Comput Mech
DOI 10.1007/s00466-014-1002-8
ORIGINAL PAPER
Isogeometric analysis of shear bands
Luc Berger-Vergiat · Colin McAuliffe · Haim Waisman
Received: 11 December 2013 / Accepted: 16 February 2014
© Springer-Verlag Berlin Heidelberg 2014
Abstract Numerical modeling of shear bands present sev-
eral challenges, primarily due to strain softening, strong non-
linear multiphysics coupling, and steep solution gradients
with fine solution features. In general it is not known a pri-
ori where a shear band will form or propagate, thus adaptive
refinement is sometimes necessary to increase the resolution
near the band. In this work we explore the use of isoge-
ometric analysis for shear band problems by constructing
and testing several combinations of NURBS elements for a
mixed finite element shear band formulation. Owing to the
higher order continuity of the NURBS basis, fine solution
features such as shear bands can be resolved accurately and
efficiently without adaptive refinement. The results are com-
pared to a mixed element formulation with linear functions
for displacement and temperature and Pian–Sumihara shape
functions for stress. We find that an element based on high
order NURBS functions for displacement, temperature and
stress, combined with gauss point sampling of the plastic
strain leads to attractive results in terms of rate of conver-
gence, accuracy and cpu time. This element is implemented
with a B-bar strain projection method and is shown to be
nearly locking free.
Keywords Shear bands · Isogeometric analysis · NURBS ·
B-bar · Volumetric locking
1 Introduction
Shear bands are highly localized zones of intense plastic
deformation which can form in materials suffering high strain
L. Berger-Vergiat · C. McAuliffe · H. Waisman (B)
Department of Civil Engineering & Engineering Mechanics,
Columbia University, New York, NY 10027, USA
e-mail: waisman@civil.columbia.edu
rate loading. Experimentally derived material models for
these loading regimes describe plastic flow as being depen-
dent on temperature, strain rate, and a hardening parameter
[1]. While several models are available, all are similar in that
increasing temperature (due to plastic work) has a soften-
ing effect, causing plastic flow to occur more readily, while
increases in strain rate and the hardening parameter have a
hardening effect. Following the experimental work of [2],
shear bands develop in three stages. In Stage 1, before local-
ization, a homogeneous distribution of plastic strain exists.
Stage 2 begins when the thermal softening effect dominates
the strain and strain rate hardening effects, resulting in strain
softening, and thus strain localization is initiated. Stage 3 is
marked by severe localization and rapid softening, a phenom-
ena termed stress collapse, which indicates a sudden and large
drop in the material’s load bearing capability [3]. Shear bands
can be modeled by a system of partial differential equations
(PDEs) describing conservation of momentum, conservation
of energy, elastic and inelastic constitutive relations [1].
Previous work by McAuliffe and Waisman [4] made use of
a mixed finite element formulations to simultaneously solve
for these four equations with the velocities, temperatures,
stresses, and equivalent plastic strains each independently
interpolated degrees of freedom. The implicit nonlinear con-
sistent (INC), or monolithic solver combined with thermal
diffusion achieved results that were insensitive to mesh size,
and capable of achieving significantly higher accuracy than
a split explicit scheme for a given amount of computational
effort. Mesh alignment sensitivity of the formulation was
also tested in [4]. This is the tendency of the mesh to be more
compliant along element edges during shear band propaga-
tion (first studied by [5]). It was found that the INC solver
improves, but does not completely eliminate the alignment
sensitivity. The shape functions employed in that work were
low order, being bilinear for velocity and temperature, the
123

Pian–Sumihara interpolations for stress [6] and gauss point
sampling was done for equivalent plastic field.
An alternative to low order FE for shear band modeling
is the meshfree family of methods, which has been stud-
ied by [7,8]. A meshfree approach can easily handle large
deformations and was shown in [7] to suppress mesh align-
ment sensitivity. Another approach is the multiscale tech-
nique called the multiresolution method, proposed by Liu
and McVeigh [9,10] and has also been used to model shear
bands [11–13]. The multiresolution method can be under-
stood as a local refinement technique that may employ differ-
ent governing equations in different local regions. However,
an embedded length scale parameter which can be related to
size, morphology of microvoids and the width of propagating
shear bands must also be assumed. The technique has been
used for adiabatic shear bands and shear bands assisted by
micro-voids. Since in practice, it is generally not known a
priori where a shear band will initiate or propagate, adaptive
local refinement techniques are likely unavoidable for large
problems. However, for the same reason, it is also desirable to
develop methods where relatively coarse meshes are capable
of furnishing good levels of accuracy. Isogemetric analysis
(IGA) is particularly well suited for this purpose, since it
employs higher order shape functions capable of efficiently
resolving the small scale solution features associated with
shear bands.
IGA [14,15] was originally developed with the intent of
streamlining the simulation and design process. This process
typically starts with discrete geometrical description of the
part or system to be analyzed with a Computer Aided Design
(CAD) package, which employ the NURBS basis func-
tions [16,17] or a T-splines basis function [18,19]. Once the
geometry is fully described it is passed on to the simulation
group, which in turn rediscretize the geometry according to
the needs of the simulation technique chosen for analysis.
IGA eliminates this rediscretization step from the simulation
work flow by using the same basis functions for both the
CAD geometrical description and analysis phases. Another
added benefit of this approach is the exact geometry used
during the analysis.
Despite the original intent, perhaps the most interesting
feature of IGA is the property of the NURBS shape functions
when used for simulation. Basis functions constructed with
NURBS are globally continuous across elements, not only
within them. This means that a solution computed with IGA
is higher order and not piecewise higher order as would be
the case with the p-ver sion of the finite elements method.
Moreover, it is also easy to raise the polynomial order of
these NURBS basis [20] which lead to more accurate sim-
ulations [21] with roughly as many nodes as for a simula-
tion with a low order basis. This last property is due to the
fact that IGA allows for a new type of refinement named
k-r e f inement, an attractive feature which raises the order
123
Comput Mech
of the shape functions without excessively increasing the size
of the linear system to be solved. Nonetheless, it should be
noted that these matrices are also much denser than stan-
dard FEM matrices since the NURBS basis is not compactly
supported.
In addition, the IGA framework retains the ability to
employ the B-bar method [22,23] which is used to reme-
diate the volumetric locking problems which appear during
the simulation of nearly incompressible deformation.
The aim of our work is to reduce the computational cost
of shear band simulations by finding shape functions which
furnish the greatest accuracy for the lowest cost. To this end
we tested multiple elements, with a focus on IGA based ele-
ments. The main evaluation method of these discretizations
is study of the convergence rates, volumetric locking and
memory usage. This paper begins by presenting the IGA for-
mulation for the mixed finite elements, then introduces the
PDE model we chose to employ for shear bands. Finally, we
study two test cases and assess the convergence properties of
the proposed elements. The formulation is shown to be free
of both mesh size and mesh alignment sensitivity, and leads
to fast rates of convergence.
2 NURBS-based IGA for mixed formulations
In this section we review the basics of NURBS functions and
how they are used in the paradigm of IGA [14,15]. More
specifically, we will focus on how mixed formulations are
discretized in IGA.
NURBS are functions designed to exactly represent curves
and were originally created for computational geometry pur-
poses. Detailed description of these functions, their proper-
ties, and how to compute them can be found in Piegl and
Tiller [24] and Rogers [25].
NURBS are parametrized curves in d-dimensional spaces
(Rd for example), constructed as a tensorial product of d
1D-NURBS curves. 1D-NURBS are themselves constructed
as a rational combination of B-splines. In order to further
describe B-splines functions we need to introduce the para-
metric space used to construct them.
The parametric space associated with a B-spline function
is called a knot vector and is composed of non-decreasing
real numbers (the knots)  = {ξ1 , ξ2 , . . . , ξn+ p+1 }. By con-
vention, the knot vector is scaled so that it spans the unit
subspace [0, 1] ⊂ R, where R is the space of all real num-
bers. The multiplicity of knots cannot exceed the degree of
the curve p + 1 where p is also the degree of the polynomial
basis used to construct the B-spline. n is the number of basis
functions defined on the knot vector, it is also the number of
control points used to describe the geometry of the B-spline.
A knot vector is said to be open if its first and last knots
have multiplicity p + 1. Open knot vectors are commonly
Comput Mech

1
N1,2 N6,2 N8,2

N3,2 N4,2
N2,2 N5,2

N7,2

0
0 1 2 3 4 5

Fig. 1 On the left a quadratic basis of NURBS functions is represented. On the right the corresponding curve (in blue) is obtained where red circles
represent the control points. (Color figure online)

used since they produce bases that are interpolatory at the These shape functions are combined with the control points
endpoints of the interval. B-spline basis of order p are con- to construct NURBS surfaces as follow
structed recursively from an underlying piecewise constant

n 
m
basis S(ξ, η) = Ri j Bi j . (6)
 i=1 j=1
1 if ξi ≤ ξ ≤ ξi+1 ,
Ni (ξ ) = (1)
0 otherwise. An example of a basis of NURBS functions and its asso-
ciated curve are presented in Fig. 1.
Bases of order p are then constructed using the following
Note how three shape functions are non zero over three
recursion formula
elements instead of two as is the case for Lagrange or Hermite
ξ −ξi ξi+ p+1 −ξ polynomials. It can also be observed that the curve is not
Ni, p (ξ ) = Ni, p−1 (ξ ) + Ni+1, p−1 (ξ )
ξi+ p −ξi ξi+ p+1 −ξi+1 interpolatory at all points. It generally is interpolatory only
(2) at the boundary.
The NURBS basis forms a partition of unity which makes
first introduced by Cox and de Boor [20].
it easy to use as a shape function basis for the finite element
The associated B-spline curve is obtained using a linear
method (FEM). NURBS k-r e f inement preserves the num-
combination of the basis function as follows
ber and geometry of the elements on the mesh for any order

n of shape functions, which allows us to use NURBS of differ-
C(ξ ) = Pi Ni (ξ ), (3) ent order in a mixed formulation. These properties will also
i=1 be used to compute different basis of shape functions for the
B-bar projection technique [22,26,27].
where Pi ∈ Rd is the ith control point.
The B-splines can then be extended to multiple dimen- 3 The mechanics and numerics of shear bands
sions using a tensorial product, in the case of d = 2, we get
the following expression for a B-spline surface In this section we introduce the governing set of equations
that describe the onset and propagation of shear bands. We

n 
m
S(ξ, η) = Ni (ξ )M j (η)Pi j , (4) then discuss the numerical discretization of the system of
i=1 j=1 PDEs. We emphasize that element formulations based on
NURBS has to be done with caution since it will strongly
where M j (η) are basis functions defined by the following influence the numerical performance of the simulations,
knot vector H = {η1 , . . . , ηm+q+1 }. which is the focus of this paper.
Finally, NURBS basis functions are created using a ratio-
nal combination of 1D B-spline basis functions, here again 3.1 Problem statement
for d = 2, the result is
In this work we employ a similar shear band formulation as
Ni (ξ )M j (η)wi j the one proposed by McAuliffe and Waisman [4]. For ther-
Ri j (ξ, η) = n m . (5)
i=1 j=1 Ni (ξ )M j (η)wi j mally driven shear band models, localization occurs when

123
 Comput Mech

thermal softening overcomes strain and strain rate harden- Following standard mixed formulation procedures, the
ing. Since this model permits an overall strain softening in the weak form is obtained by multiplying each equation in the
material, regularization is needed to avoid spurious mesh size strong form by its weight function and integrating over the
sensitivity. To this end, we employ diffusive regularization, domain, applying the divergence theorem where appropri-
where the competition between shear heating and thermal dif- ate. These equations can be regrouped into a solution vector
fusion weakly defines a length scale which control the width u with displacements, temperature, stress and plastic strain
of the shear band. This can be compared with the work that fields and a residual vector R which will be later used to
has been done for nonlocal damage [28] and for gradient elas- derive the jacobian of the system
ticity [29], plasticity [30] and damage [31]. The formulation
⎡ ⎤ ⎡ ⎤
uses an analytical derivation of the consistent tangent stiff- u Ru
ness, which is used in a fully monolithic Newton solver. This ⎢T⎥ ⎢ RT ⎥
u=⎢
⎣σ ⎦
⎥ R=⎢
⎣ Rσ ⎦ .
⎥ (10)
avoids error propagation problems which commonly arise in
operator split (or staggered) nonlinear solution schemes. γ̂p Rγ̂ p
Assuming that no body forces are applied, the governing
system of PDEs can be expressed as follows The expression for the terms Ru , RT , Rσ and Rγ̂ p can be
⎧ found in Appendix.
⎪ ρ ü = ∇ · σ
⎪
⎪
⎪
⎪

⎨ ρc Ṫ = κ∇ · ∇T + χ σ̂ g(σ̂ , T, γ̂ p ) 3.2 Linearization and discretization of the problem

  in 
⎪
⎪ σ̇ = elas : ∇ s u̇ − 3 g(σ̂ ,T,γ̂ p ) S − α Ṫ I
⎪
⎪
C 2 σ̂ In the next subsections the weak form is descritized in time
⎪
⎩
˙
γ̄ p = g(σ̂ , T, γ̂ p ) and space and consistent linearization is employed. This
(7) results in an efficient implicit, monolithic solver for the shear
with the following boundary conditions band problem presented in Sect. 3.1 that is insensitive to mesh
size.
u = u 0 (t) on u
3.2.1 Time discretization and linearization
T = T0 (t) on T
(8)
σ · n = t 0 (t) on t
The weak form is integrated in time using the Newmark
q · n = q0 (t) on q method [26,34], assuming an implicit scheme. Newton’s
method is then used to solve the resulting nonlinear system
where subscript 0 indicate initial conditions and tensorial
of equations for a given time step. To this end, we linearize
quantities are underlined according to their order. ∇· is the
the residual by computing the variation of R with respect to
divergence operator, ∇ s is the symmetric part of the gradient
the state variables u, T , σ and γ̂ p as follow
operator given by ∇ s = 21 (∇ + ∇ T ) and : is the double con-

traction operator. g is the plastic flow rule, σ̂ = 23 S : S is Jδu = R, (11)
the effective stress (assuming J2 plasticity), γ̂ p is the equiv-
alent plastic strain defined by the plastic flow law and S is where J is the jacobian (or tangent stiffness) of the problem
 
tr (σ )
the deviatoric stress tensor S = σ − 3 I . Finally the and δu is the incremental solution computed at each Newton
iteration. The jacobian is derived using variational calculus
modified Litonsky plastic flow rule which has also been used
and more precisely a Gâteaux derivative [35,36]
in previous work [4,32,33] is employed here. This law is
defined as d
Jδu = R(u + εδu). (12)
g(σ̂ , T, γ̂ p ) dε
⎡ ⎤m
σ̂ The details of the derivation will not be carried out in this
⎢ ⎥
= γ̇r e f ⎣       ⎦ , paper but we refer the reader to [4] for a more detailed expla-
γ̂ p N T −Tr e f
σr e f 1 + γr e f 1−δ exp k −1 nation of this point. The block structure of the final Jacobian
(9) matrix is given by
⎡ ⎤
with γ̇r e f , γr e f , Tr e f and σr e f are material parameters denot- Mu 0 Ku 0
⎢ 0 MT + KT + G T T GT σ GT γ̂ p ⎥
ing the reference plastic strain, plastic strain rate, temperature J=⎢ ⎣ Kσ
⎥,
⎦
and stress, respectively. The material parameters m and N are Gσ T Mσ + Gσ σ Gσ γ̂ p
the rate sensitivity and strain hardening exponent, respec- 0 Gγ̂ p T Gγ̂ p σ Mγ̂ p + Gγ̂ p γ̂ p
tively, and δ and k are thermal softening parameters. (13)

123
Comput Mech
where the terms called M∗ can be interpreted as mass matri-
ces, the terms K∗ can be interpreted as stiffness terms and the
terms G∗∗ are terms arising from the derivation of the plastic
flow rule g(σ̂ , T, γ̂ p ). Note that this final expression of the
Jacobian is non symmetric which will restrict the choice of
available solvers later on. The expressions of the terms con-
stituting the jacobian matrix can be found in the Appendix.
3.2.2 Spatial discretization
The last step is to discretize the problem in space by choosing
appropriate shape functions for the solution fields. In a gen-
eral framework, different shape functions are used for each
field so that
u = Nu u wu = Nu wu ,
T = NT T wT = NT wT ,
(14)
σ = Nσ σ wσ = Nσ wσ ,
γ̂ p = Nγ̂ p γ̂ p wγ̂ p = Nγ̂ p wγ̂ p ,
where N∗ is in practice stored as 1D arrays of shape functions
using the Voigt notation and the assembly process of the finite
element method is applied to get a global vector. At this point,
at least two different type of limitations are going to guide
our choice of shape functions.
We note that spatial derivatives of the displacement and
temperature fields appear in our formulation, therefore the
shape functions Nu and NT must be at least linear. In addition,
no derivatives of stress and equivalent plastic strain (EQPS)
appear in the weak form, hence Nσ and Nγ̂ p are only required
to be piecewise constants.
Second, to obtain a stable element, we need to ver-
ify that our mixed formulation based on the Hellinger–
Reissner variational formulation (displacement/stress for-
mulation) [37,38] satisfies the Babuška–Brezzi (BB) con-
dition [39–41]. The stress/displacement formulation (in the
case of linear elasticity for the sake of clarity) is com-
posed of the momentum equilibrium and the constitutive
relation. This formulation can be compared to the displace-
ment/pressure formulation typically used for incompressibil-
ity problems [42], which comprises the momentum equation
and the definition of pressure in a solid. Both of these for-
mulations lead to a saddle point problem, as presented in
Table 1.
We see that the stress in the Hellinger–Reissner formula-
tion plays a similar role as the displacement in the incom-
pressible formulation. More examples and details about these
classic mixed formulations for elasticity as well as detailed
explanation of the Babuška–Brezzi condition can be found
in [43].
To develop an element for shear bands that will be stable
and will not lock under certain loading conditions, it should
pass the inf-sup criteria [44]. Various methods exist in the
Table 1 comparison of mixed formulations leading to saddle point
problems with λ and μ the Lamé constants and p the pressure in the
solid
Stress/displcement Displacement/pressure

σ − C elas : ∇ s u = 0 −∇ · (− p I + 2μ∇u) = f
Mixed form
∇ · u = λp
−∇ · σ = b
         
A BT σ 0 A BT u f
Algebraic form = =
B 0 u g B 0 p 0
Appropriate boundary conditions are assumed for both formulations
literature to assess the inf-sup condition, for example, one
can analyze numerically the eigenvalues of the assembled
Jacobian, as proposed by Bathe et al. [45,46] and determine if
an element will satisfy the inf-sup condition. While such tests
have been shown to work well on some simple problems such
as Stokes flow, acoustic fluids and the bending of Reissner–
Mindlin plates, it is not trivial to extend these techniques
to complex, time dependent, multiphysics problems, as the
current shear band problem.
One engineering way to determine if an element might
pass the inf-sup criteria is via a constraints count approach,
discussed in Hughes [26] and Zienkiewicz and Taylor [34].
We emphasize that the constraints count method is a nec-
essary but not sufficient approach to pass the inf-sup con-
dition, hence even if an element passes the constraint count
method, it doesn’t guarantee it will be stable and locking free.
Nonetheless, it is an easy test to construct and allows us to
quickly filter out elements that would not be stable in any
case.
n
The idea is to construct a ratio r = neqc with n eq the num-
ber of equilibrium equations and n c the number of constraints
due to the constitutive relation used, such that r ≈ 1, would
yield a stable, non locking element. If the ratio is approach-
ing 1 from bellow, r → 1− , then the element may be subject
to stresses oscillations, which can lead to an unstable ele-
ment [34]. While if it is approaching 1 from above, r → 1+ ,
then the element may experience severe locking.
For 2D mixed formulations, given in the left column of
Table 1, we have three independent stresses σx x , σ yy and σx y
as well as two independent displacements u x and u y where at
least three displacements degrees of freedom should be fixed
to avoid rigid body motion. These considerations lead us to
inequality (15), which has to be satisfied in order to have a
stable element
3n σ − 2n u + 3 ≥ 0, (15)
where n σ and n u are the number of stress and displacement
degrees of freedom in the problem.
Now we can express the number of degrees of freedom
of each field as a function of the polynomial orders Nσ and
123
 Comput Mech

Table 2 Choices of shape

functions using the constraints Order of r -Ratio Associated Roots Smallest acceptable
count method Pu /Pσ polynomial value for n

3(n+1)2
n/n 2(n+1)2 −3
n 2 + 2n + 4 imaginar y 0
3n 2
√
n/(n-1) 2(n+1)2 −3
n 2 − 4n + 1 2± 3 4
3(n−1)2
√
n/(n-2) 2(n+1)2 −3
n2 − 10n + 4 3 ± 21 10
3(n−2)2
√
n/(n-3) 2(n+1)2 −3
n2 − 16n + 13 8 ± 51 16

ening may occur due to under resolved solution gradients, this

issue can always be resolved by mesh refinement, whereas
volumetric locking cannot.

3.3 Near incompressibility and the B-bar method

Verifying that elements pass the BB condition is a necessity

Fig. 2 Schematic representation of the P 4 P 3 element with displace-
ment and temperature nodes represented by red circles and stress and
equivalent plastic strain nodes represented by green squares. (Color
figure online)
Nu used for the shape functions Nσ and Nu to obtain the
final inequality that relates the shape function order with the
constraint count method
3Nσ2 − 2Nu2 + 3 ≥ 0. (16)
The different choices of Pu /Pσ couples and the optimal order
of suggested shape functions, based on inequality (16), are
summarized in Table 2. It appears that the most reasonable
choice for a mixed formulation would be a P n /P n (e.g.
P 1 /P 1 or P 2 /P 2 ) element or a P 4 /P 3 element, that have
good r -ratio values but also relatively low shape functions
order. Other elements with higher order shape functions are
also a possible but might lead to large bands in the jacobian
matrix and not necessarily greater computational efficiency.
An illustration of the P 4 /P 3 NURBS element is shown
in Fig. 2. Note that even though the number of nodes in this
element is high, on larger meshes most nodes are shared
with adjacents elements, thus in practice the total number of
unknowns is comparable to that of a linear element.
The constraint count technique as described in this section
will be used to assess the suitability of more complex ele-
ments, specifically derived for shear bands, that involve addi-
tional solution fields: temperature and plastic strains. Since
shear banding involves large, near incompressible deforma-
tions, a suitable element for shear bands must be free of
locking. Otherwise, the model will be artificially too stiff,
and the shear band too diffuse. While a similar artificial stiff-
in order to have a stable element, but this does not mean
that the element has any good convergence properties. For
shear bands, volumetric locking is a severe risk in elements
since the shear banding process in metallic materials is mod-
eled with J2 plasticity which leads to nearly isochoric defor-
mations in the solid. Modeling this type of deformation is
not simple if one wants to avoid volumetric locking, there-
fore herein we implemented the volumetric strain projection
technique called B-bar [22,23,26]. The idea of the B-bar
method is to reduce the amount of strain energy in the solid
by projecting the volumetric strain onto a lower order basis
function. This technique can also be interpreted as a selective
reduced order integration [26].
The projected strain is discretized as a field depending
on N̄u , where N̄u is a shape function used for u and T but
with one order lower continuity than the “full order” shape
function basis Nu , so that
ε̄(x) = N̄u (x)α, (17)
where α is the vector of lower order stains, in practice this
vector is not computed. At the same time, the “full order”
strain is expressed as a function of the nodal displacements
u and the “full order” shape function basis Nu
ε(x) = ∇ s Nu (x)u. (18)
Thus, the relation between α and u can be obtained by enforc-
ing the equality of the strain and lower order strain in a weak
sense using the lower order shape function basis. This is done
by solving the following least square problem
 
N̄u (x)ε̄(x)d = N̄u (x)ε(x)d, (19)
 

123
Comput Mech
 
Table 3 Nomenclature of elements implemented in this study Ku = ∇wu δσ d → K̄u = ∇ w̄u δσ d, (24)
Name Abbreviation  
  
Pian–Sumihara Shear band Quad PSSQ Kσ = − wσ Celas ∇ s δ u̇ d
Mixed formulation NURBS Shear band Quad MNSQ

Hybrid NURBS Shear band Quad 1 HNSQ1   
Hybrid NURBS Shear band Quad 2 HNSQ2 → K̄σ = − wσ Celas ∇ s δ ū˙ d. (25)
Irreducible NURBS Shear band Quad INSQ 

These names may be followed by a number in parentheses which indi-

cate the order of the shape functions used (i.e. HNSQ2(3) would indi-
3.4 Mixed elements investigated
cate that the HNSQ2 element is employed with cubic NURBS shape
functions) Taking into account the difficulties explained in the previous
sections:
which leads, after replacing ε̄ and ε by their respective expres-
i. balancing the number of degrees of freedom for stresses
sions, to the following expression for α
and displacements in order to get a good r -ratio,
 ii. alleviating volumetric locking,
α = M̄−1 N̄u (x) ∇ s Nu (x)u d, (20) iii. avoiding spurious oscillations in the equivalent plastic
 strain field,

Note that a lower order mass matrix is employed in this we implement and investigate the behavior of multiple ele-
process, which is given by ments, grouped in three main categories:

M̄ = N̄u (x)N̄u (x) d, (21) • Pian and Sumihara type elements
• Mixed formulation based elements

• Irreducible elements
and the final projected strain is
The characteristics of each element in their respective fam-
⎛ ⎞ ily are presented in the following paragraphs, where Table 3

ε̄(x) = N̄u (x) ⎝M̄−1 N̄u (x) ∇ s Nu (x)u d⎠ . (22) summarizes the nomenclature and Table 4 the principal fea-
tures of each elements. In the following sections we describe

each family of methods in more details.
Next, the volumetric part of the strain is replaced by the pro-
jected volumetric strain, which is simply done by redefining 3.4.1 Pian–Sumihara type elements
the strain as follow [22]
The Pian–Sumihara (PSSQ) based element is an element
tr (ε(x)) − tr (ε̄(x)) designed to improve the convergence rate of the finite ele-
ε(x) = ε(x) − I (23) ment method when the Hellinger–Reissner variational form
3
is used. For the 2D case, σx y is constant and σx x and σ yy
In our formulation only two blocks of the jacobian matrix are modeled with 2 degrees of freedom accounting for the
contain strain terms: Ku and Kσ , hence we introduce the constant and linear part of these stresses. For more details,
projected strain in these terms by the following update see [6,47–49]. This formulation ensures that the element has

Table 4 Summary of the tested

elements, the r -ratio is Name Shape functions r -Ratio B-bar Oscillations in Convergence
calculated for the shape function the EQPS field
order that yields the most stable Nu NT Nσ Nγ p
element
PSSQ P1 P1 PS I 1 no no yes
MNSQ Pn Pn P(n−1) P(n−1) 1.021 yes yes no
HNSQ1 Pn Pn I Pn 1.596 yes yes no
HNSQ2 Pn Pn Pn I 1.596 yes no yes
INSQ Pn Pn I I 1.596 yes no yes

123
 Comput Mech

a perfectly balanced ratio of constraints: r -ratio=1. This ele- of continuity and regularity. To this end, we attempted to
ment has also no apparent problem of volumetric locking, implement a hybrid NURBS-irreducible element (HNSQ1),
which makes it an excellent element for this shear band prob- that uses NURBS shape functions to discretize the EQPS.
lem [4]. However, one significant disadvantage of the PSSQ However, this element is prone to oscillations of the EQPS
element is that it is solely based on a low order element (bilin- field that are severe enough to prevent it from converging on
ear element). Hence extensions to higher order elements, and example 2 from Sect. 4.2. On the other hand, reformulation
in particular NURBS functions, is not trivial. of HNSQ1 using the NURBS shape functions to discretize
the stress field instead of the EQPS leads to a stable element
3.4.2 Mixed formulation NURBS elements (HNSQ2) that is not subject to oscillations of the EQPS field.
Thus, in this paper we will present convergence results focus-
This family of elements is based on a standard mixed formu- ing on this element.
lation element (MNSQ) [50] devised by applying the con- One future direction to further improve the HNSQ2 ele-
straint count technique to determine which orders of shape ment could be based on a discontinuous Galerkin type formu-
functions for each field is the most likely to yield a stable ele- lation. Such formulation might be able to benefit from higher
ment. This element type takes full advantage of the NURBS order shape functions for all fields including the EQPS field
shape functions for each field and so high order convergence and remain stable.
is expected. Preliminary research in 1D confirmed this good
performance, but with 2D elements, the outlook becomes 3.5 Implementation of IGA in a finite element code
more complex. Volumetric locking is a concern, and even
after implementation of the B-bar method, using low order In this section the main modifications needed to implement
shape functions leads to slow convergence. However higher IGA in a finite element code are presented. The modifica-
order elements does indeed alleviate some of these issues and tions presented apply for single patch simulations. In the
was shown to give good results on the examples presented in case of multiple patches a connectivity and assembly of the
Sect. 4.1. patches is required. More details on the computational cost
However, a clear disadvantage is the fact that using lin- and the memory usage for the different elements studied are
ear and higher order shape functions for the equivalent plastic presented in Sect. 3.4.
strain field leads to severe oscillations when solving the prob-
lem presented in Sect. 4.2. These oscillations are so severe
that the Newton solver diverges when the gradient of the Algorithm 1 IGA-FE code layout
EQPS field becomes large. Due to this instability, the mixed 1: subroutine main()
2: call preprocessing (read input material parameters, control points,
formulation elements are thus only used for the more simple knot vectors, B-bar flag, elements type flag)
tension loading cases. 3: for t = 0, T do Adaptive time stepping loop
4: while  R > tol·  R0  do Newton solver loop
3.4.3 Irreducible elements 5: for e = 1, number of elements do Assembly loop
6: get NURBS shape functions N Eq. (5)
7: compute element residual R Appendix
In order to obtain an acceptable r -ratio, an irreducible ele- 8: if B-bar then
ment (INSQ), using the approach described by Comi et 9: compute B-bar terms for Jacobian Eqs. (24)-(25)
al. [51], was implemented for shear bands by McAuliffe et 10: end if
11: compute element Jacobian J Appendix
al. [52]. The basic idea is to treat the stresses and the equiva- 12: end for
lent plastic strain as history variables computed at the gauss 13: solve linearized system: δu = J−1 R
points and then to assemble the contribution of each gauss 14: end while
point to the global system in order to maintain a fully mono- 15: update solution: u = u + δu
16: call postprocessing (Algorithm 2 and Figure 4), to post process
lithic solver. This approach converges on all the test cases values from control points to physical domain.
and also simplifies the implementation since it only requires 17: end for
to compute shape functions for the displacement and tem-
perature fields.
Nonetheless, introducing history variables in the dis-
cretization also has its drawbacks. One probelm is the mem- 3.5.1 IGA based finite element algorithm
ory usage required for the sampling at the gauss points as
explained in Sect. 3.5.2. Another problem is that history vari- Most FE codes follow the layout presented in Algorithm 1.
ables do not provide the same rate of convergence as the high The first step in an IGA implementation is to generate/read
order NURBS fields. This is due to the fact that the interpo- input parameters used in simulations, which specifically to
lation for history variables does not have NURBS properties IGA include: control points coordinates, knot vectors and

123
Comput Mech
(a) (b)
(c) (d)
(e) (f)
Fig. 3 Schematic representation of the studied elements. Red circles
represent displacement and temperature degrees of freedom. Green
squares represent stresses and EQPS for the MNSQ element and stress
for the PSSQ element. + represents the EQPS degrees of freedoms
for PSSQ and stress and EQPS for INSQ. a Single quadratic MNSQ
element. b Patch of four quadratic MNSQ elements. c Single PSSQ
element. d Patch of four PSSQ elements. e Single quadratic INSQ ele-
ment. f Patch of four quadratic INSQ elements. (Color figure online)
other relevant information related to the mesh. Then, mem-
ory is allocated for the necessary arrays as presented in Algo-
rithm 1. This task may become more complex when B-bar is
used since it requires a lower order mesh to be generated as
well. Book keeping also needs to be planed carefully to keep
track of the nodes introduced at the Gauss points in the case
of elements HNSQ1, HNSQ2 and INSQ (Fig. 3).
The shape functions used in the residual and jacobian cal-
culations are computed using special procedures in order
to obtain the NURBS shape functions and their derivatives.
First, B-spline functions are computed using the algorithm
in [24] and then NURBS shape functions are obtained from
these B-spline functions via the methods outlined in [14]. It
is important to note that the solution of the system of equa-
tions is obtained in terms of quantities associated with control
points, Gauss points. Thus, first an L 2 -projection is employed
to transfer all the quantities to control points. This process is
illustrated in Fig. 4 by going from the mesh shown in Fig. 4a
to the mesh with only control points in Fig. 4b. Then, an inter-
polation step is required to transfer all quantities from control
points, illustrated by Fig. 4b, to nodes of the actual physical
mesh, illustrated by Fig. 4c, d, for results output. Note that
for plotting purposes it is common to resolve solution fea-
tures (e.g. shearbands) on the physical mesh by subdividing
the original mesh into a locally refined mesh. Algorithm 2
describes this post processing step.
For example, consider a quantity λgp computed at the
Gauss points that we want to express as
λgp = Nu λ, (26)
where N are the NURBS shape functions and α is the vector of
values we wish to obtain at the control points. The following
least square problem is then solved to obtain α

Mλ = Nu λgp d, (27)

where M = NuT Nu d is the mass matrix associated with
the specific NURBS shape functions (note that M is different
from M̄ used for B-bar method).
Finally, the solutions (displacement, temperature, stresses
and equivalent plastic strain) at the control points are interpo-
lated to the refined physical mesh by a standard interpolation
employing NURBS functions at any point x j
u(x j ) = Nu (x j )u. (28)
3.5.2 Memory and computational cost considerations
The computational and memory costs associated with the
Isogeometric elements are mainly influenced by the num-
ber of variables introduced in the linear system of equations
solved at each Newton iteration. The coupling between differ-
ent nodes are also leading to larger bandwidth of the jacobian
that results in increasing computations and memory usage. In
addition, some post-processing operations also require addi-
tional computational work, however, it is usually negligible
compared to the cost of the linear solver.
We compare the relative cost of k-r e f inement used for
the stresses in element HNSQ2 and gauss point sampling in
element INSQ. In the example given in Fig. 5, for a 4×4
mesh, the number of nodes in the case of NURBS elements
is 49. However, in the case of gauss point sampling the num-
ber of nodes in the mesh is 144. Denoting n as the number
123
 Comput Mech

Fig. 4 Projection step from

gauss points (a) to control 11111 11111
points (b) and evaluation of the
solution on a locally refined
physical mesh (c, d). Red circles 0.75 0.75
represents the values solved at
the control points (a, b) and

η
0.5

η
0.5
plotted values on the physical
domain (c, d). Green times
0.25 0.25
represents the values solved at
the Gauss points. ξ and η are
paremetric variables spanning 00000 00000
the knot vectors. a Solution at
00000 0.25 0.5 0.75 11111 00000 0.25 0.5 0.75 11111
Gauss and control points.
b Solution at control points ξ ξ
(Gauss point values were
(a) (b)
projected using Eq. (27)).
c Solution on the physical mesh
for the top traction example.
d Solution on the physical mesh 10.0 10.0
for the top shear example.
(Color figure online)
7.5 7.5

y [μ m]
y [μ m]

5.0 5.0

2.5 2.5

0.0 0.0

0.0 2.5 5.0 7.5 10.0 0.0 2.5 5.0 7.5 10.0
x [μ m] x [μ m]
(c) (d)

Algorithm 2 IGA Solution Output element is worse than that of the INSQ element as can be seen
1: subroutine postprocessing in Fig. 6. Also the size of the jacobian for HNSQ2 is 1,872 ×
2: Project variables from Gauss points to control points Eq. (27) 1,872 and the number of non-zero entries in it is 66,547
3: for iel = 1, number of elements do
4: for j = 1, number of plotted values per element do
whereas the size of the jacobian for INSQ is 4,896 × 4,896
5: get the solution (u, T , σ and γ p ) in the physical domain for 58,608 non-zero entries in it. In other words, even though
Eq. (28) the linear system size due to HNSQ2 is smaller than INSQ,
6: end for it also much denser and hence more memory is needed to
7: Get the physical mesh connectivity for the plotted values in this
element
store the jacobian in case of the HNSQ2 mesh. Furthermore
8: end for due to the larger bandwidth of HNSQ2 jacobian the linear
9: write plotted solution and associated connectivity to a file (.vtu for solves at every Newton iteration require more memory and
ParaView) more CPU time. Detailed studies on cpu time and memory
10: Return
usage for the HNSQ2 element can be found in Fig. 11e, f.

of element per mesh direction and p as the order of the ele-

ment, the total number of nodes in the mesh (Nk for element
HNSQ2 and N g p for element INSQ) is 4 Numerical examples

Nk = (n + p + 1)2 = n 2 + 2( p + 1)n + ( p + 1)2 In this section we study the performance of the elements on
N gp = n × ( p + 1) .
2 2
(29) two benchmark examples:

This clearly shows that the number of nodes used by HNSQ2

elements grows much slower compared to the number of • A pure tension test on a quarter of a plate with a central
nodes used by INSQ element upon mesh refinement (when imperfection
n increases). Moreover, the sparsity pattern for the HNSQ2 • A shear test on a regular plate (no imperfection added)

123
Comput Mech

Fig. 5 Stress discretization in

HNSQ2 and INSQ elements. 111 111
Red circles show the location of
stress degrees of freedom. ξ and
η are paremetric variables 0.75 0.75
spanning the knot vectors.
a HNSQ2 element. b INSQ
0.5

η
0.5

η
element . (Color figure online)

0.25 0.25

000 000

000 0.25 0.5 0.75 111 000 0.25 0.5 0.75 111
ξ ξ
(a) (b)

0 0

200 500

400 1000

600 1500

800 2000

2500
1000

3000
1200

3500
1400
4000
1600
4500
1800
0 200 400 600 800 1000 1200 1400 1600 1800 0 500 1000 1500 2000 2500 3000 3500 4000 4500
nz = 66547 nz = 58608

Fig. 6 Sparsity pattern for the HNSQ2 (left) and INSQ (right) elements for a ten by ten mesh with quadratic elements. (Color figure online)

While all the elements presented in Table 4 have been tested,

only the the HNSQ2 element that uses NURBS shape func-
tions for all the fields except the EQPS field is shown as Table 5 Material parameters used in the example problems
it was found to give the best performance. This element Physical quantity name Name Value SI base unit
provides good convergence results upon k-refinement but is
also not subject to spurious oscillations in the EQPS field Young’s modulus E 200E9 kg m−1 s−2
thus providing a good compromise between performance and Poisson’s ratio ν 0.3 −
robustness. Mass density ρ 7830 kg m−3
The physical parameters used for all benchmark examples Specific heat cp 448 m2 s−2 K−1
are summarized in Table 5. Thermal conductivity κ 803.5 kg m s−3 K
On all benchmark simulations we adopt an average accel- Thermal softening parameter δ 0.8 −
eration, Newmark scheme (with parameters γ = 0.25 and Thermal softening parameter k 500 K
β = 0.5). Adaptive time stepping is used when the nonlinear Reference temperature T0 293 K
solution is slow to converge. The NURBS meshes used for the Taylor-Quinney Coefficient χ 0.9 −
simulations employ a single patch, and k-r e f inement is used
Yield stress σ0 2E9 kg m−1 s−2
in order to maintain full continuity throughout the meshes.
Yield strain γ0 0.01 −
As a point of comparison, we use the PSSQ element to test
Reference strain rate γ̇0 0.001 s−1
the performance of the proposed NURBS based element.
Strain hardening exponent n 0.01 −
Details on the implementation and characteristics of the PS
Rate sensitivity parameter m 70 −
element can be found in [6,47–49], details on using the PSSQ

123
 Comput Mech

20 by 20, 30 by 30 and 40 by 40 elements to study the conver-

Fig. 7 The error terms contributing to the error norm are computed at
the nodes denoted by red circles, which are shared on all hierarchy of
grids . (Color figure online)
element for the simulation of shear bands can be found in
[4,52].
In the following convergence analysis, we define the error
norm as the relative distance in an Euclidian space between
some reference solution and the current computed solution,
which is expressed as follow:
n
||u r e f − u||2 (u r e f,i − u i )2
e= = i=1
n , (30)
||u r e f ||2 2
i=1 u r e f,i
where u r e f is the solution obtained on the finest mesh (40 ×
40) with the highest order element (P 4 ) and n is the number
of nodes on our coarsest mesh (10×10). The error terms for a
specific solution field are computed at nodes which are shared
by coarse and fine meshes, as illustrated by the red circles in
Fig. 7. This also ensures that all the values used to compute
the error on each mesh are actually values which have been
solved for and not interpolated. This means that the error
presented is least biased by the post-processing operations,
and the number of terms in the error norm are determined by
the coarsest grid.
On all the examples studied, the same structured dis-
cretization is employed, which is generated by a tensorial
product type of construction. We use meshes with 10 by 10,
gence with h-r e f inement. Convergence of the error is also
investigated by k-r e f inement capabilities of the NURBS
shape functions.
All the simulations in this paper are obtained using an
implementation of the NURBS shape functions in FEAP [53]
using PETSc [54] as a solver for the linearized Newton sys-
tem. The results are then post-processed using ParaView [55]
and particularly its Python scripting capabilities (pvpython).
Finally, graphs and plots are generated with the python library
Matplotlib [56].
4.1 Plate with imperfection under pure tension
The first benchmark example is of a plate with a central
imperfection subjected to a uniform tension test. Since hori-
zontal and vertical symmetry exists, only the top left quarter
of the plate is considered as shown in Fig. 10. A uniform
velocity ramp is applied at the top of the plate: the velocity
linearly increases from 0 to 5 m/s in 1 µs and remains con-
stant at 5 m/s after the ramp. With this example we consider
the following two studies:
i. the sensitivity of the formulation to mesh density and
mesh alignment
ii. the convergence of the HNSQ2 as compared to the Pian
and Sumihara element.
4.1.1 Mesh sensitivity studies
We first study the sensitivity of the HNSQ2 element to mesh
density and alignment. We consider the following mesh den-
sities: 20 × 20, 30 × 30 and 40 × 40 and also three different
element aspect ratios: 2, 1 and 0.5 using 40 × 20, 20 × 20
and 20 × 40 meshes. Figure 8 shows the 20 × 20 and 20 × 30
meshes as well as the line along which the EQPS field is
plotted. The resulting EQPS curves are shown in Fig. 9.

Fig. 8 The 40 × 20 and 10 10

20 × 40 meshes used for the
mesh sensitivity studies. The
8 8
grey line represents the line
along which the EQPS is shown
in Fig. 9. The meshes are plotted 6 6
y [μ m]

y [μ m]

on top of the EQPS field

computed at time t = 1.25 µs.
4 4
(Color figure online)

2 2

0 0
0 2 4 6 8 10 0 2 4 6 8 10
x [ μm] x [ μm]

123
Comput Mech

20x20 time is presented and Fig. 11f shows the relation between
0.5 30x30 error decrease and memory needed to solve the problem.
40x40
The simulation time for this example ranged from 16.23 s
20x40
Equivalent plastic strain

40x20 for PSSQ on a 10 by 10 mesh to 6 h 54 min and 1 s for the

0.4
P 4 HNSQ2 element on a 40 by 40 mesh. The imperfection
is induced by reducing the value of the yield stress and yield
0.3 strain using a 2D beta function given as
⎡ ⎛ ⎞⎤2
X 12 + X 22
0.2 β2D (X 1 , X 2 ) = 1 − 0.04 ⎣sech ⎝ ⎠⎦ , (31)
5 · 10−6
σ yield = σ0 β2D (X 1 , X 2 ), (32)
0.1
0 2e-06 4e-06 6e-06 8e-06 1e-05 1.2e-05 1.4e-05 1.6e-05
γ yield = γ0 β2D (X 1 , X 2 ). (33)
x[ μ m]

Fig. 9 Plot of the EQPS field for three different mesh densities as well
as three different aspect ratios
It can clearly be seen that the EQPS field converges to the
same solution on all meshes considered. This indicates that
i. the width of the shear band does not depend on the size
of the element used for the simulation,
ii. the aspect ratio of the elements in the mesh does not affect
the results, suggesting that shear band formation is not
sensitive to mesh alignment.
These results are in agreement with those reported in [4] and
show that inclusion of thermal conductivity together with
a monolithic solver guarantee the accurate representation
of the localization observed during the formation of shear
bands. In addition, use of IGA suppresses mesh alignment
sensitivity because of the nonlocality of the NURBS basis
function. Nonlocality has also been found to suppress mesh
alignment in mesh free methods [7,8] (Fig. 10).
4.1.2 Convergence studies
The total duration of the simulation is 2 µs and the error is
evaluated at time t = 1.2 µs. In Fig. 11e the convergence
of the error on the displacement as a function of the CPU
This corresponds to a smooth imperfection centered on x = 0
and y = 0 with a maximal reduction of the reference yield
parameters of 4%.
As expected, the convergence of the fields discretized with
NURBS shape functions gives fast convergence rates and
smaller initial errors on coarse meshes see Fig. 11a–d as com-
pared with the PSSQ element. Nonetheless, the convergence
rate on the EQPS field for both elements is similar.
It can also be observed that the rate of increase of the
computational time (cpu time) with respect to the size of
the mesh is of the same order for all the elements. Thus to
achieve similar accuracy the HNSQ2 element will converge
in less time than the PSSQ element. Similarly Fig. 11f shows
that the HNSQ2 element requires less memory for a given
accuracy than the memory required by the PSSQ element,
especially if high order NURBS functions are used.
Finally, Fig. 12 shows a comparison of the vertical dis-
placement (u y ) and shear strain during the deformation to
verify that the elements are not subject to volumetric lock-
ing. It can be observed that the PSSQ and HNSQ2 (with
B-bar ) allow shear deformations to occur during the simu-
lation whereas the INSQ element with B-bar only allows for
limited amount of shear deformation and vertical displace-
ment. The INSQ element without B-bar is severely resist-
ing shear deformation. This confirms the need for the B-bar
method to overcome volumetric locking for coarse mesh with
low order element.

Fig. 10 Representation of the

geometry and boundary
conditions used for example 1.
Symmetry is assumed and only
a quarter of the plate is modeled.
Points A and B are used in the
volumetric locking studies in
Fig. 12

123
 Comput Mech

P1 P2 P3 P4 PS

−2 −4
10 10
−3 −5
10 10

Relative error
Relative error
−4 −6
10 10
−5 −7
10 10
−6 −8
10 10
−7 −9
10 10
−8 −10
10 10
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
h[ μ m] h[ μ m]
(a) (b)

10 −1 10 −3

10 −4

Relative error
Relative error

10 −2

10 −5

10 −3
10 −6

10 −4 10 −7
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
h[ μ m] h[ μ m]
(c) (d)

10 −2 10 −2

10 −3 10 −3
Relative error
Relative error

10 −4 10 −4

10 −5 10 −5

10 −6 10 −6

10 −7 10 −7

10 −8 1 10 −8 7
10 10 2 10 3 10 4 10 5 10 10 8 10 9
CPU time[s] Memory usage[bytes]
(e) (f)
Fig. 11 Convergence of various solution fields the pure tension case a Displacements. b Temperature. c Equivalent plastic strain. d von
in example 1, modeled by the HNSQ2 element with different order of Mises stress. e Error against CPU usage. f Error against memory usage
NURBS basis function. CPU time and memory usage are also reported.

4.2 Plate under shear loading
The second example examines a more complex numerical
test. A shear load applied at the top of the plate triggers
the formation of a shear band at the left and right bottom
corners of the plate, see Fig. 13. In this configuration, the
plastic strain develops in a much more localized region which
develops into an arc type shear band. The equivalent plastic
strain sustained by the plate at each corner is substantially
higher than in the previous example and very steep gradients
develop around this shear band. No imperfection is used in
this plate since the intense stress concentration observed at
the bottom of the plate is sufficient to trigger the shear band.
The convergence rate of the HNSQ2 element is reported
in Fig. 14. The error norms are computed at time t = 0.8 µs
when the shear band is fully formed.
Figure 14 shows that the HNSQ2 element and the PSSQ
element have similar rates of convergence for the displace-
ment and temperature fields. However, the equivalent plas-
tic strain computed with the HNSQ2 element is much more
accurate and interestingly provides faster rate of conver-
gence than the PSSQ element. Thus the PSSQ element

123
Comput Mech

underestimates the plastic work which is occurring in the since the heat source in our formulation is the plastic
plate and this could lead to errors propagating to other work.
solution fields. This especially affects the temperature field All elements listed in Tables 3 and 4 have been tested on
this example problem, which presented convergence difficul-
ties for some of the elements. For instance, the HNSQ1 ele-
9e-06
PSSQ ment which employs NURBS shape functions for the equiv-
8e-06 HNSQ2(1) bbar alent plastic strain field did not converge on this problem due
INSQ(1) bbar to oscillations in the equivalent plastic strain field as depicted
7e-06
INSQ(1) no bbar
in Fig. 15. Nonetheless, this element type, provides fast con-
6e-06
uy (x = l/ 2, y = l)

vergence rates for all fields when used to model a plate under
5e-06 pure tension. In fact the rate of convergence for each field was
4e-06 found to increase with the order of the NURBS shape func-
tions similarly to what can be observed when NURBS are
3e-06
used for linear problems (see [14]). Such an element would
2e-06 greatly improve our ability to resolve shear bands on rel-
1e-06 atively coarse meshes and therefore reduce computational
time and memory needs, compared to elements requiring
0
0 5e-07 1e-06 1.5e-06 2e-06 more refined meshes.
Time The MNSQ element has also proven to be unstable for
(a) a plate under shear loading for the same reasons of spuri-
0.016
ous oscillations in the equivalent plastic strain field. These
PSSQ oscillations might be eliminated by the implementation
0.014 HNSQ2(1) bbar
of a Streamline Upwind/Petrov–Galerkin (SUPG) type of
INSQ(1) bbar
0.012
INSQ(1) no bbar
method [57,58] which we intend to explore in future work.
0.010

0.008
εxy

0.006 5 Conclusion
0.004
A coupled mixed finite element formulation with diffusive
0.002 regularization is used in the paper to model shear bands. IGA
0.000 with high order NURBS elements is employed to discretize
the system, and the solution at every time step is obtained with
−0.002
0 5e-07 1e-06 1.5e-06 2e-06 a monolithic Newton type solver. This approach is shown
Time to be insensitive to mesh refinement and mesh alignment,
(b) and is therefore an attractive scheme for shear band model-
Fig. 12 Illustration of volumetric locking behaviour for different ele- ing. Moreover, the results demonstrate that the use of higher
ments due to J2 plasticity in pure tension example 1. The orange line order NURBS shape functions for the four field shear band
marks the onset of plasticity. a Vertical displacement at point B, b shear problem significantly improve the convergence rate over the
strain at point A. Note that points A and B are shown in Fig. 10

Fig. 13 Plate under shear

loading leading to the formation
of an arc shaped shear band. a
Geometry and boundary
conditions. b Plastic strain at
time t = 2 , µs

(a) (b)

123
 Comput Mech

P1 P2 P3 P4 PS

10 −1 10 −2

Relative error

Relative error
10 −2 10 −3

10 −3 10 −4
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
h[ μ m] h[ μ m]
(a) (b)

10 0 10 1
Relative error

Relative error
10 0

10 −1

10 −1 10 −2
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
h[ μ m] h[ μ m]
(c) (d)

Fig. 14 Convergence rate of the HNSQ2 element for the shearing example. a Displacements. b Temperature. c Equivalent plastic strain. d von
Mises stress

101 used for the four fields, with varying degrees of success. Cer-
INSQ
HNSQ1 tain types of elements were found to be prone to oscillation
100
instabilities in the equivalent plastic strain field (MNSQ and
Equivalent plastic strain

10−1 HNSQ1), which is unphysical and fatal to the simulation.

The most successful element was the HNSQ2 element,
10−2 which was tested on two examples with different types of
loading conditions. This element was found to behave effi-
10−3
ciently in all cases compared to the reference Pian–Sumihara
10−4 element, indicating that it is a well suited alternative for
the simulation of shear bands. This becomes especially pro-
10−5 nounced when the advantages of k-r e f inement is used to
model shear bands on coarse meshes.
10−6
0 2 4 6 8 10 The remaining issue with these family of elements is the
x[ μ m] slow convergence of the equivalent plastic strain field, which
Fig. 15 Oscillations observed in the equivalent plastic strain field is not interpolated with higher order NURBS, due to the oscil-
along the bottom edge of the plate. The black line indicate the value lations mentioned previously. However, the convergence of
of the zero base line for the equivalent plastic strain. This value is non the equivalent plastic strain field can be improved by dis-
zero since the oscillations are creating negative values which cannot be cretizing this field with NURBS shape functions and deriv-
plotted on a semi-log scale
ing a SUPG type method to avoid these spurious oscillations.
Alternatively an element based on a discontinuous Galerkin
low order Pian–Sumihara element used in previous work on technique applied to the equivalent plastic strain field could
shear bands. The advantage of higher order NURBS becomes lead to a stable and fast converging element and allow for
particularly apparent when cpu time and memory usage are higher order NURBS functions. These directions will be the
accounted for. Several combinations of shape functions were topic of future research.

123
Comput Mech

Acknowledgments The financial support of the Department of Energy Kσ = −wσ C elas : ∇ s δu d, (46)
through the Early Career Research Program, No. DE-SC0008196, is

gratefully acknowledged.  "
3 d "
Gσ T = wσ C elas : g(σ̂ , T + εδT, γ̂ p )"" S
2σ̂ dε ε=0

!
Appendix: Residual and jacobian terms expressions ∂δT
+α I d, (47)
∂t

The expressions of the terms constituting the residual vector
Mσ = wσ δ σ̇ d, (48)
in Eq. (10) are as follow

    "
3 d "
Ru = wu ρ ü + ∇wu · σ d − t̄dt = 0, (34) Gσ σ = wσ C elas : g(σ̂ (σ +εδσ ), T, γ̂ p )"" S
2σ̂ dε ε=0

 t "
 3 d σ̂ (σ + εδσ ) "" 1
RT = wT [ρc Ṫ − χ σ̂ g(σ̂ , T, γ̂ p )] − " g(σ̂ , T, γ̂ p )S
2 dε " σ̂ 2
ε=0
 " !
 3 d S(σ + εδσ ) "
+ g(σ̂ , T, γ̂ p ) " d, (49)
+κ∇wT ∇T d − κ q̄dq = 0, (35) 2σ̂ dε "
ε=0
 " !
q 3 d "
  Gσ γ̂ p = wσ C elas : g(σ̂ , T, γ̂ p +εδ γ̂ p )"" S d.
3 g(σ̂ , T, γ̂ p ) 2σ̂ dε ε=0
Rσ = wσ σ̇ − C elas
: ∇ u̇ −
s
S 
2 σ̂ (50)
  "
 d "
−α Ṫ I d = 0, (36) Gγ̂ p T = −wγ̂ p g(σ̂ , T + εδT, γ̂ p )"" d, (51)
dε ε=0
 
 "
Rγ̂ p = wγ̂ p [γ̇ p − g(σ̂ , T, γ̂ p )]d = 0. (37) d "
Gγ̂ p σ = −wγ̂ p g(σ̂ (σ + εδσ ), T, γ̂ p )"" d, (52)
 dε ε=0

 "
d "
The expression for the entries of the jacobian of the system Gγ̂ p γ̂ p = −wγ̂ p g(σ̂ , T, γ̂ p + εδ γ̂ p )"" d. (53)
shown in Eq. (13) are the following dε ε=0


∂ 2 δu
Mu = wu ρ d, (38)
∂t 2


Ku = (∇wu )δσ d (39) References

 ! 1. Wright TW (2002) The physics and mathematics of adiabatic shear
∂δT bands. Cambridge University Press, Cambridge
MT = wT ρc d. (40)
∂t 2. Marchand A, Duffy J (1988) An experimental study of the for-
 mation process of adiabatic shear bands. J Mech Phys Solids 38:

251–283
Mγ̂ p = wγ̂ p δ γ̄˙ p d, (41) 3. Wright TW, Walter JW (1987) On stress collapse in adiabatic shear
 bands. J Mech Phys Solids 35(6):701–720
 4. McAuliffe C, Waisman H (2013) Mesh insensitive formulation for
KT = ∇wT κ∇δT d, (42) initiation and growth of shear bands using mixed finite elements.
 Comput Mech 51(5):807–823
 " 5. Tvergaard V, Needleman A, Lo KK (1981) Flow localization in the
d "
GT T = g(σ̂ , T + δT, γ̂ p )""
−wT χ σ̂ d, (43) plane strain tensile test. J Mech Phys Solids 29:115–142
dε ε=0 6. Pian THH, Sumihara K (1984) Rational approach for assumed

 # " stress finite elements. Int J Numer Methods Eng 20:1638–
d σ̂ (σ + εδσ ) "" 1685
GT σ = wT −χ " g(σ̂ , T, γ̂ p ) 7. Li S, Liu WK (2000) Numerical simulations of strain localization
dε "
 ε=0 in inelastic solids using mesh-free methods. Int J Numer Methods
d  "" ! Eng 48(9):1285–1309
−χ σ̂ g σ̂ (σ + εδσ ), T, γ̂ p "" d, (44) 8. Li S, Liu WK, Rosakis AJ, Belytschko T, Hao W (2002) Mesh-
dε ε=0 free Galerkin simulations of dynamic shear band propagation and
 "
d " failure mode transition. Int J Solids Struct 39:1213–1240
GT γ̂ p = −wT χ σ̂ g(σ̂ , T, γ̂ p + εδ γ̂ p )"" d, (45) 9. Liu WK, McVeigh C (2008) Predictive multiscale theory for design
dε ε=0 of heterogeneous materials. Comput Mech 42(2):147–170


123

10. McVeigh C, Liu WK (2008) Linking microstructure and properties
through a predictive multiresolution continuum. Comput Methods
Appl Mech Eng 197(41–42):3268–3290
11. McVeigh C, Liu WK (2010) Multiresolution continuum modeling
of micro-void assisted dynamic adiabatic shear band propagation.
J Mech Phys Solids 58(2):187–205
12. Tang S, Kopacz AM, Chan O’Keeffe S, Olson GB, Liu WK (2013)
Three-dimensional ductile fracture analysis with a hybrid mul-
tiresolution approach and microtomography. J Mech Phys Solids
61(11):2108–2124
13. Tang S, Kopacz AM, Chan O’Keeffe S, Olson GB, Liu WL (2013)
Concurrent multiresolution finite element: formulation and algo-
rithmic aspects. Comput Mech 52(6):1265–1279
14. Cottrell JA, Hughes TJR, Bazilevs Y (2009) Isogeometric analysis:
toward integration of CAD and FEA. Wiley, New York
15. Hughes TJR, Cottrell JA, Bazilevs Y (2005) Isogeometric analysis:
Cad, finite elements, nurbs, exact geometry and mesh refinement.
Comput Methods Appl Mech Eng 194(39–41):4135–4195
16. Bartels RH, Beatty JC, Barsky BA (1987) An introduction to splines
for use in computer graphics and geometric modeling. Morgan
Kaufmann, San Francisco
17. Farin GE (1995) NURBS curves and surfaces: from projective
geometry to practical use. A.K. Peters Ltd., Wellesley, MA
18. Bazilevs Y, Calo VM, Cottrell JA, Evans JA, Hughes TJR, Lipton
S, Scott MA, Sederberg TW (2010) Isogeometric analysis using
T-splines. Comput Methods Appl Mech Eng 199(5—-8):229–263
(Computational Geometry and Analysis)
19. Bazilevs Y, Hsu M-C, Scott MA (2012) Isogeometric fluid-
structure interaction analysis with emphasis on non-matching dis-
cretizations, and with application to wind turbines. Comput Meth-
ods Appl Mech Eng 249—-252(0):28–41 (Higher Order Finite Ele-
ment and Isogeometric Methods)
20. de Boor C (2001) A practical guide to splines. Applied mathemat-
ical sciences, vol 27. Springer, New York
21. Bazilevs Y, Beirao da Veiga L, Cottrell JA, Hughes TJR, Sangalli
G (2006) Isogeometric analysis: approximation, stability and error
estimates for h-refined meshes. Math Models Methods Appl Sci
16(07):1031–1090
22. Elguedj T, Bazilevs Y, Calo VM, Hughes TJR (2008) B̄-bar and
F-bar projection methods for nearly incompressible linear and non-
linear elasticity and plasticity using higher-order nurbs elements.
Comput Methods Appl Mech Eng 197(33–40):2732–2762
23. Bouclier R, Elguedj T, Combescure A (2012) Locking free isogeo-
metric formulations of curved thick beams. Comput Methods Appl
Mech Eng 245–246:144–162
24. Piegl LA, Tiller W (1997) The NURBS book, 2nd edn. Springer,
New York
25. Rogers DF (2001) An introduction to NURBS with historical per-
spective. Morgan Kaufmann, San Francisco
26. Hughes TJR (1987) The finite element method: linear static and
dynamic finite element analysis. Prentice-Hall, New Jersey
27. Elguedj T, Hughes TJR (2013) Isogeometric analysis of nearly
incompressible large strain plasticity. Comput Methods Appl Mech
Eng 268: 388–416
28. Bažant ZP, Pijaudier-Cabot G (1988) Nonlocal continuum damage,
localization instability and convergence. J Appl Mech Trans ASME
55(2):287–293 (cited By (since 1996) 289)
29. Mindlin RD, Eshel NN (1968) On first strain–gradient theories in
linear elasticity. Int J Solids Struct 4(1):109–124 (cited By (since
1996) 190)
30. Mühlhaus H-B, Alfantis EC (1991) A variational principle for gra-
dient plasticity. Int J Solids Struct 28(7):845–857 (cited By (since
1996) 380)
31. De Borst R, Mühlhaus H-B (1992) Gradient-dependent plasticity:
formulation and algorithmic aspects. Int J Numer Methods Eng
35(3):521–539 (cited By (since 1996) 478)
123
Comput Mech
32. Zhou M, Rosakis AJ, Ravichandran G (1996) Dynamically prop-
agating shear bands in impact-loaded prenotched plates-i. Exper-
imental investigations of temperature signatures and propagation
speed. J Mech Phys Solids 44(6):981–1006
33. Zhou M, Ravichandran G, Rosakis AJ (1996) Dynamically propa-
gating shear bands in impact-loaded prenotched plates-ii. Numer-
ical simulations. J Mech Phys Solids 44(6):1007–1032
34. Zienkiewicz OC, Taylor RL (2000) The finite element method, vol
1. Butterworth Heinemann, Oxford
35. Sokolnikoff IS (1951) Tensor analysis: theory and applications.
Applied mathematics series. Wiley, New York
36. Lovelock D, Rund H (1975) Tensors, Differential Forms, and Varia-
tional Principles. Dover Books on Mathematics Series. Dover Pub-
lications, New York
37. Hellinger E (1914) Der allgemeine Ansatz der Mechanik der Kon-
tinua, chapter 4, vol 4. Springer, Berlin, p 602
38. Reissner E (1950) On a variational theorem in elasticity. J Math
Phys 29(2):90–95
39. Babuška I (1971) Error bounds for finite element method.
NumerischeMathematik 16:322–333
40. Brezzi F (1974) On the existence, uniqueness and approximation of
saddle-point problems arising form lagrange multipliers. RAIRO
8(R2):129–151
41. Babuška I, Narasimhan R (1997) The Babuška–Brezzi condition
and the patch test: an example. Comput Methods Appl Mech Eng
140:183–199
42. Hermann LR (1965) Elasticity equations for incompressible and
nearly incompressible materials by a variational theorem. AIAA J
3(10):1896–1900
43. Auricchio F, Brezzi F, Lovadina C (2004) Mixed Finite Element
Methods, Chapter 9. Wiley, New York, pp 237–278
44. Brezzi F, Bathe K-J (1990) A discourse on the stability conditions
for mixed finite elment formulations. Comput Methods Appl Mech
Eng 82(1–3):27–57
45. Bathe K-J (2001) The inf-sup condition and its evaluation for mixed
finite element methods. Comput Struct 79(2):243–252
46. De S, Bathe K-J (2001) Displacement/pressure mixed interpola-
tion in the method of finite spheres. Int J Numer Methods Eng
51(3):275–292
47. Pian THH, Tong P (1986) Relations between incompatible dis-
placement model and hybrid stress model. Int J Numer Methods
Eng 22:173–181
48. Jiao ZP, Pian THH, Yong S (1997) A new formulation of the
isoparametric finite elements and the relationship between hybrid
stress element and incompatible element. Int J Numer Methods Eng
40:15–27
49. Zhou TX, Xie XP (2002) A unified analysis for stress/strain hybrid
methods of high performance. Comput Methods Appl Mech Eng
191:4619–4640
50. Boffi D, Brezzi F, Fortin M (2013) Mixed finite element methods
and applications. Springer series in Computational Mathematics,
vol 44. Springer, New York
51. Comi C, Perego U (1995) A unified approach for variationally
consistent finite elements in elastoplasticity. Comput Methods Appl
Mech Eng 121(1–4):323–344
52. McAuliffe C, Waisman H (2013) A pian-sumihara type element for
modeling shear bands at finite deformation. Comput Mech 1–16
53. Taylor RL (2011) FEAP—A Finite Element Analysis Program
54. Balay S, Brown J, Buschelman K, Eijkhout V, Gropp WD, Kaushik
D, Knepley MG, Curfman McInnes L, Smith BF, Zhang H (2013)
PETSc users manual. Technical Report ANL-95/11—Revision 3.4,
Argonne National Laboratory
55. Henderson A (2008) ParaView guide, a parallel visualization appli-
cation. Kitware Inc., Santa Fe, NM
56. Hunter JD (2007) Matplotlib: a 2d graphics environment. Comput
Sci Eng 9(3):90–95
Comput Mech

57. Hughes TJR, Franca LP, Hulbert GM (1989) A new finite ele- 58. Brezzi F, Bristeau MO, Franca LP, Mallet M, Rogé G (1992) A rela-
ment formulation for computational fluid dynamics: viii. The tionship between stabilized finite element methods and the galerkin
galerkin/least-squares method for advective-diffusive equations. method with bubble functions. Comput Methods Appl Mech Eng
Comput Methods Appl Mech Eng 73(2):173–189 96(1):117–129

123