Numerical Modelling in Damage Mechanics

Numerical Modelling in
Damage Mechanics
This page intentionally left blank
INNOVATIVE TECHNOLOGY SERIES
Numerical Modelling in
Damage Mechanics
edited by Khemais Saanouni
London and Sterling, VA
First published in 2001 by Hermes Science Publications. Paris First published in Great Britain and the United States in 2003 by Kogan Page Science, an imprint of Kogan Page Limited Derived from Revue europeenne des elements finis, Numerical Modelling in Damage Mechanics, NUMEDAMW. Vol. 10, no. 2-3-4. Apart from any fair dealing for the purposes of research or private study, or criticism or review, as permitted under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced, stored or transmitted, in any form or by any means, with the prior permission in writing of the publishers, or in the case of reprographic reproduction in accordance with the terms and licences issued by the CLA. Enquiries concerning reproduction outside these terms should be sent to the publishers at the undermentioned addresses: 120 Pentonville Road London N1 9JN UK www.koganpagescience.com Hermes Science Publishing Limited, 2001 Kogan Page Limited, 2003 The right of Khemais Saanouni to be identified as the editor of this work has been asserted by him in accordance with the Copyright, Designs and Patents Act 1988. ISBN 1 9039 9619 8 22883 Quicksilver Drive Sterling VA 20166-2012 USA
British Library Cataloguing-in-Publication Data A CIP record for this book is available from the British Library.
Printed and bound in Great Britain by Biddies Ltd, Guildford and King's Lynn www. biddies. co. uk
Contents
Foreword Khemais Saanouni 1. Some Aspects of a Gradient Damage Formulation Tina Liebe, Paul Steinmann and Ahmed Benallal On the Numerical Modelling of Ductile Damage with an Implicit Gradient-enhanced Formulation Marc G.D. Geers, Roy A.B. Engelen and Rene J.M. Ubachs Adaptive Analysis based on Error Estimation for Nonlocal Damage Models Antonio Rodriguez-Ferran, Ivan Arbos and Antonio Huerta Mathematical and Numerical Aspects of an Elasticity-based Local Approach to Fracture R.H.J. Peerlings, W.A.M. Brekelmans, M.G.D. Geers and R. de Borst Numerical Aspects of Nonlocal Damage Analyses Claudia Comi and Umberto Perego Computational Issues and Applications for 3D Anisotropic Damage Modelling: Coupling Effects of Damage and Frictional Sliding Damien Halm, Andre Dragon and Pierre Badel Energy Dissipation Regarding Transient Response of Concrete Structures: Constitutive Equations Coupling Damage and Friction Frederic Ragueneau, Jacky Mazars and Christian La Borderie Numerical Analysis of Failure in Sheet Metal Forming with Experimental Validation Michel Brunet, Fabrice Morestin and Helene Walter
vii
2.
19
3.
39
4.
55
5. 6.
75
93
7.
111
8.
127
vi
Numerical Modelling in Damage Mechanics
9.
Damage in Sheet Metal Forming: Prediction of Necking Phenomenon Nathalie Boudeau, Arnaud Lejeune and Jean-Claude Gelin
147
10. Anisotropic Damage Applied to Numerical Ductile Rupture Patrick Croix, Franck Lauro, Jerome Oudin and Jens Christlein 11. Numerical Aspects of Finite Elastoplasticity with Isotropic Ductile Damage for Metal Forming Khemais Saanouni, Abdelhakim Cherouat and Youssef Hammi 12. 3D Nonlocal Simulation of Ductile Crack Growth: A Numerical Realization Herbert Baaser and Dietmar Gross 13. On the Theory and Computation of Anisotropic Damage at Large Strains Andreas Menzel and Paul Steinmann 14. On the Numerical Implementation of a Finite Strain Anisotropic Damage Model based upon the Logarithmic Rate Otto Timme Bruhns and Christian Ndzi Bongmba 15. Ductile Rupture of Aluminium Sheet Materials Jacques Besson, Wolfgang Brocks, Olivier Chabanet and Dirk Steglich 16. On Identification of Small Defects by Vibration Tests Yitshak M. Ram and George Z. Voyiadjis 17. Multi-scale Non-linear FE2 Analysis of Composite Structures: Damage and Fiber Size Effects Frederic Feyel and Jean-Louis Chaboche Index
165
183
209
225
241
259
275
291 317
Foreword
This publication contains seventeen selected papers derived from the thirty four papers presented during the Euromech Colloquium "Numerical Modelling in Damage Mechanics" held in Troyes at the University of Technology of Troyes, October 2000, with Professors JL Chaboche, K Saanoumi and P Steinmann as co-chairmen. Damage mechanics has now reached a high degree of maturity and is currently used for many different applications connected with numerical simulation techniques. Many attempts have been made to build efficient numerical tools for damage initiation and growth simulation in mechanical structures under both small and large deformation hypotheses. The objectives of the colloquium published here were to gather recent advances in numerical and computational aspects of damage mechanics, and also to stimulate current research and future challenges in this field. We invite the reader to make his own explorations. Simply, we hope that the reader will find this publication of much intertest and a stimulus to further research. Finally, we should like to thank all the contributors to the Euromech 417 colloquium, on whose proceedings this publication is based.
Khemais Saanouni GSM/LASMIS Universite de Technologie de Troyes
Chapter 1
Some Aspects of a Gradient Damage Formulation

Tina Liebe and Paul Steinmann
Chair for Applied Mechanics, Department of Mechanical Engineering, University of Kaiserslautern, Germany
Ahmed Benallal
Laboratoire de Mecanique et Technologie, LMT-Cachan, Cachan, France
Aspects of a Gradient Damage Formulation 3 1. Introduction Softening at the continuum level due to damage accumulation mimics deterioration processes within the material at the micro scale. As a consequence of softening, damage tends to accumulate within narrow bands, so called localized zones. In experiments these localization zones display a finite width which is related to the micro structure of the material. Upon further loading localized zones then most often form a precursor to the final rupture of the material. In a standard continuum description and in particular in the corresponding numerical solution schemes no finite width is obtained, instead pathologically mesh dependent solutions are observed upon refinement of the discretization. Among the most effective remedies against the unphysical behavior displayed by a softening standard continuum and its numerical computation nonstandard continuum theories have been proposed which incorporate higher gradients of those quantities which are responsible for softening. Physically motivated gradient models in crystal plasticity were proposed, e.g. by Steinmann [STE 96] and Menzel & Steinmann [MEN 00]. Gradient dependent models, whereby the gradient dependence is essentially incorporated in the loading surface by the Laplacian of an internal variable, were treated by e.g. Comi [COM 96], de Borst, Benallal & Heeres [BOR 96a], Benallal & Tvergaard [BEN 95]. The well-posed initial boundary value problem for a continuum model was studied by Benallal, Billardon & Geymonat [BEN 93]. A variety of numerical strategies, different from the one proposed in this work, were investigated e.g. by Sluys, de Borst & Muhlhaus [SLU 93], Parnin [PAM 94], de Borst & Pamin [BOR 96b], Peerlings et. al [PEE 96], Steinmann [STE 99], Comi [COM 99]. In this contribution the essential ingredient of gradient damage is an additional equation represented by the damage condition containing the quasi-nonlocal energy release rate. A noteworthy feature from the numerical point of view is thus the treatment of the damage field as an independent variable.
2. A gradient damage formulation As a simple phenomenological measure of micro defect interactions we might consider the gradient of the damage field d = Vzd, which we include in the free Helmholtz energy \I> = \&(d, c,d) of the standard local damage model. Moreover, the model is based on a dissipation potential and the postulate of maximum dissipation. Therefore healing processes are excluded and a thermodynamically consistent approach is envisioned. Thereby, due to the extension of the classical local theory with the damage gradient contribution, the local dissipation inequality T> = Yd + Y d + P > 0 for the whole body B incorporates the nonlocality residual P, which, according to the arguments by Polizzotto & Borino [POL 98] satisfies the insulation condition fBd Pd V = 0 for the actively damaged part of the whole body
4 Numerical Modelling in Damage Mechanics
(1) Free Energy, (2) Dissipation Inequality, (3) Insulation Condition of Nonlocality Residual, (4) Macroscopic Stress, (5) Energy Release Rate, (6) Damage Flux, (7) Quasi-Nonlocal Energy Release Rate, (8) Constitutive Boundary Conditions, (9) Continuity Boundary Conditions, (10) Damage Condition, (11) Kuhn-Tucker Conditions Table 1. Key Ingredients of Gradient Damage Bd C B. Thereby, the assumption of a bilinear form for the dissipation power T> = Yd determines the quasi-nonlocal energy release rate Y = Y (e, d, d) as conjugated to the evolution of the independent arbitrary damage variable field in Bd C B. Moreover, applying the insulation condition, integration by parts and invoking Gauss theorem on the nonlocality residual yields a constitutive boundary condition (homogeneous Neumann b.c.) on dBdxt C dB for the vector field Y Y (d} which is thermodynamically conjugated to the gradient of the damage variable d and which we tend to denote as the damage flux. In addition to that, it results also in the so-called continuity boundary condition d = 0, which is imposed on dBfnt with dBd = dBdxt U dBfnt. Thus, compatibility between the evolution of the damage variable and its gradient is automatically assured. The quasi-nonlocal energy release rate essentially contains the divergence of the damage flux divY in addition to the local energy release rate Y. Finally it can be stated that the damage condition and the Kuhn-Tucker conditions retain the same structure as for the local case. Therefore, we end up with a coupled problem for the two primary unknown fields x and d which have to satisfy a partial differential equation and an inequality constraint simultaneously, as will be shown in the sequel. The key ingredients of our gradient damage formulation are summarized in Table 1.
Aspects of a Gradient Damage Formulation 5 3. Strong form of the coupled problem To set the stage for the following developments we first summarize the pertinent set of equations for the solution of the coupled boundary value problem in strong form. Let B denote the configuration occupied by a solid body. Then the displacement field u = u(x) and the damage field d = d(x) are parameterized in terms of the placements x B. These two primary fields are determined by the simultaneous solution of a partial differential equation and a set of Kuhn-Tucker-complementary conditions. The boundary dB to B with outward normal n is subdivided into disjoint parts whereby either Neumann or Dirichlet boundary conditions for the two solution fields u(x) and d ( x ) are prescribed. The residua of the resulting coupled problem in strong form are displayed in Table 2.
(1) Balance of Linear Momentum, (2) Kuhn-Tucker Conditions, (3) Additional Completeness and Non-Overlapping Requirements, (4) Elastic Solution Domain, (5) Damaged Solution Domain Table 2. Strong Form of the Coupled Problem
4. Weak form of the coupled problem As a prerequisite for a finite element discretization the coupled nonlinear boundary value problem has to be reformulated in weak form. Therefore, the equations in strong form are tested by the corresponding virtual quantities to render the virtual work expression, see Table 3. Note that the decomposition of the solution domain B into an active damaged and an inactive elastic domain B = Bd U Be and 0 = Be n Bd is indeed a quite implicit definition at this stage since one has to test for all possible combinations of
(1) Weak Form of Balance of Linear Momentum, (2) Weak Form of Kuhn-Tuck Conditions, (3) Additional Completeness and Non-Overlapping Requirements, ( Elastic Solution Domain, (5) Damaged Solution Domain Table 3. Weak Form of the Coupled Problem
supports with all admissible test functions. Furthermore, it is remarkable that the above decomposition corresponds to the pointwise complementary condition d(p = 0.
5. Discretized form of the coupled problem The above set of equations has to be discretized in time and space, whereby we apply without loss of generality the implicit Euler backward method and resort to the standard Bubnov-Galerkin finite element method. Then the temporal integration of the primary variables u and d renders a discretized temporal update for the values un+1 and dn+1. Furthermore, based on the iso-parametric concept, the displacement field uh \Bf - Y^kNxuk HI(&) together with its variation (W11Be = fc A^foi* e Hi(B) is elementwise expanded in terms of the nodal values Uk and 6uk by the same shape functions as the geometry xh\sf = ^k N%Xk. Moreover, the damage field dh\Be = ^kNddk e HI(&) together with its variation 6dh\Be = ^kN^6dk Hi (B) is elementwise expanded by independent shape functions in terms of the nodal values dk and Sdk- Likewise, the test function 5<p\Be = Xlfc^d^fc I/2(#) is discretized by the same shape functions as for the damage field in terms of nodal
Aspects of a Gradient Damage Formulation
values 6ipk. The corresponding discrete algorithmic equations of the coupled problem are given in Table 4.
(1) Discrete Algorithmic Balance of Linear Momentum, (2) Discrete Algorithmic Kuhn-Tucker Conditions, (3) Additional Completeness and Non-Overlapping Conditions, (4) Discrete Algorithmic Elastic Solution Domain, (5) Discrete Algorithmic Damaged Solution Domain Table 4. Discretized Form of the Coupled Problem Note that now the discrete algorithmic decomposition of the node point set is indeed a complete explicit definition since one only has to check separately all node points K G B. Furthermore, it is remarkable that the above discrete algorithmic decomposition corresponds to the discrete algorithmic complementary condition &RdKR% = 0 Vtf in B. The initially unknown decomposition of the discretization node point set into active and inactive subsets B = B+I U B^+1 at time step tn+i is determined iteratively by an active set search. Thereby, the strategy is borrowed from convex nonlinear programming as is frequently used e.g. in multi-surface and crystal plasticity.
8 Numerical Modelling in Damage Mechanics An efficient algorithm for the solution of the coupled problem stated in the above sections is offered by a monolithic iterative strategy. Here, the discrete algorithmic balance of linear momentum together with the discrete algorithmic Kuhn-Tucker conditions are solved simultaneously within a typical Newton-Raphson scheme.
6. Constitutive update Typically, a strain-driven constitutive update algorithm has to provide the updated dependent variables, like stress, damage flux, etc at time tw+1. Moreover, its consistent linearization is essential in order to set up the appropriate global iteration matrix for the quadratically converging global Newton-Raphson strategy. The constitutive update of the simplest geometrically linear damage prototype model for given e n+ i, dn+\ is summarized in Table 5. Note that despite its implicit character the constitutive update does not rely on local iterations usually employed in standard update algorithms.
(1) Nominal Stress, (2) Effective Stress, (3) Damage Flux, (4) Local Energy Release Rate, (5) Internal Variable Update Table 5. Update Algorithm for Gradient Damage
Note that the damage variable d is a given input for the update of the internal variable K. Thereby, for convenience of exposition we use here a simple exponentialtype evolution law for the damage evolution, which allows a closed form update for the internal variable K. Otherwise, an additional local iteration for K = 0"1 (d) would become necessary but does not limit the generality of the formulation proposed here. It is remarkable that the linearization of the constitutive update, i.e. the tangent operator results in a symmetric global iteration matrix. Note that the update algorithm in the local case varies significantly. Here, only the strains en+1 are given and in a first step the local energy release rate Yn+1 is computed. Based on this the history variable kn+1 is determined from the maximum of the new
local energy release rate Yn+1, the old value Kn or the initial damage threshold KQ, respectively. Finally, the updated damage variable dn+1 is computed from the new history variable KN+1. Thus in contrast to the gradient update algorithm the damage variable d is a dependent variable in the local case.
7. Examples In the above sections the theory as well as the numerics were outlined for a gradient damage formulation. This is now applied to computational examples showing the performance of the elaborated model by modifying the gradient parameter as well as discretization density in deterioration processes. 7.1. One-dimensional bar under uniaxial tension
Figure 1. 1-D-Model Problem: Bar under Uniaxial Tension
As a model problem we will examine in the sequel the bar in Fig. 1 under uniaxial tension for the sake of demonstration. The problem statement, which includes a slight material imperfection in the middle of the bar, is taken from Feedings, de Borst, Brekelmans & de Vree [PEE 96], whereby homogeneous Neumann boundary conditions for the damage flux were prescribed at the boundary. The material is modeled based on a linear elastic gradient damage formulation with a simple exponential-type evolution law for the damage evolution. The material parameters for the following examples are summarized in Table 6. The total bar is discretized with 80, 160, 320, 640, 1280 and 2560 elements. Thereby, due to the symmetry in the problem statement only one half of the bar is considered. The load is applied using arclength control enabling one to trace the post-peak branch of the load-deflection curves. The main objective is to show the performance of the gradient model. Therefore as a comparison, the local model is also addressed. For different possibilities of discretization techniques for the local and quasi-local case we refer to Liebe and Steinmann [LIE 01]. Likewise, a two-field finite element formulation for elasticity coupled to local damage was proposed by Florez-Lopez et. al [FLO 94]. In this work, we focus on the classical approach in local damage with
(1) Elastic Modulus, (2) Reduced Elastic Modulus, (3) Initial Damage Threshold, (4) Exponential Hardening Modulus Table 6. Material Parameters
linear element expansions for the displacement. Hereby, the local damage variable field is not separately discretized. The element type for gradient damage reflects a continuous expansion in both the displacement as well as the damage variable field. Hereby, it appears that the choice of linear expansions in both discretized fields yields the most effective and efficient results. This can be explained by considering the discretized Kuhn-Tucker conditions, which seem to be mainly affected by the choice of discretization order. Using quadratic expansions for the displacements renders piecewise linear strains and would result in a quadratic expansion of the elastically stored energy Y. This quantity would then be coupled with a highly nonlinear history variable expression AC and a piecewise constant damage gradient, which causes oscillations in both the damage variable distribution as well as in the load-deflection curves. Therefore, we use linear-linear approximations (P1P1) for the following examples in gradient damage, which give stable results. The different element formulations are described in Table 7.
Damage Formulation local gradient
Discretization Variable u u, d
Continuity of Approximation
Element Type P 1 P1P1 Expansion Expansion
C CQ/C
Table 7. Classification of Element Formulations
Firstly, as a reference for the gradient model we investigate the local damage case. Here, in order to trigger localization we additionally introduced a graded imperfection in the middle of the bar. Hereby, the first element has the lowest elastic modulus and the neighboring elements a slightly increased elastic modulus Eg = 9500.0-/V/mm2 compared to the rest of the bar elements with the highest elastic modulus. The re-
11
suiting load-deflection curves for the classical local Pl element type are displayed in Fig. 2. The typical deficiency in terms of a quasi-lack of convergence in the postpeak branch of the curves can be observed upon mesh refinement. This is even more emphasized in Fig. 3 depicting the corresponding distribution of the damage variable, whereby a concentration of damage evolution is accumulated in only one element.
Figure 2. Load versus Deflection
(Pl)
Secondly, to overcome the lack of discretization invariance the following examples are based on the incorporation of the gradient regularization in the constitutive model as described in the previous sections. First we show the quasi-mesh independence for a constant gradient parameter c = 100.0 upon mesh densification, see Fig. 4 and Fig. 5. Clearly, also for different gradient parameters the solution converges upon mesh densification. Thereby, higher values of the gradient parameter render a somewhat more ductile post-peak behavior, see Fig. 6 and Fig. 7. In any case, the corresponding distribution of the localized zone is convergent. Note that the influence of modifying the gradient parameter results in a variation of ductility in the load-deflection curves, see Fig. 6 as well as in the damage variable distribution, see Fig. 7. Hereby, the regularizing effect of the incorporation of gradients into the damage model is obvious as the jumps in the damage variable distribution in the local model are smoothed out in the gradient one. Nevertheless, the overall solution shows a shrinkage of the localized band width upon further loading into a crack line mode, i.e. a gradual transition from a damaged zone into a line crack.
Figure 3. Damage Variable Distribution (P1)
Figure 4. Load versus Deflection for c = 100.0 (P1P1)
13
Figure 5. Damage Variable Distribution for c = 100.0 (P1P1)
Figure 6. Load versus Deflection for constant mesh discretization (640 elements) and varied c = 0.0,0.1,1.0,10.0,100.0
Figure 7. Damage Variable Distribution for constant mesh discretization (640 elements) and varied c = 0.0,0.1,1.0,10.0,100.0
7.2. Two-dimensional panel under uniaxial tension Finally, in order to show the performance of the damage gradient formulation in 2d we investigate the panel in Fig. 8 under tension. Again we have included a slight material imperfection in the center elements. The material is modeled by analogy to the Id example, see Table 6. The bar is discretized with 20x10 and 40x20 Q1Q1-elements (continuous approximation of both displacement field and damage field). Again, we focus here on the damage variable distribution which emphasizes the convergent performance of the gradient damage formulation as displayed in Fig. 9 and Fig. 10.
Figure 8. 2-D-Model Problem: Panel with Center-Imperfection
15
Figure 9. Damage Distribution shortly before reaching d=I, c=100, coarse mesh
Figure 10. Damage Distribution shortly before reaching d-l, c=100, fine mesh
8. Conclusion We have derived a theoretical formulation and the corresponding discretized algorithmic format of a gradient damage model. Based on a positive domain dissipation and the postulate of maximum dissipation we end up with algorithmic Kuhn-Tucker conditions in dependence on the quasi-nonlocal energy release rate, which is conjugated to the damage evolution. On the numerical side, due to this special structure, an active set search becomes necessary for the monolithic iterative solution of the coupled problem within a typical Newton-Raphson strategy. Nevertheless only standard FE-data structures and corresponding FE-modules are involved. Moreover, we end up with a symmetric iteration matrix avoiding the use of a secant stiffness matrix as typ-
16 Numerical Modelling in Damage Mechanics ically adopted in nonlocal models. In addition to that, other gradient damage models usually result in a non-symmetric tangent operator, see e.g. Peerlings et al. [PEE 96]. Considering a model problem of an one-dimensional bar under uniaxial tension we firstly investigated the classical local element formulation with only continuous element expansions for the displacement. Here, the local theory resulted in spurious mesh dependence in particular for the damage variable distribution. This could only be remedied by using the gradient formulation with gradient parameters c > 0. For verification we investigated the behavior for c 1.0,10.0,100.0. Thereby, it could be noted that with increasing gradient parameter the solution becomes somewhat more ductile. In any case, mesh densification renders mesh objective results and convergent distributions of the damage variable field in both Id as well as 2d computations. It is remarkable that a gradual transition from a damaged zone into a line crack can be observed in the load-deflection curves in contrast to standard gradient models. Therefore, it was emphasized that based on the theory and numerics underlying the gradient model advocated here, the regularization effect in damage is considerable. Moreover, the simultaneous solution of the discrete algorithmic Kuhn-Tucker conditions in addition to the discretized algorithmic balance of linear momentum offers an elegant solution strategy in the numerical treatment of gradient damage. In particular it is notable that the additional discrete algorithmic loading and unloading conditions complemented by an active set search are implemented on a nodal basis, which is in contrast to alternative gradient-enhanced formulations.
9. References [BEN 93] BENALLAL, A., BILLARDON, R. AND GEYMONAT, G. Bifurcation and localization in rate-independent materials: some general considereations. In CISM Lecture Notes N 327, pages 1-44. Springer-Verlag, 1993. [BEN 95] BENALLAL, A. AND TVERGAARD, V. Nonlocal continuum effects on bifurcation in the plane strain tension-compression test. J. Mech. Phys. Solids, 43:741-770,1995. [BOR 96a] DE BORST, R., BENALLAL, A. AND HEERES, O. A gradient-enhanced damage approach to fracture. J. Phys. IV, 6:491-502,1996. [BOR 96b] DE BORST, R. AND PAMIN, J. Some novel developments in finite element procedures for gradient-dependent plasticitiy and finite elements. Int. J. Num. Meth. Eng., 39:2477-2505,1996. [COM 96] COMI, C. A gradient damage model for dynamic localization problems. Rend. Sc. Istituto Lombardo, A 130:119-141,1996. [COM 99] COMI, C. Computational modelling of gradient-enhanced damage in quasi-brittle materials. Mech. Cohes.-Frict. Mater., 4:17-36, 1999.
17
[FLO 94] FLOREZ-LOPEZ, J., BENALLAL, A., GEYMONAT, G. AND BILLARDON, R. A two-field finite element formulation for elasticity coupled to damage. Com/?. Meth. Appl. Mech. Eng., 114:193-212,1994. [LIE 01] LIEBE, T. AND STEINMANN, P. Theory and numerics of a thermodynamically consistent framework for geometrically linear gradient plasticity. Int. J. Num. Meth. Eng., in press 2001. [MEN 00] MENZEL, A. AND STEINMANN, P. On the continuum formulation of higher gradient plasticity for single and polycrystals. J. Mech. Phys. Solids, 48:1777-1796,2000. [PAM 94] PAMIN, J. Gradient-dependent plasticity in numerical simulation of localization phenomena. Diss., Delft University of Technology, 1994. [PEE 96] PEERLINGS, R., DE BORST, R., BREKELMANS, W. AND DE VREE, J. Gradient enhanced damage for quasi-brittle materials. Int. J. Num. Meth. Eng., 39:3391-3403,1996. [POL 98] POLIZZOTTO, C. AND BORING, G. A thermodynamics-based formulation of gradient-dependent plasticity. Eur. J. Mech. A/Solids, 17:741-761,1998. [SLU 93] SLUYS, L., DE BORST, R. AND MUEHLHAUS, H. Wave propagation, localization and dispersion in a gradient-dependent medium. Int. J. Solids Struct., 30:1153-1171,1993. [STE 96] STEINMANN, P. Views on multiplicative elastoplasticity and the continuum theory of dislocations. Int. J. Eng. Sci., 34:1717-1735,1996. [STE 99] STEINMANN, P. Formulation and computation of geometrically nonlinear gradient damage. Int. J. Num. Meth. Eng., 46:757-779,1999.
Chapter 2
On the Numerical Modelling of Ductile Damage with an Implicit Gradient-enhanced Formulation

Marc G.D. Geers, Roy A.B. Engelen and Rene J.M. Ubachs
Department of Mechanical Engineering, Eindhoven University of Technology, The Netherlands
Gradient-enhanced Ductile Damage
21
1. Introduction The numerical modelling of ductile damage and fracture in engineering materials and particularly in metals has gained considerable interest in recent years. Many of the available solution strategies focus on the initiation process, which is of great importance in manufacturing processes where any damage is to be avoided. However, there is also an important class of practical problems, in which damage, failure and crack propagation are essential steps in the production process, e.g., cutting, blanking, drilling, etc. Analysis of these processes is mainly based on phenomenological knowledge. Lengthy trial and error procedures along with a variety of empirical guidelines are used to develop and optimize the process. Nowadays, it is expected that numerical techniques may be a versatile alternative, provided that they are capable to give a reliable and adequate description of the ductile fracture process, i.e., initiation and propagation of a crack in a ductile material. It is known from the analysis of damage and fracture in the infinitesimal deformation theory, that the adequate modelling of the evolution of damage requires a higher-order continuum in which either a gradient or a nonlocal approach is used. In the context of ductile damage, explicit gradient-enhanced small deformation theories have been developed in the past ten years, see for instance [DB 92, PAM 94, SVE 97, RAM 98]. So far, only few extensions to a large deformation framework are available, e.g., [MIK 99] where explicit gradients have been incorporated in the constitutive model. Integral nonlocal approaches applied to softening plasticity have been studied in [NIL 98] on the basis of thermodynamical considerations. Steinmann [STE 99] investigated a geometrically nonlinear gradient damage formulation, applicable to rubberlike materials. In contrast to the explicit gradient enhancements used in the cited gradient plasticity models, this paper is based on the use of an implicit (and hence nonlocal) gradient formulation. The nonlocal character of this gradient formulation has been proven recently by Peerlings et al. [PEE 99], whereas nonlocal damage models have been used with considerable success by several authors [BAZ 88, TVE 95]. The small deformation solution of the implicit gradient version has now been elaborated and tested by Engelen et al. [ENG 01]. Only basic features of this solution strategy will be reviewed here. Several finite plasticity formulations nowadays exist [NAG 90, MIE 98b, MIE 98a, ALF 98]. Many models are based on a hypoelastic stress response, which is commonly obtained by generalization of the corresponding infinitesimal framework. It is known that elastic deformations need to be small in order to apply such models, since no stored energy function exists that ensures true elastic behaviour (i.e., reversible deformations without energy dissipation). This restriction does not hold for a hyperelastic stress response, where a stored energy function does exist, that depends on the invariants of the right Cauchy-Green deformation tensor for isotropic materials. On the basis of this approach, hyperelasto-plasticity frameworks have been proposed and implemented by Simo et al. [SIM 85, SIM 88a] and Simo [SIM 88b]. The volumetric and deviatoric response is fully decoupled on the level of the stored energy potential. This hyperelasto-J2-plasticity model is taken as the point of departure for the incor-
22
poration of a ductile damage evolution. A major difficulty in numerical descriptions of ductile damage is the adequate incorporation of physical mechanisms. In metals, ductile damage starts if the number of dislocation barriers prevents further plastification and leads to void initiation, growth and coalescence, the underlying damage mechanisms of the frequently applied Gurson model [GUR 77] . A lot of empirical and micromechanical research has been performed in this area, which has to embedded in the damage initiation and evolution in a later stage. This paper subsequently presents the small and large deformation framework, the incorporation of a nonlocally driven ductile damage variable in the field function, some computational aspects for the large deformation model, as well as several numerical examples for different cases. A discussion on the use of nonlocal models in the presence of large deformations is also made, where a material or a spatial framework do not lead to the same interpretation of the underlying 'material' length scale that is mostly used as a constant in the small deformation context. Note that the ductile damage parameter appears as an additional internal state variable, a concept that is widely used in internal state variable plasticity and damage [LEM 90, KRA 96]. 2. Underlying elasto-plasticity formulations 2.1. Small deformation elasto-plasticity model The elasto-plastic framework used within the infinitesimal deformation assumption is a standard isotropic von Mises elasto-plastic model. The constitutive relation for the stress rate tensor versus the elastic strain rate tensor is typically given by
The plastic state is characterized with a yield function / where aeq is the von Mises equivalent stress
and cry the yield stress. A linear or a nonlinear hardening rule may be considered, e.g.,
In here, sp is the effective plastic strain, h is the linear hardening modulus and /IQQ, ho, 8 > 0 are nonlinear hardening parameters. The effective plastic strain is defined by
Gradient-enhanced Ductile Damage 23 The plastic strain rate tensor ep is extracted from an associate flow rule, given by
It can be easily shown that the equations [2], [3], [6] and [7] lead to the following identity
The set of equations is completed with the standard Kuhn-Tucker loading/unloading relations
and the consistency condition
The solution of this elasto-plastic problem follows standard rules, which can be found in many textbooks.
2.2. Large deformation hyperelasto-plasticity model The large deformation formulation which is used in this paper is based on a rateindependent hyperelasto-plastic model presented by [SIM 88a, SIM 88b, SIM 98], in which a correction has to be made in order to comply with the assumed isochoric plastic flow. The model is presented in a format that tends towards the infinitesimal solution of the previous section for small deformations. This large deformation model presents many features which are well-known in the classical infinitesimal theory of plasticity, although it has not been obtained through ad hoc extensions of the small strain theory. It is based on the adequate implementation of finite deformation kinematics in the elastic part and through the application of general principles of associative plasticity in the plastic part. A hyperelastic stress-strain relation is used for the elastic predictor. In this section, the essential algorithmic steps are highlighted, as well as the modifications made with respect to Simo's original model. The numerical implementation then follows identical lines as given in [SIM 88b]. Based on micromechanical considerations of crystallographic slip, a multiplicative decomposition of the deformation gradient tensor F is performed according to
Such a multiplicative split implies the existence of an intermediate state which is obtained if the current state 2 is relaxed to a (local) stress-free configuration Qp,
where only plastic deformations exist. Note that Fp and Fe are only defined up to an arbitrary rigid body motion of the intermediate state. The different configurations and the well-known associated kinematic tensors (total, elastic, plastic right or left Cauchy-Green deformation tensors, stretch tensors and strain tensors) are shown in Figure 1. The following pull-back push-forward relations are then valid
Figure 1. The multiplicative split and associated deformation tensors
The elastic left Cauchy-Green deformation tensor Be is the push-forward (with F) of the inverse plastic right Cauchy-Green deformation tensor C~l. The (objective and covariant) Lie-derivative of Be is the push-forward of the material time derivative of C-1p, which also holds for their corresponding spatial and material deviatoric parts, respectively given by
25
Assuming that the plastic flow is isochoric means that the volume change ratio J depends on elastic deformations only
The kinematic constraint which follows from this assumption can be rephrased as
which is different from the facilitating assumption made by [SIM 88a], where
was forwarded. Unfortunately, satisfaction of [19] does not lead to an isochoric plastic flow. Equation [18] should have been used instead. The Kirchhoff stress tensor i is computed from the elastic deformations, using an isotropic hyperelastic stress-strain response.
where K and G equal the bulk and shear modulus respectively. The stored energy function W that corresponds to this hyperelastic relation is given by
The yield function is the classical von Mises-Huber function, formulated in terms of the Kirchhoff stress tensor
which closely resembles equation [2]. The von Mises equivalent stress teq is defined as in the infinitesimal case by
The earlier proposed linear and a nonlinear hardening rules now read
The effective plastic strain sp is now defined with respect to the elastic left CauchyGreen deformation tensor
where a is a coefficient that will be determined later in order to equate the large and small deformation models in the case of infinitesimal displacements. Simo [SIM 88a] derived an associative flow rule for this hyperelasto-plastic formulation, through the application of the principle of maximum plastic dissipation. The following deviatoric flow rule was obtained
The relation between the plastic multiplier y and the effective plastic strain ep may be extracted by combining equations [22], [23], [26] and [27]
The formulation of the model is again completed with the standard Kuhn-Tucker loading-unloading conditions and the consistency condition
The plastic incompressibility condition for infinitesimal displacements becomes
by means of which the following simplifications can be made if the limit towards th infinitesimal framework is taken
Consequently, the flow rule [27] in this infinitesimal limit case reads
Hence, if the coefficient a equals 3, all equations coincide with the infinitesimal elasto-plasticity framework. The algorithm is completed with an elastic predictor - plastic corrector scheme, see [SIM 88b, SIM 98] for a similar example. The algorithm can be implemented in a finite element framework without difficulty, including the desired consistent tangent operator.
27
3. Incorporating ductile damage The addition of ductile damage during plastic flow is based on a progressive reduction of the yield stress once the failure process initiates. Within an isotropic framework, this can be achieved with one single damage variable, in a similar way to the stiffness reduction in damage mechanics. Similar arguments as the one used by Kachanov can be used to motivate such an approach, since the initiation of microcracks (intergranular) and voids can be observed in this stage of the deformation. Based on the knowledge which has been acquired in the field of the computational modelling of material instabilities, it is now known that a continuum solution can only be obtained if the principle of local action is abandoned. A higher-order or a nonlocal constitutive theory [DB 92, PAM 94, SVE 97, RAM 98, GEE 00] has to be used, in order to obtain a set of well-posed partial differential equations. In this paper, an implicit gradient enrichment is used, which combines the computational efficiency of gradient type theories with the integral nonlocal concept. It has been shown that such an implicit approach presents a true nonlocal character [PEE 99], in the sense that an equivalent integral format exists, in which the nonlocal variable is a long-range weighted average of field of local variables. The solution strategy for the infinitesimal case will be highlighted briefly, after which the extension towards the hyperelastoplastic formulation is scrutinized. Computational details and algorithmic aspects for the infinitesimal elasto-plasticity framework can be found in [ENG 01]. Note that the inherent relation between this ductile damage variable and its associated thermodynamic kinematic quantity is not further considered here, where the nonlocal character of the damage requires special attention, see [NIL 98, GB 99, GAN 99].
3.1. Small deformations Assuming a fully isotropic material behaviour, a ductile damage parameter 0 ^ o)p < 1 is introduced, which leads to a gradual reduction of the yield stress in the softening stage. Void nucleation, void growth and coalescence, growing out to cracks are the underlying physical mechanisms. The yield function given in [2] is now transformed to the following damage-sensitive yield function
Evidently, damage will affect the stress tensor, but unlike damage mechanics the ductile damage does not enter the hyperelastic constitutive relation and deformations are thus not reversible. The ductile damage w>p is computed from a history variable K, which is the ultimate value of the nonlocal variable ifs in the deformation history of the considered material point. The 'implicit gradient' approach refers to the equation that is used to extract the nonlocal field variables ijr from the field of their local counterparts ty. This is done through the solution of a partial differential equation of the Helmholtz type
28
along with the Neumann boundary condition V\jr-n = 0 on the boundary F with outward normal n [35]
The parameter l is a length parameter (often called the intrinsic length) that is related to the size of the influence zone of the nonlocal averaging function. Determining ^ from such a gradient formulation is equivalent to an integral nonlocal format, see [BAZ 88], where the nonlocal variable is computed as a weighted average of the local field. Note that the Laplacian used in the nonlocality equation [34] can be defined with respect to the undeformed or deformed state. In the case of infinitesimal displacements however, this difference is not relevant. The relation between the ductile damage and the nonlocal field variable is governed by a damage evolution law U)P(K) which quantifies the damage growth in terms of the field of kinematic variables. Phenomenological examples of such laws are
where the damage grows respectively linearly or exponentially (initially fast increase) or towards its ultimate value 1 at failure. The parameters /J, /c, and KC are material parameters. These equations only influence the evolution of damage once it has initiated. The initiation of damage is controlled by the proper scalar function for the local field variable \js. Its choice should be founded on micromechanical considerations and known mechanisms in ductile damage initiation and evolution, which is still subject of future research. In this contribution, the damage controlling field variable \fr is taken equal to the effective plastic strain measure ep, i.e.,
In the case of infinitesimal displacements, the constitutive equations for elastoplasticity enriched with this ductile damage approach have been implemented in a computational strategy and solved within a finite element framework [ENG 01]. It is shown that such a formulation is well suited to solve softening and failure in ductile materials up to complete failure.
3.2. Large deformations 3.2.1. Material and spatial nonlocality Extending the damage evolution proposed in the infinitesimal case to large deformations raises a fundamental problem with respect to the nature of the nonlocality. This is easily understood from the Helmholtz equation [34], which may take two possible
Gradient-enhanced Ductile Damage 29 formats in the geometrically nonlinear case, i.e.,
The first equation corresponds to the Lagrangian (or material) averaging of the local field, while the second equation reflects the Eulerian (or spatial) averaging case. Steinmann [STE 99] already addressed this difference in his analysis for large deformation quasi-brittle damage models. He showed that only the Lagrangian averaging solution seemed to inherit the properties of the infinitesimal model. Equation [40] can be pulled back towards the Lagrangian configuration, which yields
Besides a symmetric term that depends inversely on the deformation, a second non-symmetric term appears. Clearly, the relative influence of both terms gradually changes with the deformation, which may be particularly important in localization zones. This is well illustrated with the one-dimensional counterpart of [41].
where X is the material coordinate, jc the spatial coordinate and A the stretch ratio. In localization zones |^ becomes large, while the influence of A. varies from traction to compression. The coefficients are small in tension and large in compression. In the case of locally uniform stretch, the non-symmetric term locally equals zero and the nonlocal averaging equation can be written as
The material length parameter Q = tfk now depends on the deformation, which is in contrast to the one-dimensional Lagrangian equation derived from [39]
where the material length parameter l0 is constant. Nonlocal spatial averaging (one averages over distances that are fixed in space and independent of the deformation of the material) is thus performed over a volume that is constant in space but variable in the material, while nonlocal material averaging (one averages over the initial undeformed distances between the material particles or voids) is applied over a constant material volume. In the case of tension (A. > 1, typically the case with tensile cracks) the material volume over which the spatially nonlocal kernel acts vanishes, while it becomes extremely large in the case of compression (0 < X < 1).
A spatial nonlocal formulation [40] with a constant spatial averaging volume leads to a nonlocal variable that tends to the local variable with increasing tensile deformations. It is also clear that spatial and material nonlocal formulations will behave quite differently in tension or compression. Note that equation [34] has already been used with a non-constant length scale within an infinitesimal damage-mechanics framework, see [GEE 98, GEE 00]. It was shown that a length scale that decreases with deformation leads to a solution that tends towards the local ill-posed solution upon complete damage. In fact, local deformations in the neighbourhood of cracks will always tend to large values, while they tend to zero in the surrounding unloading material. Although material nonlocality seems more relevant than spatial nonlocality, there is no obvious reason why the nonlocal kernel should be independent of the deformation. The non-trivial answer to this question should ensue from micromechanics or physics and certainly constitutes a challenge for future research. 3.2.2. Computational predictor-corrector algorithm The solution of the damage-enhanced material behaviour is performed with an elastic predictor - plastic corrector scheme during a time increment from t to t + At. The predictor corresponds to a fully elastic increment Af, i.e., the plastic flow increment is zero. The intermediate stress free configuration is thus preserved in that state.
The small star symbol is used to indicate that the considered quantity corresponds with the predictor state. The incremental deformation tensor FA [Ft+^']-[F']~l, quantifies the deformation at time t + Af with respect to the configuration at time t. The predictor of the isochoric left Cauchy-Green deformation tensor is then obtained through
and hence the trial Kirchhoff stress tensor as
Note that the ductile damage does not evolve in the predictor phase, since the hyperelastic relation is independent from wp. The next step consists in the integration of the flow rule, which is performed with a pull-back and push-forward procedure in order to ensure incremental objectivity. The objective Eulerian flow rule is pulled-back to the invariant Lagrangian configuration, after which a time discretization is carried out using an implicit Euler backward scheme. The discretized Lagrangian flow rule is then pushed forward to the spatial
Gradient-enhanced Ductile Damage 31 description again. Following this procedure with respect to [27], followed by a multiplication with [jt+At]~2/3 and substitution of the elastic predictor [46] permits one to rephrase the discretized flow rule in the current state as
which may also be written in the following format
Repeating the pull-back/integration/push-forward scheme to rate equation [28] gives for a = 3
The increment of the plastic flow Ay may now be determined from equation [49] by taking the square root of the double inner product of this equation with itself, and making use of the definition of the equivalent Kirchhoff stress req and the deviatoric part of r in equation [20], which yields
Using a linear hardening law[24] enriched with ductile damage, i.e., the form ry = (TyO + /ie p )(l o)p), with a constant hardening modulus h, and making use of the trial value / = req (Tyo + he'p)(l a)'p) of the yield function permits to rewrite equation [51] and extract the plastic multiplier by means of [50]
For the nonlinear hardening rule [25], equation [51] has to be solved iteratively for Ay with a local Newton scheme. Once Ay is determined, the equivalent von Mises stress can be extracted directly from [51]. Alternatively, it can be noticed from the equations [49] and [51] that Nd = *Nd, which permits one to compute [-Bg]t+At and thus Tt+At directly from equation [48]. The spherical part of the flow rule is fully determined through the isochoric plastic flow assumption. Plastic incompressibility is enforced by computing the spherical part such that det(fle) = J2 holds. Using the principal invariants jf = 0, J%, J$ of Bde, which can be computed after the return mapping of the deviatoric part, it is easy to show that the trace of Be is the solution of the following cubic polynomial equation
The solution tr (Be) of this equation, and the result of the return mapping scheme in terms of Bde, permits the determination of the complete elastic left Cauchy-Green deformation tensor. Note that the elastic deformation tensor Fe is not computed explicitly in this algorithm, which means that the unknown rotation tensor, up to which the intermediate configuration is defined, does not have to be quantified. The solution strategy presented above can be implemented in a finite element framework without great difficulty, since it is a natural combination of the small deformation framework [ENG 01] and Simo's work [SIM 88a, SIM 88b]. In any case, i.e., material or spatial nonlocality, a consistent tangent operator can be determined.
4. Examples and comparisons 4.1. Small deformation analysis Several examples for small deformation gradient-enhanced ductile damage were already given in [ENG 01]. Only one example is therefore presented here. A twodimensional plate is loaded in compression as indicated in Figure 2, where the material parameters are also given. The problem has been investigated with several discretiz-
Figure 2. Compression of a plate with an initial imperfection
ations, where the expected mesh-objectivity of the result has been confirmed. The development of the shear bands and the localization process in the softening branch is illustrated in Figure 3, where the effective plastic strain is depicted. It can be noticed that the softening branch is modelled up to complete failure, where the deformation
33
Figure 3. Shear band development and localization in a softening plate
localizes in a zone that expands to certain width determined by l, after which the further localization takes place in a smaller band that narrows progressively.
4.2. Large deformation analysis A first trivial case, that validates the assumptions made in the infinitesimal framework and its extensions to the geometrically nonlinear framework, is found by comparing the plate compression example for both frameworks. The associated force displacement curves are depicted in Figure 4. It can be noticed that results almost overlap, which is essentially due to the fact that deformations remain small in this example. If applications towards metals are envisaged, deformations inevitably become much larger. To illustrate this, an axisymmetric tensile bar has been modelled, for which the characteristics are listed in Table 1 (geometry was taken from Simo [SIM 88b]). No imperfection was used, since physical softening will be triggered automatically by the cross-section reduction in the necking area. The tensile bar fails due to geometrical and physical softening in the necking zone. The analysis was again made with both the infinitesimal and the large deformation framework. Results are shown in Figure 5. Large differences now appear, soon after the initial yield point. Note that this strongly influences the localization of deformation and the softening behaviour of the material in the failure stage.
34
Figure 4. Plate compression solved with the infinitesimal and geometrically nonlinear approach
Length L R Radius K Bulk modulus Shear modulus G Initial yield stress TyO Residual flow stress fyoo h Hardening modulus Initial Ki (wp = 0) Ki Critical KC (wp =.;1) KC Softening slope parameter ft t Length scale
53.3 mm 6.4mm 164 GPa SOGPa 450 MPa 715 MPa 129 MPa 5% 150% 1 10mm
Table 1. Characteristics of the axisymmetric tensile bar
35
Figure 5. Axisymmetric tensile bar under failure
The same example was used to make a comparison between material and spatial averaging. The force-displacement curve is given in Figure 6. In contrast to the observations made by Steinmann [STE 99], differences are rather small in elasto-plasticity. This is mainly due to the fact that deformations are irreversible here. Furthermore, the pull-back analysis of the spatial averaging solution points out that differences can be expected if deformations are very large, i.e., at the end of the crack initiation process which is at the very end of the failure stage of a tensile bar. No singularities are present in this example and no crack propagation occurs, which means that these results may not be generalized ad hoc. Future work will undoubtedly clarify this point.
5. Conclusions A small and a large deformation elasto-plasticity framework, enhanced with an isotropic ductile damage variable, has been presented. The solution strategy has been emphasized, where it has been shown that this formulation is particularly well suited to model damage initiation and evolution in real engineering problems. In spite of the phenomenological character of the framework, it overcomes the well-known problems in continuum modelling of damage which were not solved at this level yet for large deformation elasto-plastic behaviour. Future research must address issues like damage initiation and evolution in terms of the complex deformation history (including the
Figure 6. Axisymmetric tensile bar with material/spatial nonlocality
influence of hydrostatic pressure, micromechanical theories, e.g., Gurson, etc.), as well the use of advanced remeshing techniques that allow the transition of smooth damage zones into discrete cracks.
6. References
[ALF 98] ALFANO G., ROSATI L., VALOROSO N., "A displacement-like finite element model for J2 elastoplasticity: variational formulation and finite-step solution", Computer Methods in Applied Mechanics and Engineering, vol. 155, 1998, pp. 325-358. [BAZ 88] BAANT Z. P., PIJAUDIER-CABOT G., "Nonlocal continuum damage, localization instability and convergence", Journal of Applied Mechanics, vol. 55, 1988, pp. 287-293. [DB 92] DE BORST R., MUHLHAUS H. B., "Gradient-dependent plasticity: Formulation and algorithmic aspects", International Journal for Numerical Methods in Engineering, vol. 35, 1992, pp. 521-539. [ENG01] ENGELEN R. A. B., GEERS M. G. D., BAAIJENS F.P.T., "Nonlocal implicit gradient-enhanced softening plasticity", International Journal of Plasticity. Accepted. [GAN 99] GANGHOFFER J.F., SLUYS L.J., DE BORST R., "A reappraisal of nonlocal mechanics", European Journal of Mechanics - A/Solids, vol. 18 1-2, 1999, pp. 1746.
37
[GB 99] G. BORING P. FUSCHI, POLIZZOTTO C., 'Thermodynamic approach to nonlocal plasticity related variational principles", Journal of Applied Mechanics, vol. 66 4, 1999, pp. 952-963. [GEE 98] GEERS M. G. D., DE BORST R., BREKELMANS W. A. M., PEERLINGS R. H. J., "Strain-based transient-gradient damage model for failure analyses", Computer Methods in Applied Mechanics and Engineering, vol. 160 1-2, 1998, pp. 133-154. [GEE 00] GEERS M. G. D., PEERLINGS R. H. J., BREKELMANS W. A. M., DE BORST R., "A comparison of phenomenological nonlocal approaches based on implicit gradientenhanced damage", Acta Mechanica. Accepted. [GUR77] GURSON A.L., "Continuum theory of ductile rupture by void nucleation and growth: Part 1 - yield criterion and flow rules for porous ductile media", Transactions of ASME Journal of Engineering Materials Technology, vol. 99 1, 1977, pp. 2-15. [KRA 96] KRAUSZ A.S., KRAUSZ K., eds., Unified Constitutive laws of plastic deformation, San Diego, Academic Press, 1996. [LEM 90] LEMAITRE J., CHABOCHE J.L., Mechanics of Solid Materials, Cambridge University Press, 1990. [MIE 98a] MlEHE C., "A constitutive frame for elastoplasticity at large strains based on the notion of a plastic metric", International Journal of Solids and Structures, vol. 30, 1998, pp. 3859-3897. [MIE 98b] MlEHE C., "A formulation of finite elastoplasticity based on dual co- and contravariant eigenvector triads normalized with respect to a plastic metric", Computer Methods in Applied Mechanics and Engineering, vol. 159, 1998, pp. 223-260. [MIK 99] MiKKELSEN L. P., "Necking in rectangular tensile bars approximated by a 2-d gradient dependent plasticity model", European Journal of Mechanics A/Solids, vol. 18, 1999, pp. 805-818. [NAG 90] NAGHDI P. M., "A critical review of the state of finite plasticity", Journal of Applied Mathematics and Physics, vol. 41, 1990, pp. 315-387. [NIL 98] NlLSSON C., "On nonlocal rate-independent plasticity", International Journal of Plasticity, vol. 14 6, 1998, pp. 551-575. [PAM 94] PAMIN J., Gradient-dependent plasticity in numerical simulation of localization phenomena, Ph.D. thesis, Delft University of Technology, Delft, 1994. [PEE 99] PEERLINGS R. H. J., GEERS M. G. D., DE BORST R., BREKELMANS W. A. M., "A critical comparison of nonlocal and gradient-enhanced continua", International Journal of Solids and Structures. Submitted. [RAM 98] RAMASWAMY S., AVARAS N., "Finite element implementation of gradient plasticity models. part ii: Gradient-dependent evolution equations", Computer Methods in Applied Mechanics and Engineering, vol. 163, 1998, pp. 33-53. [SIM 85] SlMO J. C., ORTIZ M., "A unified approach to finite deformation plasticity based on the use of hyperelastic constitutive equations", Computer Methods in Applied Mechanics and Engineering, vol. 49, 1985, pp. 221-245.
38
[SIM88a] SlMO J. C., "A framework for finite strain elastoplasticity based on maximum plastic dissipation and the multiplicative decomposition: Part i. continuum formulation", Computer Methods in Applied Mechanics and Engineering, vol. 66, 1988, pp. 199-219. [SIM885] SlMO J. C., "A framework for finite strain elastoplasticity based on maximum plastic dissipation and the multiplicative decomposition: Part ii. computational aspects", Computer Methods in Applied Mechanics and Engineering, vol. 68, 1988, pp. 1-31. [SIM 98] SlMO J. C., HUGHES T. J. R., Computational Inelasticity, Interdisciplinary Applied Mathematics, Springer-Verlag, 1998. [STE 99] STEINMANN P., "Formulation and computation of geometrically non-linear gradient damage", International Journal for Numerical Methods in Engineering, vol. 46, 1999, pp. 757-779. [SVE 97] SVEDBERG T., RUNESSON K., "A thermodynamically consistent theory of gradientregularized plasticity coupled to damage", International Journal of Plasticity, vol. 13 6-7, 1997, pp. 669-696. [TVE 95] TVERGAARD V., NEEDLEMAN A., "Effects of nonlocal damage in porous plastic solids", International Journal of Solids and Structures, vol. 32 8-9, 1995, pp. 1063-1077.
Chapter 3
Adaptive Analysis based on Error Estimation for Nonlocal Damage Models

Antonio Rodriguez-Ferran, Ivan Arbos and Antonio Huerta
Departament de Matematica Aplicada III, Universitat Politecnica de Catalunya, Barcelona, Spain
Error Estimation for Nonlocal Damage Models
41
1. Introduction Damage models are nowadays a standard approach to model the failure of concrete and other quasi-brittle materials [LEM 90]. To avoid the pathological mesh dependence of numerical simulations carried out with local models, various regularisation techniques may be used [BOR 93]. One possibility, considered in this work, is the use of nonlocal damage models [PIJ 87, BA2 88, MAZ 89]. The basic idea of nonlocal models is that the damage parameter that describes the loss of stiffness depends on the strain state in a neighbourhood (associated to a characteristic length) of the point under consideration. To ensure the quality of the finite element solution, an adaptive strategy based on error estimation was recently proposed by the authors [ROD 00]. A salient feature of the approach proposed in that reference is the extension of an existing residualtype nonlinear error estimator [DIE 98, HUE 00] to the context of nonlocal damage models, where tangent stiffness matrices are not readily available. Attention is focused here on the fact that the error estimator is based on local computations over elements and so-called patches. It will be shown that it is important to account for the nonlocality of the damage model when solving these local problems. These leads to a slight modification of the nonlocal damage model to be used during error estimation. An outline of paper follows. Nonlocal damage models are briefly reviewed in section 2. Section 3 is devoted to the error estimator. After reviewing its main ingredients in section 3.1, section 3.2 discusses the required modification of the damage model so that its nonlocality is taken into account during error estimation. The resulting adaptive strategy is illustrated in section 4 by means of two numerical examples involving the single-edge notched beam test. The concluding remarks of section 5 close the paper. 2. Nonlocal damage models The basic features of nonlocal damage models are briefly reviewed in this section. For the sake of clarity, only isotropic elastic-damage models are considered. These simple models are sufficient to illustrate how the error estimator based on local computations takes into account the nonlocality of the model. However, the approach can be extended to more complex nonlocal damage models, incorporating, for instance, anisotropy and/or coupling with plasticity [MAZ 89]. The loss of stiffness associated with mechanical degradation of the material is represented by a parameter D, according to
where a and e are respectively the Cauchy stress tensor and the small strain tensor, and C is the tensor of elastic moduli (E: Young's modulus; v: Poisson's coefficient).
Parameter D ranges between 0 (virgin material, with elastic stiffness) and 1 (completely damaged material, with no stiffness). For computational purposes, an upper bound of .Dmax = 0.99999 is set. In this manner, zero stiffness is avoided and it is not necessary to remove the fully damaged elements from the mesh. It is assumed that D depends on a state variable Y, which in turn depends on the strains: The basic idea of nonlocal damage models is averaging the state variable Y in the neighbourhood of each point. In this manner, the nonlocal state variable Y is obtained:
The weight function a, which depends on the distance d to the point under consideration, is typically the Gaussian
where the characteristic length lc is a material parameter of the nonlocal damage model, which acts as a localization limiter and can be associated to the grain size [PIJ 91]. The nonlocal state variable drives the evolution of damage,
Damage starts above a threshold Y0 (that is, D = 0 for Y < YQ) and it cannot decrease (that is, D > 0). To define a particular model, it is necessary to specify the definition of the state variable, equation [2], and the evolution law for damage, equation [5]. In the modified von Mises model [VRE 95] Y depends on the first strain invariant /i, the second deviatoric strain invariant J-2 and the ratio k of compressive strength to tensile strength. Regarding the damage evolution for Y > YQ, an exponential law [PIJ 91, ASK 00] is used. The modified von Mises model is summarized in Table 1. More details about this model can be found in [VRE 95, PEE 98]. 3. The error estimator In order to control the finite element discretization errors, an adaptive strategy is employed [ROD 00]. It is based on the combination of a residual-type error estimator [DIE 98, HUE 00] and /i-remeshing. The error distribution is computed with the error estimator and translated into desired element sizes with a so-called optimality criterion [DIE 99]. An unstructured quadrilateral mesh generator [SAROO] is then used to build a mesh with the desired sizes. This iterative process stops (typically after 2 to 4 iterations) when the relative error of the solution (i.e. energy norm of the error divided by energy norm of the solution) is below a prescribed threshold set a priori.
43
State variable: Damage evolution:
Table 1. Modified von Mises model 3.1. A residual-type error estimator based on local computations The error estimator used in this work was first developed for linear problems [DIE 98] and later extended to nonlinear problems [HUE 00, DIE 00]. A detailed presentation and analysis can be found in these references. Here, only a brief review is given. Using a mesh of characteristic size H, the finite element method provides the discrete nonlinear equilibrium equation
where the unknown is the nodal displacement vector uH, f int H (u H ) is the vector of nodal internal forces associated with uH and fextH is the discretized external force term. To estimate the error in uH, a finer mesh of size h (h <g. H) is used as reference. On this finer mesh, the problem reads
The error in displacements is defined as the difference between the two solutions:
Note, however, that computing u^ is computationally much more expensive than computing uH, because it involves solving the nonlinear problem over the fine mesh, see equation [7]. For this reason, the basic idea of the error estimator is to approximate eu by low-cost local computations over subdomains. This is a standard strategy in residual-type error estimators. The proposed approach consists of two phases. First, a simple residual problem is solved inside each element of the coarse mesh (interior estimate). Note that elements are the natural subdomains for the local computations. To avoid the expensive flux-splitting procedures of other residual-type estimators, homogeneous Dirichlet boundary conditions are prescribed for each element (that is, eu = 0 in the element boundary).
44
Figure 1. Patch associated to a node of the coarse mesh subdivided into 4 x 4 elements of the fine mesh Of course, the error is not really zero in the element boundary. For this reason, a second set of simple problems is solved. The idea is to use a different partition of the computational domain into subdomains. A natural choice is to associate these subdomains, called patches, to the nodes of the coarse mesh. Since four-noded quadrilateral elements are used, a patch consists of one-fourth of each element sharing the node, see Figure 1. By combining equations [7] and [8], the nonlinear problem to be solved on every element and every patch can be recast as
These local problems are solved over the fine mesh of size h. Every element of the coarse mesh H is subdivided into 4 x 4 elements of size h (i.e. the fine mesh is nested into the coarse mesh, with h = H/4). Due to the patch definition, every patch also consists of 4 x 4 elements of size h, see Figure 1. For the iterative solution of equation [9], UH is taken as an initial approximation to uh (that is, the initial approximation for the error is eu = 0). This means that the final state obtained with the coarse mesh of size H is taken as the initial state for solving the local problem over each element/patch with the fine mesh of size h. As a consequence, information must be projected from the coarse mesh H to the fine mesh h. To ensure the consistency between the various projected fields, the following projection strategy is employed: (1) Displacements and damage are projected over the fine mesh. To project the nodal field of displacements, the finite element approximation based on mesh H is used. The damage field must be projected from the Gauss points of the coarse mesh to the Gauss points of the fine mesh. A very simple and efficient strategy is used: the value of damage at each Gauss point of the coarse mesh is assigned to all the Gauss points of the four corresponding elements of the fine mesh, see Figure 2. With this projection strategy, the risk of unrealistic values of damage (i.e. D < 0 or D > 1) is precluded; (2) strains and stresses are not projected, but computed from the projected displacements and damage. In this manner, the consistency between all the "projected" quantities (displacements, damage, strains and stresses) is guaranteed. To keep the notation simple, these fields are still denoted with an H subscript denoting they are associated with the solution of the global problem over the mesh of size H, even though they are now supported by the mesh of size h.
Error Estimation for Nonlocal Damage Models 45
Figure 2. Projection strategy for the damage field. The value at each Gauss point of the element of size H is assigned to the four associated elements of size h 3.2. Accounting for the nonlocality of the model At this point, it is important to remark that the proposed error estimator for nonlocal damage models is based on local computations over subdomains (i.e. elements and patches). The nonlocality of the damage model must be accounted for when solving the local problems. Note that, upon mesh refinement, the element size may become smaller that the characteristic length lc. With the proposed approach, the interaction between adjacent elements is considered (thanks to the loop over patches, that overlap elements), but not the interaction between more distant elements. However, this is not regarded as a significant drawback of the suggested approach; due to the weighting function a of the nonlocal average, see equation [4], the error in one element has a limited influence on the error in distant elements. Moreover, accounting for the interaction between distant elements during the error estimation would destroy the most attractive feature of the suggested approach: it consists of solving independent local problems. Note, for instance, that the error estimation algorithm has a computational cost of O(N) (with N the number of elements) and can be parallelized. On the other hand, it is essential that these local problems are solved taking into account the current mechanical properties (i.e. the damaged stiffness) of each element/patch. As discussed in the following, this implies that the nonlocal damage model must be slightly modified. The standard and the modified nonlocal damage models are summarized in Table 2. In the standard model that is, the one used for solving the global problem, see equation [6], the error in strains e is computed as the symmetrized gradient of the error in displacements eu and added to the strains H to produce the strains h over the element/patch. After that, the local state variable Yh is computed and averaged into the nonlocal state variable Yh. Finally, damage Dh is obtained. Note that the nonlocal average that transforms Yh into Yh is over a local support (the element/patch under consideration). This fact leads to non-physical responses, especially in zones
46
Standard model
Modified model
e e -V s (e u )
eh = H + e
Yh = Y(eh)
Error in strains Strain Local variable
ee = V s (e u ) h = H + ee Yh *YH+ ^(H)e. ey: Error in Y
Yh-^Yh
Nonlocal variable
ey > e~
Yh = YH + e~
(- >: nonlocal average over local support) Dh = D(Yh)
Damage
Dh = D(Yh)
Table 2. Standard and modified nonlocal damage models
of large damage gradients. Assume, for instance, that the error in strains is small and e/i H- A small variation in Y is also expected. However, it may happen that Yh ^ YH=, because Yh. contains no information about nearby zones. This point is illustrated in Figure 3, which depicts the local state variable, the nonlocal state variable and the damage parameter for a given time increment in a zone of the coarse mesh with large gradients. The circled element has a very small local state variable YH, see Figure 3(a), below the threshold Y0. However, since the elements to the right have large values of YH , it has a relatively large (above Y0) nonlocal state variable YH, see Figure 3(b), which leads to damage, see Figure 3(c). If the standard model is used to solve the local problem on the circled element during error estimation, a small error in strains leads to a small variation in the local state variable which, after nonlocal averagingôver the element, results in a low value of the nonlocal state variable (that is, Yh <& YH). As a consequence, damage cannot increase in the circled element during error estimation. When estimating the error for the circled element, the nonlocal state variable YH, rather than the local state variable YH, is representative of its mechanical properties. For this reason, a modification of the nonlocal damage model is proposed here, see Table 2. The difference resides in the way the nonlocal state variable Yh is computed. By means of a first-order Taylor expansion, the local state variable Yh is expressed as
47
YH plus an error term ey. The derivative ^ the chain rule
is computed analytically by means of
In equation [10], EI denotes a component of the strain vector. All the derivatives in the RHS are very simple to compute from the definition of the local state variable Y, see Table 1 and of the strain invariants I1 and J2. The error term ey is averaged over the element/patch into e~. As a consequence, Yh is computed as the addition of a reference value YH, which describes the real damaged stiffness, and an error term e~. For doing so, it is necessary to project by means of the same projection strategy used for the damage field, see Figure 2YH and ^(CH) into the fine mesh. With this modified model, a small variation in strains does result in a small variation in the nonlocal state variable (that is, Yh ~ YH). Going back to Figure 3, this means that the damage level of the circled element may either remain constant (for Yh < YH) or increase (for Yh > YH) during error estimation. To sum up: the standard model is not capable of capturing the spread of the damaged zone associated to error estimation.
4. Numerical examples: the single-edge notched beam The proposed adaptive strategy is illustrated here by means of the single-edged notched beam (SENB) test [CAR 93]. The geometry, loads and supports are shown in Figure 4. A plane stress analysis is performed. The concrete beam is modelled with the modified von Mises model with exponential damage evolution, see Table 1. The steel loading platens are assumed to be elastic. Two sets of material parameters are used, see Table 3. For material 1, there is a significant post-peak softening in the stress-strain law for concrete. For material 2, on the contrary, the softening is very slight, so the residual strength almost coincides with the peak strength [PEE 98]. Material 1 Concrete Steel 28000MPa 280000MPa 0.1 0.2 1.5 x 10 -4 0.8 9000 10mm Material 2 Concrete Steel 35000MPa 350000MPa 0.2 0.2 6.0 x 10 -5 0.08 8200 10mm
Parameter E
V
Y0 A B
lc
Table 3. The two sets of material parameters: (a) large softening; (b) very slight softening
48
Figure 3. Fields in a zone of large gradients: (a) local state variable Y; (b) nonlocal state variable Y; (c) damage. The damage threshold is Y0 1.5 x 10 -4
Figure 4. Single-edge notched beam: problem statement. All distances in mm
49
Figure 5. SENB test with material 1, initial approximation in the adaptive process. (a) Mesh 0: 659 elements and 719 nodes; (b) final damage distribution; (c) final deformed mesh (x 300); (d) error distribution. The global relative error is 3.96%
4.1. Test with material 1 The results with material 1 are shown in Figures 5 to 7. The initial mesh is shown in Figure 5(a). Note that this mesh is relatively coarse, with only one element in the notch width. The final damage distribution and deformed mesh (amplified 300 times), corresponding to a CMSD (crack-mouth sliding displacement) of 0.08 mm, is depicted in Figure 5(b). The curved crack pattern observed in experiments [CAR 93] is clearly captured. The error estimation procedure discussed in section 3.2 is employed to compute the error field of Figure 5(d). The error is larger in the damaged zone and near the loading platens. The global relative error (i.e. energy norm of the error in displacements over the energy norm of displacements) is 3.96%, above a threshold set a priori of 2%, so adaptivity is required. The error field of Figure 5(d) is translated into the mesh of Figure 6(a). Note the element concentration in the crack and the central supports. This finer mesh leads to a better definition of the damaged zone, see Figure 6(b). The error estimator now detects that the largest errors are associated to the edges of the cracked zone, see Figure 6(d). The global relative error of 2.11% is still slightly above the error goal, so another adaptive iteration is performed. The outcome of this second iteration is shown in Figure 7. The qualitative results of iteration 1 are confirmed: (1) small elements are needed to control the error in the damaged zones and close to the loading platens and (2) error is larger in the edges than in the centre of the crack. The global relative error of 1.77% is below the threshold of 2%, so the adaptive iterative process stops. The relation between damage and error is illustrated by Figure 8, which depicts profiles of these two fields along the crack. Note that the two error peaks are associated to the edges of the damaged zone (i.e. large damage gradients). This indicates that the damage gradient is a good error indicator [HUE 99] for these models.
50
Figure 6. SENB test with material 1, after one iteration in the adaptive process. (a) Mesh 1: 1155 elements and 1228 nodes; (b) final damage distribution; (c) final deformed mesh (x300); (d) error distribution. The global relative error is 2.11%
Figure 7. SENB test with material 1, after two iterations in the adaptive process. (a) Mesh 2: 1389 elements and 1469 nodes; (b) final damage distribution; (c) final deformed mesh (x300); (d) error distribution. The global relative error is 1.77% 4.2. Test with material 2 The SENB test is now reproduced with material 2, see Table 3. The small value of parameter A leads to a stress-strain law with almost no softening. A very similar law has been employed to simulate the SENB test with gradient-enhanced damage models [PEE 98]. The results are shown in Figures 9 to 11. The initial mesh is the same as before, see Figure 9(a). The change in the material parameters lead to a completely different failure pattern, dominated by bending of opposite sign in the two halves of the beam, see Figures 9(b) and 9(c). A crack at the notch tip is also initiated, but it is only a
51
Figure 8. SENB test with material 1. Profiles of damage (dashed line) and error (solid line) across the crack. The two error peaks are associated to large damage gradients
Figure 9. SENB test with material 2, initial approximation in the adaptive process. (a) Mesh 0: 659 elements and 719 nodes; (b) final damage distribution; (c) final deformed mesh (x300); (d) error distribution. The global relative error is 3.66%
secondary mechanism. The error estimation procedure has no difficulties in reflecting the change in the failure mode, see Figure 9(d). The global relative error is 3.66%, so adaptivity is required. Figures 10 and 11 illustrate the adaptive process. Note that meshes 1 and 2 are quite different from the ones obtained with material 1. The global relative errors are 2.46% and 2.13%. This value is still slightly above the threshold of 2%. However, an additional iteration is considered not necessary for the illustrative purposes of this test. A final comparison between the two sets of material parameters is offered by Figure 12, where the total load is plotted versus the CMSD for meshes 0 and 2. The results
Figure 10. SENB test with material 2, after one iteration in the adaptive process, (a) Mesh 1: 776 elements and 848 nodes; (b) final damage distribution; (c) final deformed mesh (x300); (d) error distribution. The global relative error is 2.46%
Figure 11. SENB test with material 2, after two iterations in the adaptive process, (a) Mesh 2: 870 elements and 954 nodes; (b) final damage distribution; (c) final deformed mesh (x300); (d) error distribution. The global relative error is 2.13% obtained with material 1 a peak load of around 60 kN and post-peak structural softening, see Figure 12(a) are in good agreement with the experiments [CAR 93]. With material 2, on the other hand, the peak load is quite higher and no softening is observed, see Figure 12(b).
5. Concluding remarks An adaptive strategy based on error estimation for nonlocal damage models has been presented. The constitutive model has been slightly modified in order to account for its nonlocality during the error estimation procedure, see Table 2. The basic idea
53
Figure 12. Total load versus crack-mouth sliding displacement (CMSD) for meshes 0 (solid line) and 2 (dashed line): (a) with material 1; (b) with material 2 (see Table 3) of the modification is that the error in the local state variable, rather than the variable itself, is averaged. By doing so, the error estimation takes into account the real mechanical properties of the damaged material, while retaining its most attractive feature: it consists in solving simple, independent problems over elements and patches. The resulting adaptive strategy has been illustrated by means of the single-edge notched beam test. With two sets of material parameters leading to very different failure modes, h-remeshing concentrates elements where needed according to the error estimator, until the global relative error falls below an error threshold. By keeping the discretization error under control, it is possible to ensure the quality of the FE solution and assess the influence of the material parameters in an objective way. Acknowledgements The partial financial support of the Ministerio de Ciencia y Tecnologia (grant numbers: TAP98-0421, 2FD97-1206) is gratefully acknowledged. 6. References
[ASK 00] ASKES H., SLUYS L., "Remeshing strategies for adaptive ALE analysis of strain localisation", European Journal of Mechanics A/Solids, vol. 19, 2000, p. 447-467. [BAZ 88] BA2ANT Z., PIJAUDIER-CABOT G., "Nonlocal continuum damage localization instability and convergence", Journal of Applied Mechanics, vol. 55, 1988, p. 287-293.
54 Numerical Modelling in Damage Mechanics [BOR93] DE BORST R., SLUYS L., MCLHAUS H.-B., PAMIN J., "Fundamental issues in finite element analysis of localization of deformation", Engineering Computations, vol. 10, 1993, p. 99-121. [CAR 93] CARPINTERI A., VALENTE S., FERRARA G., MELCHIORRI G., "Is mode II fracture energy a real material property?", Computers and Structures, vol. 48, 1993, p. 397413. [DIE 98] DfEZ P., EGOZCUE J., HUERTA A., "A posteriori error estimation for standard finite element analysis", Computer Methods in Applied Mechanics and Engineering, vol. 163, 1998, p. 141-157. [DIE 99] DfEZ P., HUERTA A., "A unified approach to remeshing strategies for finite element /i-adaptivity", Computer Methods in Applied Mechanics and Engineering, vol. 176, 1999, p. 215-229. [DIE 00] DfEZ P., ARROYO M., HUERTA A., "Adaptivity based on error estimation for viscoplastic softening materials", Mechanics of Cohesive-Frictional Materials, vol. 5, 2000, p. 87-112. [HUE 99] HUERTA A., RODRIGUEZ-FERRAN A., DfEZ P., SARRATE J., "Adaptive finite element strategies based on error assessment", International Journal for Numerical Methods in Engineering, vol. 46, 1999, p. 1803-1818. [HUE 00] HUERTA A., DfEZ P., "Error estimation including pollution assessment for nonlinear finite element analysis", Computer Methods and Applied Mechanics in Engineering, vol. 181, 2000, p. 21-41. [LEM 90] LEMAITRE J., CHABOCHE J.-L., Mechanics of solid materials, Cambridge University Press, Cambridge, 1990. [MAZ89] MAZARS J., PIJAUDIER-CABOT G., "Continuum damage theory: application to concrete", Journal of Engineering Mechanics, vol. 115, 1989, p. 345-365. [PEE 98] PEERLINGS R., DE BORST R., BREKELMANS W., GEERS M., "Gradient-enhanced damage modelling of concrete fracture", Mechanics of Cohesive-Frictional Materials, vol. 3, 1998, p. 323-342. [PIJ 87] PIJAUDIER-CABOT G., ZANT Z. B., "Nonlocal damage theory", Journal of Engineering Mechanics, vol. 118, 1987, p. 1512-1533. [PIJ 91] PIJAUDIER-CABOT G., MAZARS J., "Steel-concrete bond analysis with nonlocal continuous damage", Journal of Structural Engineering, vol. 117, 1991, p. 862-882. [ROD 00] RODRfGUEZ-FERRAN A., HUERTA A., "Error estimation and adaptivity for nonlocal damage models", International Journal of Solids and Structures, vol. 37, 2000, p. 7501-7528. [SAR 00] SARRATE J., HUERTA A., "Efficient unstructured quadrilateral mesh generation", International Journal for Numerical Methods in Engineering, vol. 49, 2000, p. 1327-1350. [VRE95] DE VREE J., BREKELMANS W., VAN GILS M., "Comparison of nonlocal approaches in continuum damage mechanics", Computers and Structures, vol. 55, 1995, p. 581-588.
Chapter 4
Mathematical and Numerical Aspects of an Elasticity-based Local Approach to Fracture

R.H.J. Peerlings, W.A.M. Brekelmans and M.G.D. Geers
Department of Mechanical Engineering, Eindhoven University of Technology, The Netherlands
R. de Borst
Department of Aerospace Engineering, Delft University of Technology, The Netherlands
An Elasticity-based Local Approach to Fracture
57
1. Introduction Component failure due to the formation and growth of cracks is traditionally modelled using fracture mechanics. Fracture mechanics theory uses global criteria to determine under which conditions a pre-existing crack will grow and thus lead to complete fracture of the component. Additional criteria predict the rate and direction of crack growth. In situations where the surrounding material is relatively unaffected by the presence of a crack, this type of modelling is highly successful. In many engineering materials, however, the concentration of deformation and stress near the crack tip produces irreversible changes in the microstructure of the material and in the microstructural processes which govern its behaviour. Examples are plastic flow and/or void formation in ductile materials, microcracking in concrete and fibre pull-out or delamination in fibre-reinforced polymers. In return, these changes may have a considerable influence on - or indeed govern - the crack growth process. In situations where interactions between crack and microstructural damage play an important role, fracture mechanics may not be the most suitable modelling tool. A more natural treatment of these problems is provided by the so-called local approach to fracture, in which the change of material behaviour is modelled explicitly [LEM 86, CHA 88]. The development and growth of a crack is regarded as the ultimate consequence of the local degradation process. A crack is represented by a region in which the material integrity has been completely lost and which therefore cannot sustain any stress. The internal boundary which describes the crack contour expands when material in front of the crack tip fails. As a result, no separate fracture criteria are needed: the rate and direction of crack growth follow from the constitutive behaviour. As an additional advantage, crack initiation and crack growth can be described within the same framework, so that it is not necessary to define an - often arbitrary - initial crack. In local approaches to fracture, the degradation of material properties is often modelled using continuum damage mechanics [KAC 58, RAB 69, CHA 88, LEM 96]. In their standard form, this and other types of degradation modelling (e.g., softening plasticity) are usually not suited to describe crack initiation and crack growth, because they cannot properly describe the accompanying localised deformations. As a consequence, the deformation and damage growth are often observed to localise in a surface (i.e., in a vanishing volume) right at - or even before - the onset of fracture. In numerical analyses, this localisation results in an extreme sensitivity to the spatial discretisation of the problem. Upon refinement of the discretisation, the solution converges to one in which the fracture process is instantaneous and does not dissipate any energy. It is emphasised that this pathological behaviour is not due to the numerical treatment, but to a shortcoming of the underlying continuum modelling. It can be removed by introducing nonlocality in the constitutive modelling, either by integral terms or by higher-order gradients [BAZ 84, PC 87, FRE 96, PEE 96]. The enriched continuum formulations which are thus obtained preclude the localisation of deformation in a vanishing volume and the resulting instability. As a result, crack growth rate predictions remain finite and a positive amount of energy is dissipated.
Finite element analyses using nonlocal continuum models are mesh-insensitive once the discretisation is sufficiently fine to accurately capture the solution. However, these methods require some modifications of the standard algorithms. The nonlocality introduces either an additional integration step or an extra set of equations. Special attention must be paid to the treatment of the additional terms near boundaries. This is particularly true in fracture problems, where boundary conditions must also be applied at the internal boundary between the crack and the remaining material. Furthermore, frequent remeshing and an adaptive step size selection may be needed to accurately describe the crack path and special solution control techniques may be required to deal with instabilities and bifurcations [PC 91, GEE 99, PEE 99, PEE 00]. This paper summarises the algorithmic ingredients which are essential for reliable and accurate fracture analyses using the continuum approach. An elasticity-based damage framework is used here (Section 2), but most issues are equally relevant in other degradation models (e.g., softening plasticity). Emphasis is on preventing pathological localisation and mesh sensitivity. Since these phenomena find their origin in the mathematical (continuum) modelling, it is useful to first study the difficulties arising at this level; this is done in Section 3. Section 4 shows how these difficulties can be avoided by using an enriched formulation. Some aspects of the finite element implementation are discussed in Section 5 and results are given in Section 6. 2. Constitutive modelling and the local approach to fracture Continuum damage mechanics uses a set of continuous damage variables to represent microstructural defects (microcracks, microvoids) in a material. If it is assumed that the development of damage does not introduce anisotropy, a single, scalar damage variable can be used to describe the local damage state. This damage variable D is defined such that 0 < D < 1, where D = 0 represents the initial, undamaged material and D = 1 represents a state of complete loss of material strength. If the influence of damage is added to standard linear elasticity, the classical stress-strain relation of elasticity-based damage mechanics is obtained [LEM 90]:
where cr,7 denotes the Cauchy stress components, Cijkl the standard elasticity tensor and Skl the linear strains. An important application of relation [1] is to quasibrittle fracture (e.g., concrete, fibre-reinforced polymers). Here, however, we will use it to model high-cycle fatigue, in which plastic deformations also remain negligible [PAA 93, PEE 99, PEE 00]. In the fatigue model, the growth of damage is related to the local deformation of the material. For this purpose a damage loading function is introduced in terms of the strain components:
An Elasticity-based Local Approach to Fracture 59
with s a positive, scalar equivalent measure of the actual strain state and K a threshold variable, which is taken constant here: K = K0. For the equivalent strain the von Mises strain, scaled such that it equals the axial strain in uniaxial tension, will be used:
where
The damage variable remains constant when / < 0; the behaviour is then linear elastic. Notice that positive values of K0 therefore imply the existence of a fatigue threshold. Strain states for which / > 0 lead to damage growth only for continued deformation and as long as the critical value D = 1 has not been reached, i.e., if f > 0 and D < 1. When these three conditions are satisfied, the damage rate is governed by an evolution law which reads in its most general form
where the dependence of the damage growth rate on the equivalent strain rate has been taken to be linear in order to avoid rate effects. It is immediately clear from equation [1] that no stresses can be transferred for D = 1. In the local approach to fracture this critical state is used to represent a crack by a region of completely damaged material (fic in Figure 1). In the remaining part of the domain, and particularly next to the crack, some (noncritical) damage may have been developed (fid in the figure), while other areas may still be unaffected by damage (fi 0 ). In the latter region the material has retained its virgin stiffness. Under the influence of further straining the damage variable will increase in those parts of the body where the conditions for damage growth are met. This will often be the case particularly near the strain concentration in front of the crack tip. When the critical value D = 1 is reached in this region, the completely damaged zone fic will start to expand, thus simulating crack growth. The direction and rate of crack growth are governed by the damage growth locally near the crack tip, hence the term local approach to fracture. For reasons of computational efficiency, the stress-strain behaviour is sometimes partly uncoupled from the growth of damage. The material in the damaged zone fid then retains its virgin stiffness until the damage variable becomes critical. Upon reaching D = 1, the elastic stiffness is then suddenly decreased from its virgin value to zero. This method is sometimes referred to as uncoupled or semi-coupled approach, while the full model, in which the stresses are governed by relation [1], may be called coupled or fully coupled [PAA 93, LEM 96]. It is obvious that the uncoupled approach can only be followed if the influence of damage prior to failure is relatively small, which may be true in the high-cycle fatigue case considered here. However, this also cancels part of the advantage of the local approach, namely that it can account for the influence of material damage on crack growth.
Figure 1. Damage distribution in a continuum Both in the coupled and uncoupled approach it is important to realise that the local, complete loss of strength in 2C implies that stresses are zero for any arbitrary deformation (see equation [1]). As a consequence, the equilibrium equations are meaningless in this region. This can be seen for the elasticity-based damage model by substituting relations [1] in the standard equilibrium equations
Making use of the right minor symmetry of the elasticity tensor (i.e., Cijkl = Cijlk) the equilibrium equations can then be written as the system of second-order partial differential equations
For a given damage field D(x) < 1, the displacement components Uk can be determined from this differential system and the corresponding kinematic and dynamic boundary conditions. In a crack however, where D = 1, both terms in [7] vanish. Consequently, the partial differential equations degenerate and the boundary value problem becomes ill-posed. This indefiniteness must be avoided by limiting the equilibrium problem to the subdomain 2 = 0 Ud where D < 1. At the boundary between crack and remaining material the natural boundary condition nicij = 0 must be applied, with the vector n normal to the boundary. A free boundary problem is thus obtained, in which the position of the internal boundary (the crack front and crack faces) follows from the growth of damage, see also [BUI 80].
A nElasticity-based Local Approach to Fracture 61
3. Localisation and mesh sensitivity Analyses using the local approach as described in the previous section will usually quickly result in a situation where all further growth of damage is concentrated in a surface. Since this means that the volume of material that participates in the damage development vanishes, no work is needed for the crack to propagate, even if the specific work needed by the damage process is positive. Furthermore, the crack traverses the remaining cross-section instantaneously, instead of by a small increment in every loading cycle. From a mathematical point of view, two phenomena play a role in this pathological localisation of damage growth: loss of ellipticity of the rate equilibrium equations and singularity of the damage rate. The latter cause is often not recognised in the literature, but is actually the most important in crack growth analyses, since it is responsible for the instantaneous and perfectly brittle crack growth. Loss of ellipticity, on the other hand, may result in premature initiation of cracks, before the damage variable has reached the critical level D = 1 in a stable way. The rate equilibrium equations lose ellipticity when the acoustic tensor n i C ijkl n l , where C ijkl denotes the tangent stiffness, becomes singular for some unit vector n, i.e., when
For the fully coupled fatigue damage model the tangential stiffness Cijkl is given
by
As the damage variable grows, the second term in this expression may become of the same order as the first term, so that condition [8] may indeed be met at some point. The vector n is then the normal to a characteristic surface segment of the set of rate equilibrium equations. Since solutions of linear partial differential equations with smooth coefficients can have discontinuities or discontinuous derivatives only across characteristic surfaces, this opens the possibility of jumps in the velocity solution. A discontinuity of the velocity across a characteristic surface results in a strain rate singularity on this surface, which in turn renders the damage rate singular (see equation [5]). This means that for continued loading the damage variable immediately becomes critical on the characteristic segment and thus that instantaneous failure of this surface is predicted. In order to follow the stress drop resulting from this instantaneous loss of stiffness, material adjacent to the characteristic segment must unload elastically, so that the growth of damage indeed localises in the surface segment. Loss of ellipticity plays an important role in localisation of deformation and damage in static fracture, where damage growth is fast right at the onset [PEE 96]. In fatigue, however, the initial growth of damage is usually slow and the ellipticity of the rate equilibrium problem is therefore preserved until near the end of the fatigue life.
Indeed, it can be easily seen that for the uncoupled approach loss of ellipticity cannot occur before the damage variable becomes critical. Since the tangential stiffness tensor Cijkl equals the elasticity tensor Cijkl in this case, the characteristic determinant equals
where X, u denote Lame's constants. Expression [10] is always positive and the problem therefore remains elliptic until D = 1. As a result, the damage growth remains stable and affects a finite volume throughout the crack initiation phase in fatigue. Once a crack has been initiated, however, the continued growth of damage nevertheless localises in a surface and the predicted crack growth becomes nonphysical again. This localisation during crack growth is not due to loss of ellipticity, but is related to the strain singularity which is inevitably present at the tip of the crack. Once the damage variable becomes critical in a certain point and a crack is thus initiated, the displacement and velocity must become discontinuous across this crack. This implies that the strain (rate) field at the crack tip becomes singular. Since the damage growth rate is directly related to the equivalent strain, the strain singularity at the tip results in an infinite damage rate. For continued deformation all stiffness is therefore lost instantaneously at the most critical point in front of the crack tip and the crack thus starts to propagate. Since the material adjacent to the crack must unload elastically in order to follow the resulting stress drop, the width of the crack remains zero. This implies that the strain and damage growth rate at the crack tip remain singular as the crack grows and consequently that the crack grows at an infinite rate. No work is needed in this fracture process, since it involves damage growth in a vanishing volume. It is emphasised once more that this mechanism of damage localisation and instantaneous crack growth is activated even if the rate equilibrium equations remain elliptic until D = 1. Loss of ellipticity may cause premature initiation of a crack and thus result in perfectly brittle crack growth, but the pathological propagation behaviour is nevertheless due to the singularity of the damage rate at the crack tip. This is true even in models of static fracture, but the problem is aggravated in this case by the fact that a crack is initiated shortly after the onset of damage as a result of loss of ellipticity. Finite element solutions try to follow the nonphysical behaviour of the continuum model as described above, but are limited in doing so by their finite spatial resolution. In standard finite element methods the displacement field must be continuous. The displacement jumps and singular strains of the actual solution can therefore only be approximated by high, but finite displacement gradients in the finite element solution. As a consequence, a finite volume is involved in the damage process, and a positive amount of energy is dissipated. Also, because the damage growth rate at the tip of the damage band remains finite, the crack propagates at a finite velocity. When the spatial discretisation grid is refined, however, the finite element approximation becomes more accurate in the sense that the displacement gradients which describe the discontinuities become stronger. Consequently, the predicted fracture energy becomes smaller and the crack propagates faster. In the limit of vanishingly small elements, the actual solution
An Elasticity-based Local Approach to Fracture 63 is retrieved, i.e., a vanishing fracture energy and an infinite crack growth rate. This convergence of the finite element approximation to the actual, nonphysical solution of the problem is the origin of the apparent mesh sensitivity of damage models and other continuous descriptions of fracture. An example of the apparent mesh sensitivity is given in Figure 2. The diagram shows the steady-state fatigue crack growth rate predicted by a finite element analysis versus the size of the elements which were used in the analysis. The problem geometry, loading conditions and modelling for which these results have been obtained will be detailed in Section 6. Fully coupled as well as uncoupled analyses have been done. The dependence of the crack growth per cycle da/dN on the element size h is quite strong in both approaches: a decrease of the element size by roughly one decade leads to an increase of the crack growth rate by almost three decades. In the limit h > 0 the crack growth rate clearly goes to infinity, as predicted by the discussion above.
Figure 2. Predicted fatigue crack growth rate versus element size
4. Nonlocal modelling An effective method to avoid pathological localisation of damage is to add nonlocal terms to the constitutive model. This approach has been successfully applied to damage models of a number of failure mechanisms [BAZ 84, PC 87, SAA 89, TVE95]. The spatial interactions resulting from the nonlocality prevent the damage growth from localising in a surface. Instead, the damage growth occupies a finite band, the width of which is related to the internal length scale provided by the nonlocality. In its traditional, integral form, nonlocality can be introduced in the fatigue model of Section 2 by rewriting the loading function [2] and the damage evolution law
64 Numerical Modelling in Damage Mechanics [5] in terms of a new field variable, the nonlocal equivalent strain e :
The nonlocal equivalent strain is defined by (cf. [PC 87])
with VK(y; x) a weight function which usually decays rapidly with the distance |y x|. The nonlocality is apparent from equations [12] and [13]: the damage growth and thus the stresses in x are influenced by the value of the strain in other points y = x. Instead of the integral definition [13] of the nonlocal strain, we will use the boundary value problem given by
to define the nonlocal equivalent strain s [PEE 96]. The solution of this problem can formally be written as:
where G(y; x) denotes the Green's function associated with it [PEE 99, PEE 01]. Expression [16] takes exactly the same form as equation [13] for the nonlocal model. This means that the enhanced damage model based on the differential equation [14] is a member of the class of nonlocal models defined by [13]. The parameter c in equation [14], which is of the dimension length squared, sets the internal length scale of the model and thus determines the degree to which damage growth localises. It should be noted that, in the presence of cracks, equation [14] is defined only in the domain Q. where the damage variable has not yet become critical. This is not only natural, since the equilibrium problem is defined only on 2, but also necessary because the right-hand side s is not uniquely defined in the cracked region as a result of the indefiniteness of the displacement field (Section 2). The boundary condition [15] associated with equation [14], as well as the standard boundary conditions for the equilibrium problem, must therefore be defined on the boundary f of 2. This means that they are imposed not only at the boundary of the problem domain, but also at the internal boundary which represents the crack contour and that they move with the crack contour as the crack grows. In terms of the original nonlocal formulation, the integration in equation [13] must be limited to 2 and must therefore be re-evaluated
as the crack grows. The necessity of this identical treatment of internal and external boundaries does not seem to have been recognised in the literature. If the - nonphysical - strains in the crack are included in the integral in [13] the computed nonlocal equivalent strain and damage rate at the crack tip are too high, resulting in an overprediction of the rate of crack growth. Furthermore, since the nonlocal strain at the crack faces continues to increase as the crack opens, the damage variable continues to grow at the crack faces. As a consequence, the width of the crack region continues to increase along the entire crack surface, until it finally occupies the entire domain. Both effects can also be observed in numerical analyses if the cracked zone is not properly separated from the remaining material. It has been argued in Section 3 that the standard, local damage model predicts the immediate initiation of a crack when the displacement field becomes discontinuous upon loss of ellipticity of the rate equilibrium problem. It can be easily shown for the nonlocal model that this loss of ellipticity no longer occurs. The characteristic determinant associated to the set of rate equilibrium equations and equation [14] is given by [PEE 99]
where C*ijkl (i, l = 1,2, 3, j, k = 1, 2, 3, 4) contains the coefficients of the secondorder derivatives in the combined set of equations. Expression [17] is positive for all n as long as D < 1. This means that the partial differential system is elliptic throughout the initiation phase. However, when D = 1 somewhere in the domain, and a crack is initiated, a strain singularity may be unavoidable at the crack tip. It is important that the damage growth rate remains finite, because the crack growth would otherwise be instantaneous (see Section 3). Since the damage growth rate depends on the nonlocal equivalent strain e in the nonlocal formulation, this implies that e must remain finite at the crack tip in order to have a finite crack growth rate. An analytical expression for the nonlocal strain can be obtained for a crack in an infinite, linear elastic medium. It is emphasised that this situation is not entirely representative of a crack in a damaged medium, because the influence of damage on the deformation near the crack tip is not accounted for. However, the analysis is illustrative of the way in which the nonlocality removes the damage rate singularity and thus localisation of damage in a surface. A plane crack in an infinite medium is considered, which is loaded in mode I. Furthermore, a plane stress state is assumed throughout the medium. The asymptotic local equivalent strain field can then be determined from linear fracture mechanics and shows the usual r -1/2 -singularity:
The nonlocal equivalent strain is now obtained by solving the boundary value problem [14]-[15] for this source term. This results for the nonlocal strain at the crack tip

in [PEE 99]:
with r(a) the gamma function. This expression is indeed finite for c > 0, so that the damage growth rate at the crack tip remains finite in the enhanced model. This in turn means that in this simplified situation a finite crack growth rate is obtained instead of the instantaneous fracture predicted by the standard, local damage model. As was mentioned earlier, this does not necessarily imply that singularities are avoided also in the full, coupled model, where the singularity of e may be stronger. However, numerical simulations (see Section 6) seem to indicate that this is indeed the case.
5. Aspects of the finite element implementation In mathematical terms the essential difference between the nonlocal damage formulation introduced in Section 4 and the classical, local damage models consists of the additional linear partial differential equation [14]. This equation must be solved simultaneously with the standard equilibrium equations. For finite element implementations this means that e must be discretised in addition to the displacement components. The discrete form of [14] follows from the standard transition to a weak form and Galerkin discretisation of the nonlocal equivalent strain [PEE 96]:
where the matrices N and B contain the interpolation functions of the nonlocal equivalent strain and their derivatives and the column matrix e contains the nodal values of s. The discrete form of the equilibrium equations follows in the standard way:
with the matrices N and B containing the displacement interpolation functions and their derivatives, respectively, and the column matrices a and t the Cauchy stresses and boundary tractions. The finite element interpolations of the displacements and the nonlocal strain need to satisfy only the standard, Co-continuity requirements. The order of each of the interpolations can be selected independently, although some combinations may result in stress oscillations [PEE 99]. Since damage growth is defined by relation [12] in a rate format, it must be integrated over each time increment of the numerical analysis in order to obtain the damage at the end of the increment. Standard integration rules may be used for this
purpose, but in high-cycle fatigue analyses it may be advantageous to use a more sophisticated integration, which takes into account the cyclic character of the loading [PEE 99, PEE 00]. After discretisation in time by either method, [20] and [21] become a set of nonlinear algebraic equations, which can be solved for instance using a Newton-Raphson scheme, see references [PEE 99, PEE 00] for details. It is at this point that the gain in efficiency of an uncoupled approach becomes apparent: if the effect of damage growth on the stiffness is neglected while the damage variable is noncritical, the problem will often remain linear as long as no additional elements fail (equation [20] may be nonlinear for nonproportional loading and for some equivalent strain definitions). As a result, the tangent stiffness matrix remains constant and one iteration suffices in each increment to reach equilibrium. It has already been argued that the cracked region, 2C, should not be part of the equilibrium problem domain because the equilibrium equations are not meaningful in it. Accordingly, the equilibrium equations and the additional equation [14] are defined only on the domain & = Q \ 2C and boundary conditions are provided at the boundary f of 2. For the finite element formulation this means that the discretisation of the equilibrium problem must also be limited to the noncritical domain 2. The difficulty is, however, that this effective domain will gradually shrink as the predicted crack growth progresses. Consequently, the problem domain must be redefined in the numerical analysis for each increment of crack growth and a new finite element discretisation must be defined. This remeshing is often avoided by using the original domain Q even if this domain contains a crack. The material in the crack is then given a small residual stiffness in order to avoid singularity of the discrete equilibrium equations. It is then argued that the stresses which are still transferred by the crack influence equilibrium only marginally if the residual stiffness is sufficiently small. This may indeed be true in local damage models, in which the nonphysical strains in the crack do not influence the surrounding material. But if this approach is followed for nonlocal damage models, the nonlocal equivalent strain maps the large strains which may be computed in the cracked region onto the surrounding material in which the damage variable is not (yet) critical. This does not only result in faster growth of damage in front of the crack and consequently in higher predicted crack growth rates, but also in damage growth at the faces of the crack, thus causing the thickness of the crack region to increase unboundedly. The numerical implementation should therefore reflect the mathematical separation of the crack and the remaining domain by adapting the finite element mesh to the growth of the crack. This can be achieved without full remeshing of the problem if completely damaged elements are removed from an otherwise fixed finite element mesh. However, the crack contour always follows the (initial) grid lines in this approach, which means that a fine discretisation is needed in a relatively large region.
68 Numerical Modelling in Damage Mechanics 6. Application The nonlocal damage formulation has been used to model crack initiation and growth due to fatigue. Reference is made to [PEE 99, PEE 00] for details of the damage modelling. The problem geometry of Figure 3 has been considered. The thickness of the specimen is 0.5 mm. The lower edge of the specimen is fixed in all directions, while fully reversed vertical displacement cycles with an amplitude of 0.0048 mm are forced upon its top edge. Because of symmetry, only half of the specimen has been modelled in the finite element analyses. The reference mesh contains a regular grid of elements with an edge length h = 0.04 mm in an area of approximately 0.65 x 0.12 mm2 at the notch tip (indicated in Figure 3). The discretisation has been successively refined in this area to h = 0.02, 0.01 and 0.005 mm. Quadrilateral planestress elements with bilinear displacement and nonlocal strain interpolations and a constant damage variable have been used. Both fully coupled and uncoupled analyses have been done. In these analyses, elements were removed when the damage variable exceeded 0.999999, after which led to this critical damage value was recomputed starting from the converged state in the previous increment [PEE 99]. Figure 4 shows the crack initiation and growth process as simulated using the finest of the four meshes and the coupled approach. The area which is shown in this figure is the refined area indicated in Figure 3. The stress concentration at the notch tip initially leads to a concentration of damage at the tip. After 4210 cycles a crack is initiated, i.e., the damage variable becomes critical in an element which is then removed from the mesh. For continued cycling the crack grows along the symmetry axis. The crack width decreases as the damage zone which was formed before crack initiation is traversed. Beyond this damage zone the crack width becomes stationary at 0.04 mm, which is of the same order as the internal length ^fc = 0.1 mm.
Figure 3. Problem geometry and loading conditions of the fatigue problem (dimensions in mm)
Figure 4. Damage and crack growth at the notch tip in the h = 0.005 mm mesh (coupled approach) The influence of the finite element discretisation on the crack shape is shown in Figure 5, in which the final crack pattern has been plotted for the four discretisations. The coarsest mesh (Figure 5(a)) gives a rather crude approximation of the crack shape and necessarily overestimates the width of the steady-state part of the crack because this width is smaller than the element size. But the h = 0.02 and 0.01 mm meshes give a good approximation of the crack shape in the finest discretisation. The steady-state width of the crack does not vary between the three finest discretisations. The final damage and crack patterns obtained with the uncoupled approach are almost identical to the ones shown in Figure 5. However, there is a slight difference in the number of loading cycles needed to reach these states. This is illustrated in Figure 6, which shows the length of the crack, a, versus the number of loading cycles, N, for the four meshes in the coupled as well as the uncoupled approach. For an increasingly refined discretisation the growth curves converge to a response with a finite number of
Figure 5. Final crack pattern in the (a) h = 0.04 mm, (b) h = 0.02 mm, (c) h = 0.01 mm and (d) h = 0.005 mm meshes (coupled approach) cycles to crack initiation and a finite growth rate, instead of the instantaneous growth predicted by the local model. In the uncoupled model, the crack is initiated slightly later and grows slightly slower than in the fully coupled model. This is due to the fact that the damage in front of the crack tip has no influence on the deformation, which is therefore smaller. This results in a smaller damage rate and thus in slower crack growth. The steady-state crack growth rate obtained in both approaches has been plotted versus the element size h in Figure 7. In contrast with the local damage model (Figure 2) the growth rate in the nonlocal models becomes practically constant as the element size is reduced.
71
Figure 6. Influence of the element size on the predicted crack growth in the nonlocal damage model
Figure 7. Influence of the element size on the steady-state crack growth rate 7. Discussion and concluding remarks A key issue in the development of fracture models based on a continuum damage approach is their ability to correctly describe the localised deformations which are typical of fracture problems. If this issue is not properly addressed, the damage process which represents the initiation and growth of cracks tends to localise in a vanishing volume. A perfectly brittle response is then obtained, even if the constitutive relations
have been designed to show a gradual loss of strength. This pathological localisation of damage is not so much caused by loss of ellipticity of the rate equilibrium equations, but rather due to singularities at the crack tip. It therefore occurs not only in fully coupled analyses, but also in uncoupled analyses, in which the damage variable does not immediately affect the constitutive behaviour. As a result of the singularities, the material in front of the crack fails immediately and in a vanishing volume, even if the rate equilibrium equations do not first lose ellipticity. The crack traverses the remaining cross section at an infinite growth rate and the thickness of the corresponding damage band is zero. The nonphysical behaviour of the standard models can be effectively removed by the introduction of nonlocality in the constitutive relations. This can be achieved by including an additional partial differential equation in the equilibrium problem. As a result, the localisation of damage is limited to the scale of the intrinsic length which is introduced by the nonlocality. Crack growth is no longer instantaneous and a positive volume takes part in the damage process which describes the crack growth. This also means that a positive amount of work is needed for the crack growth and that the fracture process is thus no longer perfectly brittle. Additional boundary conditions must be provided in the nonlocal model, not only at the boundary of the problem domain, but also at the internal boundary which describes the crack contour. The latter ensures that the crack is well separated from the remaining part of the continuum and that nonphysical deformations which may be computed in the cracked region do not affect the growth of damage at the crack faces. The numerical implementation of the nonlocal model must reflect this separation. This means that the spatial discretisation of the equilibrium problem must be adapted for each increment of crack growth. If this separation is not made rigorously, the damage growth rate may be overestimated and nonphysical damage growth may be predicted at the faces of the crack. In this contribution, a rigorous but crude approach has been followed: completely damaged elements are removed from the finite element mesh. Meaningful, mesh-objective numerical solutions have been obtained with this technique for the nonlocal formulation of the coupled as well as the uncoupled problem. Although reliable and useful for development purposes, the approach is not very suitable for practical problems. The location of crack initiation and the direction of crack growth are usually not known in advance. In this case, adaptive spatial discretisation techniques are needed to follow the free boundary which represents the crack contour and to accurately describe the high deformation gradients at its tip.
8. References
[BAZ84] BAANT Z.P., BELYTSCHKO T., CHANG T.P., "Continuum theory for strainsoftening", J. Eng. Mech., vol. 110, 1984, p. 1666-1692. [BUI 80] BuI H.D., EHRLACHER A., "Propagation of damage in elastic and plastic solids", D. Francois, et al., eds., Advances in Fracture Research, 1980, p. 533-551.
73
[CHA 88] CHABOCHE J.L., "Continuum damage mechanics. Part I - General concepts; Part II - Damage growth, crack initiation, and crack growth", J. Appl. Mech., vol. 55, 1988, p. 59-64, 65-72. [FRE 96] FREMOND M., NED JAR B., "Damage, gradient of damage and principle of virtual power", Int. J. Solids Struct., vol. 33, 1996, p. 1083-1103. [GEE 99] GEERS M.G.D., "Enhanced solution control for physically and geometrically nonlinear problems, parts I and II", Int. J. Num. Meth. Eng., vol. 46,1999, p. 177-204, 205-230. [KAC 58] KACHANOV L.M., "On the time to failure under creep conditions", Izv. Akad. Nauk. SSSR, Old. Tekhn. Nauk., vol. 8, 1958, p. 26-31. In Russian. [LEM 86] LEMAITRE J., "Local approach to fracture", Eng. Fract. Mech., vol. 25, 1986, p. 523-537. [LEM 90] LEMAITRE J., CHABOCHE J.-L., Mechanics of Solid Materials, Cambridge, Cambridge University Press, 1990. [LEM 96] LEMAITRE J., A Course on Damage Mechanics, Berlin, Springer, 2nd edn., 1996. [PAA93] PAAS M.H.J.W., SCHREURS P.J.G., BREKELMANS W.A.M., "A continuum approach to brittle and fatigue damage: theory and numerical procedures", Int. J. Solids Struct., vol. 30, 1993, p. 579-599. [PC 87] PIJAUDIER-CABOT G., BAZANT Z.P., "Nonlocal damage theory", /. Eng. Mech., vol. 113, 1987, p. 1512-1533. [PC 91] PIJAUDIER-CABOT G., HUERTA A., "Finite element analysis of bifurcation in nonlocal strain softening solids", Comp. Meth. Appl. Mech. Eng., vol. 90, 1991, p. 905-919. [PEE 96] PEERLINGS R.H.J., DE BORST R., BREKELMANS W.A.M., DE VREE J.H.P., "Gradient-enhanced damage for quasi-brittle materials", Int. J. Num. Meth. Eng., vol. 39, 1996, p. 3391-3403. [PEE 99] PEERLINGS R.H.J., Enhanced damage modelling for fracture and fatigue, Ph.D. thesis, Eindhoven University of Technology, Eindhoven, The Netherlands, 1999. [PEE 00] PEERLINGS R.H.J., BREKELMANS W.A.M., DE BORST R., GEERS M.G.D., "Gradient-enhanced damage modelling of fatigue", Int. J. Num. Meth. Eng., vol. 49, 2000, p.1547-1569. [PEE 01] PEERLINGS R.H.J., GEERS M.G.D., DE BORST R., BREKELMANS W.A.M., "A critical comparison of nonlocal and gradient-enhanced softening continua", Accepted for publication. [RAB 69] RABOTNOV Y.N., Creep Problems in Structural Members, Amsterdam, NorthHolland, 1969. [SAA 89] SAANOUNI K., CHABOCHE J.-L., LESNE P.M., "On the creep crack-growth prediction by a non local damage formulation", Eur. J. Mech. A/Solids, vol. 8, 1989, p. 437459. [TVE 95] TVERGAARD V., NEEDLEMAN A., "Effects of non-local damage in porous plastic solids", Int. J. Solids Struct., vol. 32, 1995, p. 1063-1077.
Chapter 5
Numerical Aspects of Nonlocal Damage Analyses

Claudia Comi and Umberto Perego
Department of Structural Engineering, Politecnico of Milan, Italy
Aspects of Nonlocal Damage Analyses 77 1. Introduction In many instances of practical interest, the initiation of fracture is preceded by a significant strain localization phase in which the material is macroscopically integer and inelastic phenomena tend to be confined in a narrow region. In this phase, the use of continuum models with softening, like e.g. damage models, is justified. However, the strain softening behavior due to the development of material damage is well known to produce unrealistic mesh sensitivity in standard finite element application. Zero energy dissipation is expected in the limit since strains tend to localize on a zero volume region as the mesh is refined. In statics, the failure of classical discretization methods can be explained, from the mathematical point of view, with the boundary value problem losing ellipticity as a consequence of the softening material behavior. The ill-posedness of the boundary value problems reflects the fact that standard continuum mechanics theories are not appropriate when the microscopic material heterogeneity is characterized by an internal length which is not negligible if compared to the typical macroscopic length of the structure, so that the range of the microscopic interaction forces has to be considered large with respect to the macroscopic scale (see e.g. [GAN 00] for a recent discussion). Among the several regularization techniques proposed in the literature, one of the most computationally convenient seems the one based on the formulation of a nonlocal continuum (see [ERI 81] for nonlocal plasticity). The idea is that the long range nature of the microscopic interaction forces is taken into account on the macroscale by expressing the material constitutive law in terms of one or more nonlocal variables defined as suitable weighted averages of their local values over the interaction domain. In the formulation of a nonlocal model, several choices have to be made such as the definition of the nonlocal variable (variables), the definition of the weight function and the definition of the interaction domain. The adopted choices have important numerical consequences in finite element implementations (see [JIR 98] for a discussion of other aspects): the corrector phase of the iterative procedure, typically carried out at each Gauss point separately, may cease to be local [STR 96]; the consistent tangent matrix becomes non-symmetric [BAZ 88], [PIJ 95], [JIR 99]. The lack of symmetry has important consequences both from the theoretical and computational point of view. In particular a non symmetric model is not suitable for variational approaches and non symmetric solvers have to be used in numerical applications, with a consequent increase in computing costs. In the present paper the discussion is confined to isotropic damage models. Within this context, it is shown that it is possible to formulate a very general isotropic local model endowed with a symmetric consistent tangent matrix. The model considered is based on the definition of two damage variables affecting the shear and bulk moduli separately. The consistent tensor of tangent elastic moduli is
78
derived and it is shown that it is symmetric provided that associative evolution equations are assumed for both damage and kinematic internal variables. A nonlocal formulation of the model is then proposed, based on the thermodynamic formulation of Borino et al. [BOR 99], [BEN 00]. While in [BEN 00] a kinematic internal variable was assumed as the primal nonlocal variable, in the model here proposed the primal nonlocal variables are the damage variables. Following [BOR 99], the nonlocality is transferred onto the conjugate variables which in the present case are the energy release rates, by means of an energy equivalence which allows one to eliminate the so called nonlocality residual [ERI81]. It is shown that, unlike in [BEN 00], the proposed model maintains the attractive feature that all constitutive computations can be performed locally, at the Gauss point level [PIJ 87], [COM 00], and that it gives rise to a symmetric finite element tangent stiffness matrix. A one-dimensional problem is studied for a simplified version of the model, with a single damage variable. The results obtained with the proposed dual nonlocal formulation and with the standard nonlocal formulation of [COM 00] are compared.
2. A "symmetric" isotropic local damage model Let e=e-l/3Ie v be the deviatoric part of the strain tensor e, ev being its volumetric part and I the second order identity tensor. The free energy density potential under isothermal conditions for the proposed damage model is defined as
where G0 and K0 are the initial elastic and shear moduli, respectively, dc and dK are shear and volumetric damage variables and ^ is a scalar variable of kinematic nature. The state equations defining the conjugate static variables are given by
where s = a-lp is the stress deviator and p = l/3crkk is the mean stress; % is a static internal variable and YG , YK represent the elastic energy release rates. The activation of damage is governed by the following activation functions and loading-unloading conditions
Aspects of Nonlocal Damage Analyses 79
7 being a scalar dissipation multiplier. The associative evolution equations are given by
Finally, the rate of dissipation density is given by
NOTE 1. - The presence of separate damage variables dG, dK adds flexibility to the model. The activation function may be defined in a form more suited for materials with non-symmetric tension-compression behavior like concrete and the separate evolution equations for deviatoric and volumetric damages allow for a varying Poisson's coefficient while preserving the isotropic nature of the model. NOTE 2. - The scalar internal variable accounts for material rearrangements at the microscale due to damage development. Damage is the only dissipation mechanism considered in this model. In finite element applications, the constitutive law is integrated within a timestep in the corrector phase of the iterative procedure, according to a Euler backward-difference scheme. This implies computing all derivatives in [4] at the end of the step. At the end of the corrector phase, a relation between stress and strain increments is implicitly obtained: Aa = Aa(As,). In the subsequent predictor phase, the consistent tangent elastic tensor is computed by differentiating this relation under the assumption of continuous loading in the increment, i.e.
with all quantities evaluated at the end of the step. Taking into account eqs. [43 and 23], one has
The explicit expression of the consistent tangent elastic tensor is then obtained as:
where II
denotes the fourth order symmetric identity tensor of components
3. Nonlocal version of the "symmetric" isotropic damage model The nonlocal version of the model is obtained substituting one of the constitutive variables by its weighted average over the whole domain Q. of the structure. The averaged quantity reflects the effect of the interaction at the microscale between the material point considered and the neighboring points. The decaying effect of the interaction with the distance is taken into account by the weighting function. In the literature, there exist several proposals concerning the choice of the nonlocal variable (see [JIR 98] and [GAN 99] for a recent discussion on the subject). From the computational standpoint, the most convenient choices are those which allow one to carry out the constitutive calculations locally at each Gauss point, without introducing any coupling at constitutive level as, e.g., in [PIJ 87] and [COM 00], where the strain invariants have been selected as nonlocal variables. In the present context, this would imply defining two nonlocal variables as follows
with:
lc being a material internal length related to the width of the localization zone. The particular definition of the weighting function W accounts for the effect of the boundary on the nonlocal interaction at the microscale and allows the reproduction in a simple way of a uniform field. In other words, if YG is constant over the body, it seems
Aspects of Nonlocal Damage Analyses
81
logic and desirable that also YG be uniform. However, the adopted definition of W in [10]| is such that Vy(x,s)= W(s,x). This lack of symmetry of the weighting function entails that also the consistent tangent operator is not symmetric for the nonlocal model [BAZ 88], [PIJ 95], [JIR 99] even if the consistent tangent operator of the underlying local model is symmetric. A more rigorous treatment of the boundary effect could be used to employ homogeneization techniques for periodic structures in the proximity of geometric boundaries (see e.g. [LEG 97]). The non-symmetric nonlocal version of model [l]-[5] is governed by eqs. [1], [2] and by the following activation conditions and evolution equations
A symmetric nonlocal formulation of the same local damage model can be achieved following the thermodynamic nonlocal approach of Borino et al. [BOR 99], [BEN 00]. An application of that theory to the present model which preserves the computational advantages of the above non-symmetric nonlocal model is obtained by assuming that the damage variables are the variables reflecting at the macroscale the microscopic interaction due to the heterogeneity of the material and, therefore, have to be considered nonlocal [BAZ 88]. Hence, one can set:
where:
The dissipation rate density takes the expression:
P being the so called nonlocality residual representing the energy exchanged between the considered material point and other points belonging to its interaction
82 Numerical Modelling in Damage Mechanics domain due to the intrinsic nonlocality of the developing damage mechanism. The fact that the system is thermodynamically isolated implies the following insulation condition [ERI 81], [BOR 99]
The insulation condition allows elimination of the nonlocality residual and to transfer the nonlocality onto the dual variables of the nonlocal damage variables defined in the model, i.e. the energy release rates YG and YK . One can write
From the insulation condition [15] it follows that-
Having in mind the definitions [13], from eq. [17] one obtains
where W * is the adjoint function of W, i.e.
Figure 1. a) One-dimensional weight functions W(x,s) and W*(x,s) for varying position x over a bar for l/L=0.2; (b) function \ W*(x,s) ds for varying characteristic length lc
83
A plot of the weight functions W and W* centered at various positions over a bar of length L is shown in Figure la where the influence of the boundary on the shape of the weight function is also evidenced. The activation function and the evolution equations [3] and [4] are now written in terms of the dual nonlocal variables as in the non-symmetric non-local model [11]
NOTE 3. - A dual nonlocal damage model based on the thermodynamic approach has been presented by Benvenuti et al. [BEN 00]. In their model, however, only the kinematic internal variable has a nonlocal nature, while the damage variable is local. While the issue of the most appropriate choice seems to be still open from the mechanical point of view, from the computational standpoint the definition of a nonlocal kinematic internal variable leads to a nonlocal constitutive problem in the corrector phase of a standard finite element implementation. On the contrary, the integration of eqs. [20] leads to the same local, and therefore computationally convenient, problem as in the non-symmetric model [11] [COM 00]. NOTE 4. - The weight function used for the definition of the dual nonlocal variables in [18] does not allow for the reproduction of a uniform field as, in general (see Figure lb),
However, in the absence of damage, it appears to be an obvious requirement that a uniform strain field generates a uniform field of strain energy release rates YG and YK. Therefore, in the applications, the weight function W in eq. [10]1 will be used for YG and YK, while the weight function W* in [19] will be used for dG and dK , satisfying in this way condition [17]. The algorithmic tangent matrix can be computed for the dual nonlocal model following the procedure proposed by Jirasek [JIR 99]. Let
be the vector of internal equivalent nodal forces and let u be the vector of nodal displacements in a finite element discretization. Let Ng be the total number of Gauss points used to carry out the numerical integration over all the elements in the mesh.
and let Nactg be the number of Gauss points where f(YG,YK, X) = 0 and y > 0 at the end of the correction phase. Let us also define the following quantities at Gauss points q and p
where A is defined in [7]2. One then obtains the symmetric elasto-damage tangent matrix Ked
where wp denotes the Gauss weight at Gauss point p and
represent the deviatoric and volumetric contributions of the same Gauss point to the initial undamaged stiffness matrix. In [25], D0 is the matrix of initial elastic moduli, BG and BK are compatibility matrices such that:
and PT = {1/3 1/3 1/3 0 0 0}. It should be noted that in [24] the index p runs over the whole set of Gauss points. This is because the global damage variable at a point varies as a consequence of the variation of the local damage at any point in the body. Thus, even though at a point one has y = 0 and the material point unloads elastically, at the same point one has d = 0 if there is at least one active point in the structure. By contrast, the index q runs only over the active Gauss points since it concerns the dependence of the nonlocal damage variables at point p on their corresponding local variables which are zero at inactive Gauss points. On the basis of these considerations and noting that, while Wpq = Wqp and W*pq = W*qp, one has WpqW*i = W^W^ , the symmetry of Ked can be easily assessed.
85
4. A simple nonlocal damage model To study the effects of the dual nonlocal regularization described in the previous Section a simple model, with only one damage variable d, is considered. Applications of the two damage variables model to concrete problems will be presented in a forthcoming paper. The simplified model is based on the following free energy density:
where DO is the undamaged elastic tensor and k, c and n are material parameters. The state equations are given by:
Figure 2. Stress-strain behavior for the simple damage model for varying n
The activation function, loading-unloading conditions and evolution equations are defined as:
and therefore the kinematic internal variable coincides with the damage variable d.
The local model is such that, in one dimension, the stress vanishes only asymptotically, for e > , but with a bounded fracture energy density. This can be seen by confining the model to one dimension. For e > E0, 0 being the strain at the linear elastic limit, from the condition/= 0, one has (Figure 2)
E denoting Young's modulus. The fracture energy density is defined as
If / denotes the integrand in [31]2, the boundedness of gf can be established noting that
The symmetric consistent tangent matrix can be computed for the simplified model following the same procedure as in the previous Section
5. One-dimensional numerical application The simplified damage model is used for the simulation of a tensile test on a prismatic bar. The problem data and geometry, together with the adopted meshes are shown in Figure 1. To trigger the damage localization, the elements at the left
Aspects of Nonlocal Damage Analyses 87 boundary have been slightly weakened. The problem has been solved adopting two nonlocal approaches: (model A) the non-symmetric approach of Comi [COM 00] based on the definition of nonlocal strain invariants (in this simple case coinciding with the energy release rate); (model B) the symmetric dual nonlocal approach of Borino et al. [BOR 99] in the form discussed in Section 3. Note that different values of lc have been adopted for the two models to obtain comparable damage accumulation in the part of the bar where unloading occurs after localization. The parameter lc can be identified using a back-analysis technique based on one dimensional tests where the width of the process zone is measured. Alternatively, an analytical approach can be pursued where lc is related to the length of the stationary harmonic localization wave (see e.g. [SLU 93]). This type of study has still to be carried out for the symmetric nonlocal model considered here.
Figure 3. One-dimensional test problem: geometry, adopted meshes and material data
As shown in Figure 4, both approaches provide an effective regularization of the problem as the results in terms of reaction force versus imposed displacement rapidly converge towards a mesh independent solution. From Figure 4, it appears that the dual regularization technique produces an initially more ductile response with a subsequent very steep drop of the reaction force. The displacement controlled analysis cannot proceed further due to a global snap-back behavior which is not observed in the analysis with model A regularization. Both behaviors can be observed in uniaxial tension tests depending on the material properties and testing conditions. Since the present numerical test does not simulate a physical experiment, it is not possible to assess which one of the two results is more realistic. The longitudinal strain evolution obtained by means of the two regularizations is shown in Figure 5. While model A regularization gives rise to a sharp strain localization, the model B technique produces a smoother profile with a much lower peak value developing at a significant distance from the boundary, where some elements have been weakened. This is a consequence of the effect of the boundary
88 Numerical Modelling in Damage Mechanics due to the particular shape of the weight function as already mentioned in Section 3 (see Figure 1).
Figure 4. Reaction per unit cross-section area versus imposed displacement with nonlocal models A and B: convergence with mesh refinement
Figure 5. Strain evolution for imposed displacement u: (a) model A; (b) model B
The difference is less pronounced in terms of local damage profiles, as illustrated in Figure 6a for a displacement u = 0.0168 mm. Again, with model B regularization the damage peak is offset with respect to the boundary. It should also be noted that for equal imposed displacement u, the model A regularization leads to a higher damage peak. The comparison between the local and nonlocal damages in model B analysis is shown in Figure 6b. It can be noted that the nonlocal damage presents a sharper peak though at almost the same value of the local one.
89
Finally, the stress profiles are shown in Figure 7. It turns out that the dual regularization has the beneficial effect of reducing the stress oscillation caused by the weighting process [JIR 99]. Furthermore, in both cases stress oscillation tends to decrease as the mesh is refined.
Figure 6. Imposed displacement u=0.0168 mm: (a) damage profiles d (x) for models A and B; (b) local d (x) and non local d (x) damage profiles for model B
Figure 7. Stress distributions along the bar for model A and model B upon mesh refinement
90 Numerical Modelling in Damage Mechanics 6. Conclusions The finite element implementation of a family of isotropic nonlocal damage models has been discussed. Attention has been focussed on the issue of the symmetry of the consistent tangent operator. A rather general isotropic local damage model based on two damage variables affecting separately the shear and bulk moduli has been presented. The explicit expression of the consistent tangent matrix has been derived and it has been shown that symmetry is obtained provided that associative evolutions are postulated for the damage and the internal variables. Then the model has been re-formulated as a nonlocal model following the approach proposed in [PIJ 87] and [COM 00] which consists of assuming as nonlocal variable the elastic energy release rate. This has the advantage that all constitutive calculations can be carried out separately at each Gauss point during the corrector phase of the standard finite element iterative procedure. The consistent tangent matrix for the considered nonlocal model is well known to be non-symmetric [BAZ 88], [JIR 99]. A nonlocal version of the same model, based on the thermodynamically founded nonlocal theory recently put forward by Borino et al. [BOR 99] and preserving the symmetry of the underlying local model has also been formulated. In this new version of the model, the nonlocal nature, originally conferred to the damage variables, is transferred to their conjugate variables, the energy release rates, on the basis of an energy equivalence which allows one to eliminate the so called nonlocality residual. The explicit expression of the finite element tangent stiffness matrix of the new nonlocal model has been derived and it has been shown to be symmetric. A one-dimensional test has been carried out for a simpler nonlocal model based on a single damage variable. The regularization property of the dual nonlocal formulation has been assessed even though the issue of the influence of the boundary conditions with the development of a significant boundary layer seems to deserve further consideration.
Acknowledgements This work has been carried out within the framework of the joint co-financing MURST and LSC-Politecnico of Milan program.
7. References [BAZ 88] BAANT Z.P., PUAUDIER-CABOT G., "Nonlocal continuum damage, localization instability and convergence", Journal of Applied Mechanics, vol. 55, 1988, p. 287-293.
91
[BEN 00] BENVENUTI E., BORINo G., TRALLI A., "A thermodynamically consistent non-local formulation for elasto-damaging materials: theory and computations", Proceedings of ECCOMAS 2000, Barcelona, Spain, 11-14 September 2000. [BOR 99] BORINO G., FUSCHI P., POLIZZOTTO C., "A thermodynamic approach to nonlocal plasticity and related variational principles", Journal of Applied MechanicsI, vol. 66, 1999, p. 952-963. [COM 00] COMI C., "A nonlocal model with tension and compression damage mechanisms", to appear in European Journal of Mechanics A/Solids, 2000. [ERI 81] ERINGEN A.C., "On nonlocal plasticity", International Journal of Engineering Science, vol. 19, 1981, p. 1461-1474. [GAN 99] GANGHOFFER J.F., SLUYS L.J., DE BORST R., "A reappraisal of nonlocal mechanics", European Journal of Mechanics A/Solids, vol. 18, 1999, p. 17-46. [GAN 00] GANGHOFFER J.F., DE BORST R., "A new framework in nonlocal mechanics", International Journal of Engineering Science, vol. 38, 2000, p. 453-486. [JIR 98] JIRASEKM., "Nonlocal models for damage and fracture: comparison of approaches", International Journal of Solids and Structures, vol. 35, 1998, p. 4133-4145. [JIR 99] JIRASEK M., "Computational aspects of nonlocal models", Proceedings of ECCM 99, Munchen, Germany, August 31-September 3, 1999. [LEG 97] LEGUILLON D., "Comparison of mached asymptotics, multiple scalings and averages in homogenization of periodic structures", Math. Models Meth. Appl. Sci., vol. 7, 1997, p. 663-680. [PIJ 87] PUAUDIER-CABOT G., BAZANT Z.P., "Non local damage theory", Journal of Engineering Mechanics, vol. 113, 1987, p. 1512-1533. [PIJ 95] PUAUDIER-CABOT G., "Non local damage", in Continuum Models for Materials with Microstructure, H.-B. Muhlhaus (ed.), New York, Wiley, 1995, p. 105-143. [SLU 93] SLUYS L.J., DE BORST R., MUHLHAUS H.-B., "Wave propagation, localization and dispersion in a gradient-dependent medium", Int. J. Solids Struct., vol. 30, 1993, p. 11531171. [STR 96] STROMBERG L. RISTINMAA M., "FE-formulation of a nonlocal plasticity theory", Computer Methods in Applied Mechanics and Engineering, vol. 136, 1996, p. 127-144.
Chapter 6
Computational Issues and Applications for 3D Anisotropic Damage Modelling: Coupling Effects of Damage and Frictional Sliding
Damien Halm and Andre Dragon
Laboratoire de Mecanique et de Physique des Materiaux, Ecole Nationale Superieure de Mecanique et d'Aerotechnique, France
Pierre Badel
Electricite de France, Division Recherche et Developpement, France
3D Anisotropic Damage Modelling
95
1. Introduction This paper addresses some issues concerning the modelling of the behaviour of quasi-brittle materials, comprising some rocks, concrete, ceramics, etc. These materials share the same damage process, namely the generation and growth of decohesion mesosurfaces (mesocracks). This phenomenon induces a degradation of the effective properties of the material. Besides, the generally oriented nature of flaws gives rise to a number of characteristic events such as induced anisotropy, volumetric dilatancy, irreversible stress/strain effects, dissymmetry between tension and compression, unilateral behaviour due to crack opening/closure, dissipative frictional sliding on closed mesocrack lips, etc. The purpose of this paper is to summarize most salient features of a model capable of taking into account most of the above phenomena concentrating specially on its numerical implementation and applications for a set of engineering problems concerning concrete structures. The postulate of combining both physical pertinence and numerical simplicity led the authors to search a third way between micromechanical and phenomenological approaches: the former propose an accurate picture of the real mechanisms but their use is frequently limited to particular loading paths due to inherent complexities encountered; the latter are generally designed to be easily implanted in finite element codes but suffer from a lack of physical motivation. Section 2 of this paper describes a 3D damage model by mesocrack growth, originally proposed by Dragon [DRA 94], and recently developed by Halm and Dragon [HAL 96], [HAL 98]. Its particularity lies in its modular nature, with two main parts: - A first step deals with the modelling of the mesocrack growth as well as with the moduli recovery phenomenon due to crack closure (unilateral effect). The emphasis is put on the stress continuity requirement when passing from open to closed cracks (and vice versa). Thus, f. ex., tension-compression cycles can be modelled. -The second level couples damage with a second dissipative phenomenon, namely frictional sliding on closed mesocracks and allows to simulate more complex loading paths (torsion, f. ex.). The purpose of the model depicted in Section 2 is to provide an efficient tool for resolving boundary-value problems involving non linear behaviour of quasi-brittle solids. Thus, great care is taken with the accuracy and simplicity of the numerical integration scheme related to both independent mechanisms as well as to the coupled model. It is worth noting that the use of an implicit integration scheme for damage leads to the resolution of a linear equation, while classical elastoplastic models require more complex numerical treatment. Moreover the low degree of coupling between the two equations governing respectively damage and sliding evolutions avoids having to solve an intricate non linear system. Details are given in Section 3. In order to illustrate the pertinence of the coupled model and the efficiency of the integration algorithm, the constitutive equations have been introduced in
Code_Aster, the Finite Element code developed by Electricite de France. Section 4 provides comments on some boundary-value problems underscoring the applicability of the model for efficient structural analyses of concrete structures.
2. Anisotropic damage and sliding model This section outlines the salient features of the anisotropic damage model by Dragon et al. [DRA 94], [HAL 96], [HAL 98]. The particularity of this model lies in its modular structure, each part dealing with a given dissipative mechanism: damage by mesocrack growth (with unilateral behaviour) and frictional sliding on closed mesocrack lips. The behaviour of the mesocracked material is assumed to be rateindependent, isothermal and restrained to small strain.
2.1. Damage by mesocrack growth and unilateral behaviour The model at stake here aims at describing the progressive mesocrack-induced anisotropic degradation and related behaviour of elastic quasi-brittle solids. It is based on a series of assumptions combining micromechanical considerations and macroscopic formulation: (i) Damage is described by a single internal variable, a second-order tensor D conveying information on crack orientation:
where ni stands for the normal of the i-th set of parallel cracks and d(i)(S) is a dimensionless scalar function proportional to the extent S of decohesion. The form [1] derives from micromechanical considerations [KAC 92]. From a macroscopic point of view, Onat and Leckie [ONA 88] prove that D must be an even function of ni, and then at least quadratic. The spectral decomposition of D leads to:
Expression [2] can be macroscopically interpreted as follows: any system of microcracks can be reduced to three equivalent orthogonal sets of cracks characterized by densities Dk and normal vectors vk. NOTE - Unlike the case of 1-d models (the value of d is then bounded by 0 and 1), values of the Dy-components within the relative tensorial representation cannot be straightforwardly interpreted in the same simplistic manner. In fact, when considering the scalar dimensionless density function d'(S) as a part of the
97
micromechanical interpretation of the damage tensor D, one can - for a particular nature of defects considered (e.g. penny-shaped microcracks) - interpret d'(S) in terms of the conventional crack density
In such case, d(S) can
theoretically vary within the interval [0,1]. So, one can state that Dij-components values take their micromechanically licit values in the interval [0,1] while the effective control of the evolution equations (including their algorithmic management) and local instability phenomena generated by the CDM model put effective limits well below this conceptual absolute bound of unity. That is why such a damage model has to be associated with tools of detection of relevant local instabilities (i.e. localisation bifurcation) in the context of computational algorithms for efficient structural analysis. This association has been achieved for the first level of the model (frictionless damage model without unilateral behaviour), see [DRA 94]: it allows one to correctly predict the incipience of localisation phenomena within 3D framework. The localisation detection is not treated in this paper. (ii) Micromechanical studies [KAC 92] show that 3D damage configurations should be rigorously described not by the single variable D_[l], but by two damage parameters, namely D and its extension to the fourth-order D :
However, when cracks are open, the influence of D can be neglected and the single variable D appears sufficient to model the degradation of solids containing cracks. Under compressive loading, favourably oriented cracks may close,Jeading to an elastic moduli recovery phenomenon. In this case, the contribution of D into the overall elastic properties can no longer be neglected. In order to maintain the macroscopic interpretation [2], the complementary fourth-order entity (named D) necessary to account for the unilateral effectjs directly built with the eigenvalues and eigenvectors of D and slightly differs from D :
Note that there is no new information in D with respect to D, so D is not considered as a new damage variable. (iii) One assumes the existence of a thermodynamic potential (free energy per unit volume w), function of strain e, damage D and the fourth-order damage parameter D, and generating a form of elastic orthotropy for D * 0, in connection with the three eigensystems [2]. Assuming linear elasticity and non interaction between cracks, the tensorial functions representation theory [BOE 78] gives the general form of the terms entering w(e,D, D (D)):
H stands for the classical Heaviside function and activates or deactivates the D term depending on whether the k-th equivalent set of mesocracks is open (vk.e. v'SO) or closed (vk.e. vk<0). The proof for the form of the opening/closure criterion vk.e. vk=0 can be found in [HAL 96]. A, and )H are the classical Lame constants; a and 3 are material constants related to modified elastic moduli for a given damage state. The factor (oc+2|3) in front of the D-term is obtained by assuming a total stiffness recovery in the direction normal to the closed crack. The linear term, reading g tr(e.D), generates residual phenomena for D*0. The elastic stress a and the damage thermodynamic force FD are determined by partial derivation:
The forms of w,CTand F respect the continuity conditions for multilinear elasticity [CUR 95], so that these functions remain continuous despite the presence of H. (iv) The evolution of D, corresponding to the brittle, splitting-like crack kinetics, has been found to follow the normality rule with respect to a criterion in the space of components of the proper thermodynamic force FD. The damage evolution is thus apparently following the principle of maximum dissipation and is related here to tensile (positive) straining e* and to actual damage pattern. It should be stressed however that the particular damage criterion proposed in [DRA 94] f(FD,D)<0 is explicitly dependent on the part FD1+ = -ge+ = FD-FD2-FD1" of the driving force FD. FD1 is the strain energy release rate term related to residual effects: FD1 = -ge, FD2 represents the remaining recoverable energy release rate. The former term is decomposed into the splitting part FD1+ = -g+, 6+ = P+:e, with P+ a positive fourthorder projection operator selecting positive eigenvalues from strain, and the nonsplitting part FDI" = -g(e-e+). The damage criterion and rate-independent damage evolution law are thus as follows:
99
Note that the damage model including the unilateral effect necessitates the identification of eight material constants only, which can be relatively easily determined, see f. ex. [HAL 01].
2.2. Fractional sliding on closed mesocrack lips Even if it takes into account the unilateral effect the previous model does not restore the shear moduli when cracks close, assuming thus that cracks are perfectly lubricated. Because of the roughness of the crack lips and the consecutive friction, this assumption appears too strong: experimental data involving loading-unloading cycles for specimens undergoing frictional sliding on the lips (torsional tests for example) exhibit a shear moduli recovery in the direction parallel to the crack plane, due to blocking of crack lips displacement. The work by Gambarotta and Lagomarsino [GAM 93] proposes a 3D micromechanical model for this phenomenon which constitutes progress with respect to some earlier 2D attempts. This section provides a macroscopic formulation suitable for boundary-value problems involving frictional sliding. It is built within the same thermodynamic framework as for damage and is based on following hypotheses: (i) Sliding occurs within the crack plane. A micromechanical study ([KAC 92], considering this time that crack displacement has no opening component) leads to the following possible expression for the sliding variable:
S1 stands for the cracked surface of the i-th set of parallel mesocracks of normal n , ' the sliding following the direction g1, V the representative volume element. As the influence of D reduces to that of three equivalent sets according to [2], y can be written in the analogous manner:
1
where vk, k = 1,2,3 are the D-eigenvectors. (ii) Frictional blocking induces a macroscopic recovery of the shear moduli. In expression [3], the degradation of the shear moduli is related to the p-term in the first line. Invariants involving Y will thus replace the previous (3-term in the free energy w(e,D,y) taking into account this additional dissipative phenomenon. Due to the particular structure of D and y and the fact that only simultaneous (Y,D)invariants enter w, two additional invariants convey useful information: tr(e.y.D) and tr(y.Y.D). (iii) According to the points (i) and (ii), the following expression is proposed for the free energy of the solid containing sliding cracks:
with L k =v k v k v k v k and D k = D k v k v k . The coefficients 4(3 and -2(3 in the last line have been calculated by assuming: (1) the continuity between the expressions of w corresponding respectively to open and closed cracks, (2) sliding Y is equal to the strain e in the crack plane at the very closure moment. The elastic stress as well as the thermodynamic force related to D contains the contribution of each equivalent set (open or closed, sliding or blocked):
The thermodynamic force related to sliding concerning a particular equivalent set

is:
(iv) The model considers frictional non-sliding/sliding phenomena on mesocrack lips on a macroscopic scale by an approach similar to that to damage. Although it is widely employed in many models [HOR 83], [GAM 93], etc., the Coulomb's criterion
101
is not suitable in this context because of its micromechanical formulation. The pertinent quantity governing sliding on an equivalent system k is the thermodynamic force FYk (which can be physically interpreted as the sliding energy release rate). One assumes that the sliding criterion explicitly depends on the norm of the tangential part F^1* of the force F^k and on the normal strain vk.e.vk. Unlike Coulomb's law, the normality rule with respect to the function defining the reversibility domain has been found to keep a strong physical sense: it indicates a connection between y and F^ indicating that sliding occurs in the crack plane (as long as damage axes do not rotate). The sliding convex reversibility domain hk can be written as:
where p is a friction coefficient in the sense meant by the above thermodynamic force (tangential component) - normal strain relationship, and:
The normality rule gives:
2.3. Damage and sliding coupling The both dissipative phenomena (damage and frictional sliding) described independently in the previous paragraphs may occur simultaneously under particular loading paths. One assumes that the splitting-like kinetics considered in Paragraph 2.1. is still valid for closed sliding cracks even when they branch: after a short transitional distance, cracks tend to grow perpendicularly to positive principal strain direction (see, f. ex., [BAR 97]). The sliding evolution law needs a rewriting, especially when the principal axes of Dk rotate (for Dk-non-proportional loading paths): in this case, sliding tends to depart from the crack plane and thus the driving force for sliding has to incorporate not only the tangential part F7 of Fyk but also a fraction of the normal part FyNk. Let be the following partition of FYk:
Fk is the appropriate part of FYk to enter the expression of the criterion hk taking into account Dk-axes rotation and avoiding discontinuities when cracks open [HAL 98]. Note that in the case of proportional loading paths (i.e. yk:Dk=0), the term Fk reduces to F^.
The normality assumption leads to:
NOTE - It was obvious to the present authors that controlling the following effects: (i) damage-induced anisotropy, (ii) damage related volumetric dilatancy, (iii) damage related residual effects, (iv) rigorous 3D treatment of the unilateral problem, (v) idem for the frictional resistance and sliding effect for closed microcracks involving the dissipative coupling with damage, should first have been embraced within the framework of classical local approach. The non locality of constitutive equations, which can be now postulated for particular purposeful aspects of the model, would allow enlarging its domain of pertinence by e.g. casting the underlying hypothesis of non interacting microcrack in the actual one and treat the problems of clustering and related enhancement vs. shielding microcracks interactions. This is planned as further work.
3. Numerical treatment In order to treat complex boundary-value problems, an accurate numerical tool has to be associated with the previous model. The strong non linearity of the damage and frictional sliding mechanisms requires a time integration algorithm for the evolution of the damage variable D and of the sliding variable Y- This section summarises the local (i.e. for each integration point of a finite element discretization) integration scheme for both evolution laws [4] and [7], After dealing separately with damage and sliding mechanisms respectively, the coupling of both is considered.
3.1. Local integration for the damage model Let I be the time interval [0,T], with the partition I = U= [t r ,t r + 1 ]. Given the mechanical state q, = (e^D^rA) at time tr and the prescribed strain increment Ae (such as eT+1 = 8T + Ae, the integration problem amounts to calculating the stat qr+i = (er+i,Dr+1,Ykr+i,ar+1). Since only damage evolution is concerned in this paragraph, Y* is considered constant (ykr = Y'H-I)- The tensors eî, Dr+1 and or+1 are determined by:
3D Anisotropic Damage Modelling 103
with G0 standing for relation [5]. This calculation comprises two steps, (i) Elastic prediction: First, the increment is assumed elastic, i.e. AD = 0. One checks whether the mechanical state (e^Dr) meets the condition:
If [9] is satisfied, the elastic prediction coincides with the solution of the problem. Then,
and CH-I is calculated by [8]. Otherwise, if f(er+i,Dr)>0, the mechanical state has to be corrected in order to determine the increment AD. (ii) Non linear correction: The incremental formulation leads to the following formulation for the damage evolution:
The increment AD depends on e+r+i and Dr+1, i.e. the value of e+ and D at the end of the integration interval [tr,tr+i]. This assumption corresponds to a fully implicit integration scheme, which is known for its unconditional stability [ORT 85] whatever Ae is. It is worth noting that unlike for most of elastoplastic models, the implicit scheme is well adapted to the damage model presented in Section 2.1: in fact, the damage multiplier increment AAD is obtained by solving the criterion f(e+r+i.Dr+i) = 0 which reduces to a linear equation whose solution is:
104 Numerical Modelling in Damage Mechanics 3.2. Local integration for the sliding model In this paragraph, damage is assumed constant (Dr+i = Dr). The sliding integration follows the same scheme as for damage. (i) Elastic prediction: The step is assumed elastic (y k r +i = Y k r )- The damage eigenvalues Dkr+1 and eigenvectors vkr+i are known, so the value of Fkr+i entering h is given by:
Then the value of hk is checked:
If hkr+1<0, y'V+i = Ykr and the mechanical state is fully determined. Otherwise (if h r+|>0), the frictional sliding evolution undergoes the following correction. (ii) Non linear correction: The sliding increment Ayk is obtained by solving the following system:
k
Equations [10] and [11] stand respectively for equations [6] and [7]. Again, this system corresponds to an implicit integration scheme. But while the damage integration reduced to a linear equation, the above system remains non-linear and its solving necessitates a Newton-Raphson algorithm.
3.3. Fully coupled model Both above dissipative mechanisms (damage and frictional sliding) may occur simultaneously along particular loading paths. The problem is then to determine the coupled increments AD and Ayk simultaneously. The integration is here facilitated by the low degree of effective connection between f on one hand and hk on the other: whereas hk is a function of D and y\ f only depends on D and the equation f = 0 can be solved without explicit reference to sliding. The general algorithm is as follows: (1) The value of vk.e.vk of the normal strain for each equivalent microcrack system is checked.
105
(2) If vk.e.vk>0, the corresponding system is open; sliding does not occur and AD is calculated as described in Paragraph 2.1. (3) If vk.e.vk<0, the corresponding system is closed and may slide. Both criteria f <0, hk<0 are checked; AD and Ayk are calculated, if necessary, by solving successively f = 0 for the damage evolution and later hk= 0 for the sliding one. NOTE- The numerical algorithm is apparently standard and this constitutes paradoxically a non negligible contribution: the model deals with two strongly coupled dissipative phenomena (damage by mesocrack growth, frictional sliding) with only nine material constants. However, the particular structure of the equations governing the evolution of the internal variables, optimised somewhat by the modelling procedure, allows a classical backward difference time integration scheme to be efficient enough in spite of the complexity of the mechanisms at stake.
4. Numerical example This section presents an application of the model for structural analysis. The three major phenomena, i.e. degradation, unilateral effect and frictional sliding are illustrated. A numerical simulation of boundary-value problem requires an efficient tool including a reliable model (in the case of damage model with unilateral effect, great care must be taken of the continuity of the response) and of an efficient integration scheme. With the implicit scheme used, the tangent operator has a great influence on the time needed by the simulation.
4.1. Geometry and loads The numerical test is carried out on a slab with a symmetrical double edge notch. The material constants are given in Table 1. This set has been identified for a Fontainebleau sandstone. The geometry of the specimen is described in Figure 1. The structure is constrained against x- and y-displacements along the lower edge Ar A2 and the lower half A2-B2 of the right-hand side. Stress is applied on the top face As-A4 and the upper half of the left-hand side A4-B4 via sheets considered to be infinitely rigid that are stuck to the test specimen. The upper face A3-A4 is maintained in the horizontal position and the left face A4-B4 in the vertical position. The test specimen is first subjected to a positive displacement of A3-A4, then a compressive force Pn is applied; finally, in addition to the compressive force, a shearing force Ps is applied. The slab is meshed by QUA4 elements under the hypothesis of plane stress. The mesh concerned appears in Figure 2: while relatively rough, it has been refined in the critical areas, i.e. in the central band and around the notches.
Figure 1. Geometry of the test specimen
X (MPa)
26250
H (MPa)
a (MPa) 1 900
P (MPa) -20 400
g (MPa) Co (MPa) Ci (MPa)
B(l)
0
p (MPa)
17500
-110
0.001
0.55
2500
Table 1. Constitutive parameters 4.2. Mesocrack growth Figure 2 shows a damage map during the first stage of the loading history, i.e. tension (by displacement imposed) on A3-A4. More precisely, the damage presented here is the component Dyy of the damage tensor. Due to the strongly brittle behaviour of sandstone, damage rapidly localizes around the notches for a low level of Dyy (maximum about 0.12) and becomes quasi-negligible in the central slab section.
107
Figure 2. Dyy map 4.3. Unilateral effect The unilateral effect is observed in Figure 3, (force Pn vs. difference of the vertical displacement of the two edge points in the right notch). After degradation in the first stage of the loading history, the unloading stage exhibits two major effects: first the appearence of residual strain-like quantity 5 for Pn = 0 and second the stiffening of the material caused by the mesocrack closure.
Figure 3. Pn vs. 8 (difference of the vertical displacement of the two edge points in the right notch)
4.4. Frictional sliding The first stage of the loading generates microcracks, principally concentrated around the notches. After crack closure (stage 2), a shear loading is applied. Figure 4a shows the intensity of the frictional sliding: even if numerous microcracks are located close to the notches, one observes a high density of sliding cracks within a band crossing the sample. Even the central zone, damaged to slighter degree than near-notch zones, is affected by this effect. However, due to the very low level of damage, the incipience of frictional sliding does not notably influence the distribution and the level of the stress (f.ex., Von Mises stress, Figure 4b): the difference between the values of Von Mises stress with or without frictional sliding does not exceed a few percent. The damage localization phenomenon acts here as an inhibitor for the sliding mechanism. This precocious influence of localization is corroborated by recent works [GIR 00]. Frictional sliding may influence more drastically the stress distribution for more ductile materials such as some concretes. Further work (some of which being under way) deals with this subject.
Figure 4a. Sliding Yxy map
Figure 4b. Von Mises stress map
5. Conclusion The purpose of the model depicted in this paper is to provide the engineer with an efficient while physically motivated tool for structural analysis: it seems that a reasonable compromise has been found between the pertinence of the 3D theoretical formalism and its applicability for industrial boundary-value problems. The constitutive equations and the required continuity of the stress-strain response stem from the tensor functions representation theory and the multilinear functions theory, the latter applied to managing unilateral effects linked to damage deactivation. Although two coupled dissipative phenomena - mesocrack growth and frictional sliding on closed mesocrack lips - are considered, the low degree of numerical coupling between the respective equations governing these two mechanisms allows convenient algorithmic treatment and FE implementation; an example of application of the model shows its capacity to
109
illustrate the mesocrack growth and the recovery of effective properties. Further works attempt to clarify whether the localization effects are premature compared to intrinsic material and structural response, i.e. whether they represent a specific excessive model tendency to be amended. It could be done, f. ex., by introducing some rate-dependance into the model which would in this manner account for genuine viscosity of engineering materials like concrete and would contribute as a regularizing factor for numerical calculations. Another aim of prospective short term research is to quantify on a broader basis the effects induced by frictional sliding.
6. References [BAR 97] BARQUINS M., CHAKER C., PETIT J.P., Influence du frottement sur le branchement de fissures a partir de defauts obliques soumis a une compression uniaxiale , C.R. Acad. Sci. Paris, L 324, Sene lib, 1997, p. 29-36. [BOE 78] BOEHLER J.P., Lois de comportement anisotrope des milieux continus , J. Meca., Vol. 17, 1978, p. 153-190. [CUR 95] CURNIER A., HE Q., ZYSSET P., Conewise linear elastic materials , J. Elasticity, Vol. 37, 1995, p. 1-38. [DRA 94] DRAGON A., CORMERY F., DESOYER T., HALM D., Localized failure analysis using damage models , [dans] Localization and bifurcation theory for soils and rocks, Ed Chambon R. et al., Balkema, Rotterdam, 1994, p. 127-140. [GAM 93] GAMBAROTTA L., LAGOMARSINO S., A microcrack damage model for brittle materials , Int. J. Solids Structures, Vol. 30, 1993, p. 177-198. [GIR 00] GIRAUD F., Etude de validation de la loi de comportement elasto-endommageable du Pr Dragon, IFP-Report 53 185, 2000. [HAL 96] HALM D., DRAGON A., A model of anisotropic damage by mesocrack growth; unilateral effect , Int. J. Damage Mech., Vol. 5, 1996, p. 384-402. [HAL 98] HALM D., DRAGON A., An anisotropic model of damage and frictional sliding for brittle materials , Ear. J. Mech. A/Solids, Vol. 17, 1998, p. 439-460. [HAL 01] HALM D., DRAGON A., Modelisation de I'endommagement par mesofissuration du granite , Revue Francaise de Genie Civil, soumis. [HOR 83] HORII H., NEMAT-NASSER S., Overall moduli of solids with microcracks: loadinduced anisotropy , J. Mech. Phys. Solids, Vol. 31, 1983, p. 151-171. [KAC 92] KACHANOV M., Effective elastic properties of cracked solids.-critical review of some basic concepts , ASME Appl. Mech. Rev., Vol. 45, 1992, p. 304-335. [ONA 88] ONAT E.T., LECKIE F.A., Representation of mechanical behavior in the presence of changing internal structures J. Appl. Mech., Vol. 55, 1988, p. 1-10. [ORT 85] ORTTZ M., POPOV E.P., Accuracy and stability of integration algorithms for elastoplastic constitutive relations , Int. J. Num. Meth. Eng., Vol. 21, 1985, p. 1561-1576.
Chapter 7
Energy Dissipation Regarding Transient Response of Concrete Structures: Constitutive Equations Coupling Damage and Friction
Frederic Ragueneau and Jacky Mazars
Laboratoire de Mecanique et Technologic, ENS de Cachan, Universite Pierre et Marie Curie, Cachan, France
Christian La Borderie
LaSAGeC, ISA du BTP, Universite de Pan et des Pays de I'A dour, Anglet, France
Energy Dissipation Regarding Concrete Structures
113
1. Introduction Transient non-linear computations of structures need the use of energy dissipation tools for both physical and numerical reasons. One of the major drawbacks in such analysis lies in the expression of the damping matrix. Several kinds of methods can be used to achieve that purpose, such as viscous or hysteretic [BAT 82]. A more realistic approach consists of better modelling of the internal dissipation. Recent experiments on reinforced concrete mock-ups subjected to seismic loading permitted appreciation of the strong interaction between state of failure and resulting global damping [QUE 98]. Most of the constitutive models are able to reproduce realistically the behaviour of concrete in the non-linear range, based on damage mechanics, on plasticity theory or using the microplane concept [JU 89] [KRA 81]. They often ensure predictive computations in the static case but do not easily take into account a main cyclic characteristic: the influence of heterogeneities and roughness of the crack surfaces. At a fixed level of damage, concrete still exhibits dissipation due to the frictional sliding between crack surfaces. This property can be experimentally observed for a specimen during cyclic solicitations through the hysteresis loops. A new constitutive relation for concrete material including residual hysteretic loops at a fixed level of damage is proposed. Derived from thermodynamics, it allows coupling of the state of cracking with the hysteretic dissipation induced by the crack surfaces sliding. Based on damage mechanics, a particular Helmholtz free energy allows introduction of a coupling of the level of damage in one direction to a frictional stress. The main assumption postulated to describe the hysteretic behaviour is that cracks surfaces created after fracture will not open anymore following a perfect surface but will slide depending on their roughness. This phenomenon induces the occurrence of a sliding stress which prevents the crack from going on opening easily. In another sense, the consumed energy during cracking is not entirely dissipated but part of it is stored in the sliding potential. This frictional stress (os) is assumed to have a plasticity-like behaviour associated with a non-linear kinematic hardening. A particular dissipative potential allows the description of dilatancy, fundamental feature of geo-materials like concrete, sand or rocks. This model has been implemented in the finite element code EFICOS based on a multilayered beam kinematics approach. Concerning the constitutive equations, we chose the implicit return mapping algorithm. This approach of structural computation allows one to take into account at the local level refined material models, avoiding at the same time heavy analysis. In order to validate our approach of structural dissipation, comparisons between computations and experiments are made. This communication focuses on the 3D expression of the local model and on numerical comparisons of energy dissipation at the structural level pointing out the influence of frictional sliding regarding global damping.
2. Crack growth and damage 2.1. State potential Initially introduced by Kachanov [KAC 58] for creep failure problems, damage mechanics formulation requires the addition of a new internal variable in order to represent the macroscopic loss of stiffness [LEM 90]. This can be achieved in many ways. The classical one is to relate the damaged material's and the intact material's elastic properties. Defining a damage variable requires reducing the rank of damage operator while maintaining as well as can be the experimental properties of the material. In order to reach this objective, the effective stress concept is introduced. It induces a relation between the stress tensor and the effective stress tensor through a strain equivalence (the effective stress which, when applied to the undamaged material, produces the same strain). Therefore, in the basic form, a scalar variable can be used for the sake of simplicity. Numerous authors have proposed several expressions of different ranks for the damage variable. The fourth-order tensor was suggested by Chaboche. Second-order tensors are more frequently introduced [DRA 94]. In this case, a problem exists in that the symmetry of the elasticity operator is no longer ensured, as it depends on the way one defines the previous relation [MUR 78]. Cordebois [COR 79] postulated a state potential which depends only on the effective stress. Therefore, the symmetry of the elasticity operator is obtained through this energy equivalence. The pursuit of a physical and realistic description of the oriented crack growth in concrete without neglecting the simplicity requirement led to a second-order damage tensor formulation. In order to make the subsequent numerical implementation of the model in a finite element code easier, Helmholtz's strainbased free energy has been chosen. Considering a particular definition for the damage variable d, an effective strain tensor is defined on the principal axis of the damage tensor as follows,
Where e is the effective strain tensor, e the second order strain tensor (symmetric part of the displacement gradient field) and d the symmetric second order tensor related to damage phenomena. Because of symmetry conditions on the resulting stress tensor, the expression in the previous relation denotes a symmetrization. This effective strain, introduced directly into the state potential, allows description of an elastodamage material exhibiting orthotropic cracks:
115
with (j, and A,, the Lame coefficients defined for the undamaged material, p is the material density and \f/d is the Helmoltz free energy.
2.2. Damage criteria and evolution laws Based on experimental investigations, damage for brittle materials such as concrete is governed principally by their tensile behavior. To take into account this asymmetry, two damage tensors must be introduced. The splitting between the tensile and the compressive damage tensors is achieved through the sign of the sliding strains (defined in the next section) expressed in their respective principal directions:
(tensorial functions), A is the tensor of strain eigenvalues and P the transformation matrix. H is the Heaviside function. The level of damage is governed by the value of positive strains. The evolution equation assumed to be expressed in its incremental form as follows, is written in the principal axes of the incremental strains:
Bt is a material parameter driving the slope of the softening branch. The associated damage criterion is also expressed in the principal axis of the strain tensor:
edo is the initial tensile yield strain, usually equal to 1.10"04. K"(,) is the hardening variable. Such a criterion is similar to the so-called St-Venant's criterion in the principal stress space. The use of a damage evolution law based on the state of total strain is consistent with the meaning of the damage variable expressed in the elasticity law. Compressive damage in a particular direction is considered only as a consequence of the tensile behavior of the material and, therefore, is taken equal to a function of the state of tensile cracking along the orthogonal directions.
ft is a material parameter connecting the damaged Young's moduli for two orthogonal directions. The comparison of apparent Young's moduli in the longitudinal and radial directions allows the measurement of p.
3. Inelasticity and friction coupled to damage 3.1. State potential The main assumption used to describe the hysteretic behavior is that crack lips arising from fracture do not open further along a perfect surface but slide depending on their rugosity. This phenomenon induces a sliding stress which prevents the crack from continuing to open easily. In other words, the energy consumed during cracking is not entirely dissipated but part of it is stored in the sliding potential. Thus, it is through the damage variable that this energy shift can be obtained. Each dissipative nonlinear phenomena needing its own internal variable, a measure of the sliding with friction will be defined thanks to a particular second order new variable: the sliding strain es. Following the same methodology as in section 2, sliding is integrated into the behavior through an equivalent strain which couples damage and elasticity of the sliding surface.
In that way, the total state potential is written as follows,
a is the internal variable associated with the kinematic hardening phenomenon and b is a material parameter. One can easily recognize a classical elasto-damage coupling and a new term allowing the energy to shift from the elasto-damageable part to the frictional sliding part. The coupling between sliding and cracking is made possible thanks to the presence of the damage variable as a multiplier in the second element of the righthand side of [8].
117
3.2. State laws In order to define the state laws, the model has to be thermodynamically admissible: it must comply with the positiveness of the dissipated energy. Starting from the Clausius-Duhem inequality, and assuming that the state potential is linearized around the current value of every state variable, the state laws are expressed as follows: The stress tensors can be derived as
And the sliding stress tensor, associated to the sliding strain:
The back stress is defined as:
We can observe that the total stress is divided into two parts: a classical elastodamage component and a sliding component. Damage is classically controlled by the elasto-damage stress and the sliding strain is linked only to the sliding part of the stress. This kind of partitioning, in conjunction with the two failure surfaces, allows the description of a hysteretic behavior at a fixed level of damage. Details of the complementary and evolution laws will be developed in the next section. At this description level, such an approach could be compared to multi-surface modeling [MRO 67], except for the fact that the surfaces are not expressed in the same space (strain space for damage and stress space for sliding). The sliding and plastic strains being different, the thermodynamic forces associated with the total strain and the sliding strain are different. Such a formulation differs greatly from the classical plasticity-damage coupling. This choice of introducing damage into the sliding stress is guided by the idea that all inelastic phenomena in concrete result from the cracks' growth. The thermal aspects are presently ignored.
3.3. Sliding criteria and plastic potential The sliding part of the constitutive relation is assumed to represent a plasticitylike behavior. In order to reproduce the hysteresis loops, nonlinear kinematic
hardening is considered. Initially introduced by Armstrong & Frederick [ARM 66] and recently developed further [CHA 93], it allows the formulation to overcome the major drawback of Prager's kinematic hardening law, i.e. the linearity of the state law defining the forces associated with kinematic hardening. The nonlinear terms are added in the dissipative potential. The sliding criterion takes the classical form:
J2\0SX) i-e. Von Mises' equivalent stress was chosen to a first approximation in order to keep simplicity and adequacy to the classical numerical algorithm for constitutive laws implementation. The specific aspects of geomaterials non-linear behaviour have been introduced in the plastic potentials Classical plasticity, in order to govern the evolution of the internal variables, requires the definition of a dissipative potential. The expectation of nonlinear kinematic hardening imposes the use of a non-associated flow rule:
with /! = o Tr[as], the first invariant of the sliding stress tensor. It enables one to take into account the dilatancy phenomenon observed experimentally on geomaterials such as concrete or rocks, a and c are material parameters. As a result of the normality rules, the evolution laws of the internal variables are expressed as follows, thus ensuring that the dissipation is positive:
A is the plastic multiplier, determined by the consistency condition.
3.4. Unilateral effects In the case of cyclic loading, a model has to take into account the crack closure phenomenon which generates the concurrent stiffness recovery. Some damage models have been extended to take into account the crack closure conditions by inducing different behaviors in tension and in compression. A general and rigorous framework concerning the unilateral condition, based on a particular decomposition of the elastic energy introducing a deviatoric behavior and a spherical one, is given by Ladeveze [LAD 83]. Other kinds of anisotropic damage models based on a mesocrack description can also deal with the crack closure condition [HAL 96]. In our case, the need to couple damage with sliding of the crack surfaces led us to the introduction of a second-order damage tensor. This was achieved through a
119
particular effective strain [1] [7] introduced into the elastic free energy for the damage part as well as for the sliding part. This enables a physical description of the material behavior (oriented cracking, inelasticity, nonlinear unloading) but is not in good agreement with the unilateral condition. This is due to the presence of tensorial damage in the volumetric part of the stress [9]. Concerning the radial loading cases in our analysis, a simplification of the previous model has to be achieved. The same formulation is retained, but an isotropic evolution of the damage tensors is now assumed. For the damage part, the simplified model boils down to two scalar damage variables allowing one to account for stiffness restoring after tensile degradation keeping at the same time the continuity of the stress-strain law. Figure 1 shows the response of the model subject to crack closure following a tensile loading path:
Figure 1. Compression/tension response. Unilateral effect and stiffness recovery. The thickness of the hysteresis loops is proportional to the state of cracking 4. Numerical implementation 4.1. Finite element code: EFiCoS The choice of using a multilayered f.e. configuration combines the advantage of using beam type finite elements with the simplicity of uniaxial behavior. Each finite element is a beam which is discretized into several layers. The basic assumption is that plane sections remain plane (Bernoulli's kinematic) allowing one to consider a uniaxial behavior of each layer. The local constitutive equations are integrated for each layer of a cross section. Concerning reinforced concrete structures, reinforcement steel bars are introduced with special layers, the behavior of which is a combination of those of concrete and steel. A mixing homogenized law is considered.
For each element a secant matrix K is assembled from the relation between O and . The non-linear behaviors appear in the second member of the equilibrium equation and are subtracted from the vector of generalized forces:
in which U is the vector of nodal displacements, Fext is the vector of external forces and Fin is the vector of inelastic forces. Such a formulation, based on the initial secant stiffness matrix algorithm, is well suited to our modeling. After reaching the peak-load, it prevents the iteration matrix from becoming singular. Concerning dynamic analysis, the same framework is used. The seismic loading is applied by the mean of an accelerogram at the basis of the structure. A double integration of this signal allows one to determine the ground displacement. In such a way, the equations of motions are solved in the global coordinates and allows one to deal with computations needing different accelerograms exciting different parts of a same structure at the same time. For stability and precision reasons, a classical Newmark algorithm has been implemented to solve the equation of motion [NEW 59]. The expression of the discretized displacements and celerities are derived as:
Choosing 7 = 1/2 and )3 = 1/4 prevent us from any numerical dissipation and ensures an unconditional scheme stability. This expressions are introduced in the equations of motions at step t + A? imposing an implicit integration scheme.
4.2. Constitutive law implementation By defining two different surfaces, one can integrate the damage tensor under traction explicitly. For uniaxial compression, the algorithm becomes implicit due to the dependence of the compressive damage on the radial extension, in which case a kind of plane stress procedure has to be introduced. Concerning the sliding stress, a classical implicit analysis has to be carried out. Among the different methods available for this purpose (Euler's backward or mid-point rules algorithm solved by an iterative Newton method), we chose the classical form of the so-called "return mapping" algorithm [ORT 86]. Indeed, it ensures convergence in the most efficient way. More details concerning the numerical implementation can be found in [RAG 00].
Energy Dissipation Regarding Concrete Structures 5. Applications 5.1. Response to uniaxial stress loading
121
The analysis is performed at the material level. Figure 3 represents the simulation of the uniaxial compression test. One can observe that the model is able to describe the volumetric response of the material satisfactorily. The material parameters used for the analysis are: E = 36 000 MPa, v = 0.24, e^ = 1.10"*, Btdbttt = 9 000, Btnduced = 300, P = 12, adirea = 5.0 10" MPa1, bdiim = 1.0 10+1 MPa, ainduce= 2.0 10*" MPa1, binduced = 1.0 10+1 MPa, c = 0.18. The direct and induced qualification are related to the cracking mode: 'direct' means that frictional stress and damage are aligned with the direction of loading. The two curves in Figure 1 show the ability of the model to describe the hysteresis loops under traction and compression loading paths. The hysteretic dissipation capability of the model can be illustrated by plotting the absorbed energy of an unloading tensile loop against the value of tensile damage in Figure 3. One can easily appreciate the effect of the coupling between the state of damage and the sliding stress. Despite the lack of physical shape of the yield surface for the 2D compression state due to the use of a St Venant's-type criterion, the use of an induced and direct damage evolution law allows a good modeling of a 3D state of confinement. Figure 3 shows such an effect induced by a lateral pressure applied proportionally to the longitudinal one (radial path).
Figure 2. Dissipated energy versus damage-radial confinement effect
Figure 3. Compression test simulation, longitudinal, orthogonal and volumetric strain 5.2. Structural applications: seismic case study This section is dedicated to the contribution of this local material model to the global dissipation at the structural level. A dynamic example of a reinforced concrete mock-up subjected to an earthquake loading confirms the relevance of our local approach in modeling damping in the nonlinear range. The main purpose of the CAMUS experimental program [QUE 98J is to demonstrate the ability of reinforced concrete bearing walls to sustain seismic loading. To reach this goal, a one-third scale model was tested on the shaking table of CEA. This mock-up was composed of two parallel walls connected by 6 square slabs. A heavily reinforced footing allowed the mock-up to be anchored to the shaking table. The mock-up plans are shown below: The mock-up is loaded through horizontal accelerations parallel to the walls. The presence of steel bracing systems at each level disposed perpendicularly to the loading direction prevents any torsion modes occurrence. The accelerograms are modified in time with a ratio of 1/V3 to take into account the similarity rules. Two types of accelerogram are imposed: Nice SI for the far field type earthquake and San-Francisco (earthquake happened in 1957) for the near field one. The response spectra may show the difference of these two kinds of earthquakes: on one hand Nice is very rich in terms of frequencies and on the other hand, San Francisco has a thicker effective band width of high accelerations. The complete experimental sequence, as shown in Figure 5, is: Nice 0.25 g, San Francisco 1.13 g, Nice 0.4 g and Nice 0.7 Ig.
123
A simple modeling of the boundary conditions at the base was obtained using a horizontal bending beam instead of assuming a perfect anchorage introducing the anchorage and contacts defects. The knowledge of the stiffness measured from the shaking table helped us adjust the vertical stiffness as well as the rotational stiffness of the elements along the boundary. This kind of modeling allowed us to take into account 2 eigenmodes, which is the minimum for a structure such as CAMUS where the second vertical mode plays a major role.
Figure 4. CAMUS mock-up
Figure 5. Accelerograms: experimental complete sequence
No structural Rayleigh damping was introduced into the analysis. The global behavior of the mock-up was well-reproduced. The benefit of such a material modeling is made more obvious by focusing our comparisons on energy balance.
Figures 6 and 7 show comparisons between the new model with no viscous damping and a classical damage model [LAB 91 ] with Rayleigh damping in terms of energy equilibrium.
Figure 6. Computations with a classical damage model: dissipation through a global viscous matrix
Figure 7. Hysteretical energy dissipation: computations without any viscous Rayleigh damping
125
We can notice that for a "traditional" model of damage, the external Rayleigh damping takes part in height of 60 % to the total dissipation of energy. In the case of the second model, all energy is dissipated within the material. From a quantitative point of view, no adjustment is necessary at the level of the structural analysis and from a qualitative point of view, dissipation is relocated in the zones of strong degradations (in the three lowest level) and no more distributed on the whole of the structure.
6. Conclusions Analyzing a specific case study (the response of a reinforced concrete model subjected to seismic loading) allowed us to point out the major features of local nonlinear mechanisms which should be integrated in the analysis. Local constitutive equations based on damage mechanics for concrete have been developed and implemented in a finite element code dedicated to civil engineering large scale computations. The use of the simplified finite element method, allowing parametrical studies, helped us to understand and analyse the experimental responses by pointing out the influences of physical material features such as frictional sliding between crack faces This approach of damage mechanics coupling cracking and frictional sliding emphasizes the importance of a refined material modeling with regards to the accuracy and the predictive ability of structural computation tools. The effect of modelling the hysteresis loops was analyzed through its contribution to overall damping. The use of this kind of local model for dynamic and, in particular, seismic loading may, in the non-linear range, make the explicit expression of an often arbitrarily-defined damping matrix unnecessary. The role played by this global viscous damping matrix in the analysis is considerably lessened. More investigations in the identification of material parameters at the specimen scale subjected to complex loading paths would allow us to treat the cases of 3D structures bearing complex solicitations combining shear and torsion effects.
7. References [ARM 66] ARMSTRONG P.J. and FREDERICK CO., A Mathematical Representation of the Multiaxial Bauschinger Effect, G.E.G.B., Report RD/B/N, 731, 1966.
[BAT 82] BATHE K.J., Finite element procedures in engineering analysis, Prentice Hall (eds), Inc. Englewood Cl.iffs, New Jersey, 1982. [CHA 93] CHABOCHE J.L., "Cyclic Viscoplastic Constitutive Equations, Part I: A Thermodynamically Consistent formulation", J. Appl. Mech., Vol 60, p. 813-821, Dec. 1993. [COR 79] CORDEBOIS J.P. and SIDOROFFF., "Damage Induced Elastic Anisotropy", Colloque Euromech, 115, Villars de Lans, p. 761-774, 1979.
126 Numerical Modelling in Damage Mechanics [DRA 94] DRAGON A., CORMERY T., DESOYER T. and HALM D., "Localised Failure Analysis Using Damage Models", In Chambon R. et al. (eds), Localisation and Bifurcation Theory for Soils and Rocks, p. 127-140. Rotterdam:Balkema, 1994. [HALM 96] HALM D. and DRAGON A., "A model of anisotropic damage by mesocrack growth; unilateral effect", In. J. Damage Mechs., Vol. 5, Oct. 1996, p. 384-402, 1996. [JU 89] Ju J.W., "On energy-based coupled elastoplastic damage theories: constitutive modelling and computational aspects", Int. J. Solids and Structures, 25(7), p. 803-833, 1989. [KAC 58] KACHANOV L.M., "Time of the rupture process under creep conditions", Izv. Akad. Nauk. S.S.R., Otd. Tekh. Nauk., n 8, p. 26-31, 1958. [KRA 81] KRACJINOVIC D. et FONSEKA G.U., "The continuous damage theories of brittle materials, Part I and II", J. ofAppl. Mech., ASME, Vol. 48, p. 809-824, 1981. [LAB 91] LA BORDERIE CH., Phenomenes unilateraux dans un materiau endommageable.modelisation et application a 1'analyse de structures en b&on, Ph. D. thesis, Univ. Paris VI, 1991. [LAD 83] LADEVEZE P., "On an anisotropic damage theory", Proc. CNRS Int. Coll. 351, Villars-de-Lans, Ed. by J.P. Boehler, p. 355-363, 1983. [LEM 90] LEMAITRE J. and CHABOCHE J.L., Mechanics of solids material, Cambridge University Press, 1990. [MRO 67] MROZ Z., "On the description of anisotropic workhardening", J. Mech. Phys. Solids, 15, p. 163, 1967. [MUR 78] MURAKAMI S. and OHNO N., "A constitutive equation of creep damage in polycristalline metals", I.U.T.A.M. Colloquium Euromech 111, Marienbad, 1978. [NEW 59] NEWMARK N.M., "A method of computation for structural dynamics", A.S.C.E. Journal of Engineering Mechanics Division., Vol. 85, p. 67-94, 1959. [ORT 86] ORTIZ M. and SIMO J.C., "An analysis of a new class of integration algorithms for elastoplastic constitutive relations", Int. J. Numer. Meth., Eng. Vol. 23, p. 353-366, 1986. [QUE 98] QUEVAL J.C., COMBESCURE D., SOLLOGOUB P., COIN A. et MAZARS J., "CAMUS experimental program. In-plane tests of 1/3 scaled R/C bearing walls", Proc. Xlth ECEE98, CD-ROM eds Bisch P., Labbe P. et Pecker A., Paris/CNIT La defense, 1998. [RAG 00] RAGUENEAU F., LA BORDERIE Ch. and MAZARS J., "Damage model for concrete like materials coupling cracking and friction, contribution towards structural damping: first uniaxial application", Mechanics Cohesive Frictional Materials (to appear), 2000.
Chapter 8
Numerical Analysis of Failure in Sheet Metal Forming with Experimental Validation

Michel Brunet, Fabrice Morestin and Helene Walter
Laboratoire de Mecanique des Solides, Villeurbanne, France
Failure in Sheet Metal Forming
129
1. Introduction The increasing use of sheet-metals with high elastic-limits and with a limited formability such as aluminium or titanium alloys leads to new problems in the simulation of the sheet forming processes of these materials. In the experiments conducted by the authors to determine their Forming Limit Diagrams (FLD), it is currently observed that necking is immediately follows by failure and crack always appears. Moreover, the necking of the sheet is hardly visible and consequently, plastic-instability theories alone fail to predict the failure of these sheet-metals. There are several ways to achieve analysis of failure occurrence in sheet-metal forming. One way consists of carrying out a conventional F.E. simulation and by post-processing the F.E. results, using an experimental necking-failure curve to detect the zones where risks of cracks can occur. On the other hand, a large number of macroscopic fracture criteria for failure which occurs after necking have been evaluated by many authors consisting of products, integrals and sums of macroscopic stresses and strains. To determine the values of these criteria at the onset of failure, both experiments and F.E. simulations are needed. When applying these criteria, it was found that the main factor affecting the accuracy is the mode in which failure takes place, mainly under deep-drawing or under stretching conditions. The equivalent Mises-stress was judged best for the prediction of both deep-drawing and stretch-drawing cracks but the locus of maximum equivalent stress does not necessarily coincide with the locus of failure in the sheet. Moreover, the thickness distribution may also indicate the wrong locus of failure since this parameter is operation dependant and there is no material dependant critical sheet thickness reduction. Also there is a need in the simulation process to achieve better localization of the onset of failure. This can be expected by the coupled approach where the damage process is incorporated into the constitutive relations and necking criterion. Many investigations have shown that ductile fracture involves four successive damage processes which are the nucleation of voids from inclusions, void growth, void coalescence and cracking propagation. One constitutive equation to account for these processes is Gurson's model [GUR 77], which was derived in an attempt to model a porous isotropic plastic material containing randomly disposed voids. As suggested by Doege and co-workers [DOE 93], we have already extended the Gurson model to anisotropic matrix behaviour and implemented with our shell finite-elements suitable for simulating sheet-metal forming processes [BRU 96,97]. In these papers, the onset of necking may be found numerically by mathematical considerations due to the fact that the strain state gradually drifts to plane strain after the onset of load instability [BRU 97,98]. In this paper where a titanium alloy sheet is tested, a refined approach is presented due to the fact that failure occurs just after a very small necking. If the Gurson damage model demonstrates the softening effect of the material, the model itself does not constitute a fracture criterion. Therefore, a criterion of void
coalescence which determines a critical porosity has to be used to simulate the initiation of material failure. Tvergaard and Needleman [TVE 84] have introduced the so-called critical void volume fraction at which voids coalesce, which in the present study is first determined by fitting numerically the load-displacement or engineering strain curve of the tensile test. However, the critical void volume is not unique. It depends on the choice of void nucleation model and the corresponding parameters. Moreover, to the authors' knowledge there is no sound theory or method at present available in the literature for the choice of void nucleation model. As suggested by Zhang and Neimi [ZHA 94, 95], a second method to determine the critical porosity is tested by using the modified Thomason plastic limit-load model of internal necking [THO 85,90]. Fully compatible with the Gurson damage model, the main feature of the Thomason void coalescence model is that the material failure initiation is a natural process where the void coalescence is not needed to be fitted beforehand. The finite element analysis of necking-failure of our Nakazima tests on a titanium sheet-alloy for different strain-paths will show the potential advantage of this criterion. 2. Damage model The coupled approach where the damage process is incorporated into the constitutive relations and necking criteria is expected to achieve better localization of the onset of necking and failure. For example, it is frequently observed in actual production processes that steel and aluminium sheets exhibit different forming limit curves even if both have the similar n-hardening coefficients. 2.1. Extension of Gurson-Tvergaard damage model Failure in metal forming is mainly due to the development of ductile damage. Needleman and Triantafyllidis [NEE 78] found that the predictions of forming limit for voided sheets based on the Gurson damage model are qualitatively in accord with experimental results. By finite element analysis using a membrane theory, Chu [CHU 80a] examined the effects of void growth on forming limit under punch stretching, also Chu and Needleman [CHU 80b] examined the influence of void nucleation on the forming curves. A primary extension of the Gurson-Tvergaard model has been used in the context of plane-stress and orthotropic materials implemented in shell finite elements in order to simulate our Marciniack tests by Brunet, Sabourin and Mguil-Touchal [BRU 96]. The model of Gurson is based on the observation that the nucleation and growth of voids in a ductile metal may macroscopically be described by extending classical plasticity to cover effects of plastic dilatancy and pressure sensitivity of plastic flow. Tvergaard [TVE 81,82] has proposed a first modified form of the original Gurson's yield criterion by introducing three coefficients q^q^qj in order to better fit the corresponding three-dimensional finite element solutions:
131
In Eq. (1) for micro-voided material, f*(f) is the damage function of the microvoid volume fraction or porosity f and Tvergaard's constants q j = 1.5, q 2 = 1 and q3 =q^ as coefficients of the void volume fraction and pressure terms, instead of Qi
=
Q2 = <b = 1 in the original Gurson model. ay describes the hardening of the
fully dense matrix material by ay = h ( e p ) and p is the macroscopic hydrostatic stress, q is the effective Von-Mises stress of the macroscopic Cauchy stress tensor CJ which is expected to be replaced here by the quadratic orthotropic such as Hill [HIL48] effective stress or non quadratic as: Hill [HIL 79,90] or Barlat and Lian [BAR 89]. The 3/2 factor in the 'cosh' term stands for isotropic material; it must be slightly modified here in order to be consistent with the original paper of Gurson [GUR 77]. Liao, Pan and Tang [LIA 97] have established the modified yield criterion for porous sheet metals containing spherical voids based on Hill's quadratic yield criterion to describe the matrix normal anisotropy and planar isotropy. The closed-form yield criterion is a function of the anisotropy parameter r which represents the mean ratio of the transverse plastic strain rate to the through thickness plastic strain rate under in-plane uniaxial loading conditions. For all possible planestress conditions, the anisotropic yield function is expressed as:
As anisotropic yield criterion is approximate in nature, it is possible to maintain the Tvergaard coefficients in Eq. [2] but in the following it is assumed that qt = 1.45. In this case, it is worth noticing that the modified Gurson model only differs from the original one Eq. [1] by:
In sheet metal forming applications, we are generally concerned with plane stress conditions. Consider x, y to be the 'rolling' and 'cross' directions in the plane of the sheet, z is the thickness direction. Based on the Hill quadratic yield function, the yield function q is defined in the orthotropic axes x,y as:
where:
The parameters f, g, h and n are the dimensionless Hill material coefficients which are defined in terms of the Lankford coefficients r 0, r45, r 90 as:
The Lankford parameters are determined by three experiments in the various directions as pointed out by their different indices. If f = g = h = 1/2 and n = 3/2, the Von Mises isotropic yield function is recovered. The equivalent stress function q gives the current size of the yield surface but due to the anisotropy, the direct Eulerian constitutive law based on this criterion is not objective. In order to assure the objectivity, rotating frame formalism is applied. The axes of orthotropy of the Hill criterion can be updated by a rotation which can be chosen as the material spin rate co (co-rotational stress-rate) or from the polar decomposition F = RU (Green-Nagdi stress-rate). Since the elastic strain are assumed to be small and from practical sheet forming applications, the differences between these different rotations are very small. The flow rule is derived from the yield potential Eq. [1] or [2], the presence of the hydrostatic pressure in the yield function results in non-deviatoric plastic strains:
The hardening of the fully dense matrix material is described through a = h(e p ) The evolution of ep is assumed to be governed by the equivalent plastic work relation:
2.2. Damage evolution The damage model takes into account the three main phases of damage evolution: nucleation, growth and coalescence:
Failure in Sheet Metal Forming 133 The micro-void volume fraction increment due to nucleation may be expressed by the normal distribution model of Chu and Needleman [CHU 80b]:
In this strain controlled nucleation model, the normal distribution of the nucleation strain has a mean value N , a standard deviation SN and fN is the volume fraction of voids which could nucleate if sufficiently high strains are reached. With the normal distribution, the major part of voids nucleates between the effective plastic strain values: e p = e N - S N and e p = e N + S N - However, a continuous nucleation model with one constant can also be chosen in place or combined with Eq. [9].
Growth of existing voids is based on the apparent volume change and law of conservation of mass and is expressed as:
Finally, the modification of the yield condition to account for coalescence and final material failure is introduced trough the function f*(f) specified by Tvergaard and Needleman [TVE 84]:
With the accelerator ratio:
f * = i/q{ is the ultimate value of f* at ductile rupture, fc is a critical value of the void volume fraction when the coalescence of micro-voids occurs and the stress-carrying capability of the material sharply drops and finally, ff is the void volume fraction for which the stress-capability totally vanishes (final failure). The analysis of equations [8] to [13] shows that the material damage behaviour depends at least on the values assumed by four to six damage parameters, depending on the choice of the nucleation model. Consequently, the predictive capability of the damage mechanics model depends on the effectiveness of finding these parameters. A optimisation procedure is needed to match the experimental and numerical finite element results as regards the loads vs. displacement curve in a
tensile test. In this paper, such a first choice has been carried out by means of an inverse identification approach which will be described in paragraph 4. However, whether the critical porosity is a material constant, or whether the critical porosity is independent of the stress state, is questionable. Moreover, we have found that if the value of fc is dependent on the choice of the nucleation model, then the set of damage parameters must be considered as a whole in this case and not as a set of independent material parameters.
3. Computational aspects 3.1. Explicit solution procedure The four node quadrilateral shell element with five degrees of freedom per node and plane-stress state is adopted for the spatial discretization of the sheet. The through thickness shearing stresses are also taken account and in order to avoid the well known shear locking of this kind of element, the assumed strain field method of Dvorkin and Bathe is used [DVO 84]. A large number of analyses have shown that sheet forming processes can be analysed successfully by both the implicit static method and explicit dynamic procedure if the latter is run at a relatively low speed (<10 m/s). With the use of a lumped mass matrix, the advantages of the explicit dynamic algorithm is that the stiffness matrix does not need to be formed and the contact conditions are modelled accurately in a simple manner because of the requirements of small time steps. Moreover the material behaviour can be complex which is the case with internal damage variable leading to softening of the material.
3.2. Integration of constitutive equations It is known that one of the best algorithm for integrating constitutive equations is the backward Euler or implicit scheme. However, in the case of the plane-stress condition, the out of plane component of strain is not defined cinematically and must be added as an extra unknown in the local Newton iteration scheme. This fact and the presence of 'cosh' terms in the yield function and flow rule may lead to numerical difficulties when the damage variable increases rapidly. The authors have chosen a sub-stepping scheme on the modified Euler algorithm which incorporates error control. This approach is suitable with explicit dynamic analysis since it takes advantage of the small time step required by the overall stability limit. Then on each sub-step, the following set of incremental forms of equations are used to compute the plane-stress increments:
135
where iae} is the elastic stress increment vector and [D] the elastic (3 x 3) matrix satisfying the plane-stress assumption. From the normality of the flow rule of plastic strain increments, the plastic multiplier dA, is eliminated with the following set of equations:
Eq. [1] and Eq. [3] are used to yield:
Defining the gradient vector {a} so that:
where it is found that for plane-stress state:
The equivalent plastic work Eq. [7] gives the effective strain increment:
Use of h' the hardening modulus of the matrix in Eq. [18] and in Eq. [19] leads to:
The plastic out-of-plane strain increment can be now written as:
Notice that if the void volume fraction f = 0 then kj =0 incompressibility is recovered.
and the plastic
For the first order Euler algorithm the stress at the end of a sub-step is given by:
and it is the same for each internal state variable, the effective strain and porosity f where all quantifies have been evaluated at the stress state {a}k . A more accurate estimate of {cT/k+j and state variables may be obtained from the modified Eulerscheme which gives:
where {da}2 > and all quantities are evaluated at the stress state {c?}k+1 - The global error in the solution may be controlled by ensuring that the relative error for each sub-step is less than some specified tolerance:
The size of each sub-step is continually updated during the integration procedure to satisfy Eq.[24] where TOL is a small positive number in the range l.E-03 to l.E-05.
4. Damage parameters identification To determine the constitutive and damage material parameters of the proposed damage model, an identification technique must be used. First, the anisotropic coefficients are evaluated separately by our digital image correlation method (D.I.C.) for strain measurements. These Lankford coefficients r0, r 45 , r^ are determined from uniaxial tension tests in the three directions 0, 45 and 90 degrees to the rolling direction of the sheet. ra is defined as the ratio of width to thickness strain at a stabilized state of strain measured by the D.I.C. and taking account of the plastic incompressibility:
In this paper, a titanium alloy has been tested where the sheet thickness used was 1.2 mm, the D.I.C. method gives the following anisotropic Lankford's values:
The stress-strain curve can only be described to the value of the homogeneous limit strain and is expected to give ay = h(e ) as the hardening law of the pure matrix material. The tensile tests, carried out on a set of tensile specimens prepared according
137
to initial length 140 mm, initial width 20 mm, have allowed us to obtain the flow stress expressions in terms of Swift's law:
where: The material parameters (nucleation and coalescence) of the previous damage model are very difficult to quantify by direct experimental measurements. An inverse identification is needed by comparing some numerical and experimental results and searching for a suitable matching between them. This technique is based on the determination of the damage parameters minimising the cost function representative of the correlation between the load vs. displacement or engineering axial-strain during a tensile test and numerical finite element simulation. The previously described tensile tests were performed until the ductile rupture of the specimens and as an example, the load vs engineering axial-strain curve of the titanium alloy is displayed in Figure 1. The cost function expressed by the least square approximation is:
where p are the damage parameters, FjS1Itl and Fjexp are the simulated and experimental load responses and n is the number of points considered. Assuming that such response function in an assigned region of the input parameters has a regular behaviour, for instance it has a unique minimum and it is locally quadratic, it is possible to use known numerical techniques to search for such a minimum. Then six or four coefficients remain to be determined depending on the nucleation model chosen. In the present paper, the three-dimensional numerical analysis has been carried out for the simulation of the tensile tests with our specific explicit finite element code where the modified Gurson model Eq. [2] has been implemented. The explicit formulation permits a significant advantage in terms of CPU times for each call by the statistical analysis, but this approach requires a particular attention in order to avoid the occurrence of unacceptable inertial effects when the velocity is artificially increased. Then we have followed the procedure presented by Fratini, Lombardo and Micari [FRA 96], where the response function has been calculated for 26 different sets of input damage parameters all around a given starting point, the initial porosity being fixed. Two to three steps have been required to obtain the new starting point to develop the so called central composite design method with smaller incremental values for the damage parameters.
Figure 1. Experimental and numerical points of load vs. engineering axial strain
Figure 2. F.E. mesh and void volume fraction distribution at coalescence
139
However, in this fast inverse method it is necessary to select the starting point not far from the absolute minimum. In this way, the following sets of damage parameters have been obtained: Normal distribution model: f 0 =o.0001 f c = 0.044, SN =0.012 eN =0.076 fN = 0.098
Continuous model: fo =o.0001 A0 = 0.14, fc = 0.0274 In Figure 2 the void volume fraction distribution just at coalescence is reported on the mesh used, showing a clear shear band where one quarter of the specimen is analysed making use of the symmetry. Due to the assumed symmetries, Figure 2 represents two localised necks crossing each other at the centre of the strip but not observed experimentally since only one localised neck grows in reality [TVE 93].
5. Void coalescence criterion by plastic limit-load model In line with pure numerical convenience, a constant critical void volume fraction is almost always used in numerical analysis and practical applications using the Gurson model. When void nucleation is taken into account, the critical value depends on the choice of the nucleation models and parameters as it has been observed in the previous paragraph. Figure 1 shows that the two set of parameters give virtually identical prediction of the load-displacement curve but a different critical void volume fraction. It is interesting to find that the simple continuous nucleation model works as equally well as the more complicated normal distribution model in this example. Therefore, a criterion of void coalescence which determines a critical void volume fraction would be useful. As suggested by Zhang and Niemi [ZHA 95], a modified version of the coalescence model by Thomason [THO 85,90] is tested. Thomason has developed a 3D micro-mechanical model of the internal necking of the inter-void matrix called plastic limit-load model. What is interesting in the plastic limit-load criterion is that void coalescence is not only related to void volume fraction but also to void-matrix geometry and stress triaxiality. Assuming that the material containing voids consists of a rigid-plastic nonhardening and isotropic material and using the Rice-Tracey [RIC 69] void growth equations of spherical shape initial voids, the variation in the geometry of the intervoid matrix is calculated using assumed velocity fields. Then the upper bound theorem is applied to obtain an overestimate of the ratio between the mean stress and the uniaxial yield stress of the matrix. This ratio a n /CTy is called the plastic constraint factor by Thomason. If we note An the net area fraction of the inter-void matrix in the maximum principal stress direction such that:
This gives the virtual maximum principal stress to initiate the localised necking of the inter-void matrix material, which represents the strong dilational plastic behaviour. If we note al the macroscopic maximum principal stress calculated by any numerical method, the critical condition to initiate the internal necking in a unit cell with a current ellipsoidal void can be postulated as:
By approximating the ellipsoidal void by the equivalent square-prismatic void and assuming two velocity fields parallel and triangular in the inter-void matrix of the unit cell, from the upper-bound theorem Thomason obtained the following type of empirical relation:
where F and G are constants, N and M are exponents, R X , R Z are the radii of the ellipsoidal void and X denotes half the current length of the cell. For an isotropic non-hardening material with the following values, F = 0.1, G = 1.2, N = 2, M = 0.5, the empirical results have been found to represent a good approximation to the upper-bound constraint factor. The stress triaxiality used in the original model ranges from 0.5 to 3, which is greater than the range 0.33 to 1.0 currently observed in sheet-metal forming. Zhang and Niemi [ZHA 94] have found that the original Thomason criterion gave too large predictions at low stress triaxiality and proposed a modification which uses the mean void radius R in Eq. [30]. This modification greatly decreases the prediction at low stress triaxiality, while for the high stress triaxiality the predictions are almost the same. As mentioned by Thomason [THO 90], it is interesting to note that there is no theoretical basis for the validity of plane-stress models of ductile fracture. This is due to the fact that only very small void-growth strains would be needed to initiate localised plane-stress necking at a row of holes. In this paper where we are concerned with the location of a necking-failure forming limit, that is hardly preceded by necking, this location is assumed to depend on the critical void volume fraction given by the 3D modified coalescence model of Thomason. As originally suggested by Thomason [THO 85] and already tested by Zhang and Niemi [ZHA 94,95], but not in the context of anisotropic sheet-metals forming, the modified Gurson model Eq. [2] is used to characterise the macroscopic behaviour assuming that the void grows spherically and to calculate the void and matrix geometry changes using the current strain and void volume fraction. During the F.E. analysis, the maximum principal stress is calculated and the normal strains in the directions of the principal stress-axes X,Y,Z are evaluated in
141
order to calculate the current void and matrix dimensions where the current void volume fraction f is an output of the modified Gurson model. If the initial and current volumes of a unit void containing cell are 1 and V, it is readily shown that:
Denoting Z the direction of the maximum principal stress a, and the current half intervoid distance X in the direction perpendicular to Z is calculated by:
The net area fraction of the intervoid matrix is evaluated according to the cell model as:
Once the equality [30] is satisfied, the void coalescence starts to occur and the void volume fraction at this point is the critical value fc in the modified Gurson model and then in the combined necking-failure criterion as explained in the next section.
6. Necking-failure criterion for anisotropic sheet-metals and example The strain ratio: (3 = Ae 2 /Ae, has an evident influence on the internal damage of sheet metals. At the same level of deformation, it is generally noted that the damage increment is the greatest at plane strain such that Ae22 = 0 when the localised necking occurs. The formulation follows our previous works [BRU 97,98], the criterion is formulated in terms of the principal stresses and their orientation with respect to the orthotropic axes leading to an intrinsic formulation. At the onset of load instability (dF, < o ) , the plane stress assumption and plastic incompressibility give the maximum force criterion or diffuse necking:
For a given material, if damage starts just after load instability as it is generally observed, we assume the inequality in terms of the effective stresses:
If the major principal stress a, is calculated with internal damage coupled as previously explained, the inequality in the major principal stress-axis is:
When a sheet is strained under a biaxial tension stress state, development of damage will make the strain state gradually drift to plane strain. Similarly, when a sheet is deformed under tension-compression condition, in the centre of diffuse neck, the final strain state at the local necking can also approach the plane strain. These observations have been earlier mentioned by Hecker [HEC 72] and discussed by Graf and Hosford [GRA 93] in the context of a theoretical analysis. Since the state of strain evolves towards the plane strain state, due to the related stress state change, there is an additional hardening-softening effect. The major stress is a function of many variables and different possibilities may be considered in connection with instability. The induced stress rate may be expressed as:
where { is the normal strain on the major stress axis, (3 is the strain increment ratio, t is the effective strain rate and 0 is the temperature. If we consider here only the effects of the normal strain j and of the strain ratio P, according to Eq. [36], the localised necking condition coupled with damage is given by:
An analytical and intrinsic form of the left-hand side of Eq. [38] can be formulated with [HIL 48] quadratic yield surface and the Gurson damage model and can be found in [BRU 98]. It is worth noticing that the stress state can be evaluated by any others quadratic or non-quadratic flow rules for the coupled plastic-damage F.E. analysis of the sheet forming process. As an example, the comparison between the proposed necking-failure criterion and our experiments has been obtained from the titanium alloy. Experimental Nakazima tests (hemispherical punch) have been carried out on notched sheets with various radii and depths in order to obtain different strain ratios as it can be seen in Figure 3. By a direct implementation into our F.E.-code where the principal stresses and their orientation with respect to the orthotropic axes are calculated at each time step, the calculation was stopped once Eq. [38] was satisfied. The F.E. analysis of the experiments have been done for the complete formed part as it can be seen in Figure 4 where the meshes of the two blank-holders and the hemispherical punch are not represented.
143
Figure 3. Specimens tested by hemispherical punch for necking-failure analysis
Figure 4. Example of major strain distribution at the onset of necking-failure
Unlike the yield stress and other material constants, the critical void volume fraction is an indirect material parameter which depends on the mathematical form of the constitutive equations and is not a material constant. It depends strongly on stress triaxiality and strain state as it can be seen in Figure 5.
Figure 5. Evolution of the critical void volume fraction using the modified Thomason formula
Figure 6. Ductile-fracture forming limit for a titanium alloy
The predictions using the present criterion Eq. [38] with the modified Gurson damage model Eq. [2] for anisotropic sheet-metals and the modified Thomason coalescence model Eq. 30 are shown in Figure 6 where reasonable agreement is observed. However, the agreement between experiment and theory is poorest for
145
plane straining where in this case the theoretical curves are above the experimental points which vary rapidly in a narrow range across the pure plane strain state.
7. Conclusion A formability analysis of anisotropic sheet-metals presenting ductile-fracture forming limits has been carried out based on the anisotropically extended form of the Gurson damage model. The modified model takes into account the anisotropic properties of the matrix material and has been implemented in our explicit F.E.code. For shell elements and plane-stress state, a second order explicit integration scheme with error control has been found to be accurate and robust. The loaddisplacement curve of the tensile test has been used to identify the damage parameters using an inverse method. It is difficult to determine the critical void volume fraction which appears to be not a material constant. A more promising approach is to introduce more micromechanism in the damage analysis. To this end, a modified form of Thomason's coalescence model has been tested in conjunction with our necking-failure criterion. The results emphasise that the extension of Gurson's model combining the Thomason coalescence mechanism by internal necking is a good method for failure assessment in the design of metal forming processes of metal sheets with ductile-fracture forming limits.
8. References
[BAR 89] BARLAT F., LIAN J., "Plastic behaviour and stretch ability of sheet metals", Int. J. of Plasticity, Vol. 5, 1989, p. 51-67. [BRU 96] BRUNEI M, SABOURIN F., MGUIL-TOUCHAL S., "The prediction of necking and failure in 3D. sheet forming analysis using damage variable", Journal de Physique, Vol. 6, 1996, p. 473-482. [BRU 97] BRUNET M., MGUIL-TOUCHAL S., MORESTIN F., "Numerical and experimental analysis of necking in 3D. sheet forming processes using damage variable", Advanced Method in Materials Processing Defects, Studies in Applied Mechanics, 45, Elsevier Science B.V., 1997, p. 205-214. [BRU 98] BRUNET M., MGUIL-TOUCHAL S., MORESTIN F., "Analytical and experimental studies of necking in sheet metal forming processes", J. of Mat. Proc. Tech., Vol. 80-81, 1998, p. 40-46. [CHU 80a] CHU C.C., "An analysis of localised necking in punch stretching", Int. J. Solids Structure, Vol. 16, 1980, p. 913-921. [CHU 80b] CHU C.C., NEEDLEMAN A., "Void Nucleation Effects in Bi-axially Stretched Sheets", J, ofEng. Mat. Tech., Vol. 102, 1980, p. 249-256. [DOE 93] DOEGE E., EL-DSOKI T.t SEIBERT D., "Prediction of necking and wrinkles in sheet metal forming", NUMISHEET'93, 1-3 Sept. 1993, Tokyo, Japan, p. 187-197.
146 Numerical Modelling in Damage Mechanics [DVO 84] DVORKIN E.N., BATHE K.J., "A continuum mechanics based four-node shell element for general non-linear analysis", Eng. Comput., Vol. 1, 1984, p. 77-88. [FRA 96] FRATINI L, LOMBARDO A., MICARI F., "Material characterisation for the prediction of ductile fracture: an inverse approach", /. of Mat. Proc. Tech., Vol. 60, 1996, p. 311316. [GRA 93] GRAF A.F., HOSFORD W.F., "Calculation of Forming Limit Diagrams for Changing Strain Paths", Metallurgical Transactions, Vol. 24A, 1993, p. 2497-2501. [GUR 77] GURSON A.L., "Continuum Theory of Ductile Rupture by Void Nucleation and Growth", J. of Eng. Mat. Tech., Vol. 99, 1977, p. 2-15. [HEC 72] HECKER S.S., "Sheet Metal Forming and Formability", Proceedings of the 7th' Congress of IDDRG, Amsterdam, Oct. 1972, Paper 5C. [HIL 48] HILL R., "A theory of the yielding and plastic flow of anisotropic metals", Proc. Roy. Soc., London, 1948, p. 281-297. [HIL 79] HILL R., "Theoretical plasticity of textured aggregates", Math. Proc. Cambrige Phil Soc., 85, 1979, p. 179-186. [HIL 90] HILL R., "Constitutive modelling of orthotropic plasticity in sheet metals", J. of Mech. Phys. Solids, Vol. 38, 1990, p. 405-417. [LIA 97] LIAO K.L., PAN J., TANG S.C., "Approximate criteria for anisotropic porous ductile sheet metals", Mechanics of Materials, 26, 1997, p. 213-226. [NEE 78] NEEDLEMAN A., TRIANTAFYLLIDIS N., "Void growth and local necking in bi-axially stretched sheets", J. of Eng. Mater. Tech., Vol. 100, 1978, p. 164-172. [RIC 69] RICE J.R., TRACEY D.M., "On the ductile enlargement of voids in triaxial stress fields", J. of Mech. Phys. Solids, 17, 1969, p. 201-207. [THO 85] THOMASON P.F., "A three dimensional model for ductile fracture by the growth and coalescence of microvoids", Acta metall, Vol. 33, 1985, p. 1087-1095. THO 90] THOMASON P.F., Ductile fracture of metals, Oxford, Pergamon Press. [TVE 81] TVERGAARD V., "Influence of voids on shear band instabilities under plane strain conditions", Int. J. of Fracture, Vol. 17, 1981, p. 389-407. [TVE 82] TVERGAARD V., "On localisation in ductile materials containing spherical voids", Int. J. of Fracture, Vol. 18, 1982, p. 237-249. [TVE 84] TVERGAARD V., NEEDLEMAN A., "Analysis of the cup-cone fracture in a round tensile bar", Acta Metall, Vol. 32, 1984, p. 157-169. [TVE 93] TVERGAARD V., "Necking in tensile bars with rectangular cross-section", Comp. Meth. App. Mech. Eng., Vol. 103, 1993, p. 273-290. [ZHA 94] ZHANG Z.L., NIEMI E., "Studies on the ductility predictions by different local failure criteria", Eng. Fracture Mech., Vol. 48, 1994, p. 529-540. [ZHA 95] ZHANG Z.L., NIEMI E., "A new failure criterion for the Gurson-Tvergaard dilatational constitutive model", Int. J. of Fracture, Vol. 70, 1995, p. 321-334.
Chapter 9
Damage in Sheet Metal Forming: Prediction of Necking Phenomenon

Nathalie Boudeau, Arnaud Lejeune and Jean-Claude Gelin
Laboratoire de Mecanique Applicquee R. Chaleat, Universite de FrancheComte/CNRS, Besanôn, France
Damage in Sheet Metal Forming
149
1. Introduction Necking in sheet metal forming arises quite naturally due to loading conditions, boundary conditions or material inhomogeneities that develop when plastic straining increases. In the field, pioneer works have been carried out on the early stages of the development of mathematical plasticity [HIL 52]. Lots of experimental work has been carried out to characterize the forming ability of the sheet with the introduction of the Forming Limit Diagrams (FLD) [KEE 65] [GOO 68]. Unfortunately, these curves are not intrinsic and depend strongly on the strain path that can be very complex in industrial forming processes, especially in cases of multi-stage forming processes. Arrieux introduced then Stress Limit Curves (SLC) [ARR 89]. These curves seem not to be dependent on the strain path. Concurrently, theoretical works were developed and numerous models were proposed. These models can be divided into two classes: models based on homogeneous continuum and models based on heterogeneous continuum. The development of numerical methods to simulate the forming processes creates the necessity to predict necking phenomena from FE results. The commercial codes that Pamstamp and Optris propose plot the principal strain state calculated during the FE simulation on the experimental or calculated FLD without taking into account the fact that the necking state depends strongly upon the strain path. The early works based on the proposition of a criterion independent of the strain path and able to predict necking from FE results are due to two of the authors in [BOU 94], Since then, several approaches have been proposed ([BRU 95], [BRU 97], [HOR 96], [KNO 00], [FRO 98]). After this short introduction, the second section is devoted to the presentation of the linear perturbation and the numerous possibilities of this approach. Finally, section 3 will present the results obtained in two cases: FLD computations and necking prediction during FE simulations.
2. The perturbation technique: theory and applications 2.1. Theory Necking phenomenon is considered as an instability of the local equilibrium. To study this instability, the perturbations technique and its first order Taylor's development are employed [MOL 85]. Because necking occurs only in regions where plastic strains are quite important the local problem is strongly non-linear. Moreover, the hypothesis of a rigid-plastic material is applied. The local equilibrium is defined by: - the plastic yield locus, - the hardening law,
- the constitutive law, - the local equilibrium relationship, - the compatibility of the deformation, - the plastic incompressibility. The hardening law is Hollomon with strain rate dependence:
but other hardening laws can be used. 0 is defined as <3Y = (J0 (0, ) with (7y the initial yield stress. The local equilibrium conditions are:
where / corresponds to the wideness of the necking band (Figure 1). The strain rate compatibility conditions and the plastic incompressibility are:
The local equilibrium can take the following form:
where A is a non linear operator and U a vector. Vector U describes the strain and stress state for the mechanical equilibrium:
A perturbation vector 8 U is introduced to provide a way to detect plastic instability:
where 8U is the amplitude of the perturbation, is the spatial part of the perturbation and r] the temporal part of the perturbation, x and n are vectors that correspond respectively to the spatial location where necking could occur and to the normal vector associated with the necking band which could develop at the point in the material corresponding to the endpoint of the n vector as shown in Figure 1.
151
Thus if U is the solution of [5] the perturbed vector:
has also to satisfy equation [5], In the small perturbations case, the Taylor development to the first order gives:
A non trivial solution for 8U is needed meaning that:
where P,<p,e,uQ (7?) is a polynomial of 7]. Finally instability or necking occurs when:
Generally, condition [11] is too severe and a positive value for the threshold is used. So the necking criterion becomes:
The expression of equations [2] and [3] are different for 2D- or 3D-stress states. For 3D stress state under consideration, the stress tensor and the strain rate tensors are:
Equation [4] remains the same as in the 2D case. The vector position and normal vector to the localised band are given in Figure 1:
The local equilibrium is governed by equation [2] with:
where / corresponds to the thickness of the necking band as shown in Figure 1.
The case (p = 0 is ignored because it corresponds to the case where the band is parallel to the sheet plane. The perturbation of equation [2] taking into account equation [15] followed by a first order series expansion is given in Appendix 1. In case of 3D-stress state, there are six equations of compatibility. Only three equations are linearly independent. These three independent equations have been determined with the formal computation software Mathematica. The perturbation followed by Taylor's development of these equations is also given in Appendix 1.
Figure 1. Orientation of the necking band compared to the position vector in the reference frame. Model for necking in the case of3D-stress state
2.2. Applications The aim of this section is to demonstrate the high flexibility of the linearized perturbations technique. The application of the method to several plastic yield loci and to damaged material is presented. The numerical results will be presented in the following paragraphs. 2.2.1. Plastic yield locus There is a large variety of plastic yield loci that can be used to represent the sheet metal behaviour during stamping or deep drawing. They have been developed to give a better representation of the material flow in the case of large plastic deformation. The first anisotropic model is due to Hill [HIL 48] who proposed a quadratic model suitable for the behaviour of standard steel. It allows one to model isotropic, transverse isotropic and orthotropic material that is well suited to describe the behaviour in sheet metal forming. The use of aluminium alloys in sheet metal-forming to decrease the weight of automotive parts revealed a lack of the Hill 48 criterion to describe the so called anomaly of aluminium and several others criteria were developed [(BAS 77], [BAN 99], [BAR 91], [GOT 77], [HIL 79], [HIL 93], [HOS 79], KAR 93]).
Damage in Sheet Metal Forming 153 Banabic's model is a yield criterion for orthotropic sheet metals under plane stress conditions. It has been developed to take into account the particularity of the aluminium alloys which are BCC materials whereas is FCC one [BAN 00]. The yield stress is expressed as following:
where F and *F are functions of the stress tensor components defined below:
The coefficients M,N,P,Q and R are defined as following:
The above equations show that the shape of the yield locus is defined by eight material parameters: a, b, c, d, e, /, g and k . The k -parameter can take only two integer values: 3 or 4 depending on the crystallographic structure of the material. For steel and all other FCC materials, k = 3 . To the contrary, for BCC materials such as aluminium alloys, k = 4. Then only seven material parameters define the Banabic yield criterion. The constitutive law is given for the first component:
The perturbation and the linearization of equation [20] is given in Appendix 2. 2.2.2. Damaged material The damage model chosen in this study is an extension of the model proposed by Gelin and Predeleanu [GEL 92]. In that case the plastic flow is governed by the following equation:
154 Numerical Modelling in Damage Mechanics with:
3 The z - parameter can take only two values: z = in case of isotropic material and
z = 1 in case of orthotropic material described with the Hill 48 plastic criterion. The combination of this damage model with Banabic's yield locus has not been studied yet. P_ is a fourth order tensor and is a deviatoric operator. In case of isotropic material, P_: G_ = devq_ and in case of orthotropic material described with the Hill's fourth order tensor, P : cr = H_: <7 . 9d is the volume change associated with damage evolution and cr0 is the strain rate-dependent hardening law [1]. 9d is related to the void volume fraction / by:
The material is then compressible and the mass conservation condition is expressed as: trace The equations obtained by the perturbation of [21] followed by a first order Taylor development can be found in [BOU 00]. 2.2.3. Post-processing of FE results The necking criterion established with the linearized perturbation technique can be easily interfaced to FE programs for necking prediction from FE results. Figure 2 summarizes the method.
155
At each step of a FE calculation the program gives the stress and the strain state corresponding to the current mechanical equilibrium. The calculation of the r] -roots can be done very quickly and integration points where necking occurs can be determined very easily without complicated software developments. The necking zones can be revealed by post-processing the FE results.
Figure 2. Implementation of necking criterion based on perturbations technique in a FE program
3. Simulations and prediction of necking Two kinds of results will be presented in this section. In the first subsection local simulations are carried out to improve the different models used. The 3D- and the ID-stress state approaches are compared. The classical FLD or the (e *, p) diagram will be used for the illustrations where * is the effective strain at necking and p = 2 / j is the strain path. In the second subsection FE simulations of deep-drawing processes are performed and necking predictions are presented and compared to experimental observations.
3.1. Local simulations 3.1.1. 3D modelling of necking Figure 3 shows the necking strain calculated from 2D or 3D modelling. For the 3D modelling no pressure was imposed meaning that <T33 = 0 along the normal to the sheet metal. For the 2D case, an optimum threshold (different in the thinning domain
156 Numerical Modelling in Damage Mechanics than in the expansion one) has been used and for the 3D case an unique threshold has been used. That explains the difference in shape between the 2D and the 3D FLD in the expansion domain first and secondly the too high level for the 3D FLD in the expansion domain. The strain level is superior in the case of 3D modelling for negative strain path. This curve demonstrates that 2D and 3D-modelling are strongly different. The difference is due to the fact that necking is searched not only in the plane but also in the thickness of the sheet. The increases of the strain level at necking in the thinning domain are in agreement with [ITO 00]. The results in the expansion domain are somewhat different compared to results presented in [ITO 00].
Figure 3. Comparison between the 2D and the 3D modelling of necking
3.1.2. Influence of the yield surface The Banabic yield criterion allows for the distinguishing of materials with a BCC or FCC structure. Necking predictions have been made with the 2D-stress state modelling working with Hill 48 criterion or Banabic criterion. First the relationship between the strain path and the stress path has been built in Figure 4 for steel and aluminium (Table 1). Figure 4 shows that the (OC, p) relationship obtained for steel (FCC structure) is closer to the linear relationship obtained for Hill 48 than the one obtained for aluminium (BCC structure).
157
Ct = (J2 / Cf{ is the stress ratio. That proves that the use of Banabic's plastic yield surface will allow necking prediction for a larger range of materials.
Figure 4. Illustration of the effect of the plasticity model on the p = f(CC) curves in comparison with the Hill quadratic model (from [LES 00])
Figure 5. Comparison of necking predictions using the Hill quadratic plastic criterion and the Banabic one for aluminium (from [LES 00])
Secondly the diagram ("*, p) has been built for aluminium with the Hill 48 criterion and with the Banabic criterion (Figure 5). The strain level at necking for the extreme strain paths are very different and it is shown that the Hill criterion underestimates the formability. On the contrary, strain level at necking for strain path close to plane strain is diminished a little. This study shows clearly that the yield surface modelling is very important for necking evaluation.
A
Aluminium Steel 0.6512 0.2115
b
0.9521 0.9941
C
0.0987 0.8390
M
0.4881 0.5811
N
0.5659 0.5571
p
5.2209 0.5923
q
-5.2598 -0.5803
R
98.672 1.1548
k
4 3
Table 1. Material parameters for Banabic's yield locus
3.2. FE simulations and necking prediction The simulations presented in this section have been performed with Stampform, the FE code developed in our laboratory and dedicated to the simulation of deepdrawing processes. Several hardening laws and yield criteria are available. In the present cases the material is considered isotropic and a Swift hardening law is chosen. 3.2.1. Stamping of an industrial automotive part This part is obtained in two operations and necking is known to appear in the first stage. The part is a 200 x 250 mm2 sheet metal with 3 mm of thickness. The steel parameters are summarised in Table 2. Necking zones predicted with the perturbations technique are shown in Figure 6 and are in good agreement with experiments. Necking is detected for 28 mm punch displacement as ductile fracture is observed for 29 mm of displacement. The comparison of necking zones with critical thinning zones shows that necking criteria based only on geometrical aspects (critical thinning) are not efficient. For the present example, a necking criterion based on the perturbations technique is in good agreement with experiment.
159
Automotive part Young modulus Poisson coefficient Yield stress Hardening law
Cross tool
E = 206 800 MPa
E = 70000 MPa v=0.3

0Y = 259 MPa K= 419 MPa
v=0.3
0Y = 259 MPa
K= 562.3 MPa
n = 0.241
n = 0.237 . = 0.0001
/0 = 0 (no damage) Isotropic
e0 = 0.0256
Damage Anisotropy
/Q = 0 (no damage)
H = 0.6154 F = 0.3846 P = 1.6153
Table 2. Material parameters for deep-drawing simulations
Figure 6. Necking zones predicted in an industrial automotive part
3.2.2. Deep-drawing with a cross-tool The geometry of both cross-tool and sheet metal is visible in Figure 7. The aluminium parameters are given in Table 2.
Figure 7. Geometry of the sheet and of the cross-tool
Figure 8 shows the necking zones predicted for a 50 mm punch displacement are in good agreement with experiment.
Figure 8. Necking zones in the cross-tool deep-drawing
4. Conclusions This paper reveals several important tendencies concerning necking prediction. First the linear perturbations technique is a very flexible method that covers a very large domain of applications. The technique initially developed for ID-stress has been extended to 3D-stress states. The adaptation to more accurate plasticity criterion such as Banabic's yield locus can be done without problem. As the linear stability analysis analyses the current state, its implementation in a FE code does not need specific and complex developments. All these possibilities have been demonstrated in the paper.
161
Moreover, this approach can be integrated into quality functions to optimise the deepdrawing processes. An apparent inconvenience of the approach is the complexity of the analytical calculations as it is shown in the appendices. But nowadays this is no more a major problem with all the formal calculation software and the numerous interfaces existing between these software and the programming languages.
5. References
[ARR 89] ARRIEUX R., Boivin M., "Theoretical determination of the forming limit stress curve for isotropic sheet materials", Annals of the CIRP, 38/1, 1989, p. 261-264. [BAN 99] BANABIC D., BALAN T., POHLANDT K., "Analytical and experimental investigation on anisotropic yield criteria", 6th Int. Conf. ICTP'99, Nuremberg, 1999, p. 1411-1416. [BAN 00] BANABIC D., COMSA D.S., BALAN T., "A new yield criterion for orthotropic sheet metals under plane-stress conditions", Proc. of TPR 2000, International section, ClujNapoca, 11-12 May, 2000, p. 217-224. [BAR 91] BARLAT F., LEGE D.J., BREM J.C., "A six-component yield function for anisotropic materials", Int. J. of Plasticity, Vol. 7, 1991, p. 693-712. [BAS 77] BASSANI J.L., "Yield characterisation of metals with transversally isotropic plastic properties", Int. J. Mech. ScL, Vol. 19, 1977, p. 651-654. [BOU 94] BOUDEAU N., GEUN J.C., "Prediction of the localized necking in 3D sheet metal forming processes from FE simulations", J. of Mat. Process. Tech., V. 45,1994, p. 229-235. [BOU 00] BOUDEAU N., GELIN J.C., "Necking in sheet metal forming. Influence of macroscopic and microscopic properties of materials", Int. J. Mech. Sci., Vol. 42, 2000, p. 2209-2232. [BRU 95] BRUNEI M., ARRIEUX R., NGUYEN NHAT T., "Necking prediction using forming limit stress surfaces in 3D sheet metal forming simulation", Proc. ofNumiform'95, 1995. [BRU 97] BRUNEI M., MGUIL-TOUCHAL S;, MORESTIN F., "Numerical and experimental analysis of necking in 3D sheet metal forming processes using damage variable", in Advanced Methods in Materials Processing Defects, edited by M. Predeleanu and P. Gilormini, 1997, p. 205-215. [FRO 98] FROMENTDM S., Etablissement d'un critere de striction intrinseque des toles et validation numerique par simulations d'emboutissage, Ph.D. thesis, University of Metz, France, 1998. [GEL 92] GELIN J.C., PREDELEANU M., "Recent advances in damage mechanics: modelling and computational aspects", Fourth International Conference on Numerical Methods in Industrial Forming Processes, edited by Chenot J.L. et al., 1992, p. 89-98. [GOO 68] GOODWIN G.M., "Application of the strain analysis to steel metal forming in press shop", La Metallurgia Italiana, n 8, 1968, p. 767-772.
162 Numerical Modelling in Damage Mechanics [GOT 77] GOTOH M., "A theory of plastic anisotropy based on a yield function of fourth order", Int. J. Mech. Sci., Vol. 19, 1977, p. 505-520. [HIL 48] HILL R., "A theory of the yielding and plastic flow of anisotropic metals", Proc. Roy. Soc., London, A 193, 1948, p. 281-297. [HIL 52] HILL R., "On discontinuous plastic states, with special references to localized necking in thin sheets", J. Mech. Phys. Solids, Vol. 1, 1952, p. 19-30. [HIL 79] HILL R., "Theoretical plasticity of textured aggregates", Math. Proc. Cambridge Philosophical Soc., Vol. 85, 1979, p. 179-191. [HIL 93] HILL R., "A User-friendly theory of orthotropic plasticity in sheet metals", Int. J. Mech. Sci., Vol. 15, 1993, p. 19-25. [HOR 96] HORA P., TONG L., REISSNER J., "Prediction method for ductile sheet metals failure in FE simulation", Third Int. Conf. In Numerical Simulation of 3D Sheet Metal Forming Processes, edited by J.K. Lee et al., 1996, p. 252-256. [HOS 79] HOSFORD W.F., "On Yield loci of anisotropic cubic metals", Proc. 7th North American Metalworking Conf., SME, Dearborn, MI, 1979, p. 191-197. [ITO 00] ITO K., SATOH K., GOYA M., YOSHIDA T., "Prediction of limit strain in sheet metalforming processes by 3D analysis of localized necking", Int. J. Mech. Sci., Vol 42, n 11, 2000, p. 2233-2248. [KAR 93] KARAFILLIS A.P., BOYCE M.C., "A general anisotropic yield criterion using bounds and a tranformation weighting tensor", J. Mech. Phys. Solids, Vol. 41, 1993, p. 1859-1886. [KEE 65] KEELER S.P., "Determination of the forming limits in automotive stamping", Sheet metal industries, Vol 42, n 461, 1965, p. 683-691. [KNO 00] KNOCKAERT R., MASSONI E., CHASTEL Y., "Prediction of strain localization during sheet forming operations", Third ESAFORM Conference on Material Forming, Stuttgart, Deutschland, 2000, II, p. 7-10. [LES 00] LESTRIEZ P., Prediction de striction localisee par 1'analyse linearisee de stabilite: application au critere de plasticite de BANABIC, Rapport de DBA, Universite de Franche-Comte - ENSMM, France, Septembre 2000. [MOL 85] MOLINARI A., "Instabilite thermoviscoplastique en cisaillement simple", Journal de Mecanique Theorique etAppliquee, Vol 4, n 5, 1985, p. 659-684.
6. Appendix 6.1. Appendix 1 The perturbed equations for the local equilibrium in case of 3D-stress state are:
Damage in Sheet Metal Forming 163 With: The perturbation of the compatibility equations in case of SD-stress state gives:
6.2. Appendix 2 The perturbed equation for the Banabic yield criterion is the
Chapter 10
Anisotropic Damage Applied to Numerical Ductile Rupture

Patrick Croix, Franck Lauro and Jerome Oudin Laboratory for Automation, Mechanical Engineering, Information Sciences and Human-machine Systems, University of Valenciennes and Hainaut-Cabresis, France Jens Christlein AUDIAG, Germany
Numerical Ductile Rupture
167
1. Introduction Description of damage is required for numerical simulations of sheet metal forming processes, sheet metal stamping or vehicle crash tests in which internal deterioration plays a significant role. Damage of porous material can be defined as a collection of permanent microstructural changes based on the description of the growth, nucleation and coalescence of the micro voids. This microscopic approach is described by several damage models which are applied to static loading (eg: extrusion, forging, etc) [BEN 93, 95] and dynamic loading (eg: crash, stamping, etc) [LAU 97, 98]. The Gurson damage model [GUR 77], modified by Tvergaard and Needleman [TVE 81, 82, 84], is based on this damage process for isotropic materials. To introduce the anisotropy of the material, this damage model has been modified by introducing Hill's yield stress instead of the von Mises into Gurson-Tvergaard potential [DOE 95]. The micro void shape, which is frequently expected to be at the origin of the anisotropic ductility, is taken into account in order to accurately predict an anisotropic damage. This model, based on the improved version of the GologanuLeblond-Devaux model [GOL 93, 94, 97], extends the Gurson-Tvergaard model to take the void shape effect into account. In order to emphasise the role of anisotropic void growth on ductile rupture, the microvoid shape is taken into account in its growth evolution. The description of the porous material is based on three internal variables: the microvoid volume fraction, shape and orientation. The microvoid volume fraction is defined as the ratio of the microvoid volume to the material volume. The microvoid shape corresponds to the difference between the Napierian logarithms of the minor and major semi axes of the microvoids and the orientation of microvoids changes with the rotation of the material. The sensitivity of the damage evolution is analysed in the case of prolate and spherical cavities with different loading directions. The anisotropy of the material is also introduced in the Gologanu-Leblond-Devaux model by means of the Hill 48 norm. Consequently some of the equations are modified. The new damage model for anisotropic microvoided material has been integrated into the three dimensional explicit finite element code for non-linear dynamic analysis of structures, PAM-SOLID, in the case of convected coordinates shell elements. This paper will describe the constitutive damage model and its numerical implementation in the finite element code. The failure prediction is shown in the case of a non-axisymmetric double Vnotched tensile specimen. Different anisotropic ductile ruptures are obtained according to the initial shapes of the cavities and their evolution.
2. Constitutive damage models 2.1. Gurson's model The Gurson Tvergaard Needleman GTN model (1984) is a quadratic formulation of plastic potential which can also be used as a yield function as follows:
where:
in which aeq is the macroscopic equivalent stress, QM is the elasto-viscoplastic flow stress, am is the mean stress (am = akk / 3), qi and q2 are the two "material" parameters introduced by Tvergaard [TVE 81] in order to converge the model with full numerical analyses of periodic arrays of voids, f* is the Tvergaard and Needleman's coalescence function. The original GTN model used the von Mises norm: ai, = 1.5a'ija'ij where a'is the macroscopic stress deviator. In this study it is replaced by the Hill 48 norm in order to take the anisotropic plastic behaviour into account[DOE 95], and in plane stress condition take the following form:
with F, H and N the Hill anisotropic parameters.
2.2. Gologanu 's model This is based on the studies of a material unit cell formed by two confocal spheroids. The Gologanu, Leblond and Devaux's model (GLD) extends the GTN model to take the microvoid shape effect into account, and is therefore interesting as it becomes the GTN potential for spherical microvoids. The void shape is defined by the parameter S = ln (a t /bi) where ai and b1 are the major and minor axis of the ellipsoidal void. The GLD model is based on the analysis of an ellipsoidal cavity embedded in a medium Q which has the shape of a confocal ellipsoid of minor and major semi-axes a2 and b2, the axis of the void is always collinear to the unit vector ex (Figure 1.). This cavity follows the rolling direction and is defined in the local frame of each element. The plastic potential takes the following form:
Numerical Ductile Rupture
169
Figure 1. The microstructure of a prolate cavity where:
The parameters K, r|, C and X are given by:
, with ex, Cy, ez the axis of the frame of the void and where the eccentricities ei, e2 and the parameter cc2 are deduced from the void shape parameter S and the void volume fraction f by the following equations:
The notation | is used in the original GLD potential to express the calculation of the von Mises norm applied to the deviator stress cr'on
which TICTHX is added. In this paper the material anisotropy is introduced with the Hill 48 norm and then a transformation of the deviatoric expression a' + rj an X is required. Moreover, considering plane stress strate , and assuming the microvoid direction ex follows the rolling direction, the deviatoric stress tensor a' + TJ O~H ^ is expressed as:
Then the Hill 48 norm presented for Gurson's model is modified and expressed as follows:
For the particular case of spherical voids (S = 0) the parameter K and a^ are respectively equal to 3/2 and 1/3 and the GLD model is identical to the GTN model. This model is completed by the equation of the evolution of the internal shape parameter S given by:
with HT(T,C) a function, dependent on the triaxialityT = aidc/vcfeq) according to the sign of = akk a'u , given by:
Numerical Ductile Rupture 171
and (Xi and a? are obtained by:
The evolution of the micro structural damage is represented by the current void volume fraction f, defined by f = 1-VM/VA, where VA, VM are respectively the elementary apparent volume of the material and the corresponding volume of the matrix. f* is a function of the void volume fraction f
where fc is the critical microvoid volume fraction at coalescence onset, fF is the microvoid volume fraction when ductile fracture occurs. This specific function f* inside the microvoid material potential describes the rapid loss of the stress carrying capacity due to the coalescence of the neighbouring microvoids, when f reaches the limit 1/qj. The microvoids volume fraction evolution has two main phases: the nucleation of the new voids and the growth of existing voids. The microvoids volume fraction rate is expressed by:
The constitution from void nucleation is controlled by plastic strain [CHU 80], and takes the form:
where fN is the nucleated microvoid volume fraction, SN is the Gaussian standard deviation, EN is the mean effective plastic strain for nucleation and eM is the effective plastic strain.
172 Numerical Modelling in Damage Mechanics The growth of existing voids is given by:
the equivalence between the plastic work dissipated into the porous material and the ductile matrix is expressed as follows:
which leads to the following expression of the effective plastic strain rate
in which o is the Cauchy stress tensor,CTMis the elasto-viscoplastic flow stress and Dp is the macroscopic plastic strain rate tensor defined in the case of the associated plasticity by:
The plastic multiplier A is deduced from the consistency condition <|)evp = 0 and 6 =0 leading to solve: Tevp
The plastic multiplier is finally expressed by
with
and
where Ce is the isotropic material tensor and I is the second order identity tensor.
Numerical Ductile Rupture 173 2.3. Numerical implementation The previous constitutive damage model is implemented in the finite element code PAM-SOLIDTM This code uses an explicit process in which the solution is advanced along the time axis, along which the velocities are discretised at half time intervals, n"1/2t, n+1/2 t... and the displacements and accelerations are discretised at full time intervals, n "'t, nt, n+1 t... where n is the number of the time increment. Considering a given time n t, the program calculates from the known quantities, which are the nodal velocity"" 1/2 V the nodal displacement "X and the Cauchy stress n"'a; the updated or unknown quantities "a, n+1/2v, n+1X, "a the nodal acceleration and Fint, Fext, the internal and external nodal forces, respectively. The new development consists of the stress calculation described below (Figure 2.). Assuming the strain rate tensor n ~ 1/2 Eij computed at the previous time increment as elastic, the corresponding Cauchy stress tensor is updated by central finite difference
in which n Q^ depends on
, in the case of shell elements. then
The corresponding potentials are obtained by all the derivation comes from this equation.
3. Numerical examples 3.1. Unit cells tensile test In order to study the anisotropic damage evolution several unit cell tensile tests containing ellipsoidal void shapes are carried out. Two loading directions L and T are considered which are respectively along the major and minor axis of the voids. To take only the anisotropic effect of the void shape into account, the isotropic von Mises criterion is considered and for this the G, F, N Hill's parameter are respectively equal to 1, 1 and 3 in this study. Tests with the GTN model and the OLD model are performed on a unit cell model with an initial spherical void. Two tests are performed on a unit cell model with an initial prolate void according to the loading directions L and T to exhibit the anisotropic damage evolution.
Figure 2. The modified stress elastic prediction and plastic correction flowchart
Et (MPa) aM (MPa)
Table 1. Flow stress
65000 195
647 206
600 218
450 227
300.5 518
This analysis carried out with a standard elastoplastic material; the flow stress is described by successive tangent moduli (Table 1). The usual material parameters are, q1=1.5, q2=l. for the elasto-viscoplastic potential, fo=10-4 for the initial void volume fraction, fN=0.04, SN=0.1 and eN=0.2 for the nucleation and the coalescence is not considered with fc = 1. fp = 1. These parameters are obtained from the literature [TVE 82], [NEE 85, 87]. The damage evolution of these tensile tests based on the plastic strain is presented in Figure 3.
Figure 3. Damage evolution based on the plastic strain for the tensile tests
The GTN model could be considered as the reference but it is generally well known that it overestimates the damage in the structure. It is directly compared to the GLD model with initial spherical voids. The GTN model has spherical voids which stay spherical during the deformation. With the GLD model the initial spherical voids become ellipsoidal in the direction of the deformation and consequently the damage evolution is more difficult and increases slower than with the GTN model, as seen in Figure 3.
176 Numerical Modelling in Damage Mechanics Consequently, the damage value is more realistic. In the case of prolate voids, the damage value is greater for the loading direction T than for the loading direction L (Figure 3.). This is really due to the combination of the shape of the voids and the loading direction. For the direction L, the prolate voids tend to become more and more prolate with an increase of the parameter S which corresponds to the ratio between the major and minor axis of the ellipse. For the direction T, on the contrary, the prolate voids tend to become oblate voids with an inversion of the ratio between the major and minor axis of the ellipse. The parameter S then evolves to S positive to S negative. Consequently, the area of the voids for this case is greater than with the loading direction L and the evolution of the parameter S is quicker and leads to more damage and of course to a quicker rupture (Figures 3 and 4).
Figure 4. Shape evolution according longitudinal and transverse loading directions
3.2. Non axisymmetrical double V-notched specimen The aim of the numerical simulation is to confirm that the OLD model is more accurate in predicting crack propagation than the GTN model. Thus, the numerical description of crack propagation for a non-axisymmetric double V-notched tensile test specimen is performed using both the GTN and OLD models. This crack propagation is obtained by elimination of successive elements which occurs at the complete loss of the stress carrying capacity. To avoid numerical divergence the stress in the element is put to zero when it reaches the damage value f = 0.9 fF. Figure 6a. presents the initial mesh size.
The porous elasto-viscoplastic material is described as was the previous material for the unit cell tensile test. The values of the material parameters are the result of an industrial requirement and are obtained using an inverse method [LAU 99]. This method consists in the identification of the damage parameters by correlating an experimental and numerical macroscopic measurement strongly dependent on the parameters. The macroscopic measurement is given by means of a tensile test on a notched specimen. It corresponds to the variation of the inner radius of the specimen as a function of the elongation. In order to take the anisotropic aspect of the microvoid shape parameter S into account, three experimental measurements are considered. They correspond to three tensile tests specimen at 0, 45 and 90 with the rolling direction. An optimiser is used to find the damage parameters minimising the gap between numerical and experimental measurements in all the directions at the same time. Due to the symmetry of the specimen, one quarter of its finite element model is used and the result for only one direction (0) is presented in Figure 5. This identification process is applied for Gurson's parameters as well as Gologanu's parameters.
Figure 5. a) Finite element modelling, boundary conditions for one quarter of the notched tensile specimen, b) Comparison of experimental-numerical width evolution at the bottom of the notch according to the tensile specimen elongation, for a specimen sample along the rolling direction
The material has an anisotropic behaviour which is introduced by Hill's parameters G, F and N. The initial void shape used with the GLD model is prolate with S = ln(4) = 1.3863.
Figure 6. Tensile test of a non-axisytnmetrical double V-notched specimen, a) initial finite element modelling; b) and c) are the damage distribution at the end of the process using the GTN and GLD models respectively
Figure 7. Tensile test of a non-axisymmetrical double V-notched specimen, orientation and shape evolution S at the end of the process for an initial prolate void shape using the GLD model
Figures 6b. and 6c. show the results from the computations after rupture with the GTN model and the GLD model, respectively. The GTN model shows that cracks are initiated at the bottom of the notches and propagated in opposite directions without ever coming into contact. The result with the GLD model differs from that obtained with the GTN model. The rupture initiation is identical for both models but with the GLD model the cracks tend to meet each other and the final rupture occurs in the middle of the test specimen which is the experimental result. This difference is due to the shape of the voids and their orientations. Due to the shape of the specimen, the main plastic strain direction is different in the middle of the Vnotches. Consequently, the voids rotate to follow the main axis of deformation and this leads to a better estimation of the crack propagation. The orientation of voids is described by vectors and presented in Figure 7. The length of these vectors is proportional to the shape parameter value. 4. Conclusion The GTN and GLD models are used to predict the damage evolution occurring with plastic strain. Both models are based for the description of the porosity of the material by the microvoid volume fraction and the prediction of the porous material flow by a modified yield surface. The evolution of the microvoid volume fraction due to the growth of existing microvoids and the nucleation of new microvoids is taken into account. The ductile rupture is finally predicted. These models have been modified by introducing Hill's potential into the elasto-plastic potential to take the anisotropy of the material into account. Moreover, the GLD model defines different initial shape of voids (prolate, spherical, and oblate) which change of shape, and orientation during the deformation. First of all the two damage models are presented in this paper. The implementation of the GLD model in the explicit finite element code PAMSOLID is then explained in more detail. Two different elementary tests are presented. First of all, a unit cell computation highlights the effect of the shape of the voids on the damage evolution and moment of rupture. These tests illustrate that: - the damage evolution is better predicted with the GLD model than the GTN model which generally overestimates the damage; -in the case of prolate voids, the damage evolution differs if the loading direction corresponds to the direction of the major axis or the minor axis of the ellipse. Secondly, tensile tests of non-axisymmetrical V-notched specimens are performed with the GTN and GLD models. These tests show that the direction of the propagation of the crack is better predicted with the GLD model due to the change of orientation of the voids according to the direction of the main plastic strain during deformation. Finally, two different ruptures are obtained with the GTN and GLD models and the GLD model gives closer results to the experimental ones.
5. References
[BEN 93] BENNANI B., PICART P., OUDIN J., Some basic finite element analysis of microvoid nucleation, growth and coalescence , Engineering Computations, vol. 10, 1993, p. 409-421. [BEN 95] BENNANI B., OUDIN J., Backward can extrusion of steels effects of punch design on flow mode and void volume fraction , International Journal of Machine Tools and Manufacture, vol. 35, 1995, p. 903-911. [CHU 80] CHU C., NEEDLEMAN A., Void nucleation effects in biaxially stretched sheets , Journal of Engineering Materials and Technology, vol. 102, 1980, p. 249-256. [DOE 95] DOEGE E., EL-DSOKI T., SEIBERT D., Prediction in necking and wrinkling in sheet-metal forming, Journal of Materials Processing Technology, vol. 50, 1995, p. 197-206. [GOL 93] GOLOGANU M., LEBLOND J.B., DEVAUX J., Approximate models for ductile metals containing non-spherical voidsCase of Axisymmetric prolate ellipsoidal cavities , Journal of the mechanics and physics of solids, vol. 41,1993, p. 1723-1754. [GOL 94] GOLOGANU M., LEBLOND J.B., DEVAUX J., Numerical and theoretical Study of coalescence of cavities in periodically Voided solids, Computational Material Modelling, vol. 42, 1994, p. 223-244. [GOL 97] GOLOGANU M., LEBLOND J.B., PERRIN G., DEVAUX J., Recent's extensions of Gurson model for porous ductile metals , Continuum micromechanics, 1997, P. Suquet ed., New York. [GUR 77] GURSON A. L., Continuum theory of ductile rupture by void nucleation and growth: Part I Yield criteria and flow rules for porous ductile media , Engineering Material Technology, vol. 99,1977, p. 2-15. [LAU 97] LAURO F., BENNANI B., DRAZETIC P., OUDDM J., Ni X., Damage occurrence under dynamic loading for strain rate sensitive materials , Communications in Numerical Methods in Engineering, vol. 13, 1997, p. 113-126. [LAU 98] LAURO F., BENNANI B., OUDIN J., Ni X., Damage occurrence under dynamic loading for anisotropic strain rate sensitive materials , Shock and Vibration, vol. 5, 1998, p. 43-51. [LAU 99] LAURO F., BENNANI B., CROK P., OUDIN J., Identification of the damage parameters for anisotropic materials by inverse technique: application to an aluminium , Proceedings of the International Conference on Advances in Materials and Processing Technologies, AMPT'99, Dublin, 3-6 August 1999, Editors Profs. M.S.J. Hashmi, and Dr L. Looney. [NEE 85] NEEDLEMAN A., TVERGAARD V., Material strain rate sensitivity in round tensile bar , Proc. Considere Mem. Symp., Salengon J. Ed., Presse de I'ecole Nationale des Ponts et Chaussees, 1985, p. 251-262. [NEE 87] NEEDLEMAN A., A continuum model for void nucleation by inclusion debonding , Journal of Applied Mechanics, vol. 54,1987, p. 525-531.
Numerical Ductile Rupture 181 [TVE 81] TVERGAARD V., Influence of voids on shear band instabilities under plane strain conditions , International Journal of fracture, vol. 17, 1981, p. 389-407. [TVE 82] TVERGAARD V., On localization in ductile materials containing spherical voids , International Journal of fracture, vol. 18, 1982, p. 237-252. [TVE 84] TVERGAARD V., NEEDLEMAN A., Analysis of the cup-cone fracture in around tensile bar , Acta Metallurgica, vol. 32,1984, p. 157-169.
Chapter 11
Numerical Aspects of Finite Elastoplasticity with Isotropic Ductile Damage for Metal Forming
Khemais Saanouni, Abdelhakim Cherouat and Youssef Hammi
Universite de Technologie de Troyes, France
Aspects of Finite Elastoplasticity
185
1. Introduction Displacement-based finite element codes that are industrially utilized for the static or dynamic analysis of mechanical structures require accurate and efficient constitutive equation subroutines. This accuracy concerns both the description of the physical phenomena taken into account by those constitutive equations, as well as their numerical discretization with respect to time and space. This leads, in general, to highly non linear algebraic systems to be solved at both local and global levels. At the global level, the spatial discretization of the principle of virtual work (or power) leads to a non linear system for displacement (or velocity) fields in the form of partial differential equations (PDE). This algebraic system is usually linearized to be solved for each load increment by either an implicit iterative Newton-type strategy, or a dynamic explicit or implicit one. Through linearization, many terms arise which can be classified into two classes: the first contains the material non linearity related to the material behavior (stress, internal variables) or the friction behavior; and the second contains the geometrical non linearities related to the finite deformations and rotations as well the evolving contact conditions. Particularly, the derivative of the stress tensor with respect to the total deformation tensor is needed. Generally, this "incremental" stress differential differs from the "continuous" differential given directly by the constitutive equations of rate type. It has been shown (see [NAG 82], [SIM 85]) that the use of the incremental stress differential consistent with the time discretization scheme of the stress tensor leads to quadratic convergence. For the explicit strategy to solve the system of PDE, only the stress increment at each time step is needed. The calculation of the stress increment at each time step needs the local integration of the overall constitutive equations representing the coupled physical phenomena. There exist various explicit or implicit integration schemes for ordinary differential equations (ODE). Experience has shown that implicit time integration schemes have the advantage of stability and are suitable for application to those constitutive equations involving yield and loading-unloading conditions. These conditions are generally modeled by constructing a special procedure as the elastic predictor and plastic corrector scheme. When a metallic material is formed by large straining processes such as forging, stamping, hydroforming and deep drawing, it experiences large irreversible deformations, leading to the formation of high strain localization zones caused by the nucleation and growth of micro defects (voids) generally referred to as ductile isotropic damage. Accordingly, to increase the efficiency and the predictive capabilities of the virtual forming tools, an accurate theoretical and numerical modeling of the damage initiation and growth under finite transformations should be taken into account. This can be achieved by using the coupled approach in the sense that the damage evolution equation is directly incorporated and fully coupled with the elastoplastic constitutive equations. This kind of approach has been employed by
many authors using damage models based either on Gurson's theory ([GEL 85], [ARA 86], [ONA 88], [BON 91], [BRU 96], among many others), or on Continuum Damage Mechanics (CDM) in the Kachanov's sense ([MAT 87], [ZHU 92], [SAA 99], [HAM 00], [SAA 00], etc. These fully coupled approaches allow the prediction of not only the large transformation of the processed workpiece as large deformations, rotations, and evolving boundary conditions, but they can also indicate where and when the damaged zones can appear inside the formed part during the process ([SAA 99], [HAM 00], [SAA 00]). In the present work, fully coupled constitutive equations accounting for both combined isotropic and kinematic hardening as well as the ductile damage in the CDM framework are presented. The particular case of the fully isotropic and isothermal flow considering small elastic strains, large plastic strains, isotropic and kinematic hardening, isotropic damage and the evolving contact with friction is implemented into ABAQUS/STD. The associated numerical aspects concerning both the local integration of the coupled constitutive equations as well as the (global) equilibrium integration schemes are presented. For fully implicit resolution strategy, special care is given to the consistent stiffness matrix calculation. The integration of the coupled constitutive equations is realized thanks to the backward Euler scheme together with the asymptotic integration procedure pioneered by Freed and Walker [FRE 86], The efficiency of this integration procedure in the 3D isotropic case is enhanced by reducing the number of the constitutive equations from 14 to 2 as proposed by Simo and Taylor [SIM 85] and widely used since then (see [HAR 93], [DOG 93], [CHA 96], [HAM 00] among many others). The numerical implementation of the damage is made in such a manner that calculations can be executed with or without damage effect, i.e. coupled or uncoupled calculations.
2. Kinematical background The transformation gradient F between the initial (undeformed and undamaged) and the current (deformed and damaged) configuration is multiplicatively decomposed so that the following classical definitions are used:
where B is the total Eulerian left Cauchy-Green deformation tensor associated with the Cauchy stress tensor a; L is the spatial velocity gradient in the current configuration, D and W are respectively the pure strain rate and the material spin
Aspects of Finite Elastoplasticity 187
tensors. The superimposed dot () denotes the usual time derivative. To satisfy the objectivity requirement, the so-called Rotated Frame Formulation (RFF) is used. This leads to the expression of the constitutive equations in a rotated configuration obtained from the current one by an orthogonal rotation tensor Q defined by [DOG 89]:
Accordingly, for any symmetric second order tensor T, the objective rotational derivative with respect to the rotating frame is given by:
from which the classical Jaumann and Green-Nagdhi rotational derivatives can be easily obtained. On the other hand the objective rotated tensor TQ by the rotation Q is given by:
Its time and rotational derivatives are related by:
Consequently, the constitutive equations are formulated in the same way as under small strain hypothesis and their generalization to the large strain case is simply achieved by replacing all the tensorial variables by their rotated corresponding quantities by using Eq. [6]. The second main question posed by the finite transformation aspect is how the total strain rate can be decomposed into elastic (reversible) and plastic (irreversible) parts. For metallic materials, dealing with large plastic strain but small elastic strain, the total Eulerian strain rate tensor decomposition can be approximated by:
where e^ is the Jaumann derivative (rotational objective derivative) of the elastic strain tensor (for simplicity the subscript J will be removed), and Dp is the plastic strain rate tensor defined by the constitutive equations as will be shown in the next section.
3. Coupled constitutive equations for metal forming 3.1. State variables versus effective state variables Finite thermo-elastoplastic constitutive equations coupled with continuous damage is developed in the framework of the classical thermodynamics of irreversible processes with state variables. For the sake of simplicity, this will be presented hereafter using the classical small strain notations keeping in mind that the generalization to the finite strain hypothesis is made according to the RFF formulation presented above. This formulation uses a unified yield surface for both plasticity and damage as in [SAA 94]. A more general formulation using two different (but coupled) yield surfaces can be found in [HAM 00]. Limiting ourselves to the simple first displacement gradient theory, two couples of external state variables are used, namely: (1) the total strain associated with the Cauchy stress tensors (e,o) and (2) the absolute temperature associated with the specific entropy (T,s). Five couples of internal variables are taken into account: (1) the (small) elastic strain representing the inelastic flow associated with the Cauchy stress tensor (ee,a); (2) the normalized heat flux vector associated with the gradient of the absolute temperature ( q / T , g = grad(T)); (3) the isotropic hardening variables (r, R) representing the size of the yield surface in strain space (r) and stress space (R); (4) the tensorial (deviatoric) kinematic hardening variables (a, X) representing the displacement of the center of the yield surface in strain space (a) and stress space (X), (5) the isotropic damage variables (D, Y), in Chaboche's sense [CHA78]. Suppose that the current configuration contains some isotropic ductile damage distribution i.e. a given homogeneous distribution of micro-defects such as voids and/or micro-cracks ; the concept of the effective stress ([CHA 78], [LEM 85]) together with the hypothesis of total energy equivalence [SAA 94] are used to define the effective state variables by:
where, for simplicity, it has been assumed that the damage effect on the elastic behavior is the same than on the hardening variables (both isotropic and kinematic).
189
These effective state variables are used in the state and dissipation potentials to derive the complete set of fully coupled constitutive equations for metal forming processes (see [CHA 78], [LEM 85], [LEM 92], [SAA 94] among others).
3.2. State potential: state relations The Helmoltz free energy \|/(ee, cc,r,D,T) is taken as a state potential. It is supposed to be a convex function of all the deformation-like state variables defined above and additively decomposed into thermo-elastic/damage and plastic/damage contributions:
where p is the material density in the current undamaged configuration and the variable T in the last term *Fpd acts as a simple parameter. In this work, only isotropic phenomena are considered, and have the following state potential:
where K and n are the classical Lame's constants of elasticity, a is the coefficient of thermal expansion, C is the kinematic hardening modulus, Q is the scalar isotropic hardening modulus, T0 is the reference absolute temperature, Cv is the classical specific heat parameter and I being the second order unit tensor. By using the Clausius-Duhem Inequality (GDI) one can easily derive, after some algebraic manipulations, both the state relations (Eq. [15 to 19]) and the residual inequality (Eq. [20]) defining the volumetric dissipation: - state relations:
- residual inequality:
Note that in the state relations above, the main material properties are affected by the damage according to: - elasticity properties of damaged material: - kinematic hardening modulus of damaged material: - isotropic hardening modulus of damaged material: - thermal expansion of the damaged material:
3.3. Dissipation potentials: complementary relations The volumetric total dissipation given above (Eq. [20]) should be identically verified for any selected dissipative phenomenon. In this equation the force-like variables namely:CT,X, R,Y are given by the state relations (Eq. [15 to 19]), and the flux variables should be defined by using the generalized standard materials [HAL 75]. This is achieved by introducing both yield functions and dissipation potentials for each class of dissipative phenomena. As a first approximation, the total dissipation is additively decomposed into two terms, namely: mechanical dissipation 3>m (plasticity, hardening and damage) and thermal dissipation 0th, each of them being supposed independently positive or zero:
191
Each of these dissipations will be analyzed to derive the flux variables associated with each selected dissipative phenomenon. 3.3.1. Thermal dissipation: fully coupled heat equation Classically, the heat equation is derived from Fourier's dissipation potential which is a quadratic scalar function of the force g j. For a thermoelastoplastic medium with mixed hardening and damage (strong coupling) the final form of the generalized heat equation can be written as follows [SAA 94]:
where A(T) stands for the Laplacian of the temperature and the prime (X') indicates the derivative of X with respect to temperature. The weak form of the partial differential equation Eq. [26] can be discretized with respect to time (Finite Difference Method) and space (Finite Elements Method) and solved together with the discretized weak form of the equilibrium problem thanks to a sequential methods. 3.3.2. Mechanical dissipation: fully coupled constitutive equations In the present case of time independent flow, a yield function in the stress space f(o, 2L R; D, T) and a plastic potential (non associative theory [LEM 85]) F(a, X, R; D, T) are introduced to derive the constitutive equations for plasticity with damage effect:
where the temperature dependent material constants a and b are the non linearity parameters for kinematic (a) and isotropic (b) hardening respectively; while S, s and P characterize the ductile damage evolution and the parameter oy represents the initial size of the plastic yield surface. The notation o-X effective stress according to: defines the norm of the
where gd is the deviatoric part of the stress tensor. The generalized normality rule allows the derivation of the complementary relations for plasticity, with hardening including the damage effect:
The tensor n represents the outward normal to the yield surface in the stress space given by:
The accumulated plastic strain rate p can be calculated from Eq. [30] using the following norm:
which indicates that the isotropic hardening strain r is not equal to the accumulated plastic strain p unless the hardening is linear (b = 0) as clearly indicated by the Eq. [32]. The plastic multiplier A, is given by the classical consistency condition applied to the yield function f : f > 0 , A, > 0, Af = 0. This gives for the fully flow:
giving
193
where <(.)> stands for the positive part of (.) and HH > 0 is the tangent plastic modulus given by :
and HT represents the thermal effect given by:
If the variation of the Poisson's ration v with the temperature is neglected, Eq. [39.a] writes under the following simpler form:
Note that, in this unified formulation, a single yield function is taken for both plasticity and damage, leading to a single plastic multiplier. This restrictive choice is justified in the case of metal forming where the damage develops only at material points with large plastic deformation. However, for some other materials such as concrete or composite structures, damage can develop without plasticity and vice versa. In those cases the use of multisurface formulation should be preferred : one yield function for plasticity with damage effect (coupling) and another one for the damage yielding [HAM 00]. Finally, the direct time derivative of the stress tensor (Eq. [15]) gives with the help of the Eq. [37]
with:
and:
where use has been made of the following notations (i being the fourth order unit tensor):
is the classical fourth order symmetric operator of the isotropic elastic properties affected by the damage, and:
is the thermoelastic contribution in the tangent operator. It is clear from the equations [40] and [41] that the continuous tangent elastoplastic-damage operator is non symmetric for the coupled problem i.e. if Y is non zero.
4. Numerical implementation In metal forming, the large deformation and damage behavior experienced by metallic materials are described by nonlinear equilibrium, the coupled thermo-elastoplastic-damage constitutive equations and the contact conditions with frictional constitutive equations presented above. For the sake of simplicity, in this paper we limit ourselves to solving the equilibrium problem associated with elastoplastic-damaged solids without thermal effect nor the contact/friction conditions (see [HAM 00] for more details).
4.1. Finite element formulation The velocity (displacement) based finite element formulation starts with the principle of virtual power (work) which states that, among all the kinematically admissible velocity (displacement) fields u'(u'), the solution of the equilibrium problem minimizes the functional G (weak form) given here in continuous form limited to the quasi-static case using the classical updated Lagrangian formulation:
where V is the volume of the current configuration, FF is the boundary of the solid where external forces F (including the contact forces) are prescribed, f represents the
Aspects of Finite Elastoplasticity 195 vector of volumetric applied forces, D* is the virtual strain rate tensor and o is the stress tensor given by the coupled constitutive equations discussed above. By applying the minimization principle to the spatially discretized form of Eq. [45], one can obtain for the overall structure:
where 5R is called the equilibrium residual vector; Fint and Fext are the internal and external force vectors written here using the natural coordinates as:
where V0 and FFo are the volume and its boundary of the reference solid element, Jv is the Jacobian determinant of the isotropic transformation between global and natural coordinates for the solid element, Js is the Jacobian determinant for the surface element, N is the matrix of interpolation functions and B is the matrix of strain (rate) interpolation. Note that the matrices B, N and the Jacobians, Jv and Js are functions of the displacements (geometrical non linearities). The most widely used implicit iterative method to solve the system [46] is the Newton-Raphson method, which consists in linearizing Eq. [46], for the (n+1)* load increment and at the iteration (p+1), as follows :
where (u^jis the approximation of the solution at the iteration (p). The current tangent stiffness matrix Kpn+l is defined by:
The second term, (K^, / describes the dependence of the external loads on the geometry and will not be discussed here. The first, {K+1}"' , represents the variation of the internal forces with displacements. As shown by Eq. [47], this variation is due to the stress o (material non linearities) given by the fully coupled constitutive equations, and the fact that the matrix B as well as the Jacobien determinant Jv are displacement dependent (geometrical non linearities). For the sake of simplicity, only the term related to the material non linearities will be discussed hereafter.
4.2. Time integration procedure In order to calculate the internal forces \Fnp+1}"" and the tangent stiffness matrix JK^, / , we must first compute updated stresses at the end of the current load increment. This can be achieved by integrating the overall set of coupled constitutive equations discussed above. The implicit Euler integration scheme (Backward method) is used since it contains the property of absolute stability and the possibility of appending further equations to the existing system of nonlinear equations. Let us consider the system of ordinary differential equations given above (Eq. [30-35]) formally represented by y = f(y,t). The implicit method is defined by (for clarity the iteration subscript (p) is omitted):
with the abbreviations y n + 1 = y(t n + At) and yn = y(t n ). When applied to the stress tensor for example, Eq. [51] reads:
Using the elasticity relation (Eq. [15]) and the decomposition of the strain tensor we get:
where we have incorporated the fact that the plastic strain rate and damage rate only occur if the field condition is satisfied, i.e. during the time interval Atp < At. In the following, the subscript (n+1) will be omitted and the variables, which do not contain the subscript (n+1), are computed at tn + At. For the calculations of hardening variables ot and T; the AI 'Asymptotic Integration' procedure proposed by Freed and Walker [FRE 86], for a better integration of first-order ordinary equations is used. The AI procedure is mainly based on the fact that the above discussed constitutive equations have the following form:
where Y denotes here a set of state variables to be considered and A(Y) and ()>(Y) are given functions depending on the concerned constitutive equation. One can integrate Eq. [53] exactly over the time step and obtain the following recursive integral equation:
197
where is the parameter of time integration. Freed and Walker [FRE 92] have considered several discretization schemes of this exact solution. We retain here the asymptotic integration scheme at time t + At:
Applied to the kinematic and isotropic hardening evolution equations this gives:
where AX = XAt is related to the accumulated plastic strain increment according to Eq. [35]. By using the complete set of constitutive equations we end up with a system of 14 nonlinear scalar equations for 15 unknowns : six stresses, six back-stresses for kinematic hardening, one isotropic hardening stress, one isotropic damage variable and the plastic multiplier. The 14 first equations are:
where: The remaining (15th) equation is given by the yield condition Eq. [27], which must be satisfied at the end of each time step. Before solving iteratively (Newton's method) the above system of 15 equations, it is very helpful to reduce the size of this system by eliminating some equations among them. Following an idea originally proposed by Simo and Taylor [SIM 85] and widely used since that, we derive from Eq. [58] and [59] the deviatoric tensorial quantity cf'-X between t and t+At:
where the deviatoric tensor Z at tn is given by :
The multiplication of the yield function (Eq. [27]) by p^-X gives
This implies:
with the notation
Hence, the unknown tensor n is replaced by the tensor Z, which depends only on one scalar unknown, namely AA, as shown by the Eq. [67]. Furthermore, the system of 15 equations is now restricted to two scalar equations, namely:
where the expression of the damage release rate Y (scalar) is given by :
This small system (Eq. [68-69]) is solved iteratively thanks to the Newton-
199
Raphson numerical integration procedure to compute the two unknowns : AX and D (see [HAM 00] for details). Tables (1) and (2) summarize schematically the proposed stress calculation. NOTE - For plane stress hypothesis the total strain component e33 is not defined by the kinematics but by a new constraint namely : G 3 (AX,D,e 33 J=a 33 =0. This leads to an additional scalar equation with the new unknown e33 to be determined together with AX and D by solving the three equations Gi, G2 and G3 [HAM 00].
4.3. Consistent elastoplastic-damage tangent operator As discussed in paragraph 4.1 the quasi-static tangent stiffness matrix for large deformation is viewed as relating the rate of internal nodal forces to the nodal velocities. This gives rise to three main contributions: the stress contribution, the contact/friction contribution and geometry variation contribution. Only the first contribution is discussed here (see [HAM 00] for more details). The computation of this term needs the calculation of the tangent operator representing the stress variation with respect to the total strain for each load increment. The continuous form of this operator is given by Eq. [40] including thermal contribution. As reported by many authors, the equality of the global/local convergence of a Newton-Rahpson method is greatly improved when using a tangent stiffness matrix consistent with the discretized incrementation of the local constitutive equations ([NAG 82], [SIM 85]). This consistent operator is given here (thermal contribution being neglected) by differentiating with respect to the total strain the time discretized expression of the stress as follows:
This needs the calculation of the derivatives of D and AX with respect to the total strain e. These are obtained by solving equations [68-69] and the final expression of the consistent tangent operator is [HAM 00]:
The above discussed constitutive equations and the corresponding local time integration have been implemented in the general-purpose finite element code ABAQUS/STD thanks to the user's material subroutines UMAT for static implicit solving procedure.
(I)
Calculate elastic predictor : a1' =(l-D n )K(e:l)l + 2^(l-D n )(-^ If F<0 set 2 = 2""', X = X n , R = R n > D = D n and X = X n Else if F > 0 continue with (II) otherwise EXIT Calculate (AX,D) and hence n = Z/||ZJ , according to Table 2. Calculate stresses with plastic corrector : a" = (l - Dn )b* - 2^1(1 - D)AXn and a = (l - D)K( : 1)1 + o" Calculate the hardening stresses :a Eq. [56] and r Eq.[57]
(II) (III) (IV)
Table 1. Computation of the Cauchy and internal stresses
Otherwise start form the beginning
Table 2. Computation of AX and D by Newton-Raphson method ((p) stands for iterations)
5. Numerical examples 5.1. Accuracy contemplation Let us now investigate some effects of both hardening and damage on the numerical accuracy of the proposed stress algorithm. The convergence properties of the algorithm will be studied in a large domain, covering ranges of relative errors between 1 and 10'7. The selected loading conditions are a two-steps loading path under strain control:
201
- first step: uniaxial strain path with eu = 5% e,2 = 0 - the second step: multiaxial strain path with en = 5% e,2 = 5% This defines a non-proportional tensile-shear loading path making an angle of 90 between the actual normal to the yield surface and the stress increment. The relative error measure is defined by
where n is the number of increments and yref is the reference solution calculated with n = 100 000 load increments. These increments are constant for each path and equally distributed on the two loading steps. The Newton iterations are stopped for a maximum relative error of 10"10. Figures 1 and 2 summarize the obtained results. First the equivalent stress error is plotted versus the equivalent (cumulated) plastic strain error (Figure 1) where it is clear that the error is less than 0.001 for n= 1 000 increments. Figure 2 shows the damage error versus the equivalent plastic strain error which is still lower than 0.001 for the same number of iterations. From this figure, it is clear that the convergence rate is linear and better for accumulated plastic strain than for damage.
Figure 1. Equivalent stress versus cumulated plastic strain relative errors
202
Figure 2. Damage versus cumulated plastic strain relative errors
Figure 3. Modeling of the quarter of the plane strain notched bar I DOG 93]
203
Figure 4. Stress-strain response in the vicinity of the notch root during the first 4 loading cycles
Figure 5. Evolution of the back stress tensor components during the first 4 loading cycles
204 Numerical Modelling in Damage Mechanics 5.2. Notched bar under bending cyclic strain conditions The selected example is a notched bar subjected to a cyclic 4-points bending load, assuming plane strain state condition, already investigated by Doghri [DOG 93]. The calculation is achieved using the same mesh used by Doghri with the material parameters E = 210.0 GPa, v = 0.3, ay = 200.0 MPa, Q = 520.0 MPa, b = 0.26, C = 25 500.0 MPa, a = 81.0, and the boundary conditions shown in Figure 3. Number of iterations Abaqus Increm. Umat Number
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Nb
load step
1 1 1 1 1 1
1
Number of iterations Increm. Abaqus Nbload Number step

2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48
Umat
1
3 3 3 2 3 3 3 2
1
3 3 3 2 3 3 3 3
1 1 1
1 1
2 2 2 1 1 2 2 2 2 3 3 1 1 1 2 2 2 2 2 1 1 2 2 2 2 3
2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
1 1 1 1
1 1 2 2
1 1
1
1
2 2 2 2 3 5 1 1 1
1 2 2 1 1 2 2 2 2 3 5 1 1 1
1 1 2 2 2 1 1 2 2 2 2 3 3 1 1 1 2 2 2 2 2 1 1 2 2 2 2 3
Table 3. Comparison of the iteration number for both ABAQUS and Umat for the calculation shown in Figures 4 & 5 First we start with a comparison between our model (Umat) without damage and the standard nonlinear isotropic/kinematic hardening available in ABAQUS/STD. Figure 4 shows that the local material response (of element 166 located at the notch root) in terms of the first component of the Cauchy stress versus the first component of plastic strain obtained by our Umat compares well with the one
205
obtained by the ABAQUS/STD. The same remark applies to the variation of the three components of the back stress tensor (kinematic hardening) as shown in Figure 5. Careful examination of Table 3 shows that the consistent tangent operator proposed in this study gives the same numerical performance as the ABAQUS/STD one. Note that, in Table 3, the change between the first and the second load steps corresponds to the rotation of the outward normal to the actual yield surface. At that point the proposed Umat needs one additional iteration compared to ABAQUS/SDT (see the highlighted cells in Table 3). These results show that the proposed stress computation procedure based on a consistent tangent operator possess a good numerical properties compared to the similar model available in ABAQUS/STD. 5.3. Fracture prediction during hydraulic deep drawing The last example concerns the hydraulic deep drawing of a spherical box. Starting from a circular thin sheet ( 3.0 mm Thickness and 245.0 mm radius) fixed along its boundary on a table containing a circular hole (77.0 mm diameter), an increasing hydrostatic pressure is applied on the top of the system table/sheet giving rise to a vertical displacement of the table (2 mm/s) aiming to maintain the sheet to a fixed hemispherical punch of 72.25 mm of radius (Figure 7a). At the initial configuration the circular sheet is tangential to the top of the hemispherical punch as shown by Figure 7a. The Aluminum alloy sheet is characterized by the following material parameters : E = 84.0 GPa, v = 0.3, oy = 120.0 MPa, Q = 600.0 MPa, b = 3.0 (kinematic hardening being neglected), S = 200.0 MPa, s = (3 = 1.0. Figure 6 shows the material response of the^ aluminium used in both coupled and uncoupled cases. The contact and friction between the sheet and the table is supposed to be of Coulomb type with friction coefficient of 0.3. Numerical simulation aims to predict where and when damaged zones can be initiated inside the sheet formed during the process.
Figure 6. Local response of the used material in both uncoupled and coupled cases
206
Figure 7. Damaged zones prediction during a hydraulic deep drawing process
Figure 7 shows the comparison between the predicted (Figures 7 b,c) and the experimentally observed (Figure 7d) damaged zones at the end of the process. Figure 7b gives the numerically predicted damaged zones with the uncoupled formulation (i.e. no coupling between the damage and the elastoplastic behavior); while Figure 7c gives the same numerical result obtained with the fully coupled formulation. From these figures one can note that only the coupled formulation gives a result close to the experimentally observed one concerning the fully damaged zones at the end of the process. As expected, the uncoupled formulation is unable to predict correctly the location of the fully damaged zones. This shows the capability of the proposed coupled approach to predict the damage initiation location (in space and time) during metal forming processes. Many other results are available in [HAM 00]. 6. Conclusion The main purpose of this paper is to derive a fully implicit stress algorithm and the associated consistent tangent operator for a finite elastoplastic constitutive equations accounting for nonlinear isotropic/kinematic hardening and ductile isotropic damage. A problem-optimized procedure, which reduces the fully nonlinear system to only two scalar equations (three equations for the plane stress hypothesis) has been proposed. It has been shown that using an asymptotic integration procedure, in conjunction with the backward Euler method, leads to very good accuracy. The results obtained in the prediction of damaged zones for a 3D hydroforming process
Aspects of Finite Elastoplasticity 207 have shown the capacity of this coupled approach to optimize metal forming processes with respect to damage initiation. It is worth noting that because the present formulation is local, the results of coupled calculations are mesh dependent. A generalization of the present model to a damage gradient formulation is under progress and will be published later. 7. References [ARA 86] ARAVAS N., The Analysis of Void Growth that Leads to Central Burst During Extrusion , 7. Mech. Phys. Solids, 34, p. 55-79, 1986. [BON 91] BONTCHEVA N. and IANKOV R., Numerical Investigation of the Damage Process in Metal Forming , Eng. Frac. Mech., 40, p. 387-393, 1991. [BRU 96] BRUNEI M., SABOURIN F. and MGUIL-TOUCHAL S., The prediction of Necking and Failure in 3D Sheet Forming Analysis Using Damage Variable, Journal de Physique III, 6, p. 473-482, 1996. [CHA 78] CHABOCHE, J.L., Description Thermodynamique et Phenomenologique de la viscoplasticite cyclique avec endommagement, These de doctoral, Univ. Paris VI, 1978. [CHA 96] CHABOCHE J.L. et GAILLETAUD G., "Integration methods for complex plastic constitutive equations", Comput. Methods Appl. Mech. Eng., 133, p. 125-155, 1996. [DOG 89] DOGUI, A., Plasticite anisotrope en grandes deformations, These de doctoral essciences, Universil6 de Claude Bernard, Lyon 1, 1989. [DOG 93] DOGHRI I., "Fully implicit integration and consistent tangent modulus in elastoplasticity", Int. J. Numer. Methods Eng., 36, p. 3915-3932, 1993. [FRE 86] FREED A.D., WLKER K.P., "Exponential integration algorithm applied to viscoplasticity", NASA TM 104461,3rd Int. Conf. On Comput Plasticity, Barcelona, 1992. [GEL 85] GELIN J.C., OUDIN J. and RAVALARD Y., An Imposed Finite Element Method for the Analysis of Damage and Ductile Fracture in Cold Metal Forming Processes , Annals of the CIRP, 34(1), p. 209-213, 1985. [HAL 75] HALPHEN B., NGUYEN Q. S., "Sur les mat6riaux standards g6neralise"s", Journal de Mecanique, 14 (39), 1975. [HAM 00] HAMMI Y., Simulation nume'rique de 1'endommagement dans les proce'de's de mise en forme, These de doctoral, Universite" de Technologic de Troyes, Avril 2000. [HAR 93] HARTMANN S., HAUPT P., "Slress compulation and consistenl langenl operator using nonlinear kinematic hardening models", Int. J. Numer. Methods. Eng., 36, p. 38013814, 1993. [LEM 85] LEMAITRE J. and CHABOCHE J.L., Mecanique des Milieux Solides, Dunod, Paris, French edition 1985, Cambridge Univ. Press, English edition, 1990. [LEM 92] LEMAITRE J., A course on Damage Mechanics, Springer Verlag, 1992.
208 Numerical Modelling in Damage Mechanics [MAT 87] MATHUR K. and DAWSON P., Damage Evolution Modeling in Bulk Forming Processes, Computational Methods for Predicting Material Processing Defects, Edt; Predeleanu, Elsevier, 1987. [NAG 82] NAGTEGAAL J. C., "On the implementation of inelastic constitutive equations with special reference to large deformation problems", Comput. Methods Appl. Mech. Eng., 33 (1982), p. 494-484. [ONA 88] ONATE E. and KLEIBER M., Plastic and Viscoplastic Flow of Void Containing Metal - Applications to Axisymmetric Sheet Forming Problem , Int. J. Num. Meth. In Engng. 25, p. 237-251, 1988. [SAA 94] SAANOUNI K., FORSTER C. and BEN HATTRA F., On the Anelastic Flow with Damage , Int. J. Dam. Mech., 3, p. 140-169, 1994. [SAA 99] SAANOUNI K. and FRANQUEVLLLE Y., Numerical Prediction of Damage During Metal Forming Processes , Numisheet 99, Besanc,on, September, France, p. 13-17, 1999. [SAA 00] SAANOUNI K., NESNAS K. and HAMMI Y., Damage modelling in metal forming processes , Int. J. of Damage Mechanics, Vol 9, n 3, p. 196-240, July 2000. [SAA 00] SAANOUNI K., HAMMI Y., Numerical simulation of damage in metal forming processes , in Continuous Damage and Fracture, Editor A. Benallal, Elsevier, p. 353363, 2000. [SIM 85] SIMO J.C., TAYLOR R., "Consistent tangent operators for rate independent elastoplasticity", Comput. Methods Appl. Mech. Eng., 48 (1985), p. 101-118. [ZHU 92] ZHU Y.Y., CESCOTTO S. and HABRAKEN A.M., A Fully Coupled Elastoplastic Damage Modeling and Fracture Criteria in Metal forming Processes , /. Met. Proc. Tech., 32, p. 197-204, 1992.
Chapter 12
3D Nonlocal Simulation of Ductile Crack Growth: A Numerical Realization

Herbert Baaser and Dietmar Gross
Institute of Mechanics, Darmstadt University of Technology, Germany
3D Nonlocal Simulation of Ductile Crack Growth
211
1. Introduction Many structural parts of technical applications require a detailed computational analysis of their load-carrying capacity either during the design phase or later on while operating in a larger system or in a machine. Today a computational evaluation and simulation of such structural parts is also able to consider the influence of damage and failure occurance by continuum damage models implemented in the framework of the FEM. A well-known disadvantage in the numerical treatment of solid mechanics problems, where softening material behavior occurs, is the so-called mesh-dependence of numerical results. In a considerable number of investigations different methods have been proposed to overcome mesh-dependence of finite element results. The common idea is the introduction of a characteristic or internal length (scale) into the constitutive model or its evaluation. We consider four different types of models. COSSERAT models consider in addition to the displacement of a material point also its rotation as an independent kinematic variable, see [EHL 98]. The internal length introduced by this was shown to determine the width of shear bands especially in soil materials, where the additional rotational degrees of freedom are activated during the deformation history. For such shear dominated problems this method seems to be a suitable regularization technique. Another advanced regularization method is the introduction of higher displacement gradients as additional degrees of freedom. A consistent formulation in terms of small strains is available, see e.g. [BOR 99]. A totally different approach to model the failure occurrence is the discrete representation of the actual failure or damage mode. Such models are able to represent failure occurrence by special finite elements either capable of showing displacement jumps internally within the element structure, see [OLI96], or between the element edges through a specific cohesive law defining the stress-strain behavior, see e.g. [HOH 96] or [BAA 97]. Another type of regularizing approaches is known as non-local, based on a spatial smoothing of certain quantities over the volume or structure of interest, see [BAZ 88]. The main goal of this type of model is an additional evaluation of a volume integral for the internal variables like plastic strain or damage, convoluted e.g. by a bell-like kernel function, see [PIJ 93]. Many authors used this smoothing technique as in [LEB 94], [TVE 97] or [BAA 98]. They computed the smoothing of the damage parameter as an additional process applied to the actual result of the local solution of the constitutive equations. In this paper we describe a new approach of evaluating the volume integral for the increment of the damage parameter used. The idea is to compute the smoothing during the iteration of the set of constitutive equations, which is advantageous due to the iterative character of the equation solver applied. In this study 20-noded brick 3D elements are used with quadratic shape functions along the element edges. [MAT 94] have found that this element type well represents localization in the 3D case. As constitutive model we use the ductile damage model of [ROU 89]. An advantage of this model is the description of material softening behavior due to damage by the influence of just three material parameters. Nevertheless, there are only a few articles treating the calibration of the ROUSSELIER parameters to experimental data, see [ROU 89] and [LI 94]. A second advantage is related to the numerical implementation of the constitutive law by means of an implicit integration scheme. The
type of constitutive equations leads to symmetric tangent material moduli, which is advantageous in computing and storing the matrix expressions.
2. Nonlinear solution & nonlocal formulation 2.1. Classical iterative solution procedure In order to approximate the real nonlinear behavior of a structure, the final load, leading to an unknown displacement result u, is usually divided into smaller load steps. Thus, for each load level the structural response is computed by finding the actual equilibrium, using an iterative solution procedure. The accumulation of the incremental solutions Au of the displacement field results in the total answer u.
Weak Form of Equilibrium
Consistent Linearization
FIG. 1. Classical iterative FEM solution procedure for a given load level Starting from the weak form of equilibrium <?(u, Su) = 0, where Su indicates the vector of test functions which is identified with the virtual displacements, an iterative solution method can be constructed to determine for the unknown displacements u at
213
a given load, with prescribed boundary displacements u. The displacements describe the difference between the current and the initial configuration by u = x X. The representation of g (u, 8u) = 0 in terms of a TAYLOR series from a known position u with Ail = u - u leads to a first order approximation of the weak form. Thus expression is the basis for the global iteration loop to find the increments Au and thus the new displacements u. This procedure is schematically described in Fig. 1, where # describes the body considered in the current configuration with the volume v and with tboun as traction vector applied on the boundary dBa. In the next load step the new solution of the system has to be determined by repeating this iteration.
2.2. Nonlocal formulation In the usual iterative procedures all quantities are stored locally on the level of integration points. However, the nonlocal approach which is proposed here acts on the level of internal quantities on the integrations points. [BAZ 88] have shown that a nonlocal treatment of the damage quantities leads to mesh-insensitivity. For this purpose we apply the nonlocal smoothing integral to the increment A/? of ROUSSELIERS damage quantity
where the kernel function, chosen as
for an arbitrary position vector p is responsible for the nonlocal smoothing of the local quantity A/?'0"' with respect to the reference state X. Formally the characteristic length lc describes the standard deviation of the normal distribution. For a discussion of the factor k depending on the spatial dimension of the problem considered see [BAZ 88]. Following their arguments k = (6^/n)1/3 ~ 2.2 for the 3D case considered here. For the numerical implementation the volume integral [1] is applied for each integration point considering just the surrounding by a sphere of radius 3lc, which is sufficient for a tolerable integration error in this application. A possible crossing of symmetry lines or planes during the integration loop is respected by accounting for the virtual contributions of the volume integration. But the whole computation of the respective influence of the neighbourhood of each integration point can be calculated once as a preprocessing procedure and takes just a few minutes for the problems considered. The main issue of the new solution algorithm is the assembling of the global stiffness matrix and the global residual vector. In Fig. 1 the global equilibrium iteration is illustrated, with the use of the residual p(u, <5u) and the actual system stiffness T>g(u, 6u). In the FE representation the quantity #(11,6u) results in the residual vector
and the system stiffness T>g(a, 6u) in the global stiffness matrix, respectively. During the assembling procedure for these terms the actual increments for the local description of the constitutive behavior are computed. The numerical procedure for this kind of solution is described in the Appendix (see eqn. [23]), where the calculation of the local solution of the constitutive equations is demonstrated. In contrast, the new idea consists of a global computation of the increments Ax in [23] by a NEWTON-like iteration scheme. The specific algorithmic treatment of this iteration is described shortly in Section [4.2] and in an earlier version in [BAA 00]. Due to the iterative character of the new solving scheme it is possible to modify the actual solution of the increment of the damage parameter /? by a nonlocal approach. This modification is not possible in the traditional iterative solution procedure, where the evaluation of the constitutive equations is treated in each integration point violating the yield condition, without considering its spatial position. On the contrary, with this new approach a "communication" between the integration points is enabled due to eqn. [1], where the local increment of the damage parameter is modified depending on its location. The described nonlocal modification of the increment A/3 of the damage parameter influences the solution convergence of the set of constitutive equations, but this intervention acts in a moderate way so that convergence is obtained. In contrast to the method described e.g. in [BAA 98], we here operate in small steps, which influences the solution of the constitutive equations by the nonlocal character, but guarantees the global quadratic convergence behavior.
3. Finite strain plasticity and damage model 3.1. Finite strain plasticity In elastic-plastic solids under sufficiently high load finite deformations occur, where the plastic part of the strains usually is large compared with the elastic part. The description of finite plastic deformations in conjunction with damage models is often carried out using the additive decomposition of the elastic and plastic strain rates, [TVE 89]. Here however, we use the framework of multiplicative elastoplasticity which is widely accepted in plasticity. Its kinematic key assumption is the multiplicative split of the deformation gradient
into an elastic and a plastic part, providing the basis of a geometrically exact theory and avoiding linearization of any measure of deformation. As a further advantage, fast and numerically stable iterative algorithms, proposed and described by [SIM 92], can be used. In the following, only a brief summary of the algorithm in the context of a FE-implementation is given. An essential aspect of [3] is the resulting additive structure of the current logarithmic principal strains within the return mapping scheme as
215
Here, e; = In \i (i = 1,2,3) and A? are the eigenvalues of an elastic trial state, described by the left CAUCHY-GREEN tensor b^. The elastic strains eel are defined by HOOKE'S law and the plastic strain corrector AepZ can be derived by the normality rule of plastic flow. The elastic left CAUCHY-GREEN tensor can be specified with the decomposition [3] as
which clearly shows the "connection" between the elastic and plastic deformation measure by the occurance of the plastic right CAUCHY-GREEN tensor Cp/ = F^ Fp/. By means of the relative deformation gradient (see [SIM 92])
which relates the current configuration x to the configuration belonging to the previous time step at n -i> an elastic m'a/-state is calculated for the current configuration at time tn
with frozen internal variables at state n -i- If the condition $ < 0 (see eqn. [10]) is fulfilled by the current stress state T, this state is possible as is the solution. If, on the other hand, $ < 0 is violated by the trial-state, the trial stresses must be projected back on the yield surface $ = 0 in an additional step. This "return mapping" procedure is used as the integration algorithm for the constitutive equations described in Section 3.2. It should be mentioned that the algorithmic treatment in terms of principal axes has some advantages concerning computational aspects like time and memory saving. Based on this, the integration procedure of the constitutive equations for large and for small deformations is very similar, [ARA 87].
3.2. The Rousselier damage model First, some notations and characters, which will be used in this description of the constitutive law and later on in the algorithmic setting, are specified. Following the ideas of [ARA 87] we decompose the stress and strain tensors in scalar values, which is of great advantage for the numerical implementation. Thus, we write the KlRCHHOFF stress tensor r as weighted CAUCHY stress tensor as follows
where J := detF = ^
= ^ and p = \Tij6 ij defines the hydrostatic pres-
sure, q = \l\tijiij the equivalent stress and tij = r^ 4- p6ij are the components of the stress deviator. In this notation an additional important quantity is the normalized
stress deviator nr = ^t. The second order identity tensor 1 is defined as the KRONECKER symbol by its components (5y in the Cartesian frame. In an analogous way the plastic strain rate can be written as
where AP and A? describe scalar rate quantities which are defined later on. The constitutive model used in this study is the damage model proposed by [ROU 89]. Here, the yield function taking ductile damage processes into account may be written as
where cr* represents the material hardening in terms of a power law, and the last part of [10] describes the damage (softening) behavior by the function B((3) and an exponential assumption. Furthermore, E is YOUNGS modulus, a$ the initial yield stress, N the material hardening exponent, and D and cr\ are damage material parameters. Note that the formulation of [10] in quantities p and q as functions of the KlRCHHOFF stress tensor r is comparable to the original formulation F = ^- + ... = 0 in [ROU 89], because we can write ^^ = ^^Q J = ^Ô = q for po = 1 in the case of 8 = - > 0, P P P * / see also comments from the authors on eqn. (17) in [ROU 89]. The function of B(j3) is the conjugate force to the damage parameter /?, defined by
r U
Here, the initial void volume fraction /o is the third damage depending material parameter used in this constitutive set of equations. The set of constitutive equations is complemented by the evolution equations for the plastic strain epelqv and the damage parameter /?. The macroscopic plastic strain rate ep is determined by the classical associated flow rule
Note that epl coincides with the plastic increment Aepi for the algorithmic setting in [4] written in principal axes. The last bracket on the right hand side of [12] shows a further advantage of this formulation following [ARA 87], since it is very easy to determine the derivatives of $ with respect to the scalar quantities q and p. One can see with [9] and [12] that
217
These two equations allow the algebraic elimination of A if
is fulfilled. Thus, the increment of the plastic strain can be expressed by the two scalar quantities AP and Ae g . Then the equivalent plastic strain e%lqv can be incremented directly by Aeq. The evolution equation for the damage parameter fi is given by
which is obviously dependent on the deviatoric part of the strain rate Aeg and the actual hydrostatic pressure p. With this the whole set of constitutive equations is completed. The evaluation of the material model for a given load level requires the solution of the three equations [10], [14] and [15] for the unknowns Aep, Aeq and A/?, respectively. Classically, this evaluation would be done pointwise at the local level of the integrations points by an implicit EULER backward integration rule, which is described in the Appendix. In Section 2.2 we discussed a new approach, which allows for nonlocal formulation; its algorithmic treatment is shown in Section 4.2. The exact linearization of the set of equations follows the description in [ARA 87]. At this point we just mention the starting point of the linearization
where C is the elastic modulus defined by the LAME constants. The variational expression for [16] is found as
Some extended algebraic manipulations on [17], as described in [ARA 87], lead to the expressions <5Aep and <5Ae9 and finally to
which is the material modulus for the implicit integration procedure at the end of the considered time interval [, t + Ai]. 4. Finite element formulation 4.1. 3D-element The starting point is the weak form of equilibrium g(u, <5u), see Fig. 1, formulated in the current configuration, where u is the displacement and ti are the prescribed
218 Numerical Modelling in Damage Mechanics tractions acting on the loaded boundary dBa of the body in the current configuration B. Linearization with respect to the current deformation state, and rearrangement leads (with dv = J dV) to the following representation of the element stiffness
where J = det F and BQ denotes the reference configuration. As for the global resi-
FlG. 2. 20-noded solid element duum vector resulting from g(u, 6u) the elementwise results from [19] are assembled to the global stiffness matrix K. For further explanations on the implementation of the consistent linearization of the algorithm used, see [SIM 92] and the modifications in [REE 97] concerning the determination of the eigenvalue decomposition. The discretization chosen in this paper is based on a 20-node-displacement element formulation with shape functions A^, (i = 1,2,..., 20), so that quadratic functions describe the element edges. Fig. 2 shows such an element in an arbitrary configuration. As in [MAT 94] we use a 2 x 2 x 2-integration scheme, which means an underintegration with respect to the quadratic shape functions Ni. It shall be pointed out that again no hourglassing modes were detected like in the case of an 8-nodedisplacement element formulation and a 1 x 1 x 1-integration scheme, see [BAA 98].
4.2. The algorithmic treatment of the new approach As described above, the iterative solution procedure used for solving nonlinear problems by the finite element method requires an evaluation of the constitutive equations on the level of integration points. This is known as the lowest level of iteration in contrast to the global load/time incrementation and the subsequent global iteration fulfilling the weak form of equilibrium. In our new approach the originally local evaluation of the set of constitutive equations on each integration point is shifted from the lowest level to a global solution while assembling the system stiffness matrix and the right hand side residual vector.
219
Instead of solving iteratively for each integration point on the element level, we assemble a large system of equations where the set of equations from each integration point enters blockwise along the main diagonal of the new system matrix, see also [BAA 00]. The global iteration scheme to find the current equilibrium requires in every iteration loop the assembling of the (global) system matrix and the right hand side vector. Parallel to this assembling procedure we solve the new global system of constitutive equations by an NEWTON-like iteration scheme. Due to the blockwise structure of the resulting system matrix it is possible to invert this matrix block-by-block and reassemble the resulting 3 x 3-submatrices for the use in the NEWTON-like iteration scheme. The character of a full JACOBIAN matrix for a real NEWTON scheme is lost because of the coupling of the quantities AP, Aeg and A/? through [1]. But this coupling is very weak in comparison to the dependence of the constitutive equations [20]-[22] in the local evaluation. So the rate of convergence is nearly quadratic as assumed for a full NEWTON scheme, see also time consumption in Tab. 2. Having determined the nonlocal results, the global stiffness and right hand side for the global equilibrium iteration can be assembled following the same scheme as for the classical solution technique.
5. Example and results 5.1. Model of a CT specimen As an example we examine a three dimensional model of a CT specimen discretized by 20-node solid elements as shown in Fig. 3. Due to symmetry just a quarter of
FIG. 3. Model of CT specimen the structure is modeled, where the edge lengths used are shown in the right hand part of
220
Tab. 1. The loading is applied by the displacement up of the marked nodes and the respective nodes located in the middle of each element edge length, see Fig. 2 and Fig. 3. For simplicity applying an uniform numerical postprocessing procedure and for getting easier comparable results we just take the results of the marked nodes into account and sum up their reaction force to get a global answer of the structure. This process can also be justified by the long distance between the loaded nodes and the zone of interest in front of the crack tip. So the resulting reaction forces on the loaded nodes are uniformly distributed and just scaled by a factor resulting from the quadratic shape functions along the loaded element edges. Here, we distinguish two different mesh types named by the typical element edge length e in front of the crack tip. The first mesh is characterized by e = 1.0 mm, the other one by e 0.5 mm. The set of material parameters used is shown in Tab. 1, where the first four parameters
E I v I cr0 I N II D I ai I /Q II fF III r I / I b I h ImJOOO | 0.2 | 460 | 7 I 3 | 300 | 0.01 || 0.19 ||| 6 | 5 | 3 | 5~

TAB. 1. Material parameters, stress dimension [MPa], and geometry in [mm] can be obtained by simple tensile tests, and D, a\ and /o are responsible for the damage representation of the constitutive model. Ductile crack growth is represented by reaching the threshold value of fp = 0.19 by the local value of the volume fraction / expressed in terms of the damage parameter /? by / = -~, see Eqn. [11], in any integration point. This final value fp as threshold for the evolution of / is not urgent in this formulation using the ROUSSELIER damage model and its damage parameter ft, but in order to get comparable results to the use of e.g. the GURSON damage model, see e.g. [BAA 98], it is very attractive. The maximum distance of the respective integration points from the initial crack tip defines the actual crack growth Aa. To show the effect of the nonlocal regularization technique introduced here, the characteristic length scale is fixed to lc = 1.0 mm for the two different mesh types. So lc corresponds to the coarse discretization with e = 1.0 mm.
5.2. Results The response of the nodal reaction force vs. the applied displacement up is plotted in Fig. 4. It can be seen from the curves (1) and (3) that the 3D model behaves in the usual way depending on the discretization for the traditional local approach with lc = 0.0 mm. The finer the used FE mesh is, the earlier the onset of strain localization is reached. Afterwards the negative slope of the respective curves gets steeper with increasing refinement of the discretization. In contrast to these classical mesh sensitive results, the responses of the nonlocal approaches show a different behavior for the two types of discretization. Although the typical element edge length e is varied from e = 0.5 mm to e = 1.0 mm, the curves (2) and (4) characterizes nearly the same behavior of the regularized calculations with
221
FlG. 4. Reaction force vs. end displacement for different discretization the identical internal length lc 1.0 mm, while the onset of global softening can be observed for all curves at a displacement of about up 0.38 mm on the loaded nodes. Fig. 5 represents the maximum values for the void volume fraction / in the first integration point in front of the initial crack tip in the center of the specimen vs. the number of load steps. These values are plotted for the cases (2), (3) and (4) to show the effect of the regularization technique. In addition, the threshold value of fp = 0.19 is marked by a dashed line indicating the transition of this value by the different calculations at different load steps. The crack advance Aa, which corresponds to this maximal value of / in the front of the crack tip, is marked by dots in the plot. This shows that the local analysis (3) responses with a larger crack advance of Aa ~ 1.7 mm at load step 90, while the nonlocal representations (2) and (4) for both discretization result in a crack growth of about Aa ~ 0.8 mm (the accurate position of the respective integration points varies due to the different element edge lengths in front of the crack tip) at this load step 90. So, these regularized results show nearly the same crack advance, although different discretizations are used. Tab. 2 shows DOF 850 7545 Time : Equiv. Inv. 0.08 sees. 8.5 sees. Time : gbl. NLoc 0.61 sees. 10.1 sees.
"e = 1.0 mm e = 0.5 mm
TAB. 2. Comparison of computational costs iterating the global equilibrium at load level 71 a comparison of time consumption for calculations with two different discretizations e = 1.0 mm and e = 0.5 mm and the resulting degrees of freedom. The third column shows the time for inverting the FE system matrix by an advanced GAUSS elimination during the iteration of the equilibrium state on an IBM RS6000/397 workstation. Wi-
FIG. 5. Maximum values of f = -^-
in the center of the specimen
thin the considered load level of step 71 about seven iteration steps are needed to reach a converged situation. This time consumption is independent from the local or nonlocal evaluation procedure. The last column represents the duration for solving the set of constitutive equations on the global level within the proposed nonlocal approach. The classicaly local solution needs in contrast to that a vanishing time to assemble the FE stiffness matrix, indeed without nonlocal smoothing. So one can see that the time consumption for solving the nonlocal system increases obviously with increasing number of integration points, but the amount of time is in tolerable limits respecting the advantage of a nonlocal evaluation. These first results for the application of a new nonlocal solution approach demonstrate the capability of this technique in smoothing the classically local results on the level of integration points to a nonlocal distribution of the damage parameter j8, which is responsible for the softening behaviour of a structure. In this sense it can be seen that the applied modification of the classical FE solution approach, shifting the level of iteration of the constitutive equations to a more global, spatially dependent one, affects the structural response of a loaded specimen. Keeping these results in mind, attractive future work could be a modification of the classical FE solution approach with respect to a useful possibility to enable a "communication" between the integration points depending on their spatial position.
6. Summary In this contribution we show a study of a 3D simulation of CT specimen using the ROUSSELIER damage model combined with a 3-dimensional finite element formulation based on 20-node-solid elements. The main attention is put to a new algorithmic
223
approach to the iteration of the constitutive equations applying a nonlocal regularization to avoid mesh sensitive results. The first results for this treatment are shown here and underline the algorithmic applicability of the new approach. The additional time consumption for the nonlocal evaluation is in tolerable limits, although some further investigation concerning a faster realization of the new approach have to follow. Studies especially about the calibration of the model to experimental data are in preparation.
7. Bibliography [ARA 87] ARAVAS N., On the numerical Integration of a Class of pressure-dependent Plasticity Models , International Journal for Numerical Methods in Engineering, vol. 24, 1987, p. 1395-1416. [BAA 97] BAASER H., HOHE J., GROSS D., Ductile crack growth analysis using the Gurson damage model , KOSINSKIW., DE BOER R., GROSS D., Eds., Problems of Environmental and Damage Mechanics, n ISBN 83-906354-1-0, Warszawa, Poland, 1997, p. 139-147. [BAA 00] BAASER H., TVERGAARD V., A New Algorithmic Approach treating Nonlocal Effects at Finite Rate-independent Deformation using the ROUSSELIER Damage Model , submitted to Computer Methods in Applied Mechanics and Engineering, , 2000, see also DCAMM Report 647, TU Denmark, Lyngby. [BAA 98] BAASER H., GROSS D., Damage and Strain Localisation during Crack Propagation in Thin-Walled Shells , BERTRAM A., SIDOROFF F., Eds., Mechanics of Materials with Intrinsic Length Scale, n ISBN 2-86883-388-8, Magdeburg, Germany, 1998A, EDP Sciences, p. 13-17, Journal de Physique IV, 8. [BAZ 88] BAZANT Z., PIJAUDIER-CABOT G., Nonlocal Continuum Damage, Localisation Instability and Convergence , Journal of Applied Mechanics, vol. 55, 1988, p. 287-293. [BOR 99] DE BORST R., PAMIN J., GEERS M., On coupled gradient-dependent plasticity and damage theories with a view to localization analysis , European Journal of Mechanics -A/Solids, vol. 18, 1999, p. 939-962. [EHL 98] EHLERS W., DIEBELS S., VOLK W., Deformation and Compatibility for Elastoplastic Micropolar Materials with Applications to Geomechanical Problems , BERTRAM A., FOREST S., SIDOROFF F., Eds., Mechanics of Materials with Intrinsic Length Scale, Magdeburg, Germany, 1998, p. 120-127. [HOH 96] HOHE J., BAASER H., GROSS D., Analysis of ductile crack growth by means of a cohesive damage model , International Journal of Fracture, vol. 81, 1996, p. 99-112. [LEB 94] LEBLOND J., PERRIN G., DEVAUX J., Bifurcation Effects in Ductile Metals with Nonlocal Damage , ASME Journal of Applied Mechanics, vol. 61, 1994, p. 236-242. [LI 94] Li Z., BiLBY B., HOWARD I., A study of the internal parameters of ductile damage theory , Fatigue & Fracture of Engineering Materials & Structures, vol. 17, n 9, 1994, p. 1075-1087. [MAT 94] MATHUR K., NEEDLEMAN A., TVERGAARD V., Ductile failure analyses on massively parallel computers , Computer Methods in Applied Mechanics and Engineering, vol. 119, 1994, p. 283-309.
224 Numerical Modelling in Damage Mechanics [OLI 96] OLIVER J., Modelling strong Discontinuities in Solid Mechanics via Strain Softening Constitutive Equations : Part I and II , International Journal for Numerical Methods in Engineering, vol. 39, 1996, p. 3575-3623. [PIJ 93] PIJAUDIER-CABOT G., BENALLAL A., Strain localization and bifurcation in a nonlocal continuum , International Journal of Solids and Structures, vol. 30, n 13,1993, p. 1761-1775. [REE 97] REESE S., WRIGGERS P., A material model for rubber-like polymers exhibiting plastic deformation : computational aspects and a comparison with experimental results , Computer Methods in Applied Mechanics and Engineering, vol. 148, 1997, p. 279-298. [ROU 89] ROUSSELIER G., DEVAUX J.-C., MOTTET G., DEVESA G., A Methodology for Ductile Fracture Analysis based on Damage Mechanics : An Illustration of a Local Approach of Fracture , Nonlinear Fracture Mechanics, vol. 2, 1989, p. 332-354. [SIM 92] SiMO J., Algorithms for Static and Dynamic Multiplicative Plasticity that preserve the classical Return Mapping Schemes of the infinitesimal Theory , Computer Methods in Applied Mechanics and Engineering, vol. 99, 1992, p. 61-112. [TVE 89] TVERGAARD V., Material Failure by Void Growth to Coalescence , Advances in Applied Mechanics, vol. 27, 1989, p. 83-151. [TVE 97] TVERGAARD V., NEEDLEMAN A., Nonlocal Effects on Localization in a VoidSheet , International Journal of Solids and Structures, vol. 34, n 18, 1997, p. 22212238. Appendix : Iteration of the set of constitutive equations The unknowns Aep, Aeq and A/3 are determined by a NEWTON-like iteration scheme. The elastic trial-stress state is determined as described in Section 3.1, where also the quantities ptr and qtr are obtained as described below [8]. The "return mapping" iteration described here is carried out for integration points violating the yield condition [10], either on the local level or as shown in Sections 2.2 and 4.2 with the new global algorithm.
For having a better condition of the iteration system the function r\ is weighted by YOUNGS modulus, so that a dimensionless expression results as for TI and ra. Evaluate for i, j 1,2,3
where Ax is either calculated directly on each integration point by Arti = J^1 r, or calculated using a global iteration scheme where the nonlocal smoothing [1] is applied after every iteration step as described in the previous sections. So one gets [xi, x%, ZS]T = [Aep, Ae q , A/?]T. The quantities ej^r and /3tr are used to store the accumulated history of the actual integration point in addition to the tensorial quantity needed in [7].
Chapter 13
On the Theory and Computation of Anisotropic Damage at Large Strains

Andreas Menzel and Paul Steinmann
Chair of Applied Mechanics, Department of Mechanical Engineering, University of Kaiserslautern, Germany
Anisotropic Damage at Large Strains 227 1. Introduction The main objective of this work is the development of a phenomenological, geometrically nonlinear formulation of anisotropic tensorial second order continuum damage. In fact, it is the evolution of microscopic internal structure of materials by nucleation and growth of distributed microcracks or microvoids which in turn leads to the deterioration of the mechanical properties of the material. Especially the shape, orientation and evolution of these microdefects show a significant dependence on the direction of stress and strain. Obviously, the nature of damage is anisotropic and thus a continuum damage theory should provide sufficient freedom to capture these anisotropic damage effects. The appropriate choice of the physical nature of mechanical variables describing the damage state of a material and their tensorial representation is since long under discussion, for an overview see e.g. Lemaitre [LEM 96] or Krajcinovic and Lemaitre [KRA 87] and the literature cited therein. Following the attempts of Betten [BET 82] and Murakami [MUR 88], the well-known concept of deformed and reference, or rather undeformed, macroscopic configurations of a material body within the geometrically nonlinear continuum theory is supplemented by the concept of fictitious undamaged microscopic configuration. Nevertheless, in strong contrast to the classical approaches mentioned above, the present damage theory, fully outlined in Steinmann and Carol [STE 98] and further exploited in Menzel and Steinmann [MEN 00], is based on the notion of a second order damage metric tensor and its effects on the stored strain energy. Thereby, as the fundamental assumption, the storage of strain energy due to either nominal or effective strains is measured by either the damage or the energy metric based on the hypothesis of strain energy equivalence between microscopic and macroscopic configurations, see e.g. Sidoroff [SID 81]. The framework of standard dissipative materials, as proposed by Halphen and Nguyen [HAL 75] is strictly applied. Another approach to formulate anisotropic damage based on the introduction of an internal second order damage tensor similar to structural tensors has been given in Menzel and Steinmann [MEN 99].
2. Anisotropic hyper-elasticity based on a fictitious configuration The reference and spatial configuration of the body of interest are denoted by BQ C E3 and Bt C E3. Let <p(X,t) : B0 x R+ -> Bt represent the non-linear map of material points X G BQ onto spatial points x = <p(X,t) G Bt. In terms of convected coordinates 9l(x) and Ql(X) the natural and dual base vectors are given by the derivatives 0; = d^x TBt, g{ = 6*0* T*Bt, G, = d&X TB0 and Gl = dx1 T*BQ. Now, the spatial and material metric tensors follow straightforward - g* = gijgi <g> g*, g* = g*Jg. <g> 9j, Gb = GijGi Gj, G" = Gli Gi <8> GJ - and in addition we introduce the mixed-variant identity-tensors 9^ 9i gl and G*5 = G, <8> G\ On this basis, the linear tangent map of the direct motion reads F^ = dx<P = dgi<f> <8> d\0l = g{ <8> Gl GL+ and, for notational
simplicity, the gradient of the inverse motion 3>(x,) = (p~l : Bt x 1+ -> BO reads f^ dx& dQ*3> <S> dxQl = Gi <g> gl GL+. Several kinematic tensors can be introduced, e.g. the Finger tensors b" = F^ G" [F^Y = Gij g{ <8> #., or the right Cauchy-Green tensor Cb = [F^f 0b F^ = gij Gi <8> Gj which enters the definition of the Green-Lagrange strain tensor E* = [C G ]/2. A graphical representation is given in Figure 1 and in view of a detailed outline on non-linear kinematics we refer to the work of Marsden and Hughes [MAR 94]. 2.1. The fictitious configuration Now, in addition to the physical and material space we introduce a fictitious isotropic configuration with natural tangent space TBo and corresponding dual space T*BQ. In analogy to the intermediate configuration within the multiplicative decomposition of elasto-plasticity, the fictitious configuration is generally incompatible. Mathematically speaking, we have a non-vanishing Riemann-Christoffel tensor which means that the conditions of compatibility are not fulfilled and the corresponding directfictitiouslinear tangent map - denoted by F - takes the interpretation as a non-holonomic Pfaffian, see e.g. Haupt [HAU 00]. Nevertheless, within the proposed multiplicative composition F allows the interpretation as pre-stretch and defines the fictitious base vectors Gi G TBo and G1 G T*Bo which are obviously not derivable from position vectors. Next, the fictitious metric and identity-tensors follow as
The linear tangent maps due to the direct and inverse fictitious motion read
see Figure 1 for a graphical representation. Without loss of generality usual pushforward and pull-back operations in terms of the fictitious linear tangent map hold, e.g. E* = [F13]4 E* F* = [9ij - Gtj\/2 & Gj . 2.2. Energy metric tensors In the sequel we incorporate a contra-variant energy metric tensor which reads within the fictitious and undeformed setting as follows
whereby the push-forward operation
is implied. As a key idea, the fictitious energy metric tensor is chosen equal to the fictitious contra-variant metric tensor and replaces this metric within the construction
Anisotropic Damage at Large Strains 229
Figure 1. Non-linear point map (f> and linear tangent maps F and F^ of the free Helmholtz energy function i/V Consequently, due to the principle of strain energy equivalence the relation
holds, which means that the free Helmholtz energy remains invariant under the action ofF*. NOTE 2.1 - Note that eq. (5) includes all assumptions of the proposed framework. In particular isotropy is included if the energy metric tensor A^ uis spherical whereas u otherwise anisotropy is considered. Since the relation A = G is incorporated, the fictitious configuration is isotropic and thus standard isotropic constitutive equation can be applied to model anisotropic material behaviour. NOTE 2.2 - For conceptual simplicity we focus here on the composition F^ F and do not consider the spatial fictitious tangent space TBt2.3. Hyper-elasticity Since the fictitious configuration is isotropic three (basic) invariants in terms of u E and A enter the formulation. Application of the standard pull-back operations Eb = [F^ Eb F* and A* = /"" A* [/"]* renders two corresponding sets of invariants
L
with n = 1,2,3. Thus the usual hyper-elastic framework yields e.g the second PiolaKirchhoff stress tensor
230

E E
and the Hessian isotropy. 2.4.
B d^>Ê\>^Q takes a similar format compared to standard
Representation of the energy metric tensor
In order to demonstrate the nature of the energy metric tensor we choose the following ansatz for the fictitious contra-variant vectors & with respect to the anisotropic reference configuration BQ (with AN^,AN^ 6 S"2)
With this representation in hand the two-field tensor F = Gi <S> G* reads
tl
Now, straightforward computations due to eq.(4) with A = G render the symmetric energy metric tensor
whereby the abbreviated notations /30 = o%, fi\ = 2 a0 a\ + o%, 02 = 2 aQ a2 + ot\, /3s = ai ct2 have been applied.
u
NOTE 2.3 - The rank one tensors A\ )2 allow similar interpretation to structural tensors. Indeed, the incorporation of eq.(10) into eq.(6) yields a set of invariants which can be expressed as a function of the set of invariants for general orthotropy as given e.g. by Spencer [SPE 84]. NOTE 2.4 - Within geometrically linear orthotropic hyper-elasticity based on structural tensors up to nine independent material parameters are included, see e.g. Spencer [SPE 84J. Contrary, the formulation based on the fictitious configuration incorporates four independent parameters (thereby the two Lame constants are taken into account, and thus the additional isotropic parameter ctQ is not independent.) This underlines that we deal with a reduced formulation. For the sake of clarity, we consider the constant tangent operator of a material of linear St.-Venant Kirchhofftype ^ k = d2EÊ^0 = A A8 <g> A* -I- /z[A t t A8 + A8 A*]. Now, referring to a Cartesian frame Ci, we choose without loss of generality A = @Q I + /?i e\ <8> e\ + 02 2 <8> 62- Then, the relevant coefficients of the Hessian read
Anisotropic Damage at Large Strains 231 Obviously, we deal with a sub-class of rhombic symmetry. 2.5. Numerical examples
For the following numerical examples a non-linear material of Kauderer-type is applied, see Kauderer [KAU 49]. Within the general representation
compare Ogden [OGD 97], the coefficients of the implemented constitutive equation read:
The subsequent material parameters are chosen: K = 8.333 x 104, G = 3.8461 x 104, /ôi _ Kdev = 0.5 anci Kvoi _ ^dev _ g.25. In view of the energy metric tensor the following spherical coordinates define the orthogonal unit-vectors AN\^,\ ti\ - 5/6 TT, tff = 1/6 TT, $\ = 1/3 TT and tf2, = 1/2 TT, see the Appendix, and the additional scalars to compute F read a0 = 1.0, ai = 0.25 and a2 = 0.5. 2.5.1. Simple shear For the homogeneous simple shear deformation (F = I + ê\ <g> e2 with I = dij i <8> GJ ) Figures 2 and 3 highlight the anisotropic behaviour of the applied material, compare Appendix. The anisotropy measure 6 shows a strong dependence on the shear number 7. Typically, the stereographic projection with respect to the stress and strain tensors underline their non-coaxiality. Finally, the plots of the determinant of the acoustic tensor show a different shape for the anisotropic and isotropic (ai = a? = 0) setting. They are given at 7 = 0.25 in normalised form with respect to the linear isotropic case with det(g) lin ' iso = G2 [3/4 G + K}. 2.5.2. Cook's problem A three-dimensional version of Cook's problem has been discretised with 16 x 16 x 4 enhanced eight node bricks (Q1E9) as advocated by Simo and Armero [SIM 92]. Figure 4 shows the reference geometry (L = 48, HI = 44, H2 = 16, T = 4), boundary conditions and a deformed mesh at \\F\\ = 1.28 x 105 which is the amount of the conservative resultant force of a continuous shear stress in BQ . Furthermore, Figure 5 highlights the displacement of the mid point node at the top corner for the anisotropic meterial and in addition for an isotropic setting with iso A" = \pl + [/3o + 0i]2 + [0o + 02]2]1/2 G". Obviously the anisotropic case results in a non-vanishing component ^3 which indicates the "out-off-plane" deformation.
232
Figure 2. Anisotropy measure 6 and stereo-graphic projection due to the principal directions of strain E : o and stress S : with respect to a Cartesian frame
Figure 3. Determinant of the acoustic tensor for 7 = 0.25 within the anisotropic and the isotropic setting For more detailed background information on non-linear finite elements we refer e.g. to Oden [ODE 72]. Note that the proposed framework results in an efficient numerical setting. Practically, we end up with similar costs compared to isotropic hyperelasticity, since the metric tensor G^ of the standard formulation is replaced by A. Contrary, the classical approach based on the incorporation of a structural tensor yields numerous additional terms in the computation of the stress tensor and especially of the tangent operator which ends up with tremendous numerical costs. 3. Anisotropic damage based on a fictitious configuration Here, as the key idea, the energy metric tensor is introduced as an internal variable and denoted as damage metric tensor in the sequel. Then the fictitious linear tangent map F is no longer constant and allows the interpretation as damage deformation gradient. The fictitious configuration remains isotropic and un-damaged but the standard reference configuration BQ as well as the spatial one Bt can be damaged and anisotropic.
Anisotropic Damage at Large Strains
233
Figure 4. Three-dimensional Cook's problem and deformed mesh at \\F\\ = 1.28 x 105
Figure 5. Load-displacement curve of the mid point node at the top corner for the anisotropic and isotropic setting
3.1. Standard dissipative materials We adopt an additive decomposition of the free Helmholtz energy
incorporating a scalar-valued hardening variable K. Based on the theory of standard dissipative materials, see Halphen and Nguyen [HAL 75], the local form of the Clausius-Duhem inequality for the isothermal case with respect to BQ"reads
Within the standard argumentation of rational thermodynamics the nominal stress, the damage stress and the hardening stress are obtained by
234 Numerical Modelling in Damage Mechanics Next, an admissible elastic cone is introduced
whereby $ is a convex function and Y = YQ + H(K). Moreover, associated evolution equations DÂ" = A<9 A b$, D f /c = Ad#$ = A are applied. The underlying constrained optimisation problem yields the Kuhn-Tucker conditions and the Lagrange multiplier A can be computed via the consistency condition. Nevertheless, for the sake of demonstration, in the sequel the hardening contributions are assumed to be constant. NOTE 3.1 - Standard pull-back operations yield e.g. S* = /"" 5tt [ff and Ab = [F ]* A F which allow the interpretation as effective stress measures. 3.2. Construction of the damage function The specific form of the damage function significantly affects the evolution of anisotropic damage. Therefore, we will especially focus on the evolution of the eigendirections of the damage metric tensor. The most general form is of course based on the set of ten invariants in terms of the damage stress Ab and the damage metric A* itself
Nevertheless, two selected representations seem to be natural and will be highlighted in the sequel, compare Schreyer [SCH 95]. The direct formulation introduces the damage rate negative proportional to a symmetric, positive semi-definite second order tensor 2S"(A"). Obviously, the simplest choice 22* = A* results in
and the dissipation inequality reads V = A Y > 0. Indeed, the damage rate and the damage metric itself are coaxial but since the damage metric could be non-spherical we denote this type of damage as quasi isotropic. The formulation based on conjugate variables constructs the damage rate as a linear map of the damage stress via a symmetric, positive semi-definite fourth order tensor 4s"(A"). Again, a simple choice 4S" = A* A* ends up with the quadratic form
and the reduced local form of the Clausius-Duhem inequality reads T> = 2 A Y > 0. Obviously, the damage rate and the damage metric are no longer coaxial which motivates the terminology anisotropic damage.
235
NOTE 3.2 - Eqs.(19,20) can alternatively be motivated via the following quadratic form
whereby for the fourth order tensor 4S the structure of the tangent operator of linear elasticity has been adopted. Now, due to the central idea of the proposed framework, the contra-variant metric tensor is replace by the damage metric
Then the first term, incorporating the scalar rji, represents the quasi isotropic damage function (p\ of eq.(19) and the second term, incorporating the scalar 772, represents the anisotropic damage function (pi ofeq.(20). NOTE 3.3 - In contrast to isotropy the incorporation of an in-elastic potential together with the application of associated evolution equations within an anisotropic setting is an assumption since the obtained rate equations represent reduced forms of the most general tensor functions in terms of all appropriate arguments, see e.g. Betten [BET 85]. NOTE 3.4 The two introduced types of damage functions and the character of the initial damage metric tensor AQ define a general classification of the coupling of hyper-elasticity and damage. In particular, the following four categories are obtained: (1) isotropic hyper-elasticity (A\ = (2) isotropic hyper-elasticity (A\ = (3) anisotropic hyper-elasticity (AQ (4) anisotropic hyper-elasticity (A0
i|
/?0 G*) & quasi isotropic damage ((p\) /?0 G*) & anisotropic damage ((p?) ^ /?o G*) & quasi isotropic damage ((pi) ^ /?o G*) & anisotropic damage ((ft)
Assuming a material of St. Venant-Kirchhojf type category (1) is directly related to the classical [I a] damage formulation via A^ = foG^ = [I d]2 G". In this case flo represents three equal eigenvalues, which degrade for increasing damage, e.g. characterised by d. Moreover, note especially that formulations within category (2) become anisotropic within the purely elastic domain for unloading after damage evolution has taken place. 3.3. Numerical examples In the sequel we apply a compressible Mooney-Rivlin material of the form
whereby the principal invariants
Ji,2,3 are expressed in terms of the basic invari-
236

EA
ants
I\,2,3, to be specific
In order to define the initial damage metric tensor we chose the specific format Aj = 70 G" + Y%=i TCAN\ ANC Co?*1"6 Svendsen [SVE ]. Moreover, the following material parameter have been taken into account c\ = 10, c^ = 20, \p 5, 70 = 1-0, 71 = 0.5, 72 = 0.25, Y = 10 and the orthogonal unit-vectors AN\^ are define by spherical coordinates tfj = 2/3 TT, 0} = 1/3 rr, tf\ = 4/3 TT, 0} = 1/6 TT (compare Appendix). For quasi isotropic damage (<î) an exponential integration scheme is available whereas for anisotropic damage evolution (<,02) an implicit Euler backward scheme is applied. We do not focus on numerical aspects here since they are discussed in detail in Menzel and Steinmann [MEN 00]. Within the following finite element setting the tangent operator has been evaluated numerically as e.g. outlined in Miehe [MIE 96]. 3.3.1 Simple shear We consider again the homogeneous deformation of simple shear (F = I + 761 62) and take anisotropic damage (1^2) into account. Figure 6 shows the different degradations of the eigenvalues of the damage metric tensor and highlights the non-coaxiality of stress and strain, compare Appendix. In addition the stereographic projection of the principal damage directions are given which evolve during the deformation process.
Figure 6. Degradation of the eigenvalues Aî,2,3 and stereo-graphic projection due to the principal directions of strain E : o, stress S : and the damage metric A : * with respect to a Cartesian frame
Anisotropic Damage at Large Strains 237 3.3.2 Strip with a hole To give a three-dimensional finite element example a strip with a hole is discretised by 64 x 2 enhanced eight node bricks (Q1E9) as advocated by Simo and Armero [SIM 92]. The geometry of the specimen is defined by a length of 12, a width of 4, a thickness of 0.5 and a radius of 1. One end is totally clamped while the other end is subject to displacement conditions in a longitudinal direction. Figure 7 shows the deformed mesh and the anisotropy measure S for a maximal longitudinal stretch of A = 1.5 for the purely hyper-elastic solution (Y > oo). In addition the loaddisplacement curve of the mid node at the un-clamped end underlines the anisotropic behaviour since the displacement components u\ and us would be identical to zero within an isotropic setting. Now, incorporating damage evolution the degradation is concentrated (un-symmetrically) at the boundary of the hole. In this context Figure 8 highlights the smallest eigenvalue A AI of the damage metric tensor and the anisotropy measure S within quasi isotropic damage ((p\). Moreover, Figure 9 visualises the same contents for anisotropic damage (</>2)-
Figure 7. Anisotropy measure S and deformed mesh for a maximal longitudinal stretch A = 1.5 and load-displacement curve of the un-clamped mid point node at the free end within pure hyper-elasticity (Y -> oo)
4. Outlook The proposed thermodynamically consistent framework for anisotropic materials at large strains results in a manageable numerical setting. Nevertheless, concerning future work, localisation has somehow to be taken into account. Furthermore, the coupling to finite strain plasticity is an outstanding issue and finally the construction of specific damage functions for engineering material as well as the corresponding identification of material parameters are important areas constituting future research. Appendix: Visualisation of Anisotropy For three-dimensional examples especially it is not a trivial task to visualise anisotropy. In the sequel three different propositions are made.
Figure 8. Smallest eigenvalue A\i of the damage metric and anisotropy measure 6 for quasi isotropic damage (<pi) at \\F\\ = 114
Figure 9. Smallest eigenvalue A\i of the damage metric and anisotropy measure 6 for anisotropic isotropic damage (ipz) at \\F\\ = 65. In case that stress and strain tensors are not coaxial we deal with an anisotropic material. This motivates the introduction of the anisotropy measure
Within the method of stereographic projection the eigenvectors of a symmetric second order tensor - which allow interpretation of being elements of the unit-sphere S2 - are projected onto the equatorial plane by viewing from the south pole. Mathematically speaking, this method represents the homomorphism 5O(3) > SU(2), see e.g. Altmann [ALT 86]. Determinant of the acoustic tensor: Incorporating the common wave equation ansatz into the incremental equation of motion for F^ = const yields (see e.g. Antman [ANT 95])
with
whereby
239
the spatial unit-vector rfi 52 can be defined by two spherical coordinates $1>2. Referring to a Cartesian frame Si, one of three possible parametrisations to define n nl i 6 S2 reads
5. References [ALT 86] ALTMANN S. L. Rotation, Quaternions and Double Groups. Oxford University Press, 1986. [ANT 95] ANTMAN S. S. Nonlinear Problems of Elasticity. Number 107 in Applied Mathematical Sciences. Springer, 1995. [BET 82] BETTEN J. Net-stress analysis in creep mechanics . Ing. Arch., 52:405419, 1982. [BET 85] BETTEN J. The classical plastic potential theory in comparison with the tensor function theory . Eng. Fract. Mech., 21(4):641-652,1985. [HAL 75] HALPHEN B. AND NGUYEN Q.S. Sur les materiaux standards generalises . J. Mecanique, 14:39-62,1975. [HAUOO] HAUPT P. Continuum Mechanics and Theory of Materials. Advanced Texts in Physics. Springer, 2000. [KAU 49] KAUDERER H. Uber ein nichtlineares Elastizitatsgesetz . IngenieurArchiv, XVH:450-480,1949. [KRA 87] KRAJCINOVIC D. AND LEMAITRE J. Continuum Damage Mechanics. Number 295 in CISM Courses and Lectures. Springer, 1987. [LEM 96] LEMAITRE J. A Course on Damage Mechanics. Springer, 2nd edition, 1996. [MAR 94] MARSDEN I.E. AND HUGHES T.J.R. Mathematical Foundations of Elasticity. Dover, 1994. [MEN 99] MENZEL A. AND STEINMANN P. A theoretical and computational setting for geometrically nonlinear damage mechanics . In WUNDERLICH W., Ed., Proceedings of the European Conference on Computational Mechanics, number 329. ECCM, Munich, 31.08.-03.09. 1999. [MEN 00] MENZEL A. AND STEINMANN P. A theoretical and computational setting for anisotropic continuum damage mechanics at large strains . Int. J. Solids & Structures, 2000. accepted for publication.
[MIE 96] MIEHE C. Numerical computation of algorithmic (consistent) tangent moduli in large-strain computational inelasticity . Comp. Meth. Appl. Mech. Engrg., 134:223-240,1996. [MUR 88] MURAKAMI S. Mechanical modeling of material damage . ASME J. Appl. Mech., 55, 1988. [ODE 72] ODEN J.T. Finite Elements of Nonlinear Continua. Advanced Engineering Series. McGraw-Hill, 1972. [OGD 97] OGDEN R.W. Non-Linear Elastic Deformations. Dover, 1997. [SCH95] SCHREYER H.L. Continuum damage based on elastic projection tensors . Int. J. Damage Mechanics, 4:171-195,1995. [SID 81] SlDOROFF F. Description of anisotropic damage application to elasticity . In HULT J. AND LEMAITRE J., Eds, Physical Non-Linearities in Structural Analysis. IUTAM Symposium Senlis/France, Springer, 27.-30.05. 1981. [SIM 92] SlMO J.C. AND ARMERO F. Geometrically non-linear enhanced strain mixed methods and the method of incompatible modes . Int. J. Num. Meth. Eng., 33:1413-1449,1992. [SPE 84] SPENCER A. J.M. Constitutive theory of strongly anisotropic solids . In SPENCER A. J.M., Ed., Continuum Theory of the Mechanics of Fibre-Reinforced Composites, number 282 in CISM Courses and Lectures. Springer, 1984. [STE 98] STEINMANN P. AND CAROL I. A framework for geometrically nonlinear continuum damage mechanics . Int. J. Engng. Sci., 36:1793-1814,1998. [SVE ] SVENDSEN B. On the modeling of anisotropic elastic and inelastic material behaviour at large deformation. Preprint.
Chapter 14
On the Numerical Implementation of a Finite Strain Anisotropic Damage Model based upon the Logarithmic Rate
Otto Timme Bruhns and Christian Ndzi Bongmba
Institute of Mechanics, Ruhr-University Bochum, Germany
Finite Strain Anisotropic Damage Model
243
1. Introduction Materials undergoing finite deformation suffer a loss in load carrying capacity caused by the nucleation, growth, and coalescence of microdefects which is generally referred to as damage. Save the simplest, but nontrivial, special case where the material is initially isotropic and the microdefects are voids, any damaged material is anisotropic. The first damage model was presented by KACHANOV [KAC 58]. This model is generally accepted and forms the fundamental basis of Continuum Damage Mechanics (COM). Several generalizations of this one-dimensional model of the threedimensional regime exist. Here, we mention isotropic damage models of the type described in [LEM 90] and anisotropic models based upon ideas introduced in [COR 79], and [MUR 88]. For an extensive bibliography, we refer to the review article [KRA 89] and the books [LEM 96, KRA 96] on COM. The isotropic damage model is limited to the case where the material is initially isotropic and all the defects are voids. Moreover, as shown by Ju [JU 90], this model fails to predict the effects of damage on Poisson's ratio. Assuming isotropic conditions, one can easily show that the anisotropic damage model predicts only the change in a single material parameter. And, there is even a more fundamental problem with this model, since it is usually developed using a symmetric effective stress measure a. Generally, cr is derived by the symmetrization of a stress measure denoted in [MUR 88] by cr*, which is a function of the damage variable, and considered to be unsymmetric. To derive cr*, finite deformation kinematics is used. In fact, cr* is a first Piola-Kirchhoff type stress measure, a two-point tensor, and represents a bilinear map between two different vector spaces - the current and the fictitious undamaged configuration. One can therefore not expect cr* to be symmetric, since symmetry will require that we identify the two vector spaces. We mention here the fact that the nominal stress tensor, also a two-point tensor, plays a prominent role in non-linear elasticity and elastoplasticity [OGD 84, HIL 78]. We note also that using the work-conjugacy notion introduced by HILL [HIL 78], any stress measure (and its work-conjugate strain measure) can be used in formulating non linear constitutive equations. Moreover, note that one can derive, using finite deformation kinematics, a symmetric stress measure (a second Piola-Kirchhoff type tensor) that is defined in the fictitious undamaged configuration, and that none of the effective stresses employed in the literature is equal to this stress measure. The ad hoc symmetrization of cr* and damage models based upon the effective stress & are therefore questionable. Within a phenomenological framework, the anisotropy induced by damage is indistinguishable from the so-called microstructural or initial anisotropy. In general, microstructural anisotropy is modelled by the introduction of so-called structural or material tensors [DOY 56, BOE 87]. BOEHLER [BOE 87], for example, uses the following set as material tensors:
where (711,712,713) denote orthonormal vectors. Apparently, the set in Equation [1] can be used as a basis for the representation of any second order tensor.
In this paper, we present an anisotropic damage model and show its numerical implementation. Based on the above mentioned fact about structural tensors and anisotropy, we make use of an interpretation of the damage parameter as an evolving structural tensor to model damage-induced anisotropy. We assume the damage parameter to be a second order, symmetric, and positive semi-definite tensor. On the irreversible thermodynamics side, the damage parameter is treated as an internal state variable. With these interpretations, damage modelling effectively reduces to formulating thermodynamically consistent evolution equations, finding critical values for the damage variable and determining material constants. The framework of [XIA 00] is capable of effectively incorporating anisotropic material behaviour, and is therefore going to be adopted here as basis for the damage model; i. e. we use the spatial logarithmic strain and its work-conjugate stress measure. Also, we use a corotating frame and all our rate constitutive equations are formulated using the logarithmic rate. An outline of this paper is as follows. In Section 2, we briefly review the kinematic foundations of finite deformation. Strain and stress measures are presented in Section 3. In Section 4, we describe briefly the logarithmic rate and the logarithmic spin, and discuss the advantages of the logarithmic rate over other objective corotational time derivatives. In Section 5, within the framework of [XIA 00], we specify our damage model. The Gibbs and dissipation potentials, and the evolution equations are specified. Using thermodynamics with internal state variables, restrictions on the material parameters are derived. From Gurson's flow potential (which serves here as a micromechanical basis), we determine the material parameters in our Hill-type yield condition. In Section 6, the material moduli for computing the stiffness matrix and a step-by-step summary of the integration of the constitutive laws are given. We also show that our model is kinematically consistent.
2. Basic kinematic quantities Let X denote a particle of a body B undergoing finite deformation. Let X have position vectors X in the reference and undeformed configuration Boand x in the current configuration B. The motion that takes X into x is denoted by x(X,t}. The two-point tensor F = Gradx(X,t) with Jacobian J = detF > 0 is called the deformation gradient. From the polar decomposition theorem, we obtain
where R is the rotation and U and V are the right and left stretch tensors, respectively. The tensors defined as
are referred to, respectively, as the right and left Cauchy-Green tensors. Let A^ (a = 1, ,n) denote the eigenvalues of V, \a = A^ those of B, Bff their corresponding
245
orthonormal eigenprojections, and n the number of distinct eigenvalues. Then, B has the spectral representation
Here, 6ij is the Kronecker delta, and 1 denotes the second order identity tensor. Let v denote the velocity of X. The spatial velocity gradient is given by L = gradu = FF~1 and has the following unique additive decomposition:
The symmetric tensor D is called the stretching, and the antisymmetric tensor W the vorticity.
3. Strain and stress measures Following [XIA 00], we use the following form for Hill's generalized strain measures:
where Ca P?BaR. The scale function f(Xff), a smooth monotonic increasing function with the property /(I) = /'(I) 1 = 0, is given by
where for m = 0 the limiting process is understood. Note that [6] include most commonly used strain measures; Hencky's logarithmic strain measures H and h are obtained by setting m = 0; i. e. / = In A:
The spatial Hencky strain h is a forward rotation of the material Hencky strain H, and vice versa:
Let a- denote the Cauchy stress tensor, and r the Kirchhoff stress tensor, cr and r are related by r = Jcr. Again, following [XIAOO], we denote the work-conjugate stress measure to H and h in the sense of HILL [HIL 78] independent of any material
246 Numerical Modelling in Damage Mechanics symmetry by II and TT, respectively; i. e. the stress power per unit reference volume can be written as
where () R denotes the Green-Naghdi rate of (), see Section 4. The following expression for the Eulerian stress measure it is taken from [XIA 00]:
7T
The Lagrangian stress measure II is a back-rotation of TT, and vice versa.
4. Log-spin and Log-rate Inelastic material behaviour has to be formulated in rate or incremental form. In a Eulerian setting, this requires the use of objective corotational time derivatives. The general form for an objective corotational time derivative is
where G is a time-differentiable, objective, second order, Eulerian tensor, and 1 is a

o
time-dependent skew-symmetric, second order tensor chosen such that G is objective. For a general discussion of objective corotational rates and their defining spins, we refer to [XIA 98a] and the references cited therein. Well-known examples of objective corotational rates are the Zaremba-Jaumann rate G} for Q = W, and the GreenNaghdi or polar rate GR for tt = 17R = RRJ.
o
O
For the logarithmic rate (log-rate)
G'
1 08
of G, we use the the logarithmic spin (log-spin)
in Equation [12]. The time-dependent rotation tensor .RLo8 that defines the log-spin is referred to as the logarithmic rotation (log-rotation) and is the solution of the tensor differential equation
The log-rate was recently proposed by XIAO, BRUHNS & MEYERS [XIA 97, XIA 98a, XIA 98b] and has since been successfully used in formulating constitutive equations [BRU 99, XIA 00]. The log-rate has two major advantages over any other objective corotational rate. First, the log-rate of the spatial Hencky strain h is equal to the stretching D:
Finite Strain Anisotropic Damage Model 247
Actually, the validity of Equation [15] was the prime motivation for the derivation of the log-spin and the log-rate. Second, the hypo-elastic model based upon the lograte is self-consistent; i. e. it is exactly integrable to deliver an (hyper)elastic relation [BRU99].
5. Constitutive equations The damage variable T> is assumed to be a second order, symmetric, and positive semi-definite tensor. It is a member of the set a of internal state variables that also includes the hardening variables K and a:
The Gibbs potential 3> is assumed to be a function of TT and a, and the stretching D to have the additive decomposition
where De denotes the elastic part of D and Dei the elastic-inelastic part. Then, starting from the Clausius-Duhem inequality and using a standard procedure due to CoLEMAN & NOLL [COL 63], we derive the (hyper)elastic constitutive equation
and the dissipation inequality
where p0 is the density in the reference configuration BQ. Using Equation [15], and introducing the complementary hyperelastic potential as = PQ$, we obtain the following exactly-integrable rate form of the elastic constitutive law:
To separate elastic from inelastic deformations, we use, following KRONER [KRO 60] and LEE [LEE 69], the multiplicative decomposition of the deformation gradient:
where Fe denotes the elastic part and Fl the damage-plastic or simply the inelastic part. Note that [21] introduces an intermediate configuration which is assumed to be stress free and is unique only to within an arbitrary rotation. To relate the two decompositions [17] and [21], we use
where here sym() denotes the symmetric part of (). We note that [22] represents the only natural, and direct relation between the two decompositions [17] and [21] and that such a relation does not exist if we assume, as widely done, that Fe is a symmetric tensor [XIA 00].
5.1. Particular form of the Gibbs and dissipation potentials Following LEHMANN [LEH 89], we assume that the Gibbs potential takes the additive decomposition
where the elastic part 3>e is assumed to depend only on the stress, and the damage variable and the inelastic part <!>' only on the hardening variables. Using the interpretation of P as a structural tensor, it follows that <e is an isotropic tensor function of TT and 'D. Using the representation theorem of tensor functions, we postulate the following quadratic function in TT for E:
with
and
Here, I denotes the symmetric fourth order identity tensor, and 771-7/4 are assumed to be material constants. The additive split of the compliance tensor D> into an isotropic or undamaged part DP and a damage-induced part ^ is motivated by micromechanical results given in [BUD 76]. With [24] and for isotropic damage, the elastic law takes the form
where k,p depend on the damage variable. The elastic law in the form [27] is due to Hencky and is widely used in finite elastoplasticity [ANA 86, SIM 92, SCH 95]. For the inelastic part <J>! of the Gibbs potential, we use
where Hkin is the kinematic hardening modulus. The conjugate forces to X>, a, and K are respectively,
Let the function F(?r,a) = 0 denote the yield criterion in stress space. Then, using F, we define the elastic domain as the set E^ := {(7r,a) : F(7r,a) < 0} . Taking
Finite Strain Anisotropic Damage Model 249 the back stress tensor X as the centre of the yield surface (F 0), and the damage parameter as an evolving structural tensor, we derive the following anisotropic or Hilltype yield condition:
Here a 3 ,05,07,ag; 60 and 61; 7r0(the initial yield stress) are material constants and /3 denotes the determinant of T>. The fourth order tensor A is called the anisotropy tensor, and its additive decomposition into an isotropic or undamaged part A and a damageinduced part Ad is partly motivated by the additive decomposition of the compliance tensor. Note that A is an orthogonal projection that maps a second order tensor into 3/2-times its deviatoric part, and that without damage, Equation [32] reduces to the flow condition of classical J-2 isotropic plasticity. Perhaps the most generally accepted and widely used isotropic damage model is the micromechanical model of GURSON [GUR 77]. The most important ingredient of this model is the flow condition
where #1 = <fe = 1.5 and qi = I are material parameters [TVE 82], and () D refers to the deviator of (). In its original form, FQ is a function of tr, so that Gurson's flow condtion as given here in [36] is a further modification. Gurson's model has been extensively and successfully used by many researchers to model, amongst many other things, localization, shear banding and macrocracking. To put [32] on a sound micromechanical basis, its material parameters are therefore determined from [36]. To this end, we consider a material that is isotropically damaged under one-dimensional conditions; i. e. we take T2 = fl, and assume a loading case where only one normal component, say, TTH = TT of the stress tensor is different from zero. In this case, / can be regarded as the void volume fraction. From [32] and [36], the yield stresses are, respectively,
with the weighting functions w<2. and w\ given by
250
Figure 1. Weighting functions w\ andwi
Figure 2. Flow conditions FQ and F. Left: isotropic case with f = 0.10. Right: anisotropic case, where only the component T>n = 0.10 is different from zero
Considering [36] as exact, and taking the critical value of / as 0.5 [HIL 65], the material parameters
for [32] were chosen to get the best fit between w\ and u>2 as shown in Figure 1. The two-dimensional case with the set of parameters from [40] is shown in Figure 2. For the isotropic case, there is virtually no difference between the two flow conditions. Both flow conditions are inside the initial or undamaged yield surface. For the simplest case of anisotropic damage also shown in Figure 2, the flow condition [32] lies between the initial yield surface and Gurson's yield surface and predicts, as expected, a decrease in flow stress in ~K\ -direction with virtually no change in the TT^-direction.
251
5.2. Evolution equations and thermodynamic restrictions The evolution equations for the elastic-inelastic strain and the hardening variables are of the associative plasticity type:
The plastic multiplier A is determined from the consistency condition F = 0 and satisfies the loading/unloading condition
Based on the characteristics of the damage variable mentioned above, we postulate the following evolution equation:
where /?i,/?2 > 1 are material parameters and ()+ denotes the positive projection of (). Substituting the evolution equations into the dissipation inequality, we obtain the following thermodynamic restrictions on the material constants 773 and 774 and the function m:
Note that [46] is consistent with the fact that for a given stress, the complementary hyperelastic potential E of a damaged material is greater than that of an undamaged material, and that these inequalities simply state (or ensure) that damage causes softening.
6. Numerical implementation 6.1. Material moduli In standard displacement-type finite elements, the starting point of the solution process with the Newton-Raphson iteration scheme is the discretized form of the weak form of momentum balance. The Lie derivative
of the Kirchhoff stress appears naturally in the rate form of the weak form of momentum balance [SIM 98, Equation (7.2.21)]. To obtain the material moduli C11 which are
needed in computing the material part of the stiffness matrix, we proceed as follows. Using the basis-free expression for fiLog from [BRU 99], we obtain the following relation between TL and rLog:
Substituting Equation [48] into the rate form of the weak form of momentum balance, we obtain the material moduli C" as
where Op are the elastoplastic tangent moduli. SlMO & TAYLOR [SIM 85] showed that to retain the quadratic rate of asymptotic convergence of Newton's method, the elastoplastic tangent in Equation [53] should be replaced by the consistent tangent matrix. For our model, the consistent tangent matrix is currently computed numerically using the perturbation method as described, for example, in [KOJ 87].
6.2. Integration of the constitutive equations To integrate the constitutive equations from Section 5, we use the operator split method [SIM 85, SIM 98]. Tables 1 and 2 contain a step-by-step summary of the integration algorithm. In the following, we explain our notation, and comment on some aspects of the integration. First, we note the fact that in the elastic predictor step, and consistent with the use of the log-rate, the tensor internal variables are forward-rotated with -Ru8, the relative log-rotation:
where R+i and H1^8 refer to the log-rotation at time n +i and tn, respectively. Numerical solution of the tensor differential equation [14] yields an exponential map for
253
Table 1. Integration of the constitutive equations jRy8, and this is evaluated as described in [SIM 98]. Second, we mention the fact that for moderate strains the log-rotation /ZLog approximately equals the transpose of the rotation tensor R from the polar decomposition of F. Third, note that with the help of Lemma A from [XIA 00] all quantities related to the decomposition [21] can be uniquely determined; from the stress update algorithm (Table 2), we obtain the stress 7r n +i. Then using [18], the hyperelastic form of the constitutive equation, we determine Ve:
The elastic stretching >e is taken from the rate form of the elastic law. Then, with the elastic spin tensor given by
and from a tensorial differential equation similar to [14] for .Re, we determine the elastic rotation tensor JRe. From the polar decomposition and the multiplicative decomposition, Fe and Fl follow. The model presented is therefore regarded as kinematically consistent, since Fe and F1 and their related kinematic quantities can be
Table 2. Integration of the constitutive equations
uniquely determined without the use of ad hoc assumptions about Fe. Note also that with [56] our model satisfies the objectivity requirement in a general sense [NAG 90]. For details, we refer to [XIA 00].
6.3. Example To demonstrate the performance of our model, we consider the plane-stress extension of a perforated strip. This is a classical example, and here the geometry, boundary conditions, and material parameters are chosen as in [SIM 98] page 189. For damage evolution, the material parameters are taken as /^ = 0.10 and /^ = 0.0; i.e. we are considering only isotropic damage. In Figure 3, we have plotted the reactiondisplacement curves for the computation without damage, and with damage. Damage causes softening of the material. On the structural level, this softening is reflected in the reduction of the total reaction forces and is clearly, as shown in Figure 3, captured by the model. Qualitatively, these results are in a good agreement with other results reported in the literature, see for example [SOU 92].
255
Figure 3. Reaction versus displacement curves 7. Summary and conclusions A finite anisotropic damage model was presented. The elastoplasticity framework of [XIA 00] was used. Thus, the model can be regarded as kinematically, and selfconsistent. It was shown that using a dual interpretation of the damage parameter as 1) an evolving structural tensor and 2) an internal variable, damage modelling reduces to formulating thermodynamically consistent evolution equations, finding critical values for the damage variable and determining material constants. We think that the former interpretation of the damage parameter, since it ties progress in damage modelling with that in the modelling of so-called strongly anisotropic materials, will prove useful in future. On the numerical implementation side, a closed form expression for the material moduli was given. Thus, the implementation of the model in existing finite element codes can be readily accomplished. More numerical examples are presented elsewhere. Acknowledgements We thank Dr. H. Xiao, Dr.-Ing. A. Meyers and Dipl.-Ing. H. Schiitte for their interest and the fruitful discussions. This paper is abstracted from work conducted within the Collaborative Research Center SFB 398 at the Ruhr University Bochum. Financial support was provided by the Deutsche Forschungsgemeinschaft (DFG).
8. References
[ANA 86] ANAND L., Moderate deformations in extension-torsion of incompressible isotropic elastic materials , Journal of the Mechanics and Physics of Solids, vol. 34, n 3,1986, p. 293-304. [BOB 87] BOEHLER J. P., Representation for isotropic and anisotropic non-polynomial tensor functions , BOEHLER J. P., Ed., Applications of Tensor Functions in Solid Mechanics, n 292 CISM Courses and Lectures, Springer-Verlag, 1987, p. 31-53.
[BRU 99] BRUHNS O. T., XIAO H., MEYERS A., Self-consistent Eulerian rate type elastoplasticity models based upon the logarithmic stress rate , International Journal of Plasticity, vol. 15, n 5, 1999, p. 479-520. [BUD 76] BUDIANSKY B., O'CONNEL R. J., Elastic moduli of a cracked solid , International Journal of Solids and Structures, vol. 12, 1976, p. 81-97. [COL 63] COLEMAN B. D., NOLL W., The thermodynamics of elastic materials with heat conduction and viscosity , Archive for Rational Mechanics and Analysis, vol. 13, 1963, p. 167-178. [COR 79] CORDEBOIS J. P., SIDOROFF F., Damage Induced Elastic Anisotropy , BOEHLER J. P., Ed., Mechanical Behavior ofAnisotropic Solids, n 295 CNRS, Martinus Nijhoff Publishers, 1979, p. 761-774. [DOY 56] DOYLE T. C., ERICKSEN J. L., Nonlinear Elasticity, N 4 Advances in Applied Mechanics, Academic Press, New York, 1956. [GUR77] GURSON A. L., Continuum theory of ductile rupture by void nucleation and growth: Part I - Yield criteria and flow rules for porous ductile media , Journal of Engineering Materials and Technology, vol. 99, 1977, p. 2-15. [HIL 65] HILL R., A self-consistent mechanics of composite materials , Journal of the Mechanics and Physics of Solids, vol. 13, 1965, p. 213-222. [HIL 78] HILL R., Aspects of invariance in solid mechanics , YlH C.-S., Ed., Advances in Applied Mechanics, vol. 18, p. 1-75, Academic Press, New York, 1978. [JU 90] Ju J. W., Isotropic and anisotropic damage variables in continuum damage mechanics , Journal of Engineering Mechanics, vol. 116, n 12, 1990, p. 2764-2770. [KAC 58] KACHANOV L. M., On the time to failure under creep conditions , Isw. ANSSSR. Old Techn. Nauk, vol. 8, 1958, p. 26-31, Zitat aus Jansson & Stigh (1985). [KOJ 87] KOJIC M., BATHE K.-J., The effective-stress-function algorithm for thermoelasto-plasticity and creep , International Journal of Numerical Methods in Engineering, vol. 24, 1987, p. 1509-1532. [KRA 89] KRAJCINOVIC D., Damage Mechanics , Mechanics of Materials, vol. 8, 1989, p. 117-197. [KRA 96] KRAJCINOVIC D., Damage Mechanics, Applied Mathematics and Mechanics, North-Holland, Amsterdam, 1996. [KRO 60] KROENER E., Allgemeine Kontinuumstheorie der Versetzungen und Eigenspannungen , Archive for Rational Mechanics and Analysis, vol. 4, 1960, p. 273-334. [LEE 69] LEE E. H., Elastic-plastic deformation at finite strains , Journal of Applied Mechanics, vol. 36, 1969, p. 1-6. [LEH 89] LEHMANN T., Some thermodynamical considerations on inelastic deformations including damage processes , Acta Mechanica, vol. 79, 1989, p. 1-24. [LEM 90] LEMAITRE J., CHABOCHE J.-L., Mechanics of Solid Materials, Cambridge University Press, Cambridge, 1990. [LEM 96] LEMAITRE ].,A Course on Damage Mechanics, Springer-Verlag, Berlin, 2 edition, 1996. [MUR 88] MURAKAMI S., Mechanical Modeling of Material Damage , Journal of Applied Mechanics, vol. 55, 1988, p. 280-286. [NAG 90] NAGHDI P. M., A critical review of the state of finite plasticity , Journal of Applied Mathematics and Physics (ZAMP), vol. 41, 1990, p. 315-394.
257
[OGD 84] OGDEN R. W., Non-linear Elastic Deformations, Ellis Horwood Limited, Chichester, 1984. [SCH 95] SCHIECK B., STUMPF H., The appropriate corotational rate, exact formula for the plastic spin and constitutive model for finite elastoplasticity , International Journal of Solids and Structures, vol. 32, n 24, 1995, p. 3643-3667. [SIM 85] SIMO J. C., TAYLOR R. L., Consistent tangent operators for rate-independent elastoplasticity , Computer Methods in Applied Mechanics and Engineering, vol. 48, 1985, p. 101-118. [SIM 92] SIMO J. C., Algorithms for static and dynamic multiplicative plasticity that preserve the classical return mapping schemes of the infinitesimal theory , Computer Methods in Applied Mechanics and Engineering, vol. 99, 1992, p. 61-112. [SIM 98] SIMO J. C., HUGHES T. J. R., Computational Inelasticity, vol. 7 de Interdisciplinary Applied Mathematics, Springer, New York, 1998. [SOU 92] DE SOUZA NETO E., PERIC D., OWEN D. R. J., A Computational model for ductile damage at finite strains , OWEN D. R. J., ONATE E., HlNTON E., Eds., Computational Plasticity: Fundamentals and Applications, Pineridge Press, 1992, p. 1425-1441. [TVE 82] TVERGAARD V., Ductile fracture by cavity nucleation between larger voids , Journal of the Mechanics and Physics of Solids, vol. 30, n 4, 1982, p. 265-286. [XIA 97] XIAO H., BRUHNS O. T., MEYERS A., Logarithmic strain, logarithmic spin and logarithmic rate , Acta Mechanica, vol. 124, 1997, p. 89-105. [XIA 98a] XIAO H., BRUHNS O. T., MEYERS A., On objective corotational rates and their defining spin tensors , International Journal of Solids and Structures, vol. 35, n 30, 1998, p. 4001^014. [XIA 98b] XIAO H., BRUHNS O. T., MEYERS A., Strain rates and material spins , Journal of Elasticity, vol. 52, 1998, p. 1-41. [XIA 00] XIAO H., BRUHNS O. T., MEYERS A., A consistent finite elastoplasticity theory combining additive and multiplicative decomposition of the stretching and the deformation gradient , International Journal of Plasticity, vol. 16, n 2, 2000, p. 143-177.
Chapter 15
Ductile Rupture of Aluminium Sheet Materials

Jacques Besson
Ecole des Mines de Paris Centre des Materiaux, Evry, France; and Institute of Materials Research, GKSS Geesthacht, Germany
Wolfgang Brocks, Olivier Chabanet and Dirk Steglich

Institute of Materials Research, GKSS Geesthacht, Germany
261
1. Introduction A realistic assessment of the residual strength of cracked sheet materials (steel in pipeline or ships, aluminum in aircraft structures, zirconium in nuclear fuel components) requires methods to experimentally characterize crack growth resistance as well as numerical simulation tools capable of predicting crack initiation and propagation. Modeling can be done following two different strategies. On the one hand, the global approach to failure uses macroscopic rupture parameters such as the Crack Tip Opening Angle (CTOA) [GUL 99], the energy dissipation rate [TUR 92], the J-integral [RIC 68] or the JQ parameters [O'D92]. In general, these quantities suffer from a lack of transferability of fracture data from specimen to actual structures. On the other hand, the local approach to fracture provides a solution to the transferability problem by describing the degradation of the material using micromechanical state variables (e.g. void volume fraction, nucleated porosity, etc). These models have been applied successfully to predict crack growth for thick walled components of structural steels [SUN 88, GUL 00] where a high stress triaxiality triggers the growth of voids as the main failure mechanism. Their application to thin walled high strength aluminum alloys faces some specific problems: (i) the stress triaxiality ratio in sheets is much lower than in thick structures whereas models of ductile damage have been established for high triaxialities; (ii) the fracture plane often shifts from a normal to a 45 inclined orientation to the applied load (Figure 2), (iii) rolled sheets generally show an anisotropic behavior with respect to both plastic hardening and void nucleating particles. However, fracture surfaces still show dimples which might result from growth of voids nucleated at second phase particles. Because of this, the application of models of ductile tearing seems to be promising. In this work, the plastic and rupture behavior of thin aluminum sheets is characterized using smooth, blunt and sharp notched specimens. A 3D finite element modeling of fracture of small samples is performed using the Rousselier model [ROU 87]. The emphasis is put on the simulation of slant fracture. As this approach does not appear suitable to model crack extension in large structures, a 2D simulation using a cohesive zone model [NEE 90] is adopted.
2. Material and experiments In this study, sheets of 1.73 mm thickness of a 2024 Al-alloy are investigated. The surface of the sheets is protected by a layer of 1050 aluminum whose thickness varies between 50 and 80 //m. The sheets are heat treated (T351 treatment). Several types of precipitates are present in the material [HER 98]: (i) coarse secondary phases containing iron and copper (size: 5 to 30 //m), (ii) Al^M^Cu dispersoides (size: 20 to 500 nm), (iii) Al2(Cu,Mg) strengthening precipitates (size: a few nm). The samples used to characterize the plastic and fracture behavior of the material are shown in Figure 1. They include: smooth tensile bars, notched bars with U or Vnotches and Kahn cracked specimens. All tests were performed with the load applied
262
Figure 1. Geometries of the specimens: (a) smooth tensile, (b) V-notch plate, (c) Vnotch plate (notch radius: 0.25 mm), (d) Kahn specimen (notch radius: 0.06 mm, ASTM B-871-96). Gray dots indicate measurement points for the notch opening displacement A<5 along the rolling direction (L); cracks propagate along the transverse direction (T). The short transverse direction will be referred to as S. Due to bending, stable crack growth is observed on Kahn samples whereas unstable rupture is obtained on the other specimens. Post-mortem measurements on tensile specimens outside the necking region have shown that the thickness reduction along the Sdirection is larger than along the Tdirection evidencing an anisotropic plastic behavior. Fig. 2 shows the fracture surface of a V-notched specimen. Fracture is initiated at the middle of the notch where the fracture surface is normal to the loading direction (zone in Figure 2b). This region forms a small triangle. Outside this zone, slant fracture in the LS plane is observed. A closer examination of the fracture surfaces shows void growth around the iron/copper particles. Void growth is more pronounced in the flat fracture region than in the slant fracture zones ((D and in Figure 2b). Between these voids, small dimples, probably initiated on dispersoides or Al2(Cu,Mg) precipitates, are observed. Similar features were observed on Kahn specimens. On tensile and U-notched specimens, the macroscopic fracture surface is entirely slanted (Figures 4 and 5). An example of the microscopic fracture surface is shown in Figure 3. The rupture surface exhibits flat and smooth areas where small dimples
263
Figure 2. Fractography (V-notch plate): (a) Macroscopic view showing the inclined crack path, (b) Macroscopic view showing (L direction) the flat crack initiation region (triangle), (c) Microscopic fracture faces corresponding to different regions of the crack path (1,2 and 3; locations are indicated on (b))
cannot be seen. This indicates a possible change in the fracture mechanisms between cracked or severely notched samples and specimens containing smooth geometrical defects. The proportion of smooth fracture zones increases in the case of the tensile specimens.
Figure 3. Fracture surface of smooth (a) and U-notched specimens (b) exhibiting flat fracture zones
264
3. Continuous damage mechanics: Rousselier model 3.1. Constitutive equations Continuous damage mechanics (COM) can be used to describe the rupture behavior of the material. In this study, the Rousselier model has been employed [ROU 87]. The original model is extended to account for plastic anisotropy of the matrix material and damage nucleation. A similar extension was used in [GRA 00] for the Gurson model. In that case the yield surface is expressed as:
cTjfcjfc is the trace of the stress tensor a_ and an is the Hill equivalent stress defined by:
where Sij are the components of the stress deviator. The coefficients hss... are used to describe the plastic anisotropy. Using the deformation measurements and considering that the yield limit in the T-direction is 10% smaller than in the L-direction [HER 98, SIE99] one gets: h^ = 0.79, h^ = 1.16, hss = 1.68. It was assumed that: /IST = h^ hSL = 1; this assumption has however little importance as shear stresses aST, cr-^ and <TSL remain small. R(p) represent the plastic hardening of the undamaged material (shown in Figure 4) and p the von Mises accumulated plastic strain. D and a\ are material parameters related to void growth. D = 2 and a\ 260 MPa were used and correspond to the values recommended in [ROU 87]. The plastic strain rate tensor is obtained using the normality rule as: ep = \d$lda_ where A is the plastic multiplier. The evolution of the void volume fraction / is given by mass conservation:
The initial porosity /o is taken equal to the volume fraction of coarse particles which are assumed to debond at the onset of plasticity: /0 = 0.12% [HER 98]. Using these parameters, ductility is however overestimated. Based on the observation that smaller precipitates can play a role in the fracture process, strain controlled nucleation was used in the model. The nucleated effective porosity / is given by
The total effective porosity ft appearing in eq. 1 is defined as the sum ft / + fnConsidering that nucleation starts when a given amount of plastic strain has been reached, An will be written as: An = A0H(p p0) where H () is the Heavy side function and AQ and p0 are parameters to be adjusted. Different values for AQ and po will be used depending on the specimen type.
Ductile Rupture of Aluminium Sheet Materials 3.2. Computational procedures
265
FE calculations are performed using the software Zebulon [BES 98]. Constitutive equations are integrated using a fully implicit scheme which allows the calculation of the consistent tangent matrix [SIM 85]. An updated Lagrangian formulation is used; the Jauman stress rate is used to define the objective stress rate; constitutive equations are expressed in the associated corotational frame [LAD 80]. In order to reduce the number of degrees of freedom, mixed 2D/3D meshes are used. Geometrical symmetries are also accounted for. 20 nodes (3D) bricks with reduced integration (8 Gauss points) are used to model regions of crack extension. The 2D parts are freely meshed using 6 or 8 nodes plane stress elements (see 4.2.1). In order to reduce the number of elements, computations exploit the initial geometrical symmetries of the various specimens. This symmetry is however lost due to localization as evidenced by the fracture surfaces. Accounting for this effect would require twice as many elements for a given mesh size. The material is considered as broken as ft reaches 85%. The behavior is then considered as elastic with a very low Young's modulus (1 MPa) as proposed in [LIU 94]. An element containing more than 5 broken Gauss points is automatically removed by checking this condition after each time increment.
3.3. Results and discussion 3.3.1. Smooth and U-notched specimens Simulated macroscopic responses and fracture modes for smooth and U-notched specimens are compared with experimental results in Figures 4 and 5 (F: force, S0 initial minimum cross section, LQ, SQ: initial gauge length, AL: elongation, A5: notch opening displacement (see Figure 1)). All simulations were carried out using AQ~ 0.5 andpo = 0.35. Smooth specimens: The macroscopic response of the specimen and the fracture plane are well reproduced. Failure occurs after the onset of necking. The fracture surface corresponds to a plane whose normal lies in the LS plane. The angle 9 between the normal and the Ldirection is about 45. The simulation was also performed assuming an isotropic plastic behavior; in this case the normal also lies in the LS plane with a smaller value for 9 which is close to localization angle under plane stress in an isotropic incompressible material (35.26) usually observed in the LT plane. The fact that localization still occurs in the LS plane is likely related to the relatively small width of the specimen. Note however that due to the plastic anisotropy, e^ < e^ which favors localization in the LS plane [RUD 75]. U-notched specimen: Rupture of U-notched specimens is very similar to failure in smooth specimens. One single fracture plane is observed which also makes a 45
Figure 4. Simulation of the rupture of tensile specimens. Left: Macroscopic response: (I) no damage, no necking, (2) no damage, necking, (3) damage and necking. The arrow indicates the experimental rupture. The photograph shows the experimental fracture surface (9 w 45). Right: Values of the plastic deformation at Gauss points showing the localization into slant fracture
angle in the LS plane. Macroscopic behavior as well as fracture plane are well reproduced by the FE simulation (Figure 5). NOTES - For both samples, FE simulations were carried out using various values for both AQ and po as well as different mesh sizes. Very similar results are obtained using more than 5 elements to mesh the Sdirection (10 are used in Figures 4 and 5). The influence of po and A0 is limited for values of po larger than 0.3. This indicates that the onset of failure of smooth and U-notched specimens is essentially controlled by necking and void growth. Accounting for rapid damage by nucleation is only needed to model final failure. 3.3.2. V-notched and Kahn specimens In V-notched and Kahn specimens, the sharp notch generates strong stress and strain gradients so that a higher mesh size sensitivity is to be expected. Transition from flat to slanted fracture is experimentally observed. In this work, the emphasis was put on the modeling of this 3D phenomenon. Preliminary plane strain calculations have shown that a slanted crack path can be obtained provided a sufficiently fine mesh is used. In addition, elements should have a square shape at the onset of damage and strain localization. 3D meshes were designed according to these results. Figure 6 shows the crack path obtained with 7 elements in the S-direction using various values of A0 and p0 = 0.2. It is shown that increasing AQ leads to an increase
267
Figure 5. Simulation of the rupture of U-notched specimens. Left: Macroscopic response (the arrow indicates the experimental rupture). Right: Values of the plastic deformation at Gauss points showing slant fracture of the flat fracture area. High values of AQ generate a highly damaged zone which remains limited to one element so that crack deflection is not possible. A similar trend is obtained with increasing PQ. In this case the low hardening rate of the material limits the size of zones where p is higher than PQ . Figure 7 (calculation done with AQ = 0.5 and p0 = 0.2) indicates the values of the stress triaxiality ratio T ahead of the crack tip after the initiation of the flat crack area. Values as high as 1.6 are obtained showing that plane stress conditions are not met; in that case the maximum value of the stress triaxiality is equal to %. However, this maximum value remains smaller than values obtained under plane strain which lie in the range 2.53.0. Figure 7 also shows details of the flat to slanted fracture transition.
Figure 6. Effect of AQ on the flat to slanted fracture transition (PQ = 0.2)
Figure 7. Left: Gauss point value of damage (ft) and stress triaxiality (T) at crack initiation. Right: Final crack path showing details of the flat to slanted fracture transition. 10 elements are used in the Sdirection, AQ = 0.5, po = 0.2
The calculation qualitatively reproduces the experimental results shown in Figure 2 but overestimates the size of the flat region. The zone of normal fracture is smaller when using a smaller mesh size (e.g. Figure 6) but still larger than what is experimentally observed (Figure 2). Simulations of the Kahn specimen are shown in Figure 8. Calculations with 5,7 and 10 elements along the Sdirection using AQ 0.5 and po = 0.2 were performed. Decreasing the element size leads to a smaller crack growth resistance as it is expected due to the softening behavior of the material; however simulations overestimate the experimental data. For a given element size, meshing one half of the specimen (i.e. not accounting for the initial symmetries) would also reduce the simulated crack growth resistance. Faster crack growth can also be obtained by modifying the void nucleation parameters. As an example, Figure 8 shows the macroscopic response obtained for Po = 0.15. Although fractographic examinations have shown that failure mechanisms change with increasing defect severity, such a low value is unlikely to be realistic. An other fitting strategy would be to increase po and AQ (i.e. delay onset of nucleation and increase nucleation rate); this however leads to flat fracture as the mesh is then too coarse to correctly capture the deformation and damage fields. Mesh size plays a critical issue in modeling crack advance. It is often considered as a material parameter which needs to be adjusted [ROU 87, GUL 00]. Using nodular cast iron having the same volume fraction of nodules but different particle sizes, it has been shown that ductility decreases with decreasing interparticle spacing and that this effect can be modeled using meshes of decreasing sizes [STE 98]. In this
269
Figure 8. Left: Forcenotch opening curves using 5, 7 and 10 elements to mesh the thickness direction (S) using AQ = 0.5 and PQ 0.2. 7* corresponds to 7 elements withpo = 0.15. Points are experimental data. On the photograph, the normal fracture zone is within the white dashed line. Right: Simulated crack path using 10 elements in the Sdirection
study, fractographies have shown (Figure 2) that the mean distance between iron/copper particles is in the range 1020 //m which is significantly smaller than the mesh size used in the present calculations (= 80/zm). However, using such a small mesh size would require a much larger computational effort (both CPU time and memory size). Note also that the use any non-local damage model would indeed require a smaller mesh size than the mesh size needed to fit the experiments using a local theory. In order to reproduce the experimental loaddisplacement curves on Kahn specimens, nucleation parameters can be changed to obtain a faster damage kinetic; in that case the normal to slant fracture transition cannot be modeled so that the benefit of 3D calculations is lost.
4. Cohesive zone model for plane stress state In the previous section, it has been shown that the 3D-modeling of crack growth along a slanted path requires a large number of elements. This approach does not appear as suitable in the case of large panels; in this case a simplified 2D-modeling has to be developed. Several possibilities are however possible: (i) adjust Rousselier parameters in the plane stress case, (ii) use CTOA or CTOD, (iii) describe the crack path as a damageable interface. In all cases the actual stress state at the crack tip will not be accurately described (for instance the stress triaxiality will be underestimated).
It becomes indeed impossible to represent the slanted crack path. It is therefore likely that the associated material parameters will be geometry dependent [SIE 99].
4.1. Cohesive zone model In the present investigation, a simplified treatment of crack growth is proposed based on a cohesive zone model. This type of model describes the material damage and separation along a surface [NEE 92]. For all specimens, the crack is assumed to stay in the initial plane of symmetry. In that case, only normal separation is to be considered. The mechanical behavior of the interface is described using a constitutive relation relating the normal component of the stress vector on the interface (crn) to the normal opening of the interface (<). The following relation will be used in this study:
where e = exp(l). 6C and crmax are two adjustable parameters of the model. crmax represents the maximum stress carrying capacity of the interface. 8C is a characteristic length. The function ern (5W) is shown in Figure 9a. The fracture energy per unit surface F is given by:
In practice, it will be considered that the interface is broken (i.e. an = 0) when 6n > 6C so that the actual fracture energy is reduced by 4.6%.
4.2. Computational procedures 4.2.1. Plane stress elements Plane stress [along the z-direction] conditions are usually taken into account by rewriting the material constitutive equations so that the condition crzz = 0 is enforced. However, this method requires modifying the implementation of every material behavior. Another strategy, which requires more computational time but less programming, testing and maintenance efforts, relies of the implementation of plane stress elements [BES 97, BES 98]. These elements have the usual nodal degrees of freedom (displacements) and additional degrees of freedom (elzz) representing the transformation gradient along the z-direction at each integration point (i = 1.. .number of Gauss points): e\z = &W/WQ (WQ initial thickness, Au; thickness variation). The principle of virtual work is written in the current (end of time increment) configuration in order to derive the reactions associated with the different degrees of freedom. It can then be shown that the reaction R\ corresponding to each additional degree of freedom is equal to Rle = crzzul where a/ is the volume associated to the corresponding Gauss point. In the absence of prescribed values for
271
Figure 9. (a) Traction-separation law (eq. 5). (b) Cohesive zone elements with surrounding plane stress elements elzz or applied values of azz (e.g. pressure field), the force equilibrium requires that Rle = 0, so that the plane stress condition is enforced at each Gauss point. 4.2.2. Cohesive zone elements Cohesive zone elements are placed along the crack path and embedded in a surrounding of eight nodes large deformation plane stress elements with reduced integration (4 Gauss points). CZ elements have 6 nodes and 2 integration points. Due to symmetry, surface integrals have to be evaluated on the line of symmetry (Figure 9b). The virtual work for each CZ element is written as (/' denotes the end of the time increment):
where <5* denotes a virtual normal displacement. The surface integral is computed accounting for the deformation of the symmetry plane but also for the sheet thickness reduction. This is done by associating each Gauss point of the CZ elements with the nearest Gauss point of the continuum plane stress elements whose thickness reduction is then used to evaluate the surface integral. Tests have shown the importance of accounting for thickness reduction. In the case of constant thickness CZ elements a slight increase of the cohesive strength or a slight modification of the crack tip constraint can result in the localization of the plastic deformation in the above lying continuum elements with no crack advance. The calculation is then equivalent to the reference elastoplastic large deformation computation without CZ.
272
Figure 10. Simulation of the Kahn specimen using the CZM: (a) Force vs. notch opening for different mesh sizes: 1: 100 x 100 //in2. 2: 200 x 200 pm2, 4: 400 x 400 /mi2; p: no crack growth, (b) Crack extension vs. notch opening (200 x 200 /zm2 mesh). Contour map shows the stress (0-^2) in the loading direction. (Points represent experimental data) 4.3. Results The loadnotch opening curve was adjusted using the following parameters: 0max = 720 MPa, 6C = 80 /mi. Results of the simulations shows a good agreement for both the macroscopic response (Figure lOa) and crack advance (Figure lOb). The macroscopic response is also compared to the response obtained without CZ. In that case failure occurs by plastic necking. Calculations were performed with an element size of 200/mi x 200/mi and using a coarser (400/mi x 400/mi) and finer (100/mi x 100/mi) mesh. Results obtained using 200/mi and 400/mi differ only slightly whereas a faster crack advance is obtained with 100/mi. This is due to the larger deformations that are computed in that case at the crack tip so that necking (and therefore stresses) is overestimated. However, mesh size dependence remains smaller than in the case of CDM (section 3).
5. Summary Ductile tearing of Aluminum 2024 sheets has been studied. The following results were obtained: Experiments. Experiments were conducted on smooth and moderately notched specimens as well as severely notched samples. The plastic behavior is anisotropic.
273
Fractography shows that failure mechanisms change between these two sets of specimens with an increase of void growth in the second case. The crack path is slanted for moderately notched specimens whereas a transition from normal to slant fracture is observed on the second set. 3D modeling using COM: 3D modeling using an extension of the Rousselier model to plastically anisotropic material was used to represent the experimental results. Strain controlled nucleation was also modeled. On the one hand, failure of tensile and U-notched specimen is essentially controlled by necking and results depend only slightly on the nucleation parameters. In that case, experimental results can be well described. On the other hand, nucleation and mesh size play a very important role in the case of V-notched and Kahn specimens. The main features of ductile tearing (in particular the normal to slant fracture transition) are qualitatively described but the load is overestimated. Adjusting the nucleation parameters in order to obtain the experimental load leads to a flat crack path. This is attributed to the relatively coarse mesh needed to perform the 3D calculations. It is believed that using a finer mesh would allow correct description of the load and the crack path and to use material parameters closer to those obtained on moderately notched specimens. Simulations show that stress triaxiality levels reached ahead of the crack tip are much higher than those obtained under plane stress conditions. 2D modeling using CZM: In order to overcome the difficulties encountered using the 3D CDM modeling, a simplified model using CZ elements modified to account for necking in the plane stress case has been adopted. Good results are then obtained on Kahn specimens; however, based on the results presented in [SIE 99] the transferability of these parameters to other cracked geometries is questionable. Acknowledgements: This work was performed during the sabbatical leave of JB at GKSS which is acknowledged for financial support and hospitality. Materials for the study were provided by Pechiney which also financially supported OC. Many thanks to Dr. J. Heerens and his team for carrying out the experiments and to V. Heitmann for the SEM examinations.
6. References
[BBS 97] BESSON J., FOERCH R., "Large scale object-oriented finite element code design", Computer Methods in Applied Mechanics and Engineering, vol. 142, 1997, p. 165-187. [BES 98] BESSON J., FOERCH R., "Application of object-oriented programming techniques to the finite element method. Part I- General concepts", Revue europeenne des elements finis, vol. 7, num. 5, 1998, p. 535-566. [GRA 00] GRANGE M., BESSON J., ANDRIEU E., "An anisotropic Gurson model to represent the ductile rupture of hydrided Zircaloy-4 sheets", Int. J. Fracture, vol. 105, num. 3, 2000, p. 273-293.
274
[GUL99] GULLERUD A., DODDS R., HAMPTON R., DAWICKE D., "Three dimensional modeling of ductile crack growth in thin sheet metals: computational aspects and validation", Eng. Fracture Mechanics, vol. 63, 1999, p. 347-373. [GUL 00] GULLERUD A., GAO X., DODDS JR R., HAJ-ALI R., "Simulation of ductile crack growth using computational cells: numerical aspects", Eng. Fracture Mechanics, vol. 66, 2000, p. 65-92. [HER 98] HERMANN G., Dechirure ductile de toles minces d'un alliage Aluminium-Cuivre, report, 1998, Pechiney CRV. [LAD 80] LADEVEZE P., Sur la theorie de la plasticite en grandes deformations, report, 1980, Rapport interne No. 9, LMT, ENS Cachan. [LIU 94] Liu Y., MURAKAMI S., KANAGAWA Y., "Mesh-dependence and stress singularity in finite element analysis of creep crack growth by continuum damage mechanics approach", Eur. J. Mech, A/Solids, vol. 13, num. 3, 1994, p. 395-417. [NEE 90] NEEDLEMAN A., "An analysis of tensile decohesion along an interface", J. Mech. Phys. Solids, vol. 38, 1990, p. 289-324. [NEE 92] NEEDLEMAN A., "An analysis of decohesion along an imperfect interface", Int. J. Fracture, vol. 40, 1992, p. 1377-1397. [O'D92] O'Dowo N., SHIH C., "Family of crack-tip fields characterized by a triaxiality parameter-II. Fracture Applications", J. Mech. Phys. Solids, vol. 40, num. 8, 1992, p. 939963. [RIC 68] RICE J., "A path independent integral and the approximate analysis of strain concentration by notched and cracks", J. Appl. Mech., vol. 35, 1968, p. 379. [ROU 87] ROUSSELIER G., "Ductile fracture models and their potential in local approach of fracture", Nuclear Engineering and Design, vol. 105, 1987, p. 97-111. [RUD 75] RUDNICKI J., RICE J., "Conditions for the localization of deformation in pressuresensitive dilatant materials", J. Mech. Phys. Sol., vol. 23, 1975, p. 371-394. [SIE 99] SlEGMUND T., BROCKS W., "Prediction of the work of separation and implications to modelling", Int. J. Frac., vol. 99, 1999, p. 97-116. [SIM 85] SlMO J., TAYLOR R., "Consistent tangent operators for rate-independent elastoplasticity", Computer Methods in Applied Mechanics and Engineering, vol. 48,1985, p. 101-118. [STE98] STEGLICH D., BROCKS W., "Micromechanical modelling of damage and fracture of ductile materials", Fatigue Fract. Engng Mater. Struct., vol. 21, 1998, p. 1175-1188. [SUN 88] SUN D., SlEGELE D., VOSS B., SCHMITT W., "Application of local damage models to the numerical analysis of ductile rupture", Fatigue and Fracture of Engineering Materials and Structures, vol. 12, 1988, p. 201-212. [TUR 92] TURNER C., "A re-assessment of ductile tearing resistance. Part I: The geometry dependence of J-R curves in fully plastic bending. Part II: Energy dissipation rate and associated Jft-curves in fully plastic bending.", Fracture behavior and design of materials and structures, ECF 8, 1992, p. 933-968.
Chapter 16
On Identification of Small Defects by Vibration Tests

Yitshak M. Ram
Department of Mechanical Engineering, Louisiana State University, USA
George Z. Voyiadjis
Department of Civil and Environmental Engineering, Louisiana State University, USA
Identification of Small Defects
277
1. Introduction There is a wealth of literature associated with diagnostic criteria, procedures, and methods for damage detection in structures and systems. With technological progress, high-speed machinery and transport carriers are prone to fatigue and catastrophic failure. There is thus an increasing interest in developing reliable diagnostic methods allowing the detection of structural defects at an early stage. Following the early work of Cawley and Adams [CAW 79], a variety of methods for identifying the existence of damage have been proposed, see e.g. [GAS 98], [HEA 91], [LIA 92], [RAT 00], [REY 96], and [REY 00]. If the damage exceeds a certain significant level then it can be detected by changes in the spectral properties. It is a matter of debate, however, whether damage can be identified at an early stage, at which the physical parameters of the system are only slightly altered from their nominal undamaged values. An inherent difficulty associated with repeatability of modal testing results conducted in independent laboratories prevents easy confirmation or rejection of published results and procedures associated with damage detection. It appears that the fundamental problem of determining whether there exist measurable quantities that are highly sensitive to small changes in the physical parameters has not been appropriately addressed yet. In this context an analytical example demonstrating the possibility of identifying small damages in the theoretical model framework of vibrating systems may be of more significant importance than merely displaying experimental results. In this paper we furnished such an analytical example, and through it address the fundamental problem of identifying small defects in structures by means of vibration tests. For simplicity consider a conservative vibrating system modelled by:
with symmetric positive definite mass matrix M , and non-negative definite stiffness matrix K. It is well known that the eigenvalues of the system [2] are continuous functions of the elements in the system matrices, see e.g. [PAR 80]. If the eigenvalues are distinct then the mode-shapes are continuous functions of the physical parameters as well. Changes in the physical parameters of the system are related by bounds to changes in the orientation of the eigenvectors, as shown in [RAM 93]. Hence small changes in the physical parameters may produce only small variation in the spectral data. The measurable data in modal test however are rational functions, e.g.,
where H(JCO} is the frequency response function and P(JCO) and Q(JCO) are functions depending on the physical properties of the system. Hence continuity of CO with respect to the coefficients of P(jcd) and <2(jto) does not necessarily imply continuity of co with respect to the elements of H(JCO). We present in Section 2 an example demonstrating that a certain measured function is highly sensitive to changes in the physical parameters of the system. Physical interpretation of this result is given in Section 3. In Section 4, while applying the result to groove identification in a vibrating rod, we find that the frequency of excitation required in determining the location of the groove is extremely high. We therefore conclude that the problem of whether a small damage in a realistic structure can be identified by a dynamic test is still subject to debate.
2. Damage in a discrete model of a uniform vibrating rod
Figure 1. Mass-spring system
Consider the n-degree-of-freedom system shown in Figure 1, which consists of masses mi and springs of constants kf, / = l,2,...,n. Such a system represents a discrete lumped-parameter-model, or finite difference model, of an axially vibrating rod. Suppose that the harmonic excitation f ( t ) = sin(cot) is applied to they-th degree of freedom. Then, the motion of the system is described by the set of ordinary differential equations
where:
Identification of Small Defects 279
and ey is the j-th unit vector of dimension n. The system [2] has a particular solution of the form
where h; is a constant vector. Substituting [5] in [2] gives
The ^-equations defined by [6] can be assembled as follows
where I is the identity matrix and
We thus have
The matrix H(a>) is called the Frequency Response Function (FRF) matrix. Its element hy (co) represents the steady state amplitude of the harmonic response in the j'-th degree-of-freedom due to a unit sinusoidal excitation applied to the y'-th degree-offreedom. Hence, the elements of H(O;) can be determined by simple vibration tests.
Figure 2. Systems description: (a) pure system, and (b) damaged system
Consider now the uniform system of unit parameters mi=ki =1, 1 = 1,2,...,100 of dimension n = 100, shown in Figure 2(a). A harmonic exciting force with frequency ft) = 2 is applied to the system at they-th degree-of-freedom. Suppose that the constant of the p-th spring is reduced due to a damage to kp = 0.99 , as shown in Figure 2(b). Let /i,,(2) and hu(l) be the collocated frequency-response-functions at j=l,2,...,n of the pure system and the damaged system, respectively. These functions are plotted in Figure 3(a) for the case where p = 25, i.e. the damage is applied to the 25-th spring. Similar graphs for the cases when the damage is applied to the 50-th and 75-th degrees of freedom are shown in Figures 3(b) and 3(c), respectively. It is apparent that the damage and its location is clearly observable by discontinuity in slope of hu(2). In contrast, damaged applied to the 5-th spring cannot be unambiguously identified as shown in Figure 3(d). Such a result can be grasped by intuition. The fifth spring is too close to the support and hence cannot be excited easily by the harmonic force.
Figure 3. Damage identification
281
It should be note that other harmonic forces, of different frequencies, may not provide clear identification of the damage and its location. For example, Figures 4(a) and 4(b) display the functions hn (1.5) and ha (1.5) for the case where the damage is located at the p = 50 spring and the exciting frequency is (0 = 1.5. The two functions associated with the pure and damaged systems look almost identical. Figures 4(c) and 4(d) display these frequency-response-functions for the case where the exciting frequency is co = 2.5 . Here a variation between the two functions at the damage location p = 50 is observed. This variation is small, however, and cannot be considered as a reliable criterion for damage detection for realistic systems. These intriguing results deserve further considerations.
3. The interpretation Denote the stiffness matrix of the undamaged system shown in Figure 2(a) by K , and let K be the stiffness matrix of the damaged system (Figure 2(b)). Then the stiffness matrix of the undamaged system shown in Figure 2(a) is:
and the stiffness matrix of the damaged system (Figure 2(b)) is:
Both systems share the same mass matrix M = I. Define a diagonal matrix:
282
Figure 4. Excitations by various frequencies
283
Then, since D = D ', we have by [6]
which implies that
For co = 2 equation [14] takes the following explicit form:
The physical interpretation of equation [15] is that hu is the static deflection of the y'-th node due to a collocated unit static load applied to if, as shown in Figure 5(a). Note that the spring configuration of the system of Figure 5(a) is similar to that of the undamaged system shown in Figure 2(a), but with an additional spring of constant k = 2 attached between the n-th node and the ground. Following a similar process we find that for the damaged system the equation
holds, or explicitly:
284 Numerical Modelling in Damage Mechanics The element hit in equation [15] represents the static deflection of the^-th node due to a collocated unit static load applied to it, as shown in Figure 5(b). When the applied force is in the close neighbourhood of the damaged the two springs of constant 28 became dominant resulting in large change in the response. The systems of Figure 5 can be represented by using equivalent springs as shown in Figure 6, where
Figure 5. Static models for (a) pure system, and (b) damaged system
It thus follows that
In a similar manner, using:
Identification of Small Defects we obtain from Figure 6(b):
285
Figure 6. Equivalent static models for (a) pure system, and (b) damaged system
The response ^(2)
in Equation [19] describes a smooth function in j, while
hjj (2) features derivative discontinuity at j = p , which allows identification of the damage and its location. In fact, the plots in Figures 3 and 4 are precisely h^ (2) and h^ (2) given by equations [19] and [21].
4. Damage identification in an axially vibrating rod Consider an axially vibrating uniform rod of length L, modulus of elasticity E, density p and cross-sectional area A , which is fixed at one end, jc = 0, and free to oscillate at the other end, x = L. The axial vibrations of this rod are governed by the differential equation and boundary conditions
A discrete mass-spring system model of dimension n , with equal length elements of length h = L/n , leads to the eigenvalue problem:
where
and where K is given by [10].
Figure 7. An axially vibrating uniform rod with groove at x = a A corresponding damaged rod with groove at x = a is shown in Figure 7. Let a = (p - i)h for some integer 1 < p < n . Let the groove width be w , and let Al < A be the cross-sectional of the rod at the groove position. Then the eigenvalue problem associated with the damaged rod takes the form
given by [11],
Identification of Small Defects 287
and:
In order to identify the location of the groove in this system using the method presented in Section 2 it is required to excite the system with the frequency
For steel rod p=7800 kg/m3, E = 1.962 xlO11 N/m2, of length L = l m, with A = 0.01 m2, AI - 0.0095 w = 0.01 m, n = 50, and p = 25 , we obtain: 8 = 0.0256, rj =0.9750, and co =5.0154xl05 rad/s. The functions h^co) for the uniform rod, and hjj(a)) associated with the damaged rod, are shown in Figure 8. Although the damaged position is observable, its effect is less drastic than that obtained in Section 2 for the chain of mass-spring system. The reason is that the groove in the rod affects both the stiffness and the mass of the p-th element. In practice it may be difficult to excite the rod with the high frequency CO required. With current progress in materials research, however, such excitations using piezoelectric materials are not entirely beyond futuristic expectations.
Figure 8. Groove identification in an axially vibrating rod
5. Conclusions There are procedures in the literature for damage identification, accompanying experimental results. The inherent difficulty associated with the repeatability of modal analysis testing forms a barrier in achieving a scientific consensus regarding the applicability of these procedures when slightly damaged systems are considered. In the theoretical arena the fundamental problem regarding the existence of measurable quantities, which are sensitive to small changes in the physical parameters of the system, is still open. In an effort to address this issue we have presented an analytical example showing that under certain circumstances small changes in the stiffness of a uniform chain of mass-spring system can be detected. The example was restricted by the need for imposing a special frequency of excitation. It is well known that the frequency response of a damped system is smoother than that associated with its conservative counterpart. Hence damage in a damped system may less identifiable. In the context of realistic components, such as identification of a small groove in an axially vibrating steel rod, the method requires excitation with very high frequencies. We may thus argue that the fundamental question whether small defects in structures can be detected by vibration tests is still subject to debate.
6. References
[CAW 79] CAWLEY P., ADAMS R.D., "The location of defects in structures from measurements of natural frequencies", Journal of Strain Analysis, vol. 4, 1979, p. 49-57. [GAS 98] CASSER A., LADEVEZE P., PERES P., "Damage modelling for a laminated ceramic composite", Materials Science and Engineering A-Structural Materials Properties Microstructure and Processing, vol. 250, 1998, p. 249-255. [HEA 91] HEARN G., TESTA R.B., "Modal analysis for damage detection in structures", ASCE Journal of Structural Engineering, vol. 117 n 6, 1991, p. 3042-3063. [LIA 92] LIANG R.Y., Hu J., CHOY F., "Theoretical study of crack-induced eigenfrequency changes on beam structures", ASCE Journal of Engineering Mechanics, vol. 118, 1992, p. 384-396. [PAR 80] PARLETT B.N., The Symmetric Eigenvalue Problem, Englewood Cliffs, PrenticeHall, 1980. [RAM 93] RAM Y.M., BRAUN S.G., "Eigenvector error bounds and their applications to structural modification", AIAA Journal, vol. 31, 1993, p. 759-764. [RAT 00] RATCLIFFE C.P., "A frequency and curvature based experimental method for locating damage in structures", ASME Journal of Vibration and Acoustics, vol. 122, 2000, p. 324-329. [REY 96] REYNIER M., NADJAR B., "Identification of defaults in beams structures using static tests", Mecanique Industrielle el Materiaux, vol. 49, 1996, p. 28-31.
289
[REY 00] REYNIER M., NADJAR B., "Damage identification of flexible beam structures using large displacements", Inverse Problems in Engineering, vol. 8, 2000, p. 251-281.
Chapter 17
Multi-scale Non-linear FE2 Analysis of Composite Structures: Damage and Fiber Size Effects
Frederic Feyel and Jean-Louis Chaboche
ONERA, DMSE-LCME, France
Multi-scale Non-linear FE Analysis 293 1. Introduction
The inelastic analysis of structural components working under complex and severe environments, especially under high temperature cyclic loading conditions, has now an increasing impact on structural design. Most often these computations of inelastic response and stress redistributions serve to predict damage development and the component lifetime, either in uncoupled or in coupled damage simulations. The great reduction in computational costs and the considerable improvements made in parallelisation techniques and substructuring methods open new possibilities for improved numerical simulations. Several directions for enlarged representativeness of non linear structural analyses can be considered. Let us summarise them in three words: time, size and scale: - in the "time" direction, we have the capability to compute the component on its whole life, incorporating coupled inelastic and damage effects, really important for taking into account non-stationary material evolutions, especially under cyclic loading conditions. The "cycle jump technique" for instance [SAV 78, LES 89, DUN 94, NES 00] has proved to be particularly efficient to enable integration of non linear evolutions incrementally at two time scales : the real time increments within each computed cycle, and increments in number of cycles (external "time") to treat large numbers of cycles. We have also the possibility to introduce "multiphysics" coupled analyses; - in the "size" direction, we can perform the non linear analysis of large size three-dimensional finite element models , with very large numbers of degrees of freedom (we expect quite soon IMDoF for components treated in viscoplasticity), in order to improve the geometrical complexities of the structure as well as its complicated loading conditions. In this domain the progress of finite element codes in terms of parallel treatments is exceptionally powerful and promising [QUI 96, ROU 94, FAR 91, DUR 97,BJ0 86]. - for the "scale" direction, we can consider the "multiscale structural analyses", in which the material constitutive behaviour itself is built up from in-situ numerical computations. The present paper considers developments in this third class of methods. Multiscale modelling of structural components can be considered at two different levels : (i) - the "sequential multiscale analyses", successively using micromechanics models, analytical or numerical, to deliver the material constitutive responses on which more or less macroscopic models are identified and the finite element structural analysis itself, still based on these macroscopic constitutive equations. (ii) - the "integratedmultiscale analyses", in which the micromechanical local behaviors and criteria are incorporated directly into the finite element structural analysis. There is no more need for a macroscopic constitutive model that reproduces the typical responses of the micromechanics analysis. In fact, the numerical analysis of the lower
scale delivers, in situ and in real time, the appropriate material response to its specific overall loading, ie the overall strain for each local material Representative Volume Element in the structure, at each Gauss Point. Two such numerical approaches of the structural inelastic analyses were developed and exploited recently: - for polycrystalline metallic components, treated in 3D, the polycrystalline aggregate models, with uniform stress within each grain and crystal plasticity constitutive laws at the level of average slip systems inside each grain. With between 40 and 1000 grains for each local RVE, it leads to 1000 to 10000 state variables at each Gauss Point of the overall finite element model. Examples of such applications are given in [FEY 97b], but there are also other attempts in the literature [PIL 90, CAI 94]. Improved such models as "multicrystalline aggregates", with 3D third order stress fields redistributions within the grains are presently studied, but only at the level of a single RVE [QUI 99, BAR OOa, BAR 00b], not at the component level. A similar method was proposed independently in a different context, by Smit et al. [SMI 98]. - for composite structures a two-level imbricated finite element methodology, called FE2 [FEY 98, FEY 00], was proposed and applied to MMC's. This method solves the local stress equilibrium and constitutive equations, inside each RVE treated at the microstructural level by periodic homogenization and a unit-cell finite element model, as well as the overall stress equilibrium at the structural level. In that case the number of internal state variables for each macroscopic Gauss Point can be increased by one or two orders of magnitude compared with the previous case. The present paper discusses the conditions of application of this second class of methods for problems related with long fiber SiC/Ti MMC's used in the context of "bling" components (or "bladed rings"), as candidates to replace turbine and/or compressor discs in future aircraft turboengines. In that case the approach is exploited as a 2D or 2D| problem (generalised plane strain). We first recall the main lines of the FE2 multiscale approach (section 2), including some details about its implementation and use. In section 3 some recent improvements made about "relocalization" techniques are discussed, that take into account the "material lengthscale" effect induced by the presence of a "coarse grain" microstructure. Section 4 presents some results on a schematic bling, treated in cyclic elasto-viscoplasticity, the same constitutive equation being used at the macroscopic level for the pure Titanium alloy part and for the matrix inside the unit cell of the microstructural level. Moreover, additional recent exploitations of the FE2 method include damaging effects through cohesive zone models used at the lower scale, at the fibre-matrix interface. Some examples are also given on the use of the specific relocalisation technique in order to obtain stress, strain and plastic strain local fields in the extreme cases where the microstructure (fibre size) is extremely large (significantly larger than the wavelength of the overall homogenised solution).
Multi-scale Non-linear FE2 Analysis 295 2. The FE2 method 2.1. Periodic homogenization theory: summary As with other homogenization theories, the main objective of the periodic homogenization theory is to access the mechanical properties of a homogenous medium which exhibits the same mechanical response as a given heterogenous medium. This theory is based on a hypothesis to simplify the analysis: - the macroscopic and microscopic scales are supposed to be separated. In other words, the characteristic size of all heterogeneities / is supposed to be small enough relative to the macroscopic length L : 77 = l/L 1, - the spatial distribution of all heterogeneities is supposed to be periodic. A point in the heterogenous structure can then be located using two spatial coordinates : a macroscopic coordinate x (which is also the coordinate of that point in the homogeneous structure) and a microscopic coordinate y which is the location of that point around the heterogeneity (whose size goes to zero). As r; = l/L 1, it is possible to perform an asymptotic expansion of the stress and displacement fields with respect to r?:
The smaller the heterogeneities are, the smaller 77 is and the smaller the higher order terms in the previous developments are. Writing the micromechanical equilibrium diva = 0 and using the Hill-Mandel macrohomogeneity lemma leads to the definition of the macroscopic stress and strain tensors :
where < A > denotes the spatial average of \(y) over the microscopic unit cell. This also induces the definition of elastic stress and strain localization tensors (in the absence of interface discontinuities):
Another important result of the theory is that the displacement field u can be split into a periodic part v and a macroscopic contribution:
296
Figure 1. Principle ofFE2 models 2.2. FE2 principle The "FE2 method" refers to a class of models which belongs to the more general multiscale model class. We suppose in this paper that a "displacement f.e. formulation" is used, ie that (from a macroscopic point of view) the model gives the stresses at time t knowing the strain and the strain rate at that time. As soon as relevant mechanical scales are chosen, FE2 models are constructed using three main ingredients : 1. a modeling of the mechanical behavior at the lower scale (the RVE), 2. a localization rule which determines the local solutions inside the unit cell, for any given overall strain, 3. a homogenization rule giving the macroscopic stress tensor, knowing the micromechanical stress state. In the case of FE2 models (see Figure 1), a finite element computation is used to model the microscopic behavior of the RVE. Any localization / homogenization scheme can be used, but we focus in this paper on the use of the periodic homogenization because one of the current application of FE2 models is to access the mechanical behavior of long fiber SiC/Ti metal matrix composites. Using that theory, the homogenization rule is nothing but a spatial averaging of the microscopic stress distribution : E = (ff)cell. The localization rule is obtained using relation (1) which leads to a set of linear equations to be imposed on each pair of nodes on the sides of the RVE. This class of models is called "FE2" (or also "imbricated finite element") because is requires the simultaneous computation of the mechanical response at two different scales : the macroscopic scale (which is the scale of the whole structure) and the underlying microscopic representative volume element at each macroscopic integration point. Macroscopic phenomenological relations are completely useless, even in non linear cases. The mechanical behavior arises directly from what happens at the microscopic scale, phenomenological constitutive equations being written only at that scale.
Multi-scale Non-linear FE2 Analysis 297
Figure 2. Finite elements : evaluation of material response in the case of a phenomenological model (left) and in the case of an FE2 model (right) combining two finite elements scales 3. Implementation 3.1. General algorithm From an implementation point of view, FE2 models follow the classical framework of internal variables models, and are very easy to implement if modern programming techniques are used [FOE 96, BES 97]. At each macroscopic Gauss Point, such models allow computation of the stress tensor at time t knowing: (i) the strain and strain rate at that time and (ii) the mechanical history since t = 0. In classical phenomenological models, mechanical history is taken into account by the use of some internal variables. In the case of FE2 models, the internal variable set is constructed by assembling all microscopic data required by the lower finite element computation. This includes, of course, microscopic internal variables used to describe dissipative phenomena, but also all other useful quantities required by the finite element procedure.
Figure 2 compares the integration of the macroscopic constitutive equations in the case of a phenomenological model and in the case of an FE2 model. The integration of the phenomenological relations (using, for instance Runge-Kutta or Theta-method) is replaced by a finite element evaluation of the microscopic mechanic cell response.
3.2. Computation of the tangent stiffness matrix FE2 models are used in classical finite element codes, based upon a NewtonRaphson algorithm to handle all non linearities. For optimum performances, one has to compute the tangent stiffness matrix for all material models, and not only to compute the stress response. This matrix can be written as :
where all A in the right member denotes the increment of the quantity between time t and time t + At. In the case of FE2 models, this computation depends of course on the homogeneisation theory used, and on its finite element implementation. In this subsection we want to present this computation in the case of the periodic homogenization theory, in the particular framework of the ZeBuLoN finite element code. Periodic homogenization theory is implemented in ZeBuLoN through the use of specific elements named "periodic elements". These elements are classical ones, except that some degrees of freedom are added corresponding to the E (average or macroscopic strain) components. The unknown displacements are the non-periodic part v of the total displacement u on the cell: u = v -f E x x. The deformation tensor e is computed by derivation of the previous expression :
(Vs is the symmetric gradient operator) The B matrix (symmetric gradient of the shape functions after discretization) is roughly the same as usual, except that a new part comes from degrees of freedom associated with E (we suppose that these degrees of freedom are at the end of the whole degrees of freedom list):
Bstd denotes the "classical" symmetric gradient of the shape functions. Let us recall that the previous finite element discretization is at the microscopic level (ie at the cell scale). The macroscopic stiffness matrix to be computed is then :
Multi-scale Non-linear FE2 Analysis
299
(remember that E has associated degrees of freedom, and thus associated reaction give E to be multiplied by the volume of the cell). The assembled tangent stiffness matrix at the cell scale can be written as (J represents the tangent matrix given by all microscopic phenomenological constitutive equations):
leading to
Kmacro is then nothing but a condensation of the previous matrix onto the degrees of freedom associated with E (at most, 6 degrees of freedom shared across the whole microscopic mesh) inserted at the end of the list. That is
with:
This condensed matrix is then very easy to compute, and avoid the use of other approximative methods, such as the perturbation method.
3.3. Parallel computing Let us suppose that one has to compute a macroscopic structure whose finite element discretization involves K integration points in the region where FE2 modelling is used. If the microscopic discretization of the representative volume element requires k integration points, the cost in terms of global internal variables is equivalent to K x k, and increases very fast with the size of the structure to be studied. It is then necessary to use a powerful technique to solve such a big problem which is usually strongly non-linear. Parallel computing is of great interest in this area. It has been shown that the use of parallel computing can be associated with any sophisticated non-linear behavior provided that this behavior relies on the local state assumption [FEY 98, FEY 97a, FEY 97b]. Parallel computing is also used to compute structures using FE2 behavior models. The attention of the reader is focused on the fact that parallel computing procedures are fully independent of the kind of constitutive equations, and therefore that the use of FE2 models with parallel computing do not require any extra development.
FE2 models have been implemented in the finite element code ZeBuLoN, jointly developed at Onera and Ecole des Mines de Paris. This finite element code uses the FETI method [FAR 91] to split the computation of a structure into a number of subdomains. In FE2 modelings, the main bottleneck is the local stage, ie the computation for each macroscopic integration point of the corresponding microscopic finite element step. Parallel computing allows one to distribute these computations on several processors.
4. Relocalization The classical first order treatment using periodic homogenization relies on a strong assumption : it is supposed that the heterogeneities in the structure to be computed are small enough (compared to the size of the structure and to the mechanical loadings to be applied on that structure) so that macroscopic and microscopic scales are separated. From a practical point of view it is however difficult to define precisely what "small enough" means. From our experience it seems that fibers (ie heterogeneities) may be relatively big, as soon as appropriate relocalisation techniques are used in order to compute actual mechanical fields from the homogeneous solution given (at the macroscopic scale) by the FE2 computation. The purpose of this section is to explicit this techniques applied to the FE2 models. Nothing is new from a theoretical point of view, but it seems that other authors never used this technique (usually because they are interested only in the homogenous result), although it can be used to obtain also the actual mechanical fields. The goal is to compute e(x), for any x, without any reference to y (cell coordinate) because all cells have to be mapped at their real locations. An "interpolated-mapping" of the results obtained by FE2 methods is used to compute actual fields (Figure 3). Let us suppose that all heterogeneities remain elastic and that the surrounding medium is also elastic. For each macroscopic integration point whose spatial coordinate is Xi, and for each position inside the underlying unit cell, the instantaneous strain tensor (for instance) is equal to
The "interpolated-mapping" relocalization technique is nothing but a macroscopic interpolation of results coming from all microscopic computations. Let 7(x) be a mechanical component to be interpolated inside a macroscopic element; 7(x) can be computed using the shape functions of the finite element containing x :
301
Figure 3. Principle of the "interpolated-mapping" relocalization technique Ni(x) are the shape function of the current element, 7(xi) are the values of 7 at the nodes surrounding the point x. This relation can also be applied to microscopic quantities as soon as that the unit cell, resized and translated to its real location and size, is mapped onto the macroscopic mesh. For instance :
This kind of relation can be generalized and extended in non-linear cases like in plasticity or viscoplasticity. One major advantage of using FE2 techniques is that the required estimations of A (which are very difficult to estimate in non-linear cases) to compute relocated components have already been implicitly computed during the FE2 computation. Another advantage is that the computation of relocalized values can be made a posteriori in a post-processor, since it only requires already computed informations. It is then possible to restrict this extra-computation to critical zones of the structure. Note that we restrict ourselves (for practical reasons) to the average on a single macroscopic element: in a more general framework it would be necessary, for a given macroscopic node, to take into account the contribution of all elements connected to
this node. Results presented at the end of this paper will show however that this restriction is not limiting. From a practical and programing point of view, the relocalisation is performed by three steps : 1. map the unit cell onto the global mesh at its real location, 2. extract the microstructural information from all integration points near x, compute nodal values, 3. compute the relocalized component using equation (2). It is relatively easy to understand why this technique leads to continuous macroscopic fields. Let us consider the situation shown in Figure 4: the goal is to compute relocalized fields for two near points A and B. This computation will involve microscopic datas extracted from the macroscopic integration points (plain circles). Due to the periodic homogenization theory, values extracted from microscopic results concerning point A (right side of the microscopic cell) and point B (left side) are equal. Because A and B have the same spatial coordinates, the macroscopic reinterpolation leads to equal values for all mechanical fields.
Figure 4. Illustration of the continuity of all mechanical fields computed using the "interpolated-mapping" technique : thanks to the periodic homogenization theory f ( A ) f ( B ) for all couples of point A andB
5. Application to the computation of a bling disk All critical aeronautical components are subjected to specific weight optimization: new materials are studied in order to increase the performance-to-weight ratio. Long fiber SiC/Ti composites have been developed on this principle. SiC fibres are ordered periodically inside a titanium matrix (see Figure 5). Engine manufacturers are currently considering the possible replacement of some metallic turbine disks by rings whose central part are reinforced by such composites (see Figure 6). This part is a "bling" or bladed ring. A major problem is to be able to
303
Figure 5. Typical microstructure of a long fibre SiC/Ti composite. General layout (left) and detail of a single fibre (right) compute such structures. FE2 models are especially interesting for that purpose. The RVE of the microstructure consists in a fiber surrounded by a matrix part with a fibre volume fraction of 22 %. It is assumed that the average radius of the part is large enough so that the curvature of the fibres can be neglected. Generalized plane strain is then assumed as the condition applied to the unit cell.
5.1. Mesh and boundary conditions Figure 7 shows macroscopic and microscopic meshes as well as macroscopic boundary conditions. Mechanical loading consists in imposing an increasing centrifugal force. The homogeneous part is made of titanium and the reinforced kernel is made of SiC/Ti composite (the titanium in the composite is supposed to be the same as the one used in the homogeneous part).
Figure 6. An experimental "bling"
Figure 7. Macroscopic mesh, boundary conditions and domain decomposition (left microscopic mesh (right)
Figure 8. Macroscopic stress EH (top). Microscopic deformation (displacements x 10) n at point 1 (bottom,left) and 2 (bottom,right)
305
This computation was calibrated to fit in the memory of our cluster (the cluster is made of four PC Linux PIII with 512 Mb of RAM in each machine). The computation was distributed on these 4 processors and lasted about 2 hours.
5.2. Results As explained in section 2, FE2 modeling works by computing in real time the macroscopic and the microscopic scales. It is then possible to analyze mechanical results at these two scales simultaneously. Figure 8 shows for instance the macroscopic stress distribution in the 12 direction (Si2) at the end of the loading. On the same figure microscopic results are presented for two macroscopic points (denoted (1) and (2) at the macroscopic scale) at the same time. It is then possible to make a link between a macroscopic shear (12 ^ 0) and the specific shape of the corresponding unit cell (point 1).
5.3. Case of a coarse grain structure 5.3.1. Undamageable disk The relocalization technique presented in section 4 has been applied to this example. It is here supposed that the fibre are not so small. Figure 9 shows the complete mesh of the disk (ie a mesh of the actual structure, including all heterogeneities). It will serve to obtain a complete reference solution with "coarse grains", from which the FE2 method and the associated relocalization procedure will be validated. The line plotted horizontaly on that mesh highlights the points where results are plotted and compared. Figure 11 shows the stresses Sn and E22 at time t = 7.5s. Figure 10 shows the corresponding cumulated viscoplastic strain to emphasize that non-linearities are strong at that time. The plain curves in Figure 11 are obtained using a computation on the real structure whereas the dashed curves are obtained via the FE2 modelling followed by the relocalization operations exposed before in this paper. The comparison is fairly good, except on the edge of the structure where some side effects appear due to the loss of periodicity (nothing is currently done at present to handle this effect). Figure 12 shows the same comparison for the inelastic strain in the 11 direction (en), whereas Figure 13 shows the contour of this field. 5.3.2. Damageable disk The main damage mechanism in such long fiber composite is a debounding between each fiber and the surrounding matrix. The computation presented in the previous section has been run again, taking into account damage at each fiber/matrix interface using a Needleman-Tvergaard debounding model ([NEE 87]).
Figure 9. Complete mesh of the actual structure to serve as a reference to the FE2 validation
Figure 10. Cumulated viscoplastic strain at time t = 7.5s
307
Figure 11. Comparisons between the reference stresses (plain) and the relocalized stressed (dashed) at time t = 5s fern top, 022 bottom)
Figure 12. Comparisons between the reference inelastic strain (plain) and the relocalized inelastic strain (dashed) at time t = 5s in the 11 direction (E^) The presence of potential discontinuity surfaces modify the homogeneization rules into a slightly different form :
in the previous equation, [aj denotes the jump of the quantity a across the interface F and n the interface normal at a given position x along both sides of the interface, {a} denotes the symmetric part of tensor a. The second term in the homogenization rule of a turns out to be zero, because in the debounding case normal stresses along both sides of all interfaces are null. The homogenization rule for a is then not modified by the presence of a debounding mechanism at fiber/matrix interface. It is then still possible to use the implicitely computed localization tensor A, and therefore to use the relocalization technique presented before in this paper. The extra term in the E relation does not vanish, but it can be proved that it is still possible to define a localization tensor linking E and e. The relocalization technique presented before in this paper is then still valid. The FE2 computation shown in section 5.3.1 has been run again taking into account debounding at fibre/matrix interface. Mechanical fields were then relocalized. Figure 14 shows the an component at the end of the computation. One can observe that stresses are relaxed in fibers due to the partial failure of some interfaces. In these .
Figure 13. Relocalized e\\ at time t = 7.5s in the reinforced region
Figure 14. Relocalized a\\ at time t = 6.76s in the reinforced region, taking into account damage at fiber/matrix interface
311
Figure 15. Relocalized spatial distribution of damage D at time t = 6.76s in the reinforced region
312
regions, stresses are transfered to the surrounding matrix. This contour map can be correlated with Figure 15 showing the damage values D along all interfaces at the same time (this last map has been obtained the same way, although it is not clear whether or not it is possible to use this relocalization scheme for an interfacial field; but this gives a good idea of the spatial damage distribution).
6. Concluding remarks The FE2 methodology presently developed offers several interesting capabilities for the inelastic and damage analysis of structural components, especially for composite systems that can be considered as quasi- periodic at the microstructural level: - the usual macroscopic constitutive and damage equations that serve to redistribute the overall stress fields are no longer needed. All the physics of the processes is contained in the microscale constituents and in their finite element discretisation and the material response to any overall strain control (at each Gauss point) is delivered in "real time", taking into account the whole history of local state variables. - for applications to bling components made in SiC/Ti MMC's, the local constitutive equations were involving the cyclic thermo-elasto-viscoplasticity of the matrix and the damage at the fibre-matrix interface, using debonding models. The material parameters of the matrix constitutive equations were deduced from tests made on a pure matrix [BAR 95] and the debonding models by combining micromechanical tests (push-out) and tension-compression transverse tests. - efficiency of the method is greatly improved by the massive parallel computational capabilities (independently of the parallel solution strategies used for the macroscopic structural analysis); - in order to deliver correct stress and strain fields at the lower scale, application of a specific relocalisation procedure was presented in the extreme case of a very "coarse grain" microstructure, in comparison with the structural size and the macroscopic solution wavelength. Such a method was designed consistently with FE2, based on the classical finite element solution interpolating techniques, considering all local stresses, strains and displacement components in the unit cell as the internal state variables associated with each Gauss Point. - This technique was shown to give adequate results in several examples, especially for the bling analysis. The fibre size has been greatly enlarged over the actual size, in order to be able to compare to the exact reference solution in which all the microstructure has been meshed in the component. The comparisons are extremely good except near the boundary of the composite region in the bling. Some problems or difficulties have still to be treated, with the final objective of a really powerful numerical methodology. The following axes of improvements are currently under examination:
- this relocalisation method must also be generalised in cases where edge effects play a role, due to the abrupt loss of periodicity near structural boundaries or in the region near the (fictitious) interface between the composite substructure and the pure matrix part. Such a situation is specific of our present application to the SiC/Ti MMC's reinforced bling component; A method used recently in a similar context [KRU 98], but from taking overall constitutive equations for the homogeneous equivalent medium (in place of FE2), is assumed to offer good potentialities; - another way to treat the coarse grain microstructure could be to enrich the unit cell periodic homogenisation, building an overall constitutive equation in the framework of Generalised Continuum Media (with material couples or higher order gradients). Recent researches in this area could also provide interesting procedures [FOR 98]; after solving these various problems it is expected to have a method with the potential capability to predict the crack initiation at both the local (micro) level and at the macroscopic level. This capability will have to be checked by the treatment of some specific examples; - moreover, the generalisation of FE2 method should also be examined for microstructures with a lower degree of organisation and periodicity. A first attempt could be studied by introducing a more or less pronounced distortion in the fibre spatial arrangement. This is a long term objective for multiscale numerical methods. Acknowledgments The authors are very grateful to P. Suquet (CNRS-LMA Marseille, France) for a lot of stimulating discussions especially regarding relocalization techniques in periodic homogenization. 7. References
[BAR 95] BAROUMES L., ViNCON I.,Identification du comportement de 1'alliageTi 6242 , Rapport technique contrat Snecma/LMT 762 593F, LMT-Cachan, 1995. [BAR 00a] BARBE F., DECKER L., JEULIN D., CAILLETAUD G., Intergranular and transgranular behavior of polycrystalline aggregates : Part 1, generation of the f.e. model , Int. Journal of Plasticity, 2000, accepted. [BAROOb] BARBE F., FOREST S., CAILLETAUD G., Intergranular and transgranular behavior of polycrystalline aggregates : Part 2, results , Int. Journal of Plasticity, 2000 , accepted. [BBS 97] BESSON J., FOERCH R., Large scale object-oriented finite element code design , Comp. Meth. onAppl. Mechanics and Engineering, vol. 142, p. 165-187, 1997. [BJ0 86] BJ0RSTAD P., WIDLUND O.,Iterative methods for the solution of elliptic problems on regions partitioned into substructures , /. Numer. Anal, vol. 23, p. 1097-1120, 1986.
314 Numerical Modelling in Damage Mechanics [CAI94] CAILLETAUD G., PILVIN P., Utilisation de modeles polycristallins pour le calcul par e'le'ments finis , Revue Europeenne des Elements Finis, vol. 3, n. 4, p. 515-541, 1994. [DUN 94] DUNNE F. P. E., HAYHURST D. R., Efficient cycle jumping techniques for the modelling of materials and structures under cyclic mechanical and thermal loading , Eur. J. Mech., A/Solids, vol. 13, n. 5, p. 639-660, 1994. [DUR 97] DUREISSEIX D., Une approche multi-echelles pour des calculs de structures sur ordinateurs a architecture parallele , PhD thesis, Universite Pierre et Marie Curie, Paris 6,1997. [FAR 91] FARHAT C., Roux F.-X., A method of finite element tearing and interconnecting and its parallel solution algorithm , Int. J. for Numerical Methods in Engineering, vol. 32, p.1205-1227, 1991. [FEY97a] FEYEL F., CAILLETAUD G., KRUCH S., Roux F. X., Application du calcul parallele aux modeles & grand nombre de variables internes , In Colloque National en calcul de structures, may 20-23 1997 , Giens, France. [FEY 97b] FEYEL F., CALLOCH S., MARQUIS D., CAILLETAUD G., F.e. computation of a triaxial specimen using a polycrystalline model , Computational Materials Science, vol. 9, p. 141-157, 1997. [FEY 98] FEYEL F., Application du calcul parallele aux modeles a grand nombre de variables internes , These de Doctoral, Ecole Nationale Sup6rieure des Mines de Paris, 1998. [FEY 00] FEYEL F., CHABOCHE J. L., FE2 multiscale approach for modelling the elastoviscoplastic behaviour of long fibre SiC/Ti composite materials , Computer Methods in Applied Mechanics and Engineering, vol. 183, p. 309-330, 2000. [FOE 96] FOERCH R., Un environnement oriente objet pour la modelisation en mecanique des materiaux , PhD thesis, Ecole Nationale Supeiieure des Mines de Paris, 1996. [FOR 98] FOREST S., SAB K., Cosserat overall modeling of heterogeneous materials , Mechanics Research Communications, vol. 25, p. 449-454, 1998. [KRU 98] KRUCH S., FOREST S., Computation of coarse grain structures using an homogeneous equivalent medium , J. Phys. IV France, vol. 8, p. 197-205, 1998. [LES 89] LESNE P. M., SAVALLE S., An efficient "cycles jump technique" for viscoplastic structure calculations involving large number of cycles , In ET AL. O., Ed., Computational Plasticity, 2nd Int. Conf. on "Computational Plasticity", 1989 , Barcelone. [NEE 87] NEEDLEMAN A., A continuum model for void nucleation by inclusion debonding , J. ofAppl. Mechanics, vol. 54, p. 525-531, 1987. [NES 00] NESNAS K., SAANOUNI K., A cycle jumping scheme for numerical integration of a damage viscoplastic model under cyclic loading , Revue Europeenne des Elements Finis, p. in press, 2000.
Multi-scale Non-linear FE2 Analysis 315 [PIL 90] PILVIN P., Approches multiechelles pour la prevision du comportement inelastique des metaux , PhD thesis, Universite Pierre et Marie Curie, Paris 6,1990. [QUI96] QUILICI S., DEBORDES O., Parallelisation des problemes devolution non lineaires (plateforme sic) , Rapport technique, 1996 , Synthese des travaux 93-95, saut technologique Nouveaux outils de conception et de modelisation en mecanique et disciplines associees. [QUI 99] QUILICI S., CAILLETAUD G., Fe simulation of macro-,meso- and micro-scales in polycrystalline plasticity , Comp. Materials Science, vol. 16, p. 383-390, 1999. [ROU 94] Roux F. X., FARHAT C., Implicit parallel processing in structural mechanics , Computer Methods in Applied Mechanics and Engineering, vol. 2, n. 1, 1994. [SAV 78] SAVALLE S., CULIE J. P., M6thodes de calcul associees aux lois de comportement cyclique et d'endommagement , La Rech. Aerospatiale,, n. 5,1978. [SMI 98] SMIT R., BREKELMANS W., MEIJER H., Prediction of the mechanical behavior of nonlinear heterogeneous systems by multi-level finite element modeling , Comput. Methods Appl. Mech. Engrg., vol. 155, p. 181-192, 1998.
Index
3D damage modelling 93 et seq 3D ductile crack growth, nonlocal simulation 209 3D modelling of necking 155 aluminium sheet materials, ductile rupture 259 et seq anisotropic damage modelling 93 et seq damage model, finite strain 241 et seq damage modelling, 3D coupling effects of damage and frictional sliding 93 et seq damage and sliding model 96 numerical ductile rupture 165 hyper-elasticity 227 sheet-metals, necking-failure criterion and 141 anisotropy, visualisation of 237 axially vibrating rod, damage identification 284 Banabic yield criterion 156 bling disk, computation of a 302 closed mesocrack lips, frictional sliding on 99 coarse grain structure, FE2 analysis 305 cohesive zone model 270 plane stress state 269
composite structures, multi-scale non-linear FE2 analysis of damage and fiber size effects 291 et seq concrete structures, transient response of 111 et seq coupling damage and friction 111 energy dissipation 111 consistent elastoplastic-damage tangent operator 199 constitutive damage models 168 continuous damage mechanics, Rousselier model 264 Cook's problem 231 coupling damage effect, friction and sliding 93 et seq, 101 3D anisotropic damage modelling 93 et seq crack growth damage and 114 ductile, 3D nonlocal simulation 209 cross-tool, deep-drawing with 159 damage analyses, nonlocal anisotropic model, finite strain 241 modelling, 3D coupling effects of damage, frictional sliding and 93 et seq numerical ductile rupture 165 et seq sliding model and 96
318
consistent elastoplastic tangent operator 199 coupling, friction and energy dissipation, transient response of concrete 111 et seq crack growth and 114 criteria, evolution laws and 115 ductile fracture and numerical modelling 21 implicit gradientenhanced formulation 19 et seq fiber size effects and multi-scale non-linear FE 2 analysis of composite structures 291 et seq gradient formulation 1, 3 et seq key ingredients 4 identification, axially vibrating rod 285 inelasticity and friction coupled to 116 isotropic ductile, metal forming for finite elastoplasticity 183et seq mechanics, continuous Rousselier model 264 mesocrack growth, unilateral behaviour and 96 model finite strain plasticity 214 Gurson-Tvergaard 130 Rousselier, the 215 simple nonlocal 85 "symmetric" isotropic local 78 nonlocal version 80 modelling, non local 63 models constitutive 168 Gerson's 168 Gologanus' 168 nonlocal 41 adaptive analysis 39
numerical aspects of 75 numerical aspects of 75 parameter identification 136 sheet metal forming prediction of necking phenomenon 147 sliding coupling and 101 uniform vibrating rod 278 damageable disk, FE2 analysis 305 deep-drawing cross-tool 159 hydraulic, fracture prediction 205 deformation large analysis 33 hyperelasto-plasticity model 23 small analysis 32 elasto-plasticity model 22 discretized form, coupled problem 5 disk, FE2 analysis bl ing, computation of a 302 damageable 305 undamageable 305 dissipative materials, standard 233 ductile crack growth, 3D nonlocal simulation of 209 et seq damage 27 isotropic, metal forming finite elastoplasticity 183 et seq numerical modelling fracture and implicit gradient-enhanced formulation 19 rupture aluminium sheet materials 259 et seq numerical, anisotropic damage applied to 165 et seq EfiCoS, finite element code 119 elasticity-based local approach, fracture to 55 et seq
Index
319
elastoplastic-damage tangent operator, consistent 199 elasto-plasticity finite isotropic ductile damage, metal forming for 183 et seq formulations 22 model, small deformation and 22 energy dissipation, transient response of concrete structures coupling damage, friction and 111 et seq mechanical 191 metric tensors 228, 230 potentials 190 Gibbsand 248 thermal 191 transient response of concrete structures coupling damage, friction and 111 et seq error estimator, the 42 evolution laws, damage criteria and 115 failure, sheet metal form, analysis 127 FE2 method 295 multi-scale non-linear analysis of composite structures damage and fiber size effects 291 principle 296 FE simulations, necking prediction and 158 FEM solution procedure, given load level 212 fiber size effects, damage and multiscale non-linear FE2 analysis of composite structures 291 et seq finite elastoplasticity element code:EfiCoS 119 implementation 66
isotropic ductile damage, metal forming for 183 et seq strain anisotropic damage model 241 et seq plasticity, damage model and 214 fracture ductile damage and numerical modelling 21 local approach to 58 elasticity-based 55 et seq prediction, hydraulic deep drawing 205 frequency response function (FRF) 279 friction coupled to damage, inelasticity and 116 coupling damage and 111 et seq frictional sliding closed mesocrack lips on 99 coupling effects of damage and 3D anisotropic damage modelling 93 et seq Gibbs, dissipation potentials and 248 given load level, FEM solution procedure 212 Gologanu's model 168 gradient damage enhanced formulation, implicit numerical modelling of ductile damage 19 formulation 1, 3 key ingredients 4 Gurson's model 168 Gurson-Tvergaard damage model 130 hydraulic deep drawing, fracture prediction 205 hyper-elasticity 227, 229 anisotropic 227 hyperelasto-plasticity model, large deformation 23
320
implicit gradient-enhanced formulation 19 numerical modelling, ductile damage of 19 inelasticity, friction coupled to damage and 116 integrated multiscale analyses 293 "interpolated-mapping" relocalization technique 301 isotropic ductile damage, metal forming for finite elastoplasticity 183 et seq isotropic "symmetric" damage model local 78 nonlocal version 80 large deformation analysis 33 hyperelasto-plasticity model 23 localisation, mesh sensitivity and 61 material moduli 251 mechanical dissipation 191 mesh sensitivity, localisation and 61 mesocrack growth 106 unilateral behaviour and, damage by 96 lips, closed frictional sliding on 99 metal forming coupled constitutive equations for 188 isotropic ductile damage 183 et seq finite elastoplasticity 183 et seq model non-locality of 45 multi-scale analyses integrated, sequential 293 non-linear FE 2 analysis, composite structures of damage, fiber size effects and 291 et seq necking 3D modelling of 155
failure criterion, anisotropic sheetmetals 141 phenomenon, prediction of sheet metal forming, damage in 147 et seq simulations and prediction of 155 FE 158 one-dimensional bar, uniaxial tension 9 parallel computing 299 periodic homogenization theory 295 perturbation technique 149 plane stress elements 270 state, cohesive zone model 269 plastic limit-load model, void coalescence criterion 139 potential, sliding criteria and 117 yield locus 152 plasticity 214 finite strain, damage model and 214 relocalization 300 technique, "interpolated-mapping" 301 Rousselier damage model 215, 264 continuous mechanics 264 rupture, ductile aluminium sheet materials 259 et seq numerical, anisotropic damage 165 et seq seismic case study 122 sequential multiscale analyses 293 shear, simple 236 sheet materials, aluminium ductile rupture 259 et seq metal forming damage in prediction of necking phenomenon 147 et seq failure in
Index
321
numerical analysis 127 et seq metals, anisotropic necking-failure criterion 141 simple nonlocal damage model 85 sliding coupling, damage and 101 et seq criteria, plastic potential and 117 frictional 108 closed mesocrack lips 99 coupling effects of damage 3D anisotropic damage modelling 93 et seq model, anisotropic damage and 96 small defects, identification of vibration tests 275 et seq standard dissipative materials 233 strain 214,241,245 finite anisotropic damage model 241 et seq plasticity, damage model and 214 stress measures and 245 stress loading, uniaxial 121 measures, strain and 245 plane elements 270 state, cohesive zone model 269
"symmetric" isotropic local damage model 78,80 nonlocal version 80 tangent operator, consistent elastoplastic-damage 199 stiffness matrix, computation of 298 thermal dissipation 191 time integration procedure 196 two-dimensional panel, uniaxial tension 14 undamageable disk 305 uniaxial stress loading 121 tension one-dimensional bar 9 two-dimensional bar 14 uniform vibrating rod, damage 278 unit cells tensile test 173 vibrating rod axially, damage identification 285 uniform, damage in 278 vibration tests, identification of small defects 275 void coalescence criterion, plastic limit-load model 139

Numerical Modelling in Damage Mechanics - Edited by Kh. Saanouni - All937572640Limit

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Numerical Modelling in Damage Mechanics - Edited by Kh. Saanouni - All937572640Limit

Uploaded by

Copyright:

Available Formats

Numerical Modelling in

This page intentionally left blank

INNOVATIVE TECHNOLOGY SERIES

edited by Khemais Saanouni

London and Sterling, VA

Khemais Saanouni GSM/LASMIS Universite de Technologie de Troyes

This page intentionally left blank

Some Aspects of a Gradient Damage Formulation

This page intentionally left blank

4 Numerical Modelling in Damage Mechanics

6 Numerical Modelling in Damage Mechanics

Aspects of a Gradient Damage Formulation

Aspects of a Gradient Damage Formulation

Figure 1. 1-D-Model Problem: Bar under Uniaxial Tension

10 Numerical Modelling in Damage Mechanics

Damage Formulation local gradient

Element Type P 1 P1P1 Expansion Expansion

Table 7. Classification of Element Formulations

Aspects of a Gradient Damage Formulation

Figure 2. Load versus Deflection

12 Numerical Modelling in Damage Mechanics

Figure 3. Damage Variable Distribution (P1)

Figure 4. Load versus Deflection for c = 100.0 (P1P1)

Aspects of a Gradient Damage Formulation

Figure 5. Damage Variable Distribution for c = 100.0 (P1P1)

14 Numerical Modelling in Damage Mechanics

Figure 8. 2-D-Model Problem: Panel with Center-Imperfection

Aspects of a Gradient Damage Formulation

Aspects of a Gradient Damage Formulation

This page intentionally left blank

On the Numerical Modelling of Ductile Damage with an Implicit Gradient-enhanced Formulation

This page intentionally left blank

Gradient-enhanced Ductile Damage