Paper#1155 Final Revision

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 25

Accurate integration scheme for von-Mises plasticity with

mixed-hardening based on exponential maps

M. Rezaiee-Pajand 1 and Cyrus Nasirai 2


1
Professor of Civil Eng. Ferdowsi University of Mashhad, IRAN
2
PhD. Student of Structural Eng. Ferdowsi University of Mashhad, IRAN

SUMMARY

Von-Mises plasticity in small strain regime is considered. The material is supposed to have linear isotropic and
kinematic hardening. A numerical scheme for integrating model equations in plasticity is proposed. The method is
based on defining an augmented stress vector and using exponential maps to solve quasi-linear differential equations
system. The solutions obtained by this new technique gives very accurate updated stress values. Furthermore, it has
quadratic convergence rate. The new method is compared with the previous works and clearly shows the good
performance of proposed scheme. A tangent operator that is consistent with the proposed formulation is presented
and its second order convergence is demonstrated by the numerical tests.

KEY WORDS: von-Mises plasticity; integration; exponential map

NOMENCLATURE

A & = AX
matrix of coefficients of ordinary differential equation X
a0 , a1 , a , a scalar parameters utilized in matrix exponentials, defined in Eq.(27),
Eq.(35), Eq.(54) and Eq.(59), respectively
b0 , b1 , b ,b scalar parameters utilized in matrix exponentials, defined in Eq.(28),
Eq.(36), Eq.(55) and Eq.(60), respectively
Enσ , E σ error of the updated stress tensor and mean value of error
e, e e , e p deviatoric strain tensor, elastic and plastic parts of the deviatoric strain
tensor
G shear modulus
Gn , Ge , G p matrix exponentials of A ( fully elastic or fully plastic load steps )
G n , Ge , G p matrix exponentials of A ( load steps with elastic and plastic parts)
H iso , H kin hardening parameters utilized in Eq.(8 ) and Eq.(11 )
H' monotonic tensorial hardening modulus
I,i identity matrix and second rank identity vector
K material bulk modulus
n direction of the deviatoric plastic strain rate
p volumetric stress
R0 , R initial and current values of radius of the yield surface
s the deviatoric stress tensor
X non-dimensional augmented deviatoric stress vector
X0 an integrating factor which is a function of γ and introduced in Eq.(16)
Xs non-dimensional deviatoric stress vector defined by Eq.(14)
α center of the yield surface in the deviatoric plane or the deviatoric part of
the backstress tensor

1
α scalar that denotes pure elastic part of a load step
β scalar parameter defined by Eq.(39)
γ& positive scalar factor of proportionality
∆ e, ∆ ê strain rate vector and its unitary direction
ε total strain tensor
ε y,0 uniaxial yield strain in tension
θ volumetric strain
Σ shifted or active deviatoric stress vector
Σ non-dimensional shifted stress vector
σ total stress tensor
σ y,0 uniaxial yield stress in tension
0 9×9 , 01×9 null matrix and null vector

1. INTRODUCTION

Nonlinear finite element analysis is an essential tool in structural analysis and design.
Nowadays, testing of prototypes is increasingly being replaced by simulation with nonlinear
finite element analysis, because this provides a more rapid, accurate, reliable and less expensive
way to evaluate design concepts and details. For example in the field of automotive design,
simulation of crashes is replacing full scale tests[1]. In order to determine a clear factor of safety
under a certain service load, finding out the ultimate load capacity of a structure is necessary.
The most important and hardest to model source of nonlinearity is caused by material
properties.
A nonlinear finite element analysis of structures is usually based on an iterative solution of
equilibrium equations. With these equations, incremental strain histories are generated. Then,
updated stresses due to given strains are obtained. Finally, the equilibrium equations are tested
for the computed stress distribution and, if violated, the iteration process continues. Updating
stresses plays a key role in this three step process due to the large consumption of computation
time. The efficiency and accuracy of the calculations of nonlinear finite element analysis is
strongly influenced by the efficiency and accuracy of stress updating schemes. Due to the
nonlinear nature of model equations in plasticity, there are difficulties in developing exact
solution methods. Updating the stresses is normally performed by integration of elastic-plastic
constitutive equations at each stress point.
One of the constitutive models that adequately predicts the behavior of many ductile
metals is based on the von-Mises yield function with linear mixed hardening. Although, this
model is one of the simplest, it is nevertheless complex enough that its exact solution is not
available. The three common procedures used to integrate this kind of constitutive equation are
particular cases of the generalized trapezoidal rule: tangent stiffness-radial corrector; mean
normal and elastic predictor-radial corrector. Despite their availability, there is still a need for
better methods to improve both efficiency and reliability.
In the past several years, further attention is directed to develop the integration techniques
based on the internal symmetries of simple constitutive models. Internal symmetry group of the
constitutive model ensures that the plastic consistency condition is exactly satisfied at each time
step if the numerical procedure can take it into account [2-5]. Auricchio and Beirao da Veiga
[6], derived a quasi-linear system X & = A X for an augmented stress vector X. Then, by
employing an exponential map, exp(A n ∆t ) , as an approximation to the above system, a new
numerical method was developed. They showed that their method has a quadratic error
convergence, while radial return shows a linear error convergence behavior. However, they

2
obtained an A which is not symmetric and hence their method does not guarantee to preserve
the consistency condition in the plastic loading regime. By defining a non dimensional stress
vector, E.Artioli et al.[7] developed a new exponential map that has symmetry and thus
automatically ensure yield consistency of the numerical technique. Unfortunately, this method
had only a linear convergence order.
In this paper a new numerical scheme based on exponential maps is proposed. Numerical
examples show that this method gives very precise results even with very large load steps with
no additional cost. The tangent operator which is used in the global finite element analysis will
also be presented. This tangent operator is consistent with the proposed numerical updating
scheme.
To simplify the descriptive and numerical part of the work, all second-order tensors are
considered as 9-component column vectors by ordering the tensor components in a vector
format. Due to the symmetry of second order tensors, the number of independent components
may be reduced to 6. Clearly, the definition of the trace operator and the Euclidean norm must
be modified.

2. BASIC MODEL

An associated von-Mises plasticity model with linear kinematic and isotropic hardening in the
small strain domain is adopted. The total strain and stress, ε and σ are decomposed into
deviatoric and volumetric components as follows

σ = s + pi with p = 13 tr(σ) (1)

ε = e + 13 θ i with θ = tr(ε) (2)

Here, tr( ) indicates the trace operator and i is the vector corresponding to the second order
identity tensor. s, p, e and θ are the deviatoric stress, the volumetric stress, the deviatoric strain
and the volumetric strain, respectively. The volumetric component is treated elastically which
means

p = K ⋅θ (3)

where K is the material bulk modulus. The deviatoric strain is decomposed into elastic and
plastic parts

e = e p + ee (4)

The elastic deviatoric part, e e ,is related to the deviatoric stress by the elastic shear modulus

s = 2G ee = 2G ( e − e p ) (5)

The shifted or active stress, Σ , is given by the following equation

Σ=s−α (6)

Here, α is the deviatoric part of the backstress and locates the center of the yield surface in the
deviatoric plane. The von-Mises yield surface is given by

3
F = Σ -R=0 (7)

where R is the radius of the yield surface. In a linear isotropic hardening mechanism, R expands
by the following rule

R = R0 + H iso .γ (8)

where R0 is initial radius of the yield surface, H iso is a material constant and γ is a scalar
quantity. γ& is a proportionality factor which defines the length of the deviatoric plastic strain
rate as follows

e& p = γ& n (9)

Direction of deviatoric plastic strain rate, n, is a vector normal to the yield surface at the contact
point. This can be expressed as

∂F Σ Σ
n= = = (10)
∂Σ Σ R

The linear kinematic hardening rule is defined by

α& = H kin e& p (11)

which is the Prager's hardening rule and assumes that the center of the yield surface moves in
the direction of the deviatoric plastic strain rate. H kin is also a material constant.
The Kuhn-Tucker loading-unloading conditions are as follows

γ& ≥ 0 ; F ≤ 0 ; γ&F = 0 (12)

The material behaves plastically if γ& > 0 and elastically when γ& = 0 .

3. AUGMENTED DIFFERENTIAL EQUATIONS SYSTEM

E.Artioli et al.[7] converted the original differential problem of associated von-Mises plasticity
with mixed linear hardening into the following nonlinear differential system of equations

& =AX
X (13)

where X is an augmented generalized stress vector with n+1 dimensions as follows

⎧⎪ X 0 Σ ⎫⎪ ⎧⎪X s ⎫⎪
X = ⎨ 0 ⎬= ⎨ 0⎬ (14)
⎪⎩ X ⎪⎭ ⎪⎩ X ⎪⎭

Σ is non-dimensional shifted stress vector defined by

4
Σ
Σ= (15)
R

and X 0 is an integrating factor which is a continuous function of γ and defined by the


following equation
⎧ 2 G + H kin + H iso
⎛ H ⎞
⎪ ⎜1 + iso γ ⎟ H iso
⎪ if H iso ≠ 0
X 0 (γ) = ⎨ ⎝ R0 ⎠ (16)
⎪ 2G + H kin
⎪ exp( γ) if H iso = 0
⎩ R0

Matrix A in differential system of equations of (13) has different definition in elastic and
plastic phases as follows

2G ⎡0 9× 9 e& 9×1 ⎤
A= ⎢ ⎥ elastic phase (17)
R ⎣ 01×9 0 ⎦

2G ⎡ 0 9× 9 e& 9×1 ⎤
A= ⎢& T ⎥ plastic phase (18)
R ⎢⎣e1×9 0 ⎥⎦

Note that if only kinematic hardening exists, R will be constant and therefore matrix A will
only depend on e& . On the other hand, in a mixed hardening model, R is not constant in the
plastic phase and matrix A will depend on X. Therefore in this case, system of equations of
(13) is not linear. The complete procedure for converting the constitutive model to the
differential equation system is presented in Appendix A1.
For a given state to be plastic, the following two conditions, which are deduced from the Kuhn-
Tucker conditions (12), must be fulfilled simultaneously :

1- Σ must be on the yield surface. Referring to Eq.(7), this means

Σ =R (19)

Using Eq.(14) and Eq.(15), the following can be easily shown

Xs = X 0 (20)

2- e& must be outward with respect to the yield surface, i.e.

Σ T e& > 0 (21)

This can be rewritten in the following form

( X s ) T e& > 0 (22)

5
4. STRESS UPDATING ALGORITHM

For calculation purposes, one may approximate the specified controlled-strain path by a
rectilinear strain path, such that e& is constant during each time step, denoted by e& n at a discrete
time t = tn . In conventional finite element analysis, the constitutive quantities at time tn which
are s n , e n , γn and α n are all known and the updated strain e n + 1 is also known at time t = tn + 1 . The
suggested numerical algorithm must integrate the plasticity rate equations over the time
increment to determine the updated stress and constitutive quantities. This means solving the
system of differential equations (13) with the following initial value

⎧ Σ0 ⎫
⎪ ⎪
X0 = ⎨ R0 ⎬ (23)
⎪1 ⎪
⎩ ⎭

The solution of the dynamical system (13) is available in power series form or in a more
compact exponential map form as follows

X n + 1 = exp(A . ∆t ) X n (24)

As was mentioned before, if H iso ≠ 0 the control matrix A is not constant in the plastic phase,
but one may assume that A is constant during each ∆t time step. The known value of A at the
beginning of each time step, A n , is the value considered throughout the time interval [6-7].
Thus, this can be expressed as

X n + 1 = exp(A n . ∆t ) X n = Gn X n (25)

Here, Gn is the matrix exponential and can be evaluated for any real matrix A n . ∆t . Defining
the strain rate vector, ∆ e = e n + 1 − e n , and its unitary direction as ∆ ê = ∆ e , Gn in the elastic
∆e
and plastic phases will have the following forms

⎧ ⎡ 2G ⎤
⎪ G = ⎢ I R ∆ e⎥ elastic phase
⎪⎪ e
⎢ n

Gn = ⎨ ⎣ 0 1 ⎦ (26)
⎪ ⎡I + (a0 − 1)∆ ê ∆ ê T b0 ∆ ê⎤
⎪ Gp = ⎢ ⎥ plastic phase
⎪⎩ ⎣ b0 ∆ ê T a0 ⎦

where I is the identity matrix and the scalars a0 and b0 are

2G
a0 = cosh( ∆e ) (27)
Rn
2G
b0 = sinh( ∆e ) (28)
Rn

6
Like many predictor-corrector integration algorithms, each step is started by computing a trial
value of the augmented stress vector assuming an elastic behavior

XTR
n + 1 = Ge X n (29)

if the trial estimate is admissible, i.e.

X ns ,+TR1 ≤ X n0,+TR1 (30)

then the variables at time tn + 1 are taken as the trial values. Note that in Eq.(30) X n0,+TR1 is equal to
X n0 since in the elastic prediction X 0 is frozen. If the trial solution is not admissible, i.e.
Eq.(30) is violated, then the step is plastic. Therefore, this step can be divided in two parts: an
elastic deformation, followed by a plastic deformation. Parameter α which denotes elastic and
plastic parts of such a step is presented in Appendix A2. Note that α ∆ e denotes a fully elastic
step and (1 − α ) ∆ e denotes a fully plastic step and α ∈ [0,1) . Thus the updated augmented
stress vector, X n+ 1 , is computed as follows

X n + α = Ge X n (31)
X n +1 = G p Xn + α (32)
or simply
X n+1 = G p G e X n (33)

Here, G e and G p are still given by Eq.(26), where ∆ e in the elastic phase is replaced by α ∆ e
and in the plastic phase by the (1 − α ) ∆ e . Rewriting Eq.(26) through Eq.(28) , the result is

⎧ ⎡ 2G ⎤
⎪ G = ⎢ I α R ∆ e⎥ elastic phase
⎪⎪ e
⎢ n

Gn = ⎨ ⎣ 0 1 ⎦ (34)
⎪ ⎡I + (a1 − 1)∆ ê ∆ ê T b1∆ ê⎤
⎪ Gp = ⎢ ⎥ plastic phase
⎪⎩ ⎣ b1∆ ê T a1 ⎦

where
a1 = cosh( g ) (35)
b1 = sinh( g ) (36)
and
2G
g= (1 − α ) ∆ e (37)
Rn

It is obvious that when α = 0 , a fully plastic step, G e = I and G p = Gp . Finally, at the end of
each step the only parameter that must be updated is R, the radius of the yield surface. Utilizing
Eq.(8) and Eq.(16), one can show that

Rn + 1 = R0 ( X n0+ 1 ) β (38)
where

7
H iso
β= (39)
2G + H kin + H iso

The updated stress and center of the yield surface can be computed as

Rn + 1 s
Σ n +1 = Xn +1 (40)
X n0+ 1
H kin
α n +1 = (2G e n + 1 − Σ n + 1 ) (41)
2G + H kin
s n +1 = Σn +1 + α n +1 (42)

σ n + 1 = s n + 1 + K ⋅ θn + 1 i (43)

It is worth mentioning that X 0 is not a history variable and does not need to be updated and
can be set to one at the start of each time step. This means that X may be scaled at the beginning
of each time step and G p can be divided by any positive scalar without changing the final result.
Scaling G p by the parameter a avoids the risk of overflow errors when time step is very
large[7].

5. CONSISTENCY OF THE NUMERICAL ALGORITHM

The above numerical technique is fully consistent with the yield surface condition. If, at t = tn
the stress state is on the yield surface, then referring to Eq.(20), it is obvious that

2
X ns = ( X n0 ) 2 (44)

Using Eq.(25) and Eq.(26), one can get stress state X s and X 0 at tn +1 as follows

X ns +1 = X ns + (a1 − 1)∆ ê ∆ ê T X ns + b1∆ ê X n0 (45)


X n0+ 1 = b1∆ ê X ns + a1 X n0 (46)

2
s
Now, X n+ 1 can be computed as
2
X ns +1 = ( X ns +1 ) T X ns +1 (47)

Utilizing Eq.(45) and noting ∆ ê = 1 , one can prove that

2 2
X ns +1 = a12 X ns + b12 ( X n0 ) 2 + 2a1b1 X n0 X ns ∆ ê (48)

On the other hand, Eq.(46) gives

2
( X n0+1 ) 2 = a12 ( X n0 ) 2 + b12 X ns + 2a1b1 X n0 X ns ∆ ê (49)

8
Comparing Eq.(48) and Eq.(49), and referring to Eq.(44), it is clear that

2
X ns +1 = ( X n0+1 ) 2 (50)

This means that the stress state at t = tn + 1 remains exactly on the yield surface.

6. IMPROVING THE NUMERICAL ALGORITHM

The evolution law presented in Eq.(25) is an explicit method, because A n is evaluated using
Rn which is the radius of the yield surface at time t = tn . It should be mentioned that G e is
constant during each time step, but G p depends on R and it will change during a time step, if
isotropic hardening exists. In order to increase the accuracy and convergence of the numerical
scheme, an improved R has to be obtained for each step. R is assumed to be constant and equal
to its value at the middle of a plastic load step

R=R 1+ α (51)
n+
2

s
To find this, first X n+ α will be computed. Recalling Eq.(47) and Eq.(50) yields

2G 0
X ns + α = X ns + α X ∆e (52)
Rn n

1+ α
Using Eq.(32), X 0 can be obtained at time t = tn + ( )∆t , i.e. at the middle of the plastic
2
load step
2G 0
X0 = b ∆ ê ( X ns + α X ∆ e) + a X n0 (53)
n +1+α Rn n
2

where a and b still have their definition of a1 and b1 in Eq.(35) and Eq.(36) but with argument
g
, that is
2
g
a = cosh( ) (54)
2
g
b = sinh( ) (55)
2

On the other hand, with some manipulation of Eq.(38), the radius of the yield surface at the
middle of the plastic load step is calculated as

β
⎛ X0 ⎞
⎜ n+1+α ⎟
R 1+ α = Rn ⎜ 0
2 ⎟ (56)
n+
2 ⎜⎜ X n ⎟⎟
⎝ ⎠
Substituting Eq.(53) in Eq.(56), leads to

9
R 1+α = Rn Z β (57)
n+
2
where
g ∆ ê T X ns g
Z = sinh( )( + α 2G ∆ e ) + cosh( ) (58)
2 Xn 0 Rn 2

Now, the evolution laws (33) and (34) can be used, but with a new definition for a1 and b1 in
the following form
g
a = cosh( β ) (59)
Z
g
b = sinh( β ) (60)
Z

where g is still defined in Eq.(37). Numerical tests presented in section 8 show that the new
method gives very accurate updated stresses with a second order accuracy.

7. CONSISTENT TANGENT OPERATOR

An elasto-plastic tangent operator is needed to develop the structure stiffness matrix in a finite
element analysis. To achieve the quadratic rate of asymptotic convergence for Newton's
technique, the tangent operator must be consistent with the numerical method employed to
integrate the plasticity rate equations. Consistency implies that the updated stress predicted by
the tangent operator matches the updated stress predicted by integration procedure to the first
order. By linearization of the discrete time procedure, the elasto-plastic tangent operator will be
∂σ
developed. This means that ( ) n+ 1 is needed. From the definition of X s in Eq.(14) and Eq.(15)
∂ε
we have
X0
X ns +1 = n +1 Σ n +1 (61)
Rn +1

Taking derivative of Eq.(61) with respect to e n +1 , yields

∂ X ns +1 X n0+1 ∂ Σ n +1 Σ n +1 ∂X n0+1 ∂ 1
= + + X n0+1 Σ n +1 ( ) (62)
∂ e n +1 Rn +1 ∂ e n +1 Rn +1 ∂ e n +1 ∂ e n +1 Rn +1

Recalling Eq.(38) and taking the derivative, one can find

∂ 1 − β ∂X n0+1
( )= 0 (63)
∂ e n +1 Rn +1 X n +1Rn +1 ∂ e n +1

From Eq.(62) and Eq.(63), the following equation is obtained

∂ Σ n +1 Rn +1 ∂ X ns +1 1 − β s ∂X n0+1
= ( − X ) (64)
∂ e n +1 X n0+1 ∂ e n +1 X n0+1 n +1 ∂ e n +1

10
s 0
The derivatives of X n+ 1 and X n +1 are presented in next section. The derivative of deviatoric
stress may be obtained from Eq.(6) as follows

∂ s n +1 ∂ Σ n +1 ∂ α n +1
= + (65)
∂ e n +1 ∂ e n +1 ∂ e n +1

and using Eq.(41) gives


∂ s n +1 2G ∂ Σ n +1 2G ⋅ H kin
= + I (66)
∂ e n +1 2G + H kin ∂ e n +1 2G + H kin

Also, the following equation holds


∂ s n +1 ∂ s n +1 ∂ e n +1 ∂ s n +1
= = I (67)
∂ ε n +1 ∂ e n +1 ∂ ε n +1 ∂ e n +1 dev

where, I dev is a projector to the deviatoric plane


1
I dev = I − i i T (68)
3

With this definition, one can replace Eq.(1) and Eq.(2) with the following,

s = I dev σ (69)
and
e = I dev ε (70)

The result of taking the derivative of Eq.(1) and using Eq.(2) and Eq.(3) is

∂ σ n+1 ∂ s n+1
= + K ⋅ i iT (71)
∂ ε n+1 ∂ ε n+1

Putting Eq.(66) in Eq.(67) and taking into account Eq.(71) the following equation is obtained

∂ σ n +1 2G ∂ Σ n +1 2G ⋅ H kin
= I dev + I + K ⋅ i iT (72)
∂ ε n +1 2G + H kin ∂ e n +1 2G + H kin dev

To completely define tangent operator, one must compute the derivative of the relative
deviatoric stress Σ with respect to deviatoric strain e at time tn +1 , that is presented in Eq.(64).
Using Eq(5) and Eq.(70), it is obvious that in elastic phase Eq.(71) simplifies to

∂ σ n+1
= 2G I dev + K ⋅ i i T (73)
∂ ε n+1

7.1. Derivatives of X ns +1 and X n0+1

It is intended to compute the derivatives of X ns +1 and X n0+1 with respect to e n +1 which are
appeared in Eq.(64). Recalling Eq.(33), the following equation holds

11
⎧⎪X s ⎫⎪ ⎧⎪X s ⎫⎪
⎨ 0⎬ = G G
p e ⎨ 0⎬ (74)
⎪⎩ X ⎪⎭n +1 ⎪⎩ X ⎪⎭n

where G p and Ge are defined in (34), and it can be shown that

2G 0
X ns +1 = X ns + (a − 1)∆ ê ∆ eˆ T X ns + a ⋅ α X ∆ e+ b ⋅ X n0 ∆ ê (75)
Rn n
2G 0
X n0+ 1 = b∆ êX ns + b ⋅ α X ∆ e + a ⋅ X n0 (76)
Rn n

Now, the derivative of Eq.(75) and Eq.(76) can be computed. Note that α in these relationships,
is a function of ∆e . Thus the derivative can be expressed as

∂ X ns +1 ∂ X ns +1 ∂ X ns +1 T
= + v (77)
∂ e n +1 ∂∆ e ∂α
∂X n0+1 ∂X n0+1 ∂X n0+1
= + ⋅v (78)
∂ e n +1 ∂∆ e ∂α

Here, v is the derivation of α with respect to the deviatoric strain increment and are presented
in Appendix A.2. To obtain more compact equations, new vectors and scalars are introduced
and displayed in Appendix A3. Using these new definitions, the derivatives of the generalized
stress vector with respect to ∆e and α are as follows

∂ X ns +1 (a − 1) b ⋅ X n0 (a − 1) (a − 1)
= (q X ns − s− )(∆ ê ∆ ê T ) + sI+ ∆ ê ( X ns ) T +
∂∆ e ∆e ∆e ∆e ∆e
(79)
b ⋅ X n0 2G 0
I+α X (q ∆eˆ T + a I) + X n0 r ∆ ê T
∆e Rn n
∂ X ns +1
= u∆ ê ∆ êT X ns + [(a + u ⋅ α )k + w] X n0 ∆ ê (80)
∂α
∂X n0+1 b b 2G
= s r+ Xs − ∆ ê ∆ ê T X ns + X n0 (α ⋅ k r + q + b ⋅ α ∆ ê) (81)
∂∆ e ∆e n ∆e Rn
∂X n0+1
= s ⋅ w + X n0 (α ⋅ k ⋅ w + b ⋅ k + u ) (82)
∂α

Once the derivatives are computed, they can be substituted in Eq.(77) and Eq.(78). It is clear
that when a step is fully plastic, i.e. α = 0 , these equations are changed to

∂ X ns +1 ∂ X ns +1
= (83)
∂ e n +1 ∂∆ e
∂X n0+1 ∂X n0+1
= (84)
∂ e n +1 ∂∆ e

Also, Eq.(79) and Eq.(81) can be written as follows

12
∂ X ns +1 ( a − 1) b ⋅ X n0 (a − 1)
= (q X ns − s− )(∆ ê ∆ ê T ) + sI
∂∆ e ∆e ∆e ∆e
(85)
(a − 1) b ⋅ X n0
+ ∆ ê ( X ns ) T + I + X n0 r ∆ ê T
∆e ∆e

∂X n0+1 b b
= s r+ X ns − ∆ ê ∆ ê T X ns + X n0 q (86)
∂∆ e ∆e ∆e

8. NUMERICAL TESTS

Since an elastic-plastic constitutive relationship depends on deformation history, an incremental


analysis following an actual variation of external forces should be used to trace the variation of
displacement, strain and stress along with the external forces. In an incremental analysis, the
total load acting on a structure is added in gradual increments. In every load step the equilibrium
of the external force and equivalent force of stress acting on nodal points (Figure 1) must be
satisfied. The following two algorithms are therefore involved in solving nonlinear equilibrium
equation:

1. The algorithm which is used for solving nonlinear simultaneous equations. Most of the
algorithms which exists for solving nonlinear simultaneous equations are Newton like
and need tangential stiffness matrix of structure for a rapid convergence. To establish
∂σ
stiffness matrix of a structure we need at every sampling point of elements.
∂ε

2. The algorithm which is used for determining the stress increment, ∆ σ , corresponding to
a strain increment, ∆ ε , for a given stress state and deformation history. This must be
done at every sampling point of elements for every load step and corrective iterations.

Figure 1. Finite elements model, an element and the sampling point

∂σ
As mentioned in section 7, the tangent operator, , which is used in first algorithm must be
∂ε
consistent with the integration of second algorithm. In this work we focused on the second
algorithm. Therefore, the numerical presentations are all point wise. In the first example we
adopt two different nonproportional deformation histories for a sampling stress point. Then, the
corresponding stress states and the error of the numerical integration scheme are presented. In
each case the new scheme is compared with the study conducted by E.Artioli et al. (ESC)[7]. In

13
the second and third examples it is intended to test presented formulations of tangent operator
that is consistent with the new numerical integration. These examples are point wise as well.

8.1. Point wise stress-strain test

Two biaxial non-proportional strain paths are considered. For the sake of comparison, the data
are selected from [7]. The strain components histories are represented graphically in Figures 2
and 3. All the other strain and stress components are identically equal to zero. Two sets of
material parameters are considered, as follows

• Material 1 : E=100Mpa ν =0.3 H kin =10Mpa H iso =10Mpa R0 =15Mpa


• Material 2 : E=7000Mpa ν =0.3 H kin =0 H iso =225Mpa R0 =24.3Mpa

Strain at the first yield in a uniaxial test, which is used in defining the strain path is given by

3 R0
ε y0 = (87)
2 E

Now, the discussion is about why linear strain histories are chosen. Each integration method
operates under the restriction of a constant strain rate vector. For a finite size load step, this
leads to a linear approximation of a curved strain path. The use of this approximation introduces
discretization errors. These errors are in addition to the ones from the integration of the
plasticity rate equations. To eradicate these discretization errors, here linear strain paths are
chosen [8]. As there is no exact solution for the problem, one can compute the exact solutions
using very fine load steps, i.e. ∆t = 1× 10−5 sec. Such exact solutions are compared with the
corresponding numerical results, with more practical step sizes ∆t = 0.1,0.05,0.025 sec. The
error for the updated stress is defined in the following form

σ n − σ̂ n
Enσ = (88)
Rn

Figure 2. Strain components history Figure 3. Strain components history


(Path 1) (Path 2)

14
where σ̂ n is the exact stress at time tn . In order to better investigate the performance of the new
proposed scheme, a mean value of error is introduced, as


N
σ n =1
Enσ ∆t
E = (89)
N ⋅ ∆t

Figure 4. Stress relative error of Path 1 Figure 5. Stress relative error of Path 2
( Material 1 - ∆t = 0.1 ) ( Material 1 - ∆t = 0.1 )

Figure 6. Stress relative error of Path 1 Figure 7. Stress relative error of Path 2
( Material 2 - ∆t = 0.1 ) ( Material 2 - ∆t = 0.1 )

Figures 4 through 7 show the accuracy of presented scheme in comparison with the previous
study [7] for different strain paths and material parameters with the time step ∆t = 0.1 . It shows
that the integrating scheme easily reduce the error with one order of magnitude. By adopting
this scheme the size of the time step in the process of stress updating could be increased. This
significantly reduces the computation time in the global FEA. Figures 8 through 11 show the
reduction of error due to the smaller time step. In this figures, the first few time steps display no
error due to the elastic behavior of the materials. In Table I, the mean values of error, E σ , and

15
its decrease rates are presented. Time steps are increased by an order of two and as a result, the
error is roughly reduced by an order of four, i.e. a quadratic convergence rate.

Figure 8. Stress relative error of Path 1 Figure 9. Stress relative error of Path 2
Material 1 ( present scheme ) Material 1 ( present scheme )

Figure 10. Stress relative error of Path 1 Figure 11. Stress relative error of Path 2
Material 2 ( present scheme ) Material 2 ( present scheme )

Table I. Mean value of error and corresponding decrease rate

Path 1, Material 1 Path 2, Material 1 Path 1, Material 2 Path 2, Material 2


No. of σ σ σ
E decrease E decrease E decrease Eσ decrease
sub steps
× 10 5 rate × 10 5 rate × 10 5 rate × 10 5 ratio
1 10.8431 - 10.2818 - 5.3995 - 2.7367 -
2 2.6845 4.04 2.6224 3.92 2.2011 2.45 0.7003 3.91
4 1.0564 2.54 0.6604 3.97 0.3338 6.59 0.1776 3.94
8 0.1669 6.33 0.1662 3.97 0.0817 4.09 0.0447 3.97
16 0.0417 4.00 0.0415 4.00 0.0189 4.32 0.0112 3.99
32 0.0104 4.01 0.0104 3.99 0.0041 4.61 0.0028 4.00
64 0.0026 4.00 0.0026 4.00 0.0013 3.15 0.0007 4.00

16
8.2. Point wise stress-controlled path

In this example, the consistent tangent operator and second order convergence are tested. To do
so, we choose a default loading path on a sampling point. Then we try to get the resulting strain
path. This may be done through the Newton-Raphson method. It is probably the most rapidly
convergent process for the solution of a non-linear problem. Indeed, it is the only process in
which convergence is quadratic [10]. Suppose, a prescribed stress path σ is existed and one
wants to find its related strain path. This means that in Figure 12, σ n , ε n and σ n + 1 are all known
and ε n+ 1 must be determined. An iterative process based on the first order Taylor's expansion
is given by

−1
⎡⎛ ∂ σ ⎞ i ⎤
ε in++11 ≅ ε in +1 + ⎢⎜
i
⎟ ⎥ (σ n + 1 − σ n +1 ) (90)
⎢⎣⎝ ∂ ε ⎠ n +1 ⎥⎦
−1
⎡⎛ ∂ σ ⎞ i ⎤
δ ε in ≅ ⎢⎜ i
⎟ ⎥ (σ n + 1 − σ n + 1 ) (91)
⎢⎣⎝ ∂ ε ⎠ n +1 ⎥⎦

Using above relationships in a successive manner, the estimation of strain tensor at the (i+1)th.
iteration will be as follows
i
ε in++11 ≅ ε n + ∑δε
k =1
k
n (92)

Figure 12 . The Newton-Raphson method

Consider a sampling point in a structure with only two stress components of σ xx and τ xy acting
on it. The time history of these stress components and the corresponding loading paths are
presented in Figure 13 and Figure 14, respectively. The material is assumed to have the
following properties

E=2000Mpa ν =0.3 H kin =600Mpa H iso =100Mpa R0 =15Mpa

17
Figure 15 and Figure 16 display the stress-strain relationship obtained by proposed scheme and
Figure 17 demonstrate the resulting strain path. To examine the rate of convergence, the relative
Euclidean norms in successive iterations for t = 7 sec. and t = 8 sec. are presented in Table II.

Figure 13. Loading history Figure 14. Loading path

Figure 15. σ xx - ε xx diagram Figure 16. τ xy - ε xy diagram

Table II. Typical convergence values. Relative


Euclidean norms

Iteration t = 7 sec. t = 8 sec.


1 1.0000e+00 1.0000e+00
2 0.2595e+00 0.1525e+00
3 0.1613e-01 0.1050e-01
4 0.3431e-03 0.1345e-03
5 0.1430e-07 0.3782e-08

Figure 17. Strain path

18
8.3. Stress-controlled loading with an exact solution

A thin tube is first stressed in tension to the point of yielding and then twisted under constant
axial stress. At a point far enough from the two ends of the tube, the problem can be modeled as
a single material point (Figure 18). Figure 19 show the material properties and the loading path.

Figure 18. The tube and the stress state at the material point

(a) (b)

Figure 19 . (a) Uniaxial stress-strain curve and (b) The loading path

For isotropic hardening, Hill [9] gives the exact solutions of axial and engineering shear strains
as
σ y,0 3τ 2 σ y,0
ε= ln(1 + )+ (93)
2 H' σ y,0
2 E

3 σ y,0 3τ τ
γ= [τ − tan −1 ( )] + (94)
H' 3 σ y,0 G

where σ y,0 is the uniaxial yield stress in tension and H' is the uniaxial plastic modulus which by
decomposition of strains can easily be calculated as

19
E ⋅ ET
H' = (95)
E - ET

The monotonic tensorial hardening modulus can be obtained by

2
H iso = H' (96)
3

The strain path is obtained numerically using the Newton-Raphson algorithm, which is
presented in Example 8.2. Figure 20 shows the exact strain path, which is a graphical
representation of Eq.(93) and Eq.(94) with a solid line. The strain path obtained by the
numerical scheme ( ∆t = 1 sec ) is also demonstrated in this figure by hollow markers. The
significant curvature of the strain path in low values of shear strain, makes this problem a
difficult test case for numerical methods operate under the restriction of a constant strain rate
vector. Despite the complexity of this example, the numerical scheme shows a great agreement
with the exact solution. To complete the discussion, the exact and numerical path with different
load step sizes are given in Table III.

Figure 20 . Exact and numerical strain path ( ∆t = 1 sec )

Table III. Exact and Numerical strain components

Numerical strains [%]


Time τ Exact strains [%]
∆t = 2 sec ∆t = 1 sec ∆t = 0.5 sec
[sec] [Mpa]
ε γ ε γ ε γ ε γ
50 0 0.0900 0.0000 0.0900 0.0000 0.0900 0.0000 0.0900 0.0000
60 32 0.4936 0.1838 0.4940 0.1810 0.4937 0.1827 0.4936 0.1834
70 64 1.5225 1.0686 1.5282 1.0473 1.5245 1.0609 1.5231 1.0663
80 96 2.8387 2.8714 2.8549 2.8271 2.8444 2.8556 2.8404 2.8666
90 128 4.2023 5.4573 4.2305 5.3945 4.2122 5.4351 4.2053 5.4506
100 160 5.5029 8.6163 5.5388 8.5444 5.5140 8.5928 5.5070 8.6083

20
9. CONCLUSIONS

A new accurate numerical integration scheme based on exponential maps for plasticity
computations has been proposed. The widely used von-Mises yield function with mixed
isotropic-kinematic hardening have provided the constitutive relationships in this study.
Previous studies have shown effectiveness of exponential based methods with respect to
traditional radial-return map method. The proposed technique is developed to furtherer
improve the efficiency of these techniques. The significant characteristics of the proposed
method are:

• The proposed scheme easily reduce the error with one order of magnitude compare to
previous works. This would enable us to enlarge the time step in the process of stress
updating and significantly reduces the computation time in the global FEA.

• The new procedure is a fully consistent formulation and has a quadratic convergence
rate.

The consistent tangent operator which could be used to construct the stiffness matrix of a
structure and eventually guarantees the quadratic rate of convergence for solving nonlinear
simultaneous equations in a finite elements analysis, is also developed.

APPENDICES

A1. Converting the constitutive model to the differential equation system

In this section, it is intended to show how the selected constitutive model can be converted to
the following nonlinear differential system of equations

& =AX
X (A1)

Utilizing Eq.(4) and Eq.(5) leads to

Σ + α + 2G e p = 2G e (A2)

Taking the derivative in time and using Eq.(9) and Eq.(11) gives

& = 2G e& − ( 2G + H ) e& p


Σ (A3)
kin

Inserting Eq.(9) and Eq.(10) into Eq.(A3), the following differential equation is obtained

& + ( 2G + H kin ) Σ γ& = 2G e&


Σ (A4)
R

The above equation is valid for both elastic ( γ& = 0 ) and plastic phase. Now, a non-dimensional
shifted stress Σ can be defined as

21
Σ
Σ= (A5)
R

Taking the derivative of Eq.(A5) and recalling Eq.(8), the following relationship can be
obtained

& H
&=Σ
Σ − iso γ& Σ (A6)
R R

Substituting Σ& from Eq.(A4) into Eq.(A6) leads to

& + 2G + H kin + H iso γ& Σ = 2G e&


Σ (A7)
R R

Now, an integrating factor, X 0 , is introduced such that

d 2G 0
( X 0 Σ) = X e& (A8)
dt R

After taking the derivative the following equation is obtained


& + X& Σ = 2G e&
0
Σ (A9)
X0 R

The following equation is achieved by comparing Eq.(A7) and Eq.(A9)

X& 0 2G + H kin + H iso


= γ& (A10)
X0 R

Substituting Eq.(8) into Eq.(A10), the following ordinary differential equation is created

X& 0 2G + H kin + H iso


= . γ& (A11)
X0 R0 + H iso .γ

Solving above equation with the initial condition of X 0 (0) = 1 , leads to

⎧ 2 G + H kin + H iso

⎪ ⎛ H iso ⎞ H iso
⎪ ⎜1 + γ⎟ if H iso ≠ 0
0
X (γ) = ⎨ ⎝ R0 ⎠ (A12)
⎪ 2G + H kin
⎪ exp( γ) if H iso = 0
⎩ R0

Now, a new augmented stress vector X is defined as

⎧⎪ X 0 Σ ⎫⎪ ⎧⎪X s ⎫⎪
X = ⎨ 0 ⎬= ⎨ 0⎬ (A13)
⎪⎩ X ⎪⎭ ⎪⎩ X ⎪⎭

It is obvious that Eq.(A8) can be rewritten to reveal the evolution of Xs as follows

22
& s = 2G X 0 e&
X (A14)
R

Eq.(A10) in the elastic phase ( γ& = 0 ) shows that

X& 0 = 0 elastic phase (A15)

On the other hand in plastic phase X 0 is not a constant. Taking the scalar product of Eq.(A9)
with Σ , leads to

1 d 2 X& 0 2 2G
Σ + 0 Σ = e& Σ (A16)
2 dt X R

Since in plastic phase Σ = 1 , the following equation can be driven

X& 0 = 2G e& T X s plastic phase (A17)


R

Eq.(A14), Eq.(A15) and Eq.(A17) can be written in the following compact form

& =AX
X (A18)

where A is a control matrix defined as

2G ⎡0 9×9 e& 9×1 ⎤


A= elastic phase (A19)
R ⎢⎣ 01×9 0 ⎦

2G ⎡ 0 9×9 e& 9×1 ⎤


A= ⎢e& T plastic phase (A20)
R ⎢⎣ 1×9 0 ⎥⎥⎦

A2. Parameter α and its derivative with respect to ∆ e

It is intended to introduce a scalar parameter α that divide a time step ∆t into an elastic time
step, α∆t , and a plastic time step, (1 - α )∆t . Suppose that at time t = tn , a material behaves
elastically and a stress point is inside the yield surface and at the end of that time step, t = tn +1 ,
the stress point is outside the yield surface. By a geometric consideration, the elastic part of time
step, α∆t , can be found and subsequently scalar α can be derived as

C 2 − DM − C
α= (A21)
D
where

23
2GX n0 s
C= Xn ∆ e
Rn
2
⎛ 2GX n0 ∆ e ⎞
D=⎜ ⎟ (A22)
⎜ Rn ⎟
⎝ ⎠
2
M = X ns − (X n0 )2

The derivative of scalar α with respect to the deviatoric strain increment which is needed in
formulation of tangent operator is as follows

dα dC dD
v= = φ1 + φ2 (A23)
d∆ e d∆ e d∆ e
where
1 C
φ1 = ( − 1)
D C 2 − DM
1 DM
φ2 = − 2 ( + C 2 − DM − C )
D 2 C − DM
2
(A24)
dC 2G o s
= X X
d∆ e Rn n n
dD 2G o 2
= 2( X ) ∆e
d∆ e Rn n

A3. List of parameters used in the derivatives of the generalized stress vector

s = ∆ ê X ns
2G
k= ∆e
Rn
g = k (1 − α )
g s g
Z = sinh( )( 0 + α ⋅ k ) + cosh( )
2 Xn 2
∂g 2G
g= = (1 − α )∆ ê
∂∆ e Rn
∂Z 1 s g ⎡ 1 Xs 2G 1 ⎤ g
= g ( 0 + α ⋅ k )cosh( ) + ⎢ (I − ∆ ê ∆ ê T ) n0 + α ∆ ê + g ⎥sinh( )
∂∆ e 2 X n 2 ⎣ ∆e Xn Rn 2 ⎦ 2
∂Z k s g k g
= − ( 0 + αk )cosh( ) + sinh( )
∂α 2 Xn 2 2 2
∂a b ∂Z
q= = β +1 ( Z g − β ⋅ g )
∂∆ e Z ∂∆ e
∂b a ∂Z
r= = (Z g− β ⋅ g )
∂∆ e Z β +1 ∂∆ e
∂a b ∂Z
u= = (−k ⋅ Z − β ⋅ g )
∂α Z β +1 ∂α
∂b a ∂Z
w= = (−k ⋅ Z − β ⋅ g )
∂α Z β +1 ∂α

24
REFERENCES

1. Belytschko T., Wing Kam Liu, Moran B. Nonlinear Finite Elements for Continua and Structures, John-Wiley &
Sons, Ltd, 2000.
2. Hong-Ki Hong and Chein-Shan Liu. " Internal symmetry in the constitutive model of perfect elastoplasticity ",
International Journal of Non-Linear Mechanics 2000; 35:447-466.
3. Hong-Ki Hong and Chein-Shan Liu. " Internal symmetry in bilinear elastoplasticity ". International Journal of
Non-Linear Mechanics 1999; 34:279-288.
4. Chein-Shan Liu. " Symmetry groups and the pseudo-Riemann spacetimes for mixed-hardening elastoplasticity ".
International Journal of Solids and Structures 2003; 40:251-269.
5. Chein-Shan Liu. " Internal symmetry groups for the Drucker-Prager material model of plasticity and numerical
integrating methods ". International Journal of Solids and Structures 2004; 41:3771-3791.
6. Auricchio F and Beirao da Veiga L. " On a new integration scheme for von-Mises plasticity with linear
hardening ". International Journal for Numerical Methods in Engineering 2003; 56:1375-1396.
7. Artioli E., Auricchio F and Beirao da Veiga L. " Integration schemes for von-Mises plasticity models based on
exponential maps: numerical investigations and theoretical considerations ". International Journal for Numerical
Methods in Engineering 2005; 64:1133-1165.
8. Dodds R.H. " Numerical techniques for plasticity computations in finite element analysis ". Computers and
Structures 1987; Vol.26 No.5, 767-779
9. Hill R. The Mathematical Theory of Plasticity, Oxford University Press, London, 1950.
10. Zienkiewicz OC, Taylor RL. The Finite Element Mathod (5th edn.) , vol II. Mcgraw-Hill: New York, 2002.

25

You might also like