Analytical DSA For Explicit Dynamics of

You might also like

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

Comput. Mech.

(2007) 39: 761–785


DOI 10.1007/s00466-006-0068-3

O R I G I NA L PA P E R

K. Wisniewski · P. Kowalczyk · E. Turska

Analytical DSA for explicit dynamics of elastic-plastic shells

Received: 2 August 2005 / Accepted: 13 March 2006 / Published online: 28 June 2006
© Springer-Verlag 2006

Abstract The paper presents an analytical constitutive of degrees of freedom (dofs), and that they are well suited
design sensitivity analysis (DSA) algorithm for explicit for parallelization, see e.g. [3, 9]. They are only conditionally
dynamics of elastic-plastic finite rotation shells. Two explicit stable, and need very small time steps repeated millions of
dynamical algorithms for finite rotation shells are presented, times, therefore computations for a single time step must be
and the DSA is developed for the one formulated  in terms very efficient.
of the rotation vector and its time derivatives, ψ, ψ̇, ψ̈ . Explicit algorithms can be applied to problems of large
The hypo-elastic constitutive model based on the Green– size and practical importance, such as metal-forming and
McInnis–Naghdi stress rate is used to derive an incremental crashworthiness of cars and ships, in which dynamic loads
algorithm in terms of ‘back-rotated’ objects. The associative and complicated material laws are involved. Such applica-
deviatoric Huber–Mises plasticity modified by plane stress tions require also the design sensitivity analysis (DSA), the
conditions is implemented in the form suitable for finite rota- purpose of which is to provide derivatives of a selected struc-
tion/small elastic strain increments. The analytical DSA is tural response w.r.t. the sizing, material, and shape parame-
developed for the above-specified problem, with the design ters. The DSA module added to structural finite element (FE)
derivatives calculated w.r.t. material parameters. Design- codes can be used in the parametric ‘what-if’ studies and gra-
differentiation of the dynamic algorithm and the scheme of dient-based optimization, which can speed up and improve
handling the history data and the predicted values in differ- the quality of design cycles.
entiation, which is crucial in computing correct derivatives, Note that the basis for the DSA for the history-dependent
are described. Besides, we show how to avoid Newton loops constitutive models already exists, see the review in [26].
in the DSA algorithm, when such a loop is present in the con- The most popular constitutive model is the deviatoric Hu-
stitutive algorithm. Numerical examples show that, despite a ber-Mises plasticity, which is applicable to metals subjected
great complexity of the solution algorithm for the finite-rota- to cyclic loading, and for modelling such phenomena as the
tion elastic-plastic shells, it is feasible to compute analytical Bauschinger effect, shakedown, ratchetting, and relaxation
design derivatives of very good accuracy. of mean stress. In [18], the bibliography until 1997 is given,
see also [16, 20]. These works are concerned with the DSA
Keywords Explicit dynamics · Finite rotation shell · for 3D elastic-plastic materials and statics, but the approach
Elastic-plastic material · Analytical Design Sensitivity presented in there can be naturally extended to the DSA for
Analysis for constitutive parameters implicit dynamics.
However, there are only few publications on the DSA for
explicit dynamics, among them [5, 6, 17], and none of them
is devoted to shells. In the conference note [25] there are a
1 Introduction few remarks on shells. Two issues are characteristic for the
DSA for explicit dynamics.
Explicit algorithms are the most frequently used algorithms
in numerical simulations of nonlinear dynamic response of 1. The first is that the DSA module is implemented on top
structures. Their basic merit is that they do not use the tangent of the dynamic algorithm, which has a specific incremen-
stiffness matrix, so the storage is proportional to the number tal character, involving predictors and updates of the ele-
ment history database. The updates are needed not only by
K. Wisniewski (B) · P. Kowalczyk · E. Turska the dynamic algorithm, but also by the history-dependent
IFTR, Polish Academy of Sciences, constitutive equation. In consequence, the calculated de-
Swietokrzyska 21, 00-049 Warsaw, Poland sign derivatives depend on the constitutive state parame-
E-mail: kwisn@ippt.gov.pl ters and their design derivatives for the previous time step.
762 K. Wisniewski et al.

This makes the design-differentiation a complicated mat- In Sect. 4, the analytical DSA is developed for the
ter, certainly not automatic. Typically, the DSA module above-specified problem, with the design derivatives calcu-
is implemented for already existing architecture of the lated w.r.t. seven constitutive parameters of the elastic-plastic
code and the storage scheme, and a good understanding material law. All equations of the problem are differentiated
of the primary algorithm is a pre-requisite to a successful with respect to design parameters. In particular, the constit-
implementation of the DSA module. utive algorithm is differentiated, and we show how to avoid
2. The second is a relatively high cost of the analytical DSA Newton loops in the DSA algorithm, although such a loop is
for explicit dynamics, comparing to the finite difference a part of the constitutive algorithm because of the ellipsoidal
DSA, although certainly the analytical method is more yield surface and nonlinear hardening. For this purpose, we
reliable, as not prone to errors due to the design perturba- use a design-differentiated yield condition, from which the
tion. This is in contrast with statics and implicit dynamics design derivative of the consistency parameter is determined.
for which the DSA is a small fraction of the total com- This, however, is not a trivial task, because of the history data
putation time, and that’s why efficiency of the analytical and the predicted values, which affect the differentiation. The
DSA computations becomes an important issue. This is schemes of handling the history data and the predicted values
the main motivation for avoiding Newton loops in the in differentiation are crucial in computing correct derivatives,
DSA algorithm, though such a loop is used by the constit- and they are given explicitly for the design derivatives of the
utive algorithm, and it causes that we have to reformulate constitutive state variables, the stress, and the yield function.
the DSA problem. This is not a trivial task, because of In Sect. 5, the numerical examples, the main purpose of
the history data and the predicted values which affect the which is to assess accuracy of the design derivatives, are
differentiation. described. First, the results of the dynamic analyses (displace-
ment and rotation components) are presented for selected
We believe that the above issues are very important and ad- points as functions of time, and compared to the results of a
dress them in the current paper. commercial code (ABAQUS Explicit [1]). Then, the design
derivatives are computed analytically and compared to the
results obtained by the finite-difference method, all shown as
Scope of the paper The objective of this paper is to describe
functions of time. The final remarks are presented in Sect. 6.
the theoretical background and algorithmic issues, and finally
assess the accuracy of the analytical constitutive DSA for
explicit dynamics of elastic-plastic shells.
2 Dynamics of shells
In Sect. 2, the basic definitions of 3D dynamics are pre-
sented, and the equations of motion are derived for shells. The
2.1 Basic definitions for dynamics
variational background of the shell element with 6 dofs/node
is given; it has the Reissner kinematics, and is valid for arbi- Velocity and angular velocity in material description
trary finite rotations. The parameterization of rotations based Consider a time-dependent deformation function χ(t), which
on the canonical rotation vector and the related transforma- maps the reference configuration of the shell-like body B
tion operators, including these for time derivatives, are spec- onto the current configuration Bχ . If x and X denote the spa-
ified explicitly. As the last part of this section, two explicit tial and material coordinate of a particle, then χ : x(t) =
dynamics algorithms for finite rotation shells are presented, χ(X, t). The Lagrangian (material) velocity and acceleration
with the rotational dofs written in terms of either ψ, ψ̇, ψ̈ are defined as,
or {ψ, ω, ω̇}; the first one of them is used as the primary . .
algorithm for the DSA. v(X, t) = χ̇(X, t), a(X, t) = v̇(X, t) = χ̈(X, t), (1)
˙ .
In Sect. 3, a plasticity model for finite rotation shells is where the time derivatives are denoted as (·) = d(·)/dt and
¨ = . 2
described. For the finite rotation/small elastic strain case, the (·) d (·)/dt 2 . The material (or right) angular velocity ten-
multiplicative elastic/plastic split of F implies the additive sor and its axial vector are defined as follows,
split of D. This split is used to formulate the rotationally .
ω̃(X, t) = RT (X, t) Ṙ(X, t) ∈ so(3),
objective constitutive equation for finite-deformation rate .
independent deviatoric plasticity, using the Green–McInnis– ω(X, t) = 21 [I × ω̃(X, t)], (2)
Naghdi stress rate. Assuming d as the strain rate measure, the where R ∈ SO(3) is the rotation tensor. The angular accel-
.
constitutive equation is transformed to an incremental form, eration vector is defined as aa (X, t) = ω̇(X, t).
the back-rotation of which yields a form analogous to that
for the small deformation case. A crucial role in the con- Hamilton’s principle The Hamilton’s principle can be writ-
stitutive algorithm plays a mid-point configuration, see [8]. ten as,
The deviatoric Huber–Mises plasticity with nonlinear iso- t2
tropic/linear kinematic hardening is modified by the plane .
δ L dt = 0, L = −W + K + Fext , (3)
stress assumptions, and parameterized in terms of in-plane
stress components, following [14, 24]. In consequence, the t1
yield surface is ellipsoidal, and the Newton method is used where L is the Lagrangian potential, and W , K denote the
to determine the consistency parameter. strain energy and the kinetic energy, respectively, and Fext
Analytical DSA for explicit dynamics of elastic-plastic shells 763

is the potential of external loads. Time-boundary conditions where γ ∈ (0, ∞) is the regularization parameter. The virtual
for the variations δx0 (t) = 0 and δψ(t) = 0 at t = t1 and work (VW) of the nominal stress P can be expressed as
t = t2 imply δL = −δ W + δ K + δ Fext = 0. P · δF = 21 S · δC, (10)
For shells, on use of the kinematical hypothesis, Eq. (12), .
the mid-surface terms and the ζ -dependent terms are sepa- where S is the 2nd Piola–Kirchhoff stress tensor, and C =
rated, and we can define the shell-type functionals, FT F is the right Cauchy–Green deformation tensor. The strain
deduced from Eq. (10), as the work conjugate to S, is the
+ 2
h .
Green strain, E = (1/2)(C−I). Assume that the strain energy
.
(·)sh = (·) μdζ, (4) density per unit non-deformed volume, W , is a function of
C, so it satisfies the objectivity requirement. The variation of
− h2
the strain energy is δ W (C) = ∂C W (C) · δC, and if 21 S · δC
where (·) concerns a 3D body, and (·)sh concerns a shell. is identified as δ W , then we obtain the constitutive law,
.
Besides, μ = det Z, where Z is the shifter tensor. In this way S = 2 ∂C W (C). (11)
we define the shell-type counterparts of L , W , K and Fext ,
for which the Lagrange potential is Reissner hypothesis The initial (reference) configuration of
.
L sh = −Wsh + K sh + (Fext )sh . (5) the shell is parameterized in terms of ξ = {ξ α , 2ζ / h} , α =
1, 2, where ξ α ∈ [−1, +1] are the natural coordinates param-
eterizing the reference surface, and ζ ∈ [−h/2, +h/2] is
2.2 Dynamic equations of motion for shells used in the direction normal to this surface; h is the initial
shell thickness. The position vector of an arbitrary point of
Extended configuration space The classical configuration
. a shell in the initial configuration is given by X (ξ α , ζ ) =
space of the non-polar Cauchy continuum is defined as: C = X0 (ξ α ) + ζ t3 (ξ α ). In the current configuration, the position
{χ : B → R }, where χ is the deformation function de-
3
vector is expressed by the Reissner kinematical hypothesis,
fined on the reference configuration of the body B. In the        
present work, we consider an extended configuration space, x ξ α , ζ, t = x0 ξ α , t + ζ Q0 ξ α , t t3 ξ α , (12)
defined in terms of the deformation function χ and rotations where x0 is a position of the reference (middle) surface in the
Q ∈ SO(3). The rotations can be treated in two different current configuration, t3 is a shell director, and Q0 ∈ SO(3)
ways: (1) remain unconstrained, as in the Cosserat-type con- is a rotation tensor, constant over ζ . Differentiating Eq. (12)
tinuum, or (2) be constrained, either by the polar decompo- w.r.t. time, we obtain a rate form of the Reissner hypothesis
       
sition of F equation, or the rotation constraint (RC) equation v ξ α , ζ, t = v0 ξ α , t + ζ Q̇0 ξ α , t t3 ξ α , (13)
  . .
skew QT F = 0, (6) where v = ẋ and v0 = ẋ0 are the translational velocities, and
. Q̇0 t3 = Q0 (ω × t3 ) by Eq. (2). Equation (13) will be used to
where F = ∇χ. Then, the extended configuration space is
defined as follows: derive the shell kinetic energy.
.
Cext = {(χ, Q) : B → R 3 × SO(3) | χ ∈ C }, (7)
Approximation of rotations Because a two-field F̃2 (χ, Q)
where C is the classical configuration space. Note that χ is of Eq. (9) is used, we also have to approximate rotations
required to belong to C , i.e. is identical as for the classical over ζ . We assume that the rotations are constant over ζ , i.e.
non-polar Cauchy continuum. This approach is used in the Q(ζ ) ≈ Q0 , where Q0 is the rotation at the shell reference
present work. surface. Consider the components of the RC equation, Eq. (6),
in the local ortho-normal basis {ti }, see Fig. 1. The transverse
Shell equations with drilling rotation for Green strain We components α3 of this equation are typically neglected, but
use the formulation developed for 3D, see e.g. [10, 21], and the components
  12 and
 21. provide the drilling RC equation.
modify it by introducing the Reissner’s shell kinematics. The Denote skew QT F 12 = (1/2)Cd , where Cd = x0,1 · a2 −
formulation is based on the following functional x0,2 · a1 . Then, the drilling RC can be expressed
 as Cd = 0,
F3 (χ, Q, Ta ) and in Eq. (9) we can use that skew QT F · skew QT F =

.   T    Cd2 . More details related to the drilling RC are given in [27–
= W F F + Ta · skew QT F dV + Fext , (8) 29].
B
where the skew-symmetric Ta is the Lagrange multiplier for Canonical parametrization of rotations: angular velocity and
the RC equation. By regularization of this functional in Ta acceleration Below, we define the canonical representation
we obtain a two-field functional of rotations, and specify the angular velocity and the angular
acceleration. Denote by ψ the canonical rotation vector, for
F̃2 (χ, Q)
 which the rotation tensor Q0 is parameterized as follows
.   γ    

= W FT F − skew QT F · skew QT F dV + Fext , . sin ω 1 − cos ω 2


2 Q0 (ψ) = I + ψ̃ + ψ̃ ,
B ω ω2
(9) ω = ψ = ψ · ψ ≥ 0, (14)
764 K. Wisniewski et al.

Initial Current Variation of Lagrange potential for shell A variation of the


ζ shell Lagrangian potential of Eq. (5), with the modification
of Eq. (9) incorporating rotations, becomes
t3 γ
δL sh = −δ Wsh + δ K sh + δCd2 + δ(Fext )sh , (19)
t2 a3 a2 2
where only the drilling term of the RC is included. Let us
t1 a1 define the vectors of nodal values,
   
. x0I . ẍ0I
q= , a= . (20)
ψI aa I
y0
Note that the difference between ψ̈ I and aa I ! The compo-
x0 nents of Eq. (19) are as follows.

1. The VW of the stress for a 3D body is δ W = V δE·S dV ,
.  
i3 where the Green strain E = (1/2) FT F − I and the 2nd
i2
Piola–Kirchhoff stress S are a work conjugate pair. For
i1 a shell, due to kinematical assumptions, the strain is a
polynomial of the thickness coordinate ζ , i.e. E(ζ ). Also
Fig. 1 Local bases on the reference surface of a shell the stress is a function of ζ , i.e. S(ζ ). Separating the inte-
gration over the reference surface Afrom the integration
. over the thickness we have δ W = A δ Wsh d A, where
where ψ̃ = ψ × I. We assume that ψ ∈ TI SO(3), i.e. the shell VW of stress is
the rotation vector belongs to the tangent space to SO(3) at h
Q0 = I. + 2
.
The material angular velocity vector of Eq. (2) can be δ Wsh = δE(ζ ) · S(ζ ) μdζ. (21)
expressed in terms of the time derivative of the rotation vec-
− h2
tor,
.
On use of the tangent operator for the strain, B = ∂E/∂q,
ω = T(ψ) ψ̇, (15) we can write
where δE(ζ ) = B(ζ )δq,
T(ψ) = s1 I + (1 − s1 )e ⊗ e − s2 ψ̃
h
+ 2

sin ω 1 sin(w/2) 2 δ Wsh = δq · BT (ζ )S(ζ )μdζ. (22)
s1 = , s2 = (16)
ω 2 (w/2) − h2
.
ω = ψ and e = ψ/ψ. For ψ → 0 we find that 2. The kinetic energy of a shell-like body in the initial
T(ψ) → I. (reference) domain B is defined as K = (1/2) B ρ(X, t)
The material angular acceleration vector can be expressed v(X, t) · v(X, t)dV , where ρ is the initial mass density.
as follows For the shell velocity field of Eq. (13), and the definition
. of Eq. (4), we obtain
aa = ω̇ = T(ψ) ψ̈ + Ṫ(ψ) ψ̇, (17) h
+ 2
where
    K sh = 1
2 ρv · vμd ζ
Ṫ(ψ) = a1 ψ̇ · e I + a2 ψ̇ ⊗ e + e ⊗ ψ̇
      − h2
+ a3 ψ̇ · e (e ⊗ e) + a4 ψ̇ · e ψ̃ + a5 ψ̇ × I ,
= 21 Aρ v0 · v0 + 21 Iρ ω · (ω), (23)
(18)
where ω is defined in Eq. (2), and we assumed that ρ =
and the coefficients are constant over ζ , and the reference surface is attached at
cos ω sin ω 1 sin ω the geometrical center of the cross-section. The cross-
a1 = − 2 , a2 = − 2 , sectional mass and inertia for μ ≈ 1 become
ω ω ω ω
h
sin ω cos ω 2 1 sin(ω/2) 2 sin ω + 2
a3 = 3 2 − − , a4 = − 2 , .
ω ω ω ω (ω/2) ω Aρ = ρ(ζ )μdζ = ρh,
2
sin(ω/2) − h2
a5 = − 21 . (24)
(ω/2) + 2
h

  . ρh 3
For ψ → 0 we find that Ṫ(ψ) → −(1/2) ψ̇ × I and ω̇ → Iρ = ρ(ζ )ζ 2 μdζ = ,
  12
ψ̈ − (1/2) ψ̇ × I ψ̇. − h2
Analytical DSA for explicit dynamics of elastic-plastic shells 765

.
where  = I − t3 ⊗ t3 = t1 ⊗ t1 + t2 ⊗ t2 is a purely 3. The drilling term of the RC equation, becomes
geometrical 2nd rank tensor. Note that for ω = ωi ti , i = γ 2
1, 2, 3, where {ti } is a local ortho-normal basis, we obtain δC = (γ Cd ) δq · bd ,
ω · (ω) = ω12 + ω22 , i.e. the drilling angular velocity ω3 2 d
.
does not contribute to K sh . where the tangent operator bd = ∂Cd /∂q.
In order to use the Hamilton’s principle, we calculate the 4. The VW of external loads on the upper and lower bound-
time integral of a variation of the kinetic energy of a shell, ing surfaces, for the shell kinematics becomes
t
i.e. t12 δ K sh dt. A variation of the components of Eq. (23)  

yields δ(Fext )sh = δq · , (31)

δ(v0 · v0 ) = δ(ẋ0 · ẋ0 )
d where p̂ and m̂ are shell-type external forces and moments.
= 2 δ ẋ0 · ẋ0 = 2 (δx0 · v0 ) − 2δx0 · ẍ0 ,
dt Collecting all terms of the variation of the shell Lagrange
(25) potential together, we obtain a weak form of the equation of
motion
δ [ω · (ω)] = 2δ ψ̇ · (Q0 ω)      
ẍ0 0 p̂
d δq · B S − M
T
− + = 0.
= 2 [δψ · (Q0 ω)] aa Iρ [ω × (ω)] m̂
dt
− 2δψ · [aa + ω × (ω)] . (26) (32)

On use of theset expressions we can calculate the


time-integral t12 ( ) dt of δ K sh . The integrals of the terms FE equation of motion for shell Let us introduce the FE
approximations in the configuration space,
with d( )/dt shall vanish due to the time-boundary condi-
tions for variations: δx0 (t) = 0 and δψ(t) = 0 at t = t1 
4 
4
and t = t2 , so, we can consider only the integrand, which x0 = N I x0I . ψ= NI ψ I , (33)
constitutes a variation of the inertial force of a shell, I =1 I =1

δ K sh = −Aρ δx0 · ẍ0 − Iρ δψ · [aa + ω × (ω)] . where N I are the bi-linear shape functions, and I is the node
(27) indicator. Besides,

This equation may be separated into two parts: one depend- (FE)
BT (q)S(q) + (γ Cd ) bd d A = f,
ing on the translational and angular accelerations, and the
other depending on the angular velocity, A
    (34)
− Aρ δx0 · ẍ0 − Iρ δψ · (aa ) Aρ I 0 (FE) p̂ (FE)
d A = M, d A = p,
   0 Iρ  m̂
Aρ I 0 ẍ0 (28)
A A
= −δq · ,
0 Iρ  aa
where f is the internal force vector including the drilling rota-
  tion constraint contribution, M is the mass matrix (symmetric
  0 and constant) obtained from the angular acceleration compo-
δψ · −Iρ [ω × (ω)] = −δq · .
Iρ [ω × (ω)] nent, Eq. (28), and p is the vector of external forces. Besides,
(29) the angular velocity component, Eq. (29), contributes to the
centrifugal inertia vector
Remark We note that for aa = aai ti , i = 1, 2, 3, where   
{ti } is a local ortho-normal basis, the angular acceleration 0 (FE)
d A = c(ω). (35)
term of Eq. (27) is equal to aa = aa1 t1 + aa2 t2 , i.e. the I ρ [ω × (ω)]
drilling angular acceleration aa3 does not affect δ K sh . On A
the other hand, for ω = ωi ti , the angular velocity term Integrating Eq. (32) over the shell area we obtain
yields .
r = f(q) + M a + c(ω) − p = 0, (36)
ω × (ω) = ω × [(I − t3 ⊗ t3 )ω]
= −(t3 · ω)t3 × ω = ω2 ω3 t1 − ω1 ω3 t2 , where r is the residual vector for the whole shell. To make
the formulation more general, we add to Eq. (36) a damping
(30) term, Cq̇,
with the nontrivial contribution of the angular drilling r = f(q) + M a + c(ω) + C q̇ − p = 0, (37)
rotation. This term vanishes for at least two important
cases; for one-directional bending, and for an in-plane where C for explicit dynamics is the Rayleigh damping ma-
deformation. trix in the form C = a1 M.
766 K. Wisniewski et al.

Table 1 Algorithm 1 for explicit dynamics of shells

2.3 Explicit algorithms for dynamics of shells where the stress Sn+1 is calculated by the constitutive algo-
rithm. The above algorithm is only conditionally stable, and
Central finite difference (CFD) operators in time Consider the critical length of the time step for C = a1 M, is
a set of time points,
 t0 < t < · · · <. t , and
1 N
 midpoints   
2 a1
of the intervals t n , t n+1 , i.e. t n+1/2 = (1/2) t n + t n+1 , tcrit = max 1 + ξi − ξi , ξi =
2 , (42)
where n = 0, . . . , N − 1. The time increments are defined i ωi 2 ωi
as follows, where ωi are natural frequencies of a discrete system. For
. increasing ξi , tcrit decreases.
tn+1/2 = t n+1 − t n ,
(38) Explicit algorithms for shells Because of the presence of the
. n+1/2  
tn = t −t n−1/2 =2 t
1 n+1
−t n−1 , rotational degrees of freedom in the equations of shells, we
can define several explicit algorithms, and two of them are
and we replace the time derivatives by the CFD operators, discussed below. We assume that ψ ∈ TI SO(3), i.e. the rota-
q̇n+1/2 − q̇n−1/2 qn+1 − qn tion vector belongs to the space tangent to SO(3) at Q0 = I.
q̈n = , q̇n+1/2 = . (39)
tn tn+1/2 Algorithm
 1 Algorithm 1 uses for the rotational part
The above operators are used as predictors in the explicit ψ, ψ̇, ψ̈ , and its main steps for a single time increment
integration scheme. are given in Table 1. This algorithm introduces one simpli-
fication; ψ̇ n+1/2 is used instead of ψ̇ n+1 in two places, in
Eqs. (47) and (49). We calculate ψ̈ from Eq. (17); this for-
Standard explicit algorithm From Eq. (37) written for t n+1 mula is exact, and can be used also for implicit algorithms,
and the damping term written at t n+1/2 , we obtain a set of but adisadvantage is the inverse of T and a complicated form
equations, 
of Ṫ ψ n+1 . Algorithm 1 is tested in numerical examples in
M an+1 = pn+1 − c(ωn+1 ) − Cq̇n+1/2 − fn+1 , (40) this work, and the DSA is developed for it.

from which an+1 is computed. The internal force vector at Algorithm 2 Algorithm 2 uses for the rotational part {ψ, ω, ω̇},
t n+1 is and its main steps for a single time increment are given in
 Table 2. This algorithm introduces two simplifications: (1)
. in Eq. (54) in T−1 , we use ψ at t n instead of t n+1/2 , (2) in
fn+1 (qn+1 ) = BTn+1 Sn+1 dV, (41)
Eq. (57) in c(ω), ω at t n+1/2 instead of t n+1 .
V
Analytical DSA for explicit dynamics of elastic-plastic shells 767

Table 2 Algorithm 2 for explicit dynamics of shells

Remark Consider Eq. (55) only for the rotational dofs. On 3 Plasticity for finite rotation shells
use of Eq. (54) we obtain
      In this section we describe a generalization of the small defor-
ψ I n+1 = ψ I n + tn+1/2 ψ̇ I n+1/2 mations Huber–Mises plasticity modified by the plane stress
   
= ψ I n + tn+1/2 T−1 ψ I n (ω I )n+1/2 , (50) assumption, to the finite rotation case, intended for shells.
The assumption that the elastic strains are small allows
or to replace the multiplicative elastic/plastic split of F by the
  additive elastic/plastic split of d. The general spatial form
ψ I = T−1 ψ I n tn+1/2 (ω I )n+1/2 . (51)
of the rate equilibrium equation involves the Lie derivative
This form can be compared with Eq. (106) of [4], in which  of the Kirchhoff stress L v τ , but for the finite rotation/small
ψ I = tn+1/2 (ω I )n+1/2 . The presence of T−1 ψ I n in elastic strain case it can be replaced by the objective stress
our algorithm is characteristic for the rotation vector belong- rate of the rotational type, such as the Jaumann rate or the
ing to the initial tangent space, i.e. ψ ∈ TI S O(3). Green–McInnis–Naghdi rate, see [7].
The back-rotated form of the 3D deviatoric Huber–Mises
Mass matrix for a shell In explicit dynamics based on the plasticity fully resembles the one for the small deforma-
CFD operator, the lumped mass matrix should be used, as tion case. A crucial role in the constitutive algorithm plays
then the induced period errors may tend to cancel; the issue the mid-point configuration, see [8]. Then, we can (1) ap-
of matching of dynamic integrators and mass matrices is dis- ply the plane stress assumptions for each shell lamina, and
cussed in [9, p. 505]. For translational dofs, the consistent and re-parameterize the problem in terms of in-plane stress
mass matrix is generated and lumped by row summing, which components, following [14, 24], (2) solve the constitutive
yields a diagonal Mltransl . equations by an elastic predictor/plastic corrector method, in
For the rotational dofs, the mass matrix is computed on which a Newton method is used to determine the consistency
use of Mltransl as follows: parameter. For other algorithms for plasticity and explicit
   dynamics see [30]. The stress determined by this algorithm is
h2 A rotated–forward to the current configuration.
Mlrot = α Mltransl , α=
max 12 , 8 for tangent rotations,
(58)
A
6 for drilling rotation,
where h is shell thickness and A is the element area. The
scaling factor α is obtained from considerations involving 3.1 Kinematics of finite deformation for small elastic strains
moments of inertia of the element. Additionally, in explicit
dynamics, the rotary mass can be scaled to permit larger time Elastic-plastic split of rate of deformation With the purpose
steps without a loss of stability; this is an idea of [15], see to eliminate finite rotations from the deformation of the shell,
.
also [12]. we use the forward-rotated ortho-normal basis {ti∗ = Rti },
768 K. Wisniewski et al.

.
where the rotation tensor R = FU−1 ∈ SO(3) is obtained by Spin increment Because finite rotations are inherent in the
.  1/2
finite deformation problems, we have to predict also the rota-
the polar decomposition F = RU, where U = FT F .
. tion increment, and for this task the spin increment can be
For the current position vector x, we can define xR = .
T helpful. The spin is defined as w̃∗ = skew∇v, and the spin
R x, and extract the rotation from the deformation gradient,
. . ∗ . ∗
i.e. F = RFR , where F = ∂x/∂X and FR = ∂xR /∂X. Then, increment as w̃ = t w̃n+1/2 . We can express the spin
the spatial gradient of velocity is increment in terms of G,
.
∇v = ḞF−1 = R (ω̃ + ∇vR ) RT , (59) w̃∗ = t skew∇vn+1/2 = skewG. (66)
. T The spin increment w̃∗ and the angular velocity increment
where ω̃ = R Ṙ ∈ so(3) is the material angular velocity of
. and ω̃∗ are in general different, which can be shown as
the rotated basis, and ∇vR = ḞR (FR )−1 . The spatial rate-of-
deformation is follows. By time differentiation of F = RU we obtain
.
d = sym∇v = R (sym ∇vR ) RT . (60) ∇v = ḞF−1 = ω̃∗ + d + skewδ, (67)
. . ∗ .
We also define the back-rotated rate-of-deformation, D = −1
where ṘUF = ṘR −1 −1
 = −1 ω̃ , RU̇F = RU̇U R =
T T
sym∇vR = R dR. The components of D in {ti } and d in
T
δ, and symδ = sym RU̇U RT = (1/2)F−T ĊF−1 = d.
{ti∗ } are equal. .
Hence, the spin is w̃∗ = skew∇v = ω̃∗ + skewδ, and the
For the multiplicative elastic/plastic split: FR = FRe Fp ,
R spin increment is
e p
where FR is an elastic deformation gradient and FR is a
w̃∗ = ω̃∗ + skew δ, (68)
plastic part of FR , we obtain
∗ . ∗ .
p  e −1 p where ω̃ = t ω̃n+1/2 , δ = tδ n+1/2 , and skew δ = 0.
∇vR = ∇vRe + FR e
∇vR FR ≈ ∇vRe + ∇vR , (61) However, in explicit dynamics the increments are very small,
.  −1 p . p  p −1 and this difference is typically neglected.
where ∇vRe = ḞR e Fe
R and ∇vR = ḞR FR . The last
additive form is obtained for FR e ≈ I, i.e. assuming that the
elastic deformation gradient is small (close to unity). The
3.2 Constitutive elastic-plastic equations for finite
back-rotated rate-of-deformation also splits additively,
. p
deformations
D = sym ∇vR ≈ sym ∇vRe + sym ∇vR = De + Dp , (62)
. . p Incremental constitutive equation for finite deformations
where De = sym∇vRe and Dp = sym∇vR . Concluding, for
the finite rotation/small strain case, the multiplicative elas- The constitutive equation is postulated in the spatial rate
tic-plastic split of F implies the additive split of D; this is form, involving the objective rate of the Kirchhoff stress τ ,
a generalization of the classical result to the case involving and the rate-of-deformation tensor,
finite rotations, and, therefore, important for shells. • 4 .
τ =a : d e , d e = d − d p , (69)
Strain increment If we use d as the strain rate measure, i.e. 4
. where a is a rank 4 elastic constitutive operator. The Green–
ε̇ = d, then the increment of the spatial strain can be defined McInnis–Naghdi rate of the Kirchhoff stress is defined as
and approximated as follows, follows


t n+1 • . ∂  T 
. τ =R R τ R RT
ε = d(t) dt ≈ t dn+1/2 . (63) ∂t
tn ˙
= R
RT = τ̇ + τ ω̃∗ − ω̃∗ τ , (70)
We can show that ε is the push-forward to the mid-point po- ∗ .
where ω̃ = ṘR = Rω̃R ∈ so(3) is the spatial (left) angu-
T T
sition of the Green strain increment E. From Ċ = 2FT dF,
lar velocity tensor. This rate of stress is rotationally objec-
where C is the right Cauchy–Green tensor, we calculate d, • . T
for which tive, i.e. RT τ R =
, ˙ where
= R τ R is the back-rotated
−T −1 Kirchhoff stress, and
˙ is rotationally indifferent. The com-
ε ≈ t dn+1/2 ≈ Fn+1/2 E Fn+1/2 , (64) ponents of
in {ti } and of τ in {ti∗ } are equal. More details
where we used (1/2) t Ċn+1/2 ≈ (1/2)(Cn+1 − Cn ) = on this type of a formulation and the objective stress rates
.
En+1 − En = E. We note similarity of Eq. (64) to the for- can be found e.g. in [7].
−T −1 To obtain the incremental form with subtracted rotations,
mula for the Almansi strain, en+1/2 = Fn+1/2 En+1/2 Fn+1/2 . .
we write Eq. (69) at t n+1/2 , multiply it by t = t n+1 − t n ,
The definition
 T of E and Eq. (64) imply the equation, . .
introduce τ = τ n+1/2 t and ε e = dn+1/2 t, and, next,
e
(1/2) Fn+1 Fn+1 − FnT Fn = Fn+1/2
T ε Fn+1/2 , from which
we can obtain
T
back-rotate Rn+1/2 (·)Rn+1/2 , which yields
ε = symG, (65) 4 p
.
=A : εeR , ε eR = ε R − εR , (71)
where G = ∂ u/∂xn+1/2 is the gradient of the displacement . .
increment w.r.t. the mid-point configuration. Equation (65) where
= Rn+1/2T τ Rn+1/2 , and εR = Rn+1/2 T ε
is used in numerical calculations. Rn+1/2 . The constitutive operator is defined by the relation,
Analytical DSA for explicit dynamics of elastic-plastic shells 769

4 4 the in-plane components, and, instead of P4×3 , we use the


T
Rn+1/2 (an+1/2 : ε e )Rn+1/2 =An+1/2 : ε eR . For Hooke’s
operator
elasticity the constitutive operator is constant over the time ⎡ ⎤
step, and including the plane stress conditions it is as follows 2 −1 0
⎡ ⎤ . 1
1ν 0 P = ⎣ −1 2 0 ⎦ , (78)
4 4 . E ⎣ 3 0 03
An+1/2 =A= ν 1 0 ⎦, (72)
1 − ν 2 0 0 1−ν which is invertible. Hence, for plane stress, Eqs. (73)–(76)
2
can be re-written as follows,
where E, ν are the Young’s modulus and the Poisson’s ratio.
Equation (71) has the form analogous to that for small defor- 1. The yield criterion,
mation plasticity, see e.g. [23]; thus an analogous form of the
plasticity evolution equations can be postulated. f (
, α PR , ēp ) = 21 η · (Pη) − 21 R 2 = 0, (79)
2. The associative flow rule,
Huber–Mises plasticity for finite deformation and plane stress ėp = γ Pη, (80)
To the back-rotated constitutive Eq. (71) we can apply the
Huber–Mises plasticity equations developed for small defor- 3. The isotropic and kinematic hardening rules,
mation. The plasticity evolution equations for 3D finite defor- 
mation are as follows, R = 23 κ(ēp ), α̇ PR = γ 23 Hα (ēp )η, (81)
1. The yield criterion, 4. The effective plastic strain (rate form),

1 1 .
f (
D , α R , ēp ) = φ 2 − R 2 = 0, φ = 
D −α R , ē˙p = γ 23 η · (Pη), (82)
2 2
(73) 5. The normal plastic strain,
p p p
ė33 = −ė11 − ė22 , (83)
2. The associative flow rule, .
. where η =
− α PR . The yield surface of Eq. (79) is an
ėp = γ (
D − α R ), ė = ε̇ R − 13 (tr ε̇ R )I, (74) ellipsoid, thus not a radial return algorithm but a return map
3. The isotropic and kinematic hardening rules, algorithm must be used, e.g. with a Newton loop. The above

finite-deformation form of plasticity equations is analogous
R = 23 κ(ēp ), α̇ R = 23 Hα (ēp ) ėp , (75) to that for small deformations of [24]. The following hard-
ening functions are used in our code:
4. The effective plastic strain (rate form),  p
 κ(ēp ) = σ y0 + κ1 ēp + (σ y1 − σ y0 ) 1 − e−a ē ,
 (84)
˙ēp = 2 ėp (τ ) = 2 γ φ,
3 3
(76) Hα (ēp ) = Hα ēp ,
. i.e. a non-linear ‘saturation-type’ function for the isotropic
where
D = RT τ D R is the back-rotated Kirchhoff stress
. hardening and a linear function for the kinematic hardening.
deviator, and α R = RT αR is the back-rotated back-stress,
γ ≥ 0 is the consistency parameter, κ(ēp ) is the isotro-

pic hardening function, Hα (ēp ) is the kinematic hardening
3.3 Solution of constitutive equations for shell lamina
modulus.
For shells, typically, 3D constitutive equations are modi-
Time-discrete constitutive equations For the plastic state, our
fied by the plane stress conditions. To obtain the stress devia-
objective is to solve the constitutive Eq. (71), and the evo-
tor and to account for the plane stress, we use one projection
lution equations (80), (81), (82), and (79) written at t n+1 as
operator P4×3 ,
follows:

D = P4×3
, α R = P4×3 α PR , (77) 4  

n+1 −
n − A ε R − ep = 0, (85)
where p
⎧ ⎫ ⎡ ⎤ ⎧ ⎫ ėn+1 − γn+1 Pηn+1 = 0, (86)
⎪  D11 ⎪ 2 −1 0  p 
⎨ ⎬ ⎨  ⎬
.  D22 . 1 ⎢ −1 2 0 ⎥ . 11
α̇ PR,n+1 − 23 Hα ēn+1 γn+1 ηn+1 = 0, (87)

D = , P4×3 = ⎣ ⎦,
= 22 . 
⎩  D12 ⎪
⎪ ⎭ 3 0 03 ⎩ ⎭  
ē˙n+1 − γn+1 23 ηn+1 · Pηn+1 = 0,
p
 D33 −1 −1 0 12 (88)
We can omit the equations associated with the normal trans-     p 
f n+1 = ηn+1 · Pηn+1 − 23 κ ēn+1 = 0. (89)
verse direction in Eqs. (71), (74) and (75), and recover the
normal strains as follows: the elastic strain from the zero To obtain a time-discrete form of them, the time derivatives
normal stress (ZNS) condition, and the plastic strain from are replaced by the backward Euler operators. For instance,
.
the incompressibility condition. Then, in
D remain only Eq. (86) yields, ep = λn+1 Pηn+1 , where λn+1 = γn+1 t
770 K. Wisniewski et al.

is a counterpart of the consistency parameter. On use of this Table 3 Initial computations for lamina ζ
equation, Eq. (85), can be re-written in the form
n+1 −
4 . 4

TR + λn+1 A Pηn+1 = 0, where


TR =
n + A εR is
the trial stress. Adding and subtracting α PR,n+1 , we obtain
  4

n+1 − α PR,n+1 −
TR + α PR,n+1 + λn+1 A Pηn+1 = 0,
(90)
.
which, on use of the definition ηn+1 =
n+1 − α PR,n+1 and
the discrete form of Eq. (87) yields Eq. (91). Finally, the
time-discrete version of Eqs. (85)–(89) is as follows:

4

(1 + 23 λn+1 Hα,n+1 ) I + λn+1 A P ηn+1 − ηTR = 0,(91)
  
p p
ēn+1 − ēn − λn+1 23 ηn+1 · Pηn+1 = 0, (92)
      and Rn+1/2 is used to rotate-back ε, while Rn+1 to rotate-

ηn+1 · Pηn+1 − 23 κn − 23 κn+1 λn+1 ηn+1 · Pηn+1 = 0. forward
n+1 . The initial computations for a single lamina
(93) are summarized in Table 3, see also [11, 22]. Note that the
rotations Rn must be retrieved from the history database of
.
Note that ηTR =
TR − α PR,n and all symbols with the sub- an element.
p
script n are known. Unknown are: ηn+1 , ēn+1 , λn+1 , totally
five unknowns. This set of nonlinear constitutive equations Constitutive algorithm at shell lamina The constitutive algo-
is solved by the algorithm described e.g. in [14, 24]. This rithm for finite deformation plane stress elasto-plasticity with
algorithm is strain driven, and in explicit dynamics, εR is nonlinear isotropic (saturation)/linear kinematic hardening is
computed for the predicted position vector xn+1 . presented in Table 4. All operations are performed on back-
rotated objects. For the plastic state, the following steps are
performed:
Computation of lamina rotations Consider one selected lam-
ina of the shell at ζ , for which the current position is given 1. the stress
is initially elastically predicted. For the trial
. ηTR and old values of H , ēp , and λ we find ηn+1 from
by the Reissner hypothesis, Eq. (12), and qn+1 = {(x0 )n+1 ,
ψ n+1 } is predicted by Eq. (45) in Table 1. To compensate for Eq.
p
(91)
the classical Reissner kinematics without the normal stretch 2. ēn+1 is updated on use of Eq. (92)
parameter, for large plastic strains the shell thickness is addi- 3. the yield condition, Eq. (93), is solved for the consis-
.
tionally updated. tency parameter λ = λn+1 by the Newton scheme; Steps
The shell director rotation Q0 , which appears in the 5–11 of Table 4 are repeated until a convergence of λ is
Reissner hypothesis, Eq. (12), should not be confused with obtained. In the table, f ,λ denotes a derivative of f with
the lamina rotation R, which is associated with the polar respect to λ, and this is a slope coefficient in the Newton
decomposition of F for the lamina. Consider the increment scheme. The formula for f ,λ of Step 10 is only approx-

of the lamina rotation R, corresponding to ω̃ . In explicit imate, as the condition f = 0 was used in its derivation.
dynamics increments are very small, and hence, we can use This condition is not satisfied during iterations, and hence
R ≈ r, where the iteration scheme may have a convergence rate lower
   −1
than quadratic. N.B. during the DSA, the condition f = 0
r = I + 21 w̃∗ I − 21 w̃∗ is already satisfied, and then f ,λ is exact.
 −1
p
4.
n+1 , ēn+1 , α PR,n+1 are updated. These are the constit-
= I + I − 21 w̃∗ w̃∗ (94)
utive state parameters, which are stored and updated by

is the rotation for the spin increment w̃ . Generally, w̃ ∗ the constitutive algorithm.
is different than ω̃∗ , see Eq. (68). Equation (94) is the for- Finally, the Kirchhoff stress and the 2nd Piola–Kirchhoff
mula for the semi-tangential parametrization of rotations, or stress at t n+1 are calculated,
can be obtained by the mid-point rule applied to the gener- −1 −T
ating equation, ṙ = wr, see [13]. The rotation increment to τ n+1 = Rn+1
n+1 Rn+1 , Sn+1 = Fn+1 τ n+1 Fn+1 .
T

the mid-point (m-p) configuration, i.e. from t n to t n+1/2 , is (97)


computed as
. √
( R)m-p = R ≈ 21 (I + R). (95) 4 Constitutive DSA for finite rotation shells
On use of the increments of rotations, we can compute,
The objective of the DSA is to calculate a derivative of se-
Rn+1/2 = ( R)m-p Rn , Rn+1 = R Rn , (96) lected performance functions w.r.t. some design parameters.
Analytical DSA for explicit dynamics of elastic-plastic shells 771

Table 4 Constitutive algorithm for plane stress elastic-plasticity for shell lamina

In this paper, we consider the constitutive DSA, in which where ∂m (·) is the explicit derivative. In the sequel, we also
the material parameters m are used as the design variables. denote a single design variable by m, where m ∈ m.
For the constitutive model consisting of Hooke’s elasticity,
see Eq. (72), and Huber–Mises plasticity with the hardening
functions of Eq. (84),
.   4.1 Scheme of design differentiation of explicit dynamics
m = E, ν, Hα , κ1 , σ y0 , σ y1 , a . (98) algorithm
In order to generate the DSA equations, we differentiate w.r.t.
m all equations of the dynamic algorithm, including the con- The DSA of explicit dynamics for elasto-plastic finite rota-
stitutive algorithm. For any composite (scalar, vector or ten- tion shells is difficult due to the following factors:
sor) function f (m, g(m)), the design derivative Dm (·) is cal- 1. The kinematical part of the shell element is complicated
culated using the chain rule as follows due to the parametrization of finite rotations and the rota-
∂f tional inertia terms, as well as a treatment of the trans-
Dm f = ∂m f + Dm g, (99) verse shear strains by the ANS procedure [2]. The design
∂g
772 K. Wisniewski et al.

Table 5 Design sensitivity analysis (DSA) Algorithm 1 for explicit dynamics of shells

derivatives have to be computed in a way to avoid multi- The DSA computations are performed step-by-step sim-
dimensional matrices. ilarly as the dynamics computations, and it is most efficient
2. The constitutive algorithm is incremental and uses sev- when both analyses are performed in parallel, and every step
eral history variables, for which we have the design deriv- of the DSA in Table 5 immediately follows the corresponding
atives, which also are the history variables. Besides, we step of Table 1.
have predictors for q, q̇, and their design derivatives. This Design differentiation of the formula for the update of
requires specific update schemes for the design deriva- acceleration, Eq. (48) of Algorithm 1, yields the following
tives of: (1) constitutive state variables, and (2) stresses. DSA equation:
3. The Newton’s iteration loop of the constitutive algorithm  
in the design-differentiation must be treated in a special ẍ0I
Dm
way to avoid replication of the loop on the DSA level, aa I n+1
which would be inefficient.  
= M−1 −Dm c(ωn+1 ) − C Dm q̇n+1/2 − Dm fn+1 ,
Therefore, an implementation of the DSA requires a clear
(100)
understanding of the whole solution algorithm and the data
handling scheme, introduces additional specific subroutines the components of which are discussed below.
and enlarges the data base.
The design-differentiation of Algorithm 1 of Table 1
yields the DSA algorithm presented in Table 5. The design- Design derivative of internal force The design derivative of
differentiation of the constitutive algorithm of Table 4 yields the internal force f is as follows:

the algorithm of Table 6. The closed-form design-differenti- . ∂
ated equations are extremely elaborate and that’s why only Dm f = BT S dV
∂m
some of them are presented here. However, we describe the V

update schemes, taking into account the history variables and  
the predicted quantities, the correctness of which is crucial = Dm BT S + BT Dm S dV, (101)
for accuracy of the DSA. V
Analytical DSA for explicit dynamics of elastic-plastic shells 773

Table 6 Design sensitivity analysis of the constitutive algorithm for plane stress elasto-plasticity of Table 4

where S is the 2nd Piola–Kirchhoff stress. The kinematical Design derivative of inertial term Regarding
  the inertial term
.
operator, B = ∂E/∂q,where E is the Green strain, for the c(ωn+1 ), we note that ωn+1 ≈ T ψ n+1 ψ̇ n+1/2 , i.e. it
nonlinear kinematics of finite rotation shells depends on q, depends on ψ n+1 and ψ̇ n+1/2 . Hence, the design derivative
hence its design derivative is is calculated as follows:
∂c
dB dq Dm c(ωn+1 ) = Dm ψ n+1
Dm B = = B,q Dm q, (102) ∂ψ n+1
dq dm
∂c
+ Dm ψ̇ n+1/2 . (104)
where B,q = ∂ 2 E/∂q2 is a (n s × n dof × n dof ) matrix, and n s ∂ ψ̇ n+1/2
is a number of strain components. To avoid the three-dimen-
sional matrix B,q we differentiate the vector (BT S) keeping
S constant, which yields a two-dimensional matrix (BT S),q , 4.2 Scheme of design differentiation of constitutive
i.e. algorithm
 
d BT S dq Update of design derivatives of constitutive state variables
Dm B S =
T
Let us denote by s the state variables which are incremen-
dq dm
 T  tally updated, and stored as the constitutive history data. In
. p
= B S ,q Dm q, for S = constant (103) our formulation, s = {
, ēp , α R , ε33 , ψ R }, where ψ R are
rotational parameters of the lamina rotation R. The update
The derivative of the 2nd Piola–Kirchhoff stress, Dm S, is algorithm for sn+1 can be schematically written as
discussed separately. sn+1 = sn+1 (m, sn (m), qn+1 (m), qn (m)) , (105)
774 K. Wisniewski et al.

Table 7 Results of the constitutive analysis

Total Total Back Plastic Equivalent plastic


stress stress Plastic
Step strain strain strain
flow?
ε σ (×106 ) α(×106 ) εp ēp
 &  &  &  &
0.005 1223. 0 0
1 0.001 576.9 0 0 0 no
0.006 484.6 0 0
 &  &  &  &
0.011 2474. 4.786 0.001158
2 0.003 1356. 2.624 0.00007684 0.001628 yes
0.011 768.4 1.487 0.001487
 &  &  &  &
0.020 2675. 42.23 0.009684
3 0.006 1695. 26.36 0.001748 0.01409 yes
0.021 739.8 11.84 0.01184
 &  &
0.014 1152.
4 0.004 818.4 · · · · · · · · · · · · · · · · · · · · · · · · · as in step 3· · · · · · · · · · · · · · · · · · · · · · no
0.011 −67.88
 &  &  &  &
−0.004 −2105. 2.145 0.002996
5 −0.012 −2118. −13.69 −0.004919 0.03124 yes
−0.019 −986.9 −6.792 −0.006792

and upon differentiation w.r.t. m we obtain the DSA update η2


formula,
∂sn+1 ∂sn+1
Dm sn+1 = + Dm sn
∂m ∂sn
∂sn+1 ∂sn+1
+ Dm qn+1 + D m qn . (106)
∂qn+1 ∂qn
3
The derivatives, Dm sn and Dm qn are known as they are 4 2
the history variables. Also Dm qn+1 is known, as in explicit
dynamics is predicted, see Table 5. The derivatives, 1 η1
∂sn+1 ∂sn+1 ∂sn+1 ∂sn+1
, , , , (107)
∂m ∂sn ∂qn+1 ∂qn
must be explicitly calculated. This update includes the de- 5
sign derivative of back-rotated Kirchhoff stress
, which is
included in s. Note that, alternatively, we can directly differ-
entiate the algorithm of Table 4, and the result is shown in
Table 6.

Update of design derivative of stress The algorithm to calcu-


late the 2nd Piola-Kirchhoff stress S can be written Fig. 2 Loading history and evolution of flow surface in principal values
η1 , η2 of η
schematically as
Sn+1 = Sn+1 (m, sn (m), qn+1 (m), qn (m)), (108)
and upon differentiation w.r.t. m we obtain Design derivative of yield condition The yield condition,
f = 0, is a nonlinear equation of λ, and is solved by the
∂Sn+1 ∂Sn+1
Dm Sn+1 = + Dm sn Newton method in Steps 5–11 of Table 4. When the constitu-
∂m ∂sn tive algorithm is design-differentiated then this Newton loop
∂Sn+1 ∂Sn+1 provides another Newton loop. We can avoid this Newton
+ Dm qn+1 + D m qn , (109)
∂qn+1 ∂qn loop in the DSA computations, by computing Dm λ.
To compute Dm λ, we design-differentiate the yield con-
where Dm sn , Dm qn+1 and Dm qn are known, and the deriv- dition, f = 0. The algorithm to calculate f is defined in
atives, Steps 5–9 of Table 4, and can be written schematically as
∂Sn+1 ∂Sn+1 ∂Sn+1 ∂Sn+1
, , , , (110) f = f (m, λ(m), sn (m), qn+1 (m), qn (m)). (111)
∂m ∂sn ∂qn+1 ∂qn
can be explicitly calculated. Upon differentiation w.r.t. m we obtain
Analytical DSA for explicit dynamics of elastic-plastic shells 775

(a)

(b) (c)

Fig. 3 Plate: a geometry, b material constants, c load history

Fig. 4 Plate. Explicit dynamics: deformation stages

Table 8 Maximum relative differences between analytical and semi- 7e+06


kinetic
analytical design derivatives internal
6e+06
external
i mi Maximum difference balance
5e+06
1 E 0.104E−05
Shell energy

2 ν 0.282E−06
4e+06
3 σ y0 0.320E−06
4 σ y1 0.435E−05
5 a 0.220E−06 3e+06
6 κ1 0.206E−06
7 Hα 0.195E−06 2e+06

1e+06

Dm f = f ,λ Dm λ + (Dm f )∗ , (112) 0
where 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
. ∂f ∂f time t
(Dm f )∗ = + Dm sn
∂m ∂sn Fig. 5 Plate. Explicit dynamics: total energy balance
∂f ∂f
+ Dm qn+1 + D m qn . (113)
∂qn+1 ∂qn
Note that Dm sn , Dm qn+1 , and Dm qn are known, and the can be explicitly calculated, so (Dm f )∗ is fully determined.
derivatives, Also f ,λ can be calculated, and its form is given in Step 10
∂f ∂f ∂f ∂f of Table 4. Hence, from the design-differentiated yield con-
, , , , (114) dition, Dm f = 0, on use of Eq. (112), we can calculate
∂m ∂sn ∂qn+1 ∂qn
776 K. Wisniewski et al.

0 0

-0.2 -0.2

-0.4 -0.4

-0.6 -0.6

-0.8 -0.8

-1 -1

-1.2 -1.2

-1.4 -1.4

-1.6 -1.6
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

3.5

2.5

1.5

0.5

0
0 0.2 0.4 0.6 0.8 1

2 0
1.8 -0.2
1.6
-0.4
1.4
1.2 -0.6

1 -0.8
0.8 -1
0.6
-1.2
0.4
0.2 -1.4

0 -1.6
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Fig. 6 Plate. Explicit dynamics: displacements and rotations at free corner

(Dm f )∗ Dm ηTR = Dm
TR − Dm α PRn
Dm λ = − , (115)
f ,λ
. 4 4
= Dm ηn + ∂ m A εR + A Dm ε R ,
and, differently than in the constitutive computations, avoid
Newton iterations. 4
Instead of computing the derivatives as in Eq. (114), we (Dm )∗ = −λ ( 23 ∂ m Hα I + ∂ m A P) ,
can determine (Dm f )∗ from Eqs. (72), (84) and Steps 5–9
of Table 4, by treating λ as if it were design-independent, (Dm η)∗ = (Dm )∗ ηTR +  Dm ηTR

Dm κn = κn Dm ēn + ∂ m κ1 ēn + ∂ m σ y1
p p 4 4
=  Dm ηn +A Dm εR +∂ m A ε eR − 23 λ ∂ m Hα η ,
   p
+ σ y1 −σ y0 ∂ m a ēn +∂ m σ y0 −∂ m σ y1 e−a ēn ,
p
   ∗  ∗
∂ m κ = ∂ m κ1 + ∂ m σ y1 − ∂ m σ y0 a Dm φ 2 = Dm ηT Pη + ηT P (Dm η)∗ ,
    ∗
+ σ y1 − σ y0 ∂ m a (1 − a ēp ) e−a ē , (Dm φ)∗ =
p
1
2φ Dm φ 2 ,
Analytical DSA for explicit dynamics of elastic-plastic shells 777

1.5e-12 1e-12
1e-12 5e-13
5e-13
0
0
-5e-13
-5e-13
-1e-12 -1e-12
-1.5e-12 -1.5e-12
-2e-12
-2e-12
-2.5e-12
-2.5e-12
-3e-12
-3.5e-12 -3e-12
-4e-12 -3.5e-12
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

3e-12
2.5e-12
2e-12
1.5e-12
1e-12
5e-13
0
-5e-13
-1e-12
-1.5e-12
-2e-12
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

3.5e-12 2e-12
3e-12 1.5e-12
2.5e-12 1e-12
2e-12 5e-13
1.5e-12 0
1e-12 -5e-13
5e-13 -1e-12
0 -1.5e-12
-5e-13 -2e-12
-1e-12 -2.5e-12
-1.5e-12 -3e-12
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

Fig. 7 Plate. DSA: design derivatives w.r.t. Young’s modulus E of displacements and rotations at free corner; lines – analytical, crosses – FD
  
(Dm κ)∗ = Dm κn + 3λ
2
∂ m κ φ + κ (Dm φ)∗ Design derivative of T operator From Eq. (15) we have
  
 '  ∗ ( Dm T(ψ) = ω1 (Dm ψ · e) (1−s1 −ω2 s2 ) I+ 3(s1 −1)+ω2 s2 e⊗e
= Dm κn + 2λ
φ ∂mκ + 1

2κ Dm φ ,
2 2

+(2s2 −s1 ) ψ̃
 ∗
Dm κ 2 = 2κ (Dm κ)∗ + ω1 (1−s1 ) (Dm ψ ⊗ e + e ⊗ Dm ψ)−s2 Dm ψ × I,
 (117)
∗ which is analogous to Eq. (18). Differentiating the time deriv-
=2 2 φ (Dm κ)
3
ative Ṫ(ψ), Eq. (18), we obtain
  ∗  
=2 Dm κn + 2λφ 2 ∂ m κ + λκ Dm φ 2 , Dm Ṫ(ψ) = 1
(Dm ψ · e) ψ̇ · e

3
ω 2
 
    · 3(s1 −1)+ω2 (3s2 −s1 ) I−(15(s1 −1)
∗ 2 ∗ 2 ∗
(Dm f ) = Dm φ − Dm κ 

1 1
2 3   
+ ω2 (7s2 − s1 ) e ⊗ e + 5s1 −1−s2 ω2 −8 ψ̃
 ∗   
= 21 γ2 Dm φ 2 − 3φ
2
Dm κn − 23 λφ 2 ∂ m κ . (116) + ω12 Dm ψ̇ · ψ + ψ̇ · Dm ψ
778 K. Wisniewski et al.

0.3 0.15
0.1
0.2
0.05
0.1
0
0 -0.05

-0.1 -0.1
-0.15
-0.2
-0.2
-0.3
-0.25
-0.4 -0.3
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

0.25
0.2
0.15
0.1
0.05
0
-0.05
-0.1
-0.15
-0.2
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

0.25 0.2

0.2 0.15
0.1
0.15
0.05
0.1 0
0.05 -0.05

0 -0.1
-0.15
-0.05
-0.2
-0.1 -0.25
-0.15 -0.3
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

Fig. 8 Plate. DSA: design derivatives w.r.t. Poisson’s ratio ν of displacements and rotations at free corner; lines – analytical, crosses – FD

   
· 1−s1 −ω2 s2 I + 3(s1 −1)+ω2 s2 e ⊗ e 5 Numerical examples

+ (2s2 −s1 ) ψ̃ First, the generic tests of the elastic plastic module, and next
    two dynamical examples are presented: (1) the bending of an
+ ω12 ψ̇ · e 3(s1 −1)+ω2 s2 (Dm ψ ⊗e+e ⊗ Dm ψ) elastic-plastic rectangular plate, and (2) the axial compres-

sion of an elastic-plastic tube. Two types of comparisons are
+(2s2 −s1 )ω Dm ψ × I
performed to assess quality of our results:
  
+ ω12 (Dm ψ · e) 3(s1 −1)+ω2 s2 ψ̇ ⊗ e + e ⊗ ψ̇ 1. for dynamics, our results are compared with results ob-

tained with a commercial code ABAQUS Explicit [1],
+ (2s2 −s1 ) ω ψ̇ × I by the element “S4R”, which is a one-point integration
 stabilized element. Our finite-rotation shell element is
+ ω12 (1−s1 ) ψ̇ ⊗ Dm ψ + Dm ψ ⊗ ψ̇ + Dm ψ̇ ⊗ ψ
 developed for the variational formulation of Eq. (9), and
+ ψ ⊗ Dm ψ̇ − s2 Dm ψ̇ × I. (118) its characteristic features are as follows:
Analytical DSA for explicit dynamics of elastic-plastic shells 779

1.4e-08 1.6e-08

1.2e-08 1.4e-08

1e-08 1.2e-08
1e-08
8e-09
8e-09
6e-09
6e-09
4e-09
4e-09
2e-09 2e-09
0 0
-2e-09 -2e-09
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

0
-2e-09
-4e-09
-6e-09
-8e-09
-1e-08
-1.2e-08
-1.4e-08
-1.6e-08
-1.8e-08
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

0 1e-08
-1e-09 9e-09
-2e-09 8e-09
-3e-09 7e-09
-4e-09 6e-09
-5e-09 5e-09
-6e-09 4e-09
-7e-09 3e-09
-8e-09 2e-09
-9e-09 1e-09
-1e-08 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

Fig. 9 Plate. DSA: design derivatives w.r.t. yield limit σ y0 of displacements and rotations at free corner; lines – analytical, crosses – FD

(a) 4-node, bi-linear, 6 dofs/node including drilling rota- 2. for DSA; the analytical derivatives are compared with
tion, the derivatives by the finite-difference method. The latter
(b) initial geometry specified by positions of nodes, are obtained by running the code twice, with each design
warped, parameter m perturbed by +δm and −δm, and computing
(c) at nodes: displacements u, and canonical rotation the approximate design derivative of a response q by the
vector ψ, both in global frame, central finite difference operator,
(d) work-conjugate pair: Green strain, 2nd Piola–Kirch-
hoff stress, q(m+δm) − q(m−δm)
(e) the ANS for the transverse shear of [2], d̃ m q = . (119)
2δm
(f) drilling RC equation at integration points, by the
penalty method, The use of the automatic differentiation program AceGen
(g) numerical integration: 2 ×2 Gauss in lamina, 5-point developed by J. Korelc, see e.g. [19], is is gratefully acknowl-
Simpson rule over thickness. edged.
780 K. Wisniewski et al.

8e-11 8e-11
7e-11 7e-11
6e-11 6e-11
5e-11 5e-11
4e-11 4e-11
3e-11 3e-11
2e-11 2e-11
1e-11 1e-11
0 0
-1e-11 -1e-11
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

2e-11
1.5e-11
1e-11
5e-12
0
-5e-12
-1e-11
-1.5e-11
-2e-11
-2.5e-11
-3e-11
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

0 1e-10
9e-11
-2e-11
8e-11
7e-11
-4e-11
6e-11
-6e-11 5e-11
4e-11
-8e-11
3e-11
2e-11
-1e-10
1e-11
-1.2e-10 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

Fig. 10 Plate. DSA: design derivatives of displacements and rotations at free corner w.r.t. ‘infinity’ yield limit σ y1 ; lines – analytical, crosses –
FD

5.1 Generic tests of an elastic-plastic module analytical method. After the DSA computations, we eval-
uate the relative difference
In order to verify the DSA implementation, an additional pro- ) )
) Dm y AN − Dm y SA )
gram is written, which enables testing outside an FE code, ) ) ) ) ,
diff = (120)
and gives rise to two types of comparisons: max ) Dm y AN ) , ) Dm y SA )
• for the constitutive algorithm, we can compare the results where AN indicates the analytical method, and SA the
yielded by the original constitutive routines and by these semi-analytical method.
modified for the constitutive DSA. This comparison plays
the role of a QA test and ensures that the DSA-related All the material constants are treated as the design parame-
modifications did not affect the constitutive results. ters, i.e. seven different design derivatives are computed. The
• for the constitutive DSA algorithm, we can compare the following elastic-plastic material constants were used in the
design derivatives computed by the analytical and semi- test:
Analytical DSA for explicit dynamics of elastic-plastic shells 781

Fig. 11 Tube: a geometry, b material constants

Fig. 12 Tube. Explicit dynamics: deformation of notch vicinity

12 1
0
10
-1
8
-2
6 -3

4 -4
-5
2
-6
0
-7
-2 -8
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5

Fig. 13 Tube. DSA: displacement at point P

E = 2.1 · 1011 σ y0 = 2.5 · 109 , κ1 = 2 · 108 , a = 50, shown in Fig. 2. Values of stresses and state parameters at the
ν = 0.3, σ y1 = 2.75 · 109 , Hα = 3 · 109 . end of steps are given in Table 7. The original constitutive
routines and the routines modified for the constitutive DSA
A five-step scheme, driven by arbitrarily assumed strain
yielded identical results.
increments, has been applied. The steps correspond to five
To test the constitutive DSA algorithm, the design deriv-
characteristic situations for elastic-plastic materials:
atives of the strain increments are assumed as a ratio of the
1. elastic loading; ε = {0.005, 0.001, 0.006}, strain increments and the value of the design parameter,
2. continuing elastic loading up to entrance into the plas-
tic range + further elastic-plastic loading; ε = d( ε) = −5 ε , i = 1 . . . 7, (121)
{0.006, 0.002, 0.005}, dm i mi
3. elastic-plastic loading; ε = {0.009, 0.003, 0.010},
where the multiplier 5 is taken arbitrarily. To verify the ana-
4. elastic unloading; ε = {−0.006, −0.002, −0.010},
lytical DSA results, seven additional semi-analytical analy-
5. continuing elastic unloading until re-entrance into the
ses are performed; in each of them, one design parameter,
plastic range + further elastic-plastic loading; ε =
m i , is perturbed by δm i = 10−8 m i . The computed values of
{−0.018, −0.016, −0.030 }.
analytical and the approximate semi-analytical design deriv-
The loading history and the evolution of the plastic flow sur- atives are compared using Eq. (120), and maximum relative
face, expressed in principal values of the relative stress η are differences for all design variables are shown in Table 8. In
782 K. Wisniewski et al.

4e-06 1.5e-06
3.5e-06
3e-06 1e-06

2.5e-06
5e-07
2e-06
1.5e-06
0
1e-06
5e-07 -5e-07
0
-5e-07 -1e-06
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5

Fig. 14 Tube. DSA: design derivatives w.r.t. Young’s modulus E at P; lines – analytical, crosses – FD

0.2 1.2

1
0.15
0.8
0.1 0.6

0.05 0.4

0.2
0
0

-0.05 -0.2
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5

Fig. 15 Tube. DSA: design derivatives w.r.t. Poisson’s ratio ν at P; lines – analytical, crosses – FD

0.009 0.004
0.008
0.003
0.007
0.006 0.002
0.005 0.001
0.004
0.003 0

0.002 -0.001
0.001
-0.002
0
-0.001 -0.003
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5

Fig. 16 Tube. DSA: design derivatives w.r.t. yield limit σ y0 at P; lines – analytical, crosses – FD

all the cases, the difference is below 10−5 , which proves that displacements and rotations at the free corner are shown in
the implementation is correct. Fig. 6, where also the results of ABAQUS Explicit are given
for comparison.
The design derivatives computed analytically (shown by
5.2 Bending of an elastic-plastic rectangular plate lines) and by the finite-difference method (shown by crosses)
are presented in Figs. 7, 8, 9, 10. The design derivatives of dis-
An elastic-plastic rectangular sheet is fixed at three corners placements and rotations at the free corner are computed w.r.t.
and loaded by a time-dependent force at the free corner. The selected material constants, and are presented as follows: for
shell is modeled with 9×12 4-node finite elements, cf. Fig. 3. m 1 = E in Fig. 7, for m 2 = ν in Fig. 8, for m 3 = σ y0 in
The explicit dynamic analysis was performed for a sub- Fig. 9, and for m 4 = σ y1 in Fig. 10. For finite-difference
critical time step, t = 0.0001, and the deformed geome- results, only selected data points are displayed. In all the fig-
try at subsequent time instants is shown in Fig. 4. In Fig. 5 ures presented, we see that the analytical and the FD results
the energy balance obtained by our code is presented. The fully coincide.
Analytical DSA for explicit dynamics of elastic-plastic shells 783

5e-05 0.00015

0 1e-04

-5e-05
5e-05
-0.0001
0
-0.00015
-5e-05
-0.0002

-0.00025 -0.0001

-0.0003 -0.00015
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5

Fig. 17 Tube. DSA: design derivatives w.r.t. ‘infinity’ yield limit σ y1 at P; lines – analytical, crosses – FD

5e-05 1e-04

0
5e-05

-5e-05
0
-0.0001
-5e-05
-0.00015

-0.0001
-0.0002

-0.00025 -0.00015
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5

Fig. 18 Tube. DSA: design derivatives w.r.t. exponent coefficient a at P; lines – analytical, crosses – FD

6e-05 1e-04
4e-05
5e-05
2e-05
0 0
-2e-05
-5e-05
-4e-05
-6e-05 -0.0001
-8e-05
-0.00015
-0.0001
-0.00012 -0.0002
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5

Fig. 19 Tube. DSA: design derivatives w.r.t. kinematic hardening modulus Hα at P; lines – analytical, crosses – FD

5.3 Elastic-plastic compression of a rectangular notched Figure 12 shows deformed shapes of the notch vicin-
tube ity at subsequent time instances. Figure 13 presents the dis-
placement of point P located on the top of the first buckling
This is a realistic example of compression of a steel thin- fold.
walled tube with a notch, a quarter of the tube is shown in Design derivatives w.r.t. the selected constitutive param-
Fig. 11. To avoid modeling of the contact and to enable clear eters for two displacement components (u 1 , u 3 ) of point P
conclusions regarding the DSA accuracy, we perform com- are presented in Figs. 14, 15, 16, 17, 18 and 19. The deriv-
putations for the first few milliseconds only. The load is a atives are computed analytically and by the finite-difference
270 kg mass with the initial velocity v3 = −7.7778 m/s, operator of Eq. (119). The results are presented as follows:
hitting the tube end at x3 = 370 mm, and compressing it Fig. 14 presents results for m 1 = E, Fig. 15 for m 2 = ν,
axially. Fig. 16 for m 3 = σ y0 , Fig. 17 for m 4 = σ y1 , Fig. 18 for
784 K. Wisniewski et al.

m 5 = a, and Fig. 19 for m 7 = H . To enable comparisons,


References
selected points of the finite-difference derivatives are shown.
In all the figures presented, we see that the analytical and the 1. ABAQUS Explicit, v. 6.4. (2004) Abaqus Inc. (http://www.aba-
FD results fully coincide. qus.com)
2. Bathe K-J, Dvorkin EN (1986) A formulation of general shell ele-
ments – The use of mixed interpolations of tensorial components.
Int J Numer Methods Eng 22:697–722
3. Belytschko T, Liu JS, Moran B (2000) Nonlinear finite elements
6 Final remarks for continua and structures. Wiley, New York
4. Benson DJ, Hallquist JO (1986) A simple rigid body algorithm for
In the paper, the computation of design derivatives for an structural dynamics programs. Int J Numer Methods Eng 22:723–
explicit dynamics algorithm for finite rotation shells and the 749
plane stress Huber–Mises deviatoric plasticity with nonlinear 5. Cho S, Choi KK (2000) Design sensitivity analysis and optimiza-
tion of nonlinear transient dynamics. Part I – sizing design. Int J
isotropic/kinematic hardening is described from an algorith- Numer Methods Eng 48:351–373
mic point of view. The material parameters are used as the 6. Cho S, Choi KK (2000) Design sensitivity analysis and optimiza-
design variables. The most important results are as follows. tion of nonlinear transient dynamics. Part II – configuration design.
Int J Numer Methods Eng 48:375–399
7. Dienes JK (1979) On the analysis of rotation and stress rate in
1. The complicated dynamic algorithm for finite rotation deforming bodies. Acta Mechanica 32:217–232
elastic-plastic shells, which includes predictors, history 8. Hughes TJR (1984) Numerical implementation of constitutive
state variables, and a Newton loop, has been successfully models: rate independent deviatoric plasticity. In: Nemat-Nasser
design-differentiated and the DSA module implemented S et al (eds) Theoretical foundations for large-scale computing for
in our code. Close proximity of analytical derivatives nonlinear material behaviour. Nijhoff, The Hague, pp 29–57
9. Hughes TJR (1987) The finite element method. Linear static and
and finite-difference approximations is obtained in all dynamic finite element analysis. Prentice-Hall, Englewood Cliffs
verification examples. This allows to draw the conclu- 10. Hughes TJR, Brezzi F (1989) On drilling degrees of freedom. Com-
sion that, despite a substantial complexity of the dynamic put Methods Appl Mech Eng 72:105–121
algorithm, the derived analytical DSA formulas and their 11. Hughes TJR, Liu WK (1981) Nonlinear finite element analysis
of shells. Part I. Three-dimensional shells. Comput Methods Appl
implementation are correct. Mech Eng 26:331–362
2. The most important issue in the design-differentiation of 12. Hughes TJR, Liu WK, Levit I (1981) Nonlinear dynamic finite ele-
a complicated algorithm is a correct grouping of variables ments of shells. In: Wunderlich W, Stein E, Bathe KJ (eds) Non-
having similar functionality, and a good understanding of linear finite element analysis of plates and shells. Springer, Berlin
Heidelberg New York
internal dependencies between them. As a result of this 13. Hughes TJR, Winget J (1980) Finite rotation effects in numerical
type of thinking, we formulate and present the schemes integration of rate constitutive equations arising in large-deforma-
of the differentiation for the state variables, the stress, tion analysis. Int J Numer Methods Eng 15:1862–1867
and the yield function. These schemes account for the 14. Jetteur P (1986) Implicit integration algorithm for plane stress elas-
necessary updates. This form of presentation is certainly toplasticity in plane stress analysis. Eng Comput 3:251–253
15. Key SW, Beisinger ZE (1971) The transient dynamic analysis of
more helpful in implementation than numerous design- thin shells in the finite element method. In: Proceedings of the 3rd
differentiated formulas, often displayed in papers on the conference on matrix methods in structural mechanics, Wrigth-
DSA. Patterson Air Force Base, Ohio
3. Efficiency of the analytical DSA computations is an 16. Kim NH, Choi KK, Chen JS (2001) Structural optimization of finite
deformation elastoplasticity using continuum-based shape design
important issue for explicit dynamics, and that’s why we sensitivity formulation. Comput Struct 79:1959–1976
show how to avoid Newton loops in the DSA algorithm, 17. Kim NH, Bae DS, Choi KK (2001) Configuration design sensi-
although such a loop is used in the constitutive algorithm. tivity analysis of dynamics for constrained mechanical systems.
The suitable scheme explaining our method is presented. Comput Methods Appl Mech Eng 190:5271–5282
18. Kleiber M, Hien TD, Antúnez H, Kowalczyk P (1997) Parameter
Note that it is not trivial due to the history data and the sensitivity in nonlinear mechanics. Wiley, New York
predicted values affecting the differentiation. 19. Korelc J (2002) Multi-language and multi-environment generation
of nonlinear finite element codes. Eng Comput 18:312–327
Finally, we would like to mention that the analytical deriva- 20. Kowalczyk P, Kleiber M (1999) Shape sensitivity in elasto-plastic
tives do not depend on the perturbation, and are more reliable computations. Comput Methods Appl Mech Eng 171:371–386
21. Simo JC, Fox DD, Hughes TJR (1992) Formulations of finite
than the finite-difference ones. Besides, they do not cause elasticity with independent rotations. Int J Numer Methods Eng
additional problems, such as e.g. involved in the identification 95:227–288
of the contact, when the thickness is perturbed. Our experi- 22. Simo JC, Hughes TJR (1998) Computational inelasticity. Springer,
ence indicates that the effects caused by non-differentiability Berlin Heidelberg New York
23. Simo JC, Taylor RL (1985) Consistent tangent operators for a rate-
of the response w.r.t. the design parameters are minor, and independent elastoplasticity. Comput Methods Appl Mech Eng
that not accuracy but efficiency of the analytical DSA is the 48:101–118
issue which should be further addressed. 24. Simo JC, Taylor RL (1986) A return mapping algorithm for plane
stress elastoplasticity. Int J Numer Methods Eng 22:649–670
25. Stillman DW (2000) Design sensitivity analysis for structures using
Acknowledgements This research was partially supported by the EC explicit time integration. In: 8th AIAA/USAF/NASA/ISSMO sym-
Growth Project G3RD-CT-2000-00276, and the Polish Committee for posium on multidisciplinary analysis and optimization, 6–8 Sep-
Scientific Research (KBN) Grant 7T11F01921. tember 2000, Long Beach, CA, AIAA-2000-4906
Analytical DSA for explicit dynamics of elastic-plastic shells 785

26. Wisniewski K, Kowalczyk P, Turska E (2003) On the computation 29. Wisniewski K, Turska E (2002) Second order shell kinemta-
of design derivatives for Huber-Mises plasticity with non-linear ics implied by rotation constraint equation. J Elasticity 67:229–
hardening. Int J Numer Methods Eng 57:271–300 246
27. Wisniewski K, Turska E (2000) Kinematics of finite rotation shells 30. Whirley RG, Hallquist JO, Goudreau GL (1989) An assesment of
with in-plane twist parameter. Comput Methods Appl Mech Eng numerical algorithms for plane stress and shell elastoplasticity on
190(8–10):1117–1135 supercomputers. Eng Comput 6:116–126
28. Wisniewski K, Turska E (2001) Warping and in-plane twist param-
eter in kinematics of finite rotation shells, Comput Methods Appl
Mech Eng 190(43–44):5739–5758

You might also like