Computer Methods in Applied Mechanics and Engineering 199 (2010) 2559–2570

Contents lists available at ScienceDirect

Computer Methods in Applied Mechanics and Engineering

j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / c m a

Mixed stabilized finite element methods in nonlinear solid mechanics

Part I: Formulation
M. Cervera ⁎, M. Chiumenti, R. Codina
International Center for Numerical Methods in Engineering (CIMNE), Technical University of Catalonia (UPC), Edificio C1, Campus Norte, Jordi Girona 1-3, 08034 Barcelona, Spain

a r t i c l e i n f o a b s t r a c t

Article history: This paper exploits the concept of stabilized finite element methods to formulate stable mixed stress/
Received 22 July 2009 displacement and strain/displacement finite elements for the solution of nonlinear solid mechanics problems.
Received in revised form 9 March 2010 The different assumptions and approximations used to derive the methods are exposed. The proposed
Accepted 13 April 2010
procedure is very general, applicable to 2D and 3D problems. Implementation and computational aspects are
Available online 18 April 2010
also discussed, showing that a robust application of the proposed formulation is feasible. Numerical examples
Keywords:
show that the results obtained compare favorably with those obtained with the corresponding irreducible
Mixed finite element interpolations formulation.
Stabilization methods © 2010 Elsevier B.V. All rights reserved.
Algebraic sub-grid scales
Orthogonal sub-grid scales
Nonlinear solid mechanics

1. Introduction Another common application of mixed methods is plate bending

and other fourth order problems [5,15–17]. Here, the motivation is the
The term mixed methods has been used in the finite element method avoidance of C1-continuity in the definition of the interpolation
literature since the mid 1960s to denote formulations in which both the functions, required if the primal variational functional is used.
displacement and stress fields are approximated as primary variables Alternatively, the mixed functional only involves second derivatives
[1]. Despite the doubtless interest of mixed methods from the and, after integration by parts, C0-continuous elements may be used.
theoretical point of view, their practical application is greatly out- Another alternative is the use of non-conforming elements.
numbered by the implementation of irreducible methods, in which only The reasons for the limited popularity of mixed methods in
the displacement field is considered primary variable of the problem computational solid mechanics are twofold: computational cost and
and the stress field is obtained a posteriori by differentiation. lack of stability [18–20]. On one hand, because mixed methods
However, there are several fields of application in computational approximate both displacements and stresses simultaneously, the
solid mechanics in which mixed methods are well established and corresponding discrete systems of equations involve many more
regularly used in practice. For instance, it is well known that standard degrees of freedom than the corresponding irreducible formulations.
irreducible low order finite elements perform miserably in nearly Concurrently, the mixed system of equations is very often indefinite,
incompressible situations, producing solutions which are almost which makes most of the direct and iterative solution methods
completely locked by the incompressibility constraint. Remedies for inapplicable. These difficulties may be avoided with a suitable
this undesirable behavior have been actively sought for decades. In implementation. On the other hand, many choices of the individual
fact, the purely incompressible problem (Stokes problem) does not interpolation fields for the mixed problem yield meaningless, not stable,
admit an irreducible formulation and, consequently, a mixed frame- results. This is due to the strictness of the inf-sup condition [19] when
work in terms of displacements and pressure is necessary for these the standard Galerkin finite element method is applied straightfor-
situations. Over the years, and particularly in the 1990s, different wardly to mixed elements, as it imposes severe restrictions on the
strategies were proposed and tested to reduce or avoid volumetric compatibility of the interpolations used for the displacement and the
locking and pressure oscillations in finite element solutions with stress fields. This difficulty, if not circumvented, is severely restrictive
different degrees of success [2–14]. Many of these methods, while (see [21–23] for the analysis of admissible elements in linear elasticity).
resembling displacement methods, have been shown to be equivalent In parallel, mixed methods have also been the focus of attention in
to more general mixed methods. computational fluid dynamics. In [24,25], the variational multiscale
(VMS) formulation was proposed as a new way of circumventing the
difficulties posed by the inf-sup condition. In the case of incompressible
⁎ Corresponding author. problems, the reasoning behind was not new, as it consisted of modifying
E-mail address: [email protected] (M. Cervera). the discrete variational form to attain control on the pressure field. The

0045-7825/$ – see front matter © 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.cma.2010.04.006
2560 M. Cervera et al. / Computer Methods in Applied Mechanics and Engineering 199 (2010) 2559–2570

result was the possibility of using equal order interpolations for 2. Mixed stabilized stress–displacement formulation in
displacements and pressures and to construct stable low order elements. linear elasticity
Since then, the sub-grid concept underlying the VMS approach has been
extensively and fruitfully used in fluid dynamics. In [26,27], the concept 2.1. Continuous problem
of orthogonal subscale stabilization (OSS) was introduced, which leads
to well sustained and better performing stabilization procedures. The The formulation of the solid mechanics problem can be written
analysis of the formulation can be found in [28] for the linearized considering the stress as an independent unknown, additional to the
incompressible Navier–Stokes equations and, in subjects closer to the displacement field. In this case, the strong form of the continuous
topic of this paper, in [29] for the stress–displacement-pressure problem can be stated as: given a field of prescribed body forces f and
formulation of the Stokes problem (equivalent to the linear elastic a constant constitutive tensor C, find the displacement field u and the
incompressible problem) and in [30] for Darcy's problem. stress field σ such that:
In previous works, the authors have applied stabilized mixed
s
displacement/pressure methods (see [31–36]) to the solution of −σ + C : ∇ u = 0 in Ω ð1aÞ
incompressible J2-plasticity and damage problems with strain
localization using linear/linear simplicial elements in 2D and 3D. ∇⋅σ + f = 0 in Ω ð1bÞ
These procedures lead to a discrete problem which is fully stable, free
of pressure oscillations and volumetric locking and, thus, results where Ω is the open and bounded domain of Rndim occupied by the
obtained are practically mesh independent. This translates in the solid in a space of ndim dimensions.
achievement of two important goals: (a) the position and orientation Eqs. (1a)–(1b) are subjected to appropriate Dirichlet and Neu-
of the localization band is independent of the directional bias of the mann boundary conditions. In the following, we will assume these,
finite element mesh and (b) the global post-peak load-deflection without loss of generality, in the form of prescribed displacements
–
curves are independent of the size of the elements in the localization u = 0 on ∂Ωu, and prescribed tractions t on ∂Ωt, respectively, being
band. Similar ideas have been used in [37–39]. ∂Ωu and ∂Ωt a partition of ∂Ω.
In the present work we apply this approach in order to derive Multiplying by the test functions and integrating by parts the
stable mixed stress–displacement and strain–displacement formula- second equation, the associated weak form of the problem (Eq. (1a)–
tions using linear/linear interpolations in triangular elements and (1b)) can be stated as:
bilinear/bilinear interpolations in quadrilateral elements. It is note-

−1  s 
worthy that, from the numerical point of view, the difficulties − τ; C : σ + τ; ∇ u = 0 ∀τ ð2aÞ
encountered in this problem are very different to those found in
incompressible situations, analyzed in previous works. The treatment  

s 
of the incompressible case in the stress/displacement formulation ∇ v; σ = ðv; f Þ + v; t ∀v ð2bÞ
∂Ωt
would require considering the pressure as an additional independent
variable and appropriate stabilization techniques (see [29]). The where v 2 V and τ 2 T are the test functions of the displacement and
incompressible limit will not be treated here, and the following stress fields, respectively, and (⋅, ⋅) denotes the inner product in L2(Ω),
formulation is limited to compressible nonlinear solid mechanics. the space of square integrable functions in Ω. Hereafter, orthogonality
The basic motivation for this work is to show that the difficulties –
will be understood with respect to this product. Likewise, (v, t )∂Ωt
encountered when solving solid mechanics problems involving the –
denotes the integral of v and t over ∂Ωt. For the sake of shortness, we
creation and propagation of strain localization bands using standard –
will write F(v) = (v, f) + (v,t )∂Ωt in the following. Eqs. (2a) and (2b) can
elements and local constitutive models are due to the approximation be understood as the stationary conditions of the classical Hellinger–
error inherent to the spatial discretization, as well as to the poor stability in Reissner functional (see [41] for a description of a broader class of mixed
the stresses and/or strains. When using the basic, irreducible, formulation methods).
of the problem, the stresses (or strains), which are the variables of most The space of stresses T consists of symmetric tensors whose
interest for the satisfaction of the highly nonlinear constitutive behavior, components are in L2(Ω). If the weak form is written as indicated in
are not the fundamental unknowns of the problem and they are Eq. (2a), the displacements and their test functions have to have
obtained by differentiation of the displacement field, a process which components in H1(Ω) (they and their derivatives have to be in L2(Ω))
entails an important loss of accuracy, particularly where strong and must vanish on ∂Ωu. This defines the space of displacements V.
displacement gradients occur. The local approximation error committed However, it is also possible to integrate the second term in Eq. (2a), by
makes propagation of the localization bands strongly dependent on the parts, obtaining (τ, ∇su) = −(∇ · τ, u), and similarly for the left-hand-
finite element mesh used. Contrariwise, when using a mixed formula- side of Eq. (2b). In this case, the components of the stresses have to
tion in which the stress (or the strain) field is selected as primary have also the divergence in L2(Ω), but the components of the
variable, together with the displacement field, the added accuracy and displacement need to be only in L2(Ω), not H1(Ω). Similar to Darcy's
stability achieved are enough to overcome the mesh dependency problem, there are two possible functional settings for the linear
problem satisfactorily. elastic problem written in mixed form (see [30]). This is not essential
The outline of the present paper is as follows. In Section 2 the mixed for our discussion, although it has some implications in the treatment
stress/displacement finite element formulation for linear elasticity is of boundary conditions on which we will not enter.
summarized. The sub-grid scale approach is used to derive two
stabilized formulations. Results concerning stability and convergence 2.2. Galerkin finite element approximation
of these schemes are discussed. In Section 3 the stabilization is
extended to nonlinear problems, proposing both stress–displacement Let us now define the discrete Galerkin finite element counterpart
and strain–displacement formulations. The later can be considered problem as:
more suitable for the implementation of nonlinear constitutive
models. Implementation and computational aspects are discussed
  
−1 s
next. Finally, some numerical benchmarks and examples are presented − τh ; C : σh + τh ; ∇ uh = 0 ∀τh ð3aÞ
to assess the present formulation and to compare its performance with
the standard irreducible elements. The problem of strain localization is  s 
discussed in a companion paper [40]. ∇ vh ; σh = Fðvh Þ ∀vh ð3bÞ
 M. Cervera et al. / Computer Methods in Applied Mechanics and Engineering 199 (2010) 2559–2570 2561

 s   s
where uh, vh 2 V h and σh, τh 2 T h are the discrete displacement and ∇ vh ; σh + ∇ vh ; σ̃Þ = Fðvh Þ ∀vh ð5cÞ
stress fields and their test functions, defined onto the finite element
spaces V h and T h, respectively. Note that the resulting system of  
− ṽ; ∇ ⋅ σh Þ− ṽ; ∇ ⋅ σ̃Þ = Fð ṽÞ ∀ ṽ ð5dÞ
equations is symmetric but non-definite. In all what follows, we will be
interested in continuous finite element spaces V h and T h and, more
specifically, in equal interpolation for stresses and displacements. where some terms have been integrated by parts and we have
Therefore, we may replace (τh, ∇svh) by −(∇ · τh, vh), for all vh 2 V h assumed that ũ and ṽ vanish on the boundary. In the following, the
and τh 2 T h. fact that the discrete variational equations need to hold for all test
As it is well known, the stability of the discrete formulation functions will be omitted.
depends on appropriate compatibility restrictions on the choice of the Due to the approximation used, Eq. (4), and the linear indepen-
finite element spaces V h and T h, as stated by the inf-sup condition [19]. dence of τh and τ̃, now the continuous Eq. (2a) unfolds in two discrete
According to this, standard Galerkin mixed elements with continuous Eqs. (5a) and (5b), one related to each scale considered. The same
equal order linear/linear interpolation for both fields are not stable. comment is applicable to the displacement splitting. Eqs. (5a) and
Lack of stability shows as uncontrollable oscillations in the displace- (5c) are defined in the finite element spaces T h and V h, respectively.
ment field that entirely pollute the solution. Fortunately, the strictness The first one enforces the constitutive equation including a stabiliza-
of the inf-sup condition can be avoided by modifying the discrete tion term S1 = −(τh, C−1:σ̃) − (∇∙τh, ũ) depending on the sub-grid
variational form, for instance, by means of introducing appropriate stresses and displacements. The second one solves the balance of
numerical techniques that can provide the necessary stability to the momentum including a stabilization term S2 = (∇s vh, σ̃) depending
desired choice of interpolation spaces. The objective of this work is on the sub-grid stresses σ̃.
precisely to present stabilization methods which allow the use of Let us define the residuals of the finite element components as
equal order continuous interpolations for displacements and stresses.
−1 s
rσ;h = C : σh −∇ uh ð6aÞ
2.3. Stabilized finite element methods
ru;h = f + ∇⋅σh ð6bÞ
2.3.1. Scale splitting
The basic idea of the sub-grid scale approach [24] is to consider These allow us to write Eqs. (5b) and (5d) as
that the continuous unknowns can be split in two components, one


−1  s 
coarse and a finer one, corresponding to different scales or levels of − τ̃; C : σ̃ + τ̃; ∇ ũ = τ̃; rσ;h ð7aÞ
resolution. The solution of the continuous problem contains compo-
nents from both scales.  

− ṽ; ∇⋅ σ̃ = ṽ; ru;h ð7bÞ
For the solution of the discrete problem to be stable it is necessary
to, somehow, include the effect of both scales in the approximation.
The coarse scale can be appropriately solved by a standard finite These equations are the projections of the finite element residuals
element interpolation, which however cannot solve the finer scale. onto the space of subscales, which cannot be resolved by the finite
Nevertheless, the effect of this finer scale can be included, at least element mesh. Therefore, to proceed it is necessary to provide an
locally, to enhance the stability of the displacement in the mixed approximate closed form solution to them. If P σ̃ and P ̃u are the
formulation. projections onto T ̃ and V ̃, respectively, note first that we may write
To this end, the stress and the displacement fields of the mixed Eqs. (7a) and (7b) as
problem will be approximated as


−1 s
P̃ σ −C : σ̃ + ∇ ũ = P̃ σ rσ;h ð8aÞ
σ =σh + σ̃; u=uh + ũ ð4Þ
 

where σh 2 T h and uh 2 V h are the components of the stresses and the P̃ u −∇⋅ σ̃ = P̃ u ru;h ð8bÞ
displacements on the (coarse) finite element scale and σ̃2 T ̃ and ũ 2 V ̃
are the enhancement of the stresses and the displacements and therefore the problem is to approximate the operators on the left-
corresponding to the (finer) sub-grid scales. Let us also consider the hand-side of these equations. The way we motivate such an
corresponding test functions τ̃ 2 T ̃ and ṽ 2 V ̃. This approximation approximation is by using an approximate Fourier analysis of the
extends the stress solution space to T ≃ T h ⊕ T ̃, and the displacement problem. Using exactly the same procedure as in [43], it can be shown
solution space to V ≃ V h ⊕ V ̃. Each particular stabilized finite element that σ̃ and ũ may be approximated within each element by
method is defined according to the way in which spaces T ̃ and V ̃ are
  
s
chosen. In particular, the Galerkin method corresponds to taking T ̃ = σ̃ = −τσ C : P̃ σ rσ;h = τσ P̃ σ C : ∇ uh −σh ð9aÞ
{0}, V ̃ = {0}.


As it has been mentioned, in what follows we will consider
ũ = τu P̃ u ru;h = τu P̃ u ðf + ∇⋅σh Þ ð9bÞ
continuous finite element interpolations. Likewise, we will assume
that the subscales vanish on the interelement boundaries. When more
where the so called stabilization parameters τσ and τu can be computed
general situations are considered, additional terms involving inter-
as
element boundary integrals need to be added (see [30,42]).
Introducing the splitting, the problem corresponding to Eqs. (2a)– h L h
(2b) is: τσ = cσ ; τ u = cu 0 ð10Þ
L0 Cmin


   
−1 −1 s
− τh ; C : σh − τh ; C : σ̃ + τh ; ∇ uh − ∇ ⋅ τh ; ũÞ = 0 ∀τh and where cσ and cu are algorithmic constants, L0 is a characteristic
ð5aÞ length of the computational domain, h is the element size and Cmin N 0
is the smallest eigenvalue of C (see below). As shown in Ref. [30], this


   
−1 −1 s s  is the choice of the parameters that yields best order of convergence
− τ̃; C : σh − τ̃; C : σ̃ + τ̃; ∇ uh + τ̃; ∇ ũ = 0 ∀ τ̃
for equal order of interpolation of stresses and displacements. In the
ð5bÞ following, and for the sake of clarity, we will consider the mesh quasi-
2562 M. Cervera et al. / Computer Methods in Applied Mechanics and Engineering 199 (2010) 2559–2570

uniform, so that a unique h can be defined for all the mesh, and thus τσ 2.3.3. Orthogonal subscale stabilization
and τu will be constant. In general situations, it is understood that It was argued in [27] that a very natural choice for the unknown
these parameters have to be evaluated element-wise. sub-grid spaces is to take them orthogonal to the finite element space.
The methods we wish to consider are completely defined up to the This amounts to saying that the projections P σ̃ and P ũ are taken as P⊥h
choice of the projections P σ̃ and P ũ . Two possible options are applied to the appropriate space of discrete functions. This also means
described next. approximating the stress solution space as T ≃ T h ⊕ T ⊥ h and, similarly,
the displacement solution space as V ≃ V h ⊕ V ⊥ h . The subsequent
2.3.2. Residual-based algebraic sub-grid scale method stabilization method is called orthogonal subscale stabilization (OSS)
The simplest choice is to take P σ̃ and P ũ as the identity when method, and it has already been successfully applied to several
applied to the residuals in Eqs. (9a) and (9b). In fact, one may also problems in fluid and solid mechanics.
think that the projection is scaled by the stabilization parameters Noting that σh is a finite element function and computing Ph⊥ = I − Ph
given by Eq. (10), which act as upscaling of the residuals onto the (I being the identity), the subscales can be now expressed as
finite element mesh. This is what is called algebraic sub-grid scale
⊥ s   s  s 
(ASGS) method in [30], for example. If the subscales resulting from σ̃ = τσ Ph C : ∇ uh = τσ C : ∇ uh −Ph C : ∇ uh ð14aÞ
these equations are then inserted into Eqs. (5a) and (5c) one gets
⊥

 ũ = τu Ph ðf + ∇ ⋅ σh Þ = τu ½f + ∇ ⋅ σh −Ph ðf + ∇ ⋅ σh Þ ð14bÞ
−1  s 
−ð1−τσ Þ τh ; C : σh + ð1−τσ Þ τh ; ∇ uh −τu ð∇ ⋅ τh ; ∇ ⋅ σh Þ = τu ð∇ ⋅ τh ; fÞ

ð11aÞ Introducing these orthogonal subscales in Eqs. (5a) and (5c) the
first component in the stabilization term S1 vanishes because of
 s   s s  orthogonality and the mixed system of equations can be written as
ð1−τσ Þ ∇ vh ; σh + τσ ∇ vh ; C : ∇ uh = Fðvh Þ ð11bÞ

  


−1 s ⊥ ⊥
− τh ; C : σh + τh ; ∇ uh −τu ∇ ⋅ τh ; Ph ð∇⋅σh Þ = τu ∇ ⋅ τh ; Ph ðf Þ
Note that the resulting system of equations is symmetric.
Particularly interesting is the case τu = 0. In this situation, Eq. ð15aÞ
(11a) represents a projection onto the discrete finite element space

 s  s ⊥ s 
that can be written as ∇ vh ; σh + τσ ∇ vh ; Ph C : ∇ uh = Fðvh Þ ð15bÞ
 s 
σh = Ph C : ∇ uh ð12Þ It is also interesting to consider the case τu = 0. Now Eq. (15a) is
identical to Eq. (11a) in the previous section and, therefore, it can be
and, therefore, the discrete balance Eq. (11b) takes the form: written as Eq. (12) once again. With this definition, the orthogonal
 s  s   s s  subscale in Eq. (14a) is identical to the residual-based subscale in
ð1−τσ Þ ∇ vh ; Ph C : ∇ uh + τσ ∇ vh ; C : ∇ uh = ðvh ; f Þ Eq. (9a) with P σ̃ = I. Therefore, the resulting stabilization terms are
also identical and the system of Eqs. (15a)–(15b) can be arranged as
Thus, for τu = 0 the method we propose can be rewritten as in system Eqs. (11a)–(11b) or system (13a) and (13b). Therefore,
    when τu = 0 the ASGS and the OSS formulations coincide.
s s
σstab = ð1−τσ ÞPh C : ∇ uh + τσ C : ∇ uh ð13aÞ
2.4. Stability and convergence results
 s 
∇ vh ; σstab = Fðvh Þ ð13bÞ
In this section we state stability and convergence results both for
This compact form of writing the problem is only possible when the OSS method given by Eqs. (15a)–(15b) and for the ASGS method
τu = 0. Otherwise, Eqs. (11a)–(11b) have to be kept as such. given by Eqs. (11a)–(11b), which, as we have seen, coincide when
Some remarks are in order: τu = 0. The proof of these results can be done adapting the analysis
presented in [30]. To simplify the exposition, we will consider the
1. The stabilization term S2 in Eq. (5a) is computed in an element by –
boundary tractions t = 0.
element manner and within each element. Its magnitude depends The constitutive tensor C is assumed to be constant, symmetric and
on the difference between the continuous (projected) stresses σh positive definite. Let Cmax N 0 and Cmin N 0 be such that
and the discontinuous (elemental) stresses C: ∇suh.
2. This means that the term added to secure a stable solution Cmin γ : γ ≤ γ : C : γ ≤ Cmax γ : γ ð16Þ
decreases upon mesh refinement, as the finite element scale
becomes finer and the residual reduces. for all symmetric second order tensors γ.
3. In other words, σ̃ is “small” compared to σh. Let ∥∙∥ denote the standard norm in L2(Ω). For the continuous
4. With this definition, σ̃ is discontinuous across element boundaries. problem Eqs. (2a)–(2b) it can be shown that
For linear elements, σ̃ is piece-wise linear.
5. Even if defined element-wise, σ̃ cannot be condensed at the 1 Lh0 Cmin L20
element level, because σh is interelement continuous. Cmax
σ ∥ ∥2
+ ∥
Cmax
∇⋅σ ∥
2
+ ∥∥
L20
u
2
∥ s
+ Cmin ∇ u ≲ ∥
2
Cmin∥∥
f
2
ð17Þ

6. In the localization process in Eq. (9a), it is necessary to neglect the

integrals over element faces involving the subscale, in front of the This result gives optimal stability in all the fields involved in the
integrals over the element volumes. This is justified in [44] problem. The symbol ≲ is used to include constants independent of
resorting to Fourier analysis and recalling that the subscale is the unknowns and the components of C (and of h, in what follows).
associated to frequencies higher than the grid scale. It is worth to For the Galerkin finite element approximation to the problem, a
mention that for “bubble”-type enhancements these boundary bound similar to Eq. (17) can be proved provided the appropriate inf-
terms are null by construction [45,46]. See also [42] for a possible sup conditions between the interpolating spaces are met. Moreover, in
generalization. general it is not possible to bound both ||∇ · σh||2 and ||∇s uh||2, but
7. Eq. (9a) must not be interpreted point-wise, as the values of σ̃ are only one of these two terms.
not used in the stabilization procedure; only the integrals S1 and S2 Stabilized finite element methods aim precisely at providing
in Eqs. (5a)–(5c) are needed. stability estimates without relying on compatibility conditions. In
 M. Cervera et al. / Computer Methods in Applied Mechanics and Engineering 199 (2010) 2559–2570 2563

particular, for the methods given by Eqs. (11a)–(11b) and by Eqs. constitutive equation is written in rate form. To extend the present
(15a)–(15b) it can be shown that formulation to such models will be the subject of future research.
The misbehavior encountered when irreducible formulations are
2
1 L0 h Cmin Cmin h s L0
Cmax ∥σ ∥ h
2
+
Cmax ∥
∇ ⋅ σh ∥
2
+
L20
∥ ∥
uh
2
+
L0 ∥
∇ uh ∥ 2
≲
Cmin
f∥∥ 2 used is well known, and has been described already in Section 1. The
numerical problems found can be attributed to poor stability and/or
ð18Þ accuracy in the computation of the stresses. Since they are used to
evaluate the constitutive law (Eq. (23)), it is not surprising that a
where the divergence of the stresses in the left-hand-side has to be failure in calculating the stresses leads to a global failure of the overall
dropped if τu = 0. This estimate resembles very much Eq. (17) for the numerical approximation.
continuous problem. The only difference is the factor h instead of L0 in Our proposal in this work is simple: numerical instabilities present in
two terms of the left-hand-side. This however does not prevent from nonlinear solid mechanics using the irreducible formulation (i.e.,
obtaining the error estimates approximating Eq. (22)) could be at least alleviated if stability and/or
accuracy in the calculation of the stresses are improved. And this
1 L0 h
Cmax
σ−σh∥ ∥
2
+
Cmax ∥
∇ ⋅ ðσ−σh Þ ∥ 2
+
improvement can be achieved by using a mixed formulation. However,
the price to be paid is to use interpolations for the stresses and the
Cmin Cmin h s displacements that satisfy the inf-sup compatibility condition, and this
+
L20
u−uh ∥ ∥ 2
+
L0 ∥
∇ ðu−uh Þ ∥2
ð19Þ very often leads to non-standard (if not directly exotic) interpolating
pairs. The tool to overcome this is to resort to stabilized formulations, as
L0 2k + 1 2 C 2k + 1 2
≲ h j σ j k + 1 + max h j uj k + 1 we have shown so far.
Cmin L0
Even though we do not have the analysis for nonlinear problems,
the results presented in Section 4 suggest that success is possible. In
when interpolations of degree k are used for both the stresses and the
particular:
displacements.
The symbol | · |k + 1 denotes the L2(Ω) norm of the derivatives of • Stress stability is improved. From estimate Eq. (18) it is observed that
order k + 1 of the unknowns, which have been assumed sufficiently in the linear case stress stability is obtained without relying on the
regular. stability obtained for the displacement field.
The L2(Ω) estimates given in Eq. (19) can be improved using • Stress accuracy is improved, as it is clearly seen from Table 1 in linear
duality arguments. The analysis in [30] can be adapted to obtain elasticity. As a particular case, consider k = 1 (linear interpolation).
In the irreducible formulation the stresses are approximated with
h
∥σ−σ ∥ ≲ h∥∇ ⋅ ðσ−σ Þ∥ + C
h h max
L0 ∥
s
∇ ðu−uh Þ ∥ ð20Þ order h in the L2(Ω) norm. Without additional conditions on the
regularity of the solution and the shape of the elements of the finite
element mesh, point-wise estimates are expected to have one order
L0 h
∥u−u ∥ ≲ C ∥∇ ⋅ ðσ−σ Þ∥ + h∥∇ ðu−u Þ∥
h
min
h
s
h ð21Þ less of convergence. This means that no convergence order can be
guaranteed for the stresses that are used to evaluate the constitutive
law (Eq. (23)) point-wise. For the mixed stabilized formulation we
The results given by Eqs. (19)–(21) have been collected in Table 1, can formally expect order h convergence in the worst situation
indicating only the order of convergence. This order is compared with (order h1/2 if the assumptions of duality arguments do not apply).
what would be obtained in an irreducible formulation, where the
differential equation to be solved is In the following we describe how to formulate mixed stabilized
methods in the nonlinear case. The first point to keep in mind is that
 s 
−∇ ⋅ C : ∇ u = f ð22Þ results will be different depending on whether stresses or strains are
used as independent variables to be interpolated. In the linear case
It is clear from Table 1 that the stresses are approximated with a there is obviously no difference, since for constant constitutive tensors
better accuracy using the mixed stabilized formulation. C the space for the discrete strains εh = C−1:σh is the same as the
space for the discrete stresses σh, and formulating the mixed methods
3. Nonlinear problem presented in Section 2.4 in strains is trivial.

3.1. Motivation 3.2. Stress/displacement formulation

All the discussion presented heretofore is restricted to the mixed For the sake of conciseness, in this subsection we assume that ũ = 0.
stabilized formulation of the linear elasticity problem. In this work we Including displacement subscales in the following discussion is straight-
are interested in nonlinear constitutive behavior of materials of the form forward. The only remarks to be made are that the ASGS and the OSS
methods will not yield the same methods, as we have seen, and stability
s and convergence for the divergence of the stresses will be lost if ũ = 0.
C = CðσÞ or C = CðεÞ; ε = ∇ u ð23Þ

which in particular can be used to model damage. Note that plasticity- 3.2.1. General formulation
type models do not fall within this framework, because in that case the Introducing the scale splitting as described in Subsection 2.3.1 we
arrive at problem Eq. (5a)–Eq. (5d) also in the nonlinear case. In the
case ũ = 0, we may rewrite this problem as
Table 1
Order of convergence of different terms in the irreducible and mixed stabilized


  
−1 −1 s
formulations when interpolations of degree k are used.
− τh ; C : σh − τh ; C : σ̃ + τh ; ∇ uh = 0 ð24aÞ

Term Irreducible Mixed  s   s 

∥∇s(u − uh)∥ hk hk
∇ vh ; σh + ∇ vh ; σ̃ = ðvh ; f Þ ð24bÞ
∥ u − uh∥ hk + 1 with duality hk + 1/2 without duality, hk + 1 with duality
∥ σ − σh∥ hk (σh = C:∇suh) hk + 1/2 without duality, hk + 1 with duality

  s 
−1 −1
∥∇⋅(σ − σh)∥ hk − 1 (σh = C:∇uh) hk (if cu N 0 in Eq. (10)) −P̃ σ C : σ̃ = P̃ σ C : σh − P̃ σ ∇ uh ð24cÞ
2564 M. Cervera et al. / Computer Methods in Applied Mechanics and Engineering 199 (2010) 2559–2570

where Eq. (24c) corresponds to Eq. (8a). Let us see how to particu- • PC− 1≈Ph. At the computational level, it is much easier to deal with the
larize this general framework to the ASGS and the OSS methods. standard L2(Ω) projection than with the weighted one. In particular,
simpler numerical integration rules may be used. Likewise, lumping of
3.2.1.1. ASGS method. In this case P ̃σ = I when applied to the residual the matrix resulting from the projection is possible.
scaled by τσ, and we may approximate
3.3. Strain/displacement formulation
 s 
σ̃ = τσ C : ∇ uh −σh ð25Þ
3.3.1. General formulation
Note that if τσ ≠ 1 then σ̃+ σh ≠ C:∇suh. As it has been mentioned The formulation of the mixed solid mechanics problem in terms of
previously, the scaling of the residual by τσ can be understood as the stress and displacement fields, σ/u, is classical and it has been used
the upscaling of σ̃ to the finite element mesh. many times in the context of linear elasticity, where the constitutive
From Eqs. (24a)–(25) it follows that Eqs. (11a)–(11b) are still valid tensor C is constant. However, it is not the most convenient format for
in the nonlinear case, that is to say, the nonlinear problem. The reason for this is that most of the algorithms

 used for nonlinear constitutive equations in solid mechanics have been
−1  s 
− τh ; C : σh + τh ; ∇ uh = 0 ð26aÞ derived for the irreducible formulation. This means that these
procedures are usually strain driven, and they have a format in which
 s   s  the stress σ is computed in terms of the strain ε, with ε = ∇su.
ð1−τσ Þ ∇ vh ; σh + τσ ∇ vh ; C : ∇uh = ðvh ; f Þ ð26bÞ
Therefore, in order to be able to use the existing technology
available for the integration of nonlinear constitutive equations, it is
Even though the discrete problem is already given by Eqs. (26a)–
convenient to derive a mixed strain/displacement, ε/u, stabilized
(26b), it is suggestive to write it in a form similar to Eqs. (13a)–(13b).
formulation for the nonlinear solid mechanics problem. In view of the
Let PC− 1 denote the L2(Ω) projection onto the finite element space of
previous developments this is easily accomplished.
stresses weighted by C−1. Since (τh, ∇suh) = (τh, C−1:C:∇suh), we may
In this case, the strong form of the continuous problem can be
write Eq. (26a) as
stated as: for given prescribed body forces f, find the displacement
 s  field u and the strain field ε such that:
σh = PC −1 C : ∇ uh ð27Þ
s
−C : ε + C : ∇ u = 0 in Ω ð31aÞ
from where it follows that, similar to Eqs. (13a)–(13b), the ASGS
formulation can be expressed as ∇⋅ðC : εÞ + f = 0 in Ω ð31bÞ
 s   s 
σstab = ð1−τσ ÞPC −1 C : ∇ uh + τσ C : ∇ uh ð28aÞ
Eq. (31a) represents strain compatibility, while Eq. (31b) is the
 s  Cauchy equation. Eqs. (31a)–(31b) are subjected to appropriate
∇ vh ; σstab = Fðvh Þ ð28bÞ
Dirichlet and Neumann boundary conditions.
If V is, as before, the space of displacements and G the space of
Clearly, for constant constitutive tensors C there is no difference strains, following the standard procedure the associated weak form of
between Eqs. (28a)–(28b) and (13a)–(13b), but the weighted L2(Ω) the problem Eqs. (31a)–(31b) can be stated as:
projection should be in principle taken into account in nonlinear
constitutive models or simply when the medium is not homogeneous.  s 
−ðγ; C : εÞ + γ; C : ∇ u = 0 ∀γ ð32aÞ

3.2.1.2. OSS method. The first option would be to take P σ̃ = P⊥

h . In this
 s 
∇ v; C : ε = ðv; f Þ ∀v ð32bÞ
case, Eq. (24c) becomes



⊥ −1 ⊥ −1 ⊥ s  where v 2 V and γ 2 G are the test functions of the displacements and
−Ph C : σ̃ = Ph C : σh −Ph ∇ uh ð29Þ
strain fields, respectively. Eqs. (31a)–(31b) can be understood as the
stationary conditions of the classical (reduced) Hu–Washizu func-
However, it is not computationally simple to obtain an expression for tional [41].The discrete Galerkin finite element counterpart problem is:
the sub-grid stresses from this equation. To construct a basis for the
 s 
orthogonal to the space of stresses is required to invert the left-hand- −ðγh ; C : εh Þ + γh ; C : ∇ uh = 0 ∀γh ð33aÞ
side. A simpler and perhaps more natural option is to take T ̃ orthogonal
to T h with respect to PC− 1. From Eq. (24c) it immediately follows that  s 
∇ vh ; C : εh = Fðvh Þ ∀vh ð33bÞ
⊥  s 
σ̃ = τσ PC −1 C : ∇ uh ð30Þ where uh, vh 2 V h and εh, Υh 2 Gh are the discrete displacement and
strain fields and their test functions, defined onto the finite element
and, as for the linear elasticity problem, it can be shown that the OSS and spaces V h and Gh, respectively. Note that the resulting system of
the ASGS formulations coincide and are given by Eqs. (28a)–(28b). equations is symmetric but non-definite.
Stability considerations for the mixed ε/u are analogous to those
3.2.2. Simplifications of the σ/u format, so we proceed to present a stabilization method, using
System (28a)–(28b) can be approximated as is, but there are two the residual-based sub-grid scale approach, which allows in particular
approximations that simplify its numerical implementation: the use of linear/linear interpolations for displacements and strains. To
this end, the strain field of the mixed problem is approximated as
• C(σ) ≈ C(σh). Even though we have not explicitly indicated it earlier,
the dependence of C on the stresses needs to be approximated. One ε=εh + ε̃ ð34Þ
possibility is to use σstab given by Eqs. (28a)–(28b), although, since
the subscales are expected to be much smaller than the finite element where εh 2 Gh is the strain component of the (coarse) finite element
scales, C can be evaluated also with σh. This simplifies the scale and ε̃ 2 G̃ is the enhancement of the strain field corresponding to
implementation when the displacement subscales are accounted for the (finer) sub-grid scale. Let us also consider the corresponding test
(see Eqs. (11a)–(11b)). functions γh 2 Gh and γ̃ 2 G̃, respectively. The strain solution space is
 M. Cervera et al. / Computer Methods in Applied Mechanics and Engineering 199 (2010) 2559–2570 2565

G ≃ Gh ⨁ G̃. For simplicity, no subscale will be considered for the Similar to the stress–displacement formulation Eqs. (28a)–(28b),
displacement field for the moment. Its inclusion is considered in we can finally write the method we propose for the strain–
Subsection 3.4. Thus, considering only the strain subscale, the discrete displacement approach as
problem corresponding to Eqs. (32a) and (32b) is now:
 s   s 
   s  εstab = ð1−τε ÞPC ∇ uh + τε ∇ uh ð43aÞ
−ðγh ; C : εh Þ− γh ; C : ε̃ + γh ; C : ∇ uh = 0 ∀γh ð35aÞ
 
   s  s
∇ vh ; C : εstab = Fðvh Þ ð43bÞ
−ðγ̃; C : εh Þ− γ̃; C : ε̃ + γ̃; C : ∇ uh = 0 ∀γ̃ ð35bÞ
 s   s  This approach is of straightforward implementation.
∇ vh ; C : εh + ∇ vh ; C : ε̃ = Fðvh Þ ∀vh ð35cÞ
As in the previous section, some remarks are relevant:
As for the stress–displacement approach, the fact that the discrete 1. The stabilization term S2 is computed in an element by element
variational equations need to hold for all test functions will be omitted manner and, within each element, its magnitude depends on the
in the following. difference between the continuous (projected) and the discontin-
Due to the approximation used in Eq. (34), and the linear uous (elemental) strain fields. This means that the term added to
independence of εh and ε̃, the continuous Eq. (32a) unfolds in two secure a stable solution decreases upon mesh refinement, as the
discrete Eqs. (35a) and (35b), one related to each scale considered. finite element scale becomes finer and the residual (or the
Eqs. (35a) and (35c) are defined in the finite element spaces Gh and V h, projection of the residual) reduces (ε̃ is “small” compared to εh).
respectively. The first one enforces the constitutive equation including 2. With the definition in Eq. (38), the subscale ε̃ is discontinuous
a stabilization term S1 = (γh, C: ε̃) depending on the sub-grid strains ε̃. across element boundaries. For linear elements, ε̃ is piece-wise
The second one solves the balance of momentum including a linear. Therefore, even if defined element-wise, ε̃ cannot be
stabilization term S2 = (∇svh, C:ε̃) depending on the sub-grid stresses condensed at element level, because εh is interelement continuous.
σ̃ = C: ε̃. On the other hand, Eq. (35b) is defined in the sub-grid scale
space G̃ and, hence, it cannot be solved by the finite element mesh. The OSS formulation can be developed using the same reasoning as for
Following the same arguments introduced in the previous section, the stress–displacement approach. In this case, it is easy to show that if the
we can write Eq. (35b) as strain subscale is taken orthogonal to the finite element space with respect
to the L2(Ω) inner product weighted by C, the resulting formulation is
 
− γ̃; C : ε̃ = ðγ̃; rh Þ ð36Þ identical to the ASGS method. Details of the derivation are omitted.

where the residual of the constitutive equation in the finite element 3.3.2. Simplifications
scale is defined as: Analogously to the stress–displacement formulation, system Eqs.
(43a)–(43b) can be approximated as is, but there are two approxima-
s
rh = rh ðεh ; uh Þ = C : εh −C : ∇ uh ð37Þ tions that simplify the implementation:
• C(ε) ≈ C(εh).
In the case of the residual-based ASGS formulation, the subscale • PC ≈ Ph
stress can be localized within each finite element, and be expressed as
The same remarks as for the stress–displacement formulation are
−1  s 
ε̃ = τε C : rh = τε ∇ uh −εh ð38Þ applicable to these approximations.

where τε is computed in terms of an algorithmic constant cε as 3.4. Comparison between the σ/u and the ε/u formulations and final
numerical schemes
h
τε = cε ð39Þ
L0 As it has been mentioned, the stress–displacement and the strain–
displacement formulations will lead to (slightly) different results in
Introducing the strain subscale from Eq. (38) in Eq. (35a), the the nonlinear case. If we assume in both cases that σh = C:εh, we have
mixed system of equations can be written as obtained
 s 
−ð1−τε Þðγh ; C : εh Þ + ð1−τε Þ γh ; C : ∇ uh = 0 ð40aÞ −1
Stress  displacement : σh = PC −1 ðC : ∇uh Þ; εh = C : PC −1 ðC : ∇uh Þ
 s   s s 
ð1−τε Þ ∇ vh ; C : εh + τε ∇ vh ; C : ∇ uh = Fðvh Þ ð40bÞ
Strain  displacement : σh = C : PC ð∇uh Þ; εh = PC ð∇uh Þ
where the terms depending on τε represent the stabilization. Note
that the resulting system of equations is symmetric. and for the simplified formulations:
If PC is the L2(Ω) projection weighted by C, the projection involved
in Eq. (40a) can be written as −1
Stress  displacement : σh = Ph ðC : ∇uh Þ; εh = C : Ph ðC : ∇uh Þ
 s 
εh = PC ∇ uh ð41Þ
Strain  displacement : σh = C : Ph ð∇uh Þ; εh = Ph ð∇uh Þ
and, therefore, the weak form of the balance Eq. (40b), can be finally
written as: It is observed that only when C is constant both formulations coincide.
 s  s   s  For completeness, let us finally state the expression of the σ/u and
s
ð1−τε Þ ∇ vh ; C : PC ∇ uh + τε ∇ vh ; C : ∇ uh = Fðvh Þ ð42Þ ε/u mixed forms:

Stress  displacement : ð44aÞ

Eq. (38) does not need to be interpreted point-wise, as the values of
ε̃ are not used in the stabilization procedure; only the integral S2 in


−1 −1
− τh ; C : σh −τσ τh ; C : P̃ σ ðC : ∇uh −σh ÞÞ
Eq. (35c) is needed.
2566 M. Cervera et al. / Computer Methods in Applied Mechanics and Engineering 199 (2010) 2559–2570


 where BÂ is the matrix arising from applying the divergence operator
 s 
+ τh ; ∇ uh −τu ∇⋅τh ; P̃ u ð∇ ⋅ σh Þ = τu ∇ ⋅ τh ; P̃ u ðf Þ to the matrix product CtanNA.
When considering the efficient solution of system (Eq. (46)) three
 
 remarks have to be made:
∇s vh ; σh + τσ ∇s vh ; P̃ σ ðC : ∇uh −σh Þ = Fðvh Þ ð44bÞ • The monolithic solution of Eq. (46) can be substituted by an iterative
procedure, such as
Strain–displacement:
h i
ði Þ ði + 1Þ ðiÞ ðiÞ ði Þ

 −Mτ δE = −R1 − Gτ δU ð50aÞ
−ðγh ; C : εh Þ−τε γh ; C : P̃ ε ð∇uh −εh Þ ð45aÞ
h iT
 
 ðiÞ
Kτ δU
ði + 1Þ ðiÞ ðiÞ
= −R2 − Gτ δE
ði + 1 Þ
ð50bÞ
+ γh ; C : ∇s uh −τu ∇ ⋅ ðC : γh Þ; P̃ u ð∇ ⋅ ðC : εh ÞÞ


= τu ∇ ⋅ ðC : γh Þ; P̃ u ðf Þ • More efficient is to use an approximate staggered procedure, in
 s 
 which the strain projection is kept constant during the equilibrium
s
∇ vh ; C : εh + τε ∇ vh ; C : P̃ ε ð∇uh −εh Þ = Fðvh Þ ð45bÞ iterations within each time increment, taking it equal to an
appropriate prediction such as E(i + 1) ≅ E(0), computed from the
known values corresponding to the previous time steps (for
where the (simplified) projections are taken as P̃ = I for ASGS and
instance, a trivial prediction consists of taking E(0) ≅ E[n]). This
P̃ = P⊥
h for OSS and C = C(σh) or C = C(εh).
scheme leads to

4. Implementation and computational aspects


ðiÞ ði + 1Þ ð0Þ ði Þ
Kτ δU = −R2 E ; U ð51Þ
In this section, some relevant aspects concerning the implemen-
tation of the mixed strain/displacement scale stabilized method for • If τu = 0, using an appropriate integration scheme, the mass matrix
nonlinear solid mechanics formulated previously are described. Mτ can be rendered block-diagonal. The resulting lumped matrix Mτ̅
Implementation of the mixed stress/displacement scale stabilized is computationally more efficient. If τu ≠ 0, this matrix Mτ̅ is used as
method follows analogous arguments. preconditioner of the iterative strategy.
Due to the nonlinear dependence of the stresses on the strain and
displacements, the solution of the system of Eqs. (41)–(42) requires Independently of the solution strategy adopted, it is formally
the use of an appropriate incremental/iterative procedure such as the possible to express E = [M− 1 (i)
τ Gτ] U, and substitute this value in the

Newton–Raphson method. Within such a procedure, the system of equilibrium equation to obtain a reduced system of equations with
linear equations to be solved for the (i + 1)-th equilibrium iteration of the form:
the (n + 1)-th time (or load) step is: h iðiÞ
T −1
Kτ + Gτ Mτ Gτ U=F ð52Þ
" #ðiÞ  
−Mτ Gτ δE ði + 1Þ
R ðiÞ
=− 1 ð46Þ where matrices M(i) τ , Gτ
(i)
and K(i) are evaluated with a secant
Gτ
T
Kτ δU R2 τ
constitutive matrix, rather than tangent. If, as assumed in this work,
the strain field εh is interelement continuous, the elimination of the
where δE and δU are the iterative corrections to the nodal values for projection E is not feasible in practice, because the condensation
the strains and displacements, respectively, R1 and R2 are the residual procedure cannot be performed at element level; if performed at
vectors associated to the satisfaction of the kinematic and balance of global level it would yield a system reduced but with a spoiled banded
momentum equations, respectively, and the global matrices M(i) (i)
τ , Gτ structure. However, in this reduced format the overall effect of the
and K(i)
τ come from the standard assembly procedure of the elemental proposed stabilization method becomes self-evident. It is interesting
contributions. The global matrix is symmetric. Each one of the to note that it resembles the format of the enhanced assumed strain
elemental matrices to be assembled has an entry (·)AB, a sub-matrix method [5] and the more general mixed-enhanced strain method
corresponding to the local nodes A and B. Let us assume in the [47], where the enhancing fields are discontinuous and their variables
following that the same interpolation functions N are used for the can be condensed at local level.
strain and displacement fields.
Sub-matrix KABτ is obtained from the standard tangent stiffness 5. Numerical results
matrix:
In this section the formulation presented above is demonstrated in
AB T two benchmark problems and an additional illustrative example. In
Kτ = τε ∫ BA Ctan BB dΩ ð47Þ
Ωe the three cases, linear elastic constitutive behavior is assumed, with
the following material properties: Young's modulus E = 200 · 109 Pa,
where Ctan is the tangent constitutive matrix and B is the standard Poisson's ratio ν = 0.3. Examples concerning nonlinear constitutive
deformation sub-matrix. The generic term of the discrete symmetric behavior are presented in the companion paper [40].
gradient matrix operator GAB is given by: The first two tests are used in Ref. [5] to validate the Enhanced
AB T
Assumed Strain method. Performance of the formulation is tested
Gτ = ð1−τε Þ∫ BA Ctan NB dΩ ð48Þ considering 2D plane-strain quadrilateral and triangular structured
Ωe
meshes. The elements used are: P1 (linear displacement), P1P1 (linear
Finally, MAB is a “mass” matrix associated to the strain field: strain/linear displacement), Q1 (bilinear displacement), Q1Q1 (bilin-
ear strain/bilinear displacement) and Q1E4 (bilinear displacement
with enhanced strains [5]).
AB T
Mτ = ð1−τε Þ∫ NA Ctan NB dΩ + τu ∫ B̂A B̂B dΩ
T
ð49Þ When the stabilized mixed strain/displacement formulation is
Ωe Ωe used, values cε = 1.0 and cu = 0.1 are taken for the evaluation of the
 M. Cervera et al. / Computer Methods in Applied Mechanics and Engineering 199 (2010) 2559–2570 2567

stabilization parameters τε and τu, respectively. We have chosen

Cmin = E, understanding that the Young's modulus E is a characteristic
value of the elastic tensor (constants appearing in the minimum
eigenvalue of the elastic tensor may be included in the algorithmic
constant cu in Eq. (10)).
Calculations are performed with an enhanced version of the finite
element code COMET [48], developed by the authors at the
International Center for Numerical Methods in Engineering
(CIMNE). Pre and post-processing is done with GiD [49], also
developed at CIMNE. The stabilized system of equations resulting
from the mixed method, Eq. (46) is solved both in a monolithic way
and using the iterative algorithm in Eqs. (50a)–(50b).

5.1. Cook's membrane problem

The Cook membrane problem is a bending dominated example that

has been used by many authors as a reference test to check their element Fig. 2. Cook's membrane problem. Vertical displacement at point A versus number of
formulations. Here it will be used to compare results for the mixed and elements along each side.
irreducible formulations in compressible elasticity, showing the
behavior of both bilinear quadrilateral and linear triangular elements. irreducible and strain enhanced elements correspond to the continuous
The problem consists of a tapered panel, clamped on one side and projection Ph(C:∇suh) evaluated at the mesh nodes, rather than the
subjected to a shearing vertical load at the free end. Geometry of this actual discontinuous stresses C:∇suh evaluated at the integration
plane-strain problem is shown in Fig. 1 (dimensions are in mm). For points. This projection procedure yields improved stress values for
the evaluation of the stabilization parameters in the mixed formula- the irreducible formulation. Relative convergence characteristics on the
tion, L0 = 50 mm is taken as the representative length of the problem. stress values among the different elements compared are very similar
In order to test the convergence behavior of the different to those observed for the displacements, and faster for the mixed and
formulations, the problem has been discretized into structured enhanced formulations than for the irreducible one.
meshes with N finite elements along each side. Figs. 2 and 3 compare Regarding relative computational efficiency, Table 2 shows the CPU
the results obtained with five different spatial discretizations. time used by the different formulations to solve increasingly fine
Fig. 2 shows the relative convergence of the five discretizations on quadrilateral meshes. A direct solver with skyline storage has been used
the computed value of the vertical displacement at the right top in all cases. The enhanced strain formulation is only marginally more
corner of the membrane (point A in Fig. 1). Results are clearly different expensive than the standard one, because the size of the system of
for the irreducible and the mixed elements and it is evident that the equations to be solved is identical in both cases, even if the
mixed formulation performs better for coarse and fine meshes. It also corresponding matrices and vector are more elaborated for the first
shows a slightly faster convergence rate. The performance of the one. The efficiency of the mixed formulation depends very much on the
mixed Q1Q1 is very similar to the Q1E4 quadrilateral, although the implementation scheme adopted. On one hand, the relative cost of
latter seems to be more accurate on very coarse meshes and slightly solving the monolithic solution of system Eq. (46), labelled Q1Q1 (m) in
less accurate when the mesh is refined. the Table, grows quickly with the number of nodes in the mesh. Memory
Fig. 3 shows similar results on the computed value of the major requirements for direct solvers also increase rapidly in this case. On the
principal stress at the mid-side point of the bottom boundary of the other hand, the iterative solution of Eqs. (50a)–(50b), labelled Q1Q1 (i)
membrane (point B in Fig. 1). It has to be noted that in order to compare in the Table, can be obtained at a cost that compares reasonably with
stress values computed at the same point, the values reported for the those of the irreducible and enhanced strain formulations.
It has to be remarked that the relative increase in the cost of the
mixed formulation with regard to the irreducible one is smaller in
nonlinear problems, because iterations can be performed to solve the
mixed problem and the nonlinearity in a concurrent manner.

Fig. 3. Cook's membrane problem. Principal stress at point B versus number of elements
Fig. 1. Geometry for the Cook membrane problem (dimensions are in mm). along each side.
2568 M. Cervera et al. / Computer Methods in Applied Mechanics and Engineering 199 (2010) 2559–2570

Table 2
CPU time (s) for different quadrilateral meshes and formulations in Cook's membrane
problem: Q1, Q1/E4, Q1/Q1 (iterative) and Q1/Q1 (monolithic).

Mesh Q1 Q1/E4 Q1/Q1 (i) Q1/Q1 (m)

64 × 64 0.26 0.31 0.43 2.07

128 × 128 1.78 2.11 2.59 25.03
256 × 256 20.10 20.90 24.80 300.38

This relative increase would also be smaller using iterative solvers,

first because of the lower complexity (cost increase with the number
of unknowns) and, second, because the solution of a given nonlinear
iteration will be a better and better guess for the linear solver as the
iterative procedure goes on.

5.2. Clamped arch problem

Fig. 6. Clamped arch problem. Vertical displacement of point B versus number of
elements along the thickness.
As a further illustration of the performance of the stabilized mixed
ε/u formulation, we consider a clamped arch, of radius R = 10 and
thickness t = 1, vertically loaded at the top (see Fig. 2, dimensions are
in m). Because of symmetry, only one half of the structure needs to be As in the previous example, Figs. 5 and 6 compare the results
considered. The problem has been discretized into structured meshes obtained with five different spatial discretizations: Q1/Q1, P1/P1, Q1,
consisting of N finite elements along the radial direction and 10 N Q1E4 and P1. Fig. 5 shows the relative convergence of the five
elements in the circumferential direction. Length L0 = t is taken as discretizations used on the computed value of the vertical displace-
representative of the problem, for the evaluation of the stabilization ment under the point load (point A in Fig. 2). In this case, the mixed
parameters in the mixed formulation (Fig. 4). interpolations also show improved performance over their
corresponding irreducible formulations in the displacement results.
The quadrilateral mixed elements also compare well with the
quadrilaterals with enhanced strains, which are very accurate for all
meshes.
Fig. 6 shows results on the computed value of the major principal
stress at point B on the outer face of the arch (see Fig. 4). The values
reported for the irreducible and enhanced elements correspond to the
continuous projection Ph(C:∇suh) evaluated at the mesh nodes. Again,
the mixed formulations show better accuracy that the corresponding
irreducible ones. The quadrilateral mixed elements and the quadri-
lateral elements with enhanced strains show almost identical
performance in terms of stresses.

5.3. Sharp V-notched specimen under tension

For this last example, let us consider the vertical stretching of a

square V-notched specimen as the one shown in Fig. 7. Dimensions of
the sample are 2 × 2 m × m (width × height) and the V-shaped notch
has a length of 1 m and a maximum width at the boundary of 0.02 m.
Fig. 4. Geometry for the clamped arch problem. For the evaluation of the stabilization parameters in the mixed

Fig. 5. Clamped arch problem. Vertical displacement of point A versus number of

elements along the thickness. Fig. 7. Geometry for the sharp V-notched specimen under tension.
 M. Cervera et al. / Computer Methods in Applied Mechanics and Engineering 199 (2010) 2559–2570 2569

Fig. 8. Principal stresses for V-notched specimen under tension.

formulation, L0 = 1 m is taken as representative length of the problem. principal stress value is detected exactly at the tip of the notch;
Uniform vertical displacements of opposed sign are imposed at the computed stress directions are also noticeably improved. The impor-
top and bottom boundaries. tance of these two features in nonlinear solid mechanics is evident. As it
In the continuous elastic problem associated to this situation, the is shown in Part II of this work [40], they are crucial in strain localization
strain and stress fields are singular at the tip of the sharp notch. The problems where the constitutive equation depends on the principal
discrete model corresponding to the irreducible finite element formu- stress values and their directions.
lation performs satisfactorily in terms of a global error norm, but
approximates very poorly the actual behavior near the singular points. 6. Conclusions
To show this, a coarse structured mesh consisting of 8 × 8 × 2 P1
triangles with a ±45° bias is constructed. Fig. 8 (P1 coarse) depicts This paper presents the formulation of stable mixed stress/
principal stresses computed on this mesh, plotted on top of the contour displacement and strain/displacement finite elements using equal
lines for the major principal stress value. Note the strong mesh bias order interpolation for the solution of nonlinear problems is solid
dependence that is observed in front of and behind the notch tip. In fact, mechanics. The proposed stabilization is based on the sub-grid scale
the largest values of the stresses occur behind the tip (left of the tip in approach and it circumvents the strictness of the compatibility
the Figure), rather than in front of it (right of the tip in the Figure). conditions. The final method, consisting of stabilizing the standard
Computed stress directions near the tip of the crack also show strong formulation for mixed elements with the projection of the displacement
mesh bias dependence. Fig. 8 (P1 fine) depicts principal stresses symmetric gradient, yields an accurate and robust scheme, suitable
computed on a finer structured mesh consisting of 64 × 64× 2 P1 for engineering applications in 2D and 3D. Numerical examples show
triangles with the same ±45° bias. A zoom on the area around the tip of that results compare favorably with the corresponding irreducible
the crack is shown, where the same errors as in the coarse mesh are formulations, showing improved accuracy in the evaluation of the stress
displayed. Comparing the results obtained for both meshes, it can be field. This characteristic is of great importance when facing nonlinear
appreciated that the severe local errors caused by the mesh alignment problems.
are not alleviated by mesh refinement.
Fig. 8 (P1/P1 coarse) and (P1/P1 fine) show corresponding results Acknowledgment
obtained used the stabilized mixed strain/displacement formulation on
the same coarse and fine meshes. The improved accuracy with respect Financial support from the Spanish Ministry for Education and
to the irreducible formulation is clear. In particular, the maximum Science under the SEDUREC project (CSD2006-00060) is acknowledged.
2570 M. Cervera et al. / Computer Methods in Applied Mechanics and Engineering 199 (2010) 2559–2570

References [26] R. Codina, J. Blasco, A finite element method for the Stokes problem allowing equal
velocity–pressure interpolations, Comput. Meth. Appl. Mech. Eng. 143 (1997)
[1] B.X. Fraeijs de Veubeke, Displacement and equilibrium models in the finite 373–391.
element method, in: O.C. Zienkiewicz, G. Hollister (Eds.), Stress Analysis, Wiley, [27] R. Codina, Stabilization of incompressibility and convection through orthogonal
1965. sub-scales in finite element methods, Comput. Meth. Appl. Mech. Eng. 190 (2000)
[2] D.S. Malkus, T.J.R. Hughes, Mixed finite element methods — reduced and selective 1579–1599.
integration techniques: a unification of concepts, Comput. Meth. Appl. Mech. Eng. [28] R. Codina, Analysis of a stabilized finite element approximation of the Oseen
15 (1978) 63–81. equations using orthogonal subscales, Appl. Numer. Math. 58 (2008) 264–283.
[3] D.N. Arnold, F. Brezzi, M. Fortin, A stable finite element for the Stokes equations, [29] R. Codina, Finite element approximation of the three field formulation of the Stokes
Calcolo 21 (1984) 337–344. problem using arbitrary interpolations, SIAM J. Numer. Anal. 47 (2009) 699–718.
[4] J.C. Simo, R.L. Taylor, K.S. Pister, Variational and projection methods for the [30] S. Badia, R. Codina, Unified stabilized finite element formulations for the Stokes
volume constraint in finite deformation elasto-plasticity, Comput. Meth. Appl. and the Darcy problems, SIAM Journal on Numerical Analysis 17 (2009) 309–330.
Mech. Eng. 51 (1985) 177–208. [31] M. Chiumenti, Q. Valverde, C. Agelet de Saracibar, M. Cervera, A stabilized
[5] J.C. Simo, M.S. Rifai, A class of mixed assumed strain methods and the method of formulation for incompressible elasticity using linear displacement and pressure
incompatible modes, Int. J. Numer. Meth. Eng. 29 (1990) 1595–1638. interpolations, Comput. Meth. Appl. Mech. Eng. 191 (2002) 5253–5264.
[6] B.D. Reddy, J.C. Simo, Stability and convergence of a class of enhanced assumed [32] M. Cervera, M. Chiumenti, Q. Valverde, C. Agelet de Saracibar, Mixed linear/linear
strain methods, SIAM J. Num. Anal. 32 (1995) 1705–1728. simplicial elements for incompressible elasticity and plasticity, Comput. Meth.
[7] J. Bonet, A.J. Burton, A simple average nodal pressure tetrahedral element for Appl. Mech. Eng. 192 (2003) 5249–5263.
incompressible and nearly incompressible dynamic explicit applications, Com- [33] M. Chiumenti, Q. Valverde, C. Agelet de Saracibar, M. Cervera, A stabilized
mun. Numer. Meth. Eng. 1 (4) (1998) 437–449. formulation for incompressible plasticity using linear triangles and tetrahedra,
[8] O.C. Zienkiewicz, J. Rojek, R.L. Taylor, M. Pastor, Triangles and tetrahedra in Int. J. Plast. 20 (2004) 1487–1504.
explicit dynamic codes for solids, Int. J. Numer. Meth. Eng. 43 (1998) 565–583. [34] M. Cervera, M. Chiumenti, C. Agelet de Saracibar, Softening, localization and
[9] R.L.A. Taylor, Mixed-enhanced formulation for tetrahedral elements, Int. J. Num. stabilization: capture of discontinuous solutions in J2 plasticity, Int. J. Numer.
Meth. Eng. 47 (2000) 205–227. Anal. Meth. Geomechanics 28 (2004) 373–393.
[10] C.R. Dohrmann, M.W. Heinstein, J. Jung, S.W. Key, W.R. Witkowsky, Node-based [35] M. Cervera, M. Chiumenti, C. Agelet de Saracibar, Shear band localization via local J2
uniform strain elements for three-node triangular and four-node tetrahedral continuum damage mechanics, Comput. Meth. Appl. Mech. Eng. 193 (2004)
meshes, Int. J. Numer. Meth. Eng. 47 (2000) 1549–1568. 849–880.
[11] J. Bonet, H. Marriot, O. Hassan, An averaged nodal deformation gradient linear [36] M. Cervera, M. Chiumenti, Size effect and localization in J2 plasticity, International
tetrahedral element for large strain explicit dynamic applications, Commun. Numer. Journal of Solids and Structures 46 (2009) 3301–3312.
Meth. Eng. 17 (2001) 551–561. [37] M. Pastor, T. Li, X. Liu, O.C. Zienkiewicz, Stabilized low-order finite elements for
[12] J. Bonet, H. Marriot, O. Hassan, Stability and comparison of different linear failure and localization problems in undrained soils and foundations, Comput.
tetrahedral formulations for nearly incompressible explicit dynamic applications, Meth. Appl. Mech. Eng. 174 (1999) 219–234.
Int. J. Numer. Meth. Eng. 50 (2001) 119–133. [38] M. Mabssout, M.I. Herreros, M. Pastor, Wave propagation and localization
[13] E. Oñate, J. Rojek, R.L. Taylor, O.C. Zienkiewicz, Linear triangles and tetrahedra for problems in saturated viscoplastic geomaterials, Comput. Meth. Appl. Mech.
incompressible problem using a finite calculus formulation, Proceedings of Eng. 192 (2003) 955–971.
European Conference on Computational Mechanics, ECCM, 2001. [39] M. Mabssout, M. Pastor, A Taylor–Galerkin algorithm for shock wave propagation
[14] E.A. de Souza Neto, F.M.A. Pires, D.R.J. Owen, A new F-bar-method for linear and strain localization failure of viscoplastic continua, Int. J. Numer. Meth. Eng. 68
triangles and tetrahedra in the finite strain analysis of nearly incompressible (2006) 425–447.
solids, Proceedings of VII International Conference on Computational Plasticit, [40] Cervera, M., Chiumenti, M. and Codina, R.. Mixed stabilized finite element
COMPLAS, 2003. methods in nonlinear solid mechanics. Part II: strain localization, Comp. Meth. in
[15] O.C. Zienkiewicz, R.L. Taylor, J.A.W. Baynham, Mixed and irreducible formulations Appl. Mech. and Eng. (this issue).
in finite element analysis, in: S.N. Atlury, R.H. Gallagher, O.C. Zienkiewicz (Eds.), [41] J.K. Djoko, B.P. Lamichhane, B.D. Reddy, B.I. Wohlmuth, Conditions for equivalence
Hybrid and Mixed Finite Element Methods, Wiley, 1983. between the Hu–Washizu and related formulations, and computational behavior in
[16] O.C. Zienkiewicz, R.L. Taylor, The Finite Element Method, Butterworth-Heine- the incompressible limit, Comput. Meth. Appl. Mech. Eng. 195 (2006) 4161–4178.
mann, Oxford, 2000. [42] R. Codina, J. Principe, J. Baiges, Subscales on the element boundaries in the variational
[17] M. Bischoff, K.-U. Bletzinger, Improving stability and accuracy of Reissner–Mindlin two-scale finite element method, Comput. Meth. Appl. Mech. Eng. 198 (2009)
plate finite elements via algebraic subgrid scale stabilization, Comput. Meth. Appl. 838–852.
Mech. Eng. 193 (2004) 1517–1528. [43] Badia, S. and Codina, R. Stabilized continuous and discontinuous Galerkin techniques
[18] D.N. Arnold, Mixed finite element methods for elliptic problems, Comput. Meth. for Darcy flow, Comp. Meth. Appl. Mech. Eng. 199 (2010) 1654–1667. E-prints UPC:
Appl. Mech. Eng. 82 (1990) 281–300. http://upcommons.upc.edu/e-prints/handle/2117/2447.
[19] F. Brezzi, M. Fortin, Mixed and Hybrid Finite Element Methods, Spinger, New York, [44] R. Codina, Stabilized finite element approximation of transient incompressible
1991. flows using orthogonal subscales, Comput. Meth. Appl. Mech. Eng. 191 (2002)
[20] F. Brezzi, M. Fortin, D. Marini, Mixed finite element methods with continuous 4295–4321.
stresses, Math. Models Meth. Appl. Sci. 3 (1993) 275–287. [45] C. Baiocchi, F. Brezzi, L. Franca, Virtual bubbles and Galerkin/least-squares type
[21] D. Mijuca, On hexahedral finite element HC8/27 in elasticity, Comput. Mech. 33 methods (Ga.L.S.), Comput. Meth. Appl. Mech. Eng. 105 (1993) 125–141.
(2004) 466–480. [46] F. Brezzi, M.O. Bristeau, L. Franca, M. Mallet, G. Rogé, A relationship between
[22] D.N. Arnold, R. Winther, Mixed finite elements for elasticity, Numer. Math. 92 stabilized finite element methods and the Galerkin method with bubble functions,
(2002) 401–419. Comput. Meth. Appl. Mech. Eng. 96 (1992) 117–129.
[23] D.N. Arnold, G. Awanou, R. Winther, Finite elements for symmetric tensors in [47] E.P. Kasper, R.L. Taylor, A mixed-enhanced strain method. I: Geometrically linear
three dimensions, Math. Comput. 77 (2008) 1229–1251. problems. II: Geometrically nonlinear problems, Comput. Struct. 75 (2000) 237-250,
[24] T.J.R. Hughes, Multiscale phenomena: Green's function, Dirichlet-to Neumann 251-260.
formulation, subgrid scale models, bubbles and the origins of stabilized [48] Cervera, M., Agelet de Saracibar, C. and Chiumenti, M. COMET: COupled
formulations, Comput. Meth. Appl. Mech. Eng. 127 (1995) 387–401. MEchanical and Thermal analysis. Data Input Manual, Version 5.0, Technical
[25] T.J.R. Hughes, G.R. Feijoó, L. Mazzei, J.B. Quincy, The variational multiscale method— report IT-308, http://www.cimne.upc.es, 2002.
a paradigm for computational mechanics, Comput. Meth. Appl. Mech. Eng. 166 [49] GiD: The personal pre and post preprocessor. http://www.gid.cimne.upc.es, 2002.
(1998) 3–28.