Вы находитесь на странице: 1из 81
INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING, Vor, 33, 1413-1449 (1992) GEOMETRICALLY NON-LINEAR ENHANCED STRAIN MIXED METHODS AND THE METHOD OF INCOMPATIBLE MODES* 4.6 SIMO" AND F. ARMERO# Disision of Applied Mechanics, Department of Mechanical Engineering, Stanford University, Stanford, CA 94305-4040, Usa SUMMARY A class of ‘assumed strain’ mixed finite element methods for fully non-linear problems in solid mechanics is presented which, when restricted to geometrically linear problems, encompasses the classical method of !ncompatible modes as a particular case. The method relies crucially on a local multiplicative decomposition cof the deformation gradient into a conforming and an enhanced part, formulated in the context ofa theee-feld variational formulation. The resulting class of mixed methods provides a possible extension to the aon linear regime of well-known incompatible mode formulations, In addition, this class of methods includes ‘non-linear generalizations of recently proposed enhanced strain interpolations for axisymmetric problems which cannot be interpreted as incompatible modes clements, The good performance of the proposed methodology is illustrated in a number of simulations including 2-D, 3-D and axisymmetric nite {deformation problems in elasticity and elastoplastciy. Remarkably, these methods appeat to be specially well suited for problems involving localization of the deformation, as illustrated in several numerics! examples, 1. INTRODUCTION {In Simo and Rifai (Reference 9)], a class of three-field mixed finite element methods is developed which encompasses, as a particular case, the classical method of incompatible modes. ‘The key idea in the infinitesimal theory is to consider assumed strain fields e of the form = Yu a ) Within the framework of a three-field variational formulation. ‘The stress field is then effectively eliminated from the finite element method by enforcing the condition that the stress and an enhanced strain interpolations be L,-orthogonal over a typical element 3, ie. OP eayi= fi o-adv=0 @ {In the present formulation, such a requirement emanates as an Euler-Lagrange equation of the Problem the explicit enforcement of which leads to a two-field mixed method in terms of {Research supported by LLNL under Contract No, 2-DJA-192 and NSF under Grant No. 2-DIA-S44 with Stanford University "Associate Professor of Applied Mechanics ‘Graduate Research Assistant (0029-5981/92/101413-37818.50 © 1992 by John Wiley & Sons, Ltd. 1414 1.6 SIMO AND F ARMERO displacements u and enhanced strains @. Two additional conditions on the admissible interpota- tions for ensure convergence and stability of the mixed finite clement method. Within this framework, the classical method of incompatible modes, in particular, the QM6 quadrilateral of ‘Taylor et al,"? is obtained merely by a particular choice of interpolations of the enhanced strain field . This approach provides a rigorous justification for the usual practice of ignoring the contribution of the incompatible modes to the element load vector. Furthermore, it allows the correct treatment of the axisymmetric case; a problem for which an extension of the QMG6 clement was not known. Finally, it renders the method of incompatible modes a conforming (mixed) method, This paper addresses generalization of the preceding ideas within the context of non-tinear problems in solid mechanics. For the sake of concreteness, we introduce the proposed methodo- logy in the specific setting ofa representative model problem: finite elasticity. We show below that the proper generalization of (I) in a three-field variational formulation relies on an additive decomposition of the assumed displacement gradient which leads to a multiplicative decomposi tion of the assumed deformation gradient fields F of the form i ela Oe 8) ‘The development of the mixed finite element method then proceeds along similar Hines as in the physically non-linear theory. However, in contrast with this latter situation, inthe fully non-linear theory the interpretation as an incompatible modes method is lost. Furthermore, we show below that the actual numerical implementation ofthe proposed mixed formulation takes a remarkably simple form which would be lost if one were to insist on a literal interpretation as a method of incompatible modes. For instance, issues not easily resolved in the context of non-in incompatible mode formutations, such as possible updates of the current geometry of the element with incompatible modes, never arise if the present approach is adopted. Computationally, the method described below is especially attractive in general purpose non-linear inelastic analysis and, in particular, within the context of three dimensional nearly incompressible problems. It is well known that widely usedl methods for the plane problem: notably, the mean dilatation approach of Nagtegaal et al behave poorly in the three dimen: sional problem. Furthermore, optimal interpolation schemes for the pkane quasi-incompressible problem, such as the bi-quadratic clement with linear discontinuous pressures, do not have a direct counterpart for the three dimensional problem. By contrast, modified incompatible mode interpolations of the type proposed in Taylor ct al,"? furnish a convergent finite element method (sce eg. Ciarlet (Reference 2, p. 211) and Strang and Fix.'? p. 176) that exhibits excellent performance in the three dimensional nearly incompressible problem. The approach proposed herein provides one possible extension of this methodology to the fully non-linear regime. which is shown to exhibit good performance in actual numerical simulations. More importantly, the method is readily extended to accommodate complete general inelastic constitutive response: in fact, the structure of the integration algorithms at the quadrature points i identical to that arising in the standard Galerkin (displacement) finite element method. As an example, we briefly describe in the Appendix the implementation of a model of Jz-flow theory at finite strains based on ‘a hyperelastic formulation of the elastic response and a multiplicative decomposition of the deformation gradient.” We close this introduction by noting the good performance exhibited by a class of non-linear enhanced strain quadrilateral elements described below in a number of test problems. The first element, denoted by QI/E4 in what follows, is equivalent to the QM6 formulation of Taylor et al for linear isotropic elasticity (see Simo and Rifai’). The second element, denoted by Q1/ES, is the non-linear version of the axisymmetric element presented in Simo and Rifai.” These two hae his of he he nt ‘d) METHOD OF INCOMPATIBLE MODES 141s elements exhibit good accuracy in coarse meshes both for bending dominated and nearly incompressible problems. In addition, the numerical simulations presented below show an excellent performance of the element in localization problems, which appears to be at least comparable to that of elements specifically designed for this class of problems; such as the method of Ontizetal,® Nacar et al and related approaches, eg. Belytschko et al.! Qualitative arguments which attempt to explain this surprisingly good performance have been advanced in Steinmann and Willam,'* within the framework proposed in Reference 11. In spite ofthese good results, our recent numerical experiments show that spurious (non-physical) modes may arise in calculations with elements QI/E4 and Qi/ES in the presence of very large hydrostatic strains, This issue will bbe addressed in a follow-up publication. 2. MODEL PROBLEM: FINITE ELASTICITY We examine in detail the formulation of the proposed methodology in the concrete setting of non-linear elasticity. As pointed out above, the extension of this approach to accommodate inelastic constitutive response is straightforward and is described in the Appendix. We start our developments by introducing some standard notation and summarizing the usual three-field variational formulation of non-linear elasticity, Subsequently, we introduce the crucial idea underlying the method: a multiplicative decomposition of the deformation gradient into a con- forming part and an enhanced part, and describe the final variational formulation of the problem 2.1. Notation. The standard three-field variational formulation Let 1 & R%™ be the reference configuration of a continuum body with smooth boundary 28, closure B:= BU 0:8, and particles X € 4 labelled by its position vector X relative to the standard basis in Re". Heere, 1 < min < 3 is the spatial dimension of the problem, The deformation of the body from its reference configuration is described by the displacement field u: + R", which is assumed to be prescribed as u = a (given) on a part I, < 28 of the boundary. Further, we assume that the nominal traction vector is prescribed as T: [, > R"~ on a part [, & 098 of the boundary such that e8=T,OF, and Pyat= @ (4) Points xe R™™ in the current configuration of the body are obtained via the differentiable map @: A Fe, defined as x = @{X):= X + u(X) for all Xe, and subject to the local constraint condition det{De}>0 in, where De:= 14 Grade 6 One refers to Dp and Gradu as the deformation gradient and the displacement gradient, respectively. We shall denote by U the space of admissible displacement fields defined as Use (u: B+ Rm: fe WB) and w= a on Ty} © where «'(X) are the components of the displacement field relative to the standard basis in RY", For elasticity, the choice for exponent p > 2 in the Sobolev space [1’(2)}™™ is dictated by suitable growth conditions on the stored energy function; sce e.g. Ciarlet (Reference 3, p. 356). These technical considerations, however, will play no role in the developments that follow. Now let L'!" be the vector space of real maim tue matrices, denote by GLY sm the foup of invertible main x Haim Matrices with positive determinant, and let St" < GL" be the subset of symmetric positive definite matrices. In the developments that follow, we restrict our attention to clastic materials with constitutive response at each point Xe of the body characterized by means of a polyconvex stored energy function W(X,-}. GLY" + R. 1416 1.6 SIMO AND F ARMERO Frame indifference then implies the existence of a reduced stored function, denoted by WAX,-): St —» R, such that W(X, F) = W(X, FTP), forall FeGiy~ a We further assume that the elastic body is subject to dead loading with body force per unit of mass defined by B: # -» R'™, so that the potential energy of the external loading at a configura- tion @:= Identity + u takes the form Tea(u)= — | poB-udY T-uar 8) where po: 9-» R™™ is the reference density, Formally, the weak formulation of the boundary value problem for elasticity can be obtained by means of the following. 2.1. Threesfield variational formulation, Consider a formulation of non-linear elasticity of the Hu-Washizu type in which the displacement field ue, the displacement gradient denoted by H and the nominal stress tensor denoted by P are regarded as independent variables. For the formal developments that follow, it sufices to assume that Hand P are in % Mia Matrices depending on the point X €., with components in L*(#). Thus H and P lic in the space L. given, by Liz (MiB 1%": M4eL(@)} 0) Now define the three-field functional TI: U x Lx LR by the expression Tu, H, P): | (HX, 1+ AYE +H) + PGradu—H)]d¥ + Teal) (10) The stationarity conditions associated with (10) produce the equilibrium equations, identify H as the displacement gradient Grad, and provide an interpretation for P as the nominal stress tensor. Furthermore, the balance of angular momentum condition skew{PFT] =0 is also ‘obtained as an Euler-Lagrange equation. To motivate our subsequent developments, we sum. matize the formal derivation of these stationarity conditions, Let 4 be the space of displacement variations associated with the space of displacement fields U, defined in the standard fashion as Wim (duel W'(A)To: du = 0 on Ty} ay In addition, since both the displacement gradient H and nominal stress P lie in L, defined by (9) and subject, therefore, to no boundary conditions, the space of variations for both Hand P coincides with L. Now assume that (u, H, P)e U x Lx L is a stationary point of (10) character: ized as a zero of the first variation of the functional (10). The formal computation of the corresponding optimality conditions proceeds as follows. Let (bu, 0H, SP)e-d x Lx L be an arbitrary variation, Consider the curve r+(u,,H,, P,JeU x Lx, starting at (u,H, P), and defined by the expression usu edu, Hes H+ eoH, P= P+ coP wy For ¢=0 it follows that the curve (12) is tangent to the variation (bu, 5H, 6P) at the point (o, H, P). The first variation of the functional (10) at (a, H, P) in the direction (6u, 5H, SP), denoted by dl, is then obtained via the directional derivative formula as St1:= DM, H, P)-Gu, 5H, 3P)= 4] MH. PD) (3) alo J METHOD OF INCOMPATIBLE MODES 1417 Since Tea(u) is linear in ue U it follows that DI ,e(u)*6u = [eu(6u). A straightforward computa: tion then yields OM = | [Grad(6u)-P + 5P-(Gradu— Hy + bH-+(—P + 2[1 + H]éc W)] AV — G,,(5u) a4) where G,,(5u):= —Tq.(6u) is, the virtual work of the external loading and OcW’ denotes the partial derivative of 5/(X,C) with respect to the right Cauchy-Green tensor C:= [1 + H]'[1 +H]. Formally, standard arguments in the calculus of variations yield the following local Euler-Lagrange equations: DivP + pB=0) Gradu-H=0} ing (15a) —P 4214+ HJocW=0 along with the natural boundary conditions PN=T onl, (15b) where N denotes the unit outward normal to the boundary 24. By multiplying on the right (15a) by F1, where F:= [1 + H] is the deformation gradient, and noting that 2F@c WF" is symmetric, ‘one concludes that sym[PF"] + 2F@¢WF™=0 and skew[PFT] = 0 (15e) Hence, P coincides with the nominal stress and balance of angular momentum holds at a station- ary point of (10), Equations (15a, b,c) define the strong form of the local boundary value problem for non-linear elastostaties, 2.2, The enhanced displacement gradien With the preceding background in hand, the central idea underlying the proposed methodo. logy can be formulated as follows. We reparamettize the admissible displacement gradients He L in terms of the actual displacement gradient Grad wand an enhanced displacement gradient H1 by setting H= Grady + A) ina Say (16) An alternative interpretation of this reparametrization can be obtained by defining the enhanced spatial displacement gradient f(x) ata point x = @(X):= X +u(X) via the relation F(X) =:h(p(X)De(X), Now let F denote the deformation gradient defined as a function of u and T via the expression 14+ Gradu +A ay) A straightfoward computation then yields the multiplicative decomposition F=[1 + fIDp”'JD9 = [1 + i]De =TDe 18) where fix)= 1 + Bia) can be interpreted as an enhanced deformation gradient superposed onto the standard deformation gradient Dy. We remark that the local multiplicative decomposition aig. J.C. SIMO AND F. ARMERO (18) (or the additive decomposition (16)) makes only local sense and should be interpreted in the same context as, say, the local polar decomposition of the standard deformation gradient; see Figure 1. This local interpretation is crucial for the numerical implementation described below: 2.2.1. The enhanced three-field variational formulation, Consider the class of displacement gradients of the form (16), and let fi, A, be the reparametrized Hu-Washiz functional (10, Further, et 7 be the space of enhanced displacement gradient fields Fi(X) in (16) defined as Fix {i et: HeL@)} (2) Ti(u, Grad u + A, P) to) It follows that the corresponding admissible variations, denoted by 5A, also lie in f. A straight= forward manipulation then shows that the first variation ofthe reparametrized functional (19) can be written as of al) [Grad(6u)-(2F2- 17] ~ 6P-A + 5A (=P + [2FOcW])] dV ~ Gea (6u) = 0 (2la) or equivalently i) f Grad(6u)-[2F2- WAV ~ Gea(Su) = 0 | [ oti-(-p + 2racsipar =o | (21) Ja f op-ftay 0 | s—~\ | AT y ws PND L/ Figure 1. Multiplicative decomporiton of the deformation sradient-F = Dg associated with local neighbourhoods C ‘and 6, of points Ned and x = @(X}e Rm METHOD OF INCOMPATIBLE MODES 1419 where F is.the function of u and fi defined by (17). The corresponding local Euler-Lagrange equations then take the following form: Div [2217] + poB = 0) fi-o skew[PFT] = 0 sym[PF"] — 2Foc HFT = 0 ina along with the stress boundary condition (156) on T, In view of (17), it follows from the preceding Euler-Lagrange equations that forthe continuum problem the enhanced displacement gradient H vanishes in @. However, in the context of a mixed finite element method, this will no longer be the case and (18) will result in an improved representation of the strain field. We remark that the components of the admissible enhanced displacement gradients Hi: "= are requited only to le in L(@). Hence, no intrclement continuity on H need be enforced when constructing finite element approximations. Consequently, the methods constructed below are, strictly speaking, conforming, since they respect the continuity requirements inherent to the variational equations (21) The weak form of the governing equations. ‘The weak form of the governing equations (22) is defined by the variational equations (21). However, in a finite element context, the structure of (21) leads to a matrix problem in which the discrete gradient operator is not sparse, and hence not well suited for large-scale computations. As shown below, an optimal parametrization that takes full advantage of the sparsity of the discrete gradient operator is obtained by a point-wise enforcement of the angular momentum condition skew[PF™] = 0, and the introduction of the change of variables Fie 1+ Gradu +i) t= sym[PET] Se: SPF e f= fir dh: ote? J In addition, for notational convenience, we define Fu, fi): [ARACHUX. FTEVET Ip 01 Goa + (24) Observe that if Fl = 0 relations (23) and (24) define the usual change of variables from the material to the spatial descriptions of finite elasticity. In this way, the tensor field + is identified with the Kirchhoff stress tensor in the spatial strong form of the boundary value problem, With these definitions, the variational equation (21a) can be rewritten, after straightforward manipulations; in the following completely equivalent form: ofl = i (0 (6u)- #(u, hy) — de-h + oh-[—+ + eu, BJ] AV — G,,.(6u) = 0 (25) where the operator @()is defined by V(bu):= Grad(5u)F-* 26) 1420 J.C SIMO AND F. ARMERO Equation (25) is equivalent to the following three variational equations: [ contin} par 6,60 ~0 | foment +#uJav=0 } en orhav=0 | which form the basis of our finite element implementation described below. 2.3, Matrix formulation of the variational equations For our subsequent finite element formulation it proves convenient to recast the variational equations developed above in a matrix/vector notation. We adopt the following convention for the variations du, the symmetric operator sym[V(5u)], the Kirchhoff stress tensor t and the symmetric part sym [5h] of the variations associated with the enhanced displacement gradient (for the sake of concreteness, We take Msi bu = [Sus Suz du]" Gris taa tas Tra ta Tra! = [ihn Bhan Shy Shin + Bia, Bfys + Sign Bhs + Oh." (28) syml symEB(60)] = [5ay,, Bua2 Huss Bure + By, bys + by B,9 + B4y.47" Th + notation will be used for and the variations 6. Note that in (28), we have used the fact that (Gu) can be written, using index notation, as [W(6u)},, = e*6u,, where, in view of (26), Fy fou, 29) and {e,} (i= 1,2,3) denote the standard base vectors in RM", that is, e=[l 0 of [0 1 oy, e=f0 0 17 G0) Remark 1, Observe that the symbol (-),, cannot be understood, in general, as a partial derivative with respect to the co-ordinates X of some configuration defined by an enhanced displacement field, say &(X), since the local definition of F given by (17) will be, in general non-integrable, However, equation (29) is perfectly well defined since itis again a local relation. . | Next, we define the matrisdiferenial operator 6, B va its action on the variations du by the sym ¥(5u)] =:4(0, ju G1 METHOD OF INCOMPATIBLE MODES 1421 In view of (28), and (29), we have the following symbolic matrix representation of the operator bu, by Con BOs Oks Cio reson Os + IO2 FO FTO With this matrix notation in hand, a straightforward manipulation shows that the variational equations (27) can be written in the following form: b10,B)(9):= @2) 7 [, (8. B)50}¥e0.B)a¥— G(00) = 0 Oh [1 + Hu, hyJdv =0 | (33) a where (uf) i evaluated by means of expression (24) Next we turn our attention to the Finite element approximation of the non-linear variational equations summarized above 3, MIXED FINITE E! MENT APPROXIMATIONS In this section we construct mixed finite clement approximations to the three-field variational formulation of non-linear elasticity presented in the preceding section. The development ofthe interpolation for the enhanced displacement gradient Fin the Lagrangian description generalizes, the formulation presented in Simo and Rifai? sithin the context of the geometrically linear theory. Here, we shall restrict our discussion to those aspects relevant to the implementation of the fully non-linear theory, and refer to this latter reference for further details, including the convergence and stability ofthe method As pointed out above, from a computational standpoint, the formulation of the method using equations (33) is the most efficient one, since the linearized strain operator becomes Conceptually, however, it i preferable to formulate the method in the Lagrangian description and then transform the resulting equations using the change of variables (23)-(24). This is the point of view adopted in the presentation that follows. 3.1. The finite element interpolation of the field variables Let JEN B. be a finite element discretization of the closure of the reference con- figuration constructed by means of isoparametric elements on the unit cube C:=[-1,1] x" x[-1, 1], Consider the following finite element interpolations. 1422 1.€ SIMO AND F, ARMERO 3.1.1. Compatible finite element interpolations. Using standard conventions in finite element analysis, define the following approximations to the reference configuration # and the spaces U and xt= NX, Seay, out =" Nt G4) Here, noae is the number of nodes of element #, and Xj, dj, FER" (1 = 1,2,..., Mage) denote the reference nodal position vector, the nodal displacement veetor and the nodal displacement variation vector, respectively. In addition, N’: +R are the standard isoparametric shape functions associated with node I. The interpolations (34);,5 define the standard finite element spaces U" < U and %" < %, respectively. For notational simplicity, the h superscript will often be dropped in our subsequent development, when the context makes it clear that we are referring to discrete finite element quantities. Now let Jo(E) be the nay, X Main Jacobian matrix associated with the isoparametric map §e -+Xte. GradyNU= (NE Jartonams GtadgN'= (NE }a=1 G5) By the chain rule, we have the standard relation Grady N! = [Jo(@)] "Grady! (36) Consequently, the gradient of the deformation pt:= Identity + ué restricted to @, is given by 14° 5" d;@ [Grady] 67) De! Which corresponds to the conforming part of the deformation gradient F* in (17) 4.1.2. Interpolation of the enhanced displacement gradient. Next, we consider the following, discontinuous finite element interpolation for the enhanced displacement gradient Al and its variations 511. Let z.: 4, be the characteristic function of &, given by 1 ifXeg, 4 21001 tere i) Define the finite dimensional subspace f7*< HT by the expression a= fiver i SPai@atforajer| 4 where afeR™ (1 =1,2,..., nyu) are Mgin X rfyn Tocal parameters, and Gl: C) + RY are gyn Vectors of prescribed interpolation functions. Without loss of generality, we assume that the ian Vectors Gi are linearly independent If the displacement gradients FI": "+ 1%" are to enhance the ‘compatible’ displacement gradients obtained via the standard isoparametric map, itis natural to require that the corres ponding finite element subspaces have a null intersection, In fat, the linearized version of this (non-linear) condition plays a erucial role in the discrete stability analysis of the method. Accordingly, we place the following restriction of the interpolation functions G1(&): () Stability condition, Let Grad[U*} be the space of compatible displacement gradients associated with the finite element subspace U* of admissible displacement fields; ie. onaqu'y= fi =F nex ="F aoc} ao) t METHOD OF INCOMPATIBLE MODES 1423 Then, we require that the subspaces spanned by the gradients Grady N': (1 Rr" and the assumed functions G!: J -+ R%™ in (39) have null intersection. Formally, we have the requirement (0) ay — | Grad (U*) This condition ensures stability of the method for the linearized problem, and precludes interpolations for the enhanced strain field which would result in rank-deficient methods; for example, the one-point uniformly reduced integration bilinear isoparametric quadrilat- eral, se Simo and Rifai? for a detailed analysis and illustrative examples Finally, the finite element counterpart of the deformation gradient Fin (23), can be written as S Fz with F,= Dot + Fi, (42) where Dot is given by (37) and A is defined by (39). Note that F* will be discontinuous across element boundaries since Fy is also discontinuous there It remains to characterize the space P* of admissible nominal stress fields. The following two conditions are motivated by the structure ofthe variational equations (21) and the analysis ofthe geometrically linear problem, F (i) Variational consistency. In view of the variational equation (21b), Cor (33)s], we require that the finite clement subspace of admissible nominal stresses, denoted by P* cL, be Lz- orthogonal to the space H of enhanced displacement gradient fields. Restricting our attention to stress fields discontinuous across element boundaries, we define P* as P, ext (43) ref Pha {Pte L: PA tl é where P,: (+L ate prescribed frame-invariant interpolation functions, subject to the constant stress conditions introduced below. The Ly-orthogonality requirement implies the condition PLEGIEAV =0 (44) for all a y raaner (45) (ii) Consta ial stress field condition. As alluded to above, we further require that the space P* of pstant functions, Note that, for the linear theory, this requirement is closely related to satisfaction of the classical patch test (sce Taylor et al.‘*), In view of (45), this requirement is immediately satisfied if imissible nominal stress fields contain, at least, piece-wis Gi@dv=0 U=1, 46) which constitutes the non-linear version of the condition set forth in Taylor et al (Refer- cence 13) [1976] 124 1.€ SIMO AND FARMER Remarks 2. 1. Observe that P* contains the actual stress field P* and the admissible variations 5P* Similarly, both FI’ as well as any admissible variation 6H? are contained in H*. Therefore, the orthogonality condition (44) effectively eliminates the stress field from the variational equations (21) since (21b); is automatically satisfied and the term containing P* in (21b), vanishes as a result of (44). Note that the interpolation vectors G!: C+ RM are defined relative to the reference configuration. Thus, the resulting finite element interpolation automatically preserves the crucial frame invariance property of the enhanced deformation gradient . In view of Remark 2.1 above, explicit knowledge of the interpolations of the nominal stress tensor PA, the Kirchhoff stress tensor 2:= PAP" or its variations és not required in the actual implementation of the method, since these fields drop out ofthe formulation as a result of the orthogonality condition (44), Satisfaction of the constant stress condition is automaticaly ensured by enforcing (46). However, we still have to evaluate ?(u', h"). This is accomplished via the discrete counterpart of constitutive equation (24); namel E(u AY) -= [2PM WX, EMEP) (47) with F* defined by (42) 3.2, The linea 4 strain operators Asa result of the change of variables performed in Section 2.2.1 and in view of (32), the resulting symmetric operator sym[¥(-)] applied to the shape functions NV! becomes foam ] IN! uN, INS + eINY, oe INK FINS heeIN where, by (29), we have {NS} = Fe "Grady N' = [F.Jo(8)] "Grad, N! (49) If we now let 6:= [6' ... b=], the finite element counterpart of relation (31) becomes sym[V(dut)] = btu! (50) where v= (vf... vic] for any variation du. It is apparent from expression (48) that the linearized strain operator B(uf, ft) is sparse and, as pointed out above, computationally conven: ent. By contrast, the linearized strain operator in the Lagrangian description is generally filled. Note further that the enhanced spatial displacement gradient f need not be explicitly computed, since equation (33), is automatically satistied as a result of condition (44) and relation (23). Finally, we express the symmetrized variations sym (6h? ] of the enhanced spatial displacement gradient in a form particularly convenient for our actual finite element implementation. Recall METHOD OF INCOMPATIBLE MODES 1425 that, for the continuum problem, these variations are formally defined by equation (23). Using the matrix notation introduced in (28)s, we consider the nf,, vectors in R=" defined as Tol ue ) 61 and define the linearized matrix operator g!(ut, ht) by the expression: elet | let rs (2) lel + ele! ehel + elat| shai + ef eh) Introducing the matrix notation g:= [g' . .. g"J', simple vector/matrix manipulations show that the finite element expression of sym[ ht] is given by sym[dhe] = g(uS, Rf)9° (33) where 9°T:= (97... 96f.,] and 9} corresponds to the variations of the local parameters af. Remarkably, the expression of the matrix g(ut,f!) has a completely identical structure to the matrix b(uS, i) presented after (48) and, thus, it is also sparse, which makes this formulation very attractive from a computational point of view. ‘These relations along with our preceding results completely define the mixed finite element approximation to the variational equations (33). We now turn our attention to issues concerning the actual implementation of this formulation. 3.3. Implementation: Residual vector and tangent stiffness matrix ‘The mixed finite element approximation to the variational problem formulated in Section Js to the following system of equations: | 8d 86) =O, (6 12.) f Here, A isthe standard finite clement assembly operator, and fis the standard element extemal force vector. In the formulation presented so far, the element vector f" and 5, are piven by B cera...) — 2") : 64) fed aace [sot ee Rye a 65) sede J, stot Rott, Ray where the matrices b(ut, ft) and g(ut, ft) are computed from (48) and (52), respectively, and the Kirchhoff stress #(u}, ht) is obtained from (47). The non-linear system of equations (54) can be solved for the nodal displacement vector d= Al274, and the enhanced strain parameters a, (¢ = 1, ..., Meim) by an iterative solution procedure which involves the solution of a sequence of linearized problems. We outline below the 1426 1.6. SIMO AND F. ARMERO implementation of Newton's method in conjunction with a non-linear version of the classical static condensation procedure. 4.4.1. A solution procedure: Static condensation, Define the enhanced material and spatial elasticity tensor, denoted by €, and &, respectively, by the standard relations 4dbc W(X, Fe'F2) 1 x sanco$ (56) [etm FREPRE RHEE - In a computational setting, the rank four tensors (56) are converted to matrices via a standard convention in finite element analysis. Now introduce the material part of the tangent stiffness matrices: Kae BF (ut, He)e, b(ut, Ready 5 [ ooh Ra gut, Bay 6 [ate Roe.oud, Ray Further, define the geometric stiffness matrices by the expressions ‘| He bain Ke Resa Toa Kees ki = | REESE Dosim Reese Tags, 2 Where Tay, i8 the Unit Mim X7tgin Matrix, and the coefficients ki’yy, kus and Aide are given by f 5 Mla= | NNGNSAV d= 1. ) Mam | alief 0 Ud = toto) ' (59) tig | BERyNZAV (l= bees ten P= bee Med and evaluated with the Kirchhoff stress computed via (47) Finally, define kta + kes ) a= kh +B, (60) Rug = ke + A cal 56) rd iu) METHOD OF INCOMPATIBLE MODES 1427 ‘With the preceding notation in hand, the linearization of the element vectors f," and s, about an intermediate solution (d?", a’), denoted by L[fv"] and L[s,] respectively, takes the form coy oe + kt “yas. ere) =e + tet ad, + A) +} @) Lise] = 9 + (kia Ad, + [it] Aa, ‘where a superscript (k) indicates evaluation at the known intermediate solution (a). In view of (54), the linearized equation (61), can be solved at the element level for Aa, to obtain Aa, = [KS] (st? + KAA) @) Substitution of (62) into (61), yields a modified displacement-like equation which, after assembly, ccan be written in the standard form Kad = RO 6) KY (64) A eee kare) RY = Once the system (63) is solved, the solution is updated by setting dt?! = di? + Ad, and af all + da, 3.4. Example: Enhanced quadrilateral elements Q1/E4 and QUES As an illustrative particular example of the methodology developed above, we consider the formulation of two representative elements. The first clement, labelled Q4/E4, is obtained via 4 strain enhancement ofthe standard bilinear quad QI with 4 extra modes (9 modes in 3-D). For the linear theory the enhancement is such that, for non-distorted meshes, the shear strains are constant and the remaining strains complete linear polynomials. In the linear theory, the result element is equivalent to the QMG6 incompatible modes element of Taylor etal (Ret. 13) [1976} The second element, labelled QUES, is the non-linear version of the axisymmetric enhanced sirain element of Simo and Rifai.” Recall that this element cannot be interpreted as an incompat- ible mode element, even if restricted to the linear theory 3.4.1, Plane and 3-D enhanced quadrilateral QI/E4. Let Jo be the nein > Mai Jacobian matrix of the isoparametric map Be +4 Xte @, evaluated at the centroid of the element 8. Further, let jol8):= det{Jo(G)]_ and jo:= det[Jo] (65) Now consider the following enhanced deformation gradient interpolation, GA): [G? (66) where E(E) is aN Mem X Mean Mati (0 (672) ae 1428 J.C SIMO AND F_ARMERO for two-dimensional quadrilaterals, and as & 0 0 E@:=|0 & 0 (67) o 0 & for three-dimensional bricks The enhanced strain interpolation vectors gt is then obtained via (51), leading to Uee ..- 8] = ip LFOJo] EO) (68) Then, the matrix g(u, fit) is given by (52). Remarks 3 1. The interpolation scheme defined by formula (66) is a generalization to the finite deforma- tion case of the set in Simo and Rifai,” and it can be used for other choices of E(§) different from (67). Alternatively, the preceding particular interpolation can be rephrased as follows. Let N1(@):= $[E} — 1] be Wilson's incompatible shape functions (16) Then, the matrix E can be written for the 3-D case as, E(@)=[GradgN! GradgN? GradgN*]=diaglé, & 1 (69) and with a parallel expression for the 2-D case. Thus, the vectors Gi, can be defined in terms cof ‘modified’ reference gradients according to the following expression: (79) 3. Observe that in the implementation described above (including the alternative approach in Remark 3.2) the classical notion of incompatible modes is completely lost. If one insists on such an interpretation, issues immediately arise, which are not easily solved. For instance, possible updates of the current geometry with the incompatible modes. These issues never arise in the present context. = 3.4.2. Finite deformation axisymmetric element QI/ES. Consider polar axisymmettic co- ordinates {R, Z,©} and {r, 2,6} in the reference and current configurations, respectively. Fur- thermore, let the gradient of the deformation and the deformation gradient be On Ona 0 Fx Fr 0 De=| ean G2 0 | and F=| Fin Fez 0 am 0 0 rR 0 0 Feo where the (3, 3) entry of Dg, which should be 1, has been introduced to account for the fact that the metric tensor is no longer the identity for axisymmetric problems. This modification enables ‘one to operate with the deformation gradient as ifthe metric where the identity. In view of (71) we set to zero the following components of the nf,, local parameter vectors aj aj =0 (12,005 DY 5 2) 7b) 58) at et 9) 9 METHOD OF INCOMPATIBLE MODES 1429 In addition, we define interpolation vectors G! such that (P= 12,00. tay — » 3) in order to keep the structure of F shown in equation (71)2, which arises as a result of the axisymmetric nature of the problem. That is, we reduce the number of independent internal degrees of freedom and, consequently, the corresponding entries in the stiffness matrices (57)-(58) have to be deleted, Remark 4. Note that the formulation is frame indiferent in the sense of correct invariance relative to changes of observer appropriatefor the problem at hand: ie. translations and rotations around the z-axis. . ‘The geometry of the problem is discretized in the R-Z plane (two-dimensional), by means of standard isoparametric elements. We shall denote by 33"? the 2x2 Jacobian matrix of the isoparametric map evaluated at the centre of the element. The corresponding Jacobian for the axisymmetric problem, denoted by Jo, is then given by ef i where o- {i} 4) We now define G,(@) by an expression identical to (66), with Jo now given by (74) above, and E(@) defined by (ay = 3). 0 0 (75) ay doth oo STR) —| ER@)dedy | R@dédn ca — j nR(§) edn [_R@acan”? Finally, the enhanced interpolation vectors gi(E) are obtained from (68), and the matrix g(ut, ft), via formulae (52). 4. NUMERICAL SIMULATIONS. In this section, we demonstrate the accuracy and convergence characteristics ofthe formulations Presented above in the finite deformation regime. In particular, finite elastic and finite deforma- tion elastoplastic problems are considered. For these two classes of problems, the elastic response is governed by a regularized compressible neo-Hookean material model characterized by the stored energy function of the form Wim U(I) + 4ulI-Pe{C] — 3] ™ 1430 1.6 SIMO AND F. ARMERO where U(J) is a convex function of J. In the following examples, U( J) is assumed to be given by U(s):= «EJ? = 1) Ind] (78) Here x and j: correspond to the bulk modulus and shear modulus of the linear theory, respectively. In order to assess the performance of the different formulations in the nearly incompressible limit, the material constants in the elastic examples are selected so that the ratio j= 10* and quasi-incompressible response is effectively obtained, Finally, inthe elastoplastic examples, we consider finite deformation J-flow theory with flow stress denoted by oy, and a saturation isotropic hardening/softening law of the form h(a) = (6ige — oy)L1 ~ exp(~d2)] + Ho (79) where ais the hardening internal variable. A summary of the elastoplastic model at finite strains is given in the Appendix. ‘An outline of the class of problems considered in this section is as follows. In Section 4.1, we demonstrate the superior performance exhibited by the enhanced quad QI/E4 in two-dimen- sional, plane strain, finite deformation problems relative to the standard bilinear displacement) constant pressure clement (ie. the non-linear version of the mean dilatation approach). These results are the non-linear counterpart of those reported in Simo and Rifai.” Section 4.2 demon- strates the excelient performance exhibited by the axisymmetric finite deformation quadrilateral QU/ES discussed in Section 3.4.2, as well as the three-dimensional version of the enhanced quad (QI/E4 of Section 34.1, using the necking of a circular bar as a test problem. Finally, in Section 4.3, the performance of QI/E4 in problems involving localization of plastic deformation is examined in detail. Throughout these simulations, the present formulation is compared with the finite deformation version of the mean ditatation element (Q1/PO) developed in Simo et ul.” 4.1, Plane strain 2-D analysis 4.1.1. Cook's membrane problem, Consider a tapered panel clamped on one end and subjected to a shearing load on the opposite end. The linear elastic version of this simulation is often referred to as the ‘Cook's membrane problem’, and constitutes standard test for bending dominated response, Here, we shall be concerned with both finite deformation elasticity and finite deformation elastoplasticity. The material properties are summarized in Table 1. The simulation is performed using load control with steps of AF = 1-0 and F ranging from 0 to 100 in the elastic case, and AF = 05 with F from 0 to 5 in the elastoplastie case. Table |. Material properties used in the numerical simulations for both. quasi-incompressible finite clastcity and finite deformation plasticity Quasiincompressible Finite deformation finite elasticity plasticity x = 400942 x 10" = 16421 p= 80-1938 80-1938, o4s o7is 1693 o12924 METHOD OF INCOMPATIBLE MODES 1431 Figure 2 shows both the reference and deformed configurations for the elastic problem, corresponding to a finite element mesh consisting of 1024 elements. In Figures 3a)-(b), the vertical displacement of the top edge node is plotted versus the number of elements per side of the different meshes considered for the enhanced quad QU/E4 and the mean dilatation element Q1/PO. Although the two formulations converge to the same result, in both the n pressible elastic and elastoplastc finite deformation regimes, the enhanced quad exhibits much better accuracy in coarse meshes than the QI/PO element. This superior accuracy is specially evident in the elastoplastic case. As a side remark, we observe that the formulation presented in Section 2 (reflected in the finite elements equations of Section 3 which involve sparse matrices) does not appear to increase significantly the computational time required over the mixed formulation. Finally, as expected, the exact linearization of the algorithmic problem results in an asymptotic quadratic rate of convergence when Newton's method is employed. This is demon strated in Table IT where the maximum (Euclidean) norm of the residual for a typical load increment is summarized for both finite deformation elasticity and elastoplasticity. We remark that the solution strategy in the elastoplastic simulations combines Newton's method with a linear line search rly incom: 4.1.2. Double notched specimen. This classical test problem was first introduced in Nagtegaal et al? to demonstrate the behaviour exhibited by the displacement formulation in highly constrained plane strain elastoplastic problems. If perfect plasticity is assumed and attention is, restricted to the linear theory, the specimen exhibits a limit load. Here, we consider the geometrically non-linear version of this problem and demonstrate the superior performance of the present formulation over the displacement model in the non-linear regime. It should be noted, however, that in this case no analytical estimation of the limit load is available, ‘The specimen considered has total width W = 10, height L = 30 ligament thickness b = 1, and plane strain conditions are assumed, A $x 15 mesh is used to model a quarter of the specimen, using symmetry boundary conditions. The clastic constants (x andj), as well as the yield limit (cy), considered for this example are the same as in the previous problem, and are shown in Table 1. As in the infinitesimal theory, perfectly plastic response is assumed Figure 2. Cook’s membrane problem. Quasbincompressile finite deformation elasticity: a inital configuration (22x32 ‘clement mesh () final deformed configuration (QU/ES) 1432 4.C.SIMO AND F. ARMERO. 4 ayes — ey aym-- | Pas Ww as a as ao laments per side ae ' qu/Po~ ti ee Figure 3. Cook's membrane problem. Convergence of the finite element solutions: () finite deformation quasiineom resible elasticity (b) Anite deformation lastopastiity (Jy-flow theory) Table Il. Cook's membrane problem. Evolution of Euclidean norm of residual within a typical load increment Quasi-incompressible inite deformation finite elasticity plasticity 3:5078038E — 01 12303137E — 01 5:8340487E + O1 1-157S406E + 00 29872313E — 03 5808280SE — 01 734194395 — 05 152004248 — 01 25837469E — 09 24100404E — 02 27865502 — 04 49735087E — 08, 12661053E — 12 METHOD OF INCOMPATIBLE MODES 1433 The simulation is performed using displacement control, with the displacement field on the top boundary prescribed in increments of 001. Figure 4 shows the deformed configuration of the full specimen corresponding to a top displacement of 0-25. The load-deflection curves for both the enhanced quad QI/E4 and displacement clement QI are shown in Figure 5. It is apparent from these results that the displacement formulation does not exhibit a limit load, in sharp contrast with the response exhibited by the present approach. Furthermore, owing to the geometrical non-linear effects, the load-deflection curve for the enhanced quad eventually decreases, while the displacement model produces a monotonically increasing curve. 4.2. Axisymmetric and 3-D analysis: Necking of a bar We consider the problem of the necking of a circular bar, with a radius of 6413 and length of 53334, in the tension test. A small geometric imperfection (0982 per cent of the radius) is introduced at the centre of the bar and linearly extended to the top in order to trigger the necking. ‘The material is assumed to be clastoplastic, modelled by J,-flow theory with the saturation isotropic hardening expressed in equation (79), and material properties summarized in Table I ‘The simulation is performed using displacement control both with the axisymmetric quad QI/ES developed in Section 3.4.2 and the three-dimensional version of QI/E4 formulated in Section a tH oy ° sit 0 obs O10 a1 020 025 a0 Top daplocement Figure 5. Double notched specimen, Load-digplacement curves for both the enhanced and displacement formations 1434 J.€ SIMO AND F. ARMERO 4.2.1. Axisymmetric analysis. A quarter of the bar is analysed using the appropriate symmetry boundary conditions at the boundaries. The simulation of the necking in the bar is performed ‘with the axisymmetric enhanced strain element of Section 3.4.2 and the axisymmetric version of the Q1/PO clement. Two different classes of finite element meshes are employed in this simulation, leading to remarkably different results (i) First, we consider four different meshes consisting of 50, 200, 800 and 3200 elements, with 5, 10, 20 and 40 elements through the radius of the bar, respectively. The reference mesh correspond- ing to 200 elements, with 10 elements through the bar radius, is shown in Figure 6(a). This type of finite element mesh consists of two blocks, each of them with uniform vertical subdivisions. The results of the simulation are reported at two levels of deformation: (ia) A level of deformation corresponding to a top displacement of 56, obtained in 10 equal displacement increments. Figure 6(b) shows the well developed necking pattern obtained with the Q1/PO element. Similar results are also obtained with the enhanced quad Q1/ES. A comparison of these results is shown in Figure 7, where the computed total load is plotted versus the number of elements per bar radius. While both formulation converge to the same value, the enhanced element exhibits substantially better accuracy in coarse meshes. (ib) A level of deformation corresponding to a top displacement of 7, obtained in 10 more displacement increments. Figure 6(c) shows the deformed configuration computed with the QU/PO formulation. It is apparent for this result that this type of mesh is not adequate for capturing the high deformation gradients associated with the severe necking pattern. This motivates the use ofa different finite element mesh better suited for this very high levels of localized deformation, as discussed below. i) Second, to improve on the computed necking patterns at very high levels of deformation, we consider the type of finite element mesh shown in Figure 8, AS in the mesh shown in Figure 6{a), here we also consider two different blocks, but with non-uniform vertical spacing; see the detail in Figure 8, The remarkably different results obtained with the Q1/P0 formulation and the enhanced quad are displayed in Figure 9. While the necking pattern at both moderate and very high deformation levels is very well resolved by the enlranced quad, the Q1/PO element exhibits poor performance. Because of the incompressibility constraint introduced by the plastic flow, the Figure 6. Axisymmetrie necking problem: a) intial configuration (200 element mesh (b deformed configuration at an Imposed top displacement of $6; () and 7.0 (Q1/PO) try ned of 15, nd of he tal of METHOD OF INCOMPATIBLE MODES 1435 Figure 7. Axisymmetsic necking problem. Convergence ofthe finite element solutions Figure 8 Asymmetric necking problem. Initial configuration elements at the necked section exhibit the characteristic hourglass pattern that renders the simulation with the QI/PO element useless, particularly at high levels of deformation, We remark that the instability of the QI/PO element shown in Figure 9 appears in the axisymmetric problem, but does not arise in the plane strain problem. Further rescarch is needed to determine the precise origins of this phenomenon which could be attributed to 2 failure of the BB condition. Numerical experiments indicate that a smoothing of the loop strains tends to significantly alliviate the problem in the axisymmetric case, Figure 10 shows the load-displacement curve obtained in this ease for both formulations, 4.2.2, Three-imensional analysis, The same problem considered in the previous section, ic. the necking of a circular bar in the tension test, is now studied using a full 3-D analysis. The ‘geometry and material properties are as described in Section 4.2.1. Once more, only one eighth of the bar is discretized using appropriate symmetry boundary conditions. Motivated by the same considerations presented in the previous section, we consider a finite element mesh consisting of 960 4-node bricks with the same disposition of elements along the bar as the mesh in Figure 8 This mesh is shown in Figure 11. In what follows, the 3-D enhanced brick element QI/E4 and the Q1/PO mixed formulation are compared. x 1436 4.6. SIMO AND F, ARMERO ave eure 10 Figure 10. Axisymmetie necking problem. Load-displacement curves for both the enhanced and mixed formulations (Mesh in Figure 8) ‘The simulation is carried out using displacement control, Newton's method, and a line search algorithm during the iteration process. The same number of displacement increments as in the axisymmetric simulation are considered. The resulting load-displacement curves are shown in Figure 12, and the deformed configurations corresponding to two different stages of deformation METHOD OF INCOMPATIBLE MODES 1437 Figure 12. 3D necking problem. Load-dixplacement curves for both the enhanced and mixed formulations are depicted in Figures 13 and 14, We observe the same results as in the axisymmetric simulations: while the enhanced brick reproduces accurately the necking process, the QI/PO clement results again in an hourglass mode. Remarkably, the load-displacement curve for the enhanced brick (Figure 12) coincides exactly with its counterpart in the axisymmetric case (Figure 10), while the curves for the mixed formulation clearly differ. In this latter case, the appearance of the hourglass pattern makes the result useless. 4.3. Localization problems This section illustrates the excellent performance exhibited by the enhanced quad QI/E4 in problems involving localization of plastic deformation which, for the case considered below (Jz-flow theory), lead to the formation of shear bands. The results presented by Steinmann and Willam?® in the context of the infinitesimal theory suggest that the QMG6 type of interpolation is very appropriate for capturing localized states of deformation. The results presented below in the finite deformation regime for Ql/E4 are in agreement with this good performance. Two conctete 1438 1.€ SIMO AND F. ARMERO Figure 13. 3:D nocking problem. Deformed configurations at two dierent stages: a) = $6; (b) = 7-0 (longitudinal sew of half bar) examples are considered: first, an unconstrained single element test and, second, a plane strain localization problem. To trigger the formation of the shear bands, we consider the Jp-flow theory described in the appendix with a saturation hardening/softening law governed by equation (79). In particular, the choice of parameters is identical to that summarized in Column 2 of Table 1, with the exception that the linear hardening modulus is now negative with value H = —012924%10~!. This hardening/softening law results in the yield limit/equivalent plastic strain curve shown in Figure 15, Itis well known that for this type of problem standard finite element formulations produce non-objective results in the sense that the solution depends on the Iével of discretization. This issue, however, is of no concern in the investigation that follows, since its goal is to assess the ability of enhanced strain interpolation in the QI/E4 element to capture the formation and developments of localization bands. Several devices, like the introduction in the constitutive relations of a characteristic length associated with the mesh size, have been proposed to alleviate this lack of objectivity of the finite element solutions. For our subsequent developments it suffices to remark that all the comparisons are made relative to a single finite element mesh. 4.3.1, Unconstrained single element test. This example is proposed by Steinmann and Willam'? as a means of assessing the performance of a finite element method in the presence of localization of plastic deformation. A single square element with diagonals of length 5, in the METHOD OF INCOMPATIBLE MODES 1439 tions at two different stages: a) = $6 (0) = 7.0 (cross sectional Figure 15, Localization problems. Saturation hardening. configuration shown in Figure 16, is subjected to @ uniform shear before localization takes place The simulation is performed under plane strain conditions. We note that a similar test, but with more constrained boundary conditions (all four supports fixed except the imposed horizontal displacement at the top one), has been proposed in Ortiz et al.® for the infinitesimal case, However, in the geometrically non-linear regime such a test results ina locking response because of non-linear effects in all the elements tested. 1440 LC SIMO AND F. ARMERO Figure 16 Unconstrained single clement test. Mesh and final deformed configurations (a) QU/E4:(b) QU/PO. (No tagnification of the detormation) Figure 17. Unconstrained single clement test. Load P vs. imposed displacement ‘The load-displacement curves obtained for the enhanced quad Q4/E4, the mixed method Q1/PO and displacement formulations are shown in Figure 17. As expected, the displacement formulation gives an inaccurate solution: the curve does not even present a softening branch, The QI/PO clement does reflect softening, but its response appears to be too stifl. On the other hand, the enhanced quad QI/E4 easily captures the localized state and permits its development, as reflected in a steeper softening branch, Note that a steeper softening branch yields a lower value of the final dissipated energy and, therefore, results in a narrower shear band. We remark that no bifurcation points are encountered in the course of the present simulation. This feature is present in the infinitesimal case where a perturbation process is required to switch to the bifurcated path. Additionally, we remark that in the finite deformation case the two lowest eigenvalues of the 3 x 3 METHOD OF INCOMPATIBLE MODES 14a stiffness matrix inerease along the softening branch, while the largest one decreases. The combina- tion of these two effets leads to the softening response, but no bifurcation point appears in the problem. 43.2. Plane strain localization problem. This final example is concerned with the plane strain tensile test of a rectangular bar, and has been studied by a number of authors, in particular, ‘Tvergaard et al.'* In a fist stage of the deformation process, the specimen is maintained in an essentially homogeneous state of deformation. Subsequently, while the load reaches a maximum value, a diffuse necking starts to develop. Finally, in a further stage, the deformation mode changes to a pattern involving shear bands forming 45° with the axial direction of loading, The plastic deformation localizes along these bands and, eventually, produces the failure of the specimen, ‘The specimen considered in this study possesses a width of 12-826 and a length of 53-334 and is subject to ideal plain strain loading conditions; hence, the third dimension is assumed to be infinite. We consider an initial geometric imperfection in the form of a linear reduction of the Width from its initial value at the top to 0982 per cent of this value at the centre ofthe specimen, A quarter of the specimen is discretized, imposing symmetry boundary conditions along the boundaries. The 200 element mesh is identical to the one used for the necking ofthe circular bar in Figure 6(a). The simulation is again performed with displacement control. Figure 18 shows the load-displacement curves for the three different formulations. As expected, the result with the enhanced quad Q1/E4 lead to a softer descending branch, since the deforma tion process is localized inside the shear bands. The kinks in the curve of the QI/E4 element should be noted; they are associated with the highly non-homogeneous state of deformation evident in Figure 19. On the other hand, the curve associated with the Q1/PO element is smooth, which is a reflection of the deformation pattern associated to the diffusive necking. Figure 19 depicts the deformed configurations for both QU/E4 and QI/PO at two different stages. The first stage corresponds to a state shortly after the formation of the shear bands (observe that a small amount of diffusive necking has already occurred). In the second stage, on the other hand, the bands are fully developed, at least for the enhanced formulation. Figure 20 shows the distribution ofthe equivalent plastic strain « for both cases. We observe a well localized tt] ues a Top diplacenent Figure 18. Plane strain localization problem. Load-diplacement curves for the enhanced, mised and displacement formulations 1442 1.€ SIMO AND F. ARMERO Figure 19. Plane strain focatzation problem, Deformed configurations at two different stages: () = 39; (0) d= 50 pattern for Q1/E4, even from the formation of the band, which should be contrasted with solution obtained with QU/PO in which the plastic deformation extends throughout all the diffuse neck. Clearly, the enhanced element better captures the formation and development of the shear bands, while in the mixed element solution, the main mechanism involved in the deformation process is diffusive necking, The saturation hardening/softening law shown in Figure 15 and employed in the preceding example allows the development of true finite deformations prior to localization, and thus provides a good test problem for the proposed nonlinear finite element method. Next, we examine the results obtained with a steeper hardening/softening kaw, the double parabola depicted in Figure 21, which is analogous to the law considered by Ortiz, Leroy and Needleman.” Steep softening laws of this type are often used in the numerical simulation of localization problems. Figure 22 shows the load/displacement curves obtained for the double-parabola harden ing/softening law. When the present formulation is employed, the curve exhibits the sudden drop which is typically found in localization problems of this type. By contrast, the Q1/PO clement does not exhibit a comparable performance, as expected. These results are confirmed by the deforma- tion patterns shown in Figure 23. The QI/E4 element method exhibits a very well defined line of continuity in the specimen at 45° with the load direction while the Q1/P0 element shows a very diffuse localization pattern. Figure 23b depicts the distribution of the equivalent plastic strain for both formulations at the same imposed top displacement. Once more, the clear shear band METHOD OF INCOMPATIBLE MODES 1443 Figure 20, Pane strain localization problem. Distribution of the equivalent plastic strtin at two diferent stapes a= 39, 0)a=50 os = ou 03 | oz Donte paabota— 4 aaa Equivalent plate tenin Figure 21. Plan strain localization problem, Alternative hardening/softening law (Ortiz, Leroy and Needleman?) exhibited by the Q1/E4 formulation should be contrasted with the diffuse shear pattern exhibited by the mixed element. ‘The preceding results appear to confirm the superior performance of the enhanced formulation, in problems involving the formation of localization bands, and they agree with the ones exhibit in the linear case. 1444 J.C SIMO AND FARMERO aye Qy/P0— Dish wove == ‘Top displacement Figure 22. Plane stain localization problem (Double-parabols herdening/softening law), Loaddisplacement curve use ayro Figure 23, Plane strain localization problem (Double parabola coltening law) Results tan imposed top displacement of {= 4.24 a) Deformed configuration, b) Distribution ofthe equivalent plastic sirain 5. CONCLUDING REMARKS We have presented a class of mixed assumed strain methods for non-linear problems in solid mechanics which, when restricted to the linearized theory, encompasses the classical method of i incompatible modes as a particuar case. As a specific example of a member of this class of METHOD OF INCOMPATIBLE MODES 144s methods, we have considered the formulation and implementation of an enhanced strain interpo- lation with the following noteworthy characteristics: 1. The method can incorporate completely general inelastic constitutive equations within the standard (displacement-like) format of strain driven integration algorithms. As an example, the actual implementation of a model of finite deformation plasticity is summarized in an Appendix and extensively illustrated in a number of numerical simulations. 2. For plane strain and three-dimensional problems the enhanced strain interpolation in the QU/BA element generalizes to the finite deformation regime the QM6 incompatible mode clement, We remark that for finite deformations the methodology can no longer be interpreted as an incompatible mode method. Nevertheless, its numerical implementation is straight forward. 3. For axisymmetric problems, the enhanced strain interpolation in the QI/ES element generalizes tothe finite deformation regime of our previous workin the infinitesimal theory. We remark that the resulting interpolation cannot be interpreted as an incompatible mode method, even in the linearized theory. The good performance exhibited by the enhanced strain quadrilaterals has been illustrated in 4 comprehensive number of 2-D/3-D/axisymmetric numerical simulations, both for non-linear elasticity and finite strain plasticity, which include nearly incompressible response, bending dominated problems, necking bifurcation and strain localization problems. Throughout these numerical tests, the performance of this method has been compared with the widely used Q1/PO (mean dilatation) element This element exhibits « strong instability in three dimensional and axisymmetric finite defor- mation problems, but notin plane strain problems, when meshes with very high aspect ratios are employed. Although this unstable behavior is strongly reminiscent of instabilities arising from failure of the BB condition, the ultimate reasons are presently not understood. By contrast, the specific enhanced elements described in Section 3.4 did not exhibit this type on instability in any of the simulations presented in Section 4.2 ‘The numerical simulations presented also demonstrate the excellent performance exhibited by the enhanced strain method in strain localization problems. By contrast with specialized formula- tions specifically designed to capture localized deformation, no special device is introduced in the present approach. It is somewhat surprising that an interpolation, originally designed to improve con the bending performance of the standard QI element, also works remarkably well both in strain localization problems. While the reasons for this behaviour in the nearly incompressible problem are presently understood, no comparable analysis currently exists for localization problems. The excellent performance exhibited by Qt/E4 in localization problems is confirmed by our recent results with G, Mestke, obtained in the more general setting of coupled thermomechanical problems in soils. Very recent results also indicate that Q4/E4 may exhibit undesirable modes in the presence of faily high strains, specially inthe compressive regime. These eases are presently uunder investigation. In summary, we believe that the framework described in this paper offers an attractive methodology for the systematic development for enhanced nonlinear finite element methods that improve upon the standard displacement approach in the nonlinear regime. ACKNOWLEDOEMENTS ‘We are indebted to M. S. Rifai for his involvement in the initial phases of this work. Support for this research was provided by LLNL under Contract No. 2-DJA~792 with Stanford University, 1446 1.6. SIMO AND F. ARMERO Box I, Return mapping algorithm for Jp-flow theory with isotropic hardening 1. Deformation gradient and Jacobian (at a quadrature point in clement e) Fur 14 5) )@ [Grady] +S af @ GE Jes = det[Fae 2. Compute elastic predictor pedev B54] 3. Check for plastic loading Esti — JAloy + Maa) IF Qt! <0 THEN Set (Yar = CMP ELSE 4, Perform return mapping algorithm Set Tevs= AuTbe1) f= ae t n= s/s Solve for Ay (Newton's method): OAH — 2jidy — J3tH, + V/ Ha,)] = 0 | Return map: ENDIF 5. Addition of the elastic hydrostatic pressure Pressure: Kira ero Sener 6. Update of intermediate configuration METHOD OF INCOMPATIBLE MODES 1447 and NSF under Grant No, 2-DIA2-544 with Stanford University. This support is gratefully acknowledged, APPENDIX: FINITE DEFORMATION PLASTICITY ‘This Appendix summarizes the implementation of the model of finite deformation J,-flow theory used in the numerical simulations of Section 4. Further details and background are given in ‘Simo? and Simo and Miche.* 1. General model of Jz-flow theory Assume hyperelastic isotropic response relative to a local intermediate configuration defined by the multiplicative decomposition F = F*F?. This isotropy assumption dietates the following form for the stored energy function, w (b*), where bt:= PART = FOF (0) and Cr:= FPF, As a result, the (spatial) Kirchhoff stress «is given by the constitutive equation = 2d, WAWIDY = 2b a, 1719) @ Classical J,-flow theory is characterized by a Mises yield condition formulated in terms of the Kirchhoff stress tensor. If, for simplicity, attention is restricted to non-linear isotropic hardening, the yield criterion takes the form (0, 2):= Idov[t}] — /Floy + Mad] <0 (82) where « > Os the equivalent plastic strain, h(a) is a hardening function with h(0) = O and deve] denotes the deviatoric part of the Kirchhoff stress. The model is completed by assuming. ‘associative plastic flow or, equivalently, by postulating maximum plastic dissipation. This classical assumption uniquely defines the flow rule as eStore 2 evEe/Idev 1) and 720, $(t,a) <0, 7h(u a) =0 | Consistent with the in erpretation assigned to the hardening variable 2, the hardening law becomes & = Fy 2. Return mapping a Following standaed treatments of metal plasticity, assume isochorie plastic flow and uncoupled volumetrie/deviatorie response. Thus, denoting by B*:— J-2B* the volume preserving part of b, we set Wb ):= U(J) + $ulelb] — 3] > r= [JU] + pdev [BJ (84) where j > is interpreted as the shear modulus, and U(J)is a function describing the volumetric response. Combining (84) with the general flow rule (83), and neglecting a term of the order oy/jt (which for metal plasticity is ofthe order = 10~?-10"*), one obtains the following simplified flow 1448. J.€ SIMO AND F. ARMERO Box 2, Algorithmic moduli for the return mapping algorithm in Box 1 1. Spatial elasticity tensor C, for hyperelastic model (77): =(JUYIT@I—2WUI14 E Pf — 1@1) - Hsh[n@1 +1981 é s = ndev[b*], sist f= whee] 2, Scaling factors see Be -[ -z/3 isi aot hs =f 1 py seul | t-[z | A 3. Consistent (algorithmic) moduli cit — p C24 — 2Apyn@n— pala @ devtn]} rule PE tC IRE —y3utbeIn (3) ‘The resulting model can now be integrated over a typical time step [tq ta+1] using a backward Euler difference scheme leading to the closed-form return mapping algorithm summarized in Box 1 Remarks 6. 1. In the absence of plastic flow the algorithm in Box 1 is exact. Thus, the elastic predictor is exact and is computed by mere function evaluation, The so-called incrementally objective algorithms play no role in this development. 2. The plastic corrector is closed-form for the flow rule in (85). The complete return mapping algorithm in Box 1 can be linearized exactly, leading to the closed-form expression for the algorithmic tangent moduli suramarized in Box 2, 3. In the implementation summarized in Box 1, the internal variables {C?,2} are stored in METHOD OF INCOMPATIBLE MODES 1449 aan element data base during the iteration process, Note that the rotation part of F? plays no role in the algorithm, a result to be expected because of the isotropy assumption on the clastic response. REFERENCES ‘, Belvtschko, J, Fish and B. Engelman, ‘A finite element with embedded localization zones’ Comp. Methods Appl Mech. Eng, 70, 9-89 (1988). 2. P.G. Clarke, The Finite Element Method for Eligtic Problems, North-Holland, Amsterdam, 1978 3. P.G.Ciarlet, Mathematical Elasticity. Vl I: Thee-Dimensional Elasticity, North-Holland, Amsleedam, 1988, 4 A. Nacar, A. Needleman and M. Ortiz,’ finite element method for analyzing localization in Fate Sependent solids st Finite strains, Comp. Methods App. ech, Eng, 13, 235-258 (1989) 5... Nagtewnal, D. M. Parks and J. Rice, "On numerically accurate finite element solutions in the fully plastic range’, Comp. Biethods Appl. Mech, Eng. 4153-177 (1970, 6M. Ortiz, 1 Leroy and A. Needleman, ‘A finite element method for localized failure analysis, Comp. Methads Appl Mech Eng, 61, 189-214 (1988) 7. 1.C.Simo,"A framework fr finite strain elastoplastcty based on maximum plastic dissipation and the mahilicative decomposition: Parts | and If, Comp. Methods Appl. Mech, Eng, 6 199-219, 68, 1-3 (1988) 8 .C.Simo and C. Miche, “Associative theemoplasticty a rite strains: Formulation and Numerical Analysis, "Comp Meth. Appl. Mech. Engng, In pres. 9. J.C. Simo and M.S. Rift,’A class of assumed strain methods and the method of incompatible modes, nj. muner methods eng. 29, 1595-1638 (1990), 10.4°C. Simo, R. L. Taylor and KS. Piste, ‘Variational end projction methods for the volume consteaint i finite deformation elastoplasticity’, Comp, Methods Appl. Mech. Eng, Sly 177-208 (1988) 1. P. Scinmann and K.J. Wiliam, ‘Finite elements for capturing loclied flute, Ind. Archiv, to appeas. 1. G_ Strang and G. Fix, An Analysis ofthe Pinte Element Method, Prentice-Hall, Englewood- Cis, New etsy, 1973. 1B. REL Taylor PJ. Bewesford and E.L. Wilson, °A non-conforming element forstess analysis In} numer: methods eng. 10, 121-1219 (1976) 14 RL Taylog,1.C. Simo, O. C.Zienkiewiez and A.C. Chan, The patch test; A condition for esesing finite element jonvergence Inj numer. metho eng. 2, 39-62 (1986), 18. V-Tverguard, A’ Needleman and K. K-Lo, Fiow localization inthe plane stein ese test’ J. Mech Phys. Sel, 29, 115-142 (198. 16, EL. Wion, R. L Taylor, W. P. Doherty and J. Ghabouss ‘Incompatible displacement models in $.J. Fenves ea (ds Numerical and Computer Models b Structural Mechanics, Academic Press, New York, 1973 INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING, VOL. 29, 1595-1638 (1990) A CLASS OF MIXED ASSUMED STRAIN METHODS AND THE METHOD OF INCOMPATIBLE MODES* ‘0 wrongs do make a right in California’ o.srrane(1973) ‘wo rights make a right even in California’ x taxon (1989) Dicision of Applied Mechanics, Department of Mechanical Englnering, Stanford Universi SUMMARY [A three-field mixed formulation in terms of displacements, stresses and an enhanced strain field is presented which encompasses, as a particular case, the classical method of incompatible modes. Within this frame- work, incompatible elements arise as particular ‘compatible’ mixed approximations of the enhanced strain field. The conditions that the stress interpolation contain piece-wise constant functions and be L,-ortho- onal to the enhanced strain interpolation, ensure satisfaction ofthe patch test and allow the elimination of the stress field from the formulation. The preceding conditions are formulated in a form particularly convenient for element design. As anillstration of the methodology three new elements are developed and shown to exhibit good performance: a plane 3D elastic/plastic QUAD, an axisymmetric element and a thick plate bending QUAD, The formulation described herein is suitable for non-linear analysis. 1. INTRODUCTION AND MOTIVATION In recent years, considerable attention has been devoted to. the development of low order quadrilateral elements (typically bilinear) which exhibit high accuracy in coarse meshes, particu- larly in bending dominated situations. Furthermore, many of these low order elements do not exhibit the well-known ‘spurious locking’ in the nearly incompressible limit and are, therefore, particularly attractive in general purpose finite element analysis programs. It should be pointed out that, despite the enhanced accuracy of these quadrilaterals, from a numerical analysis standpoint no improvement on the standard (asymptotic) interpolation error estimates is typically obtained (see e.g. the analysis in Strang and Fix (Reference 20, p. 174)). Broadly speaking, recent design and development of low order quadriaterals with enhanced ‘coarse mesh accuracy falls within the scope of two alternative approaches, referred to as assumed strain and assumed stress methods in what follows. Representative of the first approach is the pioneering work of Nagtegaal et al"? and Willan,2* the related B-bar methods of Hughes® and Simo et al,'? and the mode-decomposition and Hu-Washizu methods of Belytschko and co-workers; ee e.g. Belytschko and Bachrach,‘ among others. As noted in Simo and Hughes,"> all of these assumed strain methods can be cast into a three-feld variational framzwork, and Research supported by AFOSR under contract nos, 2DIA-S44 and 2-DIA-T71 with Stanford University assosate Professor of Applied Mechanics Graduate Research Assistant 0029-$981/90/081595~44822.00 Received 5 October 1989 © 1990 by John Wiley & Sons, Ltd. 1596 1.6. SIMO AND M.S. RIFAE formulated as a B-bar method in which the discrete strain-displacement operator B of the standard displacement model is replaced by an assumed strain (B-bar) operator. On the other hand, assumed stress methods are based on a two-field variational formulation in which stresses and displacements are interpolated independently. Representative of this approach is the quadri- lateral element of Pian and Sumihara'! and Puch and Atluri!? ‘An alternative approach to the development of low order elements with enhanced performance in coarse meshes is the classical method of incompatible modes, originally introduced by Bazely et al. in the context of plate bending problems and Wilson et al in the context of plane elasticity. In finite element textbooks, this class of methods is often presented as an example of either ‘pariational crimes'”-?° or ‘useful tricks’.?° We show that these methods arise merely as particular interpolations within a certain class of mixed methods described below. It is also interesting to note that recently proposed stress based mixed formulations, eg. Pian and Sumihara,'? exhibit essentially the same coarse mesh accuracy and distortion insensitivity as well-known modified incompatible mode elements; ¢. Taylor et a2? ‘The class of mixed assumed strain methods presented in this paper allows the systematic development of low order elements: with enhanced accuracy for coarse meshes. Within the proposed three-field mixed finite element framework, the classical method of incompatible modes rises asa particular conforming approximation ofthe strain field. Furthermore, by exploiting this nixed finite element framework, one can improve on existing incompatible elements and design new elements. As concrete examples, we present a new element for elastoplastic stress analysis, a new axisymmetric element and a new plate bending clement. Noteworthy features of the present approach are as follows. 1. Three simple conditions are proposed which guarantee satisfaction of the patch test and stability of the interpolation, and allow the complete elimination of the independent stress field from the finite element equations. Hence, the three-feld formulation collapses to a tworfield mixed method in terms of displacements and an assumed enhanced strain field 2. The method of incompatible modes becomes simply a special instance of a ‘conforming’ mixed method, The fact, often justified on physical grounds, that incompatible modes do not participate in the definition of the element traction and body force load vectors is 2 result of the variational structure of the method. Enhanced strain approximations derived from incompatible mode fields automatically satisfy the Taylor et al* patch test condition. 3. The implementation of inelastic constitutive models; in particular, the classical strain driven return mapping algorithms of plasticity (see eg Simo and Hughes") is the same as in standard displacement models. This situation is in sharp contrast with two-feld stress driven mixed interpolations for which the return mappings take a substantially different format (see Simo et al.'*) 4, The present approach can also be formulated as a B-bar method which satisfies the ‘variational consistency condition’ set forth in Simo and Hughes."® Such a condition, however, does not guarantee convergence. In fact, we show that the original incompatible clement of Wilson et al2* can be formulated as a variationally consistent B-bar method Which, as is well known, fails to converge for certain distorted meshes. ‘An outline of the paper is as follows. In Section 2 we introduce the proposed mixed approx- imation in the context of physically non-linear elasticity, and introduce the notion of enhanced strain fields. The straightforward extension of the methodology to incorporate plasticity is discussed in Appendix IIL. In Section 3 we address issues related to the convergence and stability of the method. In Section 4 we reformulate the two crucial conditions introduced in Section 2 as conditions on the strain field in the isoparametric space (sometimes referred to as covariant CLASS OF MIXED ASSUMED STRAIN: METHODS 1597 strains). This reformulation is particularly convenient from an element design point of view. This point of view is illustrated with the development of two new clements for plane/3D and axisymmetric problems. To further illustrate the general applicability of the methodology, we consider in Section S its application to thick plates. As an example, a new plate bending element is developed. Numerical examples are presented in Section 6, As illustrated by a numerical simulation presented in this section, the proposed approach can be extended to geometrically non-linear problems. A detailed discussion, however, is deferred to a subsequent publication, 2, MODEL PROBLEM I: PHYSICALLY NON-LINEAR ELASTICITY We introduce the proposed class of mixed assumed strain methods within the context of Physically non-linear elasticity. As alluded to above, this class of mixed methods includes, as a particular case, the classical method of incompatible modes. The central idea'is as follows. We start with a chree-ield variational formulation of elasticity, and consider strain fields of the form Where Vu is the symmetric gradient of the displacement field, In a finite element context, we refer to€ as the enhanced part of the strain field. As discussed below, the finite element interpolation of theenhanced field & is nor subject to any interelement continuity requirement and, in particular, can be derived consistent with any given ‘incompatible mode’ field. We complete the formulation by eliminating the stress field via an orthogonality condition analogous to that considered in Simo and Hughes.'? Two additional conditions on the mixed interpolations guarantee convergence and stability of the method, as discussed in detail in Section 3. 2.1. Three-field variational formulation of physically non-linear elasticity Let 9 R° be the domain occupied by an elastic body, with particles labelled by xe, displacement field u:4? - R', and stored energy function W(x, #), where ¢ denotes the infinites. {mal strain tensor. In addition, 6 denotes the actual stress tensor and 1 < yim < 3 isthe spatial dimension of the problem. As usual, we assume that I(x, 6) is conver. Further, let u be specified on a part 2, of the boundary of @ as a, and let the traction vector be specified on 2,2 as oa =T (given) on 3,8 @ where fi is the normal field to 0, As usual, we assume that 2, and 2,2 are disjoint with 08 =1,900,8. Now let V be the space of admissible displacement variations, defined in the standard fashion as Vix (ne LH) Imag = 0} 8) Further, we shall denote by & and $ the spaces of admissible strain and stress variations, respectively. We have =S=[La)y~ 4 1598 J.C. SIMO AND M.S. RIFAI For plane or three dimensional elasticity we have Ms = Moin X (Main + 1)/2, whereas for axisym- metic problems faim = 2 and Meg = 4. With this notation at hand, we consider the following three standard variational equations: Vin-od¥ ~ Gault) = 0 | [Vn —e]dv=0 6) i r[-6 +4, o1d for all variations (n, 7, )€V x & x S. Here (u, ¢,) denote the actual displacement strain and stress fields, respectively, and Gza(n) is the virtual work of the external loading given by Gault) | oobenay + | tener © Relations (5) are the Euler equations associated with the well-known Hu-Washizu principle for physically non-linear elasticity. 2.1. ‘Enhanced strain field, and modified variational formulation. Next, we introduce a repara- mettization of the strain fields in the form given in (1). Consequently, any ‘admissible strain variation’ is also written as y= Vn 73 with nev o ‘The last variational equation (§), then takes the form (pee + [ Vin [—0 +0,W(x, Vu t+ OAV 6) By substituting (1) into (5),, and combining (5), and (8) we arrive at the following modified three-field variational problem: f Vind, W(x, Vu + 8d — Gault) = 0 [rt-eramaves oer -0 O) for all variations (ms) x & x S, with & being the space of enhanced strain fields CLASS OF MIXED ASSUMED STRAIN METHODS 1599 Remarks 2.1. 1. Note that the local Euler-Lagrange equations associated with (9) are the standard equilib- rium equations in ; namely, div [6, W(x, V'u + 8) + pob = » | (19) o= 0,W(x, Vu +a) along with the stress boundary condition (2) on 2,2. Although & = in @ for the continuum problem, in general #* + 0, when we introduce finite element approximations, Equations (9)-(10) are the Euler equations of tie functional nto [foverenn—enar—[ nvaer-[ tudr (ny as a straightforward manipulation shows. 3. Observe that the space of enhanced strain fields is in [Z.,(#)]%. Hence, no interelement continuity on need be enforced when constructing finite element approximations, . 2.2. Mixed finite element approximations Let # & be a finite element approximation constructed by means of standard ‘Soparametric elements. Denoting by C= [= 1, 1] x". x [ ~ 1,1] the unit cube in isopara- metric space, we thus have, for a typical element d beO x= Yn4gx,e4, (12) where N4(@) (A + Modes) are the standard isoparametric shape functions satisfying NAGs) = of (13) and &y (B = 1,.... Mioses) denote the vertices of C1. Using vector notation and standard conven- tions in finite element analysis (see e.g. Zienkiewicz and Taylor?”) we write W=N@d, Vut= Bed, (14) where d. R' **om js the vector of element nodal displacements, and B,(&) is the discrete strain operator. 2.2.1. The enhanced strain field interpolation. We consider the following discontinuous finite clement approximation for the space & of enhanced strain fields, Let z,:3, » R be the character. istic function of #, defined as 1 it xe, nm { conan a Then set ¥ UO. with 4@ = Ge), anda,e ro) 1600 1.6. SIMO AND M.S. RIFAL Here, a, are n-local element strain parameters, and G(@) is a matrix of Rt x R prescribed functions with linearly independent columns, which define the enhanced strain interpolation. If the space 4 is to ‘enhance’ the standard strain field derived from the displacement approx- imation, itis natural to require that compatible strains of the form (14), are not included in More precisely, let > ve {ree 16x. with 1.) = Bipa,} an denote the space of standard strain fields. Then we require Condition (j),. The enhanced strain interpolation * and the standard strain interpolation defined by V'V* are independent in the sense that (18) In addition, we assume that the columns of G(G) are linearly independent. Requirement (18) is the crucial condition which ensures stability of the enhanced strain approximation. In particular, as shown in Section 4, violation of (18) leads to a singular system of | equations. The admissible choices of functions G(é) are intimately related to the structure of the approximating stress subspace S*, as discussed below. 2.2.2. The assumed discrete stress field. We consider finite element stress fields t*¢* also discontinuous across element boundaries; ie, of the form tm YOre (a9) where ,(6) is the assumed stress field over a typical clement @,. We then eliminate the explicit appearance of the stress field in our finite element approximation by choosing S" orthogonal to & ie. Condition (i). S*and & are Ly-orthogonal. Since any (',*)¢4* x $*is discontinuous across clement boundaries (16) and (19) imply eae [100-.@ar=2-[ o@@ar 20 for any element ¢ = 1,2...» tim As a result of this condition, equation (9), is identically satisfied, and the first term in (9) vanishes, Hence, condition (i) (above) effectively eliminates the stress field from the finite clement equations. However, given an interpolation space &* of enhanced strain fields, the orthogonality condition (ji) (above) may produce an overly constrained space S* which would preclude ‘convergence of the method. We prevent such a situation by requiring (CLASS OF MIXED ASSUMED STRAIN METHODS 1601 Condition (ii). The space S* of stress fields of the form (19) must include at least piece-wise constant functions after enforcing the orthogonality condition (20), ic. ta Y sfz.€5%, with seca | a1) We show below that, for linéar elasticity, conditions (i) and (i) imply satisfaction of the patch test in the sense of Taylor et a”? In fact, the original incompatible modes element of Wilson et al?! fails to satisfy condition (i) when recast inthe format ofthe present formulation (see Appendix 0. 2.3. Finite element equations. Solution procedure © Substitution of the preceding interpolations into equations (9), and (9)y yields the discrete non-linear system of equations Atteid..2.) -£")=0 ) ant 22) hed,,2,)=0, (€=1,2,...57, of where A\ denotes the standard assembly operator, and | ena, = |) Bren Bd, + Gav * (23) hd. t,):= [ G70, W(x, Ba, + Ga,)aV Observe that the element external force vector f2" has the usual expression of the standard displacement model. The solution of the system of equations can be easily accomplished by a Newton procedure that incorporates static condensation of the parameters a, at the element level. Let CMs 02 W(x, Bal + Gal?) 4) be the matrix of tangent elastic moduli associated with (@®, a), and set HY: -| Gte*Gav . 25) ree f crcpay ‘Then, the iteration proceeds as follows: 2:1 i a. COMPUTE hy" by (23) and update at the element level a by setting aff?) = al) — CH]! (re aa — Boy 1602 J. SIMO AND M.S. REAL b. COMPUTE modified internal and tangent stiffness matrix by setting Fe? =e" — pay the | | kp = fc weepay— (eyo ey) t ¢. ASSEMBLE and SOLVE for a new nodal displacement increment | Ro: Aces _ | Are Ad: (ROY RH 4, SET d®*" =a” + Ad**" and GOTO a Convergence of the preceding algorithm, obtained by straightforward linearization of equa- tions (22)-(23), is attained when | R® || < TOL. Observe that at least the parameters a{t and the element residuals h{® need to be stored at the element level. 23.1. Extension to plasticity. The formulation presented above can be immediately extended to incorporate inelastic effects (e.g plasticity and viscoplastcity) such that all the standard strain | driven return mapping algorithms for plasticity and viscoplasticity carry over without any modifica- tion to the present mixed finite element context. This is in sharp contrast with stress based mixed finite clement formulations where the structure of the conventional return maps is completely lost, and a different (more cumbersome) algorithmic treatment is necessary; see Simo et al.!* As i an illustration, the treatment of rate independent plasticity is presented in Appendix III, Numer- ical simulations for J,-flow theory are presented in Section 6, 3, REMARKS ON CONVERGENCE AND STABILITY In this section we examine convergence and stability conditions for the discrete problem emanating from the mixed formulation discussed in the preceding section. First, we show that condition (ii) above ensures satisfaction of the patch test and includes, in particular, the Taylor et | al? patch test requirement. To illustrate the role played by this additional consistency condition, ‘we consider in Appendix I a detailed analysis of the original incompatible mode element of Wilson et al?® We explicitly show why this element does pass the patch test for certain distorted configurations, and give an explicit characterization of those stress fields for which the patch test i 4 is violated i Next, we make use of an interpretation of the methodology developed above as a B-bar ate method, to show that condition () ensures a unique Solution for the system of equations governing at the discrete problem; hence the stability of the discrete problem. The fact that violation of (18) results in lack of uniqueness of the discrete problem is illustrated by means of a simple example. ‘Throughout the present discussion, attention is restricted to linear elasticity so that C = constant i in expression (24), i (CLASS OF MIXED ASSUMED STRAIN METHODS 1603 3.1. Consistency and the patch test Since C is positive definite, the assumption of linear independence on the rows of Gi equivalent to the assumption that the matrix Hex | GE COWIE az 6) be positive definite; ie. aTHa,>0, for any a,eR | e With this observation in mind, we examine the implications of conditions (ii) and (ii) on the structure of G(E) and their relation to the patch test. Let GO as (28) where (G) is the Jacobian determinant of the isoparametric map. Choosing #€ = constant, for €= 1,2... sam, condition Gi) implies 0 29) Now, condition (i) on $* requires that any piece-wise constant stress be in $* Consequently, (29) must hold for any +2 eR; and we have G0) ‘We show that conditions (27) and (30) do, in fact, imply satisfaction of the patch test. In particular, (30) is satisfied if G = 0, a condition set forth in Taylor et al.?® 3.1.1. The patch test. Consider a nodal displacement vector d, such that 02:= CBd, = constant Bt) for e= 1,2, m- Then, equation (22), gives he= | GTECBd,J/(G)d5 + Ha, S762 + Ha, =0 (32) If 27) and (30) hold, then (32) gives a, = 0 ( reduces to the standard displacement model, ic =| f wrenyos i, re) which, by construction, satisfies the patch test. Hence, conditions (27) and (30) imply satisfaction of the patch test 1,2,...,main) and the internal force vector 1604 1. SIMO AND M.S. RIFAr 3.2, Stability of the discrete problem We examine solvability and uniqueness of the solutions for the discrete finite element problem ‘governed by the following system of (n + m) x (n + m) equations: K rT 6a) _ fee os) ro Hjles~lo Here 1:= Motes X Main ~ Moe ANd M:= Rai X Ney Where Mo, i the number of nodal boundary conditions. Recall that, by assumption, K and H are spmumetric and positive definite. The static condensation procedure of Section 2.3 then yields the reduced system Rd=F* 39) K:=K-r'H'r ey Since H is invertible, if (35) has a unique solution, then « are also uniquely defined by the expression a = — H~'I'd. We show below that condition (18) ensures positive definiteness of the reduced stiffness matrix K; and hence uniqueness of solution for the symmetrie system (35). The argument exploits the following interpretation of K. 3.2.1, Interpretation as a B-bar method, The system of equations (34) can be formulated as a B-bar method in which the stiffness matrix Ris given by R= Al BICB.dV (36) with B defined by the expression B,@):=B,@ - GUT 67) Furthermore, the B-bar method is variationally consistent in the sense that the following condition (set forth in Simo and Hughes") holds: { BIC[B, — B,]dV - 38) ‘The preceding result follows from a straightforward algebraic manipulation. First, substitution of the expression a, = — H~1T, into (1) defines, in view of (37), the total strain field as £, = Bd, +G,2,= Bd, 39) Consequently, the internal force veetor in (23), reduces to fre f BIC[B,4, + G,a,]4V = Lf, wzcb,ar Ja, (40) A ditet computation using defitons (25) then yields f [B, -B,]'CB,dV= FH i GICB, - G.H-'F]av =TH-'[r HHT) <0 an which proves (38). Assembly of (40) and use of (38) yields (26) (CLASS OF MIXED ASSUMED STRAIN METHODS 1605 It should be noted that the fact that the scheme can be formulated as a variationally consistent B-bar method holds independent of condition (ii) equation (20)) or, equivalently, independent of condition (30), and by no means implies convergence of the method. In fact, as shown in Appendix I, the original incompatible mode clement of Wilson et ai.?® furnishes an example of a B-bar method, which is variationally consistent and, nevertheless, fails to converge in general for distorted meshes. 3.2.2. Stability analysis. Let us denote by ker [B,J and ker (B,.] the null spaces of B, and B,. respectively. Recall that ker [B,] consists of all the nodal infinitesimal rigid body variations; ie a vector di in ker [B,} satisfies B.d:# = 0 <> 4: = nodal rigid body variation (42a) It is precisely this condition, along with the requirement that V* does not contain rigid body variations, that renders the stiffness matrix K of the displacement method positive definite Similarly, in view of (36), positive definiteness of K holds if ker [B,] consists only of nodal infinitesimal rigid body variations; hence K positive definite <> ker [B,] = ker [B,] (42) It is clear from expression (37) and the structure of that if condition B,d, = 0 holds, then condition B,d, = 0 also holds. Consequently, the inclusion ker [B,] < ker[B,] always holds. We claim that the reverse inclusion holds if and only if condition (18) is satisfied; explicitly ker [B, To prove this claim we proceed in two steps, as follows: ker[B.J = Pavyr=g (3) (a) First, we prove that 3 °V" # GF implies that ker [B,J # ket (B,J. Assume that there is #40 which isin #* VV" # Qs. This implies that there are d, and &, such that #= Ba, - 6.4.40 “4 In particular, it follows that d,¢ker [B, , However, condition (44) along with the defni- tions of H and F imply that d-eker[B.1, since Ba, = Ba. G.n-'[ rere,a.j4r GICIG,4.]4V (49) Consequently, ker [B,] # ker [B,]. (©) Conversely, we show that ker [B,] # ker [B, implies #* V°V" x Gi Infact, suppose that there is €eker[B,] such that d, #0 with d.¢ker[B,]. Then, condition B.a. = 0 and expression (37) imply that = Bd, = G.(Hra,] 40 (46) © Setting 4, = H™4P A, equation (46) implies that eda VV" ¥ gs . 1606 4.6, SIMO AND M.S. RIFAL To summarize our stability analysis, since K is positive definite, the results above show that the system (34) (or the reduced system (35)) is uniquely solvable if and only ifthe following conditions hole: (i) His positive definite; equivalently, the columns of G are linearly independent. (i) &aVV" = gi equivalently, the enhanced strains (generated by G) are independent of the standard strain field generated by the displacement approximation The simple example below illustrates a choice of interpolation functions G(E) for which conditions (i) and (ii) of Section 2 are satisfied, but the stability condition (j) of Section 2 is violated, leading to an unstable formulation. In fact, the example yields the one-point uniformly reduced bilinear quadrilateral which is rank deficient. Example 3.1. Consider a square bilinear element with bi-unit sides, so that no distinction needs to be made between Cartesian and isoparametric co-ordinates. Further, consider the following interpolation (restricted for simplicity to square elements}: 1 000 G@=|0 6 00 “ oogn [A straightforward calculation shows that B, is constant and given by B, = B,(@)|,~»- Clearly, the stiffness matrix R is that associated with the one-point uniformly reduced integration element which is rank deficient, and the method is unstable. Observe that for the square element, the standard strain-field V'n* associated with any meV" is of the form 1007000 vente spanfO 10 0 ¢ 0 0 (43) oo0100E% which contains functions in the span of (47) Consequently, *V°V" # @ and condition () of Section 2s violated, . 3.3, Stress recovery It is apparent from the basic equations (22)~(23) that the stress field 6* 0. We denote by (x) and (x) the vertical deflection and two rotations at a point xe@. As usual, we set wi=[o @"]=[o ©, ©] 7) The bending and transverse shear strains are k=V', e=Vw-0 (88) respectively. A variational formulation of the (linear elastic) plate equations in which (88) is enforced via Lagrange multipliers (corresponding to transverse sheat strains) is obtained via the following classical functional: Tu, 6, €) f IV'O-C,V'OdY + Tate) +f GGhk\e|? + o-[Vo—O —e)dV 9) where C, is the matrix of bending moduli, and TI, (u) is the potential energy of the external loading with a standard expression. Note that for simplicity we have assumed isotropy, with G denoting the shear modulus and k the shear coefficient. CLASS OF MIXED ASSUMED STRAIN METHODS 1615 5.1.1. Modified variational formulation: Enhanced shear strain field, As in Section 2 we consider transverse shear strain fields of the form e=[Vo-O]+ 90) Substitution of (90) into (89) yields naen=ne+[toume-est-enw on deflections and rotations, the condition for stationarity of the functional in (91) associated with the ‘displacement’ variables u = (c, @) take the familiar expressions DI)" [ v-cnetve 0 + ear 0 tees (92) DIL,(w)- +| Vn: Ghk [Vo —~@ +3]4V=0 for all variations (y, ¥)< V. Equations (92) are completely standard except for the appearance of the enhanced strain field &. As in Section 2 we shall denote by # and S the spaces of enhanced shear fields 7 and transverse shear forces t, respectively. Then, associated with (91) we have the ‘additional stationarity conditions [eterno oy 3) FL ~ 6 + Ghk(Vo — ©) + Ghka}av for all variations (r, 7)€S x of the transverse shear force and enhanced transverse shear strain field. Itis apparent that, aside from the change in notation and the different physical meaning of the fields (t, 7JeS x &, equations (92) constitute the counterpart of (9),, and equations (93),,3 are essentially identical to (9). 5.2. Mixed finite element interpolation ‘The interpolation subspaces #* and S* are again of the form (16)-(19) and are subject to the same two conditions (20) and (21). Following our developments in Section 4, we consider a conerete realization ofthe spaces S* and é for the four-node bi-linear isoparametric element. We summarize the relevant definitions needed, We set a= fr = L1On178 = [ frees. with aer'} 04) 1616 J.C SIMO AND M.S, REAL and foi : a {t= X 1.6) z01 4.6) = JoSOB, with Bye? x RS 05) ‘The interpolation functions $(€) and E(G) for the stress field and enhanced strain field, respect- ively, are subject to the two conditions: a otede ~o=f ST@EGag=0 6) (ii) S* contains piece-wise constant element transverse shear stresses; so that E@ds=0 7) Observe that we have employed the same notation as in Section 4, 5.2.1. The choice of isoparametric stress and enhanced strain fields. We consider the following structure of stress interpolants: 1070 so-[} a ‘| (98) Furthermore, we set (99) Clearly, (98) and (99) satisfy conditions (96) and (97) above. Further, observe that n, = 4 > 2. Our ‘numerical simulations show that the resulting element exhibits no locking response in the thin plate limit. Remarks 5.1 1. As noted in Setion 3.2.1, the scheme can be formulated as a B-bar method. We note that, for Mindlin-Reissner linear isotropic plate theory, expression (37) for B becomes independent of the constitutive properties since the factor Ghik cancels out. Infact, B nal | orcas] ik Grae (100) 2. A mechanical motivation for the choice for interpolations (98) and (98) is as follows. The standard isoparametric interpolations of the displacements (w, ©, and @,) are a, 4b, + den a (101) ay + bse tes + dyin ©, = ay + bE + eam +d, | Thus, for a square element, we have (102) CLASS OF MIXED ASSUMED STRAIN METHODS 1617 Hence, 0. ~ ©, =, 43) +d, 62) — BE ~ dh . (b, = 43) + (dy = 2)n Bad = dad _ 3 ~ ©, = (6, — 5) + (dy ~Bs)E ~~ dyn It follows that the enhanced transverse shear strain field &, emanating from (99), ie by = bay + ony a (104 By = ny + Ende provides the additional terms that balance Veo and ®. Since -[f oe iE [ otwo-0+ 908 (105) a simple calculation shows (for the square element) that by (106 The resulting enhanced shear strain field becomes Yo @ +g n fOi— a+ ds —eady ‘ton Uler = a3) +, — bse S Upon expansion of the components of these expressions, the shear strain field is seen to be precisely that of the Bathe and Dvorkin? element; sec also Bathe and Brezzi.* ‘The transformations (94) and (95) account for the case in which the element is no longer square. . 6. NUMERICAL SIMULATIONS In this section we illustrate the methodology developed in the preceding sections in a number of numerical simulations. First, we show that the interpolations (69) and (70) of Section 4 exhibit comparable performance ((69) slightly superior to (70), which is equivalent to the incompatible ‘modes element of Taylor et a.?2). We also illustrate the excellent performance of these enhanced clements in the nearly incompressible problem by means of a well-known test problem. Next, we consider elastoplastic response and assess coarse mesh accuracy in bending dominated problems. In Section 6.3 we assess the performance of the axisymmetric elements developed in Section 43 and in Section 6.4 we illustrate the performance of the quadrilateral plate bending clement developed in Section 5. We conclude this section with a simulation concerned with a geomet- rically non-linear problem. 6.1. Plane stress/plane strain elasticity ‘The incompatible modes element of Taylor et al? has been well known to produce excellent results in bending dominated situations. -It appears, however, that the excellent distortion insensitivity and coarse mesh accuracy, and the good performance of this element in the 1618 5.€.SIMO AND M.S. RIFAE Degre of Stew — 0 Figure 2 Beam bending problem: sensitivity to mesh distortion incompressible limit, are not widely appreciated. We illustrate below these characteristics, and demonstrate that the performance of this element is essentially identical to recently proposed rwosfield assumed stress elements (e.g, Pian and Sumihara'?), The five-parameter element de- veloped in Section 4 improves upon the performance of the Taylor et al? quadrilateral 6.1.1. Beam bending: Sensitivity to mesh distortion. In this standard test, a beam modelled with ‘a two-element mesh is fixed on one end and subjected to a bending moment on the other end, as shown in Figure 2. The edge separating the two elements is then gradually rotated (adistance +: a ‘on the top and bottom surfaces) to skew the mesh. The results for this test (normalized with the exact solution) are shown in Figure 2 for the quadrilateral of Taylor et al,? the five-parameter enhanced strain interpolation of Section 4, and the assumed stress element of Pian and Sumibara.1* The similar distortion sensitivity performance exhibited by all these elements is noted. 6.1.2. Incompressible elasticity: Driven cavity flow. This isa standard test problem for assessing the performance of finite element formulations in the nearly incompressible limit (ee eg. Hughes”). The ‘leaky-lid’ boundary condition is employed, and the pressure results of the enhanced strain element of Section 42 are shown in Figures 3 for 10 x 10, 20 x 20 and 40 x 40 clement mesh configurations. The ‘smoothed’ stresses, obtained via an L projection to the nodes, are also shown. Even though the results exhibit oscillations in the pressure, the magnitude of these oscillations is reduced with mesh refinement, and the smoothed distribution converges rapidly. This is in contrast with the well-known response exhibited by the bilinear displace- ‘ment/constant pressure quadrilateral element. Remark 6.1. We note that the pressures evaluated at the centre of the element are always in close agreement with the smoothed distribution, and hence more accurate. This observation is in complete agreement with the stress recovery procedure in Section 3.3. In effect, the stress field CLASS OF MIXED ASSUMED STRAIN METHODS 1619 over the element evaluated using the relation (vu +8 (108) | does not satisfy the orthogonality condition (9),. On the other hand, the stress field (54) obtained through the variational stress recovery is equivalent to evaluating (108) at the centre of the . clement. Poston —2 Figure 3a) Driven cavity Now problem. Pressure distribution at y ¥ 022.10 x 10 mesh Prosure Potion —= Figure 306). Driven cavity flow problem. Pressure distribution at y = 022.20 x 20 mesh 1620 J.C. SIMO AND MS. RIFAT ——— oF Figure 30) Driven cavity low problem. Pressure distribution at y = 022.40 x 40 mesh 6.2, Plasticity and viscoplasticity: J_-flow theory The extension of the mixed formulation presented in Section 2 to plasticity and viscoplastcity is discussed in Appendix TIL Here, we present several examples that illustrate the excellent performance of the enhanced strain elements of Section 43 in the elastoplastic regime. We focus ur attention on the interpolation (70) which, in the linear elastic regime, is equivalent to the Taylor et al? incompatible mode formulation. Several standard test problems are used to demonstrate the advantage of these enhanced elements over the standard displacement formula tion, and the constant dilatation (B-bar) approach for both plane strain and plane stress problems. 6.2.1. Extension of a double edge notched specimen. This test problem was introduced by Nagtegaal et al."° to demonstrate the spurious response of standard displacement models in highly constrained plane strain, axisymmetric or three dimensional clastoplastic problems. Although the problem at hand exhibits a limit load given analytically by ofi** = 297y, where ay is the flow stress, the displacement model produces a load-deflection curve which increases monotonically beyond the load limit. Perfect plasticity is assumed, and the values of the material constants are E=10, ‘The specimen has total width W”= 10, height L a value of Fy, * 0-7217 for the limit load. ‘The simulation is carried out by displacement controlled load steps at the top surface of the specimen. A'S x 15 mesh is used to model a quarter of the problem using symmetry boundary conditions. Each displacement step corresponds to EAS anos (119) 3, oy = 0243 (109) 30, and ligament thickness b = 1, resulting in CLASS OF MIXED ASSUMED STRAIN METHODS 1621 te Nonmallaed Verte Deboctioa — By Figure 4. Double edge notched tension specimen: clastoplastic analysis (uggested by Nagtegaal et al!) The load-defection behaviour of the incompatible modes element asymptote tothe analytical lt load, while the standard displacement formulation yields indefinitely increasing loads The load-deflection plots for both the displacement and enhanced strain formulations in plane strain are shown in Figure 4 for 0 < Ed/ey W < 15, where d is the top displacement (30 time steps). Also shown is the analytical limit load. It is apparent that the displacement formulation does not exhibit any limit load, while the enhanced strain formulation element asymptotes to the analytical solution. 6.2.2. Plane strain: Cook's membrane ‘problem. This test problem demonstrates the superiot coarse mesh accuracy in bending dominated elastoplastic problems, and insensitivity to mesh distortion exhibited by the enhanced strain elements of Section 4 over the mean dilatation approach (B-bar) of Nagtegaal et al.! We consider a tapered panel, clamped on one end, and subjected to a shearing load on the other. The elastic version of this problem is known as ‘Cook's membrane problem’. The simulation is performed using load control with steps of AF = 0-1, and F ranging from 0 to 1-8. The vertical displacement of the top edge node is plotted against different mesh configurations for both formulations in Figure 5. The mean dilatation approach appears to converge to the same answer as the enhanced element; however, much finer meshes are required. Infact, the enhanced strain formulation (with interpolation given by (69) is practically converged with an 8 x 8 mesh, while the mean dilatation method shows = 5 per cent error, even with a 64 x 64 mesh (4096 elements). Je flow theory with linear kinematic isotropic hardening is assumed. The values ofthe material constants are E=7, 243, H= 0135, K =0015 (uy where K and H denote the isotropic and kinematic hardening modulus in uniaxial stress, respectively. This same problem was re-examined in the plane stress’ case, and the results compared with the assumed stress formulation based on the interpolation of Pian and ‘Sumihara'! using the stress driven return mapping algorithm of Simo et al® The results 1622 1.6 SIMO AND M.S, RIFAL i I i Elements per Side — Figure 5, Cook’s membrane problem: plane strain elastoplastic analysis. A clamped wing is subjected to an in-plane shearing load using several mesh configurations. The incompatible modes formulation yields practically conversed results with the 8% 8 results, while the Bbar formulation is at 5 per cent error even with a 64 % 64 mesh (096 cements) (shown in Figure 6) demonstrate the superior performance of both formulations over the standard displacement model. It should be noted that both the enhanced strain element and the assumed stress element exhibit essentially identical performance. However, the assumed stress element requires a non-standard (more cumbersome) return mapping algorithm, while the enhanced strain clement uses the standard strain-driven return mappings. 6.2.3. Plane stress: Clamped arch problem. As a further illustration of the performance of the enhanced strain elements of Section 4 in plane stress analysis, we consider a clamped arch, of radius R = 10 and thickness ¢ = 1, vertically loaded at the top. Again, J,-flow theory with linear isotropic/kinematic hardening is assumed, along with the following values for the material constants: E=7, ¥ 0, K=01 (ay Load control is employed with equal force increments of AF = 0.01 to obtain values 0 y Element A Element B Element C Exact 00 m61L 372611 372611 37500 025, 4.41340 441340 441340 44531 03 453826 433 453826 45825 049 497084 497 497055 50399 0.499 493921 493921 498921 50602 0.4999 499107 499107 499107 50623, Table IL, Thick walled cylinder. Skewed mesh Displacement 10? Element A. Element B Element C Exact 372361 3.72355 372666 37500 4441005 440991 441540 44531 453463, 45M47 454076 495825 496376 496375 497625 50399 498207 4.98209 499514 50602 498388, 498391 499702 50623 6.3.2. Bending ofa thin cylinder. A cylindrical shell of median radius R = 167°5 length L = 51, thickness h= Land Young's modulus E = 11250 is subjected to an end moment M = 2000, as shown in Figure 11. The exact (shell) sélution for the deffection under the moment is given by M a me = Gt ty Dats PD aR 2-7) The cylinder is modelled with one row of 17 elements (aspect ratio 3) and the results are shown in Table III for different values of the Poisson's ratio. Again, the three elements yield almost exact results for this problem, and no degradation occurs in the incompressible limit Bt (116 6.3.3. Bending of a circular plate. A circular plate of radius R = 10, thickness h= 1 and Young's modulus E = 1875 is subjected to a uniform loading q = 1, as shown in Figure 12. The —> = Figure 11. The bending ofa thin eylinder problem is modelled using one row of 17 axisymmetric elements inthe axial iection The radius is 1675 and the thickness 1 (CLASS OF MIXED ASSUMED STRAIN METHODS 1627 Table IML. Bending of a thin cylinder Displacement under the moment Element A Element B Element C 00. 0.60637 060637 060637 025 058708 058708 058708 03 os7sas o5s7eas os784s oy 052889 052899 052899 0-499 052580 052580 04999 052548 052549 052548, HA Figure 12. The cicular plate bending problem is modelled using one row of 4 axisymmetric elements in the radial direction. The thickness is 1, and the plate is subjected to @ uniform load g= I in the -diection exact (plate) solution for the central displacement is given by ak (je: s3ere) > D\ Tey) 31 = VFR? “We ay ‘The results for this problem (listed in Table IV) exhibit some dependence for elements A and Bon the values of the Poisson's ratio used. However, the performance of element C is practically Table TV. Bending of a circular plate Central displacement ’ Element A Element B Element C Exact 00 s-16485 534791 5.16885 025 401647 427654 408232 03 3.74893, 403096 385023, 04g 255261 290857 292096 0499, 2.48632 283936, 287508 0.4999 2.47963 2.83268 287088 1628 J.C SIMO AND M.S. RIFAT Figure 13. The elect of the Poisson's ratio on the performance ofthe axisymmetric elements. Elements A and B show deterioration of performance with increasing incompresibiliy, while clement C maining excallent performance over the Whole fange of» insensitive to the value of the Poisson ratio. For element A (interpolation (77), the element is 2 per cent soft for v= Oand IT per cent stiff for y = 0-499; for element B (interpolation (79)), the element is 6 per cent soft for v= and 1 per cent soft for v = 0-4999, whereas for element C (interpolation (81), the element is 25 per cent soft for v = O.an_ 3 per cent soft for v = 04999, This dependence of elements A and B on the Poisson's ratio is illustrated in Figure 13, 64, Plate bending element In this section we ilustrate the performance of the Mindlin-Reissner plate bending element developed in Section 5. Three different eometries are used: a square plate, a rhombic plate and a circular plate. The results are compared to the quadrilateral element of Hughes and Teaduyar® (referred to as Ti in what follows). As noted in Remarks 4.1, the present element is identical for a square mesh to the TI element, and hence the results are the same for the square plate problem, ‘The thombic plate configuration results in a mesh of rhombic shaped elements, and the present clement seems to perform slightly better than T1 in this configuration, However, for the circular plate problem, the T1 element performs better. Nevertheless, the element of Section 5 is a “decent performer’ which illustrates the use of enhanced strain interpolations in the construction of locking-free plate bending elements. 64.1. Bending of a square plate. A simply supported square plate of side length L = 10, thickness h = 0-1, modulus of elasticity £ = 10:92 and Poisson's ratio of y = 0'3 is subjected to a uniform load q= 1. This problem is modelled using symmetry boundary conditions on a quadrant. Since the mesh is square, the results of our element and the T1 element (Table V) are identical, consistent with Remarks 4.1 the is the ant 99, cent 10, 1t0 are CLASS OF MIXED ASSUMED STRAIN METHODS 1629 Table V. Bending of a square plate Central displacement » 10* Mesh Present Series 2x2 39712 39712 4.0644 4x4 40836 40436 40648 8x8 40593, 4.0593 16x16 40632 4.0632 Figure I Rhombie plate mesh 6.4.2. Bending of a rhombic plate. A simply supported 30° skew plate of side length L = 100, thickness ¢= 1, modulus of elasticity £ = 10° and Poisson's ratio v= 03 is subjected to uniform loading q = 1. The exact solution to this problem includes singularities in the moments at the obtuse vertices. and the problem has been considered by some to be a non-representative numerical test. The mesh configuration shown in Figure 14 is used to model the full plate, resulting in rhombic shaped elements. The results listed in Table VI indicate that the present “enhanced strain’ element performs slightly superior to the T1 element in this problem. 6.4.3. Bending of a circular plate. simply supported circular plate of radius R = 5, thickness 0-1, modulus of elasticity £ = 1092 and Poisson’s ratio v = 0-3 is subjected to a uniform load 1. A quadrant of the problem is modelled using symmetry boundary conditions, using the q ‘mesh configuration shown in Figure 15. This mesh configuration imposes severe distortion on the elements, and as shown in Table VII, the TI element is superior to the present clement. Nevertheless, the present element is locking-free and exhibits reasonable behaviour eve in this case. 6.5. Geometrically non-linear problem: Compressible neo-Hookean material ‘The extension of the method described in Section 2 to the geometrically non-linear regime will bbe addressed in a subsequent publication. Here, we shall merely illustrate the performance of this 1630 Mesh 4x4 8x8 16 x16 32.32 1.6 SIMO AND M.§. IFAT Central displacement x 10-2 Present 393046 398412 427274 446680 Central displacement x 10* Mesh Present Helms 3.6966 48elmts 3.9140 192elmts 3.9664 768 elms 39791 Figure 15, Circular plate mesh Tt 39159 38803 41565 43882 TI 39070 39649 39789 39822 Table VI. Bending of a rhombic plate Series 4455 4455, 4435 4435 Table VII. Bending of a circular plate Series 39831 39831 39831 39831 CLASS OF MIXED ASSUMED STRAIN METHODS 1631 laments per Side — Figure 16. Cook's membrane problem: compressible neo-Hookean material mode! formulation in the context of the non-linear version of the element described in Section 4.2. We consider once more Cook's membrane problem, described in Section 6.2.2, but now in the non-linear regime for an elastic material characterized by the following stored energy function: yf Lace (E+ u)ins + futre~3) (118) where Cis the Cauchy-Green strain tensor, and J isthe Jacobian determinant of the deformation gradient. The dimensions of the problem remain the same as in Section 62.2, and the new material properties are 2 = 52-5 and = 2625. In Figure 16, we compare the results obtained with the standard bilinear element with constant volume and constant pressure, as described in Simo et al,!? and the non-linear version of the quadrilateral element in Section 4.2. The advantage of the latter formulation over the former is manifest 7. CONCLUDING REMARKS ‘We have presented a class of assumed mixed finite element methods which allows the systematic development of low order elements possessing Zood coarse mesh and distortion insensitivity 1632 1.6. SIMO AND M.S. RIFAT properties. We have shown that well-known incompatible mode element formulations arise in the j present context merely as particular (conforming) interpolations of the enhanced strain field. We | have illustrated the methodology with the development of new plane, axisymmetric and plate bending clements. Furthermore, we have shown that assumed stress low order quadrilaterals, currently perceived as possessing nearly optimal accuracy and distortion insensitivity in coarse ‘meshes, exhibit essentially the same performance as the assumed strain enhanced elements 1 considered in this paper. { Constitutive equations in non-linear solid mechanics typically define a stress tensor in terms of i a suitable conjugate strain measure. Assumed strain methods use these direct constitutive relations i in a three-field variational setting. Assumed stress methods, on the other hand, employ inverse constitutive relations which define strains in terms of stress measures in the context of a two-fcld i variational formulation, Well-known and widely used constitutive models, such as finite deforma- t tion elasticity, demonstrate that explicit expressions for these inverse relations are generally not available. Furthermore, at the constitutive level, the standard algorithmic framework in | non-linear solid mechanics is typically strain driven; this is the case for the standard return | mapping algorithms of plasticity, For these reasons, from a practical standpoint, we believe that assumed strain mixed methods are much better suited than assumed stress methods for nor-linear analysis. APPENDIX L. ANALYSIS OF WILSON'S QUAD In this Appendix, we summarize some results needed on isoparametric elements, we derive an explicit expression for the interpolation matrix G(E) of the original incompatible mode of Wilson | cet al, and we give an explicit characterization of the constant stress states that violate the patch i test We write the bilinear isoparametric interpolations as TING, yayt \ Al NG ay + aE + azn + hen J ay where § = ($4) and iy" i ' \ to Fl x=E1 om x3 4)" y=On v2 ys yal" Here, x4 = (x4. ¥4h A= L,...,4, are the co-ordinates of the nodal points of a typical element i =| 2] -pee x8] o dy ay YN Ya, on CLASS OF MIXED ASSUMED STRAIN METHODS 1633, with Nowa, tinh, Ny=ay+ ch ay Next, following Wilson et al,?* consider the following incompatible displacement interpolation: ja, + (La (as) ‘a The derivatives of the incompatible shape functions N relative to the Cartesian co-ordinates are obtained via standard transformation with F-T as ay eani (eerie sia eth —nyth tidhenoll 22 met reall “9 where J@):= det G1 = Jo tit tian an Tos constants jan) cm be easly compuedin terms oa arhand (3) sing (A) and We then consider an enhanced strain field that is derived from the incompatible displacement ficld as By = 3s + Gy) (as) where a comma indicates differentiation with respect to Cartesian co-ordinates. From (A5), (A6) and (A8), we thus obtain b6'9, 10% -6{0} a with ne 0 jo xTage —xTayy xta,g —yTa,e yan 1 f-the 0 eee 0 = xThy? (A10) FO} ayer rene? ane Observe from this expression that xth 0 xth | 40 (ai xTh o-yTh —xth yh f GHG a Hence, condition (30) is generally violated unless xT = yh = 0; i.e. for a square (or parallelogram) element. Recall that the classical convergence proof for Wilson’s element depends crucially on the assumption of square geometry; see Ciarlet (Reference 5, p. 260). 1634 4. SIMO AND M.S. RIFAT ~~ Le] Figure 17. Example ofa pateh test on a distort mesh satisfied by the Wilson er a2 incompatible modes element —_ > 1. The kernel of G* It is of interest to compute explicitly, the constant stress states for which condition (30) is Violated. From expression (A11), we easily compute that yh aint “0 | andy arn xn] oo G40 12 0) On the other hand, the vector t, orthogohal tof, sind, and thus given by { oh? y= { (th? (ais) lotion) a Gtr, = 0= ker [G7] = span {ry} (als) “This sult explains why the Wilson element passes patch tests on certain distorted meshes, such as the one showin in Figure 17 For the example in Figure 17, we have yTh in both elements. Thus, (Al4) gives ef=[1 0 0}-Since the kernel of G" contains the solution @ = [2 0 OJ the element passes this patch test APPENDIX II, ENHANCED STRAIN INTERPOLATION In this Appendix, we give an exp! space, associated with the original incompatible mode element of Wilson et al Recall that the ‘incompatible’ displacement field is cit expression for the interpolation matrix E(@) in isoparametric yas NG)a + Nas (alo) where 1V4 are the incompatible shape functions: (1 =), N?(), (A17) CLASS OF MIXED ASSUMED STRAIN METHODS 1635 The covariant basis vectors associated with the isoparametric map are ay] fxth feb} (alg) =a0 + ine, by (at) From (A16)-(A19), we conclude that ae? 0.0.0. 0 ae 2,98 1000 & be (420) ep ara | } mance? Ob n BH ~ a8 -a8 Since g?, g9 and g are generally independent, it Tollows that (A20) defines the six-parameter interpolation (65), APPENDIX Ill, EXTENSION TO PLASTICITY We present in this Appendix the extension of the formulation and solutions algorithms developed in Section 2 to include inelastic effects; in particular, plasticity and viscoplasticity. We emphasize that all the standard strain driven return mapping algorithms for plasticity and viseoplasticity carry over without any modification to the present mixed finite element context. 1. Local strain driven return mapping algorithm Let $(6, @) <0 denote the yield condition in stress space, let q be the set of generalized hardening variables and let ¢° denote the plastic train. Without loss of generality we assume that the tensor of elastic moduli C = constant, and that the matrix of generalized hardening moduli, denoted by D, is also constant. The integration of the classical equations of associative plasticity is based on the following return mapping algorithm (see Simo and Hughes" for a detailed account and further feferences to the subject). Step 1. For a given initial data {¢2,q,} and given total strain eq, compute the tral elastic state mel s= SCOR q,) } (a2) 1636 J.C. SIMO AND M:S: RIFAT Step 2. If @i'41 <0, the final state is the trial state (@!) q,) If, on the other hand, 444 > 0, find (0, 4,4s-1) by solving the constrained optimization problem 4 = ACP. 4ues) ab Bho = Oh + AVS (Cnet nei) oy (a22) Get = = AYD 3g rts dees) PO n+15Gn21) =O For plane strain and 3-D J,-flow theory, the preceding algorithm reduces to the well-known radial return method of Wilkins** and Krieg and Key.’ For plane stress J,-flow theory algorithm (A22) reduces to the return map of Simo and Taylor.'* ‘The essential point to be noted is that, within the context of standard displacement methods, the return mapping algorithm (422) is performed independently at each quadrature point of a typical element for a given total strain €, ,. A global solution procedure based on Newton's method relies crucially on the use of consistent algorithmic tangent moduli C#,, introduced in Simo and Taylor,’ and obtained by linearization of the return map. At each quadrature point one has the incremental relation Boyer = C1 Meas (023) where Ae,,, and Ac,+, are the stress and strain increments (at the quadrature point). As an example we consider perfect plasticity (q = 0), Define the tensor of moduli BELO HAP PG VORA (a24 and set for convenience Vpq+1:= 2(0,4 1)/26,» . Then, for gi" > 0 (plastic loading) one finds the following expression, CB, (A25) whereas Cz. = C for gy’? <0 (clastic loading). 2, Solution procedure ‘The finite clement equations to be solved can be recast in a form entirely analogous to (22). To this end, define the strain field +1 = BO das +6 Oa (A26) and regard the parameters a, as element degrees of freedom, stored at the element level, but otherwise entirely analogous to the degrees of freedom d,,, ‘Now define o,,, by the return map equations (A21)-(A22), and replace (23) by fe, [i BTo,.:dV oe (a7) CLASS OF MIXED ASSUMED STRAIN METHODS 1637 Observe that, for Ay = 0 in &, these equations are identical to (23) except for the presence of the plastic strain ef. The solution of (A24) is performed iteratively (in a fashion similar to Section 2.2.3) using the following algorithm: a. UPDATE nodal displacements at iteration k + 1: att) a a aa® b. UPDATE’ at the element level a! by setting af? — PHY)" aae — AP) ©. COMPUTE total enhanced sirain een ag? = Ba?) 4 Gate d.-COMPUTE of) and Cz," and update internal variables using a standard strain driven return mapping algorithrn. ©. INTEGRATE clement matrices and residuals: Het: a) Ger" Gay reine | ore Bay Bay oe J Brofs'ay §, COMPUTE modified internal load and tangent stiffness matrix by setting Tee Tee — page tng? Bye) = ee — preemyrpaa ny apr sey g. ASSEMBLE and SOLVE for a new nodal displacement increment gene Kiso —fe1 gevo.n Ragen Ad?) [RAPD] Rasy h. SET kek +1 GOTO a, "The element residual hand the matrices H and T_" are stored fom the previous iteration along with the parameters 1638 1.€.SIMO AND M.S. REAL Convergence of the preceding algorithm, obtained by straightforward linearization of equa- tions (22)-(23), i attained when | 8 < TOL. REFERENCES 1. KJ. Bathe and F. Beezi.“A simplified analysis of 1wo plate bending elements The MITC$ and MITC Proc. Conference NUMETA 87, Univesity College of Swansea, Wales, July 1987 2 KJ. Bathe and E. N. Dyorkin, ‘A continuum mechanics based four-node stl element for general nrclinear analysis, In. J. Comp. Alded Eng. Sofware, 1 (1988). 3. GP, Bazely, ¥.K. Cheung, BM. irons and O. C, Zenkiewice, Triangular elements in plate bending—Conforming tind nonconforming solutions, Proc. First Conference on Matrix Methods in Structural Mechanics, Wright-Patterson ATBEB, Ohio, 1965 g 4. T Belytschko and W. E, Bachrach, ‘Efcient implementation of quadilatrals with high coarse-mesh accuracy, Comp. Methods Appl. Mech, Eng, 23, 323-331 (1986) 5, BG. Gilet, The Finite Element Method for Eliptic Problems, North-Holland, Amsterdam, 1978 6 T.J.R, Hughes, ‘Generalization of selective integration procedures to anisotropic and nonlinear media, Int. maner methods en, 18, 1413-1418 (1980) 1. T.I.R. Hughes, The Finite Element Method, Prentice-Hall, Englewood Clif, New Jersey, 1987 8. TIER Hughes and T. E.Teauyar, Finite elements based upon Mindln plate theory with particular reference to the fout-node bilinear isoparametric element, J. App. Meck, ASME, 587-596 (198), 9. RD Kreigand SW. Key, Implementation ofa time dependent piastcity theory into structural computer programs, in JA. Strcklin and K. J, Sacalst (eds), Constitutive Equation in Viscoplastcty: Computational and Engineering Aspects, AMD-20, ASME, New York, 1976. 10, J.C. Nagtegaal, D. M. Parks and J. R. Rice, ‘On numerically accurate finite clement solutions in the fll plastic range’, Comp. Methods Appl. Mech. Eng, 4, 153-197 (1978). 11, T.-H. Pian and K Sumihara, ‘Rational approach for assumed stress int 1685-1695 (1985). 12. EF. Puch and S.N. Alluri, ‘Development and testing of stable, invariant, isoparametric curvilinear 2+ and 3D. hybrid-stress elements’, Comp. Methods Appl Mech. Eng. 47, 381~356 (1984). 13, LLC: Simo and T. R. Hughes, On the variational foundations of assumed strain methods, J. App. Mech, ASME, 13, 51-54 (1986) 14, J.C. Simo and TJ. R. Hughes, Plasticity, Viscoplasticity and Viscoelasticy: Formulation, Algorithms and Numerical “Analysis, Springer-Verlag, Beli, to appear. 15, 1.C, Simo and R. L, Taylor, ‘Consistent tangent operators for rate independent elato-pasticiy’, Comp. Methods Appl Mech. Eng. 48, 101-118 (1985). 16, 1.€. Simo and RL. Taylor,‘A return mapping algorithm for plane stress elastoplasticity’ Inj. ume. methods eng. 22, 689-670 (1986) 17. 1. Simo, J. G. Kennedy and 8. Govindjee, ‘Nom-smooth mulisurtace plasticity and vscoplaticity. Loed- ingfonloading conditions and numerical algorithms’ Ij numer. methods eng. 2, 2161-2185 (198) 16, LC Simo, .G. Kennedy and RL. Taylor, Complementary mixed finte clement formulations of elstoplstcty’, Comp. Methods Appl. Mech. Eng. appeat. 19, 1.€. Simo, R. L. Taylor and K- 3, Piste, "Variational and projection methods fr the volume constaint infinite ‘deformation elastoplastcity’, Comp. Methods Appl. Mech. Eng, St, 177-208 (1985. 20. G. Strang and G. Fi, A Analysis ofthe Finite Element Method, Prentice-Hall, Englewood Clifs, New Jersey 1973 21, RL. Taylor, (1989), Private Communication, Aug. 1989 22, RL Taylor, P. J. Beresford and EL. Wilson, A son-conforming element fr stress analysis Int. numer, methods ‘eng, 10, 1211-1219 (1976. 23, RIL. Taylor, JC Simo, O,C. Zienkiewice and A, C. Chan, ‘The patch est A condition for assessing finite element convergence’, inj. numer. methods eng, 22, 39-62 (1986). 24, M.L. Wilkin, ‘Calculation of elastic-plastic flow’, in B. Alder etal. (eds). Methods of Comp Academic Press, New York, 1964 25, E:L, Wilson, RL. Taylor, W. P. Doherty and J, Ghaboussi, "Incompatible placement model in S.J. Fenves er al (eds), Numerical and Computer Models in Stractural Mechanics, Aeademie Pres, New York, 1973, 26. 0. C Zienkiewier, [I977} The Finite Element Method, rd edo, McGraw-Hill London, 1977 21, 0. C. Zienkiewiee and RL. Taylor, The Finite Element Method, Vo. , sth eda, McGraw-Hill, London, 1989 28. K. J. Willan, Finite element analysis ofellular structures’ PhD. thesis at University of California, Berkely. ements’ Int mumer. methods eng20, putational Physics 3,

Вам также может понравиться