) = B(Q(z, r)), C(z, r) = C(Q(z, r)), and J = id.
Moreover, the elastostatic equation reduces to Du E (u, z, r) = 0. Here we see the
advantage of using general thermodynamical variables r, since the form of A(u, z, r)
strongly depends on r: we have A(u, z, r) ≡ 0 only for specific choices, see Example
7 and Sects. 4.2 and 4.3.

3.4 A Time-Incremental Minimization Procedure

If we are able to find a formulation with α ∗ Du S ≡ 0, we can take advantage of

the gradient structure derived in Theorem 4, even without eliminating u explicitly.
Hence, we start with system (8), but now under the simplifying assumption A(q) = 0.
We first construct the (primal) entropy-production potential P(u, z, r; ż, ṙ). For
this we introduce the dual potential for the heat transfer, which is quadratic, namely

∗ 1    
Pheat (q; ξr ) := ∇ C(q)ξr · κ(q)∇ C(q)ξr dx
2 Ω
Free Energy, Free Entropy, and a Gradient Structure for Thermoplasticity 151

and denote by Pheat its Legendre transform, i.e. Pheat (q; ṙ) = supξr Ω ξr ṙ dx −
∗
 ∗
Pheat (q; ξr ) . Since Pheat (q; Dr S ) corresponds to an H1 norm of C(q)Dr S = 1/Θ,
the quadratic form Pheat (q; ṙ) corresponds to an H−1 norm of ∂r E(q)ṙ.
Recall that the full dual EPP has the form P ∗ (q; ξz , ξr ) = PZ∗ (q; ξz −B(q)ξr ) +
∗
Pheat (ξr ); hence the associated primal EPP reads

P(q; ż, ṙ) = PZ (q; ż) + Pheat (q; ṙ+B(q)∗ ż).

Using the Fenchel equivalence ξ ∈ ∂P(v) ⇔ v ∈ P ∗ (ξ), we find that the system
(8) with A(q) ≡ 0 for q = (u, z, r) can be rewritten as follows:
  
Dz S (q) I B(q) ∂PZ (q; ż)
Du E (q) = 0, ∈ ∂(ż,ṙ) P(q; ż, ṙ) = .
Dr S (q) 0 I ∂Pheat (q; ṙ+B∗ ż)

We see that both relations are variational in the sense that derivatives of functionals
determine the solutions.
In particular, we can discretize the system in time such that we obtain time-
incremental minimization principles that are useful for proving existence of solutions
or for numerical simulation of concrete models.

Time-incremental minimization procedure for the case A ≡ 0: Consider a time

step τ > 0 and assume that the initial condition q0 = (u0 , z0 , r 0 ) is given such that
Du E (q0 ) = 0. We define qk for k ∈ N incrementally as follows:
⎧  k+1  
⎪
⎪ z 1 z−zk 
⎨ k+1 ∈ Arg min τP q ; k
− S (uk , z, r) ,
(TIMP) r (z,r) τ r−r k (11)
⎪
⎪
⎩ uk+1 ∈ Arg min E (u, z k+1
,r k+1
).
u

Note that we do not enforce energy conservation, which could be done as well.
However, it is better to use the errors in the energy conservation as quality control
for the numerical accuracy, see the example below.

Remark 6 It may be tempting to write a similar time-incremental minimization pro-

cedure also in the case A(q) = 0. However, we see that the term A(q)∗ u̇ needs to
be approximated. One way would  be to use the consistent tangents Dz U and Dr U
and to replace the term by A(qk ) Dz U(zk , r k )ż+Dr U(zk , r k )ṙ) before discretizing the
derivatives by time increments. However, the numerical calculation of the tangents
Dz U and Dr U seems to be very inefficient. Moreover, it is not clear whether the
update uk+1 = uk + τ Dz U(zk , r k )(zk+1 −zk ) + Dr U(zk , r k )(r k+1 −r k ) is consistent
enough with the elastostatic equation Du E (q) − A(q)Dr E (q) = 0.

To highlight the usefulness of the algorithm we return to the spring-damper model

treated in Example 5. We will take advantage of using a suitable thermodynamic
variable r, namely r = s.
152 A. Mielke

Example 7 (Continuation of Example 5) The model is originally formulated in

(u, z, θ) but ∂u S(u, z, θ) = −α does not vanish, so the model cannot be treated with
these variables. Thus, we will use the entropy density s as the thermodynamical
variable r:
 s+αu 2
s = R(z, θ) := 2cθ1/2 − αu  θ = Θ(u, s) = .
2c
Hence, we find the following relations

 1 1 y3 /(12c2 ) for y ≥ 0,
E(u, z, s) = u2 + (u−z)2 +
e(s+αu) with 
e(y) =
2 2 ∞ for y < 0,
   z−u z−u
S(u, z, s) = s, A(u, z, s) = 0, B(u, z, s) = 4c2 = .
(s+αu)2 Θ(u, s)

The full coupled system takes the form

α
0 = Du
E(u, z, s) = 2u − z + (s+αu)2 , (12a)
4c2
Θ(u, s)     Θ  u−z
ż = ∂z S − B∂s S) = 0 −B(u, z, s) 1 = , (12b)
μ μ μ
Θ  2
ṡ = −
A(u, z, s) u̇ − 
B(u, z, s)ż = 0 + B(u, z, s) . (12c)
μ
Θ
Since the dual EPP P ∗ has the form P ∗ (θ; ξz , ξs ) = 2μ (ξz −Bξs )2 the primal EPP
reads  μ
|ż|2 if ṡ+
B(u, z, s)ż = 0,
P(u, z, s; ż, ṡ) = 2Θ(u, s)
∞ else.

Using the explicit constraint ṡ = 

B(u, z, s)ż, the time-incremental minimization pro-
cedure of (11) takes the explicit form:

μ 1
(1) Find (zk+1 , sk+1 ) as minimizer of z − z k )2 − s
2Θ(u , s ) τ
k k

subject to s − 
B(uk , zk , sk )z = sk −  B(uk , zk , sk )zk ;
(2) find uk+1 as minimizer of E(u, zk+1 , sk+1 ).

Here the first minimization problem is quadratic, and the explicit solution can be
determined. In the second minimization problem the functional is cubic in u, so the
unique minimizer uk+1 = U(zk+1 , sk+1 ) can be obtained by solving (12a). Thus, we
find the incremental update formulas
Free Energy, Free Entropy, and a Gradient Structure for Thermoplasticity 153

Fig. 2 Numerical solution q(t) = (u(t), z(t), s(t)) of the ODE (12) for parameters μ = α = 1 and
c = 1/2. Here ε(t) = 106 (
E(q(t))−
E(q(0)))

Fig. 3 Numerical simulation via (13) for time steps τ = 1/30, 1/100, 1/300, 1/1000. Energy
conservation is checked via ετ (kτ ) = (
E(qk )−
E(q0 ))/(τ 
E(q0 )), hence ετ (1) ≈ 5τ

Θ(qk )  k Θ(qk )  k 2
zk+1 = zk + τ B(q ), sk+1 = sk + τ B(q ) , uk+1 = U(zk+1 , sk+1 ).
μ μ
(13)

Inserting the explicit form of  B we find the relation zk+1 = zk + μτ (uk −zk ), which
is an explicit discretization of (12b). Nevertheless, by construction of our algorithm
we know that it is entropy increasing. Figure 2 shows the numerical solution of
the ODE (12), while Fig. 3 shows numerical approximations obtain via the TIMP
(11), which yields (13). We observe that the scheme is only of first order in the
time step. However, we expect that it is stable even when treating fully nonlinear
thermomechanical systems. 

4 Gradient Structures for Thermomechanical Systems

In this section we give three examples of temperature dependent models that can be
rewritten in terms of entropic gradient flows.

4.1 Homogenization of the Penrose-Fife System

This model is without any elastic deformation, so there is no need to eliminate the
variable u and the condition Du S ≡ 0 is trivially satisfied.
154 A. Mielke

The Penrose-Fife model was introduced in [27] to resolve a long-lasting debate

concerning thermodynamically correct couplings between phase transitions models
and the heat equation, see [19, 28] and [16, Remark 4.1] for details. Typically the free
energy F(z, ∇z, θ) = −cθ log θ + ψ0 (z) + θψ1 (z) + θ α2 |∇z|2 is used which leads to

ν
E(z, θ) = cθ + ψ0 (z) and S(z, ∇z, θ) = c log θ + c − ψ1 (z) − |∇z|2 . (14)
2
The equations in the variables z and θ take the form
⎧  
⎨ ż = mδ S(z, θ)− 1 ∂ E  = m νΔz−ψ  (z)− 1 ψ  (z)
z z
(PF) θ 
1
θ 0 (15)
⎩ 
cθ̇ = −ψ0 (z)ż + div k(z, θ)∇θ .

Almost all mathematical work is restricted to the case E(z, θ) = cθ + λz, which is
physically only relevant in a small temperature range. In particular, the logarithmic
entropy σ(θ) = c log θ is only good for gases, while for solids one should have
σ(0) = 0, e.g. σ(θ) = cθα for α ∈ ]0, 1[ is more appropriate.
In [21] we consider
 the internal energy e as thermodynamic variable r = e,
namely E (z, e) = Ω e(x) dx and S (z, e) = Ω  S(z, ∇z, e) dx. Indeed, the above
case (14) can be rewritten in terms of e via s =  S(z, ∇z, e) = c log(e−ψ0 (z)) −
c log c − ψ1 (z) − α2 |∇z|2 , but much more general functions S are possible.
The Penrose-Fife system (15) can be formulated as gradient system via the EPP

1 1
P ∗ (z, e; ξz , ξe ) = ξ, K(z, e)ξ = m(x, z, e))ξz (x)2 + ∇ξe (x) · κ(z, e)∇ξe (e) dx.
2 2 Ω

In particular, one has the explicit form


ż  
(PF) ⇐⇒ = K(z, e)DS (z, e) = ∂ξ P ∗ z, e; DS (z, e)
ė
 
m 0 δz S
= .
0 − div(κ∇) ∂e S

There is one special case where the gradient system can be rewritten as an evo-
lutionary variational inequality (EVI), cf. [2, 18]. For this we have to assume that
K (or equivalently P ∗ ) does not depend on the state (z, e). Moreover, one needs
to assume that (z, e) → −S (z, e) is λ-convex, i.e. for some λ ∈ R the function
(z, e) → −S (z, e) − λP(z, e) is convex, where P is the primal EPP. Under these
assumptions, a curve q = (z, e) : [0, T ] → XPF := H1 (Ω) × L1 (Ω) is a solution of
(PF), if and only if
 t−s  
(EVI) eλ(t−s) P(q(t)−w) + P(q(s)−w) ≤ 0 eλr dr S (q(t)) − S (w)
for all 0 ≤ s < t and all w = ( e) ∈ XPF .
z,
Free Energy, Free Entropy, and a Gradient Structure for Thermoplasticity 155

This variational formulation of the Penrose-Fife model is ideal for coarse graining.
Assuming that the entropy density  S, the mobility m, and the heat conduction tensor
κ depend periodically on a microscopical variable in the form

 1
Sε = S( 1ε x, z, e) − ∇z·A( 1ε x)∇z, mε (x) = M( 1ε x), κε (x) = H( 1ε x),
2
one can pass to the homogenization limit ε → 0 using the abstract methods for
evolutionary -convergence described in [18]. In [21] it is shown that solutions
(zε , eε ) for the gradient system (XPF , Sε , Pε ) converge to the unique solution (z0 , e0 )
of the limiting gradient system (XPF , S0 , P0 ), if this is true for the initial conditions.
The effective entropy functional S0 and the effective EPP P0 are

S0 (z, e) = Seff (z, e)− 21 ∇z·Ahom ∇z dx and
Ω

∗
P0 (ξz , ξe ) = 2
1
mharm ξz2 +∇ξe ·Hhom ∇ξe dx.
Ω

Here Ahom and Hhom are the classical homogenized effective tensors obtained from
the periodic functions A and H, respectively. Moreover, mharm is the harmonic mean
of M. More interesting is the homogenization of the nonlinear function S to obtain
Seff . Here one takes advantage of the concavity of the mapping e → S(y, z, e). Doing
a partial Legendre transform of −S with dual variable τ , one obtains the free energy
evaluated at θ = −1/τ . After simply averaging F(y, z, −1/τ ) over the periodicity
cell, one can reverse the Legendre transform and obtains Seff (z, e). We refer to [21]
for more details.

4.2 A Time-Dependent Thermoplasticity Model

We consider a special case of a linearized non-isothermal elastoplastic material,

where the coupling between the strain e(u) and the temperature is only indirect via
the plastic tensor z, cf. [3]. In contrast to the theory so far, we also allow for a
time-dependent loading (t). Hence, we consider the functionals
 
1 2
E (t, u, z, θ) = |e(u)−z|C + Φ(z, θ) dx − (t), u and S (u, z, θ) = S(z, θ) dx,
Ω 2 Ω

where |e|2C = e:Ce and Gibbs’ relation ∂θ Φ = θ∂θ S. Setting E(e, z, θ) = 21 |e−z|2C +
Φ(z, θ), we will explicitly use the decouplings ∂e ∂θ E = 0 = ∂e S.
We note that Du S ≡ 0 implies that the elastic equilibrium takes the form
 
0 = Du E (t, u, z, θ) − θ ∗ Du S (u, z, θ) = Du E (t, u, z, θ) = − div C(e(u)−z) − (t).
156 A. Mielke

In particular, we are able to solve this equation for u as a nonlocal function of

z and the loading (t), namely u(t) = U(z(t), (t)), where U : L2 (Ω; Rd×d0,sym ) ×
d ∗
(HD (Ω; R )) → HD (Ω; R ) is a bounded linear operator.
1 1 d

Respecting the energy conservation, we can take the dual EPP P ∗ in the form
 ξθ ξθ 
P ∗ (q; ξu , ξz , ξθ ) = P0∗ q; ξz − ∗ Dz E (t, q), ,
∂θ Φ(z, θ) ∂θ Φ(z, θ)

which clearly satisfies P ∗ (q; ξ+λDE (t, q)) = P ∗ (q, ξ).

Defining the reduced energy E(t, z, θ) := E (t, U(z, (t)), θ) we find the relations

1
E(t, z, θ) = z:Az + Φ(z, θ) dx − z, a(t),
Ω 2

where A is a bounded, symmetric, non-negative linear operator from L2 (Ω; Rd×d 0,sym )
into itself and a(t) = K(t) for a suitable bounded linear operator K. Thus, Theorem 4
yields the gradient-flow equation

ż
= ∂ξ P ∗ (z, θ; DS (z, θ))
θ̇
   
I 0 ∗ I −1
∂θ Φ
∗ Dz E Dz S
= −1  ∂ξ P0 z, θ; .
D
∂θ Φ z
E 1
∂θ Φ 0 1
∂θ Φ Dθ S

We emphasize here that the transformation inside ∂ξ P ∗ via Dz E=Az + ∂z Φ(z, θ) −

a(t) is time-dependent and nonlocal because of A.
To simplify the gradient structure we reformulate it using the internal energy

1
e(x) := z(x) : (Az)(x) + Φ(z(x), θ(x)),
2
where we note that the nonlocal operator A has to be taken with care. This relation
can be inverted to express the temperature as function of z and e as follows. Denote by
 e) the unique solution of 
θ = Θ(z, e = Φ(z, θ) and define   e)),
e) = S(z, Θ(z,
S(z,
which gives ∂e S(z,  e) by Gibbs’ relation ∂θ Φ = θ∂θ S. Then, with 
e) = 1/Θ(z, e=
e − 21 z : Az, the total energy, total entropy, and the dual EPP read
 
E(t, z, e) = e − a(t):z dx, S(z, e) = 
S(z, e− 21 z:Az) dx, and
Ω Ω
   
∗ ∗  ∗ I a(t) ξz
P (t, z, e; ξz , ξe ) := P0 (z, e; ξz +ξe a(t), ξe ) = P0 z, e; .
0 1 ξe

The energy balance dtd E (t, z(t), e(t)) = ∂t E (t, z(t), e(t)) along solutions still follows
∗ (z, e; ξ+(0, λ)T ) = P
from the relation P ∗ (z, e; ξ) for all constants λ ∈ R.
0 0
Free Energy, Free Entropy, and a Gradient Structure for Thermoplasticity 157

 takes a similar time-dependent form

The primal EPP P

   I 0
 z, e; ż, ė) = P
P(t, 0 z, e; N(t) ż , N(t) :=
0 (z, e; ż, ė−a(t):ż) = P
ė −a(t): 1

where P0 (z, e; v, w) = ∞ if Ω w dx = 0, which enforces energy conservation.
In total, the generalized gradient flow for this simple thermoplastic model can be
written in the following two equivalent forms
  
ż −1  ∗ Dz S+De Sa(t)
= N(t) ∂ξ P0 z, e; ⇐⇒
ė De S(z, e)
   
0 0 z, e; N(t) ż Dz S(z, e)
∈ N(t)∗ ∂v P − .
0 ė De S(z, e)

We consider a specially simple case of thermo-viscoplastic gradient plasticity by

choosing

0 (z, e; v, w) = σyield vL1 + μ v2 2 + κ w2 −1 and

P L H
 2 2
1 ν
S(z, e) = 
S(z, e− z:Az) − |∇z|2 dx,
Ω 2 2

where w2H−1 = ∇φ2L2 if Δφ = w in Ω and ∇φ · n = 0 on ∂Ω. This leads to the

generalized gradient flow equation
 
0 = σyield Sign(ż) + μż + ∂z
S(z, e− 21 z:Az) − 21 Ξ (Az) + A(Ξ z) + νΔz,
  (16)
0 = ė − a(t):ż − κΔΞ, where Ξ = ∂e S z, e− 1 z:Az .
2

Here Ξ = ∂e S denotes the inverse temperature 1/θ.

In particular, the second formulation gives rise to a simple time-incremental min-
imization procedure, which is well-known in isothermal elastoplasticity (cf. [6–8,
13, 15, 24, 25], but is new for the non-isothermal case:
  
zk+1  z−zk
(TIMP)∗ ∈ Arg min (t k+1
−t )P0 z , e ; t k+1 −t k N(t )
k k k 1 k
− S(z, e).
ek+1 (z,e) e−ek

We emphasize that (TIMP)∗ is not equivalent to the one proposed in (11), since
here we eliminated u beforehand by using the nonlocal operator A. So, (TIMP)∗
should be preferable if A is available. Again, we observe that the concavity of S
implies that the minimum problem is convex. In the case of viscoplasticity, P 0 is
even strictly convex, so there is a unique minimizer in each time step. Thus, it should
be possible to show existence of solutions for the thermo-viscoplastic system in
(16). Unfortunately, the methods developed in [21] and based on the (EVI) are not
applicable because of the nonquadratic behavior of P 0 due to σyield > 0.
158 A. Mielke

4.3 A Thermoplastic Model with Thermal Expansion

Finally, we consider a classical plasticity model (see e.g. [4]) where thermal expan-
sion leads to a stronger coupling of elastostatics and heat conduction. As usual we
again start with a free energy containing a thermal expansion tensor E ∈ Rd×d
sym in the
form
1 σθ c
F(∇u, z, ∇z, θ) = |e(u)−z|2C + ψ1 (θ)E:e(u) + H(z) + |∇z|2 − θ1+α ,
2 2 α(1+α)

with c > 0 and α ∈ ]0, 1[. We obtain the energy and entropy functionals
 1+α
1 1 (θ)E:e(u) + H(z) + cθ
E (u, z, θ) = |e(u)−z|2C + ψ dx,
2 1+α
Ω
c α ν
S (u, z, θ) = θ − ψ1 (θ)E:e(u) − |∇z|2 dx,
Ω α 2

where ψ1 (θ) = ψ1 (θ) − θψ  (θ). Clearly, α ∗ Du S (u, z, θ) = − div(αψ  (θ)E) is

1 1
non-zero, so the reduction to a local gradient system is not possible, unless we
replace the temperature θ by a more convenient thermodynamically variable r. A
possible choice is
c
r = R(e(u), θ) := θα − ψ1 (θ)E:e(u).
α
Since we also need the inverse transformation θ = Θ(e(u), r), we assume ψ1 (θ) = θ
 
for notational simplicity. Then ψ1 ≡ 0 and θ = Θ(e, r) = α(r+E:e)/c 1/α , and the
functionals take the form

1 αcα  1+1/α
E (u, z, r) = |e(u)−z|2C + H(z) + r+E:e(u) dx,
2 1+α
Ω
ν
S (u, z, r) = r − |∇z|2 dx, where cα = (α/c)1/α .
Ω 2

This choice now guarantees that A(u, z, r) ≡ 0 and the reduction to a local gradient
system for (z, r) can be done as described at the end of Sect. 3.3. In particular the
solution u = U(z, r) can be obtained as the unique minimizer of the convex functional
u → E (u, z, r). The corresponding Euler-Lagrange equation reads
 
− div C(e(u)−z) + cα (r+E:e(u))1/α E = 0.

Acknowledgments The author is grateful to Michael Ortiz for many stimulating discussions on
the interaction of mathematics and mechanics. Moreover, he thanks Ulisse Stefanelli for the fruitful
collaboration concerning variational methods for evolutionary systems. The resarch was partially
supported by the European Research Council via the grant ERC-2010-AdG no.267802 Analysis of
Multiscale Systems Driven by Functionals.
Free Energy, Free Entropy, and a Gradient Structure for Thermoplasticity 159

References

1. Alber, H. -D. (1998). Materials with memory (Vol. 1682), Lecture Notes in Mathematics.
Berlin: Springer.
2. Ambrosio, L., Gigli, N., & Savaré, G. (2005). Gradient flows in metric spaces and in the space
of probability measures., Lectures in Mathematics Basel: ETH Zürich. Birkhäuser Verlag.
3. Bartels, S., & Roubíček, T. (2008). Thermoviscoplasticity at small strains. ZAMM—Journal of
Applied Mathematics and Mechanics, 88, 735–754.
4. Bartels, S., & Roubíček, T. (2011). Thermo-visco-elasticity with rate-independent plasticity
in isotropic materials undergoing thermal expansion. Mathematical Modelling and Numerical
Analysis (M2AN), 45, 477–504.
5. Carathéodory, C. (1909). Untersuchungen über die Grundlagen der Thermodynamik. Mathe-
matische Annalen, 67, 355–386.
6. Carstensen, C., Hackl, K., & Mielke, A. (2002). Non-convex potentials and microstructures
in finite-strain plasticity. Proceedings of the Royal Society of London Series A, 458(2018),
299–317.
7. Dal Maso, G., DeSimone, A., & Mora, M. G. (2006). Quasistatic evolution problems for linearly
elastic-perfectly plastic materials. Archive for Rational Mechanics and Analysis, 180(2), 237–
291.
8. Dal Maso, G., DeSimone, A., & Solombrino, F. (2011). Quasistatic evolution for cam-clay plas-
ticity: a weak formulation via viscoplastic regularization and time parametrization. Calculus
of Variations and Partial Differential Equations, 40(2), 125–181.
9. Gürses, E., Mainik, A., Miehe, C., & Mielke, A. (2006). Analytical and numerical methods
for finite-strain elastoplasticity. In R. Helmig, A. Mielke, & B. Wohlmuth (Eds.), Multifield
problems in solid and fluid mechanics (pp. 443–481). Berlin: Springer.
10. Gröger, K. (1978). Zur Theorie des quasi-statischen Verhaltens von elastisch-plastischen Kör-
pern. ZAMM—Journal of Applied Mathematics and Mechanics, 58(2), 81–88.
11. Hütter, M., & Svendsen, B. (2012). Thermodynamic model formulation for viscoplastic solids
as general equations for non-equilibrium reversible-irreversible coupling. Continuum Mechan-
ics and Thermodynamics, 24, 211–227.
12. Johnson, C. (1976). Existence theorems for plasticity problems. Journal de Mathematiques
Pures et Appliques (9), 55(4), 431–444.
13. Mainik, A., & Mielke, A. (2009). Global existence for rate-independent gradient plasticity at
finite strain. Journal of Nonlinear Science, 19(3), 221–248.
14. Miehe, C., & Stein, E. (1992). A canonical model of multiplicative elasto–plasticity. For-
mulation and aspects of numerical implementation. European Journal of Mechanics endash;
A/Solids, 11, 25–43.
15. Mielke, A. (2003). Energetic formulation of multiplicative elasto-plasticity using dissipation
distances. Continuum Mechanics and Thermodynamics, 15, 351–382.
16. Mielke, A. (2011). Formulation of thermoelastic dissipative material behavior using GENERIC.
Continuum Mechanics and Thermodynamics, 23(3), 233–256.
17. Mielke, A. (2011). On thermodynamically consistent models and gradient structures for ther-
moplasticity. GAMM—Mitteilungen, 34(1), 51–58.
18. Mielke, A. (2016). On evolutionary Γ -convergence for gradient systems. In A. Muntean,
J. Rademacher, & A. Zagaris (Eds.), Macroscopic and large scale phenomena: coarse grain-
ing, mean field limits and ergodicity, Lecture Notes in Applied Math. Mechanics, 3, 187–249.
Springer.
19. Mielke, A. (2013). Thermomechanical modeling of energy-reaction-diffusion systems, includ-
ing bulk-interface interactions. Discrete and Continuous Dynamical Systems—Series S, 6(2),
479–499.
20. Mielke, A., & Roubíček, T. (2015). Rate-independent systems: theory and application. Applied
Mathematical Sciences, 193. Springer.
21. Mielke, A., & Stefanelli, U. (2015). Homogenizing the penrose-fife system via its gradient
structure. In preparation.
160 A. Mielke

22. Moreau, J.-J. (1974). On unilateral constraints, friction and plasticity. In New Variational Tech-
niques in Mathematical Physics (Centro Internaz. Mat. Estivo (C.I.M.E.), II Ciclo, Bressanone,
1973) (pp. 171–322). Rome: Edizioni Cremonese.
23. Onsager, L. (1931). Reciprocal relations in irreversible processes, I+II. Physical Review, 37,
405–426. (part II, 38:2265–2279).
24. Ortiz, M., & Repetto, E. (1999). Nonconvex energy minimization and dislocation structures in
ductile single crystals. Journal of the Mechanics and Physics of Solids, 47(2), 397–462.
25. Ortiz, M., & Stainier, L. (1999). The variational formulation of viscoplastic constitutive updates.
Computer Methods in Applied Mechanics and Engineering, 171(3–4), 419–444.
26. Ortiz, M., Repetto, E., & Stainier, L. (2000). A theory of subgrain dislocation structures. Journal
of the Mechanics and Physics of Solids, 48, 2077–2114.
27. Penrose, O., & Fife, P. C. (1990). Thermodynamically consistent models of phase-field type
for the kinetics of phase transitions. Physica D, 43(1), 44–62.
28. Penrose, O., & Fife, P. C. (1993). On the relation between the standard phase-field model and
a "thermodynamically consistent" phase-field model. Physica D, 69(1–2), 107–113.
29. Simo, J., & Ortiz, M. (1985). A unified approach to finite deformation elastoplastic analysis
based on the use of hyperelastic constitutive relations. Computer Methods in Applied Mechanics
and Engineering, 49, 221–245.
30. Yang, Q., Stainier, L., & Ortiz, M. (2006). A variational formulation of the coupled thermo-
mechanical boundary-value problem for general dissipative solids. Journal of the Mechanics
and Physics of Solids, 54, 401–424.
Comparison of Isotropic Elasto-Plastic
Models for the Plastic Metric
Tensor C p = F pT F p

Patrizio Neff and Ionel-Dumitrel Ghiba

Abstract We discuss in detail existing isotropic elasto-plastic models based on

6-dimensional flow rules for the positive definite plastic metric tensor C p = F pT F p
and highlight their properties and interconnections. We show that seemingly different
models are equivalent in the isotropic case.

1 Introduction

Since the early days of the introduction of the multiplicative decomposition into
computational elasto-plasticity, the need was felt to reduce the level of complexity
and to discard the concept of a plastic rotation in the completely isotropic setting.
This means to consider a flow rule not for the plastic distortion F p (9-dimensional)
[5, 6, 11, 30, 34, 36, 43, 44], but to consider directly a flow rule for the plastic
metric tensor C p = F pT F p ∈ PSym(3) (6-dimensional) [1, 10, 38, 39, 45], which
is then automatically invariant under left-multiplication of F p with a plastic rotation.
The plastic distortion is in general incompatible F p = ∇ψ p , as is the plastic metric
C p = ∇ψ pT ∇ψ p . A formulation in the plastic metric C p is particular attractive
because it circumvents problems associated with the intermediate configuration
introduced by the multiplicative decomposition, which is trivially non-unique since

P. Neff · I.-D. Ghiba (B)

Lehrstuhl für Nichtlineare Analysis und Modellierung, Fakultät für Mathematik,
Universität Duisburg-Essen, Thea-Leymann Str. 9, 45127 Essen, Germany
e-mail: [email protected]; [email protected]
P. Neff
e-mail: [email protected]
I.-D. Ghiba
Department of Mathematics, Alexandru Ioan Cuza University of Iaşi,
Blvd. Carol I, No. 11, 700506 Iaşi, Romania
I.-D. Ghiba
Octav Mayer Institute of Mathematics of the Romanian Academy, Iaşi Branch,
700505 Iaşi, Romania

© Springer International Publishing Switzerland 2016 161

K. Weinberg and A. Pandolfi (eds.), Innovative Numerical Approaches
for Multi-Field and Multi-Scale Problems, Lecture Notes in Applied
and Computational Mechanics 81, DOI 10.1007/978-3-319-39022-2_8
162 P. Neff and I.-D. Ghiba

F = Fe · F p = Fe · Q T · Q · F p = Fe∗ · F p∗ , Q ∈ SO(3).

Several proposals with the aim of removing the non-uniqueness of the intermediate
configuration have been given in the literature. Our comparative study is related to
the following models: Simo’s model [41] (Reese and Wriggers [36], Miehe [21]);
Miehe’s model [22]; Lion’s model [17] (Helm [12]), (Dettmer-Reese [6]); Simo
and Hughes’ model [42]; Helm’s model [12] (Vladimirov, Pietryga and Reese [45],
Shutov and Kreißig [39], Reese and Christ [35], Brepols, Vladimirov and Reese [1],
Shutov and Ihlemann [38]); Grandi and Stefanelli’s model [10] (Frigeri and Stefanelli
[7]). All these models are given with respect to different configurations, either the
reference configuration, the intermediate configuration or the current configuration.
In order to be able to compare them, it is necessary to transform all to the same
configuration for that purpose. In our case we choose the reference configuration.
Moreover, any explicit dependence on F p instead of C p in the model formulation
must be able to be subsumed into a dependence on C p alone in the isotropic case. A
major body of our work consists in showing this for the models under consideration.
The paper is structured as follows. After a paragraph giving some definitions
which generalize the concepts from small strain-additive plasticity to finite strain
plasticity we established some auxiliary results. Then we discuss existing
6-dimensional flow rules from the literature. The main properties of the investigated
isotropic plasticity models are summarized in Figs. 1 and 2. Finally, in the appendix,
we obtain explicit formulas for some of the isotropic plasticity models.

1.1 Consistent Isotropic Finite Plasticity Model

for the Plastic Metric Tensor C p

In this paper, we use the standard Euclidean scalar product on R3×3 given by
X, Y  := tr(X Y T ), and thus the Frobenius tensor norm is X 2 = X, X . The iden-
tity tensor on R3×3 will be denoted by 1, so that tr(X ) = X, 1. We let Sym(3) and
PSym(3) denote the symmetric and positive definite symmetric tensors respectively.
We adopt the usual abbreviations of Lie-group theory. Here and in the following
the superscript T is used to denote transposition, sym X = 21 (X + X T ) denotes the
symmetric part of the matrix X ∈ R3×3 , while dev3 X = X − 13 tr(X ) · 1 represents
the deviatoric (trace free) part of the matrix X .
The classical concept of associated perfect plasticity is uniquely defined in the
case of small strain-additive plasticity. In this case the total symmetric strain is decom-
posed additively into elastic and plastic parts ε = εe + ε p and the rate-independent
evolution law for the symmetric plastic strain ε p is given in subdifferential format

d
[ε p ] ∈ ∂ χ (Σlin ), tr(ε p ) = 0,
dt
Comparison of Isotropic Elasto-Plastic Models for the Plastic … 163

where ∂ χ is the subdifferential of the indicator function χ of the convex elastic

domain
 
2  2
Ee (Σlin , σy2 ) = Σlin ∈ Sym(3)  dev3 Σlin 2 ≤ σy2 ⊂ Sym(3)
3 3

and Σlin := −Dε p [Wlin (ε − ε p )] is the thermodynamic driving stress of the plastic
process. Here, Σlin is clearly symmetric.
In such a way, the principle of maximum dissipation (equivalent to the convexity
of the elastic domain and normality of the flow direction) is satisfied. The structure
of associated flow rules in geometrically nonlinear theories is by far not as trivial as
in the geometrically linear models. However, in this work we use:
Definition 1 (geometrically nonlinear associated plastic flow) We call a plastic flow
rule for some plastic variable P (whether symmetric or not) associated, whenever
the flow rule can be written as
d √ d −1 √
[P] P −1 ∈ ∂ χ (Σ) or P [P ] P ∈ f = ∂ χ (Σ),
dt dt

where Σ is some symmetric or non-symmetric stress tensor. Here, dtd [P −1 ] P is the

correct format for the time derivative (it will lead to an exponential update, see the
implicit method based on the exponential mapping considered in [40]). Moreover,
we require that χ is the indicator function of some convex domain in the Σ-stress
space.
After liniarization (small strain-additive approximation) this condition is equiva-
lent to classical associated plasticity. Further, let us also remark that a metric is by
definition symmetric and positive definite, i.e. C p ∈ PSym(3).
Definition 2 (consistent isotropic finite plasticity model for plastic metric tensor
C p ) We say that an associated plastic flow rule, in the sense of Definition 1, for the
plastic metric tensor C p is consistent, whenever:
(i) it is thermodynamically correct, i.e. the reduced dissipation inequality is satis-
fied;
(ii) plastic incompressibility: the constraint det C p (t) = 1 for all t ≥ 0 follows from
the flow rule;
(iii) C p (t) ∈ PSym(3) for all t > 0 if C p (0) ∈ PSym(3).
As we will see from the next Lemma 4, our requirement (iii) follows if C p (t) ∈
Sym(3) for all t ≥ 0, C p (0) ∈ PSym(3) and if (ii) is satisfied.
We finish our setup of preliminaries with the following definitions:

Definition 3 (reduced dissipation inequality-thermodynamic consistency) For a

given energy W , we say that the reduced dissipation inequality along the plastic
evolution is satisfied if and only if
164 P. Neff and I.-D. Ghiba

d d  d
[W (F F p−1 (t)] = [W (C C −1
p (t)] = [Ψ (C, C p (t)] ≤ 0
dt dt dt

for all constant in time F (viz. C = F T F), depending in which format the elastic
energy is given.

Definition 4 (Loss of ellipticity in the elastic domain) We say that the elasto-plastic
formulation preserves ellipticity in the elastic domain whenever the purely elastic
response in elastic unloading of the material remains rank-one convex for arbitrary
large given plastic pre-distortion, see [9, 29].

1.2 Auxiliary Results

We consider the multiplicative decomposition of the deformation gradient [13–16,

28, 31] and we define, accordingly, the elastic and plastic strain tensors

Ce := FeT Fe ∈ PSym(3), Be := Fe FeT ∈ PSym(3),

C p := F pT F p ∈ PSym(3).

Let us also define the stress tensors

 (C)] = 2 Dlog C [W (log C)] = Dlog U [W̌ (log U )]

Σ : = 2 C DC [W
= U DU [W (U )] = F T D F [W (F)] ,
 (B)] B = 2 Dlog B [W (log B)] = Dlog V [W̌ (log V )]
τ : = 2 D B [W
= V DV [W (V )] = 2 F DC [W  (C)] F T .

The tensor Σ = C · S2 (C), where S2 = 2 DC [W (C)] is the second Piola-Kirchhoff

stress tensor, is sometimes called the Mandel stress tensor and it holds dev3 Σe =
dev3 ΣE , where ΣE is the elastic Eshelby tensor

ΣE := FeT D Fe [W (Fe )] − W (Fe ) · 1 = Dlog Ce [W (log Ce )] − W (log Ce ) · 1,

driving the plastic evolution (see e.g. [2–4, 20, 28]), while τ is the Kirchhoff stress
tensor and Σe is defined in Remark 1.
Remark 1 We also need to consider the following elasto-plastic stress tensors:

 (Ce )] = 2 Dlog Ce [W (log Ce )] = Dlog Ue [W̌ (log Ue )]

Σe : = 2 Ce DCe [W
= Ue DUe [W (Ue )] = FeT D Fe [W (Fe )] ,
 (Be )] Be = 2 Dlog Be [W (log Be )] = Dlog Ve [W̌ (log Ve )]
τe : = 2 D Be [W
=Ve DVe [W (Ve )] = 2 Fe DCe [W  (Ce )] Fe T .
Comparison of Isotropic Elasto-Plastic Models for the Plastic … 165

The following relation holds true:

Σ = F T τ F −T , Σe = FeT τe Fe−T . (1)

Note that (1) is not at variance with symmetry of Σ and Σe in case of isotropy.
Using the fact that for given Fe ∈ GL+ (3) it holds FeT S Fe−T 2 ≥ 21 S2 for all
S ∈ Sym(3), the constant being independent of Fe [30], we obtain the estimate

1
 dev3 Σe  = FeT (dev3 τe )Fe−T  ≥ √  dev3 τe ,
2

which is valid for general anisotropic materials. Indeed, since

1
dev3 Σe = dev3 (FeT τe Fe−T ) = FeT τe Fe−T − tr(FeT τe Fe−T ) · 1
3
1
= Fe (τe − tr(τe )) · 1)Fe−T ,
T
3
we have

dev3 Σe = FeT (dev3 τe )Fe−T , dev3 τe = Fe−T (dev3 Σe )FeT , tr(Σe ) = tr(τe ).

However,  dev3 Σe  =  dev3 τe  for general anisotropic materials. Let us remark

that for elastically isotropic materials we have from the representation formula for
isotropic tensor functions

 (Ce )] = α1 1 + α2 Ce + α3 Ce2 ∈ Sym(3),

DCe [W (2)
 (Ce )] = 2 Ce (α1 1 + α2 Ce + α3 Ce2 ) ∈ Sym(3),
Σe = 2 Ce · DCe [W

where
 
2 ∂W ∂W 2 ∂W
α1 = 1/2 I2 (Ce ) + I3 (Ce ) , α2 = 1/2 ,
I3 (Ce ) ∂ I 2 (C e ) ∂ I 3 (C e ) I3 (Ce ) 1 (Ce )
∂ I
1/2 ∂W
α3 = −2 I3 (Ce )
∂ I2 (Ce )

are scalar functions of the invariants of Ce , which are functions of C C −1

p , see
Lemma 2. This leads us to
Lemma 1 For the isotropic case  dev3 Σe  =  dev3 τe .
Proof For the isotropic case we have τe Be = Be τe , which implies

 dev3 Σe 2 = FeT (dev3 τe ) Fe−T , FeT (dev3 τe ) Fe−T  = Be (dev3 τe ), (dev3 τe ) Be−1 
=  dev3 τe 2 .
166 P. Neff and I.-D. Ghiba

We also consider the following tensor

Σ  (C C −1
 := 2 C DC [W  −1 −1
p )] = 2 C D[ W (C C p )] C p ∈/ Sym(3), (3)

which is not symmetric, in general. For instance, for the simplest Neo-Hooke energy
W (Fe ) = tr(Ce ) = tr(C C −1  −1 
p ) we have D W (C C p )=1 and Σ=2 C C p ∈
−1
/ Sym(3).
Lemma 2 Any isotropic and objective free energy W defined in terms of Fe can be
expressed as

 (C C −1
W (Fe ) = W  T −1
p ) = W (F F(F p F p ) ).
T
(4)

Proof It is clear that any objective elastic energy W (Fe ) which is isotropic w.r.t. Fe ,
can be expressed in terms of the invariants of Ce , i.e.

W (Fe ) = Ψ (I1 (Ce ), I2 (Ce ), I3 (Ce )),

I1 (Ce ) = tr(Ce ) = tr(Be ), I2 (Ce ) = tr(Cof Ce ) = tr(Cof Be ),
I3 (Ce ) = det Ce = det Be .

Now every invariant can be rewritten as follows

I1 (Ce ) = Ce , 1 = FeT Fe , 1 = F p−T F T (F F p−1 ), 1 = C, C −1 −1

p  = tr(C C p )
= I1 (C C −1
p ), (5)
I2 (Ce ) = Cof Ce , 1 = det Ce Ce−T , 1 = det(F p−T C F p−1 ) [F p−T C F p−1 ]−T , 1
= det C det C −1
p C
−T
, F pT F p  = det(C C −1
p ) C
−T T
C p , 1 = tr(Cof(C C −1 p ))
= I2 (C C −1
p ),
I3 (Ce ) = det Ce = det(F p−T C F p−1 ) = det C det C −1 −1
p = I3 (C C p ).

Therefore, we obtain

W (Fe ) = Ψ (I1 (Ce ), I2 (Ce ), I3 (Ce ))

= Ψ (I1 (C C −1 −1 −1  −1
p ), I2 (C C p ), I3 (C C p )) = W (C C p ),

and the proof is complete. 

Remark 2 Since the principal invariants Ik , k = 1, 2, 3 are the coefficients of the
characteristic polynomial and I1 (Ce ) = I1 (C C −1 −1
p ), I2 (C e ) = I2 (C C p ), I3 (C e ) =
I3 (C C −1 −1
p ), the eigenvalues of C e and C C p coincide. Clearly, C e ∈ PSym(3), how-
−1
ever C C p ∈ / Sym(3) in general, unless C and C −1
p commute.

 τe are related as follows

Lemma 3 The introduced stress tensors Σe , Σ,

 F pT ,
Σe = F p−T Σ  = F T τe F −T .
Σ
Comparison of Isotropic Elasto-Plastic Models for the Plastic … 167

Proof For arbitrary increment H ∈ R3×3 , we compute

D F [W (Fe )], H  = D F [W (F F p−1 )], H  = D Fe [W (Fe )], H F p−1 

= D Fe [W (Fe )] F p−T , H .

On the other hand, we deduce

 (C C −1
D F [W  T −1
p )], H  = D F [ W (F F C p )], H 

= D[W  (C C −1 −1 −1
p )], F H C p + H F C p 
T T

 (C C −1
= 2 F sym[D[W −1
p )]C p ], H ,

 (C C −1
for all H ∈ R3×3 . In view of Lemma 2 we have W (Fe ) = W p ). Therefore, we
obtain

 (C C −1
2 F sym[D[W −1 −T
p )]C p ] = D Fe [W (Fe )] F p ,

and further

 (C C −1
FeT D Fe [W (Fe )] F p−T = 2 FeT F sym[D[W −1 −T  −1 −1
p )]C p ] = 2 F p C sym[D[ W (C C p )]C p ].

The above relation implies

 (C C −1
Σe = FeT D Fe [W (Fe )] = 2 F p−T C DC [W −T  T
p )] F p = F p Σ F p .
T

Therefore, using Remark 1 the proof is complete. 

Next, we introduce a helpful lemma.

Lemma 4 If t → C p (t) ∈ R3×3 is continuous and satisfies:
⎫
det C p (t) = 1 for all t > 0, ⎬
C p (0) ∈ PSym(3), ⇒ C p (t) ∈ PSym(3) for all t > 0.
⎭
C p (t) ∈ Sym(3) for all t > 0

Proof Using Cardano’s formula and due to the symmetry of C p , the continuity of
the map t → C p (t) implies the continuity of mappings t → λi (t), i = 1, 2, 3, where
λi (t) ∈ R are the eigenvalues of C p (t). Since λi (0) > 0 and λ1 (t)λ2 (t)λ3 (t) = 1 for
all t > 0, it follows that λi (t) > 0 for all t > 0 and the proof is complete. 

We can slightly weaken the assumption in the previous lemma: det C p (t) > 0 for
all t > 0 is sufficient.
168 P. Neff and I.-D. Ghiba

2 The Simo-Miehe 1992 Spatial Model

In the remainder of this paper we discuss different proposal from the literature for
plasticity models in C p . Simo [41] (see also Reese and Wriggers [36] and Miehe [21,
p. 72, Proposition 5.25]) considered the spatial flow rule in the form

1
p ∂ τe Φ(τe ) · Be ,
− Lv (Be ) = λ+ (6)
2

where the Lie-derivative Lv (Be ) is given by Lv (Be ) := F dtd [C −1

p ] F ∈ Sym(3),
T

the tensor τe = 2 ∂ Be W (Be ) · Be is the symmetric Kirchhoff stress tensor, the yield
function Φ(τe ) =  dev3 τe  − 23 σy and the plastic multiplier λ+
p satisfies the
Karush-Kuhn-Tucker (KKT)-optimality constraints

λ+
p ≥ 0, Φ(τe ) ≤ 0, λ+
p Φ(τe ) = 0. (7)

The flow rule (6) is equivalent with

d −1
p F [∂ τe Φ(τe ) · Be ] F
[C ] = −2 λ+ −1 −T
dt p  
+ −1 dev3 τe
= −2 λp F · Be F −T , (8)
 dev3 τe 

which, in view of the properties (7) of λ+

p , can be written with a subdifferential

d −1 
[C p ] ∈ −2 F −1 ∂ τe χ (dev3 τe ) · Be F −T , (9)
dt

where χ is the indicator function of the elastic domain

 
2 2  2 2

Ee (τe , σy ) = τe ∈ Sym(3)  dev3 τe  ≤ σy = {τe ∈ Sym(3) | Φ(τe ) ≤ 0}.
2
3 3

The subdifferential ∂ χ (dev3 τe ) of the indicator function χ is the normal cone


2 0, τe ∈ int(Ee (τe , 23 σy2 ))
N (Ee (τe , σy2 ); dev3 τe ) = dev3 τe
3 {λ+p  dev3 τe 
| λ+
p ∈ R+ }, τe ∈
/ int(Ee (τe , 23 σy2 )).

We deduce (see the model Eq. (5.25) from [21]) an equivalent definition for
Lv (Be ) given by

1 dev3 τe
− Lv (Be ) = λ+ · Be .
2 p
 dev3 τe 
Comparison of Isotropic Elasto-Plastic Models for the Plastic … 169
 
Since C p = F T Be−1 F we have Lv (Be )=F d
dt
[C −1
p ] F T =F d
dt
[F −1 Be F −T ] FT .
On the other hand, from (8) it follows that
 
d −1 dev3 τe
[C p ] C p = −2 λ+ F −1
· Be F
−T
F T Be−1 F
dt p
 dev3 τe 
∈ −2 F −1 ∂ τe χ (dev3 τe ) F. (10)

Since
d d d
[det C p−1 ] = Cof C p−1 , [C p−1 ] = det C p−1 C p , [C p−1 ]
dt dt dt
−1 d −1
= det C P 1, [C p ]C p , (11)
dt

from the flow rule (8) together with det C p (0) = 1 and tr(F −1 dev3 τe F) = 0 it
follows at once that det C p (t) = 1 for all t ≥ 0.
The next step is to prove that the flow rule (6) implies dtd [W (Fe )] ≤ 0 at fixed F,
i.e. the reduced dissipation inequality is satisfied. We compute for fixed in time F

d d d
[W (F F p−1 )] = D Fe W (Fe ), F [F p−1 ] = D Fe W (Fe ), F F p−1 F p [F p−1 ]
dt dt dt
d d
= FeT D Fe W (Fe ), F p [F p−1 ] = Σe , F p [F p−1 ]
dt dt
d
= −Σe , sym( [F p ]F p−1 ), (12)
 dt  
Dp

since Σe ∈ Sym(3). We also have

d d d d
[C p ] = [F pT F p ] = [F pT ]F p + F pT [F p ]
dt dt  dt  dt 
−T d d −1
= Fp Fp
T
[F ] F p + F p
T T
[F p ]F p F p = 2 F pT D p F p ,
dt p dt
 
where D p := sym dtd [F p ]F p−1 . Hence, we easily deduce the representation D p =
1
F −T d [C p ]F p−1 . Therefore, with (12) we obtain
2 p dt

d 1 d
[W (F F p−1 )] = −Σe , F p−T [C p ]F p−1 . (13)
dt 2 dt
170 P. Neff and I.-D. Ghiba

Moreover, since Σe = FeT τe Fe−T , we deduce

d 1 d 1 d
[W (F F p−1 )] = − FeT τe Fe−T , F p−T [C p ]F p−1  = FeT τe Fe−T , F p [C −1 ]F pT 
dt 2 dt 2 dt p
1 d 1 T d −1
=  F pT FeT τe Fe−T F p , [C −1 −1
p ] =  F τe Be F, [C ]
2 dt 2 dt p
1 d
= τe , F [C −1 ] F T Be−1 . (14)
2 dt p

The flow rule (8) implies

d dev3 τe
[W (F F p−1 )] = −λ+
p τe ,  = −λ+
p  dev3 τe  ≤ 0. (15)
dt  dev3 τe 

In view of the definition of Σe = FeT τe Fe−T we have F −1 [ τe Be ]F −T = F p−1 [Σe ]

F p−T .
For the isotropic case it holds τe Be = Be τe . Hence,

F −1 [τe Be ]F −T = F −1 [ Be τe ]F −T = F p−1 Fe−1 [ Fe FeT τe ]Fe−T F p−T

= F p−1 [ FeT τe Fe−T ]F p−T = F p−1 [Σe ]F p−T .

We also observe F −1 [ tr(τe ) Be ]F −T = F p−1 [tr(Σe )]F p−T . Thus, we obtain

F −1 [dev3 τe Be ]F −T = F p−1 [dev3 Σe ]F p−T .

Together with Remark 1 this implies that

   
dev3 τe dev3 Σe
F −1 Be F −T = F p−1 F p−T . (16)
 dev3 τe   dev3 Σe 

Therefore, in the isotropic case, the flow rule (9) has a subdifferential structure:

d −1
[C ] ∈ −2 F p−1 [∂ Σe χ (dev3 Σe )] F p−T , (17)
dt p

where χ is the indicator function of the elastic domain

 
2  2
Ee (Σe , σy2 ) = Σe ∈ Sym(3)  dev3 Σe 2 ≤ σy2 .
3 3

In view of the above equivalent representations of the flow rule, we may summarize
the properties of the Simo-Miehe 1992 model:
(i) from (8) it follows, in the isotropic case (in which τe and Be commute), that
C p (t) ∈ Sym(3);
(ii) plastic incompressibility: from (10) and (11) it follows that det C p (t) = 1, since
the right hand side is trace-free;
Comparison of Isotropic Elasto-Plastic Models for the Plastic … 171

(iii) for the isotropic case, the right hand-side of (8) is a function of C −1
p and C
−1 T −1 −1 −1 T −1
alone, since Be = F C p F and F Be F = F F C p F F = C p C;
(iv) from (i) and (ii) together and using Lemma 4 it follows that C p (t) ∈ PSym(3);
(v) it is thermodynamically correct;
(vi) the right hand side in the representation (8) is not the subdifferential of the
indicator function of some convex domain in some stress space. However, this
model is an associated plasticity model in the isotropic case, see Propositions 1
and 3.

3 The Miehe 1995 Referential Model

Shutov [37] interpreted that Miehe in [22] considered the flow rule1

d −1
[C ] C p = −λ+ 
 Φ(Σ),
p DΣ (18)
dt p

where Σ  (C C −1
 = 2 C DC [W p )] and

 
2 2
 =  dev3 τe  −
Φ(Σ) σy = tr((dev3  2)
Σ) − σy .
3 3

In this model, it is important to note that it is not the Frobenius norm of


 3 Σ which is used in the yield function Φ. Instead, in the denominator F :=
dev
 2 ) is considered, see Eq. (52) from [39]. Since dev3 Σ
tr((dev3 Σ) ∈/ Sym(3), it

 
follows that F := tr((dev3 Σ) ) =  dev3 Σ. Indeed, we have
2

 2 ] =  dev3 (Σ)
tr[(dev3 Σ)  ⇔  (dev3 Σ)
dev3 Σ,  T  = dev3 Σ,
 dev3 Σ

⇔  skew(dev3 Σ)
dev3 Σ,  =0 ⇔  ∈ Sym(3)
dev3 Σ
⇔  ∈ Sym(3).
Σ

For the simplest Neo-Hooke elastic energy considered in Appendix A.2, W (Fe ) =
 (C C −1
tr(Ce ) = W −1  −1
p ) = 2 tr(C C p ), we have Σ = C C p , which is not symmetric.
1

Hence tr[(dev3 Σ)  2 ] =  dev3 Σ.
 Let us again remark that Σ
 is not necessarily
symmetric for general C p . However, using Lemma 3, we deduce

  
1 Miehe [22] only defines the elastic domain Ee (Σ, 2 σy2 ) := Σ  ∈ R3×3  tr((dev3 Σ)
 2 ) ≤ 2 σy2
3 3
  
in terms of τe , i.e. Ee (τe , 23 σy2 ) = τ ∈ Sym(3)  dev3 τ 2 ≤ 23 σy2 . He uses the same notation
for the referential quantities. Therefore, we have two interpretations at hand Φ(Σ)  =  dev3 τe  −

2 2 
σy = tr((dev3 Σ) ) − σy . On the other hand, in the isotropic case, we have also Φ(Σ)
2 2 2  =
3 3
 dev3 Σe  − 2
3 σy2 .
172 P. Neff and I.-D. Ghiba

 C p = F pT Σe F p−T C p = F pT Σe F p ∈ Sym(3)
Σ ⇒  · C p ∈ Sym(3),
dev3 Σ
(19)
C −1
p Σ
 = C −1
p F pT Σe F p−T = F p−1 Σe F p−T ∈ Sym(3) ⇒ C −1
p

dev3 Σ ∈ Sym(3).

 2 ). First,
In the following, we discuss first the sign of the quantity2 F 2 := tr((dev3 Σ)
we deduce

 2 ] = (dev3 Σ)
tr[(dev3 Σ)  (dev3 Σ),
 1 = Σ  1 = C −1
 (dev3 Σ),  
p Σ (dev3 Σ) C p , 1
= C −1   T −1  T
p Σ (dev3 Σ · C p ) , 1 = C p Σ C p (dev3 Σ) , 1 (20)
= C −1  
p Σ C p , dev3 Σ.

We further see that

C −1   −1 −1   −1  −1 
p Σ C p , dev3 Σ = U p U p Σ U p U p , dev3 Σ = U p Σ U p , U p dev3 Σ U p 
1
 U p , U p−1 Σ
= U p−1 Σ  U p − tr(Σ)  · 1
3
1
 U p , U p−1 Σ
= U p−1 Σ  U p − tr(U p−1 Σ  U p ) · 1
3
 U p , dev3 (U p−1 Σ
= U p−1 Σ  U p ) = dev3 (U p−1 Σ
 U p ), dev3 (U p−1 Σ
 U p )
 U p )2 ≥ 0,
=  dev3 (U p−1 Σ (21)

where U p2 = C p . Thus F 2 is positive and F is well defined.

1
Since DΣ Φ(Σ)  =  T the flow rule (18) becomes
(dev3 Σ)
 2]
tr[(dev3 Σ)

d −1 λ+
[C p ] C p = − 
p
T
(dev3 Σ)
dt  2]
tr[(dev3 Σ)
d −1 λ+
⇔ Cp [C p ] = − 
p

dev3 Σ. (22)
dt  2]
tr[(dev3 Σ)

 2 ) for all dev3 Σ

we are not looking for the sign of tr((dev3 Σ)
2 If  ∈ R3×3 , then considering two

particular values of dev3 Σ, e.g.
⎛ 1 ⎞ ⎛ 1 ⎞
−2 1 2 −3 0 0
dev3 Σ = ⎝ −2 − 3 ⎠
1 and dev3 Σ =⎝ 0 2 ⎠,
2 3 0
−1 −3 − 21 0 0 − 13

 2 ] = −2 and tr[(dev3 Σ)
we obtain tr[(dev3 Σ)  2 ] = 2 , respectively. Hence, tr[(dev3 Σ)
 2 ] is not
3

positive for all Σ ∈ R .
3×3
Comparison of Isotropic Elasto-Plastic Models for the Plastic … 173

Further, in view of (19), we obtain

d −1 λ+
[C p ] = − 
p

C −1 dev3 Σ
dt  2] p
tr[(dev3 Σ)
d λ+
⇔ [C p ] = 
p
 C p ∈ Sym(3).
(dev3 Σ) (23)
dt  2]
tr[(dev3 Σ)

Using Lemma 4 we obtain that C p ∈ PSym(3).

We remark that the flow rule considered by Miehe [22] (in this interpretation)
coincides with the flow rule (47) considered by Helm [12], see Proposition 2.

Remark 3 Although the flow rule considered in this interpretation of the Miehe 1995
model [22] has a subdifferential structure, the yield-function Φ is not convex. Hence,
the flow rule is not a convex flow rule. In order to see the non-convexity of Φ(Σ) 
we observe first by looking at sublevel-sets that

2 2
 =
Φ(Σ)  2) −
tr((dev3 Σ) σ is convex ⇔  Σ)
Φ(  = tr[(dev3 Σ)
 2 ] is convex.
3 y
 Σ)
The second derivative for the simpler function Φ(  is

2    H ∈ R3×3 .
 Φ(Σ).(H, H ) = (dev3 H ) , dev3 H  = tr[(dev3 H ) ], ∀ Σ,
T 2
DΣ

We know that tr[(dev3 H )2 ] is not positive for all H ∈ R3×3 , since for the previous
considered matrix H , such that
⎛ ⎞
− 21 1 2
dev3 H = ⎝ −2 − 21 3 ⎠ ,
−1 −3 − 21

 Σ)
we obtain tr[(dev3 H )2 ] = −2. Therefore Φ(  is not convex, and thus Φ(Σ)
 cannot
be convex.

4 The Lion 1997 Multiplicative Elasto-Plasticity

Formulation in Terms of the Plastic Metric C p = FpT Fp

This derivation was given by Lion [17, Eq.(47.2)] in the general form (see also [12,
Eq.(6.33)]) and by Dettmer-Reese [6] in the isotropic case. Following [6] we consider
a perfect plasticity model for the plastic metric C p based on the flow rule

d −1
[C ] ∈ −F p−1 ∂ χ (dev3 Σe ) F p−T ∈ Sym(3) for Σe ∈ Sym(3). (24)
dt p
174 P. Neff and I.-D. Ghiba

Again, in this model it is not clear from the outset, that it is a formulation in
C p alone. The goal of such a 6-dimensional formulation is to avoid any explicit
computation of the plastic distortion F p . However, the right hand side of the above
proposed flow rule is, in fact, a multivalued function in C and C −1
p alone. Hence, we
can express the flow rule (24) entirely in the form3

d −1
[C ] C p ∈ f (C, C −1
p ). (25)
dt p
In order to show this remarkable property (satisfied only for isotropic response), and
to determine the explicit form of the function f (C, C −1 p ), in view of (2) we remark
that
 
dev3 Σe 1 1
F p−1 F p−T = F p−1 Σe F p−T − tr(Σe ) C −1 ,
 dev3 Σe   dev3 Σe  3 p

tr(Σe ) = 2 1, Ce (α1 1 + α2 Ce + α3 Ce2 ),

Σe 2 = 4Ce (α1 1 + α2 Ce + α3 Ce2 ), Ce (α1 1 + α2 Ce + α3 Ce2 ),

1
 dev3 Σe  = Σe 2 − [tr(Σe )]2 .
3

It is clear that F p−1 Σe F p−T = 2 F p−1 Ce (α1 1 + α2 Ce + α3 Ce2 ) F p−T ∈ Sym(3), and

Ce = FeT Fe = F p−T F T F F p−1 = F p−T C F p−1 ,

Σe = 2 F p−T (α1 C + α2 C C −1 −1 −1 −1
p C + α3 C C p C C p C) F p ,

tr(Σe ) = 2 tr(α1 C C −1 −1 −1 −1 −1 −1
p + α2 C C p C C p + α3 C C p C C p C C p ),

Σe 2 = 4C −1   −1
p f , f C p .

where


f := α1 C + α2 C C −1 −1 −1
p C + α3 C C p C C p C.

Hence, we deduce

F p−1 Σe F p−T = 2 C −1  −1 −1
p f (C, C p ) C p ∈ Sym(3), (26)
tr(Σe ) = 2 tr( 
f (C, C −1 −1
p ) C p ), Σe 2 = 4C −1  −1  −1 −1
p f (C, C p ), f (C, C p ) C p ,

1
 dev3 Σe  = 2 tr[(  p ) C p ) ] − [tr( 
f (C, C −1 −1 2
f (C, C −1 −1 2
p )C p )] ,
3

f (C, C −1 −1  −1 / Sym(3)
p ) is not necessarily symmetric. Moreover C p f (C, C p ) ∈
3 Note carefully, that

in general.
Comparison of Isotropic Elasto-Plastic Models for the Plastic … 175

where


f (C, C −1 −1 −1 −1
p ) := α1 C + α2 C C p C + α3 C C p C C p C ∈ Sym(3), (27)

and αi = αi (I1 (Ce ), I2 (Ce ), I3 (Ce )), according to (2). Therefore, the multivalued
function f (C, C −1
p ) is given by

 −1  
−λ+ −1
p dev3 ( C p f (C, C p ))  +
f (C, C −1  λ ∈ R+ .
p ) = p
tr[(  p )C p ) ] − 3 [tr( 
f (C, C −1 −1 2 1
f (C, C −1 −1 2
p )C p )]
(28)

In Appendix A.1 we give the specific expression for the functions f (C, C −1
p ) and

f (C, C −1
p ) in case of the Neo-Hooke energy.
On the other hand, in view of Eq. (24) we also have

−1 χ
= F pT ∂ χ (dev3 Σe )F p−T .
d d −1
p = −C p [C p ] ∈ C p F p ∂ (dev3 Σe ) F p
[C p ] C −1 −T
dt dt

Hence, it follows that

d
[C p ] ∈ F pT ∂ χ (dev3 Σe ) F p−T C p = F pT ∂ χ (dev3 Σe )F p ∈ Sym(3), (29)
dt
which establishes symmetry of C p whenever C p (0) ∈ Sym(3).
Another important question is whether the solution C p of the flow rule (24) is
such that det C p (t) = 1, for all t ≥ 0. Let C p be the solution of the flow rule (24).
Then, we have

d −1 −1 dev3 Σe −1 dev3 Σe
Cp [C ] = −λ+
p C p Fp F −T = −λ+ T
p Fp Fp Fp F −T
dt p  dev3 Σe  p  dev3 Σe  p
λ+ −T
p dev3 (F p Σe F p )
T
=− , (30)
2  dev3 Σe 

which implies on the one hand

d −1 d
 [C p ] C p , 1 =  [C −1 ], C p  = 0.
dt dt p
On the other hand, the flow rule (24) together with det C p (0) = 1 leads to
det C p (t) = 1, for all t ≥ 0.
176 P. Neff and I.-D. Ghiba

Let us remark that, in view of (25) and (28) we have for the flow rule (24)

d −1 −λ+
[C ] =
p
dev3 [ C −1  −1 −1
p f (C, C p )] · C p ,
dt p tr[(    
p )C p ) ] − 3 [tr[ 
f (C, C −1 −1 2 1
f (C, C −1 −1 2
p )C p ]]
∈Sym(3)
/
  
∈Sym(3)

which is in concordance with the requirement C p ∈ Sym(3), as can be seen from

(29). Note that the above formula cannot be read as

d −1 dev3 Σ
[C p ] = −λ+ · C −1
p ,
dt p
dev3 Σ

for some Σ, since

1
[tr( 
f (C, C −1 −1 2  −1 −1 2  −1 −1 2
p )C p ) ] − [tr( f (C, C p )C p )]  =  dev3 ( f (C, C p )C p ) .
3
To see this, assume to the contrary that equality holds. Then we deduce

1
f (C, C −1
[tr(  −1 2  −1 −1 2  −1 −1 2
p )C p ) ] − [tr( f (C, C p )C p )] =  dev3 ( f (C, C p )C p )
3
⇔  f (C, C −1 −1  −1 −1 T  −1 −1  −1 −1
p )C p , ( f (C, C p )C p )  = ± f (C, C p )C p , f (C, C p )C p .
(31)

Since 
f (C, C −1
p ) ∈ Sym(3), we obtain

tr[( 
f (C, C −1 −1 2 −1  −1 −1  −1
p )C p ) ] = C p f (C, C p )C p ( f (C, C p ), 1

= C −1  −1  −1 −1
p f (C, C p ), f (C, C p )C p . (32)

Using that C p ∈ PSym(3), we further deduce that

C −1  −1  −1 −1 −1  −1 −1 −1  −1 −1
p f (C, C p ), f (C, C p )C p  = U p f (C, C p )U p , U p f (C, C p )U p 

= U p−1 
f (C, C −1 −1 2
p )U p  , (33)

where U p2 = C p . Therefore, from (31) we deduce


f (C, C −1 −1  −1 −1 T  −1 −1  −1 −1
p )C p , ( f (C, C p )C p )  =  f (C, C p )C p , f (C, C p )C p 

⇔  f (C, C −1 −1
p )C p ∈ Sym(3), (34)

which is not true, in general. However, it is an associated plasticity model in the

sense of Definition 1, see Proposition 3. We also remark that
Comparison of Isotropic Elasto-Plastic Models for the Plastic … 177

d −1 −λ+  −1 −1
p dev3 [ f (C, C p ) C p ]
Cp [C ] = . (35)
dt p tr[(  p )C p ) ] − 3 [tr[ 
f (C, C −1 −1 2 1
f (C, C −1 −1 2
p )C p ]]

In conclusion, using Lemma 4, we have

Remark 4 Any continuous solution C p ∈ Sym(3) of the flow rule (24) belongs in
fact to PSym(3).
As for the thermodynamical consistency, we remark that

d  d −1 d −1
[W (C C −1  −1
p )] = D[ W (C C p )], C [C ] = C DC [W  (C C −1
p )] C p , [C p ]
dt dt p dt
1 d 1 d
= Σ  C p , [C −1 ] = C −1  C p , C p [C −1
Σ ]
2 dt p 2 p dt p
1 d 1 λ+
= − C −1 −1 C −1
 p  
p Σ C p , [C p ] C p  = −  p Σ C p , dev3 Σ,
2 dt 4  dev3 Σ
(36)

 leads to
which, using the formula F pT Σe F p−T = Σ,

d  1 λ+
[W (C C −1 C −1 F pT Σe F p−T C p , dev3 (F pT Σe F p−T )
p
)] = −
dt p
4  dev3 (F pT Σe F p−T ) p
1 λ+p 1
=− −T
Σe , Σe − tr(Σe ) · 1
4  dev3 (F p Σe F p )
T 3
1 λ+ p
=−  dev3 Σe 2 ≤ 0. (37)
4  dev3 (F pT Σe F p−T )

Note that this proof of thermodynamical consistency may be criticized because it

involves the variable F p , which should not appear at all. However, we may also use
(20) and (21) to obtain

d  1 λ+
[W (C C −1
p )] = −
p
 2]
tr[(dev3 Σ) (38)
dt 
4  dev3 Σ
1 λ+
=−
p
 U p )2 ≤ 0,
 dev3 (U p−1 Σ (39)

4  dev3 Σ

We may summarize the properties of the Lion 1997 model:

(i) from (24) it follows that C p (t) ∈ Sym(3);
(ii) plastic incompressibility: from (24) together with det C p (0) = 1 it follows that
det C p (t) = 1;
(iii) for the isotropic case, the right hand-side of (24) is a function of C −1
p and C
alone;
178 P. Neff and I.-D. Ghiba

(iv) from (i) and (ii) together and using Lemma 4 it follows that C p (t) ∈ PSym(3);
(v) it is thermodynamically correct;
(vi) it is an associated plasticity model in the sense of Definition 1, see Proposition 3.
Remark 5 (Simo-Miehe 1992 model vs. Lion 1997 model) In the anisotropic case,
the flow rule proposed by Simo and Miehe [41] (and later by Reese and Wriggers
[36] and Miehe [21]) is not completely equivalent with the flow rule proposed by
Lion (see also [1, 6]), since  dev3 τe  =  dev3 Σe  does not hold true in general.
However, the difference is nearly absorbed by the positive plastic multipliers. The
models may differ due to different yield conditions, but the flow rules are similar,
having the same performance with respect to the thermodynamic consistency. Both
models are consistent according to our Definition 2, but we may not switch between
them, since different elastic domains are considered, namely EΣe and Eτe , respectively.
This is in fact the main difference between these two models. Having different elastic
domains we have different boundary points, since a point of the boundary of Eτe is not
necessarily on the boundary of Eτe . Hence, in these two flow rules we have a different
behaviour corresponding to the indicator function of different domains. The material
may reach the boundary of the elastic domain Eτe , while it is strictly inside the elastic
domain EΣe , for the same local response.
However, we have the following result:
Proposition 1 In the isotropic case the flow rule proposed by Simo and Miehe [41]
is equivalent with the flow rule proposed by Lion [17].
Proof We compare the flow rules (17) and (24) and the proof is complete. 

5 The Simo and Hughes 1998 Plasticity Formulation

in Terms of a Plastic Metric

The book [42] has been edited years after the untimely death of J.C. Simo. In this
book also a finite strain plasticity model is proposed. However, this model has a
subtle fundamental deficiency which we aim to describe in the interest of the reader.
The flow rule considered in [42, p. 310] is

d −1 2 devn τe Cp
[C ] = − λ+ tr(Be ) F −1 F −T , Cp = , (40)
dt p 3 p  devn τe  det C p
1/3

where Be = Fe FeT , τe = 2 Fe DCe [W (Ce )] FeT = 2 Be D Be [W (Be )] is the elastic

Kirchhoff stress tensor and λ+ p ≥ 0 is the consistency parameter. If the plas-
tic flow is isochoric then det F p = det C p = 1. However, we must always have
−1
det C p = 1 = det C p by definition of C p . Since Fe = F F p−1 , we have tr(Be ) =
F p−T F T F F p−1  = 1, C C −1 −1
p  = tr(C C p ). Moreover, note that for elastically
isotropic materials it holds
Comparison of Isotropic Elasto-Plastic Models for the Plastic … 179

DCe [W (Ce )] = α1 1 + α2 Ce + α3 Ce2 ∈ Sym(3), (41)

τe = 2 Fe [α1 1 + α2 Ce + α3 Ce2 ] FeT ,

where α1 , α2 , α3 are scalar functions of the invariants of Ce which are functions of

C C −1 −T −1
p , see Lemma 2. Since C e = F p C F p , we obtain

τe = 2 F F p−1 [α1 1 + α2 F p−T C F p−1 + α3 F p−T C F p−1 F p−T C F p−1 ]F p−T F T

= 2 F f 1 (C, C −1
p )F ,
T
(42)

with f 1 (C, C −1 −1 −1 −1 −1 −1 −1
p ) : =α1 C p +α2 C p C C p + α3 C p C C p C C p ∈ Sym(3).Thus,
for elastically isotropic materials we deduce
2
F −1 [devn τe ] F −T =2 f 1 (C, C −1 −1
p ) − tr(F f 1 (C, C p ) F ) C
T −1
3
2
=2 f 1 (C, C −1 −1
p ) −  f 1 (C, C p ), C C
−1
∈ Sym(3), (43)
 3
1
 dev3 τe  =2  f 1 (C, C −1 −1 2 −1
p ) · C, C · f 1 (C, C p ) −  f 1 (C, C p ), C .
2
9
n τe
Hence, F −1  dev
devn τe 
F −T ∈ Sym(3) is a function of C, C p−1 . Therefore the flow rule
(40) can entirely be expressed in terms of C and C p−1 alone.
Remark 6 It is not true, in general, that the right hand side of the flow rule (40) is in
concordance with det C p (t) = 1, assuming that det C p (0) = 1.
Proof From (40) we obtain by right multiplication with C p

d −1 2 −1 −1 devn τe
[C p ] C p = − (det C p )−1/3 λ+ −1
p tr(C C p ) F p Fe F −T F p . (44)
dt 3  devn τe  e

On the other hand, we have4

d −1 −1 d −1 −1 d −1 −1 d −1
[det C p ] = Cof C p , [C p ] = det C P C p , [C p ] = det C P 1, [C p ]C p .
dt dt dt dt

Hence, we deduce

d −1 2 −1 devn τe
[det C p ] = − (det C p )−1/3 λ+ −1
p tr(C C p ) Fe F −T , 1. (45)
dt 3  devn τe  e

n τe
Since Fe−1  dev F −T is not necessarily a trace free matrix, we can not conclude
devn τe  e
−1
that det C p (t) = const. for all t > 0. For instance, for elastically isotropic materials
(see (43)) we have

4 Let us remark that d [det C −1 ] −1

= det C p 1,
−1 −1
dt [C p ]C p  shows that det C p >
d
dt p 0 by direct inte-
gration of the ordinary differential equation.
180 P. Neff and I.-D. Ghiba

2
dev3 τe =2 f 1 (C, C −1 −1
p ), C p  −  f 1 (C, C p ), CC , C p 
−1

 3 
1 −1
=2 α1 C −1 p , C p  − C , CC −1
, C p 
3 p
 
1 −1
+ 2 α2 C −1 p C C −1
p , C p  − C C C −1
, CC −1
, C p 
3 p p
 
1 −1
+ 2 α3 C −1 p C C −1
p C C −1
p , C p  − C C C −1
C C −1
, CC −1
, C p  ,
3 p p p

(46)

n τe
which shows that Fe−1  dev F −T is not necessarily a trace free matrix, see Appendix
devn τe  e
A.4. 

Summarizing the properties of the flow rule (40) we have:

(i) it is thermodynamically correct;
−1
(ii) the right hand side is a function of C and C p only;
(iii) from this flow rule it follows C p (t) ∈ Sym(3) and det C p (t) > 0. Hence, it
follows that C p (t) ∈ PSym(3);
(iv) plastic incompressibility: however, it does not follow from the flow rule that
det C p (t) = 1 (which must hold by the very definition of C p , since the right
hand side is not trace-free, in general;
(v) it is not an associated plasticity model in the sense of Definition 1.

6 The Helm 2001 Model

In this section we consider the model proposed by Helm [12], Vladimirov, Pietryga
and Reese [45, Eq. 25] (see also [35] and [39, Eq. 55] and the model considered by
Brepols, Vladimirov and Reese [1, page16], Shutov and Ihlemann [38, Eq. 80]). We
prove later that this model is similar to the model considered by Miehe [22] in 1995,
provided certain interpretations are included. Vladimirov, Pietryga and Reese [45,
Eq. 25] considered the following flow rule

d 
dev3 Σ
[C p ] = λ+
p  · Cp , (47)
dt  2)
tr((dev3 Σ)

 = 2 C DC [W
where Σ  (C C −1
p )] is not necessarily symmetric for general C p ∈
d
 · C p ∈ Sym(3), see Sect. 3. Therefore, we have [C p ] ∈
PSym(3), while (dev3 Σ)
dt
Sym(3). The flow rule (47) implies
Comparison of Isotropic Elasto-Plastic Models for the Plastic … 181

d   d −1  (C C −1 d −1
[W (C C −1 −1
p )] = D[ W (C C p )], C [C p ] = C DC [W p )] C p , [C ]
dt dt dt p
1 d 1 d
= Σ  C p , [C −1 ] = C −1 Σ C p , C p [C −1 ] (48)
2 dt p 2 p dt p
1 λ+
=− 
p
 C p , dev3 Σ.
C −1 Σ 
 2) p
2 tr((dev3 Σ)

Thus, using (20) and (21) we deduce

d  1 λ+
[W (C C −1
p )] = − 
p
 2]
tr[(dev3 Σ)
dt  2)
2 tr((dev3 Σ)
λ+
=−
p
 U p ) ≤ 0,
 dev3 (U p−1 Σ (49)
2
which shows thermodynamical consistency.
Summarizing, the Helm 2001 (Reese 2008 and Shutov-Ihlemann 2014) model
has the following properties:
(i) from (49) it follows that it is thermodynamically correct;
(ii) plastic incompressibility: from (47) and (11) it follows that det C p (t) = 1;
(iii) for the isotropic case, the right hand-side of the flow rule (47) is a function of
C −1
p and C alone;
(iv) from dtd [C p ] ∈ Sym(3). it follows, in the isotropic case, that C p (t) ∈ Sym(3);
(v) from (ii) and (iii) together and using Lemma 4 it follows that C p (t) ∈ PSym(3);
(vi) it has formally subdifferential structure, see Proposition 4. However, the elastic
domain Ee (Σ,  2 σy2 ) is not convex w.r.t Σ,
 see Remark 3.
3

Moreover, we see that the following result holds:

Proposition 2 The flow rule considered by Helm [12] coincides with the flow rule
(47), i.e. with the interpretation of Miehe’s proposal [22] presented in Sect. 3.

Proof The proof follows from (47) and combined with (23). 

7 The Grandi-Stefanelli 2015 Model

In this section we present a model based on one representation used by Grandi and
Stefanelli [10] and previously used by Frigeri and Stefanelli [7, p. 7]. We start by
computing
182 P. Neff and I.-D. Ghiba

d   d −1  (C C −1 d −1
W (C C −1 −1
p ) = D W (C C p ), C [C p ] = sym[C D W p )], [C ]
dt dt dt p
 −1  −1  d 
=  C p sym[C D W  (C C −1
p )] C p , C p [C −1 p ] C p .
   dt
◦
:= 21 Σ ∈ Sym(3)
(50)

It is now easy to see that, if we choose

 d  ◦
◦ χ (dev3 Σ),
p ] C p ∈ −∂ Σ
C p [C −1 (51)
dt
◦
where χ (dev3 Σ) is the indicator function of the convex elastic domain

◦
 
◦ 1 2 ◦ ◦ 1
E e (Σ, σy ) := Σ ∈ Sym(3) |  dev3 Σ2 ≤ σy2 ,
3 3

 (C C −1
then C p ∈ Sym(3) and the reduced dissipation inequality dtd W p ) ≤ 0 is sat-
isfied. Thus, the model is thermodynamically correct. We also remark that the flow
rule (51) implies

d −1 d    d 
tr( [C p ] C p ) =  [C −1
p ] Cp C p , 1 =  C p [C −1
p ] C p , 1 = 0.
dt dt dt
Hence, we obtain det C p (t) = 1 and further C p (t) ∈ PSym(3).
Using Lemma 3, we give next some new representations of the stress-tensor
◦  −1  −1  −1  −1
 (C C −1
Σ := 2 C p sym[C D W p )] C p
 (C C −1
= 2 sym C p (C D W p )) C p

in terms of the stress tensors Σe , Σ and τe , respectively. From (3) we obtain


C DW  (C C −1
p ) = 1 
Σ C p . We also use F p = R p U p = R p C p and F = Fe F p .
2
Hence, we deduce
◦  −1   
 C p C p −1 ) = sym( C p −1 Σ
Σ = sym( C p Σ  C p ),
◦  −1   −1   −1 
Σ = sym( C p F pT Σe F p−T C p ) = sym( C p C p R Tp Σe R p C p C p)
= sym(R Tp Σe R p ),
◦  −1   −1 
Σ = sym( C p F T τe F −T C p ) = sym( C p F pT FeT τe Fe−T F p−T C p )
= sym(R Tp FeT τe Fe−T R p ).
Comparison of Isotropic Elasto-Plastic Models for the Plastic … 183

Note that Σe is symmetric in case of elastic isotropy. Hence, for the isotropic case,
we have
◦ ◦
Σ = R Tp Σe R p , Σ = R Tp FeT τe Fe−T R p . (52)

◦ ◦
However, we have Σ2 =R Tp Σe R p 2 =Σe 2 , tr(Σ) = tr(R Tp Σe R p ) = tr(Σe ).
Together, we obtain that
◦
 dev3 Σ =  dev3 Σe .

In conclusion, for isotropic elastic materials we have the equivalence of the elastic
domains
◦ ◦ 1 1
E e (Σ, σy2 ) = Ee (Σe , σy2 ). (53)
3 3
Therefore, the flow rule (51) proposed by Grandi and Stefanelli [10] has the
following properties:
(i) it is thermodynamically correct;
(ii) from this flow rule it follows C p (t) ∈ Sym(3) and det C p (t) = 1. Hence, it
follows that C p (t) ∈ PSym(3);
◦ ◦
(iii) the elastic domain E e is convex w.r.t. Σ;
(iv) it is an associated plasticity model in the sense of Definition 1;
(v) it preserves ellipticity in elastic loading if the energy is elliptic throughout the
◦
domain E e which makes it useful in association with the exponentiated Hencky
energy WeH [8, 9, 27, 29, 30, 32, 33].
We finish this section by comparing the Helm 2001 model and the Lion 1997 flow
rule with the Grandi-Stefanelli 2015 model.
Proposition 3 In the isotropic case, the Lion 1997 flow rule (i.e. the Dettmer-Reese
2004 model [6]) is equivalent with the Grandi-Stefanelli 2015 flow rule.

Proof We recall that the flow rule of the Lion 1997 model is

d −1 −1 dev3 Σe
[C p ] = −λ+
p Fp F −T , λ+
p ∈ R+ , (54)
dt  dev3 Σe  p

for Σe ∈
/ int(Ee (Σe , 13 σy2 )). Since F p = R p C p and in the isotropic case Σe =
◦
R p Σ R Tp , using (53) we rewrite the Lion’s flow rule in the form

◦
d −1  −1 T dev3 (R p Σ R Tp )  −1
[C p ] = −λ+
p Cp Rp ◦ Rp Cp , λ+
p ∈ R+ ,
dt  dev (R Σ R T ) 3 p p
184 P. Neff and I.-D. Ghiba

◦ ◦ ◦
for R p Σ R Tp ∈
/ int(E e (Σ, 13 σy2 )), which is equivalent with

◦
 d −1  dev3 Σ
C p [C p ] C p = −λ+
p ◦ , λ+
p ∈ R+ , (55)
dt  dev3 Σ
◦ ◦ ◦ ◦ ◦ ◦ ◦
for R p Σ R Tp ∈
/ int(E e (Σ, 13 σy2 )). Moreover, R p Σ R Tp ∈ int(E e (Σ, 13 σy2 )) ⇔ Σ ∈
◦ ◦
int(E e (Σ, 13 σy2 )). Therefore, the flow rule (55) becomes

 d −1  ◦
C p [C p ] C p ∈ −∂ ◦ χ (dev3 Σ), (56)
dt Σ

which coincides with the Grandi-Stefanelli 2015 flow rule (51). 

Proposition 4 In the isotropic case, the Helm 2001 flow rule is equivalent with the
Grandi-Stefanelli 2015 flow rule, i.e. it is also equivalent with the Lion 1997 flow
rule and the Dettmer-Reese 2004 model.

Proof We have
◦  −1   
 C p ) = sym( C p −1 Σ
Σ = sym( C p Σ  C p C −1
p C p)
 −1  −1
 C p ) C p ),
= sym( C p (Σ

and we recall that for isotropic materials

 C p = F pT Σe F p−T C p = F pT Σe F p ∈ Sym(3)
Σ

holds. Hence, for isotropic materials

◦  −1     ◦ 
Σ=  C p ) C p −1 = C p −1 Σ
C p (Σ  C p, =
Σ
−1
Cp Σ Cp .

Using the above identity, we may rewrite the Helm 2001-flow rule (47) in the form

 ◦  −1
d dev3 ( C p Σ C p )
[C p ] = λ+ · Cp . (57)
dt p  ◦  −1 2
tr((dev3 ( C p Σ C p )) )
Comparison of Isotropic Elasto-Plastic Models for the Plastic … 185

We also have
 ◦  −1 ◦
tr( C p Σ C p ) = tr(Σ),
 ◦  −1  ◦  −1
dev3 ( C p Σ C p ) = C p (dev3 Σ) C p ,
 ◦  −1  ◦  −1  ◦  −1
tr([dev3 ( C p Σ C p )]2 ) = tr([ C p (dev3 Σ) C p ]2 ) = tr( C p (dev3 Σ)2 C p )
◦ ◦
= tr((dev3 Σ)2 ) = (dev3 Σ)2 , 1
◦ ◦ ◦
= dev3 Σ, dev3 Σ =  dev3 Σ2 .

Hence, Helm’s flow rule (47) is equivalent with

◦
d  dev3 Σ 
[C p ] = λ+
p Cp ◦ Cp
dt  dev3 Σ
◦
 d  + dev3 Σ
⇔ C p [C −1
p ] C p = −λp ◦ , (58)
dt  dev3 Σ

and the proof is complete. 

Remark 7 The equivalence is true for an isotropic formulation only. However, the
Grandi-Stefanelli model will provide a consistent flow-rule for a plastic metric also
in the anisotropic case.

An existence proof for the energetic formulation [7] of the model given by Grandi
and Stefanelli [10] together with a full plastic strain regularization can be given along
the lines of Mielke’s energetic approach [18, 19, 24–26].

8 Summary

In isotropic elasto-plasticity it is common knowledge that a reduction to a 6-

dimensional flow rule for a plastic metric C p is in principle possible. We have
discussed several existing different models. Not all of them are free of inconsis-
tencies. This testifies to the fact that setting up a consistent 6-dimensional flow-rule
is not entirely trivial.
One problem which often occurs, is that the flow rule for C p is written in terms
of F p , which however should not appear at all. One finding of our investigation is
that, nevertheless, in the isotropic case, all consistent flow rules can be expressed
in C p alone and are equivalent. The Grandi-Stefanelli model [10] has the decisive
186 P. Neff and I.-D. Ghiba

Fig. 1 Idealized, isotropic perfect plasticity models involving a 6-dimensional flow rule for C p
w.r.t. the reference configuration are considered. By definition, the trajectory for the plastic metric
C p (t) should remain in PSym(3). λ+ p is the plastic multiplier. We have recast all flow rules in the
√ √
format dtd [P −1 ] P ∈ −∂χ or P dtd [P −1 ] P ∈ −∂χ

advantage to be operable also in the anisotropic case. In Figs. 1 and 2 we summarize

the investigated isotropic plasticity models and we indicate if the known conditions
which make them consistent are satisfied.
Comparison of Isotropic Elasto-Plastic Models for the Plastic … 187

Fig. 2 An inconsistent model and a 9-dimensional flow rule for F p . They are associative, since
both flow rules are in the format dtd [P] P −1 ∈ −∂χ or dtd [ε p ] ∈ ∂χ

Acknowledgments We would like to thank Prof. Stefanie Reese (RWTH Aachen), Prof. Jörn
Ihlemann (TU Chemnitz), Dr. Alexey Shutov (TU Chemnitz) and Prof. Alexander Mielke (WIAS-
Berlin) for in-depth discussion of flow rules in C p . I.D. Ghiba acknowledges support from the
Romanian National Authority for Scientific Research (CNCS-UEFISCDI), Project No. PN-II-ID-
PCE-2011-3-0521.

Appendix

A.1 The Lion 1997 Model for the Neo-Hooke Elastic Energy

For a quick consistency check we exhibit the consistency of this model directly for
a Neo-Hooke elastic energy and we give the concrete expression for the functions
f (C, C −1  −1
p ) and f (C, C p ). To this end, we consider the energy

   
NH (Ce ) = μ tr Ce det C p =1 Ce
W 1/3
+ h(det C) = μ tr + h(det C).
det Ce det C 1/3

We deduce Σe := 2 Ce DCe [W  (Ce )] = 2 Ce μ 1 1/3 · 1 = 2 μ 1 1/3 · Ce . Hence,

det C det C
the flow rule (24) can be written in the form

d −1 −1 dev C e
[C p ] = −λ+
p Fp F −T
dt  dev Ce  p
 
λ+ 1
F p−1 Ce F p−T − tr(Ce ) · F p−1 F p−T
p
=−
 dev Ce  3
+  
λp −1 −1 1 −1 −1
=− C p C C p − tr(C p C) · C p . (59)
 dev Ce  3
188 P. Neff and I.-D. Ghiba

We also deduce
1 1
 dev Ce 2 = Ce 2 − [tr(Ce )]2 = F p−T C F p−1 2 − [tr(F p−T C F p−1 )]2
3 3
−T −1 −T −1 1 −T −1
= F p C F p , F p C F p  − F p C F p ), 12
3
1 1
= C p C, C C p  − [tr(C p C)]2 = [tr(C −1
−1 −1 −1
p C) ] −
2
[tr(C −1p C)] .
2
3 3
Therefore, we obtain
 
d −1 λ+ 1
C −1 −1
C −1
p
[C p ] = − p C − tr(C p C) · 1 p
dt tr[(C p C)2 ] − 13 [tr(C −1
−1 2 3
p C)]
 
λ+ 1
C −1 −1 −1
p
=− p C C p − tr(C C p ) · 1 .
[tr(C −1 −1 3
p C) ] − 3 tr[(C p C)]
2 1 2

(60)

Comparing (35), (60), (25) and (28), we deduce


f (C, C −1
p ) = C,
⎧ ⎫
⎨ −λ+  ⎬
f (C, C −1 ) =
p
dev3 ( C −1 C)  λ+ ∈ R+ .
p
⎩ p p
⎭
tr[(C C −1 −1 2
p ) ] − 3 [tr(C C p )]
2 1

We clearly see that even for this simple energy, we have tr[(C C −1 −1 2
p ) ] − 3 [tr(C C p )]
2 1

=  dev3 (C C −1
p ) , since if we assume the contrary we deduce
2

1 1
tr[(C C −1 −1 2 −1 2
p ) ] − [tr(C C p )] = C C p  − [tr(C C p )]
2 −1 2
3 3
⇔ C C −1 −1 T −1 −1
p , (C C p )  = ±C C p , C C p . (61)

On the other hand, we deduce

tr[(C C −1 −1 −1 −1 −1 −1
p ) ] = (C C p ) (C C p ), 1 = C p C C p (C C p ) C p , 1
2

= C −1 −1 −1 T −1 −1
p C C p C p (C C p ) , 1 = C p C, C C p . (62)

Since from Remark 4 it follows that C p ∈ PSym(3), we further deduce that

C −1 −1 −1 −1 −1 −1 −1
p C C p C p , C C p  = U p U p C C p U p U p , C C p 

= U p−1 C C −1 −1 −1
p Up, Up C C p Up

= U p−1 C U p−1 2 ≥ 0, (63)

Comparison of Isotropic Elasto-Plastic Models for the Plastic … 189

where U p2 = C p . Hence, [tr(C C −1

p ) ] ≥ 0 and from (61) we deduce
2

C C −1 −1 T −1 −1
p , (C C p )  = C C p , C C p  ⇔ C C −1 −1
p , skew(C C p ) = 0

⇔ skew(C C −1
p )=0 ⇔ C C −1
p ∈ Sym(3). (64)

In conclusion, tr[(C C −1 −1 2 −1 2
p ) ] − 3 [tr(C C p )]  =  dev3 (C C p ) and the flow-rule
2 1

χ
p ] ∈ − ∂ (dev Σ).
does not have a subdifferential structure of the form C p dtd [C −1

A.2 The Helm 2001 Model for the Neo-Hooke Energy

 (C C −1
For the simplest Neo-Hooke elastic energy W (Fe ) = tr(Ce ) = W p )=
1
2
−1
tr(C C p ), we have

 (C C −1 1 −1
DC [W p )] = C ⇒  = C C −1
Σ p ∈/ Sym(3), (65)
2 p
and the flow rule (47) implies

d dev3 (C C −1
p )
[C p ] = λ+
p · Cp
dt tr[(dev3 (C C −1
p )) ]
2

λ+ 1
[ C − tr(C C −1
p
= p ) · C p ] ∈ Sym(3) ⇒ C p ∈ Sym(3),
tr[(dev3 (C C −1
p )) ]
2 3

and also

d dev3 (C C −1
p )
[C p ] C −1 +
p = λp ⇒ det C p = 1. (66)
dt tr[(dev3 (C C −1
p )) ]
2

The thermodynamical consistency may follow from (49). An alternative proof,

directly for the Neo-Hooke case, results from (48) and (65), since we have at fixed
in time C

d  λ+
[W (C C −1
p )] = −
p
C −1  
p Σ C p , dev3 Σ
dt −1 2
4 tr[(dev3 (C C p )) ]
λ+
C −1 −1
p
=− p C, dev3 (C C p )
tr[(dev3 (C C −1
p )) ]
2
190 P. Neff and I.-D. Ghiba

λ+
(C C −1 −1
p
=− p ) , dev3 (C C p )
T
(67)
tr[(dev3 (C C −1
p )) ]
2

λ+
C −1/2 dev3 (C C −1 1/2 −1/2
dev3 (C C −1
p
=− p )C C p )C
1/2
, 1
tr[(dev3 (C C −1
p )) ]
2

λ+
dev3 (C 1/2 C −1 ) dev3 (C 1/2 C −1
p
=− p C
1/2 T
p C
1/2
), 1
tr[(dev3 (C C −1
p )) ]
2

λ+
 dev3 (C 1/2 C −1
p
=− p C ) ,
1/2 2

tr[(dev3 (C C −1
p )) ]
2

which is negative.5 Therefore, this model is thermodynamically correct as now shown

also for the simple Neo-Hooke energy.

A.3 Another Referential Model

We recall that, in view of Lemma 2, any isotropic free energy W defined in terms
of Fe can be expressed as W (Fe ) = W  (C C −1
p ). In order to assume that the reduced
dissipation inequality is satisfied, we compute

d   d −1
W (C C −1 −1
p ) = D W (C C p ), C [C ]
dt dt p
 (C C −1 −1 d 
d −1
= C D W p ) Cp , [C −1 ] C p  = Σ, [C ] C p .
dt p dt p

Here, Σ = 2 C DC [W  (C C −1
p )], as in the Reese 2008 and Shutov-Ihlemann 2014
model. It is tempting to assume the flow rule in the associated form (see e.g. the
habilitation thesis of Miehe [21, p. 73, Satz 5.32] or [23] and also [22, Table 1])

d −1
[C ] C p ∈ − ∂ Σ χ (dev3 Σ),
 (68)
dt p

where χ (dev3 Σ)
 is the indicator function of the convex elastic domain
 
2 2 2

Ee (Σ, σy ) := Σ  2 ≤ σy2 .
 ∈ R3×3 |  dev3 Σ
3 3

and C −1
5 Surprisingly, this follows even if C
p do not commute in general. If C and C p commute, then
X =C C −1
p ∈ Sym(3) and the quantity does have a sign, since then X T , dev3 X  =  dev3 X 2
≥ 0.
Comparison of Isotropic Elasto-Plastic Models for the Plastic … 191

Note that this flow rule (68) is not the formulation which Miehe seemed to intend.
We have discussed the correct interpretation in Sect. 3.
Regarding such a formulation we can summarize our observations:
(i) this flow rule is thermodynamically correct;
(ii) the right hand side is a function of C and C −1  
p only, i.e. Σ = Σ(C, C p );
−1

(iii) plastic incompressibility: from this flow rule it follows that det C p (t) = 1, since
the right hand side is trace-free;
(iv) however, the computed tensor C p (t) will not be symmetric since Σ  C −1
p ∈/
Sym(3) in general. For instance, for the simplest Neo-Hooke energy W (Fe ) =
tr(Ce ) = tr(C C −1 
p ) we have Σ = 2 C C p ∈
−1
/ Sym(3), Σ  C −1
p = 2 C Cp ∈
−2
/
Sym(3), in general, and the flow rule becomes

d −1 λ+ 1
[C C −2 tr(C C −1 −1
p
[C p ] = −2 p − p ) · Cp ] ∈
/ Sym(3);
dt  dev(C C −1
p ) 3
(69)

(v) it is an associated plasticity model in the sense of Definition 1.

In conclusion, this model is inconsistent with the requirement for a plastic metric, i.e.
C p ∈ Psym(3). Moreover, if we are looking to the flow rule in the associated form
considered in the habilitation thesis of Miehe [21, p. 73, Satz 5.32] (see [23] and also
[22, Table 1]), since the subdifferential ∂ Σ χ (dev3 Σ)
 of the indicator function χ is
the normal cone
$
1 2 0,  ∈ int(Ee (Σ,
Σ  1 σy2 ))
 σy ); dev3 Σ)
N (Ee (Σ,  = 
3
3 {λ+ dev3 Σ
p  dev Σ
+  / int(Ee (Σ,
 | λp ∈ R+ }, Σ ∈
 1 σy2 )).
3 3

the flow rule can be written in the form

d −1 dev3 Σ
[C p ] C p = −λ+ , (70)
dt p 
 dev3 Σ

which is not equivalent with the flow rule (22) considered by Miehe in [22], since
∈
Σ / Sym(3). Let us remark that we have the symmetries dev3 Σ  · C p ∈ Sym(3),
−1 
C p dev3 Σ ∈ Sym(3), but these do not assure that the flow rule (70) implies C p ∈
Sym(3).

A.4 The Simo-Hughes 1998-Model for the

Saint-Venant-Kirchhoff Energy and for the Neo-Hooke Energy

n τe
In order to see that the quantity Fe−1  dev F −T which appears in the Simo-Hughes
devn τe  e
flow rule is not necessarily a trace free matrix, we consider two energies: the
192 P. Neff and I.-D. Ghiba

isotropic elastic Saint-Venant-Kirchhoff energy and the energy considered by Simo

and Hughes [42, p. 307]. On the one hand, the well known isotropic elastic Saint-
Venant-Kirchhoff energy is

μ λ μ λ
WSVK = Ce − 12 + [tr(Ce − 1)]2 = Be − 12 + [tr(Be − 1)]2 ,
4 8 4 8
and the corresponding Kirchhoff stress tensor is given by

λ
τeSVK (U ) = D Be [W SVK (Be )] = μ (Fe−T Ce FeT − 1) + tr(Ce − 1) · 1
2
λ
= μ (Fe FeT − 1) + tr(Fe FeT − 1) · 1.
2
Hence, we deduce

1
Fe−1 [devn τeSVK ] Fe−T = μ Fe−1 devn [ Fe FeT ] Fe−T = μ Fe−1 [ Fe FeT − tr(Fe FeT ) · 1] Fe−T
3
1 −1 −T
= μ [ 1 − tr(Fe Fe ) · Fe Fe ]
T
3
1
= μ [ 1 − tr(Fe FeT ) · Fe−1 Fe−T ],
3

and further
1
Fe−1 [devn τeSVK ] Fe−T , 1 = μ  1 − tr(Fe FeT ) · Fe−1 Fe−T ], 1
 3 
1 −1 −T
=μ 3− tr(Fe Fe ) tr(Fe Fe )
T
3
 
1 −1
=μ 3− tr(Ce ) tr(Ce )
3
 
1
=μ 3− tr(Ce ) tr(Cof Ce ) .
3 det Ce

We remark that Fe−1 [devn τeSVK ] Fe−T , 1 = 0 if and only if tr(Ce ) tr(Cof Ce ) =
9 det Ce , which does not hold true in general. Since Ce and CofCe are coaxial and
symmetric, the problem can be reduced to the diagonal case, i.e. we may assume Ce =
diag(λ1 , λ2 , λ3 ), λi > 0. Hence the condition tr(Ce ) tr(Cof Ce ) = 9 det Ce , becomes

9 λ1 λ2 λ3 = (λ1 + λ2 + λ3 )(λ1 λ2 + λ2 λ3 + λ3 λ2 )
⇔ 0 = λ1 (λ2 − λ3 )2 + λ2 (λ3 − λ1 )2 + λ3 (λ1 − λ3 )2

which is satisfied if and only if λ1 = λ2 = λ3 . Therefore, for the Saint-Venant-

Kirchhoff energy, in this model, det C p = 1 is only true for the conformal mapping
Fe = λ · SO(3) ∈ R+ · SO(3).
Comparison of Isotropic Elasto-Plastic Models for the Plastic … 193

On the other hand, the energy considered by Simo and Hughes [42, p. 307] is

μ Be κ 
WSimo (Be ) =  1/3
− 1, 1 + (det Be − 1) − log(det Be ) ,
2 det Be 4

for which the Kirchhoff stress tensor is given by

   
Be κ 1
τeSimo = μ dev3 1/3
+ Je − · 1.
det Be 2 Je

Hence, we deduce

1
Fe−1 [devn τeSimo ] Fe−T , 1 = μ 1/3
Fe−1 [dev3 Be ] Fe−T , 1
det Be
1
=μ 1/3
dev3 Be , Fe−T Fe−1 
det Be
 
1 1 1
=μ 1/3
dev3 Be , Be−1  =μ 1/3
Be , Be−1  − tr(Be ) tr(Be−1 )
det Be det Be 3
 
1 1
=μ 1/3
3 − tr(Be ) tr(Be−1 )
det Be 3
 
1 1
=μ 4/3
3 det Be − tr(Be ) tr(Cof Be ) .
det Be 3

Therefore Fe−1 [devn τeSimo ] Fe−T , 1 = 0 if and only if 9 det Be = tr(Be ) tr(Cof Be ).
Similar as above, it follows that this holds true if and only if Fe = λ · SO(3) ∈
R+ · SO(3).

References

1. Brepols, T., Vladimirov, I., & Reese, S. (2014). Numerical comparison of isotropic hypo-and
hyperelastic-based plasticity models with application to industrial forming processes. Interna-
tional Journal of Plasticity, 63, 18–48.
2. Cleja-Ţigoiu, S. (2003). Consequences of the dissipative restrictions in finite anisotropic elasto-
plasticity. International Journal of Plasticity, 19(11), 1917–1964.
3. Cleja-Ţigoiu, S., & Iancu, L. (2013). Orientational anisotropy and strength-differential effect
in orthotropic elasto-plastic materials. International Journal of Plasticity, 47, 80–110.
4. Cleja-Ţigoiu, S., & Maugin, G. A. (2000). Eshelby’s stress tensors in finite elastoplasticity.
Acta Mechanica, 139(1–4), 231–249.
5. Cuitino, A., & Ortiz, M. (1992). A material-independent method for extending stress update
algorithms from small-strain plasticity to finite plasticity with multiplicative kinematics. Engi-
neering Computations, 9(4), 437–451.
6. Dettmer, W., & Reese, S. (2004). On the theoretical and numerical modelling of Armstrong-
Frederick kinematic hardening in the finite strain regime. Computer Methods in Applied
Mechanics and Engineering, 193(1), 87–116.
194 P. Neff and I.-D. Ghiba

7. Frigeri, S., & Stefanelli, U. (2012). Existence and time-discretization for the finite-strain Souza-
Auricchio constitutive model for shape-memory alloys. Continuum Mechanics and Thermo-
dynamics, 24(1), 63–77.
8. Ghiba, I. D., Neff, P., & Silhavy, M. (2015). The exponentiated Hencky-logarithmic strain
energy. International Journal of Non-Linear Mechanics, 71, 48–51.
9. Ghiba, I. D., Neff, P., & Martin, R. J. (2015). An ellipticity domain for the distortional Hencky-
logarithmic strain energy. In Proceedings of the Royal Society of London A, 471(2184). doi:10.
1098/rspa.2015.0510.
10. Grandi, D., & Stefanelli, U. (2015). Finite plasticity in P T P. PreprintarXiv:1509.08681
11. Gupta, A., Steigmann, D. J., & Stölken, J. S. (2011). Aspects of the phenomenological theory
of elastic-plastic deformation. Journal of Elasticity, 104(1–2), 249–266.
12. Helm, D. (2001). Formgedächtnislegierungen: experimentelle Untersuchung, phänomenol-
ogische Modellierung und numerische Simulation der thermomechanischen Materialeigen-
schaften. Ph.D-Thesis: Universität Kassel.
13. Kröner, E. (1955). Der fundamentale Zusammenhang zwischen Versetzungsdichte und Span-
nungsfunktionen. Zeitschrift fur Angewandte Mathematik und Physik, 142(4), 463–475.
14. Kröner, E. (1958). Kontinuumstheorie der Versetzungen und Eigenspannungen. Berlin:
Springer.
15. Kröner, E. (1959). Allgemeine Kontinuumstheorie der Versetzungen und Eigenspannungen.
Archive for Rational Mechanics and Analysis, 4(1), 273–334.
16. Lee, E. H. (1969). Elastic-plastic deformation at finite strains. Journal of Applied Mechanics,
36(1), 1–6.
17. Lion, A. (1997). A physically based method to represent the thermo-mechanical behaviour of
elastomers. Acta Mechanica, 123(1–4), 1–25.
18. Mainik, A., & Mielke, A. (2005). Existence results for energetic models for rate-independent
systems. Calculus of Variations and Partial Differential Equations, 22(1), 73–99.
19. Mainik, A., & Mielke, A. (2009). Global existence for rate-independent gradient plasticity at
finite strain. Journal of nonlinear science, 19(3), 221–248.
20. Maugin, G. (1994). Eshelby stress in elastoplasticity and ductile fracture. International Journal
of Plasticity, 10(4), 393–408.
21. Miehe, C. (1992). Kanonische Modelle multiplikativer Elasto-Plastizität. Thermodynamis-
che Formulierung und numerische Implementation. Habilitationsschrift: Universität Hannover,
Germany.
22. Miehe, C. (1995). A theory of large-strain isotropic thermoplasticity based on metric transfor-
mation tensors. Archive of Applied Mechanics, 66, 45–64.
23. Miehe, C. (1998). A constitutive frame of elastoplasticity at large strains based on the notion
of a plastic metric. International Journal of Solids and Structures, 35(30), 3859–3897.
24. Mielke, A. (2003). Energetic formulation of multiplicative elasto-plasticity using dissipation
distances. Continuum Mechanics and Thermodynamics, 15(4), 351–382.
25. Mielke, A. (2004). Existence of minimizers in incremental elasto-plasticity with finite strains.
SIAM Journal on Mathematical Analysis, 36, 384–404.
26. Mielke, A., & Müller, S. (2006). Lower semicontinuity and existence of minimizers in incre-
mental finite-strain elastoplasticity. ZAMM - Journal of Applied Mathematics and Mechanics,
86, 233–250.
27. Montella, G., Govindjee, S., & Neff, P. (2015) The exponentiated Hencky strain energy in
modelling tire derived material for moderately large deformations. To appear in Journal of
Engineering Materials and Technology-Transactions of the ASME. Preprint arXiv:1509.06541.
28. Neff, P., Chełmiński, K., & Alber, H. D. (2009). Notes on strain gradient plasticity: finite
strain covariant modelling and global existence in the infinitesimal rate-independent case.
Mathematical Models and Methods in Applied Sciences, 19, 307–346.
29. Neff, P., & Ghiba, I. D. (2014). Loss of ellipticity for non-coaxial plastic deformations in
additive logarithmic finite strain plasticity. International Journal of Non-Linear Mechanics,
81, 122–128. Preprint arXiv:1410.2819.
Comparison of Isotropic Elasto-Plastic Models for the Plastic … 195

30. Neff, P., & Ghiba, I. D. (2015). The exponentiated Hencky-logarithmic strain energy. Part III:
Coupling with idealized isotropic finite strain plasticity. Continuum Mechanics and Thermo-
dynamics, 28, 477–487. doi:10.1007/s00161-015-0449-y, the special issue in honour of D. J.
Steigmann.
31. Neff, P., & Knees, D. (2008). Regularity up to the boundary for nonlinear elliptic systems arising
in time-incremental infinitesimal elasto-plasticity. SIAM Journal on Mathematical Analysis,
40(1), 21–43.
32. Neff, P., Ghiba, I. D., & Lankeit, J. (2015). The exponentiated Hencky-logarithmic strain
energy. Part I. Journal of Elasticity, 121, 143–234.
33. Neff, P., Ghiba, I. D., Lankeit, J., Martin, R., & Steigmann, D. J. (2015). The exponentiated
Hencky-logarithmic strain energy. Part II: Coercivity, planar polyconvexity and existence of
minimizers. Zeitschrift fur Angewandte Mathematik und Physik, 66, 1671–1693.
34. Ortiz, M., & Simo, J. C. (1986). An analysis of a new class of integration algorithms for elasto-
plastic constitutive relations. International Journal for Numerical Methods in Engineering,
23(3), 353–366.
35. Reese, S., & Christ, D. (2008). Finite deformation pseudo-elasticity of shape memory alloys-
Constitutive modelling and finite element implementation. International Journal of Plasticity,
24(3), 455–482.
36. Reese, S., & Wriggers, P. (1997). A material model for rubber-like polymers exhibiting plastic
deformation: computational aspects and a comparison with experimental results. Computer
Methods in Applied Mechanics and Engineering, 148, 279–298.
37. Shutov, A.V. (2014). Personal comunication. 8/2014.
38. Shutov, A. V., & Ihlemann, J. (2014). Analysis of some basic approaches to finite strain elasto-
plasticity in view of reference change. International Journal of Plasticity, 63, 183–197.
39. Shutov, A. V., & Kreißig, R. (2008). Finite strain viscoplasticity with nonlinear kinematic
hardening: Phenomenological modeling and time integration. Computer Methods in Applied
Mechanics and Engineering, 197(21), 2015–2029.
40. Shutov, A. V., Landgraf, R., & Ihlemann, J. (2013). An explicit solution for implicit time
stepping in multiplicative finite strain viscoelasticity. Computer Methods in Applied Mechanics
and Engineering, 265, 213–225.
41. Simo, J.C. (1993). Recent developments in the numerical analysis of plasticity. In E. Stein
(ed.), Progress in computational analysis of inelastic structures (pp. 115–173). Springer.
42. Simo, J.C., & Hughes, J.R. (1998). Computational Inelasticity., volume 7 of Interdisciplinary
Applied Mathematics. Springer, Berlin.
43. Simo, J. C., & Ortiz, M. (1985). A unified approach to finite deformation elastoplastic analysis
based on the use of hyperelastic constitutive equations. Computer Methods in Applied Mechan-
ics and Engineering, 49, 221–245.
44. Steigmann, D. J., & Gupta, A. (2011). Mechanically equivalent elastic-plastic deformations
and the problem of plastic spin. Theoretical and Applied Mechanics, 38(4), 397–417.
45. Vladimirov, I., Pietryga, M., & Reese, S. (2008). On the modelling of non-linear kinematic
hardening at finite strains with application to springback-comparison of time integration algo-
rithms. International Journal for Numerical Methods in Engineering, 75(1), 1–28.
Quasi-Static Evolutions in Brittle Fracture
Generated by Gradient Flows: Sharp Crack
and Phase-Field Approaches

Matteo Negri

Abstract In this paper we will describe how gradient flows, in a suitable norm,
are natural and helpful to generate quasi-static evolutions in brittle fracture. First,
we will consider the case of a brittle crack running along a straight line according
to Griffith’s law. Then, we will see how the same approach leads to quasi-static
evolutions in the phase field setting, taking into account the alternate minimization
scheme. In the latter, the norm associated to the gradient flow is not “user supplied”,
however, the algorithm itself together with the separate quadratic structure of the
energy defines a family of norms which, in the limit, characterizes the quasi-static
evolution. Mathematically speaking, all of these evolutions are (parametrized) BV -
evolutions.

1 Introduction

The idea of using monotone descent paths (among which the gradient flow) for
quasi-static crack propagation goes back to the foundation of fracture mechanics:
according to Griffith’s principle [12] “the system can pass from the unbroken to the
broken condition by a process involving a continuous decrease of potential energy”.
Choosing the gradient flow, in a suitable norm, as optimal and most common descent
path, Griffith’s criterion would be: “the system follows the gradient flow of the
potential energy”.
As a matter of fact, gradient flows usually refer to time dependent problems while
Griffith’s law does not make any reference to time, neither directly or indirectly
(e.g. in terms of velocities). In our rate-independent setting, the gradient flows will
provide in fact a parametrization of the path connecting “the unbroken” with “the
broken condition”. Such a parametrization will appear both in the construction of the
solution, by time discretization, and in the quasi-static evolution itself, specifically
in the instantaneous “catastrophic” propagations.

M. Negri (B)
Department of Mathematics, University of Pavia, Via Ferrata 1, 27100 Pavia, Italy
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 197

K. Weinberg and A. Pandolfi (eds.), Innovative Numerical Approaches
for Multi-Field and Multi-Scale Problems, Lecture Notes in Applied
and Computational Mechanics 81, DOI 10.1007/978-3-319-39022-2_9
198 M. Negri

Our construction of quasi-static propagations follows closely a well known

scheme in computational mechanics: we employ a uniform time discretization, say
tk = kΔt, together with an incremental law based on Griffith’s principle: at each
discrete time tk the crack advances along a decreasing path of the energy, if any, until
it reaches a stationary point of the energy; decreasing paths will then be written as
suitable gradient flows.
In particular, for a straight crack this scheme provides in the limit, as Δt → 0, a
quasi-static evolution which can be described rigorously in several equivalent ways:
by means of Karush-Kuhn-Tucker (KKT) conditions (cf. Theorem 1 and [23]) by
BV -solutions [18] (cf. Corollary 1 and [19]) or by parametrized BV -solutions [9]
(cf. Corollary 2 and [22]). At this point it is important to remark that in the limit, as
Δt → 0, the quasi-static evolution can be discontinuous in time. This is a common
feature of BV -solutions for rate-independent systems and, most important, it is not a
pure mathematical artefact. In the rate independent setting discontinuities represent
catastrophic propagations of the crack, which can happen in real life: a numerical
example (cf. Sect. 2) shows a clear jump discontinuity in a standard ASTM com-
pact tension test. Mathematically, jump discontinuities are characterized by unstable
regimes of propagation where instantaneously the crack advances following a gra-
dient flow, which is indeed “a process involving a continuous decrease of potential
energy”.
For the general situation in which the crack path is unknown, both mathematical
and numerical models involving geometrical and topological features of the crack
become sensibly harder. Facing these problems is challenging but in practice it is
more convenient to employ regularized models [1, 5, 7, 13, 16, 17, 24, 26] which
bypass the issues related to the morphology of the crack. One of the most success-
ful choice is the phase field approach, which has been implemented in different
ways and for several problems in fracture. Here we will focus our interest on the
evolution obtained with a very efficient numerical method, known as alternate min-
imization [5]. In this scheme at each time tk the evolution is obtained by a sequence
of (quadratic) minimization problems, which produces a monotone decreasing path
of the energy, in agreement with Griffith’s criterion. Formally this scheme “defines”
a discrete evolution law for the crack (represented by the phase-field variable). Our
goal, in analogy with the straight crack problem, is to characterize the limit evolution
obtained by letting Δt → 0 and to show its main properties. First, we will see that in
the limit we get a quasi-static BV -evolution, which in general does not coincide with
the evolution obtained by global minimization problems [10]. Then, we recast the
evolution by KKT conditions where it appears an energy release rate; in this respect,
note that the alternate minimization algorithm does not employ explicitly any kind
of energy release. Finally, we show that the irreversibility constraint is thermody-
namically consistent and that the evolution of the displacement field follows a sort
of visco-elastic flow in the jumps.
Mathematically, in order to characterize the limit it is fundamental to recast alter-
nate minimization as a gradient flow, with respect to a suitable family of norms,
induced by the separately quadratic structure of the phase-field energy; clearly
this is a particular choice, which works extremely well, but other choices are also
Quasi-Static Evolutions in Brittle Fracture Generated … 199

possible and worth studying, e.g. [22]. In this work, proofs and fine mathematical
details are not included; the interested reader can make reference for instance to
[14, 19, 22, 23].

2 Sharp Crack

2.1 Setting: Compact Tension

In order to avoid technical issues as much as possible, we will state our results only
for a representative example, cf. Fig. 1. Denote by Ω the open set in Fig. 1 (obtained
removing a couple of symmetric holes from a rectangle). Let ∂Ω = ∂ D Ω ∪ ∂ N Ω
where ∂ N Ω is the boundary of the rectangle while ∂ D Ω denotes (the union of) the
boundaries of the circular holes. Assume that the initial crack K 0 is given by the
line segment (0, l0 ] × {0} for l0 > 0. In our simple setting the crack will propagate
horizontally, thus our family of admissible cracks will be given by the line segments
of the form K l = (0, l] × {0}. Clearly such a family is simply parametrized by the
scalar l ∈ [l0 , L), which gives as well the position of the crack tip.
Consider on ∂ D Ω a proportional boundary condition of the form u = ±t ê where
the sign ± is chosen as in Fig. 1.
We consider in-plane elasticity with linearised energy density

W (Du) = 1
2
Du : C[Du] = 1
2
ε(u) : σ (u)

where ε(u) = (Du + Du T )/2 and C[Du] = σ (u) = 2με(u) + λtr(ε(u))I , for
λ, μ > 0 the Lamé coefficients. For t ∈ [0, T ] and l ∈ [l0 , L) the space of admissible
configurations is

Ut,l = {u ∈ H 1 (Ω\K l , R2 ) : u = ±t ê ∂ D Ω} .

Fig. 1 An ASTM-compact
tension geometry: the set
Ω\K 0

0 l0 L 
200 M. Negri

Hence, for u ∈ Ut,l the elastic energy will be


E(u) = W (Du) d x .
Ω\K l

Before proceeding it is convenient to introduce the reduced elastic energy: for t ∈

[0, T ] and l ∈ [l0 , L) let
E (t, l) = E(u t,l ),

where u t,l ∈ argmin{E(u) : u ∈ Ut,l }. Note that in quasi-static evolutions it is not

restrictive to employ E instead of E since it is assumed that the system is always in
equilibrium and u t,l is indeed the only equilibrium point; in particular it solves the
PDE ⎧
⎨ div(σ (u t,l )) = 0 Ω\K l
u t,l = ±t ê ∂D Ω (1)
⎩
σ (u t,l ) n̂ = 0 ∂ N Ω ∪ K l± .

Note that the Neumann homogeneous boundary condition holds on the boundary
∂ N Ω of the rectangle and on both the crack faces, above denoted by K l± .
Let us now turn to dissipation. Since we are interested in brittle fracture the energy
dissipated by the crack will be provided by a potential K : [l0 , L) → R+ which is
simply of the form K (l) = G c (l − l0 ), being G c > 0 the material toughness.
In the sequel we will always work with the reduced total energy F : [0, T ] ×
[l0 , L) → R+ given by
F (t, l) = E (t, l) + K (l).

Before proceeding it is fundamental to have at our disposal the partial derivatives

of the energy F .

Lemma 1 The energy F : [0, T ] × [l0 , L) → R+ is differentiable with respect to

both its variables with

∂t F (t, l) = ∂t E (t, l) = (±ê) · σ (u t,l ) n̂ ds = P ext (t, l),
∂D Ω
∂l F (t, l) = −G(t, l) + G c ,

where P ext is the power of the external forces while G denotes as usual the energy
release rate. Moreover, G(t, ·) is non-negative and locally Lipschitz continuous in
[l0 , L).

A proof can be adapted e.g. from [19] or [23]. In this setting, by irreversibility,
equilibrium reads

∂l F (t, l) = −G(t, l) + G c ≥ 0 ⇔ G(t, l) ≤ G c . (2)

Quasi-Static Evolutions in Brittle Fracture Generated … 201

2.2 Discrete in Time Evolution

Now, we will define the discrete in time evolution by a sequence of incremental

problems. Denote by  the evolution in time. Given Δt > 0 let tk = kΔt for k =
0, ..., [T /Δt] and let (t0 ) = l0 . Knowing (tk ) we define (tk+1 ) as

(tk+1 ) = min{l ≥ (tk ) : G(tk+1 , l) ≤ G c }. (3)

In other terms, we advance the crack up to the closest equilibrium point. This is in
some sense a return mapping algorithm on the set {G(tk+1 , l) ≤ G c } of equilibrium
points at time tk+1 , which is usually at the core of many crack tracking algorithms.
Now, let us see how to recast the incremental problem as a gradient flow. By
irreversibility the crack cannot heal, for this reason it is convenient to introduce the
one sided “slope”
|∂l F (t, l)|−

where | · |− denotes the negative part. Next, let us introduce an auxiliary parameter
s ∈ R+ and an auxiliary function l : R+ → [l0 , L). We set (tk+1 ) = sups l(s) where
l solves the gradient flow

˙ = |∂l F (tk+1 , l(s))|−
l(s)
l(0) = (tk ).

In this simple setting the gradient flow boils down to an autonomous Cauchy problem
for a non-linear, first order ODE. Intuitively, l grows when ∂l F (tk+1 , l) < 0, i.e. when
G(tk+1 , l) > G c and thus when the crack is not in equilibrium. It is easy to see that
there exists a unique solution and that the definition (tk+1 ) = sups l(s) coincides
with (3) (for a proof, see [19]).
At this point we have defined (tk ) for tk = kΔt. Now consider a sequence of time
steps Δtn  0 and denote by n the corresponding discrete evolutions, defined in
the discrete points tn,k = kΔtn . Denote again by n : [0, T ] → [l0 , L) the piecewise
affine interpolate of n (tn,k ). By Helly’s Theorem it follows that (up to subsequences)
n converges pointwise to a limit evolution . Our goal is now the characterization
of : we will provide two characterizations, the first in terms of KKT conditions,
the second in terms of parametrized BV-evolutions. In order to better understand
the meaning of these characterizations it is useful to show first an explicit example,
which has been computed numerically.

2.3 Example

First, let us comment on Fig. 2. Remember that the “loading” is monotone increasing.
The set of critical points of the energy, i.e. {(t, l) : G(t, l) = G c } is represented with
a dotted curve. This curve splits the (t, l)-plane into two regions: on the left is the set
202 M. Negri

Fig. 2 Quasi-static
evolutions: of energetic type
+
(dashed) and of BV-type
(solid) G < Gc

G > Gc

−

t# t∗

Fig. 3 A detail of the energy

•
landscape F (t ∗ , l)

− +

{(t, l) : G(t, l) < G c } of stable points while on the right is the set {(t, l) : G(t, l) >
G c } of unstable points. From the picture it is clear that up to time t ∗ the crack is
not moving, since it is inside the stable region. At time t ∗ a “catastrophic evolution”
occurs: in the instantaneous transition from − to + the system crosses the unstable
region, since G(t ∗ , l) > G c for every l ∈ (− , + ). The energy landscape at time t ∗
is reported in Fig. 3: it is clear that F (t ∗ , ·) is not convex and that  makes a transition
from − to + following a descent path, in particular F (t ∗ , − ) > F (t ∗ , + ).
Figure 2 shows (in dashed bold line) also the energetic evolution, obtained by
global energy minimization; this evolution presents a discontinuity as well, however
the qualitative behaviour is quite different: the system crosses first the stable and
then the unstable region, in particular a propagation occurs even if G(t # , − ) < G c .

2.4 Characterization by Karush-Kuhn-Tucker Conditions

The next Theorem provides the first characterization of the limit evolution. For a
proof, see [19].
Theorem 1 The limit evolution , obtained letting Δtn  0, is non-decreasing and
belongs to BV (0, T ). Moreover

G(t,  − (t)) ≤ G c for t ∈ [0, T ] , (4)

 
G(t,  − (t)) − G c d(t) = 0 in the sense of measures in [0, T ]. (5)
Quasi-Static Evolutions in Brittle Fracture Generated … 203

Furthermore, for t ∈ J () (the set of jumps) we have

G(t, l) ≥ G c for every l ∈ [ − (t),  + (t)]. (6)

Let us make some comments on the above Theorem. Clearly (4) and (5) play the
role of the “classical” KKT conditions, with some technical differences: the left limit
− is used instead of  and the weak measure theoretic derivative d is used instead
of the speed . ˙ Both these technical details are due to the fact that in general the
evolution belongs to BV (0, T ) and thus it may have jump discontinuities. However
the qualitative meaning of (4) and (5) is quite clear and consistent with standard
KKT conditions. What is instead not common in the study of quasi-static evolutions
is condition (6) which characterizes the behaviour in the jumps in terms of unstable
branches of propagation (cf. also Fig. 2).
Finally, in terms of derivatives of the energy, the above Theorem reads as follows.
Corollary 1 The limit , obtained letting Δtn  0, is non-decreasing and belongs
to BV (0, T ). Moreover

∂l F (t, − (t)) ≥ 0 for t ∈ [0, T ] , (7)

−
∂l F (t,  (t)) d(t) = 0 in the sense of measures in [0, T ]. (8)

Furthermore, for t ∈ J () (the set of jumps) we have

∂l F (t, − (t)) ≤ 0 for every l ∈ [ − (t),  + (t)]. (9)

At this point it is necessary to introduce the concept of BV -solution. Here we will

give the “simplest” possible definition, for a general treatise see [18].
Corollary 2 The limit  is non-decreasing and of class BV (0, T ). Moreover

|∂l F (t, − (t))|− = 0 for t ∈ [0, T ] (10)

and for every t ∈ [0, T ] the following energy identity holds:

 t
F (t, − (t)) = F (0, l0 ) + ∂t F (τ, (τ )) dτ − diss(F (t, ·)), (11)
0 t∈J ()

where  + (t)
diss(F (t, ·)) = |∂l F (t, l)|− dl
− (t)

denotes the “energy gap” in the discontinuity points.

An evolution  which satisfies (10) and (11) is called a BV-solution.
It is important to note that equilibrium (10) and energy balance (11) provide a
very concise and mathematically convenient way of characterizing the quasi-static
204 M. Negri

evolution; they are indeed equivalent to the KKT conditions of Theorem 1. Without
entering too much into the technical details (for a complete proof the reader can
follow [20] or [22]) let us see how (7)–(9) follow from (10)–(11). First,

|∂l F (t, − (t))|− = 0 ⇔ ∂l F (t, − (t)) ≥ 0 ,

gives (7). Next, by the chain rule in BV (0, T )

 t  t
−
F (t,  (t)) = F (0, l0 ) + ∂t F (τ, (τ )) dτ + ∂l F (τ, − (τ )) dac (τ ) +
0 0

+ F (t, ·),
t∈J ()

where dac  denotes the (weak) measure theoretic derivative of  in [0, T ] \ J ().
Comparing with (11) we get
 t
∂l F (τ, − (τ )) dac (τ ) + F (t, ·) = − diss(F (t, ·)).
0 t∈J () t∈J ()

Since the measure dac  is supported in [0, T ]\J () it follows that

∂l F (t, − (t)) dac (t) = 0 in the sense of measures in [0, T ]

and that
 + (t)  + (t)
F (t, ·) = ∂l F (t, l) dl = − |∂l F (t, l)|− dl, for every t ∈ J ().
− (t) − (t)

The former leads to (8), thanks to the continuity of G, while the latter leads to

∂l F (t, l) = −|∂l F (t, l)|− for every l ∈ [− (t), + (t)],

which is in turn equivalent to

∂l F (t, l) ≤ 0 for every l ∈ [− (t), + (t)],

that is (9).

2.5 Characterization as a Graph Parametrized BV-evolution

We have seen that the limit evolution  can have jump discontinuities in time. For this
reason it is convenient, both for theoretical and numerical purposes, to represent 
Quasi-Static Evolutions in Brittle Fracture Generated … 205

by a (Lipschitz) parametrization of the form s → (t (s), l(s)) of the extended graph.

Remember that the extended graph is just the set {(t, l) : − (t) ≤ l ≤ + (t)} obtained
“completing” the jumps with a vertical line segments (see Fig. 2). Therefore, using
the map s → (t (s), l(s)) a jump of  at time t ∗ will be characterized by t (s) = t ∗
in [s1 , s2 ] together with l(s1 ) = − (t ∗ ) and l(s2 ) = + (t ∗ ). In essence, continuity
points (in time) will correspond to points with t (s) > 0 while discontinuity points (in
time) will correspond to points with t (s) = 0. The advantages of this representation,
originally suggested in [9], will become more clear in Sect. 3 and in general are quite
evident in the case of infinite dimensional systems [22].
Here for sake of simplicity we will skip any argument on the different ways
which provide existence of a parametrized evolution. We will instead assume that
s → (t (s), l(s)) is a parametrization of the extended graph of the solution , obtained
as above by letting Δtn  0. We will consider also t ≥ 0 (in order to avoid physically
meaningless cases) and we will normalize the parametrization with t (s) + l (s) = 1
(for a.e. s ∈ [0, S]). The resulting parametrization satisfies the following properties.

Theorem 2 The (normalized) parametrization s → (t (s), l(s)) satisfies t (s) ≥ 0,

l (s) ≥ 0 and t (s) + l (s) ≤ 1. Moreover it satisfies the following conditions:

|∂l F (t (s), l(s))|− = 0 for every s with t (s) > 0, (12)

 s
F (t (s), l(s)) = F (0, l(0)) + ∂t F (t (r ), l(r )) t (r ) dr +
 s 0

− |∂ F (t (r ), l(r ))|− l (r ) dr for every s. (13)

0

A (normalized) parametrization s → (t (s), (s)) which satisfies (10) and (11) is

called a parametrized BV-solution.

It is not difficult to see that, upon choosing the right parametrization, (12)–(13)
follows from (10)–(11) and vice versa (for a proof see [22]).

3 Alternate Minimization Scheme in the Phase-Field

Approach

In this section we will deal with the phase-field approach for fracture, which in the
last decade has been an effective and popular method for the simulation of crack
propagation, see e.g. [1, 5, 7, 13, 16, 17] and many others. In particular we will
consider evolutions defined by the alternate minimization scheme; we will see that
in the limit they will provide indeed parametrized BV-evolution, with respect to a
suitable family of norms.
206 M. Negri

3.1 Setting

For sake of simplicity let us consider the same geometry and the same boundary
conditions of Sect. 2.1. In the phase-field framework it is however necessary to re-
define the space of admissible displacements by

Ut = {u ∈ H 1 (Ω, R2 ) : u = ±t ê on ∂ D Ω}

and to introduce a set for the phase field variables

Z = {z ∈ H 1 (Ω) : 0 ≤ z ≤ 1}.

By linearity we can always write Ut = tU where

U = {u ∈ H 1 (Ω, R2 ) : u = ±ê on ∂ D Ω}.

In the sequel we will work indeed with the spaces U and Z which are independent
of time. For ε > 0 and ηε > 0, typically with ηε = o(ε), the phase field elastic and
dissipated energy [2] will be respectively

Eε (t, u, z) = 1
2
t 2 (z 2 + ηε )W (Du) d x,
Ω

Kε (z) = 1
G
2 c
(z − 1)2 /2ε + ε|∇z|2 d x.
Ω

The total energy Fε : [0, T ] × U × Z → R will be Fε (t, u, z) = Eε (t, u, z) +

Kε (z).
In the sequel it will be fundamental to have at hand the partial derivatives of the
energy,

∂t Fε (t, u, z) = t (z 2 + ηε )W (Du) d x, (14)
Ω

∂u Fε (t, u, z)[φ] = t 2 (z 2 + ηε )σ (u) : ε(φ) d x, (15)
Ω
 
∂z Fε (t, u, z) [ξ ] = t 2 zξ W (Du) d x + G c (z − 1)ξ/2ε + ε∇z · ∇ξ d x.
Ω Ω
(16)

For our purposes the spaces of admissible variations for Z and U will be provided
respectively by

Ξ = {ξ ∈ H 1 (Ω) : ξ ≤ 0}, Φ = {φ ∈ H 1 (Ω, R2 ) : φ = 0 on ∂ D Ω}.

Quasi-Static Evolutions in Brittle Fracture Generated … 207

To conclude this section, let us see how to define a notion of energy release,
with respect to a variation ξ , and how to get the power of external forces. To this
end, denoting u(t, z) ∈ argmin {Eε (t, u, z) : u ∈ U } we will call “energy release
functional” (with respect to a variation ξ )

Eε (t, u(t, z + hξ ), z + hξ ) − Eε (t, u(t, z), z)

Gε (t, z)[ξ ] = − lim+
h→0 h

In this way the displacement field changes “simultaneously” with the variation of
the phase field variable. Actually, by minimality of u(t, z), the derivative can be
represented explicitly as (see e.g. [15])

Gε (t, z)[ξ ] = −∂z Eε (t, u, z) [ξ ] = − t 2 zξ W (Du) d x . (17)
Ω

Now, let us introduce the set of normalized variations


Ξ̂z = ξ ∈ Ξ : (z − 1)ξ/4ε + ε∇z · ∇ξ d x ≤ 1 .
Ω

Note that these variations normalize the “variation of crack length” since

dKε (z)[ξ̂ ] = G c (z − 1)ξ̂ /4ε + ε∇z · ∇ ξ̂ d x ≤ G c .
Ω

Then, we can define the energy release as

G ε (t, z) = sup{Gε (t, z)[ξ̂ ] : ξ̂ ∈ Ξ̂z }. (18)

Finally, Green’s formula allows to rewrite (14) as


∂t Fε (t, u(t, z), z) = (±ê) · σ z (tu(t, z)) n̂ ds = Pεext (t, u(t, z), z), (19)
∂D Ω

where σ z (w) denotes the phase field stress, that is

σ z (w) = (z 2 + ηε )σ (w).

3.2 Discrete in Time Evolution by Alternate Minimization

As we did in Sect. 2.2, given Δt > 0, let tk = kΔt and set the initial conditions
u(t0 ) = u 0 and z(t0 ) = z 0 . Known u(tk−1 ) and z(tk−1 ) we will introduce a couple of
auxiliary sequences, u m and z m , with u 0 = u(tk−1 ) and z 0 = z(tk−1 ) defined recur-
sively by the following alternate minimization scheme [5]
208 M. Negri

  
u m ∈ argmin Fε tk , · , z m−1 : u ∈ U ,
(20)
z m ∈ argmin Fε (tk , u m , · ) : z ∈ Z with z ≤ z m−1 },

where the constraint z ≤ z m−1 models the irreversibility of the crack. Then we define
the updates u(tk ) = limm→+∞ u m and z(tk ) = limm→+∞ z m . More precisely, we have
the following result (for a proof see [14]).

Proposition 1 The sequence u m converges to u(tk ) strongly in H 1 (Ω, R2 ) while z m

converges to z(tk ) strongly in H 1 (Ω). Further,

∂u Fε (tk , u(tk ), z(tk ))[φ] = 0 for every φ ∈ Φ,

∂z Fε (tk , u(tk ), z(tk ))[ξ ] ≥ 0 for every ξ ∈ Ξ.

In other terms, u(tk ) is an equilibrium point for Fε (tk , ·, z(tk )) while z(tk ) is an
equilibrium point for Fε (tk , u(tk ), ·) (for the latter remember the irreversibility con-
straint).

In this way, given Δt > 0 a time discrete evolution is provided in terms of the
equilibrium configurations (u(tk ), z(tk )) for tk = kΔt. In order to understand the
limit evolution, obtained by letting Δt → 0, we have first to recast the alternate
minimization scheme as a gradient flow. This is the goal of the next section.

3.3 Minimization as a Gradient Flow

3.3.1 An Illustrative Example

In order to better understand the gradient flow structure behind (20) let us start with
an example: the minimization of a quadratic functional in a finite dimensional√setting.
Let F(x) = 21 x T Ax + b T x + c for x ∈ Rn and A T = A > 0. Let x A = x T Ax
be the “natural” norm induced by the symmetric, positive define matrix A, with
associated scalar product ·, · A .
Our problem is the following: given x0 find the increment x∗ in such a way that
x0 + x∗ is the minimizer of F. Since F is quadratic we can write

F(x0 + x∗ ) = F(x0 ) + ∇ F(x0 )x∗ + 21 x∗T Ax∗

and we can characterize x∗ by stationarity of the energy, i.e.,

∇ T F(x0 ) + Ax∗ = 0 ⇔ x∗ = −A−1 ∇ T F(x0 ) = −∇ AT F(x0 ).

Quasi-Static Evolutions in Brittle Fracture Generated … 209

In the last term of the previous row we have introduced the notation ∇ AT F(x0 ) which
denotes the gradient of F (computed in x0 ) with respect to the norm  ·  A , that is
the (unique) vector such that d F(x0 )[x ] = ∇ A F(x0 ), x  A for every x ∈ Rn . For
our purposes it is just important to remark that ∇ A F(x0 ) = ∇ F(x0 )A−1 and that

A F(x0 ) = −∇ A F(x0 )/∇ A F(x0 ) A = argmin d F(x0 )[x ] : x  A ≤ 1 .
−∇

In other terms, the normalized gradient ∇ A F(x0 ) provides the steepest descent
direction with respect to the norm  ·  A . Now, let k = ∇ A F(x0 ) A and define
x(s) = x0 − λ(s) ∇A F(x0 ) where λ(s) = 1 − e−ks for s ∈ [0, +∞). Then, by homo-
geneity, x(s) solves the gradient flow

x (s) = −∇ A F(x(s)) for s ∈ [0, +∞),
(21)
x(0) = x0 .

Note that the right hand side is evaluated in x(s), and not in x0 , and that lims→+∞
x(s) = x0 + x∗ is exactly the minimizer. In other terms, the linear interpolation of
the points x0 and x0 + x∗ with parametrization λ(s) is the solution of the gradient
flow (21).
It is now time to turn back to the phase field setting.

3.3.2 A Family of “Intrinsic Norms” for the Phase Field Energy

As we have seen above, in order to recast minimization as a gradient flow, as a first

step it is necessary to single out an “intrinsic norm” which is nothing but the quadratic
part of the energy. In our setting, for the phase field variable we will employ

 
z2t,u = z 2 G c /2ε + t 2 W (Du) + εG c |∇z|2 d x,
Ω
 
z, ξ t,u = zξ G c /2ε + t 2 W (Du) + ε∇z · ∇ξ d x,
Ω

while for the displacement field u we will employ


|u|2t,z = t 2 (z 2 + ηε )W (Du) d x,
Ω

u, φt,z = t 2 (z 2 + ηε )σ (u) : ε(φ) d x.
Ω

With the above definitions the quadratic structure of the energy looks very clear,
indeed we can write the energy as
210 M. Negri

Fε (t, u, z) = 21 |u|2t,z + cz for cz = 21 G c (z − 1)2 /2ε + ε|∇z|2 d x,
Ω
⎧ 
⎪
⎨b(z) = G c z/2ε d x,
Fε (t, u, z) = 21 z2t,u − b(z) + ct,u for  Ω
⎪
⎩ct,u = 1 ηε t 2 W (Du) + G c /2ε d x.
2
Ω

Note that cz is independent of u and vice versa ct,u is independent of z, while b(·)
is linear. As a consequence, the partial derivatives as well take a particularly simple
form, being

∂u Fε (t, u, z)[φ] = u, φt,z , ∂z Fε (t, u, z)[ξ ] = z, ξ t,u − b(ξ ).

We will see in Sect. 3.4.3 how to write more explicitly the gradient flows originating
from alternate minimization.
Finally, it will be very convenient, if not necessary, to define a couple of slopes,
with respect to the “norms” defined above, that is

|∂u Fε (t, u, z)|− −

t,z = | min{∂u Fε (t, u, z)[φ] : φ ∈ Φ, |φ|t,z ≤ 1}| ,
|∂z Fε (t, u, z)|− −
t,u = | min{∂z Fε (t, u, z)[ξ ] : ξ ∈ Ξ, ξ t,z ≤ 1}| ,

where | · |− is once again the negative part.

3.4 A Parametrized “BV-Evolution”

Consider a sequence Δtn  0. For each Δtn , let u(tn,k ) and z(tn,k ) (for tk = kΔtn )
be given by the alternate minimization scheme (20). In order to define the limit as
Δtn  0, it is natural to introduce, for every time tk , an arc length interpolation
of the alternate minimizing path (u m , z m ) which links (u(tk−1 ), z(tk−1 )) = (u 0 , z 0 )
with (u(tk ), z(tk )) = limm→+∞ (u m , z m ). In the end, the interpolation provides a
parametrization of the evolution, i.e. a map (0, Sn )  s → (tn (s), u n (s), z n (s)) with
tn (0) = 0, u n (0) = u 0 , vn (0) = v0 and tn (S) = T . In this respect, it is technically
quite delicate to show that the lengths Sn of the interpolating curve are uniformly
finite with respect to Δtn (for the detail the reader should make reference to the
forthcoming [14]). We can then apply an abstract results developed in [22] which
yields the following Theorem.

Theorem 3 Given Δtn  0, let s → (tn (s), u n (s), z n (s)) be the parameteriza-
tions of the discrete evolutions provided by the alternate minimization scheme
(20). Then, up to subsequences, there exists a limit normalized parametrization
s → (t (s), u(s), z(s)) with t (s) ≥ 0, z (s) ≤ 0 and t (s) + |u (s)|t (s),z(s) + z (s)
t (s),u(s) ≤ 1. Moreover, for every s with t (s) > 0 the following equilibrium
conditions holds
Quasi-Static Evolutions in Brittle Fracture Generated … 211

|∂u Fε (t (s), u(s), z(s))|− −

t (s),z(s) = |∂z Fε (t (s), u(s), z(s))|t (s),z(s) = 0. (22)

Finally, for every s it holds the energy balance

 s
F (t (s), u(s), z(s)) = F (0, u 0 , z 0 ) + ∂t F (t (r ), u(r ), z(r )) t (r ) dr +
 s 0

− |∂u F (t (r ), u(r ), z(r ))|−

t (r ),z(r ) |u (r )|t (r ),z(r ) dr +
 s
0

− |∂v F (t (r ), u(r ), z(r ))|−

t (r ),u(r ) z (r )t (r ),u(r ) dr. (23)
0

In the next section we will explain better the meaning of the previous Theorem,
which is by itself quite technical. However, the analogy with Theorem 2 should be
quite evident.

3.4.1 Continuity Points: Equilibrium

Let us start considering the points where t (s) > 0, which correspond in the para-
metric setting to continuity points in time. Equation (22) gives equilibrium. Indeed,
by definition of the slopes

|∂u Fε (t (s), u(s), z(s))|−

t (s),z(s) = 0 ⇔ ∂u Eε (t (s), u(s), z(s))[φ] = 0 for φ ∈ Φ.
|∂z Fε (t (s), u(s), z(s))|−
t (s),u(s) = 0 ⇔ ∂z Fε (t (s), u(s), z(s))[ξ ] ≥ 0 for ξ ∈ Ξ.

Let us write more explicitly the equilibrium conditions. Introducing the phase field
stress σ z (u) = (z 2 + ηε ) σ (u) we get

∂u Eε (t (s), u(s), z(s))[φ] = σ z(s) (u(s)) : ε(φ) d x = 0, forφ ∈ Φ,
Ω

and thus ⎧
⎪
⎨div(σ z(s) (u(s))) = 0 Ω
u(s) = ±ê ∂D Ω (24)
⎪
⎩
σ z(s) (u(s)) n̂ = 0 ∂ N Ω,

which is the phase-field counterpart of (1).

Now, let us see discuss the physical meaning to the equilibrium condition with
respect to the phase field variable z. By the definition (17) of energy release the
equilibrium condition

∂z Fε (t (s), u(s), z(s))[ξ ] ≥ 0 for ξ ∈ Ξ

212 M. Negri

reads
− Gε (t (s), z(s))[ξ ] + ∂z Kε (z(s))[ξ ] ≥ 0 for ξ ∈ Ξ (25)

where 
∂z Kε (z(s))[ξ ] = G c (z − 1)ξ/4ε + ε∇z · ∇ξ d x .
Ω

Employing the normalized set Ξ̂z(s) and taking the supremum with respect to ξ̂ ∈
Ξ̂z(s) from (25) it follows
G ε (t (s), z(s)) ≤ G c , (26)

which plays the role of (4).

To conclude this section, we remark that by the separate quadratic structure of the
energy we have the following separate minimality property


u(s) ∈ argmin Eε (t (s), u, z(s)) : u ∈ U

z(s) ∈ argmin Fε (t (s), u(s), z) : z ∈ Z , z ≤ z(s) .

Note that z(s) is only a constrained minimizer and that in general it is not true that

(u(s), z(s)) ∈ argmin Fε (t (s), u, z) : u ∈ U , z ∈ Z , z ≤ z(s)

as it would be in an energetic evolution.

3.4.2 Continuity Points: Thermodynamic Consistency

In this section we will discuss a couple of thermodynamic issues: the first is simply the
energy balance (which will also lead to a KKT condition) while the second originates
from the relationship between irreversibility constraint and dissipated energy. In both
the cases we will assume that t (s) > 0 in a parametrization interval [s1 , s2 ] (so that
the corresponding evolution is continuous in time). By (19) and (22) we can rewrite
(23) as
 s2
Fε (t (s2 ), u(s2 ), z(s2 )) = Fε (t (s1 ), u(s1 ), z(s1 )) + Pεext (t (r ), u(r ), z(r )) t (r ) dr,
s1

which is the usual energy balance in parametrized integral form. Using the chain rule
the above energy identity reads: for every s ∈ (s1 , s2 )

Fε (t (s), u(s), z(s)) = ∂t Fε (t (s), u(s), z(s)) t (s) + ∂u Fε (t (s), u(s), z(s))[u (s)] +
+∂z Fε (t (s), u(s), z(s))[z (s)]
= Pεext (t (s), u(s), z(s)) t (s) .
Quasi-Static Evolutions in Brittle Fracture Generated … 213

Again by (19) and (22) it follows that

∂z Fε (t (s), u(s), z(s))[z (s)] = 0 .

Here, note that equilibrium gives ∂z Fε (t (s), u(s), z(s))[ξ ] ≥ 0 for every ξ ∈ Ξ .
Using the notation of (25) the above identity in KKT fashion becomes
 
Gε (t (s), z(s)) − ∂z Kε (z(s)) [z (s)] = 0 , (27)

which plays the role of (5).

Now, let us study the relationship between the irreversibility constraint and the
dissipated energy Kε (z). In general if z 1 ≤ z 2 it is not true that Kε (z 1 ) ≤ Kε (z 2 )!
This is simply due to the fact that

Kε (z) = 21 G c (z − 1)2 /2ε + ε|∇z|2 d x
Ω

includes a gradient term which is not monotone. For instance, consider z 1 ≤ z 2 with
z 1 constant and with z 2 (highly) oscillating. Then
 
(z 1 − 1)2 /4ε d x ≥ (z 2 − 1)2 /4ε d x
Ω Ω

while  
ε|∇z 1 |2 d x < ε|∇z 2 |2 d x.
Ω Ω

If z 1 ≈ z 2 but the energy of ∇z 2 is big enough it may be that Kε (z 1 ) < Kε (z 2 ). Thus,

the irreversibility constraint, given by the monotonicity of z, does not always match
with the monotonicity of the dissipated energy. However, this is not what happens
in the evolution, at least in those interval [s1 , s2 ] where t (s) > 0. Indeed, z (s) ∈ Ξ
and thus by (27)

∂z Eε (t (s), u(s), z(s))[z (s)] + ∂z Kε (z(s))[z (s)] = 0.

Note that

∂z Eε (t (s), u(s), z(s))[z (s)] = t 2 (s) z(s) z (s) W (Du(s)) d x ≤ 0,
Ω

because the only negative term is z (s). It follows that ∂z Kε (z(s))[z (s)] ≥ 0 and
thus the energy s → Kε (z(s)) is non-decreasing.
214 M. Negri

3.4.3 Discontinuity Points: Which Gradient Flow?

In this section we want to collect some properties of the evolution in the jumps, i.e.,
in the parametrization intervals [s1 , s2 ] where t (s) is constant, and thus t (s) = 0. It
is fair to say that at the current stage the picture is not fully clear and detailed. In
order to understand the main qualitative features it is not too restrictive to assume
that for s ∈ [s1 , s2 ] it holds

|∂u Fε (t (s), u(s), z(s))|− −

t (s), z(s) = |∂v Fε (t (s), u(s), z(s))|t (s), u(s) = 1 .

Under these assumptions, (23) implies (by the chain rule and convexity arguments)
that for a.e. s ∈ [s1 , s2 ] we have

u (s) ∈ argmin {∂u Fε (t (s), u(s), z(s))[φ] : φ ∈ Φ, |φ|t (s), z(s) = 1},

z (s) ∈ argmin {∂v Fε (t, u, v)[ξ ] : ξ ∈ Ξ, ξ t (s), u(s) = 1}.

In other terms, u and z are the steepest descent direction for Fε with respect to
the “intrinsic norms”. If the mathematical meaning is formally clear, the physical
behaviour is understood only for the gradient flow of u. Indeed (cf. [14]) on the jumps
the displacement field evolves like a “phase-field visco-elastic flow”
⎧
⎪
⎨div(σ z(s) (u(s) + u (s))) = 0 Ω
u(s) = ±ê ∂D Ω
⎪
⎩
σ z(s) (u(s) + u (s)) n̂ = 0 ∂ N Ω.

On the contrary, it seems not easy to provide a meaningful PDE for the evolution of
the phase field variable z since Ξ is not a space, but just a convex set.

4 Open Problems

In a broader perspective, the most interesting, and probably most difficult, open
problem is the convergence of the parametrized quasi-static BV -evolutions, say
s → (tε (s), u ε (s), z ε (s)), as ε → 0. On the base of Γ -convergence [2, 6] it is
expected a sharp crack evolution, possibly in the space SBD [4] or GSBD [8]. How-
ever, Γ -convergence has been crafted to study the convergence of energies and global
minimizers but it is not enough to provide convergence of equilibrium points and
slopes which are the main ingredients for gradient flows and BV -evolutions. In gen-
eral to have convergence of BV -evolutions it is necessary to have at least a sort of
Γ -liminf inequality for the slopes [22], as it is for gradient flows [25]. However, it
is not yet known any reasonable notion of slope in SBD spaces.
Quasi-Static Evolutions in Brittle Fracture Generated … 215

In this direction some partial results have been published: for instance [11] (in the
one dimensional setting) [21, 27] (with geometrical restriction on the crack) and [3]
(with a regularized energy).

References

1. Abdollahi, A., & Arias, I. (2012). Phase-field modeling of crack propagation in piezoelectric
and ferroelectric materials with different electromechanical crack conditions. Journal of the
Mechanics and Physics of Solids, 60(12), 2100–2126.
2. Ambrosio, L., & Tortorelli, V. M. (1990). Approximation of functionals depending on jumps
by elliptic functionals via Γ -convergence. Journal of the Mechanics and Physics of Solids,
43(8), 999–1036.
3. Babadjian, J. F. & Millot, V. (2014) Unilateral gradient flow of the Ambrosio-Tortorelli func-
tional by minimizing movements. Journals Annales de l’Institut Henri Poincar.é Annales:
Analyse Non Lineaire, 31(4), 779–822.
4. Bellettini, G., Coscia, A., & Dal Maso, G. (1998). Compactness and lower semicontinuity
properties in SBD(Ω). Mathematische Zeitschrift, 228(2), 337–351.
5. Bourdin, B., Francfort, G. A., & Marigo, J.-J. (2000). Numerical experiments in revisited brittle
fracture. Journal of the Mechanics and Physics of Solids, 48(4), 797–826.
6. Chambolle, A. (2003). A density result in two-dimensional linearized elasticity and applica-
tions. Archive for Rational Mechanics and Analysis, 167(3), 211–233.
7. Conti, S., Focardi, M. & Lurlano, F. Phase-field approximation of cohesive fracture energies.
8. Dal Maso, G. (2013). Generalised functions of bounded deformation. Journal of the European
Mathematical Society (JEMS), 15(5):1943–1997.
9. Efendiev, M. A., & Mielke, A. (2006). On the rate-independent limit of systems with dry
friction and small viscosity. Archive for Rational Mechanics and Analysis, 13(1), 151–167.
10. Francfort, G. A., & Marigo, J.-J. (1998). Revisiting brittle fracture as an energy minimization
problem. Journal of the Mechanics and Physics of Solids, 46(8), 1319–1342.
11. Francfort, G. A., Le, N. Q., & Serfaty, S. (2009). Critical points of Ambrosio-Tortorelli converge
to critical points of Mumford-Shah in the one-dimensional Dirichlet case. Archive for Rational
Mechanics and Analysis, 15(3), 576–598.
12. Griffith, A. A. (1920). The phenomena of rupture and flow in solids. Journal of the Mechanics
and Physics of Solids, 18, 163–198.
13. Hesch, C., & Weinberg, K. (2014). Thermodynamically consistent algorithms for a finite-
deformation phase-field approach to fracture. Journal of the Mechanics and Physics of Solids,
99(12), 906–924.
14. Knees, D. & Negri, M. Convergence of alternate minimization schemes for phase field fracture
and damage.
15. Knees, D., Rossi, R., & Zanini, C. (2013). A vanishing viscosity approach to a rate-independent
damage model. Journal of the Mechanics and Physics of Solids, 23(4), 565–616.
16. Larsen, C. J., Ortner, C., & Süli, E. (2010). Existence of solutions to a regularized model of
dynamic fracture. Journal of the Mechanics and Physics of Solids, 20(7), 1021–1048.
17. Miehe, C., Welschinger, F., & Hofacker, M. (2010). Thermodynamically consistent phase-field
models of fracture: Variational principles and multi-field FE implementations. Journal of the
Mechanics and Physics of Solids, 83(10), 1273–1311.
18. Mielke, A. (2011). Differential, energetic, and metric formulations for rate-independent
processes. In nonlinear PDE’s and applications, volume 2028 of lecture notes in mathematics
(pp. 87–170). Heidelberg: Springer.
19. Negri, M. (2010). A comparative analysis on variational models for quasi-static brittle crack
propagation. Journal of the Mechanics and Physics of Solids, 3, 149–212.
216 M. Negri

20. Negri, M. (2010). From rate-dependent to rate-independent brittle crack propagation. Journal
of the Mechanics and Physics of Solids, 98(2), 159–178.
21. Negri, M. (2013). From phase-field to sharp cracks: Convergence of quasi-static evolutions in
a special setting. Journal of the Mechanics and Physics of Solids, 26, 219–224.
22. Negri, M. (2014). Quasi-static rate-independent evolutions: Characterization, existence,
approximation and application to fracture mechanics. Journal of the Mechanics and Physics
of Solids, 20(4), 983–1008.
23. Negri, M., & Ortner, C. (2008). Quasi-static propagation of brittle fracture by Griffith’s criterion.
Journal of the Mechanics and Physics of Solids, 18(11), 1895–1925.
24. Pandolfi, A., & Ortiz, M. (2012). An eigenerosion approach to brittle fracture. Journal of the
Mechanics and Physics of Solids, 92(8), 694–714.
25. Sandier, E., & Serfaty, S. (2004). Gamma-convergence of gradient flows with applications to
Ginzburg-Landau. Journal of the Mechanics and Physics of Solids, 57(12), 1627–1672.
26. Schmidt, B., Fraternali, F., & Ortiz, M. (2008). Eigenfracture: An eigendeformation approach
to variational fracture. Journal of the Mechanics and Physics of Solids, 7(3), 1237–1266.
27. Sicsic, P., & Marigo, J. -J. (2013). From gradient damage laws to Griffith’s theory of crack
propagation. Journal of Elasticity, 113(1), 55–74.
Improving the Material-Point Method

Deborah Sulsky and Ming Gong

Abstract The material-point method (MPM) was introduced about 20 years ago and
is a versatile method for solving problems in continuum mechanics. The flexibility
of the method is achieved by combining two discretizations of the material. One is
a Lagrangian description based on representing the continuum by a set of material
points that are followed throughout the calculation. The second is a background grid
that is used to solve the continuum equations efficiently. In its original form, some
applications of the method appeared to be second order accurate while other tests
showed poor or no convergence. This paper provides a framework for analyzing the
errors in MPM. Moreover, the analysis suggests modifications to the algorithm to
improve accuracy. The analysis also points to connections between MPM and other
meshfree methods.

1 Introduction

The 1990s saw significant advances in the development of meshfree technologies

for computational mechanics. Methods such as the Element Free Galerkin Method
(EFG) [3, 4], Reproducing Kernel Particle Method (RKPM) [13, 14], h-p clouds
[8], meshless local Petrov-Galerkin (MLPG) [2] and the Partition of Unity Method
(PU) [15] rely on meshfree approximations of functions constructed from scattered
data. Two mainly equivalent approaches to function reconstruction are used in this
literature. The first is the moving least squares method (MLS) with polynomial basis
functions and the second is reproducing kernel particle methods (RKPM). In MLS a
functional is minimized to determine the particle shape functions whereas the guiding
principle in RKPM is to determine shape functions that exactly reproduce polyno-

D. Sulsky (B) · M. Gong

Department of Mathematics and Statistics, The University of New Mexico,
Albuquerque, NM 87131, USA
e-mail: [email protected]
M. Gong
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 217

K. Weinberg and A. Pandolfi (eds.), Innovative Numerical Approaches
for Multi-Field and Multi-Scale Problems, Lecture Notes in Applied
and Computational Mechanics 81, DOI 10.1007/978-3-319-39022-2_10
218 D. Sulsky and M. Gong

mials up to a given order. These methods most often use a weak formulation of the
governing equations resulting in an implementation that is similar in structure to finite
element methods but with shape functions developed from the scattered data points.
A continuing research topic is development of efficient and stable methods for the
integration of the weak form equations [7]. More recently, the Optimal Transporta-
tion Meshfree (OTM) [12] scheme was developed based on max-ent interpolation [1].
The equations of motion are formulated from optimal transportation theory which
discretizes the inertial action in space and time within a variational framework.
The material-point method (MPM) [20] was developed contemporaneously with
the meshfree technologies but its formulation was inspired by the particle-in-cell
method [6, 10]. Unlike general meshfree methods, MPM is restricted to solving
problems in continuum mechanics. The method discretizes the continuum based on
representing it by a set of Lagrangian material points that are followed throughout the
calculation. Like meshfree methods, functions are reconstructed from these scattered
data points. However, these functions are then evaluated to provide information for
a grid where continuum equations of motion are solved efficiently.
Equations of motion are solved in an updated Lagrangian frame on the computa-
tional grid, using standard finite difference or finite element methods. Advection is
modeled by moving the material points in the computed velocity field. Each numer-
ical material point carries its material properties without error while it is advected.
Since all the properties of the continuum are assigned to the numerical material
points, the information carried by these points is enough to characterize the flow
and the grid carries no permanent information. Accuracy and consistency of MPM
have been examined with fragmentary results in the literature, e.g. [18, 19]. The goal
of this paper is to establish a framework for studying the accuracy of MPM. The
accuracy of each step in the algorithm is examined as well as how the steps interact,
in order to make improvements.
For time-dependent problems, there are four steps in the algorithm: (i) choose a
convenient computational grid; (ii) map information from the material points to the
grid; (iii) solve the field equations on the grid; and (iv) update the material points
based on the grid solution. Similar accuracy should be maintained by the function
reconstruction in step (ii) and the grid-based solution method chosen to advance the
solution in step (iii). Finally, a scheme of consistent accuracy must be employed to
update the material points in step (iv) in order to maintain the overall accuracy of the
method. Our analysis brings to light connections between MPM and other meshfree
methods. Moreover, our analysis shows how to exploit the function reconstruction
techniques used in meshfree methods to improve MPM.

2 Preliminaries

Let Ω(0) ∈ Rd denote the reference configuration of a continuum body in d dimen-

sions, with material points labeled X . The set Ω(0) is assumed open and bounded,
with a smooth boundary Γ0 = ∂Ω(0). Assume the reference boundary Γ0 is parti-
Improving the Material-Point Method 219

 
tioned into disjoint subsets, Γ0 = Γu Γt and Γu Γt = ∅, such that displacement
is prescribed on Γu and traction is prescribed on Γt .
Let the spatial (current) configuration of the same body be Ω(t) ∈ Rd , with points
labeled x. Assume there exists a smooth mapping, the motion of the body, ϕ : Ω(0) ×
[0, T] → Rd , such that Ω(t) = ϕ(Ω(0), t) and x = ϕ(X, t), where [0, T] ⊂ R is
the time interval of interest. The current boundary is Γ (t) = ∂Ω(t) with Γ (t) =
Γt (t) ∪ Γu (t) and Γt (t) ∩ Γu (t) = ∅, where Γu (t) = ϕ(Γu , t) and Γt (t) = ϕ(Γt , t).
The deformation gradient of the motion is defined as

F(X, t) = Grad ϕ. (1)

Given a reference density ρ0 : Ω(0) → R+ , the spatial density is ρ = J −1 ρ0 , where

J = det F.
MPM solves continuum mechanics problems for a body occupying Ω(t) at time
t, of the form

∇ · σ + ρb = ρa in Ω(t) (2)
σ ·n =t on Γt (t) (3)
u=u on Γu (t) (4)

with initial conditions

v(x, 0) = v0 (x), x ∈ Ω(0)

σ (x, 0) = σ0 (x), x ∈ Ω(0).

In these equations, the mass density is ρ(x, t), the acceleration is a(x, t), the velocity
is v(x, t), the displacement is u(x, t), the specific body force is b(x, t) and the
Cauchy stress is σ (x, t). Boundary conditions can consist of applied traction, t, on
a portion of the boundary denoted by Γt (t), with n being the unit outward normal
to the boundary, and prescribed displacement, u, on a portion denoted by Γu (t). A
constitutive equation for the stress is required to complete the specification. For this
work, we will use a neo-Hookean constitutive model extended to the compressible
range,
 
J σ = λ(ln J )I + μ F F T − I , (5)

where μ and λ are the Lamé constants and I is the second order identity tensor.
The philosophy of MPM is to solve the equations of motion on a background
grid, but to keep track of trajectories of a set of material points, where the material
points represent the geometry of the body Ω(t) during deformation and carry the
solution. The process is accomplished by mapping information between the grid and
the material points. The details are given in the next section.
220 D. Sulsky and M. Gong

3 Main Steps of MPM

A body occupying Ω(0) is discretized into a finite set of N p regions, Ω p (0), with
a material point at the centroid, with position x p (0), p = 1, 2, . . . , N p . Each of
these points represents a volume of material, V p (0), the volume of Ω p (0), with a
mass, m p = Ω p (0) ρ0 (X ) d V . The material points move in time and the associated
volume moves with them. Mass conservation requires that m p be constant in time.
The aim is to determine the position, x p (t), the density, ρ p (t), the velocity, v p (t),
and the stress, σ p (t), at time t associated with these points given the initial values,
ρ p (0) = ρ0 (x p (0)), v p (0) = v0 (x p (0)) and σ p (0) = σ0 (x p (0)).

3.1 Function Reconstruction from Scattered Data

Mapping the information from the material points to the grid involves reconstructing
a function from scattered data on the material points, and evaluating the reconstructed
function on the background grid. That is, given data, u p , p = 1, 2, . . . , N p at material
points, we wish to determine basis functions ψ p[r ] (x) so that the reconstructed function
can be written in the form


Np
u R (x) = ψ p[r ] (x)u p . (6)
p=1

We base this function reconstruction on the idea of reproducing polynomials exactly

up to a specified degree r , where r is a non-negative integer [11, 16]. The notation
Pr = Pr (Ω) denotes the space of polynomials of degree less than or equal to r on
Ω. The dimension of Pr is Nr ,

r +d (r + d)!
Nr = = . (7)
d r !d!

The polynomial reproducing conditions are


Np
u(x) = ψ p[r ] (x)u(x p ) ∀u ∈ Pr . (8)
p=1

In order to express these conditions concisely, introduce a multi-index, α. A

multi-index is an ordered collection of non-negative integers, α = (α1 , α2 , . . . , αd ).
d
The length of α is defined as |α| = i=1 αi . We also define, for any point x ∈ Rd ,
α α1 α2 αd
the monomial, x = x1 x2 . . . xd . A sequence of points ξα ∈ R, indexed by the
multi-indices α with |α| ≤ r , is ordered with the terms in lexical order to form a
vector ξ ∈ R Nr . Lexical order is α = (0, 0, . . . , 0), (1, 0, . . . , 0), . . . , (0, 0, . . . , 1),
Improving the Material-Point Method 221

(2, 0, 0, . . . , 0), (1, 1, 0, . . . , 0), . . . , (0, 0, . . . , 0, r ). The basis functions in (6) are
written in terms of a weight function wh (x) with compact support defined by h, and
a local polynomial basis,

ψ p[r ] (x) = wh (x − x p ) (x − x p )α aα (x), p = 1, 2, . . . , N p . (9)
|α|≤r

The reproducing conditions (8) introduce Nr conditions for the coefficients aα ,


m α+β (x)aα (x) = δ|β|,0 , |β| ≤ r (10)
|α|≤r

where the moment functions are


Np
m α (x) = wh (x − x p )(x − x p )α . (11)
p=1

The above system of equations can be written in matrix form

M(x)a(x) = h(0) (12)

by introducing the moment matrix M(x),


Np
M(x) = wh (x − x p )h(x − x p )h(x − x p )T , (13)
p=1

where h(z) = (z α )|α|≤r ∈ R Nr . Then, the shape function is

ψ p[r ] (x) = wh (x − x p )h T (0)M −1 (x)h(x − x p ). (14)

The moment matrix is symmetric and positive semi-definite and the sum of rank one
matrices. A necessary condition for M(x) to be nonsingular is that for any x there
be at least Nr nonzero terms in the sum (13). The resulting shape functions ψ p[r ] (x)
form a partition of unity and if wh ∈ C k then ψ p[r ] ∈ C k , p = 1, 2, . . . , N p .
For this work, we restrict our attention to functions u(x) in a Hilbert space and
assume the boundary of the domain is smooth. Under appropriate hypotheses on the
distribution of material points so that the moment matrix is nonsingular, we have the
following error estimates for the function reconstruction [9]. For any u ∈ H m+1 (Ω)
and w ∈ C k

u − uR H l (Ω) ≤ ch min(m+1,r +1)−l |u| H min(m+1,r +1) (Ω) ∀l ≤ min(m + 1, r + 1, k).

(15)
222 D. Sulsky and M. Gong

If m ≥ r then the above reduces to

u − uR H l (Ω) ≤ ch r +1−l |u| H r +1 (Ω) ∀l ≤ min(r + 1, k); (16)

in particular, for l = 0

u − uR L 2 (Ω) ≤ ch r +1 |u| H r +1 (Ω) . (17)

3.1.1 Standard MPM

The reconstructed function in the standard MPM algorithm is a particular case of

Shepard function interpolation [17] which corresponds to r = 0. Shepard function
interpolation is used to construct the velocity v R (x) from N p discrete, scattered values
v p as
Np
v (x) =
R
ψ p[0] (x)v p , (18)
p=1

where, from (13) and (14)

wh (x − x p )
ψ p[0] (x) = Np
, (19)
p=1 w h (x − x p )

for some weight function, w. The standard MPM uses a linear hat function weighted
by the material-point mass for the weight function

wh (x − x p ) = m p s(x − x p ). (20)

In one dimension, the linear hat function, written in terms of natural coordinates, is
s = ŝ(ξ ),
1 − |ξ | |ξ | ≤ 1
ŝ(ξ ) = , (21)
0 otherwise

where ξ = (x p − x)/ h is the natural coordinate with h being the support of s and the
mesh spacing. In two space dimensions, the weight function is the tensor product of
one-dimensional weight functions, s(ξ, η) = ŝ(ξ ) ⊗ ŝ(η), with a similar construc-
tion in three dimensions.

3.1.2 Examples in One Dimension

Consider a function, f (x) : R → R, defined on an interval [a, b]. Construct Ne

elements of equal size h = (b − a)/Ne . The I th element is the interval [x I −1 , x I ]
where the nodes are given by x I = a + I h, I = 1, 2, . . . , Nn = Ne + 1. Within each
Improving the Material-Point Method 223

(a) 1.2 (b) 1.2

1 1

0.8 0.8
f (x)

f (x)
0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x

Fig. 1 Shepard interpolation using the functions f (x) = 1 and f (x) = x (shown with solid lines)
and a randomly-placed sampling points or b equally-spaced sampling points. The circles on the x-
axis show the positions of the sampling points. The other circles in the plots show the reconstructed
values f R (x I ). With equally-spaced points, the reconstructed function is exact for constant functions
and exact at the interior nodes for linear functions

element, sample the function f (x) at two randomly selected points. The sampling
points, x p , p = 1, 2, . . . , N p = 2Ne , play the role of material points. Set f p = f (x p ),
p = 1, 2, . . . , N p . Now reconstruct the function using the material-point data to
obtain f R (x). In MPM, we are interested in the value of the reconstructed function
at nodes, f R (x I ).
By construction, Shepard interpolation reproduces constant functions. If we set
f (x) = 1 and take a = 0, b = 1 and Ne = 8, Fig. 1a shows the exact function (solid
line) and the reconstructed values at nodes, x I , (circles). The circles along the x-axis
indicate the position of the randomly selected material points. The figure shows that
f (x) = 1 is reconstructed exactly. The figure also shows the reconstructed nodal
values for f (x) = x which are not exact. However, for equally-distributed sampling
points, the Shepard interpolation reproduces linear functions except on the boundary
of the domain. Figure 1b shows the reconstruction of f (x) = 1 and f (x) = x when
the material points are equally spaced with N pe = 2 material points per element
located at x p = x I −1 + (2 p e − 1)h/N pe , p e = 1, 2, . . . , N pe , I = 1, 2, . . . , Ne and
p = (I − 1)N pe + p e . For linear functions, f R (x I ) = x I exactly in the interior of
the domain but there is an O(h) error at the boundary points.
Figure 2 shows the same type of reconstruction except using the smooth func-
tion f (x) = sin π x on [0, 1], again with Ne = 8 and N pe = 2. For this example, the
reconstruction is not exact whether or not the sampling points are equally spaced
or randomly selected. However, the errors get smaller as h gets smaller, but the rate
of convergence differs. We examine the L 2 convergence rate as given by the error
estimate (17). For each h, we compute the L 2 error excluding the endpoints of the
domain by approximating the integral
x Ne
( f (x) − f R (x))1/2 d x (22)
x1
224 D. Sulsky and M. Gong

(a) 1 (b) 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
f (x)

f (x)
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x

(c) 0
unequally spaced
equally spaced
-2

-4
log 10(L2 error)

-6
1

-8

-10 2

-12

-14
-6 -5 -4 -3 -2 -1
log 10(h)

Fig. 2 Shepard interpolation using the function f (x) = sin π x (solid line) with a randomly placed
sampling points and b equally spaced sampling points. The circles on the x-axis show the positions
of the sampling points. The other circles in the plots show the reconstructed values f R (x I ). c Shows
the convergence rate with h using random and equally spaced sampling points

using the trapezoidal rule based on the nodal values of the integrand. With randomly-
selected sampling points, the convergence rate is O(h) as expected from (17). With
equally-spaced points the convergence rate is one order higher than expected, O(h 2 ).
Since there are O(h) errors at the end points, if the end point values are included in
the error estimate, the rate would be O(h 3/2 ) for one space dimension.

3.2 Solve the Momentum Equation on the Background Grid

In MPM, Eq. (2) is solved in weak form. In order to simplify the task, assume u = 0
on Γu (t) and Γt (t) = ∅. That is, displacement is prescribed over the entire boundary
and the domain boundary is fixed in time. We also assume a nice boundary, such as a
piecewise polygonal boundary with the background mesh conforming to the domain
boundary. Under these conditions, the weak form of the governing equation is
Improving the Material-Point Method 225


σ · ∇η − ρ(b − a) · η dv = 0 (23)
Ω(t)

for all admissible η. In particular, an admissible η is zero on the domain boundary.

Shape functions are used in the MPM to construct approximations for the finite
element method on the background grid. The background grid is subdivided into
elements, Ω e , e = 1, 2, . . . , Ne . The nodes of this mesh are x I (t), I = 1, . . . , Nn .
If each element Ω e has m nodes then we can refer to the nodes belonging to an
individual element with the notation x Ie (t), I = 1, . . . , m. The nodal displacements
are given by

Nn
u h (x, t) = u I (t)N I (x). (24)
I =1

The velocity is represented by

D h  n
d
N  n N
vh (x, t) = u (x, t) = u I (t)N I (x) = v I (t)N I (x). (25)
Dt I =1
dt I =1

Likewise, the acceleration is

D h  n
d
N  n N
a h (x, t) = v (x, t) = v I (t)N I (x) = a I (t)N I (x). (26)
Dt I =1
dt I =1

Over a time step, the grid is Lagrangian. Thus, over the step, the shape function
does not change with time. This fact is significant since a time derivative of the
shape function is not required in the formulas for velocity and acceleration. Also,
the natural coordinates for a material point in an element remain constant over the
Lagrangian time step.
The approximations, (28)–(30), along with a representation for an admissible,
smooth, virtual displacement η,


Nn
ηh (x) = η I N I (x), (27)
I =1

are used to obtain semi-discrete equations of motion. Substitute the representations

(24)–(27) into each term of equation (23) to obtain

Dvh Nn Nn

dv J (t)
ηh · ρ(x, t) dv = ηI · ρ(x, t)N I (x)N J (x) dv
Ω(t) Dt I =1 J =1 Ω(t) dt

Nn 
Nn
dv J (t)
= ηI · M I J (t) , (28)
I =1 J =1
dt
226 D. Sulsky and M. Gong


Nn
− ∇ηh : σ dv = − ηI · ∇ N I (x) · σ (x, t)dv (29)
Ω(t) I =1 Ω(t)


Nn
ηh · ρ(x, t)bdv = ηI · ρ(x, t)b(x, t)N I (x)dv. (30)
Ω(t) I =1 Ω(t)

In order to complete the spatial discretization a quadrature rule must be given to

evaluate the integrals in Eqs. (28)–(30). In the standard MPM, the material points are
used as quadrature points and the integrals become sums over material points. For
example, to discretize the inertial term, Eq. (28), the components of the consistent
mass matrix M I J are written as

M I J (t) = ρ(x, t)N I (x)N J (x)dv

Ω(t)


Np
∼ ρ(x p (t), t)N I (x p (t))N J (x p (t))V p (t)
(31)
p=1


Np
= m p N I (x p )N J (x p ),
p=1

where the material-point mass is related to the density and volume through the
equation m p = ρ(x p (t), t)V p (t).
The numerical simulations use the simpler lumped-mass matrix. This matrix is
diagonal with the diagonal entries obtained by summing over the corresponding row
of the consistent mass matrix and using the property NJ =1
n
N J (x) = 1,


Np
M I (t) = ρ(x, t)N I (x)dv ∼ m p N I (x p ). (32)
Ω(t) p=1

Equation (29) provides the nodal values of the internal forces

FI int (t) = − ∇ N I (x) · σ (x, t)dv

Ω(t)


Np (33)
∼− ∇ N I (x)|x=x p σ p (t)V p (t).
p=1

In the above equation, the notation for the stress at the material point position,
σ p (t) = σ (x p (t), t) has been introduced.
Improving the Material-Point Method 227

Finally, the nodal values of the external forces arise from the body forces, Eq. (30),

FI ext (t) = ρ(x, t)b(x, t)N I (x)dv

Ω(t)


Np (34)
∼ m p b(x p (t), t)N I (x p (t)).
p=1

The weak form of the momentum balance equates (28) to the sum of the forces
(29)–(30). Since the weak form must hold for arbitrary η I , except at nodes on the
boundary where the displacement is prescribed, we obtain the semi-discrete equation
for the nodal acceleration at unconstrained nodes


Nn
dv J (t)
M I J (t) = FI int (t) + FI ext (t). (35)
J =1
dt

With the lumped mass matrix this equation becomes

dv I (t)
M I (t) = FI int (t) + FI ext (t). (36)
dt
The momentum equation is solved for the acceleration at unconstrained nodes, which
is then integrated in time to obtain the corresponding velocity and displacement.

dv I (t)
= a h (x I , t)
dt (37)
du I (t)
= vh (x I , t).
dt
Nodes constrained by the displacement boundary conditions move according to those
prescribed constraints.

3.3 Update the Material-Point Information

Once the equation of motion on the grid is solved, we need to update the material-
point information. The material points move along with the flow within an element
as it deforms in a Lagrangian manner over the time step. The material-point velocity
and position are updated as
dv p (t)
= a h (x p , t)
dt (38)
d x p (t)
= v (x p , t).
h
dt
228 D. Sulsky and M. Gong

The deformation gradient is also updated according to

d F p (t)
= L p (t)F p (t), L p (t) = ∇vh (x, t)|x=x p (t) . (39)
dt
The density and volume are obtained from the algebraic update

V p (t) = det F p (t)V p (0), ρ p (t) = ρ0 (x p (t))/V p (t). (40)

A time-integration scheme is needed to solve the semi-discrete equations numerically.

The order of the time integration must match the order of the spatial discretization
to maintain overall accuracy of the method.

3.4 Generate a New Grid

After the material-point information is updated, a new grid is generated for the next
time step. Usually in MPM the grid points that have moved during the Lagrangian
step are moved back to their previous locations. Note, however that one can generate
any convenient grid instead.

4 Time Discretization

Although the theoretical framework above provides guidance on the construction of

a method of any order, the focus in this work will be to obtain second order accuracy
in space and time. A nominally second order, leapfrog scheme is used for the time
discretization. In this scheme the velocity and displacement updates are staggered
in time. A superscript is used to denote the time level in the discrete equations. Let
1 1 1 1
t n+ 2 = 21 (t n+1 + t n ), Δt n = t n+ 2 − t n− 2 , and Δt n+ 2 = t n+1 − t n . The acceleration
is obtained from Eq. (36) evaluated at time t n

M In a nI = FIint,n + FIext,n , (41)

where the mass comes from (32). The internal force comes from Eq. (33)


Np
FI int,n = − ∇ N I (x)|x=x np σ pn V pn , (42)
p=1
Improving the Material-Point Method 229

and the external force comes from Eq. (34)

ext,n

Np
FI = m p b(x np , t n )N I (x np ). (43)
p=1

A centered difference formula for the acceleration is

n+ 21 n− 21
vI − vI 1 n+ 21 n− 1
a nI = = (v I − v I 2 ). (44)
t n+ 21
−t n− 21 Δt n

This formula can be converted into an update for the velocity

n+ 21 n− 21
vI = vI + Δt n a nI . (45)

Similarly, the velocity can be obtained by differencing the displacement

n+ 21 u n+1 − u nI 1
vI = I
= 1 (u I
n+1
− u nI ). (46)
t n+1 −t n
Δt n+ 2
Likewise, this formula can be converted into an update for the displacement

1 n+ 21
u n+1
I = u nI + Δt n+ 2 v I . (47)

The material points are transported in the same fields as the nodes. Since a nI =
n+ 21 1
a h (x I , t n ) and v I = vh (x I , t n+ 2 ), the material points are updated using

n+ 21 n− 21 n− 21 
Nn
vp = vp + Δt n a h (x np , t n ) = v p + Δt n N I (x np )a nI ,
I =1

n+ 21 
Nn
n+ 21 (48)
n+ 21 h n+ 21 n+ 21
x n+1
p = x np + Δt v (x p ,t )= x np + Δt N I (x np )v I ,
I =1
u n+1
p = x n+1
p − x 0p ,

Information from the grid is also used to update the material-point stress state. In this
work, we will consider hyperelastic models that are specified through the deformation
gradient. The deformation gradient is updated using the chain rule

∂ x n+1 n
F pn+1 = F , (49)
∂xn p
230 D. Sulsky and M. Gong

and
∂ x n+1 1 n+ 1
= I + Δt n+ 2 L p 2 , (50)
∂x n

where
n+ 21 
Nn
n+ 21
Lp = ∇ N I (x np )v I . (51)
I =1

Once the deformation gradient is advanced in time, we can compute the stress, σ pn+1
at material points from the constitutive equation. Likewise the density is ρ n+1 p =
ρ0 (x 0p )/J pn+1 with J pn+1 = det(F pn+1 ). Similarly, the volume corresponding to the
material point is V pn+1 = J pn+1 V p (0).

5 Numerical Example

To keep the presentation simple while illustrating the ideas, a one-dimensional exam-
ple will be used to demonstrate the convergence properties of MPM. The example
is constructed using the method of manufactured solutions described in [18, 19] and
outlined in the next section.

5.1 Manufactured Solution

The standard MPM and its improvements will be examined through the solution of
a one-dimensional bar problem. We would like a problem with an exact solution so
that we may accurately compute the order of convergence of the method. An example
consisting of a longitudinally vibrating bar can be obtained by writing the problem
in the reference configuration. Consider the time dependent problem for a 1D, elastic
bar of length, L, with fixed ends, and constant cross-section. The equivalent equation
to (2) for the bar problem in the reference configuration is

∂ P(X, t) ∂ V (X, t)
+ ρ0 (X )B(X, t) = ρ0 (X ) , (52)
∂X ∂t
with boundary conditions

V (0, t) = V (L , t) = 0. (53)

The initial conditions are

V (X, 0) = v0 (X ), P(X, 0) = P0 (X ). (54)

Improving the Material-Point Method 231

In these equations V is the velocity, ρ0 is the original mass density, B is the body
force per unit mass, and P is the 1st Piola-Kirchoff stress. The relationship between
P and the Cauchy stress is

Jσ = P FT . (55)

For the neo-Hookean model

  
P(X, t) = λ(ln J )I + μ F F T − I F −T . (56)

For the manufactured solution, determine the body force so that the displacement,
U , and velocity, V , are given by

A
U (X, t) = sin π X sin Cπ t, V (X, t) = A sin π X cos Cπ t. (57)
Cπ
The corresponding initial conditions are

v0 (X ) = A sin π X, P0 (X ) = 0, F(X, 0) = 1, (58)

and the boundary conditions V (0, t) = V (L , t) = 0 are satisfied.

The deformation gradient is a diagonal matrix with diag(F) = (F11 , 1, 1), where
F11 = 1 + (A/C) cos π X sin Cπ t. Note that J = det(F) = F11 and
 
∂ P11 λ μ ∂ F11
= (1 − ln F11 ) + (F 2
+ 1)
∂X 2
F11 2
F11 11
∂X
  (59)
λ μ
= 2
(1 − ln F11 ) + 2
(F11
2
+ 1) (−π 2 U ).
F11 F11

Similarly, the time derivative of the velocity is related to the displacement

∂V
= −C 2 π 2 U. (60)
∂t
Therefore, to have a solution to equation (52), the body force must be
 
λ μ
ρ0 B = π 2 U 2
(1 − ln F11 ) + 2
(F 2
11 + 1) − ρ0 C 2
. (61)
F11 F11

This solution is valid for any choice of C 2 . We will use C 2 = E/ρ0 , where E is the
Young’s modulus. With ν being Poison’s ratio, we can set μ = E/(2(1 + ν)) and
λ = Eν/((1 + ν)(1 − 2ν)) in the usual way.
If we introduce dimensionless variables so that lengths are scaled by L, density
by ρ0 , velocity by C, and stresses by E, we need only consider a bar of unit length
with dimensionless density ρ0 = 1, dimensionless modulus, E = 1 and therefore
232 D. Sulsky and M. Gong

C = 1. The only free parameter is the dimensionless amplitude of the initial velocity
 = A/C which controls the amount of deformation in the bar.
The material-point solution is an approximation to the solution of the partial
n+ 1 1
differential equations where u np ∼ U (x 0p , t n ), v p 2 ∼ V (x 0p , t n+ 2 ), σ pn ∼ P(x 0p , t n ),
and F pn ∼ F(x 0p , t n ). In one space dimension, the two-norm of the error can be
computed using one-point quadrature to approximate the integral defining the norm.
In order to measure the error in mesh-based quantities, the exact material-point
quantities are projected to the grid. For example, the exact material-point velocity,
1
V (x 0p , t n+ 2 ), is used to reconstruct a velocity field which is then evaluated at grid
n+ 21
nodes and compared with the numerical solution, v I .

6 Convergence Rate of Standard MPM

Standard MPM with leapfrog time differencing is given by the equations in Sect. 4
with function reconstruction based on ψ [0] as given in Sect. 3.1.1. To test convergence,
the method is applied to solve the problem of Sect. 5.1. Since we wish to consider
smooth solutions, the choice of the dimensionless parameter  should be less than
one to avoid the formation of shocks. The last consideration in running the problem
is stability. A necessary condition for the stability of the leapfrog scheme is that the
Courant-Friedrichs-Levy (CFL) condition be satisfied. For this difference scheme,
the CFL stability requirement is

cΔt
max ≤ 1. (62)
h
The maximum is over all x at time t and c is √ the maximum√ wave speed.√ In the
one-dimensional case, the wave speed is c = E/ρ = C ρ0 /ρ = C J√. For the
manufactured
√ solution of Sect. 5.1, we can compute the maximum of J to be
1 + A/C. To insure stability for all cases, we choose Δt = 0.7h.
Figure 3a shows a plot of the logarithm of the L 2 norm of the error in various
computed quantities versus the logarithm of the mesh size h for standard MPM.
The error is the error in the computed solution to the problem from Sect. 5.1 solved
using E = 1, ν = 0, ρ0 = 1, L = 1 and  = 0.001π . The mesh sizes considered are
h = 2−(i+2) , i = 1, 2, 3, 4. There are initially two, equally-spaced, material points
per element. The computed quantities considered in the figure are σ p , ρ p , v p , and
u p —the stress, density, velocity and displacement stored at material points, and,
v I , the velocity stored on the grid. The simulations are run while the time satisfies
t ≤ tmax and tmax is set to a dimensionless time of 2. Thus, the simulations terminate at
a time tend such that tend ≤ tmax and tend + Δt > tmax . The error norms are computed
at tend for the density, stress and displacement, and at tend − Δt/2 for the velocity.
The observed convergence rate for the deformation gradient, density and stress is
1.8. The rate is the same for these quantities since the density and stress are algebraic
Improving the Material-Point Method 233

(a) (b)

Fig. 3 Convergence of various computed quantities using the standard MPM for the problem in
Sect. 5.1 with a  = 0.001π and b  = 0.1π

functions of the deformation gradient. The figure shows data for the density only. The
observed convergence rate for the displacement is 1.7. The convergence behavior for
the velocity is a bit erratic but a least squares fit gives an overall convergence rate of
2.4.
The calculation is repeated in Fig. 3b except with a larger value of  = 0.1π .
The results shown in the plot indicate that MPM fails to converge. The difference
in these outcomes is due to the change in the parameter Â. Note that the maximum
displacement of a material point is Â/π and the distance of a material point to the
boundary of an element is initially h/4. Thus, when the mesh size is smaller than
h = 4 Â/π , material points cross element boundaries during the numerical solution
procedure. In Fig. 3a, the mesh sizes used are such that no material points cross
cell boundaries and in Fig. 3b all of the mesh sizes used result in material points
crossing cell boundaries. These results emphasize the conditional consistency of the
standard MPM and may explain seemingly contradictory reports of convergence and
of non-convergence of the method in the literature.

7 Improving the Material-Point Method

In the standard MPM, there are two main steps that account for inaccuracies of the
algorithm. The primary error corresponds to the quadrature rule used in standard
MPM to approximate the nodal forces or nodal masses. The material points are used
as the quadrature points. Since the material points move, they can be located arbi-
trarily in the computational element, degrading accuracy. Indeed, for large enough
motion relative to the grid size, the method is inconsistent. The other source of error
corresponds to the mapping from the material-point information to the grid. As we
saw in Sect. 3.1, the standard algorithm uses Shepard interpolation which is first
order for arbitrarily placed material points. The improvements considered in this
section are intended to achieve consistent second order accuracy of the method.
234 D. Sulsky and M. Gong

7.1 Function Reconstruction

Since the function reconstruction outlined in Sect. 3.1 can, in principle, be used to
reconstruct functions to any order of accuracy from scattered data, we use this prop-
erty to obtain second order accuracy in MPM. The guiding principle is to regard the
material points as sampling points, and use the function reconstruction to approxi-
mate quantities as needed. For example, use the material-point velocity to reconstruct
the velocity field and evaluate the reconstructed velocity field at grid nodes to start
a time step. Similarly, reconstruct a stress field from the material-point values and
evaluate the reconstructed stress field at quadrature points when computing internal
forces.
To obtain the nodal velocity, standard MPM uses a shape function determined
from the general form given in Eq. (9) using r = 0 with the weight function wh
chosen to be the linear hat function given in Eq. (21), weighted by material-point
mass. With r = 0, in general, only constant functions are reproduced exactly. To
increase the order of accuracy, the improved method uses ψ p[1] (x) determined from
the same general form, but using a polynomial basis consisting of the constant and
linear monomials. There is also the choice of weight function; and to construct
ψ p[1] (x) we use the linear hat function, Eq. (21), without the mass weighting. The
reconstructed function is then continuous and the consistency error is O(h 2 ).
Smoother reconstructed functions can be achieved using higher order splines for
the weight function. For example, the cubic spline would give a C 2 reconstructed
function. The cubic spline expressed in natural coordinates is
⎧2
⎨ 3 − ξ 2 + 21 |ξ |3 |ξ | ≤ 1
ŝ3 (ξ ) = 43 − 2|ξ | + ξ 2 − 16 |ξ |3 1 ≤ |ξ | ≤ 2 . (63)
⎩
0 otherwise

To distinguish a shape function constructed with the cubic spline as weight from
ψ p[1] (x) which uses the linear hat function for the weight function, introduce the
notation C [1]
p (x) to denote the former. Note that both shape functions have the same
O(h ) consistency error since they are constructed to reproduce constant and linear
2

functions. In addition to being smoother, the wider support of the shape function
C [1]
p (x) is also useful if a situation arises where there is only one material point in
an element; since, in that case, the moment matrix for ψ p[1] (x) would be singular.
The linear hat function or the cubic spline are convenient weight functions for
reproducing nodal values since it is easy to determine which nodes are in the support
of the shape function associated with a material point. To determine function values
at the centers of elements (the quadrature points for one-point quadrature), it is
numerically simpler to use a quadratic spline rather than the linear hat function or
cubic spline since the quadratic spline does not require logic to determine which
element centers are in the support of the weight function of a given material point.
The shape function Q [1]p (x) used to reconstruct values at the quadrature points is the
Improving the Material-Point Method 235

same order as ψ p[1] (x) or C [1]

p as indicated by the superscript [1], but uses a different
weight function. The weight used to determine Q [1] p (x) is

⎧3
⎨ 4 − ξ2 |ξ | ≤ 21
ŝ2 (ξ ) = 2 ( 2 − |ξ |) 21 ≤ |ξ | ≤
1 3 2 3
. (64)
⎩ 2
0 otherwise

The quadratic and cubic spline weight functions render the reconstruction nonlocal
to the element and they require the use of a structured (logically rectangular) grid.

7.1.1 Examples in One Dimension

We again consider the function f (x) = sin π x on [0, 1]. Divide the domain into
Ne = 8 elements and use N pe = 2 randomly chosen material points per element as
sampling points. Figure 4a shows the reconstructed values f R (x I ) at nodes, using

(a) 1 (b) 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
f (x)

f (x)

0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x

(c) -4
nodal values
-5 element values

-6
log 10(L2 error)

-7
-8
-9
-10 2

-11
-12
-13
-7 -6 -5 -4 -3 -2
log 10(h)

Fig. 4 Function reconstruction using the functionf (x) = sin π x (solid line) with randomly chosen
sampling points. The circles on the x-axis show the positions of the sampling points. The other
circles in the plots show the reconstructed values a f R (x I ) using ψ p[1] (x) and b f R (xe ) using
Q [1]
p (x). c Shows the convergence rate with h for nodal and element values
236 D. Sulsky and M. Gong

ψ [1] (x) as the shape function. Similarly, Fig. 4b shows function values reconstructed
at element centers, using Q [1]
p (x) as the shape function. The rate of convergence for
both methods is shown in Fig. 4c. It is observed that the errors are nearly identical
for both methods and both have O(h 2 ) consistency errors.

7.2 Summary of the IMPM Algorithm

The main steps of IMPM algorithms are summarized as follows.

(1) Reconstruct velocity, stress and density from material-point data. The velocity
is needed at grid nodes

n− 21 
Np
n− 21
ψ p[1] (x I )v p
1
vI = v R (x I , t n− 2 ) = .
p=1

The density and the stress are needed at quadrature points for the elements in
order to compute the mass matrix and the internal forces, respectively. For the
one-dimensional problem, we use one-point quadrature with the quadrature point
located at the center of the element. The reconstructed values are


Np
ρen = ρ (xe , t ) =
R n
Q [1]
p (x e )ρ p ,
n

p=1


Np
σen = σ R (xe , t n ) = Q [1]
p (x e )σ p ,
n

p=1

where xe denotes the quadrature point.

Nodal forces and nodal masses are now assembled from the element contribu-
tions as in the standard finite element method. The internal and external forces
are

Ne
f Iint,n = − σ (x, t n )∇ N I (x) d x ∼ − ∇ N I (xe )σen h, (65)
Ω e=1


Ne
f Iext,n = ρ(x, t n )b(x, t n )N I (x) d x ∼ N I (xe )ρen b(xe , t n )h. (66)
Ω e=1

Similarly, the lumped nodal masses are

Improving the Material-Point Method 237


Ne
M In = ρ(x, t n )N I (x) d x ∼ ρen N I (xe )h. (67)
Ω e=1

(2) Solve the momentum equation on the finite element grid.

M In a nI = f Iint,n + f Iext,n (68)

n+ 21 n− 21
vI = vI + Δt n a nI . (69)

(3) Update the information on the material points. The update follows the same
equations as the standard method given in Eq. (48) for the velocity, position
and displacement. The density and deformation gradient are updated following
Eq. (49).
(4) Generate a new grid.

7.3 Convergence Rate of Improved MPM

In this section we repeat the analysis of Sect. 6 using the improved method. We restrict
attention to the larger deformation case where  = 0.1π . Nodal values are obtained
using ψ p[1] and element values are obtained using Q [1] p . Figure 5 shows the logarithm
of the L 2 error as a function of the logarithm of the mesh size for the material-point
quantities, stress, velocity and displacement, and also for the grid stress, velocity
and displacement. The mesh sizes used in this figure are h = 2−(i+3) , i = 1, 2, 3, 4.

Fig. 5 Convergence of
various computed quantities
using the improved MPM for
the problem in Sect. 5.1 with
 = 0.1π . The observed
rates of convergence are σe
(1.98), σ p (1.87), v I (2.02),
v p (1.98), u i (2.01), u p
(2.01)
238 D. Sulsky and M. Gong

The observed accuracy is second order. This figure should be compared with Fig. 3b
which shows failure of the standard method to converge for this large deformation
case.
Similar behavior is obtained using C [1]
p for computing nodal values from the material
point data.

8 Conclusions

This paper has presented a framework for improving the accuracy of MPM. The
material points are viewed as providing data for a finite element method. In this
finite element method, in essence, a new grid is used each step with the configuration
at the n-th time step being the reference configuration for the step. Thus, it is necessary
to initialize the values on the grid at the beginning of each step. The initialization
is accomplished by evaluating a function at nodes that is reconstructed from the
material-point data. This function reconstruction from scattered data appears in other
contexts, notably other meshfree methods for continuum mechanics. We therefore
draw on the literature to find methods that perform the reconstruction to any desired
accuracy. A potential advantage to using MPM over other meshfree methods that
rely on the same function reconstruction arises from the use of the finite element
mesh to take gradients and perform quadrature. In MPM the shape functions for
function reconstruction are not used as the basis functions for solving the equations
of motion, avoiding costly differentiation of these functions. A potential advantage
of MPM over using the finite element method on its own is in avoiding problematic
mesh distortion when materials undergo large deformations.
To illustrate the ideas for improving MPM in one space dimension, two-noded dis-
placement elements have been used with function reconstruction based on reproduc-
ing linear polynomials. B-splines have been used as the weight function in building
the shape functions used for the function reconstruction. We have also used four-
noded quadrilaterals in two space dimensions and will report the results elsewhere.
Clearly, higher-order element formulations need to be paired with higher-order func-
tion reconstruction for overall higher-order accuracy of the method. Also, if higher-
order accuracy is needed in time, then the leapfrog scheme considered in this paper
also needs to be replaced with a method of appropriate order.
For small deformations relative to the grid size and equally-spaced, or nearly
equally-spaced material points, it is possible to obtain apparent second-order accu-
racy with the standard MPM algorithm. We have demonstrated that second-order
accuracy is possible with large deformations using the improved method. However,
there is more work to be done to achieve second-order accuracy in general. The prin-
cipal remaining issue is related to the fact that MPM represents the geometry of a
body implicitly through the distribution of material points. This representation makes
the treatment of general boundary conditions difficult, especially if one is interested
in accuracy greater than first order. We also find apparent instabilities arise if we drive
the simulations to very small mesh sizes. The source of the instability is unknown
Improving the Material-Point Method 239

at this time. Two potential sources are ill-conditioning of the moment matrices or a
ringing instability [5].
We have shown a path to obtaining methods with specified accuracy. It is of interest
to provide a mathematical analysis of the methods to make precise the conditions for
convergence at a given rate. This topic, as well as a stability analysis of the methods,
remain topics of continued research.

Acknowledgments This work was partially supported by the National Science Foundation under
grant ARC 1023667 to the University of New Mexico.

References

1. Arroyo, M., & Ortiz, M. (2006). Local maximum-entropy approximation schemes: A seamless
bridge between finite elements and meshfree methods. International Journal for Numerical
Methods in Engineering, 65(13), 2167–2202.
2. Atluri, S. N., & Zhu, T. (1998). A new meshless local Petrov-Galerkin (MLPG) approach in
computational mechanics. Computational Mechanics, 22, 117–127.
3. Beltschko, T., Lu, Y. Y., & Gu, L. (1994). Element-free Galerkin methods. International Journal
of Numerical Methods in Engineering, 37, 229–256.
4. Belytschko, T., Krongauz, Y., Organ, D., Fleming, M., & Krysl, P. (1996). Meshless meth-
ods: An overview and recent developments. Computer Methods in Applied Mechanics and
Engineering, 139, 3–47.
5. Brackbill, J. U. (1988). The ringing instability in particle-in-cell calculations of low-speed flow.
Journal of Computational Physics, 75, 469–492.
6. Brackbill, J. U., & Ruppel, H. M. (1986). FLIP: A method for adaptively zoned, particle-in-cell
calculations of fluid flows in two dimensions. Journal of Computational Physics, 65, 314–343.
7. Duan, Q., Gao, X., Wang, B., Li, X., Zhang, H., Belytschko, T., et al. (2014). Consistent
element-free Galerkin method. International Journal for Numerical Methods in Engineering,
99, 79–101.
8. Duarte, A. C., & Oden, J. T. (1996). H-p clouds–an h-p meshless method. Numerical Methods
for Partial Differential Equations, 12(6), 673–705.
9. Han, W., & Meng, X. (2001). Error analysis of the reproducing kernel particle method. Com-
puter Methods in Applied Mechanics and Engineering, 190, 6157–6181.
10. Harlow, F. H. (1957). Hydrodynamic problems involving large fluid distortions. Journal of the
Association for Computing Machinery, 4, 137.
11. Li, S., & Liu, W. K. (2002). Meshfree and particle methods and their applications. Applied
Mechanics Reviews, 55(1), 1–34.
12. Li, B., Habbal, F., & Ortiz, M. (2010). Optimal transportation meshfree approximation schemes
for fluid and plastic flows. International Journal for Numerical Methods in Engineering, 83,
1541–1579.
13. Liu, W., Jun, S., Li, S., Adee, J., & Beltschko, T. (1995). Reproducing kernel particle methods
for structural dynamics. International Journal of Numerical Methods in Engineering, 38, 1655–
1680.
14. Liu, W. K., Jun, S., & Zhang, Y. (1995). Reproducing kernel particle methods. International
Journal for Numerical Methods in Engineering, 20, 1081–1106.
15. Melenk, J. M., & Babuska, I. (1996). Partition of unity finite element method. Computer
Methods in Applied Mechanics and Engineering, 139, 314–389.
16. Nguyen, V. P., Rabczuk, T., Bordas, S., & Duflot, M. (2008). Meshless methods: A review and
computer implementation aspects. Mathematics and Computers in Simulation, 79(3), 793.
240 D. Sulsky and M. Gong

17. Shepard, D. (1968). A two-dimensional interpolation function for irregularly-spaced data. In

Proceedings—1968 ACM National Conference (pp. 517–524).
18. Steffen, M., Kirby, R. M., & Berzins, M. (2008). Analysis and reduction of quadrature errors
in the material point method (MPM). International Journal for Numerical Methods in Engi-
neering, 76(6), 922–948.
19. Steffen, M., Kirby, R. M., & Berzins, M. (2010). Decoupling and balancing of space and time
errors in the material point method (MPM). International Journal for Numerical Methods in
Engineering, 82, 1207–1243.
20. Sulsky, D., Chen, Z., & Schreyer, H. L. (1995). Application of a particle-in-cell method to solid
mechanics. Computer Physics Communications, 87, 236–252.
Meshfree Methods Applied to Consolidation
Problems in Saturated Soils

Pedro Navas, Susana López-Querol, Rena C. Yu and Bo Li

Abstract A meshfree numerical model, based on the principle of Local Maximum

Entropy, with a B-Bar based algorithm to avoid instabilities, is applied to solve con-
solidation problems in saturated soils. This numerical scheme has been previously
validated for purely elasticity problems without water (mono phase), as well as for
steady seepage in elastic porous media. Hereinafter, the model is validated for well
known consolidation theoretical problems, both static and dynamic, with known ana-
lytical solutions. For several examples, the solutions obtained with the new code are
compared to PLAXIS (commercial software). Finally, after validated, solutions for
dynamic radial consolidation and sinks, which have not been found in the litera-
ture, are presented as a novelty. This new numerical approach is demonstrated to be
feasible for this kind of problems in porous media.

1 Introduction

The settlement of saturated soils under loading is caused by a gradual interchange

between pore pressure and effective stress. Immediately after external loadings are
applied to a saturated soil domain, all the external pressure transfers to water, and

P. Navas · R.C. Yu (B)

School of Civil Engineering, University of Castilla La-Mancha, Avda. Camilo
Jose Cela s/n, 13071 Ciudad Real, Spain
e-mail: [email protected]
R.C. Yu
e-mail: [email protected]
S. López-Querol
Department of Civil, Environmental and Geomatic Engineering, University
College London, Gower Street, London WC1E 6BT, UK
e-mail: [email protected]
B. Li
Department of Mechanical and Aerospace Engineering, Case Western
Reserve University, Cleveland, OH 44106, USA
e-mail: [email protected]
© Springer International Publishing Switzerland 2016 241
K. Weinberg and A. Pandolfi (eds.), Innovative Numerical Approaches
for Multi-Field and Multi-Scale Problems, Lecture Notes in Applied
and Computational Mechanics 81, DOI 10.1007/978-3-319-39022-2_11
242 P. Navas et al.

some time is required for drainage to take place. When this drainage (i.e. dissipa-
tion of excess pore pressures) is complete, the solid phases totally takes the external
pressure. This process is known as soil consolidation [1]. The implementation of
the Biot’s equations [2] is a well-known way to solve problems in porous media
from a macro-scale point of view. The advantage of this method is the possibility of
accounting for coupling between the fluid phase and the solid skeleton. The u − pw
formulations, where u denotes the solid phase displacement, and pw is the pore fluid
pressure [3], have traditionally been employed for simulating coupled problems in
saturated porous media since the final equations work with less degrees of freedom
(three in 2D, four in 3D problems) compared with that of a complete formulation. The
recent u − w formulation, where w represents the relative fluid displacement with
respect to the solid phase, which is usually referred as the displacement-based or com-
plete formulation, has been employed in several numerical schemes (López-Querol
et al. [4], and recently adopted by Cividini and Gioda [5]). Such a methodology is
assumed in this work, first for its simplicity in imposing impervious boundary con-
ditions compared to the u − pw approaches; second, as the free surface comes out
naturally as the zero-pressure contour, no detection algorithm is necessary; third, it
facilities the modelling of large and/or nonlinear deformations of the solid phase as
well as the possible separation between the solid and fluid phases in the case of local
failure (liquefaction or slope instability). Since meshfree numerical schemes have
been known to perform particularly well in the regime of large deformations, we
endeavour to apply such schemes to coupled problems in saturated porous media,
using the u − w formulation.
There are many different flavours of meshfree methods available. The present
research has been carried out using the principle of maximum entropy [6], the shape
functions developed by Arroyo and Ortiz [7], in particular, the OTM framework [8],
for its numerous advantages in comparison with its alternatives. For example, the
exact mass transport, the satisfaction of the continuity equation, exact linear and
angular momentum conservation in order to solve different problems as spurious
modes, tensile instabilities and unknown convergence or stability properties. Since
the deformation and velocity fields are interpolated from nodal values using max-
ent shape functions, the Kronecker-delta property at the boundary makes it possible
for the direct imposition of essential boundary conditions. In the current work, an
Eulerian framework is employed to solve the Biot’s equations for porous media. In
addition, the parameters pertinent to the local maximum entropy are obtained effi-
ciently and independent of the support size through the Nelder-Mead algorithm [9].
Locking in near-incompressible materials is not unusual for the numerical meth-
ods based on either finite elements approaches [10–17] or meshfree approximation
schemes [18]. In the case of flow through saturated media, in a displacement formu-
lation, since both the undrained soil phase and the fluid phase are nearly incompress-
ible, locking may also occur. The recent approach developed by Ortiz and Suku-
mar [18], avoids locking by averaging the volumetric part of the strain tensor with
the value of the pressure. However, since the pressure term is part of the constitutive
model employed, the constitutive model is necessarily modified. As an alternative,
we implement a volumetric-strain average instead of a pressure average approach.
Meshfree Methods Applied to Consolidation Problems … 243

The proposed algorithm is independent of the constitutive model employed and it is

generically applicable to solve other locking problems. The idea behind is inspired by
that of Hughes [10] and the posterior developments of the B-Bar method [16]. The
specific strategy is analogous to that of the diamond elements of Hauret, Kuhl and
Ortiz [15] and it is the first B-Bar implementation for a pure displacement approach
within the framework of meshfree methods. It is also straightforward to be extended
for finite deformations and nonlinear applications. Extension of the formulation by
the authors in [19] to axisymmetric framework is presented in the current work.
The rest of the paper is organised as follows. The mathematical framework, includ-
ing the B-Bar based algorithm is presented next. Applications to various consolida-
tion problems are illustrated in Sect. 3. Relevant conclusions are drawn in Sect. 4.

2 Mathematical Framework

In this section, we first summarise the governing equations for unconfined seep-
age problems, in particular, the Biot’s equations, formulated in a u − w framework,
which have been successfully utilised by López-Querol et al. [20, 21], Cividini and
Gioda [5]; next, the B-bar implementation in axisymmetric framework for elastic
and porous media are given in detail.

2.1 The Biot’s Equations: A u − w Formulation

The Biot’s equations [22] are based on formulating the mechanical behaviour of a
solid-fluid mixture, the coupling between different phases, and the continuity of flux
through a differential domain of saturated porous media. For clarity, bold symbols
for notation of vectors and matrices, and regular letters to denote scalar variables, are
used. Let ρ and ρf respectively represent the mixture and fluid phase densities; b and
κ stand for the external acceleration vector and the permeability coefficient in [m3 ·
s/kg] (in civil engineering, however, the notion of hydraulic conductivity, k = κρf g, in
[m/s], is often used instead), the three equations of Biot, which represent the mixture
equilibrium, the fluid phase equilibrium and the continuity equation respectively, are
expressed as follows:

ST dσ − ρd ü − ρf d ẅ + ρdb = 0, (1)
ρf
−∇dpw − κ −1 d ẇ − ρf d ü − d ẅ + ρf db = 0, (2)
n
d ṗw
∇ · d ẇ + m · d ε̇ s + = 0, (3)
Q
244 P. Navas et al.

where S is a differential operator, u is the displacement vector of the solid skeleton

and w is the relative displacement of the fluid phase with respect to the solid one.
By denoting U as the absolute displacement of the fluid phase, w is related with U
through the soil porosity, n, as follows:

w = n(U − u). (4)

Additionally in Eq. (3), m represents the unit matrix expressed in Voigt form, whereas
Q stands for the mixture compressibility, which is calculated as
 −1
Q = Ks−1 (1 − n) + nKf−1 , (5)

where Ks and Kf are the bulk modulus of the solid grains and the compressive modulus
of the fluid phase.
A 2D approach is considered in the derivations presented hereinafter, therefore
the differential operator, S, and the unit matrix m are written as
⎛∂ ⎞ ⎛ ⎞
∂x
0 1
∂
S = ⎝ 0 ∂y ⎠ , m = ⎝1⎠ . (6)
∂ ∂
∂y ∂x
0

Assuming tensile stresses (except pore pressure pw , which is positive for com-
pression) and strains as positive, the Terzaghi’s effective stress [23] is defined as
follows
σ = σ  − pw m, (7)

where σ  and σ are the respective vectorial form in Voigt notation for the effective
and total stress tensor. For the case of linear elasticity, the incremental relationship
between stresses and strains is governed by:

dσ  = De dε s , (8)

where De denotes the elastic tensor, which under plane strain conditions, it is given
by: ⎛ ⎞
1−ν ν 0
λ
De = ⎝ ν 1 − ν 0 ⎠ (9)
ν 0 0 1−2ν2

where ν is the Poisson’s ratio, λ, the first constant of Lamé.

Rearranging the above equations, Eq. (1) can be re-written as

ST De Sdu − ∇dpw − ρd ü − ρf d ẅ + ρdb = 0. (10)

Meshfree Methods Applied to Consolidation Problems … 245

In the u − w approach, also known as the complete formulation (no additional

assumption is required under plane strain conditions), each node has four degrees of
freedom, u and w (two components each in 2D problems) and the scalar pw , the pore
pressure, is obtained afterwards. By comparison, in the traditional u − pw formula-
tion, each node has only three degrees of freedom in 2D, but results in complications
in imposing impervious boundary conditions.
Integrating Eq. (3) in time, and substituting dpw in Eqs. (10) and (2), it yields:
   
ST De Sdu + Q∇ ∇ T du + Q∇ ∇ T dw − ρd ü − ρf d ẅ + ρdb = 0, (11)
    ρf
Q∇ ∇ T du + Q∇ ∇ T dw − k −1 d ẇ − ρf d ü − d ẅ + ρf db = 0. (12)
n

Note that an isotropic medium is assumed in the above equations. The final system
of equations, once the elementary matrices have been assembled, can be expressed
as:
Kdu + Cd u̇ + Md ü = df , (13)

where K, C and M respectively denote stiffness, damping and mass matrices, du

represents the vector of unknowns (containing both the solid phase and fluid dis-
placements, u and w), expressed incrementally, and df is the increment of the exter-
nal forces vector, containing gravity acceleration, as well as boundary conditions for
nodal forces.

2.2 B-Bar Formulation in Elastic Axisymmetric Problems

In axisymmetric problems, x direction is changed by r, y changes to z. Due to this

fact, the shape function based on the principle of Local Maximum Entropy is similar
to that of the 2D case. Consequently, the new displacement vector is calculated with
the following equation:
⎡ h
⎤
ur1
 
h ⎥
⎢uz1
⎢ ⎥
ur N1 0 N2 0 · · · ⎢uh ⎥
= ⎢ r2 ⎥ , (14)
uz 0 N1 0 N2 · · · ⎢uh ⎥
⎣ z2 ⎦
..
.

where the superscript h denotes discrete nodal values. In an axisymmetric problem,

a different ε matrix is obtained according to [24]:
246 P. Navas et al.
⎡ ⎤ ⎡∂ ⎤
εr ∂r
0 
⎢ εz ⎥ ⎢ 0 ∂ ⎥ ur
⎢ ⎥ = ⎢ 1 ∂z ⎥
⎣ εθ ⎦ ⎣ 0 ⎦ uz . (15)
r
γrz ∂ ∂
∂z ∂r

Voigt notation is assumed in order to obtain the final B-bar matrix. The process to
obtain B matrix in index notation is:

εl = Slj uj = Slj Njk ukh = Blk ukh .

Thereby, the B matrix is the following one:

⎡ ∂N1 ∂N2 ⎤
∂r
0 ∂r
0
⎢ 0 ∂N1
0 ∂N 2
···⎥
B=⎢
⎣ N1
∂z
N2
∂z ⎥.
⎦ (16)
r
0 r
0
∂N1 ∂N1 ∂N2 ∂N2
∂z ∂r ∂z ∂r

If σ is required, the following equation should be employed:

⎡⎤
σr
⎢ σz ⎥
σ =⎢ ⎥
⎣ σθ ⎦ = Dε, (17)
τrz

where ⎡ ⎤
1−ν ν ν 0
λ⎢ ν 1 − ν ν 0 ⎥
D= ⎢⎣
⎥. (18)
ν ν ν 1 − ν 0 ⎦
1−2ν
0 0 0 2

Stiffness matrix is calculated by taking into account that the volume integral is
extended around the whole ring of material as follows:

Kp = 2π BT DB r dr dz, (19)

where the superscript p represents the fact that the matrix is calculated for each
material point within a patch. The final expression is written as:
T
Kp = 2π B DB r A, (20)

where A is the associated area of the material point.

Meshfree Methods Applied to Consolidation Problems … 247

The external forces in Eq. (13) are calculated in the same way:

2π rfr
f= , (21)
2π rfz

where fr and fz respectively denote radial and vertical components of the external
force.
If a B-Bar based algorithm is implemented in this problem, the starting point is
similar to the plane strain one, which is based on the transformation of the ε tensor
to the ε tensor, a tensor where the volumetric part is obtained as an average of the
neighbour integration points, as we can see in the following equation:

1 1 p
ε=ε− tr(ε)I + [tr(ε)] I, (22)
d d
p
where d is the dimension of the problem, in this case 3; and [tr(ε)] is the average
trace of ε of the neighbour integration points in a chosen patch, calculated by:

p 
Nb
[tr(ε)] = tr[ε(i) ] wi . (23)
i=1

In addition, the trace could be obtained from the strain vector as:
⎡ ⎤
εr
  ⎢ εz ⎥
εx + εy = 1 1 1 0 ⎢ ⎥
⎣ εθ ⎦ = εkk = mk εk . (24)
γrz

Rearranging the above equation using the B-matrix,

εll = ml εl = ml Blk ukh = Tk ukh , (25)

where  ∂N1 
N1 ∂N1 ∂N2 N2 ∂N2
T= ∂r
+ r ∂z ∂r
+ r ∂z
··· . (26)

Thus, the final lth-component of the tensor ε (in Voigt form) for a single integration
point i is calculated in the same way as in 2D problems:
⎡ ⎛ ⎞⎤
1 Nb
(j)
ε (i)
l = ⎣Blk(i) − ml ⎝Tk(i) − [Tk w(j) ]⎠⎦ ukh
d j=1
(i)
= Blk ukh .
248 P. Navas et al.

2.3 B-Bar Implementation in u − w Axisymmetric Problems

In order to apply the B-Bar method in a multiphase problem, we need to define a

constitutive matrix first to relate stress with strain or displacements of the different
phases. The proposed problem is the u − w problem with soil and water phases. The
effective stress tensor is calculated the same way as Eq. (7), if linear elasticity is
assumed, the relationship between stresses and strains, expressed in its incremental
form, is governed by Eq. (8), and pw is obtained with the third Biot’s equation:

d ṗw
∇ · d ẇ + mT d ε˙s + =0
Q

d ṗw = −Q [∇ · d ẇ + mT d ε˙s ]

The final stress equation will be:

σ = De ε s + Q [tr(εs ) + tr(ε w )] I.

If: ⎤ ⎡
εrs
⎢ s⎥
⎡ s ⎤ ⎢ εsz ⎥
ε ⎢ε ⎥
⎢ θs ⎥
ε=⎣ ⎦=⎢ ⎢γrz
⎥=Su
⎥
εw ⎢ εw ⎥
⎢ rw ⎥
⎣ εz ⎦
εθw

where S is the derivative matrix operator and u is a vector of displacements of both

phases:
⎡ s⎤ ⎡ ∂ ⎤
εr ∂r
0 0 0
⎢ εs ⎥ ⎢ 0 ∂z ∂
0 0⎥ ⎡ ⎤
⎢ sz ⎥ ⎢ 1 ⎥
⎢ ε ⎥ ⎢ 0 0 0 ⎥ ur
⎢ θs ⎥ ⎢ ∂r ∂ ⎥ ⎢ uz ⎥
⎢γ ⎥ = ⎢ 0 0⎥ ⎢ ⎥
⎢ rz ⎥ ⎢ ∂z ∂r ⎥⎣ ⎦.
⎢ εw ⎥ ⎢ 0 0 ∂ 0 ⎥ wr
⎢ rw ⎥ ⎢ ∂r ⎥
⎣ εz ⎦ ⎣ 0 0 0 ∂ ⎦ wz
∂z
εθw 0 0 1r 0

In addition, the summation of traces of strain could be done with a m∗ operator:

Meshfree Methods Applied to Consolidation Problems … 249
⎤ ⎡
εrs
⎢ εs ⎥
⎢ sz ⎥
⎢ ⎥
  ⎢ εθs ⎥
tr(ε ) + tr(ε ) = m ε = 1 1 1 0 1 1 1 ⎢
s w T
⎢γrz
⎥.
⎥
⎢ εw ⎥
⎢ rw ⎥
⎣ εz ⎦
εθw

Thus, in Voigt notation:

σ = De∗ ε + Q mT ε m = (De∗ + Q mT m) ε = Du−w ε

⎡ λ(1−ν) ⎤⎡ s⎤
ν
+Q λ+Q λ+Q 0 Q Q Q εr
⎢ λ + Q λ(1−ν) + Q λ + Q 0 Q Q Q⎥ ⎢ εs ⎥
⎢ ν ⎥ ⎢ sz ⎥
⎢ λ+Q λ + Q λ(1−ν) + ⎥ ⎢ε ⎥
⎢ ν
Q 0 Q Q Q ⎥ ⎢ θs ⎥
=⎢⎢ 0 0 0 μ 0 0 0⎥ ⎢ ⎥.
⎥ ⎢γrz ⎥
⎢ 0 Q Q Q⎥ ⎢
⎥ w⎥
⎢ Q Q Q ⎢ εrw ⎥
⎣ Q Q Q 0 Q Q Q⎦ ⎣ εz ⎦
Q Q Q 0 QQQ εθw

If the problem needs a B-Bar based algorithm, it is necessary to calculate the average
value of ε. Thus, the main equation yields:

1 1 1 1
ε=ε− tr(ε s )I + [tr(ε s )]p I − tr(ε w )I + [tr(ε w )]p I.
d d d d

In Voigt notation the equation, the lth-component of the strain tensor is:
⎛ ⎞
1⎝ Nb
(j)
Nb
[εkk w(j) ] mls − εkk [εkk w(j) ] mlw ⎠ ,
w(j)
εl = εl + −εkk mls + w
mlw +
d j=1 j=1

where
⎡ ⎤ ⎡ s⎤
εrs εr
⎢ εs ⎥ ⎢ εs ⎥
⎢ sz ⎥ ⎢ sz ⎥
⎢ε ⎥ ⎢ε ⎥
  ⎢ θ⎥   ⎢ θs ⎥
εkk
s
= mks εk = 1 1 1 0 0 0 0 ⎢ γ
⎢ rz
s⎥
⎥ , ε w
kk = m w
k kε = 0 0 0 0 1 1 1 ⎢γ ⎥ .
⎢ rz ⎥
⎢ εw ⎥ ⎢ εw ⎥
⎢ rw ⎥ ⎢ rw ⎥
⎣ εz ⎦ ⎣ εz ⎦
εθw εθw

Additionally, we know that the strain tensor in Voigt notation is:

εl = Slj uj = Slj Njk ukh = Blk ukh

250 P. Navas et al.

where, in this case, yields:

⎡ (1) ⎤
⎡ ⎤ ⎡ ∂N ⎤ ur(1)
∂N2 ⎢ uz ⎥
εr 0 0 0 0 0 0 ⎢ ⎥
1
∂r ∂r
⎢ εz ⎥ ⎢ ∂N1 ∂N2 ⎥ ⎢wr(1) ⎥
⎢ ⎥ ⎢ ⎥ ⎢ (1) ⎥
0 ∂z
0 0 0 ∂z
0 0
⎢ εθ ⎥ ⎢ ⎢
N1
0 0 0 N2
0 0 0 ⎥ ⎢wz ⎥
⎥ ⎢ (2) ⎥
⎢ ⎥ ⎢ ∂Nr r
⎥⎢ ⎥
⎢γrz ⎥ = ⎢ 1 ∂N1
0 0 ∂N2 ∂N2
0 0 · · · ⎥ ⎢ ur ⎥ .
⎢ w ⎥ ⎢ ∂z ∂r ∂z ∂r ⎥ ⎢ uz(2) ⎥
⎢ε ⎥ ⎢ 0 0 ∂N1
0 0 0 ∂N2
0 ⎥ ⎢ (2) ⎥
⎢ rw ⎥ ⎢ ∂r ∂r ⎥ ⎢wr ⎥
⎣ εz ⎦ ⎣ 0 0 0 ∂N1
0 0 0 ∂N2
⎦ ⎢ (2) ⎥
∂z ∂z ⎢wz ⎥
εθw 0 0 N1
0 0 0 N2
0 ⎣ ⎦
r r ..
.

In order to calculate εll , the above equation can be rearranged as follows:

εlls = mls εl = ml Blk ukh = Tks ukh ,

εllw = mlw εl = ml Blk ukh = Tkw ukh

where  ∂N1 
N1 ∂N1 ∂N2 N2 ∂N2
Ts = ∂r
+ r ∂z
00 ∂r
+ r ∂z
0 0 ··· ,
 ∂N1 N1 ∂N1 ∂N2 N2 ∂N2

Tw = 0 0 ∂r
+ r ∂z
00 ∂r
+ r ∂z
··· .

Thus, the final lth-component for the new strain tensor ε at a single integration point i
in Voigt notation is calculated as:
⎛ ⎞
1 s ⎝ s(i) h  s(j) (j) h ⎠
Nb
εl (i)
= Blk(i) ukh
− ml Tk uk − [Tk w ]uk
d j=1
⎛ ⎞
1 Nb
− mlw ⎝Tkw(i) ukh − [Tk w(j) ]ukh ⎠
w(j)
d j=1
⎡ ⎛ ⎞
1  Nb
= ⎣Blk(i) − mls ⎝Tks(i) − [Tk w(j) ]⎠
s(j)
d j=1
⎛ ⎞⎤
1  Nb
− miw ⎝Tkw(i) − [Tk w(j) ]⎠⎦ ukh
w(j)
d j=1

≡ Blk ukh .
Meshfree Methods Applied to Consolidation Problems … 251

3 Application to Consolidation of Soils

As mentioned before, the settlement of saturated soils under loading is caused by a

gradual interchange between pore pressure and effective stress. In this Section, we
apply the above developed methodology for consolidation of soils at three different
configurations: one dimensional case, radial consolidation and consolidation with
sinks. Both static and dynamic scenarios are studied. The obtained solutions are
compared with analytical or available numerical solutions.

3.1 Consolidation of a Column of Soil: Static, One

Dimensional Case

In this case, the analytical solution for this problem is available, and thus it is com-
pared with the solution proposed by the present method. Although the analytical
solution is presented in non-dimensional terms, the geometry of the problem carried
out is shown in Fig. 1. It consists of a column of 30 m of soil resting on an imperme-
able rigid base layer and loaded by a vertical, homogeneous loading at the top. The
lateral displacements are restricted for both the solid and fluid phase. At the base
layer, the solid phase is fixed, whereas the vertical movement of the fluid phase is

1m

Nodes
P P=P·e
Material Points

x
y (m)

y 5

HT=30 m

0
0 1
x (m)
Base layer, impermeable

Fig. 1 Geometry, loading condition of the consolidation column, and the discretised nodes and
material points (shown for the first 5 m only). The same geometry has been employed for both static
and dynamic simulations
252 P. Navas et al.

Fig. 2 Loading history in a P

monotonic problem

100 kPa

0.1 s t

prevented. The column is discretised into 240 nodes and 183 material points. The
external loading, in this case, is static at the end, but gradually applied as depicted
in Fig. 2. The behaviour of the consolidation is led by the vertical consolidation
coefficient, cv , which is function of the vertical permeability coefficient, kv :
kv (1 + e) kv kv
cv = = Em = (27)
ρw gav ρw g ρw gmv

where av is the compressibility coefficient and mv is the volumetric compressibility

coefficient. The porous index, e, a measure of the porosity is expressed as follows:
n
e= . (28)
1−n

In Eq. (27), Em is the oedometric modulus, related with the Young’s modulus E
according to the following equation:
 
2ν 2
E = Em 1− . (29)
1−ν

Typical values adopted for clays are 2 MPa for the Young’s modulus and 0.33 for
the Poisson’s ratio.
The basic equation for one-dimensional consolidation derived by Terzaghi in
1923 [23] is
∂ 2u ∂u
cv 2 = . (30)
∂z ∂t

The solution searched is a measure of the consolidation of the soil. It depends on the
vertical time factor, Tv , defined by:

cv t
Tv = . (31)
H2
Meshfree Methods Applied to Consolidation Problems … 253

Adopting the degree of interstitial pressure dissipation, Uv , we can compare the

analytical solution, given by the following equation, with the results obtained with
the present research:

ue (z)  2
m=∞   z 
Uv (z) = 1 − =1− sin M 1 − exp(−M 2 Tv ) (32)
u0e m=0
M H

where π
M= (2m + 1), m = 0, 1, 2, . . . , ∞. (33)
2
In Fig. 3, it is given the comparison between the analytical and the numerical solution
along the depth of the column of soil for different values of Tv . As it is seen, all the
values for Tv are dimensionless. With this comparison, we consider the current model
employed is sufficiently validated.
This solution has been also compared with PLAXIS, in order to have an idea on the
accuracy of this commercial software, since it is going to be employed hereinafter
for several other theoretical examples. The direct conclusion obtained in Fig. 4 is
that the accuracy decreases at the final stages of the consolidation. For low values
of Tv PLAXIS solution of Uv along the column of soil is similar to the calculated

Present Research
1
Analytical solution
0.9
Tv=0.01

0.8
0.03
Tv=
0.7
5
0.0
0.6 Tv=
0.1
Tv= 0.2
Tv=
z/H

0.5

0.4
0.3 .5
Tv= =0
Tv
0.3
.7

Tv=
=0

1.0
Tv

0.9

Tv=

0.2
Tv=

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Consolidation degree, Uv

Fig. 3 Analytical and computational solution of Uv for different values of Tv

254 P. Navas et al.

Present Research

Analytical solution
1
PLAXIS solution
0.9
Tv=0.01

0.8

0.7

0.6
0.2
Tv=
z/H

0.5

0.4 .5
=0
Tv
0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Consolidation degree, Uv

Fig. 4 Solutions of Uv for several values of Tv obtained in the current work compared with the
analytical ones and those obtained with PLAXIS software

in this research, but is getting noticebly different for higher values of Tv . Therefore,
PLAXIS solutions will provide a good approach of the trend of the pressure along the
consolidation without an accurate degree of precision, at least, for one dimensional,
static problems.
The solution for the monotonic loading at the upper side of the column shows
the dissipation of the pore pressure along time. Figure 5 provides the comparison
between the solution obtained with the present methodology, the u − w solution
obtained with a quadratic FEM model, and the one calculated using the software
GeHoMadrid [25]. Hardly any perceptible difference can be noticed from these
three solutions.

3.2 Consolidation of a Column of Soil: Dynamic, 1D Case

In this example, using the same geometry for the soil column, the consolidation of a
soil column vertically subjected to harmonic pressure is obtained. This problem was
first analytically solved by Zienkiewicz et al. [3] in 1980s, and recently by López-
Meshfree Methods Applied to Consolidation Problems … 255

100

80

López-Querol (2006)
60
PW (kPa)

GeHo-Madrid

Present research
40

20

0 1 2 3 4
t (s)

Fig. 5 Pore pressure evolution solutions at the top of the consolidation column

Table 1 Material parameters employed for the dynamic consolidation problem of a soil column
G (MPa) ν (–) n (–) ρ (kg/m3 ) ρf (kg/m3 ) Kf (MPa) ω (rad/s) Ks (MPa)
312.5 0.2 0.333 3003 1000 103 3.379 1034

Querol [26], among others, through employing quadratic finite element method. The
parameters for the material are those presented in Table 1, where ω, is the applied
loading frequency. A periodic surface load with the amplitude of 100 kPa, a frequency
of 3.379 rad/s, is imposed. This dynamic load as well as all the material properties
in Table 1 are chosen to be the same as those of Zienkiewicz et al. [3].
The variation of the pore pressure with depth is illustrated for different values of
π1 , a dimensionless parameter defined as follows,

Vc2
π1 = k ρf , (34)
g ρ
ω HT2

where HT is the column height and Vc is the compressive wave velocity calculated
as: 
Kf 1 2G(1 − ν)
Vc = D+ , D= , (35)
n ρ 1 − 2ν

where D the bulk modulus of the soil skeleton (dry mixture). Note that, for the
given material properties and loading frequency, π1 is proportional to the hydraulic
conductivity k. Once k (thus π1 ) is known, transient calculations can be carried out to
obtain the envelop of the pore pressure history for different points along the column
depth, or the isochrone.
256 P. Navas et al.

P/PT
0 0.5 1 1.5
0 10-2

10-1
Zienkiewicz et al. (1980)
Present research
100 López-Querol (2006)
H/HT 0.5 2
10

103 101

Fig. 6 Isochrones of the pressure in the whole column for different π1 values: comparison for
solutions taken from Zienkiewicz et al. [3], López-Querol [26] and those obtained in the present
research. The depth is normalised by the column height, HT , whereas the pore pressure is made
non-dimensional by PT , 100 kPa

After performing six different calculations for six different levels of π1 , from 10−2
to 103 , we obtain the isochrones of the pore pressure depicted in Fig. 6. Additionally
plotted are the results obtained by López-Querol [26] using quadratic finite elements
formulating the problem in displacements as well, along with the analytical solutions
provided by Zienkiewicz et al. [3]. It is noteworthy that the three different approaches
achieve quite similar isochrone maps; nevertheless, whenever more scattering is
observed, the current meshfree solution is closer to the analytical one.

3.3 Radial Consolidation: Static Axisymmetric Problems

The second problem carried out in this research is about radial consolidation. The
physical equation which governs this problem is different from the Terzaghi’s equa-
tion [23] shown in the previous section, i.e.:
 
∂ 2 ur 1 ∂ur ∂ur
ch + = , (36)
∂r 2 r ∂r ∂t

where ch is the horizontal consolidation coefficient, equivalent to the cv coefficient

in vertical consolidation:
kh (1 + e)
ch = . (37)
ρw gav
Meshfree Methods Applied to Consolidation Problems … 257

Fig. 7 Scheme of section of q

set of drains

re

rw

Impervious layer

In Fig. 7, a scheme of drains with induced radial flux is shown, where r and z
directions are defined as depicted. There is an analytical solution for this problem
given by Barron in 1948 [27] who defined the radial consolidation degree as a function
of the non-dimensional time Tr :

2Tr
Ur (Tr ) = 1 − exp − , (38)
F(nr )

where
nr2 3nr2 − 1
F(nr ) = log(nr ) − . (39)
nr2 − 1 4nr2

The coefficient nr depends on the relative extension of the drain in a particular

geometry, for the section defined in Fig. 7, it is calculated as
re
nr = . (40)
rw

where rw is the drain radius and re is the radius of influence for each type of problem.
In this case a quadrangular net of drains shown in the scheme of Fig. 8 is studied.

Fig. 8 Quadrangular net of

drains (re = 0.564 S)
S

re

Drains
258 P. Navas et al.

Since radial consolidation equation involves a second term (tangential flow), it is

not possible to solve within a plane strain formulation, as the one employed for verti-
cal, one dimensional consolidation. Therefore, the axisymmetric framework, shown
in previous sections, is employed instead and excellent results are obtained. In Fig. 9,
several solutions of the radial consolidation degree, Ur , along the non-dimensional
time Tr are shown. In addition, a comparison with a commercial software, PLAXIS,
is given, even though this program does not allow us to implement a perfect radial
consolidation due to the impossibility to neglect the vertical displacement throughout
the domain. Two alternatives are proposed instead of the original problem, allowing
for the vertical displacement: one assumes an impervious boundary condition on the
top layer; the other one allows the flux of water through this boundary. Results in
Fig. 10 offer a good trend in both cases but not the accuracy expected, as it occurs in
the vertical consolidation too.

Fig. 9 Analytical and Tr

computational solutions of 0.0001 0.001 0.01 0.1 1 10 100
Ur along the 0
non-dimensional time Tr 0.1

0.2

0.3
n r=5

0.4
Ur

0.5
n r=15

Present research
0.6
Analytical solution
0.7
n r=10

0.8
00

0.9

Fig. 10 Comparison Tr
between PLAXIS and 0.01 0.1 1 10 100
present research solutions of 0
Ut = Ur + Uv along the 0.1
non-dimensional time Tr for
case A (permeable boundary) 0.2
Present research
and B (impervious boundary) 0.3
PLAXIS
0.4
0.5
Ut

A
0.6
0.7
B
0.8
0.9
1
Meshfree Methods Applied to Consolidation Problems … 259

3.4 Radial Consolidation: Dynamic Axisymmetric Problems

A dynamic loading on the surface has been applied, aiming to simulate its effect on the
development of excess pore water pressure in the domain. The frequency of loading
is the same as for the case of the soil column, while the amplitude is 50 kN. The
geometry is the same as represented in Fig. 7. Vertical displacements of water in the
entire domain have been prevented. Figure 11 represents the evolution of pore water
pressure at the lateral, lowest corner, which clearly demonstrates the cyclic response
of this result as well. From this figure it can be concluded that the generation and
dissipation of excess pore water pressure are balanced in every cycle, and steady state
is achieved from the very beginning of the loading. Additionally, Fig. 12 represents

1.5

0.5
P/P0

−0.5

−1

−1.5
0 1 2 3 4 5 6 7 8 9 10
t (s)

Fig. 11 Evolution of normalised excess pore water pressure during external cyclic loading (for the
dynamic, radial consolidation problem)

Maximum Minimum
10 1 10 −0.8
9 0.98 9 −0.82
8 0.96 8 −0.84
7 0.94 7 −0.86
6 0.92 6 −0.88
z (m)
z (m)

5 0.9 5 −0.9
4 0.88 4 −0.92
3 0.86 3 −0.94
2 0.84 2 −0.96
1 0.82 1 −0.98
0 0.8 0 −1
0 1 2 3 4 5 0 1 2 3 4 5
p/p0 p/p0
r (m) r (m)

Fig. 12 Maximum and minimum normalised excess pore water pressures in dynamic, radial con-
solidation
260 P. Navas et al.

maximum and minimum values of excess pore water pressure in the entire domain
during the cyclic loading. This figure demonstrates the 2D nature of the problem, the
drain being clearly displayed on the left hand side of the domain. Higher excess pore
water pressures can be seen at the lower, left corner for both maximum and minimum
cases. Moreover, the figure demonstrates the alternate positive and negative values
for the pressures in the domain, as in Fig. 11.

3.5 Static Consolidation in a Soil with a Singular Point: Sink

The insertion of singular points inside the domain may vary the consolidation behav-
iour. The existence of a sink in the middle of horizontal soil layer is expected to
accelerate the consolidation of the porous media, since this means an output of water
at the permeable top boundary. To reproduce the sink, excess pore water pressure
is not allowed to develop in several nodes in the centre of the domain (see Fig. 13).
A square with one meter edge length is proposed for the study of this problem.
Material properties are given in Table 2. Figures 14 and 15 show the comparison of
results obtained with the present model as well as with PLAXIS for point (0, 0).
Figure 14 provides the evolution in time of the degree of consolidation, U, at one of
the corners at the bottom of the domain, whereas Fig. 15 represents the solutions in
the whole domain after two seconds. In spite of slight differences in the final part of
the evolution, it can be concluded that both results are fairly similar, demonstrating
the good performance of the present formulation for this kind of problems.

Fig. 13 Geometry for the q=100 kN/m

problem of the sink

0.5 m.
1 m.

Drain
0.5 m.

Impervious

1 m.

Table 2 Material parameters employed for the consolidation problem of a soil with a sink
E (MPa) ν (–) n (–) ρ ρf Kf Ks k (m/s) ksink
(kg/m3 ) (kg/m3 ) (MPa) (MPa) (m/s)
100 0.0 0.333 3003 1000 103 1034 10−3 10
Meshfree Methods Applied to Consolidation Problems … 261

t (s)
0.1 1 10 100

Consolidation degree, U
0

0.2

0.4
Present research
0.6
PLAXIS
0.8

Fig. 14 Evolution of consolidation degree at the bottom corner for the sink problem. Present model
versus PLAXIS

Present research solution p/p0 PLAXIS solution p/p0

1 0.8 1 0.8
0.9 0.9
0.7 0.7
0.8 0.8
0.6 0.6
0.7 0.7
0.6 0.5 0.6 0.5

0.5 0.4 0.5 0.4

z

0.4 0.4
0.3 0.3
0.3 0.3
0.2 0.2
0.2 0.2
0.1 0.1 0.1 0.1

0 0 0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x x

Fig. 15 Field of consolidation degrees in the domain after 2 s. Present model versus PLAXIS

3.6 Dynamic Consolidation in a Soil

with a Singular Point: Sink

As for the case of radial consolidation, a dynamic simulation has been carried out. The
geometry is the same as in Fig. 13. Figure 16 represents the evolution of normalised
excess pore water pressure at a bottom corner, clearly showing the cyclic nature of
the solution, which is steady from the beginning. Moreover, Fig. 17 shows maximum
and minimum values in the whole domain. This figure clearly demonstrates the
location of the sink, with zero water pressures in the centre of the domain for both
cases. The alternate positive and negative values are also clear from the plot. Once
again, these results demonstrate the suitability of the present formulation for dynamic
consolidation problems in saturated soils.
262 P. Navas et al.

1.5

0.5
P/P0

−0.5

−1

−1.5
0 1 2 3 4 5 6 7 8 9 10
t (s)

Fig. 16 Evolution of normalised excess pore water pressure during external cyclic loading (dynamic
consolidation in a soil with a sink)

Maximum p/p0 Minimum p/p0

1 1 1 1
0.9 0.8 0.9 0.8
0.8 0.6 0.8 0.6
0.7 0.4 0.7 0.4
0.6 0.2 0.6 0.2
z/H

z/H

0.5 0 0.5 0
0.4 −0.2 0.4 −0.2
0.3 −0.4 0.3 −0.4
0.2 −0.6 0.2 −0.6
0.1 −0.8 0.1 −0.8
0 −1 0 −1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x/L x/L

Fig. 17 Maximum and minimum normalised excess pore water pressures for dynamic soil consol-
idation with a sink

4 Conclusions

We have extended the previously developed B-bar based algorithm to meshfree

numerical schemes in axisymmetric framework for both elastic and porous media.
The methodology is applied to both static and dynamic consolidation problems in
saturated soils. In particular, static and dynamic consolidation of a soil column, sta-
tic and dynamic radial consolidation, static and dynamic consolidation with singular
points (sinks), are carried out and compared with analytical solutions (whenever
exist) or available finite element solutions. The feasibility of the current formulation
in solving consolidation problems in saturated soils has been clearly demonstrated.
Meshfree Methods Applied to Consolidation Problems … 263

Acknowledgments This research has been partially funded by the Spanish Ministry of Economy
and Competitiveness through the projects BIA2012–31678 and MAT2012–35416. The first author
also acknowledges the financial support via the fellowship No. BES2013–063924.

References

1. Biot, M. A. (1941). General theory of three-dimensional consolidation. Journal of Applied

Physics, 12(2), 155–164.
2. Biot, M. A. (1956). General solutions of the equations of elasticity and consolidation for a
porous material. Journal of Applied Mechanics, 91–96.
3. Zienkiewicz, O. C., Chang, C. T., & Bettes, P. (1980). Drained, undrained, consolidating and
dynamic behaviour assumptions in soils. Géotechnique, 30(4), 385–395.
4. López-Querol, S., & Blazquez, R. (2006). Liquefaction and cyclic mobility model in saturated
granular media. International Journal for Numerical and Analytical Methods in Geomechanics,
30, 413–439.
5. Cividini, A., & Gioda, G. (2013). On the dynamic analysis of two-phase soils. In S. Pietruszczak
& G. N. Pande (Eds.), Proceedings of the third international symposium on computational
geomechanics (ComGeo III) (pp. 452–461).
6. Ortiz, A., Puso, M. A., & Sukumar, N. (2004). Construction of polygonal interpolants: A
maximum entropy approach. International Journal for Numerical Methods in Engineering,
61(12), 2159–2181.
7. Arroyo, M., & Ortiz, M. (2006). Local maximum-entropy approximation schemes: A seamless
bridge between finite elements and meshfree methods. International Journal for Numerical
Methods in Engineering, 65(13), 2167–2202.
8. Li, B., Habbal, F., & Ortiz, M. (2010). Optimal transportation meshfree approximation schemes
for fluid and plastic flows. International Journal for Numerical Methods in Engineering, 83,
1541–1579.
9. Nelder, J. A., & Mead, R. (1965). A simplex method for function minimization. Computer
Journal, 7, 308–313.
10. Hughes, T. J. R. (1980). Generalization of selective integration procedures to anisotropic and
nonlinear media. International Journal for Numerical Methods in Engineering, 15, 1413–1418.
11. Simo, J. C., & Rifai, M. S. (1990). A class of mixed assumed strain methods and the method of
incompatible modes. International Journal for Numerical Methods in Engineering, 29, 1595–
1638.
12. Kasper, E. P., & Taylor, R. L. (2000). A mixed-enhanced strain method: Part I: Geometrically
linear problems. Computers and Structures, 75(3), 237–250.
13. De Souza Neto, E. A., Pires, F. M., & Owen, D. R. J. (1980). F-bar-based linear triangles
and tetrahedra for finte strain analysis of nearly incompressible solids. Part I: Formulation and
benchmarking. International Journal for Numerical Methods in Engineering, 62, 353–383.
14. Bonet, J., & Burton, A. J. (1998). A simple average nodal pressure tetrahedral element for
incompressible and nearly incompressible dynamic Explicit applications. Communications in
Numerical Methods in Engineering, 14(5), 437–449.
15. Hauret, P., Kuhl, E., & Ortiz, M. (2007). Diamond elements: A finite element/discrete-
mechanics approximation scheme with guaranteed optimal convergene in incompressible elas-
ticity. International Journal for Numerical Methods in Engineering, 73, 253–294.
16. Elguedj, T., Bazilevs, Y., Calo, V. M., & Hughes, T. J. R. (2008). B and F projection methods
for nearly incompressible linear and non-linear elasticity and plasticity using higher-order
NURBS elements. Computer Methods in Applied Mechanics and Engineering, 197(33–40),
2732–2762.
17. Artioli, E., Castellazzi, G., & Krysl, P. (2014). Assumed strain nodally integrated hexahedral
finite element formulations for elastoplastic applications. International Journal for Numerical
Methods in Engineering, 99(11), 844–866.
264 P. Navas et al.

18. Ortiz, A., Puso, M. A., & Sukumar, N. (2010). Maximum-entropy meshfree method for com-
pressible and near-incompressible elasticity. Computer Methods in Applied Mechanics and
Engineering, 199, 1859–1871.
19. Navas, P., López-Querol, S., Yu, R. C., & Li, B. (2016). B-bar based algorithm applied to
meshfree numerical schemes to solve unconfined seepage problems through porous media.
International Journal for Numerical and Analytical Methods in Geomechanics, 40, 962–984.
20. López-Querol, S., Navas, P., Peco, J., & Arias-Trujillo, J. (2011). Changing impermeabil-
ity boundary conditions to obtain free surfaces in unconfined seepage problems. Canadian
Geotechnical Journal, 48, 841–845.
21. Navas, P., & López-Querol, S. (2013). Generalized unconfined seepage flow model using
displacement based formulation. Engineering Geology, 166, 140–141.
22. Biot, M. A. (1956). Theory of propagation of elastic waves in a fluid-saturated porous solid. I.
Low-Frequency range. Journal of the Acoustical Society of America, 28(2), 168–178.
23. Terzaghi, K. V. (1925). Principles of Soil Mechanics. Engineering News-Record, 95, 19–27.
24. Zienkiewicz, O. C., & Taylor, R. L. (1994). El método de los elementos finitos. Vol 1: Formu-
lación básica y problemas lineales. Barcelona: CIMNE.
25. Fernández Merodo, J. A., Mira, P., Pastor, M., & Li, T. (1999). GeHoMadrid User Manual.
Madrid: CEDEX. Technical Report.
26. López-Querol, S. (2006). Modelización geomecánica de los procesos de densificación, licue-
facción y movilidad cíclica de suelos granulares sometidos a solicitaciones dinámicas. PhD
thesis, University of Castilla-La Mancha, Ciudad Real, Spain.
27. Barron, R. A. (1948). Consolidation of fine-grained soils by drain wells. Transationc ASCE,
113, 718–754.
A Multiscale Microstructural Model
of Permeability in Fractured Solids

Anna Pandolfi, Maria Laura De Bellis and Gabriele Della Vecchia

Abstract We discuss a microstructural model of permeability in fractured solids,

where fractures are modeled as recursive families of parallel, equidistant faults.
Faults are originated by the attainment of a resistance threshold under the action of a
confinement pressure over an initially undamaged, fully elastic matrix, not necessar-
ily isotropic. The initially undamaged matrix might possess a natural permeability,
which is modified by the progressive damage of the rock. The particular organiza-
tion of the micro-faults considered in the model allows to define analytically the
equivalent permeability of the solid. The model is particularly appealing to describe
the permeability of rock undergoing the process of fracking, in a form that has great
computational advantages, since the approach does not track explicitly the formation
of individual macro-faults.

1 Fluid Flow in Porous Media

Permeability is the property of materials that measures the ability for fluids (gas or
liquid) to flow through a porous solid material, essentially related to the void topology
and not to the properties of the fluid.
The void space of a porous medium, where liquid or gaseous fluids are allowed
to flow, can be thought as composed of a spatial network of interconnected random
passages, channels or tubes of varying length, cross section and orientation, and
junctions, where channels meet [1]. Considering an elementary volume with a suffi-
ciently large number of channels, experimentally it is observed that the network of
channels produces average gradients of pressure, density, viscosity and solute con-

A. Pandolfi (B) · M.L. De Bellis · G.D. Vecchia

Politecnico di Milano, Milano, Italy
e-mail: [email protected]
M.L. De Bellis
e-mail: [email protected]
G.D. Vecchia
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 265

K. Weinberg and A. Pandolfi (eds.), Innovative Numerical Approaches
for Multi-Field and Multi-Scale Problems, Lecture Notes in Applied
and Computational Mechanics 81, DOI 10.1007/978-3-319-39022-2_12
266 A. Pandolfi et al.

centration, which are independent of the geometry of a single channel. Contrariwise,

local deviations from the average at points within the void space depend strongly on
the geometry of the solid matrix.
A physical property of a material containing voids is the porosity, or void fraction
n, expressing the ratio between the volume of the voids and the total volume V of
solid VS and voids VV
VV VV
n= = . (1)
V VS + VV

In a volume of space occupied by a multispecies fluid system, every species α has

a different velocity vα , and it is possible to introduce several definitions of average
(v∗ mass or v volume) velocities. In general, the three velocities differ in direction
and magnitude. This discrepancy disappears if the fluid contains homogeneous single
species; in the following we consider a homogeneous fluid, and we denote its velocity
with v, of magnitude v.
According to the geometrical arrangement, fluids lose energy most in the channels
and not in the junctions that can be disregarded. The energy of a fluid is traditionally
measured in terms of total hydraulic head h, a simplified form of the Bernoulli’s
definition for incompressible fluids, i.e.,

p v2
h= +z+ , (2)
ρg 2g

where p is the gauge pressure, ρ is the density of the fluid, and g is the gravitational
acceleration. The first term, p/ρg, is called pressure head, the second term, z, is called
elevation head, and the third term, v2 /2g, is called velocity head. The pressure head is
the equivalent gauge pressure of a column of water at the base of the piezometer. The
elevation head is the relative potential energy in terms of an elevation. The velocity
head is related to the kinetic energy of the fluid, and in most applications, where
the fluid velocity is very small, is disregarded. In the following we will discard the
kinetic contribution.
Fluid flow across packed porous media is in general characterized by laminar
regime (Reynolds number Re ≤ 1) and a drop in the hydraulic head. The hydraulic
gradient ∇h measures the change of the hydraulic head field and is related to the
direction of the flow in the porous medium. In a cartesian orthogonal reference system
it is defined as
∂h ∂h ∂h
∇h = e1 + e2 + e3 . (3)
∂x1 ∂x2 ∂x3

Analytical models of fluid flow in rocks use constitutive relations that link the average
velocity of the fluid across the medium to the pressure drop. In particular, Darcy’s
law states that the discharge rate in a porous media is proportional to the hydraulic
head gradient and inversely proportional to the fluid viscosity
A Multiscale Microstructural Model of Permeability … 267

ρg
q = −κ ∇h, (4)
μ

where q is the discharge rate and μ the fluid viscosity, and κ the medium permeability.
In anisotropic media the permeability is described by a symmetric and positive
definite second order tensor. Symmetry of κ is the consequence of the Onsager recip-
rocal relations, while positive definiteness follows from the fact that a fluid cannot
flow against the pressure drop. These two properties render the tensor diagonaliz-
able. Eigenvalues of the permeability tensor represent the principal permeabilities,
and the corresponding eigenvectors define the principal directions of flow, i.e., the
directions where flow is parallel to the pressure drop.

1.1 Analytical Models of Permeability

Permeability is an overall important physical property of porous media. In order to

be used in applications, it has to be measured, either directly or through estima-
tions using formulas derived empirically. Relationships between permeability and
commonly measured physical variables of porous media, such as porosity and elas-
tic wave velocity, are not easily established. Permeability is also very difficult to
characterize theoretically. Difficulties derive from the complexity of the pore fab-
ric geometry and connectivity, which introduces large uncertainty on the range of
applicability and on the predictability of analytical models. However, for simple and
structured models of porous media, permeability can be estimated through analytical
relationships. Analytical models often are limited to the particular porous medium
under investigation, and apply only under a narrow range of conditions.
The class of Kozeny-Carman type models collects simple relations that, under the
assumption of laminar flow of the pore fluid, link the permeability to the microstruc-
tural characteristics of the porous medium. The original Kozeny-Carman relation
[2–4] was derived by extending Poiseuille’s law, valid for straight circular section
pipes, to the flow of a fluid in a collection of curving passages and tubes embedded
in a porous media. The original pipe radius of Poiseuille’s law is replaced by the
hydraulic radius, defined as the ratio of the pore volume to the solid-fluid interfacial
area, a new concept introduced to bypass the definition of a representative radius,
complex even in a natural homogeneous porous medium. The Kozeny-Carman equa-
tion reads  2
c n
k= 2 n , (5)
8av τ 1−n

where k is a scalar permeability, c an empirical geometric parameter, n the porosity,

av the ratio of the exposed surface of the channels to the volume of the solids (also
called specific internal surface area), and τ the tortuosity. In a simple manner, the
tortuosity can be defined as
268 A. Pandolfi et al.
 2
La
τ= , (6)
L

where La is the average length of the channels and L the macroscopic length of the
flow path. Relation (5) incorporates a characteristic microstructural length parameter
similar to that used in other analyses of permeability [5].
The estimation of the shape coefficients appearing in the equation has been pro-
moting an active research [6]. The specific internal surface area has been evaluated
with several methods, for example by combining the contributions of plastic clayey
and granular silty-sandy fractions [7], or scanning electron microscope images [8].
Simplified and more tractable tortuosity models have been introduced, for example
considering pore channels of variable shape but constant cross-section [9] or by intro-
ducing alternative definitions of tortuosity [10]. By using the analog of networks of
electrical resistors, τ can be linked empirically to the electrical conductivity of rocks
and the brine (salt solution) saturation through a quantitative relation due to Archie
[11]. Starting from Archie’s law, permeability models have been able to include the
pore connectedness in the correlation between permeability and local electric field
[12–15].
By describing the connected pore space as a bundle of tortuous leaky hydraulic
tubes, alternative permeability models have been proposed by Civan [16, 17]
 β
n
k=γn , (7)
1−n

where the parameter γ and the fractal exponent β have been originally determined
by experimental data fitting, and successively derived analytically, showing how the
two parameters γ and β can be linked to physically meaningful parameters of the
porous media.
The scalar nature of variables and parameters used in this class of models leads
to scalar definitions, and the correct tensor nature of the permeability is disregarded.
Therefore, such models are not meaningful if applied to soils characterized by the
presence of sedimentation layers or fissures. The complexity of the relationship
between the permeability tensor and a scalar property such as the porosity in rocks
has been clearly pointed out [18]. In particular, permeability depends not only on the
actual stress and on the strain during the loading history, but also on the evolution of
the crack patterns, which is anisotropic in nature. The Kozeny-Carman permeability
models are primarily intended for applications to static porous materials, whose
effective or conductive pore structure and properties remain unchanged during fluid
flow. Hence, these models do not allow for the modification of the porous medium
microstructure due to fluid-porous matrix interactions, or by the presence of a variable
confining pressure.
A Multiscale Microstructural Model of Permeability … 269

1.2 Computational Methods for Permeability

The permeability’s parameters can be evaluated using numerical approaches. Modi-

fied models of permeability are obtained by introducing the uncertainty deriving from
particle size distribution, through first-order error analysis methods [19]. Neural net-
works and genetic algorithms are employed to elaborate wide samples of data [20]
and to link, in modified models, the saturated permeability to the effective porosity.
Lattice Boltzmann simulation of flow in simplified 2D or 3D porous media have been
able to provide numerical evaluation of tortuosity [21] and permeability [22] for dif-
ferent values of the solid fraction, considering also the presence of spanning planar
fractures in the matrix [22]. Numerical results illustrate the importance of matrix-
fracture interactions, and prove the inadequacy of using simplified assumptions to
predict permeability from porosity in fractured porous rock.
More recently, advances in imaging techniques gave impulse to the numerical
simulations of physical processes occurring within sub volumes of rock samples,
in order to characterize the rock permeability and correlate it to the microstructural
features of the pores. Most of the research has been conducted using specialized
research software, but recently applications use commercial software [23].

1.3 Mechanics and Permeability

Fractures and faults represent the most ubiquitous and efficient ways for flows in
natural rock formations. The availability of fault and fracture mappings in reser-
voirs is an important recent achievement in geology, but the understanding of the
influence of these structures on fluid flows is still far from being satisfactory, in
particular when the mechanical coupling is significant. Difficulties emerge from the
complexity of the topology and geometry of faults. Each group or class of faults is
characterized by orientation, spacing, distribution, and connectivity, which affect the
entrapment of fluids, limiting or advantaging the migration and flow of fluids in a
given environment [24].
Clearly, the complexity of natural fracture networks is associated to the stress state
history, which is hardly known. Furthermore, cracks and fracture can evolve due to
the action of gravity, superposed localized pressures, and shear tractions resulting
from the viscosity of the flowing fluids.
The observation that intact rocks contain distributed flaws and cracks, arranged
between particles of various shape, has motivated the use of fracture mechanics to
study their organization and the conditions that promote their growth [25]. The most
widely used fracture mechanics models to describe the progressive microfracturing
of rocks upon increasing loading are the open crack and the sliding crack models.
However, standard approaches of fracture mechanics have limiting drawbacks related
to the explicit treatment of cracking.
270 A. Pandolfi et al.

As an alternative to fracture mechanics, continuum damage mechanics consid-

ers the averaged effect of microstructural changes, following a phenomenological
approach able to reproduce hydro-mechanical responses during rock progressive
degeneration. One of the most interesting approaches treats rock masses containing
a large number of discontinuities as homogeneous, anisotropic porous media [26].
The cracks in the medium are assumed to follow a probability distribution function
PDF (N, L, Δ) in terms of crack orientation N, size L and opening Δ. By using an
averaging procedure, a symmetric crack tensor associated to the permeability tensor
of the cracked porous medium is derived. The principal directions of the permeability
tensor are coaxial to the ones of the crack tensor. Thus, the first invariant of the crack
tensor results to be proportional to the mean permeability, while the deviatoric part
of the crack tensor is related to the anisotropic permeability.
Various methods to take into account the effects on the permeability of rocks of
several factors, such as the coupled effect of flow, stress and deformation, the propa-
gation of existing fractures, and the initiation of new fractures, have been developed
in the framework of continuum mechanics.
A simplified coupled hydro-mechanical continuum approach, based on the Biot’s
theory of fluid saturated porous media and on brittle-elastic solids with residual
strength, is considered in a finite element model that includes damage with elas-
tic unloading/reloading [27]. Hydraulic anisotropy and internal state variables are
not considered, thus the resulting permeability is treated as a scalar and directly
dependent on the stress state.
A micromechanical point of view has been taken in [28] to assess the influence
of local damage on the macroscopic hydro-mechanical response of porous media.
The damage variables are related to the degradation of elastic properties and to
the characteristics of the fracture network, thus the model has the possibility to
describe the evolution of the porous network with deformation and its influence
on permeability [29]. The model distinguishes between the natural pore network,
sensitive to the deformation of the representative volume, and the crack network,
enucleated in the damaged material, and assumes laminar flow. The natural pore
network is characterized by a Pore Size Distribution (PSD) curve, updated with the
state variables and with the evolution of the cracks, and linked to the permeability.
In spite of the anisotropic nature of the crack pattern, a scalar value of hydraulic
conductivity is defined by integration of the PSD.
Coupling between deformation and fluid flow has been accounted for in various
manners. By considering the volumetric strain as an additional controlling parame-
ter, in [30] numerical simulations of permeability reduction (increase) upon elastic
contraction (dilation) in rocks are presented. Permeability is considered there as a
scalar variable, although the approach accounts for flow in both matrix and fractures.
The dependence of rock permeability on material deformation has been enforced
alternatively in terms of crack opening. A sophisticated coupled semi-empirical
hydro-mechanical constitutive model accounting for anisotropic damage induced by
cracks and modification in permeability in brittle rocks under deviatoric compressive
stresses has been proposed in [31]. The rock is regarded as a porous medium with
embedded microcracks. Upon homogenization, the cracked material is treated as an
A Multiscale Microstructural Model of Permeability … 271

equivalent porous medium where the permeability tensor is decomposed additively

into initial permeability and crack induced permeability tensors. Since the micro-
crack distribution is orientation-dependent, the crack permeability tensor has to be
anisotropic in nature and it is regarded as a function of crack number, orientation,
radius, and average opening. The mechanical model is formulated in terms of linear
elasticity and the crack propagation conditions are based on linear elastic fracture
mechanics, without the support of a thermodynamical framework.
The effects of damage on anisotropic permeability have been discussed in [32]
by adopting a relationship between macroscopic and microscopic aspects of dam-
age, and exploiting micro-level analyses of flow through randomly generated crack
networks.
In agreement with the literature on the topic, material models that account for the
progressive evolution of defects, flaws and cracks seem to possess the right charac-
teristics to account for the variation of the permeability according to the evolution
of the damage in the considered porous material.
In the next paragraph we recall a multi-scale brittle damage material model [33]
that describes the fractures using a cohesive approach [34, 35]. The mechanical model
has been discussed in the original paper; here we analyze the kinematic aspects of
the model in view of deriving analytically the permeability.

2 Brittle Damage Model

The brittle damage model presented in [33] is characterized by a homogeneous matrix

where nested microstructures characterized by different length scales are embedded.
In each level k of the nested architecture, microstructures assume the form of families
of equidistant cohesive faults, characterized by an orientation Nk and a spacing Lk . In
line with general mathematical results pertaining to free discontinuity problems, the
constitutive model is derived within a thermodynamical approach where we assume
the existence of a free energy density which accounts for reversible and dissipative
behaviors of the material.
The key of the model is given by the kinematic assumptions. We begin by consid-
ering the particular case of a single family of fault planes of normal N and spacing L,
and later extend the behavior to nested families. The total deformation gradient F of
the material is assumed to decompose multiplicatively into a part Fm pertaining the
uniform deformation of the matrix, and a second part Ff pertaining the discontinuous
kinematics of the cohesive faults, i.e.,

F = Fm Ff . (8)

The deformation gradient Ff can be easily linked to the fault kinematic activity.
Consider a material vector dX, shorter than the system size but longer than the
internal scale L, that spans two material points P and Q in the material configuration.
The number of faults m traversed by the vector is
272 A. Pandolfi et al.

Fig. 1 Inelastic kinematics of the fault system. The opening displacement Δ applied to all the
faults at distance L leads to a deformed configuration characterized by the inelastic deformation
gradient Ff . a Reference configuration. b Spatial configuration

1
m= dX · N.
L
Let us now apply an opening displacement Δ to each fault, Fig. 1a. In the spatial
configuration the two points P and Q are joined by the vector dx given by

1 1
Δ = dX +
dx = dX + mΔ (dX · N) Δ = (I + Δ ⊗ N) dX,
L L
where we set
1
Ff ≡ I + Δ ⊗ N. (9)
L

Note that, once N and L are supplied, Ff and Δ are in one-to-one correspondence,
and the opening displacements Δ follow from Ff through the relation

Δ = L (Ff − I) N. (10)

Note that the matrix may, in turn, accommodate a second fault family:
1 2
F = Fm1 Ff , Fm1 = Fm2 Ff . (11)

This decomposition can be applied recursively for as many levels as necessary, assign-
ing the innermost level a purely elastic behavior

k k−1 2 1
F = Fe Ff Ff . . . Ff Ff . (12)

The constitutive behavior of the material is thus obtained by introducing a free

energy density that decomposes into the sum of two contributions

1
A(Fm , Δ , q) = W m (Fm ) + Δ, q),
Φ(Δ (13)
L
A Multiscale Microstructural Model of Permeability … 273

where W m is the strain-energy density per unit volume of the matrix, Φ is the cohe-
sive energy per unit fault surface, suitably divided by the scale length L to result in a
specific energy per unit of volume, and q is an internal variable describing the max-
imum opening displacement ever experienced by the faults. Note that the separation
of the variables excludes strong coupling between the two energies. The particular
form of the energy densities W m and Φ can be selected with a certain degree of
freedom according to the particular material considered.
In keeping with the irreversible nature of fracture, decohered faults permanently
damage the material. This can be expressed by the internal variable q evolution
constraint q̇ ≥ 0. Another important constraint is the impenetrability of the faults
surfaces upon closure, i.e., the component of the opening displacement along the
fault normal cannot be negative, thus N · Δ ≥ 0.
The behavior of irreversible materials can be characterized variationally by
recourse to time discretization [33, 36], that requires to consider a process of defor-
mation at distinct successive times t0 , . . . , tn+1 = tn + Δt, . . . . We assume that the
state of the material at time tn is known and the deformation Fn+1 at time tn+1 is
assigned. The problem is to determine the state of the material at time tn+1 . We
define an effective, incremental, strain-energy density taking the infimum of the
constrained energy with respect to Δ n+1 and qn+1 , as

Wn (Fn+1 ) = inf A(Fn+1 , Δ n+1 , qn+1 ). (14)

Δn+1 , qn+1
Δ n+1 · N ≥ 0
qn+1 ≥ qn

The subindex n used in Wn signifies the dependence on the initial state. The irre-
versibility and the impenetrability constraints render the effective strain-energy den-
sity Wn dependent on the initial conditions at time tn , and account for all the inelastic
behaviors, such as irreversibility, hysteresis, and path dependency. Thus, Wn (Fn+1 )
acts as a potential for the first Piola-Kirchhoff stress tensor Pn+1 at time tn+1 [36],
i.e., as
∂ Wn (Fn+1 )
Pn+1 = . (15)
∂Fn+1

The stable equilibrium configurations are the minimizers of the corresponding effec-
tive energy. The fault geometrical features N and L, until now considered as known,
can be determined with the aid of the time-discretized variational formulation.
Energy optimization will ascertain whether the insertion of faults is energetically
favorable, and the optimal orientation of the faults. For a given deformation Fn+1 ,
we test two end states of the material, one with faults and another without faults,
and choose the end state which results in the lowest incremental energy density
Wn (Fn+1 ). The orientation of the faults N and the remaining state variables are
obtained variationally from the extended minimum problem:
274 A. Pandolfi et al.

Wn (Fn+1 ) = inf A(Fn+1 , Δ n+1 , qn+1 , N). (16)

Δ n+1 , qn+1 , N
Δ n+1 · N ≥ 0
qn+1 ≥ qn
|N|2 = 1

The actual orientation of the faults is defined by the surrounding stress state. In
particular, faults can be originated under tensile stress if the maximum tensile stress
reaches the tensile resistance of the material. Under compressive stress, it is likely
that the material would fail in shear.
The length L can be computed variationally by accounting for the misfit energy
Δ, L) contained in the boundary layers that forms where the faults meet a
E mis (Δ
confining boundary. In fact, the compatibility between the faults and their container
is only on average, and this gives rise to boundary layers that penetrate into the
faulted region to a certain depth. The addition to the energy furnishes a selection
mechanism among all possible microstructures leading to the relaxed energy [33].
The total free-energy density of the faulted region becomes

A(Fn+1 , Δn+1 , qn+1 , Ln+1 )

1 (17)
= W m (Fm n+1 , Ln+1 ) + Δn+1 , qn+1 ) + E mis (Δ
Φ(Δ Δn+1 , Ln+1 ).
Ln+1

The variational update (14) now becomes

Wn (Fn+1 ) = inf A(Fn+1 , Δ n+1 , qn+1 , Ln+1 ). (18)

Δ n+1 , qn+1 , Ln+1
Δ n+1 · N ≥ 0
qn+1 ≥ qn

The optimal fault separation is determined by two competing demands. On one hand,
the cohesive energy favors a large value of L resulting in fewer faults per unit volume.
On the other hand, the misfit energy favors a small value of L resulting in a narrow
boundary layer.
Internal friction is an important dissipation mechanism in brittle materials, espe-
cially in geological applications. We assume that friction operates at the faults con-
currently with cohesion, but if faults loose cohesion completely upon the attainment
of a critical opening displacement, friction may become the sole dissipation mech-
anism at the faults. In order to retain the variational structure of the model, we
define an incremental strain energy density Wn (Fn+1 ) that includes a dual kinetic
potential, attending the frictional dissipation, with suitable convexity and regularity
properties [37].
So far we have consider either an intact material or a single family of parallel
faults. The material with a single fault family is referred to as rank-1 faulting pattern
material. More complex microstructures, can effectively be generated by applying
the previous construction recursively. In the first level of recursion, we simply replace
A Multiscale Microstructural Model of Permeability … 275

the elastic strain-energy density W (Fm ) of the matrix by Wn (Fm ), i.e., by the effec-
tive strain-energy density of a rank-1 faulting pattern. This substitution can now be
iterated, resulting in a recursive definition of Wn (Fn+1 ). The recursion stops when the
matrix between the faults remains elastic. The resulting microstructures are shown
in Fig. 1a, and consist of faults within faults. The level of recursion is the rank of the
microstructure.
According to the particular loading history, at the time tn and at the generic point
P the material is characterized by a particular opening displacement Δ , which will
be respectful of the equilibrium and compatibility conditions. The model is therefore
able to account for a variable opening of the faults.

3 Permeability of the Brittle Damage Model

Under the assumption of a perfectly impermeable matrix and considering the pres-
ence of a single fault family, the permeability tensor for the material fractured brittle
damage model can be directly derived from the particular geometry of faults. The
permeability of a particular geometry of parallel and equidistant faults has been exam-
ined by Irmay [38]. Snow [39, 40] and Parsons [41] obtained analytical expressions
for the anisotropic permeability, similar to the one described here, by considering
networks of parallel fissures.
According to Fig. 2, the opening displacement Δ decomposes into a normal com-
ponent ΔN and a sliding component ΔT computed as

ΔN = N · Δ , Δ T = (I − N ⊗ N) Δ , ΔT |,
ΔT = |Δ (19)

The fluid fills layers of variable thickness ΔN and the average fluid flow will take
place in the direction of the layer.
We begin by considering the solution of the Navier-Stokes’ equation for average
velocity vs along the direction s in a fault of constant width ΔN :

Δ2N ρg ∂h
vs = , (20)
12 μ ∂s

Fig. 2 Kinematics of the

single fault, defined by an
opening displacement Δ ,
with a component ΔN along
the normal and a component
ΔT along the fault
276 A. Pandolfi et al.

where ∂h/∂s is the hydraulic head gradient in the direction s. By considering a porous
medium where pores are in the form of parallel faults, the discharge qs in the direction
of the flow is
ΔN Δ2N ρg ∂h
qs = n vs = , (21)
L 12 μ ∂s

where
ΔN
n= (22)
L
can be intended as a measure of the local porosity. The permeability in the direction
of the fault plane becomes
ΔN Δ2N
ks = . (23)
L 12
Now we restate the above equations in vector form. By introducing the unit vector d
in the direction of the fluid flow
∂x1 ∂x2 ∂x3
d= e1 + e2 + e3 (24)
∂s ∂s ∂s
and using Eq. (3), the directional gradient can be expressed as

∂h ∂h ∂x1 ∂h ∂x2 ∂h ∂x3

= ∇h · d = + + . (25)
∂s ∂x1 ∂s ∂x2 ∂s ∂x3 ∂s

The average velocity vs (20) becomes

Δ2N ρg
vs = ∇h · d, (26)
12 μ

and the average flow velocity vector vs = vs d is

Δ2N ρg
vs = (∇h · d) d. (27)
12 μ

Since the hydraulic discharge can be written as

ΔN Δ2N ρg
qs = n vs = d⊗d ∇h, (28)
L 12 μ

the fault permeability tensor kf corresponds to

ΔN Δ2N
kf = d ⊗ d. (29)
L 12
A Multiscale Microstructural Model of Permeability … 277

To account for a generic direction of the flow in the layer of normal N, in (29) the
tensor d ⊗ d must be replaced by the projection (I − N ⊗ N), leading to

ΔN Δ2N
kf = (I − N ⊗ N) . (30)
L 12
If Q fault families are present in the porous medium, each of which character-
ized by a normal NK , a separation L K and a normal opening displacement ΔKN , the
equivalent permeability is given by the sum of the corresponding permeabilities:

Q 2
ΔKN ΔKN  
kf = K
I − NK ⊗ NK . (31)
K=1
L 12

In the brittle damage model the permeability is described by an anisotropic tensor

variable from point to point.
The model does not exclude the presence of an initial permeability of the intact
matrix. If this is the case, the resulting permeability will be given by the sum of the
intact matrix and of the faults
k = km + kf . (32)

4 Examples

We want to study the response of the brittle damage model to the action of external
loadings and to analyze the correspondent variation of the permeability. For the sake
of simplicity, we assume that the permeability of the intact matrix is null, thus the
permeability will be exclusively related to the formations of the faults.
We specialize the strain energy density W to a standard neo-Hookean material
extended to the compressible range, i.e.,

1 1  
W (Fm ) = λ log2 J m + μ (FmT Fm ) : I − 3 − 2 log J m , (33)
2 2
where λ and μ are the Lamé coefficients, and J m = det Fm is the determinant of
Fm . Following [34], the cohesive energy on a fault with orientation N is assumed to
depend on an effective opening displacement Δ defined as

Δ= Δ · N)2 + β 2 |Δ
(1 − β 2 ) (Δ Δ |2 , (34)

Δ| is the norm of the opening displacement and β a material parameter

where |Δ
expressing the ratio between shear and tensile resistance of the material. In the first
loading, the cohesive law follows a linearly decreasing law, i.e.,
278 A. Pandolfi et al.

Fig. 3 Irreversible linear

decreasing cohesive law.
The enclosed area represents
the critical energy release
rate Gc

Tc Δ (1 − 0.5 Δ/Δc ) if Δ ≤ Δc
Δ, q) = Φ(Δ, q) =
Φ(Δ , (35)
Gc = 0.5 Tc Δc otherwise

where Gc is the critical energy release rate of the material, Tc is the tensile resis-
tance and Δc the maximum opening displacement associated to the presence of
cohesive tractions, see Fig. 3. Irreversibility is enforced by recording the maximum
ever attained effective opening displacement q = Δmax , and assuming unloading and
reloading to/from the origin, see Fig. 3, with the kinetic equation

Δ̇ if Δ = q and Δ̇ ≥ 0,
q̇ = (36)
0 otherwise.

The cohesive tractions follows as [33]

∂Φ T  
T= = Δ · N) N + β 2Δ ,
1 − β 2 (Δ (37)
Δ
∂Δ Δ
where
∂Φ  
T= = 1 − β −2 (T · N)2 + β −2 |T|2 . (38)
∂Δ
The material is characterized by the constants reported in Table 1. We assume an
intact material, with no pre-existent or natural faults, and limit our attention to the
constitutive response. We feed the material with an assigned deformation gradient,
whose significant components grow according to a prescribed history. The material is
allowed to form up to three families of faults, which cannot have the same orientation
NK .

Table 1 Rock material constants adopted in the examples

λ (MPa) μ (MPa) Tc (MPa) Gc (N/mm) β
2778 4167 10 0.01 1.0
A Multiscale Microstructural Model of Permeability … 279

(a) (b)
12 9E-08

10
K11
K22
8 6E-08 K33

4 3E-08

0 0
0 0.01 0.02 0.03 0 0.01 0.02 0.03

Fig. 4 a Uniaxial response σ3 versus ε3 of the material, stress in MPa. The drop in stiffness follows
the linearly decreasing law of the cohesive layers. b Variation of the material permeability, expressed
in m2 , in direction e1 , e2 , and e3 as a function of the deformation ε3 related to the opening of the
faults. After the formation of the faults, the matrix is unstressed

A first example is a simple uniaxial tensile test in direction e3 , with a maximum

stretch λ3 = l/l0 = 1.05, where l is the final length of the specimen and l0 the orig-
inal length. The only non zero component of the stress tensor is in the direction of
loading. The material begins behaving elastically, and fails at the attainment of the
tensile strength, originating a single family of faults normal to the loading direction.
Figure 4a shows the uniaxial response of the material in direction e3 , and Fig. 4b
shows the variation of the permeability tensor components k11 , k22 and k33 with the
strain in direction e3 . As expected from the orientation of the faults, the permeability
in direction e3 is always null. Note that, after the formation of the faults, the matrix
remains unstressed.
A second example is a multistage multiaxial test, that wants to mimic the variation
of stress and permeability in the field due to a fracking job. The material is initially
compressed isotropically by applying a uniform stretch λ1 = λ2 = λ3 = 0.99, to
induce a geostatic-like stress state. Then, the material undergoes an isotropic exten-
sion λ1 = λ2 = λ3 = 1.01, as it may happen in terms of effective stress when a
high-pressurized fluid is injected. Given the isotropy of the stress state, at the load-
ing corresponding to the material strength the material fails three times in tension,
creating in sequence three families of faults, with normal in direction e1 , e2 , and
e3 , respectively. The three families of faults differ because they are characterized by
different spacings. The failed material is still able to sustain an overall compressive
stress, since the interpenetration of the faults is controlled by a contact algorithm.
In a last stage of loading, the material is compressed again with an anisotropic
stretch. A λ1 = λ2 = 0.97 stretch is applied in direction e1 and e2 , while the orig-
inal geostatic-like stretch λ3 = 0.99 is applied in direction e3 . Figure 5a shows the
mechanical response of the model in direction e1 . The material initially undergoes
a compression (black circles). The following extension induces a tensile state that
280 A. Pandolfi et al.

(a) 20 (b) 3E-08

Compression
0 Extension
2E-08 Reclosure

-20

1E-08
-40 Compression
Extension
Reclosure

-60 0
-0.008 -0.004 0 0.004 0.008 0.012 -0.008 -0.004 0 0.004 0.008 0.012

(c) (d)
3E-08 3E-08

Compression Compression
Extension Extension
2E-08 Reclosure 2E-08 Reclosure

1E-08 1E-08

0 0
-0.008 -0.004 0 0.004 0.008 0.012 -0.008 -0.004 0 0.004 0.008 0.012

Fig. 5 Multi-stage multiaxial response of the material, undergoing an isotropic compression, fol-
lowed by an isotropic extension, and by a final anisotropic compression. a Stress σ1 (MPa) versus
strain ε1 . Although the material fails, generating three nested families of faults, the material pre-
serves its ability of sustaining load. b Permeability (m2 ) in direction e1 as a function of the strain.
c Permeability (m2 ) in direction e2 as a function of the strain. d Permeability (m2 ) in direction e3
as a function of the strain

reaches the strength of the material and causes triple tensile failure (open circles);
note that the three stress components show the behavior observed in the uniaxial
loading, cf. Fig. 4a. The final compressive stretch is characterized by a null stress
until faults close completely. Afterwards, the contact algorithm provides the com-
pressive tractions that guarantee the equilibrium of the system (grey circles). The
resulting reduction of the stiffness of the material due to the damage is remarkable.
Figures 5b–d show the permeability in direction e1 , e2 , and e3 , respectively. The
permeability is null until the material fails (black circles). Then the permeability
reaches a maximum corresponding to the maximum extension imposed to the mate-
rial (open circles). The different values of the maximum permeability for the three
directions is the combined result of the different spacing of the fault families and
of the stress anisotropy deriving from the formation of faults. Upon faults closure,
permeability decreases until it goes to zero (gray circles). Note that the anisotropy
A Multiscale Microstructural Model of Permeability … 281

of the compressive loading reflects in the anisotropy of the permeability history. In

particular, the permeability reduces more quickly in the direction e3 , where no extra-
confinement is applied. In fact, the over-compression in the two directions e1 and e2
closes the faults parallel to direction e3 , while the flow is still allowed in the faults
normal to e3 .

5 Remarks on the Porous Brittle Damage Model

and Possible Applications

The brittle damage model hereby introduced may be used in applications where
porous brittle materials experience a stress history leading to the formation of dam-
aged zones characterized by micro or macro cracks. In particular, the model has the
potential to be successfully employed in numerical simulations of localized stimu-
lation of oil or gas reservoirs by means of hydraulic fracture.
Brittle materials are sensitive and fail under tensile stresses or under non-isotropic
compressive stresses characterized by high shear components. The model here illus-
trated accounts for the progressive damage that can occur in brittle porous materials
failing under tensile or shear stress. As far as deep rock formations are concerned,
while the shear critical case can occur in natural situations under non-isotropic stress
states, the tensile critical state is more likely to happen under the action of a highly
pressurized fluid acting on a localized area, which changes the compressive stress
state in a tensile stress.
The mechanical features of the damage model have been described widely in the
original paper [33]. More importantly, unlikely standard damage theory models, the
brittle damage model is based on cohesive theories of fracture, therefore it describes
the formation of cohesive surfaces within the medium in the correct physical way.
Moreover, the model is characterized by multiple scales, represented by the L K
spacings between faults in each set, and is able to describe the microstructures that
can be observed in layered geologic media. The model is characterized variationally,
by introducing a generalized free energy density, so that stress and material tangent
stiffness are obtained analytically, and preserve a symmetric structure that makes
them appealing in numerical applications. Furthermore, inclusion of friction and
other forms of dissipative phenomena is easily achieved through the expedient of
introducing dual dissipation potential in the free energy density.
The porous version of the damage model retains the same mechanical features as
the original model. Thus, the microstructures resulting from the damage evolution
can be characterized with a permeability tensor, which is of fundamental importance
for the evaluation of the permeation and of the flow of fluids within the porous
material.
282 A. Pandolfi et al.

References

1. Bear, J. (1972). Dynamics of fluids in porous media. New York: American Elsevier Publishing
Company Inc.
2. Kozeny, J. (1927). Über kapillare leitung des wassers im boden. Sitzungsberichte der Kaiser-
lichen Akademie der Wissenschaften, 136(2a), 36.
3. Carman, P. C. (1937). Fluid flow through granular beds. Transactions, Institution of Chemical
Engineers, 15, 17.
4. Carman, P. C. (1956). Flow of gases through porous media. London.
5. Blair, S. C., Berge, P. A., & Berryman, J. G. (1996). Using two-point correlation functions
to characterize microgeometry and estimate permeabilities of sandstones and porous glass.
Journal of Geophysical Research - Solid Earth, 101, 20359–20375.
6. Göktepe, A. B., & Sezer, A. (2010). Effect of particle shape on density and permeability
of sands. Proceedings of the Institution of Civil Engineers: Geotechnical Engineering, 163,
307–320.
7. Sanzeni, A., Colleselli, F., & Grazioli, D. (2013). Specific surface and hydraulic conductivity
of fine-grained soils. Journal of Geotechnical and Geoenvironmental Engineering, 139, 1828–
1832.
8. Schaap, M. G., & Lebron, I. (2001). Using microscope observations of thin sections to estimate
soil permeability with the Kozeny-Carman equation. Journal of Hydrology, 251, 186–201.
9. Lasowska, A., Cies̀licki, K., & Smolarski, A. Z. (2000). The influence of channel tortuousity
on hydraulic resistance. Proceedings of the Institution of Civil Engineers: Geotechnical Engi-
neering, 48, 161–173.
10. Saripalli, K. P., Serne, R. J., Meyer, P. D., & McGrail, B. P. (2002). Prediction of diffusion
coefficients in porous media using tortuosity factors based on interfacial areas. Ground Water,
40, 346–352.
11. Archie, G. E. (2003). The electrical resistivity log as an aid in determining some reservoir
characteristicsn. In SPE reprint series (pp. 9–16).
12. Katz, A. J., & Thompson, A. H. (1987). Prediction of rock electrical conductivity from mercury
injection measurements. Journal of Geophysical Research, 92, 599–607.
13. Schwartz, L. M., Sen, P. N., & Johnson, D. L. (1989). Novel geometrical effects in electrolytic
conduction in porous media. Physica A: Statistical Mechanics and its Applications, 157, 493–
496.
14. Revil, A., & Cathles, L. M. (1999). Permeability of shaly sand. Physica A: Statistical Mechanics
and its Applications, 35, 651–662.
15. Cosentini, R. M., Della Vecchia, G., Foti, S., & Musso, G. (2012). Estimation of the hydraulic
parameters of unsaturated samples by electrical resistivity tomography. Géotechnique, 62(7),
583–594.
16. Civan, F. (2001). Scale effect on porosity and permeability: Kinetics, model, and correlation.
AIChE Journal, 47, 271–287.
17. Civan, F. (2002). Relating permeability to pore connectivity using a power-law flow unit equa-
tion. Petrophysics, 43, 457–476.
18. Schulze, O., Popp, T., & Kern, H. (2001). Development of damage and permeability in deform-
ing rock salt. Engineering Geology, 61(2), 163–180.
19. Mishra Parker, J. C., & Singhal, S. N. (1989). Estimation of soil hydraulic properties and their
uncertainty from particle size distribution data. Journal of Hydrology, 108, 1–18.
20. Pachepsky, Y. A., Timlin, D. J., & Ahuja, L. R. (1999). Estimating saturated soil hydraulic
conductivity using water retention data and neural networks. Soil Science, 164, 552–560.
21. Latief, F. D. E., & Fauzi, U. (2012). Kozeny-Carman and empirical formula for the permeability
of computer rock models. International Journal of Rock Mechanics and Mining Sciences, 50,
117–123.
22. Dardis, O., & McCloskey, J. (1998). Permeability porosity relationships from numerical sim-
ulations of fluid flow. Geophysical Research Letters, 25(9), 1471–1474.
A Multiscale Microstructural Model of Permeability … 283

23. Bird, M. B., Butler, S. L., Hawkes, C. D., & Kotzer, T. (2014). Numerical modeling of fluid
and electrical currents through geometries based on synchrotron X-ray tomographic images of
reservoir rocks using Avizo and COMSOL. Computers and Geosciences, 73, 6–16.
24. Aydin, A. (2000). Fractures, faults, and hydrocarbon entrapment, migration and flow. Marine
and Petroleum Geology, 17(7), 797–814.
25. Yuan, S. C., & Harrison, J. P. (2006). A review of the state of the art in modelling progressive
mechanical breakdown and associated fluid flow in intact heterogeneous rocks. International
Journal of Rock Mechanics and Mining Sciences, 43(7), 1001–1022.
26. Oda, M. (1985). Permeability tensor for discontinuous rock masses. Geotechnique, 35(4),
483–495.
27. Tang, C. A., Tham, L. G., Lee, P. K. K., Yang, T. H., & Li, L. C. (2002). Coupled analysis
of flow, stress and damage (fsd) in rock failure. International Journal of Rock Mechanics and
Mining Sciences, 39(4), 477–489.
28. Arson, C., & Gatmiri, B. (2008). On damage modelling in unsaturated clay rocks. Physics and
Chemistry of the Earth, Parts A/B/C, 33, S407–S415.
29. Arson, C., & Pereira, J.-M. (2013). Influence of damage on pore size distribution and perme-
ability of rocks. 37(8), 810–831.
30. Yuan, S. C., & Harrison, J. P. (2005). Development of a hydro-mechanical local degradation
approach and its application to modelling fluid flow during progressive fracturing of heteroge-
neous rocks. International Journal of Rock Mechanics and Mining Sciences, 42(7), 961–984.
31. Shao, J.-F., Zhou, H., & Chau, K. T. (2005). Coupling between anisotropic damage and perme-
ability variation in brittle rocks. International Journal for Numerical and Analytical Methods
in Geomechanics, 29(12), 1231–1247.
32. Maleki, K., & Pouya, A. (2010). Numerical simulation of damage- permeability relationship
in brittle geomaterials. Computers and Geosciences, 37(5), 619–628.
33. Pandolfi, A., Conti, S., & Ortiz, M. (2006). A recursive-faulting model of distributed damage
in confined brittle materials. Journal of Mechanics and Physics of Solids, 54, 1972–2003.
34. Ortiz, M., & Pandolfi, A. (1999). Finite-deformation irreversible cohesive elements for three-
dimensional crack propagation analysis. International Journal for Numerical Methods in Engi-
neering, 44, 1267–1282.
35. Pandolfi, A., & Ortiz, M. (2002). An efficient adaptive procedure for three-dimensional frag-
mentation simulations. Engineering with Computers, 18(2), 148–159.
36. Ortiz, M., & Stainier, L. (1999). The variational formulation of viscoplastic constitutive updates.
Computer Methods in Applied Mechanics and Engineering, 171, 419–444.
37. Pandolfi, A., Kane, C., Marsden, J. E., & Ortiz, M. (2002). Time-discretized variational for-
mulation of non- smooth frictional contact. International Journal for Numerical Methods in
Engineering, 53(8), 1801–1829.
38. Irmay, S. (1937). Flow of liquid through cracked media. Bulletin of the Research Council of
Istrael, 1(5A), 84.
39. Snow, D. T. (1965). A parallel plate model of fractured permeable media. PhD thesis, California:
University of California at Berkeley.
40. Snow, D. T. (1969). Anisotropic permeability of fractured media. Water Resources Research,
5(6), 1263–1279.
41. Parsons, R. C. (1966). Permeability of idealized fractured rocks. Journal of the Society of
Petroleum Engineers, 6, 126–136.
Thermal Diffusion in a Polymer Blend

Kerstin Weinberg, Stefan Schuß and Denis Anders

Abstract This contribution presents a thermodynamically sound approach to model

temperature sensitive diffusion in multi-phase solids. In order to describe the phenom-
ena of thermal diffusion (thermophoresis) and to simulate the effect numerically, an
extended version of the Cahn-Hilliard phase-field model is combined with the heat-
diffusion equation. The derived model is formulated consistently with the basic laws
of thermodynamics. Its discretized version is embedded in a NURBS-based finite
element framework. Numerical simulations and a comparison to experimental results
show the effect of thermal diffusion, induced by non-uniform and non-steady tem-
perature fields, on the microstructural evolution of a binary polymer blend consisting
of polydimethylsiloxane and polyethylmethylsiloxane.

1 Introduction

Composed materials, such as metallic alloys, solid solutions and multiphase plastics,
play an increasing role in industrial design. In particular, polymer blends became
a matter of interest in recent years. In order to combine beneficial characteristics
of single polymers, specific multiphase blends are composed. Such mixtures are
subjected to a great variety of microstructural changes such as separation of phases
and coarsening processes, cf. [6, 9, 13, 17, 22, 24, 25].
When a multicomponent system at a critical composition is quenched from an
initially homogeneous state, usually at high temperature, the blend destabilizes and
becomes susceptible to external perturbations. At a critical point the mixture decom-
poses into two phases of different composition—an effect which is known from

K. Weinberg (B) · S. Schuß · D. Anders

Lehrstuhl für Festkörpermechanik, Fakultät IV, Universität Siegen, 57068 Siegen, Germany
e-mail: [email protected]
S. Schuß
e-mail: [email protected]
D. Anders
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 285

K. Weinberg and A. Pandolfi (eds.), Innovative Numerical Approaches
for Multi-Field and Multi-Scale Problems, Lecture Notes in Applied
and Computational Mechanics 81, DOI 10.1007/978-3-319-39022-2_13
286 K. Weinberg et al.

Fig. 1 Illustration of the

free energy shape in an
stable and unstable scenario

metallic alloys and corresponds to two minima in the free energy function, see
Fig. 1. Polymer blends with a close-to-critical-point composition are very suscep-
tible to external fields such as convective, strain, electric, thermal or just random
fields. Such driven systems have been studied and simulated by the authors in [1, 3,
4]. External fields allow to control the microstructure evolution in the mixture.
In this text we will focus on the effect of thermal diffusion (Ludwig-Soret effect)
induced by non-uniform temperature gradients on the microstructural evolution in
a blend consisting of polydimethylsiloxane (PDMS) and polyethyl-methylsiloxane
(PEMS). Despite of a wide range of applications, comparatively little studies have
been performed for the phase behavior of such polymer blends subjected to non-
uniform temperature fields. Lee et al. worked on spinodal decomposition in the
presence of local temperature gradients in [19–21] and Köhler et al. published out-
standing studies on periodically driven and thermally patterned polymer mixtures in
[11, 25, 26]. The experimental setup to perform thermal patterning experiments is
explained in detail by Voit in [24].
Motivated by these publications we will present here a robust numerical scheme
to approximate the coupled heat-diffusion equations at hand by means of B-spline
based finite element analyses. To do so, we first derive the required thermal dif-
fusion equation and the resulting extended Cahn-Hilliard phase-field model. Then,
the numerical discretization where we use B-splines to exactly fulfill the continu-
ity requirements of the problem is presented. Finally, two- and three-dimensional
simulations of a thermal patterning experiment in a polymer blend are illustrated.

2 Thermal Diffusion and the Ludwig-Soret Effect

To derive the thermal diffusion model the classical Thermodynamics of Irreversible

Processes (TIP) in the sense of de Groot and Mazur [10] is employed. Point of depar-
ture is the entropy balance. In this context it is important to mention that a cornerstone
Thermal Diffusion in a Polymer Blend 287

of the classical TIP is the local equilibrium hypothesis, where thermodynamic state
variables in non-equilibrium states are considered to be the same as in equilibrium.
In the thermodynamic community it is still under debate that other variables, not
found at equilibrium, are able to influence non-equilibrium processes. For detailed
remarks on limitations of classical TIP the reader is referred to [18, pp. 63–65] or [8].
For thermal diffusion in solids and liquids, however, we have no reason to assume
additional state variables and so the TIP framework is fully acceptable.
At equilibrium the entropy per unit mass s is a well-defined function depending
on state variables such as the internal energy u, the specific volume ν and the mass
fractions ck of a multicomponent system with n components. In such a situation the
total differential of s in equilibrium is given by the common Gibbs relation
 
1 
n
ds = du + pdν − μk dck (1)
T k=1

where p is the equilibrium pressure, T the absolute temperature, μk the chemical

potential and ck the concentration of component k.

2.1 Balances

We follow the local equilibrium hypothesis and make use of the substantial time
derivative d (•) /dt = ∂ (•) /∂t + v · ∇ (•) with the barycentric velocity field v and
the spatial gradient ∇ to rewrite the balance (1) in the form

dν  dck
n
ds du
T = +p − μk . (2)
dt dt dt k=1
dt

We see here that Eq. (2) already includes the mass balance of the kth component by
dck /dt and the energy balance by du/dt and, thus, compliance with the basic laws
of thermodynamics is inherently guaranteed. Introducing the total mass density of
our multicomponent system by


n
ρk
ρ= ρk with ck = , (3)
k=1
ρ

the conservation of mass can also be expressed as

dck  r
ρ = − div Jk + vk j k̂ j , (k = 1, 2, . . . , n) , (4)
dt j=1
288 K. Weinberg et al.

where Jk represents the diffusive mass current, k̂ j is a chemical reaction rate of

reaction j and vk j denotes a parameter which is proportional to the stoichiometric
coefficient that weights the contribution of component k in the chemical reaction
j. In this way we also account for mass production due to chemical reactions; the
parameter r denotes their total number. Since mass is conserved in each separate
chemical reaction, it holds


n
vk j = 0, ( j = 1, 2, . . . , r ) . (5)
k=1

Now we need to deduce an appropriate formulation for the rate of the specific internal
energy du/dt implying the conservation of the total specific energy. To this end we
start with the equation of motion

dv  n
ρ = − div P + ρk Fk , (6)
dt k=1

where P is a microscopic generalized pressure tensor due to mechanical load and Fk

are vectorial contributions arising from external forces acting on the system. In this
study we restrict the discussion to the consideration of conservative forces, which
are derived from a stationary potential ψk with

∂ψk
Fk = −∇ψk , = 0. (7)
∂t
Multiplying Eq. (6) by the barycentric velocity field v and making use of the relation
div (Pv) = div P · v + P : ∇v, we obtain the balance equation for the kinetic
energy

d 21 v2 n
ρ = − div (Pv) + P : ∇v + ρk Fk · v. (8)
dt k=1

Accordingly, from the local form of the continuity mass equation dρ/dt = −ρ div v
we get for an arbitrary scalar field Φ the identity

dΦ ∂ (ρΦ)
ρ = + div (ρΦv) , (9)
dt ∂t
which is now used for a reformulation of Eq. (8) into
  n
∂ 21 ρv2 1 2
= − div ρv · v + Pv + P : ∇v + ρk Fk · v. (10)
∂t 2 k=1
Thermal Diffusion in a Polymer Blend 289

In the next step we reformulate Eq. (4) by means of identity (9) into

∂ρk  r
= − div (Jk + ρk v) + vk j k̂ j , (k = 1, 2, . . . , n) . (11)
∂t j=1

By combining Eqs. (11) and (7) with the differential formulae (see [16])

div (Jk ψk ) = ψk div Jk + Jk · ∇ψk

(12)
div (ψk ρk v) = ψk div (ρk v) + ρk v · ∇ψk

 establish an equation for the rate of change of the potential energy density ρψ ≡
we
k ρk ψk which is of the form
⎛ ⎞
∂ρψ 
n 
n 
n 
n 
r
= − div ⎝ρψv + ψk Jk ⎠ − ρk F k · v − J k · Fk + ψk vk j k̂ j .
∂t
k=1 k=1 k=1 k=1 j=1
(13)

Since we consider only chemical reactions where the potential energy is conserved,
for each reaction it holds


n
ψk vk j = 0, ( j = 1, 2, . . . , r ) , (14)
k=1

and the last term of Eq. (13) vanishes. As a result we obtain an equation for the rate
of change of the mechanical energy as a sum of kinetic energy 21 ρv2 and potential
energy ρψ,
   
∂ρ 1 2
v +ψ 1 2 n
2
= − div ρ v + ψ v + Pv + ψk Jk
∂t 2 k=1
(15)

n
+ P : ∇v − J k · Fk .
k=1

Clearly, since a source term appears at the right-hand side of the equation (underlined
in 15), the mechanical energy is not a conserved quantity.
According to the first law of thermodynamics (conservation of energy) the total
energy within an arbitrary control volume Ω in the system can only change due to
energy fluxes Je through the boundary ∂Ω:
  
d ∂ρe
ρe dΩ = dΩ = − Je · n dΩ. (16)
dt Ω Ω ∂t ∂Ω
290 K. Weinberg et al.

Here e is the total specific energy and n denotes the unit outward normal on ∂Ω. An
application of Gauss’ theorem provides the local form of energy conservation

∂ρe
= − div Je . (17)
∂t

In general the total energy flux Je includes a convective term ρev, an energy
flux Pv
due to mechanical work performed on the system, a potential energy flux k ψk Jk
due to diffusion and finally a heat flow Jθ


n
Je = ρev + Pv + ψk Jk + Jθ . (18)
k=1

If we keep in mind that the total specific energy is defined as the sum of specific
kinetic energy 21 v2 , the specific potential energy ψ and the specific internal energy
u, we can subtract Eq. (15) from Eq. (17) and use the formula for the energy flux (18)
to obtain the balance equation for the internal energy

∂ρu  n
= − div (ρuv + Jθ ) − P : ∇v + J k · Fk . (19)
∂t k=1

This equation shows that also the internal energy u is not a conserved quantity. Again
there is a source term, which is equal but of opposite sign to the source term in the
balance equation (15) of the mechanical energy. Therefore, such a formulation of the
balance equation for the internal energy inherently guarantees the conservation of
total energy. In our notation it is convenient to write Eq. (19) in an alternative form.
For this purpose we split the total pressure tensor P into a spherical/hydrostatic part
pI = tr (P) /3I and a deviatoric part S̄,

P = pI + S̄ (20)

where I is the identity tensor. With relations (20) and (9), Eq. (19) becomes

du  n
ρ = − div Jθ − p div v − S̄ : ∇v + J k · Fk . (21)
dt k=1

Here we have used that I : ∇v = tr (∇v) = div v. Another version of the mass
continuity equation in terms of the specific volume ν ≡ ρ −1

dν
ρ = div v (22)
dt
Thermal Diffusion in a Polymer Blend 291

is now used to formulate

du dν  n
= −ν div Jθ − p − ν S̄ : ∇v + ν J k · Fk . (23)
dt dt k=1

In order to find an explicit form of the entropy balance equation we have to insert
the expressions for du/dt (23) and dck /dt (4) into Eq. (2), which becomes

1  1  1 
n n r
ds 1 1
ρ = − div Jθ − S̄ : ∇v + J k · Fk + μk div Jk − A j k̂ j .
dt T T T T T
k=1 k=1 j=1
(24)

Here we have introduced the so-called chemical affinity A j of the j-th reaction
defined by


n
Aj = vk j μk , ( j = 1, 2, . . . , r ) . (25)
k=1

Now we intend to bring Eq. (24) into the typical structure of a balance equation

ds
ρ = − div Js + πs , (26)
dt
where Js is a general entropy flux and πs is an entropy source strength. According
to the second law of thermodynamics the entropy source πs vanishes for reversible
(or equilibrium) thermodynamic processes and it holds πs > 0 for irreversible ther-
modynamic transformations. Consequently, it must hold πs ≥ 0 for a general ther-
modynamic process.
By means of the relations (12) we obtain the entropy balance equation in the
required form
⎛ 
n ⎞
Jθ − μk Jk
ds ⎜ ⎟
ρ = − div ⎜
k=1 ⎟ − 1 Jθ ∇T
dt ⎝ T ⎠ T2
(27)
1 
n  μ   1 1 
r
k
− Jk · T ∇ − Fk − S̄ : ∇v − A j k̂ j .
T k=1
T T T j=1

From comparison with (26) it is possible to identify the entropy flux and the entropy
source term as
 
1 n
Js = Jθ − μk Jk , (28)
T k=1
292 K. Weinberg et al.

1 1 
n  μ   1 1 
r
k
πs = − J θ ∇T − J k · T ∇ − F k − S̄ : ∇v − A j k̂ j ≥ 0.
T2 T k=1 T T T j=1
(29)

At the first glance the applied separation into flux quantities and entropy source con-
tribution seems to be arbitrary, but Js and πs have to satisfy a number of requirements
which determine this separation uniquely; for a discussion we refer to [2].

2.2 Sources and Fluxes

Let us now have a closer look at the expressions for the entropy flux Js (28) and
the entropy source (29). Expression (28) indicates that the entropy flow consists
of a reduced heat flow Jθ /T and a current due to diffusion. The entropy source
expression (29) demonstrates that it can be divided into four contributions which all
are connected to a flow quantity. The first term of (29) arises from heat conduction
and is connected to heat flow Jθ , the second is a weighted mass diffusion flow Jk ,
the third one connects the momentum flow/viscous pressure S̄ to gradients of the
velocity field and the fourth term is a sum of chemical rates J j multiplied by their
affinities A j . In miscellaneous physical applications the terms in the entropy source
are classified into thermodynamic fluxes, and quantities which multiply the fluxes
are called thermodynamic forces or affinities.
In order to derive the model for thermal diffusion based on our entropy source
formulation (29), it is convenient to split off all thermodynamic forces proportional
to the temperature gradient which multiply the diffusive flux Jk . By means of the
thermodynamic relation
μ  1 hk
k
d = (dμk )T =const. − 2 dT, (30)
T T T
where h k := μk − T ∂μk /∂ T is the partial specific enthalpy of component k. Please
note that relation (30) takes into account that the chemical potentials μk depend on
a spatially non-uniform temperature field. Now we introduce a generalized heat flux
Jθ as

n
Jθ = Jθ − h k Jk (31)
k=1

which is conjugate to the temperature gradient to reformulate the entropy source

1  1 
n r
1  1
πs = − 2 Jθ ∇T − Jk · ((∇μk )T =const. − Fk ) − S̄ : ∇v − A j k̂ j .
T T k=1 T T j=1
(32)
Thermal Diffusion in a Polymer Blend 293


The generalized heat flux Jθ involves with the term k h k Jk a transfer of heat due
to diffusion. Therefore the quantity Jθ can be interpreted as an irreversible heat flow.
An alternative form of the conservation of mass accounts for the fact, that the sum
of all mass fluxes is zero,


n
Jk = 0. (33)
k=1

Thus, we can eliminate Jn from Eq. (32)

1 
r
1  1
πs = − J ∇T − S̄ : ∇v − A j k̂ j (34)
T2 θ T T j=1

1   
n−1
− Jk · (∇ (μk − μn ))T =const. − Fk + Fn .
T k=1

Now we will leave the general framework in order to reduce the general form of the
entropy source (34) to a multicomponent isotropic mixture where the concentrations
and the temperature are non-uniformly distributed over the system. In this text we
consider only isobaric, mechanically equilibrated systems where no external forces
and chemical reactions are supposed to be present. In this case the entropy source
reduces to the simple equation

1 
n−1
1 
πs = − J θ ∇T − Jk · [∇ (μk − μn )]T, p=const. . (35)
T2 T k=1

An application of the Gibbs-Duhem relation


n
ρk δμk = −ρsδT + δp (36)
k=1

for constant temperature and pressure provides


n
ρk ∇μk = 0. (37)
k=1

In this context δ (•) denotes the variation with respect to spatial coordinates. With
the help of Eq. (37) we can eliminate μn from (35). This gives

1 
n−1
1 
πs = − J θ ∇T − Jk · Akm [∇μm ]T, p=const. , (38)
T2 T k=1
294 K. Weinberg et al.

where the matrix components Akm are defined as

cm
Akm = δkm + , (k, m = 1, 2, . . . , n − 1) . (39)
cn

We regard πs as a linear combination of thermodynamic fluxes J• multiplied by their

corresponding affinities X• ,

n−1
πs = Jθ Xθ + J k Xk , (40)
k=1

and assume a linear dependency between the fluxes and affinities in form of:


n−1
J• = L •θ X• + L •k Xk . (41)
k=1

This leads to a quadratic expression for the entropy source strength. To guarantee the
positive definiteness of the entropy source strength πs it is therefore appropriate to
take the following choice of the phenomenological equations for the thermodynam-
ical fluxes Jθ and Ji :

1 
n−1
L θθ
Jθ = − ∇T − L θk Akm [∇μm ]T, p=const. , (42)
T2 T k,m=1

1 
n−1
L iθ
Ji = − 2 ∇T − L ik Akm [∇μm ]T, p=const. . (43)
T T k,m=1

Since we study here isotropic systems, the phenomenological coefficients L θθ and L ik

are scalars with symmetry in their indices as a consequence of Onsager’s reciprocal
relations. Due to the positive definiteness of the entropy source strength, additionally
it holds
1
L θθ ≥ 0, L ii ≥ 0, L ii L kk ≥ (L ik + L ki )2 . (44)
4

2.3 Mixtures of a Thermophobic and a Thermophilic

Component

Let us finally focus on the case of a binary mixture (n = 2). In this situation the
coefficients Akm reduce to

c1 1
A11 = 1 + = (45)
c2 c2
Thermal Diffusion in a Polymer Blend 295

and the phenomenological equations for the thermodynamic fluxes become

L θθ L θ1
Jθ = − ∇T − [∇μ1 ]T, p=const. , (46)
T2 T c2
L 1θ L 11
J1 = − 2 ∇T − [∇μ1 ]T, p=const. . (47)
T T c2

The coupling coefficients L 1θ characterize the phenomenon of thermal diffusion or

thermophoresis (Ludwig-Soret effect), where a mass diffusion current is caused by
temperature gradients. The reciprocal phenomenon, where a heat flow is caused by
concentration gradients, depends on the coupling coefficients L θi . This phenomenon
is referred to as Dufour effect. In detail we introduce instead of the phenomenological
coefficients in (46) and (47) the following set of coefficients:

L θθ
Λ= thermal conductivity (48)
T2
L θ1
Dθ = Dufour coefficient (49)
ρc1 c2 T 2
L 1θ
DT = thermal diffusion coefficient (50)
ρc1 c2 T 2
L 11
M= mobility coefficient (51)
ρT c2

The Onsager reciprocal relations imply equality between the thermal mass diffusivity
DT and the Dufour coefficient Dθ which leads to thermodynamical fluxes in the form:

Jθ = −Λ∇T − ρ1 Dθ T [∇μ]T, p=const. , (52)

J1 = J = −ρ DT c (1 − c) ∇T − ρ M [∇μ]T, p=const. . (53)

For clarity we made here the choice c1 := c , μ1 := μ and therefore c2 := 1 −

c. The indices T, p = const. will be omitted. Equations (52) and (53) enable us
now to study diffusion phenomena which arise in a binary mixture where both the
temperature and the concentration are non-uniform within the system. Furthermore
if the concentration gradients are moderate we may consider the overall density ρ as
roughly uniform.
If the temperature gradient plays the dominant role within the irreversible heat
flux Jθ and the contribution from the Dufour cross-phenomenon can be neglected, it
is possible to derive an approximation of Eq. (52).
Let us now thoroughly elaborate on the diffusion flow J. The concept of thermal
diffusion expresses that different particle types move differently under a temperature
gradient. Therefore the applied temperature gradient induces a diffusive mass flux
JT . In the introduced notation JT takes the form

JT = −ρ DT c (1 − c) ∇T. (54)
296 K. Weinberg et al.

In a single phase it is not possible to achieve thermal diffusion, since JT vanishes for
c = 0 and c = 1. Usually the thermally activated diffusive mass current may occur
in either direction, dependent on the materials involved. Thermophilic substances
diffuse in the direction of the temperature gradient. Thermophobic materials diffuse
in the direction opposite to the temperature gradient. Typically the heavier/larger
species in a mixture exhibits a thermophobic behavior while the lighter/smaller
species exhibit thermophilic behavior. In addition to the sizes of the various types of
particles and the steepness of the temperature gradient, the heat conductivity and heat
absorption of the particles play a significant role in thermal diffusion. However JT
leads to a buildup of a concentration gradient, which is accompanied by a generalized
Fickean type mass diffusion current,

J D = −ρM∇μ. (55)

Consequently, the entire diffuse mass flux is then J = J D +JT . Note that the mobility
coefficient M may be replaced by a tensor valued mobility to account for anisotropic
effects.
At this point it should be clear that the Ludwig-Soret effect can be used in various
technical applications to influence the microstructure of materials. Thermal diffusion
is a powerful tool in pharmacology to discover and design new types of drugs [5]. As
well it is employed as a technique for manipulating single biological macromolecules
such as DNA in polymer micro- and nanochannels, cf. [23, 27].

3 Phase Decomposition and Coarsening

in a PDMS-PEMS Blend

The evolution of multiphase polymer mixtures can be deduced from a variation

of a thermodynamic energy functional with respect to a set of order parameters
φ = [φ1 , φ2 , . . . , φm ] under the constraint of mass conservation. In general the
energy functional has the following structure

kB T
E (φ) = Ψ con (φ, T ) + Ψ int (φ, T ) dx, (56)
v Ω

m κi (T )
where Ψ con is the configurational energy density and Ψ int (φ, T ) = i=1 2
∇φi 2
characterizes the contribution of interfacial energy. The temperature dependent para-
meters κi are related to surface energy density γi and length li of the transition regions
between the domains of each phase. In this context T is the temperature of the sys-
tem, k B denotes the Boltzmann constant and v characterizes a corresponding unit
volume. In (56) the free energy density Ψ = Ψ con + Ψ int is a dimensionless quantity
which is multiplied by the coefficient k B T /v to obtain an energy density J/m3 .
Thermal Diffusion in a Polymer Blend 297

The corresponding unit volume v comprises an average volume that a monomer with
Kuhn length (persistence length) l undergoes within the scenario of a random walk.
In the following we consider specifically a PDMS-PEMS mixture. Both compo-
nents are polymeric organosilica with a wide range of technical applications, such as
material for contact lenses, adhesives, coating implementations, silicone based lubri-
cants, and even as a food additive in defoaming agents. The structural formulas of
PDMS and PEMS show their similar molecular structure. Both are terminated with
trimethylsiloxane endgroups, one methyl respectively ethyl group in the repeating
unit makes the only difference.
⎡ ⎤
CH3 CH3 CH3
⎢ ⎥
⎢ ⎥
CH3 Si O⎢
⎢ Si O⎥
⎥ Si CH3 (PDMS)
⎣ ⎦
CH3 CH3 CH3
n

⎡ ⎤
CH3 CH3 CH3
⎢ ⎥
⎢ ⎥
CH3 Si O⎢
⎢ Si O⎥
⎥ Si CH3 (PEMS)
⎣ ⎦
CH3 C2 H5 CH3
n

We choose the mass concentration field c as order parameter, whereby it simply

holds cPDMS = 1 − cPEMS =: c. The densities of PDMS 0.969 g/cm3 and PEMS
0.977 g/cm3 are very similar, so that mass and volume fraction are basically the
same. According to Flory-Huggins thermodynamics of mixing [12, 15], for a binary
polymer blend the configurational energy density can be written as

c (1 − c)
Ψ con (c) = g A c + g B (1 − c) + ln (c) + ln (1 − c) + χ c (1 − c) .
NPDMS NPEMS
(57)

Here NPDMS and NPEMS represent the degrees of polymerization, and χ is a temper-
ature dependent material parameter characterizing the chemical interaction between
the constituents of the mixture. The value of χ is usually approximated by a rela-
tion of the form χ = a + bT −1 , where a and b are experimentally obtained fitting
parameters. The terms g A c and g B (1 − c) quantify the free energy of the individual
components.
In order to investigate the evolution of the concentration field c (x, t) and the
corresponding temperature field T (x, t) in a representative domain Ω within the
time t¯ we make use of the diffusion equation, which expresses the conservation of
mass in the system:
∂c
= −∇ · j + ξ (x, t) . (58)
∂t
298 K. Weinberg et al.

The flux j is given by flux (53) divided by the average density ρ, ξ (x, t) is a random
variable arising from thermal fluctuations. The chemical potential is related to Ψ =
Ψ con + Ψ int by the variational derivative:

μ = δc Ψ = ∂c Ψ con − ∇ · ∂∇c Ψ int = ∂c Ψ con − κΔc. (59)

Using this relation, Eq. (58) results in a modified Cahn-Hilliard equation ([7], cf. [2])
coupled with a heat diffusion equation with specific heat capacity c p , conductivity
  k
and heat source q. Thus,
  in the strong form the problem reads: Find c : Ω × 0, t¯ →
R and T : Ω × 0, t¯ → R such that

∂c
= ∇ · (M∇μ + DT c(1 − c)∇T ) , in Ω × [0, t¯] (60)
∂t
∂T
cpρ = ∇ · (k∇T ) + q, in Ω × [0, t¯]. (61)
∂t
Initial concentration and temperature are given by

c(x, 0) = c0 (x), T (x, 0) = T0 (x) in Ω, (62)

along with the boundary conditions

(M∇μ + DT c(1 − c)∇T ) · n = 0 on ∂Ω × [0, t¯],

∇c · n = 0 on ∂Ω × [0, t¯],
(63)
T = T̃ on ∂Ωe × [0, t¯],
∇T · n = 0 on ∂Ωn × [0, t¯],

with properties ∂Ω = ∂Ωe ∪ ∂Ωn and ∂Ωe ∩ ∂Ωn = ∅. Vector n denotes the unit
outward normal to ∂Ω. The specific heat capacity c p and thermal conductivity k are
assumed to have the form

c p (c, T ) = c PPDMS (T ) + (1 − c)PPEMS (T ),

(64)
k(c, T ) = cQ PDMS (T ) + (1 − c)Q PEMS (T ),

where PPDMS , PPEMS , Q PDMS and Q PEMS are polynomial functions of degree 3
which fit the specific heat capacity and thermal conductivity of the two species
for T ∈ [50 K, 340 K] and T ∈ [230 K, 410 K], respectively. We use an Onsager
coefficient of M = 1.815 × 10−18 m5 / (J s) and a thermal diffusivity of DT =
−2 × 10−13 m2 / (s K). Additional parameters are summarized in Table 1. The mate-
rial data as well as the experimental results are obtained from [17, 24–26].
Since PDMS is a thermophilic polymer its thermal diffusion coefficient has a
negative sign. The corresponding unit volume is assumed as v = 4π (σ A /2)3 /3.
The critical temperature of the considered polymer blend is Tcrit. = 313 K and the
critical interaction parameter χcrit. evaluates to χcrit. = 0.0084318. At this point it is
Thermal Diffusion in a Polymer Blend 299

Table 1 Material parameters for PDMS (A) and PEMS (B)

N A (–) N B (–) a (–) b (K) κ σ A (nm) σ B (nm)
219.4 257.25 −1.86 × 10−3 3.22 9× 10−9 0.583 0.64

Fig. 2 Shapes of the

configurational energy
density at different
temperatures

important to mention that the chosen value for a = −1.8557 × 10−3 significantly
differs from the value given in [25]. There the authors suggest a = 2.9× 10−3 , which
is obviously erroneous, because it does not fit χcrit. . The approximation procedure
for the values of the interaction parameter χ prescribes negative values for a and
positive values for b in such a manner that it holds
⎧
⎪
⎨χcrit. , for T = Tcrit.
χ (T ) = χ (T ) < χcrit. , for T > Tcrit. (65)
⎪
⎩
χ (T ) > χcrit. , for T < Tcrit. .

Our value for a perfectly fits these conditions. The corresponding plot of the config-
urational energy density is presented in Fig. 2.

4 Numerical Approximation

For finite element analysis we reformulate our coupled diffusion model in a vari-
ational form. Note that the mass diffusion equation involves spatial derivatives of
fourth order. Thus, we define the spaces of admissible test functions V c = {δc ∈
H 2 (Ω) | ∇δc · n = 0 on ∂Ω × (0, t˜)} and V T = {δT ∈ H 1 (Ω) | δT =
0 on ∂Ωe × (0, t˜)} where H 1 (Ω) and H 2 (Ω) are the Sobolev space of square
integrable functions with square integrable derivatives of first and of second order,
respectively. The weak forms of the boundary-value problems (60) and (61) follows
as
300 K. Weinberg et al.

Fig. 3 Two-dimensional B-spline of order p = 2

  
∂T
ρ cp δT dV + k∇T · ∇δT dV − α I δT dV = 0,
∂t
Ω Ω Ω
  
∂c
δc dV + M ∇∂c Ψ · ∇δc dV + DT c(1 − c)∇T · ∇δc dV
con
(66)
∂t
Ω Ω Ω

+ λM ΔcΔδc dV = 0,
Ω

for all δc ∈ V c , δT ∈ V T . Clearly, the variational formulation of the problem

requires approximation functions which are piecewise smooth and globally C 1 -
continuous. For this reason we decide to employ B-splines as finite element basis
(Fig. 3).
A multivariate B-spline basis of degree p = [ p1 , . . . , pd ] and dimension d ∈ N
is defined by the tensor product Θ1 ⊗ · · · ⊗ Θd of knot vectors, built by a sequence
of knots Θl = [ξ1l ≤ ξ2l ≤ · · · ≤ ξnl l + pl +1 ], l ∈ {1, . . . , d}. In the absence of
repeated knots, the partition [ξi11 , ξi11 +1 ] × · · · × [ξidd , ξidd +1 ] forms an element of the
mesh in the parametric domain. A single multivariate B-spline B A , A ∈ [1, . . . , n],
n := n1 . . . nd , is then defined by


d
B A = Bpi (ξ ) = Bpi (ξ 1 , . . . , ξ d ) = Nil , pl (ξ l ), (67)
l=1

with multi-index i=[i 1 , . . . , i d ] and supp(B A ) = [ξi11 , ξi11 + p1 +1 ]×· · ·×[ξidd , ξidd + pd +1 ].
It provides the necessary support for the required continuity. The recursive definition
of a univariate B-spline is given as follows

ξ − ξill ξill + pl +1 − ξ
Nil , pl (ξ ) = Nil , pl −1 (ξ ) + Nil +1, pl −1 (ξ ), (68)
ξill + pl − ξill ξill + pl +1 − ξill +1

starting with piecewise constant functions

1 if ξill ≤ ξ < ξill +1

Nil ,0 (ξ ) = . (69)
0 otherwise
Thermal Diffusion in a Polymer Blend 301

Linear independence as a fundamental property of finite element basis as well as

local support are given by a B-spline basis. Moreover, smoothness is related to
knot multiplicity, i.e., the number of repetitions in Θ at node i. Unfortunately, the
tensor product structure in (67) impedes standard local refinement strategies which
motivated us to introduce a specific hierarchical refinement strategy in [14].
For temporal discretization the considered time interval [0, t¯] is divided into n t
subintervals. The first order time derivative is approximated by finite differences

∂c cn+1 − cn
= (70)
∂t Δt
with time step Δt = tn+1 −tn . The time integration is performed by an implicit Crank-
Nicholson scheme, known to be second-order accurate. Now the fully discretized
problem reads
  
ρ
c p,n+1/2 (Tn+1 − Tn )B A dV + kn+1/2 ∇Tn+1/2 · ∇ B A dV − α I B A dV = 0,
Δt
Ω Ω Ω
 
1
(cn+1 − cn )B A dV + M (∂c2 Ψn+1/2
con ∇c A
n+1/2 + ∂c ∂T ∇Tn+1/2 ) · ∇ B dV
Δt
Ω Ω
 
+ DT cn+1/2 (1 − cn+1/2 )∇Tn+1/2 · ∇ B A dV + λM Δcn+1/2 ΔB A dV = 0,
Ω Ω
(71)
for all A ∈ {1, . . . , n}.

5 Simulation Results

In our numerical simulations phase separation induced by laser light absorption is

studied. The setting for simulation is arranged in such a manner that the polymer
blend is homogeneously quenched from the one-phase regime and simultaneously is
heated. In order to integrate a focused laser spot into the model, a time independent
heat source of the form q(x) = α I (x) with intensity
∞
    
iπ x1 jπ x2
I (x) = I0 (x) ai j sin sin (72)
i, j=1
l x1 l x2

is used. The laser spot points at the middle of the upper surface of an aged cuboid
of side length l x = l y = 1.5 µm and hight 0.75 µm, see Fig. 4, with the diameter of
the spot being about 3/5 of the domain size in the two-dimensional and 1/5 in the
tree-dimensional model. The laser intensity and the optical absorption coefficient are
given by I0 = 3.75 × 107 W m−2 and α = 500 m−1 , respectively. In both settings
we consider a critical PDMS-PEMS blend with a composition of 55 % PDMS and
302 K. Weinberg et al.

Fig. 4 Temperature field with laser intensity (72) in K (left) and computational mesh (right) for
the three-dimensional simulation

(a) (b) (c)

Fig. 5 Micrographs of a PDMS-PEMS polymer blend exposed to a focused laser spot. The specimen
size is about 100 µm × 100 µm. Subfigures are adapted from [24]

45 % PEMS. Therefore, the initial concentration is set c0 (x) = 0.55 including slight
randomly perturbed inhomegeneities (±1 %).
Köhler et al. studied the microstructural evolution of a PDMS-PEMS blend sub-
jected to a focused laser spot in [24, 25]. In their experimental set-up the blend
under consideration was initially quenched into the spinodal regime for approxi-
mately 120 min before exposing it to the laser, see Fig. 5a. At this moment the blend
has already completed phase decomposition and reached an intermediate state of
coarsening. The recorded experimental observations indicate the following scenario.
Starting from a homogeneous one-phase configuration the early stages of microstruc-
tural evolution are driven by spinodal decomposition of phases. After the sample is
exposed to laser illumination, thermal forcing asserts itself gradually against the sup-
pression of phase decomposition. Over time the inhomogeneous temperature field
increasingly coins the microstructure. Figure 5b, c show the sample after an exposure
of 100 and 200 s, respectively. Due to the local heating, a circle of PDMS evolves in
the middle of the micrograph. Since a PDMS enrichment takes place at the heated
domain within the sample an outer ring of PEMS-rich material surrounds the concen-
Thermal Diffusion in a Polymer Blend 303

tric PDMS circle. The structures slowly propagate through the sample like spherical
waves in a medium. Outside the heated spot a more irregular spinodal pattern unfolds
due to the decreasing impact of thermal diffusion, see Fig. 5 and [24, 25].
The simulation setting is also arranged in such a manner that the polymer blend
is at first homogeneously quenched from the one-phase regime, see Fig. 6. Here
and below the reddish areas denote the PDMS-rich α-phase and the blue domains
represent the PEMS-rich β-phase. Heating with the laser spot starts at a system time
t = 0.03 s. Similar to the experimental observations a spherical PDMS-rich phase
emerges in the center of the domain, where the laser induces an enrichment of the
thermophilic PDMS. The PDMS is attracted from the surrounding area, and—like
in the experiments—a second ring of a PEMS-rich phase evolves around the heated
spot. It seems that the PEMS-ring spreads like a spherical wave starting at the spot
in the middle of the domain Ω. In the outer part of Ω the thermal effect decreases
and more irregular spinodal pattern of phases become prevalent. Note that since
the relaxation into thermal equilibrium occurs on a time scale that is by a factor
of 1000 faster then the time scale at which diffusion takes place, we simplified the
two-dimensional model by assuming quasistatic equilibrium for the heat equation.
As we continue in time, a typical coarsening process dominates the evolution of
phases outside the laser spot. The microstructure within the exposure of the laser spot
is characterized by a PDMS-rich concentric ring structure, which grows towards the
center of the laser illumination, see Fig. 7. After the PDMS enrichment within the

Fig. 6 Phase decomposition and early stages of coarsening for a critical PDMS-PEMS polymer
blend subjected to a focused laser spot
304 K. Weinberg et al.

Fig. 7 Phase coarsening of the PDMS-PEMS polymer blend subjected to a focused laser spot (red
PDMS, blue PEMS)

heated spot is finished and the PEMS material disappeared from the center of the
spot at t = 6 s, the phase coarsening process takes the leading role in microstructural
evolution. In the very late stages additional PDMS material from the surroundings
is deposited at the PDMS circle.
In order to integrate the focused laser spot into the three-dimensional simulation,
a laser intensity I0 (x) = 3.2 × 1011 W m−2 is focused on a domain Ω̄ which is
defined by
!  2  2 "
3 3 3 29 3
Ω̄ := x | x1 − µm + x2 − µm ≤ µm, µm ≤ x3 ≤ µm .
4 4 20 40 4

Outside the domain Ω̄ the laser intensity is zero.

In order to locally refine the computational mesh around the domain Ω̄, a hierar-
chical refinement scheme based on B-spline subdivision is used, see [14] for details.
The resulting mesh is shown in Fig. 4. Periodic boundary conditions on the left and
right, and front and rear surfaces of the computational domain are applied.
At the beginning of the simulation the laser intensity I0 (x) is set for approximately
2.5 s to zero so that a typical coarsening scenario can take place, see the pictures in the
upper row of Fig. 8. Starting at time t = 2.4691 s the middle of the upper surface, i.e.,
the region Ω̄, is heated by the laser spot. As can be seen in the second row of Fig. 8,
Thermal Diffusion in a Polymer Blend 305

Fig. 8 Decomposition, phase coarsening and thermophoresis in a cuboid of critical PDMS-PEMS

polymer blend, (red PDMS, blue PEMS). First row phase evolution without heat supply at start
t = 0 and t = 1.75 s, t = 2.4691 s; second and third row results with a focused laser spot at times
t = [2.4697 s, 2.4703 s, 2.4709 s, 2.4715 s, 2.4721 s, 2.4727 s]; fourth row profile of the heated
region at times t = 2.4697 s, t = 2.4715 s and t = 2.4727 s

a ring of PDMS forms around the heated region. The subsequent pictures show an
enrichment of PDMS in the heated center whereas the microstructural evolution in the
remaining region is mainly driven by spinodal decomposition of phases. The last row
of Fig. 8 illustrates the distribution across the depth where we see that thermophoresis
mainly takes place at the heated surface, whereas the decomposition of phases in the
remaining domain is a slower effect.
306 K. Weinberg et al.

6 Conclusions

In this contribution we have presented an extended version of the commonly known

Cahn-Hilliard phase-field model in order to capture diffusion phenomena induced by
local non-uniform temperature gradients. Our diffusion model was formulated con-
sistently with the basic laws of thermodynamics. Its discrete version was embedded
into the isogeometric finite element concept in order to perform numerical simu-
lations. The simulations were compared to experimental studies of PDMS-PEMS
blends by the application of realistic material parameters.

Acknowledgments The authors gratefully acknowledge the support of the Deutsche Forschungs-
gemeinschaft (DFG) under the grants WE2525/2-3, WE2525/4-1 and WE2525/8-1.

References

1. Anders, D., & Weinberg, K. (2011). A variational approach to the decomposition of unstable
viscous fluids and its consistent numerical approximation. ZAMM—Zeitschrift für angewandte
Mathematik und Mechanik, 91(8), 609–629.
2. Anders, D., & Weinberg, W. (2012). Thermophoresis in binary blends. Mechanics of Materials,
47, 33–50.
3. Anders, D., Reichert, R., & Weinberg, K. (2011). Isogeometric analysis of thermal diffusion
in binary blends. Computational Materials Science, 52(1), 182–188.
4. Anders, D., Hoffmann, A., Scheffler, H.-P., & Weinberg, K. (2011). Application of operator-
scaling anisotropic random fields to binary mixtures. Philosophical Magazine, 91(29), 3766–
3792.
5. Baaske, P., Wienken, C. J., Reineck, P., Duhr, S., & Braun, D. (2010). Optical thermophoresis
for quantifying the buffer dependence of aptamer binding. Angewandte Chemie International
Edition, 49, 2238–2241.
6. Baumgärtner, A., & Heermann, D. W. (1986). Spinodal decomposition of polymer films. Poly-
mer, 27(11), 1777–1780.
7. Cahn, J. W. (1961). On spinodal decomposition. Acta Metallurgica, 9(9), 795–801.
8. Cimmelli, V. A., Jou, D., Ruggeri, T., & Ván, P. (2014). Entropy principle and recent results
in non-equilibrium theories. Entropy, 16(3), 1756.
9. de Gennes, P. G. (1980). Dynamics of fluctuations and spinodal decomposition in polymer
blends. Journal of Chemical Physics, 72(9), 4756–4763.
10. de Groot, S. R., & Mazur, P. (1962). Non-equilibrium thermodynamics. Amsterdam: North-
Holland.
11. Enge, W., & Köhler, W. (2004). Thermal diffusion in a critical polymer blend. Physical Chem-
istry Chemical Physics, 6, 2373–2378.
12. Flory, P. J. (1942). Thermodynamics of high polymer solutions. Journal of Chemical Physics,
10(1), 51–61.
13. Hashimoto, T., Kumaki, J., & Kawai, H. (1983). Time-resolved light scattering studies on
kinetics of phase separation and phase dissolution of polymer blends. 1. Kinetics of phase
separation of a binary mixture of polystyrene and poly(vinyl methyl ether). Macromolecules,
16(4), 641–648.
14. Hesch, C., Schuß, S., Dittmann, M., Franke, M., & Weinberg, K. (2016). Isogeometric analysis
and hierarchical refinement for higher-order phase-field models. Computer Methods in Applied
Mechanics and Engineering, 303, 185–207.
Thermal Diffusion in a Polymer Blend 307

15. Huggins, M. L. (1942). Theory of solutions of high polymers. Journal of the American Chemical
Society, 64(7), 1712–1719.
16. Itskov, M. (2007). Tensor algebra and tensor analysis for engineers (with applications to
continuum mechanics). Berlin: Springer.
17. Krekhov, A. P., & Kramer, L. (2004). Phase separation in the presence of spatially periodic
forcing. Physical Review E, 70, 061801.
18. Lebon, G., Jou, D., & Casas-Vázquez, J. (2007). Understanding non-equilibrium thermody-
namics. In Foundations, Applications, Frontiers. Springer: Berlin.
19. Lee, K.-W. D., Chan, P. K., & Feng, X. (2002). A computational study of the thermal-induced
phase separation phenomenon in polymer solutions under a temperature gradient. Macromole-
cular Theory and Simulations, 11, 996–1005.
20. Lee, K.-W. D., Chan, P. K., & Feng, X. (2003). A computational study of the polymerization-
induced phase separation phenomenon in polymer solutions under a temperature gradient.
Macromolecular Theory and Simulations, 12(6), 413–424.
21. Lee, K.-W. D., Chan, P. K., & Feng, X. (2004). Morphology development and characterization of
the phase-separated structure resulting from the thermal-induced phase separation phenomenon
in polymer solutions under a temperature gradient. Chemical Engineering Science, 59(7), 1491–
1504.
22. Strobl, G. R. (1985). Structure evolution during spinodal decomposition of polymer blends.
Macromolecules, 18(3), 558–563.
23. Thamdrup, L. H., Larsen, N. B., & Kristensen, A. (2010). Light-induced local heating for
thermophoretic manipulation of DNA in polymer micro- and nanochannels. Nano Letters,
10(2), 826–832.
24. Voit, A. (2007). Photothermische Strukturierung binärer Polymermischungen. Ph.D. thesis,
University of Bayreuth.
25. Voit, A., Krekhov, A., Enge, W., Kramer, L., & Köhler, W. (2005). Thermal patterning of a
critical polymer blend. Physical Review Letters, 94, 214501.
26. Voit, A., Krekhov, A., & Köhler, W. (2007). Quenching a UCST polymer blend into phase
separation by local heating. Macromolecules, 40, 9–11.
27. Würger, A. (2007). Das Salz in der DNA-Suppe. Physik Journal, 7(22–24), 2014.