Download as pdf or txt
Download as pdf or txt
You are on page 1of 30

INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING

Int. J. Numer. Meth. Engng 2000; 00:1–6 Prepared using nmeauth.cls [Version: 2002/09/18 v2.02]

An adaptive stabilization strategy for enhanced strain methods in


nonlinear elasticity

Alex Ten Eyck and Adrian J. Lew∗


Department of Mechanical Engineering, Stanford University, CA 94305-4040.

SUMMARY

This paper proposes and analyzes an adaptive stabilization strategy for enhanced strain methods
applied to quasistatic nonlinear elasticity problems. The approach is formulated for any type of
enhancements or material models, and it is distinguished by the fact that the stabilization term
is solution-dependent. The stabilization strategy is first constructed for general linearized elasticity
problems, and then extended to the nonlinear elastic regime via an incremental variational principle.
A heuristic choice of the stabilization parameters is proposed, which in the numerical examples proved
to provide stable approximations for a large range of deformations, different problems and material
models. We also provide explicit lower bounds for the stabilization parameters that guarantee that
the method will be stable. These are not advocated, since they are generally larger than the ones
based on heuristics, and hence prone to deteriorate the locking-free behavior of enhanced strain
methods. Numerical examples with two different nonlinear elastic models in thin geometries and near
incompressible situations show that the method remains stable and locking-free over a large range of
deformations. Finally, the method is strongly based on earlier developments for discontinuous Galerkin
methods, and hence throughout the paper we offer a perspective about the similarities between the
two. Copyright c 2000 John Wiley & Sons, Ltd.

key words: enhanced strain methods, discontinuous Galerkin methods, nonlinear elasticity,
stabilization, adaptivity.

1. Introduction

This paper proposes and analyzes an adaptive stabilization technique for enhanced strain
methods applied to problems in nonlinear elasticity. In this approach the stabilization term is
solution-dependent, and hence it is termed adaptive stabilization. The idea and applicability
of the method is not specific to any particular material model or type of enhancement.
Enhanced strain (ES) methods construct approximations to solutions of elasticity problems
by enriching the space that approximates strains within each element. The added strain fields
are not required to satisfy any kinematic compatibility condition across element boundaries;
the method is hence incompatible† . The extra degrees of freedom provide unquestionable

∗ Correspondence to: lewa@stanford.edu


† Incompatibility in the sense that the strain field is not the weak derivative of any displacement field.

Received
Copyright c 2000 John Wiley & Sons, Ltd. Revised
2 A. TEN EYCK AND A. LEW

advantages, such as the absence of volumetric and shear locking for near incompressible and
very thin geometries, respectively, when low order elements are adopted (see, e.g., [1]). However,
the use of incompatible approximations exposes the method to the appearance of numerical
instabilities. These instabilities would not be observed in the exact solution, or in approximate
ones obtained with most conforming methods.
Enhanced strain methods are not the only incompatible finite element methods. For example,
discontinuous Galerkin (DG) methods also suffer from this artifact [2]. This defect not only
degrades the benefits of incompatible methods, but their overall versatility and reliability as
well. The heart of the problem can be easily described, at least in the linearized elasticity
case. In the exact linearized elasticity problem we seek a minimizer of the potential energy
functional among all kinematically compatible strain fields, up to some additional conditions
needed to satisfy Dirichlet boundary conditions. Consequently, if a unique solution exists, its
potential energy is lower than that of any other compatible strain field. This property can not
longer be guaranteed if some incompatible strain fields are included in the set over which a
potential energy minimizer is sought. In fact, over this set, the potential energy functional‡
may not longer have a minimum and a minimizer. If this is the case, an incompatible unstable
mode will be observed. This explains why spurious unstable modes may be observed in this
class of methods: incompatible strain fields are considered in the set over which the potential
energy minimizer is sought.
Developing stable ES methods for nonlinear elasticity has been a topic of interest for fifteen
or more years by now (see, e.g., [1, 3, 4, 5, 6, 7, 8, 9, 10]). Many of the ideas for stabilization
have been motivated by the observation that some numerical instabilities in ES methods for
nonlinear elasticity exhibit similar deformation modes to those seen in conforming methods
with reduced integration. These are the so-called hourglass instabilities (see, e.g., [11]), because
their manifestation can be identified by some pairs of quadrilateral elements forming hourglass
shapes. Inspired by the techniques of hourglass stabilization for conforming methods with
reduced integration, authors in [4, 6, 5] have proposed one-quadrature point formulations of ES
methods for nonlinear elasticity. These methods are essentially based in providing alternative
ways of computing the integrals that involve the parts of the strain approximation that are
non-constant over an element. With some mild simplifications, this is often done by providing
analytical expressions for them. Because without them the method based on one-quadrature
point per element would be unstable, these receive the name of stabilization techniques in the
literature.
In the case of ES methods, however, hourglass instabilities appear even if the integration is
performed exactly (see, e.g. [8], and references therein), as a consequence of the method being
incompatible. The methods proposed in [8, 7] are based on reduced integration techniques, but
have an additional stabilization parameter to tune the amount of stabilization added. Despite
the good numerical performance shown, additional discussion and analysis would be needed
to conclude that the proposed stabilization terms will guarantee some form of stability for
arbitrary loading situations.
Another class of stabilization strategies for ES methods consist in constructing a term
that penalizes or adds stiffness to all strain enhancements. These terms generally involve

‡ extended to incompatible strain fields by considering their absolutely continuous part only.

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
AN ADAPTIVE STABILIZATION STRATEGY FOR ENHANCED STRAIN METHODS 3

one or more stabilization parameters. Such an approach was proposed in [1, 12, 3]. It was
observed therein that it is possible to stabilize general ES methods and avoid locking in this
way, provided the stabilization parameters were adequately chosen. This may not be an easy
task, since the onset of the instabilities and hence the appropriate values of the stabilization
parameters depend on the material model, as recognized early on in [1, 13]. More importantly,
numerical solutions sometimes display a strong sensitivity to the choice of the stabilization
parameters. This sensitivity to the choice of stabilization parameters is particularly important
in the presence of near incompressible behavior. Large values of the stabilization parameter
result in locking, as commented in [6, 7].
This provides another common feature with DG methods, in which a standard strategy
to restore stability is to penalize the discontinuities across element boundaries. These
discontinuities are responsible for the incompatible strain fields that appear in the method
(see Sec. 3). It was shown in [2] that such an approach stabilizes the method and avoids
locking if the stabilization parameter is properly chosen. However, the numerical examples
in [14, 15] show that DG methods are prone to display the same sensitivity to the choice
of the stabilization parameters as ES methods. Furthermore, it was argued and showcased
with numerical examples therein that this dependence could be ameliorated by constructing
a solution-dependent, spatially-varying stabilization term. Such a strategy was proposed in
[14, 15] and termed adaptive stabilization, because of its dependence on the solution sought.
The design of solution-dependent stabilization terms for ES methods is not new, see
[4, 6, 8, 7, 9]. Among these, the strategy proposed in [9] explicitly highlights this feature
and shares a number of similarities with the one proposed here, as discussed in Sec. 5. In
principle the strategy in [9] seems to be widely applicable. However, the manuscript crafts a
stabilization term only for a two-dimensional ES element made of a neo-Hookean material.
In this manuscript we construct an adaptive stabilization strategy for ES methods, building
on the ideas for DG methods in [14, 15]. The contribution is not conceptually far from some
of the strategies proposed by other authors, but it combines a number of desirable features
in a way that no other existing strategy does. The method is independent of the hyperelastic
material model and the type of strain enhancement chosen, the stabilization term is solution-
dependent, for most problems only one stabilization constant needs to be selected with a value
that is O(1) over a wide range of deformations, it performs well in situations where volumetric
or shear locking would otherwise be expected, and the strategy is guaranteed to stabilize the
ES method provided the stabilization parameters are chosen large enough. In this manuscript
we are only concerned with instabilities arising from the use of an incompatible method. We
leave the possibility of including reduced integration techniques to improve its efficiency open
for the future.
To design the stabilization strategy, we consider first the linearized elasticity problem at any
given nonlinear elastic deformation, see Sec. 5. Such problem is not necessarily coercive, i.e.,
there may be buckling modes. We then propose a stabilization strategy for an ES discretization
designed so that the resulting method is coercive whenever the exact linearized elasticity
problem is, see Sec. 6. To this end and following [15], we prove a proposition in which explicit
expressions for the minimum values of the set of stabilization parameters for the method are
provided. However, as we discuss, these lower bounds are often far from optimal, resulting
in methods that are over-stabilized and that may display locking. Instead, we choose the
stabilization parameters in a heuristic way and demonstrate through numerical examples that
this choice is a robust one in many situations, holding for a wide range of deformations. A

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
4 A. TEN EYCK AND A. LEW

similar behavior was observed for the analog stabilization strategy proposed for DG methods
in [14]. Based on the proposition, it is possible to see that for large enough values of the
stabilization parameters a stable method will be recovered. While this feature is shared by
other stabilization strategies for ES methods in the literature, this seems to be the first
time an explicit albeit simple proof is provided. Finally, we note that the construction of
the stabilization term requires solving for the minimum eigenvalue of the elastic moduli at
each quadrature point in the mesh. This computation can be performed very efficiently with
off-the-shelf eigenvalue solvers, such that the overwhelming majority of the computational
time is still spent in solving the linear systems of equations. The stabilization strategy is then
applied to the nonlinear elasticity problem by formulating an incremental variational principle
along a prescribed loading path, see Sec. 7. For small enough loading steps, a stabilization term
adapted to the solution of the last computed load provides stable discretizations to obtain the
solution upon the next load increment. The adoption of an incremental variational principle
introduces a mild loading-path dependency of the numerical solution that can be easily removed
if desired, as illustrated in [14]. We demonstrate the performance of the stabilization strategy
with numerical examples in two- and three-dimensions, see Sec. 8. These include cases in which
common ES elements are unstable without the adaptive stabilization and stable with it, as
well as examples showing that the stabilization strategy does not significantly deteriorate the
performance in the face of volumetric or shear locking.
Because of the close resemblance between the DG and ES methods, throughout the paper
special sections highlighting the relations between the two have been included. These also
serve to motivate the stabilization strategy we introduce here. The paper begins next by
briefly stating the variational formulation of the nonlinear elasticity problem in Sec. 2, by
reviewing the definition of the ES derivative in Sec. 3, and the variational formulation of the
ES method in Sec. 4. Throughout the paper k · k denotes a norm on a finite dimensional space,
k · ks,Ω stands for the W 2,s (Ω)-norm, and we write | · |s,Ω for the W 2,s (Ω) s-seminorm.

2. Variational formulation of the nonlinear elasticity problem

We begin by describing the variational formulation of the nonlinear elasticity problem. Consider
a body that, when unstressed, occupies a connected open domain B0 ⊂ Rd , the reference
configuration with d = 2, 3. The deformation of the body is described through the deformation
mapping ϕ : B0 7→ Rd . The body deforms due to the presence of body forces f on B0 , tractions
T on ∂τ B0 ⊆ ∂B0 , and the prescribed deformation mapping ϕ on ∂d B0 ⊆ ∂B0 . Here ∂τ B0 and
∂d B0 are relatively open sets in ∂B0 that stand for the Neumann and Dirichlet boundaries,
respectively, and are assumed to satisfy ∂τ B0 ∩ ∂d B0 = ∅, ∂τ B0 ∪ ∂d B0 = ∂B0 and ∂d B0 is
nonempty.
The mechanical behavior of the material is assumed to be hyperelastic and defined by a
strain energy density W : Rd×d 7→ R, which depends on the deformation gradient ∇ϕ (see,
e.g. [16]). Under these conditions the potential energy of the body when deformed with a
deformation mapping ϕ is given by the functional
Z Z Z
I[ϕ] = W (∇ϕ) dV − T · ϕ dS − f · ϕ dV. (1)
B0 ∂τ B0 B0

The nonlinear elasticity problem consists in finding those deformations for which the body is in

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
AN ADAPTIVE STABILIZATION STRATEGY FOR ENHANCED STRAIN METHODS 5

mechanical equilibrium. These are the deformation mappings ϕ that satisfy ϕ(X) = ϕ(X) for
X ∈ ∂d B0 and that render the potential energy stationary with respect to all admissible
variations within some suitable set of admissible deformations. More precisely, the first
variation of the potential energy,
Z Z Z
hδI, δϕi = P(∇ϕ) : ∇δϕ dV − T · δϕ dS − f · δϕ dV, (2)
B0 ∂τ B0 B0

should be identically equal to zero for all admissible variations δϕ. Here P = ∂W
∂F (F), is the
first Piola-Kirchhoff stress tensor.
For the sake of simplicity and because they are not essential to the upcoming discussion,
we will henceforth assume that displacements are identically equal to zero on the Dirichlet
boundary, i.e., ϕ(X) = X, for any X ∈ ∂d B0 .

3. Definition of the enhanced strain derivative

The formulation of enhanced strain methods begins with the definition of the way derivatives
are enhanced. This has traditionally been done weakly via the Hu-Washizu variational
principle, or strongly by simply prescribing its value. We follow the latter approach next.
Let Th be a finite element mesh over the reference configuration B0 . The ES method
constructs an approximation vh to v ∈ H∂1 (B0 ) within a finite dimensional space Vh ⊂ H∂1 (B0 ),
where
H∂1 (B0 ) = v ∈ H 1 (B0 ) : v = 0 on ∂d B0 .

(3)
In contrast, the derivative of v, ∇v, is approximated within a finite dimensional space Wh . This
E
space takes the form Wh = ∇Vh ⊕ Wenh, so that ∇Vh ∩ Wenh = {0}. Here Wenh = ΠE∈Th Wenh ,
E
where Wenh denotes the space of enhancements in element E. Henceforth, we will indicate
elements of Wenh in the text with an overline, to facilitate their identification. The ES derivative
DES (·) : Vh × Wenh → Wh for any given vh ∈ Vh and specific enhancement wh ∈ Wenh is
defined as
DES (vh , wh ) = ∇vh + wh . (4)
Notice that in the absence of enhancements the approximation of ∇v reduces to that of a
conforming method, i.e., DES (vh , 0) = ∇vh . Additional conditions, not mentioned here, are
E
needed on the form of Wenh to guarantee consistency and convergence of the method, see
[17, 18].
A standard ES method over meshes of quadrilaterals, known as Q1/E4 [17], is constructed
by adopting
Vh = vh ∈ H∂1 (B0 ) : vh |E ∈ Q1 (E), ∀E ∈ Th

(5)
to approximate each component of the displacement field, and
n o
E
Wenh = span NE,enh
1 , N E,enh
2 (6)

to enhance the approximation of the gradient of each component of the displacement field.
Here NE,enh
i satisfies that
det ∇φE (0, 0)
NE,enh
i (φE (ξ1 , ξ2 )) = [∇φE (0, 0)]−T ∇N̂ienh (ξ1 , ξ2 ), i = 1, 2 (7)
det ∇φE (ξ1 , ξ2 )

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
6 A. TEN EYCK AND A. LEW

where N̂ienh (ξ1 , ξ2 ) = ξi2 /2 and φE : [−1, 1]2 → E is the standard isoparametric map
with bilinear shape functions for a quadrilateral. The element Q1/E9 over a hexahedron
is constructed in an analog way, see [17]. In either case, full integration of the elemental
contributions is considered here.

Relation to some discontinuous Galerkin methods. Some discontinuous Galerkin


methods can be cast in a form that enables the definition of a DG derivative operator, see
e.g. [2]. In this case, Vh,DG and Wh,DG are finite dimensional spaces of functions that may
be discontinuous across element boundaries, and are required to satisfy that ∇Vh,DG |E =
Wh,DG |E for any element E. Let Γ = ∪E∈Th ∂E. The DG derivative DDG (vh ) : Vh,DG → Wh,DG
is defined as
DDG (vh ) = ∇vh + R([[vh − v̂h (vh )]]) + L({vh − v̂h (vh )}) in every element E (8)
where R and L are two linear operators that map functions defined on Γ to Wh,DG , [[·]] and {·}
are the jump and average operators along Γ§ , respectively, and v̂h (vh ) is the so-called numerical
flux, which maps the values of vh on both sides of Γ to a (possibly bi-valued) function on Γ.
If the numerical flux is consistent, i.e., if v̂h (vh ) = vh whenever vh is continuous, then R and
L are different than zero if and only if vh is discontinuous. In particular, in [2] we adopted the
numerical flux from [19], i.e., we set v̂h (vh ) = {vh } , in which case the DG derivative adopts
the simpler form
DDG (vh ) = ∇vh + R([[vh ]]) in every element E. (9)
For simplicity, we shall only consider this particular form in subsequent discussions about the
relations between the two methods, ES and DG.
Comparing (8) to (4) reveals the similarities between the two methods. The discontinuities
across element boundaries induce enhancements in the approximation of derivatives through
R and L. The choices of the numerical flux v̂h and the space Wh,DG are the tools through
which the type of enhancements in these DG methods are decided, inasmuch the choice of
Wenh does that for ES methods.
A standard example of one such DG method is constructed by adopting (9) for the
approximation of derivatives and Vh,DG = Wh,DG = ΠE∈Th Q1 (E). Essential boundary
conditions may be imposed either strongly, as restrictions on Vh,DG , or weakly as discontinuities
along the boundary, and hence accounted for in the value of R and L.

4. Variational formulation of the ES method

With the definition of the enhanced strain derivative, the formulation of the ES method is
straightforward. The ES method seeks a pair (ϕh , Fh ) ∈ Vhd × Wenh
d
that renders the discrete
potential energy functional
Z Z Z
IES [ϕh , Fh ] = W (DES (ϕh , Fh )) dV − T · ϕh dS − f · ϕh dV (10)
B0 ∂τ B0 B0

§ with special definitions along ∂B0

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
AN ADAPTIVE STABILIZATION STRATEGY FOR ENHANCED STRAIN METHODS 7

stationary in Vhd × Wenh


d
. Equivalently, the first variation

hδIES , (δϕh , δFh )i


Z Z Z
= P(DES (ϕh , Fh )) : δDES (ϕh , Fh ) dV − T · δϕh dS − f · δϕh dV (11)
B0 ∂τ B0 B0

should be identically zero for any admissible variation (δϕh , δFh ), where δDES (ϕh , Fh ) =
∇δϕh + δFh . The same method can be obtained starting from the Hu-Washizu variational
principle, see e.g. [17].
Under suitable conditions, it was shown in [18] that in the linear elastic case ϕh converges
in H k (B0 ) with optimal rate to ϕ ∈ H k+1 (B0 ), for k ≥ 1. Additionally, Fh converges to
zero in H k−1 (B0 ) at the same rate than ∇ϕh − ∇ϕ, so both ∇ϕh and DES (ϕh , Fh ) are
approximations to ∇ϕ of the same order.

Relation to some discontinuous Galerkin methods. A class of discontinuous Galerkin


methods for nonlinear elasticity is obtained in a similar way [2]. In this case, we seek a
d
stationary point ϕh ∈ Vh,DG of the potential energy functional
Z Z Z
IDG [ϕh ] = W (DDG (ϕh )) dV − T · ϕh dS − f · ϕh dV (12)
B0 ∂τ B0 B0
d
in Vh,DG .Numerical examples computed with this method are reported in [14, 15, 2] for the
nonlinear elastic case. Discontinuous Galerkin methods for nonlinear elasticity can also be
derived from a Hu-Washizu-like variational principle, see [2, 20].

5. Stabilization strategies for enhanced strain methods

We discuss next the stability of enhanced strain methods for a linearized elasticity problem.
The adaptive stabilization strategy will be formulated for this case, and it serves as the basis
of the strategy for the nonlinear elastic case.
In the absence of tractions and with homogeneous boundary conditions, the linearized
d
elasticity problem from (2) consists in finding u ∈ H01 (B0 ) (for ∂d B0 ≡ ∂B0 ) such that

Z Z
B(u, v) := ∇u : A : ∇v dV = f · v dV, (13)
B0 B0
d
for all v ∈ H01 (B0 ) , where A = ∂ 2 W/∂F2 is the spatially-varying fourth order elasticity


tensor. If kA(∇ϕ(X))k ∈ L∞ (B0 ) and the domain B0 has a smooth boundary (C 1 ), then this
problem has a unique solution provided the bilinear form B is coercive in [H01 (B0 )]d (see, e.g.,
Thm. 4.2 in [21]). More precisely, provided that there exists κ > 0 such that
Z
2
∇u : A : ∇u dV ≥ κ kuk1,B0 (14)
B0

for all u ∈ [H01 (B0 )]d . Furthermore, if the elastic moduli A and the domain are smooth enough
then u ∈ [H 2 (B0 ) ∩ H01 (B0 )]d , see [21]. If (14) is satisfied, the linearized elasticity problem is
said to be stable.

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
8 A. TEN EYCK AND A. LEW

Consider next a conforming finite element method to approximate this problem that consists
d
in finding uh ∈ Vh,conf ⊆ H01 (B0 )d such that
Z
B(uh , vh ) = f · vh dV (15)
B0
d d
for all vh ∈ Vh,conf . Equation (14) is in particular valid for all uh ∈ Vh,conf , and hence it
automatically provides a stability estimate for the method.
This “inherited stability” does not hold for ES methods. The ES method in the linearized
elasticity case follows by adopting the following strain energy density function, i.e.,
1 T 
W (F) = F − I : A : (F − I) (16)
2
where I is the identity tensor in Rd×d . Consequently, the ES method (11) consists in finding
(uh , zh ) ∈ Vhd × Wenh
d
such that
Z Z
Bh (uh , zh , vh , wh ) := DES (uh , zh ) : A : DES (vh , wh ) = f · vh dV, (17)
B0 B0

Vhd
for all (vh , wh ) ∈ ×d
Wenh . Notice that the function DES (uh , zh ) is not required to be the
distributional derivative of any function in [H01 (B0 )]d . Consequently, (14) does not imply the
coercivity of Bh in Vhd × Wenh d
as it did for the bilinear form of the conforming method. Since
(14) does imply the coercivity of Bh in Vhd × {0}, it follows that any numerical instability will
necessarily involve nonzero enhancements.
The ES method is said to be unstable whenever Bh is not coercive in Vhd × Wenh d
while
1 d
B is coercive in [H0 (B0 )] . Roughly speaking, we would like the discretization to be stable
whenever the exact problem is. This condition may be relaxed by requesting Bh to be coercive
in Vhd × Wenh d
for h small enough if B is coercive in [H01 (B0 )]d .
It is important to recall that the elastic moduli A at any point in the domain, when
regarded as linear operator over the space of d × d second-order tensors, do not need to
be positive definite. For example, in the standard isotropic linear elastic case A : ω = 0 for any
skew-symmetric second-order tensor ω. In more general cases, A may have several negative
eigenvalues, a situation often encountered in the linearized problem formulated on a body
subjected to predominantly compressive stresses, such as hydrostatic compression. Even if A
has negative eigenvalues everywhere in the domain, (14) may still hold. For example, in the
standard isotropic linear elastic case, coercivity follows from one of Korn’s inequalities.
This has important consequences for the design of stabilization methods. For instance, it
means that it is not possible to simply examine the elemental stiffness matrices and, barring
rigid body motions, expect them to be positive definite.
Out of a number of stabilization strategies proposed for ES methods, we comment here on
two [9, 3] that have common features with the one we propose in the next section. These
solve the linearized problem in (17) but with a modified bilinear form Bh . They add to
Bh (uh , zh , vh , wh ) the term Z
zh : Astab : wh dV, (18)
B0

where Astab is a spatially-varying fourth-order tensor. If properly chosen, this extra term adds
additional stiffness to some or all the enhanced strains, hence acting as a stabilization term to
restore coercivity to Bh .

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
AN ADAPTIVE STABILIZATION STRATEGY FOR ENHANCED STRAIN METHODS 9

In [3] the simplest choice Astab = βI is made, where β ≥ 0 is a stabilization parameter and I is
the fourth order identity tensor, i.e., I : g = g for all g in Rd×d . The beauty of this formulation
is its simplicity. On the other hand, the value of β strongly depends on the material model,
the geometry of the problem, and most importantly, on the loading conditions. Consequently,
when considering a quasistatic loading program, the value of β needs to be progressively
tuned at each loading step. This is possible, but clearly not very efficient. Additionally, the
uniform penalization of enhanced modes throughout the entire domain often results in some
computational efficiency loss and in a higher propensity to observe locking, at least when an
analog approach is adopted for a discontinuous Galerkin method [14, 15].
In [9] the choice of Astab is (essentially)
n
X
Astab = τi (∇ϕ) PT
i : A : Pi . (19)
F=I
i=0
d d
Here Pi : Wenh → Wenh is a projection operator over the space of strain enhancements, n is
the desired number of these projections, and τi (∇ϕ) is a scalar function of the deformation
gradient. The authors provide specific expressions of each Pi and τi for a specific choice of ES
element and material model. This concept and idea is very general, and relatively simple to
implement. However, while guiding principles to construct the stabilization parameters τi for
any material model and enhanced strain element are discussed, a general procedure has yet to
be formulated.
Finally, we also mention the stabilization technique in [7]. Building on hourglass stabilization
techniques for reduced integration methods, the method also introduces a term to add stiffness
to the enhanced modes that is sensitive to the local deformation, and it is scaled by a
stabilization parameter. It has the appealing feature of utilizing only one quadrature point
per element. However, the construction of the stabilization term is quite sophisticated.
As it is common to all of these methods, the design challenge lies in adding enough stiffness
to the enhanced modes to avoid instabilities. Too much added stiffness, however, produces the
strain enhancements to contribute very little to the resulting solution. In fact, when the added
stiffness is too large spurious effects such as incompressible and shear locking creep back into
ES methods, e.g., [7, 3].

Relation to some discontinuous Galerkin methods. The formulation of the


discontinuous Galerkin method in [22, 2] for linearized elasticity consists in finding uh ∈ Vh,DG
such that Z Z
Bh,DG (uh , vh ) := DDG (uh ) : A : DDG (vh ) = f · vh dV, (20)
B0 B0

for all vh ∈ Vh,DG .


Stabilization strategies for DG methods are essentially based on adding stiffness to the
discontinuities across element boundaries, see e.g. [23]. Under mild conditions, this can be
accomplished by adding to Bh,DG a term of the form
Z
R([[uh ]]) : Astab : R([[vh ]]) dV, (21)
B0

for a suitably chosen fourth-order tensor Astab . This follows from a result in [24], in which
it is shown that the value of R([[uh ]]) is in a one-to-one correspondence with [[uh ]] for some

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
10 A. TEN EYCK AND A. LEW

triangular and quadrilateral elements in two-dimensions. This result is easily generalizable to


three-dimensions.
The adaptive stabilization strategy in [14, 15] follows this format, as we describe next. To
this end, for any point X ∈ B0 set
λX : = λ(A(F(X))), (22)
where  
g:A:g
λ(A) : = max 0, − min . (23)
06=g∈Rd×d g:g
If A(F(X)) as a linear operator in Rd×d is positive semi-definite then λX is 0. In contrast, if
A(F(X)) has one or more negative eigenvalues then λX is equal to the absolute value of its
smallest eigenvalue. The description of the stabilization term is completed by choosing
Astab = β1 A + β2 λX I + β3 κ∗ I, (24)
for some stabilization parameters β1 , β2 , β3 ≥ 0, where κ∗ > 0 is some constant of the order
of some characteristic stiffness of the material placed to make β3 a non-dimensional quantity.
Notice that if β1 < β2 and β3 > 0 then Astab is a positive definite linear operator in Rd×d
and (21) adds a positive stiffness to the discontinuities, i.e., acts as a stabilization term. In
[15] β2 was allowed to depend on X, or to change from element to element, and lower bounds
for each one of the stabilization parameters were provided that guarantee that the stabilized
bilinear form is coercive. However, not only was the calculation of these values computationally
costly, but it was also observed therein that the provided lower bounds were sometimes greatly
overestimated. Instead, the numerical examples in [14] suggest that it is very often possible
to obtain a coercive bilinear form by choosing β1 = β3 = 0 and β2 = O(1). From a practical
perspective, this resulted in a very robust method. However, there are cases in which this is not
sufficient, and larger values of β1 , β2 and β3 need to be selected. In this context, setting β1 = 0
is often convenient, since the computation of the resulting term involves fewer operations.

6. Adaptive stabilization of ES methods for linearized elasticity

The adaptive stabilization strategy for discontinuous Galerkin methods outlined in the last
section can be immediately translated into ES methods. The idea consists in combining the
stabilization term (18) with the choice of Astab in (24). Precisely, the ES method for the
linearized elasticity problem with stabilization consists in finding (uh , zh ) ∈ Vhd × Wenh
d
such
that

Bstab
h (uh , zh , vh , wh ) :=
Z Z
DES (uh , zh ) : A : DES (vh , wh ) + zh : (β1 A + β2 λX I + β3 κ∗ I, ) : wh
B0 B0
Z
= f · vh dV, (25)
B0

for all (vh , wh ) ∈ Vhd × Wenh


d
.
The resulting method is guaranteed to be stable provided the choice of the stabilization
parameters β1 , β2 and β3 is done according to the following proposition, the proof of which is

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
AN ADAPTIVE STABILIZATION STRATEGY FOR ENHANCED STRAIN METHODS 11

given in the appendix. The stabilization parameter β2 is allowed to be a function of the spatial
position X, indicated next with the notation β2X .
Theorem 6.1. Let λX be defined as in (22) and set
Λ = kλX k∞,B0 .
Assume that Λ < ∞, and that for some κh > 0
Z
∇vh : A : ∇vh dV ≥ κh kvk21,B0 for all vh ∈ Vhd ⊆ [H∂1 (B0 )]d . (26)
B0

For any κh ∈ (0, κh ), let η > 0 be such that


κh − κh
η≤ , (27)
κh + Λ
let the stabilization parameters β1 and β3 be such that
1
β1 ≥ − 1, β3 κ∗ ≥ κh , (28)
η
and let the spatially-varying stabilization parameter β2X be such that
λX
β2X ≥ β1 + 1 + (29)
κh − κh − η(κh + λX )
for almost every X ∈ B0 . Then,
2 2
Bstab
h (vh , wh , vh , wh ) ≥ κh (|vh |1,B0 + kwh k0,B0 )

for all (vh , wh ) ∈ Vhd × Wenh


d
.
Notice that (26) requires the conforming discretization in Vhd × {0} to be stable. This
condition is automatically satisfied if the exact linearized elasticity problem is stable, i.e.,
satisfies (14), since necessarily κh ≥ κ. This is the “inherited” stability of a conforming
discretization. Consequently, if the exact problem is stable, and the stabilization parameters
satisfy the conditions in Theorem 6.1, the ES method is stable.
If Λ = 0, i.e., the elastic moduli have non-negative eigenvalues almost anywhere in B0 , then
the value of β2 is unimportant, and β1 and β3 can be chosen arbitrarily small for Bstab h to be
coercive. It is interesting to compare this with the result of Reddy and Simo [18]. Therein,
the elastic moduli were assumed to satisfy Aijkl = Ajikl = Aklij and to constitute a positive
definite linear operator on the set of symmetric second-order tensors. Under these conditions
it is clear that Λ = 0, since A : ω = 0 for any skew-symmetric second order tensor ω. However,
they showed that no stabilization is required, since the maximum angle between members of
Wenh and the symmetric parts of ∇Vh is required to be bounded away from π/2 independently
of h, as measured with an L2 -projection (see condition (II) in [18], pp. 1712). We did not make
use of this property in our proof, since it is not apparent that an equivalent condition in the
more general case dealt with here can be formulated. Consequently, the lower bounds for the
stabilization parameters are overestimated in many situations. Numerical experiments with a
similar result for DG methods [15] confirmed that this is the case.
Very large values for the stabilization parameters result in numerical solutions in which
the strain enhancements do not play any substantial role. Consequently, if the conforming

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
12 A. TEN EYCK AND A. LEW

discretization resulting from seeking solutions in Vh × {0} displays some form of locking, the
same features will be observed in the ES method when the stabilization parameters are overly
large.
A heuristic alternative that we found to be very robust, and whose performance is
demonstrated in the numerical examples next, consists in choosing β1 = β3 = 0, and β2 = O(1).
Occasionally, additional stiffness in regions where λX = 0 is also needed. A positive value for
β3 is then chosen, leaving for simplicity β1 = 0. It is then easy to see that, given h, Bstab h
will be coercive for β3 large enough. One way to see this, albeit not the only one, stems from
noticing that for κh > 0, Theorem 6.1 provides minimum values for β1 , β2 and β3 for Bstab h to
be coercive. It is then possible to choose β2∗ , β3∗ > 0 such that
Z Z

zh : (β1 A + β2 λX I + β3 κ I, ) : zh dV ≤ zh : (β2∗ λX I + β3∗ κ∗ I, ) : zh dV (30)
B0 B0

d
for all zh ∈ Wenh , given that the square roots of both sides of the inequality multiplied by any
d
positive scalar are norms in the finite dimensional space Wenh , and hence are equivalent.
A rough explanation of why β2 should be of order 1 in many circumstances is that in this
way the value of stabilization term is comparable to that of the rest of the bilinear form for
all unstable modes, due to the presence of λX . Larger values of β2 would generally make the
former much larger than the latter, and hence make all the enhanced degrees of freedom stiff,
preventing them from playing an important role in the numerical approximation. Values of β2
of order 1 were adopted in all the numerical examples that follow, and stable dicretizations
were recovered. Of course, there are likely situations in which β2 needs to be larger, or in which
no stability can be accomplished unless β3 is greater than zero.
That the choice of the stabilization parameters do not strongly depend on the particular
material model is a key advantage of this method. Additionally, the addition of stabilization
terms mostly in regions where λX > 0 aims at minimizing the stiffness added to the enhanced
modes. For DG discretizations, this has proved to enhance the accuracy and robustness of
solutions, see [14]. For example, it is much more difficult for the method to display locking as
a consequence of a poor choice of a stabilization parameter. We have not conducted similar
studies for ES stabilization strategies.

7. Application to nonlinear elasticity problems

We next apply the proposed stabilization strategy to the nonlinear elasticity problem. To
this end, we construct a quasistatic loading path given by a continuous one-parameter family
of body forces and imposed tractions on the boundary. This path takes the body from the
reference, unstressed configuration to the deformed one under the prescribed loads. Known also
as the incremental method [25], the consideration of loading paths can ease the computation
of strongly nonlinear problems.
Let then f k and Tk for k = 0, . . . , n be the values of body forces and imposed tractions
obtained by sampling the loading path, with f 0 = T0 = 0, f n = f and Tn = T. For example,
it is often convenient to choose f k = (k/n)f and Tk = (k/n)T. The proposed method for
nonlinear elasticity is based on a sequence of incremental variational principles to determine
k
(ϕkh , Fh ) ∈ Vhd × Wenh
d
, k = 0, . . . , n, the sequence of ES approximations to the deformation

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
AN ADAPTIVE STABILIZATION STRATEGY FOR ENHANCED STRAIN METHODS 13

k k+1
mapping through the loading path. Given (ϕkh , Fh ), k = 1, . . . , n − 1, we find (ϕk+1h , Fh ) as
a stationary point of the functional
Z Z Z
k+1
IES [ϕh , Fh ] = W (DES (ϕh , Fh )) dV − Tk+1 · ϕh dS − f k+1 · ϕh dV
B0 ∂τ B0 B0
1
Z
Fh : β1 Ak + β2 λkX I + β3 κ∗ I : Fh dV. (31)

+
2 B0

among all (ϕh , Fh ) ∈ Vhd × Wenh


d
, where
  k

Ak = A DES ϕkh , Fh (32)

λkX λ Ak

= . (33)
X
0
The definition is completed by specifying ϕ0h (X) = X and Fh = I. The dependence of the
k+1 k
stabilization term in IES on (ϕkh , Fh ) is what makes it an incremental variational principle.
k+1
Notice as well that IES reduces to IES in (10) when β1 = β2 = β3 = 0, and the standard ES
method is recovered.
With this formulation and the suitable choice of the stabilization parameters described
k+1 k
earlier, the method given by IES is linearly stable at (ϕkh , Fh ), provided the linearized
elasticity problem therein is stable. More precisely, under these conditions the bilinear form
k+1 k
corresponding to the second variation of IES at (ϕkh , Fh ), i.e.,
D h k
i E
k+1
δ 2 IES ϕkh , Fh , (uh , zh ) , (vh , wh ) =
Z
= DES (uh , zh ) : Ak : DES (vh , wh ) dV
B0
Z
zh : β1 Ak + β2 λkX I + β3 κ∗ I : wh dV, (34)

+
B0

is coercive in Vhd × Wenh


d
. This follows from the discussion in Sec. 6 because (34) is the bilinear
k
form of the ES method for the linearized elasticity problem at (ϕkh , Fh ).
The adoption of a loading path enables the method to constantly adapt to the evolving
material properties. Therefore, if the linearized elasticity problem at each deformation in a
k
neighborhood of (ϕkh , Fh ) is linearly stable, and the size of the loading steps are chosen small
enough, the method should remain linearly stable at each step of a Newton-Raphson iteration.
Furthermore, under the same conditions the method should be linearly stable at a stationary
k+1
point of IES as well, provided that the solution at step k + 1 is not too far from that at step
k. This was in fact the case for all the following numerical examples.
The use of Ak and λkX in the stabilization term induces certain loading-path dependency
k+1
in the value of the solution (ϕk+1
h , Fh ). This is because both Ak and λkX are computed at
k
configuration (ϕkh , Fh ).
However, this dependency can easily be removed as mentioned in [8]
for ES methods, and in [14] for DG methods.
As we shall see below in the numerical examples, the proposed method is remarkably robust
with respect to the choice of stabilization parameters. By adopting β2 = O(1) and β1 = β3 = 0,

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
14 A. TEN EYCK AND A. LEW

we were able to avoid the appearance of instabilities even when the body was undergoing quite
large deformations, with different material models.
Finally, it is worth mentioning that the notion of a linearly stable method embraced here
may not be needed to obtain some form of convergence of the method. For example, in [26]
the convergence of a class of DG methods for a simplified nonlinear elastic model requires only
that sets of deformations with bounded potential energy be compact. An equivalent analysis
in the case of ES methods is yet to be performed.

8. Numerical results

In Sec. 8.1 we analyze the performance of the method with selected two- and three-dimensional
problems. In the first subsection we study how the choice of stabilization parameters affects
the overall coercivity of an adaptively stabilized ES formulation. Sec. 8.2 and 8.3 examines
the behavior of stabilized ES methods for problems subject to volumetric and shear locking,
respectively.
To test the performance and to showcase that the method does not depend on the choice of
the material model, we will perform numerical examples with both a compressible neo-Hookean
material and a compressible Ogden material. The strain energy density for a compressible neo-
Hookean material is
tr(FT F − I)
 
E ν 2
W (F) = log(det(F)) − log(det(F)) + , (35)
2(1 + ν) 1 − 2ν 2
where E and ν are material properties. All forthcoming simulations that make use of a neo-
Hookean model will adopt the value of E = 1 N/m2 . We also consider an Ogden model for
rubber. The strain energy density for the Ogden material is
ρ
W (F) = 2 9 log(det(F)) + det(F)−9 − 1

9
3 3
X X µp αp α α 
− µp log(det(F)) + λ1 + λ2 p + λ3 p − 3 (36)
p=1
α
p=1 p
where λ1 , λ2 and λ3 are the principal stretches of F, i.e., the square roots of the eigenvalues of
FT F. We chose the coefficients as in [27], namely, µ1 = 6.3 × 105 N/m2 , µ2 = 1.2 × 103 N/m2 ,
µ3 = −104 N/m2 , α1 = 1.3, α2 = 5 and α3 = −2. By adjusting the value of ρ it is possible to
tailor the compressibility of the material, so its value will be specified later.
The choice of two sets of material constants differing by several orders of magnitude serves
to illustrate that, because of the way the method is formulated, the value of the stabilization
parameters do not depend on the units chosen to solve the problem. In other words, the
stabilization parameters are dimensionless.
In the following examples we will compare the ES method with adaptive stabilization defined
in Sec. 7, with a conforming method obtained by seeking stationary points of IES (see (10))
in Vhd × {0}, where Vh is made of Q1 elements as defined in (5)¶ . The specification of the ES

¶ Standard modifications to the definition of the spaces to account for the more general boundary conditions
adopted in the forthcoming examples should of course be made.

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
AN ADAPTIVE STABILIZATION STRATEGY FOR ENHANCED STRAIN METHODS 15

E E
method is completed by detailing the choice of Wenh . In two-dimensions, we chose Wenh to be
E
defined as in (6), so that Q1/E4 elements are used. In three dimensions, we chose Wenh so as
to obtain Q1/E9 elements, which are the natural extension of Q1/E4 elements to hexahedra;
see [17] for their definition.
For simplicity, all length quantities have been intentionally stripped of their units, but they
are expressed in meters. Contour plots of the values of λX shown in this section were obtained
by plotting the contour plot of a Q1 nodal interpolant. Its value at each node is that of λX at
the nearest quadrature point. The same node may have different values on different elements,
and hence the interpolant is not necessarily continuous across element boundaries.

8.1. Coercivity
We examine two boundary value problems for which a conforming method is coercive, but an
enhanced strain method is not. These problems have been extensively studied in [8, 3]. The
domain of the first problem is the two-dimensional block B0 = (−1, 1) × (0, 1). A nonzero
constant force per unit length normal to the face is imposed on (−0.75, 0.75) × {1}, while the
rest of the upper face remains traction free. The two side faces, {±1} × (0, 1) also remain
traction free. On the bottom surface, (−1, 1) × {0}, displacements normal to the face and
tractions tangent to it are set to zero. Taking advantage of the symmetry of the problem with
respect to the axis {0} × (0, 1), only the right half of the domain is simulated. Consequently,
normal displacements and tractions tangential to the surface {0} × (0, 1) are set to zero.
In the following, a deformation (ϕh , Fh ) ∈ Vhd × Wenh d
is said to be linearly stable if and
d
only if the bilinear form (34) is coercive in (a) Vh × {0} for the conforming method and, (b)
Vhd × Wenh
d
for the ES method. Of course, for a conforming method only deformations in which
Fh = 0 are possible. Testing the linear stability of a deformation amounts to evaluating the
sign of the smallest eigenvalue of the stiffness matrix of the method at the given deformation;
stable deformations have a positive smallest eigenvalue. This is how we concluded on the
stability of deformations below.
Under these loading conditions, a P block made of either a neo-Hookean material with ν = 0.25
3
or an Ogden material with ρ = p=1 µp , displays linear instabilities for a large enough
compressive force per unit length on the top face. In the following examples we selected the
compressive load to be small enough such that the conforming approximation is stable, yet
large enough so that we observe the (numerical) instabilities arising from the ES method with
no stabilization.
For the conforming method, normal compressive forces per unit length resulted in a loss of
linearized stability at the resulting deformation whenever these were in excess of ∼ 0.33 N/m
when the block is made of a neo-Hookean material, and of ∼ 7.8 × 105 N/m when it is built
with the Ogden material instead. In contrast, for the ES method with no stabilization the
loss of stability of the linearized elasticity problem occurred for normal compressive forces per
unit length above ∼ 0.22 N/m (neo-Hookean) and ∼ 3 × 105 N/m (Ogden). These values were
obtained for the meshes shown in Figs. 1, and were confirmed to be essentially the same for
meshes with one-half the mesh size.
Figure 1 shows the reference and deformed configuration of the block made of the neo-
Hookean material for a compressive load slightly larger than that needed for the onset of
the instability in the ES method. The values of λX at the configuration shown are indicated
with a contour plot. Notice that the classical pattern of hourglass-shaped elements seen when

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
16 A. TEN EYCK AND A. LEW

instabilities set in can be clearly identified in the enlarged area. A similar result was found for
the block made of the Ogden material.

0.3
0.26
0.22 8
0.18
0.14
0.1
0.06
0.02

6
1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Figure 1. Example of a nonlinear elastic deformation for which the ES method displays a numerical
instability, which disappears once the adaptive stabilization term is included. The colors are the
contour plots of λX and the material is neo-Hookean. Regions in white have exactly λX = 0. The
reference configuration is shown on the left, while the resulting deformed configuration when Q1/E4
elements are adopted, a force per unit length of ∼ 0.22 N/m is imposed, and no adaptive stabilization
term is added is displayed in the middle. The latter displays numerical instabilities, a fact we confirmed
by computing the minimum eigenvalue of the stiffness matrix at that deformation. These instabilities
are also manifested in the hourglass shape some of the elements adopt, as highlighted in the figure.
Notice as well that the location of these instabilities coincides with that of the maximum values of
λX . Once the adaptive stabilization term is added, in this case with β2 = 2 and β1 = β3 = 0,
the deformation of the block under the same load is no longer unstable, and the hourglass-shaped
elements disappear. The figure on the right shows the deformed configuration obtained with the
stabilized method at the end of the loading path. This corresponds to a force per unit length of 0.28
N/m. At this larger load the resulting configuration is stable.

We next consider the class of adaptively stabilized ES methods in which β1 = β3 = 0, and


study the dependence of the onset of the numerical instability on the value of β2 . To this
end, we plot the eigenvalue of the stiffness matrix that corresponds to bilinear form (34) as a
function of β2 , for a fixed prescribed load on the upper surface. The value of this eigenvalue is
normalized by the smallest eigenvalue of the stiffness matrix that corresponds to the conforming
method, for the same prescribed load on the upper surface. Since Vhd × {0} ⊆ Vhd × Wenh d
, this
ratio is strictly less than or equal to one.
The chart in Fig. 2 shows the ratio of the eigenvalues as a function of β2 , measured at a final
force per unit length of 0.28 N/m when the block is made of the neo-Hookean material, and at
a final force per unit length of 6 × 105 N/m when it is made of the Ogden material. These two
configurations at the end of the loading path were linearly stable. The loading path consisted
of n = 30 uniform increments of the compressive force per unit length on the upper surface,
and the results were tested to be essentially insensitive to the choice of any larger n. Only non-
negative values for the eigenvalue ratio are shown. Any significant reduction in the minimum
value of β2 for which data is shown in each curve would result in a negative eigenvalue ratio,

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
AN ADAPTIVE STABILIZATION STRATEGY FOR ENHANCED STRAIN METHODS 17

and hence, in a deformation at the end of the loading path that is not linearly stable. Notice
that the smallest value of β2 needed to stabilize the problem is O(1), regardless of the fact
that the stiffness of the two materials vary by up to five orders of magnitude.
The deformation at the end of the loading path when β2 = 2 is shown in Fig. 1, on the right,
for a block made of the neo-Hookean material. Notice the stabilization term has penalized
the enhancements mostly in the region where the hourglass-shaped elements were observed.
However, no hourglass-shaped elements can be found in the solution with the stabilized
method.

0.1
Eigenvalue Ratio

0.01 neo-Hookean,Q1/E4
Ogden,Q1/E4

0.001
1 10 100 1000 10000
β2

Figure 2. The chart shows the minimum eigenvalue of the stiffness matrix of the ES method as a
function of the stabilization parameter β2 , in normalized units (see Sec. 8.1). All values shown were
obtained for a fixed load on the top face and the same mesh. Any significant decrease in the value of
β2 from the smallest value shown in each curve results in a negative eigenvalue, and hence in a loss
of linear stability of the ES method at this deformation. Notice that the minimum value of β2 that
renders a stable discretization for this problem is of order 1 for both material models. Finally, since
λX = 0 in parts of the domain, the eigenvalue ratio is not expected to asymptotically approach 1 as
β2 grows.

A similar set of results is found in an analogous three-dimensional problem. In this case the
reference configuration is the block B0 = (−1, 1) × (0, 1) × (0, 1). A constant, possibly non-zero
traction normal to the upper surface is applied on (−0.5, 0.5) × (0, 1) × {1}, while the rest of
the top face is traction free. Normal displacements and tangent tractions to the bottom face
(−1, 1) × (0, 1) × {0} are set to zero, while the remaining faces are traction free. Because of the
symmetry of the problem only the block (0, 1)3 is simulated, and hence symmetry boundary
conditions are imposed on the face {0} × (0, 1) × (0, 1).
Deformations computed with the conforming method ceased to be linearly stable for
compressive traction values exceeding ∼ 0.95N/m2 and ∼ 1.65 × 108 N/m2 for a block made
out of the neo-Hookean or Ogden material, respectively. When computed with the ES method
with no stabilization, the onset of instabilities was detected for compressive tractions larger
than ∼ 0.6 N/m2 for the neo-Hookean and ∼ 8.8 × 105 N/m2 for the Ogden.

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
18 A. TEN EYCK AND A. LEW

The reference and a deformed configuration of a block just after the onset of the instability
are shown in Fig. 3. In this case the block is made of an Ogden material, and the solution is
computed with no stabilization. The mesh for this example is made of rectangular hexahedra
in the reference configuration. In this case it is more difficult to tell whether elements have
adopted an hourglass shape or not. It is simple to see, however, that the most distorted
elements lie precisely where the value of λX is large, as in the two-dimensional example. The
computation of the minimum eigenvalue of the stiffness matrix indicates that this is an unstable
configuration.

5.0E+05
4.5E+05
4.0E+05
3.5E+05
3.0E+05
2.5E+05
2.0E+05
1.5E+05
1.0E+05
5.0E+04

1 1
0.8

0.8 0.8
0.6

0.6 0.6
0.4

0.4 0.4
0.2

1
1 0.2 1 0.2
0.8 0

0.6 0.8 0.8


1
0.4 0.8 0.6 0 0.6 0
0.6
0.2 0.4 1 0.4 1
0.4
0.8 0.8
0 0.2 0.2 0.6 0.2 0.6
0
0.4 0.4
0 0
0.2 0.2
0 0

Figure 3. Example of a three-dimensional nonlinear elastic deformation for which the ES method
displays a numerical instability, which disappears once the adaptive stabilization term is included. The
colors are the contour plots of λX and the material is Ogden. Regions in white have exactly λX = 0.
The reference configuration is shown on the left, while the resulting deformed configuration when
Q1/E9 elements are adopted, a traction of ∼ 8.8×105 N/m2 is imposed, and no adaptive stabilization
term is added is displayed in the middle. The latter displays numerical instabilities, a fact we confirmed
by computing the minimum eigenvalue of the stiffness matrix at that deformation. In contrast to the
two-dimensional case, it is more difficult to identify this numerical instability by the deformed shape
of some of the elements. The fact that this is a numerical instability follows after noticing that once
the adaptive stabilization term is added, in this case with β2 = 3 and β1 = β3 = 0, the deformation
of the block under the same load is no longer unstable. The figure on the right shows the deformed
configuration obtained with the stabilized method at the end of the loading path. This corresponds to
an imposed traction of ∼ 1.235×106 N/m2 . At this larger load the resulting configuration is stable.

We also studied the onset of the numerical instability as a function of β2 in this case, for
the adaptively stabilized method with β1 = β3 = 0. The normalized smallest eigenvalue at the
end of the loading path as a function of β2 for the two different material types are shown in
Fig. 4. The computation of the normalized eigenvalue is performed exactly as explained for
the two-dimensional case. The loading paths consisted of n = 30 uniform increments of the
compressive traction on the top surface, and the results were tested to be essentially insensitive
to the choice of any larger n. The maximum values of the traction at the end of the loading
paths are 0.8075 N/m2 and 1.235 × 106 N/m2 , for the neo-Hookean and the Ogden materials,
respectively. Only the positive smallest eigenvalues are shown. Instabilities are observed to
appear whenever β2 is chosen to be somewhat smaller than the smallest values shown for each

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
AN ADAPTIVE STABILIZATION STRATEGY FOR ENHANCED STRAIN METHODS 19

curve. Just as in the two-dimensional case, a value of β2 of order 1 is what the ES method for
this problem needs to render a stable discretization.
The deformation at the end of the loading path obtained with a stabilized ES formulation
with β2 = 3 is displayed in Fig. 3. As before, the contours of λX indicate that the stabilization
term is stiffer precisely where the most distorted elements were observed.

1
neo-Hookean,Q1/E9
Ogden,Q1/E9

0.1
Eigenvalue Ratio

0.01

0.001
1 10 100 1000
β2

Figure 4. The chart shows the minimum eigenvalue of the stiffness matrix of the ES method as a
function of the stabilization parameter β2 , in normalized units (see Sec. 8.1). All values shown were
obtained with the same mesh and a fixed load on the top face. Any significant decrease in the value of
β2 from the smallest values shown in each curve results in a negative eigenvalue, and hence in a loss
of linear stability of the ES method at this deformation. Notice that the minimum value of β2 that
renders a stable discretization for this problem is of order 1 for both material models. Finally, since
λX = 0 in parts of the domain, the eigenvalue ratio is not expected to asymptotically approach 1 as
β2 grows.

8.2. Volumetric locking


In this section we examine the performance of the ES method with adaptive stabilization
for materials that display a near-incompressible constitutive behavior. It is known that, for
isotropic and initially stress-free linear elastic materials, low order conforming methods display
volumetric locking as the ratio of bulk to shear modulus becomes arbitrarily large. In contrast,
ES methods do not lock, as proved in [28]. One of the challenging aspects of creating robust
stabilization strategies for ES methods is to preserve the desirable non-locking properties in
the context of nonlinear elasticity problems.
In the following example we study the behavior of the numerical solutions as the bulk
modulus of a material is progressively increased keeping all shear moduli bounded. In the limit
of infinitely large bulk modulus, a smooth deformation mapping should satisfy the nonlinear
constraint det(F) = 1 almost everywhere in B0 . It is simple to see from (35) and (36) that
regardless of the state of deformation, the bulk modulus scales with ν/(1 − 2ν) for a neo-

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
20 A. TEN EYCK AND A. LEW

P3
Hookean material, and with ρ in an Ogden material for a constant value of p=1 µp . We
P3
therefore define a scalar parameter ξ = ν/(1 − 2ν) for the former and ξ = ρ/( p=1 µp ) for the
latter, so that near incompressible behavior corresponds to large values of ξ in either material.
All simulations were performed with β1 = β3 = 0.
We first consider a two-dimensional example in which the reference configuration is B0 =
(0, 0.25) × (0, 1). Dirichlet boundary conditions are imposed on the surface (0, 0.25) × {0},
constraining the displacement field to be zero there. A constant force per unit length tangent
to the surface is imposed on the top face (0, 0.25) × {1}, with magnitude equal to 6.25 × 10−3
N/m for the neo-Hookean material and to 1.875 × 104 N/m for the Ogden material. The other
faces are traction free. The domain is partitioned into rectangular elements with edge lengths
h = 1/8 and h = 1/32, and both the ES elements Q1/E4 and the conforming elements Q1
are used. To compare, simulations for the Ogden material using a conforming triangular mesh
made of quadratic (P2) elements with h = 1/32 have also been performed, since these are not
known to lock under near incompressibility conditions.
The deformed configuration with ξ = 5.0 × 105 as well as the contour plot for λX obtained
with β2 = 1 are shown in Fig. 5(a), for a block made of a neo-Hookean material. The chart
in Fig. 5(b) shows the Euclidean norm of the displacement at the tip (0.25, 1) of the beam
normalized by the maximum value of the same quantity observed for each material, as a
function of ξ. The sharp drop in the value of the tip displacement as ξ grows clearly shows that
a low order conforming method displays volumetric locking for both materials, as expected. In
contrast, by keeping an essentially constant tip displacement for all values of ξ the adaptively
stabilized ES method does not display signs of locking.
In three-dimensions we consider a reference configuration B0 = (0, 0.25) × (0, 0.5) × (0, 1).
Dirichlet boundary conditions are imposed on the face (0, 0.25) × (0, 0.5) × {0}. On the upper
face (0, 0.25) × (0, 0.5) × {1} a constant traction tangent to the surface in the direction (1, 0, 0)
is applied. Its magnitude is set to be 1.0 × 10−2 N/m2 for the neo-Hookean material and
1.875 × 104 N/m2 for the Ogden material. The remaining faces are traction free. The domain
is partitioned into cubic elements with edge sizes h = 1/8 and h = 1/16 when a neo-Hookean
material is used, with both Q1 and Q1/E9 elements. For an Ogden material, an edge size equal
to h = 1/28 was chosen for Q1 elements, and equal to h = 1/16 when Q1/E9 is adopted.
The deformed configuration and the contour plot of λX are shown in Fig. 6 for a beam made
of an Ogden material, with the choice β2 = 2 and ξ = 1.0 × 104 . The chart in Fig. 6 shows
the Euclidean norm of the displacement of the tip (0.25, 0.5, 1) normalized by the maximum
value of the same quantity observed for each material. As before, the results obtained with
a conforming method clearly display volumetric locking, an effect that is ameliorated when
finer meshes are adopted. In contrast, the adaptively stabilized ES method displays either no
locking or very mild evidence of it with either material, regardless of the mesh refinement.

8.3. Shear locking


This next example examines the behavior of the adaptively stabilized method for the simulation
of thin structures with three-dimensional elements. For example, for the simulation of thin
beams and shells. It is known that under appropriate conditions low-order conforming methods,
such as Q1 elements, are prone to display shear locking for these problems (see, e.g., [29]).
Consider first the problem of the deflection of a two-dimensional thin beam clamped at
one end due to a force on the other end. Its reference configuration is B0 = (0, t) × (0, 1),

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
AN ADAPTIVE STABILIZATION STRATEGY FOR ENHANCED STRAIN METHODS 21

0.8

0.13
0.6 0.11
0.09
0.07
0.05
0.03
0.4 0.01

0.2

0
0 0.2 0.4 0.6 0.8 1

(a) Deformed configuration of a near incompressible


beam, with ξ = 5.0 × 105 and β2 = 2. The colors show
the values of λX .

1
Normalized Tip Displacement

0.1

Q1/E4-neo-Hookean,h=1/8,β2=1
Q1/E4-Ogden,h=1/8,β2=1
P2-Ogden,h=1/32
Q1-neo-Hookean,h=1/8
Q1-neo-Hookean,h=1/32
Q1-Ogden,h=1/8

0.01
1 10 100 1000 10000 100000 1e+06
ξ
(b) Euclidean norm of the displacement of the upper rightmost corner
as a function of ξ, in a normalized scale (see text).

Figure 5. Performance of the ES method with adaptive stabilization and Q1/E4 elements under near
incompressible conditions. The parameter ξ measures the ratio of bulk to shear moduli. As expected,
conforming Q1 discretizations display volumetric locking as ξ grows. The adaptively stabilized ES
method with β2 = 1, however, remains essentially locking free for the range of values of ξ shown for
both materials, as the comparison with the curve obtained with a conforming P2 discretization shows.
The latter is not known to lock under these conditions.

where t is the (non-dimensional) beam thickness. We denote by X1 the thin direction of the
beam, and by X2 the longitudinal one. The surface X2 = 0 is fixed. A force per unit length

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
22 A. TEN EYCK AND A. LEW

(a) Deformed beam made of an Ogden material with ξ = 1.0 × 104


and β2 = 2. The contour levels of λX are also shown.

1
Normalized Tip Displacement

0.1 Q1/E9-neo-Hookean,β2=1,h=1/8
Q1/E9-Ogden,β2=2,h=1/16
Q1-neo-Hookean,h=1/8
Q1-neo-Hookean,h=1/16
Q1-Ogden,h=1/28

0.01
1 10 100 1000 10000 100000 1e+06
ξ

(b) Displacement of one of the tips normalized by the maximum


value observed for each material, as a function of ξ.

Figure 6. Performance of the adaptively stabilized ES method with Q1/E9 elements under near
incompressible conditions in three dimensions. The method does not display significant traces of
locking as either material becomes increasingly more incompressible, i.e. larger ξ. These results were
obtained with β2 = 1 and β2 = 2 for the neo-Hookean and Ogden materials, respectively.

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
AN ADAPTIVE STABILIZATION STRATEGY FOR ENHANCED STRAIN METHODS 23

along the X1 -direction is applied on the surface X2 = 1, with magnitude equal to t2 N/m
and 3 × 105 t2 N/m for the neo-Hookean and Ogden materials, respectively.P The material
constants are chosen as E = 1 and ν = 0.15 for the former, and ρ = p µp for the latter. All
ES simulations in this section were performed with β1 = β3 = 0.
The domain is discretized into identical rectangular elements with sides parallel to X1 and
X2 , and mesh sizes denoted by hX1 and hX2 , respectively. For each of the following simulations
we adopted hX1 /hX2 = 4t, for all values of t. By fixing this ratio, the elements in each mesh
become flatter as t → 0. Under these conditions, shear locking is detected by a sharp decrease in
the value of the tip displacement as t → 0 and hX2 is kept constant. In addition to simulations
with Q1/E4 ES elements and Q1 conforming elements, we have also performed simulations with
a mesh of conforming P2 triangular elements constructed by dividing each one of the rectangles
along the same diagonal, since these elements are not known to display shear locking.
A deformation of the beam when made of the Ogden material is shown in Fig. 7(a), obtained
with Q1/E4 elements and β2 = 1 for 1/t = 10. The contour plot of λX is shown as well,
indicating the local magnitude of the stabilization term.
The chart in Fig. 7(b) shows the displacement of the point (t, 1), normalized by the maximum
value recorded for that material. The leftmost point of each curve is evaluated at 1/t = 4, so
that for each curve hX1 /hX2 = 1 on the vertical axis. For 1/t > 40 the linear conforming
method with Q1 elements clearly displays shear locking for the meshes we considered. For a
given thickness, this effect is ameliorated as finer meshes are considered. The stabilized ES
method with Q1/E4 elements shows similar performance to that of the conforming method
with P2 elements for this range of thicknesses, regardless of the material. Only mild signs of
locking may be detected for 1/t ∼ 100. The test for the Ogden material uses more highly
refined meshes than the test for the neo-Hookean material, but the results are qualitatively
similar.
We consider next a three dimensional version of the problem just studied. The domain is
B0 = (0, t) × (0, 0.5) × (0, 1), and the coordinate axes are labeled X1 , X2 and X3 , with X1
oriented along the thin direction of the beam. The surface X3 = 0 is fixed, while a constant
traction in the X1 direction is applied on the surface X3 = 1. The magnitude of the traction
vector is 0.2 t2 N/m2 for the neo-Hookean material and 4 t2 ×105 N/m2 for the Ogden material.
In this case we chose ν = 0.25 for the neo-Hookean material, while the same coefficients as for
the two-dimensional case are adopted for the Ogden material. For each simulation we chose
a mesh of identical rectangular hexahedra with hX1 /hX3 = 4t, and hX2 = 0.125. Simulations
with both conforming Q1 elements and ES Q1/E9 elements are shown. Shear locking is in this
case identified by a sharp decrease in the tip displacement as t → 0 and hX3 and hX1 are kept
constant.
An example deformation is shown in Fig. 8(a) for the adaptively stabilized ES method with
Q1/E9 elements, for 1/t = 10, β2 = 2 and a beam made of an Ogden material. The colored
areas indicate the contours of λX , and hence the magnitude of the stabilization term.
Fig. 8 shows the norm of the displacement of the point (t, 0.5, 1) normalized by the maximum
value observed for each material, as a function of 1/t. The leftmost point in the chart is
evaluated at 1/t = 4, so for each curve hX1 /hX3 = 1. As in the two-dimensional case, when
1/t > 40 the appearance of shear locking in the conforming Q1 method is evident, for both
materials and this range of mesh sizes. Mild signs of shear locking, although more evident than
in the two-dimensional case, are observed for the ES method with adaptive stabilization and
Q1/E9 elements, regardless of the material. As illustrated by the simulations with conforming

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
24 A. TEN EYCK AND A. LEW

8.5E+04
6.2E+04
0.6 4.5E+04
3.3E+04
2.4E+04
1.8E+04
1.3E+04
9.4E+03
0.4 6.8E+03
5.0E+03

0.2

0
0 0.2 0.4 0.6 0.8

(a) Deformed configuration of a beam made of a


compressible Ogden material when testing for shear
locking. This solution was obtained with Q1/E4
elements, β2 = 1 and t = 1/10. The contour levels of
λX are also shown.

1
Normalized Tip Displacement

0.1 Q1/E4-Ogden,β2=1,hX =1/20


P2-Ogden,hX2=1/80
Q1/E4-neo-Hookean, β2=1,hX2=1/12
P2-neo-Hookean,hX2=1/80
Q1-neo-Hookean,h2X =1/8
2
Q1-neo-Hookean,hX =1/20
Q1-Ogden,hX 2=1/32
2

0.01
10 100
1/t

(b) Norm of the displacement of the right upper corner of the beam
as a function of its thickness t. For each curve, the mesh is always
the same regardless of t; it is stretched however to fit the changing
thickness. Results for different mesh sizes are shown.

Figure 7. Study of shear locking in two dimensions for the stabilized ES method. The tip displacement
in Fig. 7(b) was normalized by the maximum value observed for each material. The appearance of
shear locking is evident for a conforming Q1 method when 1/t > 40 for this range of mesh sizes,
regardless of the material. The method with adaptive stabilization with Q1/E4 elements used β2 = 1
for both materials. Both display a performance similar to that of conforming P2 triangles, not known
to lock under these circumstances.

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
AN ADAPTIVE STABILIZATION STRATEGY FOR ENHANCED STRAIN METHODS 25

(a) Deformed configuration of a three-dimensional


beam made of a compressible Ogden material when
testing for shear locking. The solution was obtained
with Q1/E9 elements, β2 = 2 and t = 1/10. The contour
plots of λX are also shown.

1
Normalized Tip Displacement

0.1

0.01 Q1-neo-Hookean,hX =1/8


3
Q1-neo-Hookean,hX =1/16
Q1-Ogden,hX3=1/16
3
Q1/E9-neo-Hookean,β2=1,hX =1/8
3
Q1/E9-Ogden,β2=2,hX =1/28
3

0.001
1 10 100 1000
1/t

(b) Norm of the displacement of one of the corners of the beam


as a function of its thickness t. For each curve, the mesh is always
the same regardless of t; it is stretched however to fit the changing
thickness. Results for different mesh sizes are shown.

Figure 8. Three-dimensional study of shear locking for the stabilized ES method with Q1/E9 elements.
Evidence of shear locking in the conforming Q1 method is seen for 1/t > 40 for the meshes studied.
Mild signs of locking are observed for the ES method with Q1/E9 elements, and β2 = 1 and β2 = 2
for the neo-Hookean and the Ogden material, respectively. Finer meshes would ameliorate this effect.

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
26 A. TEN EYCK AND A. LEW

elements, this effect can be ameliorated with the use of finer meshes. Values of β2 = 2 and
β2 = 1 were adopted for the Ogden and neo-Hookean materials, respectively.

9. Summary

We have introduced a stabilization strategy for ES methods in nonlinear elasticity problems,


and some of its defining characteristics are that the stabilization term depends on the solution
sought and that the method is independent of the type of enhancement or material model. It is
simple to see that the proposed scheme is able to stabilize any linearized elasticity problem that
is itself stable, provided that the stabilization parameters are chosen large enough. Explicit
lower bounds for the stabilization parameters that guarantee the stability of the method were
provided. However, a brief discussion and references to related numerical examples with DG
methods [15] indicate that these lower bounds are generally too large, and may result in
methods that display a locking behavior similar to standard conforming methods. Based on
these considerations, a heuristic set of values for the stabilization parameters was proposed
and used for the numerical examples later.
The performance of the method was demonstrated with a number of numerical examples.
The first examples tested how large the stabilization parameters need to be to stabilize the
problem. It was found that our heuristic choice of stabilization parameter did stabilize the
method for these examples, in two and three dimensions and for two different nonlinear elastic
materials. The same choice of parameters was then used to examine the behavior of the method
for the appearance of volumetric and shear locking. The stabilization strategy provided stable
ES solutions under these conditions that were significantly better for the same mesh size than
those stemming from conforming approximations. Only mild signs of locking were detectable
in some circumstances. The locking traces in both conforming and ES approximations can be
ameliorated by adopting finer meshes. However, significantly finer meshes are needed for the
former to obtain the same accuracy, making it overly impractical.
The paper strongly draws on previous work of the authors on DG methods to build the
stabilization scheme. We therefore found it worthwhile to describe these similarities through
short separate paragraphs. These serve to present both methods under a similar perspective
as well as to motivate the ideas herein.
A major drawback of this as well as other stabilization strategies is that it is difficult, if not
impossible, to ascertain what the value of the onset of buckling of a structure is (at what load
the linearized elasticity problem ceases to be stable). Because the method uses incompatible
strain fields, one way to test this is, roughly, by raising the value of the stabilization parameters
to get closer and closer to the underlying conforming or compatible discretization. It is then
possible to determine the buckling load if the latter does converge to it as finer meshes are
considered. However, if locking is a problem for the underlying conforming method, there does
not seem to be a way by which the buckling load could be determined with these schemesk .
This is apparently an open question, and momentarily it seems that higher order methods are

k The lower bounds in Thm. 6.1 use the coercivity constant κ of the conforming method, since it is the only
h
one we can easily compute, and hence would miss the exact buckling load as well if the conforming scheme
does.

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
AN ADAPTIVE STABILIZATION STRATEGY FOR ENHANCED STRAIN METHODS 27

better suited instead when such load is of interest. See [30] for a related discussion.
Finally, it may be possible to combine the basic stabilization strategy introduced here with
reduced integration methods to reduce the computational cost.

APPENDIX

II. Analysis of the adaptive method

We provide here the proof of Thm. 6.1. The idea behind the proof is identical to the one
for the adaptive stabilization strategy for DG methods in [15]. It is only included here for
completeness.
Proof:(Theorem 6.1) We begin by constructing a lower bound for

Bh (vh , wh , vh , wh ) =
Z Z Z
∇vh : A : ∇vh dV + 2∇vh : A : wh dV + wh : A : wh dV. (37)
B0 B0 B0

Applying a generalized version of Young’s inequality (see [15]) to the integrand in the second
term on the right-hand side we get
2∇vh : A : wh ≥ −∇vh : [ηA + λX (η + ξX )I] : ∇vh
(38)
   
1 1 1
− wh : A + λX + I : wh
η η ξX
which is valid for any η, ξX > 0. We will choose η > 0 independent of X ∈ B0 , but let ξX
change from point to point. In (37) it yields

Bh (vh , wh , vh , wh ) ≥ T1 + T2 ,

where Z
T1 = ∇vh : [(1 − η)A − (λX (η + ξX ))I] : ∇vh dV,
B0
    
1 1 1
Z
T2 = wh : 1− A − λX + I : wh dV.
B0 η η ξX
Given that η ≤ 1 from (27), it follows from the stability hypothesis (26) that
Z
T1 ≥ [κh (1 − η) − λX (η + ξX )] ∇vh : ∇vh dV. (39)
B0

A lower bound for Bstab


h then ensues as
Z
Bstab
h (vh , wh , vh , wh ) ≥ Bm
h (vh , vh ) + BM
h (wh , w h ) + β3 κ∗ wh : wh dV. (40)
B0

where Z
Bm
h (vh , vh ) := m(η, ξX )∇vh : ∇vh dV,
B0

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
28 A. TEN EYCK AND A. LEW

and Z
BM
h (wh , wh ) := wh : M(η, ξX ) : wh dV. (41)
B0
Here
m(η, ξX ) = κh (1 − η) − λX (η + ξX )
and    
1 1 1
M(η, ξX ) = β1 − + 1 A + λX β2X − − I.
η η ξX
We now investigate the positivity of m(η, ξX ) and M(η, ξX ). Since λX ≤ Λ, we have that
κh − κh κh − κh

κh + Λ κh + λX
and hence
κh − κh
κh − κh − (κh + λX ) ≥ 0.
κh + Λ
It then follows that
κh − κh − η(κh + λX ) ≥ 0 (42)
with equality only when η = 1. When η < 1 the left hand side of (42) is strictly positive, and
hence it is possible to find ξX > 0 such that
κh − κh − η(κh + λX ) ≥ λX ξX (43)
for almost every X ∈ B0 . Similarly, if η = 1 then necessarily 0 ≤ λX ≤ Λ = 0, and (43) holds
for any ξX > 0. In either case, we can always choose ξX > 0 such that (43) is satisfied.
Assuming that ξX obeys these bounds we see that
m(η, ξX ) = κh (1 − η) − λX (η + ξX ) ≥ κh (44)
for almost every X in B0 . Thus
ess inf m(η, ξX ) ≥ κh . (45)
X∈B0

We will next see that M(η, ξX ) is a positive semi-definite tensor. We know, by the definition
of λX , that A + λX I is a positive semi-definite tensor. As a consequence, α1 A + α2 λX I is
positive semi-definite for any α2 ≥ α1 ≥ 0. Based on hypothesis (29) on β2X we conclude that
1 1 λX 1 1
β2X − − ≥ β1 + 1 + − − (46)
η ξX κh − κh − η(κh + λX ) η ξX
     
1 1 1 1 λX ξX
β2X − − − β1 + 1 − ≥ −1 (47)
η ξX η ξX κh − κh − η(κh + λX )
We next consider the possible values of ξX that satisfy (43). If λX 6= 0, then it is enough to
choose ξX so that the equality holds in (43) to conclude that
1 1 1
β2X − − ≥ β1 + 1 − ≥ 0, (48)
η ξX η
where we have used hypothesis (28) on β1 . If λX = 0, then inequality (47) holds for all ξX > 0,
from where we conclude that (48) is satisfied as well. Consequently, (48) implies that M is a
positive semi-definite tensor.

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
AN ADAPTIVE STABILIZATION STRATEGY FOR ENHANCED STRAIN METHODS 29

From (40), (45), (48) and (28), we conclude that


 
2
Bstab 2
h (vh , wh , vh , w h ) ≥ κh |vh |1,B0 + kwh k0,B0

for all (vh , wh ) ∈ Vhd × Wenh


d
, which completes the proof.

REFERENCES

1. Glaser S, Armero F. On the formulation of enhanced strain finite elements in finite deformations.
Engineering Computations 1997; 14(7):759–791.
2. Ten Eyck A, Lew A. Discontinuous Galerkin methods for nonlinear elasticity. Internat. J. Numer. Methods
Engrg. 2006; 67:1204–1243.
3. Areias P, César de Sá J, Conceiçao António C, Fernandes A. Analysis of 3D problems using a new enhanced
strain hexahedral element. Int. J. Numer. Meth. Eng 2003; 58:1637–1682.
4. Reese S, Küssner M, Reddy B. A new stabilization technique for finite elements in non-linear elasticity.
Internat. J. Numer. Methods Engrg. 1999; 44:1617–1652.
5. Küssner M, Reddy B. The equivalent parallelogram and parallelepiped, and their application to stabilized
finite elements in two and three dimensions. Comput. Methods Appl. Mech. Engrg. 2001; 190(15-17):1967–
1983.
6. Puso M. A highly efficient enhanced assumed strain physically stabilized hexahedral element. Int. J.
Numer. Meth. Engng 2000; 49:1029–1064.
7. Reese S. On a physically stabilized one point finite element formulation for three-dimensional finite elasto-
plasticity. Comput. Methods Appl. Mech. Engrg. 2005; 194(45-47):4685–4715.
8. Reese S, Wriggers P. A stabilization technique to avoid hourglassing in finite elasticity. Internat. J. Numer.
Methods Engrg. 2000; 48:79–109.
9. Wall W, Bischoff M, Ramm E. A deformation dependent stabilization technique, exemplified by EAS
elements at large strains. Comput. Methods Appl. Mech. Engrg. 2000; 188(4):859–871.
10. Lovadina C, Auricchio F. On the enhanced strain technique for elasticity problems. Computers and
Structures 2003; 81(8-11):777–787.
11. Flanagan D, Belytschko T. A uniform strain hexahedron and quadrilateral with orthogonal hourglass
control. Internat. J. Numer. Methods Engrg. 1981; 17:679–706.
12. César de Sá J, Areias P, Jorge R. Quadrilateral elements for the solution of elasto-plastic finite strain
problems. Int. J. Numer. Meth. Engng 2001; 51:883–917.
13. Armero F. On the locking and stability of finite elements in finite deformation plane strain problems.
Computers and Structures 2000; 75(3):261–290.
14. Ten Eyck A, Celiker F, Lew A. Adaptive stabilization of discontinuous galerkin methods for nonlinear
elasticity: Motivation, formulation, and numerical examples. Comput. Methods Appl. Mech. Engrg. 2008;
197(45-48):3605–3622.
15. Ten Eyck A, Celiker F, Lew A. Adaptive stabilization of discontinuous galerkin methods for nonlinear
elasticity: Analytical estimates. Comput. Methods Appl. Mech. Engrg. 2008; 197(33-40):2989–3000.
16. Marsden JE, Hughes T. Mathematical Foundations of Elasticity. Dover: Mineola, New York, 1994.
17. Simo J, Armero F. Geometrically non–linear enhanced strain mixed methods and the method of
incompatible modes. Comput. Methods Appl. Mech. Engrg. 1992; 33:1413–1449.
18. Reddy B, Simo J. Stability and convergence of a class of enhanced strain methods. SIAM J. Numer. Anal
1995; 32(6):1705–1728.
19. Bassi F, Rebay S. High-order accurate discontinuous finite element method for the numerical solution of
the compressible navier-stokes equations. J. Comput. Phys. 1997; 131:267–279.
20. Noels L, Radovitzky R. A general discontinuous Galerkin method for finite hyperelasticity. Formulation
and numerical applications. Internat. J. Numer. Methods Engrg. 2006; .
21. Valent T. Boundary Value Problems of Finite Elasticity: Local Theorems on Existence, Uniqueness, and
Analytic Dependence on Data. Springer-Verlag, 1988.
22. Lew A, Neff P, Sulsky D, Ortiz M. Optimal BV estimates for a discontinuous Galerkin method in linear
elasticity. Applied Mathematics Research Express 2004; 3:73–106.
23. Brezzi F, Cockburn B, Marini L, Suli E. Stabilization mechanisms in discontinuous Galerkin finite element
methods. Comput. Methods Appl. Mech. Engrg. 2005; .
24. Brezzi F, Hughes T, Marini L, Masud A. Mixed discontinuous Galerkin methods for Darcy flow. Journal
of Scientific Computing 2005; 22(1):119–145.

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls
30 A. TEN EYCK AND A. LEW

25. Ciarlet P. Mathematical Elasticity-Volume I: Three-Dimensional Elasticity. Studies in Mathematics and


its Applications 1988; 20.
26. Buffa A, Ortner C. Variational convergece of IP–DGFEM. Technical Report 07/10, Oxford University
Computing Laboratory, Numerical Analysis Group, Wolfson Building, Parks Road, Oxford, England OX1
3QD April 2007.
27. Holzapfel G. Nonlinear solid mechanics. Wiley, 2000.
28. Braess D, Carstensen C, Reddy B. Uniform convergence and a posteriori error estimators for the enhanced
strain finite element method. Numerische Mathematik 2004; 96(3):461–479.
29. Chapelle D, Bathe K. The Finite Element Analysis of Shells. Springer, 2003.
30. Auricchio F, Beirão da Veiga L, Lovadina C, Reali A. A stability study of some mixed finite elements for
large deformation elasticity problems. Comput. Methods Appl. Mech. Engrg. 2005; 194(9-11):1075–1092.

ACKNOWLEDGEMENTS

We gratefully acknowledge the support of the National Institutes of Health through the NIH Roadmap
for Medical Research, Grant U54GM072970, the Department of the Army Research Grant W911NF-
07-2-0027, an NSF-Career Award, NSF Grant No. CMMI-0747089, and an ONR Young Investigator
Award, ONR Grant No. N000140810852.

Copyright c 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:1–6
Prepared using nmeauth.cls

You might also like