Sie sind auf Seite 1von 24

COMMUNICATIONS IN NUMERICAL METHODS IN ENGINEERING Commun. Numer. Meth. Engng 2003; 19:809832 (DOI: 10.1002/cnm.

622)

Mixed variational methods for nite element analysis of geometrically non-linear, inelastic BernoulliEuler beams
K. D. Hjelmstad1; and E. Taciroglu2
1 Department

of Civil and Environmental Engineering; University of Illinois at Urbana-Champaign; Urbana; IL; 61801; U.S.A. 2 Center for Simulation of Advanced Rockets; University of Illinois at Urbana-Champaign; Urbana; IL; 61801; U.S.A.

SUMMARY BernoulliEuler beam theory has long been the standard for the analysis of reticulated structures. The need to accurately compute the non-linear (material and geometric) response of structures has renewed interest in the application of mixed variational approaches to this venerable beam theory. Recent contributions in the literature on mixed methods and the so-called (but quite related) non-linear exibility methods have left open the question of what is the best approach to the analysis of beams. In this paper we present a consistent computational approach to one-, two-, and three-eld variational formulations of non-linear BernoulliEuler beam theory, including the e ects of non-linear geometry and inelasticity. We examine the question of superiority of methods through a set of benchmark problems with features typical of those encountered in the structural analysis of frames. We conclude that there is no clear winner among the various approaches, even though each has predictable computational strengths. Copyright ? 2003 John Wiley & Sons, Ltd.
KEY WORDS:

beam theory; frame analysis; nite elements; mixed variational principles; HuWashizu; HellingerReissner; non-linear exibility methods

1. INTRODUCTION Virtually all analysis of reticulated structures done today is based upon BernoulliEuler beam theory. Most commercial programs available to do this sort of analysis allow only the consideration of prismatic beams (i.e. the properties do not change along the length of the beam). However, most of these programs do not include capabilities to do geometrically non-linear analysis or inelastic analysis. There is a growing awareness among structural engineers that these limitations are not compatible with the limit-states philosophies, particularly
Correspondence

to: Prof. K. D. Hjelmstad, Department of Civil and Environmental Engineering, University of Illinois at Urbana-Champaign, Urbana, IL 61801, U.S.A.

Contract=grant sponsor: The Department of Energy, Center for Simulation of Advanced Rockets

Copyright ? 2003 John Wiley & Sons, Ltd.

Received 18 July 2002 Accepted 15 January 2003

810

K. D. HJELMSTAD AND E. TACIROGLU

performance-based design, commonly used to design structures for extreme environments in which non-linear response is likely. It is rather straightforward to obtain an exact solution to the BernoulliEuler beam equations for linear, elastic, prismatic beams. The simplicity of treating the problem with classical methods has probably had more to do with the ubiquity of prismatic beam analysis than has the actual incidence of prismatic beams. That said, the importance of correctly computing the sti ness coe cients of non-prismatic beams has long been recognized. The sti ness coe cients for non-prismatic beams are generally obtained with the classical principle of virtual forces because that approach a ords an exact solution [1]. Recently, the merits of the so-called exibility approach to determining sti ness coe cients has turned the attention of structural analysts to non-linear exibility methods [25]. The primary goal of these investigations has been to improve the accuracy of analysis with BernoulliEuler beam theory when the constitutive equations are non-linear. Some of the methods proposed have a consistent variational structure while others do not. The connection between the non-linear exibility methods and certain mixed variational methods has been recognized [6] and yet it seems that mixed methods have not been completely and systematically explored for the BernoulliEuler beam. In particular, non-linear geometric e ects have yet to be included and the relative merits of the various approaches is still unclear. The sampling of methods of analysis that have appeared in the literature have included contributions that are far from transparent and the literature is spotted with unsubstantiated claims of superiority of one method over another. The purpose of the present paper is to exhaust the range of possibilities of variational approaches to the BernoulliEuler beam and to attempt, therefore, to provide a context to make comparisons among them. We shall attempt to show in this paper that there is no clear winner among the three viable variational possibilities (one-, two-, and three-eld formulations) in terms of computational e ciency, at least for the types of problems typically encountered in non-linear structural analysis. By that same token, we shall see that each approach has a rather predictable strength in its ability to approximate the response of the beam. In this paper we extend our earlier work [7] in two respects. First, we extend the formulation of mixed methods for BernoulliEuler beams to geometrically non-linear problems and coupled (momentaxial interaction) inelasticity. Second, we study the performance of the various formulations across a set of benchmark problems specically designed to excite certain features of response. We make an attempt to do some cost accounting of the methods with h- and p-renement to clarify the question of which formulation is the best for general use. This paper also extends the work of Petrangeli and Ciampi [8] through the incorporation of new approaches for three-eld problems and a more complete discussion of the computational treatment (e.g. condensation and recovery) of the mixed variables. These extensions draw on the work in mixed methods for three-dimensional nite element technology [816]. This paper answers the question of how to extend non-linear exibility approaches to geometrically non-linear problems. 2. NON-LINEAR BERNOULLIEULER BEAM THEORY The classical equilibrium and straindisplacement equations of BernoulliEuler beam theory are n = p; U = u + 1 (w )2 2 m (nw ) = q;
Copyright ? 2003 John Wiley & Sons, Ltd.

=w

(1)

Commun. Numer. Meth. Engng 2003; 19:809832

MIXED VARIATIONAL METHODS FOR FINITE ELEMENT ANALYSIS

811

m n q n u p
(a) (b)

Figure 1. Planar BernoulliEuler beam: (a) denition of position and loading; and (b) denition of positive axial and transverse displacement and positive axial force and bending moment.

where m is the bending moment eld, n is the axial force eld, q and p are the applied transverse and axial load per unit length, respectively, is the curvature eld, U is the axial strain eld, w is the transverse displacement eld, and u is the axial displacement eld. Each of these elds is a function of the axial coordinate , which is measured from the left end of the beam, as shown in Figure 1. This gure also indicates the convention for positive values of the eld variables. A prime denotes di erentiation with respect to . Throughout this paper lowercase characters represent eld variables that depend upon . Let us dene generalized displacement, stress, strain, and load vectors, respectively, as u [u; w]T ; s [n; m]T ; e [U; ]T ; p [p; q]T (2)

Let us also dene a straindisplacement operator


U(u) [u + 1 (w )2 ; w ]T 2

(3)

For the variational methods described herein, it is important to maintain a distinction between strain as an independent eld, e, and strain computed as derivatives of the displacement eld, U(u). Variational methods rely on the introduction of virtual elds that represent variations of the independent elds. Let the virtual displacements, stresses, and strains be denoted u [u; w]T ; s [n; m]T , and e [U; ]T , respectively. When the virtual strain is to be computed from the virtual displacement then we have d [U(u + au)]a=0 = [u + w w ; w ]T (4) da where DU(u) u stands for the directional derivative of the function U(u) in the direction u. Let us also dene the matrix
U(u) = DU(u) u

(u) and the di erential operator

1 0

w 0

0 1

(5)

B(u) [u ; w ; w ]T

(6)

With this notation DU(u) u = (u)B(u). We will often have need of B( u) and B(u) in our formulations. These vectors can be obtained by using the incremental displacement eld and the virtual displacement eld, respectively, in Equation (6).
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

812

K. D. HJELMSTAD AND E. TACIROGLU

m mo f (s, ) = 0

no

Figure 2. The yield surface for stress resultants.

In BernoulliEuler beam theory the bending moment m and axial force n play the role of the internal stress resultants, while the curvature and axial strain U are the corresponding strain resultants. In general, we can express the constitutive hypothesis as g(s; e) = 0. It will be convenient, however, to also think of the constitutive equations in a strain-driven or stressdriven format (i.e. the stress computed given the strain or the strain computed given the stress, respectively). Thus, we shall write the constitutive equations alternatively (and equivalently) as s = s(e); e = e(s) (7)

For example, for a linear, elastic beam with centroidal axes, the constitutive function takes the form s(e) = De, where D = diag[EA; EI] is the sti ness of the cross section. Models of inelasticity generally include internal variables (e.g. plastic strain and hardening variables). We shall always assume that equations for the internal variables are satised pointwise (at the Gauss points of numerical integration). In Section 3 we present a simple model of inelasticity to make denite the treatment of the constitutive equations for history dependent response.

3. SIMPLE MODEL FOR INELASTICITY IN STRESS RESULTANTS One of the most common sources of material non-linearity is inelasticity. Inelasticity provides a good benchmark for comparing competing algorithms for beam problems because cyclic inelasticity can lead to curvature and residual stress elds that are much more complicated than typically arise in elastic problems. This section describes a very simple model of inelasticity and allows us to show clearly how the constitutive equations of inelasticity manifest in the variational formulations. Let the stress be given from the elastic strains as s = D[e e p ], where e p [U p ; p ]T is the plastic strain and D diag[EA; EI] is the elastic modulus. Let the yield function be dened as f(s; ) s Qs k(), where the metric Q is dened through the relationship Q1 diag[n2 ; m2 ] and where n0 and m0 are the axial yield force and the yield moment, re0 0 spectively. Figure 2 illustrates the initial yield surface. For simplicity in the examples we will consider a linear strain hardening function k() = 1 + , where is the hardening modulus. More general models of hardening present no additional di culty.
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

MIXED VARIATIONAL METHODS FOR FINITE ELEMENT ANALYSIS

813

States are elastic inside the yield surface. Plastic strains accrue for states on the yield surface as e p q(s; ), where is the consistency parameter and is the internal variable governing (isotropic) strain hardening. Assuming associative ow, we have q @f = @s Qs s Qs (8)

The evolution of the hardening variable is given by e p Q1 e p = , which is an ordinary strain hardening law. The denition of the plasticity formulation is completed with the classical complementarity conditions ( 0; f(s; )60, and f(s; ) = 0) and the consistency condition, f(s; ) = 0. 3.1. Return-mapping algorithm for strain-driven formulations The constitutive rate equations can be discretized in time by integrating with the backwardEuler method. For the time interval [tn ; tn1 ], let us dene tn tn tn1 and let the discrete consistency parameter be n tn n . Note that a subscript n indicates that the variable is evaluated at time tn . Let qn Qsn sn Qsn
nqn]

(9)

where sn is the stress at time tn . Integrating the constitutive equations we have [17] sn = sn1 + D[en en1
p p en = en1 + nqn n n,

(10)

n = n1 +

from which it is clear that the new state depends upon the consistency parameter is subject to the discrete complementarity conditions
n 0;

which (11)

fn 60;

n fn

=0

A trial stress can be dened by assuming the process for the current time interval is elastic trial (i.e. assuming n = 0). To wit, sn sn1 + D[en en1 ]. The yield function, evaluated at trial trial the trial stress is fn f(sn ; n1 ). The complementarity conditions require that if fntrial 60 then the trial state is the actual state and n = 0. The new state can then be computed from Equation (10). On the other hand, if fntrial 0 then the trial state is not the actual state. It follows from the complementarity conditions that since n 0 the actual state must necessarily satisfy fn = 0. Dening R( ) k(n1 + )Q1 + D, we can compute the following recursion: q = R1 (
trial 1 )sn ;

q =

q ; q Q1 q

trial q sn k(n1 + q Dq

1 )

(12)

The iteration can be started with 0 = 0 and should continue until the residual of Eq. (10)a trial falls below a certain tolerance, i.e. |R( )q sn |tol, where tol is some predened tolerance. Upon convergence, update the consistency parameter as n = .
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

814

K. D. HJELMSTAD AND E. TACIROGLU

The consistent tangent matrix can be computed knowing the consistency parameter. Let %n n =kn , where kn k(n1 + n ), and let Dn [D1 + %n Q]1 . Then the tangent sti ness matrix is given by the expression [17] Dn Dn where, with kn k (n1 +
n n ),

n Dn q n
n

Dn q n has the expression n q n Dq n

(13)

the parameter

1 kn %n ; kn + (1 kn %n )n

(14)

Note that, if n = 0 (i.e. an elastic state) then @s=@e = D. If n 0, then the consistent tangent is @s=@e = Dn . The method to obtain the stress and the consistent tangent described in this section is apropos to both the classical sti ness and the mixed-enhanced formulations. 3.2. Return-mapping algorithm for stress-driven formulations The constitutive update procedure for stress-driven problems is slightly di erent from the strain-driven case. Given the stress vector sn (from the interpolation), we can evaluate a trial yield function fntrial f(sn ; n1 ). If fntrial 60 then the trial state is the actual state and the state can be updated as en = en1 + D1 [sn sn1 ]; n = n1 ;
p p en = en1 n 0

(15) and that

On the other hand, if fntrial 0 then the complementarity conditions imply that f(sn ; n ) = 0. This equation can be solved iteratively as
+1 n

1 k (n1 +
n)

( sn Qsn k(n1 +

n ))

(16)

Now, with

known, we can easily obtain the strain as en = en1 +


nqn

+ D1 [sn sn1 ]
1

(17)

Note that, if n = 0 (i.e. an elastic state) then @e=@s = D . If n 0, then the consistent tangent is @e=@s = D1 (see Equation (13)). This format is apropos to the HellingerReissner n formulation.

4. VARIATIONAL APPROACHES TO NUMERICAL APPROXIMATION Within the context of variational approaches to the problem of determining the unknown elds u; s, and e, given the loading p, there are three basic formulations [18, 19]. The classical sti ness approach allows variations of u and satises the equilibrium equations weakly. The straindisplacement and constitutive equations are strongly enforced. The HellingerReissner approach allows variations of u and s and satises the equilibrium and straindisplacement equations weakly. The constitutive equations are strongly enforced. Finally, the mixedenhanced (HuWashizu) approach allows variations of u, s, and e and satises all of the eld equations weakly. This section outlines the approach to computation based upon these di erent variational formulations.
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

MIXED VARIATIONAL METHODS FOR FINITE ELEMENT ANALYSIS

815

4.1. The one-eld (Classical Sti ness) approach The classical sti ness approach is based upon the displacement-based functional S(u; u)
0

[U(u)T s(U(u)) uT p] d

(18)

By the fundamental theorem of the calculus of variations, equilibrium is satised if S(u; u) = 0 for all variations u [20]. Let u i represent a conguration of the beam (not necessarily an equilibrium conguration). This functional can be linearized at the state u i to give L[S] = Si + Si , where Si = S(u i ; u) and Si d [S(u + d u; u)]u=u i =
0

BT (u)[

T i Di

i + Gi ]B( u i ) d

(19)

where i = (u i ); Di = @ s=@e is the tangent sti Gi = G(s i ) is the geometric sti ness 0 G(s) = 0 0

ness evaluated at the state u i , and the matrix 0 n 0 0

(20)

evaluated at the state s i . Note that the argument of the matrix G is the stress s. It should be understood that in the case of a displacement-based formulation the stress is evaluated from the displacement as s = s(U(u)). For the mixed formulations the stress s will be an independent eld and is used in G directly. In Newtons method we set L[S] = 0 for all u to provide a means of estimating an incremental state u i that, when added to u i , will yield a state that comes closer to satisfying S(u; u) = 0. The process can be repeated to convergence. In the classical sti ness method, we interpolate the displacement eld as a linear combination of known base functions h (which are functions of ) and unknown displacement parameters U as u = hT U. The incremental displacement and virtual displacement elds are interpolated similarly as u = hT U and u = hT U, respectively, where U and U are the discrete incremental and virtual displacement parameters, respectively. The matrix h interpolates both the axial displacement u and the transverse displacement w. The specic form of this matrix is hT g1 0 g2 0

gM 0

0 h1

0 h2

0 hN

(21)

The set of functions {g1 ; : : : ; gM } interpolate the axial displacement eld, where M is the total number of terms in the interpolation. The set of functions {h1 ; : : : ; hN } interpolate the transverse displacement eld, where N is the number of terms in the interpolation. This organization of h implies that the displacement unknowns are organized in the array U with the M axial displacement parameters followed by the N transverse displacement parameters. The minimal interpolation is two terms for the axial displacement, e.g. g1 = = and g2 = 1 = (to assure continuity of the displacement), and four terms for the transverse displacement, e.g. the cubic hermitian polynomial functions (to assure continuity of the displacement and
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

816

K. D. HJELMSTAD AND E. TACIROGLU

its rst derivative). Both interpolations can be augmented with additional functions as shown in Section 5. Substitution of the interpolated elds into the linearized classical sti ness functional yields T the discrete version L[S] = U [Ki Ui ru (u i )]. The condition L[S] = 0 for all U implies the discrete equation Ki Ui ru (u i ) = 0 where the tangent sti ness Ki can be computed as Ki
0

(22)

a[

T i Di

+ Gi ]aT d

(23)

and the residual force ru (u i ) is given from the general expression ru (u i ) =


0

[a

T i s(U(u i ))

hp] d

(24)

The matrix a is simply the result of interpolating B() g1 g2 gM 0 0 aT 0 0 0 h1 h2 0 0 0 h1 h2

and has the particular form 0 hN hN

(25)

The sti ness matrix and residual force can be assembled in the standard manner from element contributions to give the linearized (incremental) global equilibrium equations. Equilibrium is established for the non-linear case by iteratively solving the equation Ki Ui = ru (u i ) and updating the estimate of the displacements with the equation Ui+1 = Ui + Ui until the force residual is small enough, i.e. ru (u i ) tol. The iterative computation must be started with an estimate U0 and the tolerance tol on the satisfaction of equilibrium must be specied a priori. 4.2. The two-eld (HellingerReissner) approach Consider the following the two-eld (HellingerReissner type) functional for the Bernoulli Euler beam H (u; s; u; s)
0

[U(u)T s uT p sT ( (s) U(u))] d e

(26)

This functional depends upon the two elds s and u (and their virtual counterparts) and has the property that if H = 0 for all variations u and s then the state (u; s) satises the classical equilibrium and straindisplacement equations. This functional can be linearized about the state (u i ; s i ) to give L[H ] = Hi + Hi , where Hi = H (u i ; s i ; u; s) and Hi d [H (u + d u; s + s; u; s)](u=u i ; s=s i ) (27)

Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

MIXED VARIATIONAL METHODS FOR FINITE ELEMENT ANALYSIS

817

is the directional derivative of the functional in the direction of an increment of the two elds. The quantity Hi has the particular expression Hi =
0

[BT (u)(

T i

s + Gi B( u)) sT (D1 s i

i B(

u))] d

(28)

e where i and Gi are given by Equations (5) and (20), respectively, and D1 = @ =@s is the i tangent compliance evaluated at the state s i . As in the classical sti ness formulation, the terms with B() can be interpolated with the displacement interpolation h given in Equation (21), with the result being the matrix a. The stresses s must be interpolated independently. This interpolation will be expressed as s = bT S, where S is the vector of force parameters. We will organize the interpolation as bT b1 0 b2 0

bA 0

0 b1

0 b2

0 bB

(29)

The incremental and virtual stress elds are also interpolated with b as s = bT S and s = bT S, respectively. Although it is not required, we shall interpolate the axial force n and the bending moment m with the same base functions bi . The number of terms in the axial interpolation is A and the number in the bending interpolation is B. As will be discussed later, these functions will be selected as orthogonal polynomials, starting with the constant function. Again, we substitute the interpolated elds into the linearized functional L[H ]. The condition L[H ] = 0 for all U and S implies the discrete equations Gi Ui + Bi Si + ru (u i ; s i ) = 0 BiT Ui Fi Si + rs (u i ; s i ) = 0 (30)

where the matrix Bi , the geometric sti ness matrix Gi and the exibility matrix Fi are given by the expressions BiT
0

b i aT d ;

Gi
0

aGi aT d ;

Fi
0

bD1 bT d i

(31)

The residuals are dened as ru (u i ; s i )


0

[a

T i si

hq] d ;

rs (u i ; s i )
0

b[ (s i ) U(u i )] d e

(32)

There are basically three choices for implementing the global solution scheme for the mixed formulation. First, one can retain both Ui and Si as global variables. One would then form an element sti ness and residual for the combined unknowns as Ki Gi BiT Bi Fi ; r ru (u i ; s i ) rs (u i ; s i ) (33)

and assemble a combined global sti ness and residual in the standard manner, taking care to include the stress degrees of freedom in the global equation numbering scheme. The
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

818

K. D. HJELMSTAD AND E. TACIROGLU

drawback of this choice is a larger global system of equations and potential ill-conditioning of the system matrix. The second alternative involves elimination of the stress variables at the element level. Provided that Fi is invertible, we can express the incremental generalized stress in terms of the incremental displacement using Equation (30). To wit, Si = Fi1 [rs (u i ; s i ) + BiT Ui ] (34)

This result can be substituted into the rst equation to produce and incremental equations for the displacement alone as Ki Ui + ri = 0 (35)

where the reduced sti ness and residual (of dimension equal to the number of displacement degrees of freedom) are given by Ki Bi Fi1 BiT + Gi ; ri ru (u i ; s i ) + Bi Fi1 rs (u i ; s i ) (36)

As usual, the equations would be assembled into global equations for the incremental displacement. Upon solving the global equations for Ui , the incremental stresses can be computed from Equation (34). Note that the computation of the increment in stress requires either the storage of Fi ; Bi , and rs (u i ; s i ) or their recomputation. Finally, the state is updated in the usual manner as Ui+1 = Ui + Ui and Si+1 = Si + Si . The third alternative was suggested by Simo et al. [14]. The displacement state u i+1 is known upon solving the global equations and updating. Now the functional H can be regarded as a single-eld functional H (s; s)
0

sT (e(s) U(u i+1 )) d

(37)

Using the same procedure of linearization, but viewing only the eld s as a variable, we obtain the incremental equation F Si+1 = rs (s i+1 ) (38)

where the exibility matrix F is the same as before (computed at the stress state s i+1 ), and the revised residual is rs (s)
0

b[ (s) U(u i+1 )] d e

(39)

Equation (38) is solved repeatedly and the stress state is updated in the usual manner as +1 Si+1 = Si+1 + Si+1 . The iteration can start with S0 = Si . Upon convergence to rs tol, i+1 the stress state can be updated as Si+1 = Si+1 . Note that this nal stress state is di erent from the ones generated by the rst two methods. In particular, it exactly forces the residual in Equation (39) to zero. If the function U() is linear, then a single iteration of Equation (38) is equivalent to the previous two methods. For a linearly elastic beam, the stress converges exactly in one iteration. Because the updated stress is di erent from the Newton estimate, this third approach will a ect the convergence of Newtons method. Because the new algorithm is consistent with the original one, the quadratic asymptotic rate of convergence is preserved.
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

MIXED VARIATIONAL METHODS FOR FINITE ELEMENT ANALYSIS

819

Formation and solution of Equation (38) requires repeated computation of F and the new residual rs (s i+1 ). Thus, it is not clear which of the three implementations is the most economical. 4.3. The three-eld (mixed-enhanced) approach Consider the following three eld (HuWashizu type) functional for the BernoulliEuler beam J (u; s; e; u; s; e)
0

[U(u)T s uT p sT (e U(u)) eT (s s(e))] d

(40)

This functional depends upon the three elds u, s and e (and their virtual counterparts) and has the property that if J = 0 for all variations u, s, and e then the state (u; s; e) satises the classical equilibrium, the strain-displacement equations, and the constitutive equations. This functional can be linearized about the state (ui ; si ; ei ) to give L[J ] = Ji + Ji , where Ji = J (ui ; si ; ei ; u; s; e) and Ji d [J (u + d u; s + s; e + e; u; s; e)](u=ui ; s=si ; e=ei ) (41)

is the directional derivative of the functional in the direction of an increment of the two elds. The quantity Ji has the particular expression Ji =
0

[BT (u)(

T i

s + Gi B( u)) sT ( e

i B(

u)) eT ( s Di e)] d

(42)

where i , and Gi are given by Equations (5) and (20), respectively, and Di = @ =@e is the s tangent sti ness evaluated at the state ei . As in the classical sti ness formulation, the terms with B() can be interpolated with the displacement interpolation h given in Equation (21), with the result being the matrix a. The stresses s can be interpolated as s = bT S, as in the previous section. The strains must also be interpolated in this mixed formulation. The idea behind the mixedenhanced formulation is to let the strain be interpolated as e = bT E0 + cT Ee . The strain parameters E0 represent the part of the strain that is interpolated identically to the stresses and Ee are the enhanced strains. The enhanced strain interpolation is organized as cT c1 0 c2 0

cA 0

0 c1

0 c2

0 cB

(43)

Again, although it is not required, we shall interpolate the axial strain and the curvature with the same base functions ci . The number of terms in the axial interpolation is A and the number in the bending interpolation is B (not necessarily the same number as in the stress interpolation). If we substitute our interpolations of displacement, stress, and strain into the linearized 0 functional we get the discrete version of the functional. The condition L[J ] = 0 for all U, S, E ,
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

820
e

K. D. HJELMSTAD AND E. TACIROGLU

and E implies the following set of equations: Gi Ui + Bi Si + ru (ui ; si ; ei ) = 0 BiT Ui P Ei0 C Eie + rs (ui ; si ; ei ) = 0
P Si + Ji11 Ei0 + Ji12 Eie + r0 (ui ; si ; ei ) = 0 CT Si + Ji21 Ei0 + Ji22 Eie + re (ui ; si ; ei ) = 0

(44)

where the matrices P and C are integrals of only interpolation functions and have the expressions P
0

bbT d ;

C
0

bcT d

(45)

The (sti ness) matrices Ji11 ; Ji12 = [Ji21 ]T , Ji22 are given by the expressions Ji11
0

bDi bT d ;

Ji12
0

bDi cT d ;

Ji22
0

cDi cT d

(46)

and the residuals are dened as ru (ui ; si ; ei )


0

[a
0

T i si

hq] d

rs (ui ; si ; ei )
0

b[ei (ui )] d (47) c[si s(ei )] d

r0 (ui ; si ; ei )

b[si s(ei )] d

re (ui ; si ; ei )
0

The base functions bi and ci can be chosen to be orthogonal on [0; ]. If this choice is made (and it will be in the sequel) then the matrix C is identically zero. The same three choices are available for implementing the global solution scheme for the three-eld mixed formulation as were available for the two-eld mixed formulation. First, one can retain all of the parameters Ui , Si , Ei0 , and Eie in the global solution. One would then form an element sti ness and residual for the combined unknowns as 0 0 G i Bi ru (ui ; si ; ei ) T B rs (ui ; si ; ei ) 0 P 0 i (48) Ki ; ri 0 P J11 J12 r0 (ui ; si ; ei ) i i re (ui ; si ; ei ) 0 0 Ji21 Ji22 and assemble a combined global sti ness and residual in the standard manner. The drawback of this choice, again, is a larger global system of equations. Numerical conditioning of those equations is also an issue as the unknowns can have vastly di erent scales. The second alternative is to condense the system to only displacement unknowns. Provided that Ji22 is invertible, the incremental generalized stress and strain can be eliminated
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

MIXED VARIATIONAL METHODS FOR FINITE ELEMENT ANALYSIS

821
i

from Equation (44) in favour of the displacement increment. Dening, the matrix we nd Ei0 =
i T

Bi P1

Ui P1 rs (49)

Eie = [Ji22 ]1 [re Ji21 Ei0 ] Si = P1 [Ji11 Ei0 + Ji12 Eie r0 ]

These results can be substituted into the rst equation to produce and incremental equations for the displacement alone as Ki Ui + ri = 0 (50)

where the reduced sti ness and residual (of dimension equal to the number of displacement degrees of freedom) are given by Ki i Ji
i T

+ Gi ;

ri ru + i [r0 + Ji P1 rs Ji12 [Ji22 ]1 re ]

(51)

where Ji Ji11 Ji12 [Ji22 ]1 Ji21 . As usual, the equations would be assembled into global equations for the incremental displacement. Upon solving the global equations for Ui , the incremental stresses and strains can be computed from Equation (49). Again, the computation of the increment in stress and strain requires either the storage or recomputation of the element matrices and residuals. Finally, the state is updated in the standard fashion. The third alternative is similar to the one described in the previous section. However, the presence of more elds complicates the recovery of the additional elds. This recovery can be done sequentially, somewhat like the sequence implicit in Equation (49). First, we recognize that the displacement state can be updated upon solving the global equations obtained by assembling the element contributions from Equation (50). We can determine the strain 0 parameters Ei+1 by recognizing that the residual rs should go to zero at the new state. To wit, rs =
0 0 e Substituting ei+1 = bT Ei+1 + cT Ei+1 , recognizing the orthogonality of b and c we nd that 0 Ei+1 = P1 0 0 By similar reasoning, noting that we now know Ui+1 and Ei+1 , the equations re = 0 leads us to the local iteration e e [Ei+1 ] +1 = [Ei+1 ] [J22 ]1 0 0 e where ei+1 = bT Ei+1 + cT [Ei+1 ] is the strain associated with the current estimate of the ene e hanced strain Ei+1 . This local Newton iteration can be started with [Ei+1 ]0 = Eie and can continue until the residual (the integral term in Equation (54)) reduces in absolute value to e within an acceptable tolerance. Now the enhanced strain Ei+1 is known and hence the strain

b(ei+1 U(ui+1 )) d = 0

(52)

bU(ui+1 ) d

(53)

c (ei1 ) d s

(54)

Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

822

K. D. HJELMSTAD AND E. TACIROGLU

Table I. Base functions for interpolation.


Minimal displacement interpolation functions Axial g1 = x g2 = 1 x Transverse h1 = 1 3x2 + 2x3 h2 = x2 (3 2x) h3 = x(1 2x + 2x2 ) h4 = x2 (x 1) Legendre polynomials p1 = 1 p2 3(1 2x) = p3 = 5(1 6x + 6x2 ) p4 = 7(1 12x + 30x2 20x3 )

ei+1 is completely known. Finally, recognizing that r0 = 0 must be satised, we can determine the stress parameters from Si+1 = P1
0

b (ei+1 ) d s

(55)

The economies of the third method should be evident. Assuming that no element level information is saved (except, of course, the state variables), then the recovery requires the evaluation of residuals and evaluation and inversion of the matrices P and J22 (repeatedly). Remark Nodeless (bubble) displacement degrees of freedom can appear in any of the three types of beam elements. These unknowns are not bound by any inter-element continuity requirements and therefore need not appear as unknowns in the global equations. They can easily be condensed out and recovered at the element level much like the stress and curvature parameters in the mixed formulations.

5. CHOICE OF INTERPOLATION FUNCTIONS For the classical sti ness method and the mixed methods, the minimal interpolation of the axial displacement eld has two terms and the minimal interpolation of the transverse displacement has four terms. These interpolations are accomplished with the two linear and four cubic Hermitian polynomials in Table I, in which =. This interpolation is necessary because it is the simplest means of enforcing the interelement continuity of the displacement eld. Enhancement of the displacement eld is possible for all formulations. This enhancement can be accomplished using the so-called bubble functions for the axial and transverse base functions. Let g x(1 x); h x2 (1 x)2 (56)

A sequence of enhanced displacement interpolation functions can then be dened as g2+i = g pi ; h4+i = h pi (57)

where the functions pi are the (orthonormal) Legendre polynomials, the rst four of which are given in Table I.
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

MIXED VARIATIONAL METHODS FOR FINITE ELEMENT ANALYSIS

823

For the mixed methods the stress eld is interpolated with the Legendre polynomials, i.e., bi = pi . Because the Legendre polynomials are orthogonal on [0; 1], they are the right choice to interpolate the enhanced curvatures in the mixed-enhanced method. The functions ci must be orthogonal to the functions bi . Hence, the rst enhanced strain interpolation function is the Legendre polynomial following the last one used for stress interpolation. For example, if we use a two term approximation for the stress and a three term approximation for the enhanced strain then the appropriate interpolation matrices are bT p1 0 p2 0 0 p1 0 p2 ; cT p3 0 p4 0 p5 0 0 p3 0 p4 0 p5 (58)

Note that one of the ramications of selecting the orthonormal Legendre polynomials is that the matrix P = I. This choice leads to savings in computation as this matrix must be inverted in both of the mixed methods. In the classical sti ness method there is one choice that needs to be madethe number of base functions to use for the displacement eld (above the minimum required to enforce continuity). We will refer to the classical sti ness methods as CSi with the subscript i indicating the number of transverse displacement interpolation functions used (e.g. CS4 is the usual method used in structural analysis). The method based upon the HellingerReissner type of functional has two choicesthe number of displacement base functions and the number of stress base functions. We will refer to the HellingerReissner type methods as HR ij with the subscript i indicating the number of transverse displacement interpolation functions and the subscript j indicating the number of stress interpolation functions. Similarly, we will refer to the mixed-enhanced methods as MEijk where i, j, and k indicate the number of transverse displacement, stress, and enhanced strain interpolation functions, respectively. Note that the axial and transverse elds can be interpolated with a di erent number of base functions. Hjelmstad and Taciroglu [7] have shown that the best mixed elements have i = j + 2 for the transverse elds and i = j+1 for the axial elds, where i is the number of displacement parameters and j is the number of stress parameters. These elements are all stable and convergent. Throughout the examples selection of the number of axial parameters follows this convention. 6. ELEMENT PERFORMANCE IN BENCHMARK EXAMPLES In this section, we compare the performance of the classical sti ness, HellingerReissner and the mixed-enhanced beam formulations through a set of benchmark problems. Each problem is solved for di erent levels of mesh renement and order of interpolation. GaussLobatto quadrature is used throughout for numerical integration of the nite element matrices. Gauss Lobatto quadrature uses N sampling points to exactly integrate polynomial integrands of degree up to 2N 3. The computational cost of each element (type and order) are reported, in an approximate sense, for a comparison of their performance. 6.1. Transversely loaded non-prismatic cantilever beam A cantilever beam of length = 1 is transversely loaded at its free end with a loading of P = 1, H = 0. The linearly varying depth gives rise to a bending modulus that varies as D( ) = (1 + )3 . The beam and its loading are shown in Figure 3.
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

824

K. D. HJELMSTAD AND E. TACIROGLU

Figure 3. Cantilever beam with a transverse load and variable bending sti ness. Table II. Normalized displacement for the HR elements for the cantilever beam problem.
N 4 5 6 7 8 Exact HR 42 0.910 0.995 1.000 1.000 1.000 1.000 HR 53 2.792 1.041 1.002 1.000 1.000 1.000 HR 64 1.570 1.469 1.055 1.011 1.001 1.000

The number of quadrature points required to fully integrate the nite element matrices and vectors di ers among the three approaches. Because the bending sti ness varies cubically along the length of the beam, the full integration of the element matrices of the CSi elements require N i sampling points with the GaussLobatto quadrature. The ME421 , ME422 , ME431 , ME432 , ME531 , and ME532 elements require, respectively 5, 6, 6, 7, 6, and 7 point quadrature for exact integration of the system matrices for this problem. The elements of the exibility matrix F in the HR elements involve rational fractions. Thus, it is not possible to determine the number of sampling points required for exact integration. Table II shows the normalized tip displacement for HR 42 , HR 53 , and HR 64 elements obtained with di erent orders of quadrature. The values in this table are normalized by the exact tip displacement. The error measure used throughout the rest of this paper is dened as follows. Let us denote the square integral of a function u over the length of the domain as I (u)
0

u2 d

(59)

The error in the eld u is then dened to be Err(u) I (u u)=I (u) (60)

where u is the exact value of the eld and u is the nite element approximation. The error can be computed for any of the computed elds. Figure 4 shows how the numerical errors vary with mesh renement for the cantilever beam problem. The displacement errors are essentially identical across all three approaches when the number of displacement parameters is equal. The two- and three-eld mixed methods
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

MIXED VARIATIONAL METHODS FOR FINITE ELEMENT ANALYSIS

825

100 Err (w)

(a) 10-8
100

(d)

(g)

Err (m)

(b) 10-8
100

(e)

(h)

Err (x)

(c) 10-7
1 5 10

(f)
1 5 10 Number of Elements

(i)
1 5 10

Figure 4. Displacement, moment, and curvature errors for the cantilever beam under mesh renement (a, b, c) Classical Sti ness [ , , +, and for CS4 , CS5 , CS6 and CS7 ], (d, e, f) HellingerReissner [ , , + for HR 42 , HR 53 , and HR 64 ], (g, h, i) mixed-enhanced [ , , +, and for ME421 , ME422 , ME531 and ME532 ].

capture the moment exactly, as expected, while the CS elements do not. The curvature errors are identical for CS and HR elements with the same number of displacement parameters. Evidently, the performance of the ME elements with i displacement parameters and k enhanced strain parameters is identical to the CS and HR elements with i + k displacement parameters.

6.2. Linearized buckling analysis of non-prismatic cantilever beam A second interesting class of problems are the matrix eigenvalue problems that arise from linearized buckling. The non-prismatic cantilever beam of Figure 3, with P = 0, H = illustrates the performance of the three approaches on buckling eigenvalue problems. Table III displays the two lowest buckling eigenvalues obtained for di erent elements for a single element mesh. The CS elements converge from above, as is well-known. The HR and ME elements appear to converge from below, except for known bad elements like ME431 and ME432 [7]. The performance in linearized buckling appears to be dominated by the displacement interpolation indicating that the primary di culty is getting the geometric sti ness right.
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

826

K. D. HJELMSTAD AND E. TACIROGLU

Table III. Approximations of the lowest two of the buckling eigenvalues for the CS, HR and ME elements.
CS4
1 2

CS5 4.6347 67.882 HR 53 4.6005 57.681 ME422 4.4453 151.40 ME532 4.6006 57.759

CS6 4.6128 62.592 HR 64 4.6118 61.868 ME431 4.9973 178.38 ME642 4.6118 61.872

CS7 4.6121 62.592 HR 75 4.6121 62.536 ME432 4.9940 178.27 ME643 4.6118 61.868

CS8 4.6121 62.438 HR 86 4.6121 62.393

CS12 4.6121 62.397

5.0149 178.98 HR 42

1 2

4.4413 151.16 ME421

1 2

4.4944 154.01 ME531

ME752 4.6121 62.537

ME753 4.6121 62.536

1 2

4.6028 58.669

P (t)

(a)

1
^

2
^

200

P
^ w

2.5

5.0 t

0.75

0.75 w

(b) -200

(c)

Figure 5. Cyclically loaded two-span beam example: (a) problem geometry; (b) transverse load versus time; and (c) the transverse displacement under the load versus load.

6.3. Continuous beam with non-linear material behaviour A two-span continuous prismatic beam of length = 4 is transversely loaded at the middle of its rst span with a transverse point load as shown in Figure 5(a). The beam is elasto-plastic and behaves in accord with the model presented in Section 3. The uniform bending and axial elastic sti nesses are EI = 3000 and EA = 3000, respectively. The isotropic hardening modulus for the linear hardening rule is = 0:2. The yield moment and the yield axial force are of equal magnitude and are chosen as m0 = n0 = 20. The continuous beam is discretized with 3 beam nite elements, two elements of unit length for the left span, and a single element for the right span. Figure 5(b) displays the variation of the load with respect to time given
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

MIXED VARIATIONAL METHODS FOR FINITE ELEMENT ANALYSIS


Err (w)
0.015 0.15

827

Err ()
40

Err (m)

CS Elements

-0.015 0.015

-0.15 0.15

-40 40

HR Elements

-0.015 0.015

-0.15 0.15

-40 40

ME Elements

-0.015 0.05

-0.15 0.2

-40 75

Fine Mesh Solution

-0.05

-0.2

-75

Figure 6. Errors in the transverse displacement, curvature, and moment elds in the two-span continuous beam example at the end of the cyclic loading regime along with the ne mesh (exact) solution (.....CS4 , HR 42 , ME421 , - - - CS5 , HR 53 , ME531 , CS7 , HR 64 , ME532 ).

by the formula P(t) = 200 sin(t=4) sin(2t). Figure 5(c) displays the load versus transverse displacement as computed with (three) CS4 beam elements. The response of the beam is computed with di erent types of elements at time t = 3:95 s when the load reaches P(t) = P = 167. The exact solution was obtained using a very ne mesh consisting of 40 CS, HR or ME elements (all gave the same result). For the three element mesh, the displacement, moment and curvature error diagrams are shown in Figure 6.
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

828

K. D. HJELMSTAD AND E. TACIROGLU

Figure 7. Geometrically nonlinear response: (a) geometry and loading; (b) response for 20 CS4 elements; (c) convergence for CS elements [ , , and + for CS4 , CS5 , CS6 ]; (d) convergence for HR elements [ , , + for HR 42 , HR 53 , and HR 64 ]; and (e) convergence for ME elements [ , , +, and for ME421 , ME422 , ME531 and ME532 ].

Again, as this gure indicates, the moment eld is captured almost exactly by even the lowest order mixed elements, but not by the CS elements. Also, the displacement elds captured by all the elements, with the exception of CS4 , have similar accuracy. The complex curvature eld, however, proves di cult for all elements to capture with this coarse mesh. However, the HR and ME elements seem to do better than the CS elements. 6.4. Geometrically nonlinear analysis A prismatic cantilever beam of length = 1 is loaded with transverse and axial loads of equal magnitude at its free end, as shown in the cartoon in Figure 7(a). The axial and the bending sti ness of the beam are EI = 200, EA = 100, respectively. The beams response, obtained by a geometrically non-linear analysis made with 20 CS4 beam nite elements, is shown in Figure 7(b). The results of the mesh renement study for this problem is presented in Figure 7(c)(e). Here, we see that most of the elements perform well, even for the coarse meshes.

7. WHICH ELEMENT IS BEST? From the example problems it should be evident that there is no clear winner among the three di erent formulations. Each has merit in certain circumstances. All converge with mesh
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

MIXED VARIATIONAL METHODS FOR FINITE ELEMENT ANALYSIS

829

Table IV. Floating point operation counts for di erent beam elements.
NQ 10 Min NQ 10 Min NQ 10 Min CS4 7500 2900 HR 42 8900 5400 ME421 12200 5700 CS5 33100 16100 HR 53 45500 32700 ME422 13000 7200 CS6 46900 28100 HR 64 72200 59300 ME531 55400 32600 ME532 58000 39300 CS7 64700 45900

renement. Each has a di erent relative computational cost. In this section we attempt to make an approximate accounting for the computational e ort associated with each approach. We present a computational cost analysis for all the elements based on the error trends we have observed in the examples and the operation counts required to compute the nite element matrices for each element. Assuming that we use 10-point GaussLobatto quadrature, three plastic iterations and, whenever present, three Newton iterations for recovering both the bubble degrees of freedom and the mixed eld variables each, we obtain the ( oating point) operation counts for various formulations as presented in the rst row(s) of Table IV. In the second row(s), the operation counts are computed assuming that the material behavior is linear elastic (i.e. 1 plastic iteration) and the beam element has a cubic variation of bending sti ness. Thus, we use the minimum number of quadrature points to exactly integrate the element matrices. The operation counts presented in Table IV are approximate. Using the operation counts presented in Table IV, we can compare the cost versus performance of di erent element formulations. For example, for a given mesh and element formulation (say, two CS4 elements) we can compute the total error dened as etot Err(w) + Err(m) + Err() and compute the e ort dened as the product of the number of elements in the mesh and the operation count (2 2900 = 5800 for two CS4 elements with minimum order of quadrature). We consistently ignore the additional e ort required for solving a larger system of equations for ner meshes for all types of elements. Using the data in Table IV and the total errors, we can generate a graph containing total error versus the corresponding e ort. Figure 8 contains such graphs generated for the rst benchmark problem which involves the transverse loading of a non-prismatic cantilever beam (see Figure 3). We can see that the smallest total errors are attained by the CS7 and ME532 type elements but at great cost. Among the two types however, CS7 is more costly. On the other hand, while attaining similar total error levels, HR 64 , CS6 and ME531 require respectively less e ort. The trend at higher amounts of total error (and correspondingly lower levels of e ort) is about the same. Considering the approximate nature of the operation counts and the trends of the results in Figure 8, we conclude that, at least for the rst benchmark problem, one cannot clearly identify a preferred element type. In other words, to attain a specic level of accuracy, all
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

830

K. D. HJELMSTAD AND E. TACIROGLU

106

105

104

(a) 103

106 Effort (log)

105

104

(b) 103

106

105

104

(c)

103 10-7

10-5

10-3 Total Error

10-1

Figure 8. E ort versus Total Error for bending of the nonprismatic cantilever beam benchmark problem: (a) CS elements [ , , +, and for CS4 , CS5 , CS6 and CS7 ]; (b) HR elements [ , , + for HR 42 , HR 53 , and HR 64 ]; (c) ME elements [ , , +, and for ME421 , ME422 , ME531 and ME532 ].

element types (CS, HR, or ME) require similar amounts of computational e ort. The apparent advantage of low-order elements (of any variety) with rened meshes will be o -set by the additional cost of solving the global equations (which is not re ected in these costs).

8. CONCLUSIONS The formulations of various methods for geometrically nonlinear planar BernoulliEuler beam nite elements have been presented. In particular, one-eld (CS), two-eld (HR), and threeeld (ME) variational formulations have been formulated and studied. We have examined strategies for condensing and recovering the nodeless variables at the element level and the
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

MIXED VARIATIONAL METHODS FOR FINITE ELEMENT ANALYSIS

831

choice of interpolation (we have conned our attention to convergent and stable mixed methods). Extending these ideas to three dimensional BernoulliEuler beams is straightforward, as is the extension to Timoshenko beams. For Timoshenko beams the mixed formulations have the additional merit of curing shear locking (which is not an issue for BernoulliEuler beams). Simple benchmark problems that are representative of those commonly encountered in the structural analysis of frames have been used to compare the performance and e ciency of each of the di erent formulations. Mixed methods and p-renement for the classical sti ness method clearly provide high coarse-mesh accuracy and, with proper enrichment of the mixed elds the accuracy can easily be increased, though with added computational cost. Similar improvements in accuracy can also be attained with h-renement with a comparable computational cost. The elements with high coarse-mesh accuracy may be attractive for general frame analysis when adaptive mesh renement procedures are not available. In certain circumstances, one of the response elds may be more important to capture than the others. In problems where an accurate resolution of the curvature eld is essential (e.g. inelastic response) ME elements are best. In problems, where an accurate resolution of the moment eld is necessary, HR elements are best. On the other hand, if only an accurate displacement eld is needed (e.g., geometrically nonlinear but materially linear analysis of nonprismatic beams) p-rened CS elements are best. In contrast to other recent papers on mixed methods for beam elements, we conclude that, despite certain di erences in performance, there is no clear advantage over all to mixed formulations over the classical displacement-based methods for BernoulliEuler beams.
ACKNOWLEDGEMENTS

The research reported herein was supported by the Department of Energy through the Center for Simulation of Advanced Rockets. This support is gratefully acknowledged. The opinions expressed in this paper are those of the authors and do not necessarily re ect those of the sponsor.
REFERENCES 1. Argyris JH, Kelsey S. Energy Theorems and Structural Analysis. Butterworths: London, 1960. 2. Ciampi V, Carlesimo L. A nonlinear beam element for seismic analysis of structures. Proceedings of the 8th European Conference on Earthquake Engineering. Laboratorio Nacional de Engenharia Civil, Lisbon, Portugal, 1986. 3. Spacone E, Ciampi V, Filippou FC. Mixed formulation of nonlinear beam nite element. Computers and Structures 1996; 58:7183. 4. Spacone E, Filippou FC, Taucer FF. Fibre beam-column model for non-linear analysis of R=C frames: Formulation. Earthquake Engineering and Structural Dynamics 1996; 25:711725. 5. Neuenhofer A, Filippou FC. Evaluation of nonlinear frame nite elements models. Journal of Structural Engineering, ASCE 1997; 123(7):958 966. 6. Petrangeli M, Ciampi V. Equilibrium based iterative solutions for the non-linear beam problem. International Journal for Numerical Methods in Engineering 1997; 40:423 437. 7. Hjelmstad KD, Taciroglu E. Mixed methods and exibility approaches for nonlinear frame analysis. Journal of Constructional Steel Research 2002; 58:967993. 8. Brezzi F, Fortin M. Mixed and Hybrid Finite Element Methods. Springer-Verlag: New York, 1991. 9. Pian THH, Sumihara K. Rational approach for assumed stress nite elements. International Journal for Numerical Methods in Engineering 1985; 20:1685 1695. 10. Wilson EL, Taylor RL, Doherty WP, Ghaboussi J. Incompatible displacement models. In Numerical and Computer Models in Structural Mechanics, Fenves SJ et al. (eds). Academic Press: New York, 1973. 11. Simo JC, Hughes TJR. On the variational foundations of assumed strain methods. Journal of Applied Mechanics ASME 1986; 53:51 54.
Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

832

K. D. HJELMSTAD AND E. TACIROGLU

12. Simo JC, Kennedy JG, Taylor RL. Complementary mixed nite element formulations for elastoplasticity. Computer Methods in Applied Mechanics and Engineering 1989; 74:177206. 13. Simo JC, Rifai MS. A class of mixed assumed strain methods and the method of incompatible modes. International Journal for Numerical Methods in Engineering 1990; 29:1595 1638. 14. Simo JC, Armero F, Taylor RL. Improved versions of assumed enhanced strain tri-linear elements for 3D nite deformation problems. Computer Methods in Applied Mechanics and Engineering 1993; 110:359 386. 15. Kasper EP, Taylor RL. A mixed-enhanced strain method: Part 1geometrically linear problems. Computers and Structures 2000; 75:237250. 16. Kasper EP, Taylor RL. A mixed-enhanced strain method: Part 2geometrically nonlinear problems. Computers and Structures 2000; 75:251260. 17. Simo JC, Hughes TJR. Computational Inelasticity. Springer-Verlag: New York, 1998. 18. Hjelmstad KD. Fundamentals of Structural Mechanics. Prentice-Hall: Upper Saddle River, NJ, 1997. 19. Washizu K. Variational Methods in Elasticity and Plasticity. Pergamon Press: Oxford, England, 1982. 20. Gelfand IM, Fomin SV. Calculus of Variations. Prentice-Hall: Englewood Cli s, NJ, 1963.

Copyright ? 2003 John Wiley & Sons, Ltd.

Commun. Numer. Meth. Engng 2003; 19:809832

Das könnte Ihnen auch gefallen